{"id":"action","type":"choice","probabilities":{"continue":0.017314544455845064,"human_review":0.006597015089721065,"observe":0.9755867662926122,"stop":0.0005016741618211027},"predicted":"observe","confidence":0.9755867662926122,"entropy_confidence":0.9053013539987245,"candidate_mass":0.9999419959017044,"accepted":true,"accepted_predicted":"observe","input_tokens":210,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":4.05620841069446,"brier_hard":1.9174840153883579,"soft_accuracy":0.09983448338990716,"brier_soft":1.1862659952982337,"tvd":0.8855867662926126,"kl":3.091724337079579,"case_id":"typed:0/r0","base_id":"typed:0","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.436401}
{"id":"needs_review","type":"binary","probabilities":{"false":1.769746807572699e-06,"true":0.9999982302531932},"predicted":"true","confidence":0.9999982302531932,"entropy_confidence":0.9999636304300193,"candidate_mass":0.9999985179044982,"accepted":true,"accepted_predicted":"true","input_tokens":172,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.24467406817977,"brier_hard":1.9999929210190355,"soft_accuracy":0.35666750732623864,"brier_soft":0.8277501436381732,"tvd":0.6433312302531928,"kl":5.557609725081707,"case_id":"typed:0/r0","base_id":"typed:0","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.436401}
{"id":"outcome","type":"choice","probabilities":{"failure":6.121078692242111e-07,"harmful":3.0327093229733745e-07,"partial":3.305481888726736e-06,"success":0.9999957791393091},"predicted":"success","confidence":0.9999957791393091,"entropy_confidence":0.9999572641166583,"candidate_mass":0.9999993758468648,"accepted":true,"accepted_predicted":"success","input_tokens":215,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":4.220869598708699e-06,"brier_hard":2.920852479021559e-11,"soft_accuracy":0.6066655319786489,"brier_soft":0.26037216990652895,"tvd":0.39332877913930947,"kl":3.319536225802795,"case_id":"typed:0/r0","base_id":"typed:0","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.436401}
{"id":"risk","type":"ordinal","probabilities":{"0":0.4034547494756449,"1":0.5953015536649433,"2":0.0009184898789647096,"3":0.00032520698044766173},"predicted":"1","confidence":0.5953015536649433,"entropy_confidence":0.5065823818280296,"candidate_mass":0.9999877142898793,"accepted":false,"accepted_predicted":null,"input_tokens":223,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.5186871889405217,"brier_hard":0.3265575167237021,"score_mae":0.19188584563578437,"within_one":true,"case_id":"typed:0/r0","base_id":"typed:0","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.436401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00025391758070972937,"1":0.9990417446806782,"2":0.0006481576678155918,"3":5.6180070797865394e-05},"predicted":"1","confidence":0.9990417446806782,"entropy_confidence":0.993963725048705,"candidate_mass":0.999996182767254,"accepted":true,"accepted_predicted":"1","input_tokens":211,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.27850082899728,"brier_hard":1.9975770601918945,"score_mae":0.44383960022870306,"within_one":true,"case_id":"typed:0/r0","base_id":"typed:0","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.436401}
{"id":"action","type":"choice","probabilities":{"continue":0.00034803853048056145,"human_review":0.0010177109729965607,"observe":0.9983213798986844,"stop":0.00031287059783946427},"predicted":"observe","confidence":0.9983213798986844,"entropy_confidence":0.9899114357110111,"candidate_mass":0.999970775611403,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0016800305626780488,"brier_hard":4.072519898789849e-06,"soft_accuracy":0.43292618502950303,"brier_soft":0.4667052178142615,"tvd":0.5649883798986839,"kl":3.132980217488241,"case_id":"typed:1/r0","base_id":"typed:1","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.5859}
{"id":"needs_review","type":"binary","probabilities":{"false":3.1275159391199826e-07,"true":0.9999996872484075},"predicted":"true","confidence":0.9999996872484075,"entropy_confidence":0.9999927907085161,"candidate_mass":0.9999994103669179,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.977856591171136,"brier_hard":1.9999987489938227,"soft_accuracy":0.2500001563757973,"brier_soft":1.124999061745416,"tvd":0.7499996872484067,"kl":6.561181766890075,"case_id":"typed:1/r0","base_id":"typed:1","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.5859}
{"id":"outcome","type":"choice","probabilities":{"failure":0.002168770749338104,"harmful":0.0008047097985901336,"partial":0.9794896807718511,"success":0.017536838680222044},"predicted":"partial","confidence":0.9794896807718511,"entropy_confidence":0.9204758876663418,"candidate_mass":0.9999494089708102,"accepted":true,"accepted_predicted":"partial","input_tokens":211,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":false,"nll_hard":4.043451543161375,"brier_hard":1.924639249213418,"soft_accuracy":0.23992969236590578,"brier_soft":0.9507204902424764,"tvd":0.7461564474386171,"kl":2.460396508240304,"case_id":"typed:1/r0","base_id":"typed:1","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.5859}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0025548104544058144,"1":0.5185660496057718,"2":0.4711494997903587,"3":0.007729640149464586},"predicted":"1","confidence":0.5185660496057718,"entropy_confidence":0.4604665631892464,"candidate_mass":0.999970031505137,"accepted":false,"accepted_predicted":null,"input_tokens":219,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":null,"nll_hard":5.969777244220936,"brier_hard":1.4858492524409275,"score_mae":0.9273869696348829,"within_one":true,"case_id":"typed:1/r0","base_id":"typed:1","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.5859}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0022290939954413102,"1":0.014517952162556054,"2":0.9730982657627436,"3":0.010154688079258227},"predicted":"2","confidence":0.9730982657627436,"entropy_confidence":0.8930953439996095,"candidate_mass":0.9999922543940212,"accepted":true,"accepted_predicted":"2","input_tokens":207,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.232369315168583,"brier_hard":1.918203187990369,"score_mae":1.271178547925818,"within_one":false,"case_id":"typed:1/r0","base_id":"typed:1","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.5859}
{"id":"action","type":"choice","probabilities":{"continue":0.0010688932328266811,"human_review":0.003866222063677315,"observe":0.9945917679204606,"stop":0.0004731167830336414},"predicted":"observe","confidence":0.9945917679204606,"entropy_confidence":0.9727280203556891,"candidate_mass":0.9999672852060424,"accepted":true,"accepted_predicted":"observe","input_tokens":223,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.841131527673934,"brier_hard":1.9870913123947733,"soft_accuracy":0.3131081279344905,"brier_soft":0.6821463279250962,"tvd":0.6812584545871482,"kl":3.181245623120774,"case_id":"typed:2/r0","base_id":"typed:2","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.27700000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.8707486732125847e-06,"true":0.999997129251326},"predicted":"true","confidence":0.999997129251326,"entropy_confidence":0.9999430075016765,"candidate_mass":0.9999983460627746,"accepted":true,"accepted_predicted":"true","input_tokens":185,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.760937700464709,"brier_hard":1.999988517021788,"soft_accuracy":0.4066675358711715,"brier_soft":0.7040812845547914,"tvd":0.5933301292513264,"kl":5.098897103277167,"case_id":"typed:2/r0","base_id":"typed:2","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.27700000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":3.658385888158909e-05,"harmful":1.553050409314514e-05,"partial":0.0039710989083502814,"success":0.9959767867286735},"predicted":"success","confidence":0.9959767867286735,"entropy_confidence":0.9808728393319162,"candidate_mass":0.9999915474558141,"accepted":true,"accepted_predicted":"success","input_tokens":228,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":5.528712419475049,"brier_hard":1.9840433330917884,"soft_accuracy":0.3967449450267848,"brier_soft":0.5602953039222313,"tvd":0.5993101833952775,"kl":3.112858381298493,"case_id":"typed:2/r0","base_id":"typed:2","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.27700000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.8606158621573164,"1":0.13768080318999396,"2":0.0010240695263903604,"3":0.0006792651262985707},"predicted":"0","confidence":0.8606158621573164,"entropy_confidence":0.701228727899243,"candidate_mass":0.9999846370720076,"accepted":true,"accepted_predicted":"0","input_tokens":236,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":1.9828172933509975,"brier_hard":1.4842555695033417,"score_mae":0.5582332623783296,"within_one":true,"case_id":"typed:2/r0","base_id":"typed:2","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.27700000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.353889055246798e-05,"1":0.9952866640365026,"2":0.004609953270424179,"3":3.98438025210812e-05},"predicted":"1","confidence":0.9952866640365026,"entropy_confidence":0.9779849802878318,"candidate_mass":0.9999962563330038,"accepted":true,"accepted_predicted":"1","input_tokens":224,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0047244787583941876,"brier_hard":4.347282977950484e-05,"score_mae":0.31204089801508594,"within_one":true,"case_id":"typed:2/r0","base_id":"typed:2","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.27700000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.001976336932468186,"human_review":0.01714853104526837,"observe":0.980537480782001,"stop":0.0003376512402617138},"predicted":"observe","confidence":0.980537480782001,"entropy_confidence":0.9249799944633177,"candidate_mass":0.9999548827037461,"accepted":true,"accepted_predicted":"observe","input_tokens":225,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.065842762375766,"brier_hard":1.9274547811608174,"soft_accuracy":0.28421074243278505,"brier_soft":0.6983303350524324,"tvd":0.697204197448718,"kl":2.596512909091028,"case_id":"typed:3/r0","base_id":"typed:3","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.085101}
{"id":"needs_review","type":"binary","probabilities":{"false":6.494432018045544e-05,"true":0.9999350556798181},"predicted":"true","confidence":0.9999350556798181,"entropy_confidence":0.9990029043119243,"candidate_mass":0.9999907708349616,"accepted":true,"accepted_predicted":"true","input_tokens":187,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.641980267779253,"brier_hard":1.9997402311548047,"soft_accuracy":0.4000129888640355,"brier_soft":0.7198441420670947,"tvd":0.5999350556798189,"kl":5.112202472229957,"case_id":"typed:3/r0","base_id":"typed:3","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.085101}
{"id":"outcome","type":"choice","probabilities":{"failure":1.46525613857006e-05,"harmful":7.610142609280358e-06,"partial":0.0002988585154615177,"success":0.9996788787805423},"predicted":"success","confidence":0.9996788787805423,"entropy_confidence":0.9978364930649025,"candidate_mass":0.9999957033917366,"accepted":true,"accepted_predicted":"success","input_tokens":230,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":8.115540289000855,"brier_hard":1.9987602332380232,"soft_accuracy":0.3500297612780612,"brier_soft":0.6731875054908238,"tvd":0.6496864889231522,"kl":4.752823560029488,"case_id":"typed:3/r0","base_id":"typed:3","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.085101}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9618442598011985,"1":0.03642039584628137,"2":0.00099036683487705,"3":0.0007449775176429013},"predicted":"0","confidence":0.9618442598011985,"entropy_confidence":0.8771676628159981,"candidate_mass":0.9999881240704462,"accepted":true,"accepted_predicted":"0","input_tokens":238,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":3.312626335815839,"brier_hard":1.8536315694715217,"score_mae":0.6993639379310358,"within_one":true,"case_id":"typed:3/r0","base_id":"typed:3","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.085101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003993269138401608,"1":0.9986268790842899,"2":0.000896283583562891,"3":7.75104183077325e-05},"predicted":"1","confidence":0.9986268790842899,"entropy_confidence":0.9916898634567228,"candidate_mass":0.9999961201809296,"accepted":true,"accepted_predicted":"1","input_tokens":226,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0013740645101130346,"brier_hard":2.8542551603881426e-06,"score_mae":0.01731897750633893,"within_one":true,"case_id":"typed:3/r0","base_id":"typed:3","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.085101}
{"id":"action","type":"choice","probabilities":{"continue":0.00010888391292123352,"human_review":0.0004793885934886023,"observe":0.9992675123445913,"stop":0.00014421514899966664},"predicted":"observe","confidence":0.9992675123445913,"entropy_confidence":0.9951920422815461,"candidate_mass":0.9999724103391147,"accepted":true,"accepted_predicted":"observe","input_tokens":215,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":9.125228262342187,"brier_hard":1.9983180558686449,"soft_accuracy":0.36654883455779314,"brier_soft":0.5948827101349009,"tvd":0.6326005123445909,"kl":4.268880997762998,"case_id":"typed:4/r0","base_id":"typed:4","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.7524}
{"id":"needs_review","type":"binary","probabilities":{"false":3.68407295013162e-08,"true":0.9999999631592693},"predicted":"true","confidence":0.9999999631592693,"entropy_confidence":0.999999037100551,"candidate_mass":0.9999997507086382,"accepted":true,"accepted_predicted":"true","input_tokens":177,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.116661823873812,"brier_hard":1.9999998526370824,"soft_accuracy":0.30000001473629145,"brier_soft":0.9799998968459585,"tvd":0.6999999631592699,"kl":6.086046430137767,"case_id":"typed:4/r0","base_id":"typed:4","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.7524}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00028577600931189263,"harmful":0.00033256263920536155,"partial":0.9993001719985319,"success":8.148935295235513e-05},"predicted":"partial","confidence":0.9993001719985319,"entropy_confidence":0.9953384968190714,"candidate_mass":0.9999927520083618,"accepted":true,"accepted_predicted":"partial","input_tokens":220,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":false,"nll_hard":9.415038184872834,"brier_hard":1.998438053956742,"soft_accuracy":0.2666155562321221,"brier_soft":0.7937030301988542,"tvd":0.7326334386652645,"kl":5.2859270749705844,"case_id":"typed:4/r0","base_id":"typed:4","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.7524}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0003741405393499681,"1":0.003312484771261301,"2":0.994880096341246,"3":0.0014332783481425789},"predicted":"2","confidence":0.994880096341246,"entropy_confidence":0.9737729907583378,"candidate_mass":0.999991197340838,"accepted":true,"accepted_predicted":"2","input_tokens":228,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.890879057471967,"brier_hard":1.9890512918405934,"score_mae":1.6973725124981809,"within_one":false,"case_id":"typed:4/r0","base_id":"typed:4","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.7524}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00038474261120076755,"1":0.000947505779814421,"2":0.9912324626693132,"3":0.00743528893967158},"predicted":"2","confidence":0.9912324626693132,"entropy_confidence":0.9604740315117702,"candidate_mass":0.9999912733538417,"accepted":true,"accepted_predicted":"2","input_tokens":216,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.862935989586237,"brier_hard":1.9818286391427657,"score_mae":1.7057182979374554,"within_one":false,"case_id":"typed:4/r0","base_id":"typed:4","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.7524}
{"id":"action","type":"choice","probabilities":{"continue":0.0009643072791433988,"human_review":0.004121546658154726,"observe":0.9943785112464475,"stop":0.0005356348162553233},"predicted":"observe","confidence":0.9943785112464475,"entropy_confidence":0.9718891437205335,"candidate_mass":0.9999517283614209,"accepted":true,"accepted_predicted":"observe","input_tokens":226,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.944100559846955,"brier_hard":1.9868782130104548,"soft_accuracy":0.31580890523602884,"brier_soft":0.8244113948746841,"tvd":0.6782471460627023,"kl":3.7481636483194976,"case_id":"typed:5/r0","base_id":"typed:5","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.4556}
{"id":"needs_review","type":"binary","probabilities":{"false":5.483128676233606e-08,"true":0.9999999451687118},"predicted":"true","confidence":0.9999999451687118,"entropy_confidence":0.9999985983412112,"candidate_mass":0.9999996410882805,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.719004879588866,"brier_hard":1.9999997806748562,"soft_accuracy":0.1500000383819005,"brier_soft":1.4449998135736286,"tvd":0.8499999451687126,"kl":7.5442213266715665,"case_id":"typed:5/r0","base_id":"typed:5","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.4556}
{"id":"outcome","type":"choice","probabilities":{"failure":4.183701966354237e-05,"harmful":2.4310612205073697e-05,"partial":0.0034762227031038487,"success":0.9964576296650284},"predicted":"success","confidence":0.9964576296650284,"entropy_confidence":0.9827613251859998,"candidate_mass":0.9999898290661665,"accepted":true,"accepted_predicted":"success","input_tokens":231,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.003548659385256046,"brier_hard":2.463485321374171e-05,"soft_accuracy":0.7976896289884715,"brier_soft":0.06256063620632756,"tvd":0.196481940277233,"kl":0.7433272743976718,"case_id":"typed:5/r0","base_id":"typed:5","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.4556}
{"id":"risk","type":"ordinal","probabilities":{"0":0.19219702091251428,"1":0.755436260270668,"2":0.04644855909007902,"3":0.005918159726740246},"predicted":"1","confidence":0.755436260270668,"entropy_confidence":0.4937758431756897,"candidate_mass":0.9999625515350183,"accepted":false,"accepted_predicted":null,"input_tokens":239,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.2804598684384593,"brier_hard":0.09894361089413767,"score_mae":0.10057914236895327,"within_one":true,"case_id":"typed:5/r0","base_id":"typed:5","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.4556}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006011481034522109,"1":0.03293734166462186,"2":0.9654183224518671,"3":0.0010431877800588173},"predicted":"2","confidence":0.9654183224518671,"entropy_confidence":0.8860146538231,"candidate_mass":0.9999907516653067,"accepted":true,"accepted_predicted":"2","input_tokens":227,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.416669225414488,"brier_hard":1.9319165592145915,"score_mae":1.0502365499085324,"within_one":false,"case_id":"typed:5/r0","base_id":"typed:5","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.4556}
{"id":"action","type":"choice","probabilities":{"continue":0.0010269729007439212,"human_review":0.005108245307866117,"observe":0.9922179581572087,"stop":0.0016468236341803276},"predicted":"observe","confidence":0.9922179581572087,"entropy_confidence":0.9622529933307581,"candidate_mass":0.9999497028225317,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.276899317715536,"brier_hard":1.9743098467454743,"soft_accuracy":0.14216667436556182,"brier_soft":1.1046150664080832,"tvd":0.8522179581572092,"kl":3.8930366197393145,"case_id":"typed:6/r0","base_id":"typed:6","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.326301}
{"id":"needs_review","type":"binary","probabilities":{"false":8.62984396985297e-08,"true":0.9999999137015616},"predicted":"true","confidence":0.9999999137015616,"entropy_confidence":0.999997850410699,"candidate_mass":0.9999997164400309,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.629844211907068e-08,"brier_hard":1.489484116387855e-14,"soft_accuracy":0.8999999309612494,"brier_soft":0.019999965480639266,"tvd":0.09999991370156093,"kl":0.8262096507741723,"case_id":"typed:6/r0","base_id":"typed:6","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.326301}
{"id":"outcome","type":"choice","probabilities":{"failure":1.5050638737046617e-05,"harmful":0.9998956652110331,"partial":8.203366073564754e-05,"success":7.250489495694716e-06},"predicted":"harmful","confidence":0.9998956652110331,"entropy_confidence":0.9991855536258638,"candidate_mass":0.9999980669214562,"accepted":true,"accepted_predicted":"harmful","input_tokens":214,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":9.408380898154583,"brier_hard":1.999627280994956,"soft_accuracy":0.3333338756500989,"brier_soft":0.6481235936829078,"tvd":0.6665623318776991,"kl":5.463832323966231,"case_id":"typed:6/r0","base_id":"typed:6","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.326301}
{"id":"risk","type":"ordinal","probabilities":{"0":1.3884995593741512e-05,"1":1.4361633220215913e-05,"2":0.03032879293414644,"3":0.9696429604370387},"predicted":"3","confidence":0.9696429604370387,"entropy_confidence":0.9017336725726508,"candidate_mass":0.9999975861695526,"accepted":true,"accepted_predicted":"3","input_tokens":222,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.495657755826766,"brier_hard":1.8804697209367036,"score_mae":0.6262678288126291,"within_one":true,"case_id":"typed:6/r0","base_id":"typed:6","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.326301}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005342322705373116,"1":0.0005698034417589252,"2":0.8877858598293134,"3":0.11111010445839213},"predicted":"2","confidence":0.8877858598293134,"entropy_confidence":0.7416961702694278,"candidate_mass":0.9999899213154471,"accepted":true,"accepted_predicted":"2","input_tokens":210,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.197233637251731,"brier_hard":1.578289589388925,"score_mae":0.12386116352443821,"within_one":true,"case_id":"typed:6/r0","base_id":"typed:6","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.326301}
{"id":"action","type":"choice","probabilities":{"continue":0.00019772149969669688,"human_review":0.0005212224456118351,"observe":0.9991345912884549,"stop":0.00014646476623708437},"predicted":"observe","confidence":0.9991345912884549,"entropy_confidence":0.9943846533245947,"candidate_mass":0.999972005563334,"accepted":true,"accepted_predicted":"observe","input_tokens":219,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0008657833938485491,"brier_hard":1.0811507950188529e-06,"soft_accuracy":0.4498299753990352,"brier_soft":0.4486103129296344,"tvd":0.5491345912884547,"kl":3.2057695069655905,"case_id":"typed:7/r0","base_id":"typed:7","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7451}
{"id":"needs_review","type":"binary","probabilities":{"false":1.0148987344575868e-08,"true":0.9999999898510126},"predicted":"true","confidence":0.9999999898510126,"entropy_confidence":0.9999997158609932,"candidate_mass":0.9999999124352751,"accepted":true,"accepted_predicted":"true","input_tokens":181,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.014898747973886e-08,"brier_hard":2.060038899388918e-16,"soft_accuracy":0.5166669996616936,"brier_soft":0.4672215581566382,"tvd":0.48333298985101264,"kl":4.110476926635574,"case_id":"typed:7/r0","base_id":"typed:7","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7451}
{"id":"outcome","type":"choice","probabilities":{"failure":5.628326425194341e-05,"harmful":2.5176282305767183e-05,"partial":0.01719312359914936,"success":0.9827254168542914},"predicted":"success","confidence":0.9827254168542914,"entropy_confidence":0.9366644306624083,"candidate_mass":0.9999849981424841,"accepted":true,"accepted_predicted":"success","input_tokens":224,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":4.063245765847487,"brier_hard":1.9316586050338886,"soft_accuracy":0.2879038280060927,"brier_soft":0.7976593339980019,"tvd":0.6993924168542921,"kl":2.856037271934938,"case_id":"typed:7/r0","base_id":"typed:7","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7451}
{"id":"risk","type":"ordinal","probabilities":{"0":0.8288968985669778,"1":0.16842053473077784,"2":0.0017590882562204498,"3":0.0009234784460254599},"predicted":"0","confidence":0.8288968985669778,"entropy_confidence":0.6586822252596438,"candidate_mass":0.9999838944578026,"accepted":true,"accepted_predicted":"0","input_tokens":232,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":1.7812912444014941,"brier_hard":1.3785984227153334,"score_mae":0.6252908534187049,"within_one":true,"case_id":"typed:7/r0","base_id":"typed:7","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7451}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00010763636989483928,"1":0.988929663328869,"2":0.010901166893136163,"3":6.153340809844636e-05},"predicted":"1","confidence":0.988929663328869,"entropy_confidence":0.955384635711767,"candidate_mass":0.9999942370269294,"accepted":true,"accepted_predicted":"1","input_tokens":220,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.011132068868409508,"brier_hard":0.00024140316559263203,"score_mae":0.3724164026605632,"within_one":true,"case_id":"typed:7/r0","base_id":"typed:7","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7451}
{"id":"action","type":"choice","probabilities":{"continue":0.001206337975579235,"human_review":0.01534425534871634,"observe":0.9816662350429526,"stop":0.0017831716327523055},"predicted":"observe","confidence":0.9816662350429526,"entropy_confidence":0.9266743278991744,"candidate_mass":0.9999493037441564,"accepted":true,"accepted_predicted":"observe","input_tokens":217,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.1770141193791375,"brier_hard":1.9332201674505627,"soft_accuracy":0.18863199192438243,"brier_soft":0.9616448109660888,"tvd":0.7983330517097691,"kl":2.908365976573823,"case_id":"typed:8/r0","base_id":"typed:8","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":197.4778}
{"id":"needs_review","type":"binary","probabilities":{"false":3.841768759126736e-08,"true":0.9999999615823122},"predicted":"true","confidence":0.9999999615823122,"entropy_confidence":0.9999989982069669,"candidate_mass":0.9999997276618389,"accepted":true,"accepted_predicted":"true","input_tokens":179,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.841768854672659e-08,"brier_hard":2.9518374564321177e-15,"soft_accuracy":0.7333329820717712,"brier_soft":0.14222253679908484,"tvd":0.2666669615823123,"kl":2.2286466867756345,"case_id":"typed:8/r0","base_id":"typed:8","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":197.4778}
{"id":"outcome","type":"choice","probabilities":{"failure":1.6982615133763348e-05,"harmful":0.9998758898965492,"partial":9.994219724402306e-05,"success":7.1852910714391594e-06},"predicted":"harmful","confidence":0.9998758898965492,"entropy_confidence":0.9990505002449677,"candidate_mass":0.999997836377481,"accepted":true,"accepted_predicted":"harmful","input_tokens":222,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00012411780574699848,"brier_hard":2.5731798193097813e-08,"soft_accuracy":0.4666370066413995,"brier_soft":0.40370017002009767,"tvd":0.53320888989655,"kl":4.199009241813773,"case_id":"typed:8/r0","base_id":"typed:8","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":197.4778}
{"id":"risk","type":"ordinal","probabilities":{"0":2.4816849830333326e-06,"1":1.366908748359131e-06,"2":0.0002168837415044319,"3":0.9997792676647657},"predicted":"3","confidence":0.9997792676647657,"entropy_confidence":0.9984845506320021,"candidate_mass":0.9999993687175713,"accepted":true,"accepted_predicted":"3","input_tokens":230,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00022075670020174062,"brier_hard":9.576934834684833e-08,"score_mae":0.5597729373860543,"within_one":true,"case_id":"typed:8/r0","base_id":"typed:8","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":197.4778}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010041625077904781,"1":0.001752020508110862,"2":0.7418462167737407,"3":0.2553976002103565},"predicted":"2","confidence":0.7418462167737407,"entropy_confidence":0.5757190113236523,"candidate_mass":0.9999829522129268,"accepted":false,"accepted_predicted":null,"input_tokens":218,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.3649337316284829,"brier_hard":1.1047726210322109,"score_mae":0.051637254686661915,"within_one":true,"case_id":"typed:8/r0","base_id":"typed:8","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":197.4778}
{"id":"action","type":"choice","probabilities":{"continue":0.016585103488912106,"human_review":0.033766239966399395,"observe":0.948918288184676,"stop":0.0007303683600133906},"predicted":"observe","confidence":0.948918288184676,"entropy_confidence":0.8287338010077189,"candidate_mass":0.9999482615272793,"accepted":true,"accepted_predicted":"observe","input_tokens":212,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.3882937932059747,"brier_hard":1.8343291957756849,"soft_accuracy":0.09928087724492334,"brier_soft":1.151077476774637,"tvd":0.8655852881846756,"kl":3.3822852878524468,"case_id":"typed:9/r0","base_id":"typed:9","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.76350100000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.762193314113409e-07,"true":0.9999997237806673},"predicted":"true","confidence":0.9999997237806673,"entropy_confidence":0.9999935833207354,"candidate_mass":0.9999994311176029,"accepted":true,"accepted_predicted":"true","input_tokens":174,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.762193708884801e-07,"brier_hard":1.5259423882461104e-13,"soft_accuracy":0.8933327827076423,"brier_soft":0.022755579924202598,"tvd":0.10666672378066788,"kl":0.8816753527088484,"case_id":"typed:9/r0","base_id":"typed:9","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.76350100000002}
{"id":"outcome","type":"choice","probabilities":{"failure":1.9916583227496897e-05,"harmful":0.9998641854640831,"partial":0.00010717506638697696,"success":8.722886303729846e-06},"predicted":"harmful","confidence":0.9998641854640831,"entropy_confidence":0.9989665314149989,"candidate_mass":0.9999979079494102,"accepted":true,"accepted_predicted":"harmful","input_tokens":217,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00013582375954616173,"brier_hard":3.040484205431064e-08,"soft_accuracy":0.6665958114381951,"brier_soft":0.1570927562346178,"tvd":0.33319718546408245,"kl":2.430337745934734,"case_id":"typed:9/r0","base_id":"typed:9","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.76350100000002}
{"id":"risk","type":"ordinal","probabilities":{"0":2.806356399307499e-06,"1":2.2967415540263073e-06,"2":0.0006697461851464669,"3":0.9993251507169015},"predicted":"3","confidence":0.9993251507169015,"entropy_confidence":0.9959350388672612,"candidate_mass":0.9999991819145684,"accepted":true,"accepted_predicted":"3","input_tokens":225,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0006750770963747659,"brier_hard":9.039946580747655e-07,"score_mae":0.4159842412625516,"within_one":true,"case_id":"typed:9/r0","base_id":"typed:9","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.76350100000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009732727389378194,"1":0.0017667816332830122,"2":0.9192217917790919,"3":0.07803815384868742},"predicted":"2","confidence":0.9192217917790919,"entropy_confidence":0.7876257144649287,"candidate_mass":0.9999854472863343,"accepted":true,"accepted_predicted":"2","input_tokens":213,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.5505574199820913,"brier_hard":1.6949864170174649,"score_mae":0.14234217326247123,"within_one":true,"case_id":"typed:9/r0","base_id":"typed:9","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.76350100000002}
{"id":"action","type":"choice","probabilities":{"continue":0.0018172262143702933,"human_review":0.002661601136976394,"observe":0.9950359688303145,"stop":0.00048520381833885845},"predicted":"observe","confidence":0.9950359688303145,"entropy_confidence":0.9741022318792399,"candidate_mass":0.9999489426290705,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.310443998489539,"brier_hard":1.986472748691814,"soft_accuracy":0.13460594143291094,"brier_soft":1.4337622740327325,"tvd":0.8618541726486534,"kl":4.920319106829,"case_id":"typed:10/r0","base_id":"typed:10","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.171}
{"id":"needs_review","type":"binary","probabilities":{"false":4.067016924604815e-06,"true":0.9999959329830755},"predicted":"true","confidence":0.9999959329830755,"entropy_confidence":0.9999213020443676,"candidate_mass":0.9999973302130029,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.412600769600878,"brier_hard":1.9999837319653833,"soft_accuracy":0.1000032536135397,"brier_soft":1.619985358772153,"tvd":0.8999959329830756,"kl":8.05904823218168,"case_id":"typed:10/r0","base_id":"typed:10","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.171}
{"id":"outcome","type":"choice","probabilities":{"failure":2.632141245364154e-05,"harmful":1.5966311929196244e-05,"partial":0.00036095482998960436,"success":0.9995967574456263},"predicted":"success","confidence":0.9995967574456263,"entropy_confidence":0.9973178324314564,"candidate_mass":0.9999939662194698,"accepted":true,"accepted_predicted":"success","input_tokens":211,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00040332387851553964,"brier_hard":2.938406868208575e-07,"soft_accuracy":0.8993292773260848,"brier_soft":0.01553605385776962,"tvd":0.09994572375755612,"kl":0.4933564085296596,"case_id":"typed:10/r0","base_id":"typed:10","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.171}
{"id":"risk","type":"ordinal","probabilities":{"0":0.020850296118708974,"1":0.9769942885499094,"2":0.0012804343909297003,"3":0.0008749809404510016},"predicted":"1","confidence":0.9769942885499094,"entropy_confidence":0.9147892658224144,"candidate_mass":0.9999849773747366,"accepted":true,"accepted_predicted":"1","input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.023274472862368566,"brier_hard":0.0009664027114383065,"score_mae":0.2985131001531218,"within_one":true,"case_id":"typed:10/r0","base_id":"typed:10","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.171}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0034703993733332912,"1":0.3065676434581033,"2":0.6729046663733919,"3":0.017057290795170457},"predicted":"2","confidence":0.6729046663733919,"entropy_confidence":0.4819785774690598,"candidate_mass":0.9999855670022679,"accepted":false,"accepted_predicted":null,"input_tokens":207,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.182316851518216,"brier_hard":0.9339521179674154,"score_mae":1.0698818485903985,"within_one":false,"case_id":"typed:10/r0","base_id":"typed:10","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.171}
{"id":"action","type":"choice","probabilities":{"continue":0.006485598036470688,"human_review":0.002893482655554516,"observe":0.9904070895136661,"stop":0.0002138297943074658},"predicted":"observe","confidence":0.9904070895136661,"entropy_confidence":0.9560392825922589,"candidate_mass":0.9999731553121638,"accepted":true,"accepted_predicted":"observe","input_tokens":229,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":5.038171247042968,"brier_hard":1.9679854878329393,"soft_accuracy":0.18342991664150327,"brier_soft":1.2560523884008743,"tvd":0.8104070895136668,"kl":3.5334465877650505,"case_id":"typed:11/r0","base_id":"typed:11","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.13569999999999}
{"id":"needs_review","type":"binary","probabilities":{"false":2.16146773547149e-08,"true":0.9999999783853222},"predicted":"true","confidence":0.9999999783853222,"entropy_confidence":0.9999994184331455,"candidate_mass":0.9999998757096022,"accepted":true,"accepted_predicted":"true","input_tokens":191,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.649893245854027,"brier_hard":1.9999999135412905,"soft_accuracy":0.0633330188768326,"brier_soft":1.7546900567949801,"tvd":0.9366669783853224,"kl":8.45226405163443,"case_id":"typed:11/r0","base_id":"typed:11","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.13569999999999}
{"id":"outcome","type":"choice","probabilities":{"failure":7.556634604868704e-07,"harmful":4.5435672853229577e-07,"partial":6.867745252333843e-06,"success":0.9999919222345576},"predicted":"success","confidence":0.9999919222345576,"entropy_confidence":0.9999228062181966,"candidate_mass":0.9999994114006894,"accepted":true,"accepted_predicted":"success","input_tokens":234,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":8.077798067749853e-06,"brier_hard":1.1319368669609766e-10,"soft_accuracy":0.8566600427665698,"brier_soft":0.03459624985043961,"tvd":0.14332577890070142,"kl":1.1876282771010507,"case_id":"typed:11/r0","base_id":"typed:11","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.13569999999999}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9993087876517697,"1":0.0005860500686000456,"2":6.535638352269868e-05,"3":3.9805896107748304e-05},"predicted":"0","confidence":0.9993087876517697,"entropy_confidence":0.9956102652030604,"candidate_mass":0.9999978127884798,"accepted":true,"accepted_predicted":"0","input_tokens":242,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0006914513456238155,"brier_hard":8.270851594843122e-07,"score_mae":0.41583081947603134,"within_one":true,"case_id":"typed:11/r0","base_id":"typed:11","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.13569999999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012800111326060628,"1":0.9994639946610929,"2":0.0003852697854492889,"3":2.2734440198939338e-05},"predicted":"1","confidence":0.9994639946610929,"entropy_confidence":0.9964256593083729,"candidate_mass":0.9999971116020069,"accepted":true,"accepted_predicted":"1","input_tokens":230,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.963471596733992,"brier_hard":1.9986724397313353,"score_mae":0.5436357375525884,"within_one":true,"case_id":"typed:11/r0","base_id":"typed:11","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.13569999999999}
{"id":"action","type":"choice","probabilities":{"continue":0.0023257194073975566,"human_review":0.033477028144432375,"observe":0.9630825300853323,"stop":0.0011147223628375955},"predicted":"observe","confidence":0.9630825300853323,"entropy_confidence":0.876197186842407,"candidate_mass":0.999954964826691,"accepted":true,"accepted_predicted":"observe","input_tokens":216,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.03761616983294405,"brier_hard":0.002490262574991669,"soft_accuracy":0.34735624705083695,"brier_soft":0.5394984064219822,"tvd":0.6130825300853324,"kl":1.9970761117927984,"case_id":"typed:12/r0","base_id":"typed:12","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.502601}
{"id":"needs_review","type":"binary","probabilities":{"false":1.4980952110511748e-08,"true":0.9999999850190485},"predicted":"true","confidence":0.9999999850190485,"entropy_confidence":0.9999995889977225,"candidate_mass":0.9999999058906619,"accepted":true,"accepted_predicted":"true","input_tokens":178,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.016486302097956,"brier_hard":1.9999999400761932,"soft_accuracy":0.38333300349556565,"brier_soft":0.760556340824966,"tvd":0.6166669850190483,"kl":5.312153658779427,"case_id":"typed:12/r0","base_id":"typed:12","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.502601}
{"id":"outcome","type":"choice","probabilities":{"failure":7.650972609659451e-05,"harmful":5.080224961592198e-05,"partial":0.0018872167289989234,"success":0.9979854712952877},"predicted":"success","confidence":0.9979854712952877,"entropy_confidence":0.9891236451845474,"candidate_mass":0.9999823613100505,"accepted":true,"accepted_predicted":"success","input_tokens":221,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0020165605969921824,"brier_hard":7.628347491076643e-06,"soft_accuracy":0.4996342015584689,"brier_soft":0.376710594488602,"tvd":0.49798497129478814,"kl":2.590415238281699,"case_id":"typed:12/r0","base_id":"typed:12","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.502601}
{"id":"risk","type":"ordinal","probabilities":{"0":8.776924114813898e-06,"1":1.2084371946240905e-05,"2":0.9998850051862579,"3":9.413351768061519e-05},"predicted":"2","confidence":0.9998850051862579,"entropy_confidence":0.9991151139484671,"candidate_mass":0.9999979352313524,"accepted":true,"accepted_predicted":"2","input_tokens":229,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00011500142615263589,"brier_hard":2.2307992780761243e-08,"score_mae":0.500064495297504,"within_one":true,"case_id":"typed:12/r0","base_id":"typed:12","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.502601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007971752809474101,"1":0.7088125184234453,"2":0.28953822733886087,"3":0.0008520789567474348},"predicted":"1","confidence":0.7088125184234453,"entropy_confidence":0.5567095838041431,"candidate_mass":0.9999768461440961,"accepted":false,"accepted_predicted":null,"input_tokens":217,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.2394679446108903,"brier_hard":1.0071724782135723,"score_mae":0.10711220997140924,"within_one":true,"case_id":"typed:12/r0","base_id":"typed:12","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.502601}
{"id":"action","type":"choice","probabilities":{"continue":0.0005879190135227274,"human_review":0.0027588144197394,"observe":0.9964791482392976,"stop":0.00017411832744011323},"predicted":"observe","confidence":0.9964791482392976,"entropy_confidence":0.9814953976590979,"candidate_mass":0.9999653037467693,"accepted":true,"accepted_predicted":"observe","input_tokens":223,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":7.4389213516557895,"brier_hard":1.9918028418716316,"soft_accuracy":0.3826870181766905,"brier_soft":0.5648267213232963,"tvd":0.6131461482392977,"kl":3.2093812974905123,"case_id":"typed:13/r0","base_id":"typed:13","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.6395}
{"id":"needs_review","type":"binary","probabilities":{"false":9.725583284405573e-06,"true":0.9999902744167138},"predicted":"true","confidence":0.9999902744167138,"entropy_confidence":0.9998240401585162,"candidate_mass":0.99999627661898,"accepted":true,"accepted_predicted":"true","input_tokens":185,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.540750692427455,"brier_hard":1.999961097856033,"soft_accuracy":0.233338186984235,"brier_soft":1.1755267528321316,"tvd":0.7666572744167147,"kl":6.721699422093478,"case_id":"typed:13/r0","base_id":"typed:13","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.6395}
{"id":"outcome","type":"choice","probabilities":{"failure":6.789299264022866e-05,"harmful":3.3442187576713835e-05,"partial":0.011950153314015643,"success":0.9879485115057691},"predicted":"success","confidence":0.9879485115057691,"entropy_confidence":0.9524788579910066,"candidate_mass":0.9999828021684634,"accepted":true,"accepted_predicted":"success","input_tokens":228,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":4.427011171062176,"brier_hard":1.9522847666505003,"soft_accuracy":0.3670430787027901,"brier_soft":0.5837652992068432,"tvd":0.6212818781724015,"kl":2.8924866717339675,"case_id":"typed:13/r0","base_id":"typed:13","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.6395}
{"id":"risk","type":"ordinal","probabilities":{"0":0.5472523811691445,"1":0.45021334265636287,"2":0.0017210389057219857,"3":0.0008132372687698842},"predicted":"0","confidence":0.5472523811691445,"entropy_confidence":0.49077642253774334,"candidate_mass":0.9999788590237956,"accepted":false,"accepted_predicted":null,"input_tokens":236,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":null,"nll_hard":0.6028451914344153,"brier_hard":0.4076760835925955,"score_mae":0.0972378677258835,"within_one":true,"case_id":"typed:13/r0","base_id":"typed:13","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.6395}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.597081743972425e-05,"1":0.9953596074254722,"2":0.004537603356935292,"3":3.6818400154544935e-05},"predicted":"1","confidence":0.9953596074254722,"entropy_confidence":0.9782711741089091,"candidate_mass":0.9999954859803449,"accepted":true,"accepted_predicted":"1","input_tokens":224,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":9.626298073723586,"brier_hard":1.990629402011279,"score_mae":0.3478782693398065,"within_one":true,"case_id":"typed:13/r0","base_id":"typed:13","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.6395}
{"id":"action","type":"choice","probabilities":{"continue":0.048003909970179316,"human_review":0.28415254523528527,"observe":0.10258457716526853,"stop":0.5652589676292673},"predicted":"stop","confidence":0.5652589676292673,"entropy_confidence":0.23583916054036325,"candidate_mass":0.9998513408890759,"accepted":false,"accepted_predicted":null,"input_tokens":224,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":null,"nll_hard":3.0364728136796764,"brier_hard":1.3170805203532394,"soft_accuracy":0.09741397466581715,"brier_soft":0.6280608772295263,"tvd":0.7327453961984357,"kl":1.5556980618890668,"case_id":"typed:14/r0","base_id":"typed:14","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":226.12460000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.0900276434945955e-05,"true":0.9999790997235644},"predicted":"true","confidence":0.9999790997235644,"entropy_confidence":0.9996449294168321,"candidate_mass":0.9999936071366263,"accepted":true,"accepted_predicted":"true","input_tokens":186,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.775748172528541,"brier_hard":1.999916399767902,"soft_accuracy":0.17668051549816094,"brier_soft":1.3556856271024498,"tvd":0.8233120997235647,"kl":7.276930618283064,"case_id":"typed:14/r0","base_id":"typed:14","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":226.12460000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":9.601462816259184e-05,"harmful":4.0148026869028245e-05,"partial":0.002782641418967056,"success":0.9970811959260024},"predicted":"success","confidence":0.9970811959260024,"entropy_confidence":0.9851523327134009,"candidate_mass":0.9999878174050061,"accepted":true,"accepted_predicted":"success","input_tokens":229,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0029230720896329314,"brier_hard":1.6273341161818554e-05,"soft_accuracy":0.6654693829914908,"brier_soft":0.1828182547661848,"tvd":0.33041419592600185,"kl":1.374198803538582,"case_id":"typed:14/r0","base_id":"typed:14","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":226.12460000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.2608339858470319,"1":0.7231906479132428,"2":0.007190937765733253,"3":0.008784428473990406},"predicted":"1","confidence":0.7231906479132428,"entropy_confidence":0.5224825712737955,"candidate_mass":0.9999496362871345,"accepted":false,"accepted_predicted":null,"input_tokens":237,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.32408240152367546,"brier_hard":0.14478666134510523,"score_mae":0.2560741911333194,"within_one":true,"case_id":"typed:14/r0","base_id":"typed:14","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":226.12460000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0021249055822609073,"1":0.8378372985269593,"2":0.1555194036436204,"3":0.004518392247158333},"predicted":"1","confidence":0.8378372985269593,"entropy_confidence":0.6572638712389645,"candidate_mass":0.9999864238403008,"accepted":true,"accepted_predicted":"1","input_tokens":225,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.17693135185291456,"brier_hard":0.05050795775093455,"score_mae":0.30909828255567506,"within_one":true,"case_id":"typed:14/r0","base_id":"typed:14","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":226.12460000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.00026277068310554805,"human_review":0.0016745601333920514,"observe":0.9974621669237382,"stop":0.0006005022597628638},"predicted":"observe","confidence":0.9974621669237382,"entropy_confidence":0.9856744313305342,"candidate_mass":0.9999626108110181,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.392204755122464,"brier_hard":1.9915848879804514,"soft_accuracy":0.1339355858857312,"brier_soft":1.0403963298094465,"tvd":0.8641290335906056,"kl":4.833976077931507,"case_id":"typed:15/r0","base_id":"typed:15","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.849601}
{"id":"needs_review","type":"binary","probabilities":{"false":1.198656221941197e-07,"true":0.9999998801343787},"predicted":"true","confidence":0.9999998801343787,"entropy_confidence":0.9999970711106522,"candidate_mass":0.9999996557965705,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.1986562850137839e-07,"brier_hard":2.873553455781237e-14,"soft_accuracy":0.876666909701152,"brier_soft":0.030421998644481835,"tvd":0.12333288013437826,"kl":1.0205450357303332,"case_id":"typed:15/r0","base_id":"typed:15","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.849601}
{"id":"outcome","type":"choice","probabilities":{"failure":9.341526146187277e-05,"harmful":0.9993154034087576,"partial":0.0005551261929844955,"success":3.605513679706747e-05},"predicted":"harmful","confidence":0.9993154034087576,"entropy_confidence":0.9956132165808798,"candidate_mass":0.9999953126913472,"accepted":true,"accepted_predicted":"harmful","input_tokens":211,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00068483103449427,"brier_hard":7.868639668416535e-07,"soft_accuracy":0.5664132768089976,"brier_soft":0.25747152339712837,"tvd":0.43264897007519043,"kl":2.595988116800438,"case_id":"typed:15/r0","base_id":"typed:15","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.849601}
{"id":"risk","type":"ordinal","probabilities":{"0":8.512816925313464e-06,"1":7.989070374240971e-06,"2":0.09366172168576174,"3":0.9063217764269382},"predicted":"3","confidence":0.9063217764269382,"entropy_confidence":0.7755621963856288,"candidate_mass":0.9999968013323616,"accepted":true,"accepted_predicted":"3","input_tokens":219,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0983608743749791,"brier_hard":0.017548127817238936,"score_mae":0.5296297617227128,"within_one":true,"case_id":"typed:15/r0","base_id":"typed:15","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.849601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002943256732449325,"1":0.0002949022147833475,"2":0.47794625470196594,"3":0.521464517410007},"predicted":"3","confidence":0.521464517410007,"entropy_confidence":0.4970978847205907,"candidate_mass":0.9999841663055807,"accepted":false,"accepted_predicted":null,"input_tokens":207,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.7382569906742873,"brier_hard":0.5444655294922742,"score_mae":0.320580963848736,"within_one":true,"case_id":"typed:15/r0","base_id":"typed:15","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.849601}
{"id":"action","type":"choice","probabilities":{"continue":0.0087263022714553,"human_review":0.034011510075852944,"observe":0.905565264642655,"stop":0.051696923010037},"predicted":"observe","confidence":0.905565264642655,"entropy_confidence":0.7119349406238933,"candidate_mass":0.9999462302163854,"accepted":true,"accepted_predicted":"observe","input_tokens":227,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.3810562800121495,"brier_hard":1.7559309313932943,"soft_accuracy":0.1089135213017852,"brier_soft":1.1000161867194298,"tvd":0.8222322646426549,"kl":2.1427878934197144,"case_id":"typed:16/r0","base_id":"typed:16","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.505}
{"id":"needs_review","type":"binary","probabilities":{"false":2.200972400829591e-08,"true":0.9999999779902746},"predicted":"true","confidence":0.9999999779902746,"entropy_confidence":0.9999994083790849,"candidate_mass":0.9999998067825304,"accepted":true,"accepted_predicted":"true","input_tokens":189,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.2009725617796442e-08,"brier_hard":9.688559620299147e-16,"soft_accuracy":0.8166669860604924,"brier_soft":0.06722196163756551,"tvd":0.18333297799027526,"kl":1.5231647995244022,"case_id":"typed:16/r0","base_id":"typed:16","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.505}
{"id":"outcome","type":"choice","probabilities":{"failure":6.562901815519259e-06,"harmful":0.9999442484327963,"partial":4.6468471447037154e-05,"success":2.7201939414830326e-06},"predicted":"harmful","confidence":0.9999442484327963,"entropy_confidence":0.9995437226984388,"candidate_mass":0.9999985200857191,"accepted":true,"accepted_predicted":"harmful","input_tokens":232,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":5.5753121380109495e-05,"brier_hard":5.318027219614576e-09,"soft_accuracy":0.6499713151798566,"brier_soft":0.18050134960190667,"tvd":0.3499442484327961,"kl":2.7690145167028164,"case_id":"typed:16/r0","base_id":"typed:16","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.505}
{"id":"risk","type":"ordinal","probabilities":{"0":3.819178973703384e-05,"1":1.950136098817398e-05,"2":0.3488992347079583,"3":0.6510430721413148},"predicted":"3","confidence":0.6510430721413148,"entropy_confidence":0.5330028566647627,"candidate_mass":0.9999949516960092,"accepted":false,"accepted_predicted":null,"input_tokens":240,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.42917947591875105,"brier_hard":0.24350161531928646,"score_mae":0.1176141872008496,"within_one":true,"case_id":"typed:16/r0","base_id":"typed:16","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.505}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009195180190865267,"1":0.0011079875144737778,"2":0.5927721125144974,"3":0.405200381951941},"predicted":"2","confidence":0.5927721125144974,"entropy_confidence":0.502267527860178,"candidate_mass":0.9999877732232093,"accepted":false,"accepted_predicted":null,"input_tokens":228,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":0.9033735639823348,"brier_hard":0.7051674361547364,"score_mae":0.05225335839929146,"within_one":true,"case_id":"typed:16/r0","base_id":"typed:16","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.505}
{"id":"action","type":"choice","probabilities":{"continue":0.0042649447191195275,"human_review":0.013536269134399787,"observe":0.9591853739759924,"stop":0.023013412170489765},"predicted":"observe","confidence":0.9591853739759924,"entropy_confidence":0.8497557723264654,"candidate_mass":0.9999329743238198,"accepted":true,"accepted_predicted":"observe","input_tokens":208,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.302382593455427,"brier_hard":1.8936950808559296,"soft_accuracy":0.3122298178708143,"brier_soft":0.6668636611611005,"tvd":0.6425183739759917,"kl":1.9514622436742446,"case_id":"typed:17/r0","base_id":"typed:17","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.4787}
{"id":"needs_review","type":"binary","probabilities":{"false":4.5701425520810664e-07,"true":0.9999995429857457},"predicted":"true","confidence":0.9999995429857457,"entropy_confidence":0.9999897153730433,"candidate_mass":0.9999992650866893,"accepted":true,"accepted_predicted":"true","input_tokens":170,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.570143587602483e-07,"brier_hard":4.177240581234971e-13,"soft_accuracy":0.6933328232881266,"brier_soft":0.18808873717365587,"tvd":0.30666654298574525,"kl":2.57057810929545,"case_id":"typed:17/r0","base_id":"typed:17","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.4787}
{"id":"outcome","type":"choice","probabilities":{"failure":0.002070323155029384,"harmful":0.5896907956508051,"partial":0.366636527275089,"success":0.04160235391907707},"predicted":"harmful","confidence":0.5896907956508051,"entropy_confidence":0.405321018259181,"candidate_mass":0.999949691002207,"accepted":false,"accepted_predicted":null,"input_tokens":213,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":null,"nll_hard":3.1795985287225683,"brier_hard":1.400687911859037,"soft_accuracy":0.17163950063164363,"brier_soft":0.5407465848345709,"tvd":0.5696607095925071,"kl":1.5550487472004,"case_id":"typed:17/r0","base_id":"typed:17","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.4787}
{"id":"risk","type":"ordinal","probabilities":{"0":3.8152114148449785e-06,"1":6.415204384214337e-06,"2":0.9999826954309976,"3":7.074153202842429e-06},"predicted":"2","confidence":0.9999826954309976,"entropy_confidence":0.9998373337123176,"candidate_mass":0.9999989494465897,"accepted":true,"accepted_predicted":"2","input_tokens":221,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":1.7304718728177995e-05,"brier_hard":4.0520243732719453e-10,"score_mae":0.116660028525988,"within_one":true,"case_id":"typed:17/r0","base_id":"typed:17","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.4787}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008379988263170943,"1":0.0013330455466343235,"2":0.8577810794879603,"3":0.14004787613908792},"predicted":"2","confidence":0.8577810794879603,"entropy_confidence":0.6958416059833425,"candidate_mass":0.9999902553012291,"accepted":true,"accepted_predicted":"2","input_tokens":209,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.1534063641629473,"brier_hard":0.039842108215141475,"score_mae":0.17037183293981895,"within_one":true,"case_id":"typed:17/r0","base_id":"typed:17","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.4787}
{"id":"action","type":"choice","probabilities":{"continue":0.0016784682802409188,"human_review":0.0015837406077801306,"observe":0.9965630209328038,"stop":0.00017477017917451798},"predicted":"observe","confidence":0.9965630209328038,"entropy_confidence":0.9813313172278143,"candidate_mass":0.9999680010701497,"accepted":true,"accepted_predicted":"observe","input_tokens":216,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.389873639404655,"brier_hard":1.9897862741649301,"soft_accuracy":0.18405278397731828,"brier_soft":1.2153048299715596,"tvd":0.8132298375996209,"kl":4.493024169519032,"case_id":"typed:18/r0","base_id":"typed:18","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.876501}
{"id":"needs_review","type":"binary","probabilities":{"false":1.1584916961197369e-08,"true":0.999999988415084},"predicted":"true","confidence":0.999999988415084,"entropy_confidence":0.9999996778712822,"candidate_mass":0.9999999104799078,"accepted":true,"accepted_predicted":"true","input_tokens":178,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.273561846873918,"brier_hard":1.9999999536603343,"soft_accuracy":0.10666700911346037,"brier_soft":1.5960876563812476,"tvd":0.8933329884150836,"kl":7.989175616407969,"case_id":"typed:18/r0","base_id":"typed:18","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.876501}
{"id":"outcome","type":"choice","probabilities":{"failure":1.3477528750831743e-06,"harmful":6.878129372805646e-07,"partial":9.097059126967092e-06,"success":0.9999888673750607},"predicted":"success","confidence":0.9999888673750607,"entropy_confidence":0.9998956175641146,"candidate_mass":0.9999992664187358,"accepted":true,"accepted_predicted":"success","input_tokens":221,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":1.1132686907476235e-05,"brier_hard":2.0898134724869828e-10,"soft_accuracy":0.7933255249607782,"brier_soft":0.06561508281554605,"tvd":0.2066558673750607,"kl":1.7198159606392251,"case_id":"typed:18/r0","base_id":"typed:18","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.876501}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9954917192724402,"1":0.004226870899567641,"2":0.00018936078192201042,"3":9.204904606846817e-05},"predicted":"0","confidence":0.9954917192724402,"entropy_confidence":0.978300408145023,"candidate_mass":0.9999953018631353,"accepted":true,"accepted_predicted":"0","input_tokens":229,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.004518473671758715,"brier_hard":3.823536325271129e-05,"score_mae":0.29845126039838293,"within_one":true,"case_id":"typed:18/r0","base_id":"typed:18","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.876501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006930991153006848,"1":0.9689580385771599,"2":0.030106753658165175,"3":0.00024210864937344707},"predicted":"1","confidence":0.9689580385771599,"entropy_confidence":0.8967918397986927,"candidate_mass":0.9999914133771549,"accepted":true,"accepted_predicted":"1","input_tokens":217,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.274337545479348,"brier_hard":1.9384004379115107,"score_mae":0.5265648718416105,"within_one":true,"case_id":"typed:18/r0","base_id":"typed:18","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.876501}
{"id":"action","type":"choice","probabilities":{"continue":0.0014889326873379221,"human_review":0.002774152760229321,"observe":0.9953429160723613,"stop":0.0003939984800704854},"predicted":"observe","confidence":0.9953429160723613,"entropy_confidence":0.9756473693822874,"candidate_mass":0.9999584133753203,"accepted":true,"accepted_predicted":"observe","input_tokens":221,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.509695732925405,"brier_hard":1.9877397232796425,"soft_accuracy":0.39915904804863644,"brier_soft":0.6123994925570456,"tvd":0.5957369145524323,"kl":2.902107647908377,"case_id":"typed:19/r0","base_id":"typed:19","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.8125}
{"id":"needs_review","type":"binary","probabilities":{"false":1.1295947897468255e-07,"true":0.9999998870405192},"predicted":"true","confidence":0.9999998870405192,"entropy_confidence":0.9999972301899385,"candidate_mass":0.9999996308120565,"accepted":true,"accepted_predicted":"true","input_tokens":183,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.996236675688369,"brier_hard":1.999999548162106,"soft_accuracy":0.1833330715410783,"brier_soft":1.3338896087769074,"tvd":0.8166668870405202,"kl":7.210765590358926,"case_id":"typed:19/r0","base_id":"typed:19","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.8125}
{"id":"outcome","type":"choice","probabilities":{"failure":4.752555894254366e-05,"harmful":2.7515206984894665e-05,"partial":0.002000108129074269,"success":0.9979248511049983},"predicted":"success","confidence":0.9979248511049983,"entropy_confidence":0.9889887843296375,"candidate_mass":0.9999900898728562,"accepted":true,"accepted_predicted":"success","input_tokens":226,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.002077304999811989,"brier_hard":8.309691229784016e-06,"soft_accuracy":0.6159903675871135,"brier_soft":0.238876793394491,"tvd":0.38125846777138167,"kl":1.7961981675917873,"case_id":"typed:19/r0","base_id":"typed:19","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.8125}
{"id":"risk","type":"ordinal","probabilities":{"0":1.149060233427911e-05,"1":1.4430543668858907e-05,"2":0.9999527482316376,"3":2.1330622360021026e-05},"predicted":"2","confidence":0.9999527482316376,"entropy_confidence":0.9995901242702054,"candidate_mass":0.9999984791711167,"accepted":true,"accepted_predicted":"2","input_tokens":234,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":11.14616350959863,"brier_hard":1.999876638403937,"score_mae":1.349983918874024,"within_one":false,"case_id":"typed:19/r0","base_id":"typed:19","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.8125}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002059701656338048,"1":0.9495166203482232,"2":0.050115258969786115,"3":0.000162150516357854},"predicted":"1","confidence":0.9495166203482232,"entropy_confidence":0.8540231286030114,"candidate_mass":0.9999904763558509,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.487779226683504,"brier_hard":1.903681479884352,"score_mae":0.6502335898368689,"within_one":true,"case_id":"typed:19/r0","base_id":"typed:19","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.8125}
{"id":"action","type":"choice","probabilities":{"continue":0.0010505510277220177,"human_review":0.0064541512051677526,"observe":0.9920248570016338,"stop":0.0004704407654767578},"predicted":"observe","confidence":0.9920248570016338,"entropy_confidence":0.9629939128440245,"candidate_mass":0.9999506277197415,"accepted":true,"accepted_predicted":"observe","input_tokens":217,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.008007114549872817,"brier_hard":0.00010658394559922814,"soft_accuracy":0.46456633380224316,"brier_soft":0.4226016259003804,"tvd":0.5253578570016336,"kl":2.2465072522950513,"case_id":"typed:20/r0","base_id":"typed:20","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.538}
{"id":"needs_review","type":"binary","probabilities":{"false":8.254105919339265e-07,"true":0.9999991745894081},"predicted":"true","confidence":0.9999991745894081,"entropy_confidence":0.9999821289699362,"candidate_mass":0.9999987234258162,"accepted":true,"accepted_predicted":"true","input_tokens":179,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.007384887224287,"brier_hard":1.999996698358995,"soft_accuracy":0.28333335767847345,"brier_soft":1.0272208116012318,"tvd":0.7166661745894081,"kl":6.243426812459479,"case_id":"typed:20/r0","base_id":"typed:20","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.538}
{"id":"outcome","type":"choice","probabilities":{"failure":5.126935307294164e-05,"harmful":3.79854294049787e-05,"partial":0.0015891308370569574,"success":0.9983216143804634},"predicted":"success","confidence":0.9983216143804634,"entropy_confidence":0.9907585820189623,"candidate_mass":0.9999883322368592,"accepted":true,"accepted_predicted":"success","input_tokens":222,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0016797956866589855,"brier_hard":5.346386544564179e-06,"soft_accuracy":0.8487348044994356,"brier_soft":0.03386788392660008,"tvd":0.1483216143804643,"kl":0.6103732018164649,"case_id":"typed:20/r0","base_id":"typed:20","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.538}
{"id":"risk","type":"ordinal","probabilities":{"0":2.2183944687895174e-06,"1":2.890353328712427e-06,"2":0.9999890225385238,"3":5.868713679847678e-06},"predicted":"2","confidence":0.9999890225385238,"entropy_confidence":0.9998936619695206,"candidate_mass":0.9999992216765543,"accepted":true,"accepted_predicted":"2","input_tokens":230,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":1.0977521728929368e-05,"brier_hard":1.6822187710052336e-10,"score_mae":0.4833315415714159,"within_one":true,"case_id":"typed:20/r0","base_id":"typed:20","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.538}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001242132927605706,"1":0.0022018539634580388,"2":0.8542532362529637,"3":0.142302776855972},"predicted":"2","confidence":0.8542532362529637,"entropy_confidence":0.687069972408984,"candidate_mass":0.9999909370345104,"accepted":true,"accepted_predicted":"2","input_tokens":218,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":6.118455562836392,"brier_hard":1.7456013550777525,"score_mae":1.3676166570373014,"within_one":false,"case_id":"typed:20/r0","base_id":"typed:20","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.538}
{"id":"action","type":"choice","probabilities":{"continue":0.009111720385903478,"human_review":0.00561731679530259,"observe":0.9848878448724907,"stop":0.0003831179463037972},"predicted":"observe","confidence":0.9848878448724907,"entropy_confidence":0.935130336897804,"candidate_mass":0.9999614289263599,"accepted":true,"accepted_predicted":"observe","input_tokens":207,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":4.698193739666358,"brier_hard":1.9518953506815029,"soft_accuracy":0.2560055147412337,"brier_soft":0.9877086078390046,"tvd":0.7315545915392371,"kl":2.7559078417014358,"case_id":"typed:21/r0","base_id":"typed:21","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.109501}
{"id":"needs_review","type":"binary","probabilities":{"false":1.340339089120457e-06,"true":0.99999865966091},"predicted":"true","confidence":0.99999865966091,"entropy_confidence":0.9999719176649362,"candidate_mass":0.9999990548276071,"accepted":true,"accepted_predicted":"true","input_tokens":169,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.522587924431175,"brier_hard":1.9999946386472347,"soft_accuracy":0.08666810801275336,"brier_soft":1.6683494410779063,"tvd":0.9133316596609105,"kl":8.200147989385023,"case_id":"typed:21/r0","base_id":"typed:21","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.109501}
{"id":"outcome","type":"choice","probabilities":{"failure":6.9001530171067615e-06,"harmful":2.869315705692075e-06,"partial":6.895301826070462e-05,"success":0.9999212775130152},"predicted":"success","confidence":0.9999212775130152,"entropy_confidence":0.9993810471362596,"candidate_mass":0.9999975720874468,"accepted":true,"accepted_predicted":"success","input_tokens":212,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":7.872558576242547e-05,"brier_hard":1.1007593768614147e-08,"soft_accuracy":0.7766180616318391,"brier_soft":0.08111795832594611,"tvd":0.22325427751301583,"kl":1.5912215950209116,"case_id":"typed:21/r0","base_id":"typed:21","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.109501}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9711100558200818,"1":0.0274537497733828,"2":0.0008860074211011812,"3":0.0005501869854325505},"predicted":"0","confidence":0.9711100558200818,"entropy_confidence":0.9007941121272894,"candidate_mass":0.9999892112332723,"accepted":true,"accepted_predicted":"0","input_tokens":220,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.029315474353492018,"brier_hard":0.0015894249662074908,"score_mae":0.08912367442811718,"within_one":true,"case_id":"typed:21/r0","base_id":"typed:21","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.109501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004213284411343635,"1":0.9986492609247081,"2":0.0008704099259785008,"3":5.900070818063608e-05},"predicted":"1","confidence":0.9986492609247081,"entropy_confidence":0.991825428245551,"candidate_mass":0.9999968057205777,"accepted":true,"accepted_predicted":"1","input_tokens":208,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.772097883224724,"brier_hard":1.996458628075375,"score_mae":0.270567082901207,"within_one":true,"case_id":"typed:21/r0","base_id":"typed:21","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.109501}
{"id":"action","type":"choice","probabilities":{"continue":5.728248582606353e-05,"human_review":0.0002072750529987618,"observe":0.9996286337074207,"stop":0.00010680875375480044},"predicted":"observe","confidence":0.9996286337074207,"entropy_confidence":0.9973558932362835,"candidate_mass":0.9999769653036774,"accepted":true,"accepted_predicted":"observe","input_tokens":213,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.000371435266117737,"brier_hard":1.9556526392078436e-07,"soft_accuracy":0.4598890742854811,"brier_soft":0.4187682921871433,"tvd":0.5396286337074205,"kl":3.8220721899341643,"case_id":"typed:22/r0","base_id":"typed:22","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.2574}
{"id":"needs_review","type":"binary","probabilities":{"false":2.7916222131624216e-07,"true":0.9999997208377777},"predicted":"true","confidence":0.9999997208377777,"entropy_confidence":0.9999935192245238,"candidate_mass":0.9999994112459688,"accepted":true,"accepted_predicted":"true","input_tokens":175,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.091472785685394,"brier_hard":1.9999988833512685,"soft_accuracy":0.23333314888670392,"brier_soft":1.1755557216803032,"tvd":0.7666667208377782,"kl":6.721697217920685,"case_id":"typed:22/r0","base_id":"typed:22","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.2574}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00013454429748084153,"harmful":0.00014890835696303143,"partial":0.9996425062956752,"success":7.404104988086008e-05},"predicted":"partial","confidence":0.9996425062956752,"entropy_confidence":0.9974225462076193,"candidate_mass":0.9999930574226941,"accepted":true,"accepted_predicted":"partial","input_tokens":218,"reused_prefix_tokens":0,"expected":"partial","correct":true,"accepted_correct":true,"nll_hard":0.0003575576204326478,"brier_hard":1.7355969245737547e-07,"soft_accuracy":0.5665153737225735,"brier_soft":0.27014361648389584,"tvd":0.4329755062956752,"kl":2.8504603928872334,"case_id":"typed:22/r0","base_id":"typed:22","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.2574}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0003905318117458043,"1":0.004992018402702831,"2":0.9936879599284033,"3":0.0009294898571484795},"predicted":"2","confidence":0.9936879599284033,"entropy_confidence":0.9694849105649914,"candidate_mass":0.9999889518393197,"accepted":true,"accepted_predicted":"2","input_tokens":226,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.848001127864773,"brier_hard":1.986660634797404,"score_mae":1.581823407830955,"within_one":false,"case_id":"typed:22/r0","base_id":"typed:22","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.2574}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004052989588128552,"1":0.001073784306227838,"2":0.996400487362523,"3":0.0021204293724374572},"predicted":"2","confidence":0.996400487362523,"entropy_confidence":0.980412947192667,"candidate_mass":0.999988332534527,"accepted":true,"accepted_predicted":"2","input_tokens":214,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.8108855932628565,"brier_hard":1.9920091467993535,"score_mae":1.376903047148586,"within_one":false,"case_id":"typed:22/r0","base_id":"typed:22","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.2574}
{"id":"action","type":"choice","probabilities":{"continue":0.0005486039490073868,"human_review":0.000989180877378873,"observe":0.9982744933087895,"stop":0.0001877218648235291},"predicted":"observe","confidence":0.9982744933087895,"entropy_confidence":0.989686501015732,"candidate_mass":0.9999759518187649,"accepted":true,"accepted_predicted":"observe","input_tokens":218,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":7.50813378115021,"brier_hard":1.9954560707775053,"soft_accuracy":0.2466814815923162,"brier_soft":0.9934793532688877,"tvd":0.7516074933087898,"kl":4.719399862866488,"case_id":"typed:23/r0","base_id":"typed:23","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.1311}
{"id":"needs_review","type":"binary","probabilities":{"false":2.983444686320838e-07,"true":0.999999701655531},"predicted":"true","confidence":0.999999701655531,"entropy_confidence":0.9999931025108427,"candidate_mass":0.9999993549782167,"accepted":true,"accepted_predicted":"true","input_tokens":180,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.025017083012482,"brier_hard":1.9999988066223027,"soft_accuracy":0.10000023867557488,"brier_soft":1.6199989259600902,"tvd":0.8999997016555312,"kl":8.059047855313612,"case_id":"typed:23/r0","base_id":"typed:23","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.1311}
{"id":"outcome","type":"choice","probabilities":{"failure":4.983276253073454e-05,"harmful":3.091113333847875e-05,"partial":0.002625421273682781,"success":0.9972938348304475},"predicted":"success","confidence":0.9972938348304475,"entropy_confidence":0.9862087146380646,"candidate_mass":0.9999946408668131,"accepted":true,"accepted_predicted":"success","input_tokens":223,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0027098334539985875,"brier_hard":1.4219605591590755e-05,"soft_accuracy":0.7119408610210364,"brier_soft":0.12505398255849015,"tvd":0.28396012149673444,"kl":1.2256063033542874,"case_id":"typed:23/r0","base_id":"typed:23","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.1311}
{"id":"risk","type":"ordinal","probabilities":{"0":0.006350218097994828,"1":0.992669373594781,"2":0.000784100505420599,"3":0.00019630780180488096},"predicted":"1","confidence":0.992669373594781,"entropy_confidence":0.9663030686309126,"candidate_mass":0.9999917464063776,"accepted":true,"accepted_predicted":"1","input_tokens":231,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":5.059266120526434,"brier_hard":1.972733027697313,"score_mae":0.3614934980110368,"within_one":true,"case_id":"typed:23/r0","base_id":"typed:23","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.1311}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0013646802180083827,"1":0.007400758631328999,"2":0.9846983051430492,"3":0.006536256007613612},"predicted":"2","confidence":0.9846983051430492,"entropy_confidence":0.9326435153569735,"candidate_mass":0.9999923840381201,"accepted":true,"accepted_predicted":"2","input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.906172766279436,"brier_hard":1.9549285911119492,"score_mae":1.1264061369402683,"within_one":false,"case_id":"typed:23/r0","base_id":"typed:23","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.1311}
{"id":"action","type":"choice","probabilities":{"continue":0.0009893047887838684,"human_review":0.259260053635754,"observe":0.7391761999023395,"stop":0.0005744416731211889},"predicted":"observe","confidence":0.7391761999023395,"entropy_confidence":0.5783684793207973,"candidate_mass":0.9999221969028493,"accepted":false,"accepted_predicted":null,"input_tokens":222,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":null,"nll_hard":1.3499236529300873,"brier_hard":1.0950784313489703,"soft_accuracy":0.3921633802816963,"brier_soft":0.1976050080572573,"tvd":0.37250956656897355,"kl":0.7970635284766658,"case_id":"typed:24/r0","base_id":"typed:24","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.638701}
{"id":"needs_review","type":"binary","probabilities":{"false":6.5752590402678e-08,"true":0.9999999342474112},"predicted":"true","confidence":0.9999999342474112,"entropy_confidence":0.999998336388977,"candidate_mass":0.9999997626769426,"accepted":true,"accepted_predicted":"true","input_tokens":184,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.575259097983695e-08,"brier_hard":8.646806080949126e-15,"soft_accuracy":0.5333329956165386,"brier_soft":0.4355560550397538,"tvd":0.46666693424741046,"kl":3.9629038699480943,"case_id":"typed:24/r0","base_id":"typed:24","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.638701}
{"id":"outcome","type":"choice","probabilities":{"failure":0.005999870189758693,"harmful":0.019860041083378206,"partial":0.5974752561069774,"success":0.37666483261988387},"predicted":"partial","confidence":0.5974752561069774,"entropy_confidence":0.4344431263836849,"candidate_mass":0.999943244374975,"accepted":false,"accepted_predicted":null,"input_tokens":227,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":null,"nll_hard":0.9763995251064657,"brier_hard":0.745953832227023,"soft_accuracy":0.38777021637789016,"brier_soft":0.13096534248901046,"tvd":0.29747525610697834,"kl":0.3650107122501575,"case_id":"typed:24/r0","base_id":"typed:24","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.638701}
{"id":"risk","type":"ordinal","probabilities":{"0":1.8282041111277582e-05,"1":2.869174605296289e-05,"2":0.999913554173141,"3":3.9472039694839725e-05},"predicted":"2","confidence":0.999913554173141,"entropy_confidence":0.9992885933216147,"candidate_mass":0.9999977850468812,"accepted":true,"accepted_predicted":"2","input_tokens":235,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":8.644956351485401e-05,"brier_hard":1.0188372217772173e-08,"score_mae":0.4333072162114193,"within_one":true,"case_id":"typed:24/r0","base_id":"typed:24","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.638701}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008141456983734872,"1":0.005755064027133856,"2":0.9903571404116996,"3":0.0030736498627917538},"predicted":"2","confidence":0.9903571404116996,"entropy_confidence":0.9546625775583307,"candidate_mass":0.9999932188731814,"accepted":true,"accepted_predicted":"2","input_tokens":223,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.157675111390795,"brier_hard":1.9693403684288249,"score_mae":0.6623572944389082,"within_one":true,"case_id":"typed:24/r0","base_id":"typed:24","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.638701}
{"id":"action","type":"choice","probabilities":{"continue":0.023929585432332794,"human_review":0.03860004906404617,"observe":0.9341542395564288,"stop":0.0033161259471904945},"predicted":"observe","confidence":0.9341542395564288,"entropy_confidence":0.7853956593162043,"candidate_mass":0.9999403028492863,"accepted":true,"accepted_predicted":"observe","input_tokens":217,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.254501731422881,"brier_hard":1.7975176306911653,"soft_accuracy":0.15541294480522186,"brier_soft":0.9836472769868139,"tvd":0.7941542395564297,"kl":2.314504888555744,"case_id":"typed:25/r0","base_id":"typed:25","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":186.944}
{"id":"needs_review","type":"binary","probabilities":{"false":2.135068961423335e-07,"true":0.999999786493104},"predicted":"true","confidence":0.999999786493104,"entropy_confidence":0.9999949608296916,"candidate_mass":0.9999994613786919,"accepted":true,"accepted_predicted":"true","input_tokens":179,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.1350691884441885e-07,"brier_hard":9.117038936201513e-14,"soft_accuracy":0.7933328747427633,"brier_soft":0.08542232127877242,"tvd":0.20666678649310394,"kl":1.7198072980934076,"case_id":"typed:25/r0","base_id":"typed:25","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":186.944}
{"id":"outcome","type":"choice","probabilities":{"failure":2.7234000088516976e-05,"harmful":0.9998218190255959,"partial":0.00013850471063052671,"success":1.2442263684574364e-05},"predicted":"harmful","confidence":0.9998218190255959,"entropy_confidence":0.9986759588772756,"candidate_mass":0.9999966825754742,"accepted":true,"accepted_predicted":"harmful","input_tokens":222,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00017819685051986972,"brier_hard":5.182851519287281e-08,"soft_accuracy":0.4832871877967036,"brier_soft":0.3760467898422998,"tvd":0.5164888190255961,"kl":3.890070089996255,"case_id":"typed:25/r0","base_id":"typed:25","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":186.944}
{"id":"risk","type":"ordinal","probabilities":{"0":2.0806424880766194e-05,"1":1.1945572462131697e-05,"2":0.02174513087306158,"3":0.9782221171295963},"predicted":"3","confidence":0.9782221171295963,"entropy_confidence":0.9241524617459702,"candidate_mass":0.9999974217772324,"accepted":true,"accepted_predicted":"3","input_tokens":230,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.02201852110427179,"brier_hard":0.0009471274746076174,"score_mae":0.5781685587073739,"within_one":true,"case_id":"typed:25/r0","base_id":"typed:25","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":186.944}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00020757311595084144,"1":0.000295350250701907,"2":0.0025866826440113524,"3":0.9969103939893353},"predicted":"3","confidence":0.9969103939893353,"entropy_confidence":0.983657643022195,"candidate_mass":0.9999913564201065,"accepted":true,"accepted_predicted":"3","input_tokens":218,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0030943886969332377,"brier_hard":1.6366910771019965e-05,"score_mae":0.6595328975067312,"within_one":true,"case_id":"typed:25/r0","base_id":"typed:25","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":186.944}
{"id":"action","type":"choice","probabilities":{"continue":0.001408555018527171,"human_review":0.0032289840900226285,"observe":0.9951275485403317,"stop":0.00023491235111804161},"predicted":"observe","confidence":0.9951275485403317,"entropy_confidence":0.9750477860818678,"candidate_mass":0.9999633665211397,"accepted":true,"accepted_predicted":"observe","input_tokens":214,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.73558771443318,"brier_hard":1.9838333352331516,"soft_accuracy":0.20131806974589211,"brier_soft":1.1015436416994127,"tvd":0.795127548540332,"kl":3.893235911247966,"case_id":"typed:26/r0","base_id":"typed:26","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3382}
{"id":"needs_review","type":"binary","probabilities":{"false":2.0660093997651086e-08,"true":0.9999999793399044},"predicted":"true","confidence":0.9999999793399044,"entropy_confidence":0.9999994427709629,"candidate_mass":0.9999998510507138,"accepted":true,"accepted_predicted":"true","input_tokens":176,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.0660095847703726e-08,"brier_hard":8.536790356095371e-16,"soft_accuracy":0.8166669869152587,"brier_soft":0.06722196262729219,"tvd":0.18333297933990517,"kl":1.523164798422204,"case_id":"typed:26/r0","base_id":"typed:26","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3382}
{"id":"outcome","type":"choice","probabilities":{"failure":5.007153709548125e-05,"harmful":0.9996794313653723,"partial":0.0002510876885753145,"success":1.940940895855301e-05},"predicted":"harmful","confidence":0.9996794313653723,"entropy_confidence":0.9977577925889195,"candidate_mass":0.999996300869369,"accepted":true,"accepted_predicted":"harmful","input_tokens":219,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.0003206200277360975,"brier_hard":1.6869316084438953e-07,"soft_accuracy":0.4999026605538556,"brier_soft":0.3606866547169444,"tvd":0.4996799313648715,"kl":3.5287608308600094,"case_id":"typed:26/r0","base_id":"typed:26","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3382}
{"id":"risk","type":"ordinal","probabilities":{"0":2.566184062989419e-06,"1":1.7925499051726078e-06,"2":0.00044263342992766086,"3":0.9995530078361028},"predicted":"3","confidence":0.9995530078361028,"entropy_confidence":0.9971708740778724,"candidate_mass":0.9999992586421781,"accepted":true,"accepted_predicted":"3","input_tokens":227,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.722768602409616,"brier_hard":1.9982211445484968,"score_mae":0.7528790829180685,"within_one":true,"case_id":"typed:26/r0","base_id":"typed:26","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3382}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007295316888438776,"1":0.0011512245389360847,"2":0.5148474803875795,"3":0.4832717633846389},"predicted":"2","confidence":0.5148474803875795,"entropy_confidence":0.49052470300581663,"candidate_mass":0.9999880727345862,"accepted":false,"accepted_predicted":null,"input_tokens":215,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":0.7271761264506722,"brier_hard":0.5320778561114838,"score_mae":0.2873284754680121,"within_one":true,"case_id":"typed:26/r0","base_id":"typed:26","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3382}
{"id":"action","type":"choice","probabilities":{"continue":0.001602916237413862,"human_review":0.003210123241240493,"observe":0.9947681171978817,"stop":0.00041884332346223715},"predicted":"observe","confidence":0.9947681171978817,"entropy_confidence":0.9731493184002703,"candidate_mass":0.9999678912646471,"accepted":true,"accepted_predicted":"observe","input_tokens":224,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.005245617025789022,"brier_hard":4.042225907282731e-05,"soft_accuracy":0.6309357062339512,"brier_soft":0.19381599974293398,"tvd":0.36143511719788257,"kl":1.2775190425709666,"case_id":"typed:27/r0","base_id":"typed:27","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.906501}
{"id":"needs_review","type":"binary","probabilities":{"false":1.7563949339953772e-06,"true":0.9999982436050671},"predicted":"true","confidence":0.9999982436050671,"entropy_confidence":0.999963885630775,"candidate_mass":0.9999982106433287,"accepted":true,"accepted_predicted":"true","input_tokens":186,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.252247182574454,"brier_hard":1.999992974426436,"soft_accuracy":0.2666678196497986,"brier_soft":1.0755494256947067,"tvd":0.7333312436050665,"kl":6.401778673128366,"case_id":"typed:27/r0","base_id":"typed:27","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.906501}
{"id":"outcome","type":"choice","probabilities":{"failure":3.9932785585984525e-05,"harmful":2.6600538711940284e-05,"partial":0.0012618186457530832,"success":0.9986716480299478},"predicted":"success","confidence":0.9986716480299478,"entropy_confidence":0.992472709827972,"candidate_mass":0.9999925360971371,"accepted":true,"accepted_predicted":"success","input_tokens":229,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0013292350116103353,"brier_hard":3.3590074671361555e-06,"soft_accuracy":0.6993672397500496,"brier_soft":0.1475009533452635,"tvd":0.29869824856866034,"kl":1.4639228481329232,"case_id":"typed:27/r0","base_id":"typed:27","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.906501}
{"id":"risk","type":"ordinal","probabilities":{"0":0.037579894018996245,"1":0.9519181569761107,"2":0.008967285856469038,"3":0.0015346631484256757},"predicted":"1","confidence":0.9519181569761107,"entropy_confidence":0.8395474937639813,"candidate_mass":0.9999816626201078,"accepted":true,"accepted_predicted":"1","input_tokens":237,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.049276217449275336,"brier_hard":0.003806879469663687,"score_mae":0.3911237181343258,"within_one":true,"case_id":"typed:27/r0","base_id":"typed:27","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.906501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004398440119604591,"1":0.91571020703916,"2":0.08330113454334064,"3":0.0005488144055392009},"predicted":"1","confidence":0.91571020703916,"entropy_confidence":0.7870720201081962,"candidate_mass":0.9999898563145002,"accepted":true,"accepted_predicted":"1","input_tokens":225,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.08805533222405512,"brier_hard":0.014044342873595603,"score_mae":0.11729191934245875,"within_one":true,"case_id":"typed:27/r0","base_id":"typed:27","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.906501}
{"id":"action","type":"choice","probabilities":{"continue":0.009546372334110119,"human_review":0.005765921904766434,"observe":0.9844477193862393,"stop":0.00023998637488574235},"predicted":"observe","confidence":0.9844477193862393,"entropy_confidence":0.9339500206780347,"candidate_mass":0.9999651766114641,"accepted":true,"accepted_predicted":"observe","input_tokens":207,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":4.651594056957304,"brier_hard":1.950169004210161,"soft_accuracy":0.2854533353297656,"brier_soft":0.8866887215526486,"tvd":0.7013544224278409,"kl":2.526424124572301,"case_id":"typed:28/r0","base_id":"typed:28","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.6842}
{"id":"needs_review","type":"binary","probabilities":{"false":3.4383437566452544e-08,"true":0.999999965616564},"predicted":"true","confidence":0.999999965616564,"entropy_confidence":0.9999990979021751,"candidate_mass":0.9999998336284431,"accepted":true,"accepted_predicted":"true","input_tokens":169,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.18569085460057,"brier_hard":1.9999998624662554,"soft_accuracy":0.12333302590221275,"brier_soft":1.5370899372067048,"tvd":0.8766669656165632,"kl":7.816281868760351,"case_id":"typed:28/r0","base_id":"typed:28","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.6842}
{"id":"outcome","type":"choice","probabilities":{"failure":5.390839700294614e-07,"harmful":2.4207762055990143e-07,"partial":2.8526619376283287e-06,"success":0.9999963661764731},"predicted":"success","confidence":0.9999963661764731,"entropy_confidence":0.99996283393113,"candidate_mass":0.9999994910165152,"accepted":true,"accepted_predicted":"success","input_tokens":212,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":3.633830129213409e-06,"brier_hard":2.1691566655878433e-11,"soft_accuracy":0.6666653355387397,"brier_soft":0.18055133907515844,"tvd":0.333329608254093,"kl":2.7998029358728393,"case_id":"typed:28/r0","base_id":"typed:28","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.6842}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9991364279312629,"1":0.0007464628474940727,"2":7.083133459967975e-05,"3":4.6277886643280054e-05},"predicted":"0","confidence":0.9991364279312629,"entropy_confidence":0.9946789401814721,"candidate_mass":0.9999972499166155,"accepted":true,"accepted_predicted":"0","input_tokens":220,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0008639451619067493,"brier_hard":1.3101222213451846e-06,"score_mae":0.33564004082337673,"within_one":true,"case_id":"typed:28/r0","base_id":"typed:28","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.6842}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00016783186176772523,"1":0.9995813202729323,"2":0.00022603120707354136,"3":2.481665822771642e-05},"predicted":"1","confidence":0.9995813202729323,"entropy_confidence":0.9970871052644038,"candidate_mass":0.9999974713257661,"accepted":true,"accepted_predicted":"1","input_tokens":208,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.692547902529732,"brier_hard":1.99882723198855,"score_mae":0.5467748326617625,"within_one":true,"case_id":"typed:28/r0","base_id":"typed:28","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.6842}
{"id":"action","type":"choice","probabilities":{"continue":0.0006739876313252136,"human_review":0.001776229009250576,"observe":0.9972521792542177,"stop":0.00029760410520661814},"predicted":"observe","confidence":0.9972521792542177,"entropy_confidence":0.9846125783266426,"candidate_mass":0.999951933203717,"accepted":true,"accepted_predicted":"observe","input_tokens":203,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0027516029353158566,"brier_hard":1.1248335874869978e-05,"soft_accuracy":0.4661357521286492,"brier_soft":0.4333556581430119,"tvd":0.5305851792542177,"kl":2.4622068669492,"case_id":"typed:29/r0","base_id":"typed:29","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":214.727}
{"id":"needs_review","type":"binary","probabilities":{"false":5.86082387459462e-07,"true":0.9999994139176115},"predicted":"true","confidence":0.9999994139176115,"entropy_confidence":0.9999870211540707,"candidate_mass":0.9999990502719915,"accepted":true,"accepted_predicted":"true","input_tokens":165,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.860825602558958e-07,"brier_hard":6.869851310112369e-13,"soft_accuracy":0.5499999413917608,"brier_soft":0.40499894505238854,"tvd":0.449999413917612,"kl":3.815843139319099,"case_id":"typed:29/r0","base_id":"typed:29","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":214.727}
{"id":"outcome","type":"choice","probabilities":{"failure":9.781538810597193e-06,"harmful":5.3890510084433e-06,"partial":7.882992007268566e-05,"success":0.9999059994901098},"predicted":"success","confidence":0.9999059994901098,"entropy_confidence":0.9992663867727973,"candidate_mass":0.9999962386294504,"accepted":true,"accepted_predicted":"success","input_tokens":208,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":9.44821793681298,"brier_hard":1.999654354315047,"soft_accuracy":0.11670825780218878,"brier_soft":1.2291730541068149,"tvd":0.8832389994901091,"kl":7.730519313972571,"case_id":"typed:29/r0","base_id":"typed:29","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":214.727}
{"id":"risk","type":"ordinal","probabilities":{"0":0.586453170089303,"1":0.408401850235009,"2":0.0027112520438969147,"3":0.002433727631790953},"predicted":"0","confidence":0.586453170089303,"entropy_confidence":0.48830145506309197,"candidate_mass":0.9999608759147046,"accepted":false,"accepted_predicted":null,"input_tokens":216,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.8955036622537768,"brier_hard":0.6939289654309848,"score_mae":0.3455414627818243,"within_one":true,"case_id":"typed:29/r0","base_id":"typed:29","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":214.727}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00039553816086574024,"1":0.98737453995997,"2":0.011953488735773088,"3":0.00027643314339005973},"predicted":"1","confidence":0.98737453995997,"entropy_confidence":0.9489109761169006,"candidate_mass":0.9999899597662545,"accepted":true,"accepted_predicted":"1","input_tokens":204,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.01270583841996141,"brier_hard":0.0003025209998981129,"score_mae":0.03788918313831369,"within_one":true,"case_id":"typed:29/r0","base_id":"typed:29","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":214.727}
{"id":"action","type":"choice","probabilities":{"continue":0.0017674465249328222,"human_review":0.03492681023143111,"observe":0.9625248618181825,"stop":0.0007808814254526853},"predicted":"observe","confidence":0.9625248618181825,"entropy_confidence":0.8768546185161546,"candidate_mass":0.9999489775248853,"accepted":true,"accepted_predicted":"observe","input_tokens":223,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.35450054321581,"brier_hard":1.8578241048712107,"soft_accuracy":0.2915847865837981,"brier_soft":0.726174982166755,"tvd":0.6791915784848996,"kl":1.9000555087011881,"case_id":"typed:30/r0","base_id":"typed:30","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.323001}
{"id":"needs_review","type":"binary","probabilities":{"false":1.6635305466576469e-06,"true":0.9999983364694546},"predicted":"true","confidence":0.9999983364694546,"entropy_confidence":0.9999656647061425,"candidate_mass":0.9999984948574466,"accepted":true,"accepted_predicted":"true","input_tokens":185,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6635319290944786e-06,"brier_hard":5.534667755228435e-12,"soft_accuracy":0.743332190416244,"brier_soft":0.13175418988995602,"tvd":0.25666533646945394,"kl":2.1435108116400787,"case_id":"typed:30/r0","base_id":"typed:30","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.323001}
{"id":"outcome","type":"choice","probabilities":{"failure":0.003159471913374696,"harmful":0.0006388053831915874,"partial":0.25582891019890097,"success":0.7403728125045325},"predicted":"success","confidence":0.7403728125045325,"entropy_confidence":0.571372320384637,"candidate_mass":0.9999230776233246,"accepted":false,"accepted_predicted":null,"input_tokens":228,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":null,"nll_hard":1.3632463774631702,"brier_hard":1.101952902726716,"soft_accuracy":0.30534788959717196,"brier_soft":0.42733678170817396,"tvd":0.5303728125045326,"kl":1.061420869214462,"case_id":"typed:30/r0","base_id":"typed:30","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.323001}
{"id":"risk","type":"ordinal","probabilities":{"0":5.153246454617853e-06,"1":5.944275509709659e-06,"2":0.9999753124046283,"3":1.3590073408849144e-05},"predicted":"2","confidence":0.9999753124046283,"entropy_confidence":0.9997754782330008,"candidate_mass":0.9999987016574936,"accepted":true,"accepted_predicted":"2","input_tokens":236,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":12.033081900729705,"brier_hard":1.9999387371142952,"score_mae":0.6166643393049931,"within_one":true,"case_id":"typed:30/r0","base_id":"typed:30","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.323001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0013089062842573693,"1":0.06349293541164784,"2":0.9306793104524782,"3":0.004518847851616053},"predicted":"2","confidence":0.9306793104524782,"entropy_confidence":0.8016378951156541,"candidate_mass":0.9999844963265743,"accepted":true,"accepted_predicted":"2","input_tokens":224,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.7568266326324817,"brier_hard":1.7432315941497594,"score_mae":0.45507509987145234,"within_one":true,"case_id":"typed:30/r0","base_id":"typed:30","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.323001}
{"id":"action","type":"choice","probabilities":{"continue":0.0010946624471056458,"human_review":0.0018990525247668789,"observe":0.9967030217073055,"stop":0.00030326332082270804},"predicted":"observe","confidence":0.9967030217073055,"entropy_confidence":0.9818861548630943,"candidate_mass":0.9999583446135404,"accepted":true,"accepted_predicted":"observe","input_tokens":226,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.266400188362891,"brier_hard":1.9896237050859464,"soft_accuracy":0.3333398155163267,"brier_soft":0.7172978568808269,"tvd":0.6633700217073052,"kl":3.2380856627925287,"case_id":"typed:31/r0","base_id":"typed:31","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.7592}
{"id":"needs_review","type":"binary","probabilities":{"false":8.292532583711224e-08,"true":0.9999999170746746},"predicted":"true","confidence":0.9999999170746746,"entropy_confidence":0.9999979296608607,"candidate_mass":0.999999577999166,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.292532882253602e-08,"brier_hard":1.3753219255271826e-14,"soft_accuracy":0.6499999751224026,"brier_soft":0.24499988390455787,"tvd":0.34999991707467437,"kl":2.943610288633032,"case_id":"typed:31/r0","base_id":"typed:31","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.7592}
{"id":"outcome","type":"choice","probabilities":{"failure":2.13608780833104e-05,"harmful":1.243036280290911e-05,"partial":0.0006071430173392352,"success":0.9993590657417729},"predicted":"success","confidence":0.9993590657417729,"entropy_confidence":0.996026960415897,"candidate_mass":0.9999940546134267,"accepted":true,"accepted_predicted":"success","input_tokens":231,"reused_prefix_tokens":0,"expected":"failure","correct":false,"accepted_correct":false,"nll_hard":10.753949435948009,"brier_hard":1.9986761897575471,"soft_accuracy":0.06357199723368887,"brier_soft":1.2965751137133044,"tvd":0.9360260024087105,"kl":7.255548326732753,"case_id":"typed:31/r0","base_id":"typed:31","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.7592}
{"id":"risk","type":"ordinal","probabilities":{"0":0.6375452669662552,"1":0.35512837051655294,"2":0.0035940325062518955,"3":0.0037323300109400755},"predicted":"0","confidence":0.6375452669662552,"entropy_confidence":0.4981372696805111,"candidate_mass":0.9999628238437183,"accepted":false,"accepted_predicted":null,"input_tokens":239,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.0352759477819262,"brier_hard":0.8223502333006765,"score_mae":0.5431535744381231,"within_one":true,"case_id":"typed:31/r0","base_id":"typed:31","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.7592}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007676400515331591,"1":0.7316673317391714,"2":0.26646775387544763,"3":0.0010972743338482358},"predicted":"1","confidence":0.7316673317391714,"entropy_confidence":0.5715325654508061,"candidate_mass":0.999983268711979,"accepted":false,"accepted_predicted":null,"input_tokens":227,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.31242933315846594,"brier_hard":0.14300927799361446,"score_mae":0.03210533750838862,"within_one":true,"case_id":"typed:31/r0","base_id":"typed:31","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.7592}
{"id":"action","type":"choice","probabilities":{"continue":0.0005079816203079028,"human_review":0.006181805791869905,"observe":0.9923150066469659,"stop":0.0009952059408557831},"predicted":"observe","confidence":0.9923150066469659,"entropy_confidence":0.9640556548829451,"candidate_mass":0.9999589973736551,"accepted":true,"accepted_predicted":"observe","input_tokens":220,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.086144850862748,"brier_hard":1.972364924036068,"soft_accuracy":0.26825380917114544,"brier_soft":0.8548873706112035,"tvd":0.7256482733136995,"kl":2.9960117646174216,"case_id":"typed:32/r0","base_id":"typed:32","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.44320000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":1.8489647630304373e-08,"true":0.9999999815103531},"predicted":"true","confidence":0.9999999815103531,"entropy_confidence":0.9999994983499373,"candidate_mass":0.9999998594828795,"accepted":true,"accepted_predicted":"true","input_tokens":182,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.848964709556892e-08,"brier_hard":6.837341128904992e-16,"soft_accuracy":0.799999988906212,"brier_soft":0.07999998520828283,"tvd":0.1999999815103527,"kl":1.6635532481355861,"case_id":"typed:32/r0","base_id":"typed:32","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.44320000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":9.178224445855518e-06,"harmful":0.9998978560605363,"partial":8.78173008251605e-05,"success":5.1484141915189475e-06},"predicted":"harmful","confidence":0.9998978560605363,"entropy_confidence":0.9992126340144954,"candidate_mass":0.9999982227469443,"accepted":true,"accepted_predicted":"harmful","input_tokens":225,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00010214915651117548,"brier_hard":1.82560086660524e-08,"soft_accuracy":0.383324205675275,"brier_soft":0.5453694968045313,"tvd":0.6165648560605369,"kl":4.987732185272971,"case_id":"typed:32/r0","base_id":"typed:32","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.44320000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":3.855634762183949e-06,"1":2.911282453338189e-06,"2":0.0009360695453550374,"3":0.999057163537431},"predicted":"3","confidence":0.999057163537431,"entropy_confidence":0.9945498247440335,"candidate_mass":0.9999989390833669,"accepted":true,"accepted_predicted":"3","input_tokens":233,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0009432812124394653,"brier_hard":1.765190130375727e-06,"score_mae":0.7990465409854561,"within_one":true,"case_id":"typed:32/r0","base_id":"typed:32","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.44320000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005717085710494085,"1":0.0019748159944476843,"2":0.9951894413838309,"3":0.0022640340506719587},"predicted":"2","confidence":0.9951894413838309,"entropy_confidence":0.9746410832323684,"candidate_mass":0.9999847398518155,"accepted":true,"accepted_predicted":"2","input_tokens":221,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.004822166595474182,"brier_hard":3.249407328433841e-05,"score_mae":0.05085419908587441,"within_one":true,"case_id":"typed:32/r0","base_id":"typed:32","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":194.44320000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.01987158866032202,"human_review":0.07528954869055705,"observe":0.9031183957992892,"stop":0.0017204668498323813},"predicted":"observe","confidence":0.9031183957992892,"entropy_confidence":0.7290789505597144,"candidate_mass":0.9999496177622753,"accepted":true,"accepted_predicted":"observe","input_tokens":223,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":2.5864139494194234,"brier_hard":1.6711100956340617,"soft_accuracy":0.08970563651470068,"brier_soft":1.1194999977637743,"tvd":0.8729899844596108,"kl":3.015650830879328,"case_id":"typed:33/r0","base_id":"typed:33","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.344301}
{"id":"needs_review","type":"binary","probabilities":{"false":7.059320684009868e-07,"true":0.9999992940679303},"predicted":"true","confidence":0.9999992940679303,"entropy_confidence":0.9999845565628587,"candidate_mass":0.9999990356062615,"accepted":true,"accepted_predicted":"true","input_tokens":185,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.059323188591023e-07,"brier_hard":9.966801722122081e-13,"soft_accuracy":0.9333323881926768,"brier_soft":0.0088887895295037,"tvd":0.06666629406793098,"kl":0.5496327633916608,"case_id":"typed:33/r0","base_id":"typed:33","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.344301}
{"id":"outcome","type":"choice","probabilities":{"failure":2.875775632391873e-06,"harmful":0.9999690463445511,"partial":2.5961929703215307e-05,"success":2.115950113322539e-06},"predicted":"harmful","confidence":0.9999690463445511,"entropy_confidence":0.9997335186160305,"candidate_mass":0.9999989257543739,"accepted":true,"accepted_predicted":"harmful","input_tokens":228,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":3.095413452319651e-05,"brier_hard":1.6448979099349275e-09,"soft_accuracy":0.7666458637056708,"brier_soft":0.08386904470065855,"tvd":0.23330416229466439,"kl":1.8506973041391148,"case_id":"typed:33/r0","base_id":"typed:33","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.344301}
{"id":"risk","type":"ordinal","probabilities":{"0":2.9369350375788087e-06,"1":1.9926819394343714e-06,"2":0.00011342271970159305,"3":0.99988164766332},"predicted":"3","confidence":0.99988164766332,"entropy_confidence":0.9991255186790742,"candidate_mass":0.9999992444794911,"accepted":true,"accepted_predicted":"3","input_tokens":236,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00011835934087045125,"brier_hard":2.688458531085021e-08,"score_mae":0.29987378111130214,"within_one":true,"case_id":"typed:33/r0","base_id":"typed:33","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.344301}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008021638903268509,"1":0.0028941187958825645,"2":0.9781299003360875,"3":0.018173816977702067},"predicted":"2","confidence":0.9781299003360875,"entropy_confidence":0.9155301702810423,"candidate_mass":0.9999877943036782,"accepted":true,"accepted_predicted":"2","input_tokens":224,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":4.007773348316416,"brier_hard":1.920729774990131,"score_mae":0.5029916295988364,"within_one":true,"case_id":"typed:33/r0","base_id":"typed:33","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.344301}
{"id":"action","type":"choice","probabilities":{"continue":0.0013859193043354947,"human_review":0.033269519465648566,"observe":0.9618783764449536,"stop":0.003466184785063438},"predicted":"observe","confidence":0.9618783764449536,"entropy_confidence":0.8706178727321321,"candidate_mass":0.9999340824948525,"accepted":true,"accepted_predicted":"observe","input_tokens":224,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.03886726412099864,"brier_hard":0.002574054317230176,"soft_accuracy":0.4555307103584194,"brier_soft":0.3680475420462984,"tvd":0.495211376444953,"kl":1.7707820936457916,"case_id":"typed:34/r0","base_id":"typed:34","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.58309999999997}
{"id":"needs_review","type":"binary","probabilities":{"false":9.513676419659126e-07,"true":0.9999990486323589},"predicted":"true","confidence":0.9999990486323589,"entropy_confidence":0.9999795967906357,"candidate_mass":0.9999985812086021,"accepted":true,"accepted_predicted":"true","input_tokens":186,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.865365264493704,"brier_hard":1.9999961945312439,"soft_accuracy":0.38333322198641767,"brier_soft":0.7605540310716927,"tvd":0.6166660486323585,"kl":5.312154017727519,"case_id":"typed:34/r0","base_id":"typed:34","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.58309999999997}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0004621114747289898,"harmful":0.9446184840679539,"partial":0.05132939437291457,"success":0.003590010084402622},"predicted":"harmful","confidence":0.9446184840679539,"entropy_confidence":0.8340898896342608,"candidate_mass":0.9999715048249893,"accepted":true,"accepted_predicted":"harmful","input_tokens":229,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":false,"nll_hard":5.629600267460095,"brier_hard":1.8877718687201455,"soft_accuracy":0.04822284263430591,"brier_soft":1.2468401669548173,"tvd":0.9112854507349205,"kl":3.8767286524608133,"case_id":"typed:34/r0","base_id":"typed:34","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.58309999999997}
{"id":"risk","type":"ordinal","probabilities":{"0":1.9655936074630756e-05,"1":3.0750883099896854e-05,"2":0.999925941432835,"3":2.365174799222196e-05},"predicted":"2","confidence":0.999925941432835,"entropy_confidence":0.9993807238277941,"candidate_mass":0.9999977295865835,"accepted":true,"accepted_predicted":"2","input_tokens":237,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":7.406130963611355e-05,"brier_hard":7.376049188017606e-09,"score_mae":0.30662058899274647,"within_one":true,"case_id":"typed:34/r0","base_id":"typed:34","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.58309999999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001565996746845543,"1":0.0036817232559176565,"2":0.9435464234812538,"3":0.0512058565159819},"predicted":"2","confidence":0.9435464234812538,"entropy_confidence":0.8284946794407614,"candidate_mass":0.9999894878872206,"accepted":true,"accepted_predicted":"2","input_tokens":225,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.604374360343737,"brier_hard":1.88555445392591,"score_mae":0.8277251397663712,"within_one":true,"case_id":"typed:34/r0","base_id":"typed:34","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.58309999999997}
{"id":"action","type":"choice","probabilities":{"continue":0.0008240949184055062,"human_review":0.0017877534179588789,"observe":0.9971849652254339,"stop":0.0002031864382031858},"predicted":"observe","confidence":0.9971849652254339,"entropy_confidence":0.984345843469781,"candidate_mass":0.9999723199635464,"accepted":true,"accepted_predicted":"observe","input_tokens":229,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":7.101224842449227,"brier_hard":1.9927335815142853,"soft_accuracy":0.3330621916975396,"brier_soft":0.7243685435120172,"tvd":0.6638519652254331,"kl":3.6108893501038426,"case_id":"typed:35/r0","base_id":"typed:35","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":208.3467}
{"id":"needs_review","type":"binary","probabilities":{"false":7.383380455289803e-08,"true":0.9999999261661962},"predicted":"true","confidence":0.9999999261661962,"entropy_confidence":0.9999981442731003,"candidate_mass":0.9999996675147023,"accepted":true,"accepted_predicted":"true","input_tokens":191,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.421449153943403,"brier_hard":1.9999997046647942,"soft_accuracy":0.2333330393780785,"brier_soft":1.1755563513542464,"tvd":0.7666669261661958,"kl":6.721697170010781,"case_id":"typed:35/r0","base_id":"typed:35","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":208.3467}
{"id":"outcome","type":"choice","probabilities":{"failure":1.354504310651531e-05,"harmful":7.904529493937181e-06,"partial":0.0003266585267994077,"success":0.9996518919006016},"predicted":"success","confidence":0.9996518919006016,"entropy_confidence":0.9976810846382395,"candidate_mass":0.9999952502674995,"accepted":true,"accepted_predicted":"success","input_tokens":234,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00034816870308771244,"brier_hard":2.281309917768367e-07,"soft_accuracy":0.5665851790612633,"brier_soft":0.31615625158766825,"tvd":0.4329848919006008,"kl":2.8350565852891565,"case_id":"typed:35/r0","base_id":"typed:35","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":208.3467}
{"id":"risk","type":"ordinal","probabilities":{"0":0.992362271865273,"1":0.00726158575498494,"2":0.00024065317293838569,"3":0.000135489206802919},"predicted":"0","confidence":0.992362271865273,"entropy_confidence":0.966396136591361,"candidate_mass":0.9999941773222486,"accepted":true,"accepted_predicted":"0","input_tokens":242,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.007667044951574548,"brier_hard":0.0001111417900118054,"score_mae":0.5951836402787295,"within_one":true,"case_id":"typed:35/r0","base_id":"typed:35","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":208.3467}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002445597899868649,"1":0.9063603546156097,"2":0.0932443231079849,"3":0.00015076248641843041},"predicted":"1","confidence":0.9063603546156097,"entropy_confidence":0.7737151440418018,"candidate_mass":0.9999898944586949,"accepted":true,"accepted_predicted":"1","input_tokens":230,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.316050739120758,"brier_hard":1.8296945591696445,"score_mae":0.4033012882908349,"within_one":true,"case_id":"typed:35/r0","base_id":"typed:35","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":208.3467}
{"id":"action","type":"choice","probabilities":{"continue":0.0013003521263454279,"human_review":0.002787171098560121,"observe":0.9955357904015012,"stop":0.0003766863735920219},"predicted":"observe","confidence":0.9955357904015012,"entropy_confidence":0.9765841734475674,"candidate_mass":0.9999644617283758,"accepted":true,"accepted_predicted":"observe","input_tokens":220,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.882728140705389,"brier_hard":1.9855267689042306,"soft_accuracy":0.3167340023579466,"brier_soft":0.6898554841634577,"tvd":0.6788687904015018,"kl":3.0873096923448484,"case_id":"typed:36/r0","base_id":"typed:36","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.635601}
{"id":"needs_review","type":"binary","probabilities":{"false":1.1886047866497753e-07,"true":0.999999881139523},"predicted":"true","confidence":0.999999881139523,"entropy_confidence":0.9999970942270894,"candidate_mass":0.9999996487521279,"accepted":true,"accepted_predicted":"true","input_tokens":182,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.1886048408027632e-07,"brier_hard":2.8255626385026426e-14,"soft_accuracy":0.6833329564179048,"brier_soft":0.20055582722126453,"tvd":0.3166668811395222,"kl":2.656416190780962,"case_id":"typed:36/r0","base_id":"typed:36","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.635601}
{"id":"outcome","type":"choice","probabilities":{"failure":1.958205232909873e-05,"harmful":1.0180216457799663e-05,"partial":0.00021675759834614237,"success":0.9997534801328665},"predicted":"success","confidence":0.9997534801328665,"entropy_confidence":0.9982655023867084,"candidate_mass":0.9999956256125524,"accepted":true,"accepted_predicted":"success","input_tokens":225,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":8.436730887111178,"brier_hard":1.9990735533120354,"soft_accuracy":0.20005509066414426,"brier_soft":0.9643968871804395,"tvd":0.7997636603493246,"kl":6.338981293658507,"case_id":"typed:36/r0","base_id":"typed:36","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.635601}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9966314955044004,"1":0.0031335376720345543,"2":0.00014245267530189085,"3":9.251414826248308e-05},"predicted":"0","confidence":0.9966314955044004,"entropy_confidence":0.9830119456615939,"candidate_mass":0.9999963533562682,"accepted":true,"accepted_predicted":"0","input_tokens":233,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0033741906797501472,"brier_hard":2.119473251126355e-05,"score_mae":0.39630401453257424,"within_one":true,"case_id":"typed:36/r0","base_id":"typed:36","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.635601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00019211825686077665,"1":0.9980328951959263,"2":0.0017123852743783693,"3":6.260127283439721e-05},"predicted":"1","confidence":0.9980328951959263,"entropy_confidence":0.9890912272968205,"candidate_mass":0.9999944393655714,"accepted":true,"accepted_predicted":"1","input_tokens":221,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.001969042095716165,"brier_hard":6.842592982097398e-06,"score_mae":0.034978469563186176,"within_one":true,"case_id":"typed:36/r0","base_id":"typed:36","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.635601}
{"id":"action","type":"choice","probabilities":{"continue":0.03055567651387952,"human_review":0.06505297597541228,"observe":0.8825239723649105,"stop":0.02186737514579848},"predicted":"observe","confidence":0.8825239723649105,"entropy_confidence":0.6550319047759354,"candidate_mass":0.9999353603603197,"accepted":true,"accepted_predicted":"observe","input_tokens":223,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":2.73255332602454,"brier_hard":1.6543863309941624,"soft_accuracy":0.23454592317981493,"brier_soft":0.7340445921413571,"tvd":0.6658569723649101,"kl":1.272665828625764,"case_id":"typed:37/r0","base_id":"typed:37","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.1829}
{"id":"needs_review","type":"binary","probabilities":{"false":2.4506424695716097e-08,"true":0.9999999754935768},"predicted":"true","confidence":0.9999999754935768,"entropy_confidence":0.9999993450667882,"candidate_mass":0.9999998344213056,"accepted":true,"accepted_predicted":"true","input_tokens":185,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.450642352432901e-08,"brier_hard":1.2011296306028952e-15,"soft_accuracy":0.7599999872566603,"brier_soft":0.11519997647383419,"tvd":0.23999997549357605,"kl":2.001909705165828,"case_id":"typed:37/r0","base_id":"typed:37","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.1829}
{"id":"outcome","type":"choice","probabilities":{"failure":6.5712275654257414e-06,"harmful":0.9999446192668252,"partial":4.528451455783927e-05,"success":3.5249910510800797e-06},"predicted":"harmful","confidence":0.9999446192668252,"entropy_confidence":0.9995448215520492,"candidate_mass":0.9999987174995812,"accepted":true,"accepted_predicted":"harmful","input_tokens":228,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":5.53822667442163e-05,"brier_hard":5.1733194593435165e-09,"soft_accuracy":0.5666486566493641,"brier_soft":0.27590348596424147,"tvd":0.4332776192668254,"kl":3.528104944277405,"case_id":"typed:37/r0","base_id":"typed:37","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.1829}
{"id":"risk","type":"ordinal","probabilities":{"0":6.634734185442181e-05,"1":3.542954988978306e-05,"2":0.8328086196294707,"3":0.16708960347878668},"predicted":"2","confidence":0.8328086196294707,"entropy_confidence":0.6737159048495196,"candidate_mass":0.9999934465421113,"accepted":true,"accepted_predicted":"2","input_tokens":236,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":1.7892250626785078,"brier_hard":1.387309931219492,"score_mae":0.2030785207548087,"within_one":true,"case_id":"typed:37/r0","base_id":"typed:37","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.1829}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009758700575351598,"1":0.0022253984137398535,"2":0.9476637629802903,"3":0.049134968548435104},"predicted":"2","confidence":0.9476637629802903,"entropy_confidence":0.8417708964774957,"candidate_mass":0.9999889224216126,"accepted":true,"accepted_predicted":"2","input_tokens":224,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.013184307273068,"brier_hard":1.8022168204238185,"score_mae":0.2717091699803742,"within_one":true,"case_id":"typed:37/r0","base_id":"typed:37","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":206.1829}
{"id":"action","type":"choice","probabilities":{"continue":0.0003360424244247341,"human_review":0.000929679902178277,"observe":0.998580740056465,"stop":0.00015353761693188522},"predicted":"observe","confidence":0.998580740056465,"entropy_confidence":0.991384152517219,"candidate_mass":0.9999715058635529,"accepted":true,"accepted_predicted":"observe","input_tokens":220,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0014202680468820038,"brier_hard":3.015101818663496e-06,"soft_accuracy":0.5162255255891001,"brier_soft":0.3286025218145485,"tvd":0.4819137400564651,"kl":2.5158618201803757,"case_id":"typed:38/r0","base_id":"typed:38","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9127}
{"id":"needs_review","type":"binary","probabilities":{"false":1.0354502150486704e-07,"true":0.9999998964549786},"predicted":"true","confidence":0.9999998964549786,"entropy_confidence":0.999997448036292,"candidate_mass":0.9999995207317991,"accepted":true,"accepted_predicted":"true","input_tokens":182,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.08325932843669,"brier_hard":1.9999995858199355,"soft_accuracy":0.36666702761193676,"brier_soft":0.8022211154641048,"tvd":0.6333328964549785,"kl":5.465335095754064,"case_id":"typed:38/r0","base_id":"typed:38","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9127}
{"id":"outcome","type":"choice","probabilities":{"failure":9.767869810191905e-06,"harmful":4.662174706446601e-06,"partial":0.00017429002517517885,"success":0.999811279930308},"predicted":"success","confidence":0.999811279930308,"entropy_confidence":0.9986531973199544,"candidate_mass":0.9999965256600477,"accepted":true,"accepted_predicted":"success","input_tokens":225,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00018873787956505782,"brier_hard":6.610942473372122e-08,"soft_accuracy":0.5166339898828215,"brier_soft":0.37988774580478535,"tvd":0.4831447965967915,"kl":3.526033111444345,"case_id":"typed:38/r0","base_id":"typed:38","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9127}
{"id":"risk","type":"ordinal","probabilities":{"0":0.998377130687842,"1":0.0014735910145774825,"2":0.00010219407463410578,"3":4.708422294551059e-05},"predicted":"0","confidence":0.998377130687842,"entropy_confidence":0.9908839024987296,"candidate_mass":0.9999975300124739,"accepted":true,"accepted_predicted":"0","input_tokens":233,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0016241875910161434,"brier_hard":4.817835835528535e-06,"score_mae":0.5081807681673178,"within_one":true,"case_id":"typed:38/r0","base_id":"typed:38","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9127}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004088657279698257,"1":0.7773021412860142,"2":0.2218012558998151,"3":0.00048773708620006013},"predicted":"1","confidence":0.7773021412860142,"entropy_confidence":0.6128102756864929,"candidate_mass":0.999986833932789,"accepted":false,"accepted_predicted":null,"input_tokens":221,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":null,"nll_hard":7.80212374929388,"brier_hard":1.652577089569267,"score_mae":0.40236786434424465,"within_one":true,"case_id":"typed:38/r0","base_id":"typed:38","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9127}
{"id":"action","type":"choice","probabilities":{"continue":0.00038127594297323285,"human_review":0.0014172038393540423,"observe":0.997905224908027,"stop":0.000296295309644578},"predicted":"observe","confidence":0.997905224908027,"entropy_confidence":0.9878837666268775,"candidate_mass":0.9999679067827327,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.559069475864263,"brier_hard":1.9929826718490093,"soft_accuracy":0.13413545243321098,"brier_soft":1.2942133746621287,"tvd":0.8645720915748942,"kl":5.045348548961593,"case_id":"typed:39/r0","base_id":"typed:39","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.858501}
{"id":"needs_review","type":"binary","probabilities":{"false":1.352351331409607e-08,"true":0.9999999864764854},"predicted":"true","confidence":0.9999999864764854,"entropy_confidence":0.9999996269856665,"candidate_mass":0.9999998963133994,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.3523514732898573e-08,"brier_hard":3.6577086061420436e-16,"soft_accuracy":0.8999999891811882,"brier_soft":0.019999994590594768,"tvd":0.099999986476486,"kl":0.8262095852767377,"case_id":"typed:39/r0","base_id":"typed:39","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.858501}
{"id":"outcome","type":"choice","probabilities":{"failure":2.4614588676932845e-05,"harmful":0.9997720157720076,"partial":0.00019177038500993774,"success":1.1599254303953707e-05},"predicted":"harmful","confidence":0.9997720157720076,"entropy_confidence":0.9983680260045901,"candidate_mass":0.999996765762227,"accepted":true,"accepted_predicted":"harmful","input_tokens":214,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00022801022034716346,"brier_hard":8.949310945630057e-08,"soft_accuracy":0.5332619976646544,"brier_soft":0.3130203523748892,"tvd":0.4664384824384751,"kl":3.388437045344789,"case_id":"typed:39/r0","base_id":"typed:39","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.858501}
{"id":"risk","type":"ordinal","probabilities":{"0":7.347888688412992e-06,"1":4.9356849446053326e-06,"2":0.22445484927594486,"3":0.7755328671504212},"predicted":"3","confidence":0.7755328671504212,"entropy_confidence":0.6157777803524003,"candidate_mass":0.9999965170192457,"accepted":false,"accepted_predicted":null,"input_tokens":222,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.2542049153804306,"brier_hard":0.10076547317155002,"score_mae":0.3088462356880979,"within_one":true,"case_id":"typed:39/r0","base_id":"typed:39","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.858501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001550217461543428,"1":0.003838618434865857,"2":0.9648234197128813,"3":0.02978774439071038},"predicted":"2","confidence":0.9648234197128813,"entropy_confidence":0.8769408319385912,"candidate_mass":0.9999927151258308,"accepted":true,"accepted_predicted":"2","input_tokens":210,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.03581017914039458,"brier_hard":0.00214183968224897,"score_mae":0.04381830896724015,"within_one":true,"case_id":"typed:39/r0","base_id":"typed:39","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.858501}
{"id":"action","type":"choice","probabilities":{"continue":0.0009487075118296215,"human_review":0.0016963778820110115,"observe":0.9970128984920101,"stop":0.0003420161141504907},"predicted":"observe","confidence":0.9970128984920101,"entropy_confidence":0.98331005527435,"candidate_mass":0.999948930627558,"accepted":true,"accepted_predicted":"observe","input_tokens":215,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.379259958640883,"brier_hard":1.9906458587143012,"soft_accuracy":0.2832890361370206,"brier_soft":0.714682819982282,"tvd":0.7136798984920095,"kl":3.7290859344118155,"case_id":"typed:40/r0","base_id":"typed:40","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.009}
{"id":"needs_review","type":"binary","probabilities":{"false":1.7144819766048346e-07,"true":0.9999998285518008},"predicted":"true","confidence":0.9999998285518008,"entropy_confidence":0.9999958992301962,"candidate_mass":0.9999995355812116,"accepted":true,"accepted_predicted":"true","input_tokens":177,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7144821389048015e-07,"brier_hard":5.878896948763172e-14,"soft_accuracy":0.7666669085608457,"brier_soft":0.10888841775996885,"tvd":0.2333328285518016,"kl":1.9453709658283862,"case_id":"typed:40/r0","base_id":"typed:40","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.009}
{"id":"outcome","type":"choice","probabilities":{"failure":1.3633851290763943e-05,"harmful":8.989828945147979e-06,"partial":0.00015238342450213922,"success":0.9998249928952615},"predicted":"success","confidence":0.9998249928952615,"entropy_confidence":0.998722131018733,"candidate_mass":0.9999963428779719,"accepted":true,"accepted_predicted":"success","input_tokens":220,"reused_prefix_tokens":0,"expected":"failure","correct":false,"accepted_correct":false,"nll_hard":11.20295479231773,"brier_hard":1.9996227722028352,"soft_accuracy":0.04337405643092412,"brier_soft":1.4660136425995687,"tvd":0.9564919928952618,"kl":8.207062642034408,"case_id":"typed:40/r0","base_id":"typed:40","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.009}
{"id":"risk","type":"ordinal","probabilities":{"0":0.854199229505149,"1":0.14145167326524855,"2":0.002115975547366396,"3":0.0022331216822359225},"predicted":"0","confidence":0.854199229505149,"entropy_confidence":0.6841023074690451,"candidate_mass":0.9999723741110187,"accepted":true,"accepted_predicted":"0","input_tokens":228,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.1575908225827298,"brier_hard":0.04127590473139556,"score_mae":0.2709500105933109,"within_one":true,"case_id":"typed:40/r0","base_id":"typed:40","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.009}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000636285462058782,"1":0.9051702444665086,"2":0.09368520572067833,"3":0.000508264350755345},"predicted":"1","confidence":0.9051702444665086,"entropy_confidence":0.7687711283328686,"candidate_mass":0.9999842502476247,"accepted":true,"accepted_predicted":"1","input_tokens":216,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.09963223754109406,"brier_hard":0.017770263497307034,"score_mae":0.17260155103986863,"within_one":true,"case_id":"typed:40/r0","base_id":"typed:40","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.009}
{"id":"action","type":"choice","probabilities":{"continue":0.0011974249855390459,"human_review":0.0021532139432430186,"observe":0.9963380409376372,"stop":0.0003113201335818013},"predicted":"observe","confidence":0.9963380409376372,"entropy_confidence":0.980200975280601,"candidate_mass":0.9999716331586671,"accepted":true,"accepted_predicted":"observe","input_tokens":229,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.727581873197451,"brier_hard":1.9903008089254777,"soft_accuracy":0.33300621081847015,"brier_soft":0.7689056150376155,"tvd":0.6633163610712185,"kl":3.509280849214026,"case_id":"typed:41/r0","base_id":"typed:41","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.99200000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":9.028506700059395e-08,"true":0.999999909714934},"predicted":"true","confidence":0.999999909714934,"entropy_confidence":0.9999977569909905,"candidate_mass":0.999999753640017,"accepted":true,"accepted_predicted":"true","input_tokens":191,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.220293761177786,"brier_hard":1.9999996388597503,"soft_accuracy":0.1833330571806028,"brier_soft":1.3338896828466789,"tvd":0.8166669097149335,"kl":7.210765586201957,"case_id":"typed:41/r0","base_id":"typed:41","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.99200000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":2.8929191731454315e-05,"harmful":1.193646770691758e-05,"partial":0.0009643813866971108,"success":0.9989947529538655},"predicted":"success","confidence":0.9989947529538655,"entropy_confidence":0.9941289029650523,"candidate_mass":0.9999930653333139,"accepted":true,"accepted_predicted":"success","input_tokens":234,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0010057526458098062,"brier_hard":1.941532460165477e-06,"soft_accuracy":0.616338822492145,"brier_soft":0.24253638023390106,"tvd":0.3823396894215719,"kl":2.088332297237995,"case_id":"typed:41/r0","base_id":"typed:41","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.99200000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9891180165709256,"1":0.009868519251623197,"2":0.0006600747773541192,"3":0.00035338940009650667},"predicted":"0","confidence":0.9891180165709256,"entropy_confidence":0.9538035117453518,"candidate_mass":0.9999925449125691,"accepted":true,"accepted_predicted":"0","input_tokens":242,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.010941625286548639,"brier_hard":0.00021636581835010775,"score_mae":0.32108416299337905,"within_one":true,"case_id":"typed:41/r0","base_id":"typed:41","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.99200000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005384667920630032,"1":0.9974960963346303,"2":0.0018881198839351984,"3":7.73169893705788e-05},"predicted":"1","confidence":0.9974960963346303,"entropy_confidence":0.986201839641078,"candidate_mass":0.9999957152008991,"accepted":true,"accepted_predicted":"1","input_tokens":230,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.526784730747522,"brier_hard":1.993925389539799,"score_mae":0.5681712870706126,"within_one":true,"case_id":"typed:41/r0","base_id":"typed:41","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.99200000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.0018178701896716812,"human_review":0.003549874546927082,"observe":0.9942482057381269,"stop":0.0003840495252742562},"predicted":"observe","confidence":0.9942482057381269,"entropy_confidence":0.9709651281195038,"candidate_mass":0.9999575739959669,"accepted":true,"accepted_predicted":"observe","input_tokens":217,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.31008968859701,"brier_hard":1.9849098079895044,"soft_accuracy":0.28353070558640764,"brier_soft":0.7492616927520328,"tvd":0.7109152057381269,"kl":3.180593765869849,"case_id":"typed:42/r0","base_id":"typed:42","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.780801}
{"id":"needs_review","type":"binary","probabilities":{"false":1.8556874523493488e-08,"true":0.9999999814431251},"predicted":"true","confidence":0.9999999814431251,"entropy_confidence":0.9999994966231398,"candidate_mass":0.9999999046951028,"accepted":true,"accepted_predicted":"true","input_tokens":179,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.80242552228765,"brier_hard":1.9999999257725019,"soft_accuracy":0.4000000037113748,"brier_soft":0.7199999554635014,"tvd":0.5999999814431253,"kl":5.159687937858798,"case_id":"typed:42/r0","base_id":"typed:42","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.780801}
{"id":"outcome","type":"choice","probabilities":{"failure":1.460527833039797e-05,"harmful":8.23527623966154e-06,"partial":0.00015924288504188456,"success":0.9998179165603869},"predicted":"success","confidence":0.9998179165603869,"entropy_confidence":0.9986772746303566,"candidate_mass":0.9999962925336323,"accepted":true,"accepted_predicted":"success","input_tokens":222,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00018210001881517653,"brier_hard":5.879380934766069e-08,"soft_accuracy":0.4166406321873134,"brier_soft":0.5102436053179562,"tvd":0.5831591518366271,"kl":4.51884869816549,"case_id":"typed:42/r0","base_id":"typed:42","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.780801}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9968447656776567,"1":0.00280992017736344,"2":0.00021920932790200405,"3":0.0001261048170771455},"predicted":"0","confidence":0.9968447656776567,"entropy_confidence":0.9836711664669575,"candidate_mass":0.9999950489508788,"accepted":true,"accepted_predicted":"0","input_tokens":230,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00316022256964741,"brier_hard":1.7915110186376838e-05,"score_mae":0.3930403467156011,"within_one":true,"case_id":"typed:42/r0","base_id":"typed:42","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.780801}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022554383383131975,"1":0.9988586153018603,"2":0.0008620981867238737,"3":5.374267758361733e-05},"predicted":"1","confidence":0.9988586153018603,"entropy_confidence":0.9930418306029047,"candidate_mass":0.9999950443558216,"accepted":true,"accepted_predicted":"1","input_tokens":218,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.396996032850584,"brier_hard":1.9972682426666668,"score_mae":0.5274110397080589,"within_one":true,"case_id":"typed:42/r0","base_id":"typed:42","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.780801}
{"id":"action","type":"choice","probabilities":{"continue":0.009754993003342984,"human_review":0.02682555440263289,"observe":0.9593696159320518,"stop":0.004049836661970914},"predicted":"observe","confidence":0.9593696159320518,"entropy_confidence":0.852603161765408,"candidate_mass":0.9999444079110893,"accepted":true,"accepted_predicted":"observe","input_tokens":208,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.618400323292814,"brier_hard":1.8675701226028394,"soft_accuracy":0.22498753850877395,"brier_soft":0.8790237099465571,"tvd":0.7427026159320524,"kl":2.015462846193637,"case_id":"typed:43/r0","base_id":"typed:43","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.9652}
{"id":"needs_review","type":"binary","probabilities":{"false":8.912472791934435e-08,"true":0.9999999108752715},"predicted":"true","confidence":0.9999999108752715,"entropy_confidence":0.9999977841548138,"candidate_mass":0.9999996516924166,"accepted":true,"accepted_predicted":"true","input_tokens":170,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.912473245782063e-08,"brier_hard":1.5886434354438127e-14,"soft_accuracy":0.6999999643501084,"brier_soft":0.17999989305034209,"tvd":0.29999991087527184,"kl":2.513429713223055,"case_id":"typed:43/r0","base_id":"typed:43","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.9652}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00041412496261115174,"harmful":0.9789390281440747,"partial":0.0193168180980798,"success":0.0013300287952327599},"predicted":"harmful","confidence":0.9789390281440747,"entropy_confidence":0.9212932088175096,"candidate_mass":0.9999786427427709,"accepted":true,"accepted_predicted":"harmful","input_tokens":213,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":3.9467791584777956,"brier_hard":1.920063064565021,"soft_accuracy":0.09007791048744208,"brier_soft":1.0846522353422965,"tvd":0.8956060281440755,"kl":3.9659051036136646,"case_id":"typed:43/r0","base_id":"typed:43","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.9652}
{"id":"risk","type":"ordinal","probabilities":{"0":1.871454106303237e-06,"1":2.5323815489202874e-06,"2":0.9999735548744466,"3":2.204128989766951e-05},"predicted":"2","confidence":0.9999735548744466,"entropy_confidence":0.9997690967034893,"candidate_mass":0.9999991368452464,"accepted":true,"accepted_predicted":"2","input_tokens":221,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":2.6445475231884713e-05,"brier_hard":1.1950784226688252e-09,"score_mae":0.30001576600013524,"within_one":true,"case_id":"typed:43/r0","base_id":"typed:43","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.9652}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010449590383932077,"1":0.001833509039976322,"2":0.39765628365102096,"3":0.5994652482706104},"predicted":"3","confidence":0.5994652482706104,"entropy_confidence":0.5006914743649368,"candidate_mass":0.9999913311892117,"accepted":false,"accepted_predicted":null,"input_tokens":209,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.9221672557397511,"brier_hard":0.7221809902040356,"score_mae":0.8455418211538497,"within_one":true,"case_id":"typed:43/r0","base_id":"typed:43","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.9652}
{"id":"action","type":"choice","probabilities":{"continue":0.00438253053537549,"human_review":0.014450019490293693,"observe":0.9574125919798031,"stop":0.023754857994527687},"predicted":"observe","confidence":0.9574125919798031,"entropy_confidence":0.8445257166562431,"candidate_mass":0.9999335732903196,"accepted":true,"accepted_predicted":"observe","input_tokens":218,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.237059515615353,"brier_hard":1.8885311352164011,"soft_accuracy":0.2345785362379042,"brier_soft":0.89438545727718,"tvd":0.7240795919798031,"kl":2.271390704761491,"case_id":"typed:44/r0","base_id":"typed:44","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.8961}
{"id":"needs_review","type":"binary","probabilities":{"false":9.9384123522002e-08,"true":0.9999999006158773},"predicted":"true","confidence":0.9999999006158773,"entropy_confidence":0.9999975447048611,"candidate_mass":0.9999996210029857,"accepted":true,"accepted_predicted":"true","input_tokens":180,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.938412764786958e-08,"brier_hard":1.9754407854927255e-14,"soft_accuracy":0.7166669569334808,"brier_soft":0.16055506514281268,"tvd":0.2833329006158768,"kl":2.370840142111595,"case_id":"typed:44/r0","base_id":"typed:44","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.8961}
{"id":"outcome","type":"choice","probabilities":{"failure":0.008203766329867128,"harmful":0.24138405167688184,"partial":0.6566506749980577,"success":0.093761506995193},"predicted":"partial","confidence":0.6566506749980577,"entropy_confidence":0.36476551120588463,"candidate_mass":0.9999471945715509,"accepted":false,"accepted_predicted":null,"input_tokens":223,"reused_prefix_tokens":0,"expected":"partial","correct":true,"accepted_correct":null,"nll_hard":0.4206030990095826,"brier_hard":0.1850135413592416,"soft_accuracy":0.31693356208771045,"brier_soft":0.19778130051366952,"tvd":0.43136826000847306,"kl":1.046384281626025,"case_id":"typed:44/r0","base_id":"typed:44","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.8961}
{"id":"risk","type":"ordinal","probabilities":{"0":2.1712013860725537e-05,"1":3.37896686289469e-05,"2":0.9998982947695503,"3":4.620354795891616e-05},"predicted":"2","confidence":0.9998982947695503,"entropy_confidence":0.9991748235990598,"candidate_mass":0.9999976779030919,"accepted":true,"accepted_predicted":"2","input_tokens":231,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00010171040277737759,"brier_hard":1.4091874996765371e-08,"score_mae":0.3999689898516061,"within_one":true,"case_id":"typed:44/r0","base_id":"typed:44","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.8961}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0012009563313161418,"1":0.005029427780079677,"2":0.9791041327644687,"3":0.014665483124133726},"predicted":"2","confidence":0.9791041327644687,"entropy_confidence":0.9153920958518047,"candidate_mass":0.9999909137436993,"accepted":true,"accepted_predicted":"2","input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.292449062757381,"brier_hard":1.9488278610714722,"score_mae":0.6905671426814182,"within_one":true,"case_id":"typed:44/r0","base_id":"typed:44","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.8961}
{"id":"action","type":"choice","probabilities":{"continue":0.0010576597343085054,"human_review":0.0022934230563706817,"observe":0.996296486838893,"stop":0.0003524303704281683},"predicted":"observe","confidence":0.996296486838893,"entropy_confidence":0.9800300542233483,"candidate_mass":0.9999590540107457,"accepted":true,"accepted_predicted":"observe","input_tokens":213,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0037103881456203988,"brier_hard":2.0218650329562725e-05,"soft_accuracy":0.6315171429010539,"brier_soft":0.18957929319332134,"tvd":0.3629628535052595,"kl":1.4548763052650988,"case_id":"typed:45/r0","base_id":"typed:45","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.10150099999998}
{"id":"needs_review","type":"binary","probabilities":{"false":5.578247319022226e-07,"true":0.9999994421752688},"predicted":"true","confidence":0.9999994421752688,"entropy_confidence":0.9999876071541166,"candidate_mass":0.9999992988708477,"accepted":true,"accepted_predicted":"true","input_tokens":175,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.399221024438656,"brier_hard":1.9999977687016963,"soft_accuracy":0.1833333532893689,"brier_soft":1.3338881555504223,"tvd":0.8166664421752685,"kl":7.210765671917434,"case_id":"typed:45/r0","base_id":"typed:45","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.10150099999998}
{"id":"outcome","type":"choice","probabilities":{"failure":1.3882327231925168e-05,"harmful":9.835138873765038e-06,"partial":0.0002054789412466327,"success":0.999770803592646},"predicted":"success","confidence":0.999770803592646,"entropy_confidence":0.9983824591636458,"candidate_mass":0.9999950609872497,"accepted":true,"accepted_predicted":"success","input_tokens":218,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0002292226768645379,"brier_hard":9.504203740584502e-08,"soft_accuracy":0.6498973938133199,"brier_soft":0.1869691923786898,"tvd":0.3497806387315206,"kl":2.4509720433705984,"case_id":"typed:45/r0","base_id":"typed:45","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.10150099999998}
{"id":"risk","type":"ordinal","probabilities":{"0":0.012194354214140229,"1":0.9873133179757724,"2":0.00031912527801295344,"3":0.00017320253207397146},"predicted":"1","confidence":0.9873133179757724,"entropy_confidence":0.9492079175792542,"candidate_mass":0.9999930041795329,"accepted":true,"accepted_predicted":"1","input_tokens":226,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.01276784516596053,"brier_hard":0.00030978601554396357,"score_mae":0.25513817612802014,"within_one":true,"case_id":"typed:45/r0","base_id":"typed:45","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.10150099999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00014851299150736816,"1":0.9992252538255082,"2":0.0005812544992716537,"3":4.4978683713927714e-05},"predicted":"1","confidence":0.9992252538255082,"entropy_confidence":0.9950484519179098,"candidate_mass":0.9999946707775075,"accepted":true,"accepted_predicted":"1","input_tokens":214,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.814838118649998,"brier_hard":1.99815444383562,"score_mae":0.8171896988751932,"within_one":true,"case_id":"typed:45/r0","base_id":"typed:45","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.10150099999998}
{"id":"action","type":"choice","probabilities":{"continue":0.017675228730323015,"human_review":0.006815902587513864,"observe":0.9750852181094217,"stop":0.00042365057274132444},"predicted":"observe","confidence":0.9750852181094217,"entropy_confidence":0.9038997547964458,"candidate_mass":0.9999657186043452,"accepted":true,"accepted_predicted":"observe","input_tokens":222,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":4.035591126388013,"brier_hard":1.9157997748334121,"soft_accuracy":0.17678820100849166,"brier_soft":1.2560845458330747,"tvd":0.8084182181094217,"kl":2.7920511860054984,"case_id":"typed:46/r0","base_id":"typed:46","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.298}
{"id":"needs_review","type":"binary","probabilities":{"false":2.830951589759138e-07,"true":0.9999997169048395},"predicted":"true","confidence":0.9999997169048395,"entropy_confidence":0.9999934336348429,"candidate_mass":0.9999995574034222,"accepted":true,"accepted_predicted":"true","input_tokens":184,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.077482745024522,"brier_hard":1.9999988676195213,"soft_accuracy":0.09666722836323936,"brier_soft":1.6320199948613605,"tvd":0.9033327169048402,"kl":8.094143605457944,"case_id":"typed:46/r0","base_id":"typed:46","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.298}
{"id":"outcome","type":"choice","probabilities":{"failure":2.1689728941883323e-06,"harmful":1.1221320030282832e-06,"partial":1.760802364358978e-05,"success":0.9999791008714577},"predicted":"success","confidence":0.9999791008714577,"entropy_confidence":0.9998143854926544,"candidate_mass":0.9999989148970155,"accepted":true,"accepted_predicted":"success","input_tokens":227,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":2.089934693213627e-05,"brier_hard":7.527796941090141e-10,"soft_accuracy":0.9133151738688691,"brier_soft":0.012549472535956837,"tvd":0.08664610087145846,"kl":0.6396688986662151,"case_id":"typed:46/r0","base_id":"typed:46","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.298}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9939161468469835,"1":0.005796891730662982,"2":0.0001766629821502262,"3":0.00011029844020349542},"predicted":"0","confidence":0.9939161468469835,"entropy_confidence":0.9722616772250483,"candidate_mass":0.9999956724272135,"accepted":true,"accepted_predicted":"0","input_tokens":235,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.006102435192879375,"brier_hard":7.06605984796714e-05,"score_mae":0.08685188698442609,"within_one":true,"case_id":"typed:46/r0","base_id":"typed:46","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.298}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00014201334886107707,"1":0.9996816014816985,"2":0.00016261101406366932,"3":1.377415537620834e-05},"predicted":"1","confidence":0.9996816014816985,"entropy_confidence":0.9977282307995726,"candidate_mass":0.9999985946504387,"accepted":true,"accepted_predicted":"1","input_tokens":223,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.859589498717384,"brier_hard":1.9990793244431517,"score_mae":0.8167151459759545,"within_one":true,"case_id":"typed:46/r0","base_id":"typed:46","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.298}
{"id":"action","type":"choice","probabilities":{"continue":0.0017944536201614843,"human_review":0.0016963867675174082,"observe":0.9963166546892592,"stop":0.00019250492306054296},"predicted":"observe","confidence":0.9963166546892592,"entropy_confidence":0.9801689883695986,"candidate_mass":0.9999714990372756,"accepted":true,"accepted_predicted":"observe","input_tokens":230,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0036901455305873537,"brier_hard":1.970188268347739e-05,"soft_accuracy":0.5818857596869381,"brier_soft":0.24943666966532582,"tvd":0.4129836546892599,"kl":1.6587770805426485,"case_id":"typed:47/r0","base_id":"typed:47","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":205.4297}
{"id":"needs_review","type":"binary","probabilities":{"false":4.0038828849721824e-08,"true":0.9999999599611699},"predicted":"true","confidence":0.9999999599611699,"entropy_confidence":0.999998958320999,"candidate_mass":0.9999997174580928,"accepted":true,"accepted_predicted":"true","input_tokens":192,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.033416132434613,"brier_hard":1.9999998398446852,"soft_accuracy":0.2500000200194141,"brier_soft":1.1249998798835148,"tvd":0.7499999599611705,"kl":6.561181698711873,"case_id":"typed:47/r0","base_id":"typed:47","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":205.4297}
{"id":"outcome","type":"choice","probabilities":{"failure":1.101593486574329e-05,"harmful":6.90858171157502e-06,"partial":0.0003963343427091111,"success":0.9995857411407137},"predicted":"success","confidence":0.9995857411407137,"entropy_confidence":0.9973118171925355,"candidate_mass":0.9999955244274158,"accepted":true,"accepted_predicted":"success","input_tokens":235,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00041434468819191066,"brier_hard":3.288603930300433e-07,"soft_accuracy":0.6764768816336928,"brier_soft":0.16361761974177388,"tvd":0.322919417807037,"kl":2.0316481818882925,"case_id":"typed:47/r0","base_id":"typed:47","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":205.4297}
{"id":"risk","type":"ordinal","probabilities":{"0":0.8502008436082613,"1":0.1399403372060296,"2":0.008267577839503496,"3":0.0015912413462051372},"predicted":"0","confidence":0.8502008436082613,"entropy_confidence":0.6659654471239835,"candidate_mass":0.9999779597223281,"accepted":true,"accepted_predicted":"0","input_tokens":243,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.16228267081110198,"brier_hard":0.04209397012536798,"score_mae":0.395417783076348,"within_one":true,"case_id":"typed:47/r0","base_id":"typed:47","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":205.4297}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004925003752099733,"1":0.8449201110130361,"2":0.1541378057661234,"3":0.00044958284562990363},"predicted":"1","confidence":0.8449201110130361,"entropy_confidence":0.684179878019662,"candidate_mass":0.999988570927374,"accepted":true,"accepted_predicted":"1","input_tokens":231,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.616015335504759,"brier_hard":1.7366639010916112,"score_mae":0.4045444710821726,"within_one":true,"case_id":"typed:47/r0","base_id":"typed:47","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":205.4297}
{"id":"action","type":"choice","probabilities":{"continue":0.00043499507466570296,"human_review":0.0009783507364556523,"observe":0.9983604233975962,"stop":0.00022623079128337328},"predicted":"observe","confidence":0.9983604233975962,"entropy_confidence":0.9901292523116035,"candidate_mass":0.9999671244594792,"accepted":true,"accepted_predicted":"observe","input_tokens":213,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.001640922179106528,"brier_hard":3.885782684581328e-06,"soft_accuracy":0.41630307801829164,"brier_soft":0.5302301320972937,"tvd":0.5816934233975958,"kl":3.3179222255380236,"case_id":"typed:48/r0","base_id":"typed:48","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.2908}
{"id":"needs_review","type":"binary","probabilities":{"false":5.647076161493848e-07,"true":0.9999994352923838},"predicted":"true","confidence":0.9999994352923838,"entropy_confidence":0.9999874642322646,"candidate_mass":0.9999990226541315,"accepted":true,"accepted_predicted":"true","input_tokens":175,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.386957733286877,"brier_hard":1.999997741170173,"soft_accuracy":0.18333335764853334,"brier_soft":1.3338881330663388,"tvd":0.8166664352923838,"kl":7.210765673179294,"case_id":"typed:48/r0","base_id":"typed:48","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.2908}
{"id":"outcome","type":"choice","probabilities":{"failure":3.7997441096339314e-05,"harmful":2.3112952906941906e-05,"partial":0.0007433399950804922,"success":0.9991955496109156},"predicted":"success","confidence":0.9991955496109156,"entropy_confidence":0.9950999695583214,"candidate_mass":0.9999902461407669,"accepted":true,"accepted_predicted":"success","input_tokens":218,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0008047741329341918,"brier_hard":1.2016727909062374e-06,"soft_accuracy":0.49989891054759966,"brier_soft":0.4185944797994229,"tvd":0.4992186625638229,"kl":2.9565959768356334,"case_id":"typed:48/r0","base_id":"typed:48","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.2908}
{"id":"risk","type":"ordinal","probabilities":{"0":0.5433421329847696,"1":0.45185595153249924,"2":0.0027146968413164787,"3":0.0020872186414163454},"predicted":"0","confidence":0.5433421329847696,"entropy_confidence":0.48111913431860587,"candidate_mass":0.9999633817786726,"accepted":false,"accepted_predicted":null,"input_tokens":226,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.7943918412425504,"brier_hard":0.5956942974073782,"score_mae":0.4364529988606188,"within_one":true,"case_id":"typed:48/r0","base_id":"typed:48","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.2908}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008385368348470596,"1":0.39201843045801954,"2":0.6057079665222369,"3":0.0014350661848976899},"predicted":"2","confidence":0.6057079665222369,"entropy_confidence":0.5050725228274745,"candidate_mass":0.9999811209305178,"accepted":false,"accepted_predicted":null,"input_tokens":214,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.9364464238223746,"brier_hard":0.7365264921702117,"score_mae":0.5910725620571864,"within_one":true,"case_id":"typed:48/r0","base_id":"typed:48","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.2908}
{"id":"action","type":"choice","probabilities":{"continue":0.006563884725917884,"human_review":0.011945386342696457,"observe":0.9796318766784519,"stop":0.0018588522529337885},"predicted":"observe","confidence":0.9796318766784519,"entropy_confidence":0.9150788604386699,"candidate_mass":0.9999533126714368,"accepted":true,"accepted_predicted":"observe","input_tokens":211,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.02057841392880638,"brier_hard":0.0006040926169114472,"soft_accuracy":0.47753061984214196,"brier_soft":0.3731401096231696,"tvd":0.4962983933449685,"kl":1.4434890278484065,"case_id":"typed:49/r0","base_id":"typed:49","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.090901}
{"id":"needs_review","type":"binary","probabilities":{"false":2.0396257573708405e-07,"true":0.9999997960374246},"predicted":"true","confidence":0.9999997960374246,"entropy_confidence":0.9999951726367862,"candidate_mass":0.9999994985339115,"accepted":true,"accepted_predicted":"true","input_tokens":173,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.405329312200816,"brier_hard":1.999999184149781,"soft_accuracy":0.3433330639084098,"brier_soft":0.8624225620361128,"tvd":0.6566667960374244,"kl":5.681084977083111,"case_id":"typed:49/r0","base_id":"typed:49","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.090901}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0007793071587715856,"harmful":0.0004871564111842636,"partial":0.03319872794394817,"success":0.9655348084860956},"predicted":"success","confidence":0.9655348084860956,"entropy_confidence":0.8873203436279353,"candidate_mass":0.9999664680215505,"accepted":true,"accepted_predicted":"success","input_tokens":216,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.035073125473331884,"brier_hard":0.002290849604203061,"soft_accuracy":0.720850277306189,"brier_soft":0.07794878974201629,"tvd":0.2255348084860958,"kl":0.4154544272422662,"case_id":"typed:49/r0","base_id":"typed:49","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.090901}
{"id":"risk","type":"ordinal","probabilities":{"0":2.931961386579569e-06,"1":4.087358199457698e-06,"2":0.9999815368382747,"3":1.144384213938893e-05},"predicted":"2","confidence":0.9999815368382747,"entropy_confidence":0.9998292290759132,"candidate_mass":0.9999991193348313,"accepted":true,"accepted_predicted":"2","input_tokens":224,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":1.846333217153684e-05,"brier_hard":4.971527584277378e-10,"score_mae":0.406668492561167,"within_one":true,"case_id":"typed:49/r0","base_id":"typed:49","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.090901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0015986684302412668,"1":0.9883883921801919,"2":0.009767959348394504,"3":0.0002449800411713108},"predicted":"1","confidence":0.9883883921801919,"entropy_confidence":0.9501647107768367,"candidate_mass":0.999994135142131,"accepted":true,"accepted_predicted":"1","input_tokens":212,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.438584227332168,"brier_hard":1.9738123057218646,"score_mae":0.3219922510004948,"within_one":true,"case_id":"typed:49/r0","base_id":"typed:49","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.090901}
{"id":"action","type":"choice","probabilities":{"continue":0.004246320300410864,"human_review":0.03250010689841477,"observe":0.9616812995185615,"stop":0.0015722732826131103},"predicted":"observe","confidence":0.9616812995185615,"entropy_confidence":0.8685138495161222,"candidate_mass":0.9999462581173196,"accepted":true,"accepted_predicted":"observe","input_tokens":203,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.426511900469862,"brier_hard":1.860907468274657,"soft_accuracy":0.399800813605883,"brier_soft":0.5013060548597206,"tvd":0.5616812995185614,"kl":1.3206012313450008,"case_id":"typed:50/r0","base_id":"typed:50","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.53379999999999}
{"id":"needs_review","type":"binary","probabilities":{"false":8.80236451877908e-08,"true":0.9999999119763533},"predicted":"true","confidence":0.9999999119763533,"entropy_confidence":0.9999978099515816,"candidate_mass":0.9999997336745564,"accepted":true,"accepted_predicted":"true","input_tokens":165,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.80236505789859e-08,"brier_hard":1.5496324491375926e-14,"soft_accuracy":0.5999999823952701,"brier_soft":0.319999859162182,"tvd":0.3999999119763541,"kl":3.3776408273450698,"case_id":"typed:50/r0","base_id":"typed:50","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.53379999999999}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00040512697161064015,"harmful":0.00017047757393200437,"partial":0.05302999971428221,"success":0.9463943957401736},"predicted":"success","confidence":0.9463943957401736,"entropy_confidence":0.8466920812689848,"candidate_mass":0.9999482765935479,"accepted":true,"accepted_predicted":"success","input_tokens":208,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":2.9368974932209504,"brier_hard":1.792414726920007,"soft_accuracy":0.21942375948779472,"brier_soft":0.875182985150982,"tvd":0.7465648733141064,"kl":2.5145865436130572,"case_id":"typed:50/r0","base_id":"typed:50","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.53379999999999}
{"id":"risk","type":"ordinal","probabilities":{"0":1.1389441925342234e-05,"1":1.7763583826155595e-05,"2":0.9998361469321028,"3":0.00013470004214610943},"predicted":"2","confidence":0.9998361469321028,"entropy_confidence":0.9987821512255834,"candidate_mass":0.9999968923389886,"accepted":true,"accepted_predicted":"2","input_tokens":216,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001638664932777112,"brier_hard":4.543719351121915e-08,"score_mae":0.6000941575744703,"within_one":true,"case_id":"typed:50/r0","base_id":"typed:50","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.53379999999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0014795576375281113,"1":0.6774293508082806,"2":0.3171948346956987,"3":0.0038962568584920785},"predicted":"1","confidence":0.6774293508082806,"entropy_confidence":0.5244180304337029,"candidate_mass":0.999982834095403,"accepted":false,"accepted_predicted":null,"input_tokens":204,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.3894500109155104,"brier_hard":0.20468175678590905,"score_mae":0.32350779077515424,"within_one":true,"case_id":"typed:50/r0","base_id":"typed:50","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.53379999999999}
{"id":"action","type":"choice","probabilities":{"continue":0.0018375685465229145,"human_review":0.0022599504733192555,"observe":0.9956156762919655,"stop":0.00028680468819390376},"predicted":"observe","confidence":0.9956156762919655,"entropy_confidence":0.9768750060049038,"candidate_mass":0.9999634691506897,"accepted":true,"accepted_predicted":"observe","input_tokens":228,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.00439396304017595,"brier_hard":2.778858561103003e-05,"soft_accuracy":0.5319066620238259,"brier_soft":0.30911274712236847,"tvd":0.4622821429584314,"kl":1.8459891599434468,"case_id":"typed:51/r0","base_id":"typed:51","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3699}
{"id":"needs_review","type":"binary","probabilities":{"false":6.04813159520077e-08,"true":0.9999999395186858},"predicted":"true","confidence":0.9999999395186858,"entropy_confidence":0.9999984624663528,"candidate_mass":0.9999996966162995,"accepted":true,"accepted_predicted":"true","input_tokens":190,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.62093134684611,"brier_hard":1.999999758074747,"soft_accuracy":0.31666702217644277,"brier_soft":0.9338878124624934,"tvd":0.6833329395186849,"kl":5.92959271862006,"case_id":"typed:51/r0","base_id":"typed:51","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3699}
{"id":"outcome","type":"choice","probabilities":{"failure":1.7959126234608296e-05,"harmful":8.863913846247814e-06,"partial":0.0008922102578390412,"success":0.9990809667020811},"predicted":"success","confidence":0.9990809667020811,"entropy_confidence":0.9946022243441428,"candidate_mass":0.9999947484186131,"accepted":true,"accepted_predicted":"success","input_tokens":233,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":7.0218087380945065,"brier_hard":1.9963791539509304,"soft_accuracy":0.0803450036629355,"brier_soft":1.2522955049187374,"tvd":0.9190809667020806,"kl":6.879747675558826,"case_id":"typed:51/r0","base_id":"typed:51","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3699}
{"id":"risk","type":"ordinal","probabilities":{"0":0.7247932643347829,"1":0.2745286272423727,"2":0.00047303756612781275,"3":0.00020507085671756949},"predicted":"0","confidence":0.7247932643347829,"entropy_confidence":0.5718552429558054,"candidate_mass":0.9999890798577269,"accepted":false,"accepted_predicted":null,"input_tokens":241,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":null,"nll_hard":0.32186881743288603,"brier_hard":0.15110498034968156,"score_mae":0.003910085055219004,"within_one":true,"case_id":"typed:51/r0","base_id":"typed:51","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3699}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00042208358831274836,"1":0.8833459706572768,"2":0.11589209919019752,"3":0.000339846564214206},"predicted":"1","confidence":0.8833459706572768,"entropy_confidence":0.7364763040043298,"candidate_mass":0.9999863381665338,"accepted":true,"accepted_predicted":"1","input_tokens":229,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.770307186972342,"brier_hard":1.7928872090047745,"score_mae":0.6328167087303145,"within_one":true,"case_id":"typed:51/r0","base_id":"typed:51","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.3699}
{"id":"action","type":"choice","probabilities":{"continue":0.05871403780714339,"human_review":0.05381227314688692,"observe":0.8833615726019293,"stop":0.004112116444039025},"predicted":"observe","confidence":0.8833615726019293,"entropy_confidence":0.6711678032437808,"candidate_mass":0.9999381580568542,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.12402068019765976,"brier_hard":0.019964531224398185,"soft_accuracy":0.537285023960334,"brier_soft":0.13378463850818698,"tvd":0.3000279892683466,"kl":0.3108598331907899,"case_id":"typed:52/r0","base_id":"typed:52","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.00460099999998}
{"id":"needs_review","type":"binary","probabilities":{"false":7.145070707498029e-08,"true":0.9999999285492923},"predicted":"true","confidence":0.9999999285492923,"entropy_confidence":0.9999982007874934,"candidate_mass":0.9999996979456073,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.4542580365386,"brier_hard":1.9999997141971806,"soft_accuracy":0.1900000442994383,"brier_soft":1.3121997684997182,"tvd":0.8099999285492926,"kl":7.14483678558023,"case_id":"typed:52/r0","base_id":"typed:52","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.00460099999998}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0017229576684108156,"harmful":0.001463639194951216,"partial":0.042496263428894246,"success":0.9543171397077427},"predicted":"success","confidence":0.9543171397077427,"entropy_confidence":0.8561935444156292,"candidate_mass":0.9999599857624066,"accepted":true,"accepted_predicted":"success","input_tokens":211,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.046759231217499996,"brier_hard":0.0038979669527200085,"soft_accuracy":0.5564737227898723,"brier_soft":0.258451209722184,"tvd":0.38765070637417653,"kl":0.7119819177294526,"case_id":"typed:52/r0","base_id":"typed:52","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.00460099999998}
{"id":"risk","type":"ordinal","probabilities":{"0":5.257671044071197e-05,"1":9.759473575272243e-05,"2":0.9997538308727365,"3":9.599768107011398e-05},"predicted":"2","confidence":0.9997538308727365,"entropy_confidence":0.9981580110310707,"candidate_mass":0.9999954243094593,"accepted":true,"accepted_predicted":"2","input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":9.234687002959689,"brier_hard":1.9993125543778043,"score_mae":0.9298932495244361,"within_one":true,"case_id":"typed:52/r0","base_id":"typed:52","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.00460099999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001095894255619205,"1":0.046237122710267586,"2":0.9510799727507814,"3":0.001587010283333104},"predicted":"2","confidence":0.9510799727507814,"entropy_confidence":0.8502954275078348,"candidate_mass":0.9999897679098052,"accepted":true,"accepted_predicted":"2","input_tokens":207,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":3.0739722816780493,"brier_hard":1.8142204602494751,"score_mae":1.0364910990618297,"within_one":false,"case_id":"typed:52/r0","base_id":"typed:52","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.00460099999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0007478485156724439,"human_review":0.0013800422993702722,"observe":0.9976060652085513,"stop":0.00026604397640578754},"predicted":"observe","confidence":0.9976060652085513,"entropy_confidence":0.9862563029444095,"candidate_mass":0.999947526574522,"accepted":true,"accepted_predicted":"observe","input_tokens":218,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.002396804834720766,"brier_hard":8.26549733353486e-06,"soft_accuracy":0.4827228759479531,"brier_soft":0.38921879957453004,"tvd":0.5142730652085514,"kl":2.4624528529626626,"case_id":"typed:53/r0","base_id":"typed:53","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.1544}
{"id":"needs_review","type":"binary","probabilities":{"false":1.171262043754955e-06,"true":0.9999988287379546},"predicted":"true","confidence":0.9999988287379546,"entropy_confidence":0.9999752322598768,"candidate_mass":0.9999987770387566,"accepted":true,"accepted_predicted":"true","input_tokens":180,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.657428720624914,"brier_hard":1.9999953149545655,"soft_accuracy":0.4000002342524081,"brier_soft":0.7199971889738367,"tvd":0.5999988287379554,"kl":5.159688398941141,"case_id":"typed:53/r0","base_id":"typed:53","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.1544}
{"id":"outcome","type":"choice","probabilities":{"failure":1.7707450422959352e-05,"harmful":1.0050842647221734e-05,"partial":0.00031512547216017584,"success":0.9996571162347684},"predicted":"success","confidence":0.9996571162347684,"entropy_confidence":0.997696775786937,"candidate_mass":0.9999951896043365,"accepted":true,"accepted_predicted":"success","input_tokens":223,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0003429425633108347,"brier_hard":2.1728791290198472e-07,"soft_accuracy":0.48329030434623177,"brier_soft":0.4271777966209861,"tvd":0.516324116234769,"kl":3.521262123552174,"case_id":"typed:53/r0","base_id":"typed:53","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.1544}
{"id":"risk","type":"ordinal","probabilities":{"0":0.7711199672321654,"1":0.2261325300836412,"2":0.0016384695687836602,"3":0.0011090331154091993},"predicted":"0","confidence":0.7711199672321654,"entropy_confidence":0.5999014154691182,"candidate_mass":0.9999778237471509,"accepted":false,"accepted_predicted":null,"input_tokens":231,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.4866340352854017,"brier_hard":1.1935007793958612,"score_mae":0.7339304314325639,"within_one":true,"case_id":"typed:53/r0","base_id":"typed:53","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.1544}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010023521124157366,"1":0.39647616333452157,"2":0.6010105407626143,"3":0.0015109437904473422},"predicted":"2","confidence":0.6010105407626143,"entropy_confidence":0.5026084106158762,"candidate_mass":0.9999850125972255,"accepted":false,"accepted_predicted":null,"input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.9251393574078897,"brier_hard":0.7254579791920843,"score_mae":0.636363076231092,"within_one":true,"case_id":"typed:53/r0","base_id":"typed:53","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.1544}
{"id":"action","type":"choice","probabilities":{"continue":0.0009404753213528922,"human_review":0.018036782855351013,"observe":0.9805353932831896,"stop":0.00048734854010608477},"predicted":"observe","confidence":0.9805353932831896,"entropy_confidence":0.9264450208490679,"candidate_mass":0.9999656362446748,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.019656536819059732,"brier_hard":0.0007053184528408016,"soft_accuracy":0.39619770401288723,"brier_soft":0.5293806969934456,"tvd":0.5810227418232959,"kl":2.5436609293797905,"case_id":"typed:54/r0","base_id":"typed:54","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.9573}
{"id":"needs_review","type":"binary","probabilities":{"false":5.814734478959375e-08,"true":0.9999999418526552},"predicted":"true","confidence":0.9999999418526552,"entropy_confidence":0.9999985184983158,"candidate_mass":0.9999998172375892,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.660285620424187,"brier_hard":1.9999997674106273,"soft_accuracy":0.25000002907367236,"brier_soft":1.1249998255579723,"tvd":0.7499999418526552,"kl":6.561181703239002,"case_id":"typed:54/r0","base_id":"typed:54","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.9573}
{"id":"outcome","type":"choice","probabilities":{"failure":6.754228339792019e-06,"harmful":3.228218363545285e-06,"partial":6.603504781516386e-05,"success":0.9999239825054831},"predicted":"success","confidence":0.9999239825054831,"entropy_confidence":0.9993992249732403,"candidate_mass":0.9999972297475284,"accepted":true,"accepted_predicted":"success","input_tokens":211,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":7.602038399308227e-05,"brier_hard":1.0195328006848382e-08,"soft_accuracy":0.4666548876501476,"brier_soft":0.4354271776839991,"tvd":0.5332602107238459,"kl":4.3286726435980984,"case_id":"typed:54/r0","base_id":"typed:54","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.9573}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9975317309936257,"1":0.0020099419537713678,"2":0.00023751969384061558,"3":0.00022080735876082923},"predicted":"0","confidence":0.9975317309936257,"entropy_confidence":0.9864478827317976,"candidate_mass":0.9999941576296182,"accepted":true,"accepted_predicted":"0","input_tokens":219,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0024713202041368888,"brier_hard":1.0237390040003374e-05,"score_mae":0.19685259658226492,"within_one":true,"case_id":"typed:54/r0","base_id":"typed:54","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.9573}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002716788642128183,"1":0.9984495558970913,"2":0.0012208404882112812,"3":5.792475048347026e-05},"predicted":"1","confidence":0.9984495558970913,"entropy_confidence":0.9909580796521782,"candidate_mass":0.9999949407970563,"accepted":true,"accepted_predicted":"1","input_tokens":207,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.210889835456559,"brier_hard":1.9963597255588525,"score_mae":0.7677320111249641,"within_one":true,"case_id":"typed:54/r0","base_id":"typed:54","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.9573}
{"id":"action","type":"choice","probabilities":{"continue":0.000579059254025666,"human_review":0.0007314189055806903,"observe":0.9985301456284953,"stop":0.0001593762118966483},"predicted":"observe","confidence":0.9985301456284953,"entropy_confidence":0.9910120136739493,"candidate_mass":0.9999750363248089,"accepted":true,"accepted_predicted":"observe","input_tokens":216,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":7.454105747068503,"brier_hard":1.9959052289048247,"soft_accuracy":0.21683871744577,"brier_soft":0.9165191929214221,"tvd":0.7818633622952795,"kl":4.670106022481076,"case_id":"typed:55/r0","base_id":"typed:55","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":184.841401}
{"id":"needs_review","type":"binary","probabilities":{"false":4.353485412849971e-08,"true":0.999999956465147},"predicted":"true","confidence":0.999999956465147,"entropy_confidence":0.9999988726236447,"candidate_mass":0.9999998323504719,"accepted":true,"accepted_predicted":"true","input_tokens":178,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.949703975343326,"brier_hard":1.9999998258605896,"soft_accuracy":0.3400000139311537,"brier_soft":0.8711998850679903,"tvd":0.6599999564651464,"kl":5.712029375439674,"case_id":"typed:55/r0","base_id":"typed:55","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":184.841401}
{"id":"outcome","type":"choice","probabilities":{"failure":1.6371966775079187e-06,"harmful":1.058005178436926e-06,"partial":1.3614419271001605e-05,"success":0.9999836903788729},"predicted":"success","confidence":0.9999836903788729,"entropy_confidence":0.9998519654588686,"candidate_mass":0.9999991052117909,"accepted":true,"accepted_predicted":"success","input_tokens":221,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":1.630975413040315e-05,"brier_hard":4.551559413141632e-10,"soft_accuracy":0.5666624082295345,"brier_soft":0.33217560035429605,"tvd":0.4333172570453063,"kl":3.6662223372396943,"case_id":"typed:55/r0","base_id":"typed:55","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":184.841401}
{"id":"risk","type":"ordinal","probabilities":{"0":0.8865619945626758,"1":0.11277456595947072,"2":0.0004431414697965521,"3":0.00022029800805584672},"predicted":"0","confidence":0.8865619945626758,"entropy_confidence":0.7416581797986193,"candidate_mass":0.9999903141164863,"accepted":true,"accepted_predicted":"0","input_tokens":229,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.12040422408820113,"brier_hard":0.025586528710520008,"score_mae":0.22234525707676864,"within_one":true,"case_id":"typed:55/r0","base_id":"typed:55","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":184.841401}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.118383481053571e-05,"1":0.9938786958406,"2":0.006067855443934896,"3":2.2264880654555884e-05},"predicted":"1","confidence":0.9938786958406,"entropy_confidence":0.9728488545618984,"candidate_mass":0.9999966359659236,"accepted":true,"accepted_predicted":"1","input_tokens":217,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":10.37561071246375,"brier_hard":1.9877693147140358,"score_mae":0.4394142013704334,"within_one":true,"case_id":"typed:55/r0","base_id":"typed:55","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":184.841401}
{"id":"action","type":"choice","probabilities":{"continue":0.01025452021229801,"human_review":0.002603070245208014,"observe":0.987002195406292,"stop":0.0001402141362024229},"predicted":"observe","confidence":0.987002195406292,"entropy_confidence":0.9447346034391543,"candidate_mass":0.9999562213668545,"accepted":true,"accepted_predicted":"observe","input_tokens":211,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":4.580036674286262,"brier_hard":1.953776244131734,"soft_accuracy":0.3027523036938614,"brier_soft":0.8635358149466071,"tvd":0.6870021954062918,"kl":2.4681464216782043,"case_id":"typed:56/r0","base_id":"typed:56","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.0411}
{"id":"needs_review","type":"binary","probabilities":{"false":2.5514610292041223e-08,"true":0.9999999744853912},"predicted":"true","confidence":0.9999999744853912,"entropy_confidence":0.9999993196070901,"candidate_mass":0.9999998490958014,"accepted":true,"accepted_predicted":"true","input_tokens":173,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.484014596227652,"brier_hard":1.9999998979415632,"soft_accuracy":0.13333301871073142,"brier_soft":1.5022232893273206,"tvd":0.8666669744853904,"kl":7.713644664790544,"case_id":"typed:56/r0","base_id":"typed:56","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.0411}
{"id":"outcome","type":"choice","probabilities":{"failure":9.9457088256715e-07,"harmful":4.392176502301746e-07,"partial":5.3857084576328205e-06,"success":0.9999931805030083},"predicted":"success","confidence":0.9999931805030083,"entropy_confidence":0.9999333958835424,"candidate_mass":0.9999991963127212,"accepted":true,"accepted_predicted":"success","input_tokens":216,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":6.819520244536614e-06,"brier_hard":7.669347819461918e-11,"soft_accuracy":0.7166634904197029,"brier_soft":0.14368207802130453,"tvd":0.28332618050300895,"kl":2.3699252098910786,"case_id":"typed:56/r0","base_id":"typed:56","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.0411}
{"id":"risk","type":"ordinal","probabilities":{"0":0.981856493801079,"1":0.016762369707373277,"2":0.0007412702702357043,"3":0.000639866221310392},"predicted":"0","confidence":0.981856493801079,"entropy_confidence":0.9303459420318793,"candidate_mass":0.9999879824423904,"accepted":true,"accepted_predicted":"0","input_tokens":224,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.0886188032681865,"brier_hard":1.930799370953203,"score_mae":0.7231684910882241,"within_one":true,"case_id":"typed:56/r0","base_id":"typed:56","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.0411}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003128133509738755,"1":0.9991577661126807,"2":0.00048787953087266086,"3":4.1541005473113364e-05},"predicted":"1","confidence":0.9991577661126807,"entropy_confidence":0.9945858179899892,"candidate_mass":0.9999968043214595,"accepted":true,"accepted_predicted":"1","input_tokens":212,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.069903868062429,"brier_hard":1.997690952485619,"score_mae":0.38025814819084547,"within_one":true,"case_id":"typed:56/r0","base_id":"typed:56","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.0411}
{"id":"action","type":"choice","probabilities":{"continue":0.0014122836016822703,"human_review":0.0012150612132138074,"observe":0.9972575600333121,"stop":0.00011509515179328024},"predicted":"observe","confidence":0.9972575600333121,"entropy_confidence":0.9847020969633951,"candidate_mass":0.9999773224889562,"accepted":true,"accepted_predicted":"observe","input_tokens":225,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.00274620734462164,"brier_hard":1.1005142588290603e-05,"soft_accuracy":0.5490820083850642,"brier_soft":0.3130282417734009,"tvd":0.4472581100327614,"kl":2.0021577866515,"case_id":"typed:57/r0","base_id":"typed:57","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.625}
{"id":"needs_review","type":"binary","probabilities":{"false":1.0395517701173735e-07,"true":0.9999998960448214},"predicted":"true","confidence":0.9999998960448214,"entropy_confidence":0.999997438520524,"candidate_mass":0.9999996616556044,"accepted":true,"accepted_predicted":"true","input_tokens":187,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.079306020979796,"brier_hard":1.9999995841793103,"soft_accuracy":0.3333330346517945,"brier_soft":0.8888895005640756,"tvd":0.6666668960448222,"kl":5.7740259560641665,"case_id":"typed:57/r0","base_id":"typed:57","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.625}
{"id":"outcome","type":"choice","probabilities":{"failure":3.519021567212779e-06,"harmful":2.045379422728169e-06,"partial":4.2418509187858916e-05,"success":0.9999520170898204},"predicted":"success","confidence":0.9999520170898204,"entropy_confidence":0.9996061211791978,"candidate_mass":0.9999984715723641,"accepted":true,"accepted_predicted":"success","input_tokens":230,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":4.798406139626094e-05,"brier_hard":4.118256680797492e-09,"soft_accuracy":0.4166648267764792,"brier_soft":0.536574211411916,"tvd":0.5832854337564046,"kl":4.997076803908984,"case_id":"typed:57/r0","base_id":"typed:57","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.625}
{"id":"risk","type":"ordinal","probabilities":{"0":0.2952781499035513,"1":0.7032294459917167,"2":0.0009280508054037227,"3":0.0005643532993289264},"predicted":"1","confidence":0.7032294459917167,"entropy_confidence":0.5538609089272315,"candidate_mass":0.9999838683244727,"accepted":false,"accepted_predicted":null,"input_tokens":238,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.3520720592106876,"brier_hard":0.17526312730979138,"score_mae":0.2765543924994891,"within_one":true,"case_id":"typed:57/r0","base_id":"typed:57","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.625}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007254330163507009,"1":0.6626175060502454,"2":0.33578924188960846,"3":0.0008678190437939044},"predicted":"1","confidence":0.6626175060502454,"entropy_confidence":0.5307605139373095,"candidate_mass":0.9999883407057619,"accepted":false,"accepted_predicted":null,"input_tokens":226,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.411557369226525,"brier_hard":0.22658264155550817,"score_mae":0.12013244696084402,"within_one":true,"case_id":"typed:57/r0","base_id":"typed:57","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.625}
{"id":"action","type":"choice","probabilities":{"continue":0.0005079367321254847,"human_review":0.0015957237855321888,"observe":0.9976528239300322,"stop":0.00024351555230931455},"predicted":"observe","confidence":0.9976528239300322,"entropy_confidence":0.9866547310597136,"candidate_mass":0.9999677305550794,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.002349935005703638,"brier_hard":8.372869451201567e-06,"soft_accuracy":0.40974835637510504,"brier_soft":0.5210173079793056,"tvd":0.5876528239300326,"kl":2.8993827050924637,"case_id":"typed:58/r0","base_id":"typed:58","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.57330000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.0505660039549134e-07,"true":0.9999997949433986},"predicted":"true","confidence":0.9999997949433986,"entropy_confidence":0.9999951483261018,"candidate_mass":0.999999430774919,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.0505662238971296e-07,"brier_hard":8.409641912936894e-14,"soft_accuracy":0.5733329699251681,"brier_soft":0.3640891078145452,"tvd":0.4266667949433991,"kl":3.610809813055278,"case_id":"typed:58/r0","base_id":"typed:58","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.57330000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":1.8958082427458668e-05,"harmful":1.212302850882629e-05,"partial":0.0002479853543924854,"success":0.9997209335346704},"predicted":"success","confidence":0.9997209335346704,"entropy_confidence":0.9980659151022595,"candidate_mass":0.9999936970204046,"accepted":true,"accepted_predicted":"success","input_tokens":211,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":8.30214086841211,"brier_hard":1.9989460362417606,"soft_accuracy":0.156744278263012,"brier_soft":1.0357088882025216,"tvd":0.8430539335346708,"kl":6.657766161714192,"case_id":"typed:58/r0","base_id":"typed:58","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.57330000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.4832399724482107,"1":0.5117756853618443,"2":0.0030640209032283816,"3":0.0019203212867165895},"predicted":"1","confidence":0.5117756853618443,"entropy_confidence":0.47774316240481374,"candidate_mass":0.9999731174799236,"accepted":false,"accepted_predicted":null,"input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.6698688644704873,"brier_hard":0.471896928233584,"score_mae":0.23300230897154917,"within_one":true,"case_id":"typed:58/r0","base_id":"typed:58","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.57330000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000969831097113748,"1":0.13854812949795028,"2":0.8575622294758165,"3":0.002919809929119183},"predicted":"2","confidence":0.8575622294758165,"entropy_confidence":0.6902608280383362,"candidate_mass":0.9999900514751918,"accepted":true,"accepted_predicted":"2","input_tokens":207,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":1.9765375083203978,"brier_hard":1.4775217684773922,"score_mae":0.679099018236941,"within_one":true,"case_id":"typed:58/r0","base_id":"typed:58","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":193.57330000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.015287829028218087,"human_review":0.05203974813943897,"observe":0.9236220060707367,"stop":0.009050416761604698},"predicted":"observe","confidence":0.9236220060707367,"entropy_confidence":0.7592893444611178,"candidate_mass":0.9999397152784348,"accepted":true,"accepted_predicted":"observe","input_tokens":207,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":2.9557474650989874,"brier_hard":1.752021876965625,"soft_accuracy":0.19045057763182624,"brier_soft":0.9635332079813821,"tvd":0.7569551727375707,"kl":1.7723108874192914,"case_id":"typed:59/r0","base_id":"typed:59","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.54840000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":3.7625450315572433e-08,"true":0.9999999623745495},"predicted":"true","confidence":0.9999999623745495,"entropy_confidence":0.9999990177345326,"candidate_mass":0.9999998018905193,"accepted":true,"accepted_predicted":"true","input_tokens":169,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.7625451240822666e-08,"brier_hard":2.83134903925974e-15,"soft_accuracy":0.8033329771739184,"brier_soft":0.07735578817926503,"tvd":0.19666696237454961,"kl":1.6354513950868013,"case_id":"typed:59/r0","base_id":"typed:59","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.54840000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":4.831537538461142e-06,"harmful":0.9999524622971496,"partial":3.844387705044757e-05,"success":4.262288260455337e-06},"predicted":"harmful","confidence":0.9999524622971496,"entropy_confidence":0.9996031038645227,"candidate_mass":0.9999986890693466,"accepted":true,"accepted_predicted":"harmful","input_tokens":212,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":4.7538832802781914e-05,"brier_hard":3.779275731161293e-09,"soft_accuracy":0.4999886229488763,"brier_soft":0.35492805914322073,"tvd":0.4999519622966501,"kl":4.172153745211989,"case_id":"typed:59/r0","base_id":"typed:59","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.54840000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":1.4376882577110568e-05,"1":1.3401073145042775e-05,"2":0.014022277147221894,"3":0.9859499448970567},"predicted":"3","confidence":0.9859499448970567,"entropy_confidence":0.9465509465313473,"candidate_mass":0.9999980321640649,"accepted":true,"accepted_predicted":"3","input_tokens":220,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.014149691492615848,"brier_hard":0.0003940286910727593,"score_mae":0.5525747900587588,"within_one":true,"case_id":"typed:59/r0","base_id":"typed:59","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.54840000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004897284660544001,"1":0.0006957790248354809,"2":0.20633978024039554,"3":0.7924747122687138},"predicted":"3","confidence":0.7924747122687138,"entropy_confidence":0.6257876705147684,"candidate_mass":0.9999909860214915,"accepted":false,"accepted_predicted":null,"input_tokens":208,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.2325946825489922,"brier_hard":0.08564357390002972,"score_mae":0.45746647631176796,"within_one":true,"case_id":"typed:59/r0","base_id":"typed:59","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.54840000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.0007934065503439672,"human_review":0.0012967733759371844,"observe":0.9976991949337299,"stop":0.00021062513999026502},"predicted":"observe","confidence":0.9976991949337299,"entropy_confidence":0.9867515543574599,"candidate_mass":0.9999760882365866,"accepted":true,"accepted_predicted":"observe","input_tokens":221,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":7.139174793875817,"brier_hard":1.993819225948817,"soft_accuracy":0.21348423609452744,"brier_soft":1.1585490024164502,"tvd":0.7843661949337293,"kl":4.880058338424849,"case_id":"typed:60/r0","base_id":"typed:60","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.6546}
{"id":"needs_review","type":"binary","probabilities":{"false":2.7310243456034846e-06,"true":0.9999972689756536},"predicted":"true","confidence":0.9999972689756536,"entropy_confidence":0.9999455848329576,"candidate_mass":0.9999982752894484,"accepted":true,"accepted_predicted":"true","input_tokens":183,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.810833800974365,"brier_hard":1.999989075917533,"soft_accuracy":0.09666920302448469,"brier_soft":1.6320111496952547,"tvd":0.9033302689756539,"kl":8.09414384209227,"case_id":"typed:60/r0","base_id":"typed:60","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.6546}
{"id":"outcome","type":"choice","probabilities":{"failure":1.5190624022863335e-06,"harmful":8.657969574352751e-07,"partial":3.356283154783504e-05,"success":0.9999640523090941},"predicted":"success","confidence":0.9999640523090941,"entropy_confidence":0.9997012524057498,"candidate_mass":0.9999990147104989,"accepted":true,"accepted_predicted":"success","input_tokens":226,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":3.5948337039594146e-05,"brier_hard":2.4217572979257694e-09,"soft_accuracy":0.7499793939790218,"brier_soft":0.10025829685990194,"tvd":0.24996405230909327,"kl":1.977337065929289,"case_id":"typed:60/r0","base_id":"typed:60","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.6546}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9943139924801825,"1":0.005533108599977616,"2":9.72297276192181e-05,"3":5.566919222239638e-05},"predicted":"0","confidence":0.9943139924801825,"entropy_confidence":0.9741259874471092,"candidate_mass":0.9999963538189791,"accepted":true,"accepted_predicted":"0","input_tokens":234,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.197005487574415,"brier_hard":1.9776247262853832,"score_mae":0.6207724243681167,"within_one":true,"case_id":"typed:60/r0","base_id":"typed:60","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.6546}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.2414927688871e-05,"1":0.9972993328604509,"2":0.0026203179455197055,"3":2.7934266341918346e-05},"predicted":"1","confidence":0.9972993328604509,"entropy_confidence":0.9862345839625453,"candidate_mass":0.9999966614043589,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":9.856319127649147,"brier_hard":1.994507999062306,"score_mae":0.4192907715505162,"within_one":true,"case_id":"typed:60/r0","base_id":"typed:60","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":207.6546}
{"id":"action","type":"choice","probabilities":{"continue":0.006717910694426198,"human_review":0.013911858401860613,"observe":0.9729258016829213,"stop":0.0064444292207916235},"predicted":"observe","confidence":0.9729258016829213,"entropy_confidence":0.8901413262843128,"candidate_mass":0.9999320136217695,"accepted":true,"accepted_predicted":"observe","input_tokens":216,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.275013680111581,"brier_hard":1.9190410995729075,"soft_accuracy":0.42851179738225437,"brier_soft":0.4748413582790082,"tvd":0.5395923683494881,"kl":1.4519309070356696,"case_id":"typed:61/r0","base_id":"typed:61","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.5222}
{"id":"needs_review","type":"binary","probabilities":{"false":2.853804039438837e-07,"true":0.9999997146195961},"predicted":"true","confidence":0.9999997146195961,"entropy_confidence":0.9999933839389901,"candidate_mass":0.9999994459833175,"accepted":true,"accepted_predicted":"true","input_tokens":178,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.069442795985303,"brier_hard":1.9999988584785473,"soft_accuracy":0.47000001712282424,"brier_soft":0.5617993949937067,"tvd":0.529999714619596,"kl":4.526619916665512,"case_id":"typed:61/r0","base_id":"typed:61","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.5222}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00043350864093799244,"harmful":0.00029364905439728723,"partial":0.013617026871534306,"success":0.985655815433131},"predicted":"success","confidence":0.985655815433131,"entropy_confidence":0.9433809468984231,"candidate_mass":0.9999711168162189,"accepted":true,"accepted_predicted":"success","input_tokens":221,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.014448056887996358,"brier_hard":0.00039145321121740744,"soft_accuracy":0.6769645131448191,"brier_soft":0.15124129138877465,"tvd":0.3023221320994474,"kl":0.8009790487859686,"case_id":"typed:61/r0","base_id":"typed:61","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.5222}
{"id":"risk","type":"ordinal","probabilities":{"0":2.902605259366089e-05,"1":4.4889296648739756e-05,"2":0.9998611306383952,"3":6.49540123635944e-05},"predicted":"2","confidence":0.9998611306383952,"entropy_confidence":0.9989051528614556,"candidate_mass":0.9999975404867042,"accepted":true,"accepted_predicted":"2","input_tokens":229,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00013887900484736145,"brier_hard":2.636128399743975e-08,"score_mae":0.54996201261053,"within_one":true,"case_id":"typed:61/r0","base_id":"typed:61","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.5222}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00123517367448338,"1":0.0055996992982638456,"2":0.9852262259246644,"3":0.00793890110258914},"predicted":"2","confidence":0.9852262259246644,"entropy_confidence":0.9348172034305067,"candidate_mass":0.9999904395148388,"accepted":true,"accepted_predicted":"2","input_tokens":217,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.185042379421354,"brier_hard":1.9595672260901837,"score_mae":0.5832018544553601,"within_one":true,"case_id":"typed:61/r0","base_id":"typed:61","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.5222}
{"id":"action","type":"choice","probabilities":{"continue":0.002893711244954304,"human_review":0.0016877387378732053,"observe":0.9952651722654615,"stop":0.00015337775171228867},"predicted":"observe","confidence":0.9952651722654615,"entropy_confidence":0.9756471893631982,"candidate_mass":0.9999698970578346,"accepted":true,"accepted_predicted":"observe","input_tokens":220,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":5.845215432891791,"brier_hard":1.9847765861862414,"soft_accuracy":0.3666429804922525,"brier_soft":0.6861668254696448,"tvd":0.6287515500171731,"kl":2.824229051937856,"case_id":"typed:62/r0","base_id":"typed:62","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.9024}
{"id":"needs_review","type":"binary","probabilities":{"false":2.4597765342316836e-08,"true":0.999999975402233},"predicted":"true","confidence":0.999999975402233,"entropy_confidence":0.9999993427577305,"candidate_mass":0.9999997998800958,"accepted":true,"accepted_predicted":"true","input_tokens":182,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.52061023787749,"brier_hard":1.9999999016089363,"soft_accuracy":0.15000001721843548,"brier_soft":1.4449999163675962,"tvd":0.8499999754022338,"kl":7.544221322136538,"case_id":"typed:62/r0","base_id":"typed:62","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.9024}
{"id":"outcome","type":"choice","probabilities":{"failure":1.8861627342675503e-05,"harmful":1.0543005059361982e-05,"partial":0.0008516257286039696,"success":0.9991189696389928},"predicted":"success","confidence":0.9991189696389928,"entropy_confidence":0.9947873627735385,"candidate_mass":0.9999939495664346,"accepted":true,"accepted_predicted":"success","input_tokens":225,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0008814186963625947,"brier_hard":1.5019477945783682e-06,"soft_accuracy":0.6330185128854391,"brier_soft":0.2205360587887005,"tvd":0.3657958793104194,"kl":2.056114035834507,"case_id":"typed:62/r0","base_id":"typed:62","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.9024}
{"id":"risk","type":"ordinal","probabilities":{"0":0.012493731110062562,"1":0.9871740603479949,"2":0.0002402959879151083,"3":9.191255402843974e-05},"predicted":"1","confidence":0.9871740603479949,"entropy_confidence":0.9492500998967447,"candidate_mass":0.9999935862473011,"accepted":true,"accepted_predicted":"1","input_tokens":233,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.012908902160841468,"brier_hard":0.00032066423508681716,"score_mae":0.40459738998591044,"within_one":true,"case_id":"typed:62/r0","base_id":"typed:62","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.9024}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001247487680101073,"1":0.9984833777818677,"2":0.001348223734644368,"3":4.364971547787878e-05},"predicted":"1","confidence":0.9984833777818677,"entropy_confidence":0.9913543296544045,"candidate_mass":0.9999958058391262,"accepted":true,"accepted_predicted":"1","input_tokens":221,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.989208699051893,"brier_hard":1.9967213933454593,"score_mae":0.48464377439759,"within_one":true,"case_id":"typed:62/r0","base_id":"typed:62","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.9024}
{"id":"action","type":"choice","probabilities":{"continue":0.00034279322302764224,"human_review":0.001632940698373515,"observe":0.9974749261326947,"stop":0.0005493399459054924},"predicted":"observe","confidence":0.9974749261326947,"entropy_confidence":0.9856741901906285,"candidate_mass":0.9999628187458702,"accepted":true,"accepted_predicted":"observe","input_tokens":224,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.417372780182582,"brier_hard":1.9916934326435722,"soft_accuracy":0.10085637811992366,"brier_soft":1.2191134762008826,"tvd":0.897474826132594,"kl":5.125157252218775,"case_id":"typed:63/r0","base_id":"typed:63","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.9338}
{"id":"needs_review","type":"binary","probabilities":{"false":3.985375288378531e-08,"true":0.9999999601462485},"predicted":"true","confidence":0.9999999601462485,"entropy_confidence":0.9999989628696823,"candidate_mass":0.9999998122620467,"accepted":true,"accepted_predicted":"true","input_tokens":186,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.9853752315317995e-08,"brier_hard":3.1766431292319675e-15,"soft_accuracy":0.859999971305299,"brier_soft":0.03919997768190195,"tvd":0.13999996014624783,"kl":1.1597400011791508,"case_id":"typed:63/r0","base_id":"typed:63","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.9338}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00012795750644384467,"harmful":0.99906633330215,"partial":0.0007683156345108845,"success":3.739355689525903e-05},"predicted":"harmful","confidence":0.99906633330215,"entropy_confidence":0.994249967465267,"candidate_mass":0.9999945319954227,"accepted":true,"accepted_predicted":"harmful","input_tokens":229,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.0009341028360943055,"brier_hard":1.479813818460591e-06,"soft_accuracy":0.742777034340577,"brier_soft":0.09855743329296462,"tvd":0.25573333330215,"kl":1.2442137069462578,"case_id":"typed:63/r0","base_id":"typed:63","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.9338}
{"id":"risk","type":"ordinal","probabilities":{"0":8.724680135064379e-06,"1":8.842808429933746e-06,"2":0.010527301712058285,"3":0.9894551307993753},"predicted":"3","confidence":0.9894551307993753,"entropy_confidence":0.9577054785303438,"candidate_mass":0.9999981712190962,"accepted":true,"accepted_predicted":"3","input_tokens":237,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.010600860294200543,"brier_hard":0.000222018502110293,"score_mae":0.5060958386306722,"within_one":true,"case_id":"typed:63/r0","base_id":"typed:63","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.9338}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00040483495046710646,"1":0.0005446828909155762,"2":0.939970181555923,"3":0.0590803006026939},"predicted":"2","confidence":0.939970181555923,"entropy_confidence":0.8322314030448505,"candidate_mass":0.9999898584837652,"accepted":true,"accepted_predicted":"2","input_tokens":225,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.8288577332700555,"brier_hard":1.7688742834989806,"score_mae":0.32560705218915675,"within_one":true,"case_id":"typed:63/r0","base_id":"typed:63","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.9338}
{"id":"action","type":"choice","probabilities":{"continue":0.0005972070070761776,"human_review":0.0011480286684161537,"observe":0.9980788648393921,"stop":0.0001758994851149349},"predicted":"observe","confidence":0.9980788648393921,"entropy_confidence":0.9887144604763283,"candidate_mass":0.9999731319850825,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0019229829076543414,"brier_hard":5.396326966993811e-06,"soft_accuracy":0.4163127607163396,"brier_soft":0.4868709245731088,"tvd":0.5814122815059758,"kl":3.040475749861831,"case_id":"typed:64/r0","base_id":"typed:64","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.66920000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":1.824770493851065e-08,"true":0.9999999817522963},"predicted":"true","confidence":0.9999999817522963,"entropy_confidence":0.9999995045674267,"candidate_mass":0.9999998848965741,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.819226521619896,"brier_hard":1.9999999270091835,"soft_accuracy":0.33333300608258093,"brier_soft":0.8888897291174316,"tvd":0.6666669817522957,"kl":5.7740259274950345,"case_id":"typed:64/r0","base_id":"typed:64","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.66920000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":7.350268066223591e-05,"harmful":4.262358043806482e-05,"partial":0.01007615059185044,"success":0.989807723147051},"predicted":"success","confidence":0.989807723147051,"entropy_confidence":0.9584542459421997,"candidate_mass":0.9999829440597726,"accepted":true,"accepted_predicted":"success","input_tokens":214,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":4.597583975004962,"brier_hard":1.9596685636480118,"soft_accuracy":0.2095530264617962,"brier_soft":0.9273588874641202,"tvd":0.7831407231470502,"kl":4.034819740538914,"case_id":"typed:64/r0","base_id":"typed:64","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.66920000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.046205018854882006,"1":0.949687527047684,"2":0.0031109740711062553,"3":0.0009964800263289907},"predicted":"1","confidence":0.949687527047684,"entropy_confidence":0.8442350233491202,"candidate_mass":0.9999872828741871,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":3.074666853572051,"brier_hard":1.8116419362196754,"score_mae":0.5888989152688835,"within_one":true,"case_id":"typed:64/r0","base_id":"typed:64","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.66920000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009800546841985916,"1":0.16752463323817682,"2":0.8296444799565301,"3":0.0018508321210958373},"predicted":"2","confidence":0.8296444799565301,"entropy_confidence":0.6590321502667429,"candidate_mass":0.9999900136819319,"accepted":true,"accepted_predicted":"2","input_tokens":210,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.927902187653785,"brier_hard":1.7144187425822541,"score_mae":0.9656990895145247,"within_one":true,"case_id":"typed:64/r0","base_id":"typed:64","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":188.66920000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.006606519610715017,"human_review":0.02992551409447815,"observe":0.9615152951318161,"stop":0.0019526711629900186},"predicted":"observe","confidence":0.9615152951318161,"entropy_confidence":0.8643224601113799,"candidate_mass":0.9999353137127525,"accepted":true,"accepted_predicted":"observe","input_tokens":202,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.509043848254482,"brier_hard":1.8656036300033234,"soft_accuracy":0.08231888069295175,"brier_soft":1.216927452362376,"tvd":0.8948482951318164,"kl":3.2833414468365443,"case_id":"typed:65/r0","base_id":"typed:65","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.88920000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":7.653494550426664e-08,"true":0.9999999234650538},"predicted":"true","confidence":0.9999999234650538,"entropy_confidence":0.9999980803504103,"candidate_mass":0.9999996247391387,"accepted":true,"accepted_predicted":"true","input_tokens":164,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.65349490872455e-08,"brier_hard":1.1715195866817348e-14,"soft_accuracy":0.9266669346901283,"brier_soft":0.010755435327862979,"tvd":0.07333292346505416,"kl":0.604846074872058,"case_id":"typed:65/r0","base_id":"typed:65","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.88920000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":4.770117393750953e-06,"harmful":0.9999531243891915,"partial":3.769601590716552e-05,"success":4.409477508693083e-06},"predicted":"harmful","confidence":0.9999531243891915,"entropy_confidence":0.9996078250042955,"candidate_mass":0.9999986708928064,"accepted":true,"accepted_predicted":"harmful","input_tokens":207,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":4.687670950429581e-05,"brier_hard":3.6605100157946347e-09,"soft_accuracy":0.7999661973487328,"brier_soft":0.06052943551942746,"tvd":0.19995753386669962,"kl":1.5378166649912266,"case_id":"typed:65/r0","base_id":"typed:65","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.88920000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":3.2058846767726922e-06,"1":3.0256813055411824e-06,"2":0.0005927398806747899,"3":0.9994010285533423},"predicted":"3","confidence":0.9994010285533423,"entropy_confidence":0.9963338923158876,"candidate_mass":0.9999990464919754,"accepted":true,"accepted_predicted":"3","input_tokens":215,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0005991509017171636,"brier_hard":7.101267924974565e-07,"score_mae":0.34939159110268214,"within_one":true,"case_id":"typed:65/r0","base_id":"typed:65","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.88920000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000978056820662076,"1":0.0018571588868764538,"2":0.7973199294918805,"3":0.19984485480058184},"predicted":"2","confidence":0.7973199294918805,"entropy_confidence":0.6242916590378434,"candidate_mass":0.999987952567707,"accepted":false,"accepted_predicted":null,"input_tokens":203,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.610213939462291,"brier_hard":1.2759717319883148,"score_mae":0.15396841772761727,"within_one":true,"case_id":"typed:65/r0","base_id":"typed:65","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.88920000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.004073576647744811,"human_review":0.005819956956330426,"observe":0.9896559874487924,"stop":0.0004504789471325252},"predicted":"observe","confidence":0.9896559874487924,"entropy_confidence":0.9522962839185525,"candidate_mass":0.9999606209314568,"accepted":true,"accepted_predicted":"observe","input_tokens":221,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.010397883666708867,"brier_hard":0.00015766745261994175,"soft_accuracy":0.41504255251630473,"brier_soft":0.4749403150955952,"tvd":0.5729889874487923,"kl":2.049317348483263,"case_id":"typed:66/r0","base_id":"typed:66","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.005399}
{"id":"needs_review","type":"binary","probabilities":{"false":5.01745802213593e-08,"true":0.9999999498254203},"predicted":"true","confidence":0.9999999498254203,"entropy_confidence":0.9999987109566757,"candidate_mass":0.999999743810343,"accepted":true,"accepted_predicted":"true","input_tokens":183,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.017458097198007e-08,"brier_hard":5.034976949789474e-15,"soft_accuracy":0.5166669983274809,"brier_soft":0.46722148077388403,"tvd":0.4833329498254201,"kl":4.110476947315477,"case_id":"typed:66/r0","base_id":"typed:66","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.005399}
{"id":"outcome","type":"choice","probabilities":{"failure":4.0325771346551394e-05,"harmful":2.2521561250914894e-05,"partial":0.0017027734992272044,"success":0.9982343791681737},"predicted":"success","confidence":0.9982343791681737,"entropy_confidence":0.9904283463114116,"candidate_mass":0.9999925213096702,"accepted":true,"accepted_predicted":"success","input_tokens":226,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":6.3754968871879925,"brier_hard":1.993069230325793,"soft_accuracy":0.18379033764029334,"brier_soft":0.9994497798216607,"tvd":0.8149239007294254,"kl":5.54120350833781,"case_id":"typed:66/r0","base_id":"typed:66","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.005399}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0019292909209611003,"1":0.9978926760730489,"2":0.00012296349581914547,"3":5.506951017053171e-05},"predicted":"1","confidence":0.9978926760730489,"entropy_confidence":0.9885944050160821,"candidate_mass":0.999995749803632,"accepted":true,"accepted_predicted":"1","input_tokens":234,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.002109547458367498,"brier_hard":8.181130263058112e-06,"score_mae":0.3149708115951988,"within_one":true,"case_id":"typed:66/r0","base_id":"typed:66","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.005399}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00017560665545382883,"1":0.9983139238132269,"2":0.00144153603546644,"3":6.893349585193278e-05},"predicted":"1","confidence":0.9983139238132269,"entropy_confidence":0.9904101807638572,"candidate_mass":0.9999955854286027,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0016874992130055918,"brier_hard":4.956468573441592e-06,"score_mae":0.23473679637171552,"within_one":true,"case_id":"typed:66/r0","base_id":"typed:66","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.005399}
{"id":"action","type":"choice","probabilities":{"continue":0.0025666084307338403,"human_review":0.006962262081500082,"observe":0.9896795343135206,"stop":0.0007915951742456984},"predicted":"observe","confidence":0.9896795343135206,"entropy_confidence":0.9525254299308539,"candidate_mass":0.9999699261717171,"accepted":true,"accepted_predicted":"observe","input_tokens":222,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":5.965169928372493,"brier_hard":1.9743880509726075,"soft_accuracy":0.4207399464256443,"brier_soft":0.5721966527607866,"tvd":0.5663465343135204,"kl":2.452273542362204,"case_id":"typed:67/r0","base_id":"typed:67","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.45909999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":1.1691043896519598e-07,"true":0.9999998830895611},"predicted":"true","confidence":0.9999998830895611,"entropy_confidence":0.9999971391094488,"candidate_mass":0.9999996531352584,"accepted":true,"accepted_predicted":"true","input_tokens":184,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.961857674207199,"brier_hard":1.9999995323582715,"soft_accuracy":0.1000000935283512,"brier_soft":1.6199995791224473,"tvd":0.899999883089561,"kl":8.059047837170205,"case_id":"typed:67/r0","base_id":"typed:67","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.45909999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":2.195433800964513e-06,"harmful":9.100700825851344e-07,"partial":1.4978020417873917e-05,"success":0.9999819164756995},"predicted":"success","confidence":0.9999819164756995,"entropy_confidence":0.9998371655147266,"candidate_mass":0.9999988443936514,"accepted":true,"accepted_predicted":"success","input_tokens":227,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":1.808368780944627e-05,"brier_hard":5.57003103896446e-10,"soft_accuracy":0.8499862237244391,"brier_soft":0.03419138605952364,"tvd":0.14998191647569903,"kl":1.2030254640556426,"case_id":"typed:67/r0","base_id":"typed:67","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.45909999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9997513088041167,"1":0.00019803495596458764,"2":2.96350838893574e-05,"3":2.1021156029000732e-05},"predicted":"0","confidence":0.9997513088041167,"entropy_confidence":0.9982163190849964,"candidate_mass":0.9999984894012718,"accepted":true,"accepted_predicted":"0","input_tokens":235,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00024872212466670463,"brier_hard":1.0238528189170315e-07,"score_mae":0.09467963140816969,"within_one":true,"case_id":"typed:67/r0","base_id":"typed:67","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.45909999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.351173321480931e-05,"1":0.9997920442902449,"2":0.00012057550018280041,"3":1.3868476356013578e-05},"predicted":"1","confidence":0.9997920442902449,"entropy_confidence":0.9984485720152768,"candidate_mass":0.9999976780168176,"accepted":true,"accepted_predicted":"1","input_tokens":223,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":9.518065528910249,"brier_hard":1.9994371284943984,"score_mae":0.7600748007196785,"within_one":true,"case_id":"typed:67/r0","base_id":"typed:67","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.45909999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0043639920022604676,"human_review":0.010439958979449894,"observe":0.9841172873668392,"stop":0.0010787616514493538},"predicted":"observe","confidence":0.9841172873668392,"entropy_confidence":0.9318544627839502,"candidate_mass":0.9999474816635694,"accepted":true,"accepted_predicted":"observe","input_tokens":205,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.016010194554485897,"brier_hard":0.00038146145697659213,"soft_accuracy":0.43123606522316654,"brier_soft":0.47892166130032193,"tvd":0.5507842873668397,"kl":1.7013824914735158,"case_id":"typed:68/r0","base_id":"typed:68","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.91150000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":1.5141074481043673e-08,"true":0.9999999848589242},"predicted":"true","confidence":0.9999999848589242,"entropy_confidence":0.9999995848370018,"candidate_mass":0.999999871070885,"accepted":true,"accepted_predicted":"true","input_tokens":167,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.5141075952180842e-08,"brier_hard":4.585043139590962e-16,"soft_accuracy":0.6066669967698931,"brier_soft":0.30942167395606246,"tvd":0.3933329848589249,"kl":3.3195336743260895,"case_id":"typed:68/r0","base_id":"typed:68","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.91150000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":4.4342302339879e-05,"harmful":2.2573525144710082e-05,"partial":0.0014344091528490654,"success":0.9984986750196656},"predicted":"success","confidence":0.9984986750196656,"entropy_confidence":0.9916487767253658,"candidate_mass":0.9999901551791426,"accepted":true,"accepted_predicted":"success","input_tokens":210,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":6.54700225473368,"brier_hard":1.9941328457157512,"soft_accuracy":0.1272796039769542,"brier_soft":1.1553085970018462,"tvd":0.8718318016865393,"kl":5.832001876186539,"case_id":"typed:68/r0","base_id":"typed:68","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.91150000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.00573558668807009,"1":0.9939367426368788,"2":0.00023192159557368431,"3":9.57490794787077e-05},"predicted":"1","confidence":0.9939367426368788,"entropy_confidence":0.972247168292119,"candidate_mass":0.999994728611596,"accepted":true,"accepted_predicted":"1","input_tokens":218,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":5.161065234056746,"brier_hard":1.9764720348976381,"score_mae":0.5646878330664624,"within_one":true,"case_id":"typed:68/r0","base_id":"typed:68","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.91150000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002282037583977084,"1":0.9946996448940018,"2":0.004967309636592854,"3":0.00010484171100625709},"predicted":"1","confidence":0.9946996448940018,"entropy_confidence":0.9751052361178659,"candidate_mass":0.9999934376544843,"accepted":true,"accepted_predicted":"1","input_tokens":206,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.005314451821920629,"brier_hard":5.2830998015183456e-05,"score_mae":0.1783842106997937,"within_one":true,"case_id":"typed:68/r0","base_id":"typed:68","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.91150000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.01757477478696,"human_review":0.08011865237707747,"observe":0.04766442638640412,"stop":0.8546421464495585},"predicted":"stop","confidence":0.8546421464495585,"entropy_confidence":0.6014015739366976,"candidate_mass":0.9999282897101381,"accepted":true,"accepted_predicted":"stop","input_tokens":217,"reused_prefix_tokens":0,"expected":"stop","correct":true,"accepted_correct":true,"nll_hard":0.15707243982285493,"brier_hard":0.030128674299067886,"soft_accuracy":0.4382569067389303,"brier_soft":0.2875883114983243,"tvd":0.3879751464495585,"kl":0.5148036803324015,"case_id":"typed:69/r0","base_id":"typed:69","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.1009}
{"id":"needs_review","type":"binary","probabilities":{"false":2.903043014843193e-07,"true":0.9999997096956998},"predicted":"true","confidence":0.9999997096956998,"entropy_confidence":0.9999932769513892,"candidate_mass":0.9999994974311069,"accepted":true,"accepted_predicted":"true","input_tokens":179,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.9030434229283287e-07,"brier_hard":1.6855317414851077e-13,"soft_accuracy":0.9433327425970475,"brier_soft":0.0064222319754733,"tvd":0.0566667096956992,"kl":0.4668924235087964,"case_id":"typed:69/r0","base_id":"typed:69","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.1009}
{"id":"outcome","type":"choice","probabilities":{"failure":8.907676083386979e-06,"harmful":0.9999361937079914,"partial":5.066284131396702e-05,"success":4.235774610920294e-06},"predicted":"harmful","confidence":0.9999361937079914,"entropy_confidence":0.9994800062498207,"candidate_mass":0.9999980205034243,"accepted":true,"accepted_predicted":"harmful","input_tokens":222,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":6.380832771668347e-05,"brier_hard":6.735254869656983e-09,"soft_accuracy":0.8166196168116793,"brier_soft":0.046944796083879074,"tvd":0.18326919370799152,"kl":1.3589785725893697,"case_id":"typed:69/r0","base_id":"typed:69","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.1009}
{"id":"risk","type":"ordinal","probabilities":{"0":2.0088977755211154e-06,"1":9.267193600141896e-07,"2":7.111468192715707e-05,"3":0.9999259497009363},"predicted":"3","confidence":0.9999259497009363,"entropy_confidence":0.9994283281256453,"candidate_mass":0.9999995767128497,"accepted":true,"accepted_predicted":"3","input_tokens":230,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":7.405304092246393e-05,"brier_hard":1.0545639256070391e-08,"score_mae":0.3865880051860233,"within_one":true,"case_id":"typed:69/r0","base_id":"typed:69","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.1009}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004059822291773926,"1":0.000544514303865355,"2":0.009078190555127834,"3":0.9899713129118285},"predicted":"3","confidence":0.9899713129118285,"entropy_confidence":0.9567728498858039,"candidate_mass":0.9999914236322804,"accepted":true,"accepted_predicted":"3","input_tokens":218,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.01007931313007756,"brier_hard":0.0001834494258651919,"score_mae":0.43861483414960656,"within_one":true,"case_id":"typed:69/r0","base_id":"typed:69","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.1009}
{"id":"action","type":"choice","probabilities":{"continue":0.004336041854355079,"human_review":0.011530799647567995,"observe":0.9832716297395933,"stop":0.0008615287584847254},"predicted":"observe","confidence":0.9832716297395933,"entropy_confidence":0.9295115816148573,"candidate_mass":0.9999556072706447,"accepted":true,"accepted_predicted":"observe","input_tokens":208,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.016869869702571077,"brier_hard":0.00043234120284603027,"soft_accuracy":0.4461090599873127,"brier_soft":0.4165131184854071,"tvd":0.5332716297395926,"kl":1.7056445154477011,"case_id":"typed:70/r0","base_id":"typed:70","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.8157}
{"id":"needs_review","type":"binary","probabilities":{"false":1.2683975285486664e-06,"true":0.9999987316024723},"predicted":"true","confidence":0.9999987316024723,"entropy_confidence":0.9999733240072189,"candidate_mass":0.9999990253779838,"accepted":true,"accepted_predicted":"true","input_tokens":170,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.577756242763567,"brier_hard":1.9999949264131052,"soft_accuracy":0.3233334481679726,"brier_soft":0.915753024650217,"tvd":0.6766657316024719,"kl":5.867267391200305,"case_id":"typed:70/r0","base_id":"typed:70","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.8157}
{"id":"outcome","type":"choice","probabilities":{"failure":2.4774473281220435e-05,"harmful":1.497507673325987e-05,"partial":0.0002463717090225309,"success":0.999713878740964},"predicted":"success","confidence":0.999713878740964,"entropy_confidence":0.9980074853203503,"candidate_mass":0.9999927188076517,"accepted":true,"accepted_predicted":"success","input_tokens":213,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.00028616219953296303,"brier_hard":1.4340242132858194e-07,"soft_accuracy":0.6432165963386781,"brier_soft":0.20270554376299313,"tvd":0.35638087874096347,"kl":2.27612386683341,"case_id":"typed:70/r0","base_id":"typed:70","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.8157}
{"id":"risk","type":"ordinal","probabilities":{"0":9.34830379669912e-06,"1":1.6691827469458435e-05,"2":0.999957032245099,"3":1.692762363377662e-05},"predicted":"2","confidence":0.999957032245099,"entropy_confidence":0.9996243081241684,"candidate_mass":0.9999982313033523,"accepted":true,"accepted_predicted":"2","input_tokens":221,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":4.296867804140415e-05,"brier_hard":2.4987802912630235e-09,"score_mae":0.30998153918856897,"within_one":true,"case_id":"typed:70/r0","base_id":"typed:70","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.8157}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0011427662565099757,"1":0.008054490213799493,"2":0.98709685305427,"3":0.0037058904754221876},"predicted":"2","confidence":0.98709685305427,"entropy_confidence":0.942190583650905,"candidate_mass":0.9999906141839864,"accepted":true,"accepted_predicted":"2","input_tokens":209,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.821525552524882,"brier_hard":1.9583311312335812,"score_mae":0.560032867748606,"within_one":true,"case_id":"typed:70/r0","base_id":"typed:70","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.8157}
{"id":"action","type":"choice","probabilities":{"continue":0.0030986257896412177,"human_review":0.0062788029319204635,"observe":0.9900267306495616,"stop":0.000595840628876184},"predicted":"observe","confidence":0.9900267306495616,"entropy_confidence":0.9537723829816888,"candidate_mass":0.9999576549319087,"accepted":true,"accepted_predicted":"observe","input_tokens":207,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":5.776796559438843,"brier_hard":1.9740050556954742,"soft_accuracy":0.3324036395732969,"brier_soft":0.7248389836841629,"tvd":0.6566937306495618,"kl":2.7792743502515087,"case_id":"typed:71/r0","base_id":"typed:71","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.6398}
{"id":"needs_review","type":"binary","probabilities":{"false":5.806658463777564e-06,"true":0.9999941933415368},"predicted":"true","confidence":0.9999941933415368,"entropy_confidence":0.9998906225803058,"candidate_mass":0.9999973320885891,"accepted":true,"accepted_predicted":"true","input_tokens":169,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.056505287876494,"brier_hard":1.9999767734335807,"soft_accuracy":0.16667087110177153,"brier_soft":1.3888684223249652,"tvd":0.8333271933415365,"kl":7.376654698857019,"case_id":"typed:71/r0","base_id":"typed:71","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.6398}
{"id":"outcome","type":"choice","probabilities":{"failure":6.770132444319367e-05,"harmful":3.101984824790756e-05,"partial":0.006192895266712217,"success":0.9937083835605969},"predicted":"success","confidence":0.9937083835605969,"entropy_confidence":0.9720617639584713,"candidate_mass":0.999984417445957,"accepted":true,"accepted_predicted":"success","input_tokens":212,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.006311492068526499,"brier_hard":7.794193490535092e-05,"soft_accuracy":0.5816307482054219,"brier_soft":0.2747885904232552,"tvd":0.41040640340884466,"kl":1.6649587516361377,"case_id":"typed:71/r0","base_id":"typed:71","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.6398}
{"id":"risk","type":"ordinal","probabilities":{"0":0.1817487281966174,"1":0.7959775611377875,"2":0.01504293933535999,"3":0.007230771330234028},"predicted":"1","confidence":0.7959775611377875,"entropy_confidence":0.5741805167673009,"candidate_mass":0.9999730495408212,"accepted":false,"accepted_predicted":null,"input_tokens":220,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":null,"nll_hard":1.705130160135674,"brier_hard":1.303393995720591,"score_mae":0.3644227537992095,"within_one":true,"case_id":"typed:71/r0","base_id":"typed:71","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.6398}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007426939957402941,"1":0.9978227247268151,"2":0.0013324617756084367,"3":0.00010211950183776387},"predicted":"1","confidence":0.9978227247268151,"entropy_confidence":0.9875304541883299,"candidate_mass":0.9999956803971878,"accepted":true,"accepted_predicted":"1","input_tokens":208,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.205226447689153,"brier_hard":1.9941671394669123,"score_mae":0.5507940067835453,"within_one":true,"case_id":"typed:71/r0","base_id":"typed:71","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.6398}
{"id":"action","type":"choice","probabilities":{"continue":0.002779571358640434,"human_review":0.5369022324111445,"observe":0.3873953609723115,"stop":0.07292283525790194},"predicted":"human_review","confidence":0.5369022324111445,"entropy_confidence":0.3445922338919972,"candidate_mass":0.9997465544679768,"accepted":false,"accepted_predicted":null,"input_tokens":221,"reused_prefix_tokens":0,"expected":"observe","correct":false,"accepted_correct":null,"nll_hard":0.9483095028555616,"brier_hard":0.6688739168453041,"soft_accuracy":0.34824505954685736,"brier_soft":0.06772999747421236,"tvd":0.1931580676690473,"kl":0.6250523285493816,"case_id":"typed:72/r0","base_id":"typed:72","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":215.08749999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":1.260353223581967e-05,"true":0.9999873964677632},"predicted":"true","confidence":0.9999873964677632,"entropy_confidence":0.9997766843354148,"candidate_mass":0.9999944777670955,"accepted":true,"accepted_predicted":"true","input_tokens":183,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.260361166193437e-05,"brier_hard":3.1769804966225766e-10,"soft_accuracy":0.5833308994196958,"brier_soft":0.3472017721918329,"tvd":0.4166543964677637,"kl":3.5232374647776634,"case_id":"typed:72/r0","base_id":"typed:72","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":215.08749999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":0.016629112235206703,"harmful":0.029151787198104177,"partial":0.10503369521064845,"success":0.8491854053560391},"predicted":"success","confidence":0.8491854053560391,"entropy_confidence":0.6056429315474108,"candidate_mass":0.9999343042081383,"accepted":true,"accepted_predicted":"success","input_tokens":226,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":2.253474073536097,"brier_hard":1.5232068934485818,"soft_accuracy":0.32775523953417823,"brier_soft":0.47840804035752216,"tvd":0.5350031925541441,"kl":0.7989758675159535,"case_id":"typed:72/r0","base_id":"typed:72","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":215.08749999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":3.8454602430228195e-05,"1":3.08231784634828e-05,"2":0.9996649838284503,"3":0.0002657383906560772},"predicted":"2","confidence":0.9996649838284503,"entropy_confidence":0.9976672469961062,"candidate_mass":0.9999956045466585,"accepted":true,"accepted_predicted":"2","input_tokens":234,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00033507230200409676,"brier_hard":1.8528155224698548e-07,"score_mae":0.3501580060073324,"within_one":true,"case_id":"typed:72/r0","base_id":"typed:72","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":215.08749999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.003244012718467381,"1":0.0050178936052081,"2":0.3760627445033587,"3":0.6156753491729666},"predicted":"3","confidence":0.6156753491729666,"entropy_confidence":0.48670843184941537,"candidate_mass":0.9999827315776606,"accepted":false,"accepted_predicted":null,"input_tokens":222,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.9779992758349607,"brier_hard":0.7683895372506863,"score_mae":1.2375024301308253,"within_one":false,"case_id":"typed:72/r0","base_id":"typed:72","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":215.08749999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0010024596690836084,"human_review":0.0037514255308345027,"observe":0.9938919298614012,"stop":0.0013541849386804576},"predicted":"observe","confidence":0.9938919298614012,"entropy_confidence":0.9690473487490523,"candidate_mass":0.9999570105012374,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.585619369679446,"brier_hard":1.9803352291177012,"soft_accuracy":0.32683038470182246,"brier_soft":0.7359327285537253,"tvd":0.6672249298614014,"kl":2.8654618130305476,"case_id":"typed:73/r0","base_id":"typed:73","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.14329999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":3.4956683074660275e-06,"true":0.9999965043316933},"predicted":"true","confidence":0.9999965043316933,"entropy_confidence":0.9999315943407969,"candidate_mass":0.9999978183591688,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.4956744165704714e-06,"brier_hard":2.4439393826343622e-11,"soft_accuracy":0.6266661144283655,"brier_soft":0.2787498376090951,"tvd":0.37332950433169293,"kl":3.145659649523585,"case_id":"typed:73/r0","base_id":"typed:73","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.14329999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":0.006715750781167496,"harmful":0.016185491960821117,"partial":0.9648084787908073,"success":0.012290278467204676},"predicted":"partial","confidence":0.9648084787908073,"entropy_confidence":0.8636846214105814,"candidate_mass":0.999948047072389,"accepted":true,"accepted_predicted":"partial","input_tokens":214,"reused_prefix_tokens":0,"expected":"success","correct":false,"accepted_correct":false,"nll_hard":4.398946697629583,"brier_hard":1.9067329662155923,"soft_accuracy":0.2660462714891933,"brier_soft":0.8889318157276153,"tvd":0.698141478790807,"kl":2.358734493528185,"case_id":"typed:73/r0","base_id":"typed:73","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.14329999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":7.152354711425861e-06,"1":1.316618830692203e-05,"2":0.9999382788469497,"3":4.1402610032286703e-05},"predicted":"2","confidence":0.9999382788469497,"entropy_confidence":0.9994862111102356,"candidate_mass":0.9999982314332286,"accepted":true,"accepted_predicted":"2","input_tokens":222,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":6.172305787899987e-05,"brier_hard":5.748181543789838e-09,"score_mae":0.32334693171230344,"within_one":true,"case_id":"typed:73/r0","base_id":"typed:73","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.14329999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0017400956619851651,"1":0.0019113508680591357,"2":0.8430131998907888,"3":0.15333535357916703},"predicted":"2","confidence":0.8430131998907888,"entropy_confidence":0.6721416321768304,"candidate_mass":0.9999901417285271,"accepted":true,"accepted_predicted":"2","input_tokens":210,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.17077266286943177,"brier_hard":0.04816326726083129,"score_mae":0.5179438113871377,"within_one":true,"case_id":"typed:73/r0","base_id":"typed:73","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.14329999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0001249873240127517,"human_review":0.0002624366636188894,"observe":0.9995392595563648,"stop":7.331645600408032e-05},"predicted":"observe","confidence":0.9995392595563648,"entropy_confidence":0.9967929753348503,"candidate_mass":0.9999836177586737,"accepted":true,"accepted_predicted":"observe","input_tokens":214,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":8.987298233702088,"brier_hard":1.9988288466165967,"soft_accuracy":0.34992685507231625,"brier_soft":0.7842251111199897,"tvd":0.6496125760123687,"kl":4.981095544350291,"case_id":"typed:74/r0","base_id":"typed:74","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.4592}
{"id":"needs_review","type":"binary","probabilities":{"false":4.906261171040432e-08,"true":0.9999999509373877},"predicted":"true","confidence":0.9999999509373877,"entropy_confidence":0.9999987379381162,"candidate_mass":0.9999997715741825,"accepted":true,"accepted_predicted":"true","input_tokens":176,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.830168564506415,"brier_hard":1.9999998037495568,"soft_accuracy":0.08333304088554241,"brier_soft":1.6805565978816954,"tvd":0.9166669509373879,"kl":8.235740022667377,"case_id":"typed:74/r0","base_id":"typed:74","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.4592}
{"id":"outcome","type":"choice","probabilities":{"failure":9.061691875422653e-05,"harmful":3.738005105675173e-05,"partial":0.01117285924876893,"success":0.9886991437814202},"predicted":"success","confidence":0.9886991437814202,"entropy_confidence":0.9547894088229327,"candidate_mass":0.9999907768396851,"accepted":true,"accepted_predicted":"success","input_tokens":219,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.011365196083854612,"brier_hard":0.00025255174375999646,"soft_accuracy":0.6445332034619037,"brier_soft":0.17247331016079306,"tvd":0.3387365238324769,"kl":1.5734052817895399,"case_id":"typed:74/r0","base_id":"typed:74","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.4592}
{"id":"risk","type":"ordinal","probabilities":{"0":0.004229294620008703,"1":0.9780590765960097,"2":0.016424126447924584,"3":0.0012875023360555053},"predicted":"1","confidence":0.9780590765960097,"entropy_confidence":0.9128108873053058,"candidate_mass":0.9999852626401095,"accepted":true,"accepted_predicted":"1","input_tokens":227,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":5.465720056328486,"brier_hard":1.9484302645966474,"score_mae":0.7981028365000253,"within_one":true,"case_id":"typed:74/r0","base_id":"typed:74","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.4592}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007074479719205052,"1":0.007140284931872517,"2":0.9909230656694308,"3":0.0012292014267763877},"predicted":"2","confidence":0.9909230656694308,"entropy_confidence":0.9583840201150841,"candidate_mass":0.9999939081072744,"accepted":true,"accepted_predicted":"2","input_tokens":215,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.253846469072474,"brier_hard":1.980566621219551,"score_mae":1.8593410205510632,"within_one":false,"case_id":"typed:74/r0","base_id":"typed:74","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.4592}
{"id":"action","type":"choice","probabilities":{"continue":0.0007930887622923762,"human_review":0.00610194779143956,"observe":0.9927560993892913,"stop":0.00034886405697855376},"predicted":"observe","confidence":0.9927560993892913,"entropy_confidence":0.9662612101416831,"candidate_mass":0.9999716078826189,"accepted":true,"accepted_predicted":"observe","input_tokens":227,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0072702650568459305,"brier_hard":9.045855882240563e-05,"soft_accuracy":0.4479653685056505,"brier_soft":0.4618938981041039,"tvd":0.5427560993892904,"kl":2.5579658098661024,"case_id":"typed:75/r0","base_id":"typed:75","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.84570000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":8.712572532891554e-08,"true":0.9999999128742731},"predicted":"true","confidence":0.9999999128742731,"entropy_confidence":0.9999978310032382,"candidate_mass":0.9999997479719729,"accepted":true,"accepted_predicted":"true","input_tokens":189,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.25591364274783,"brier_hard":1.9999996514971108,"soft_accuracy":0.23333304646711123,"brier_soft":1.175556310592339,"tvd":0.7666669128742739,"kl":6.721697173112226,"case_id":"typed:75/r0","base_id":"typed:75","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.84570000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":3.943980117279918e-05,"harmful":1.628720595868193e-05,"partial":0.0015046161993602994,"success":0.998439656793508},"predicted":"success","confidence":0.998439656793508,"entropy_confidence":0.991403362410403,"candidate_mass":0.9999912244707163,"accepted":true,"accepted_predicted":"success","input_tokens":232,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0015615618097442015,"brier_hard":4.700361600417689e-06,"soft_accuracy":0.4665947831484907,"brier_soft":0.47925002542963496,"tvd":0.5317889439994667,"kl":2.8828296370179634,"case_id":"typed:75/r0","base_id":"typed:75","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.84570000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.996705833337086,"1":0.0029402134003366316,"2":0.00021892808950728658,"3":0.00013502517307090303},"predicted":"0","confidence":0.996705833337086,"entropy_confidence":0.9830656608585578,"candidate_mass":0.9999944024580915,"accepted":true,"accepted_predicted":"0","input_tokens":240,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.829273115117207,"brier_hard":1.9875508024236468,"score_mae":0.679549854901436,"within_one":true,"case_id":"typed:75/r0","base_id":"typed:75","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.84570000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00019083839829412823,"1":0.9983110460902765,"2":0.001449538593213389,"3":4.857691821633429e-05},"predicted":"1","confidence":0.9983110460902765,"entropy_confidence":0.9904210171986207,"candidate_mass":0.9999944591376739,"accepted":true,"accepted_predicted":"1","input_tokens":228,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.564083570276928,"brier_hard":1.9962454078904184,"score_mae":0.43468885403135227,"within_one":true,"case_id":"typed:75/r0","base_id":"typed:75","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.84570000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.0036251722194330567,"human_review":0.018321273160757434,"observe":0.9750094471808215,"stop":0.003044107438988717},"predicted":"observe","confidence":0.9750094471808215,"entropy_confidence":0.9019203438556977,"candidate_mass":0.9999520029320157,"accepted":true,"accepted_predicted":"observe","input_tokens":230,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.999692426476333,"brier_hard":1.9143589532842882,"soft_accuracy":0.33375918914933916,"brier_soft":0.618483157973863,"tvd":0.6416761138474879,"kl":1.8722892511003704,"case_id":"typed:76/r0","base_id":"typed:76","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.0176}
{"id":"needs_review","type":"binary","probabilities":{"false":4.3140316079435355e-08,"true":0.9999999568596829},"predicted":"true","confidence":0.9999999568596829,"entropy_confidence":0.9999988822739676,"candidate_mass":0.9999997384244765,"accepted":true,"accepted_predicted":"true","input_tokens":192,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.3140318026945266e-08,"brier_hard":3.722173830611682e-15,"soft_accuracy":0.6166669899338968,"brier_soft":0.29388831162957585,"tvd":0.38333295685968344,"kl":3.2325145040321406,"case_id":"typed:76/r0","base_id":"typed:76","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.0176}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0014003570455788392,"harmful":0.0005001676825840795,"partial":0.1898736718047719,"success":0.8082258034670651},"predicted":"success","confidence":0.8082258034670651,"entropy_confidence":0.6389358546739493,"candidate_mass":0.999953140874211,"accepted":true,"accepted_predicted":"success","input_tokens":235,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":1.661396313193908,"brier_hard":1.309535828192631,"soft_accuracy":0.24576322681210022,"brier_soft":0.6064007937339745,"tvd":0.6415588034670652,"kl":1.6123109923021568,"case_id":"typed:76/r0","base_id":"typed:76","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.0176}
{"id":"risk","type":"ordinal","probabilities":{"0":1.4554121804636508e-05,"1":1.6867192490938665e-05,"2":0.9999006885726655,"3":6.78901130399725e-05},"predicted":"2","confidence":0.9999006885726655,"entropy_confidence":0.999207699773423,"candidate_mass":0.9999974950597345,"accepted":true,"accepted_predicted":"2","input_tokens":243,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":9.93163590408663e-05,"brier_hard":1.4968151691835915e-08,"score_mae":0.36668891467694165,"within_one":true,"case_id":"typed:76/r0","base_id":"typed:76","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.0176}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008196031620694618,"1":0.012197461825602602,"2":0.9816904700082164,"3":0.005292465004112933},"predicted":"2","confidence":0.9816904700082164,"entropy_confidence":0.9239308048336218,"candidate_mass":0.9999905808814634,"accepted":true,"accepted_predicted":"2","input_tokens":231,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.01847922397240613,"brier_hard":0.00051269889867009,"score_mae":0.2581227968543742,"within_one":true,"case_id":"typed:76/r0","base_id":"typed:76","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":204.0176}
{"id":"action","type":"choice","probabilities":{"continue":0.05001211478987834,"human_review":0.05799393686594172,"observe":0.13956273365032026,"stop":0.7524312146938593},"predicted":"stop","confidence":0.7524312146938593,"entropy_confidence":0.4201792842151134,"candidate_mass":0.999931376236753,"accepted":false,"accepted_predicted":null,"input_tokens":231,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":null,"nll_hard":2.8474168106941775,"brier_hard":1.4755071240767101,"soft_accuracy":0.13668799338791515,"brier_soft":0.9053412888107635,"tvd":0.6920060631340581,"kl":1.6835577208684456,"case_id":"typed:77/r0","base_id":"typed:77","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":217.2803}
{"id":"needs_review","type":"binary","probabilities":{"false":4.5097538367968175e-08,"true":0.9999999549024611},"predicted":"true","confidence":0.9999999549024611,"entropy_confidence":0.9999988344509145,"candidate_mass":0.9999998025303061,"accepted":true,"accepted_predicted":"true","input_tokens":193,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.50975399423478e-08,"brier_hard":4.067575983983179e-15,"soft_accuracy":0.8933329645232995,"brier_soft":0.02275567853632743,"tvd":0.10666695490246131,"kl":0.8816751462400898,"case_id":"typed:77/r0","base_id":"typed:77","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":217.2803}
{"id":"outcome","type":"choice","probabilities":{"failure":3.368369549360768e-06,"harmful":0.9999734016959546,"partial":2.193108073824416e-05,"success":1.2988537583061043e-06},"predicted":"harmful","confidence":0.9999734016959546,"entropy_confidence":0.9997677868804912,"candidate_mass":0.9999992288508028,"accepted":true,"accepted_predicted":"harmful","input_tokens":236,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":2.6598657786550433e-05,"brier_hard":1.2014750149448553e-09,"soft_accuracy":0.6333215633010868,"brier_soft":0.19483736772512847,"tvd":0.366639768362321,"kl":3.0634842748360374,"case_id":"typed:77/r0","base_id":"typed:77","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":217.2803}
{"id":"risk","type":"ordinal","probabilities":{"0":1.8589430749602136e-06,"1":8.705141168106317e-07,"2":0.0001171393127054026,"3":0.9998801312301044},"predicted":"3","confidence":0.9998801312301044,"entropy_confidence":0.9991221916110821,"candidate_mass":0.9999995799668605,"accepted":true,"accepted_predicted":"3","input_tokens":244,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00011987595473072128,"brier_hard":2.8094354041552118e-08,"score_mae":0.5732085428298412,"within_one":true,"case_id":"typed:77/r0","base_id":"typed:77","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":217.2803}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000308277783967918,"1":0.0003859124952452358,"2":0.04725923252003277,"3":0.9520465772007541},"predicted":"3","confidence":0.9520465772007541,"entropy_confidence":0.8582187268187512,"candidate_mass":0.9999909115469626,"accepted":true,"accepted_predicted":"3","input_tokens":232,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.049141319756547126,"brier_hard":0.00453320978019183,"score_mae":0.7177111091375732,"within_one":true,"case_id":"typed:77/r0","base_id":"typed:77","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":217.2803}
{"id":"action","type":"choice","probabilities":{"continue":0.005815850110593939,"human_review":0.014462368882772044,"observe":0.9761113056003157,"stop":0.0036104754063199735},"predicted":"observe","confidence":0.9761113056003157,"entropy_confidence":0.9025410505248457,"candidate_mass":0.9999450434880455,"accepted":true,"accepted_predicted":"observe","input_tokens":225,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.024178656448135914,"brier_hard":0.0008266894789914195,"soft_accuracy":0.38096672034598733,"brier_soft":0.5211158866543714,"tvd":0.5927779222669315,"kl":1.6620692953471505,"case_id":"typed:78/r0","base_id":"typed:78","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":210.4048}
{"id":"needs_review","type":"binary","probabilities":{"false":9.71307602302437e-07,"true":0.999999028692396},"predicted":"true","confidence":0.999999028692396,"entropy_confidence":0.9999791982211745,"candidate_mass":0.9999989009281975,"accepted":true,"accepted_predicted":"true","input_tokens":187,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.713080757336295e-07,"brier_hard":1.8868769199061083e-12,"soft_accuracy":0.6333327409852859,"brier_soft":0.26888795319410713,"tvd":0.36666602869239684,"kl":3.087847985671784,"case_id":"typed:78/r0","base_id":"typed:78","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":210.4048}
{"id":"outcome","type":"choice","probabilities":{"failure":0.005904624074743204,"harmful":0.017620455063087563,"partial":0.28262127455547625,"success":0.6938536463066922},"predicted":"success","confidence":0.6938536463066922,"entropy_confidence":0.48625668336685457,"candidate_mass":0.9999510702933311,"accepted":false,"accepted_predicted":null,"input_tokens":230,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":null,"nll_hard":0.36549422499885503,"brier_hard":0.1739457197331641,"soft_accuracy":0.4863706668721688,"brier_soft":0.024644793802544876,"tvd":0.1271872129731259,"kl":0.16791745376327477,"case_id":"typed:78/r0","base_id":"typed:78","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":210.4048}
{"id":"risk","type":"ordinal","probabilities":{"0":6.245810125248827e-06,"1":7.490073897992007e-06,"2":0.9999780295315659,"3":8.234584412328312e-06},"predicted":"2","confidence":0.9999780295315659,"entropy_confidence":0.9997968551804849,"candidate_mass":0.99999881227245,"accepted":true,"accepted_predicted":"2","input_tokens":238,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":2.1970709788383828e-05,"brier_hard":6.45621214775899e-10,"score_mae":0.21665525289026677,"within_one":true,"case_id":"typed:78/r0","base_id":"typed:78","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":210.4048}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0016667143292210386,"1":0.0042986583025970086,"2":0.973608775767211,"3":0.020425851600969776},"predicted":"2","confidence":0.973608775767211,"entropy_confidence":0.8992976793655416,"candidate_mass":0.9999915074435163,"accepted":true,"accepted_predicted":"2","input_tokens":226,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.026745723619768593,"brier_hard":0.001134968529987903,"score_mae":0.2127937646399285,"within_one":true,"case_id":"typed:78/r0","base_id":"typed:78","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":210.4048}
{"id":"action","type":"choice","probabilities":{"continue":0.08653068881250672,"human_review":0.015979876388540995,"observe":0.8971245113803852,"stop":0.0003649234185683685},"predicted":"observe","confidence":0.8971245113803852,"entropy_confidence":0.7272272092436587,"candidate_mass":0.9999478590315775,"accepted":true,"accepted_predicted":"observe","input_tokens":224,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":2.4472561440056353,"brier_hard":1.6395140610193428,"soft_accuracy":0.2649761523489092,"brier_soft":0.8814452317245379,"tvd":0.6741564347989529,"kl":1.3080540438934665,"case_id":"typed:79/r0","base_id":"typed:79","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.94219999999999}
{"id":"needs_review","type":"binary","probabilities":{"false":4.345667554363985e-06,"true":0.9999956543324473},"predicted":"true","confidence":0.9999956543324473,"entropy_confidence":0.9999163255502816,"candidate_mass":0.9999981305273091,"accepted":true,"accepted_predicted":"true","input_tokens":186,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.346331173679925,"brier_hard":1.9999826173675554,"soft_accuracy":0.08333662139252589,"brier_soft":1.6805408436956126,"tvd":0.9166626543324465,"kl":8.235740380717143,"case_id":"typed:79/r0","base_id":"typed:79","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.94219999999999}
{"id":"outcome","type":"choice","probabilities":{"failure":9.971049215647867e-07,"harmful":5.180573061749721e-07,"partial":7.937042503160986e-06,"success":0.9999905477952679},"predicted":"success","confidence":0.9999905477952679,"entropy_confidence":0.9999105955471382,"candidate_mass":0.9999991771600608,"accepted":true,"accepted_predicted":"success","input_tokens":229,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":9.452249404509684e-06,"brier_hard":1.5360341959238336e-10,"soft_accuracy":0.7899940144698718,"brier_soft":0.07841518458239556,"tvd":0.20999106585257463,"kl":1.7479583019103828,"case_id":"typed:79/r0","base_id":"typed:79","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.94219999999999}
{"id":"risk","type":"ordinal","probabilities":{"0":0.9999373789023461,"1":4.3786920303002266e-05,"2":1.0936034656690768e-05,"3":7.898142695803827e-06},"predicted":"0","confidence":0.9999373789023461,"entropy_confidence":0.9994807784412951,"candidate_mass":0.9999991658751132,"accepted":true,"accepted_predicted":"0","input_tokens":237,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":6.262305843666362e-05,"brier_hard":6.0206737730526055e-09,"score_mae":0.5865776465822963,"within_one":true,"case_id":"typed:79/r0","base_id":"typed:79","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.94219999999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.469318397074283e-05,"1":0.9998380984674861,"2":7.896790159777542e-05,"3":8.24044694683284e-06},"predicted":"1","confidence":0.9998380984674861,"entropy_confidence":0.9987635624361528,"candidate_mass":0.9999983750645128,"accepted":true,"accepted_predicted":"1","input_tokens":225,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00016191463998182205,"brier_hard":3.809501241068966e-08,"score_mae":0.19335375561152213,"within_one":true,"case_id":"typed:79/r0","base_id":"typed:79","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":202.94219999999999}
{"id":"action","type":"choice","probabilities":{"continue":0.0006985818826002664,"human_review":0.003753160316487299,"observe":0.9941486261773463,"stop":0.0013996316235672039},"predicted":"observe","confidence":0.9941486261773463,"entropy_confidence":0.9703741341843629,"candidate_mass":0.9999611352912781,"accepted":true,"accepted_predicted":"observe","input_tokens":228,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.585157042851826,"brier_hard":1.98084170349502,"soft_accuracy":0.25092889250366734,"brier_soft":0.92149023912066,"tvd":0.7441486261773458,"kl":3.374670404370799,"case_id":"typed:80/r0","base_id":"typed:80","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.3721}
{"id":"needs_review","type":"binary","probabilities":{"false":1.1713815096943977e-07,"true":0.9999998828618499},"predicted":"true","confidence":0.9999998828618499,"entropy_confidence":0.9999971338659965,"candidate_mass":0.9999996325934113,"accepted":true,"accepted_predicted":"true","input_tokens":190,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.1713815692931445e-07,"brier_hard":2.744269261404268e-14,"soft_accuracy":0.7833329336217932,"brier_soft":0.09388907625814086,"tvd":0.21666688286184954,"kl":1.804289038131106,"case_id":"typed:80/r0","base_id":"typed:80","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.3721}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0006692990126086642,"harmful":0.011057783018939017,"partial":0.9881317399903248,"success":0.00014117797812903375},"predicted":"partial","confidence":0.9881317399903248,"entropy_confidence":0.9511269913584709,"candidate_mass":0.9999812412306459,"accepted":true,"accepted_predicted":"partial","input_tokens":233,"reused_prefix_tokens":0,"expected":"failure","correct":false,"accepted_correct":false,"nll_hard":7.309279643118494,"brier_hard":1.9751884800087731,"soft_accuracy":0.4058452521161944,"brier_soft":0.5872588115796016,"tvd":0.5781317399903241,"kl":3.323900642133246,"case_id":"typed:80/r0","base_id":"typed:80","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.3721}
{"id":"risk","type":"ordinal","probabilities":{"0":5.182540813754547e-06,"1":7.067781758397855e-06,"2":0.9999175230602217,"3":7.0226617206861e-05},"predicted":"2","confidence":0.9999175230602217,"entropy_confidence":0.999350064477334,"candidate_mass":0.9999982345911058,"accepted":true,"accepted_predicted":"2","input_tokens":241,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":8.248034118807734e-05,"brier_hard":1.1811035627775167e-08,"score_mae":0.4833857937538224,"within_one":true,"case_id":"typed:80/r0","base_id":"typed:80","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.3721}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00031198980436226666,"1":0.0002922733792665332,"2":0.8761139341323148,"3":0.12328180268405733},"predicted":"2","confidence":0.8761139341323148,"entropy_confidence":0.7267285102331981,"candidate_mass":0.9999828511401391,"accepted":true,"accepted_predicted":"2","input_tokens":229,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.072540049014808,"brier_hard":1.7821502316064746,"score_mae":1.2056985496960682,"within_one":false,"case_id":"typed:80/r0","base_id":"typed:80","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.3721}
{"id":"action","type":"choice","probabilities":{"continue":0.0006555179143125744,"human_review":0.001179991897961861,"observe":0.9979563345448743,"stop":0.00020815564284969824},"predicted":"observe","confidence":0.9979563345448743,"entropy_confidence":0.9880494662807856,"candidate_mass":0.9999688059860105,"accepted":true,"accepted_predicted":"observe","input_tokens":218,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0020457565889095805,"brier_hard":6.041981879364596e-06,"soft_accuracy":0.4329625881766909,"brier_soft":0.46054901249624625,"tvd":0.5646233345448751,"kl":2.8276268447987727,"case_id":"typed:81/r0","base_id":"typed:81","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.799}
{"id":"needs_review","type":"binary","probabilities":{"false":6.265904692106712e-08,"true":0.9999999373409543},"predicted":"true","confidence":0.9999999373409543,"entropy_confidence":0.9999984103025783,"candidate_mass":0.9999997279576377,"accepted":true,"accepted_predicted":"true","input_tokens":180,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.265904769285186e-08,"brier_hard":7.852312172822372e-15,"soft_accuracy":0.5499999937340959,"brier_soft":0.40499988721372443,"tvd":0.44999993734095367,"kl":3.815842851436167,"case_id":"typed:81/r0","base_id":"typed:81","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.799}
{"id":"outcome","type":"choice","probabilities":{"failure":3.6243209317053204e-05,"harmful":2.2773120148755884e-05,"partial":0.002369595503322426,"success":0.9975713881672117},"predicted":"success","confidence":0.9975713881672117,"entropy_confidence":0.9874745290459563,"candidate_mass":0.9999900355366727,"accepted":true,"accepted_predicted":"success","input_tokens":223,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":6.045036012105129,"brier_hard":1.9904151002982475,"soft_accuracy":0.24067312865335644,"brier_soft":0.8898972517761795,"tvd":0.7575713881672117,"kl":4.470223800652713,"case_id":"typed:81/r0","base_id":"typed:81","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.799}
{"id":"risk","type":"ordinal","probabilities":{"0":0.02573331629051225,"1":0.9720756358193119,"2":0.001427427746348683,"3":0.0007636201438279718},"predicted":"1","confidence":0.9720756358193119,"entropy_confidence":0.901502082630533,"candidate_mass":0.9999869943958734,"accepted":true,"accepted_predicted":"1","input_tokens":231,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.028321662920177466,"brier_hard":0.0014445943478983462,"score_mae":0.18722135174349308,"within_one":true,"case_id":"typed:81/r0","base_id":"typed:81","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.799}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0011623774919709373,"1":0.13483031278372173,"2":0.8613883711656084,"3":0.0026189385586977963},"predicted":"2","confidence":0.8613883711656084,"entropy_confidence":0.6955074409500344,"candidate_mass":0.9999917132986923,"accepted":true,"accepted_predicted":"2","input_tokens":219,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.0037382335011102,"brier_hard":1.490516723617861,"score_mae":0.6487968707910319,"within_one":true,"case_id":"typed:81/r0","base_id":"typed:81","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":201.799}
{"id":"action","type":"choice","probabilities":{"continue":0.002182757512928769,"human_review":0.005130706165446329,"observe":0.9918929376858575,"stop":0.0007935986357664209},"predicted":"observe","confidence":0.9918929376858575,"entropy_confidence":0.9609279330206993,"candidate_mass":0.999962053176328,"accepted":true,"accepted_predicted":"observe","input_tokens":228,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.272511975189808,"brier_hard":1.9736219058750988,"soft_accuracy":0.3169263695762921,"brier_soft":0.7172529568314073,"tvd":0.675225937685858,"kl":2.7152781332987743,"case_id":"typed:82/r0","base_id":"typed:82","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.9265}
{"id":"needs_review","type":"binary","probabilities":{"false":1.7779626807184452e-08,"true":0.9999999822203735},"predicted":"true","confidence":0.9999999822203735,"entropy_confidence":0.9999995166093781,"candidate_mass":0.9999998325336319,"accepted":true,"accepted_predicted":"true","input_tokens":190,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7779626706865943e-08,"brier_hard":6.322302496178438e-16,"soft_accuracy":0.6833329934808154,"brier_soft":0.2005559552571165,"tvd":0.31666698222037337,"kl":2.6564161217090763,"case_id":"typed:82/r0","base_id":"typed:82","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.9265}
{"id":"outcome","type":"choice","probabilities":{"failure":6.104077195415473e-05,"harmful":3.313495285077151e-05,"partial":0.0018232526598931538,"success":0.9980825716153003},"predicted":"success","confidence":0.9980825716153003,"entropy_confidence":0.9896492340281268,"candidate_mass":0.9999914363796738,"accepted":true,"accepted_predicted":"success","input_tokens":233,"reused_prefix_tokens":0,"expected":"failure","correct":false,"accepted_correct":false,"nll_hard":9.703968524374812,"brier_hard":1.9960500672924657,"soft_accuracy":0.03708995268614151,"brier_soft":1.4883246374657366,"tvd":0.9614156082822645,"kl":7.664049333402626,"case_id":"typed:82/r0","base_id":"typed:82","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.9265}
{"id":"risk","type":"ordinal","probabilities":{"0":0.7920286843232182,"1":0.2067537625320411,"2":0.000842697285054452,"3":0.00037485585968488757},"predicted":"0","confidence":0.7920286843232182,"entropy_confidence":0.6252736891553112,"candidate_mass":0.9999872912487434,"accepted":false,"accepted_predicted":null,"input_tokens":241,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":null,"nll_hard":0.233157670243731,"brier_hard":0.08600003712111708,"score_mae":0.12376927531879531,"within_one":true,"case_id":"typed:82/r0","base_id":"typed:82","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.9265}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000878589786918449,"1":0.3550025408432068,"2":0.6429246515072129,"3":0.0011942178626632807},"predicted":"2","confidence":0.6429246515072129,"entropy_confidence":0.5196768588516207,"candidate_mass":0.9999867771805924,"accepted":false,"accepted_predicted":null,"input_tokens":229,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.0356303322275267,"brier_hard":0.8293760279107074,"score_mae":0.6277674974456224,"within_one":true,"case_id":"typed:82/r0","base_id":"typed:82","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.9265}
{"id":"action","type":"choice","probabilities":{"continue":0.0020239986820629226,"human_review":0.009344854997913054,"observe":0.9882086769082011,"stop":0.00042246941182227623},"predicted":"observe","confidence":0.9882086769082011,"entropy_confidence":0.9486213552968904,"candidate_mass":0.9999615588153062,"accepted":true,"accepted_predicted":"observe","input_tokens":214,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.01186139209071866,"brier_hard":0.00023063666625612765,"soft_accuracy":0.51246920089324,"brier_soft":0.3572425442968617,"tvd":0.47154219357468474,"kl":1.817021731869692,"case_id":"typed:83/r0","base_id":"typed:83","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.15789999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":1.256092953765093e-07,"true":0.9999998743907036},"predicted":"true","confidence":0.9999998743907036,"entropy_confidence":0.9999969392470988,"candidate_mass":0.9999996412419141,"accepted":true,"accepted_predicted":"true","input_tokens":176,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.890089577875997,"brier_hard":1.999999497562848,"soft_accuracy":0.16666708373944636,"brier_soft":1.3888873590805462,"tvd":0.833332874390704,"kl":7.376653752010789,"case_id":"typed:83/r0","base_id":"typed:83","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.15789999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":1.1449493419795943e-05,"harmful":6.7342514179343895e-06,"partial":0.00016936035111556289,"success":0.999812455904045},"predicted":"success","confidence":0.999812455904045,"entropy_confidence":0.9986520720518681,"candidate_mass":0.9999958792371857,"accepted":true,"accepted_predicted":"success","input_tokens":219,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.000187561684548095,"brier_hard":6.403215749929702e-08,"soft_accuracy":0.5666235713776667,"brier_soft":0.3259561886409141,"tvd":0.43314545590404585,"kl":3.0608144394627854,"case_id":"typed:83/r0","base_id":"typed:83","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.15789999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0025686416867356132,"1":0.9945658343210863,"2":0.002240370640051666,"3":0.0006251533521272961},"predicted":"1","confidence":0.9945658343210863,"entropy_confidence":0.9718526173827128,"candidate_mass":0.9999912642246694,"accepted":true,"accepted_predicted":"1","input_tokens":227,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.005448984466774168,"brier_hard":4.1538154059201615e-05,"score_mae":0.010922035657571438,"within_one":true,"case_id":"typed:83/r0","base_id":"typed:83","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.15789999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002587956803699787,"1":0.8672614072545985,"2":0.1321734206593433,"3":0.00030637640568660895},"predicted":"1","confidence":0.8672614072545985,"entropy_confidence":0.7146358689921419,"candidate_mass":0.9999878838677515,"accepted":true,"accepted_predicted":"1","input_tokens":215,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.14241483985516373,"brier_hard":0.035089507974527395,"score_mae":0.1658603777903448,"within_one":true,"case_id":"typed:83/r0","base_id":"typed:83","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.15789999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.052892293417926194,"human_review":0.00794003933809576,"observe":0.9388034526098247,"stop":0.00036421463415340494},"predicted":"observe","confidence":0.9388034526098247,"entropy_confidence":0.8153048084724608,"candidate_mass":0.9999496479727855,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":2.9394976328113387,"brier_hard":1.778428107376273,"soft_accuracy":0.417951285294014,"brier_soft":0.47282123918009766,"tvd":0.5221364526098247,"kl":0.9776364556096014,"case_id":"typed:84/r0","base_id":"typed:84","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.11839899999998}
{"id":"needs_review","type":"binary","probabilities":{"false":2.400452672202631e-07,"true":0.9999997599547332},"predicted":"true","confidence":0.9999997599547332,"entropy_confidence":0.9999943750468494,"candidate_mass":0.999999502232995,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.242438317971931,"brier_hard":1.9999990398190473,"soft_accuracy":0.1233331808342614,"brier_soft":1.537089216019059,"tvd":0.876666759954733,"kl":7.816281894125245,"case_id":"typed:84/r0","base_id":"typed:84","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.11839899999998}
{"id":"outcome","type":"choice","probabilities":{"failure":7.231493972477191e-07,"harmful":2.984597722597003e-07,"partial":3.4028403527322643e-06,"success":0.9999955755504788},"predicted":"success","confidence":0.9999955755504788,"entropy_confidence":0.9999552918979129,"candidate_mass":0.9999994015333599,"accepted":true,"accepted_predicted":"success","input_tokens":214,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":12.590900077230305,"brier_hard":1.9999843454520192,"soft_accuracy":0.3100004670395654,"brier_soft":0.8608563819877832,"tvd":0.6899958855501683,"kl":5.992072412248155,"case_id":"typed:84/r0","base_id":"typed:84","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.11839899999998}
{"id":"risk","type":"ordinal","probabilities":{"0":0.1420500411496243,"1":0.8565194400274261,"2":0.0010172238745407734,"3":0.0004132949484088436},"predicted":"1","confidence":0.8565194400274261,"entropy_confidence":0.6969567375195593,"candidate_mass":0.9999870109016412,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":1.951575880997897,"brier_hard":1.4697048885933823,"score_mae":0.7497937726217342,"within_one":true,"case_id":"typed:84/r0","base_id":"typed:84","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.11839899999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00035851068277290243,"1":0.9844754749920315,"2":0.014954231018421742,"3":0.00021178330677312383},"predicted":"1","confidence":0.9844754749920315,"entropy_confidence":0.9402086662412136,"candidate_mass":0.9999919954768239,"accepted":true,"accepted_predicted":"1","input_tokens":210,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.9335520998388525,"brier_hard":1.9686987419026711,"score_mae":0.8016862869491943,"within_one":true,"case_id":"typed:84/r0","base_id":"typed:84","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":190.11839899999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0013318781986137243,"human_review":0.0019925790129928437,"observe":0.9964178767600461,"stop":0.0002576660283476448},"predicted":"observe","confidence":0.9964178767600461,"entropy_confidence":0.9805855788236767,"candidate_mass":0.9999676072368252,"accepted":true,"accepted_predicted":"observe","input_tokens":230,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.621165153527247,"brier_hard":1.9901906393922122,"soft_accuracy":0.43259521206529966,"brier_soft":0.5294193094368402,"tvd":0.5630848767600459,"kl":2.720528202328228,"case_id":"typed:85/r0","base_id":"typed:85","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.0444}
{"id":"needs_review","type":"binary","probabilities":{"false":3.7342045058286014e-06,"true":0.9999962657954935},"predicted":"true","confidence":0.9999962657954935,"entropy_confidence":0.9999272821180095,"candidate_mass":0.9999974621308264,"accepted":true,"accepted_predicted":"true","input_tokens":192,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.49797574577764,"brier_hard":1.999985063209864,"soft_accuracy":0.18333536499867636,"brier_soft":1.333877779399523,"tvd":0.8166632657954939,"kl":7.210766254253917,"case_id":"typed:85/r0","base_id":"typed:85","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.0444}
{"id":"outcome","type":"choice","probabilities":{"failure":1.6061378961882092e-05,"harmful":8.439917663154495e-06,"partial":0.0008305512219423351,"success":0.9991449474814315},"predicted":"success","confidence":0.9991449474814315,"entropy_confidence":0.9949346695691891,"candidate_mass":0.9999937296607776,"accepted":true,"accepted_predicted":"success","input_tokens":235,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":7.093420954783511,"brier_hard":1.99663021377832,"soft_accuracy":0.433351572895014,"brier_soft":0.5448326859881767,"tvd":0.565811947481432,"kl":3.435559700239343,"case_id":"typed:85/r0","base_id":"typed:85","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.0444}
{"id":"risk","type":"ordinal","probabilities":{"0":0.995398752806797,"1":0.004163721443285695,"2":0.0002904378149953231,"3":0.00014708793492225966},"predicted":"0","confidence":0.995398752806797,"entropy_confidence":0.9775828429339696,"candidate_mass":0.9999944472176965,"accepted":true,"accepted_predicted":"0","input_tokens":243,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00461186551527248,"brier_hard":3.86140409752144e-05,"score_mae":0.4348141391219569,"within_one":true,"case_id":"typed:85/r0","base_id":"typed:85","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.0444}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00017740603627155457,"1":0.9996560225598,"2":0.00014962612134990905,"3":1.694528258041525e-05},"predicted":"1","confidence":0.9996560225598,"entropy_confidence":0.9975617357090654,"candidate_mass":0.9999982975151728,"accepted":true,"accepted_predicted":"1","input_tokens":231,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.637069462334225,"brier_hard":1.9989574055155568,"score_mae":0.3100061106502412,"within_one":true,"case_id":"typed:85/r0","base_id":"typed:85","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.0444}
{"id":"action","type":"choice","probabilities":{"continue":0.0003839763417526642,"human_review":0.0024857250446005826,"observe":0.9965779450543093,"stop":0.0005523535593375104},"predicted":"observe","confidence":0.9965779450543093,"entropy_confidence":0.981615095417854,"candidate_mass":0.9999686420198951,"accepted":true,"accepted_predicted":"observe","input_tokens":226,"reused_prefix_tokens":0,"expected":"stop","correct":false,"accepted_correct":false,"nll_hard":7.501322210677415,"brier_hard":1.9920695248112779,"soft_accuracy":0.10098990003157271,"brier_soft":1.1917497096448075,"tvd":0.8965779450543093,"kl":5.085602012079476,"case_id":"typed:86/r0","base_id":"typed:86","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.6427}
{"id":"needs_review","type":"binary","probabilities":{"false":3.3833353250827115e-08,"true":0.9999999661666484},"predicted":"true","confidence":0.9999999661666484,"entropy_confidence":0.9999991115471891,"candidate_mass":0.9999997289327508,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.383335213716378e-08,"brier_hard":2.2893914703036862e-15,"soft_accuracy":0.9166669718055179,"brier_soft":0.013888766500263262,"tvd":0.08333296616664759,"kl":0.6877652006542495,"case_id":"typed:86/r0","base_id":"typed:86","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.6427}
{"id":"outcome","type":"choice","probabilities":{"failure":1.547773735913732e-05,"harmful":0.9998666073977822,"partial":0.00011138690323221676,"success":6.527961625299409e-06},"predicted":"harmful","confidence":0.9998666073977822,"entropy_confidence":0.9989925255003004,"candidate_mass":0.9999978617861893,"accepted":true,"accepted_predicted":"harmful","input_tokens":231,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00013340149980220557,"brier_hard":3.048280317483541e-08,"soft_accuracy":0.7665815920484342,"brier_soft":0.07972607895949907,"tvd":0.23319960739778278,"kl":1.5766827484572379,"case_id":"typed:86/r0","base_id":"typed:86","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.6427}
{"id":"risk","type":"ordinal","probabilities":{"0":2.5919161335809127e-06,"1":1.8666931035607141e-06,"2":0.0006397501859644991,"3":0.9993557912047977},"predicted":"3","confidence":0.9993557912047977,"entropy_confidence":0.9960996989931923,"candidate_mass":0.999999249162561,"accepted":true,"accepted_predicted":"3","input_tokens":239,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.000644416386847933,"brier_hard":8.242954748300324e-07,"score_mae":0.4426817406794257,"within_one":true,"case_id":"typed:86/r0","base_id":"typed:86","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.6427}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003171182083841378,"1":0.007955038860847644,"2":0.9908660268323983,"3":0.0008618160983712615},"predicted":"2","confidence":0.9908660268323983,"entropy_confidence":0.9594728595988509,"candidate_mass":0.9999942806659933,"accepted":true,"accepted_predicted":"2","input_tokens":227,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":7.0564686530182,"brier_hard":1.9801559768681039,"score_mae":0.3743944591792421,"within_one":true,"case_id":"typed:86/r0","base_id":"typed:86","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.6427}
{"id":"action","type":"choice","probabilities":{"continue":0.0006325025209561173,"human_review":0.0009292743891837577,"observe":0.9982327526482281,"stop":0.00020547044163239382},"predicted":"observe","confidence":0.9982327526482281,"entropy_confidence":0.9894276141187722,"candidate_mass":0.999971222664573,"accepted":true,"accepted_predicted":"observe","input_tokens":209,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.0017688107756157118,"brier_hard":4.428991634138153e-06,"soft_accuracy":0.4163326624136877,"brier_soft":0.4960269872387148,"tvd":0.5815657526482279,"kl":3.090135418003666,"case_id":"typed:87/r0","base_id":"typed:87","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.09990000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.264187283849181e-08,"true":0.9999999773581261},"predicted":"true","confidence":0.9999999773581261,"entropy_confidence":0.9999993923119094,"candidate_mass":0.9999998662513062,"accepted":true,"accepted_predicted":"true","input_tokens":171,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.603464864484526,"brier_hard":1.9999999094325076,"soft_accuracy":0.3333330075473057,"brier_soft":0.8888897173996418,"tvd":0.6666669773581266,"kl":5.774025928959757,"case_id":"typed:87/r0","base_id":"typed:87","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.09990000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":3.005610201450452e-05,"harmful":1.3137252171862896e-05,"partial":0.004120227801758426,"success":0.9958365788440549},"predicted":"success","confidence":0.9958365788440549,"entropy_confidence":0.980348312889159,"candidate_mass":0.9999873606615651,"accepted":true,"accepted_predicted":"success","input_tokens":214,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":5.491846825460293,"brier_hard":1.9834670135134096,"soft_accuracy":0.11202231266746349,"brier_soft":1.2067518815599994,"tvd":0.885836578844055,"kl":5.3310339303127465,"case_id":"typed:87/r0","base_id":"typed:87","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.09990000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.0674593914874259,"1":0.9307953496574007,"2":0.0013113836095151328,"3":0.0004338752456579063},"predicted":"1","confidence":0.9307953496574007,"entropy_confidence":0.811943701829252,"candidate_mass":0.9999892423574978,"accepted":true,"accepted_predicted":"1","input_tokens":222,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":2.6962294697358047,"brier_hard":1.736013877443545,"score_mae":0.6680527426134047,"within_one":true,"case_id":"typed:87/r0","base_id":"typed:87","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.09990000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.311515947809155e-05,"1":0.9844887494265617,"2":0.01540617056881029,"3":4.196484515150931e-05},"predicted":"1","confidence":0.9844887494265617,"entropy_confidence":0.9417776000550754,"candidate_mass":0.9999948897767804,"accepted":true,"accepted_predicted":"1","input_tokens":210,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":9.67054957197675,"brier_hard":1.9693292232646862,"score_mae":0.017906014900363143,"within_one":true,"case_id":"typed:87/r0","base_id":"typed:87","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":185.09990000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.006314310671130677,"human_review":0.03657631327410398,"observe":0.9481896249772538,"stop":0.008919751077510743},"predicted":"observe","confidence":0.9481896249772538,"entropy_confidence":0.8228874988764592,"candidate_mass":0.9999534424266057,"accepted":true,"accepted_predicted":"observe","input_tokens":226,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":3.308354426388263,"brier_hard":1.827368197537559,"soft_accuracy":0.08440096031492013,"brier_soft":1.1567184201234681,"tvd":0.8815226916441875,"kl":2.791858237715909,"case_id":"typed:88/r0","base_id":"typed:88","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.8537}
{"id":"needs_review","type":"binary","probabilities":{"false":5.2130786943972966e-08,"true":0.9999999478692145},"predicted":"true","confidence":0.9999999478692145,"entropy_confidence":0.9999986635759757,"candidate_mass":0.9999998163567829,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.213078682498511e-08,"brier_hard":5.435237740718334e-15,"soft_accuracy":0.8799999603806032,"brier_soft":0.028799974977228054,"tvd":0.1199999478692138,"kl":0.9927475235835357,"case_id":"typed:88/r0","base_id":"typed:88","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.8537}
{"id":"outcome","type":"choice","probabilities":{"failure":2.2168434082634453e-05,"harmful":0.9998536734570098,"partial":0.00011609116579975177,"success":8.066943107851748e-06},"predicted":"harmful","confidence":0.9998536734570098,"entropy_confidence":0.9988960381406004,"candidate_mass":0.9999971526209707,"accepted":true,"accepted_predicted":"harmful","input_tokens":231,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00014633724976324396,"brier_hard":3.5445131000984685e-08,"soft_accuracy":0.6832460567731198,"brier_soft":0.15932582436891096,"tvd":0.3165206734570098,"kl":2.439749091430958,"case_id":"typed:88/r0","base_id":"typed:88","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.8537}
{"id":"risk","type":"ordinal","probabilities":{"0":2.88252109274545e-06,"1":1.697456494136241e-06,"2":0.00025457891787644465,"3":0.9997408411045372},"predicted":"3","confidence":0.9997408411045372,"entropy_confidence":0.9982505015595166,"candidate_mass":0.9999994345389851,"accepted":true,"accepted_predicted":"3","input_tokens":239,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.000259192482932465,"brier_hard":1.3198494881103788e-07,"score_mae":0.38306637860585857,"within_one":true,"case_id":"typed:88/r0","base_id":"typed:88","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.8537}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006960781352302969,"1":0.0008886772749931766,"2":0.05071693488517198,"3":0.9476983097046053},"predicted":"3","confidence":0.9476983097046053,"entropy_confidence":0.8460457230089399,"candidate_mass":0.999992369244893,"accepted":true,"accepted_predicted":"3","input_tokens":227,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.05371906608402104,"brier_hard":0.00530894856397159,"score_mae":0.5120844761591528,"within_one":true,"case_id":"typed:88/r0","base_id":"typed:88","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":198.8537}
{"id":"action","type":"choice","probabilities":{"continue":0.0005300378880749397,"human_review":0.0016382433375715182,"observe":0.9975334481726884,"stop":0.00029827060166490887},"predicted":"observe","confidence":0.9975334481726884,"entropy_confidence":0.986012730788594,"candidate_mass":0.9999686789311741,"accepted":true,"accepted_predicted":"observe","input_tokens":215,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":6.4141307468488975,"brier_hard":1.9917995472948982,"soft_accuracy":0.3831683733037741,"brier_soft":0.578183842918493,"tvd":0.6142004481726885,"kl":3.056843228309126,"case_id":"typed:89/r0","base_id":"typed:89","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.9469}
{"id":"needs_review","type":"binary","probabilities":{"false":5.681522595782978e-08,"true":0.9999999431847736},"predicted":"true","confidence":0.9999999431847736,"entropy_confidence":0.9999985505389231,"candidate_mass":0.999999766154757,"accepted":true,"accepted_predicted":"true","input_tokens":177,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.681522802641723e-08,"brier_hard":6.455939852935188e-15,"soft_accuracy":0.5333329962123559,"brier_soft":0.43555607172284194,"tvd":0.46666694318477386,"kl":3.9629038651815036,"case_id":"typed:89/r0","base_id":"typed:89","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.9469}
{"id":"outcome","type":"choice","probabilities":{"failure":7.115873198649503e-05,"harmful":3.63556251359537e-05,"partial":0.0030416412395388817,"success":0.996850844403338},"predicted":"success","confidence":0.996850844403338,"entropy_confidence":0.9842581533440101,"candidate_mass":0.9999872408777776,"accepted":true,"accepted_predicted":"success","input_tokens":220,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":5.7953580278202885,"brier_hard":1.987637581475297,"soft_accuracy":0.20091117702807235,"brier_soft":0.9541208876762299,"tvd":0.7968508444033383,"kl":4.651572389738241,"case_id":"typed:89/r0","base_id":"typed:89","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.9469}
{"id":"risk","type":"ordinal","probabilities":{"0":0.04765432621137028,"1":0.9308440871394162,"2":0.015724473270422142,"3":0.005777113378791846},"predicted":"1","confidence":0.9308440871394162,"entropy_confidence":0.7786705683924757,"candidate_mass":0.9999706467454732,"accepted":true,"accepted_predicted":"1","input_tokens":228,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":3.043781861578868,"brier_hard":1.773713631044956,"score_mae":0.6129573738166361,"within_one":true,"case_id":"typed:89/r0","base_id":"typed:89","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.9469}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005245154262661341,"1":0.37477834788543984,"2":0.6239914357632376,"3":0.0007057009250571734},"predicted":"2","confidence":0.6239914357632376,"entropy_confidence":0.5158433614920703,"candidate_mass":0.9999820299501816,"accepted":false,"accepted_predicted":null,"input_tokens":216,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":null,"nll_hard":7.55303571918553,"brier_hard":1.5287758642273024,"score_mae":1.0915453221870868,"within_one":false,"case_id":"typed:89/r0","base_id":"typed:89","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":195.9469}
{"id":"action","type":"choice","probabilities":{"continue":0.00013185793588635516,"human_review":0.0003783776788633086,"observe":0.9993402554091063,"stop":0.00014950897614514068},"predicted":"observe","confidence":0.9993402554091063,"entropy_confidence":0.9955738931365413,"candidate_mass":0.9999791013121626,"accepted":true,"accepted_predicted":"observe","input_tokens":212,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":8.933785458228769,"brier_hard":1.9984174131184822,"soft_accuracy":0.14003911161444887,"brier_soft":1.3624256035393572,"tvd":0.8594897643852509,"kl":6.950225468074547,"case_id":"typed:90/r0","base_id":"typed:90","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.5113}
{"id":"needs_review","type":"binary","probabilities":{"false":3.3990393452197634e-06,"true":0.9999966009606531},"predicted":"true","confidence":0.9999966009606531,"entropy_confidence":0.9999333477816776,"candidate_mass":0.9999980021788976,"accepted":true,"accepted_predicted":"true","input_tokens":174,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.592017711789264,"brier_hard":1.9999864038657227,"soft_accuracy":0.10666967390868537,"brier_soft":1.5960755519050425,"tvd":0.893329600960654,"kl":7.989175977738187,"case_id":"typed:90/r0","base_id":"typed:90","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.5113}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0014428308128524227,"harmful":0.00019664956950417518,"partial":0.14605167032564975,"success":0.8523088492919941},"predicted":"success","confidence":0.8523088492919941,"entropy_confidence":0.6910513435646266,"candidate_mass":0.9999678279307757,"accepted":true,"accepted_predicted":"success","input_tokens":217,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.15980631867076406,"brier_hard":0.04314588683417489,"soft_accuracy":0.6732227291869368,"brier_soft":0.013740724822289576,"tvd":0.09583849886149808,"kl":0.1414289218951016,"case_id":"typed:90/r0","base_id":"typed:90","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.5113}
{"id":"risk","type":"ordinal","probabilities":{"0":0.003711194587649889,"1":0.13493470717629133,"2":0.8563082071031575,"3":0.005045891132902318},"predicted":"2","confidence":0.8563082071031575,"entropy_confidence":0.6749881469741146,"candidate_mass":0.9999710150773405,"accepted":true,"accepted_predicted":"2","input_tokens":225,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":5.596401462910567,"brier_hard":1.7440879655602683,"score_mae":1.6393557947813133,"within_one":false,"case_id":"typed:90/r0","base_id":"typed:90","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.5113}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009465179037220208,"1":0.0018840458724939178,"2":0.985969912568646,"3":0.011199523655137393},"predicted":"2","confidence":0.985969912568646,"entropy_confidence":0.9403808899521006,"candidate_mass":0.9999866226361496,"accepted":true,"accepted_predicted":"2","input_tokens":213,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.962720671781149,"brier_hard":1.9703735075382731,"score_mae":1.8074224419751979,"within_one":false,"case_id":"typed:90/r0","base_id":"typed:90","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.5113}
{"id":"action","type":"choice","probabilities":{"continue":0.0011136336243349845,"human_review":0.00277698680044015,"observe":0.9955106269180773,"stop":0.000598752657147589},"predicted":"observe","confidence":0.9955106269180773,"entropy_confidence":0.9763097240003159,"candidate_mass":0.9999549131972059,"accepted":true,"accepted_predicted":"observe","input_tokens":206,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.886388823841205,"brier_hard":1.9854967450462266,"soft_accuracy":0.08507739392478482,"brier_soft":1.3753406863535373,"tvd":0.9121776269180774,"kl":4.867661578290822,"case_id":"typed:91/r0","base_id":"typed:91","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.97719999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":1.8599642594898658e-08,"true":0.9999999814003575},"predicted":"true","confidence":0.9999999814003575,"entropy_confidence":0.9999994955247791,"candidate_mass":0.9999998855296605,"accepted":true,"accepted_predicted":"true","input_tokens":168,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.859964266581809e-08,"brier_hard":6.91893405519604e-16,"soft_accuracy":0.9266669841282927,"brier_soft":0.01075545232213034,"tvd":0.07333298140035743,"kl":0.6048460211853214,"case_id":"typed:91/r0","base_id":"typed:91","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.97719999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":3.5158451371873415e-05,"harmful":0.9996884177176298,"partial":0.00025845580755845857,"success":1.7968023439713242e-05},"predicted":"harmful","confidence":0.9996884177176298,"entropy_confidence":0.9978334402801033,"candidate_mass":0.9999962814528802,"accepted":true,"accepted_predicted":"harmful","input_tokens":211,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.00031163083421502824,"brier_hard":1.654418897168914e-07,"soft_accuracy":0.7498045892945955,"brier_soft":0.09032330006595854,"tvd":0.2496884177176299,"kl":1.5264389884444813,"case_id":"typed:91/r0","base_id":"typed:91","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.97719999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":4.7448169070066444e-06,"1":4.604065072162973e-06,"2":0.015107837172529745,"3":0.9848828139454896},"predicted":"3","confidence":0.9848828139454896,"entropy_confidence":0.9434050072260752,"candidate_mass":0.9999979217233661,"accepted":true,"accepted_predicted":"3","input_tokens":219,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.01523261550132913,"brier_hard":0.00045677610194905797,"score_mae":0.3415357202466005,"within_one":true,"case_id":"typed:91/r0","base_id":"typed:91","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.97719999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0013308941822209408,"1":0.003465852304550675,"2":0.9778660238787066,"3":0.017337229634521157},"predicted":"2","confidence":0.9778660238787066,"entropy_confidence":0.9129807386777148,"candidate_mass":0.999989258605144,"accepted":true,"accepted_predicted":"2","input_tokens":207,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":4.0548990877225535,"brier_hard":1.9218618643302303,"score_mae":0.2487904110344723,"within_one":true,"case_id":"typed:91/r0","base_id":"typed:91","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.97719999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.0019446629726765992,"human_review":0.015584917006858079,"observe":0.9813427307756631,"stop":0.0011276892448014812},"predicted":"observe","confidence":0.9813427307756631,"entropy_confidence":0.9256058592816572,"candidate_mass":0.9999467139491347,"accepted":true,"accepted_predicted":"observe","input_tokens":212,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.24266659575575,"brier_hard":1.9593921723361034,"soft_accuracy":0.4111160150392431,"brier_soft":0.5710719659809702,"tvd":0.5646757307756635,"kl":2.5522982344296476,"case_id":"typed:92/r0","base_id":"typed:92","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.3997}
{"id":"needs_review","type":"binary","probabilities":{"false":6.058750211401353e-08,"true":0.9999999394124989},"predicted":"true","confidence":0.9999999394124989,"entropy_confidence":0.999998459920255,"candidate_mass":0.9999997582505465,"accepted":true,"accepted_predicted":"true","input_tokens":174,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.61917720088301,"brier_hard":1.9999997576500008,"soft_accuracy":0.14000004362300167,"brier_soft":1.4791997915790018,"tvd":0.8599999394124984,"kl":7.645636928489572,"case_id":"typed:92/r0","base_id":"typed:92","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.3997}
{"id":"outcome","type":"choice","probabilities":{"failure":0.0008997264177031451,"harmful":0.0005709385284694614,"partial":0.017528454099874885,"success":0.9810008809539539},"predicted":"success","confidence":0.9810008809539539,"entropy_confidence":0.9276666047754023,"candidate_mass":0.9999631275802724,"accepted":true,"accepted_predicted":"success","input_tokens":217,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.01918192140091431,"brier_hard":0.0006693487060872569,"soft_accuracy":0.46603362947124,"brier_soft":0.4687369620718561,"tvd":0.5143343476204865,"kl":1.4401853346715723,"case_id":"typed:92/r0","base_id":"typed:92","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.3997}
{"id":"risk","type":"ordinal","probabilities":{"0":9.41072466096379e-06,"1":1.3798902049624441e-05,"2":0.9999268226147717,"3":4.996775851850665e-05},"predicted":"2","confidence":0.9999268226147717,"entropy_confidence":0.999400270967072,"candidate_mass":0.9999982156428903,"accepted":true,"accepted_predicted":"2","input_tokens":225,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":7.318006282375506e-05,"brier_hard":8.130678036630544e-09,"score_mae":0.7500173474071485,"within_one":true,"case_id":"typed:92/r0","base_id":"typed:92","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.3997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001163177288164737,"1":0.0028402165326166626,"2":0.9613911313985586,"3":0.034605474780660915},"predicted":"2","confidence":0.9613911313985586,"entropy_confidence":0.871043622018731,"candidate_mass":0.9999909031874821,"accepted":true,"accepted_predicted":"2","input_tokens":213,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.756599976681059,"brier_hard":1.9231535116516218,"score_mae":1.2961059036717164,"within_one":false,"case_id":"typed:92/r0","base_id":"typed:92","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.3997}
{"id":"action","type":"choice","probabilities":{"continue":0.004146971144457259,"human_review":0.00429158913761217,"observe":0.9910975156089193,"stop":0.0004639241090127737},"predicted":"observe","confidence":0.9910975156089193,"entropy_confidence":0.9577540681314055,"candidate_mass":0.9999664079455975,"accepted":true,"accepted_predicted":"observe","input_tokens":229,"reused_prefix_tokens":0,"expected":"observe","correct":true,"accepted_correct":true,"nll_hard":0.008942348273333095,"brier_hard":0.00011508456091138994,"soft_accuracy":0.5633886646845901,"brier_soft":0.2638658364100614,"tvd":0.4244310822753519,"kl":1.3677915479028682,"case_id":"typed:93/r0","base_id":"typed:93","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.58870000000002}
{"id":"needs_review","type":"binary","probabilities":{"false":2.644491968196063e-07,"true":0.9999997355508045},"predicted":"true","confidence":0.9999997355508045,"entropy_confidence":0.999993840131727,"candidate_mass":0.9999995041565896,"accepted":true,"accepted_predicted":"true","input_tokens":191,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.145616676612011,"brier_hard":1.999998942203355,"soft_accuracy":0.3166670969645296,"brier_soft":0.9338872549506896,"tvd":0.6833327355508039,"kl":5.9295927832099675,"case_id":"typed:93/r0","base_id":"typed:93","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.58870000000002}
{"id":"outcome","type":"choice","probabilities":{"failure":4.7219695262646436e-05,"harmful":1.7774724263356414e-05,"partial":0.0024394724072417454,"success":0.9974955331732321},"predicted":"success","confidence":0.9974955331732321,"entropy_confidence":0.9871297999594242,"candidate_mass":0.9999898284321912,"accepted":true,"accepted_predicted":"success","input_tokens":234,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0025076082499673152,"brier_hard":1.2225925352517843e-05,"soft_accuracy":0.5495208540932008,"brier_soft":0.33931738186341526,"tvd":0.4474955331732322,"kl":2.12279243556265,"case_id":"typed:93/r0","base_id":"typed:93","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.58870000000002}
{"id":"risk","type":"ordinal","probabilities":{"0":0.993254475568227,"1":0.00598810573330931,"2":0.00043188721678263844,"3":0.0003255314816822504},"predicted":"0","confidence":0.993254475568227,"entropy_confidence":0.9687441975988094,"candidate_mass":0.9999925561473633,"accepted":true,"accepted_predicted":"0","input_tokens":242,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.117980155047363,"brier_hard":1.9746143916772818,"score_mae":0.6755045253880786,"within_one":true,"case_id":"typed:93/r0","base_id":"typed:93","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.58870000000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006922956654557954,"1":0.9989305548272005,"2":0.000345501581043125,"3":3.164792630038449e-05},"predicted":"1","confidence":0.9989305548272005,"entropy_confidence":0.9933726972414475,"candidate_mass":0.9999975365328918,"accepted":true,"accepted_predicted":"1","input_tokens":230,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0010700174373283855,"brier_hard":1.7433591997754468e-06,"score_mae":0.24971650176818794,"within_one":true,"case_id":"typed:93/r0","base_id":"typed:93","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":199.58870000000002}
{"id":"action","type":"choice","probabilities":{"continue":0.13264072726067486,"human_review":0.1924100277488025,"observe":0.6296704025854932,"stop":0.04527884240502886},"predicted":"observe","confidence":0.6296704025854932,"entropy_confidence":0.26677993074267203,"candidate_mass":0.9999117743693323,"accepted":false,"accepted_predicted":null,"input_tokens":210,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":null,"nll_hard":2.0201111046436964,"brier_hard":1.1878687162469024,"soft_accuracy":0.23387701400259597,"brier_soft":0.32817369831906035,"tvd":0.4463374025854935,"kl":0.5115903249732908,"case_id":"typed:94/r0","base_id":"typed:94","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.09890000000001}
{"id":"needs_review","type":"binary","probabilities":{"false":2.2322423293112864e-06,"true":0.9999977677576705},"predicted":"true","confidence":0.9999977677576705,"entropy_confidence":0.9999548735061754,"candidate_mass":0.9999983069332409,"accepted":true,"accepted_predicted":"true","input_tokens":172,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.01250394885859,"brier_hard":1.999991071040648,"soft_accuracy":0.3833335208580316,"brier_soft":0.7605508715872435,"tvd":0.6166647677576707,"kl":5.312154508729837,"case_id":"typed:94/r0","base_id":"typed:94","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.09890000000001}
{"id":"outcome","type":"choice","probabilities":{"failure":2.927362221177455e-05,"harmful":1.5903608676403225e-05,"partial":0.0006136079099331401,"success":0.9993412148591784},"predicted":"success","confidence":0.9993412148591784,"entropy_confidence":0.9959040390724829,"candidate_mass":0.9999922733439286,"accepted":true,"accepted_predicted":"success","input_tokens":215,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":0.0006590022351034051,"brier_hard":8.116223986262175e-07,"soft_accuracy":0.39988065607411616,"brier_soft":0.552810706970523,"tvd":0.5993571184678549,"kl":4.435508165520279,"case_id":"typed:94/r0","base_id":"typed:94","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.09890000000001}
{"id":"risk","type":"ordinal","probabilities":{"0":0.14038729987735613,"1":0.8305944876649547,"2":0.01863027508300326,"3":0.010387937374684407},"predicted":"1","confidence":0.8305944876649547,"entropy_confidence":0.6022160189538897,"candidate_mass":0.9999696450903813,"accepted":true,"accepted_predicted":"1","input_tokens":223,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":1.9633502481958693,"brier_hard":1.4292761935441198,"score_mae":0.5990188499550144,"within_one":true,"case_id":"typed:94/r0","base_id":"typed:94","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.09890000000001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005887109721990515,"1":0.9965376305655476,"2":0.0026695051091308987,"3":0.0002041533531241855},"predicted":"1","confidence":0.9965376305655476,"entropy_confidence":0.9816859333300622,"candidate_mass":0.999993582990695,"accepted":true,"accepted_predicted":"1","input_tokens":211,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.437575204089388,"brier_hard":1.9919173417055258,"score_mae":0.502489100843182,"within_one":true,"case_id":"typed:94/r0","base_id":"typed:94","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":187.09890000000001}
{"id":"action","type":"choice","probabilities":{"continue":0.0022868728097684096,"human_review":0.11426456163732257,"observe":0.8831824557320651,"stop":0.00026610982084354764},"predicted":"observe","confidence":0.8831824557320651,"entropy_confidence":0.7304503868964913,"candidate_mass":0.9999477791296457,"accepted":true,"accepted_predicted":"observe","input_tokens":226,"reused_prefix_tokens":0,"expected":"continue","correct":false,"accepted_correct":false,"nll_hard":6.08056997993123,"brier_hard":1.7884991951412388,"soft_accuracy":0.3764121404227264,"brier_soft":0.4848270154713226,"tvd":0.5174470173693879,"kl":2.458478002527406,"case_id":"typed:95/r0","base_id":"typed:95","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.7236}
{"id":"needs_review","type":"binary","probabilities":{"false":2.2877762879542338e-08,"true":0.9999999771222365},"predicted":"true","confidence":0.9999999771222365,"entropy_confidence":0.9999993863229115,"candidate_mass":0.9999998506957124,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.593100451459,"brier_hard":1.9999999084889484,"soft_accuracy":0.09000001875976552,"brier_soft":1.656199916724943,"tvd":0.9099999771222369,"kl":8.164694635778657,"case_id":"typed:95/r0","base_id":"typed:95","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.7236}
{"id":"outcome","type":"choice","probabilities":{"failure":1.3100815935714715e-06,"harmful":1.0604336670409565e-06,"partial":2.4822850165387583e-05,"success":0.999972806634575},"predicted":"success","confidence":0.999972806634575,"entropy_confidence":0.9997671908527935,"candidate_mass":0.9999990132062458,"accepted":true,"accepted_predicted":"success","input_tokens":231,"reused_prefix_tokens":0,"expected":"success","correct":true,"accepted_correct":true,"nll_hard":2.719373517126918e-05,"brier_hard":1.3584938468151213e-09,"soft_accuracy":0.8499796677240942,"brier_soft":0.03594185695745556,"tvd":0.14997280663457452,"kl":1.1500498720168941,"case_id":"typed:95/r0","base_id":"typed:95","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.7236}
{"id":"risk","type":"ordinal","probabilities":{"0":0.5676113191281927,"1":0.426209169653899,"2":0.004999544912718557,"3":0.0011799663051886987},"predicted":"0","confidence":0.5676113191281927,"entropy_confidence":0.48108000353801794,"candidate_mass":0.9999797080081853,"accepted":false,"accepted_predicted":null,"input_tokens":239,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.852825044640806,"brier_hard":0.6514449143615307,"score_mae":0.3469188416050978,"within_one":true,"case_id":"typed:95/r0","base_id":"typed:95","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.7236}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00026788839589829264,"1":0.8555731170381103,"2":0.1439017966971768,"3":0.000257197868815211},"predicted":"1","confidence":0.8555731170381103,"entropy_confidence":0.699373905643671,"candidate_mass":0.9999878752441551,"accepted":true,"accepted_predicted":"1","input_tokens":227,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.15598372227906412,"brier_hard":0.04156698952969935,"score_mae":0.12748130403890956,"within_one":true,"case_id":"typed:95/r0","base_id":"typed:95","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":192.7236}
{"id":"action","type":"choice","probabilities":{"continue":0.0008625230131897491,"human_review":0.011182510181501884,"observe":0.98705360537099,"stop":0.0009013614343187989},"predicted":"observe","confidence":0.98705360537099,"entropy_confidence":0.9455271429474428,"candidate_mass":0.9999593052628529,"accepted":true,"accepted_predicted":"observe","input_tokens":214,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.493404312172384,"brier_hard":1.9520364044452092,"soft_accuracy":0.1209213564251357,"brier_soft":1.1270026675139415,"tvd":0.8703866053709898,"kl":3.940849185413406,"case_id":"typed:96/r0","base_id":"typed:96","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.8416}
{"id":"needs_review","type":"binary","probabilities":{"false":1.7851927438875034e-08,"true":0.9999999821480722},"predicted":"true","confidence":0.9999999821480722,"entropy_confidence":0.9999995147481948,"candidate_mass":0.9999999193523156,"accepted":true,"accepted_predicted":"true","input_tokens":176,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7851927984231314e-08,"brier_hard":6.373826403477857e-16,"soft_accuracy":0.8933329859564954,"brier_soft":0.02275569016115436,"tvd":0.10666698214807233,"kl":0.8816751219006854,"case_id":"typed:96/r0","base_id":"typed:96","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.8416}
{"id":"outcome","type":"choice","probabilities":{"failure":3.79886304420573e-05,"harmful":0.9997110464615264,"partial":0.00023358461605835,"success":1.7380291973917137e-05},"predicted":"harmful","confidence":0.9997110464615264,"entropy_confidence":0.9979663148628234,"candidate_mass":0.9999966507227844,"accepted":true,"accepted_predicted":"harmful","input_tokens":219,"reused_prefix_tokens":0,"expected":"harmful","correct":true,"accepted_correct":true,"nll_hard":0.0002889952935909976,"brier_hard":1.3980113084749113e-07,"soft_accuracy":0.5832223081904603,"brier_soft":0.25408837189926314,"tvd":0.41637804646152604,"kl":2.8355686975787844,"case_id":"typed:96/r0","base_id":"typed:96","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.8416}
{"id":"risk","type":"ordinal","probabilities":{"0":4.782021035968664e-06,"1":4.075249961898654e-06,"2":0.0024692727411204504,"3":0.9975218699878807},"predicted":"3","confidence":0.9975218699878807,"entropy_confidence":0.9874418347269862,"candidate_mass":0.9999987618873964,"accepted":true,"accepted_predicted":"3","input_tokens":227,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.002481205658583014,"brier_hard":1.223847570239421e-05,"score_mae":0.48084123069584495,"within_one":true,"case_id":"typed:96/r0","base_id":"typed:96","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.8416}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006696361197210843,"1":0.004545017576844653,"2":0.9920010753932526,"3":0.0027842709101805045},"predicted":"2","confidence":0.9920010753932526,"entropy_confidence":0.9612220564377786,"candidate_mass":0.9999891615541074,"accepted":true,"accepted_predicted":"2","input_tokens":215,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.8837692315529715,"brier_hard":1.9785264495228165,"score_mae":0.20310001890610896,"within_one":true,"case_id":"typed:96/r0","base_id":"typed:96","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":189.8416}
{"id":"action","type":"choice","probabilities":{"continue":0.009368243641898743,"human_review":0.011507366638261152,"observe":0.9780538888474556,"stop":0.0010705008723827867},"predicted":"observe","confidence":0.9780538888474556,"entropy_confidence":0.9104398101941313,"candidate_mass":0.999956130558037,"accepted":true,"accepted_predicted":"observe","input_tokens":210,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":4.464767871454857,"brier_hard":1.9337960056611079,"soft_accuracy":0.2836183751248297,"brier_soft":0.7329074420215341,"tvd":0.6947206055141731,"kl":2.379035081202624,"case_id":"typed:97/r0","base_id":"typed:97","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.06029999999998}
{"id":"needs_review","type":"binary","probabilities":{"false":4.527190713093815e-07,"true":0.9999995472809283},"predicted":"true","confidence":0.9999995472809283,"entropy_confidence":0.9999898058642236,"candidate_mass":0.999999343987904,"accepted":true,"accepted_predicted":"true","input_tokens":172,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.527191741525215e-07,"brier_hard":4.0990911538569425e-13,"soft_accuracy":0.6666668490933408,"brier_soft":0.2222211741535849,"tvd":0.33333254728092854,"kl":2.799800815131142,"case_id":"typed:97/r0","base_id":"typed:97","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.06029999999998}
{"id":"outcome","type":"choice","probabilities":{"failure":8.4726511326933e-05,"harmful":3.3104786658779655e-05,"partial":0.0019652640241772878,"success":0.9979169046778368},"predicted":"success","confidence":0.9979169046778368,"entropy_confidence":0.9888446429200363,"candidate_mass":0.9999890945468155,"accepted":true,"accepted_predicted":"success","input_tokens":215,"reused_prefix_tokens":0,"expected":"failure","correct":false,"accepted_correct":false,"nll_hard":9.376082002602129,"brier_hard":1.9956725661563341,"soft_accuracy":0.09040947073159801,"brier_soft":1.292422729583011,"tvd":0.9079169946777469,"kl":6.7703731875267845,"case_id":"typed:97/r0","base_id":"typed:97","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.06029999999998}
{"id":"risk","type":"ordinal","probabilities":{"0":0.7888222875449719,"1":0.20520239651335923,"2":0.002962066609241605,"3":0.0030132493324265337},"predicted":"0","confidence":0.7888222875449719,"entropy_confidence":0.6055334310124607,"candidate_mass":0.999981408416922,"accepted":false,"accepted_predicted":null,"input_tokens":223,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":null,"nll_hard":0.23721422109164436,"brier_hard":0.0867219032827014,"score_mae":0.47983372227087795,"within_one":true,"case_id":"typed:97/r0","base_id":"typed:97","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.06029999999998}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004426348056422758,"1":0.9993159240818531,"2":0.0002152057754063969,"3":2.6235337097011853e-05},"predicted":"1","confidence":0.9993159240818531,"entropy_confidence":0.9955304391000097,"candidate_mass":0.9999975440589348,"accepted":true,"accepted_predicted":"1","input_tokens":211,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0006843100048392596,"brier_hard":7.108872516353592e-07,"score_mae":0.09982504164395689,"within_one":true,"case_id":"typed:97/r0","base_id":"typed:97","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.06029999999998}
{"id":"action","type":"choice","probabilities":{"continue":0.00039580644859742295,"human_review":0.9931532343214731,"observe":0.006178348373598769,"stop":0.0002726108563323943},"predicted":"human_review","confidence":0.9931532343214731,"entropy_confidence":0.9685570908899834,"candidate_mass":0.9999585684167488,"accepted":true,"accepted_predicted":"human_review","input_tokens":219,"reused_prefix_tokens":0,"expected":"observe","correct":false,"accepted_correct":false,"nll_hard":5.086704296694091,"brier_hard":1.9740350530640545,"soft_accuracy":0.39980358128719873,"brier_soft":0.5290067650148547,"tvd":0.5931532343214723,"kl":2.446982447144599,"case_id":"typed:98/r0","base_id":"typed:98","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7545}
{"id":"needs_review","type":"binary","probabilities":{"false":6.759526615879882e-07,"true":0.9999993240473384},"predicted":"true","confidence":0.9999993240473384,"entropy_confidence":0.9999851700927245,"candidate_mass":0.9999992344106179,"accepted":true,"accepted_predicted":"true","input_tokens":181,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.759528900778959e-07,"brier_hard":9.13824001461471e-13,"soft_accuracy":0.5899998783285209,"brier_soft":0.33619889143854886,"tvd":0.4099993240473384,"kl":3.464936633493841,"case_id":"typed:98/r0","base_id":"typed:98","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7545}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00011258328847179177,"harmful":4.395156034027317e-05,"partial":0.002215662187618468,"success":0.9976278029635702},"predicted":"success","confidence":0.9976278029635702,"entropy_confidence":0.9874655137109863,"candidate_mass":0.9999881098442104,"accepted":true,"accepted_predicted":"success","input_tokens":224,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":6.112203964128064,"brier_hard":1.9908348326363492,"soft_accuracy":0.08435864333012293,"brier_soft":1.2480826200854183,"tvd":0.9142947196304865,"kl":5.7693751404066145,"case_id":"typed:98/r0","base_id":"typed:98","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7545}
{"id":"risk","type":"ordinal","probabilities":{"0":0.03120256604864909,"1":0.9390598616231749,"2":0.013794415558809572,"3":0.015943156769364752},"predicted":"1","confidence":0.9390598616231749,"entropy_confidence":0.7891470618051133,"candidate_mass":0.999962223245643,"accepted":true,"accepted_predicted":"1","input_tokens":232,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.06287605140842702,"brier_hard":0.005131770741788569,"score_mae":0.23114516304888844,"within_one":true,"case_id":"typed:98/r0","base_id":"typed:98","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7545}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0018339938281841713,"1":0.9522310597374007,"2":0.04472277596659634,"3":0.0012121704678188254},"predicted":"1","confidence":0.9522310597374007,"entropy_confidence":0.8519276615947433,"candidate_mass":0.9999885752927424,"accepted":true,"accepted_predicted":"1","input_tokens":220,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.048947563831838195,"brier_hard":0.004286831234575013,"score_mae":0.15468687692594996,"within_one":true,"case_id":"typed:98/r0","base_id":"typed:98","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":200.7545}
{"id":"action","type":"choice","probabilities":{"continue":0.0007973762042048127,"human_review":0.0025161495875285337,"observe":0.996211489279457,"stop":0.0004749849288079817},"predicted":"observe","confidence":0.996211489279457,"entropy_confidence":0.9796840651654048,"candidate_mass":0.9999500503651111,"accepted":true,"accepted_predicted":"observe","input_tokens":214,"reused_prefix_tokens":0,"expected":"human_review","correct":false,"accepted_correct":false,"nll_hard":5.985025487408695,"brier_hard":1.9874122246255772,"soft_accuracy":0.3001341722257566,"brier_soft":0.707731557127121,"tvd":0.6962114892794579,"kl":3.296452157253519,"case_id":"typed:99/r0","base_id":"typed:99","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9796}
{"id":"needs_review","type":"binary","probabilities":{"false":1.786228568211482e-07,"true":0.999999821377143},"predicted":"true","confidence":0.999999821377143,"entropy_confidence":0.99999573818816,"candidate_mass":0.9999993404305102,"accepted":true,"accepted_predicted":"true","input_tokens":176,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.786228729030925e-07,"brier_hard":6.381225000393862e-14,"soft_accuracy":0.6666669404589286,"brier_soft":0.22222153961449276,"tvd":0.3333328213771431,"kl":2.799800632400183,"case_id":"typed:99/r0","base_id":"typed:99","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9796}
{"id":"outcome","type":"choice","probabilities":{"failure":0.00012143083630147132,"harmful":7.153011494370195e-05,"partial":0.0037027922902015792,"success":0.9961042467585524},"predicted":"success","confidence":0.9961042467585524,"entropy_confidence":0.9809589621371015,"candidate_mass":0.9999796091274814,"accepted":true,"accepted_predicted":"success","input_tokens":219,"reused_prefix_tokens":0,"expected":"partial","correct":false,"accepted_correct":false,"nll_hard":5.598668070927005,"brier_hard":1.9848318163627696,"soft_accuracy":0.2671482279697597,"brier_soft":0.7819631427816535,"tvd":0.7294372467585528,"kl":4.075735881425049,"case_id":"typed:99/r0","base_id":"typed:99","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9796}
{"id":"risk","type":"ordinal","probabilities":{"0":1.0071628157808581e-05,"1":1.4147871946494786e-05,"2":0.9999262389140455,"3":4.9541585848626087e-05},"predicted":"2","confidence":0.9999262389140455,"entropy_confidence":0.9993950018313056,"candidate_mass":0.999998125491521,"accepted":true,"accepted_predicted":"2","input_tokens":227,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":7.376380643714987e-05,"brier_hard":8.196666503943266e-09,"score_mae":0.31668225045758347,"within_one":true,"case_id":"typed:99/r0","base_id":"typed:99","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9796}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0017762273339769728,"1":0.015092749513564964,"2":0.9706945462948154,"3":0.012436476857643666},"predicted":"2","confidence":0.9706945462948154,"entropy_confidence":0.8860462052595528,"candidate_mass":0.9999887887233883,"accepted":true,"accepted_predicted":"2","input_tokens":215,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.193540815141798,"brier_hard":1.9124480152074195,"score_mae":0.4604582726761268,"within_one":true,"case_id":"typed:99/r0","base_id":"typed:99","repeat":0,"workflow":"agent_trace_observability","case_elapsed_ms":196.9796}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.6872096922005055e-05,"close_no_action":4.3517626272012294e-05,"escalate_to_human":0.00014065611737451561,"execute_refund":2.0123583680461972e-05,"request_information":0.9997688305757517},"predicted":"request_information","confidence":0.9997688305757517,"entropy_confidence":0.9984987989260231,"candidate_mass":0.9999922326868461,"accepted":true,"accepted_predicted":"request_information","input_tokens":320,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":10.524422098857979,"brier_hard":1.9994839932019577,"soft_accuracy":0.10003949371922534,"brier_soft":1.174947847735351,"tvd":0.8997688305757514,"kl":7.4448079633663236,"case_id":"typed:100/r0","base_id":"typed:100","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.585101}
{"id":"category","type":"choice","probabilities":{"account":2.2510951783619792e-07,"billing":2.3814466184441687e-07,"delivery":0.9999984150274364,"refund":7.054130418225788e-07,"technical":4.1630534261444516e-07},"predicted":"delivery","confidence":0.9999984150274364,"entropy_confidence":0.9999846091896523,"candidate_mass":0.9999997301974153,"accepted":true,"accepted_predicted":"delivery","input_tokens":310,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":1.5849738197099937e-06,"brier_hard":3.2904429003386044e-12,"soft_accuracy":0.9233315788948891,"brier_soft":0.008377527824384716,"tvd":0.07666541502743614,"kl":0.6324806677923965,"case_id":"typed:100/r0","base_id":"typed:100","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.585101}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.660356717826547e-05,"1":0.9997040180071007,"2":0.000245679709909949,"3":3.6987158093177376e-06},"predicted":"1","confidence":0.9997040180071007,"entropy_confidence":0.9979448963243903,"candidate_mass":0.999998599704451,"accepted":true,"accepted_predicted":"1","input_tokens":296,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.000296025804214444,"brier_hard":1.5014943295443358e-07,"score_mae":0.06646052642565148,"within_one":true,"case_id":"typed:100/r0","base_id":"typed:100","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.585101}
{"id":"needs_human","type":"binary","probabilities":{"false":1.27639190928933e-07,"true":0.9999998723608076},"predicted":"true","confidence":0.9999998723608076,"entropy_confidence":0.999996892736186,"candidate_mass":0.9999997369118452,"accepted":true,"accepted_predicted":"true","input_tokens":264,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.87405837425175,"brier_hard":1.999999489443266,"soft_accuracy":0.46666700850919357,"brier_soft":0.5688879054812606,"tvd":0.5333328723608084,"kl":4.5565132428370605,"case_id":"typed:100/r0","base_id":"typed:100","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.585101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00066085610486627,"1":0.2945782623533984,"2":0.6939302815126037,"3":0.010830600029131452},"predicted":"2","confidence":0.6939302815126037,"entropy_confidence":0.518544614796639,"candidate_mass":0.999987201573404,"accepted":false,"accepted_predicted":null,"input_tokens":290,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.2222105646637864,"brier_hard":0.9792768021722947,"score_mae":0.5149306254660002,"within_one":true,"case_id":"typed:100/r0","base_id":"typed:100","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.585101}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0017411347217419304,"close_no_action":0.00020200780624215865,"escalate_to_human":0.0008444409112467857,"execute_refund":0.007043994152758889,"request_information":0.9901684224080098},"predicted":"request_information","confidence":0.9901684224080098,"entropy_confidence":0.9605784950507149,"candidate_mass":0.9999661315693497,"accepted":true,"accepted_predicted":"request_information","input_tokens":422,"reused_prefix_tokens":0,"expected":"execute_refund","correct":false,"accepted_correct":false,"nll_hard":4.955579918450894,"brier_hard":1.9663989197197989,"soft_accuracy":0.10845123109112427,"brier_soft":1.021473094510513,"tvd":0.8835015290749034,"kl":4.101245535177814,"case_id":"typed:101/r0","base_id":"typed:101","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.1172}
{"id":"category","type":"choice","probabilities":{"account":2.335335564511522e-12,"billing":0.999999999870484,"delivery":8.572118251436783e-11,"refund":3.121323628154367e-11,"technical":1.0247195402171284e-11},"predicted":"billing","confidence":0.999999999870484,"entropy_confidence":0.9999999980158153,"candidate_mass":0.9999999997697557,"accepted":true,"accepted_predicted":"billing","input_tokens":412,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.2951595352589886e-10,"brier_hard":2.5207228272188428e-20,"soft_accuracy":0.8799991198903769,"brier_soft":0.022333601962747984,"tvd":0.1200008798696035,"kl":0.9927548154084141,"case_id":"typed:101/r0","base_id":"typed:101","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.1172}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0003696045853217132,"1":0.996874295884755,"2":0.0027108667492936867,"3":4.523278062992176e-05},"predicted":"1","confidence":0.996874295884755,"entropy_confidence":0.983757515707835,"candidate_mass":0.999996601319206,"accepted":true,"accepted_predicted":"1","input_tokens":398,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00313059933167997,"brier_hard":1.7257478302420265e-05,"score_mae":0.3809012722747678,"within_one":true,"case_id":"typed:101/r0","base_id":"typed:101","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.1172}
{"id":"needs_human","type":"binary","probabilities":{"false":7.883758316403304e-07,"true":0.9999992116241695},"predicted":"true","confidence":0.9999992116241695,"entropy_confidence":0.9999828785995721,"candidate_mass":0.9999993016703095,"accepted":true,"accepted_predicted":"true","input_tokens":366,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.883761412755593e-07,"brier_hard":1.2430729020424539e-12,"soft_accuracy":0.5566669106502141,"brier_soft":0.39308689972715377,"tvd":0.4433322116241689,"kl":3.757159383082886,"case_id":"typed:101/r0","base_id":"typed:101","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.1172}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0012311256168094085,"1":0.1532635641134642,"2":0.8060329283859142,"3":0.03947238188381158},"predicted":"2","confidence":0.8060329283859142,"entropy_confidence":0.5692869042451107,"candidate_mass":0.9999906239018318,"accepted":true,"accepted_predicted":"2","input_tokens":392,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.21563068323319717,"brier_hard":0.06267252955717165,"score_mae":0.21625343346327264,"within_one":true,"case_id":"typed:101/r0","base_id":"typed:101","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.1172}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0001280763541089847,"close_no_action":9.349129232740537e-05,"escalate_to_human":0.000162698022380118,"execute_refund":3.050275226678576e-05,"request_information":0.9995852315789161},"predicted":"request_information","confidence":0.9995852315789161,"entropy_confidence":0.9974112280089035,"candidate_mass":0.9999804526133699,"accepted":true,"accepted_predicted":"request_information","input_tokens":304,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0004148544612974622,"brier_hard":2.2457808173357986e-07,"soft_accuracy":0.41657638198477753,"brier_soft":0.49499561932235886,"tvd":0.5829182315789164,"kl":3.999258865207683,"case_id":"typed:102/r0","base_id":"typed:102","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.16760100000005}
{"id":"category","type":"choice","probabilities":{"account":3.433770355670459e-09,"billing":0.9999999068797977,"delivery":7.577602968761864e-08,"refund":1.1285086069800677e-08,"technical":2.625316535711302e-09},"predicted":"billing","confidence":0.9999999068797977,"entropy_confidence":0.9999989680785465,"candidate_mass":0.9999999650028577,"accepted":true,"accepted_predicted":"billing","input_tokens":294,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":9.312020666256587e-08,"brier_hard":1.4559414989988817e-14,"soft_accuracy":0.9666659435632792,"brier_soft":0.001838940735088078,"tvd":0.033333873545830874,"kl":0.27424515714372594,"case_id":"typed:102/r0","base_id":"typed:102","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.16760100000005}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.631611705201948e-05,"1":0.9280621778019676,"2":0.07171067318982545,"3":0.0001608328911536585},"predicted":"1","confidence":0.9280621778019676,"entropy_confidence":0.8122370146586707,"candidate_mass":0.9999972256739094,"accepted":true,"accepted_predicted":"1","input_tokens":280,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.07465654649869084,"brier_hard":0.010317501176979925,"score_mae":0.055299022855079594,"within_one":true,"case_id":"typed:102/r0","base_id":"typed:102","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.16760100000005}
{"id":"needs_human","type":"binary","probabilities":{"false":9.545668470881621e-08,"true":0.9999999045433167},"predicted":"true","confidence":0.9999999045433167,"entropy_confidence":0.9999976361800111,"candidate_mass":0.9999997849109956,"accepted":true,"accepted_predicted":"true","input_tokens":248,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.545668789153807e-08,"brier_hard":1.8223957049021544e-14,"soft_accuracy":0.5833329840906171,"brier_soft":0.3472226186834176,"tvd":0.416666904543316,"kl":3.523230168358098,"case_id":"typed:102/r0","base_id":"typed:102","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.16760100000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003992071913637237,"1":0.04854666999591487,"2":0.9482364837528515,"3":0.0028176390598687547},"predicted":"2","confidence":0.9482364837528515,"entropy_confidence":0.8435153005901447,"candidate_mass":0.9999971672130541,"accepted":true,"accepted_predicted":"2","input_tokens":274,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":3.0252296757647095,"brier_hard":1.8044239667520878,"score_mae":0.7034725546812242,"within_one":true,"case_id":"typed:102/r0","base_id":"typed:102","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.16760100000005}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0002773349271303273,"close_no_action":0.00030631851694982646,"escalate_to_human":0.0008426375035434643,"execute_refund":0.00026593117273877165,"request_information":0.9983077778796378},"predicted":"request_information","confidence":0.9983077778796378,"entropy_confidence":0.9909317059945403,"candidate_mass":0.9999655961806756,"accepted":true,"accepted_predicted":"request_information","input_tokens":456,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0016936555455586957,"brier_hard":3.815118751288184e-06,"soft_accuracy":0.4661438623636285,"brier_soft":0.4165312684181733,"tvd":0.531641244546171,"kl":2.9879527796314433,"case_id":"typed:103/r0","base_id":"typed:103","repeat":0,"workflow":"customer_service","case_elapsed_ms":348.08529999999996}
{"id":"category","type":"choice","probabilities":{"account":8.911875562620944e-10,"billing":0.9999999576953416,"delivery":2.677932111548589e-08,"refund":1.1017897772456615e-08,"technical":3.616252278982654e-09},"predicted":"billing","confidence":0.9999999576953416,"entropy_confidence":0.9999995029498435,"candidate_mass":0.9999999864713481,"accepted":true,"accepted_predicted":"billing","input_tokens":446,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":4.230465933785743e-08,"brier_hard":2.642081732516486e-15,"soft_accuracy":0.9666669591987728,"brier_soft":0.0017555147711404986,"tvd":0.033332995492560274,"kl":0.2742371385933152,"case_id":"typed:103/r0","base_id":"typed:103","repeat":0,"workflow":"customer_service","case_elapsed_ms":348.08529999999996}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0005056953457926382,"1":0.7969263663215271,"2":0.04866727824378666,"3":0.15390066008889391},"predicted":"1","confidence":0.7969263663215271,"entropy_confidence":0.5528650025928709,"candidate_mass":0.9999973643294512,"accepted":false,"accepted_predicted":null,"input_tokens":432,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":3.0227483793677195,"brier_hard":1.5638112497260976,"score_mae":0.4940370969242178,"within_one":true,"case_id":"typed:103/r0","base_id":"typed:103","repeat":0,"workflow":"customer_service","case_elapsed_ms":348.08529999999996}
{"id":"needs_human","type":"binary","probabilities":{"false":4.706358084474389e-09,"true":0.999999995293642},"predicted":"true","confidence":0.999999995293642,"entropy_confidence":0.9999998630193924,"candidate_mass":0.9999999660473803,"accepted":true,"accepted_predicted":"true","input_tokens":400,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.70635798108286e-09,"brier_hard":4.4299611761154856e-17,"soft_accuracy":0.5499999995293643,"brier_soft":0.4049999915285555,"tvd":0.44999999529364193,"kl":3.8158428195621874,"case_id":"typed:103/r0","base_id":"typed:103","repeat":0,"workflow":"customer_service","case_elapsed_ms":348.08529999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.4673650132571854e-05,"1":1.7174584065439695e-05,"2":0.0029189332113798492,"3":0.9970492185544203},"predicted":"3","confidence":0.9970492185544203,"entropy_confidence":0.9853316606276028,"candidate_mass":0.9999989269945503,"accepted":true,"accepted_predicted":"3","input_tokens":426,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00295514358440845,"brier_hard":1.7227792514419724e-05,"score_mae":0.697002696670086,"within_one":true,"case_id":"typed:103/r0","base_id":"typed:103","repeat":0,"workflow":"customer_service","case_elapsed_ms":348.08529999999996}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.053816074955849e-05,"close_no_action":6.203200173219049e-05,"escalate_to_human":0.0002269570894628042,"execute_refund":3.530360272176282e-05,"request_information":0.9996551691453355},"predicted":"request_information","confidence":0.9996551691453355,"entropy_confidence":0.9978665485484609,"candidate_mass":0.9999926473440067,"accepted":true,"accepted_predicted":"request_information","input_tokens":407,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":10.793225902285386,"brier_hard":1.9992694379031402,"soft_accuracy":0.046691129377778556,"brier_soft":1.6462832732470822,"tvd":0.9530855047497886,"kl":8.325298730483267,"case_id":"typed:104/r0","base_id":"typed:104","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.270201}
{"id":"category","type":"choice","probabilities":{"account":2.2809969981444637e-05,"billing":0.9999043299676627,"delivery":5.534395424573458e-05,"refund":7.155714163578267e-06,"technical":1.0360393945887912e-05},"predicted":"billing","confidence":0.9999043299676627,"entropy_confidence":0.9993254579746014,"candidate_mass":0.9999963915702157,"accepted":true,"accepted_predicted":"billing","input_tokens":397,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":9.567460900679328e-05,"brier_hard":1.2894545097442246e-08,"soft_accuracy":0.9599081166675127,"brier_soft":0.002823930497940011,"tvd":0.03990528996670299,"kl":0.24089095722500833,"case_id":"typed:104/r0","base_id":"typed:104","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.270201}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":9.465483140967952e-06,"1":1.0481062411344654e-05,"2":0.9930926277732175,"3":0.006887425681229378},"predicted":"2","confidence":0.9930926277732175,"entropy_confidence":0.9701368361779374,"candidate_mass":0.9999981635082792,"accepted":true,"accepted_predicted":"2","input_tokens":383,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.006931338548955802,"brier_hard":9.514862304182461e-05,"score_mae":0.006858013652534467,"within_one":true,"case_id":"typed:104/r0","base_id":"typed:104","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.270201}
{"id":"needs_human","type":"binary","probabilities":{"false":1.2157916425575667e-07,"true":0.9999998784208376},"predicted":"true","confidence":0.9999998784208376,"entropy_confidence":0.9999970317303465,"candidate_mass":0.9999997689753375,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.922700228676796,"brier_hard":1.9999995136833761,"soft_accuracy":0.15000008510541524,"brier_soft":1.4449995866308742,"tvd":0.8499998784208367,"kl":7.544221336683749,"case_id":"typed:104/r0","base_id":"typed:104","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.270201}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0011957957868911433,"1":0.14268560387351403,"2":0.8507403485383707,"3":0.005378251801222797},"predicted":"2","confidence":0.8507403485383707,"entropy_confidence":0.6743148722718965,"candidate_mass":0.9999965745728557,"accepted":true,"accepted_predicted":"2","input_tokens":377,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":1.9471116434461102,"brier_hard":1.4587774699569112,"score_mae":0.27696805635392363,"within_one":true,"case_id":"typed:104/r0","base_id":"typed:104","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.270201}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00012412414669862317,"close_no_action":0.00021118578452811686,"escalate_to_human":0.0010682169768956062,"execute_refund":0.000519084391125898,"request_information":0.9980773887007512},"predicted":"request_information","confidence":0.9980773887007512,"entropy_confidence":0.9900220228775893,"candidate_mass":0.9999840108732955,"accepted":true,"accepted_predicted":"request_information","input_tokens":434,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":6.841764397191444,"brier_hard":1.9940235104242734,"soft_accuracy":0.3499584872435343,"brier_soft":0.763464847668996,"tvd":0.6488076588764055,"kl":3.7289289594286585,"case_id":"typed:105/r0","base_id":"typed:105","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.54299999999995}
{"id":"category","type":"choice","probabilities":{"account":6.020675457848561e-13,"billing":0.9999999999443911,"delivery":4.077792753138731e-11,"refund":1.2501945154164228e-11,"technical":1.7271689271759345e-12},"predicted":"billing","confidence":0.9999999999443911,"entropy_confidence":0.9999999991247142,"candidate_mass":0.9999999998342091,"accepted":true,"accepted_predicted":"billing","input_tokens":424,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":5.5608850858971016e-11,"brier_hard":4.914827897908815e-21,"soft_accuracy":0.9299999999491231,"brier_soft":0.009355597768536308,"tvd":0.069999999985771,"kl":0.5772330817536522,"case_id":"typed:105/r0","base_id":"typed:105","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.54299999999995}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.617242569011944e-06,"1":3.6862033550483727e-06,"2":0.0004538035810279781,"3":0.999537892973048},"predicted":"3","confidence":0.999537892973048,"entropy_confidence":0.9970726604935811,"candidate_mass":0.9999986418150386,"accepted":true,"accepted_predicted":"3","input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.697846094446543,"brier_hard":1.9981685982995414,"score_mae":0.9661919722845549,"within_one":true,"case_id":"typed:105/r0","base_id":"typed:105","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.54299999999995}
{"id":"needs_human","type":"binary","probabilities":{"false":2.6521720512778295e-08,"true":0.9999999734782786},"predicted":"true","confidence":0.9999999734782786,"entropy_confidence":0.999999294231942,"candidate_mass":0.9999999095200339,"accepted":true,"accepted_predicted":"true","input_tokens":378,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.652172172632913e-08,"brier_hard":1.406803363631339e-15,"soft_accuracy":0.816666983202892,"brier_soft":0.06722195832877473,"tvd":0.18333297347827904,"kl":1.5231648032092004,"case_id":"typed:105/r0","base_id":"typed:105","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.54299999999995}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.8806235798573686e-08,"1":2.827484144224943e-08,"2":2.771845921112375e-07,"3":0.9999996357343324},"predicted":"3","confidence":0.9999996357343324,"entropy_confidence":0.9999956575749737,"candidate_mass":0.9999999085369573,"accepted":true,"accepted_predicted":"3","input_tokens":404,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.642657339742364e-07,"brier_hard":2.1377841474480712e-13,"score_mae":0.34999948984702245,"within_one":true,"case_id":"typed:105/r0","base_id":"typed:105","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.54299999999995}
{"id":"action","type":"choice","probabilities":{"answer_directly":8.098271655929493e-06,"close_no_action":2.100866405551325e-05,"escalate_to_human":6.838018238296013e-05,"execute_refund":1.7055117017850042e-05,"request_information":0.9998854577648867},"predicted":"request_information","confidence":0.9998854577648867,"entropy_confidence":0.999205437866044,"candidate_mass":0.9999968233374162,"accepted":true,"accepted_predicted":"request_information","input_tokens":378,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":9.590427506559248,"brier_hard":1.9996341737586034,"soft_accuracy":0.383323540278168,"brier_soft":0.6255011291230332,"tvd":0.6165524577648872,"kl":5.075849968704562,"case_id":"typed:106/r0","base_id":"typed:106","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.577801}
{"id":"category","type":"choice","probabilities":{"account":3.653660327588068e-08,"billing":6.079283396073778e-07,"delivery":3.2380185223240135e-07,"refund":2.2980415179051407e-07,"technical":0.9999988019290527},"predicted":"technical","confidence":0.9999988019290527,"entropy_confidence":0.9999882713259763,"candidate_mass":0.9999999144647768,"accepted":true,"accepted_predicted":"technical","input_tokens":368,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":1.1980716650267178e-06,"brier_hard":1.963943372024135e-12,"soft_accuracy":0.953332825300966,"brier_soft":0.0027331187265588323,"tvd":0.04666484859509955,"kl":0.3842514732844725,"case_id":"typed:106/r0","base_id":"typed:106","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.577801}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.220007593930042e-05,"1":0.9929819659242701,"2":0.006924142796333618,"3":5.1691203456858623e-05},"predicted":"1","confidence":0.9929819659242701,"entropy_confidence":0.969443280767466,"candidate_mass":0.9999971061321723,"accepted":true,"accepted_predicted":"1","input_tokens":354,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.97274101849354,"brier_hard":1.9722128472644522,"score_mae":0.726347674872692,"within_one":true,"case_id":"typed:106/r0","base_id":"typed:106","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.577801}
{"id":"needs_human","type":"binary","probabilities":{"false":1.7827103056832128e-07,"true":0.9999998217289682},"predicted":"true","confidence":0.9999998217289682,"entropy_confidence":0.9999957460753999,"candidate_mass":0.9999997523031134,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.782710477157211e-07,"brier_hard":6.356112112799795e-14,"soft_accuracy":0.7433329132415498,"brier_soft":0.1317557147529005,"tvd":0.2566668217289688,"kl":2.143509707596652,"case_id":"typed:106/r0","base_id":"typed:106","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.577801}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00011480860674384943,"1":0.0003475185439335437,"2":0.9982587400172176,"3":0.0012789328321059473},"predicted":"2","confidence":0.9982587400172176,"entropy_confidence":0.9898512914377146,"candidate_mass":0.999997810394901,"accepted":true,"accepted_predicted":"2","input_tokens":348,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.001742777738073154,"brier_hard":4.801605671238078e-06,"score_mae":0.09929820292531355,"within_one":true,"case_id":"typed:106/r0","base_id":"typed:106","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.577801}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0005148619771927854,"close_no_action":5.222947456955811e-05,"escalate_to_human":0.0004431417534253105,"execute_refund":0.8438576255603034,"request_information":0.1551321412345075},"predicted":"execute_refund","confidence":0.8438576255603034,"entropy_confidence":0.7264986878450409,"candidate_mass":0.9999392192837282,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":324,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":7.721620853944756,"brier_hard":1.7352758541388131,"soft_accuracy":0.3033743580611497,"brier_soft":0.5425019593013644,"tvd":0.5989897667948116,"kl":3.7361289004709297,"case_id":"typed:107/r0","base_id":"typed:107","repeat":0,"workflow":"customer_service","case_elapsed_ms":314.448}
{"id":"category","type":"choice","probabilities":{"account":1.0160438289825034e-06,"billing":0.9999283314945266,"delivery":6.778866572573229e-06,"refund":6.311943533749899e-05,"technical":7.541597354028016e-07},"predicted":"billing","confidence":0.9999283314945266,"entropy_confidence":0.9995107644691761,"candidate_mass":0.9999983484732655,"accepted":true,"accepted_predicted":"billing","input_tokens":314,"reused_prefix_tokens":0,"expected":"refund","correct":false,"accepted_correct":false,"nll_hard":9.670481827323783,"brier_hard":1.9997304332863701,"soft_accuracy":0.13337814775482537,"brier_soft":1.5019897544253944,"tvd":0.866603880564663,"kl":7.7136542175078855,"case_id":"typed:107/r0","base_id":"typed:107","repeat":0,"workflow":"customer_service","case_elapsed_ms":314.448}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.2486270252080334e-07,"1":4.204169239571703e-08,"2":1.4167215172286558e-06,"3":0.9999984163740867},"predicted":"3","confidence":0.9999984163740867,"entropy_confidence":0.9999831480720477,"candidate_mass":0.9999999362987302,"accepted":true,"accepted_predicted":"3","input_tokens":300,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.5836271672297036e-06,"brier_hard":4.532329089011614e-12,"score_mae":0.13666512460698677,"within_one":true,"case_id":"typed:107/r0","base_id":"typed:107","repeat":0,"workflow":"customer_service","case_elapsed_ms":314.448}
{"id":"needs_human","type":"binary","probabilities":{"false":5.362197374258297e-08,"true":0.999999946378027},"predicted":"true","confidence":0.999999946378027,"entropy_confidence":0.9999986275297412,"candidate_mass":0.9999998280191532,"accepted":true,"accepted_predicted":"true","input_tokens":268,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.36219744106417e-08,"brier_hard":5.750632053565647e-15,"soft_accuracy":0.7499999731890137,"brier_soft":0.12499994637803238,"tvd":0.24999994637802664,"kl":2.086823578871691,"case_id":"typed:107/r0","base_id":"typed:107","repeat":0,"workflow":"customer_service","case_elapsed_ms":314.448}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.0938010389848405e-06,"1":6.355311375902083e-07,"2":1.4206518565537391e-05,"3":0.9999840641492591},"predicted":"3","confidence":0.9999840641492591,"entropy_confidence":0.9998567491826535,"candidate_mass":0.9999995619462396,"accepted":true,"accepted_predicted":"3","input_tokens":294,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.5935977717907513e-05,"brier_hard":4.573768091285853e-10,"score_mae":0.5499812410160456,"within_one":true,"case_id":"typed:107/r0","base_id":"typed:107","repeat":0,"workflow":"customer_service","case_elapsed_ms":314.448}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0007211797686345557,"close_no_action":0.001184395634770871,"escalate_to_human":0.0071542866736639225,"execute_refund":0.001357908515235921,"request_information":0.9895822294076948},"predicted":"request_information","confidence":0.9895822294076948,"entropy_confidence":0.9578306209014925,"candidate_mass":0.9999231606905071,"accepted":true,"accepted_predicted":"request_information","input_tokens":394,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.010472415413826036,"brier_hard":0.0001634805707369665,"soft_accuracy":0.4319829076315302,"brier_soft":0.4993399596790663,"tvd":0.5576071379229307,"kl":2.0494409613660096,"case_id":"typed:108/r0","base_id":"typed:108","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.619601}
{"id":"category","type":"choice","probabilities":{"account":0.9999990244811234,"billing":8.723429559430687e-07,"delivery":8.336809017724569e-08,"refund":1.2701427118032441e-08,"technical":7.106403778547055e-09},"predicted":"account","confidence":0.9999990244811234,"entropy_confidence":0.9999907609460539,"candidate_mass":0.9999994474322057,"accepted":true,"accepted_predicted":"account","input_tokens":384,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":9.75519352385843e-07,"brier_hard":1.719781377007321e-12,"soft_accuracy":0.8000001454512004,"brier_soft":0.06251561976362685,"tvd":0.1999982244803232,"kl":1.6635472669666456,"case_id":"typed:108/r0","base_id":"typed:108","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.619601}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0001238819133443958,"1":0.9991379803727576,"2":0.0006782865840671846,"3":5.98511298305242e-05},"predicted":"1","confidence":0.9991379803727576,"entropy_confidence":0.9945849723346155,"candidate_mass":0.9999982855757305,"accepted":true,"accepted_predicted":"1","input_tokens":370,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0008623913798153219,"brier_hard":1.222079414072543e-06,"score_mae":0.44932589306961646,"within_one":true,"case_id":"typed:108/r0","base_id":"typed:108","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.619601}
{"id":"needs_human","type":"binary","probabilities":{"false":4.809463947157569e-08,"true":0.99999995190536},"predicted":"true","confidence":0.99999995190536,"entropy_confidence":0.9999987614551247,"candidate_mass":0.9999998726117606,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.809464120483791e-08,"brier_hard":4.626188747275491e-15,"soft_accuracy":0.7333329775558666,"brier_soft":0.1422225264769914,"tvd":0.2666669519053602,"kl":2.2286466938720633,"case_id":"typed:108/r0","base_id":"typed:108","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.619601}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.3243085943389195e-06,"1":3.4442438638823017e-06,"2":0.9999921285209764,"3":3.102926564727762e-06},"predicted":"2","confidence":0.9999921285209764,"entropy_confidence":0.9999217519238389,"candidate_mass":0.9999996617390184,"accepted":true,"accepted_predicted":"2","input_tokens":364,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":7.871510003844782e-06,"brier_hard":8.52049443318603e-11,"score_mae":0.07000298993448895,"within_one":true,"case_id":"typed:108/r0","base_id":"typed:108","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.619601}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.1003073734506117e-05,"close_no_action":3.0977741341547895e-05,"escalate_to_human":0.00010086531600176523,"execute_refund":3.681203570137531e-05,"request_information":0.9998203418332212},"predicted":"request_information","confidence":0.9998203418332212,"entropy_confidence":0.9988002889985886,"candidate_mass":0.9999921609596514,"accepted":true,"accepted_predicted":"request_information","input_tokens":388,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":9.201724435964955,"brier_hard":1.999438997921122,"soft_accuracy":0.36665797289363405,"brier_soft":0.7065241690336446,"tvd":0.6331537084998544,"kl":4.996825268455556,"case_id":"typed:109/r0","base_id":"typed:109","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.2419}
{"id":"category","type":"choice","probabilities":{"account":0.9999261880307937,"billing":6.034867815350068e-05,"delivery":6.726050487533194e-06,"refund":1.6439535979619643e-06,"technical":5.093286967709331e-06},"predicted":"account","confidence":0.9999261880307937,"entropy_confidence":0.999487900193207,"candidate_mass":0.9999979813930525,"accepted":true,"accepted_predicted":"account","input_tokens":378,"reused_prefix_tokens":0,"expected":"billing","correct":false,"accepted_correct":false,"nll_hard":9.715371513668053,"brier_hard":1.999731687869334,"soft_accuracy":0.08335734654718571,"brier_soft":1.2458273619551967,"tvd":0.9165931046977102,"kl":8.1029544952086,"case_id":"typed:109/r0","base_id":"typed:109","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.2419}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.649659969205152e-05,"1":1.2338808322057957e-05,"2":0.3888739569129903,"3":0.611087207678996},"predicted":"3","confidence":0.611087207678996,"entropy_confidence":0.5176492429171675,"candidate_mass":0.9999949316892947,"accepted":false,"accepted_predicted":null,"input_tokens":364,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.9445000060874698,"brier_hard":0.7469026167824141,"score_mae":0.3343548756712904,"within_one":true,"case_id":"typed:109/r0","base_id":"typed:109","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.2419}
{"id":"needs_human","type":"binary","probabilities":{"false":1.4413788591451627e-08,"true":0.9999999855862108},"predicted":"true","confidence":0.9999999855862108,"entropy_confidence":0.9999996037552907,"candidate_mass":0.9999999362216576,"accepted":true,"accepted_predicted":"true","input_tokens":332,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4413789266105125e-08,"brier_hard":4.155146195721782e-16,"soft_accuracy":0.8333329903908168,"brier_soft":0.05555576816878862,"tvd":0.16666698558621112,"kl":1.3831249075665977,"case_id":"typed:109/r0","base_id":"typed:109","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.2419}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.7188108040999394e-05,"1":4.678226875537297e-05,"2":0.9998356710162448,"3":0.00010035860695938544},"predicted":"2","confidence":0.9998356710162448,"entropy_confidence":0.998742483205321,"candidate_mass":0.9999982419901093,"accepted":true,"accepted_predicted":"2","input_tokens":358,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001643424872419939,"brier_hard":3.955987662076756e-08,"score_mae":0.2499807998778767,"within_one":true,"case_id":"typed:109/r0","base_id":"typed:109","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.2419}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0008953974675586055,"close_no_action":0.0002791975691995353,"escalate_to_human":0.21098740379175981,"execute_refund":1.2504856996196248e-05,"request_information":0.7878254963144854},"predicted":"request_information","confidence":0.7878254963144854,"entropy_confidence":0.6738755863086735,"candidate_mass":0.9999243660273901,"accepted":false,"accepted_predicted":null,"input_tokens":391,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":null,"nll_hard":1.5559568449574321,"brier_hard":1.2432107694627117,"soft_accuracy":0.3402799536094862,"brier_soft":0.44407062538325887,"tvd":0.5211710011714817,"kl":0.9097077312856987,"case_id":"typed:110/r0","base_id":"typed:110","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.1863}
{"id":"category","type":"choice","probabilities":{"account":3.2586363428291373e-09,"billing":4.5610705108844845e-08,"delivery":5.189511451273634e-08,"refund":2.651044268545903e-08,"technical":0.999999872725101},"predicted":"technical","confidence":0.999999872725101,"entropy_confidence":0.9999985740965105,"candidate_mass":0.9999999838143571,"accepted":true,"accepted_predicted":"technical","input_tokens":381,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":1.272749070528174e-07,"brier_hard":2.1685761516599552e-14,"soft_accuracy":0.9799998758221365,"brier_soft":0.0005999938059507113,"tvd":0.019999944846249,"kl":0.16440827899776345,"case_id":"typed:110/r0","base_id":"typed:110","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.1863}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00023748055503348741,"1":0.0008052149868313401,"2":0.919410402474401,"3":0.07954690198373231},"predicted":"2","confidence":0.919410402474401,"entropy_confidence":0.793452539820559,"candidate_mass":0.9999957115446857,"accepted":true,"accepted_predicted":"2","input_tokens":367,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0840226812576574,"brier_hard":0.012823097612736587,"score_mae":0.04493372588683053,"within_one":true,"case_id":"typed:110/r0","base_id":"typed:110","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.1863}
{"id":"needs_human","type":"binary","probabilities":{"false":2.3989330917067867e-07,"true":0.9999997601066898},"predicted":"true","confidence":0.9999997601066898,"entropy_confidence":0.999994378388503,"candidate_mass":0.9999996887025859,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.398933389909704e-07,"brier_hard":1.1509760007152024e-13,"soft_accuracy":0.7999998560640137,"brier_soft":0.0799998080854673,"tvd":0.19999976010669027,"kl":1.6635534252585398,"case_id":"typed:110/r0","base_id":"typed:110","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.1863}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0021012122016017976,"1":0.039664059597080595,"2":0.941309250629692,"3":0.01692547757162481},"predicted":"2","confidence":0.941309250629692,"entropy_confidence":0.8074475738998456,"candidate_mass":0.9999941299146881,"accepted":true,"accepted_predicted":"2","input_tokens":361,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.06048355297306433,"brier_hard":0.00530872856911281,"score_mae":0.08972599357133904,"within_one":true,"case_id":"typed:110/r0","base_id":"typed:110","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.1863}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00017813408571472946,"close_no_action":0.00043465142368844423,"escalate_to_human":0.001823066886048868,"execute_refund":0.00018269367331685005,"request_information":0.9973814539312317},"predicted":"request_information","confidence":0.9973814539312317,"entropy_confidence":0.9872075541755895,"candidate_mass":0.999968444406971,"accepted":true,"accepted_predicted":"request_information","input_tokens":417,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0026219804572385065,"brier_hard":1.0434386976147777e-05,"soft_accuracy":0.48281737939674163,"brier_soft":0.4363605612339563,"tvd":0.5146657990282367,"kl":2.562770169639026,"case_id":"typed:111/r0","base_id":"typed:111","repeat":0,"workflow":"customer_service","case_elapsed_ms":322.658901}
{"id":"category","type":"choice","probabilities":{"account":8.927650048234523e-07,"billing":0.9999840248291727,"delivery":8.229232581115175e-06,"refund":2.463843370528582e-06,"technical":4.389329871673762e-06},"predicted":"billing","confidence":0.9999840248291727,"entropy_confidence":0.999869070773059,"candidate_mass":0.9999995937613883,"accepted":true,"accepted_predicted":"billing","input_tokens":407,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.597529843169864e-05,"brier_hard":3.490601220661483e-10,"soft_accuracy":0.9233183180432091,"brier_soft":0.011753071698987245,"tvd":0.07666610723499562,"kl":0.6324939548539914,"case_id":"typed:111/r0","base_id":"typed:111","repeat":0,"workflow":"customer_service","case_elapsed_ms":322.658901}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00033050079368782354,"1":0.8577670741864728,"2":0.12686354205446376,"3":0.015038882965375052},"predicted":"1","confidence":0.8577670741864728,"entropy_confidence":0.6686867002093327,"candidate_mass":0.9999972824347494,"accepted":true,"accepted_predicted":"1","input_tokens":393,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.064643242194517,"brier_hard":1.4983579049837201,"score_mae":0.6100561928084745,"within_one":true,"case_id":"typed:111/r0","base_id":"typed:111","repeat":0,"workflow":"customer_service","case_elapsed_ms":322.658901}
{"id":"needs_human","type":"binary","probabilities":{"false":1.8774677051921327e-08,"true":0.9999999812253224},"predicted":"true","confidence":0.9999999812253224,"entropy_confidence":0.9999994910310516,"candidate_mass":0.9999999306362006,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.8774677766616143e-08,"brier_hard":7.049770170263533e-16,"soft_accuracy":0.6333329949934317,"brier_soft":0.26888935024178223,"tvd":0.36666698122532265,"kl":3.08784738240095,"case_id":"typed:111/r0","base_id":"typed:111","repeat":0,"workflow":"customer_service","case_elapsed_ms":322.658901}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.676877453538279e-06,"1":2.160709545420097e-06,"2":3.369638714520025e-05,"3":0.9999614660258546},"predicted":"3","confidence":0.9999614660258546,"entropy_confidence":0.9996767815445423,"candidate_mass":0.9999993872705412,"accepted":true,"accepted_predicted":"3","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.853471659809313e-05,"brier_hard":2.6321480087214504e-09,"score_mae":0.13328695156139947,"within_one":true,"case_id":"typed:111/r0","base_id":"typed:111","repeat":0,"workflow":"customer_service","case_elapsed_ms":322.658901}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.37309314024723e-05,"close_no_action":5.388904692490942e-05,"escalate_to_human":0.00017054955538699283,"execute_refund":3.604545953887182e-05,"request_information":0.9997257850067468},"predicted":"request_information","confidence":0.9997257850067468,"entropy_confidence":0.9982564708526905,"candidate_mass":0.9999934956872603,"accepted":true,"accepted_predicted":"request_information","input_tokens":535,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":8.676484656039442,"brier_hard":1.999110579575576,"soft_accuracy":0.3499741048429127,"brier_soft":0.6239475245565246,"tvd":0.6497257850067468,"kl":4.938006528562437,"case_id":"typed:112/r0","base_id":"typed:112","repeat":0,"workflow":"customer_service","case_elapsed_ms":418.9714}
{"id":"category","type":"choice","probabilities":{"account":4.230779450534793e-12,"billing":0.9999999998569429,"delivery":1.0201742387232257e-10,"refund":3.0741220349740863e-11,"technical":6.06726277302464e-12},"predicted":"billing","confidence":0.9999999998569429,"entropy_confidence":0.9999999978242927,"candidate_mass":0.9999999996202309,"accepted":true,"accepted_predicted":"billing","input_tokens":525,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.4305712171449243e-10,"brier_hard":3.187262864475984e-20,"soft_accuracy":0.9629999998625821,"brier_soft":0.002006877766721497,"tvd":0.036999999958960535,"kl":0.3044756958026797,"case_id":"typed:112/r0","base_id":"typed:112","repeat":0,"workflow":"customer_service","case_elapsed_ms":418.9714}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00010753130927738541,"1":0.9992222917214726,"2":0.0006612525583778796,"3":8.924410873264004e-06},"predicted":"1","confidence":0.9992222917214726,"entropy_confidence":0.995163338128984,"candidate_mass":0.9999981646905314,"accepted":true,"accepted_predicted":"1","input_tokens":511,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.321374705796785,"brier_hard":1.9971231320539293,"score_mae":0.7827614299291519,"within_one":true,"case_id":"typed:112/r0","base_id":"typed:112","repeat":0,"workflow":"customer_service","case_elapsed_ms":418.9714}
{"id":"needs_human","type":"binary","probabilities":{"false":3.50328922164946e-07,"true":0.9999996496710796},"predicted":"true","confidence":0.9999996496710796,"entropy_confidence":0.9999919818533285,"candidate_mass":0.9999997313436785,"accepted":true,"accepted_predicted":"true","input_tokens":479,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.5032898179393467e-07,"brier_hard":2.454607061940221e-13,"soft_accuracy":0.7599998178289619,"brier_soft":0.11519966368448101,"tvd":0.23999964967107867,"kl":2.0019099527909723,"case_id":"typed:112/r0","base_id":"typed:112","repeat":0,"workflow":"customer_service","case_elapsed_ms":418.9714}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012101080334817406,"1":0.9994295651636521,"2":0.00042663330639632475,"3":2.2790726603823513e-05},"predicted":"1","confidence":0.9994295651636521,"entropy_confidence":0.9962375550526009,"candidate_mass":0.999998033849604,"accepted":true,"accepted_predicted":"1","input_tokens":505,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.759585680909186,"brier_hard":1.998006386289424,"score_mae":0.7329817960437437,"within_one":true,"case_id":"typed:112/r0","base_id":"typed:112","repeat":0,"workflow":"customer_service","case_elapsed_ms":418.9714}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00021545960768498307,"close_no_action":0.0003202449495324924,"escalate_to_human":0.0011303626063740924,"execute_refund":0.00021233415970635033,"request_information":0.9981215986767028},"predicted":"request_information","confidence":0.9981215986767028,"entropy_confidence":0.990221386604855,"candidate_mass":0.9999692037850711,"accepted":true,"accepted_predicted":"request_information","input_tokens":304,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":8.442737101325175,"brier_hard":1.9958172783146546,"soft_accuracy":0.29986351868528965,"brier_soft":0.7237431379374453,"tvd":0.6983339328364088,"kl":4.2484354857861035,"case_id":"typed:113/r0","base_id":"typed:113","repeat":0,"workflow":"customer_service","case_elapsed_ms":282.326101}
{"id":"category","type":"choice","probabilities":{"account":0.9999925590755592,"billing":6.9299924455379305e-06,"delivery":4.2714338239137627e-07,"refund":6.248718748817968e-08,"technical":2.1301425903873493e-08},"predicted":"account","confidence":0.9999925590755592,"entropy_confidence":0.999939454528772,"candidate_mass":0.9999993722489594,"accepted":true,"accepted_predicted":"account","input_tokens":294,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":7.440952124642486e-06,"brier_hard":1.0357896169777422e-10,"soft_accuracy":0.8799941929619515,"brier_soft":0.025841247886794195,"tvd":0.11999255907555893,"kl":0.9919271858995876,"case_id":"typed:113/r0","base_id":"typed:113","repeat":0,"workflow":"customer_service","case_elapsed_ms":282.326101}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00017770431647841579,"1":0.9997108742096142,"2":0.00010104383306532015,"3":1.037764084164528e-05},"predicted":"1","confidence":0.9997108742096142,"entropy_confidence":0.9979280582685535,"candidate_mass":0.9999989176795938,"accepted":true,"accepted_predicted":"1","input_tokens":280,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.199956144524016,"brier_hard":1.9992197862431962,"score_mae":0.33338890520173026,"within_one":true,"case_id":"typed:113/r0","base_id":"typed:113","repeat":0,"workflow":"customer_service","case_elapsed_ms":282.326101}
{"id":"needs_human","type":"binary","probabilities":{"false":2.3293427887692225e-06,"true":0.9999976706572107},"predicted":"true","confidence":0.9999976706572107,"entropy_confidence":0.9999530536358661,"candidate_mass":0.9999986099638958,"accepted":true,"accepted_predicted":"true","input_tokens":248,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.969924395080366,"brier_hard":1.9999906826396956,"soft_accuracy":0.45000023293427865,"brier_soft":0.6049948754567158,"tvd":0.549997670657211,"kl":4.70635978949438,"case_id":"typed:113/r0","base_id":"typed:113","repeat":0,"workflow":"customer_service","case_elapsed_ms":282.326101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006664784320843524,"1":0.5246975833823078,"2":0.47367047727152295,"3":0.0009654609140864342},"predicted":"1","confidence":0.5246975833823078,"entropy_confidence":0.49222946497724274,"candidate_mass":0.9999961131627593,"accepted":false,"accepted_predicted":null,"input_tokens":274,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.7472433946772945,"brier_hard":0.5523316968110973,"score_mae":0.09160192066761308,"within_one":true,"case_id":"typed:113/r0","base_id":"typed:113","repeat":0,"workflow":"customer_service","case_elapsed_ms":282.326101}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00029327331351543744,"close_no_action":0.00044479973621374194,"escalate_to_human":0.0027852289633934877,"execute_refund":0.00032446487805643345,"request_information":0.9961522331088198},"predicted":"request_information","confidence":0.9961522331088198,"entropy_confidence":0.9821974534257981,"candidate_mass":0.9999561490132802,"accepted":true,"accepted_predicted":"request_information","input_tokens":390,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.883425195914519,"brier_hard":1.9867569602347788,"soft_accuracy":0.06855734976879599,"brier_soft":1.4454116048926111,"tvd":0.9294852997757537,"kl":5.081429435047601,"case_id":"typed:114/r0","base_id":"typed:114","repeat":0,"workflow":"customer_service","case_elapsed_ms":313.4353}
{"id":"category","type":"choice","probabilities":{"account":0.9999902238338171,"billing":8.589654527431717e-06,"delivery":6.025983617340404e-07,"refund":1.7206602998648583e-07,"technical":4.1184726326945674e-07},"predicted":"account","confidence":0.9999902238338171,"entropy_confidence":0.999920879411132,"candidate_mass":0.9999986115997227,"accepted":true,"accepted_predicted":"account","input_tokens":380,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":9.776213969941724e-06,"brier_hard":1.6991793980915335e-10,"soft_accuracy":0.54999542121228,"brier_soft":0.3425954031909922,"tvd":0.4499902238338173,"kl":3.8037002026361244,"case_id":"typed:114/r0","base_id":"typed:114","repeat":0,"workflow":"customer_service","case_elapsed_ms":313.4353}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0004202627283942249,"1":0.8189605367975651,"2":0.1700130463343003,"3":0.010606154139738692},"predicted":"1","confidence":0.8189605367975651,"entropy_confidence":0.627574484162071,"candidate_mass":0.9999940862308385,"accepted":true,"accepted_predicted":"1","input_tokens":366,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.1997193809083413,"brier_hard":0.061792390286891506,"score_mae":0.14252790811461824,"within_one":true,"case_id":"typed:114/r0","base_id":"typed:114","repeat":0,"workflow":"customer_service","case_elapsed_ms":313.4353}
{"id":"needs_human","type":"binary","probabilities":{"false":1.1614715027157829e-08,"true":0.9999999883852838},"predicted":"true","confidence":0.9999999883852838,"entropy_confidence":0.9999996770857627,"candidate_mass":0.9999999493506286,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.161471625627441e-08,"brier_hard":2.6980323730900674e-16,"soft_accuracy":0.783332993418335,"brier_soft":0.09388916771189798,"tvd":0.21666698838528445,"kl":1.8042889554711128,"case_id":"typed:114/r0","base_id":"typed:114","repeat":0,"workflow":"customer_service","case_elapsed_ms":313.4353}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.435245591033511e-05,"1":2.1614187324029196e-05,"2":0.08049826503754362,"3":0.9194457683192232},"predicted":"3","confidence":0.9194457683192232,"entropy_confidence":0.7975746084235356,"candidate_mass":0.9999973963626589,"accepted":true,"accepted_predicted":"3","input_tokens":360,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.08398421621032168,"brier_hard":0.012968956562999203,"score_mae":0.2560224492200809,"within_one":true,"case_id":"typed:114/r0","base_id":"typed:114","repeat":0,"workflow":"customer_service","case_elapsed_ms":313.4353}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00022676357874515832,"close_no_action":0.0002730693220626842,"escalate_to_human":0.03662266497486945,"execute_refund":9.219803983961175e-05,"request_information":0.9627853040844823},"predicted":"request_information","confidence":0.9627853040844823,"entropy_confidence":0.8989534347706386,"candidate_mass":0.9999666625133317,"accepted":true,"accepted_predicted":"request_information","input_tokens":355,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.307087968629709,"brier_hard":1.8550515658902256,"soft_accuracy":0.18916439025536896,"brier_soft":1.1927454708852263,"tvd":0.7962105021243223,"kl":2.341827797675305,"case_id":"typed:115/r0","base_id":"typed:115","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.191601}
{"id":"category","type":"choice","probabilities":{"account":1.6378874638916425e-07,"billing":2.9485258325137356e-06,"delivery":1.8300787043075868e-06,"refund":1.240804026544888e-06,"technical":0.9999938168026918},"predicted":"technical","confidence":0.9999938168026918,"entropy_confidence":0.9999457316547187,"candidate_mass":0.9999996701930283,"accepted":true,"accepted_predicted":"technical","input_tokens":345,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":6.183216424206021e-06,"brier_hard":5.1841342986365736e-11,"soft_accuracy":0.9816599768572125,"brier_soft":0.0004205791030380136,"tvd":0.018327798468709355,"kl":0.13111822372364731,"case_id":"typed:115/r0","base_id":"typed:115","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.191601}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":9.293475741813119e-05,"1":0.00024748544058338285,"2":0.861102706647152,"3":0.13855687315484685},"predicted":"2","confidence":0.861102706647152,"entropy_confidence":0.7074628316683425,"candidate_mass":0.9999966934814732,"accepted":true,"accepted_predicted":"2","input_tokens":331,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":8.304158803611568,"brier_hard":1.7602009774982452,"score_mae":0.934790518199428,"within_one":true,"case_id":"typed:115/r0","base_id":"typed:115","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.191601}
{"id":"needs_human","type":"binary","probabilities":{"false":2.4747622642432525e-06,"true":0.9999975252377344},"predicted":"true","confidence":0.9999975252377344,"entropy_confidence":0.999950339015221,"candidate_mass":0.9999985237303183,"accepted":true,"accepted_predicted":"true","input_tokens":299,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.4747653277877264e-06,"brier_hard":1.2248896535554823e-11,"soft_accuracy":0.8599982181711686,"brier_soft":0.03919861414538056,"tvd":0.13999752523773512,"kl":1.159742095203106,"case_id":"typed:115/r0","base_id":"typed:115","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.191601}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.4493465593519318e-06,"1":2.7525987536032613e-06,"2":5.380475972833254e-05,"3":0.9999399932949603},"predicted":"3","confidence":0.9999399932949603,"entropy_confidence":0.9995184726352684,"candidate_mass":0.9999993516335307,"accepted":true,"accepted_predicted":"3","input_tokens":325,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.0008505514046874e-05,"brier_hard":6.515231610729359e-09,"score_mae":0.566597342003091,"within_one":true,"case_id":"typed:115/r0","base_id":"typed:115","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.191601}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0011743206393065326,"close_no_action":0.00014846835046553156,"escalate_to_human":0.0003362650130333212,"execute_refund":0.005756674604288559,"request_information":0.992584271392905},"predicted":"request_information","confidence":0.992584271392905,"entropy_confidence":0.9695552626574064,"candidate_mass":0.9999355554259296,"accepted":true,"accepted_predicted":"request_information","input_tokens":345,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":6.747065477909718,"brier_hard":1.9829095479864445,"soft_accuracy":0.33160462262631857,"brier_soft":0.796116298813656,"tvd":0.6616746093308574,"kl":3.6885107962819337,"case_id":"typed:116/r0","base_id":"typed:116","repeat":0,"workflow":"customer_service","case_elapsed_ms":306.80429999999996}
{"id":"category","type":"choice","probabilities":{"account":2.378719987326265e-10,"billing":1.640142421955744e-10,"delivery":0.9999999989865529,"refund":3.5599838074014553e-10,"technical":2.5556229175488333e-10},"predicted":"delivery","confidence":0.9999999989865529,"entropy_confidence":0.9999999854795156,"candidate_mass":0.9999999986252774,"accepted":true,"accepted_predicted":"delivery","input_tokens":335,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":1.0134471045281684e-09,"brier_hard":1.302605724116299e-18,"soft_accuracy":0.9700009690249279,"brier_soft":0.001183258959706751,"tvd":0.02999902898558301,"kl":0.24675678729906217,"case_id":"typed:116/r0","base_id":"typed:116","repeat":0,"workflow":"customer_service","case_elapsed_ms":306.80429999999996}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.381703570464196e-05,"1":0.9999392432932915,"2":6.385517919503544e-06,"3":5.541530850926866e-07},"predicted":"1","confidence":0.9999392432932915,"entropy_confidence":0.9995137141677013,"candidate_mass":0.9999996487750642,"accepted":true,"accepted_predicted":"1","input_tokens":321,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":9.829920492134015,"brier_hard":1.9997708591439065,"score_mae":0.5366206767883859,"within_one":true,"case_id":"typed:116/r0","base_id":"typed:116","repeat":0,"workflow":"customer_service","case_elapsed_ms":306.80429999999996}
{"id":"needs_human","type":"binary","probabilities":{"false":1.0614780633045551e-07,"true":0.9999998938521933},"predicted":"true","confidence":0.9999998938521933,"entropy_confidence":0.9999973876900589,"candidate_mass":0.9999997724980157,"accepted":true,"accepted_predicted":"true","input_tokens":289,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.0584333147624,"brier_hard":1.9999995754087965,"soft_accuracy":0.15000007430346438,"brier_soft":1.4449996390974804,"tvd":0.8499998938521935,"kl":7.544221334369046,"case_id":"typed:116/r0","base_id":"typed:116","repeat":0,"workflow":"customer_service","case_elapsed_ms":306.80429999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.86939916959297e-05,"1":2.7793917476034144e-05,"2":0.9986092924167647,"3":0.001344219674064951},"predicted":"2","confidence":0.9986092924167647,"entropy_confidence":0.9922291058042929,"candidate_mass":0.9999993623621578,"accepted":true,"accepted_predicted":"2","input_tokens":315,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.001391675514536721,"brier_hard":3.7421160813857552e-06,"score_mae":0.11872096222679973,"within_one":true,"case_id":"typed:116/r0","base_id":"typed:116","repeat":0,"workflow":"customer_service","case_elapsed_ms":306.80429999999996}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.982640395487036e-05,"close_no_action":9.655213943955048e-05,"escalate_to_human":0.9997546769916249,"execute_refund":4.91446165994115e-05,"request_information":4.9799848381532285e-05},"predicted":"escalate_to_human","confidence":0.9997546769916249,"entropy_confidence":0.9983767437024861,"candidate_mass":0.9999973728256193,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":369,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.0002453531049866785,"brier_hard":7.688358283916342e-08,"soft_accuracy":0.8531338350210699,"brier_soft":0.028309257625951576,"tvd":0.14642082365777143,"kl":0.8273664062323295,"case_id":"typed:117/r0","base_id":"typed:117","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.565601}
{"id":"category","type":"choice","probabilities":{"account":1.4734582197858037e-05,"billing":7.430025042946868e-05,"delivery":7.245460359748709e-05,"refund":2.3997581077606575e-05,"technical":0.9998145129826993},"predicted":"technical","confidence":0.9998145129826993,"entropy_confidence":0.9987562419348555,"candidate_mass":0.999994145432044,"accepted":true,"accepted_predicted":"technical","input_tokens":359,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.0001855042221450439,"brier_hard":4.596862219358289e-08,"soft_accuracy":0.9431616879526333,"brier_soft":0.003995721852915234,"tvd":0.056480569648755086,"kl":0.27993137294215337,"case_id":"typed:117/r0","base_id":"typed:117","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.565601}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":8.442909749581225e-08,"1":4.816609990496651e-08,"2":2.4736623686966126e-06,"3":0.9999973937424328},"predicted":"3","confidence":0.9999973937424328,"entropy_confidence":0.9999735067625568,"candidate_mass":0.9999998707467069,"accepted":true,"accepted_predicted":"3","input_tokens":345,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.6062609635010487e-06,"brier_hard":1.2921032266607735e-11,"score_mae":0.25333017671813574,"within_one":true,"case_id":"typed:117/r0","base_id":"typed:117","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.565601}
{"id":"needs_human","type":"binary","probabilities":{"false":2.0753695150801352e-08,"true":0.9999999792463046},"predicted":"true","confidence":0.9999999792463046,"entropy_confidence":0.9999994403817659,"candidate_mass":0.9999999032886546,"accepted":true,"accepted_predicted":"true","input_tokens":313,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.0753695644223568e-08,"brier_hard":8.614317363665135e-16,"soft_accuracy":0.8866669839504616,"brier_soft":0.025688728369686678,"tvd":0.11333297924630473,"kl":0.9371820832747431,"case_id":"typed:117/r0","base_id":"typed:117","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.565601}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.6497809156475626e-05,"1":1.665483293510763e-05,"2":0.00024206043095525306,"3":0.9997247869269515},"predicted":"3","confidence":0.9997247869269515,"entropy_confidence":0.9980844060615668,"candidate_mass":0.9999976050383141,"accepted":true,"accepted_predicted":"3","input_tokens":339,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00027525095111609327,"brier_hard":1.3488504897807556e-07,"score_mae":0.38967513647570007,"within_one":true,"case_id":"typed:117/r0","base_id":"typed:117","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.565601}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0015936096985009903,"close_no_action":0.00037292942406751085,"escalate_to_human":0.13956096643261534,"execute_refund":0.8578858153693493,"request_information":0.0005866790754653329},"predicted":"execute_refund","confidence":0.8578858153693493,"entropy_confidence":0.7366121163728891,"candidate_mass":0.9999380565483798,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":413,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":1.9692537378336041,"brier_hard":1.4763264255588722,"soft_accuracy":0.3941508595998367,"brier_soft":0.34196891700242943,"tvd":0.47455281536935007,"kl":0.8783488824743644,"case_id":"typed:118/r0","base_id":"typed:118","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.3681}
{"id":"category","type":"choice","probabilities":{"account":2.954777141826935e-12,"billing":0.9999999998998721,"delivery":6.55899615109039e-11,"refund":2.8387225919823874e-11,"technical":3.1955553783700404e-12},"predicted":"billing","confidence":0.9999999998998721,"entropy_confidence":0.9999999984525857,"candidate_mass":0.9999999996947508,"accepted":true,"accepted_predicted":"billing","input_tokens":403,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.0012790597148672e-10,"brier_hard":1.5152417481790112e-20,"soft_accuracy":0.8833338832482606,"brier_soft":0.022672036427039924,"tvd":0.11666611656598896,"kl":0.9649552865732789,"case_id":"typed:118/r0","base_id":"typed:118","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.3681}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.074966450680297e-07,"1":1.4968823142042134e-07,"2":1.2014271836874146e-05,"3":0.9999876285432857},"predicted":"3","confidence":0.9999876285432857,"entropy_confidence":0.9998888898659808,"candidate_mass":0.9999998284426189,"accepted":true,"accepted_predicted":"3","input_tokens":389,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.2371533241390306e-05,"brier_hard":2.97461130428162e-10,"score_mae":0.16332006386176223,"within_one":true,"case_id":"typed:118/r0","base_id":"typed:118","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.3681}
{"id":"needs_human","type":"binary","probabilities":{"false":5.774357377087246e-09,"true":0.999999994225643},"predicted":"true","confidence":0.999999994225643,"entropy_confidence":0.999999833638527,"candidate_mass":0.9999999580943681,"accepted":true,"accepted_predicted":"true","input_tokens":357,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.774357009232169e-09,"brier_hard":6.668640179585505e-17,"soft_accuracy":0.7599999969973344,"brier_soft":0.11519999445661716,"tvd":0.2399999942256428,"kl":2.0019096909294576,"case_id":"typed:118/r0","base_id":"typed:118","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.3681}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.7076040782344436e-06,"1":1.5104199306570285e-06,"2":5.966661419622674e-05,"3":0.9999371153617943},"predicted":"3","confidence":0.9999371153617943,"entropy_confidence":0.999505035697275,"candidate_mass":0.999999623508774,"accepted":true,"accepted_predicted":"3","input_tokens":383,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.28866155274528e-05,"brier_hard":7.51977985195755e-09,"score_mae":0.2332651897337059,"within_one":true,"case_id":"typed:118/r0","base_id":"typed:118","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.3681}
{"id":"action","type":"choice","probabilities":{"answer_directly":7.470134147171452e-07,"close_no_action":4.3884008931278045e-06,"escalate_to_human":1.1831034331131605e-05,"execute_refund":2.7747862297481973e-06,"request_information":0.9999802587651325},"predicted":"request_information","confidence":0.9999802587651325,"entropy_confidence":0.999842093472733,"candidate_mass":0.999999057414204,"accepted":true,"accepted_predicted":"request_information","input_tokens":622,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":11.344784450902385,"brier_hard":1.999936856018808,"soft_accuracy":0.16667120547809772,"brier_soft":1.0805070849092748,"tvd":0.8333132587651318,"kl":7.367490376481251,"case_id":"typed:119/r0","base_id":"typed:119","repeat":0,"workflow":"customer_service","case_elapsed_ms":517.6776010000001}
{"id":"category","type":"choice","probabilities":{"account":2.0353761414203377e-10,"billing":0.9999999960387438,"delivery":2.6562289749925832e-09,"refund":8.521180717456906e-10,"technical":2.4937212530471586e-10},"predicted":"billing","confidence":0.9999999960387438,"entropy_confidence":0.9999999476442984,"candidate_mass":0.99999999647933,"accepted":true,"accepted_predicted":"billing","input_tokens":612,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.9612562197831e-09,"brier_hard":2.357682236964741e-17,"soft_accuracy":0.9333329963423269,"brier_soft":0.005977854948833687,"tvd":0.06666699603874351,"kl":0.5496321082189026,"case_id":"typed:119/r0","base_id":"typed:119","repeat":0,"workflow":"customer_service","case_elapsed_ms":517.6776010000001}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.42861153072679e-05,"1":0.016261197713477147,"2":0.9835816567357278,"3":0.00012285943548936902},"predicted":"2","confidence":0.9835816567357278,"entropy_confidence":0.938886692919146,"candidate_mass":0.9999968220831666,"accepted":true,"accepted_predicted":"2","input_tokens":598,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.016554617923667123,"brier_hard":0.000534004816598857,"score_mae":0.08287391050859894,"within_one":true,"case_id":"typed:119/r0","base_id":"typed:119","repeat":0,"workflow":"customer_service","case_elapsed_ms":517.6776010000001}
{"id":"needs_human","type":"binary","probabilities":{"false":2.293553023047454e-09,"true":0.9999999977064462},"predicted":"true","confidence":0.9999999977064462,"entropy_confidence":0.9999999308666603,"candidate_mass":0.9999999926188838,"accepted":true,"accepted_predicted":"true","input_tokens":566,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.2935537939628643e-09,"brier_hard":1.0520774463266579e-17,"soft_accuracy":0.7499999988532229,"brier_soft":0.12499999770644661,"tvd":0.2499999977064466,"kl":2.0868235403753754,"case_id":"typed:119/r0","base_id":"typed:119","repeat":0,"workflow":"customer_service","case_elapsed_ms":517.6776010000001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007142338748782591,"1":0.05476592239354341,"2":0.9415735599302747,"3":0.002946283801302501},"predicted":"2","confidence":0.9415735599302747,"entropy_confidence":0.8282428426024399,"candidate_mass":0.9999942693255581,"accepted":true,"accepted_predicted":"2","input_tokens":592,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.06020280336842992,"brier_hard":0.006422145873102663,"score_mae":0.11341889365800029,"within_one":true,"case_id":"typed:119/r0","base_id":"typed:119","repeat":0,"workflow":"customer_service","case_elapsed_ms":517.6776010000001}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.4767615538009221e-05,"close_no_action":8.422819904630702e-06,"escalate_to_human":4.4441328193805815e-05,"execute_refund":5.1809453106909824e-06,"request_information":0.9999271872910528},"predicted":"request_information","confidence":0.9999271872910528,"entropy_confidence":0.9994756524434875,"candidate_mass":0.9999981484805626,"accepted":true,"accepted_predicted":"request_information","input_tokens":609,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":10.021340706290097,"brier_hard":1.9997654995183087,"soft_accuracy":0.11669179136029029,"brier_soft":1.2854263972321158,"tvd":0.8832601872910528,"kl":7.809346844045135,"case_id":"typed:120/r0","base_id":"typed:120","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.670301}
{"id":"category","type":"choice","probabilities":{"account":3.674851469885468e-05,"billing":0.00024551741121421746,"delivery":3.550403399914278e-05,"refund":2.2000032143851734e-05,"technical":0.9996602300079444},"predicted":"technical","confidence":0.9996602300079444,"entropy_confidence":0.9979151414493047,"candidate_mass":0.9999963731824844,"accepted":true,"accepted_predicted":"technical","input_tokens":599,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.00033982772695749546,"brier_hard":1.788174378879421e-07,"soft_accuracy":0.926359847227834,"brier_soft":0.006762669933658611,"tvd":0.07299323000794414,"kl":0.3480107146996496,"case_id":"typed:120/r0","base_id":"typed:120","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.670301}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.106132337657041e-06,"1":9.470950456906607e-06,"2":0.003926420092606647,"3":0.9960600028245994},"predicted":"3","confidence":0.9960600028245994,"entropy_confidence":0.9813566428737914,"candidate_mass":0.9999995606468955,"accepted":true,"accepted_predicted":"3","input_tokens":585,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.003947779412325064,"brier_hard":3.0940459045015316e-05,"score_mae":0.4427093196094689,"within_one":true,"case_id":"typed:120/r0","base_id":"typed:120","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.670301}
{"id":"needs_human","type":"binary","probabilities":{"false":9.332897118115074e-09,"true":0.9999999906671029},"predicted":"true","confidence":0.9999999906671029,"entropy_confidence":0.999999737580185,"candidate_mass":0.9999999842688397,"accepted":true,"accepted_predicted":"true","input_tokens":553,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.332897165345788e-09,"brier_hard":1.7420593730331702e-16,"soft_accuracy":0.8199999940269459,"brier_soft":0.06479999328031429,"tvd":0.17999999066710293,"kl":1.495131504855142,"case_id":"typed:120/r0","base_id":"typed:120","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.670301}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.3997620349740616e-06,"1":1.0928657284414578e-05,"2":0.9999692294570177,"3":1.744212366426964e-05},"predicted":"2","confidence":0.9999692294570177,"entropy_confidence":0.9997274890405914,"candidate_mass":0.9999993308660287,"accepted":true,"accepted_predicted":"2","input_tokens":579,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":3.077101640515293e-05,"brier_hard":1.3762484012089247e-09,"score_mae":0.06333128605768756,"within_one":true,"case_id":"typed:120/r0","base_id":"typed:120","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.670301}
{"id":"action","type":"choice","probabilities":{"answer_directly":3.7375511084023735e-05,"close_no_action":2.457917859352742e-05,"escalate_to_human":0.0001210851101100513,"execute_refund":1.4955601497901629e-05,"request_information":0.9998020045987153},"predicted":"request_information","confidence":0.9998020045987153,"entropy_confidence":0.9986963735540996,"candidate_mass":0.9999917648944263,"accepted":true,"accepted_predicted":"request_information","input_tokens":359,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.00019801500496186697,"brier_hard":5.608851768555464e-08,"soft_accuracy":0.5166032679875088,"brier_soft":0.3369534070889306,"tvd":0.4831499602002128,"kl":3.5245208469233726,"case_id":"typed:121/r0","base_id":"typed:121","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.30379999999997}
{"id":"category","type":"choice","probabilities":{"account":6.039194181886902e-07,"billing":1.5668848529169874e-05,"delivery":7.620781623920737e-06,"refund":6.0358260513013324e-06,"technical":0.9999700706243771},"predicted":"technical","confidence":0.9999700706243771,"entropy_confidence":0.9997674467857194,"candidate_mass":0.9999982632454962,"accepted":true,"accepted_predicted":"technical","input_tokens":349,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":2.9929823515549628e-05,"brier_hard":1.236152566748453e-09,"soft_accuracy":0.9566386420364238,"brier_soft":0.00237459396805953,"tvd":0.043303070624377264,"kl":0.2944922284884018,"case_id":"typed:121/r0","base_id":"typed:121","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.30379999999997}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0003604798256180728,"1":0.9838350100599661,"2":0.011478362240227948,"3":0.0043261478741875095},"predicted":"1","confidence":0.9838350100599661,"entropy_confidence":0.9323979881993913,"candidate_mass":0.9999970996149795,"accepted":true,"accepted_predicted":"1","input_tokens":335,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.467291560268958,"brier_hard":1.9451252008400897,"score_mae":0.7302298218370156,"within_one":true,"case_id":"typed:121/r0","base_id":"typed:121","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.30379999999997}
{"id":"needs_human","type":"binary","probabilities":{"false":8.655280617725722e-07,"true":0.9999991344719399},"predicted":"true","confidence":0.9999991344719399,"entropy_confidence":0.9999813196449805,"candidate_mass":0.9999993128838669,"accepted":true,"accepted_predicted":"true","input_tokens":303,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.95992604024157,"brier_hard":1.9999965378892546,"soft_accuracy":0.3833332019571254,"brier_soft":0.7605542428091274,"tvd":0.6166661344719391,"kl":5.312153984822345,"case_id":"typed:121/r0","base_id":"typed:121","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.30379999999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006725242423154361,"1":0.8353827191072023,"2":0.1631617063802141,"3":0.0007830502702690855},"predicted":"1","confidence":0.8353827191072023,"entropy_confidence":0.6706444026204872,"candidate_mass":0.9999927545011977,"accepted":true,"accepted_predicted":"1","input_tokens":329,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.8130135062630366,"brier_hard":1.3981636825080002,"score_mae":0.5026117173215621,"within_one":true,"case_id":"typed:121/r0","base_id":"typed:121","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.30379999999997}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.098031713139377e-05,"close_no_action":4.4793578440549456e-05,"escalate_to_human":0.000151861533377963,"execute_refund":3.5922798735003606e-05,"request_information":0.9997464417723155},"predicted":"request_information","confidence":0.9997464417723155,"entropy_confidence":0.9983652986776904,"candidate_mass":0.9999931167424807,"accepted":true,"accepted_predicted":"request_information","input_tokens":390,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0002535903790068191,"brier_hard":9.109078599179451e-08,"soft_accuracy":0.4665777099188699,"brier_soft":0.4491153103536773,"tvd":0.5331153645710502,"kl":4.177416989827956,"case_id":"typed:122/r0","base_id":"typed:122","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.0099}
{"id":"category","type":"choice","probabilities":{"account":0.999999999943407,"billing":3.864791964684271e-11,"delivery":1.4446163043472048e-11,"refund":1.5705009596254827e-12,"technical":1.9283881368612394e-12},"predicted":"account","confidence":0.999999999943407,"entropy_confidence":0.9999999991061952,"candidate_mass":0.999999998491349,"accepted":true,"accepted_predicted":"account","input_tokens":380,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":5.659295254805396e-11,"brier_hard":4.911300751702994e-21,"soft_accuracy":0.8866661132842533,"brier_soft":0.022289066519925595,"tvd":0.11333388660952037,"kl":0.9371894514076688,"case_id":"typed:122/r0","base_id":"typed:122","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.0099}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00018784474934679945,"1":0.0009814260870038825,"2":0.047747260010773826,"3":0.9510834691528746},"predicted":"3","confidence":0.9510834691528746,"entropy_confidence":0.8547574868611687,"candidate_mass":0.9999982620806417,"accepted":true,"accepted_predicted":"3","input_tokens":366,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":6.926503853243489,"brier_hard":1.9048777124432097,"score_mae":1.5497263535671753,"within_one":false,"case_id":"typed:122/r0","base_id":"typed:122","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.0099}
{"id":"needs_human","type":"binary","probabilities":{"false":2.824268315838826e-08,"true":0.9999999717573176},"predicted":"true","confidence":0.9999999717573176,"entropy_confidence":0.9999992509971913,"candidate_mass":0.9999999220044694,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.382431416143987,"brier_hard":1.9999998870292703,"soft_accuracy":0.3000000112970735,"brier_soft":0.9799999209204896,"tvd":0.6999999717573171,"kl":6.086046427558352,"case_id":"typed:122/r0","base_id":"typed:122","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.0099}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.7307692566117112e-06,"1":1.199306800875034e-06,"2":4.3504917626757195e-05,"3":0.9999535650063144},"predicted":"3","confidence":0.9999535650063144,"entropy_confidence":0.9996229869381351,"candidate_mass":0.9999997706703759,"accepted":true,"accepted_predicted":"3","input_tokens":360,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":4.643607182331284e-05,"brier_hard":4.0533203953165575e-09,"score_mae":0.44328190416099744,"within_one":true,"case_id":"typed:122/r0","base_id":"typed:122","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.0099}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0001778042711973617,"close_no_action":2.0481966390590953e-05,"escalate_to_human":0.0001919818280513271,"execute_refund":0.9993911886526087,"request_information":0.00021854328175342094},"predicted":"execute_refund","confidence":0.9993911886526087,"entropy_confidence":0.9963651551180354,"candidate_mass":0.9999849053657395,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":302,"reused_prefix_tokens":0,"expected":"execute_refund","correct":true,"accepted_correct":true,"nll_hard":0.0006089967482729403,"brier_hard":4.873033148171751e-07,"soft_accuracy":0.4997950968916785,"brier_soft":0.36752560082594693,"tvd":0.4994121706184986,"kl":3.162779163517705,"case_id":"typed:123/r0","base_id":"typed:123","repeat":0,"workflow":"customer_service","case_elapsed_ms":267.0711}
{"id":"category","type":"choice","probabilities":{"account":4.9839165942242074e-12,"billing":0.999999999680087,"delivery":2.1543172551999006e-10,"refund":8.885109995155827e-11,"technical":1.0645869772469974e-11},"predicted":"billing","confidence":0.999999999680087,"entropy_confidence":0.9999999952964131,"candidate_mass":0.9999999995311573,"accepted":true,"accepted_predicted":"billing","input_tokens":292,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.199129850429603e-10,"brier_hard":1.5678783825730875e-19,"soft_accuracy":0.816666999755027,"brier_soft":0.0672219776281201,"tvd":0.1833329999111487,"kl":1.523164781811048,"case_id":"typed:123/r0","base_id":"typed:123","repeat":0,"workflow":"customer_service","case_elapsed_ms":267.0711}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00028033435698724904,"1":0.03345178878663543,"2":0.9639693950604282,"3":0.00229848179595055},"predicted":"2","confidence":0.9639693950604282,"entropy_confidence":0.88076956649644,"candidate_mass":0.9999984544489713,"accepted":true,"accepted_predicted":"2","input_tokens":278,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":3.3976500174061606,"brier_hard":1.8634578008188407,"score_mae":0.6349530242953434,"within_one":true,"case_id":"typed:123/r0","base_id":"typed:123","repeat":0,"workflow":"customer_service","case_elapsed_ms":267.0711}
{"id":"needs_human","type":"binary","probabilities":{"false":3.2599973506615785e-07,"true":0.9999996740002642},"predicted":"true","confidence":0.9999996740002642,"entropy_confidence":0.9999925048355491,"candidate_mass":0.99999956385411,"accepted":true,"accepted_predicted":"true","input_tokens":246,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.936369268260531,"brier_hard":1.9999986960012708,"soft_accuracy":0.33333310866679544,"brier_soft":0.8888889084451501,"tvd":0.6666666740002646,"kl":5.774026030078961,"case_id":"typed:123/r0","base_id":"typed:123","repeat":0,"workflow":"customer_service","case_elapsed_ms":267.0711}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0011564306481477479,"1":0.0021839345840862084,"2":0.7876953687526339,"3":0.20896426601513285},"predicted":"2","confidence":0.7876953687526339,"entropy_confidence":0.6131183405312925,"candidate_mass":0.9999960749050034,"accepted":false,"accepted_predicted":null,"input_tokens":272,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":6.126627173889538,"brier_hard":1.6597682961595304,"score_mae":0.6878004701347527,"within_one":true,"case_id":"typed:123/r0","base_id":"typed:123","repeat":0,"workflow":"customer_service","case_elapsed_ms":267.0711}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0012352590670971471,"close_no_action":0.0016099388313407313,"escalate_to_human":0.0026772872177807383,"execute_refund":0.05320678042688835,"request_information":0.9412707344568916},"predicted":"request_information","confidence":0.9412707344568916,"entropy_confidence":0.8461950994923422,"candidate_mass":0.9998376801974551,"accepted":true,"accepted_predicted":"request_information","input_tokens":501,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.922951229283701,"brier_hard":1.8834782682276998,"soft_accuracy":0.10225721524470407,"brier_soft":1.0925180224412245,"tvd":0.887810621550674,"kl":4.568625016401098,"case_id":"typed:124/r0","base_id":"typed:124","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.3677}
{"id":"category","type":"choice","probabilities":{"account":6.751700665148674e-12,"billing":0.9999999998138254,"delivery":1.2567531342533108e-10,"refund":4.291349566800711e-11,"technical":1.0833842363969229e-11},"predicted":"billing","confidence":0.9999999998138254,"entropy_confidence":0.9999999971897907,"candidate_mass":0.9999999995661693,"accepted":true,"accepted_predicted":"billing","input_tokens":491,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.8617463128435254e-10,"brier_hard":5.2459803444653995e-20,"soft_accuracy":0.9733329998199244,"brier_soft":0.000911135543802042,"tvd":0.026666999813825507,"kl":0.21930391476607192,"case_id":"typed:124/r0","base_id":"typed:124","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.3677}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.178942199614537e-05,"1":4.358516241024891e-05,"2":0.21061521203400013,"3":0.7892894133815948},"predicted":"3","confidence":0.7892894133815948,"entropy_confidence":0.627934605873123,"candidate_mass":0.9999948376917537,"accepted":false,"accepted_predicted":null,"input_tokens":477,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.557722450010516,"brier_hard":1.246106126130199,"score_mae":0.9058092493751948,"within_one":true,"case_id":"typed:124/r0","base_id":"typed:124","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.3677}
{"id":"needs_human","type":"binary","probabilities":{"false":6.999965022291505e-09,"true":0.999999993000035},"predicted":"true","confidence":0.999999993000035,"entropy_confidence":0.9999998002720968,"candidate_mass":0.9999999703865581,"accepted":true,"accepted_predicted":"true","input_tokens":445,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.9999650765455645e-09,"brier_hard":9.79990210431672e-17,"soft_accuracy":0.6266669982266708,"brier_soft":0.27875504732472833,"tvd":0.373332993000035,"kl":3.1456574632864323,"case_id":"typed:124/r0","base_id":"typed:124","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.3677}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.57366346868002e-07,"1":1.459443655814153e-07,"2":1.4472799404901244e-06,"3":0.9999980494093473},"predicted":"3","confidence":0.9999980494093473,"entropy_confidence":0.999979071863525,"candidate_mass":0.999999886452325,"accepted":true,"accepted_predicted":"3","input_tokens":471,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.9505925550706495e-06,"brier_hard":6.048433584132705e-12,"score_mae":0.0666641887322883,"within_one":true,"case_id":"typed:124/r0","base_id":"typed:124","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.3677}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.765103956276649e-06,"close_no_action":9.780370730278022e-06,"escalate_to_human":3.062795485530946e-05,"execute_refund":7.083130637909673e-06,"request_information":0.9999477434398188},"predicted":"request_information","confidence":0.9999477434398188,"entropy_confidence":0.9996111744951718,"candidate_mass":0.9999973972935421,"accepted":true,"accepted_predicted":"request_information","input_tokens":514,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":10.393597408682439,"brier_hard":1.9998342348072793,"soft_accuracy":0.09001351025549377,"brier_soft":1.2422903479840026,"tvd":0.9099548265704576,"kl":8.135206550976648,"case_id":"typed:125/r0","base_id":"typed:125","repeat":0,"workflow":"customer_service","case_elapsed_ms":374.5377}
{"id":"category","type":"choice","probabilities":{"account":0.9999999978556406,"billing":1.702254453672765e-09,"delivery":3.94412961694681e-10,"refund":2.74113860529515e-11,"technical":2.0281225804907808e-11},"predicted":"account","confidence":0.9999999978556406,"entropy_confidence":0.9999999712809192,"candidate_mass":0.9999999944592215,"accepted":true,"accepted_predicted":"account","input_tokens":504,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":2.1443593590474207e-09,"brier_hard":7.652671572480933e-18,"soft_accuracy":0.7799992186715794,"brier_soft":0.0885336079703747,"tvd":0.22000077785486033,"kl":1.8324816211136716,"case_id":"typed:125/r0","base_id":"typed:125","repeat":0,"workflow":"customer_service","case_elapsed_ms":374.5377}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.14072295772213e-05,"1":0.09408648697550916,"2":0.9051545056121598,"3":0.0006976001827539435},"predicted":"2","confidence":0.9051545056121598,"entropy_confidence":0.770437533937603,"candidate_mass":0.9999961596838339,"accepted":true,"accepted_predicted":"2","input_tokens":490,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.09964962542064275,"brier_hard":0.017848425253929302,"score_mae":0.10351170125190912,"within_one":true,"case_id":"typed:125/r0","base_id":"typed:125","repeat":0,"workflow":"customer_service","case_elapsed_ms":374.5377}
{"id":"needs_human","type":"binary","probabilities":{"false":5.592672104176849e-09,"true":0.9999999944073288},"predicted":"true","confidence":0.9999999944073288,"entropy_confidence":0.999999838615002,"candidate_mass":0.9999999666933933,"accepted":true,"accepted_predicted":"true","input_tokens":458,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.592671230665698e-09,"brier_hard":6.255595258422627e-17,"soft_accuracy":0.6333329985086251,"brier_soft":0.26888936957540743,"tvd":0.3666669944073283,"kl":3.08784737405235,"case_id":"typed:125/r0","base_id":"typed:125","repeat":0,"workflow":"customer_service","case_elapsed_ms":374.5377}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001394101171464697,"1":0.9213893661641118,"2":0.07819039951283345,"3":0.00028082420590733686},"predicted":"1","confidence":0.9213893661641118,"entropy_confidence":0.7992869010107719,"candidate_mass":0.9999886159606043,"accepted":true,"accepted_predicted":"1","input_tokens":484,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.548608407345615,"brier_hard":1.6986914019280288,"score_mae":0.5047203621924994,"within_one":true,"case_id":"typed:125/r0","base_id":"typed:125","repeat":0,"workflow":"customer_service","case_elapsed_ms":374.5377}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0013267633277703208,"close_no_action":0.00036032900568322147,"escalate_to_human":0.002717319929776178,"execute_refund":0.00024584116843171764,"request_information":0.9953497465683375},"predicted":"request_information","confidence":0.9953497465683375,"entropy_confidence":0.9786363426072483,"candidate_mass":0.9999645147290525,"accepted":true,"accepted_predicted":"request_information","input_tokens":505,"reused_prefix_tokens":0,"expected":"execute_refund","correct":false,"accepted_correct":false,"nll_hard":8.310824887339649,"brier_hard":1.9902387700601913,"soft_accuracy":0.15044334462430667,"brier_soft":1.0070660409264414,"tvd":0.8457100755740212,"kl":4.849605523323847,"case_id":"typed:126/r0","base_id":"typed:126","repeat":0,"workflow":"customer_service","case_elapsed_ms":353.28900000000004}
{"id":"category","type":"choice","probabilities":{"account":2.6487979846979573e-08,"billing":0.9999994135831557,"delivery":2.7293517375344525e-07,"refund":1.2044245741691626e-07,"technical":1.6655123337024773e-07},"predicted":"billing","confidence":0.9999994135831557,"entropy_confidence":0.9999939779484955,"candidate_mass":0.9999999210055637,"accepted":true,"accepted_predicted":"billing","input_tokens":495,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":5.864170162483618e-07,"brier_hard":4.613256363197073e-13,"soft_accuracy":0.8799994985599525,"brier_soft":0.028022127824867814,"tvd":0.11999971300630925,"kl":0.9927479937554177,"case_id":"typed:126/r0","base_id":"typed:126","repeat":0,"workflow":"customer_service","case_elapsed_ms":353.28900000000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":8.235857344983953e-05,"1":0.06897555476783748,"2":0.9303374405442495,"3":0.0006046461144613025},"predicted":"2","confidence":0.9303374405442495,"entropy_confidence":0.8147045290067366,"candidate_mass":0.9999965713523035,"accepted":true,"accepted_predicted":"2","input_tokens":481,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0722079193502019,"brier_hard":0.009610871725315277,"score_mae":0.1647973741997204,"within_one":true,"case_id":"typed:126/r0","base_id":"typed:126","repeat":0,"workflow":"customer_service","case_elapsed_ms":353.28900000000004}
{"id":"needs_human","type":"binary","probabilities":{"false":1.572620003205887e-08,"true":0.9999999842737993},"predicted":"true","confidence":0.9999999842737993,"entropy_confidence":0.9999995696533266,"candidate_mass":0.9999999486878859,"accepted":true,"accepted_predicted":"true","input_tokens":449,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.9679377236272,"brier_hard":1.999999937095199,"soft_accuracy":0.4833330005242169,"brier_soft":0.5338895452771655,"tvd":0.5166669842737996,"kl":4.40727195833795,"case_id":"typed:126/r0","base_id":"typed:126","repeat":0,"workflow":"customer_service","case_elapsed_ms":353.28900000000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003744969848279838,"1":0.003950050183595733,"2":0.9811560191877644,"3":0.014519433643810862},"predicted":"2","confidence":0.9811560191877644,"entropy_confidence":0.9243090990597163,"candidate_mass":0.9999962810635398,"accepted":true,"accepted_predicted":"2","input_tokens":475,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.019023791096326955,"brier_hard":0.0005816527106335,"score_mae":0.19315338949055705,"within_one":true,"case_id":"typed:126/r0","base_id":"typed:126","repeat":0,"workflow":"customer_service","case_elapsed_ms":353.28900000000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0003463395136079528,"close_no_action":0.0002851123571063753,"escalate_to_human":0.19393928138810745,"execute_refund":2.936995656331065e-05,"request_information":0.8053998967846133},"predicted":"request_information","confidence":0.8053998967846133,"entropy_confidence":0.6907017604198074,"candidate_mass":0.9999011373171064,"accepted":true,"accepted_predicted":"request_information","input_tokens":578,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":1.6402101514435934,"brier_hard":1.2984030779324955,"soft_accuracy":0.26249118300764085,"brier_soft":0.6186098702494289,"tvd":0.6320668967846141,"kl":1.6384346134990386,"case_id":"typed:127/r0","base_id":"typed:127","repeat":0,"workflow":"customer_service","case_elapsed_ms":512.2863}
{"id":"category","type":"choice","probabilities":{"account":0.9999995241754003,"billing":4.071062316347737e-07,"delivery":6.016586809249267e-08,"refund":5.0628718861377335e-09,"technical":3.489628034258846e-09},"predicted":"account","confidence":0.9999995241754003,"entropy_confidence":0.9999952585597095,"candidate_mass":0.9999998885248056,"accepted":true,"accepted_predicted":"account","input_tokens":568,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":4.7582471292618005e-07,"brier_hard":3.9580227539498136e-13,"soft_accuracy":0.4933329046565839,"brier_soft":0.3940441145940286,"tvd":0.5066665241754004,"kl":4.318001029738174,"case_id":"typed:127/r0","base_id":"typed:127","repeat":0,"workflow":"customer_service","case_elapsed_ms":512.2863}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":7.007040460843429e-06,"1":8.304703490744955e-06,"2":0.9905293584183127,"3":0.009455329837733997},"predicted":"2","confidence":0.9905293584183127,"entropy_confidence":0.9612788342528,"candidate_mass":0.999997872740106,"accepted":true,"accepted_predicted":"2","input_tokens":554,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.009515773284490997,"brier_hard":0.00017909643237584363,"score_mae":0.10723398894668179,"within_one":true,"case_id":"typed:127/r0","base_id":"typed:127","repeat":0,"workflow":"customer_service","case_elapsed_ms":512.2863}
{"id":"needs_human","type":"binary","probabilities":{"false":3.8766635901653124e-08,"true":0.9999999612333647},"predicted":"true","confidence":0.9999999612333647,"entropy_confidence":0.9999989896133755,"candidate_mass":0.9999999313473525,"accepted":true,"accepted_predicted":"true","input_tokens":522,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.876663609772501e-08,"brier_hard":3.0057040752042585e-15,"soft_accuracy":0.7399999813920152,"brier_soft":0.135199959682702,"tvd":0.2599999612333644,"kl":2.1718707567410167,"case_id":"typed:127/r0","base_id":"typed:127","repeat":0,"workflow":"customer_service","case_elapsed_ms":512.2863}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00024407630820856503,"1":0.0075210379342069665,"2":0.9910072365163164,"3":0.0012276492412678862},"predicted":"2","confidence":0.9910072365163164,"entropy_confidence":0.9596123228605418,"candidate_mass":0.9999967353791275,"accepted":true,"accepted_predicted":"2","input_tokens":548,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.009033442442365869,"brier_hard":0.00013900250258506788,"score_mae":0.12655145869064333,"within_one":true,"case_id":"typed:127/r0","base_id":"typed:127","repeat":0,"workflow":"customer_service","case_elapsed_ms":512.2863}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.0923229306671293e-06,"close_no_action":3.6311804256466378e-06,"escalate_to_human":1.4090670411590458e-05,"execute_refund":3.5109859408802554e-06,"request_information":0.9999766748402914},"predicted":"request_information","confidence":0.9999766748402914,"entropy_confidence":0.9998150534031975,"candidate_mass":0.9999990742066367,"accepted":true,"accepted_predicted":"request_information","input_tokens":258,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":2.3325431744366073e-05,"brier_hard":7.725003768865356e-10,"soft_accuracy":0.743318070727407,"brier_soft":0.09231777366581981,"tvd":0.256642931506548,"kl":2.0785613625979473,"case_id":"typed:128/r0","base_id":"typed:128","repeat":0,"workflow":"customer_service","case_elapsed_ms":216.5257}
{"id":"category","type":"choice","probabilities":{"account":2.2463406406999576e-09,"billing":1.5251483200254908e-09,"delivery":0.9999999912350788,"refund":3.874104376838943e-09,"technical":1.1193288368425388e-09},"predicted":"delivery","confidence":0.9999999912350788,"entropy_confidence":0.999999886564077,"candidate_mass":0.9999999943539812,"accepted":true,"accepted_predicted":"delivery","input_tokens":248,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":8.76492127166271e-09,"brier_hard":1.0045754966485732e-16,"soft_accuracy":0.9338559919882087,"brier_soft":0.00566661961774027,"tvd":0.06614399123507829,"kl":0.5453021677398714,"case_id":"typed:128/r0","base_id":"typed:128","repeat":0,"workflow":"customer_service","case_elapsed_ms":216.5257}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00011509925733231705,"1":0.9998062281415416,"2":7.66266557260249e-05,"3":2.045945398493992e-06},"predicted":"1","confidence":0.9998062281415416,"entropy_confidence":0.9985640644756959,"candidate_mass":0.9999993247608719,"accepted":true,"accepted_predicted":"1","input_tokens":234,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00019379063465050533,"brier_hard":5.6671202429189566e-08,"score_mae":0.05670138071081088,"within_one":true,"case_id":"typed:128/r0","base_id":"typed:128","repeat":0,"workflow":"customer_service","case_elapsed_ms":216.5257}
{"id":"needs_human","type":"binary","probabilities":{"false":1.3918583655776984e-08,"true":0.9999999860814165},"predicted":"true","confidence":0.9999999860814165,"entropy_confidence":0.9999996166667932,"candidate_mass":0.9999999388015209,"accepted":true,"accepted_predicted":"true","input_tokens":202,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.0900409360835,"brier_hard":1.9999999443256662,"soft_accuracy":0.27666700621695817,"brier_soft":1.0464212175069172,"tvd":0.7233329860814164,"kl":6.30664245798855,"case_id":"typed:128/r0","base_id":"typed:128","repeat":0,"workflow":"customer_service","case_elapsed_ms":216.5257}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003367029173447788,"1":0.9985035651834581,"2":0.0011103416464578976,"3":4.93902527408769e-05},"predicted":"1","confidence":0.9985035651834581,"entropy_confidence":0.9911770624852246,"candidate_mass":0.9999968832587034,"accepted":true,"accepted_predicted":"1","input_tokens":228,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0014975555933745414,"brier_hard":3.5879839836320257e-06,"score_mae":0.3091275807654035,"within_one":true,"case_id":"typed:128/r0","base_id":"typed:128","repeat":0,"workflow":"customer_service","case_elapsed_ms":216.5257}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.002151618919391228,"close_no_action":0.00034024080568869535,"escalate_to_human":0.0024659071011954702,"execute_refund":0.993573968310265,"request_information":0.0014682648634611005},"predicted":"execute_refund","confidence":0.993573968310265,"entropy_confidence":0.9709690990149362,"candidate_mass":0.999971978955547,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":250,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":6.141534734430753,"brier_hard":1.9828989743923462,"soft_accuracy":0.002007067527878126,"brier_soft":1.4531217037091555,"tvd":0.9935739683102642,"kl":5.277234369024181,"case_id":"typed:129/r0","base_id":"typed:129","repeat":0,"workflow":"customer_service","case_elapsed_ms":199.0384}
{"id":"category","type":"choice","probabilities":{"account":0.012639124903172346,"billing":0.9873267535268921,"delivery":2.9849604232697346e-05,"refund":3.888867828622159e-06,"technical":3.8309787388864455e-07},"predicted":"billing","confidence":0.9873267535268921,"entropy_confidence":0.9576232154122934,"candidate_mass":0.9999830547926358,"accepted":true,"accepted_predicted":"billing","input_tokens":240,"reused_prefix_tokens":0,"expected":"account","correct":false,"accepted_correct":false,"nll_hard":4.370958125004055,"brier_hard":1.9496956168081947,"soft_accuracy":0.2788005846876381,"brier_soft":0.9915940732409368,"tvd":0.713994026859619,"kl":2.639525469772302,"case_id":"typed:129/r0","base_id":"typed:129","repeat":0,"workflow":"customer_service","case_elapsed_ms":199.0384}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.3607956427236783e-07,"1":7.634901168449699e-08,"2":2.878113545464624e-06,"3":0.9999968094578785},"predicted":"3","confidence":0.9999968094578785,"entropy_confidence":0.9999677094876621,"candidate_mass":0.999999875938333,"accepted":true,"accepted_predicted":"3","input_tokens":226,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.1905472112909465e-06,"brier_hard":1.8524659341908766e-11,"score_mae":0.16999626094973808,"within_one":true,"case_id":"typed:129/r0","base_id":"typed:129","repeat":0,"workflow":"customer_service","case_elapsed_ms":199.0384}
{"id":"needs_human","type":"binary","probabilities":{"false":2.8393421274182722e-08,"true":0.9999999716065777},"predicted":"true","confidence":0.9999999716065777,"entropy_confidence":0.9999992472176259,"candidate_mass":0.9999998965317776,"accepted":true,"accepted_predicted":"true","input_tokens":194,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.377108363888418,"brier_hard":1.9999998864263144,"soft_accuracy":0.2533330140074398,"brier_soft":1.1150231329762774,"tvd":0.7466669716065781,"kl":6.5292182320248955,"case_id":"typed:129/r0","base_id":"typed:129","repeat":0,"workflow":"customer_service","case_elapsed_ms":199.0384}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.435670543873788e-05,"1":0.00013994127503813764,"2":0.9995603707231532,"3":0.00022533129636855773},"predicted":"2","confidence":0.9995603707231532,"entropy_confidence":0.9969121908710561,"candidate_mass":0.999999312765853,"accepted":true,"accepted_predicted":"2","input_tokens":220,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0004397259421296477,"brier_hard":2.691605742869958e-07,"score_mae":0.1332696766104502,"within_one":true,"case_id":"typed:129/r0","base_id":"typed:129","repeat":0,"workflow":"customer_service","case_elapsed_ms":199.0384}
{"id":"action","type":"choice","probabilities":{"answer_directly":8.808231434724915e-05,"close_no_action":8.351818312465419e-05,"escalate_to_human":0.0008677699494113654,"execute_refund":0.0009028079496632581,"request_information":0.9980578216034524},"predicted":"request_information","confidence":0.9980578216034524,"entropy_confidence":0.9900629213025507,"candidate_mass":0.9999824519403916,"accepted":true,"accepted_predicted":"request_information","input_tokens":286,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0019440668705747815,"brier_hard":5.354877584106171e-06,"soft_accuracy":0.7154051874745607,"brier_soft":0.11253330091336732,"tvd":0.28147433978657754,"kl":1.4736495616655598,"case_id":"typed:130/r0","base_id":"typed:130","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.74850000000004}
{"id":"category","type":"choice","probabilities":{"account":3.463003038199088e-07,"billing":5.820120526537593e-07,"delivery":0.999980814853124,"refund":1.0798003909064014e-06,"technical":1.71770341290044e-05},"predicted":"delivery","confidence":0.999980814853124,"entropy_confidence":0.9998533694538568,"candidate_mass":0.999999529861935,"accepted":true,"accepted_predicted":"delivery","input_tokens":276,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":1.9185330913328186e-05,"brier_hard":6.647449929384245e-10,"soft_accuracy":0.8333179889161387,"brier_soft":0.04054733031671537,"tvd":0.16664816115342762,"kl":1.3214341156184695,"case_id":"typed:130/r0","base_id":"typed:130","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.74850000000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.172965929347431e-05,"1":0.9990674475255843,"2":0.0008859923762144277,"3":4.83043890671474e-06},"predicted":"1","confidence":0.9990674475255843,"entropy_confidence":0.9944892480421225,"candidate_mass":0.9999989366492952,"accepted":true,"accepted_predicted":"1","input_tokens":262,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0009329875719963892,"brier_hard":1.6564013058537219e-06,"score_mae":0.2175209235947334,"within_one":true,"case_id":"typed:130/r0","base_id":"typed:130","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.74850000000004}
{"id":"needs_human","type":"binary","probabilities":{"false":4.434088208156002e-08,"true":0.999999955659118},"predicted":"true","confidence":0.999999955659118,"entropy_confidence":0.9999988529243361,"candidate_mass":0.999999896124118,"accepted":true,"accepted_predicted":"true","input_tokens":230,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.931358739371095,"brier_hard":1.999999822636476,"soft_accuracy":0.2666670206923821,"brier_soft":1.0755544477114758,"tvd":0.7333329556591179,"kl":6.401778216579637,"case_id":"typed:130/r0","base_id":"typed:130","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.74850000000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010205234346440306,"1":0.11454603908058159,"2":0.8755074474685719,"3":0.008925990016202867},"predicted":"2","confidence":0.8755074474685719,"entropy_confidence":0.7015463071995791,"candidate_mass":0.9999965931653473,"accepted":true,"accepted_predicted":"2","input_tokens":256,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.166778448763562,"brier_hard":1.550622722246671,"score_mae":0.4423389040663339,"within_one":true,"case_id":"typed:130/r0","base_id":"typed:130","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.74850000000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.9121405173272116e-05,"close_no_action":0.00015511380714692812,"escalate_to_human":0.0006968179594997183,"execute_refund":0.00017853319501695612,"request_information":0.9989204136331626},"predicted":"request_information","confidence":0.9989204136331626,"entropy_confidence":0.9940768524080219,"candidate_mass":0.999979874321161,"accepted":true,"accepted_predicted":"request_information","input_tokens":372,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.001080169539960708,"brier_hard":1.7094092994792145e-06,"soft_accuracy":0.3831239363412214,"brier_soft":0.561038819549182,"tvd":0.6157659468281799,"kl":4.237601731771976,"case_id":"typed:131/r0","base_id":"typed:131","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.1951}
{"id":"category","type":"choice","probabilities":{"account":0.6280725492647751,"billing":0.003520917018750082,"delivery":0.0011408081059164083,"refund":0.0014410868330092864,"technical":0.3658246387775492},"predicted":"account","confidence":0.6280725492647751,"entropy_confidence":0.5669059388733892,"candidate_mass":0.9997479718404998,"accepted":false,"accepted_predicted":null,"input_tokens":362,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":null,"nll_hard":1.0056011894228476,"brier_hard":0.796669290952627,"soft_accuracy":0.3611627350229196,"brier_soft":0.747194967531683,"tvd":0.6182604529505117,"kl":0.9398800347654254,"case_id":"typed:131/r0","base_id":"typed:131","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.1951}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00012652956814979688,"1":0.993646089823033,"2":0.0050867395377559655,"3":0.001140641071061509},"predicted":"1","confidence":0.993646089823033,"entropy_confidence":0.9696585845202628,"candidate_mass":0.9999958366399699,"accepted":true,"accepted_predicted":"1","input_tokens":348,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.006374182180852737,"brier_hard":6.75641654465426e-05,"score_mae":0.5427585078882706,"within_one":true,"case_id":"typed:131/r0","base_id":"typed:131","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.1951}
{"id":"needs_human","type":"binary","probabilities":{"false":3.787400628160342e-07,"true":0.9999996212599356},"predicted":"true","confidence":0.9999996212599356,"entropy_confidence":0.9999913742013559,"candidate_mass":0.9999995993005285,"accepted":true,"accepted_predicted":"true","input_tokens":316,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.786415717256371,"brier_hard":1.9999984850400323,"soft_accuracy":0.43333305049892684,"brier_soft":0.6422221193003044,"tvd":0.5666666212599364,"kl":4.856821913283096,"case_id":"typed:131/r0","base_id":"typed:131","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.1951}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.005519956687746594,"1":0.9921378849230936,"2":0.002213296404372813,"3":0.00012886198478781724},"predicted":"1","confidence":0.9921378849230936,"entropy_confidence":0.9630553360141251,"candidate_mass":0.9999965748194133,"accepted":true,"accepted_predicted":"1","input_tokens":342,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.007893184458158205,"brier_hard":9.719806170184993e-05,"score_mae":0.5530489363137974,"within_one":true,"case_id":"typed:131/r0","base_id":"typed:131","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.1951}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0005352533991248336,"close_no_action":0.00014929338838407397,"escalate_to_human":0.0014742389490265278,"execute_refund":0.9902440064210674,"request_information":0.007597207842398697},"predicted":"execute_refund","confidence":0.9902440064210674,"entropy_confidence":0.9616381273087284,"candidate_mass":0.9999648359260991,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":416,"reused_prefix_tokens":0,"expected":"execute_refund","correct":true,"accepted_correct":true,"nll_hard":0.009803895090129766,"brier_hard":0.00015537914290869595,"soft_accuracy":0.3491312160817223,"brier_soft":0.5501586153775987,"tvd":0.6402440064210666,"kl":2.5568498533220825,"case_id":"typed:132/r0","base_id":"typed:132","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.5865}
{"id":"category","type":"choice","probabilities":{"account":5.6292971985772145e-09,"billing":0.9999995085446525,"delivery":7.051848035968916e-08,"refund":4.0529052532242457e-07,"technical":1.0017044428215323e-08},"predicted":"billing","confidence":0.9999995085446525,"entropy_confidence":0.9999950855646664,"candidate_mass":0.9999999741031672,"accepted":true,"accepted_predicted":"billing","input_tokens":406,"reused_prefix_tokens":0,"expected":"refund","correct":false,"accepted_correct":false,"nll_hard":14.718661680534753,"brier_hard":1.999998206508665,"soft_accuracy":0.030000378388149136,"brier_soft":1.8817982603134173,"tvd":0.9699995947094745,"kl":8.828833438640961,"case_id":"typed:132/r0","base_id":"typed:132","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.5865}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.300484195662599e-05,"1":7.153566877209076e-05,"2":0.7173092160829136,"3":0.28257624340635634},"predicted":"2","confidence":0.7173092160829136,"entropy_confidence":0.5696716582250854,"candidate_mass":0.9999938812993784,"accepted":false,"accepted_predicted":null,"input_tokens":392,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.3322482676360181,"brier_hard":0.1597634196160735,"score_mae":0.4490856980536684,"within_one":true,"case_id":"typed:132/r0","base_id":"typed:132","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.5865}
{"id":"needs_human","type":"binary","probabilities":{"false":4.860579308581031e-09,"true":0.9999999951394205},"predicted":"true","confidence":0.9999999951394205,"entropy_confidence":0.9999998587568163,"candidate_mass":0.9999999597732011,"accepted":true,"accepted_predicted":"true","input_tokens":360,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.860579506260464e-09,"brier_hard":4.7250464236852947e-17,"soft_accuracy":0.5666669993519194,"brier_soft":0.37555496935300214,"tvd":0.4333329951394206,"kl":3.6692869498592717,"case_id":"typed:132/r0","base_id":"typed:132","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.5865}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.997950879628701e-05,"1":1.9691528592203525e-05,"2":0.7908858522141564,"3":0.20905447674845432},"predicted":"2","confidence":0.7908858522141564,"entropy_confidence":0.6296853276921488,"candidate_mass":0.9999912641822752,"accepted":false,"accepted_predicted":null,"input_tokens":386,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.23460162982746266,"brier_hard":0.08743250303888708,"score_mae":0.142287826202268,"within_one":true,"case_id":"typed:132/r0","base_id":"typed:132","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.5865}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.1231683588911769e-05,"close_no_action":1.0391790578130524e-05,"escalate_to_human":8.878602434249219e-05,"execute_refund":1.019680308040235e-05,"request_information":0.9998793936984115},"predicted":"request_information","confidence":0.9998793936984115,"entropy_confidence":0.9991839693841228,"candidate_mass":0.9999971783803776,"accepted":true,"accepted_predicted":"request_information","input_tokens":365,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":11.396771881580484,"brier_hard":1.999736346796598,"soft_accuracy":0.3666387845935217,"brier_soft":0.6345477515103435,"tvd":0.6332127603650441,"kl":5.1643115677901905,"case_id":"typed:133/r0","base_id":"typed:133","repeat":0,"workflow":"customer_service","case_elapsed_ms":317.37579999999997}
{"id":"category","type":"choice","probabilities":{"account":1.4426063967487484e-05,"billing":1.0271607680447758e-05,"delivery":0.9991924589612805,"refund":1.5619908763034277e-05,"technical":0.0007672234583079974},"predicted":"delivery","confidence":0.9991924589612805,"entropy_confidence":0.995798564735735,"candidate_mass":0.9999958658588333,"accepted":true,"accepted_predicted":"delivery","input_tokens":355,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":0.0008078672756290558,"brier_hard":1.2413119629899444e-06,"soft_accuracy":0.8659766303505076,"brier_soft":0.02587795408950862,"tvd":0.1325254589612808,"kl":0.9175130287334943,"case_id":"typed:133/r0","base_id":"typed:133","repeat":0,"workflow":"customer_service","case_elapsed_ms":317.37579999999997}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":9.390938441445412e-05,"1":0.5227498095407842,"2":0.47705795269587825,"3":9.832837892377236e-05},"predicted":"1","confidence":0.5227498095407842,"entropy_confidence":0.4994282116689639,"candidate_mass":0.9999958840688316,"accepted":false,"accepted_predicted":null,"input_tokens":341,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.6486523050143802,"brier_hard":0.45535205301118314,"score_mae":0.4104937000693121,"within_one":true,"case_id":"typed:133/r0","base_id":"typed:133","repeat":0,"workflow":"customer_service","case_elapsed_ms":317.37579999999997}
{"id":"needs_human","type":"binary","probabilities":{"false":2.7719627486511217e-08,"true":0.9999999722803711},"predicted":"true","confidence":0.9999999722803711,"entropy_confidence":0.999999264121173,"candidate_mass":0.9999999333976697,"accepted":true,"accepted_predicted":"true","input_tokens":309,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.40112510115255,"brier_hard":1.9999998891214887,"soft_accuracy":0.3000000110878505,"brier_soft":0.9799999223850426,"tvd":0.6999999722803718,"kl":6.086046427401436,"case_id":"typed:133/r0","base_id":"typed:133","repeat":0,"workflow":"customer_service","case_elapsed_ms":317.37579999999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003279209803649637,"1":0.01118094807404521,"2":0.988049572886391,"3":0.00044155805919961384},"predicted":"2","confidence":0.988049572886391,"entropy_confidence":0.9508309579435978,"candidate_mass":0.9999977486304145,"accepted":true,"accepted_predicted":"2","input_tokens":335,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.012022407506766098,"brier_hard":0.00026812881372118534,"score_mae":0.011395231975574127,"within_one":true,"case_id":"typed:133/r0","base_id":"typed:133","repeat":0,"workflow":"customer_service","case_elapsed_ms":317.37579999999997}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.297413373417403e-05,"close_no_action":8.183150721406358e-05,"escalate_to_human":0.0006040593346430551,"execute_refund":5.243978240159538e-05,"request_information":0.9992186952420058},"predicted":"request_information","confidence":0.9992186952420058,"entropy_confidence":0.9956648014702234,"candidate_mass":0.999991711502718,"accepted":true,"accepted_predicted":"request_information","input_tokens":506,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":7.4118381286903166,"brier_hard":1.9972302584326327,"soft_accuracy":0.21687783522889056,"brier_soft":1.0074379855781372,"tvd":0.7825519119087898,"kl":5.184146953105208,"case_id":"typed:134/r0","base_id":"typed:134","repeat":0,"workflow":"customer_service","case_elapsed_ms":355.3363}
{"id":"category","type":"choice","probabilities":{"account":4.161514857862205e-10,"billing":2.8155237430711206e-10,"delivery":0.9999999978115355,"refund":7.023357051821631e-10,"technical":7.884250598950206e-10},"predicted":"delivery","confidence":0.9999999978115355,"entropy_confidence":0.9999999697423061,"candidate_mass":0.9999999975848652,"accepted":true,"accepted_predicted":"delivery","input_tokens":496,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":2.1884645231084393e-09,"brier_hard":6.1567202748671116e-18,"soft_accuracy":0.7333329985877871,"brier_soft":0.10464469400349678,"tvd":0.2666669978115354,"kl":2.228646660207549,"case_id":"typed:134/r0","base_id":"typed:134","repeat":0,"workflow":"customer_service","case_elapsed_ms":355.3363}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.779216696179876e-05,"1":0.0006120941495780887,"2":0.026104114440546507,"3":0.973215999242913},"predicted":"3","confidence":0.973215999242913,"entropy_confidence":0.908556039839836,"candidate_mass":0.9999988662545395,"accepted":true,"accepted_predicted":"3","input_tokens":482,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.64566233567734,"brier_hard":1.8956229563470397,"score_mae":0.6724683207594104,"within_one":true,"case_id":"typed:134/r0","base_id":"typed:134","repeat":0,"workflow":"customer_service","case_elapsed_ms":355.3363}
{"id":"needs_human","type":"binary","probabilities":{"false":2.775591060526789e-09,"true":0.9999999972244105},"predicted":"true","confidence":0.9999999972244105,"entropy_confidence":0.9999999171007221,"candidate_mass":0.9999999791442422,"accepted":true,"accepted_predicted":"true","input_tokens":450,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.775589539837949e-09,"brier_hard":1.5407803007551207e-17,"soft_accuracy":0.7999999983346466,"brier_soft":0.07999999777952774,"tvd":0.19999999722440964,"kl":1.66355323556434,"case_id":"typed:134/r0","base_id":"typed:134","repeat":0,"workflow":"customer_service","case_elapsed_ms":355.3363}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.1784437751359483e-05,"1":6.5057444619205635e-06,"2":0.9964836334271656,"3":0.003498076390620221},"predicted":"2","confidence":0.9964836334271656,"entropy_confidence":0.9830446324348917,"candidate_mass":0.9999990123608927,"accepted":true,"accepted_predicted":"2","input_tokens":476,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.65554206425362,"brier_hard":1.9859957156265975,"score_mae":0.6965319982293465,"within_one":true,"case_id":"typed:134/r0","base_id":"typed:134","repeat":0,"workflow":"customer_service","case_elapsed_ms":355.3363}
{"id":"action","type":"choice","probabilities":{"answer_directly":3.853902906594746e-05,"close_no_action":1.2656958137015736e-05,"escalate_to_human":0.9996474546843334,"execute_refund":6.629631159391953e-06,"request_information":0.0002947196973037506},"predicted":"escalate_to_human","confidence":0.9996474546843334,"entropy_confidence":0.9979111371395544,"candidate_mass":0.9999979506797099,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":367,"reused_prefix_tokens":0,"expected":"request_information","correct":false,"accepted_correct":false,"nll_hard":8.12948583196168,"brier_hard":1.998705682811366,"soft_accuracy":0.4066619678427055,"brier_soft":0.5697040857213806,"tvd":0.592980861350927,"kl":4.065936337508489,"case_id":"typed:135/r0","base_id":"typed:135","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.817}
{"id":"category","type":"choice","probabilities":{"account":1.954960964828039e-11,"billing":0.9999999993620463,"delivery":4.422952691414393e-10,"refund":1.590235511807073e-10,"technical":1.7086096022457467e-11},"predicted":"billing","confidence":0.9999999993620463,"entropy_confidence":0.9999999908923967,"candidate_mass":0.999999998387958,"accepted":true,"accepted_predicted":"billing","input_tokens":357,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":6.379536900783408e-10,"brier_hard":6.285726272746443e-19,"soft_accuracy":0.8466661528162389,"brier_soft":0.0362668205596301,"tvd":0.15333384602819924,"kl":1.2713303418572526,"case_id":"typed:135/r0","base_id":"typed:135","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.817}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.5369648540894807e-05,"1":9.830718243179984e-05,"2":0.9969183783429576,"3":0.002967944826070191},"predicted":"2","confidence":0.9969183783429576,"entropy_confidence":0.9845433747847765,"candidate_mass":0.99999869530072,"accepted":true,"accepted_predicted":"2","input_tokens":343,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0030863796304245205,"brier_hard":1.831498905596351e-05,"score_mae":0.4361718983465577,"within_one":true,"case_id":"typed:135/r0","base_id":"typed:135","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.817}
{"id":"needs_human","type":"binary","probabilities":{"false":1.277208424296384e-09,"true":0.9999999987227923},"predicted":"true","confidence":0.9999999987227923,"entropy_confidence":0.9999999604230748,"candidate_mass":0.9999999788959146,"accepted":true,"accepted_predicted":"true","input_tokens":311,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.2772076737717673e-09,"brier_hard":3.2625207989516842e-18,"soft_accuracy":0.7433329993784267,"brier_soft":0.13175589646673136,"tvd":0.2566669987227919,"kl":2.1435095760312897,"case_id":"typed:135/r0","base_id":"typed:135","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.817}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.74641529836153e-05,"1":0.0002032199711245353,"2":0.9335766875103285,"3":0.06615262836556235},"predicted":"2","confidence":0.9335766875103285,"entropy_confidence":0.8224049773724835,"candidate_mass":0.9999947779974889,"accepted":true,"accepted_predicted":"2","input_tokens":337,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.06873216885462692,"brier_hard":0.008788272531541354,"score_mae":0.1341855199115316,"within_one":true,"case_id":"typed:135/r0","base_id":"typed:135","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.817}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0033880341261169613,"close_no_action":0.0001741539828059078,"escalate_to_human":0.2869062666351748,"execute_refund":0.7095054934493861,"request_information":2.6051806516537574e-05},"predicted":"execute_refund","confidence":0.7095054934493861,"entropy_confidence":0.6130474994065795,"candidate_mass":0.9999574168845861,"accepted":false,"accepted_predicted":null,"input_tokens":342,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":null,"nll_hard":5.687505429599327,"brier_hard":1.578948692600703,"soft_accuracy":0.02178998347470051,"brier_soft":1.1749221494595359,"tvd":0.9297447600845608,"kl":5.3225881821817955,"case_id":"typed:136/r0","base_id":"typed:136","repeat":0,"workflow":"customer_service","case_elapsed_ms":332.3391}
{"id":"category","type":"choice","probabilities":{"account":0.9999995992359366,"billing":3.481340813486002e-07,"delivery":4.719858649565378e-08,"refund":3.0884478347944874e-09,"technical":2.3429473788676005e-09},"predicted":"account","confidence":0.9999995992359366,"entropy_confidence":0.9999959731153191,"candidate_mass":0.9999998524572897,"accepted":true,"accepted_predicted":"account","input_tokens":332,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":4.0076414368557146e-07,"brier_hard":2.8405190757262416e-13,"soft_accuracy":0.8999996686639875,"brier_soft":0.017044256700182163,"tvd":0.09999959923593676,"kl":0.8262099337933042,"case_id":"typed:136/r0","base_id":"typed:136","repeat":0,"workflow":"customer_service","case_elapsed_ms":332.3391}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":7.587040514957253e-07,"1":1.6548909042182637e-07,"2":1.2657869426645603e-05,"3":0.9999864179374299},"predicted":"3","confidence":0.9999864179374299,"entropy_confidence":0.9998776572280803,"candidate_mass":0.9999998045890037,"accepted":true,"accepted_predicted":"3","input_tokens":318,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.3582154807119226e-05,"brier_hard":3.452971005561709e-10,"score_mae":0.13331773504023348,"within_one":true,"case_id":"typed:136/r0","base_id":"typed:136","repeat":0,"workflow":"customer_service","case_elapsed_ms":332.3391}
{"id":"needs_human","type":"binary","probabilities":{"false":1.5952331199150372e-08,"true":0.9999999840476691},"predicted":"true","confidence":0.9999999840476691,"entropy_confidence":0.9999995637938326,"candidate_mass":0.9999999453404357,"accepted":true,"accepted_predicted":"true","input_tokens":286,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.95366086170886,"brier_hard":1.9999999361906762,"soft_accuracy":0.14333301137934026,"brier_soft":1.467756643114658,"tvd":0.856666984047669,"kl":7.61175816713608,"case_id":"typed:136/r0","base_id":"typed:136","repeat":0,"workflow":"customer_service","case_elapsed_ms":332.3391}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.018106035657171622,"1":0.9702292142875845,"2":0.01099135850206191,"3":0.000673391553182507},"predicted":"1","confidence":0.9702292142875845,"entropy_confidence":0.8871439239389135,"candidate_mass":0.9999947485708638,"accepted":true,"accepted_predicted":"1","input_tokens":312,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.510645905643507,"brier_hard":1.919811103198103,"score_mae":1.3391008940487443,"within_one":false,"case_id":"typed:136/r0","base_id":"typed:136","repeat":0,"workflow":"customer_service","case_elapsed_ms":332.3391}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.005698520836785979,"close_no_action":0.0010070004373290825,"escalate_to_human":0.3423767268107027,"execute_refund":0.0004904191086132562,"request_information":0.6504273328065685},"predicted":"request_information","confidence":0.6504273328065685,"entropy_confidence":0.5732215587286659,"candidate_mass":0.9997155329804629,"accepted":false,"accepted_predicted":null,"input_tokens":429,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":null,"nll_hard":1.0718436077878017,"brier_hard":0.8555578124025819,"soft_accuracy":0.25979226711071757,"brier_soft":0.38961590958055226,"tvd":0.5504273328065687,"kl":1.5919681766139466,"case_id":"typed:137/r0","base_id":"typed:137","repeat":0,"workflow":"customer_service","case_elapsed_ms":330.7426}
{"id":"category","type":"choice","probabilities":{"account":1.249956157865412e-09,"billing":7.221542535163215e-10,"delivery":0.9999999963072366,"refund":1.1672692961225287e-09,"technical":5.533828862197384e-10},"predicted":"delivery","confidence":0.9999999963072366,"entropy_confidence":0.9999999500931583,"candidate_mass":0.999999996734056,"accepted":true,"accepted_predicted":"delivery","input_tokens":419,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":3.692763435193907e-09,"brier_hard":1.7389149128837378e-17,"soft_accuracy":0.9999999963072366,"brier_soft":1.7389149128837378e-17,"tvd":3.6927630110498282e-09,"kl":3.6927634215574052e-09,"case_id":"typed:137/r0","base_id":"typed:137","repeat":0,"workflow":"customer_service","case_elapsed_ms":330.7426}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00010813466155052875,"1":0.9985183776888321,"2":0.001273687614194146,"3":9.980003542321148e-05},"predicted":"1","confidence":0.9985183776888321,"entropy_confidence":0.9914321134032521,"candidate_mass":0.9999965415789316,"accepted":true,"accepted_predicted":"1","input_tokens":405,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.665838952688318,"brier_hard":1.9944932192872395,"score_mae":0.8486348469765101,"within_one":true,"case_id":"typed:137/r0","base_id":"typed:137","repeat":0,"workflow":"customer_service","case_elapsed_ms":330.7426}
{"id":"needs_human","type":"binary","probabilities":{"false":2.2295924488233185e-06,"true":0.999997770407552},"predicted":"true","confidence":0.999997770407552,"entropy_confidence":0.9999549232548434,"candidate_mass":0.9999988662149054,"accepted":true,"accepted_predicted":"true","input_tokens":373,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.229594933544743e-06,"brier_hard":9.942164972027732e-12,"soft_accuracy":0.733331959525011,"brier_soft":0.1422201995530244,"tvd":0.26666477040755154,"kl":2.2286482936382175,"case_id":"typed:137/r0","base_id":"typed:137","repeat":0,"workflow":"customer_service","case_elapsed_ms":330.7426}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00036877412372073227,"1":0.02218998108560146,"2":0.9768859286457894,"3":0.0005553161448891003},"predicted":"2","confidence":0.9768859286457894,"entropy_confidence":0.9174599831803291,"candidate_mass":0.9999981883885806,"accepted":true,"accepted_predicted":"2","input_tokens":399,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":3.8081143944218887,"brier_hard":1.9104189950458978,"score_mae":0.4442947868118474,"within_one":true,"case_id":"typed:137/r0","base_id":"typed:137","repeat":0,"workflow":"customer_service","case_elapsed_ms":330.7426}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00016366795839792714,"close_no_action":0.00022693636364751893,"escalate_to_human":0.0006805504194464413,"execute_refund":0.0001910984843026965,"request_information":0.9987377467742051},"predicted":"request_information","confidence":0.9987377467742051,"entropy_confidence":0.9930461700485848,"candidate_mass":0.9999588167354253,"accepted":true,"accepted_predicted":"request_information","input_tokens":404,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0012630505384087901,"brier_hard":2.171238023892784e-06,"soft_accuracy":0.665950025343985,"brier_soft":0.16168896965446425,"tvd":0.33226184525850794,"kl":1.7930554639353988,"case_id":"typed:138/r0","base_id":"typed:138","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.1424}
{"id":"category","type":"choice","probabilities":{"account":0.9999999999476508,"billing":4.086497580588792e-11,"delivery":9.240510590072121e-12,"refund":1.1737352116081905e-12,"technical":1.0703524769450055e-12},"predicted":"account","confidence":0.9999999999476508,"entropy_confidence":0.9999999991758656,"candidate_mass":0.9999999992174295,"accepted":true,"accepted_predicted":"account","input_tokens":394,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":5.23492360584956e-11,"brier_hard":4.498299108117547e-21,"soft_accuracy":0.916330916283197,"brier_soft":0.012461122040972078,"tvd":0.08366908361673425,"kl":0.6905538356033767,"case_id":"typed:138/r0","base_id":"typed:138","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.1424}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00028121365101026164,"1":0.9996019596585427,"2":0.00010620014660051182,"3":1.0626543845421045e-05},"predicted":"1","confidence":0.9996019596585427,"entropy_confidence":0.9972655713114632,"candidate_mass":0.9999980831353814,"accepted":true,"accepted_predicted":"1","input_tokens":380,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00039811958054160806,"brier_hard":2.4890862551403027e-07,"score_mae":0.16651323958328001,"within_one":true,"case_id":"typed:138/r0","base_id":"typed:138","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.1424}
{"id":"needs_human","type":"binary","probabilities":{"false":0.0001501556378338693,"true":0.9998498443621664},"predicted":"true","confidence":0.9998498443621664,"entropy_confidence":0.9978762226148473,"candidate_mass":0.9999866680290989,"accepted":true,"accepted_predicted":"true","input_tokens":348,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.803838216228726,"brier_hard":1.9993994225420963,"soft_accuracy":0.2167520880946688,"brier_soft":1.2267507354064264,"tvd":0.7831828443621663,"kl":6.373712529149179,"case_id":"typed:138/r0","base_id":"typed:138","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.1424}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001806715033181423,"1":0.9997402870725828,"2":6.746946362806988e-05,"3":1.1571960470610793e-05},"predicted":"1","confidence":0.9997402870725828,"entropy_confidence":0.9981271225659862,"candidate_mass":0.9999983507265064,"accepted":true,"accepted_predicted":"1","input_tokens":374,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0002597466586599751,"brier_hard":1.0477903557025336e-07,"score_mae":0.3834230581187492,"within_one":true,"case_id":"typed:138/r0","base_id":"typed:138","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.1424}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0002973726866279579,"close_no_action":0.0003087089219319201,"escalate_to_human":0.0007655834304529003,"execute_refund":0.0001463470467323001,"request_information":0.9984819879142552},"predicted":"request_information","confidence":0.9984819879142552,"entropy_confidence":0.9917908301594619,"candidate_mass":0.9999788169058451,"accepted":true,"accepted_predicted":"request_information","input_tokens":279,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0015191654334359187,"brier_hard":3.095627852771144e-06,"soft_accuracy":0.7090253910278456,"brier_soft":0.13213864717867194,"tvd":0.2884819879142551,"kl":1.4878113572948004,"case_id":"typed:139/r0","base_id":"typed:139","repeat":0,"workflow":"customer_service","case_elapsed_ms":237.7626}
{"id":"category","type":"choice","probabilities":{"account":5.904393140793211e-07,"billing":9.897880861075855e-06,"delivery":5.1123246067464205e-06,"refund":3.8061489287997756e-06,"technical":0.9999805932062875},"predicted":"technical","confidence":0.9999805932062875,"entropy_confidence":0.9998436010281946,"candidate_mass":0.9999990713828821,"accepted":true,"accepted_predicted":"technical","input_tokens":269,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":1.9406982026741704e-05,"brier_hard":5.155629388755277e-10,"soft_accuracy":0.9799820680684637,"brier_soft":0.000515628494405456,"tvd":0.0199796132053084,"kl":0.12435270989247328,"case_id":"typed:139/r0","base_id":"typed:139","repeat":0,"workflow":"customer_service","case_elapsed_ms":237.7626}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.0581593415078782e-05,"1":0.9999708848359229,"2":7.9777894865284e-06,"3":5.557811765172111e-07},"predicted":"1","confidence":0.9999708848359229,"entropy_confidence":0.9997454595631363,"candidate_mass":0.9999996571954509,"accepted":true,"accepted_predicted":"1","input_tokens":255,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":2.9115587931725936e-05,"brier_hard":1.3352487845481402e-09,"score_mae":0.26667849224157447,"within_one":true,"case_id":"typed:139/r0","base_id":"typed:139","repeat":0,"workflow":"customer_service","case_elapsed_ms":237.7626}
{"id":"needs_human","type":"binary","probabilities":{"false":3.1271571712025227e-06,"true":0.9999968728428296},"predicted":"true","confidence":0.9999968728428296,"entropy_confidence":0.9999383030352138,"candidate_mass":0.9999982491565456,"accepted":true,"accepted_predicted":"true","input_tokens":223,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.67538621813496,"brier_hard":1.9999874913908748,"soft_accuracy":0.28333435510352584,"brier_soft":1.0272142132761657,"tvd":0.7166638728428292,"kl":6.243427464621531,"case_id":"typed:139/r0","base_id":"typed:139","repeat":0,"workflow":"customer_service","case_elapsed_ms":237.7626}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.0339348406732117e-05,"1":0.9998442303518114,"2":0.00011938639792340495,"3":6.0439018582319985e-06},"predicted":"1","confidence":0.9998442303518114,"entropy_confidence":0.9988296564112552,"candidate_mass":0.9999989910064758,"accepted":true,"accepted_predicted":"1","input_tokens":249,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00015578178154031173,"brier_hard":3.9474300117355816e-08,"score_mae":0.216565865146767,"within_one":true,"case_id":"typed:139/r0","base_id":"typed:139","repeat":0,"workflow":"customer_service","case_elapsed_ms":237.7626}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0006052582032076737,"close_no_action":0.0003138595139930625,"escalate_to_human":0.9782366217028725,"execute_refund":0.0007937113907487343,"request_information":0.020050549189177657},"predicted":"escalate_to_human","confidence":0.9782366217028725,"entropy_confidence":0.9300406693202364,"candidate_mass":0.9999623701920926,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":502,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.02200369372970098,"brier_hard":0.0008767639807503911,"soft_accuracy":0.46751685383134906,"brier_soft":0.4789160531908338,"tvd":0.5026776692741378,"kl":2.875580917409258,"case_id":"typed:140/r0","base_id":"typed:140","repeat":0,"workflow":"customer_service","case_elapsed_ms":364.7703}
{"id":"category","type":"choice","probabilities":{"account":4.266225320856793e-08,"billing":0.9999996069135895,"delivery":2.602998848737469e-07,"refund":5.446378461842014e-08,"technical":3.5660488398464746e-08},"predicted":"billing","confidence":0.9999996069135895,"entropy_confidence":0.9999959078398073,"candidate_mass":0.9999998517769221,"accepted":true,"accepted_predicted":"billing","input_tokens":492,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.9308648779130024e-07,"brier_hard":2.283309983274708e-13,"soft_accuracy":0.9433326316385457,"brier_soft":0.00572776610631559,"tvd":0.0566666069135892,"kl":0.46614964095379147,"case_id":"typed:140/r0","base_id":"typed:140","repeat":0,"workflow":"customer_service","case_elapsed_ms":364.7703}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.132348402299012e-05,"1":6.884630480867645e-05,"2":0.2014535749732142,"3":0.7984362552379523},"predicted":"3","confidence":0.7984362552379523,"entropy_confidence":0.6367487107152476,"candidate_mass":0.9999971388464107,"accepted":false,"accepted_predicted":null,"input_tokens":478,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.6021963212961658,"brier_hard":1.2751768530489085,"score_mae":0.631617761965094,"within_one":true,"case_id":"typed:140/r0","base_id":"typed:140","repeat":0,"workflow":"customer_service","case_elapsed_ms":364.7703}
{"id":"needs_human","type":"binary","probabilities":{"false":2.7843451815641704e-10,"true":0.999999999721565},"predicted":"true","confidence":0.999999999721565,"entropy_confidence":0.9999999907602506,"candidate_mass":0.9999999928428274,"accepted":true,"accepted_predicted":"true","input_tokens":446,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.784350528305553e-10,"brier_hard":1.5505185952416426e-19,"soft_accuracy":0.6999999998886258,"brier_soft":0.17999999966587832,"tvd":0.29999999972156527,"kl":2.513429651030647,"case_id":"typed:140/r0","base_id":"typed:140","repeat":0,"workflow":"customer_service","case_elapsed_ms":364.7703}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.5144394257658225e-08,"1":3.12045700423505e-08,"2":4.1373821154278344e-07,"3":0.9999994999128254},"predicted":"3","confidence":0.9999994999128254,"entropy_confidence":0.9999941988010816,"candidate_mass":0.9999998778808393,"accepted":true,"accepted_predicted":"3","input_tokens":472,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.000872996033708e-07,"brier_hard":4.2528111925933487e-13,"score_mae":0.133332358419469,"within_one":true,"case_id":"typed:140/r0","base_id":"typed:140","repeat":0,"workflow":"customer_service","case_elapsed_ms":364.7703}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0008059345303998443,"close_no_action":0.0002548106914538749,"escalate_to_human":0.9900367914119237,"execute_refund":0.00015028369679213888,"request_information":0.008752179669430414},"predicted":"escalate_to_human","confidence":0.9900367914119237,"entropy_confidence":0.9623733357060655,"candidate_mass":0.9999565691260254,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":275,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.01001317350168303,"brier_hard":0.00017660321848080016,"soft_accuracy":0.7110264442613559,"brier_soft":0.11261991529761657,"tvd":0.2737748858001697,"kl":0.8538061993618307,"case_id":"typed:141/r0","base_id":"typed:141","repeat":0,"workflow":"customer_service","case_elapsed_ms":245.83089999999999}
{"id":"category","type":"choice","probabilities":{"account":1.7319846298495073e-06,"billing":1.8324527909107737e-05,"delivery":1.0756142687587374e-05,"refund":5.0674523049187825e-06,"technical":0.9999641198924687},"predicted":"technical","confidence":0.9999641198924687,"entropy_confidence":0.9997243985726045,"candidate_mass":0.9999985650887196,"accepted":true,"accepted_predicted":"technical","input_tokens":265,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":3.588075123776111e-05,"brier_hard":1.7675438886861024e-09,"soft_accuracy":0.9166348406548667,"brier_soft":0.009748038020748001,"tvd":0.0833078760351562,"kl":0.6147609905252864,"case_id":"typed:141/r0","base_id":"typed:141","repeat":0,"workflow":"customer_service","case_elapsed_ms":245.83089999999999}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.172155806212213e-08,"1":7.520575010918127e-09,"2":7.738315201869031e-07,"3":0.9999991969263453},"predicted":"3","confidence":0.9999991969263453,"entropy_confidence":0.9999911877913186,"candidate_mass":0.9999999656391628,"accepted":true,"accepted_predicted":"3","input_tokens":251,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":8.030739771509454e-07,"brier_hard":1.2442709016204507e-12,"score_mae":0.48333214596265117,"within_one":true,"case_id":"typed:141/r0","base_id":"typed:141","repeat":0,"workflow":"customer_service","case_elapsed_ms":245.83089999999999}
{"id":"needs_human","type":"binary","probabilities":{"false":1.6049035420847426e-09,"true":0.9999999983950971},"predicted":"true","confidence":0.9999999983950971,"entropy_confidence":0.9999999507975722,"candidate_mass":0.9999999763960975,"accepted":true,"accepted_predicted":"true","input_tokens":219,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6049028785491948e-09,"brier_hard":5.1514286248378755e-18,"soft_accuracy":0.8766669987909721,"brier_soft":0.030422056986249908,"tvd":0.12333299839509682,"kl":1.0205449320550575,"case_id":"typed:141/r0","base_id":"typed:141","repeat":0,"workflow":"customer_service","case_elapsed_ms":245.83089999999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.733869361580702e-05,"1":1.5222456126939393e-05,"2":0.002114450987054359,"3":0.9978529878632018},"predicted":"3","confidence":0.9978529878632018,"entropy_confidence":0.9888000263983991,"candidate_mass":0.9999993758827614,"accepted":true,"accepted_predicted":"3","input_tokens":245,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0021493202716768613,"brier_hard":9.081096445680952e-06,"score_mae":0.5878030880198408,"within_one":true,"case_id":"typed:141/r0","base_id":"typed:141","repeat":0,"workflow":"customer_service","case_elapsed_ms":245.83089999999999}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00010241894710916332,"close_no_action":0.0001628830361234153,"escalate_to_human":0.009543938246040091,"execute_refund":8.760156277397098e-05,"request_information":0.9901031582079548},"predicted":"request_information","confidence":0.9901031582079548,"entropy_confidence":0.9643200208495549,"candidate_mass":0.9999762165398559,"accepted":true,"accepted_predicted":"request_information","input_tokens":362,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":4.651849064655895,"brier_hard":1.9613075188530884,"soft_accuracy":0.13855216052229363,"brier_soft":1.2060684298565814,"tvd":0.8569330412440775,"kl":3.9224701350645503,"case_id":"typed:142/r0","base_id":"typed:142","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.5091}
{"id":"category","type":"choice","probabilities":{"account":4.998672081955998e-08,"billing":0.9999982978002758,"delivery":1.0529739248664933e-06,"refund":3.916541152099446e-07,"technical":2.0758496352923335e-07},"predicted":"billing","confidence":0.9999982978002758,"entropy_confidence":0.9999838404085738,"candidate_mass":0.9999998565714542,"accepted":true,"accepted_predicted":"billing","input_tokens":352,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.7022011729175529e-06,"brier_hard":4.2052211227292575e-12,"soft_accuracy":0.809998693354609,"brier_soft":0.06975464667353927,"tvd":0.18999955835916407,"kl":1.5792820140927464,"case_id":"typed:142/r0","base_id":"typed:142","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.5091}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.807890330288425e-07,"1":1.861710415007727e-07,"2":0.00027401033010546953,"3":0.9997255227098197},"predicted":"3","confidence":0.9997255227098197,"entropy_confidence":0.9981756488460521,"candidate_mass":0.9999996900809426,"accepted":true,"accepted_predicted":"3","input_tokens":338,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00027451496596593236,"brier_hard":1.5041955733134693e-07,"score_mae":0.6330577749607116,"within_one":true,"case_id":"typed:142/r0","base_id":"typed:142","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.5091}
{"id":"needs_human","type":"binary","probabilities":{"false":6.600251756047e-09,"true":0.9999999933997472},"predicted":"true","confidence":0.9999999933997472,"entropy_confidence":0.9999998111171168,"candidate_mass":0.9999999575994076,"accepted":true,"accepted_predicted":"true","input_tokens":306,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.6002528105393116e-09,"brier_hard":8.712666011870455e-17,"soft_accuracy":0.8166669958198354,"brier_soft":0.06722197293782387,"tvd":0.18333299339974768,"kl":1.5231647869399942,"case_id":"typed:142/r0","base_id":"typed:142","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.5091}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00013011240497195638,"1":0.0001655567487761978,"2":0.006059110130428974,"3":0.9936452207158236},"predicted":"3","confidence":0.9936452207158236,"entropy_confidence":0.9712333632696608,"candidate_mass":0.9999964610854327,"accepted":true,"accepted_predicted":"3","input_tokens":332,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00637505684598594,"brier_hard":7.714037359825724e-05,"score_mae":0.5765524391571049,"within_one":true,"case_id":"typed:142/r0","base_id":"typed:142","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.5091}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0017620276007234556,"close_no_action":0.00019214374344872807,"escalate_to_human":0.0022840723617389156,"execute_refund":7.124787360666034e-05,"request_information":0.9956905084204809},"predicted":"request_information","confidence":0.9956905084204809,"entropy_confidence":0.9803101587968337,"candidate_mass":0.9999849502271791,"accepted":true,"accepted_predicted":"request_information","input_tokens":477,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.004318804203101904,"brier_hard":2.6935440970956818e-05,"soft_accuracy":0.3658472076507631,"brier_soft":0.5336022336477391,"tvd":0.6290238750871149,"kl":3.470797418042662,"case_id":"typed:143/r0","base_id":"typed:143","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.36830000000003}
{"id":"category","type":"choice","probabilities":{"account":1.352973617444139e-06,"billing":0.9999883759671148,"delivery":8.314235291589091e-06,"refund":1.0962358246176868e-06,"technical":8.605881511134855e-07},"predicted":"billing","confidence":0.9999883759671148,"entropy_confidence":0.9999041739403252,"candidate_mass":0.9999992471271489,"accepted":true,"accepted_predicted":"billing","input_tokens":467,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.1624100444831747e-05,"brier_hard":2.0801753155951076e-10,"soft_accuracy":0.7666576290681084,"brier_soft":0.10166089067360193,"tvd":0.23333155310446452,"kl":1.9453862444864252,"case_id":"typed:143/r0","base_id":"typed:143","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.36830000000003}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.78379499409737e-05,"1":0.0007495043461324153,"2":0.9643593958979148,"3":0.03482326180601271},"predicted":"2","confidence":0.9643593958979148,"entropy_confidence":0.8860555976688079,"candidate_mass":0.9999984506407061,"accepted":true,"accepted_predicted":"2","input_tokens":453,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.03629123652921257,"brier_hard":0.0024834785823240027,"score_mae":0.06606191843999998,"within_one":true,"case_id":"typed:143/r0","base_id":"typed:143","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.36830000000003}
{"id":"needs_human","type":"binary","probabilities":{"false":2.6664503328720104e-09,"true":0.9999999973335498},"predicted":"true","confidence":0.9999999973335498,"entropy_confidence":0.9999999202061343,"candidate_mass":0.9999999854478078,"accepted":true,"accepted_predicted":"true","input_tokens":421,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.6664501753282274e-09,"brier_hard":1.4219913896222848e-17,"soft_accuracy":0.7333329987556584,"brier_soft":0.14222257493378285,"tvd":0.2666669973335497,"kl":2.2286466605580717,"case_id":"typed:143/r0","base_id":"typed:143","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.36830000000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005727075786404601,"1":0.9243940593628884,"2":0.074495607728908,"3":0.0005376253295624347},"predicted":"1","confidence":0.9243940593628884,"entropy_confidence":0.8020175763835634,"candidate_mass":0.9999935491393459,"accepted":true,"accepted_predicted":"1","input_tokens":447,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.5970151119935885,"brier_hard":1.7110633741334482,"score_mae":0.9083348491906083,"within_one":true,"case_id":"typed:143/r0","base_id":"typed:143","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.36830000000003}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00013471141515472096,"close_no_action":4.390506243057552e-05,"escalate_to_human":0.9962578426697675,"execute_refund":3.241483666179498e-05,"request_information":0.0035311260159845274},"predicted":"escalate_to_human","confidence":0.9962578426697675,"entropy_confidence":0.9840636528679628,"candidate_mass":0.9999890114197499,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":592,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.003749176718216011,"brier_hard":2.649371796649166e-05,"soft_accuracy":0.5657935629441029,"brier_soft":0.31151093094729576,"tvd":0.4296671625688603,"kl":1.8245977276036764,"case_id":"typed:144/r0","base_id":"typed:144","repeat":0,"workflow":"customer_service","case_elapsed_ms":502.2607}
{"id":"category","type":"choice","probabilities":{"account":7.512738918775474e-08,"billing":5.459963064266878e-08,"delivery":0.9999995107469892,"refund":1.6908427478849876e-07,"technical":1.904417159013555e-07},"predicted":"delivery","confidence":0.9999995107469892,"entropy_confidence":0.9999948937949582,"candidate_mass":0.9999998691692419,"accepted":true,"accepted_predicted":"delivery","input_tokens":582,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":false,"nll_hard":15.473919642466631,"brier_hard":1.9999986406108594,"soft_accuracy":0.08333313344475077,"brier_soft":1.6505549323827897,"tvd":0.9166666404740091,"kl":8.235740059349773,"case_id":"typed:144/r0","base_id":"typed:144","repeat":0,"workflow":"customer_service","case_elapsed_ms":502.2607}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":9.515713271892058e-05,"1":0.0003771024914449544,"2":0.0045351516304734584,"3":0.9949925887453616},"predicted":"3","confidence":0.9949925887453616,"entropy_confidence":0.9759647634259105,"candidate_mass":0.9999988446242507,"accepted":true,"accepted_predicted":"3","input_tokens":568,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.39589676036794,"brier_hard":1.9809606672587297,"score_mae":1.4110921719884764,"within_one":false,"case_id":"typed:144/r0","base_id":"typed:144","repeat":0,"workflow":"customer_service","case_elapsed_ms":502.2607}
{"id":"needs_human","type":"binary","probabilities":{"false":5.789445849700117e-08,"true":0.9999999421055409},"predicted":"true","confidence":0.9999999421055409,"entropy_confidence":0.9999985245774115,"candidate_mass":0.9999999315270595,"accepted":true,"accepted_predicted":"true","input_tokens":536,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.789446079152144e-08,"brier_hard":6.703536720953176e-15,"soft_accuracy":0.7999999652633245,"brier_soft":0.07999995368443963,"tvd":0.19999994210554115,"kl":1.663553279659437,"case_id":"typed:144/r0","base_id":"typed:144","repeat":0,"workflow":"customer_service","case_elapsed_ms":502.2607}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.338208386880316e-06,"1":2.8473423334851804e-06,"2":0.9835096847697288,"3":0.01648012967955114},"predicted":"2","confidence":0.9835096847697288,"entropy_confidence":0.9393076025587824,"candidate_mass":0.9999990584440922,"accepted":true,"accepted_predicted":"2","input_tokens":562,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0166277939521642,"brier_hard":0.0005435252326051982,"score_mae":0.21687039407955577,"within_one":true,"case_id":"typed:144/r0","base_id":"typed:144","repeat":0,"workflow":"customer_service","case_elapsed_ms":502.2607}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.002188234721736342,"close_no_action":0.000444549890237543,"escalate_to_human":0.0031536674387695734,"execute_refund":0.2941580220910453,"request_information":0.7000555258582107},"predicted":"request_information","confidence":0.7000555258582107,"entropy_confidence":0.5995025891410029,"candidate_mass":0.9998757391300129,"accepted":false,"accepted_predicted":null,"input_tokens":337,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":null,"nll_hard":6.124660123331388,"brier_hard":1.5722451434157758,"soft_accuracy":0.05200492271982788,"brier_soft":1.2819901429755929,"tvd":0.9208805479492563,"kl":5.29095430535253,"case_id":"typed:145/r0","base_id":"typed:145","repeat":0,"workflow":"customer_service","case_elapsed_ms":321.9001}
{"id":"category","type":"choice","probabilities":{"account":5.236786141002226e-10,"billing":0.9999999850642977,"delivery":1.0276310666747417e-08,"refund":3.433344939675489e-09,"technical":7.023671775120612e-10},"predicted":"billing","confidence":0.9999999850642977,"entropy_confidence":0.9999998155493992,"candidate_mass":0.999999993747636,"accepted":true,"accepted_predicted":"billing","input_tokens":327,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.4935702453105095e-08,"brier_hard":3.4123318177311885e-16,"soft_accuracy":0.9733329855943715,"brier_soft":0.0009055767178526997,"tvd":0.026666985064298127,"kl":0.21930392912227403,"case_id":"typed:145/r0","base_id":"typed:145","repeat":0,"workflow":"customer_service","case_elapsed_ms":321.9001}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.023657931984046792,"1":0.9762276968522053,"2":0.00010662091915034567,"3":7.750244597185712e-06},"predicted":"1","confidence":0.9762276968522053,"entropy_confidence":0.9183935939852766,"candidate_mass":0.9999926163666998,"accepted":true,"accepted_predicted":"1","input_tokens":313,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":3.744056829829045,"brier_hard":1.9062643613071162,"score_mae":0.8597971894242976,"within_one":true,"case_id":"typed:145/r0","base_id":"typed:145","repeat":0,"workflow":"customer_service","case_elapsed_ms":321.9001}
{"id":"needs_human","type":"binary","probabilities":{"false":7.093128534455158e-08,"true":0.9999999290687133},"predicted":"true","confidence":0.9999999290687133,"entropy_confidence":0.9999982131205031,"candidate_mass":0.9999998483152845,"accepted":true,"accepted_predicted":"true","input_tokens":281,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.461554240586135,"brier_hard":1.999999716274866,"soft_accuracy":0.08000005958227958,"brier_soft":1.6927997389728777,"tvd":0.919999929068714,"kl":8.271454856347932,"case_id":"typed:145/r0","base_id":"typed:145","repeat":0,"workflow":"customer_service","case_elapsed_ms":321.9001}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.827552789653448e-06,"1":8.431150831051055e-06,"2":0.00039071245034291615,"3":0.9995950288460365},"predicted":"3","confidence":0.9995950288460365,"entropy_confidence":0.9973744643950181,"candidate_mass":0.9999989369238345,"accepted":true,"accepted_predicted":"3","input_tokens":307,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0004050531769266818,"brier_hard":3.1676289907137933e-07,"score_mae":0.18624194258962623,"within_one":true,"case_id":"typed:145/r0","base_id":"typed:145","repeat":0,"workflow":"customer_service","case_elapsed_ms":321.9001}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.456149525908192e-05,"close_no_action":0.00014946765241846457,"escalate_to_human":0.0006868532257236017,"execute_refund":7.032055002499768e-05,"request_information":0.9990487970765752},"predicted":"request_information","confidence":0.9990487970765752,"entropy_confidence":0.9947877282357795,"candidate_mass":0.9999828768508473,"accepted":true,"accepted_predicted":"request_information","input_tokens":459,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":7.283389933802251,"brier_hard":1.9967252935273443,"soft_accuracy":0.08368768421366539,"brier_soft":1.2879458093294607,"tvd":0.915865264728993,"kl":6.4628255069019875,"case_id":"typed:146/r0","base_id":"typed:146","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.795703}
{"id":"category","type":"choice","probabilities":{"account":7.739832945900509e-13,"billing":0.9999999999654587,"delivery":2.4141213328167283e-11,"refund":8.781909109918574e-12,"technical":8.443679255067267e-13},"predicted":"billing","confidence":0.9999999999654587,"entropy_confidence":0.9999999994449277,"candidate_mass":0.9999999998568203,"accepted":true,"accepted_predicted":"billing","input_tokens":449,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.454125874273442e-11,"brier_hard":1.8543306713961975e-21,"soft_accuracy":0.9166669999690691,"brier_soft":0.0138887777707795,"tvd":0.0833329999912182,"kl":0.6877651696719952,"case_id":"typed:146/r0","base_id":"typed:146","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.795703}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0005513034352705383,"1":0.016476980941988375,"2":0.3737383226326251,"3":0.6092333929901171},"predicted":"3","confidence":0.6092333929901171,"entropy_confidence":0.46510020031371235,"candidate_mass":0.9999968628101716,"accepted":false,"accepted_predicted":null,"input_tokens":435,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":4.10579096655081,"brier_hard":1.4781634938909625,"score_mae":1.20832080517759,"within_one":false,"case_id":"typed:146/r0","base_id":"typed:146","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.795703}
{"id":"needs_human","type":"binary","probabilities":{"false":1.8553947784288112e-07,"true":0.9999998144605223},"predicted":"true","confidence":0.9999998144605223,"entropy_confidence":0.9999955833319224,"candidate_mass":0.9999998002328341,"accepted":true,"accepted_predicted":"true","input_tokens":403,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.8553949488308862e-07,"brier_hard":6.884979561250201e-14,"soft_accuracy":0.699999925784209,"brier_soft":0.17999977735269562,"tvd":0.29999981446052226,"kl":2.513429780713389,"case_id":"typed:146/r0","base_id":"typed:146","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.795703}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.3342367806078624e-05,"1":8.365890763694495e-05,"2":0.9956576642681385,"3":0.004225334456419186},"predicted":"2","confidence":0.9956576642681385,"entropy_confidence":0.9793979417417272,"candidate_mass":0.999998100966165,"accepted":true,"accepted_predicted":"2","input_tokens":429,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0043517910537148,"brier_hard":3.671744140312221e-05,"score_mae":0.24592500918682836,"within_one":true,"case_id":"typed:146/r0","base_id":"typed:146","repeat":0,"workflow":"customer_service","case_elapsed_ms":339.795703}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00020667237680409428,"close_no_action":0.0003612624097282087,"escalate_to_human":0.0030101426822502476,"execute_refund":0.00047517037475743924,"request_information":0.99594675215646},"predicted":"request_information","confidence":0.99594675215646,"entropy_confidence":0.9815004009506506,"candidate_mass":0.9999485880565593,"accepted":true,"accepted_predicted":"request_information","input_tokens":225,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.004061484516967122,"brier_hard":2.5888787933729967e-05,"soft_accuracy":0.6312180959822077,"brier_soft":0.1955939566924383,"tvd":0.36308892253121744,"kl":1.7895148777077896,"case_id":"typed:147/r0","base_id":"typed:147","repeat":0,"workflow":"customer_service","case_elapsed_ms":196.405401}
{"id":"category","type":"choice","probabilities":{"account":7.636321738400976e-07,"billing":0.9999939953241909,"delivery":3.328600162775034e-06,"refund":1.4819760301265707e-06,"technical":4.304674427604264e-07},"predicted":"billing","confidence":0.9999939953241909,"entropy_confidence":0.9999472205218447,"candidate_mass":0.9999992749655532,"accepted":true,"accepted_predicted":"billing","input_tokens":215,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":6.004693837272729e-06,"brier_hard":5.0100399886504505e-11,"soft_accuracy":0.5833303710207641,"brier_soft":0.29166083199066234,"tvd":0.4166652225668003,"kl":3.5232285116325635,"case_id":"typed:147/r0","base_id":"typed:147","repeat":0,"workflow":"customer_service","case_elapsed_ms":196.405401}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.662250273481329e-05,"1":9.746666128184285e-05,"2":0.062271648012523585,"3":0.937594262823459},"predicted":"3","confidence":0.937594262823459,"entropy_confidence":0.8307916522852936,"candidate_mass":0.999998202474532,"accepted":true,"accepted_predicted":"3","input_tokens":201,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":9.236000174004612,"brier_hard":1.8827658373440554,"score_mae":1.3707565511567064,"within_one":false,"case_id":"typed:147/r0","base_id":"typed:147","repeat":0,"workflow":"customer_service","case_elapsed_ms":196.405401}
{"id":"needs_human","type":"binary","probabilities":{"false":6.965654577310493e-09,"true":0.9999999930343471},"predicted":"true","confidence":0.9999999930343471,"entropy_confidence":0.9999998012016907,"candidate_mass":0.9999999623952582,"accepted":true,"accepted_predicted":"true","input_tokens":169,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.782274253181473,"brier_hard":1.999999972137385,"soft_accuracy":0.3000000027862623,"brier_soft":0.9799999804961694,"tvd":0.6999999930343461,"kl":6.086046421175243,"case_id":"typed:147/r0","base_id":"typed:147","repeat":0,"workflow":"customer_service","case_elapsed_ms":196.405401}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.6360341480683686e-05,"1":4.312099316581055e-05,"2":0.0003512976198589587,"3":0.999559221045496},"predicted":"3","confidence":0.999559221045496,"entropy_confidence":0.9970201686622346,"candidate_mass":0.9999974070051582,"accepted":true,"accepted_predicted":"3","input_tokens":195,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0004408761261025527,"brier_hard":3.2170480576602583e-07,"score_mae":0.5994233793693717,"within_one":true,"case_id":"typed:147/r0","base_id":"typed:147","repeat":0,"workflow":"customer_service","case_elapsed_ms":196.405401}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0010901948777343488,"close_no_action":0.0014028850197337049,"escalate_to_human":0.002198217045206511,"execute_refund":0.003873083424957555,"request_information":0.9914356196323673},"predicted":"request_information","confidence":0.9914356196323673,"entropy_confidence":0.962630798935875,"candidate_mass":0.9999442051595016,"accepted":true,"accepted_predicted":"request_information","input_tokens":358,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":6.120108681194079,"brier_hard":1.9785711433300477,"soft_accuracy":0.08447334579647826,"brier_soft":1.426799041383504,"tvd":0.9119757030573251,"kl":4.964134882897224,"case_id":"typed:148/r0","base_id":"typed:148","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.832903}
{"id":"category","type":"choice","probabilities":{"account":1.3491745500919983e-11,"billing":8.193441712715977e-12,"delivery":0.9999999999598064,"refund":1.3306499997979858e-11,"technical":5.202419270711516e-12},"predicted":"delivery","confidence":0.9999999999598064,"entropy_confidence":0.9999999993442197,"candidate_mass":0.99999999984726,"accepted":true,"accepted_predicted":"delivery","input_tokens":348,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":4.019362620591723e-11,"brier_hard":2.0688153797418395e-21,"soft_accuracy":0.9866679866284629,"brier_soft":0.00022217822301657007,"tvd":0.013332013291819496,"kl":0.10954963622731105,"case_id":"typed:148/r0","base_id":"typed:148","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.832903}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.001316944553250901,"1":0.06758425769372836,"2":0.7474067612624712,"3":0.18369203649054788},"predicted":"2","confidence":0.7474067612624712,"entropy_confidence":0.4808449705566231,"candidate_mass":0.9999958723670297,"accepted":false,"accepted_predicted":null,"input_tokens":334,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.2911457155942685,"brier_hard":0.10211547475692763,"score_mae":0.08014088969031441,"within_one":true,"case_id":"typed:148/r0","base_id":"typed:148","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.832903}
{"id":"needs_human","type":"binary","probabilities":{"false":2.601249681534413e-07,"true":0.9999997398750314},"predicted":"true","confidence":0.9999997398750314,"entropy_confidence":0.9999939346695937,"candidate_mass":0.9999995532259017,"accepted":true,"accepted_predicted":"true","input_tokens":302,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.6012500239770623e-07,"brier_hard":1.353299983278755e-13,"soft_accuracy":0.7999998439250189,"brier_soft":0.07999979190016061,"tvd":0.1999997398750316,"kl":1.6635534414438704,"case_id":"typed:148/r0","base_id":"typed:148","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.832903}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009618412600133268,"1":0.04392754271834231,"2":0.946547917346707,"3":0.008562698674937965},"predicted":"2","confidence":0.946547917346707,"entropy_confidence":0.8292402589731604,"candidate_mass":0.9999928160254556,"accepted":true,"accepted_predicted":"2","input_tokens":328,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.05493368378567002,"brier_hard":0.004860999096453508,"score_mae":0.20395552656342986,"within_one":true,"case_id":"typed:148/r0","base_id":"typed:148","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.832903}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.7795642980986685e-06,"close_no_action":6.947443788855535e-06,"escalate_to_human":1.585698523785557e-05,"execute_refund":6.468841214896878e-06,"request_information":0.9999689471654596},"predicted":"request_information","confidence":0.9999689471654596,"entropy_confidence":0.999757883884108,"candidate_mass":0.9999984541512579,"accepted":true,"accepted_predicted":"request_information","input_tokens":438,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":3.1053316689660555e-05,"brier_hard":1.3090022447817365e-09,"soft_accuracy":0.4999891173484886,"brier_soft":0.3755152387209443,"tvd":0.4999823634504637,"kl":4.258612122366464,"case_id":"typed:149/r0","base_id":"typed:149","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.124203}
{"id":"category","type":"choice","probabilities":{"account":2.2806687451659554e-09,"billing":0.999999907513783,"delivery":5.3931019004730445e-08,"refund":2.484348535377195e-08,"technical":1.143104385128718e-08},"predicted":"billing","confidence":0.999999907513783,"entropy_confidence":0.999998953361677,"candidate_mass":0.9999999842748847,"accepted":true,"accepted_predicted":"billing","input_tokens":428,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":9.248622129503437e-08,"brier_hard":1.221532412720152e-14,"soft_accuracy":0.883332919637498,"brier_soft":0.027222353530582062,"tvd":0.1166669885689561,"kl":0.9649627302326079,"case_id":"typed:149/r0","base_id":"typed:149","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.124203}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":7.376039492672486e-06,"1":1.3668923338267348e-05,"2":0.017559592631748416,"3":0.9824193624054216},"predicted":"3","confidence":0.9824193624054216,"entropy_confidence":0.9360568418222193,"candidate_mass":0.9999979889110698,"accepted":true,"accepted_predicted":"3","input_tokens":414,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.042154889691844,"brier_hard":1.9303369579002165,"score_mae":0.9657239414031,"within_one":true,"case_id":"typed:149/r0","base_id":"typed:149","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.124203}
{"id":"needs_human","type":"binary","probabilities":{"false":1.1498617674078591e-08,"true":0.9999999885013829},"predicted":"true","confidence":0.9999999885013829,"entropy_confidence":0.99999968014687,"candidate_mass":0.9999999589896769,"accepted":true,"accepted_predicted":"true","input_tokens":382,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.281039011056112,"brier_hard":1.9999999540055307,"soft_accuracy":0.48333300038329524,"brier_soft":0.5338895540141757,"tvd":0.5166669885013826,"kl":4.407271956294619,"case_id":"typed:149/r0","base_id":"typed:149","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.124203}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.92078218348488e-08,"1":1.3267903108543468e-08,"2":2.1902039474799564e-07,"3":0.9999997385038797},"predicted":"3","confidence":0.9999997385038797,"entropy_confidence":0.9999968496224805,"candidate_mass":0.9999999521611553,"accepted":true,"accepted_predicted":"3","input_tokens":408,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.614961545030149e-07,"brier_hard":1.1737928836349969e-13,"score_mae":0.2999996668203315,"within_one":true,"case_id":"typed:149/r0","base_id":"typed:149","repeat":0,"workflow":"customer_service","case_elapsed_ms":336.124203}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.458820746964836e-06,"close_no_action":7.781039454300338e-06,"escalate_to_human":0.9999656937591875,"execute_refund":3.3555284514470118e-06,"request_information":1.871085215935508e-05},"predicted":"escalate_to_human","confidence":0.9999656937591875,"entropy_confidence":0.9997348354231235,"candidate_mass":0.9999997534546934,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":592,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":3.4306829285003264e-05,"brier_hard":1.6186993758434968e-09,"soft_accuracy":0.7166460595999727,"brier_soft":0.1298387702045652,"tvd":0.2832994104254711,"kl":2.3423831800537833,"case_id":"typed:150/r0","base_id":"typed:150","repeat":0,"workflow":"customer_service","case_elapsed_ms":510.922505}
{"id":"category","type":"choice","probabilities":{"account":0.9999999999262279,"billing":5.839564323565793e-11,"delivery":1.2887425851094516e-11,"refund":1.0843770409884329e-12,"technical":1.40416098678857e-12},"predicted":"account","confidence":0.9999999999262279,"entropy_confidence":0.9999999988560355,"candidate_mass":0.9999999994386979,"accepted":true,"accepted_predicted":"account","input_tokens":582,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":7.377209954301357e-11,"brier_hard":9.021607106200353e-21,"soft_accuracy":0.733332999946275,"brier_soft":0.14222257773790595,"tvd":0.2666669999985956,"kl":2.2286466586567752,"case_id":"typed:150/r0","base_id":"typed:150","repeat":0,"workflow":"customer_service","case_elapsed_ms":510.922505}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.8923578578946798e-05,"1":6.73862719954281e-05,"2":0.410693177306247,"3":0.5892105128431785},"predicted":"3","confidence":0.5892105128431785,"entropy_confidence":0.5108508238115259,"candidate_mass":0.999992020636696,"accepted":false,"accepted_predicted":null,"input_tokens":568,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.8899088705115261,"brier_hard":0.694451565095811,"score_mae":0.28908527941402484,"within_one":true,"case_id":"typed:150/r0","base_id":"typed:150","repeat":0,"workflow":"customer_service","case_elapsed_ms":510.922505}
{"id":"needs_human","type":"binary","probabilities":{"false":1.4629210364539932e-10,"true":0.9999999998537064},"predicted":"true","confidence":0.9999999998537064,"entropy_confidence":0.9999999950095171,"candidate_mass":0.9999999967378592,"accepted":true,"accepted_predicted":"true","input_tokens":536,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4629364387634795e-10,"brier_hard":4.2803209824485023e-20,"soft_accuracy":0.816666999907347,"brier_soft":0.06722197767071872,"tvd":0.1833329998537071,"kl":1.5231647816692588,"case_id":"typed:150/r0","base_id":"typed:150","repeat":0,"workflow":"customer_service","case_elapsed_ms":510.922505}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00023666246424768595,"1":0.000647583473728379,"2":0.9556632195783438,"3":0.04345253448367899},"predicted":"2","confidence":0.9556632195783438,"entropy_confidence":0.8655837195919355,"candidate_mass":0.9999967895939524,"accepted":true,"accepted_predicted":"2","input_tokens":562,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.1360860979321323,"brier_hard":1.8282757184140204,"score_mae":0.3576683739185471,"within_one":true,"case_id":"typed:150/r0","base_id":"typed:150","repeat":0,"workflow":"customer_service","case_elapsed_ms":510.922505}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0004708067616376642,"close_no_action":0.00022851692764323423,"escalate_to_human":0.9827585792313097,"execute_refund":0.0006114683640955459,"request_information":0.015930628715315424},"predicted":"escalate_to_human","confidence":0.9827585792313097,"entropy_confidence":0.9421634457406652,"candidate_mass":0.9999712216032658,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":627,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.017391784897554927,"brier_hard":0.00055169929394157,"soft_accuracy":0.7873267995748446,"brier_soft":0.04655418638487186,"tvd":0.18275857923130892,"kl":0.6058202155698521,"case_id":"typed:151/r0","base_id":"typed:151","repeat":0,"workflow":"customer_service","case_elapsed_ms":529.0737049999999}
{"id":"category","type":"choice","probabilities":{"account":1.2370873129586052e-09,"billing":1.0282805123854957e-09,"delivery":0.9999999915247777,"refund":4.284618752623465e-09,"technical":1.925235987132562e-09},"predicted":"delivery","confidence":0.9999999915247777,"entropy_confidence":0.9999998904448233,"candidate_mass":0.9999999949987721,"accepted":true,"accepted_predicted":"delivery","input_tokens":617,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":8.475222337655056e-09,"brier_hard":9.648163035743503e-17,"soft_accuracy":0.7500007441910712,"brier_soft":0.10374966883615772,"tvd":0.24999924152402755,"kl":2.0868171714888017,"case_id":"typed:151/r0","base_id":"typed:151","repeat":0,"workflow":"customer_service","case_elapsed_ms":529.0737049999999}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.4723845926938585e-09,"1":8.512528128549805e-10,"2":3.6320907792171574e-08,"3":0.9999999603554536},"predicted":"3","confidence":0.9999999603554536,"entropy_confidence":0.9999994744051837,"candidate_mass":0.9999999915734357,"accepted":true,"accepted_predicted":"3","input_tokens":603,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.964454720931243e-08,"brier_hard":2.897735720895473e-15,"score_mae":0.2166669545594293,"within_one":true,"case_id":"typed:151/r0","base_id":"typed:151","repeat":0,"workflow":"customer_service","case_elapsed_ms":529.0737049999999}
{"id":"needs_human","type":"binary","probabilities":{"false":1.4518501154034531e-09,"true":0.9999999985481498},"predicted":"true","confidence":0.9999999985481498,"entropy_confidence":0.999999955279887,"candidate_mass":0.999999991734601,"accepted":true,"accepted_predicted":"true","input_tokens":571,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4518501966757077e-09,"brier_hard":4.215737748124002e-18,"soft_accuracy":0.8833329988869157,"brier_soft":0.02722237710046798,"tvd":0.11666699854814981,"kl":0.9649626498189434,"case_id":"typed:151/r0","base_id":"typed:151","repeat":0,"workflow":"customer_service","case_elapsed_ms":529.0737049999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.533721744223026e-08,"1":3.8403167545648924e-08,"2":5.614083954014717e-07,"3":0.999999324851221},"predicted":"3","confidence":0.999999324851221,"entropy_confidence":0.9999923199835046,"candidate_mass":0.9999999507900389,"accepted":true,"accepted_predicted":"3","input_tokens":597,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.751490069077003e-07,"brier_hard":7.781557598147137e-13,"score_mae":0.24999913577362154,"within_one":true,"case_id":"typed:151/r0","base_id":"typed:151","repeat":0,"workflow":"customer_service","case_elapsed_ms":529.0737049999999}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0017887799581669849,"close_no_action":0.0007705121479798827,"escalate_to_human":0.9965251270709999,"execute_refund":0.0005052910726053068,"request_information":0.0004102897502467162},"predicted":"escalate_to_human","confidence":0.9965251270709999,"entropy_confidence":0.9830105390262847,"candidate_mass":0.9999789689266368,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":493,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.0034809243225524316,"brier_hard":1.6291821328834053e-05,"soft_accuracy":0.4986448117209439,"brier_soft":0.3796658002994409,"tvd":0.4965251270710005,"kl":2.5445407236688866,"case_id":"typed:152/r0","base_id":"typed:152","repeat":0,"workflow":"customer_service","case_elapsed_ms":363.044103}
{"id":"category","type":"choice","probabilities":{"account":0.4919841378712151,"billing":0.5068806143166178,"delivery":0.0008888453822987127,"refund":0.00021257634541524737,"technical":3.3826084454659064e-05},"predicted":"billing","confidence":0.5068806143166178,"entropy_confidence":0.5639631278893746,"candidate_mass":0.9999805083044555,"accepted":false,"accepted_predicted":null,"input_tokens":483,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":null,"nll_hard":0.6794797778401539,"brier_hard":0.4852159568326593,"soft_accuracy":0.48259325524585134,"brier_soft":0.017791369642097432,"tvd":0.10865075453783098,"kl":0.15342210924793256,"case_id":"typed:152/r0","base_id":"typed:152","repeat":0,"workflow":"customer_service","case_elapsed_ms":363.044103}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.5574778292543823e-08,"1":5.177364893596927e-09,"2":2.2776048882738863e-07,"3":0.9999997414873697},"predicted":"3","confidence":0.9999997414873697,"entropy_confidence":0.9999969068827006,"candidate_mass":0.9999999723055847,"accepted":true,"accepted_predicted":"3","input_tokens":469,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.5851266371669467e-07,"brier_hard":1.1938449468865828e-13,"score_mae":0.21666668516045196,"within_one":true,"case_id":"typed:152/r0","base_id":"typed:152","repeat":0,"workflow":"customer_service","case_elapsed_ms":363.044103}
{"id":"needs_human","type":"binary","probabilities":{"false":1.252964937126212e-09,"true":0.999999998747036},"predicted":"true","confidence":0.999999998747036,"entropy_confidence":0.9999999611396675,"candidate_mass":0.9999999770127597,"accepted":true,"accepted_predicted":"true","input_tokens":437,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.2529639555969623e-09,"brier_hard":3.139839805725828e-18,"soft_accuracy":0.8499999991229253,"brier_soft":0.04499999924822135,"tvd":0.14999999874703557,"kl":1.2434099667883385,"case_id":"typed:152/r0","base_id":"typed:152","repeat":0,"workflow":"customer_service","case_elapsed_ms":363.044103}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.6196756989032563e-07,"1":1.1145174057939699e-07,"2":2.0981533349026695e-06,"3":0.9999975284273529},"predicted":"3","confidence":0.9999975284273529,"entropy_confidence":0.9999742780364149,"candidate_mass":0.9999998013086879,"accepted":true,"accepted_predicted":"3","input_tokens":463,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.4715757014410027e-06,"brier_hard":1.0591967264809975e-11,"score_mae":0.1666638930404689,"within_one":true,"case_id":"typed:152/r0","base_id":"typed:152","repeat":0,"workflow":"customer_service","case_elapsed_ms":363.044103}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0025490610561819268,"close_no_action":0.00016859828959821997,"escalate_to_human":0.004950356103709443,"execute_refund":0.6538291809120426,"request_information":0.3385028036384687},"predicted":"execute_refund","confidence":0.6538291809120426,"entropy_confidence":0.572858454126325,"candidate_mass":0.9997705585834398,"accepted":false,"accepted_predicted":null,"input_tokens":236,"reused_prefix_tokens":0,"expected":"request_information","correct":false,"accepted_correct":null,"nll_hard":1.0832229041280108,"brier_hard":0.8651021707694836,"soft_accuracy":0.206557170337548,"brier_soft":0.5395716729273251,"tvd":0.6338291809120421,"kl":1.8711927899569962,"case_id":"typed:153/r0","base_id":"typed:153","repeat":0,"workflow":"customer_service","case_elapsed_ms":194.436802}
{"id":"category","type":"choice","probabilities":{"account":1.9957418052358553e-12,"billing":0.999999999928443,"delivery":4.939200024546562e-11,"refund":1.7768745132358704e-11,"technical":2.400554165931397e-12},"predicted":"billing","confidence":0.999999999928443,"entropy_confidence":0.9999999988806508,"candidate_mass":0.999999999721668,"accepted":true,"accepted_predicted":"billing","input_tokens":226,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":7.155698256412049e-11,"brier_hard":7.885445390798978e-21,"soft_accuracy":0.9533329999320415,"brier_soft":0.0030055877708030894,"tvd":0.046666999928443,"kl":0.38425820394685756,"case_id":"typed:153/r0","base_id":"typed:153","repeat":0,"workflow":"customer_service","case_elapsed_ms":194.436802}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0001464181853528017,"1":0.0008273933704530122,"2":0.9720571319190284,"3":0.02696905652516554},"predicted":"2","confidence":0.9720571319190284,"entropy_confidence":0.904670556761435,"candidate_mass":0.9999978161319526,"accepted":true,"accepted_predicted":"2","input_tokens":212,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.028340698554565284,"brier_hard":0.0015088399045226255,"score_mae":0.0925158267840065,"within_one":true,"case_id":"typed:153/r0","base_id":"typed:153","repeat":0,"workflow":"customer_service","case_elapsed_ms":194.436802}
{"id":"needs_human","type":"binary","probabilities":{"false":1.6702367897767353e-08,"true":0.9999999832976326},"predicted":"true","confidence":0.9999999832976326,"entropy_confidence":0.9999995443916878,"candidate_mass":0.9999999441424234,"accepted":true,"accepted_predicted":"true","input_tokens":180,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6702367517536513e-08,"brier_hard":5.579381694237841e-16,"soft_accuracy":0.5666669977730069,"brier_soft":0.3755549488272522,"tvd":0.43333298329763237,"kl":3.6692869565696222,"case_id":"typed:153/r0","base_id":"typed:153","repeat":0,"workflow":"customer_service","case_elapsed_ms":194.436802}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007622684496118966,"1":0.0013479374785170235,"2":0.9489892336585273,"3":0.04890056041334487},"predicted":"2","confidence":0.9489892336585273,"entropy_confidence":0.8473278724657942,"candidate_mass":0.999992118016605,"accepted":true,"accepted_predicted":"2","input_tokens":206,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.01796642217138,"brier_hard":1.8051731075704835,"score_mae":0.27063891396439343,"within_one":true,"case_id":"typed:153/r0","base_id":"typed:153","repeat":0,"workflow":"customer_service","case_elapsed_ms":194.436802}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0003684957421029883,"close_no_action":0.00031301338230972495,"escalate_to_human":0.012728814023976319,"execute_refund":4.170897410645424e-05,"request_information":0.9865479678775062},"predicted":"request_information","confidence":0.9865479678775062,"entropy_confidence":0.9535440152412415,"candidate_mass":0.9999865284936876,"accepted":true,"accepted_predicted":"request_information","input_tokens":485,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":4.363887034619591,"brier_hard":1.9479815230878694,"soft_accuracy":0.07066355052391318,"brier_soft":1.6473120500879956,"tvd":0.9269026902339216,"kl":3.8521700469320024,"case_id":"typed:154/r0","base_id":"typed:154","repeat":0,"workflow":"customer_service","case_elapsed_ms":352.19120300000003}
{"id":"category","type":"choice","probabilities":{"account":0.9999332859418979,"billing":2.129418133463157e-05,"delivery":7.977421050685467e-06,"refund":7.15713964645851e-06,"technical":3.0285316072027128e-05},"predicted":"account","confidence":0.9999332859418979,"entropy_confidence":0.9995095625158993,"candidate_mass":0.9999970072049307,"accepted":true,"accepted_predicted":"account","input_tokens":475,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":6.671628358385734e-05,"brier_hard":5.93627197128422e-09,"soft_accuracy":0.8332819308756644,"brier_soft":0.046079871624739,"tvd":0.16660744308154352,"kl":1.205236990011226,"case_id":"typed:154/r0","base_id":"typed:154","repeat":0,"workflow":"customer_service","case_elapsed_ms":352.19120300000003}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.790362562474489e-07,"1":1.9744164964461155e-07,"2":1.1843551281618067e-05,"3":0.9999876799708138},"predicted":"3","confidence":0.9999876799708138,"entropy_confidence":0.999888963532091,"candidate_mass":0.9999998623252703,"accepted":true,"accepted_predicted":"3","input_tokens":461,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.2320105078346498e-05,"brier_hard":2.9216967054555473e-10,"score_mae":0.06665392445665441,"within_one":true,"case_id":"typed:154/r0","base_id":"typed:154","repeat":0,"workflow":"customer_service","case_elapsed_ms":352.19120300000003}
{"id":"needs_human","type":"binary","probabilities":{"false":6.870497368504846e-09,"true":0.9999999931295029},"predicted":"true","confidence":0.9999999931295029,"entropy_confidence":0.9999998037811146,"candidate_mass":0.9999999598254775,"accepted":true,"accepted_predicted":"true","input_tokens":429,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.870497083764826e-09,"brier_hard":9.44074639443399e-17,"soft_accuracy":0.953332993770754,"brier_soft":0.004355616495498123,"tvd":0.0466669931295028,"kl":0.3842582104285116,"case_id":"typed:154/r0","base_id":"typed:154","repeat":0,"workflow":"customer_service","case_elapsed_ms":352.19120300000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.0437811026910377e-06,"1":1.4026310815191839e-06,"2":9.311097082697082e-05,"3":0.9999024426169895},"predicted":"3","confidence":0.9999024426169895,"entropy_confidence":0.9992646949181795,"candidate_mass":0.999999396206712,"accepted":true,"accepted_predicted":"3","input_tokens":455,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":9.75621420415496e-05,"brier_hard":1.8198327845557777e-08,"score_mae":0.08322795242370384,"within_one":true,"case_id":"typed:154/r0","base_id":"typed:154","repeat":0,"workflow":"customer_service","case_elapsed_ms":352.19120300000003}
{"id":"action","type":"choice","probabilities":{"answer_directly":4.180752366230597e-05,"close_no_action":1.3672404797778096e-05,"escalate_to_human":0.00011554099772857774,"execute_refund":0.9997639281657423,"request_information":6.505090807044173e-05},"predicted":"execute_refund","confidence":0.9997639281657423,"entropy_confidence":0.998455800878583,"candidate_mass":0.9999856333443361,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":330,"reused_prefix_tokens":0,"expected":"execute_refund","correct":true,"accepted_correct":true,"nll_hard":0.00023609970359932957,"brier_hard":7.52460574144124e-08,"soft_accuracy":0.6498725666130204,"brier_soft":0.16700497612950138,"tvd":0.34977760057053936,"kl":2.3502252961868413,"case_id":"typed:155/r0","base_id":"typed:155","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.202603}
{"id":"category","type":"choice","probabilities":{"account":3.9487192036957705e-11,"billing":0.9999999975674161,"delivery":1.6432910195920449e-09,"refund":6.689276336786622e-10,"technical":8.087801585505402e-11},"predicted":"billing","confidence":0.9999999975674161,"entropy_confidence":0.9999999673007512,"candidate_mass":0.9999999939573385,"accepted":true,"accepted_predicted":"billing","input_tokens":320,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":2.4325839150407416e-09,"brier_hard":9.073434535274623e-18,"soft_accuracy":0.533332999014789,"brier_soft":0.4355561748832539,"tvd":0.46666699933107236,"kl":3.962903836177445,"case_id":"typed:155/r0","base_id":"typed:155","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.202603}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.222377739836222e-05,"1":6.078618047811932e-05,"2":0.24623490073486026,"3":0.7536720893072624},"predicted":"3","confidence":0.7536720893072624,"entropy_confidence":0.596657849916958,"candidate_mass":0.9999946006292894,"accepted":false,"accepted_predicted":null,"input_tokens":306,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.4014693176226416,"brier_hard":1.1361834478042914,"score_mae":0.5868798555719859,"within_one":true,"case_id":"typed:155/r0","base_id":"typed:155","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.202603}
{"id":"needs_human","type":"binary","probabilities":{"false":1.9187850267127285e-06,"true":0.9999980812149728},"predicted":"true","confidence":0.9999980812149728,"entropy_confidence":0.9999607914140995,"candidate_mass":0.9999988178574231,"accepted":true,"accepted_predicted":"true","input_tokens":274,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.9187868680698997e-06,"brier_hard":7.363471959342162e-12,"soft_accuracy":0.5333328720822772,"brier_soft":0.4355525960507549,"tvd":0.4666650812149731,"kl":3.962904858232424,"case_id":"typed:155/r0","base_id":"typed:155","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.202603}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.2660856628140109e-06,"1":2.13194925740499e-06,"2":0.9999940125975446,"3":2.589367535359826e-06},"predicted":"2","confidence":0.9999940125975446,"entropy_confidence":0.9999391686371231,"candidate_mass":0.9999996332796461,"accepted":true,"accepted_predicted":"2","input_tokens":300,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":5.987420379953054e-06,"brier_hard":4.870199293768767e-11,"score_mae":0.09999792524695272,"within_one":true,"case_id":"typed:155/r0","base_id":"typed:155","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.202603}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.007204997007589949,"close_no_action":0.00029209861831890256,"escalate_to_human":0.02708732871308423,"execute_refund":0.965387411353323,"request_information":2.8164307685127404e-05},"predicted":"execute_refund","confidence":0.965387411353323,"entropy_confidence":0.8943912588055816,"candidate_mass":0.9999620427241437,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":353,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":4.932980464854825,"brier_hard":1.9183485814578112,"soft_accuracy":0.021416535125561974,"brier_soft":1.327214624244668,"tvd":0.9487203946863058,"kl":5.656668873525025,"case_id":"typed:156/r0","base_id":"typed:156","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.679603}
{"id":"category","type":"choice","probabilities":{"account":0.9999995903909586,"billing":2.519249114683935e-07,"delivery":1.092996687429101e-07,"refund":2.3148486627055052e-08,"technical":2.5235974815993726e-08},"predicted":"account","confidence":0.9999995903909586,"entropy_confidence":0.9999957514000427,"candidate_mass":0.9999994498744499,"accepted":true,"accepted_predicted":"account","input_tokens":343,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":4.0960912524170564e-07,"brier_hard":2.443648522209262e-13,"soft_accuracy":0.9299996263230852,"brier_soft":0.006844383691991452,"tvd":0.06999959039095847,"kl":0.5772334626384225,"case_id":"typed:156/r0","base_id":"typed:156","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.679603}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":8.923233158457656e-09,"1":1.5806886001728646e-09,"2":7.338253705325748e-08,"3":0.9999999161135413},"predicted":"3","confidence":0.9999999161135413,"entropy_confidence":0.9999989274962787,"candidate_mass":0.9999999849244859,"accepted":true,"accepted_predicted":"3","input_tokens":329,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":8.388646218677219e-08,"brier_hard":1.2504057358732533e-14,"score_mae":0.3633328966863867,"within_one":true,"case_id":"typed:156/r0","base_id":"typed:156","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.679603}
{"id":"needs_human","type":"binary","probabilities":{"false":8.971712833495336e-07,"true":0.999999102828718},"predicted":"true","confidence":0.999999102828718,"entropy_confidence":0.9999806831777028,"candidate_mass":0.9999988991314253,"accepted":true,"accepted_predicted":"true","input_tokens":297,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.924019041779228,"brier_hard":1.999996411316479,"soft_accuracy":0.4100001614908315,"brier_soft":0.6961978826773827,"tvd":0.5899991028287173,"kl":5.068545958399989,"case_id":"typed:156/r0","base_id":"typed:156","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.679603}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.334324592523374e-06,"1":3.4035978583562108e-06,"2":1.6102738797512956e-05,"3":0.9999711593387502},"predicted":"3","confidence":0.9999711593387502,"entropy_confidence":0.9997421035026112,"candidate_mass":0.9999983421096766,"accepted":true,"accepted_predicted":"3","input_tokens":323,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.884107714961883e-05,"brier_hard":1.189796032083846e-09,"score_mae":0.5966160870917041,"within_one":true,"case_id":"typed:156/r0","base_id":"typed:156","repeat":0,"workflow":"customer_service","case_elapsed_ms":312.679603}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.9682622546862761,"close_no_action":0.00343163691974862,"escalate_to_human":0.0010933581307986978,"execute_refund":0.0029168175241424195,"request_information":0.024295932739035304},"predicted":"answer_directly","confidence":0.9682622546862761,"entropy_confidence":0.8971675905216361,"candidate_mass":0.9998349497309713,"accepted":true,"accepted_predicted":"answer_directly","input_tokens":375,"reused_prefix_tokens":0,"expected":"answer_directly","correct":true,"accepted_correct":true,"nll_hard":0.03225230413622887,"brier_hard":0.0016190562136788383,"soft_accuracy":0.7587172224833585,"brier_soft":0.079843105553825,"tvd":0.21223985520241523,"kl":0.9589141943444063,"case_id":"typed:157/r0","base_id":"typed:157","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.92310299999997}
{"id":"category","type":"choice","probabilities":{"account":1.8871425408848072e-10,"billing":0.9999999965865953,"delivery":2.394822992615769e-09,"refund":6.997056497443859e-10,"technical":1.30161368991993e-10},"predicted":"billing","confidence":0.9999999965865953,"entropy_confidence":0.9999999547116356,"candidate_mass":0.9999999954598433,"accepted":true,"accepted_predicted":"billing","input_tokens":365,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.4134046758530412e-09,"brier_hard":1.79286516552841e-17,"soft_accuracy":0.9966659999325038,"brier_soft":1.667327682970997e-05,"tvd":0.0033339963471274987,"kl":0.02737881478267241,"case_id":"typed:157/r0","base_id":"typed:157","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.92310299999997}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.1314807512188342e-05,"1":3.584223208897604e-05,"2":0.000783884251973009,"3":0.9991589587084244},"predicted":"3","confidence":0.9991589587084244,"entropy_confidence":0.9949198391359831,"candidate_mass":0.99999867481526,"accepted":true,"accepted_predicted":"3","input_tokens":351,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0008413951652314683,"brier_hard":1.3235639612468118e-06,"score_mae":0.14908048686130826,"within_one":true,"case_id":"typed:157/r0","base_id":"typed:157","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.92310299999997}
{"id":"needs_human","type":"binary","probabilities":{"false":0.001923956415886117,"true":0.998076043584113},"predicted":"true","confidence":0.998076043584113,"entropy_confidence":0.979869616169592,"candidate_mass":0.9999388321992582,"accepted":true,"accepted_predicted":"true","input_tokens":319,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.253371579861046,"brier_hard":1.992311577553034,"soft_accuracy":0.4900384791283173,"brier_soft":0.5162825321281718,"tvd":0.5080760435841134,"kl":2.4972159852073683,"case_id":"typed:157/r0","base_id":"typed:157","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.92310299999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.5995040630662554e-05,"1":3.094021619554099e-05,"2":0.440792163859437,"3":0.559130900883737},"predicted":"3","confidence":0.559130900883737,"entropy_confidence":0.5044830185784639,"candidate_mass":0.9999866011359613,"accepted":false,"accepted_predicted":null,"input_tokens":345,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.5813716635021428,"brier_hard":0.38866329734831084,"score_mae":0.32432502941371943,"within_one":true,"case_id":"typed:157/r0","base_id":"typed:157","repeat":0,"workflow":"customer_service","case_elapsed_ms":311.92310299999997}
{"id":"action","type":"choice","probabilities":{"answer_directly":5.0425630635960735e-05,"close_no_action":7.730215565461084e-05,"escalate_to_human":0.00016173485055667616,"execute_refund":4.8100690196671534e-05,"request_information":0.999662436672956},"predicted":"request_information","confidence":0.999662436672956,"entropy_confidence":0.9978511447030876,"candidate_mass":0.9999892817671356,"accepted":true,"accepted_predicted":"request_information","input_tokens":378,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":8.729552288588295,"brier_hard":1.999001554584004,"soft_accuracy":0.30000655137304627,"brier_soft":0.913467479317025,"tvd":0.699662436672956,"kl":5.343997476903535,"case_id":"typed:158/r0","base_id":"typed:158","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.896703}
{"id":"category","type":"choice","probabilities":{"account":1.490161110783438e-08,"billing":3.334153560324998e-07,"delivery":2.478474318414771e-07,"refund":1.5596884954486722e-07,"technical":0.9999992478667523},"predicted":"technical","confidence":0.9999992478667523,"entropy_confidence":0.9999924149489401,"candidate_mass":0.9999999146211246,"accepted":true,"accepted_predicted":"technical","input_tokens":368,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":7.521335305357704e-07,"brier_hard":7.628469114214578e-13,"soft_accuracy":0.9933312608037701,"brier_soft":5.556513155874798e-05,"tvd":0.006667241198758488,"kl":0.05295280647758231,"case_id":"typed:158/r0","base_id":"typed:158","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.896703}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.148061961444183e-07,"1":2.895090482401749e-07,"2":9.611611132902412e-06,"3":0.9999896840736225},"predicted":"3","confidence":0.9999896840736225,"entropy_confidence":0.9999049200994778,"candidate_mass":0.9999998473122399,"accepted":true,"accepted_predicted":"3","input_tokens":354,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.0315979586992157e-05,"brier_hard":1.99057285264634e-10,"score_mae":0.19998856495218176,"within_one":true,"case_id":"typed:158/r0","base_id":"typed:158","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.896703}
{"id":"needs_human","type":"binary","probabilities":{"false":2.379616888473853e-08,"true":0.9999999762038317},"predicted":"true","confidence":0.9999999762038317,"entropy_confidence":0.9999993630386569,"candidate_mass":0.9999998534822191,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.3796168541223313e-08,"brier_hard":1.1325152773585422e-15,"soft_accuracy":0.8333329841359037,"brier_soft":0.05555576191385703,"tvd":0.16666697620383142,"kl":1.383124915385244,"case_id":"typed:158/r0","base_id":"typed:158","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.896703}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.0900651547219064e-06,"1":1.6156826066058349e-06,"2":0.9998249561456458,"3":0.00017133810659195818},"predicted":"2","confidence":0.9998249561456458,"entropy_confidence":0.9987666894289982,"candidate_mass":0.9999992902728,"accepted":true,"accepted_predicted":"2","input_tokens":348,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001750591763176944,"brier_hard":6.00040765203224e-08,"score_mae":0.14983445770632597,"within_one":true,"case_id":"typed:158/r0","base_id":"typed:158","repeat":0,"workflow":"customer_service","case_elapsed_ms":328.896703}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0016128149046845627,"close_no_action":0.0002345811709395881,"escalate_to_human":0.021670585141706018,"execute_refund":0.9678597699248391,"request_information":0.008622248857832058},"predicted":"execute_refund","confidence":0.9678597699248391,"entropy_confidence":0.8956336749306981,"candidate_mass":0.999923546166514,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":440,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.8317994614123627,"brier_hard":1.8939579775915434,"soft_accuracy":0.20363109838057217,"brier_soft":0.789125668891811,"tvd":0.7678597699248384,"kl":2.4162630614505463,"case_id":"typed:159/r0","base_id":"typed:159","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.562502}
{"id":"category","type":"choice","probabilities":{"account":3.447296760271226e-07,"billing":2.660245504404956e-07,"delivery":0.9999697699109217,"refund":9.523426199744216e-07,"technical":2.8666992230693384e-05},"predicted":"delivery","confidence":0.9999697699109217,"entropy_confidence":0.9997810164103305,"candidate_mass":0.9999997842070114,"accepted":true,"accepted_predicted":"delivery","input_tokens":430,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":3.0230546016671977e-05,"brier_hard":1.736751293314603e-09,"soft_accuracy":0.5666511832229721,"brier_soft":0.3244155306686505,"tvd":0.4333031146405983,"kl":3.597578676347961,"case_id":"typed:159/r0","base_id":"typed:159","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.562502}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.309910502264708e-06,"1":1.5106359138660833e-05,"2":0.06066017662443374,"3":0.9393194071059257},"predicted":"3","confidence":0.9393194071059257,"entropy_confidence":0.8347887025560466,"candidate_mass":0.999998198745216,"accepted":true,"accepted_predicted":"3","input_tokens":416,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.802467865014208,"brier_hard":1.7646802526014649,"score_mae":0.705960680925783,"within_one":true,"case_id":"typed:159/r0","base_id":"typed:159","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.562502}
{"id":"needs_human","type":"binary","probabilities":{"false":1.4374022655947707e-09,"true":0.9999999985625969},"predicted":"true","confidence":0.9999999985625969,"entropy_confidence":0.9999999557041711,"candidate_mass":0.999999983359107,"accepted":true,"accepted_predicted":"true","input_tokens":384,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4374030864799949e-09,"brier_hard":4.132252903189337e-18,"soft_accuracy":0.6266669996358566,"brier_soft":0.2787550556314806,"tvd":0.3733329985625974,"kl":3.145657459800558,"case_id":"typed:159/r0","base_id":"typed:159","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.562502}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.076598429459146e-07,"1":2.666875131958034e-07,"2":9.707934229203907e-06,"3":0.999989717718416},"predicted":"3","confidence":0.999989717718416,"entropy_confidence":0.9999055130365732,"candidate_mass":0.9999997227165658,"accepted":true,"accepted_predicted":"3","input_tokens":410,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.0282334446970602e-05,"brier_hard":2.0013507837886173e-10,"score_mae":0.5166558357112199,"within_one":true,"case_id":"typed:159/r0","base_id":"typed:159","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.562502}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.000530266603455783,"close_no_action":0.00016432251672722823,"escalate_to_human":0.0020489477381981265,"execute_refund":0.37773894924305335,"request_information":0.6195175138985661},"predicted":"request_information","confidence":0.6195175138985661,"entropy_confidence":0.5759405274003009,"candidate_mass":0.9999051214692789,"accepted":false,"accepted_predicted":null,"input_tokens":435,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":null,"nll_hard":0.47881430728985214,"brier_hard":0.28745814237656797,"soft_accuracy":0.2589419533207611,"brier_soft":0.348053619088178,"tvd":0.5805894631416192,"kl":3.0292495914834063,"case_id":"typed:160/r0","base_id":"typed:160","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.885604}
{"id":"category","type":"choice","probabilities":{"account":2.2861401343496384e-10,"billing":0.9999999881354517,"delivery":7.361211907826269e-09,"refund":3.4284795273567174e-09,"technical":8.462434952971964e-10},"predicted":"billing","confidence":0.9999999881354517,"entropy_confidence":0.9999998513167375,"candidate_mass":0.9999999964503168,"accepted":true,"accepted_predicted":"billing","input_tokens":425,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.1864548414171117e-08,"brier_hard":2.074778124439665e-16,"soft_accuracy":0.9799990083996487,"brier_soft":0.000550045041474461,"tvd":0.020000968134471334,"kl":0.16441623182024223,"case_id":"typed:160/r0","base_id":"typed:160","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.885604}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.864201454934461e-05,"1":6.084433290393769e-05,"2":0.03682258941280046,"3":0.9630779242397466},"predicted":"3","confidence":0.9630779242397466,"entropy_confidence":0.8854570414239248,"candidate_mass":0.9999986046557652,"accepted":true,"accepted_predicted":"3","input_tokens":411,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.3016437793924993,"brier_hard":1.85522981761864,"score_mae":0.8962727958777448,"within_one":true,"case_id":"typed:160/r0","base_id":"typed:160","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.885604}
{"id":"needs_human","type":"binary","probabilities":{"false":7.420460901259135e-09,"true":0.9999999925795393},"predicted":"true","confidence":0.9999999925795393,"entropy_confidence":0.9999997888987283,"candidate_mass":0.9999999685995857,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.4204607178865764e-09,"brier_hard":1.1012647684421872e-16,"soft_accuracy":0.5000000000000001,"brier_soft":0.4999999851590785,"tvd":0.4999999925795392,"kl":4.2585965994183494,"case_id":"typed:160/r0","base_id":"typed:160","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.885604}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.080573506729336e-08,"1":2.934174846654722e-08,"2":3.757131920766794e-07,"3":0.9999995441393255},"predicted":"3","confidence":0.9999995441393255,"entropy_confidence":0.9999946789563039,"candidate_mass":0.9999999378088181,"accepted":true,"accepted_predicted":"3","input_tokens":405,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":4.5586077844646135e-07,"brier_hard":3.524115182130938e-13,"score_mae":0.09999941318610928,"within_one":true,"case_id":"typed:160/r0","base_id":"typed:160","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.885604}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0010972460485587034,"close_no_action":0.0011995028113795938,"escalate_to_human":0.003047943540298703,"execute_refund":0.00041499591975924964,"request_information":0.9942403116800022},"predicted":"request_information","confidence":0.9942403116800022,"entropy_confidence":0.9737935147397354,"candidate_mass":0.9999457456996974,"accepted":true,"accepted_predicted":"request_information","input_tokens":505,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":6.814951830624832,"brier_hard":1.9863314102097542,"soft_accuracy":0.08443856727723402,"brier_soft":1.2868044655304038,"tvd":0.9113223075997622,"kl":5.028396944162512,"case_id":"typed:161/r0","base_id":"typed:161","repeat":0,"workflow":"customer_service","case_elapsed_ms":358.274203}
{"id":"category","type":"choice","probabilities":{"account":3.965312368530644e-05,"billing":0.9999292896974841,"delivery":2.958650302305649e-05,"refund":1.372060206782605e-06,"technical":9.861560225166199e-08},"predicted":"billing","confidence":0.9999292896974841,"entropy_confidence":0.9995021541098342,"candidate_mass":0.9999968080994424,"accepted":true,"accepted_predicted":"billing","input_tokens":495,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":7.071280260722703e-05,"brier_hard":7.449570535278003e-09,"soft_accuracy":0.9566004818258063,"brier_soft":0.002569358748926094,"tvd":0.04326228969748332,"kl":0.26069689874817403,"case_id":"typed:161/r0","base_id":"typed:161","repeat":0,"workflow":"customer_service","case_elapsed_ms":358.274203}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.037755431963275e-05,"1":0.002309429388025379,"2":0.03272036404407589,"3":0.964909829013579},"predicted":"3","confidence":0.964909829013579,"entropy_confidence":0.883885095994929,"candidate_mass":0.9999988988008847,"accepted":true,"accepted_predicted":"3","input_tokens":481,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.03572062344577982,"brier_hard":0.002307279432580456,"score_mae":0.4191466445169141,"within_one":true,"case_id":"typed:161/r0","base_id":"typed:161","repeat":0,"workflow":"customer_service","case_elapsed_ms":358.274203}
{"id":"needs_human","type":"binary","probabilities":{"false":5.438992949111609e-09,"true":0.9999999945610085},"predicted":"true","confidence":0.9999999945610085,"entropy_confidence":0.9999998428310092,"candidate_mass":0.9999999853919911,"accepted":true,"accepted_predicted":"true","input_tokens":449,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":19.02967191286392,"brier_hard":1.9999999782440312,"soft_accuracy":0.4833330001813041,"brier_soft":0.5338895665374088,"tvd":0.5166669945610078,"kl":4.407271953365802,"case_id":"typed:161/r0","base_id":"typed:161","repeat":0,"workflow":"customer_service","case_elapsed_ms":358.274203}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.397900020423456e-07,"1":2.4530043427385767e-06,"2":0.9999948874921663,"3":2.0197134878275387e-06},"predicted":"2","confidence":0.9999948874921663,"entropy_confidence":0.9999477677615306,"candidate_mass":0.9999996454225044,"accepted":true,"accepted_predicted":"2","input_tokens":475,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":5.112520902661554e-06,"brier_hard":3.664354047526336e-11,"score_mae":0.016668712870861357,"within_one":true,"case_id":"typed:161/r0","base_id":"typed:161","repeat":0,"workflow":"customer_service","case_elapsed_ms":358.274203}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.06854025820734838,"close_no_action":0.0006245084702661516,"escalate_to_human":0.0031629073180017963,"execute_refund":0.0002479190881507349,"request_information":0.9274244069162344},"predicted":"request_information","confidence":0.9274244069162344,"entropy_confidence":0.8269834113380006,"candidate_mass":0.9999032496984315,"accepted":true,"accepted_predicted":"request_information","input_tokens":344,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.756263637015916,"brier_hard":1.8584984383604615,"soft_accuracy":0.11956077439636911,"brier_soft":1.0545914819817268,"tvd":0.8282968344746505,"kl":3.1269357525467374,"case_id":"typed:162/r0","base_id":"typed:162","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.962503}
{"id":"category","type":"choice","probabilities":{"account":1.2727599469162365e-10,"billing":1.1824176647186383e-10,"delivery":0.9999999995023554,"refund":1.9618948656533465e-10,"technical":5.5937946078695385e-11},"predicted":"delivery","confidence":0.9999999995023554,"entropy_confidence":0.9999999926644451,"candidate_mass":0.9999999991281907,"accepted":true,"accepted_predicted":"delivery","input_tokens":334,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":4.976445923249659e-10,"brier_hard":3.1944980276044136e-19,"soft_accuracy":0.9766669995185445,"brier_soft":0.0010888577456215414,"tvd":0.023332999803810837,"kl":0.19184622907456772,"case_id":"typed:162/r0","base_id":"typed:162","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.962503}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.0601036998711e-05,"1":0.9966746109245453,"2":0.0032220840402720245,"3":4.2703998185104994e-05},"predicted":"1","confidence":0.9966746109245453,"entropy_confidence":0.9835349145772718,"candidate_mass":0.999995322885076,"accepted":true,"accepted_predicted":"1","input_tokens":320,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0033309304699784856,"brier_hard":2.1445534182875165e-05,"score_mae":0.030086109000355243,"within_one":true,"case_id":"typed:162/r0","base_id":"typed:162","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.962503}
{"id":"needs_human","type":"binary","probabilities":{"false":5.3846789744000415e-09,"true":0.9999999946153206},"predicted":"true","confidence":0.9999999946153206,"entropy_confidence":0.9999998443225356,"candidate_mass":0.9999999578849641,"accepted":true,"accepted_predicted":"true","input_tokens":288,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.384679383555562e-09,"brier_hard":5.798953956490664e-17,"soft_accuracy":0.5833329991025569,"brier_soft":0.3472227688035276,"tvd":0.41666699461532086,"kl":3.5232301158161228,"case_id":"typed:162/r0","base_id":"typed:162","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.962503}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00025255007493954273,"1":0.4754290114593988,"2":0.5237415594031825,"3":0.0005768790624805744},"predicted":"2","confidence":0.5237415594031825,"entropy_confidence":0.4960469531613143,"candidate_mass":0.9999927696365153,"accepted":false,"accepted_predicted":null,"input_tokens":314,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.743537700551169,"brier_hard":0.5494803396355339,"score_mae":0.12464276745320557,"within_one":true,"case_id":"typed:162/r0","base_id":"typed:162","repeat":0,"workflow":"customer_service","case_elapsed_ms":320.962503}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.401376431398584e-05,"close_no_action":4.4604849829976805e-05,"escalate_to_human":8.190656785398519e-05,"execute_refund":3.512934506152647e-05,"request_information":0.999824345472942},"predicted":"request_information","confidence":0.999824345472942,"entropy_confidence":0.9988131729360723,"candidate_mass":0.9999937083417305,"accepted":true,"accepted_predicted":"request_information","input_tokens":366,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":11.175470546054754,"brier_hard":1.9996207044005037,"soft_accuracy":0.3266218848244941,"brier_soft":0.8149607200581437,"tvd":0.6731573454729413,"kl":5.617271538623925,"case_id":"typed:163/r0","base_id":"typed:163","repeat":0,"workflow":"customer_service","case_elapsed_ms":346.364603}
{"id":"category","type":"choice","probabilities":{"account":8.20191265454954e-11,"billing":0.9999999963544433,"delivery":2.5849601130474475e-09,"refund":7.663444120125985e-10,"technical":2.1223420889347523e-10},"predicted":"billing","confidence":0.9999999963544433,"entropy_confidence":0.9999999518611492,"candidate_mass":0.9999999988664285,"accepted":true,"accepted_predicted":"billing","input_tokens":356,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":3.6455567520136365e-09,"brier_hard":2.0611157024115593e-17,"soft_accuracy":0.9233329966852397,"brier_soft":0.007577854894406968,"tvd":0.07666699635444271,"kl":0.6324792076998275,"case_id":"typed:163/r0","base_id":"typed:163","repeat":0,"workflow":"customer_service","case_elapsed_ms":346.364603}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.925091539674165e-05,"1":0.9999014536587588,"2":5.6785030725522007e-05,"3":2.51039512038013e-06},"predicted":"1","confidence":0.9999014536587588,"entropy_confidence":0.9992178582203921,"candidate_mass":0.9999995163385866,"accepted":true,"accepted_predicted":"1","input_tokens":342,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":9.855119725089574e-05,"brier_hard":1.4482857529663998e-08,"score_mae":0.243310445094429,"within_one":true,"case_id":"typed:163/r0","base_id":"typed:163","repeat":0,"workflow":"customer_service","case_elapsed_ms":346.364603}
{"id":"needs_human","type":"binary","probabilities":{"false":7.468399321027968e-08,"true":0.9999999253160069},"predicted":"true","confidence":0.9999999253160069,"entropy_confidence":0.9999981241381931,"candidate_mass":0.9999998615255048,"accepted":true,"accepted_predicted":"true","input_tokens":310,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.410000048755975,"brier_hard":1.9999997012640385,"soft_accuracy":0.1333330547683115,"brier_soft":1.5022231188734019,"tvd":0.8666669253160069,"kl":7.7136446713464455,"case_id":"typed:163/r0","base_id":"typed:163","repeat":0,"workflow":"customer_service","case_elapsed_ms":346.364603}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001093139453139111,"1":0.9997213747158592,"2":0.00015273104687625377,"3":1.65802919519917e-05},"predicted":"1","confidence":0.9997213747158592,"entropy_confidence":0.9979800851817332,"candidate_mass":0.9999989219429202,"accepted":true,"accepted_predicted":"1","input_tokens":336,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.000278664107376871,"brier_hard":1.1318326636376332e-07,"score_mae":0.13659042231453244,"within_one":true,"case_id":"typed:163/r0","base_id":"typed:163","repeat":0,"workflow":"customer_service","case_elapsed_ms":346.364603}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.09331029857838685,"close_no_action":0.0003349889797250468,"escalate_to_human":0.04530490826742245,"execute_refund":5.297435111973495e-05,"request_information":0.8609968298233442},"predicted":"request_information","confidence":0.8609968298233442,"entropy_confidence":0.6933295537817703,"candidate_mass":0.9999359884899296,"accepted":true,"accepted_predicted":"request_information","input_tokens":503,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.0943399020883895,"brier_hard":1.6614651859888097,"soft_accuracy":0.10288495937760106,"brier_soft":1.306304230436399,"tvd":0.8213612250664433,"kl":2.3647640947820956,"case_id":"typed:164/r0","base_id":"typed:164","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.924803}
{"id":"category","type":"choice","probabilities":{"account":0.054727003312180694,"billing":0.0004796953255185973,"delivery":0.00021389706095545816,"refund":0.00012568544831208797,"technical":0.9444537188530329},"predicted":"technical","confidence":0.9444537188530329,"entropy_confidence":0.863567126091788,"candidate_mass":0.9999812732085304,"accepted":true,"accepted_predicted":"technical","input_tokens":493,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.057148593926129324,"brier_hard":0.006080725897179281,"soft_accuracy":0.9296246436851135,"brier_soft":0.0029604540110180475,"tvd":0.038879281146966996,"kl":0.019852932375786576,"case_id":"typed:164/r0","base_id":"typed:164","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.924803}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.518487203079628e-05,"1":0.9988868560933478,"2":0.0009984358000114163,"3":6.952323460848859e-05},"predicted":"1","confidence":0.9988868560933478,"entropy_confidence":0.9934150308992802,"candidate_mass":0.9999976942865424,"accepted":true,"accepted_predicted":"1","input_tokens":479,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.001113763911476406,"brier_hard":2.2428385564721824e-06,"score_mae":0.23224070260280394,"within_one":true,"case_id":"typed:164/r0","base_id":"typed:164","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.924803}
{"id":"needs_human","type":"binary","probabilities":{"false":4.7591489548555163e-07,"true":0.9999995240851045},"predicted":"true","confidence":0.9999995240851045,"entropy_confidence":0.9999893178580422,"candidate_mass":0.9999996070263738,"accepted":true,"accepted_predicted":"true","input_tokens":447,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.759150087400095e-07,"brier_hard":4.529899754966414e-13,"soft_accuracy":0.9166666034039365,"brier_soft":0.013888619140789043,"tvd":0.0833325240851045,"kl":0.6877656058959155,"case_id":"typed:164/r0","base_id":"typed:164","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.924803}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.290519348924251e-06,"1":7.423576293008671e-06,"2":0.999967128661537,"3":2.3157242819277074e-05},"predicted":"2","confidence":0.999967128661537,"entropy_confidence":0.9997132947617641,"candidate_mass":0.9999995789375473,"accepted":true,"accepted_predicted":"2","input_tokens":473,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":3.287187873725565e-05,"brier_hard":1.6771387512039906e-09,"score_mae":0.1666558473721751,"within_one":true,"case_id":"typed:164/r0","base_id":"typed:164","repeat":0,"workflow":"customer_service","case_elapsed_ms":370.924803}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.2679759989276742e-05,"close_no_action":5.23262987068946e-05,"escalate_to_human":0.00027978470293228594,"execute_refund":3.2550180094882175e-05,"request_information":0.999612659058277},"predicted":"request_information","confidence":0.999612659058277,"entropy_confidence":0.9976569302306197,"candidate_mass":0.999994453603298,"accepted":true,"accepted_predicted":"request_information","input_tokens":397,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":8.18149016866315,"brier_hard":1.9986659813351018,"soft_accuracy":0.30005393651524137,"brier_soft":0.8591176777104835,"tvd":0.6996975355370786,"kl":5.080468367298802,"case_id":"typed:165/r0","base_id":"typed:165","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.608603}
{"id":"category","type":"choice","probabilities":{"account":0.9999973704602998,"billing":1.9578802902358904e-06,"delivery":4.17055041995244e-07,"refund":9.764853440323813e-08,"technical":1.569558335593789e-07},"predicted":"account","confidence":0.9999973704602998,"entropy_confidence":0.9999760630152892,"candidate_mass":0.9999990253987459,"accepted":true,"accepted_predicted":"account","input_tokens":387,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":false,"nll_hard":15.667301386082979,"brier_hard":1.9999944270198884,"soft_accuracy":0.43333244283786526,"brier_soft":0.6066634685895783,"tvd":0.5666644518304429,"kl":4.856818968442134,"case_id":"typed:165/r0","base_id":"typed:165","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.608603}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.165013278707707e-06,"1":1.1978328568809096e-06,"2":6.647161884417485e-05,"3":0.9999311655350193},"predicted":"3","confidence":0.9999311655350193,"entropy_confidence":0.9994658747807367,"candidate_mass":0.9999994413747114,"accepted":true,"accepted_predicted":"3","input_tokens":373,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.883683418124517e-05,"brier_hard":9.159451740442322e-09,"score_mae":0.233260637675603,"within_one":true,"case_id":"typed:165/r0","base_id":"typed:165","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.608603}
{"id":"needs_human","type":"binary","probabilities":{"false":7.818097588939947e-09,"true":0.9999999921819018},"predicted":"true","confidence":0.9999999921819018,"entropy_confidence":0.9999997781753158,"candidate_mass":0.9999999430098806,"accepted":true,"accepted_predicted":"true","input_tokens":341,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.818098199416033e-09,"brier_hard":1.2224530888803788e-16,"soft_accuracy":0.7666669958303424,"brier_soft":0.10888857048111922,"tvd":0.23333299218190212,"kl":1.9453708403785763,"case_id":"typed:165/r0","base_id":"typed:165","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.608603}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.7046008148798072e-05,"1":6.427267313172687e-06,"2":0.8247697090599548,"3":0.17520681766458251},"predicted":"2","confidence":0.8247697090599548,"entropy_confidence":0.6650567562445597,"candidate_mass":0.9999947152960812,"accepted":true,"accepted_predicted":"2","input_tokens":367,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":1.7417881876263301,"brier_hard":1.360528866941704,"score_mae":0.22483370161902938,"within_one":true,"case_id":"typed:165/r0","base_id":"typed:165","repeat":0,"workflow":"customer_service","case_elapsed_ms":340.608603}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.4088288086687021e-05,"close_no_action":3.94257988946908e-05,"escalate_to_human":0.00015543260602432266,"execute_refund":3.8746544949603954e-05,"request_information":0.9997523067620443},"predicted":"request_information","confidence":0.9997523067620443,"entropy_confidence":0.9984084554528952,"candidate_mass":0.9999937980868492,"accepted":true,"accepted_predicted":"request_information","input_tokens":471,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":8.769298322063523,"brier_hard":1.9991938370774431,"soft_accuracy":0.21668236778378677,"brier_soft":0.8761397133891152,"tvd":0.7830855234288279,"kl":6.184830547811621,"case_id":"typed:166/r0","base_id":"typed:166","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.632303}
{"id":"category","type":"choice","probabilities":{"account":0.9848566434228677,"billing":0.01498089386968539,"delivery":6.278927144789582e-05,"refund":3.5359779744542193e-05,"technical":6.431365625584149e-05},"predicted":"account","confidence":0.9848566434228677,"entropy_confidence":0.9505707874697327,"candidate_mass":0.9999964511110077,"accepted":true,"accepted_predicted":"account","input_tokens":461,"reused_prefix_tokens":0,"expected":"billing","correct":false,"accepted_correct":false,"nll_hard":4.200979631799775,"brier_hard":1.9402052568649746,"soft_accuracy":0.22313156773943485,"brier_soft":1.0071031693934722,"tvd":0.7681898600896504,"kl":3.085499272576764,"case_id":"typed:166/r0","base_id":"typed:166","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.632303}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00016592845035343876,"1":0.47787397580544333,"2":0.5184839595972319,"3":0.0034761361469696542},"predicted":"2","confidence":0.5184839595972319,"entropy_confidence":0.48455692512053317,"candidate_mass":0.9999963131517903,"accepted":false,"accepted_predicted":null,"input_tokens":447,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.6568461879584325,"brier_hard":0.4602333449720245,"score_mae":0.1747296965591838,"within_one":true,"case_id":"typed:166/r0","base_id":"typed:166","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.632303}
{"id":"needs_human","type":"binary","probabilities":{"false":4.250604160420675e-07,"true":0.9999995749395845},"predicted":"true","confidence":0.9999995749395845,"entropy_confidence":0.9999903900116022,"candidate_mass":0.9999996666603068,"accepted":true,"accepted_predicted":"true","input_tokens":415,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.671034522732015,"brier_hard":1.999998299758698,"soft_accuracy":0.5000000000000002,"brier_soft":0.4999991498795298,"tvd":0.49999957493958425,"kl":4.258596808238372,"case_id":"typed:166/r0","base_id":"typed:166","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.632303}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010007725029240982,"1":0.007647098400574602,"2":0.9577105099695041,"3":0.03364161912699846},"predicted":"2","confidence":0.9577105099695041,"entropy_confidence":0.8559653661885618,"candidate_mass":0.9999975729559197,"accepted":true,"accepted_predicted":"2","input_tokens":441,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.391991314030939,"brier_hard":1.8511174208490875,"score_mae":0.34267402427942173,"within_one":true,"case_id":"typed:166/r0","base_id":"typed:166","repeat":0,"workflow":"customer_service","case_elapsed_ms":351.632303}
{"id":"action","type":"choice","probabilities":{"answer_directly":7.222359303513104e-08,"close_no_action":1.1215996224748703e-07,"escalate_to_human":3.5132524684210673e-07,"execute_refund":1.6234598273315917e-07,"request_information":0.9999993019452148},"predicted":"request_information","confidence":0.9999993019452148,"entropy_confidence":0.9999928920042701,"candidate_mass":0.9999998956484842,"accepted":true,"accepted_predicted":"request_information","input_tokens":479,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":6.980550288379321e-07,"brier_hard":6.5486223483776e-13,"soft_accuracy":0.44999977096597,"brier_soft":0.44879906195914443,"tvd":0.549999301945215,"kl":4.7063590554136665,"case_id":"typed:167/r0","base_id":"typed:167","repeat":0,"workflow":"customer_service","case_elapsed_ms":371.046104}
{"id":"category","type":"choice","probabilities":{"account":1.3118311930820795e-06,"billing":7.542025773353845e-07,"delivery":0.9999932410717894,"refund":2.3822184618743012e-06,"technical":2.3106759782991357e-06},"predicted":"delivery","confidence":0.9999932410717894,"entropy_confidence":0.9999403577383679,"candidate_mass":0.9999993846656575,"accepted":true,"accepted_predicted":"delivery","input_tokens":469,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":6.758951052217769e-06,"brier_hard":5.898702143908387e-11,"soft_accuracy":0.9066610267580439,"brier_soft":0.011042804242477994,"tvd":0.09332624107178943,"kl":0.7634679658980866,"case_id":"typed:167/r0","base_id":"typed:167","repeat":0,"workflow":"customer_service","case_elapsed_ms":371.046104}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.899963049344633e-06,"1":0.9999304513467064,"2":6.245748427872255e-05,"3":1.1912059657941341e-06},"predicted":"1","confidence":0.9999304513467064,"entropy_confidence":0.9994507037798885,"candidate_mass":0.9999995692528469,"accepted":true,"accepted_predicted":"1","input_tokens":455,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":6.955107191329003e-05,"brier_hard":8.774181053011368e-09,"score_mae":0.2499410600668388,"within_one":true,"case_id":"typed:167/r0","base_id":"typed:167","repeat":0,"workflow":"customer_service","case_elapsed_ms":371.046104}
{"id":"needs_human","type":"binary","probabilities":{"false":6.829169568057409e-09,"true":0.9999999931708299},"predicted":"true","confidence":0.9999999931708299,"entropy_confidence":0.9999998049019766,"candidate_mass":0.9999999704476061,"accepted":true,"accepted_predicted":"true","input_tokens":423,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.829170141613097e-09,"brier_hard":9.32751214938854e-17,"soft_accuracy":0.4999999999999997,"brier_soft":0.49999998634166043,"tvd":0.49999999317083016,"kl":4.258596599122704,"case_id":"typed:167/r0","base_id":"typed:167","repeat":0,"workflow":"customer_service","case_elapsed_ms":371.046104}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022724109919432546,"1":0.96149316517475,"2":0.03818780541130219,"3":9.178831475529763e-05},"predicted":"1","confidence":0.96149316517475,"entropy_confidence":0.8808277907619512,"candidate_mass":0.999996658123755,"accepted":true,"accepted_predicted":"1","input_tokens":449,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.2652390444049963,"brier_hard":1.8495518644008981,"score_mae":0.5951888590583796,"within_one":true,"case_id":"typed:167/r0","base_id":"typed:167","repeat":0,"workflow":"customer_service","case_elapsed_ms":371.046104}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.4965135764463439,"close_no_action":0.0005591992140068154,"escalate_to_human":0.0012018837121606089,"execute_refund":0.0001594258786942752,"request_information":0.5015659147487944},"predicted":"request_information","confidence":0.5015659147487944,"entropy_confidence":0.5604769750985544,"candidate_mass":0.9998740567634922,"accepted":false,"accepted_predicted":null,"input_tokens":403,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":null,"nll_hard":0.7001444515679509,"brier_hard":0.5050687281854758,"soft_accuracy":0.4142940842329683,"brier_soft":0.28737370354662733,"tvd":0.435058407294422,"kl":0.9976634825186044,"case_id":"typed:168/r0","base_id":"typed:168","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.519603}
{"id":"category","type":"choice","probabilities":{"account":2.01428077207817e-11,"billing":1.2416182018305646e-11,"delivery":0.9999999999380815,"refund":1.9763159589288276e-11,"technical":9.597144847688013e-12},"predicted":"delivery","confidence":0.9999999999380815,"entropy_confidence":0.9999999990056303,"candidate_mass":0.9999999998219984,"accepted":true,"accepted_predicted":"delivery","input_tokens":393,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":6.191847035289148e-11,"brier_hard":4.876478915571773e-21,"soft_accuracy":0.9699999999403851,"brier_soft":0.0011555577733927958,"tvd":0.029999999938081133,"kl":0.24676477995917914,"case_id":"typed:168/r0","base_id":"typed:168","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.519603}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.5058194017019422e-05,"1":0.9994049266290154,"2":0.000569593341263872,"3":1.0421835703689787e-05},"predicted":"1","confidence":0.9994049266290154,"entropy_confidence":0.9962945470069493,"candidate_mass":0.9999976970268294,"accepted":true,"accepted_predicted":"1","input_tokens":379,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.470587888131693,"brier_hard":1.997671345459758,"score_mae":0.6494246211813457,"within_one":true,"case_id":"typed:168/r0","base_id":"typed:168","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.519603}
{"id":"needs_human","type":"binary","probabilities":{"false":9.221636826631665e-09,"true":0.9999999907783632},"predicted":"true","confidence":0.9999999907783632,"entropy_confidence":0.9999997405490171,"candidate_mass":0.9999999580705508,"accepted":true,"accepted_predicted":"true","input_tokens":347,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.50171328513106,"brier_hard":1.999999963113453,"soft_accuracy":0.2666670043034244,"brier_soft":1.0755545507278779,"tvd":0.7333329907783632,"kl":6.401778207214494,"case_id":"typed:168/r0","base_id":"typed:168","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.519603}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.77971941612976e-06,"1":0.0001916127026709722,"2":0.999760057995123,"3":3.854958279125711e-05},"predicted":"2","confidence":0.999760057995123,"entropy_confidence":0.998279776050866,"candidate_mass":0.9999993634703231,"accepted":true,"accepted_predicted":"2","input_tokens":373,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00023997079556532198,"brier_hard":9.586930677449763e-08,"score_mae":0.4298273774412906,"within_one":true,"case_id":"typed:168/r0","base_id":"typed:168","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.519603}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0002585256110558784,"close_no_action":0.00024292534261698349,"escalate_to_human":0.0031643775161125003,"execute_refund":0.004016501208857621,"request_information":0.9923176703213573},"predicted":"request_information","confidence":0.9923176703213573,"entropy_confidence":0.967576227944225,"candidate_mass":0.9999371510475722,"accepted":true,"accepted_predicted":"request_information","input_tokens":319,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.755798920103146,"brier_hard":1.9783918752150198,"soft_accuracy":0.10162329800722784,"brier_soft":1.2203403631666923,"tvd":0.8923177703212571,"kl":4.513059556664772,"case_id":"typed:169/r0","base_id":"typed:169","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.305204}
{"id":"category","type":"choice","probabilities":{"account":1.2858502144812926e-11,"billing":0.9999999989025472,"delivery":7.401438189416604e-10,"refund":3.118070371426023e-10,"technical":3.264291204721354e-11},"predicted":"billing","confidence":0.9999999989025472,"entropy_confidence":0.9999999847188719,"candidate_mass":0.9999999989897468,"accepted":true,"accepted_predicted":"billing","input_tokens":309,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.0974527959089095e-09,"brier_hard":1.850670039840031e-18,"soft_accuracy":0.8833338824003442,"brier_soft":0.02502218082850996,"tvd":0.11666611556866414,"kl":0.9649552874542499,"case_id":"typed:169/r0","base_id":"typed:169","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.305204}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":7.551519745120725e-06,"1":2.671383995163013e-06,"2":0.9928393536438965,"3":0.007150423452363288},"predicted":"2","confidence":0.9928393536438965,"entropy_confidence":0.9692810151073742,"candidate_mass":0.999998922193063,"accepted":true,"accepted_predicted":"2","input_tokens":295,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0071864068323251075,"brier_hard":0.00010240347594702778,"score_mae":0.04046564902887817,"within_one":true,"case_id":"typed:169/r0","base_id":"typed:169","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.305204}
{"id":"needs_human","type":"binary","probabilities":{"false":9.22509368857622e-09,"true":0.9999999907749064},"predicted":"true","confidence":0.9999999907749064,"entropy_confidence":0.9999997404567463,"candidate_mass":0.9999999618934589,"accepted":true,"accepted_predicted":"true","input_tokens":263,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.225093620475958e-09,"brier_hard":1.7020470508447786e-16,"soft_accuracy":0.7833329947724531,"brier_soft":0.09388916978290676,"tvd":0.21666699077490642,"kl":1.8042889535992426,"case_id":"typed:169/r0","base_id":"typed:169","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.305204}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.235882659620309e-05,"1":3.8376103928397904e-05,"2":0.48267271343189866,"3":0.5171965516375766},"predicted":"3","confidence":0.5171965516375766,"entropy_confidence":0.4995007030245553,"candidate_mass":0.9999948758378934,"accepted":false,"accepted_predicted":null,"input_tokens":289,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":0.7284164669545383,"brier_hard":0.5351198044565931,"score_mae":0.26697345788045546,"within_one":true,"case_id":"typed:169/r0","base_id":"typed:169","repeat":0,"workflow":"customer_service","case_elapsed_ms":316.305204}
{"id":"action","type":"choice","probabilities":{"answer_directly":3.41487929812829e-05,"close_no_action":6.127965235425949e-05,"escalate_to_human":0.0001887468773876053,"execute_refund":4.18819734384733e-05,"request_information":0.9996739427038381},"predicted":"request_information","confidence":0.9996739427038381,"entropy_confidence":0.997941917925988,"candidate_mass":0.999988635887673,"accepted":true,"accepted_predicted":"request_information","input_tokens":320,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.000326110464399629,"brier_hard":1.4861417965779528e-07,"soft_accuracy":0.5831700701042112,"brier_soft":0.2908081960805802,"tvd":0.4163822413436934,"kl":3.24510258535733,"case_id":"typed:170/r0","base_id":"typed:170","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.45110600000004}
{"id":"category","type":"choice","probabilities":{"account":3.502403795152509e-08,"billing":4.689146958379605e-07,"delivery":3.9624898037257287e-07,"refund":2.800074282482752e-07,"technical":0.9999988198048585},"predicted":"technical","confidence":0.9999988198048585,"entropy_confidence":0.9999883928706788,"candidate_mass":0.9999998621228261,"accepted":true,"accepted_predicted":"technical","input_tokens":310,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":1.1801958379344156e-06,"brier_hard":1.8493856615564127e-12,"soft_accuracy":0.9866668261408168,"brier_soft":0.0002221388902622055,"tvd":0.013330833136871405,"kl":0.10783638489969216,"case_id":"typed:170/r0","base_id":"typed:170","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.45110600000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.3668808584344774e-05,"1":0.9998608336582995,"2":8.997025920084439e-05,"3":5.5272739143282545e-06},"predicted":"1","confidence":0.9998608336582995,"entropy_confidence":0.9989305154852262,"candidate_mass":0.999998329657006,"accepted":true,"accepted_predicted":"1","input_tokens":296,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00013917602623437413,"brier_hard":2.9399433803074597e-08,"score_mae":0.26660964400155573,"within_one":true,"case_id":"typed:170/r0","base_id":"typed:170","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.45110600000004}
{"id":"needs_human","type":"binary","probabilities":{"false":8.609464998707071e-07,"true":0.9999991390535012},"predicted":"true","confidence":0.9999991390535012,"entropy_confidence":0.9999814119347712,"candidate_mass":0.9999992183930501,"accepted":true,"accepted_predicted":"true","input_tokens":264,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.965233471649384,"brier_hard":1.999996556215485,"soft_accuracy":0.2500004304732502,"brier_soft":1.1249974171619845,"tvd":0.7499991390535007,"kl":6.561181903938882,"case_id":"typed:170/r0","base_id":"typed:170","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.45110600000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.956884353397727e-05,"1":0.9997694453901449,"2":0.0001297318916216487,"3":1.1253874700209057e-05},"predicted":"1","confidence":0.9997694453901449,"entropy_confidence":0.9983014458534465,"candidate_mass":0.999998258121437,"accepted":true,"accepted_predicted":"1","input_tokens":290,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00023058119165495034,"brier_hard":7.813501925695289e-08,"score_mae":0.31660432920251114,"within_one":true,"case_id":"typed:170/r0","base_id":"typed:170","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.45110600000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":6.7519023875174244e-06,"close_no_action":1.372499681476536e-05,"escalate_to_human":6.419787224696458e-05,"execute_refund":1.1520828553804982e-05,"request_information":0.9999038043999954},"predicted":"request_information","confidence":0.9999038043999954,"entropy_confidence":0.9993283429921149,"candidate_mass":0.9999964466253531,"accepted":true,"accepted_predicted":"request_information","input_tokens":409,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":9.620022709809254e-05,"brier_hard":1.3741653475268999e-08,"soft_accuracy":0.62661476072002,"brier_soft":0.22113477465947684,"tvd":0.3732361777323695,"kl":2.9326752358332695,"case_id":"typed:171/r0","base_id":"typed:171","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.719806}
{"id":"category","type":"choice","probabilities":{"account":0.9999999996694182,"billing":2.407485000434917e-10,"delivery":7.309856222776145e-11,"refund":8.656995536341839e-12,"technical":8.07733754681249e-12},"predicted":"account","confidence":0.9999999996694182,"entropy_confidence":0.9999999951564494,"candidate_mass":0.9999999968064817,"accepted":true,"accepted_predicted":"account","input_tokens":399,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":3.3058178422386817e-10,"brier_hard":1.7272774305102716e-19,"soft_accuracy":0.896666103041076,"brier_soft":0.016800159658141783,"tvd":0.10333389633552172,"kl":0.8539394434324995,"case_id":"typed:171/r0","base_id":"typed:171","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.719806}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.6191997038676526e-05,"1":0.9992604979875298,"2":0.0006552744135939858,"3":2.8035601838776974e-05},"predicted":"1","confidence":0.9992604979875298,"entropy_confidence":0.9953931173121854,"candidate_mass":0.9999971080221715,"accepted":true,"accepted_predicted":"1","input_tokens":385,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0007397755789603796,"brier_hard":9.801913190600668e-07,"score_mae":0.2660118463797658,"within_one":true,"case_id":"typed:171/r0","base_id":"typed:171","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.719806}
{"id":"needs_human","type":"binary","probabilities":{"false":7.529179976743316e-09,"true":0.9999999924708192},"predicted":"true","confidence":0.9999999924708192,"entropy_confidence":0.9999997859638198,"candidate_mass":0.9999999620477247,"accepted":true,"accepted_predicted":"true","input_tokens":353,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.704479701895636,"brier_hard":1.9999999698832787,"soft_accuracy":0.18333300476848552,"brier_soft":1.3338899531826676,"tvd":0.8166669924708196,"kl":7.210765571030071,"case_id":"typed:171/r0","base_id":"typed:171","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.719806}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00032941266325642386,"1":0.9670677761932787,"2":0.03237089401735471,"3":0.00023191712610924986},"predicted":"1","confidence":0.9670677761932787,"entropy_confidence":0.8932302114644838,"candidate_mass":0.9999950262355264,"accepted":true,"accepted_predicted":"1","input_tokens":379,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.43049559272427,"brier_hard":1.8715263327944427,"score_mae":0.9741616843936844,"within_one":true,"case_id":"typed:171/r0","base_id":"typed:171","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.719806}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0002419638977023594,"close_no_action":0.0005712526912274488,"escalate_to_human":0.01907802926848999,"execute_refund":0.0003281748691204329,"request_information":0.9797805792734604},"predicted":"request_information","confidence":0.9797805792734604,"entropy_confidence":0.9350947337055063,"candidate_mass":0.9999257176446021,"accepted":true,"accepted_predicted":"request_information","input_tokens":387,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.9592179061115402,"brier_hard":1.9221783887601367,"soft_accuracy":0.2865787247097059,"brier_soft":0.7332883534337048,"tvd":0.6967757541425805,"kl":2.7487426171340266,"case_id":"typed:172/r0","base_id":"typed:172","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.277806}
{"id":"category","type":"choice","probabilities":{"account":1.325151754010985e-07,"billing":0.9999983775101331,"delivery":1.1646842945402675e-06,"refund":2.2223963311363207e-07,"technical":1.0305076355014179e-07},"predicted":"billing","confidence":0.9999983775101331,"entropy_confidence":0.9999846550690529,"candidate_mass":0.999999782177645,"accepted":true,"accepted_predicted":"billing","input_tokens":377,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.622491183141828e-06,"brier_hard":4.066533060260345e-12,"soft_accuracy":0.9766654220024977,"brier_soft":0.0007110065753371154,"tvd":0.02333137751013327,"kl":0.18767768910241414,"case_id":"typed:172/r0","base_id":"typed:172","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.277806}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.4772762236032945e-07,"1":1.6801340068030301e-07,"2":6.158271631639618e-05,"3":0.999938101542659},"predicted":"3","confidence":0.999938101542659,"entropy_confidence":0.9995211022414148,"candidate_mass":0.999999769101505,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.190037312951481e-05,"brier_hard":7.62390002204805e-09,"score_mae":0.4566046380740105,"within_one":true,"case_id":"typed:172/r0","base_id":"typed:172","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.277806}
{"id":"needs_human","type":"binary","probabilities":{"false":6.495069741747779e-09,"true":0.9999999935049306},"predicted":"true","confidence":0.9999999935049306,"entropy_confidence":0.999999813976638,"candidate_mass":0.9999999642873696,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.4950693923192e-09,"brier_hard":8.437185708720875e-17,"soft_accuracy":0.7766669964060574,"brier_soft":0.09975525197574661,"tvd":0.22333299350493047,"kl":1.8606753440441721,"case_id":"typed:172/r0","base_id":"typed:172","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.277806}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.09997739513518e-07,"1":7.005194076013203e-07,"2":2.2534875362647634e-05,"3":0.999975854607489},"predicted":"3","confidence":0.999975854607489,"entropy_confidence":0.999792350300699,"candidate_mass":0.9999996195409309,"accepted":true,"accepted_predicted":"3","input_tokens":357,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.4145684015736728e-05,"brier_hard":1.0921394104493021e-09,"score_mae":0.4333063340926002,"within_one":true,"case_id":"typed:172/r0","base_id":"typed:172","repeat":0,"workflow":"customer_service","case_elapsed_ms":331.277806}
{"id":"action","type":"choice","probabilities":{"answer_directly":3.0501876646352686e-05,"close_no_action":6.682380981180887e-05,"escalate_to_human":0.000246664395393802,"execute_refund":5.4668889246435255e-05,"request_information":0.9996013410289002},"predicted":"request_information","confidence":0.9996013410289002,"entropy_confidence":0.9975495611199028,"candidate_mass":0.9999928763197197,"accepted":true,"accepted_predicted":"request_information","input_tokens":475,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":10.39772234685515,"brier_hard":1.9991419064612803,"soft_accuracy":0.3998824636899651,"brier_soft":0.5893042697686792,"tvd":0.5996564099177474,"kl":4.73892056318754,"case_id":"typed:173/r0","base_id":"typed:173","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.514406}
{"id":"category","type":"choice","probabilities":{"account":0.7875006804544403,"billing":0.0010358536328367101,"delivery":0.00021935055150170554,"refund":0.00018152709694592304,"technical":0.21106258826427526},"predicted":"account","confidence":0.7875006804544403,"entropy_confidence":0.6725648388235103,"candidate_mass":0.9999674860293696,"accepted":false,"accepted_predicted":null,"input_tokens":465,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":null,"nll_hard":1.5556005626462088,"brier_hard":1.242580715411971,"soft_accuracy":0.20919597179179672,"brier_soft":1.1683491090933276,"tvd":0.779167672121432,"kl":1.528924272259162,"case_id":"typed:173/r0","base_id":"typed:173","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.514406}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00020604429450650142,"1":0.0009701562731548142,"2":0.017699952890937967,"3":0.9811238465414008},"predicted":"3","confidence":0.9811238465414008,"entropy_confidence":0.9288883503224894,"candidate_mass":0.999998904314303,"accepted":true,"accepted_predicted":"3","input_tokens":451,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.034193300934824,"brier_hard":1.9275183684601054,"score_mae":1.2297416016792333,"within_one":false,"case_id":"typed:173/r0","base_id":"typed:173","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.514406}
{"id":"needs_human","type":"binary","probabilities":{"false":4.817003641320563e-09,"true":0.9999999951829963},"predicted":"true","confidence":0.9999999951829963,"entropy_confidence":0.9999998599604946,"candidate_mass":0.9999999752625616,"accepted":true,"accepted_predicted":"true","input_tokens":419,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":19.15111375336742,"brier_hard":1.9999999807319853,"soft_accuracy":0.48333300016057,"brier_soft":0.5338895678228527,"tvd":0.5166669951829963,"kl":4.407271953065175,"case_id":"typed:173/r0","base_id":"typed:173","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.514406}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.5141145839113155e-06,"1":2.083897617073473e-06,"2":0.7053592677714634,"3":0.294633134216336},"predicted":"2","confidence":0.7053592677714634,"entropy_confidence":0.5626130051903676,"candidate_mass":0.999994291596572,"accepted":false,"accepted_predicted":null,"input_tokens":445,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.3490480062958525,"brier_hard":0.17362184490105778,"score_mae":0.04462002208955207,"within_one":true,"case_id":"typed:173/r0","base_id":"typed:173","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.514406}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.07142496645571483,"close_no_action":0.00393361033800702,"escalate_to_human":0.016183683483715357,"execute_refund":0.054489138225797595,"request_information":0.8539686014967658},"predicted":"request_information","confidence":0.8539686014967658,"entropy_confidence":0.645603611392229,"candidate_mass":0.9994577075641634,"accepted":true,"accepted_predicted":"request_information","input_tokens":454,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":4.123751736666112,"brier_hard":1.7052429822940938,"soft_accuracy":0.05329653797279789,"brier_soft":1.2073500873167244,"tvd":0.8717907763895265,"kl":2.923069857962146,"case_id":"typed:174/r0","base_id":"typed:174","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.85260600000004}
{"id":"category","type":"choice","probabilities":{"account":2.3956223391977884e-11,"billing":1.803729151786705e-11,"delivery":0.99999999990467,"refund":4.018897247322874e-11,"technical":1.3147950683979541e-11},"predicted":"delivery","confidence":0.99999999990467,"entropy_confidence":0.999999998497096,"candidate_mass":0.9999999996576854,"accepted":true,"accepted_predicted":"delivery","input_tokens":444,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":9.532996614779774e-11,"brier_hard":1.1775069085023086e-20,"soft_accuracy":0.9600009599103819,"brier_soft":0.002199911058316192,"tvd":0.03999903990370979,"kl":0.3292165791078064,"case_id":"typed:174/r0","base_id":"typed:174","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.85260600000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.45318845220785e-05,"1":7.001585343649648e-05,"2":0.0005820524439953963,"3":0.9993133998180448},"predicted":"3","confidence":0.9993133998180448,"entropy_confidence":0.9956382722549266,"candidate_mass":0.9999980239520154,"accepted":true,"accepted_predicted":"3","input_tokens":430,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.448950004334769,"brier_hard":1.9974635110476269,"score_mae":1.3991743201955615,"within_one":false,"case_id":"typed:174/r0","base_id":"typed:174","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.85260600000004}
{"id":"needs_human","type":"binary","probabilities":{"false":1.0580363517606102e-07,"true":0.9999998941963659},"predicted":"true","confidence":0.9999998941963659,"entropy_confidence":0.999997395664425,"candidate_mass":0.9999996234014928,"accepted":true,"accepted_predicted":"true","input_tokens":398,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.0580363967340075e-07,"brier_hard":2.2388818200198573e-14,"soft_accuracy":0.7966669372231068,"brier_soft":0.0826885317245406,"tvd":0.20333289419636535,"kl":1.6916690722683936,"case_id":"typed:174/r0","base_id":"typed:174","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.85260600000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.445983942941569e-06,"1":4.7421813104759184e-06,"2":0.0003541157318944322,"3":0.9996336961028518},"predicted":"3","confidence":0.9996336961028518,"entropy_confidence":0.9976007579472124,"candidate_mass":0.9999990297427015,"accepted":true,"accepted_predicted":"3","input_tokens":424,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.945886772003785,"brier_hard":1.9985594203963422,"score_mae":0.7162810619536546,"within_one":true,"case_id":"typed:174/r0","base_id":"typed:174","repeat":0,"workflow":"customer_service","case_elapsed_ms":347.85260600000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0004302225090216244,"close_no_action":0.00029768090580765266,"escalate_to_human":0.00893269337166095,"execute_refund":0.0003338895275224835,"request_information":0.9900055136859861},"predicted":"request_information","confidence":0.9900055136859861,"entropy_confidence":0.9624007854558644,"candidate_mass":0.9999500200651216,"accepted":true,"accepted_predicted":"request_information","input_tokens":251,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":4.718037320200164,"brier_hard":1.962325708583749,"soft_accuracy":0.11622390371606982,"brier_soft":1.4293425113628027,"tvd":0.8800056236858766,"kl":3.7657578901823103,"case_id":"typed:175/r0","base_id":"typed:175","repeat":0,"workflow":"customer_service","case_elapsed_ms":224.761704}
{"id":"category","type":"choice","probabilities":{"account":9.459091039043536e-06,"billing":2.6947727675588085e-05,"delivery":0.0007459600915494804,"refund":1.7812311637875517e-05,"technical":0.9991998207780981},"predicted":"technical","confidence":0.9991998207780981,"entropy_confidence":0.9958003057587064,"candidate_mass":0.999996668400549,"accepted":true,"accepted_predicted":"technical","input_tokens":241,"reused_prefix_tokens":0,"expected":"delivery","correct":false,"accepted_correct":false,"nll_hard":7.200838455773972,"brier_hard":1.9969089192492755,"soft_accuracy":0.0672856453365263,"brier_soft":1.678874024315322,"tvd":0.9325328207780981,"kl":6.429213229123883,"case_id":"typed:175/r0","base_id":"typed:175","repeat":0,"workflow":"customer_service","case_elapsed_ms":224.761704}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.5805525640733183e-08,"1":8.655908780469e-09,"2":5.641235812283433e-07,"3":0.9999994114149855},"predicted":"3","confidence":0.9999994114149855,"entropy_confidence":0.9999933998137217,"candidate_mass":0.9999999552709592,"accepted":true,"accepted_predicted":"3","input_tokens":227,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.885851877210997e-07,"brier_hard":6.649924735951874e-13,"score_mae":0.13666637114802782,"within_one":true,"case_id":"typed:175/r0","base_id":"typed:175","repeat":0,"workflow":"customer_service","case_elapsed_ms":224.761704}
{"id":"needs_human","type":"binary","probabilities":{"false":3.495182654313048e-09,"true":0.9999999965048172},"predicted":"true","confidence":0.9999999965048172,"entropy_confidence":0.9999998967708984,"candidate_mass":0.9999999678254353,"accepted":true,"accepted_predicted":"true","input_tokens":195,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.4951828159509965e-09,"brier_hard":2.443260466123173e-17,"soft_accuracy":0.9199999970640466,"brier_soft":0.012799998881541538,"tvd":0.07999999650481723,"kl":0.6601161527497356,"case_id":"typed:175/r0","base_id":"typed:175","repeat":0,"workflow":"customer_service","case_elapsed_ms":224.761704}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.2029748842905155e-06,"1":8.341616883938725e-07,"2":1.0850144890325952e-05,"3":0.9999871127185364},"predicted":"3","confidence":0.9999871127185364,"entropy_confidence":0.9998809835452106,"candidate_mass":0.9999994594045387,"accepted":true,"accepted_predicted":"3","input_tokens":221,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.2887364505372367e-05,"brier_hard":2.8595064195894767e-10,"score_mae":0.19665087260707814,"within_one":true,"case_id":"typed:175/r0","base_id":"typed:175","repeat":0,"workflow":"customer_service","case_elapsed_ms":224.761704}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.018567222444647234,"close_no_action":0.0027965931875577822,"escalate_to_human":0.0053939376237886185,"execute_refund":0.000366588389669913,"request_information":0.9728756583543369},"predicted":"request_information","confidence":0.9728756583543369,"entropy_confidence":0.9078679838047822,"candidate_mass":0.9998662535446342,"accepted":true,"accepted_predicted":"request_information","input_tokens":431,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.222479618330606,"brier_hard":1.9360809630037097,"soft_accuracy":0.04309730864143437,"brier_soft":1.2695638831923335,"tvd":0.9395426250213033,"kl":3.5670263992920597,"case_id":"typed:176/r0","base_id":"typed:176","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.16130699999997}
{"id":"category","type":"choice","probabilities":{"account":1.3334122755505137e-12,"billing":8.766967846685003e-13,"delivery":0.9999999999954845,"refund":1.7309747517293547e-12,"technical":5.749435637233673e-13},"predicted":"delivery","confidence":0.9999999999954845,"entropy_confidence":0.9999999999202218,"candidate_mass":0.9999999999453664,"accepted":true,"accepted_predicted":"delivery","input_tokens":421,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":4.515499085765632e-12,"brier_hard":2.6263151234886497e-23,"soft_accuracy":0.9866679866635464,"brier_soft":0.0002221782242057229,"tvd":0.013332013327497599,"kl":0.10954963619210865,"case_id":"typed:176/r0","base_id":"typed:176","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.16130699999997}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.126159099072727e-05,"1":3.17125489496473e-05,"2":0.011781799363109233,"3":0.9881752264969489},"predicted":"3","confidence":0.9881752264969489,"entropy_confidence":0.9534465947130053,"candidate_mass":0.9999986347932979,"accepted":true,"accepted_predicted":"3","input_tokens":407,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.441199365130572,"brier_hard":1.9530654914648196,"score_mae":0.7481209907660147,"within_one":true,"case_id":"typed:176/r0","base_id":"typed:176","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.16130699999997}
{"id":"needs_human","type":"binary","probabilities":{"false":3.205004747672359e-08,"true":0.9999999679499544},"predicted":"true","confidence":0.9999999679499544,"entropy_confidence":0.9999991558724776,"candidate_mass":0.9999998896859087,"accepted":true,"accepted_predicted":"true","input_tokens":375,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.2050046131062666e-08,"brier_hard":2.0544109673414993e-15,"soft_accuracy":0.6399999910259879,"brier_soft":0.25919995384793504,"tvd":0.3599999679499535,"kl":3.030098808741267,"case_id":"typed:176/r0","base_id":"typed:176","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.16130699999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.8645189672305334e-05,"1":7.033624818044324e-05,"2":0.019598498727429314,"3":0.9802825198347164},"predicted":"3","confidence":0.9802825198347164,"entropy_confidence":0.9294921128635965,"candidate_mass":0.9999976943486223,"accepted":true,"accepted_predicted":"3","input_tokens":401,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.019914463318183803,"brier_hard":0.0007728874899796881,"score_mae":0.6034478932071883,"within_one":true,"case_id":"typed:176/r0","base_id":"typed:176","repeat":0,"workflow":"customer_service","case_elapsed_ms":345.16130699999997}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00035859317471870146,"close_no_action":0.00017374698533290727,"escalate_to_human":0.502047148622015,"execute_refund":5.5973759531576055e-05,"request_information":0.4973645374584011},"predicted":"escalate_to_human","confidence":0.502047148622015,"entropy_confidence":0.5661754348251435,"candidate_mass":0.9998877764683648,"accepted":false,"accepted_predicted":null,"input_tokens":449,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":null,"nll_hard":0.6890612421429623,"brier_hard":0.49532868722681644,"soft_accuracy":0.44450701602956244,"brier_soft":0.1891195279677218,"tvd":0.35403153745840144,"kl":0.71122736116442,"case_id":"typed:177/r0","base_id":"typed:177","repeat":0,"workflow":"customer_service","case_elapsed_ms":350.91900599999997}
{"id":"category","type":"choice","probabilities":{"account":0.9999999563684617,"billing":3.7915903381475203e-08,"delivery":4.822895224996788e-09,"refund":5.910988898277247e-10,"technical":3.0164130470780727e-10},"predicted":"account","confidence":0.9999999563684617,"entropy_confidence":0.9999995010275127,"candidate_mass":0.999999974378721,"accepted":true,"accepted_predicted":"account","input_tokens":439,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":4.3631539223104917e-08,"brier_hard":3.365027564874416e-15,"soft_accuracy":0.7266669774222364,"brier_soft":0.13282199567045416,"tvd":0.27333295636846144,"kl":2.2854747042131374,"case_id":"typed:177/r0","base_id":"typed:177","repeat":0,"workflow":"customer_service","case_elapsed_ms":350.91900599999997}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.5055113394702313e-07,"1":4.2638620182724047e-07,"2":1.9938546609723597e-05,"3":0.9999792845160538},"predicted":"3","confidence":0.9999792845160538,"entropy_confidence":0.9998211260210377,"candidate_mass":0.9999997458296729,"accepted":true,"accepted_predicted":"3","input_tokens":425,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.0715698514821743e-05,"brier_hard":8.269816073248769e-10,"score_mae":0.31664515702758234,"within_one":true,"case_id":"typed:177/r0","base_id":"typed:177","repeat":0,"workflow":"customer_service","case_elapsed_ms":350.91900599999997}
{"id":"needs_human","type":"binary","probabilities":{"false":3.3662566183969874e-09,"true":0.9999999966337434},"predicted":"true","confidence":0.9999999966337434,"entropy_confidence":0.9999999003961604,"candidate_mass":0.9999999741039133,"accepted":true,"accepted_predicted":"true","input_tokens":393,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.366256612461856e-09,"brier_hard":2.2663367163699338e-17,"soft_accuracy":0.8299999977782705,"brier_soft":0.05779999771094556,"tvd":0.16999999663374343,"kl":1.411104316131005,"case_id":"typed:177/r0","base_id":"typed:177","repeat":0,"workflow":"customer_service","case_elapsed_ms":350.91900599999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.32771166040445e-06,"1":3.5301985315578995e-06,"2":6.190184183572648e-05,"3":0.9999302402479735},"predicted":"3","confidence":0.9999302402479735,"entropy_confidence":0.9994464727730359,"candidate_mass":0.9999994914201139,"accepted":true,"accepted_predicted":"3","input_tokens":419,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.976218535116668e-05,"brier_hard":8.729452415341772e-09,"score_mae":0.39325105462612386,"within_one":true,"case_id":"typed:177/r0","base_id":"typed:177","repeat":0,"workflow":"customer_service","case_elapsed_ms":350.91900599999997}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.000488161210894442,"close_no_action":0.00035041560709997687,"escalate_to_human":0.005279320948594833,"execute_refund":0.0012120282272056998,"request_information":0.9926700740062041},"predicted":"request_information","confidence":0.9926700740062041,"entropy_confidence":0.9691587566031354,"candidate_mass":0.9999273456039628,"accepted":true,"accepted_predicted":"request_information","input_tokens":342,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":5.243957797755389,"brier_hard":1.9748649352648604,"soft_accuracy":0.06385752271958797,"brier_soft":1.5117308095008744,"tvd":0.9326700740062046,"kl":4.52496696943921,"case_id":"typed:178/r0","base_id":"typed:178","repeat":0,"workflow":"customer_service","case_elapsed_ms":303.030806}
{"id":"category","type":"choice","probabilities":{"account":4.04316902928498e-05,"billing":0.0005472502596776906,"delivery":0.0006214128720521466,"refund":7.321659220863768e-05,"technical":0.9987176885857678},"predicted":"technical","confidence":0.9987176885857678,"entropy_confidence":0.993111844928682,"candidate_mass":0.9999925358113803,"accepted":true,"accepted_predicted":"technical","input_tokens":332,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.0012831342790348933,"brier_hard":2.3369547582941252e-06,"soft_accuracy":0.8755736404575342,"brier_soft":0.019268768767225643,"tvd":0.12205068858576826,"kl":0.5681150563653679,"case_id":"typed:178/r0","base_id":"typed:178","repeat":0,"workflow":"customer_service","case_elapsed_ms":303.030806}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":4.9955255250647176e-05,"1":6.147942427160515e-05,"2":0.9788659807655059,"3":0.021022584554970622},"predicted":"2","confidence":0.9788659807655059,"entropy_confidence":0.9255622895265162,"candidate_mass":0.999998926567916,"accepted":true,"accepted_predicted":"2","input_tokens":318,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.02136053983114299,"brier_hard":0.0008886021056219917,"score_mae":0.14247180537980508,"within_one":true,"case_id":"typed:178/r0","base_id":"typed:178","repeat":0,"workflow":"customer_service","case_elapsed_ms":303.030806}
{"id":"needs_human","type":"binary","probabilities":{"false":3.540725907450489e-09,"true":0.999999996459275},"predicted":"true","confidence":0.999999996459275,"entropy_confidence":0.9999998954919257,"candidate_mass":0.9999999790355025,"accepted":true,"accepted_predicted":"true","input_tokens":286,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.5407250527597555e-09,"brier_hard":2.5073473806542533e-17,"soft_accuracy":0.8399999975923071,"brier_soft":0.05119999773393574,"tvd":0.15999999645927454,"kl":1.3271976172887854,"case_id":"typed:178/r0","base_id":"typed:178","repeat":0,"workflow":"customer_service","case_elapsed_ms":303.030806}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.330198431658069e-05,"1":1.9790853503703404e-05,"2":0.001057671589254059,"3":0.9988992355729244},"predicted":"3","confidence":0.9988992355729244,"entropy_confidence":0.9936450050517602,"candidate_mass":0.9999986964571411,"accepted":true,"accepted_predicted":"3","input_tokens":312,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00110137071319722,"brier_hard":2.3312861749858397e-06,"score_mae":0.3754998407507846,"within_one":true,"case_id":"typed:178/r0","base_id":"typed:178","repeat":0,"workflow":"customer_service","case_elapsed_ms":303.030806}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.014784646087191789,"close_no_action":0.001732004170714289,"escalate_to_human":0.1511870027528033,"execute_refund":0.005942637954033517,"request_information":0.8263537090352555},"predicted":"request_information","confidence":0.8263537090352555,"entropy_confidence":0.6601185669860915,"candidate_mass":0.9999161201415259,"accepted":true,"accepted_predicted":"request_information","input_tokens":308,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":1.8892377795611388,"brier_hard":1.403600857276317,"soft_accuracy":0.2586887686381999,"brier_soft":0.7197082810615236,"tvd":0.6430207090352563,"kl":0.9957750109198754,"case_id":"typed:179/r0","base_id":"typed:179","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.672905}
{"id":"category","type":"choice","probabilities":{"account":3.3661255918568035e-11,"billing":0.9999999985404102,"delivery":9.61788920982916e-10,"refund":4.1795560131264564e-10,"technical":4.618440421860883e-11},"predicted":"billing","confidence":0.9999999985404102,"entropy_confidence":0.9999999798902872,"candidate_mass":0.9999999987395398,"accepted":true,"accepted_predicted":"billing","input_tokens":298,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.4595897834362414e-09,"brier_hard":3.2333932253390806e-18,"soft_accuracy":0.946666998624925,"brier_soft":0.004511075386970427,"tvd":0.05333299854041,"kl":0.4393302689557359,"case_id":"typed:179/r0","base_id":"typed:179","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.672905}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.1859633206096178e-05,"1":1.0458267202050434e-05,"2":0.18256999166662186,"3":0.8174076904329716},"predicted":"3","confidence":0.8174076904329716,"entropy_confidence":0.6569700632237501,"candidate_mass":0.9999971489688412,"accepted":true,"accepted_predicted":"3","input_tokens":284,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.7006216635015576,"brier_hard":1.3363471511528975,"score_mae":0.6673735128993608,"within_one":true,"case_id":"typed:179/r0","base_id":"typed:179","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.672905}
{"id":"needs_human","type":"binary","probabilities":{"false":3.078942446143413e-08,"true":0.9999999692105743},"predicted":"true","confidence":0.9999999692105743,"entropy_confidence":0.9999991872920526,"candidate_mass":0.9999998520347976,"accepted":true,"accepted_predicted":"true","input_tokens":252,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.0789426149686804e-08,"brier_hard":1.8959773921053465e-15,"soft_accuracy":0.8599999778316133,"brier_soft":0.03919998275792387,"tvd":0.13999996921057495,"kl":1.1597399933838304,"case_id":"typed:179/r0","base_id":"typed:179","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.672905}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002110349500859409,"1":0.0004000708593919208,"2":0.03050049715237254,"3":0.9688883970381512},"predicted":"3","confidence":0.9688883970381512,"entropy_confidence":0.8975788649115023,"candidate_mass":0.9999963459146712,"accepted":true,"accepted_predicted":"3","input_tokens":278,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.4900122954238277,"brier_hard":1.8786742165293975,"score_mae":0.76806625627859,"within_one":true,"case_id":"typed:179/r0","base_id":"typed:179","repeat":0,"workflow":"customer_service","case_elapsed_ms":275.672905}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00029162493078405284,"close_no_action":8.219479654075223e-05,"escalate_to_human":0.022962778075446326,"execute_refund":7.125560955931936e-05,"request_information":0.9765921465876701},"predicted":"request_information","confidence":0.9765921465876701,"entropy_confidence":0.9294052005897139,"candidate_mass":0.9999773328946258,"accepted":true,"accepted_predicted":"request_information","input_tokens":452,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.7738807185328866,"brier_hard":1.9083340506812096,"soft_accuracy":0.32972249444081675,"brier_soft":0.6308597299824689,"tvd":0.6496691465876698,"kl":2.473610314244577,"case_id":"typed:180/r0","base_id":"typed:180","repeat":0,"workflow":"customer_service","case_elapsed_ms":359.91850600000004}
{"id":"category","type":"choice","probabilities":{"account":7.695679561401338e-11,"billing":6.178017102714734e-11,"delivery":0.9999999995514894,"refund":2.0614415462644128e-10,"technical":1.0362989408303185e-10},"predicted":"delivery","confidence":0.9999999995514894,"entropy_confidence":0.9999999933685251,"candidate_mass":0.9999999992318109,"accepted":true,"accepted_predicted":"delivery","input_tokens":442,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":4.485105621239107e-10,"brier_hard":2.6413542960411434e-19,"soft_accuracy":0.6333329997566607,"brier_soft":0.23611155514565738,"tvd":0.366666999628446,"kl":3.0878473707943837,"case_id":"typed:180/r0","base_id":"typed:180","repeat":0,"workflow":"customer_service","case_elapsed_ms":359.91850600000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":3.15574489456895e-06,"1":2.921892664977964e-06,"2":0.113826277630104,"3":0.8861676447323368},"predicted":"3","confidence":0.8861676447323368,"entropy_confidence":0.744265268317676,"candidate_mass":0.9999966703235696,"accepted":true,"accepted_predicted":"3","input_tokens":428,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.1730818733120607,"brier_hard":1.5705969608078707,"score_mae":1.0694914113498835,"within_one":false,"case_id":"typed:180/r0","base_id":"typed:180","repeat":0,"workflow":"customer_service","case_elapsed_ms":359.91850600000004}
{"id":"needs_human","type":"binary","probabilities":{"false":3.769293390981009e-09,"true":0.9999999962307076},"predicted":"true","confidence":0.9999999962307076,"entropy_confidence":0.999999889085702,"candidate_mass":0.9999999701909776,"accepted":true,"accepted_predicted":"true","input_tokens":396,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.76929244083448e-09,"brier_hard":2.8415138118272596e-17,"soft_accuracy":0.7999999977384248,"brier_soft":0.07999999698456567,"tvd":0.19999999623070705,"kl":1.6635532363593024,"case_id":"typed:180/r0","base_id":"typed:180","repeat":0,"workflow":"customer_service","case_elapsed_ms":359.91850600000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.021086044582973e-09,"1":4.90169732055706e-09,"2":1.2961126102710844e-07,"3":0.9999998584659544},"predicted":"3","confidence":0.9999998584659544,"entropy_confidence":0.9999982524581781,"candidate_mass":0.9999999552232072,"accepted":true,"accepted_predicted":"3","input_tokens":422,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.4153405561052636e-07,"brier_hard":3.6904287333274355e-14,"score_mae":0.4999998395220828,"within_one":true,"case_id":"typed:180/r0","base_id":"typed:180","repeat":0,"workflow":"customer_service","case_elapsed_ms":359.91850600000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":3.9284690614487615e-06,"close_no_action":9.089638200789825e-06,"escalate_to_human":3.839803721952562e-05,"execute_refund":1.3334756508982441e-05,"request_information":0.9999352490990092},"predicted":"request_information","confidence":0.9999352490990092,"entropy_confidence":0.9995282451843172,"candidate_mass":0.9999971975852918,"accepted":true,"accepted_predicted":"request_information","input_tokens":525,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":6.475299742087149e-05,"brier_hard":5.942958564372292e-09,"soft_accuracy":0.4166485337664251,"brier_soft":0.5325959343861268,"tvd":0.5832815838555182,"kl":4.860551246671214,"case_id":"typed:181/r0","base_id":"typed:181","repeat":0,"workflow":"customer_service","case_elapsed_ms":425.690508}
{"id":"category","type":"choice","probabilities":{"account":0.9999943360707788,"billing":4.44178652556099e-06,"delivery":5.836242333178191e-07,"refund":2.7568384417785327e-07,"technical":3.628346182064488e-07},"predicted":"account","confidence":0.9999943360707788,"entropy_confidence":0.9999513318438381,"candidate_mass":0.9999978559011637,"accepted":true,"accepted_predicted":"account","input_tokens":515,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":false,"nll_hard":14.829318703724924,"brier_hard":1.999987946524679,"soft_accuracy":0.3166652056764421,"brier_soft":0.911657837508333,"tvd":0.6833285120455396,"kl":5.913260279351913,"case_id":"typed:181/r0","base_id":"typed:181","repeat":0,"workflow":"customer_service","case_elapsed_ms":425.690508}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.5743778510793474e-05,"1":0.999863375413573,"2":7.446981211548255e-05,"3":6.410995801270109e-06},"predicted":"1","confidence":0.999863375413573,"entropy_confidence":0.998941700198308,"candidate_mass":0.9999980781567068,"accepted":true,"accepted_predicted":"1","input_tokens":501,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.00013663392041603377,"brier_hard":2.736050024270139e-08,"score_mae":0.16996845197479216,"within_one":true,"case_id":"typed:181/r0","base_id":"typed:181","repeat":0,"workflow":"customer_service","case_elapsed_ms":425.690508}
{"id":"needs_human","type":"binary","probabilities":{"false":3.20717638393362e-08,"true":0.9999999679282366},"predicted":"true","confidence":0.9999999679282366,"entropy_confidence":0.9999991553318551,"candidate_mass":0.9999999309471038,"accepted":true,"accepted_predicted":"true","input_tokens":469,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.25528982508629,"brier_hard":1.9999998717129477,"soft_accuracy":0.26666701496680184,"brier_soft":1.0755544837008717,"tvd":0.7333329679282364,"kl":6.4017782133078684,"case_id":"typed:181/r0","base_id":"typed:181","repeat":0,"workflow":"customer_service","case_elapsed_ms":425.690508}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00015331616167644154,"1":0.9993071418613805,"2":0.00048285625581575545,"3":5.6685721128998906e-05},"predicted":"1","confidence":0.9993071418613805,"entropy_confidence":0.9954696077576737,"candidate_mass":0.9999960633165944,"accepted":true,"accepted_predicted":"1","input_tokens":495,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0006930982757466975,"brier_hard":7.399216804428321e-07,"score_mae":0.03622408846360092,"within_one":true,"case_id":"typed:181/r0","base_id":"typed:181","repeat":0,"workflow":"customer_service","case_elapsed_ms":425.690508}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0010169511281562873,"close_no_action":0.0006356299313500585,"escalate_to_human":0.0021324638320397156,"execute_refund":0.0002777449483902505,"request_information":0.9959372101600653},"predicted":"request_information","confidence":0.9959372101600653,"entropy_confidence":0.9806571253875453,"candidate_mass":0.9999419290070435,"accepted":true,"accepted_predicted":"request_information","input_tokens":383,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":6.150477239098791,"brier_hard":1.9876320616765928,"soft_accuracy":0.08440434440361055,"brier_soft":1.1841991560894511,"tvd":0.9126042101600645,"kl":4.898127116714705,"case_id":"typed:182/r0","base_id":"typed:182","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.854406}
{"id":"category","type":"choice","probabilities":{"account":0.9873758895635625,"billing":0.012544046056293978,"delivery":7.15477420020996e-05,"refund":7.744308260095532e-06,"technical":7.723298829064393e-07},"predicted":"account","confidence":0.9873758895635625,"entropy_confidence":0.9575918702363954,"candidate_mass":0.99999378862941,"accepted":true,"accepted_predicted":"account","input_tokens":373,"reused_prefix_tokens":0,"expected":"billing","correct":false,"accepted_correct":false,"nll_hard":4.378509143801267,"brier_hard":1.949980413449961,"soft_accuracy":0.46662992048101076,"brier_soft":0.47958702015652743,"tvd":0.5207804373055638,"kl":1.9238209147655345,"case_id":"typed:182/r0","base_id":"typed:182","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.854406}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0001106113975264112,"1":0.9940602183744828,"2":0.005741246733824452,"3":8.792349416673546e-05},"predicted":"1","confidence":0.9940602183744828,"entropy_confidence":0.973038823721607,"candidate_mass":0.9999945858523789,"accepted":true,"accepted_predicted":"1","input_tokens":359,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.1600788912206745,"brier_hard":1.9767062061665561,"score_mae":1.0441935176753678,"within_one":false,"case_id":"typed:182/r0","base_id":"typed:182","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.854406}
{"id":"needs_human","type":"binary","probabilities":{"false":5.099723758613805e-07,"true":0.9999994900276249},"predicted":"true","confidence":0.9999994900276249,"entropy_confidence":0.9999886042734435,"candidate_mass":0.9999996279200668,"accepted":true,"accepted_predicted":"true","input_tokens":327,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.099725051427045e-07,"brier_hard":5.201436475137195e-13,"soft_accuracy":0.716666779011631,"brier_soft":0.16055459981050782,"tvd":0.28333249002762445,"kl":2.3708404363667355,"case_id":"typed:182/r0","base_id":"typed:182","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.854406}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005888339526781937,"1":0.9909433652276175,"2":0.008293605333896718,"3":0.00017419548580783613},"predicted":"1","confidence":0.9909433652276175,"entropy_confidence":0.9605799500696468,"candidate_mass":0.9999951616167616,"accepted":true,"accepted_predicted":"1","input_tokens":353,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.792270502821509,"brier_hard":1.9654507033797675,"score_mae":0.9252798376471656,"within_one":true,"case_id":"typed:182/r0","base_id":"typed:182","repeat":0,"workflow":"customer_service","case_elapsed_ms":329.854406}
{"id":"action","type":"choice","probabilities":{"answer_directly":2.9573040234362486e-05,"close_no_action":7.785820631432503e-05,"escalate_to_human":0.0004294781901388343,"execute_refund":9.269057042678847e-05,"request_information":0.999370399992885},"predicted":"request_information","confidence":0.999370399992885,"entropy_confidence":0.9963559567997409,"candidate_mass":0.9999824335650417,"accepted":true,"accepted_predicted":"request_information","input_tokens":616,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0006297982884291087,"brier_hard":5.963756916093462e-07,"soft_accuracy":0.4498494499503298,"brier_soft":0.4481980542388019,"tvd":0.5493703999928853,"kl":3.8325205293741464,"case_id":"typed:183/r0","base_id":"typed:183","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.655809}
{"id":"category","type":"choice","probabilities":{"account":0.9999999999690647,"billing":2.337567195274292e-11,"delivery":6.128241180967121e-12,"refund":7.921179210272832e-13,"technical":6.39516505079425e-13},"predicted":"account","confidence":0.9999999999690647,"entropy_confidence":0.99999999950206,"candidate_mass":0.9999999997915285,"accepted":true,"accepted_predicted":"account","input_tokens":606,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":3.093525435863386e-11,"brier_hard":1.5420037735792051e-21,"soft_accuracy":0.8833329999729372,"brier_soft":0.023711195548255157,"tvd":0.11666699996906459,"kl":0.9649626485638028,"case_id":"typed:183/r0","base_id":"typed:183","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.655809}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00034630565808982605,"1":0.8648655787830494,"2":0.13420176051165805,"3":0.0005863550472008946},"predicted":"1","confidence":0.8648655787830494,"entropy_confidence":0.7098615246573722,"candidate_mass":0.999992401225574,"accepted":true,"accepted_predicted":"1","input_tokens":592,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.14518118437693775,"brier_hard":0.03627188806191886,"score_mae":0.04830483505203187,"within_one":true,"case_id":"typed:183/r0","base_id":"typed:183","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.655809}
{"id":"needs_human","type":"binary","probabilities":{"false":1.1768438323072511e-06,"true":0.9999988231561684},"predicted":"true","confidence":0.9999988231561684,"entropy_confidence":0.9999751222982507,"candidate_mass":0.9999996451550385,"accepted":true,"accepted_predicted":"true","input_tokens":560,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.652674421320842,"brier_hard":1.9999952926274422,"soft_accuracy":0.4666670784554713,"brier_soft":0.5688856671821643,"tvd":0.5333318231561681,"kl":4.556513732466561,"case_id":"typed:183/r0","base_id":"typed:183","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.655809}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.004546152379481115,"1":0.9924380332810834,"2":0.002718341375665754,"3":0.00029747296377007553},"predicted":"1","confidence":0.9924380332810834,"entropy_confidence":0.963552019570973,"candidate_mass":0.9999936269832834,"accepted":true,"accepted_predicted":"1","input_tokens":586,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.907733373028947,"brier_hard":1.9795247125229496,"score_mae":1.1345658650762749,"within_one":false,"case_id":"typed:183/r0","base_id":"typed:183","repeat":0,"workflow":"customer_service","case_elapsed_ms":525.655809}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0008289013286510106,"close_no_action":0.0008216558206884298,"escalate_to_human":0.05677939330541421,"execute_refund":8.522884420546926e-05,"request_information":0.9414848207010407},"predicted":"request_information","confidence":0.9414848207010407,"entropy_confidence":0.8557494544467548,"candidate_mass":0.9998898684031975,"accepted":true,"accepted_predicted":"request_information","input_tokens":475,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":2.8685818129994303,"brier_hard":1.7760601499634292,"soft_accuracy":0.18315648548161073,"brier_soft":0.8395949633890364,"tvd":0.7748178207010408,"kl":3.271420875311532,"case_id":"typed:184/r0","base_id":"typed:184","repeat":0,"workflow":"customer_service","case_elapsed_ms":397.794808}
{"id":"category","type":"choice","probabilities":{"account":6.991499868980504e-07,"billing":0.999979830912468,"delivery":3.358518096113553e-06,"refund":1.5251934016971588e-05,"technical":8.594854316650342e-07},"predicted":"billing","confidence":0.999979830912468,"entropy_confidence":0.9998424483466408,"candidate_mass":0.9999993810138681,"accepted":true,"accepted_predicted":"billing","input_tokens":465,"reused_prefix_tokens":0,"expected":"refund","correct":false,"accepted_correct":false,"nll_hard":11.090804242166907,"brier_hard":1.9999291586088228,"soft_accuracy":0.0433465803707238,"brier_soft":1.8113782172914092,"tvd":0.9566476903978999,"kl":8.66719739081685,"case_id":"typed:184/r0","base_id":"typed:184","repeat":0,"workflow":"customer_service","case_elapsed_ms":397.794808}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":2.879848194425405e-05,"1":8.963410156042712e-05,"2":0.006078421353362063,"3":0.9938031460631324},"predicted":"3","confidence":0.9938031460631324,"entropy_confidence":0.9723490735826413,"candidate_mass":0.9999990234905106,"accepted":true,"accepted_predicted":"3","input_tokens":451,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.1030102625464835,"brier_hard":1.975524806488029,"score_mae":0.9869889149976814,"within_one":true,"case_id":"typed:184/r0","base_id":"typed:184","repeat":0,"workflow":"customer_service","case_elapsed_ms":397.794808}
{"id":"needs_human","type":"binary","probabilities":{"false":1.384418751006691e-09,"true":0.9999999986155821},"predicted":"true","confidence":0.9999999986155821,"entropy_confidence":0.9999999572619318,"candidate_mass":0.9999999913500464,"accepted":true,"accepted_predicted":"true","input_tokens":419,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.384417914599607e-09,"brier_hard":3.833228237749858e-18,"soft_accuracy":0.6099999996954284,"brier_soft":0.30419999784030743,"tvd":0.38999999861558166,"kl":3.290511989608191,"case_id":"typed:184/r0","base_id":"typed:184","repeat":0,"workflow":"customer_service","case_elapsed_ms":397.794808}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.225109031689179e-05,"1":3.605029426587019e-05,"2":0.19908865883803262,"3":0.8008030397773835},"predicted":"3","confidence":0.8008030397773835,"entropy_confidence":0.6391253841763352,"candidate_mass":0.9999978410509023,"accepted":true,"accepted_predicted":"3","input_tokens":445,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.2221402550625271,"brier_hard":0.07931572955970095,"score_mae":0.5539554873024821,"within_one":true,"case_id":"typed:184/r0","base_id":"typed:184","repeat":0,"workflow":"customer_service","case_elapsed_ms":397.794808}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00016421555456227127,"close_no_action":6.297274085795393e-05,"escalate_to_human":0.9996977680765334,"execute_refund":1.1734159838106209e-05,"request_information":6.330946820984345e-05},"predicted":"escalate_to_human","confidence":0.9996977680765334,"entropy_confidence":0.9980815669464542,"candidate_mass":0.999996645859459,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":456,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.000302277604738806,"brier_hard":1.2642222928577762e-07,"soft_accuracy":0.7764584999953809,"brier_soft":0.07072349814053445,"tvd":0.22303076807653263,"kl":1.2954273323218646,"case_id":"typed:185/r0","base_id":"typed:185","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.740206}
{"id":"category","type":"choice","probabilities":{"account":1.7609034231837412e-12,"billing":1.5811454081500736e-12,"delivery":0.999999999994678,"refund":1.6349403632879526e-12,"technical":3.456061106195596e-13},"predicted":"delivery","confidence":0.999999999994678,"entropy_confidence":0.9999999999066548,"candidate_mass":0.9999999999541664,"accepted":true,"accepted_predicted":"delivery","input_tokens":446,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":5.321965090857312e-12,"brier_hard":3.671658767085315e-23,"soft_accuracy":0.9333339333289935,"brier_soft":0.008466550266327803,"tvd":0.06666606666408643,"kl":0.5496243771441706,"case_id":"typed:185/r0","base_id":"typed:185","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.740206}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":5.109108797867307e-07,"1":8.700012861930651e-07,"2":6.26182649231235e-05,"3":0.9999360008229116},"predicted":"3","confidence":0.9999360008229116,"entropy_confidence":0.9995025678616509,"candidate_mass":0.9999996389443572,"accepted":true,"accepted_predicted":"3","input_tokens":432,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.400122512315243e-05,"brier_hard":8.017959702144566e-09,"score_mae":0.24993410899986745,"within_one":true,"case_id":"typed:185/r0","base_id":"typed:185","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.740206}
{"id":"needs_human","type":"binary","probabilities":{"false":2.2290167941115494e-09,"true":0.999999997770983},"predicted":"true","confidence":0.999999997770983,"entropy_confidence":0.9999999327201584,"candidate_mass":0.9999999833107794,"accepted":true,"accepted_predicted":"true","input_tokens":400,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.2290169734846776e-09,"brier_hard":9.937032525439213e-18,"soft_accuracy":0.8799999983059471,"brier_soft":0.0287999989300719,"tvd":0.11999999777098311,"kl":0.9927474796699782,"case_id":"typed:185/r0","base_id":"typed:185","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.740206}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.7052142053135403e-05,"1":4.755221209417954e-05,"2":0.004810843270742389,"3":0.9951145523751113},"predicted":"3","confidence":0.9951145523751113,"entropy_confidence":0.9774173151793494,"candidate_mass":0.9999983483537276,"accepted":true,"accepted_predicted":"3","input_tokens":426,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.004897420435078416,"brier_hard":4.7014804502443065e-05,"score_mae":0.44834589587891305,"within_one":true,"case_id":"typed:185/r0","base_id":"typed:185","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.740206}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.921376878477879,"close_no_action":0.000860628341471371,"escalate_to_human":0.02121480942978223,"execute_refund":0.00027221421085983063,"request_information":0.0562754695400089},"predicted":"answer_directly","confidence":0.921376878477879,"entropy_confidence":0.7965556433357257,"candidate_mass":0.9999432341290614,"accepted":true,"accepted_predicted":"answer_directly","input_tokens":440,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":3.8530557831736516,"brier_hard":1.8101235447268849,"soft_accuracy":0.06265599428512318,"brier_soft":1.151129952794203,"tvd":0.8713768784778783,"kl":3.649893733172603,"case_id":"typed:186/r0","base_id":"typed:186","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.720306}
{"id":"category","type":"choice","probabilities":{"account":4.07773234504046e-12,"billing":0.9999999998413571,"delivery":1.0998675113298719e-10,"refund":3.739534986150756e-11,"technical":7.183296919117247e-12},"predicted":"billing","confidence":0.9999999998413571,"entropy_confidence":0.9999999975955499,"candidate_mass":0.9999999995859934,"accepted":true,"accepted_predicted":"billing","input_tokens":430,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.5864287662574134e-10,"brier_hard":3.8731287571857737e-20,"soft_accuracy":0.8599999998682472,"brier_soft":0.03508881772421991,"tvd":0.13999999995852705,"kl":1.1597399670413566,"case_id":"typed:186/r0","base_id":"typed:186","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.720306}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00012243516432831992,"1":0.02482438943234282,"2":0.9743247578912051,"3":0.0007284175121225989},"predicted":"2","confidence":0.9743247578912051,"entropy_confidence":0.910944320290829,"candidate_mass":0.9999949611855331,"accepted":true,"accepted_predicted":"2","input_tokens":416,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.026010603921005078,"brier_hard":0.00127601395047528,"score_mae":0.05767384224887895,"within_one":true,"case_id":"typed:186/r0","base_id":"typed:186","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.720306}
{"id":"needs_human","type":"binary","probabilities":{"false":4.737174391183355e-08,"true":0.9999999526282555},"predicted":"true","confidence":0.9999999526282555,"entropy_confidence":0.9999987790362668,"candidate_mass":0.999999884822159,"accepted":true,"accepted_predicted":"true","input_tokens":384,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.737174564911051e-08,"brier_hard":4.488164300786267e-15,"soft_accuracy":0.7566669756824728,"brier_soft":0.11842185166956995,"tvd":0.2433329526282558,"kl":2.030196804039041,"case_id":"typed:186/r0","base_id":"typed:186","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.720306}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00030834999217796975,"1":0.04463365665494454,"2":0.9442374093745064,"3":0.010820583978372659},"predicted":"2","confidence":0.9442374093745064,"entropy_confidence":0.8236838144652506,"candidate_mass":0.9999952943138639,"accepted":true,"accepted_predicted":"2","input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.05737765147422638,"brier_hard":0.005218809937008553,"score_mae":0.11776277266092472,"within_one":true,"case_id":"typed:186/r0","base_id":"typed:186","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.720306}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.007950561966542688,"close_no_action":0.00029799441468342014,"escalate_to_human":0.7649416407577684,"execute_refund":9.536962124610175e-05,"request_information":0.22671443323975954},"predicted":"escalate_to_human","confidence":0.7649416407577684,"entropy_confidence":0.6376570443947085,"candidate_mass":0.9999686607737168,"accepted":false,"accepted_predicted":null,"input_tokens":529,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":null,"nll_hard":0.2679557346567236,"brier_hard":0.10671517582049507,"soft_accuracy":0.6523406595577856,"brier_soft":0.06478243915612576,"tvd":0.22014280952799894,"kl":0.6005478376541348,"case_id":"typed:187/r0","base_id":"typed:187","repeat":0,"workflow":"customer_service","case_elapsed_ms":492.104509}
{"id":"category","type":"choice","probabilities":{"account":0.9999999929278882,"billing":5.6997642538541396e-09,"delivery":1.0183235168284613e-09,"refund":1.7312274389176305e-10,"technical":1.8090083690073885e-10},"predicted":"account","confidence":0.9999999929278882,"entropy_confidence":0.999999910339024,"candidate_mass":0.9999999850519129,"accepted":true,"accepted_predicted":"account","input_tokens":519,"reused_prefix_tokens":0,"expected":"technical","correct":false,"accepted_correct":false,"nll_hard":22.433072097221036,"brier_hard":1.9999999854939747,"soft_accuracy":0.43333299703793116,"brier_soft":0.6422229695579144,"tvd":0.566666999819099,"kl":4.856821752227076,"case_id":"typed:187/r0","base_id":"typed:187","repeat":0,"workflow":"customer_service","case_elapsed_ms":492.104509}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.3123287913382564e-05,"1":0.9892917595548336,"2":0.01061086863874086,"3":8.424851851354548e-05},"predicted":"1","confidence":0.9892917595548336,"entropy_confidence":0.9568458768392185,"candidate_mass":0.9999912635608222,"accepted":true,"accepted_predicted":"1","input_tokens":505,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.545876459892046,"brier_hard":1.9575890460489191,"score_mae":0.9225667576121439,"within_one":true,"case_id":"typed:187/r0","base_id":"typed:187","repeat":0,"workflow":"customer_service","case_elapsed_ms":492.104509}
{"id":"needs_human","type":"binary","probabilities":{"false":1.596781256300009e-08,"true":0.9999999840321864},"predicted":"true","confidence":0.9999999840321864,"entropy_confidence":0.9999995633928485,"candidate_mass":0.9999999567244702,"accepted":true,"accepted_predicted":"true","input_tokens":473,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.5967813754379237e-08,"brier_hard":5.099421100703157e-16,"soft_accuracy":0.8666669882902592,"brier_soft":0.03555536926185483,"tvd":0.13333298403218694,"kl":1.104021881133344,"case_id":"typed:187/r0","base_id":"typed:187","repeat":0,"workflow":"customer_service","case_elapsed_ms":492.104509}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.349364259242108e-05,"1":0.006200110235163224,"2":0.9934742612617927,"3":0.00025213486045266593},"predicted":"2","confidence":0.9934742612617927,"entropy_confidence":0.9705622856837555,"candidate_mass":0.9999970493858302,"accepted":true,"accepted_predicted":"2","input_tokens":499,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00654712446044318,"brier_hard":8.109560631087185e-05,"score_mae":0.3272380373401067,"within_one":true,"case_id":"typed:187/r0","base_id":"typed:187","repeat":0,"workflow":"customer_service","case_elapsed_ms":492.104509}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0008133472975000556,"close_no_action":0.0005853551521196047,"escalate_to_human":0.010113765134901588,"execute_refund":0.0010208044877429352,"request_information":0.9874667279277358},"predicted":"request_information","confidence":0.9874667279277358,"entropy_confidence":0.9527229622721873,"candidate_mass":0.9999414417439902,"accepted":true,"accepted_predicted":"request_information","input_tokens":261,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":7.1143523594350775,"brier_hard":1.9735681786307957,"soft_accuracy":0.18464080699881244,"brier_soft":0.9736908147841707,"tvd":0.8051545324154787,"kl":3.9248947209384024,"case_id":"typed:188/r0","base_id":"typed:188","repeat":0,"workflow":"customer_service","case_elapsed_ms":234.62940500000002}
{"id":"category","type":"choice","probabilities":{"account":1.5817387494207332e-07,"billing":0.9999986268019074,"delivery":9.809795377439198e-07,"refund":1.8376563112093523e-07,"technical":5.0279049545849216e-08},"predicted":"billing","confidence":0.9999986268019074,"entropy_confidence":0.9999868793885455,"candidate_mass":0.999999623945525,"accepted":true,"accepted_predicted":"billing","input_tokens":251,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.3731990354064473e-06,"brier_hard":2.9093106196263437e-12,"soft_accuracy":0.9499987033705057,"brier_soft":0.004999846865712564,"tvd":0.04999984182612547,"kl":0.41178969346971983,"case_id":"typed:188/r0","base_id":"typed:188","repeat":0,"workflow":"customer_service","case_elapsed_ms":234.62940500000002}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0002992353000640885,"1":0.059917049775166674,"2":0.9389798913075185,"3":0.0008038236172512402},"predicted":"2","confidence":0.9389798913075185,"entropy_confidence":0.829812689482873,"candidate_mass":0.9999984091452674,"accepted":true,"accepted_predicted":"2","input_tokens":237,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.06296121501114625,"brier_hard":0.007314242192774517,"score_mae":0.2236213032419574,"within_one":true,"case_id":"typed:188/r0","base_id":"typed:188","repeat":0,"workflow":"customer_service","case_elapsed_ms":234.62940500000002}
{"id":"needs_human","type":"binary","probabilities":{"false":5.6958735056665914e-08,"true":0.9999999430412653},"predicted":"true","confidence":0.9999999430412653,"entropy_confidence":0.9999985470850439,"candidate_mass":0.9999998576418706,"accepted":true,"accepted_predicted":"true","input_tokens":205,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.68093877766186,"brier_hard":1.9999997721650669,"soft_accuracy":0.4000000113917472,"brier_soft":0.7199998632990428,"tvd":0.599999943041265,"kl":5.159687953219542,"case_id":"typed:188/r0","base_id":"typed:188","repeat":0,"workflow":"customer_service","case_elapsed_ms":234.62940500000002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007106344752376438,"1":0.9918840617383068,"2":0.007286938557600751,"3":0.00011836522885573179},"predicted":"1","confidence":0.9918840617383068,"entropy_confidence":0.9638109317953638,"candidate_mass":0.9999961612862651,"accepted":true,"accepted_predicted":"1","input_tokens":231,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.921671772112063,"brier_hard":1.9693137333005066,"score_mae":0.8931869654599245,"within_one":true,"case_id":"typed:188/r0","base_id":"typed:188","repeat":0,"workflow":"customer_service","case_elapsed_ms":234.62940500000002}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0007906905454768488,"close_no_action":0.0002012596254554089,"escalate_to_human":0.021999128375919573,"execute_refund":0.0001748397174644074,"request_information":0.9768340817356831},"predicted":"request_information","confidence":0.9768340817356831,"entropy_confidence":0.9280905081928724,"candidate_mass":0.999977975307363,"accepted":true,"accepted_predicted":"request_information","input_tokens":244,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":7.142603886114756,"brier_hard":1.9531081000646442,"soft_accuracy":0.027344067691455923,"brier_soft":1.7849829324756274,"tvd":0.9501671084026567,"kl":6.590144709256573,"case_id":"typed:189/r0","base_id":"typed:189","repeat":0,"workflow":"customer_service","case_elapsed_ms":208.59740299999999}
{"id":"category","type":"choice","probabilities":{"account":0.9999945115580304,"billing":3.6469501144587103e-06,"delivery":1.2268242904969662e-06,"refund":2.9067983580669913e-07,"technical":3.23987727242537e-07},"predicted":"account","confidence":0.9999945115580304,"entropy_confidence":0.9999521144656672,"candidate_mass":0.9999989969818252,"accepted":true,"accepted_predicted":"account","input_tokens":234,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":5.488457031122727e-06,"brier_hard":4.5117801044789496e-11,"soft_accuracy":0.9433279837365013,"brier_soft":0.0050213434661762505,"tvd":0.05666151155803123,"kl":0.4624094464302402,"case_id":"typed:189/r0","base_id":"typed:189","repeat":0,"workflow":"customer_service","case_elapsed_ms":208.59740299999999}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00019136706815269045,"1":0.0001713836801269301,"2":0.0014149311943387418,"3":0.9982223180573812},"predicted":"3","confidence":0.9982223180573812,"entropy_confidence":0.9897687460698661,"candidate_mass":0.9999981044998544,"accepted":true,"accepted_predicted":"3","input_tokens":220,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0017792638942459623,"brier_hard":5.228177094413183e-06,"score_mae":0.08766820024094768,"within_one":true,"case_id":"typed:189/r0","base_id":"typed:189","repeat":0,"workflow":"customer_service","case_elapsed_ms":208.59740299999999}
{"id":"needs_human","type":"binary","probabilities":{"false":6.473584571900249e-07,"true":0.9999993526415415},"predicted":"true","confidence":0.9999993526415415,"entropy_confidence":0.9999857570610429,"candidate_mass":0.9999995573362972,"accepted":true,"accepted_predicted":"true","input_tokens":188,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.250365666203272,"brier_hard":1.9999974105670069,"soft_accuracy":0.05000058262261141,"brier_soft":1.8049975400386982,"tvd":0.9499993526415421,"kl":8.60003677206761,"case_id":"typed:189/r0","base_id":"typed:189","repeat":0,"workflow":"customer_service","case_elapsed_ms":208.59740299999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0022972450902120356,"1":0.9972889462673408,"2":0.0003801562905264225,"3":3.36523519215075e-05},"predicted":"1","confidence":0.9972889462673408,"entropy_confidence":0.98556883498634,"candidate_mass":0.999997053319227,"accepted":true,"accepted_predicted":"1","input_tokens":214,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.002714735294276265,"brier_hard":1.2772798631885505e-05,"score_mae":0.04815021590415813,"within_one":true,"case_id":"typed:189/r0","base_id":"typed:189","repeat":0,"workflow":"customer_service","case_elapsed_ms":208.59740299999999}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0029947684651369924,"close_no_action":0.0021574416956250675,"escalate_to_human":0.9846466046016785,"execute_refund":0.001960091415528104,"request_information":0.008241093822032537},"predicted":"escalate_to_human","confidence":0.9846466046016785,"entropy_confidence":0.9393279078744773,"candidate_mass":0.999968812081776,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":386,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.015472479240067125,"brier_hard":0.0003211075288277695,"soft_accuracy":0.7239240849718901,"brier_soft":0.10352247474547896,"tvd":0.25131287126794455,"kl":0.5787969380300948,"case_id":"typed:190/r0","base_id":"typed:190","repeat":0,"workflow":"customer_service","case_elapsed_ms":349.757907}
{"id":"category","type":"choice","probabilities":{"account":0.9999915482859235,"billing":5.985480505995144e-06,"delivery":1.6609294005765274e-06,"refund":3.751050504927704e-07,"technical":4.301991181541633e-07},"predicted":"account","confidence":0.9999915482859235,"entropy_confidence":0.9999289227934218,"candidate_mass":0.9999986852961813,"accepted":true,"accepted_predicted":"account","input_tokens":376,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":8.451749792421695e-06,"brier_hard":1.1034190927216856e-10,"soft_accuracy":0.733327477140551,"brier_soft":0.09743889795708678,"tvd":0.2666585482859241,"kl":2.228652856549706,"case_id":"typed:190/r0","base_id":"typed:190","repeat":0,"workflow":"customer_service","case_elapsed_ms":349.757907}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.328415243510956e-07,"1":4.27486463209419e-08,"2":5.682352928260565e-06,"3":0.9999941420569005},"predicted":"3","confidence":0.9999941420569005,"entropy_confidence":0.9999442261363941,"candidate_mass":0.9999999202229258,"accepted":true,"accepted_predicted":"3","input_tokens":362,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.857960257301081e-06,"brier_hard":6.662410647547419e-11,"score_mae":0.15666083362520444,"within_one":true,"case_id":"typed:190/r0","base_id":"typed:190","repeat":0,"workflow":"customer_service","case_elapsed_ms":349.757907}
{"id":"needs_human","type":"binary","probabilities":{"false":4.331376542444346e-08,"true":0.9999999566862359},"predicted":"true","confidence":0.9999999566862359,"entropy_confidence":0.9999988780307943,"candidate_mass":0.9999997966077452,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.331376506895559e-08,"brier_hard":3.752164438432207e-15,"soft_accuracy":0.8299999714129159,"brier_soft":0.05779997054664373,"tvd":0.16999995668623524,"kl":1.411104349287437,"case_id":"typed:190/r0","base_id":"typed:190","repeat":0,"workflow":"customer_service","case_elapsed_ms":349.757907}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.147965205200172e-05,"1":3.560868006555602e-05,"2":0.00016538986202484338,"3":0.9997375218058567},"predicted":"3","confidence":0.9997375218058567,"entropy_confidence":0.9980787484105723,"candidate_mass":0.9999974606735886,"accepted":true,"accepted_predicted":"3","input_tokens":356,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00026251264757344914,"brier_hard":1.0129633457375386e-07,"score_mae":0.36624595382168534,"within_one":true,"case_id":"typed:190/r0","base_id":"typed:190","repeat":0,"workflow":"customer_service","case_elapsed_ms":349.757907}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0002803740823268989,"close_no_action":0.0007015795865439098,"escalate_to_human":0.002081042304758035,"execute_refund":0.0005141978837913523,"request_information":0.9964228061425814},"predicted":"request_information","confidence":0.9964228061425814,"entropy_confidence":0.9827870087847027,"candidate_mass":0.999936860043108,"accepted":true,"accepted_predicted":"request_information","input_tokens":318,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":6.174886402724876,"brier_hard":1.9887014899516204,"soft_accuracy":0.18417753970836911,"brier_soft":1.1556194507003983,"tvd":0.8130898061425806,"kl":4.347703284512385,"case_id":"typed:191/r0","base_id":"typed:191","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.023405}
{"id":"category","type":"choice","probabilities":{"account":3.655440594129143e-10,"billing":0.9999999865587644,"delivery":8.491271371384623e-09,"refund":4.146739856926595e-09,"technical":4.376804884888948e-10},"predicted":"billing","confidence":0.9999999865587644,"entropy_confidence":0.9999998330747097,"candidate_mass":0.9999999916733042,"accepted":true,"accepted_predicted":"billing","input_tokens":308,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":1.3441235659345074e-08,"brier_hard":2.7028914123456003e-16,"soft_accuracy":0.9699990169946417,"brier_soft":0.0012667332647162942,"tvd":0.030000956557794282,"kl":0.24677278650518744,"case_id":"typed:191/r0","base_id":"typed:191","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.023405}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.339233790222971e-05,"1":3.631068503217825e-05,"2":0.6107322392423974,"3":0.3891680577346687},"predicted":"2","confidence":0.6107322392423974,"entropy_confidence":0.5171234622688087,"candidate_mass":0.9999954765434066,"accepted":false,"accepted_predicted":null,"input_tokens":294,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.4930966495088,"brier_hard":0.3029811720632669,"score_mae":0.13900496237383297,"within_one":true,"case_id":"typed:191/r0","base_id":"typed:191","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.023405}
{"id":"needs_human","type":"binary","probabilities":{"false":2.1132487300372344e-07,"true":0.999999788675127},"predicted":"true","confidence":0.999999788675127,"entropy_confidence":0.9999950091977496,"candidate_mass":0.9999998030523062,"accepted":true,"accepted_predicted":"true","input_tokens":262,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.1132489537419751e-07,"brier_hard":8.931640391756467e-14,"soft_accuracy":0.8766668408017881,"brier_soft":0.03042195352476707,"tvd":0.12333278867512701,"kl":1.0205451159096541,"case_id":"typed:191/r0","base_id":"typed:191","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.023405}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007088822859108452,"1":0.0004850070108082466,"2":0.011349171332889969,"3":0.9874569393703919},"predicted":"3","confidence":0.9874569393703919,"entropy_confidence":0.9479659368636966,"candidate_mass":0.9999964640784483,"accepted":true,"accepted_predicted":"3","input_tokens":288,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.478610548038173,"brier_hard":1.9525024058808012,"score_mae":0.9355541677877639,"within_one":true,"case_id":"typed:191/r0","base_id":"typed:191","repeat":0,"workflow":"customer_service","case_elapsed_ms":301.023405}
{"id":"action","type":"choice","probabilities":{"answer_directly":1.2488118726454696e-05,"close_no_action":2.6641852429266606e-05,"escalate_to_human":9.389485298921087e-05,"execute_refund":3.3384484325493825e-05,"request_information":0.9998335906915304},"predicted":"request_information","confidence":0.9998335906915304,"entropy_confidence":0.9988798318222416,"candidate_mass":0.9999935776826941,"accepted":true,"accepted_predicted":"request_information","input_tokens":353,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":false,"accepted_correct":false,"nll_hard":9.273334986998808,"brier_hard":1.999479430165649,"soft_accuracy":0.39998142385425567,"brier_soft":0.5996381186972093,"tvd":0.5998331906911301,"kl":4.686633001145312,"case_id":"typed:192/r0","base_id":"typed:192","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.816706}
{"id":"category","type":"choice","probabilities":{"account":7.554778501431699e-07,"billing":6.763456184439273e-07,"delivery":0.9999961290394981,"refund":1.6594831440834952e-06,"technical":7.796538876265975e-07},"predicted":"delivery","confidence":0.9999961290394981,"entropy_confidence":0.9999644720091793,"candidate_mass":0.9999995942301332,"accepted":true,"accepted_predicted":"delivery","input_tokens":343,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":3.8709679941040915e-06,"brier_hard":1.9374269875036454e-11,"soft_accuracy":0.8866628184528865,"brier_soft":0.01802120292764262,"tvd":0.11333001570561224,"kl":0.9362636624896256,"case_id":"typed:192/r0","base_id":"typed:192","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.816706}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00014202246022766265,"1":0.008878618473681895,"2":0.9888176037910992,"3":0.0021617552749916215},"predicted":"2","confidence":0.9888176037910992,"entropy_confidence":0.9512457155917577,"candidate_mass":0.9999981513754718,"accepted":true,"accepted_predicted":"2","input_tokens":329,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.011245389250394339,"brier_hard":0.0002085692072222082,"score_mae":0.05966609188085514,"within_one":true,"case_id":"typed:192/r0","base_id":"typed:192","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.816706}
{"id":"needs_human","type":"binary","probabilities":{"false":1.831803231212017e-07,"true":0.9999998168196765},"predicted":"true","confidence":0.9999998168196765,"entropy_confidence":0.9999956361085279,"candidate_mass":0.999999685589216,"accepted":true,"accepted_predicted":"true","input_tokens":297,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.831803402428326e-07,"brier_hard":6.711006168364533e-14,"soft_accuracy":0.7433329108523646,"brier_soft":0.13175570971269093,"tvd":0.2566668168196767,"kl":2.143509711245891,"case_id":"typed:192/r0","base_id":"typed:192","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.816706}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007367524955281189,"1":0.07525651257895778,"2":0.9205512982185958,"3":0.003455436706919087},"predicted":"2","confidence":0.9205512982185958,"entropy_confidence":0.7866382142957973,"candidate_mass":0.9999928840483517,"accepted":true,"accepted_predicted":"2","input_tokens":323,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.08278255121583905,"brier_hard":0.011988121747372517,"score_mae":0.2832745808630932,"within_one":true,"case_id":"typed:192/r0","base_id":"typed:192","repeat":0,"workflow":"customer_service","case_elapsed_ms":327.816706}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00022631554239005034,"close_no_action":0.00022770412704194176,"escalate_to_human":0.9974237447223905,"execute_refund":8.810748629550267e-05,"request_information":0.002034128121882493},"predicted":"escalate_to_human","confidence":0.9974237447223905,"entropy_confidence":0.9876901504829895,"candidate_mass":0.9999834121964571,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":393,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.0025795795338865064,"brier_hard":1.0885599294984642e-05,"soft_accuracy":0.7816036375860417,"brier_soft":0.06762831542799252,"tvd":0.21409074472239029,"kl":0.8173047798565645,"case_id":"typed:193/r0","base_id":"typed:193","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.282317}
{"id":"category","type":"choice","probabilities":{"account":0.0031524911346952323,"billing":0.003796441197601319,"delivery":0.000984248224873401,"refund":0.0006876640353945316,"technical":0.9913791554074353},"predicted":"technical","confidence":0.9913791554074353,"entropy_confidence":0.962891998336802,"candidate_mass":0.9999567761241789,"accepted":true,"accepted_predicted":"technical","input_tokens":383,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.008658219027800788,"brier_hard":0.00010011175400407075,"soft_accuracy":0.8101078838873244,"brier_soft":0.03982095035022581,"tvd":0.17471215540743537,"kl":0.428568026561749,"case_id":"typed:193/r0","base_id":"typed:193","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.282317}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":1.655062241623134e-07,"1":1.2659843154182676e-07,"2":6.082537363412003e-06,"3":0.9999936253579805},"predicted":"3","confidence":0.9999936253579805,"entropy_confidence":0.999939391401317,"candidate_mass":0.9999998097798265,"accepted":true,"accepted_predicted":"3","input_tokens":369,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.374662337585587e-06,"brier_hard":7.767674112678564e-11,"score_mae":0.16666016774710002,"within_one":true,"case_id":"typed:193/r0","base_id":"typed:193","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.282317}
{"id":"needs_human","type":"binary","probabilities":{"false":5.117863696747562e-09,"true":0.9999999948821348},"predicted":"true","confidence":0.9999999948821348,"entropy_confidence":0.9999998516612459,"candidate_mass":0.9999999575518084,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.117865255806013e-09,"brier_hard":5.238507346102305e-17,"soft_accuracy":0.8666669962468952,"brier_soft":0.03555537504847918,"tvd":0.13333299488213557,"kl":1.1040218717300518,"case_id":"typed:193/r0","base_id":"typed:193","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.282317}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.982348604059769e-07,"1":4.115110741764034e-07,"2":5.595635109566813e-06,"3":0.9999935946189543},"predicted":"3","confidence":0.9999935946189543,"entropy_confidence":0.9999379674251655,"candidate_mass":0.9999995957018797,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.405401560222885e-06,"brier_hard":7.266797098801285e-11,"score_mae":0.2566593866381566,"within_one":true,"case_id":"typed:193/r0","base_id":"typed:193","repeat":0,"workflow":"customer_service","case_elapsed_ms":344.282317}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00011106789433954067,"close_no_action":0.0003236407211318905,"escalate_to_human":0.0013332171183441713,"execute_refund":0.00021687426827023113,"request_information":0.9980151999979124},"predicted":"request_information","confidence":0.9980151999979124,"entropy_confidence":0.9899029442796158,"candidate_mass":0.9999664126106336,"accepted":true,"accepted_predicted":"request_information","input_tokens":427,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.0019867723278252387,"brier_hard":5.881012774698256e-06,"soft_accuracy":0.7187391003249842,"brier_soft":0.10369043676037314,"tvd":0.27801591999719333,"kl":1.3266799693465046,"case_id":"typed:194/r0","base_id":"typed:194","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.21551900000003}
{"id":"category","type":"choice","probabilities":{"account":1.8566451368759966e-12,"billing":0.999999999932017,"delivery":4.7670862930861746e-11,"refund":1.5538147771389163e-11,"technical":2.9173827919510732e-12},"predicted":"billing","confidence":0.999999999932017,"entropy_confidence":0.999999998934218,"candidate_mass":0.9999999997102744,"accepted":true,"accepted_predicted":"billing","input_tokens":417,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":6.798295260299783e-11,"brier_hard":7.147585306564845e-21,"soft_accuracy":0.8833329999414299,"brier_soft":0.02057787553717424,"tvd":0.11666699993201698,"kl":0.9649626485965281,"case_id":"typed:194/r0","base_id":"typed:194","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.21551900000003}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.00011071640707402492,"1":0.0005493442398204852,"2":0.22981715334182526,"3":0.7695227860112801},"predicted":"3","confidence":0.7695227860112801,"entropy_confidence":0.6071000370032528,"candidate_mass":0.9999966867997895,"accepted":false,"accepted_predicted":null,"input_tokens":403,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.470471271697793,"brier_hard":1.1853472495142685,"score_mae":0.9154190089573111,"within_one":true,"case_id":"typed:194/r0","base_id":"typed:194","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.21551900000003}
{"id":"needs_human","type":"binary","probabilities":{"false":2.5626419661482325e-08,"true":0.9999999743735802},"predicted":"true","confidence":0.9999999743735802,"entropy_confidence":0.9999993167871498,"candidate_mass":0.9999999017517472,"accepted":true,"accepted_predicted":"true","input_tokens":371,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.562642009235885e-08,"brier_hard":1.313426774587248e-15,"soft_accuracy":0.5599999969248296,"brier_soft":0.38719995489750253,"tvd":0.43999997437358024,"kl":3.727851420678668,"case_id":"typed:194/r0","base_id":"typed:194","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.21551900000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.8813793165105904e-06,"1":1.4026925791660714e-06,"2":1.7055731604128372e-05,"3":0.9999796601965015},"predicted":"3","confidence":0.9999796601965015,"entropy_confidence":0.9998187236991681,"candidate_mass":0.9999995382550146,"accepted":true,"accepted_predicted":"3","input_tokens":397,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.0340010355078443e-05,"brier_hard":7.101127215126899e-10,"score_mae":0.17330749474529172,"within_one":true,"case_id":"typed:194/r0","base_id":"typed:194","repeat":0,"workflow":"customer_service","case_elapsed_ms":343.21551900000003}
{"id":"action","type":"choice","probabilities":{"answer_directly":6.1009036510624315e-06,"close_no_action":1.219465876008205e-05,"escalate_to_human":6.28419374862267e-05,"execute_refund":2.0202675745746188e-05,"request_information":0.9998986598243566},"predicted":"request_information","confidence":0.9998986598243566,"entropy_confidence":0.9992923373279206,"candidate_mass":0.9999965312744812,"accepted":true,"accepted_predicted":"request_information","input_tokens":490,"reused_prefix_tokens":0,"expected":"request_information","correct":true,"accepted_correct":true,"nll_hard":0.00010134531090599114,"brier_hard":1.4813019141390183e-08,"soft_accuracy":0.599950787769737,"brier_soft":0.20742845052346776,"tvd":0.39989925982375674,"kl":3.062890594064198,"case_id":"typed:195/r0","base_id":"typed:195","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.358119}
{"id":"category","type":"choice","probabilities":{"account":3.2125728242136616e-06,"billing":2.8668651232265193e-06,"delivery":0.9999823718708918,"refund":4.72662093980486e-06,"technical":6.822070222231909e-06},"predicted":"delivery","confidence":0.9999823718708918,"entropy_confidence":0.9998546325334213,"candidate_mass":0.9999992394723428,"accepted":true,"accepted_predicted":"delivery","input_tokens":480,"reused_prefix_tokens":0,"expected":"delivery","correct":true,"accepted_correct":true,"nll_hard":1.762828448550981e-05,"brier_hard":3.9817206326724157e-10,"soft_accuracy":0.8399850780877202,"brier_soft":0.03399324663348426,"tvd":0.15998321187005113,"kl":1.2916356064495318,"case_id":"typed:195/r0","base_id":"typed:195","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.358119}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.70723012710018e-05,"1":0.9994354908281626,"2":0.0004875490257121924,"3":9.887844855795053e-06},"predicted":"1","confidence":0.9994354908281626,"entropy_confidence":0.9963637203889097,"candidate_mass":0.9999987551002225,"accepted":true,"accepted_predicted":"1","input_tokens":466,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0005646685671294717,"brier_hard":5.609711206351e-07,"score_mae":0.3662267475858456,"within_one":true,"case_id":"typed:195/r0","base_id":"typed:195","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.358119}
{"id":"needs_human","type":"binary","probabilities":{"false":9.78051680851625e-10,"true":0.9999999990219486},"predicted":"true","confidence":0.9999999990219486,"entropy_confidence":0.9999999693164991,"candidate_mass":0.9999999916431399,"accepted":true,"accepted_predicted":"true","input_tokens":434,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.780514181110339e-10,"brier_hard":1.9131696659501043e-18,"soft_accuracy":0.643332999719626,"brier_soft":0.25442269638264503,"tvd":0.3566669990219484,"kl":3.001254913278344,"case_id":"typed:195/r0","base_id":"typed:195","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.358119}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003459043994308557,"1":0.3599154841302294,"2":0.6392392020561837,"3":0.000499409414156234},"predicted":"2","confidence":0.6392392020561837,"entropy_confidence":0.5236290575553759,"candidate_mass":0.9999930691468023,"accepted":false,"accepted_predicted":null,"input_tokens":460,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.0218860413988224,"brier_hard":0.8183353139612815,"score_mae":0.2732251164850654,"within_one":true,"case_id":"typed:195/r0","base_id":"typed:195","repeat":0,"workflow":"customer_service","case_elapsed_ms":357.358119}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.09776591750544411,"close_no_action":0.000667024211538317,"escalate_to_human":0.0036308961176626173,"execute_refund":0.7883635133171466,"request_information":0.10957264884820916},"predicted":"execute_refund","confidence":0.7883635133171466,"entropy_confidence":0.5760298999038515,"candidate_mass":0.9999167771850463,"accepted":false,"accepted_predicted":null,"input_tokens":423,"reused_prefix_tokens":0,"expected":"execute_refund","correct":true,"accepted_correct":null,"nll_hard":0.23779598417943498,"brier_hard":0.06636797082467187,"soft_accuracy":0.3703348600804684,"brier_soft":0.2882248723654525,"tvd":0.41303568721575445,"kl":1.6690118513129815,"case_id":"typed:196/r0","base_id":"typed:196","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.192418}
{"id":"category","type":"choice","probabilities":{"account":6.240572280602544e-08,"billing":0.9999989806752357,"delivery":4.506530488704539e-07,"refund":3.879121036707824e-07,"technical":1.1835388964525172e-07},"predicted":"billing","confidence":0.9999989806752357,"entropy_confidence":0.9999899007895625,"candidate_mass":0.999999880940571,"accepted":true,"accepted_predicted":"billing","input_tokens":413,"reused_prefix_tokens":0,"expected":"refund","correct":false,"accepted_correct":false,"nll_hard":14.762487059917477,"brier_hard":1.9999971855276746,"soft_accuracy":0.060000300431943356,"brier_soft":1.7424201982659955,"tvd":0.939999099029125,"kl":8.487909952479246,"case_id":"typed:196/r0","base_id":"typed:196","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.192418}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":6.642185169870035e-07,"1":7.154556286236053e-07,"2":2.7682201094884722e-05,"3":0.9999709381247582},"predicted":"3","confidence":0.9999709381247582,"entropy_confidence":0.9997553545184642,"candidate_mass":0.9999997272871165,"accepted":true,"accepted_predicted":"3","input_tokens":399,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.906229754624312e-05,"brier_hard":1.611849913020379e-09,"score_mae":0.13330189423209315,"within_one":true,"case_id":"typed:196/r0","base_id":"typed:196","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.192418}
{"id":"needs_human","type":"binary","probabilities":{"false":8.36275378950301e-07,"true":0.9999991637246219},"predicted":"true","confidence":0.9999991637246219,"entropy_confidence":0.9999819095128663,"candidate_mass":0.9999993742010536,"accepted":true,"accepted_predicted":"true","input_tokens":367,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.362757277361981e-07,"brier_hard":1.3987130173841008e-12,"soft_accuracy":0.6833326933662525,"brier_soft":0.2005549184961376,"tvd":0.31666616372462153,"kl":2.6564166810144725,"case_id":"typed:196/r0","base_id":"typed:196","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.192418}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.24608158673294e-05,"1":4.2635501553988765e-05,"2":0.03358326878383167,"3":0.9663416348987467},"predicted":"3","confidence":0.9663416348987467,"entropy_confidence":0.8933686456598435,"candidate_mass":0.9999976359506012,"accepted":true,"accepted_predicted":"3","input_tokens":393,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.034237847976861274,"brier_hard":0.002260724354986909,"score_mae":0.44956707776545723,"within_one":true,"case_id":"typed:196/r0","base_id":"typed:196","repeat":0,"workflow":"customer_service","case_elapsed_ms":335.192418}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.00025067931955492746,"close_no_action":0.00011959896914128634,"escalate_to_human":0.9979155486497825,"execute_refund":0.0001039171148564771,"request_information":0.0016102559466632063},"predicted":"escalate_to_human","confidence":0.9979155486497825,"entropy_confidence":0.9897168172179001,"candidate_mass":0.9999871909038248,"accepted":true,"accepted_predicted":"escalate_to_human","input_tokens":500,"reused_prefix_tokens":0,"expected":"escalate_to_human","correct":true,"accepted_correct":true,"nll_hard":0.0020866268425976235,"brier_hard":7.025804446620088e-06,"soft_accuracy":0.8650285454865819,"brier_soft":0.027537149908847713,"tvd":0.1313524657646398,"kl":0.44070660693586816,"case_id":"typed:197/r0","base_id":"typed:197","repeat":0,"workflow":"customer_service","case_elapsed_ms":366.92422000000005}
{"id":"category","type":"choice","probabilities":{"account":9.709295422182785e-05,"billing":0.00023982795606755785,"delivery":6.671440224066592e-05,"refund":2.5712134060880968e-05,"technical":0.9995706525534105},"predicted":"technical","confidence":0.9995706525534105,"entropy_confidence":0.9973663532800147,"candidate_mass":0.9999937972206496,"accepted":true,"accepted_predicted":"technical","input_tokens":490,"reused_prefix_tokens":0,"expected":"technical","correct":true,"accepted_correct":true,"nll_hard":0.0004294396425948158,"brier_hard":2.5639564546829235e-07,"soft_accuracy":0.9329427270201469,"brier_soft":0.006077725240172574,"tvd":0.06626336468747067,"kl":0.28072616717587806,"case_id":"typed:197/r0","base_id":"typed:197","repeat":0,"workflow":"customer_service","case_elapsed_ms":366.92422000000005}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":9.594166708202622e-07,"1":9.94836538431078e-07,"2":5.5366824097218836e-05,"3":0.9999426789226918},"predicted":"3","confidence":0.9999426789226918,"entropy_confidence":0.9995476841404908,"candidate_mass":0.999999664690342,"accepted":true,"accepted_predicted":"3","input_tokens":476,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.732272022389359e-05,"brier_hard":6.353101294466752e-09,"score_mae":0.2832727652528084,"within_one":true,"case_id":"typed:197/r0","base_id":"typed:197","repeat":0,"workflow":"customer_service","case_elapsed_ms":366.92422000000005}
{"id":"needs_human","type":"binary","probabilities":{"false":2.7861275794151727e-09,"true":0.9999999972138731},"predicted":"true","confidence":0.9999999972138731,"entropy_confidence":0.9999999168012539,"candidate_mass":0.9999999749130989,"accepted":true,"accepted_predicted":"true","input_tokens":444,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.7861268886385742e-09,"brier_hard":1.5525009906745395e-17,"soft_accuracy":0.8766669979011159,"brier_soft":0.030422056403514308,"tvd":0.1233329972138728,"kl":1.0205449330905976,"case_id":"typed:197/r0","base_id":"typed:197","repeat":0,"workflow":"customer_service","case_elapsed_ms":366.92422000000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.159905521214495e-06,"1":2.117248083269371e-06,"2":0.006612149754377009,"3":0.9933825730920183},"predicted":"3","confidence":0.9933825730920183,"entropy_confidence":0.9712553614108288,"candidate_mass":0.9999992548389974,"accepted":true,"accepted_predicted":"3","input_tokens":470,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.018846450402496,"brier_hard":1.9736283575530071,"score_mae":0.6667071360328922,"within_one":true,"case_id":"typed:197/r0","base_id":"typed:197","repeat":0,"workflow":"customer_service","case_elapsed_ms":366.92422000000005}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0009711762888286491,"close_no_action":7.523855836152306e-05,"escalate_to_human":0.00041566949112338585,"execute_refund":0.9911904727365176,"request_information":0.007347442925169603},"predicted":"execute_refund","confidence":0.9911904727365176,"entropy_confidence":0.9654791080811704,"candidate_mass":0.9999521798012648,"accepted":true,"accepted_predicted":"execute_refund","input_tokens":332,"reused_prefix_tokens":0,"expected":"execute_refund","correct":true,"accepted_correct":true,"nll_hard":0.008848560561140986,"brier_hard":0.0001327143134951624,"soft_accuracy":0.46353509813278565,"brier_soft":0.39711004685445084,"tvd":0.5245991779614121,"kl":2.5695226233954727,"case_id":"typed:198/r0","base_id":"typed:198","repeat":0,"workflow":"customer_service","case_elapsed_ms":307.62831600000004}
{"id":"category","type":"choice","probabilities":{"account":8.475664164235553e-12,"billing":0.9999999995071391,"delivery":3.077761072159889e-10,"refund":1.5919562290304985e-10,"technical":1.74137035886076e-11},"predicted":"billing","confidence":0.9999999995071391,"entropy_confidence":0.9999999928715861,"candidate_mass":0.9999999990959463,"accepted":true,"accepted_predicted":"billing","input_tokens":322,"reused_prefix_tokens":0,"expected":"billing","correct":true,"accepted_correct":true,"nll_hard":4.928608633540919e-10,"brier_hard":3.6335628298663375e-19,"soft_accuracy":0.8666669995890558,"brier_soft":0.02888877761416661,"tvd":0.13333299983232885,"kl":1.1040218677217128,"case_id":"typed:198/r0","base_id":"typed:198","repeat":0,"workflow":"customer_service","case_elapsed_ms":307.62831600000004}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.0001277797945696664,"1":0.00040508927888007764,"2":0.9268904438806126,"3":0.07257668704593688},"predicted":"2","confidence":0.9268904438806126,"entropy_confidence":0.8088024647956235,"candidate_mass":0.9999952096479546,"accepted":true,"accepted_predicted":"2","input_tokens":308,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0759199039165637,"brier_hard":0.010612563123537483,"score_mae":0.0885830381779158,"within_one":true,"case_id":"typed:198/r0","base_id":"typed:198","repeat":0,"workflow":"customer_service","case_elapsed_ms":307.62831600000004}
{"id":"needs_human","type":"binary","probabilities":{"false":5.077965573306437e-08,"true":0.9999999492203435},"predicted":"true","confidence":0.9999999492203435,"entropy_confidence":0.9999986962897622,"candidate_mass":0.9999998401513949,"accepted":true,"accepted_predicted":"true","input_tokens":276,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.077965776899153e-08,"brier_hard":5.157146948565355e-15,"soft_accuracy":0.5333329966147231,"brier_soft":0.43555608298924614,"tvd":0.4666669492203439,"kl":3.962903861962535,"case_id":"typed:198/r0","base_id":"typed:198","repeat":0,"workflow":"customer_service","case_elapsed_ms":307.62831600000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.6606324028972296e-06,"1":1.3273585156605459e-06,"2":3.0990174592748536e-05,"3":0.9999660218344897},"predicted":"3","confidence":0.9999660218344897,"entropy_confidence":0.9997145086220883,"candidate_mass":0.9999987095278313,"accepted":true,"accepted_predicted":"3","input_tokens":302,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.3978742781259906e-05,"brier_hard":2.1194262333422315e-09,"score_mae":0.18329437321117048,"within_one":true,"case_id":"typed:198/r0","base_id":"typed:198","repeat":0,"workflow":"customer_service","case_elapsed_ms":307.62831600000004}
{"id":"action","type":"choice","probabilities":{"answer_directly":0.0036776300011722098,"close_no_action":0.0010689864613013427,"escalate_to_human":0.011668150432650428,"execute_refund":0.00019688754930015514,"request_information":0.983388345555577},"predicted":"request_information","confidence":0.983388345555577,"entropy_confidence":0.939100167526088,"candidate_mass":0.9999030959983296,"accepted":true,"accepted_predicted":"request_information","input_tokens":451,"reused_prefix_tokens":0,"expected":"answer_directly","correct":false,"accepted_correct":false,"nll_hard":5.60548675569277,"brier_hard":1.9598482303658966,"soft_accuracy":0.2883552658094711,"brier_soft":0.8273155165272986,"tvd":0.6935852331048766,"kl":2.9302779814692856,"case_id":"typed:199/r0","base_id":"typed:199","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.605318}
{"id":"category","type":"choice","probabilities":{"account":0.9999848373931554,"billing":1.36667068797561e-05,"delivery":9.25918332088785e-07,"refund":2.5257260463021455e-07,"technical":3.174090272314027e-07},"predicted":"account","confidence":0.9999848373931554,"entropy_confidence":0.9998821409729717,"candidate_mass":0.9999934728602342,"accepted":true,"accepted_predicted":"account","input_tokens":441,"reused_prefix_tokens":0,"expected":"account","correct":true,"accepted_correct":true,"nll_hard":1.516272179812826e-05,"brier_hard":4.1770538943126627e-10,"soft_accuracy":0.9199867042028369,"brier_soft":0.00895184457634247,"tvd":0.0799848373931558,"kl":0.6099859160232439,"case_id":"typed:199/r0","base_id":"typed:199","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.605318}
{"id":"churn_risk","type":"ordinal","probabilities":{"0":0.001104279580709196,"1":0.9907402044146906,"2":0.006901964978287707,"3":0.0012535510263140786},"predicted":"1","confidence":0.9907402044146906,"entropy_confidence":0.9571121971143328,"candidate_mass":0.9999969107267864,"accepted":true,"accepted_predicted":"1","input_tokens":427,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.808562119791919,"brier_hard":1.9794080214263738,"score_mae":0.3916377874502083,"within_one":true,"case_id":"typed:199/r0","base_id":"typed:199","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.605318}
{"id":"needs_human","type":"binary","probabilities":{"false":1.2653174176440883e-05,"true":0.9999873468258246},"predicted":"true","confidence":0.9999873468258246,"entropy_confidence":0.999775876514269,"candidate_mass":0.9999960203763647,"accepted":true,"accepted_predicted":"true","input_tokens":395,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.27760245122786,"brier_hard":1.999949387623502,"soft_accuracy":0.40000253063483576,"brier_soft":0.7199696327021834,"tvd":0.5999873468258241,"kl":5.159692991737739,"case_id":"typed:199/r0","base_id":"typed:199","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.605318}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.928606779847455e-06,"1":2.281964777576876e-06,"2":0.0005362303531641799,"3":0.9994565590752801},"predicted":"3","confidence":0.9994565590752801,"entropy_confidence":0.9966302329401172,"candidate_mass":0.9999989189883339,"accepted":true,"accepted_predicted":"3","input_tokens":421,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.000543588642258843,"brier_hard":5.82900528843057e-07,"score_mae":0.22277741989694588,"within_one":true,"case_id":"typed:199/r0","base_id":"typed:199","repeat":0,"workflow":"customer_service","case_elapsed_ms":334.605318}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.021702842932691587,"1":0.0023137646356438737,"2":0.07959538596031068,"3":0.8963880064713547},"predicted":"3","confidence":0.8963880064713547,"entropy_confidence":0.7138708751975642,"candidate_mass":0.9999599119981821,"accepted":true,"accepted_predicted":"3","input_tokens":390,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.530799153134584,"brier_hard":1.6511324785893893,"score_mae":1.547335555970329,"within_one":false,"case_id":"typed:200/r0","base_id":"typed:200","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.254618}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9975343893781109,"hold":0.0017479786625716532,"manual_review":0.00037823461049768157,"reject":0.00033939734881802624},"predicted":"approve","confidence":0.9975343893781109,"entropy_confidence":0.9861121058770785,"candidate_mass":0.9999708294110481,"accepted":true,"accepted_predicted":"approve","input_tokens":394,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0024686552453587643,"brier_hard":9.392917124541192e-06,"soft_accuracy":0.4393818519759824,"brier_soft":0.4700032454993815,"tvd":0.5575343893781117,"kl":3.0099522760754795,"case_id":"typed:200/r0","base_id":"typed:200","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.254618}
{"id":"duplicate","type":"binary","probabilities":{"false":0.008504079031231435,"true":0.9914959209687699},"predicted":"true","confidence":0.9914959209687699,"entropy_confidence":0.9292956203089382,"candidate_mass":0.9999536683887591,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.767209344567114,"brier_hard":1.9661283225954158,"soft_accuracy":0.04454744089955516,"brier_soft":1.8233965366367686,"tvd":0.9548289209687693,"kl":4.435520204080305,"case_id":"typed:200/r0","base_id":"typed:200","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.254618}
{"id":"matches_order","type":"binary","probabilities":{"false":7.554259386633524e-08,"true":0.9999999244574067},"predicted":"true","confidence":0.9999999244574067,"entropy_confidence":0.999998103818238,"candidate_mass":0.9999997862199377,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.39856918259826,"brier_hard":1.999999697829637,"soft_accuracy":0.48000000302170404,"brier_soft":0.5407998428714168,"tvd":0.5199999244574065,"kl":4.437071825649565,"case_id":"typed:200/r0","base_id":"typed:200","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.254618}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.8173832157517875e-05,"1":4.2289924975928826e-05,"2":0.00011000442533255982,"3":0.9997895318175342},"predicted":"3","confidence":0.9997895318175342,"entropy_confidence":0.9984084541416169,"candidate_mass":0.9999976017890881,"accepted":true,"accepted_predicted":"3","input_tokens":389,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.114989962685076,"brier_hard":1.999359116354865,"score_mae":1.0162978942282432,"within_one":false,"case_id":"typed:200/r0","base_id":"typed:200","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.254618}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0008571012373441159,"1":0.0011540599472855053,"2":0.9603598190476547,"3":0.03762901976771537},"predicted":"2","confidence":0.9603598190476547,"entropy_confidence":0.8729521258322803,"candidate_mass":0.9999885024511945,"accepted":true,"accepted_predicted":"2","input_tokens":398,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.279979723939693,"brier_hard":1.8484509521113854,"score_mae":0.7552392426542589,"within_one":true,"case_id":"typed:201/r0","base_id":"typed:201","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.80341699999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9935470940072321,"hold":0.004770970280142367,"manual_review":0.000941982510826336,"reject":0.0007399532017999835},"predicted":"approve","confidence":0.9935470940072321,"entropy_confidence":0.968382321253049,"candidate_mass":0.9999857010623124,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":6.967523849562028,"brier_hard":1.9852760600077686,"soft_accuracy":0.03561043285126877,"brier_soft":1.3287169148828837,"tvd":0.9602140940072317,"kl":5.065522808866836,"case_id":"typed:201/r0","base_id":"typed:201","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.80341699999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.005911416308967402,"true":0.9940885836910328},"predicted":"true","confidence":0.9940885836910328,"entropy_confidence":0.9477389247691397,"candidate_mass":0.9999489124227416,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.130869830087313,"brier_hard":1.9764242244496866,"soft_accuracy":0.035556731330429366,"brier_soft":1.8589335944067626,"tvd":0.9640885836910327,"kl":4.842379435741774,"case_id":"typed:201/r0","base_id":"typed:201","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.80341699999997}
{"id":"matches_order","type":"binary","probabilities":{"false":0.00035456663061285995,"true":0.9996454333693874},"predicted":"true","confidence":0.9996454333693874,"entropy_confidence":0.9954246385649441,"candidate_mass":0.9999801440208453,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.94461427305891,"brier_hard":1.9985819849125401,"soft_accuracy":0.05031910996755159,"brier_soft":1.8036528982386626,"tvd":0.9496454333693872,"kl":7.348886047535303,"case_id":"typed:201/r0","base_id":"typed:201","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.80341699999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00015822528516498867,"1":0.939555434270004,"2":0.06015365362394841,"3":0.00013268682088369226},"predicted":"1","confidence":0.939555434270004,"entropy_confidence":0.8339225826590223,"candidate_mass":0.9999904633248097,"accepted":true,"accepted_predicted":"1","input_tokens":397,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.8108530965283833,"brier_hard":1.7660756115037421,"score_mae":0.689739198019448,"within_one":true,"case_id":"typed:201/r0","base_id":"typed:201","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.80341699999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999612452624379,"1":1.7654015241171146e-05,"2":1.5875118640107886e-05,"3":5.22560368239119e-06},"predicted":"0","confidence":0.9999612452624379,"entropy_confidence":0.9996602779529111,"candidate_mass":0.9999971831194995,"accepted":true,"accepted_predicted":"0","input_tokens":398,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.87554885463413e-05,"brier_hard":2.0929202633254122e-09,"score_mae":0.026601918936431438,"within_one":true,"case_id":"typed:202/r0","base_id":"typed:202","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.942118}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999976129839464,"hold":1.312115632703405e-06,"manual_review":8.728356668108539e-07,"reject":2.0206475476609872e-07},"predicted":"approve","confidence":0.9999976129839464,"entropy_confidence":0.999974428026653,"candidate_mass":0.9999985074475771,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.387018902522835e-06,"brier_hard":8.222165340083033e-12,"soft_accuracy":0.9633307375987291,"brier_soft":0.0019109063346566181,"tvd":0.03666461298394606,"kl":0.30173145466148693,"case_id":"typed:202/r0","base_id":"typed:202","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.942118}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0059493316584738,"true":0.9940506683415268},"predicted":"true","confidence":0.9940506683415268,"entropy_confidence":0.9474587651250223,"candidate_mass":0.9999638528005507,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.124476392040595,"brier_hard":1.976273462460471,"soft_accuracy":0.022418016636970246,"brier_soft":1.9105576702814782,"tvd":0.9773836683415265,"kl":4.954398762770221,"case_id":"typed:202/r0","base_id":"typed:202","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.942118}
{"id":"matches_order","type":"binary","probabilities":{"false":2.960336711762592e-09,"true":0.999999997039664},"predicted":"true","confidence":0.999999997039664,"entropy_confidence":0.9999999118580873,"candidate_mass":0.9999999621047138,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.9603359807359393e-09,"brier_hard":1.7527182539906005e-17,"soft_accuracy":0.9866669971186044,"brier_soft":0.00035553762011936174,"tvd":0.013332997039663695,"kl":0.10955775323176194,"case_id":"typed:202/r0","base_id":"typed:202","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.942118}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000915524997874149,"1":0.9730849727827492,"2":0.02542871694886631,"3":0.0005707852705112498},"predicted":"1","confidence":0.9730849727827492,"entropy_confidence":0.905800268748449,"candidate_mass":0.9999935136633419,"accepted":true,"accepted_predicted":"1","input_tokens":397,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.671876155053514,"brier_hard":1.8966847139853835,"score_mae":0.9743452375079844,"within_one":true,"case_id":"typed:202/r0","base_id":"typed:202","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.942118}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.011623022933759976,"1":0.0017389574981072147,"2":0.21722920296534057,"3":0.7694088166027926},"predicted":"3","confidence":0.7694088166027926,"entropy_confidence":0.569946052902494,"candidate_mass":0.9999813655194382,"accepted":false,"accepted_predicted":null,"input_tokens":424,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.5268022480019212,"brier_hard":1.2048581663916846,"score_mae":0.7444238132371659,"within_one":true,"case_id":"typed:203/r0","base_id":"typed:203","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.71531799999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9994008441077178,"hold":0.00041994682160965437,"manual_review":0.00010758499262100147,"reject":7.16240780516464e-05},"predicted":"approve","confidence":0.9994008441077178,"entropy_confidence":0.9960103436131045,"candidate_mass":0.999985797638317,"accepted":true,"accepted_predicted":"approve","input_tokens":428,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":7.775382469918352,"brier_hard":1.9979623466198717,"soft_accuracy":0.03359163493296044,"brier_soft":1.3693969259531702,"tvd":0.9660678441077177,"kl":7.106116617187874,"case_id":"typed:203/r0","base_id":"typed:203","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.71531799999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.009072412517819236,"true":0.9909275874821806},"predicted":"true","confidence":0.9909275874821806,"entropy_confidence":0.9254208784756053,"candidate_mass":0.9999427378889248,"accepted":true,"accepted_predicted":"true","input_tokens":376,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.702517061489619,"brier_hard":1.9638749672665097,"soft_accuracy":0.048346619516393685,"brier_soft":1.8085265532693606,"tvd":0.9509275874821806,"kl":4.346836783994704,"case_id":"typed:203/r0","base_id":"typed:203","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.71531799999997}
{"id":"matches_order","type":"binary","probabilities":{"false":2.121039116610644e-08,"true":0.9999999787896086},"predicted":"true","confidence":0.9999999787896086,"entropy_confidence":0.9999994287331413,"candidate_mass":0.9999998128457948,"accepted":true,"accepted_predicted":"true","input_tokens":396,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.668774626007753,"brier_hard":1.9999999151584356,"soft_accuracy":0.08333301767534008,"brier_soft":1.6805567000065378,"tvd":0.9166669787896087,"kl":8.235740020346368,"case_id":"typed:203/r0","base_id":"typed:203","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.71531799999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001938147535709895,"1":0.00012633326241774615,"2":0.0004614666584430033,"3":0.9992183853255682},"predicted":"3","confidence":0.9992183853255682,"entropy_confidence":0.9948663364151572,"candidate_mass":0.9999960955108504,"accepted":true,"accepted_predicted":"3","input_tokens":423,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.681100752691904,"brier_hard":1.9975147147314782,"score_mae":0.765371422556008,"within_one":true,"case_id":"typed:203/r0","base_id":"typed:203","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":327.71531799999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.778656536940982,"1":0.006462354083549737,"2":0.06303561013062033,"3":0.15184549884484716},"predicted":"0","confidence":0.778656536940982,"entropy_confidence":0.5038310165454393,"candidate_mass":0.999903371187335,"accepted":false,"accepted_predicted":null,"input_tokens":391,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.8848917300446502,"brier_hard":1.3296873105155098,"score_mae":2.211929929120668,"within_one":false,"case_id":"typed:204/r0","base_id":"typed:204","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.88811699999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999676712325959,"hold":1.8761134095991224e-05,"manual_review":1.1099089926679372e-05,"reject":2.468543382717422e-06},"predicted":"approve","confidence":0.9999676712325959,"entropy_confidence":0.9997150544710043,"candidate_mass":0.9999951439596506,"accepted":true,"accepted_predicted":"approve","input_tokens":395,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":11.408647441594677,"brier_hard":1.9999131458117512,"soft_accuracy":0.02001371016545724,"brier_soft":1.42153014143869,"tvd":0.9799676712325952,"kl":8.943684129563328,"case_id":"typed:204/r0","base_id":"typed:204","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.88811699999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.019293608927596826,"true":0.9807063910724021},"predicted":"true","confidence":0.9807063910724021,"entropy_confidence":0.8625442184755324,"candidate_mass":0.9999252832329566,"accepted":true,"accepted_predicted":"true","input_tokens":343,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.947981381559295,"brier_hard":1.9235700509805127,"soft_accuracy":0.0353174757676043,"brier_soft":1.858743895078498,"tvd":0.9640393910724027,"kl":3.797737650138957,"case_id":"typed:204/r0","base_id":"typed:204","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.88811699999997}
{"id":"matches_order","type":"binary","probabilities":{"false":2.7194128576310814e-08,"true":0.9999999728058707},"predicted":"true","confidence":0.9999999728058707,"entropy_confidence":0.9999992773208023,"candidate_mass":0.999999884561706,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.420264748110878,"brier_hard":1.9999998912234858,"soft_accuracy":0.05666702411210917,"brier_soft":1.7797541951655242,"tvd":0.943332972805871,"kl":8.525751808069025,"case_id":"typed:204/r0","base_id":"typed:204","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.88811699999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00025098394929266296,"1":0.7515694830495329,"2":0.24785218322575187,"3":0.0003273497754209101},"predicted":"1","confidence":0.7515694830495329,"entropy_confidence":0.5923779928411175,"candidate_mass":0.9999830708268127,"accepted":false,"accepted_predicted":null,"input_tokens":390,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3949227458968885,"brier_hard":1.1305831962804282,"score_mae":0.7350771011727006,"within_one":true,"case_id":"typed:204/r0","base_id":"typed:204","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.88811699999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.021721684113340623,"1":0.002117002168944613,"2":0.06172925543520054,"3":0.9144320582825157},"predicted":"3","confidence":0.9144320582825157,"entropy_confidence":0.7475776774024151,"candidate_mass":0.999969709226241,"accepted":true,"accepted_predicted":"3","input_tokens":417,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.7849973042852234,"brier_hard":1.7170142925798844,"score_mae":1.0355386878868926,"within_one":false,"case_id":"typed:205/r0","base_id":"typed:205","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.957318}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998409707882889,"hold":8.80605382507359e-05,"manual_review":5.084440542024351e-05,"reject":2.012426804119126e-05},"predicted":"approve","confidence":0.9998409707882889,"entropy_confidence":0.998772568390894,"candidate_mass":0.9999834251308027,"accepted":true,"accepted_predicted":"approve","input_tokens":421,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.886740462756634,"brier_hard":1.9995802888008256,"soft_accuracy":0.21667784578009666,"brier_soft":1.0013256827192445,"tvd":0.7831741874550717,"kl":6.564476423740458,"case_id":"typed:205/r0","base_id":"typed:205","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.957318}
{"id":"duplicate","type":"binary","probabilities":{"false":0.027720161208863704,"true":0.9722798387911348},"predicted":"true","confidence":0.9722798387911348,"entropy_confidence":0.8171735521165808,"candidate_mass":0.9999659881575113,"accepted":true,"accepted_predicted":"true","input_tokens":369,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.5855952890617715,"brier_hard":1.890656169839433,"soft_accuracy":0.07494814508797726,"brier_soft":1.7012002020812058,"tvd":0.9222798387911355,"kl":3.2092058620611135,"case_id":"typed:205/r0","base_id":"typed:205","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.957318}
{"id":"matches_order","type":"binary","probabilities":{"false":1.954963445885284e-08,"true":0.9999999804503654},"predicted":"true","confidence":0.9999999804503654,"entropy_confidence":0.9999994711632652,"candidate_mass":0.9999998974949981,"accepted":true,"accepted_predicted":"true","input_tokens":389,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.75030924844673,"brier_hard":1.9999999218014626,"soft_accuracy":0.26666700912314967,"brier_soft":1.0755545204324322,"tvd":0.7333329804503654,"kl":6.401778209968629,"case_id":"typed:205/r0","base_id":"typed:205","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.957318}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00791871684636306,"1":0.6304597607288436,"2":0.32362483784701734,"3":0.03799668457777422},"predicted":"1","confidence":0.6304597607288436,"entropy_confidence":0.40956796640332527,"candidate_mass":0.9999918108092463,"accepted":false,"accepted_predicted":null,"input_tokens":416,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.128170341962651,"brier_hard":0.8564693239911401,"score_mae":0.4749675098437991,"within_one":true,"case_id":"typed:205/r0","base_id":"typed:205","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.957318}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.02807117745811135,"1":0.01179770335387763,"2":0.744146383311976,"3":0.21598473587603426},"predicted":"2","confidence":0.744146383311976,"entropy_confidence":0.49246431826919357,"candidate_mass":0.999977206486962,"accepted":false,"accepted_predicted":null,"input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.2955175117387263,"brier_hard":0.11303765611209346,"score_mae":0.6980446776059324,"within_one":true,"case_id":"typed:206/r0","base_id":"typed:206","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.756317}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9992871462337904,"hold":0.0005200616210450566,"manual_review":0.0001339612570233784,"reject":5.8830888140925e-05},"predicted":"approve","confidence":0.9992871462337904,"entropy_confidence":0.9953741411307181,"candidate_mass":0.9999834944738057,"accepted":true,"accepted_predicted":"approve","input_tokens":414,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":8.917959926509749,"brier_hard":1.9983071699848076,"soft_accuracy":0.38320544710523463,"brier_soft":0.5949415538443852,"tvd":0.6159541462337906,"kl":4.190532927088464,"case_id":"typed:206/r0","base_id":"typed:206","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.756317}
{"id":"duplicate","type":"binary","probabilities":{"false":0.00667242883372266,"true":0.9933275711662777},"predicted":"true","confidence":0.9933275711662777,"entropy_confidence":0.9421804453905119,"candidate_mass":0.9999677052763126,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.009771342351666,"brier_hard":1.9733993272781922,"soft_accuracy":0.032983561514304906,"brier_soft":1.8688653196950278,"tvd":0.9666605711662775,"kl":4.753396101464731,"case_id":"typed:206/r0","base_id":"typed:206","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.756317}
{"id":"matches_order","type":"binary","probabilities":{"false":1.927554013041646e-07,"true":0.9999998072445998},"predicted":"true","confidence":0.9999998072445998,"entropy_confidence":0.999995422170965,"candidate_mass":0.9999996811319154,"accepted":true,"accepted_predicted":"true","input_tokens":382,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.461843802565294,"brier_hard":1.9999992289784714,"soft_accuracy":0.3833330449763892,"brier_soft":0.7605559023144954,"tvd":0.6166668072445993,"kl":5.312153726926247,"case_id":"typed:206/r0","base_id":"typed:206","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.756317}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00037983119359675166,"1":0.9011601575601581,"2":0.09813297574761219,"3":0.00032703549863442004},"predicted":"1","confidence":0.9011601575601581,"entropy_confidence":0.7639672434006228,"candidate_mass":0.9999894355470236,"accepted":true,"accepted_predicted":"1","input_tokens":409,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.10407228183171918,"brier_hard":0.01939964660656721,"score_mae":0.4349257844487142,"within_one":true,"case_id":"typed:206/r0","base_id":"typed:206","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.756317}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.8123420560812767,"1":0.002497637004708765,"2":0.025107589608623065,"3":0.1600527173053926},"predicted":"0","confidence":0.8123420560812767,"entropy_confidence":0.5891442927849879,"candidate_mass":0.9999071815636369,"accepted":true,"accepted_predicted":"0","input_tokens":386,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":1.8322520348572624,"brier_hard":1.3660476830309731,"score_mae":2.370462031861867,"within_one":false,"case_id":"typed:207/r0","base_id":"typed:207","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.99851800000005}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9982793490033696,"hold":0.0010095059299793582,"manual_review":0.0003964154961678354,"reject":0.0003147295704823013},"predicted":"approve","confidence":0.9982793490033696,"entropy_confidence":0.9896659064953178,"candidate_mass":0.9999712975064996,"accepted":true,"accepted_predicted":"approve","input_tokens":390,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.833047664022686,"brier_hard":1.9957701029564263,"soft_accuracy":0.013805144940814949,"brier_soft":1.5757309996231321,"tvd":0.98494634900337,"kl":6.844670740747605,"case_id":"typed:207/r0","base_id":"typed:207","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.99851800000005}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0021062012659823384,"true":0.9978937987340192},"predicted":"true","confidence":0.9978937987340192,"entropy_confidence":0.9782380636982603,"candidate_mass":0.9999634011276416,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.162869301943825,"brier_hard":1.9915840671036193,"soft_accuracy":0.012064077240662708,"brier_soft":1.9518683151542584,"tvd":0.9878937987340184,"kl":6.045260158793813,"case_id":"typed:207/r0","base_id":"typed:207","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.99851800000005}
{"id":"matches_order","type":"binary","probabilities":{"false":3.884989259710967e-08,"true":0.9999999611501068},"predicted":"true","confidence":0.9999999611501068,"entropy_confidence":0.9999989875636721,"candidate_mass":0.9999998514654361,"accepted":true,"accepted_predicted":"true","input_tokens":358,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.063560524689738,"brier_hard":1.9999998446004315,"soft_accuracy":0.01666703755487027,"brier_soft":1.933887424968476,"tvd":0.9833329611501072,"kl":8.98859152216208,"case_id":"typed:207/r0","base_id":"typed:207","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.99851800000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00014240575372361525,"1":9.023693262382651e-05,"2":0.0002703283367379189,"3":0.9994970289769137},"predicted":"3","confidence":0.9994970289769137,"entropy_confidence":0.9965191559544363,"candidate_mass":0.9999968552986181,"accepted":true,"accepted_predicted":"3","input_tokens":385,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.215873275778097,"brier_hard":1.9984537557597142,"score_mae":0.8924549805368409,"within_one":true,"case_id":"typed:207/r0","base_id":"typed:207","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.99851800000005}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9648267886040736,"1":0.0005592304847680149,"2":0.004027580960313593,"3":0.03058639995084588},"predicted":"0","confidence":0.9648267886040736,"entropy_confidence":0.8790973636468378,"candidate_mass":0.9999789144198736,"accepted":true,"accepted_predicted":"0","input_tokens":372,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.48719981487001,"brier_hard":1.8706699941154379,"score_mae":2.6929594077420673,"within_one":false,"case_id":"typed:208/r0","base_id":"typed:208","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.398619}
{"id":"disposition","type":"choice","probabilities":{"approve":0.999537030025087,"hold":0.0003291604248258761,"manual_review":9.09951583690534e-05,"reject":4.281439171796031e-05},"predicted":"approve","confidence":0.999537030025087,"entropy_confidence":0.9968406953314486,"candidate_mass":0.9999871300598617,"accepted":true,"accepted_predicted":"approve","input_tokens":376,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.304704257598523,"brier_hard":1.9988924025344097,"soft_accuracy":0.016826998976760238,"brier_soft":1.4100649504536273,"tvd":0.982870030025087,"kl":7.814742299846699,"case_id":"typed:208/r0","base_id":"typed:208","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.398619}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004799207887376465,"true":0.9952007921126227},"predicted":"true","confidence":0.9952007921126227,"entropy_confidence":0.9561246069296175,"candidate_mass":0.9999531819191806,"accepted":true,"accepted_predicted":"true","input_tokens":324,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.339304398149398,"brier_hard":1.980849233243185,"soft_accuracy":0.04111426277616356,"brier_soft":1.8375740612436107,"tvd":0.9585337921126231,"kl":4.986501492010888,"case_id":"typed:208/r0","base_id":"typed:208","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.398619}
{"id":"matches_order","type":"binary","probabilities":{"false":1.633637545087719e-09,"true":0.9999999983663628},"predicted":"true","confidence":0.9999999983663628,"entropy_confidence":0.9999999499584806,"candidate_mass":0.9999999371390287,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":20.232456685746186,"brier_hard":1.9999999934654504,"soft_accuracy":0.08333300136136573,"brier_soft":1.6805567717879941,"tvd":0.9166669983663627,"kl":8.235740018714978,"case_id":"typed:208/r0","base_id":"typed:208","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.398619}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022522177660778766,"1":0.931813763044422,"2":0.06729378336582763,"3":0.0006672318131431803},"predicted":"1","confidence":0.931813763044422,"entropy_confidence":0.8166462408293124,"candidate_mass":0.9999912306770288,"accepted":true,"accepted_predicted":"1","input_tokens":371,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.698687418576796,"brier_hard":1.7382182714701793,"score_mae":0.5982639747844931,"within_one":true,"case_id":"typed:208/r0","base_id":"typed:208","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.398619}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.999895338732569,"1":4.872105634076664e-05,"2":3.968438715816624e-05,"3":1.6255823931315183e-05},"predicted":"0","confidence":0.999895338732569,"entropy_confidence":0.9991561214616588,"candidate_mass":0.9999955613202787,"accepted":true,"accepted_predicted":"0","input_tokens":398,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":10.13455271819744,"brier_hard":1.9997113238576463,"score_mae":1.3331561426975491,"within_one":false,"case_id":"typed:209/r0","base_id":"typed:209","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":313.93701699999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999836787791749,"hold":9.97000165657343e-06,"manual_review":4.932008044769508e-06,"reject":1.4192111232092615e-06},"predicted":"approve","confidence":0.9999836787791749,"entropy_confidence":0.999848146852777,"candidate_mass":0.9999964384959809,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":11.51592980783473,"brier_hard":1.9999474179471588,"soft_accuracy":0.25666811798830425,"brier_soft":0.8288311042405567,"tvd":0.7433183546570418,"kl":6.497295688374843,"case_id":"typed:209/r0","base_id":"typed:209","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":313.93701699999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0011955565056923744,"true":0.9988044434943084},"predicted":"true","confidence":0.9988044434943084,"entropy_confidence":0.9866696235669162,"candidate_mass":0.9999734094424692,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.729143506860929,"brier_hard":1.9952206326879487,"soft_accuracy":0.02447276466579775,"brier_soft":1.9030890741457378,"tvd":0.9754714434943079,"kl":6.461418902409184,"case_id":"typed:209/r0","base_id":"typed:209","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":313.93701699999997}
{"id":"matches_order","type":"binary","probabilities":{"false":9.883008436984316e-09,"true":0.9999999901169918},"predicted":"true","confidence":0.9999999901169918,"entropy_confidence":0.999999722928894,"candidate_mass":0.9999999542109342,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.883008238440895e-09,"brier_hard":1.9534770664128233e-16,"soft_accuracy":0.6333329973645379,"brier_soft":0.26888936328290813,"tvd":0.36666699011699166,"kl":3.087847376769562,"case_id":"typed:209/r0","base_id":"typed:209","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":313.93701699999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00019735786250670365,"1":0.9635678981111009,"2":0.036074487129579266,"3":0.0001602568968116865},"predicted":"1","confidence":0.9635678981111009,"entropy_confidence":0.8855295047166003,"candidate_mass":0.9999928566660666,"accepted":true,"accepted_predicted":"1","input_tokens":397,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.03711232333444773,"brier_hard":0.0026287313021041755,"score_mae":0.09713535693930542,"within_one":true,"case_id":"typed:209/r0","base_id":"typed:209","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":313.93701699999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999746128691624,"1":1.3717469364394222e-05,"2":8.588418733109979e-06,"3":3.0812427388222793e-06},"predicted":"0","confidence":0.9999746128691624,"entropy_confidence":0.9997704200648345,"candidate_mass":0.9999981050590114,"accepted":true,"accepted_predicted":"0","input_tokens":385,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":2.5387453096211083e-05,"brier_hard":9.159303710767983e-10,"score_mae":0.04029286196495292,"within_one":true,"case_id":"typed:210/r0","base_id":"typed:210","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.55661699999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999946402712003,"hold":3.2360340609997224e-06,"manual_review":1.8025401281513326e-06,"reject":3.211546114800375e-07},"predicted":"approve","confidence":0.9999946402712003,"entropy_confidence":0.9999459641462469,"candidate_mass":0.999998168562872,"accepted":true,"accepted_predicted":"approve","input_tokens":389,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":5.359743163067564e-06,"brier_hard":4.2550900448031354e-11,"soft_accuracy":0.8999954038018305,"brier_soft":0.014154030759290481,"tvd":0.09999464027119984,"kl":0.8262143968744216,"case_id":"typed:210/r0","base_id":"typed:210","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.55661699999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003125739922774634,"true":0.9968742600772263},"predicted":"true","confidence":0.9968742600772263,"entropy_confidence":0.9694864503432372,"candidate_mass":0.999960480596056,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.768084248532734,"brier_hard":1.9875165808090332,"soft_accuracy":0.01637538894199394,"brier_soft":1.9347068205485942,"tvd":0.9835412600772258,"kl":5.620411001791245,"case_id":"typed:210/r0","base_id":"typed:210","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.55661699999996}
{"id":"matches_order","type":"binary","probabilities":{"false":2.538970262326553e-09,"true":0.9999999974610301},"predicted":"true","confidence":0.9999999974610301,"entropy_confidence":0.9999999238415451,"candidate_mass":0.9999999693021562,"accepted":true,"accepted_predicted":"true","input_tokens":357,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.5389699264168743e-09,"brier_hard":1.2892738263860737e-17,"soft_accuracy":0.9766669975795137,"brier_soft":0.0010888575410328606,"tvd":0.023332997461029955,"kl":0.19184623106826315,"case_id":"typed:210/r0","base_id":"typed:210","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.55661699999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022463510898008992,"1":0.9913801021812015,"2":0.008278133462025683,"3":0.00011712924779305526},"predicted":"1","confidence":0.9913801021812015,"entropy_confidence":0.9630550327789243,"candidate_mass":0.9999935314630362,"accepted":true,"accepted_predicted":"1","input_tokens":384,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.794137763289104,"brier_hard":1.9663468317505663,"score_mae":0.8650452431513678,"within_one":true,"case_id":"typed:210/r0","base_id":"typed:210","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.55661699999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999266175540318,"1":3.581559100984084e-05,"2":2.5522537632368286e-05,"3":1.2044317325197557e-05},"predicted":"0","confidence":0.9999266175540318,"entropy_confidence":0.9993894669286445,"candidate_mass":0.9999963294191232,"accepted":true,"accepted_predicted":"0","input_tokens":412,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":7.3385138591634e-05,"brier_hard":7.464205442688266e-09,"score_mae":0.00854400638174983,"within_one":true,"case_id":"typed:211/r0","base_id":"typed:211","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.56861799999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999989493161379,"hold":5.673095863280124e-07,"manual_review":4.0003516068342006e-07,"reject":8.33391152193044e-08},"predicted":"approve","confidence":0.9999989493161379,"entropy_confidence":0.9999881254697731,"candidate_mass":0.9999990553328665,"accepted":true,"accepted_predicted":"approve","input_tokens":416,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.0506844141047017e-06,"brier_hard":1.5927502828012825e-12,"soft_accuracy":0.983331974622231,"brier_soft":0.00040618316740655754,"tvd":0.01666594931613781,"kl":0.1368338991350214,"case_id":"typed:211/r0","base_id":"typed:211","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.56861799999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.013602218101757631,"true":0.9863977818982426},"predicted":"true","confidence":0.9863977818982426,"entropy_confidence":0.8961762744459127,"candidate_mass":0.9999773569369359,"accepted":true,"accepted_predicted":"true","input_tokens":364,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.2975224038213184,"brier_hard":1.9459611682675453,"soft_accuracy":0.016844545715891314,"brier_soft":1.932832730817278,"tvd":0.9830647818982424,"kl":4.260905929324682,"case_id":"typed:211/r0","base_id":"typed:211","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.56861799999996}
{"id":"matches_order","type":"binary","probabilities":{"false":1.2546245849457566e-09,"true":0.9999999987453752},"predicted":"true","confidence":0.9999999987453752,"entropy_confidence":0.9999999610905881,"candidate_mass":0.9999999736027764,"accepted":true,"accepted_predicted":"true","input_tokens":384,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.254624849243884e-09,"brier_hard":3.1481663595156665e-18,"soft_accuracy":0.9966669987537384,"brier_soft":2.2217761273346832e-05,"tvd":0.0033329987453753138,"kl":0.027370626336032464,"case_id":"typed:211/r0","base_id":"typed:211","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.56861799999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.018601195523633554,"1":0.008486104770747875,"2":0.0007170137770144718,"3":0.972195685928604},"predicted":"3","confidence":0.972195685928604,"entropy_confidence":0.8938208024948688,"candidate_mass":0.9999911960473062,"accepted":true,"accepted_predicted":"3","input_tokens":411,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.769325185972789,"brier_hard":1.928610774754538,"score_mae":1.7765071901105887,"within_one":false,"case_id":"typed:211/r0","base_id":"typed:211","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.56861799999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999907105247343,"1":4.929002810048072e-06,"2":3.2758624900913922e-06,"3":1.0846099657408049e-06},"predicted":"0","confidence":0.9999907105247343,"entropy_confidence":0.9999092612098275,"candidate_mass":0.9999991278831574,"accepted":true,"accepted_predicted":"0","input_tokens":410,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":9.289518413126145e-06,"brier_hard":1.224970732449808e-10,"score_mae":1.0499852654423125,"within_one":false,"case_id":"typed:212/r0","base_id":"typed:212","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.568717}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9977071385818874,"hold":0.000984709962987843,"manual_review":0.00044524428381221973,"reject":0.000862907171310937},"predicted":"approve","confidence":0.9977071385818874,"entropy_confidence":0.9865602681581404,"candidate_mass":0.9999623386075254,"accepted":true,"accepted_predicted":"approve","input_tokens":414,"reused_prefix_tokens":0,"expected":"reject","correct":false,"accepted_correct":false,"nll_hard":7.055203437775383,"brier_hard":1.9936956325396054,"soft_accuracy":0.256699214203242,"brier_soft":0.8291120162537433,"tvd":0.7410401385818882,"kl":4.0853563627857765,"case_id":"typed:212/r0","base_id":"typed:212","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.568717}
{"id":"duplicate","type":"binary","probabilities":{"false":6.005246688734794e-06,"true":0.9999939947533123},"predicted":"true","confidence":0.9999939947533123,"entropy_confidence":0.9998871732104244,"candidate_mass":0.9999991214448536,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.005264719267495e-06,"brier_hard":7.212597557270904e-11,"soft_accuracy":0.7299972375865239,"brier_soft":0.1457935144057027,"tvd":0.26999399475331176,"kl":2.2570574405438335,"case_id":"typed:212/r0","base_id":"typed:212","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.568717}
{"id":"matches_order","type":"binary","probabilities":{"false":2.026851393240195e-09,"true":0.9999999979731484},"predicted":"true","confidence":0.9999999979731484,"entropy_confidence":0.9999999385442286,"candidate_mass":0.9999999722001917,"accepted":true,"accepted_predicted":"true","input_tokens":382,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.0268515775191798e-09,"brier_hard":8.216253879245144e-18,"soft_accuracy":0.9766669980677334,"brier_soft":0.0010888575888299101,"tvd":0.023332997973148563,"kl":0.19184623056809413,"case_id":"typed:212/r0","base_id":"typed:212","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.568717}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002702468884505638,"1":0.0001421861877703097,"2":0.0005517559646922251,"3":0.999035810959087},"predicted":"3","confidence":0.999035810959087,"entropy_confidence":0.9938085627535669,"candidate_mass":0.9999957039714695,"accepted":true,"accepted_predicted":"3","input_tokens":409,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.502404702387945,"brier_hard":1.9969694373342335,"score_mae":0.8150201309944154,"within_one":true,"case_id":"typed:212/r0","base_id":"typed:212","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.568717}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.009817466035956351,"1":0.0018040825455704769,"2":0.043590696410423964,"3":0.9447877550080487},"predicted":"3","confidence":0.9447877550080487,"entropy_confidence":0.8218166237134916,"candidate_mass":0.9999851653679808,"accepted":true,"accepted_predicted":"3","input_tokens":424,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.05679497458935499,"brier_hard":0.005048178163795391,"score_mae":0.05668174039056417,"within_one":true,"case_id":"typed:213/r0","base_id":"typed:213","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":333.884518}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9996109188664661,"hold":0.00024613475058873896,"manual_review":0.00010327158137565188,"reject":3.967480156987843e-05},"predicted":"approve","confidence":0.9996109188664661,"entropy_confidence":0.9972702533923353,"candidate_mass":0.9999796667018356,"accepted":true,"accepted_predicted":"approve","input_tokens":428,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.178148327377464,"brier_hard":1.9990155187757341,"soft_accuracy":0.016774231984136297,"brier_soft":1.6995624757482128,"tvd":0.9829439188664659,"kl":8.420206650544033,"case_id":"typed:213/r0","base_id":"typed:213","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":333.884518}
{"id":"duplicate","type":"binary","probabilities":{"false":0.002138793327660256,"true":0.9978612066723413},"predicted":"true","confidence":0.9978612066723413,"entropy_confidence":0.9779487449323873,"candidate_mass":0.9999780306392599,"accepted":true,"accepted_predicted":"true","input_tokens":376,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.147513474559677,"brier_hard":1.991453975563159,"soft_accuracy":0.041967689861447506,"brier_soft":1.8349961824955847,"tvd":0.9578612066723405,"kl":5.733754431195622,"case_id":"typed:213/r0","base_id":"typed:213","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":333.884518}
{"id":"matches_order","type":"binary","probabilities":{"false":5.11113989714704e-08,"true":0.9999999488886018},"predicted":"true","confidence":0.9999999488886018,"entropy_confidence":0.9999986882527961,"candidate_mass":0.9999998081061754,"accepted":true,"accepted_predicted":"true","input_tokens":396,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.789258292764842,"brier_hard":1.9999997955544109,"soft_accuracy":0.01666704940765161,"brier_soft":1.9338873767399056,"tvd":0.9833329488886013,"kl":8.988591522366441,"case_id":"typed:213/r0","base_id":"typed:213","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":333.884518}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008855277253618985,"1":0.25261029943595775,"2":0.7455636724002458,"3":0.0009405004384358185},"predicted":"2","confidence":0.7455636724002458,"entropy_confidence":0.5821551629686563,"candidate_mass":0.9999874201509877,"accepted":false,"accepted_predicted":null,"input_tokens":423,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.3759072961995926,"brier_hard":1.1144582228125768,"score_mae":0.5132261455517568,"within_one":true,"case_id":"typed:213/r0","base_id":"typed:213","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":333.884518}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9994105659261268,"1":0.0002374248292257011,"2":0.0002026466619026356,"3":0.0001493625827460152},"predicted":"0","confidence":0.9994105659261268,"entropy_confidence":0.9959533855841137,"candidate_mass":0.9999872104337053,"accepted":true,"accepted_predicted":"0","input_tokens":379,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":8.345659493112336,"brier_hard":1.9983467493717297,"score_mae":1.2155761940987309,"within_one":false,"case_id":"typed:214/r0","base_id":"typed:214","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.798618}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999882813872569,"hold":6.773463253414521e-06,"manual_review":4.050067818432297e-06,"reject":8.950816729372633e-07},"predicted":"approve","confidence":0.9999882813872569,"entropy_confidence":0.9998881235532363,"candidate_mass":0.9999965917584099,"accepted":true,"accepted_predicted":"approve","input_tokens":383,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.1718681406551137e-05,"brier_hard":2.0040990960258477e-10,"soft_accuracy":0.48333051790231063,"brier_soft":0.4405375049483026,"tvd":0.516656176468929,"kl":4.407277614762399,"case_id":"typed:214/r0","base_id":"typed:214","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.798618}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0026309081206164777,"true":0.9973690918793834},"predicted":"true","confidence":0.9973690918793834,"entropy_confidence":0.9736619288331685,"candidate_mass":0.9999509619852215,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.940426199382308,"brier_hard":1.989490210872612,"soft_accuracy":0.052367817308554825,"brier_soft":1.795016392496735,"tvd":0.9473690918793833,"kl":5.445021364819395,"case_id":"typed:214/r0","base_id":"typed:214","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.798618}
{"id":"matches_order","type":"binary","probabilities":{"false":2.562923162952027e-09,"true":0.9999999974370777},"predicted":"true","confidence":0.9999999974370777,"entropy_confidence":0.9999999231577785,"candidate_mass":0.9999999610076465,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.5629223221004486e-09,"brier_hard":1.313714595148204e-17,"soft_accuracy":0.5999999994874159,"brier_soft":0.31999999589932365,"tvd":0.39999999743707726,"kl":3.377640776068633,"case_id":"typed:214/r0","base_id":"typed:214","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.798618}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0056147916504289954,"1":0.7395759167479241,"2":0.2373137207354112,"3":0.01749557086623583},"predicted":"1","confidence":0.7395759167479241,"entropy_confidence":0.5207792304397247,"candidate_mass":0.9999909680557475,"accepted":false,"accepted_predicted":null,"input_tokens":378,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.4383722970037205,"brier_hard":1.1290005180972078,"score_mae":0.8666429291825459,"within_one":true,"case_id":"typed:214/r0","base_id":"typed:214","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.798618}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999859883392076,"1":7.446154195579869e-06,"2":4.964670837028478e-06,"3":1.6008357603391436e-06},"predicted":"0","confidence":0.9999859883392076,"entropy_confidence":0.9998673212350386,"candidate_mass":0.9999986204473567,"accepted":true,"accepted_predicted":"0","input_tokens":397,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.4011758956637281e-05,"brier_hard":2.789824821172874e-10,"score_mae":0.023310821996849345,"within_one":true,"case_id":"typed:215/r0","base_id":"typed:215","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.87701699999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999953722624789,"hold":2.72829101020299e-06,"manual_review":1.5683678942187061e-06,"reject":3.310786159369223e-07},"predicted":"approve","confidence":0.9999953722624789,"entropy_confidence":0.9999527631941145,"candidate_mass":0.999998516731305,"accepted":true,"accepted_predicted":"approve","input_tokens":401,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":4.627748229089352e-06,"brier_hard":3.14289173019039e-11,"soft_accuracy":0.9466627295271536,"brier_soft":0.004110361058079546,"tvd":0.05332837226247926,"kl":0.43872574338856934,"case_id":"typed:215/r0","base_id":"typed:215","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.87701699999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0007016743795873269,"true":0.999298325620414},"predicted":"true","confidence":0.999298325620414,"entropy_confidence":0.9916366732776223,"candidate_mass":0.9999785207082204,"accepted":true,"accepted_predicted":"true","input_tokens":349,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.26204110829034,"brier_hard":1.9971942871755233,"soft_accuracy":0.02067360740440386,"brier_soft":1.9180504211258902,"tvd":0.9792983256204133,"kl":7.018775211258166,"case_id":"typed:215/r0","base_id":"typed:215","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.87701699999997}
{"id":"matches_order","type":"binary","probabilities":{"false":1.2914229145484863e-09,"true":0.9999999987085779},"predicted":"true","confidence":0.9999999987085779,"entropy_confidence":0.9999999600032305,"candidate_mass":0.9999999776625439,"accepted":true,"accepted_predicted":"true","input_tokens":369,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.291422081218904e-09,"brier_hard":3.3355441339268763e-18,"soft_accuracy":0.9933329987257977,"brier_soft":8.88977435603476e-05,"tvd":0.006666998708577481,"kl":0.054760614542704575,"case_id":"typed:215/r0","base_id":"typed:215","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.87701699999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000582334733135707,"1":0.9465771358018268,"2":0.05243350162276891,"3":0.0004070278422692827},"predicted":"1","confidence":0.9465771358018268,"entropy_confidence":0.8455813414719381,"candidate_mass":0.9999885703412221,"accepted":true,"accepted_predicted":"1","input_tokens":396,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.9482095480024384,"brier_hard":1.793891047655083,"score_mae":0.7806677774258277,"within_one":true,"case_id":"typed:215/r0","base_id":"typed:215","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.87701699999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9514030446858673,"1":0.0039418031442280035,"2":0.01163547497140472,"3":0.03301967719849902},"predicted":"0","confidence":0.9514030446858673,"entropy_confidence":0.8314511711013449,"candidate_mass":0.9999565044073623,"accepted":true,"accepted_predicted":"0","input_tokens":379,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.4106516164767946,"brier_hard":1.8403696202126718,"score_mae":2.7670612153174656,"within_one":false,"case_id":"typed:216/r0","base_id":"typed:216","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.767219}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999458388574332,"hold":3.097947284348061e-05,"manual_review":1.817523048893633e-05,"reject":5.006439235493142e-06},"predicted":"approve","confidence":0.9999458388574332,"entropy_confidence":0.9995417368146534,"candidate_mass":0.9999934929662795,"accepted":true,"accepted_predicted":"approve","input_tokens":383,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.915450852977987,"brier_hard":1.999855331502449,"soft_accuracy":0.016685872009067453,"brier_soft":1.547075015723292,"tvd":0.9832788388574326,"kl":8.901009564263463,"case_id":"typed:216/r0","base_id":"typed:216","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.767219}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003916958805817007,"true":0.996083041194183},"predicted":"true","confidence":0.996083041194183,"entropy_confidence":0.9630399050083677,"candidate_mass":0.9999710682022902,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.542439741122443,"brier_hard":1.984362849909305,"soft_accuracy":0.030375051724867562,"brier_soft":1.879534921849204,"tvd":0.969416041194183,"kl":5.271785960753392,"case_id":"typed:216/r0","base_id":"typed:216","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.767219}
{"id":"matches_order","type":"binary","probabilities":{"false":2.6490520890784493e-08,"true":0.9999999735094786},"predicted":"true","confidence":0.9999999735094786,"entropy_confidence":0.9999992950172087,"candidate_mass":0.999999902559764,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.446478870179487,"brier_hard":1.9999998940379167,"soft_accuracy":0.020000025430900046,"brier_soft":1.9207998961571584,"tvd":0.9799999735094789,"kl":8.947893104957906,"case_id":"typed:216/r0","base_id":"typed:216","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.767219}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.019812630509983636,"1":0.010243448378963972,"2":0.0010224118978257601,"3":0.9689215092132281},"predicted":"3","confidence":0.9689215092132281,"entropy_confidence":0.8829608073543619,"candidate_mass":0.9999937537099763,"accepted":true,"accepted_predicted":"3","input_tokens":378,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.885590837255194,"brier_hard":1.9372625811088944,"score_mae":0.9190527998142999,"within_one":true,"case_id":"typed:216/r0","base_id":"typed:216","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.767219}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.021141493630357926,"1":0.0039571085824753405,"2":0.8407140146257727,"3":0.13418738316139314},"predicted":"2","confidence":0.8407140146257727,"entropy_confidence":0.625758394976289,"candidate_mass":0.9999565556308934,"accepted":true,"accepted_predicted":"2","input_tokens":409,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.17350373079335318,"brier_hard":0.04384090039759707,"score_mae":0.042052712681799775,"within_one":true,"case_id":"typed:217/r0","base_id":"typed:217","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.757004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999946876590675,"hold":2.9837029763936176e-06,"manual_review":1.6190350731483071e-06,"reject":7.096028829497508e-07},"predicted":"approve","confidence":0.9999946876590675,"entropy_confidence":0.9999459662536432,"candidate_mass":0.9999978458793921,"accepted":true,"accepted_predicted":"approve","input_tokens":413,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":13.33368022003937,"brier_hard":1.999986137288237,"soft_accuracy":0.05000171978593581,"brier_soft":1.3472085135645115,"tvd":0.9499946876590675,"kl":8.600037005317427,"case_id":"typed:217/r0","base_id":"typed:217","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.757004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0019343936660743842,"true":0.9980656063339238},"predicted":"true","confidence":0.9980656063339238,"entropy_confidence":0.9797755241911604,"candidate_mass":0.9999832065664839,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.247961352470984,"brier_hard":1.9922699090934095,"soft_accuracy":0.028498224714287927,"brier_soft":1.8872305047749824,"tvd":0.9713986063339247,"kl":5.958440403078419,"case_id":"typed:217/r0","base_id":"typed:217","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.757004}
{"id":"matches_order","type":"binary","probabilities":{"false":3.5310455058491946e-08,"true":0.9999999646895457},"predicted":"true","confidence":0.9999999646895457,"entropy_confidence":0.9999990749358392,"candidate_mass":0.9999998600924664,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.159086739564607,"brier_hard":1.9999998587581838,"soft_accuracy":0.09000002895457321,"brier_soft":1.6561998714699477,"tvd":0.9099999646895454,"kl":8.1646946368976,"case_id":"typed:217/r0","base_id":"typed:217","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.757004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00037928673194453417,"1":0.6333780681682245,"2":0.36583141983187817,"3":0.00041122526795408945},"predicted":"1","confidence":0.6333780681682245,"entropy_confidence":0.5215132117856147,"candidate_mass":0.9999817422130554,"accepted":false,"accepted_predicted":null,"input_tokens":408,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.45668777104034675,"brier_hard":0.268244581600917,"score_mae":0.2629415836358433,"within_one":true,"case_id":"typed:217/r0","base_id":"typed:217","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.757004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.018526460749203996,"1":0.0008501528406601224,"2":0.07650511477845776,"3":0.9041182716316766},"predicted":"3","confidence":0.9041182716316766,"entropy_confidence":0.7347722526467316,"candidate_mass":0.9999728333332423,"accepted":true,"accepted_predicted":"3","input_tokens":382,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.10079509569667877,"brier_hard":0.015390290929906233,"score_mae":0.04954819729260551,"within_one":true,"case_id":"typed:218/r0","base_id":"typed:218","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.817802}
{"id":"disposition","type":"choice","probabilities":{"approve":0.877860456473332,"hold":0.08291824775051916,"manual_review":0.004089260671252804,"reject":0.035132035104895154},"predicted":"approve","confidence":0.877860456473332,"entropy_confidence":0.6674962991283957,"candidate_mass":0.9999760310965412,"accepted":true,"accepted_predicted":"approve","input_tokens":386,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":5.4993910902422485,"brier_hard":1.7705868774505267,"soft_accuracy":0.07454455696118502,"brier_soft":1.0545431612709901,"tvd":0.8378604164732925,"kl":2.9087627983476794,"case_id":"typed:218/r0","base_id":"typed:218","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.817802}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0005585291823676476,"true":0.9994414708176338},"predicted":"true","confidence":0.9994414708176338,"entropy_confidence":0.9931589269186343,"candidate_mass":0.9999660505614212,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.490203689490354,"brier_hard":1.9977665071802273,"soft_accuracy":0.02719574058695529,"brier_soft":1.8925803421490521,"tvd":0.9727744708176331,"kl":7.167519127722529,"case_id":"typed:218/r0","base_id":"typed:218","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.817802}
{"id":"matches_order","type":"binary","probabilities":{"false":3.903439443537393e-08,"true":0.9999999609656055},"predicted":"true","confidence":0.9999999609656055,"entropy_confidence":0.9999989830223274,"candidate_mass":0.9999997615127914,"accepted":true,"accepted_predicted":"true","input_tokens":354,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.058822670870335,"brier_hard":1.9999998438624254,"soft_accuracy":0.05666703461047038,"brier_soft":1.7797541504882732,"tvd":0.9433329609656056,"kl":8.525751808739976,"case_id":"typed:218/r0","base_id":"typed:218","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.817802}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00010234109328685814,"1":6.238207796662057e-05,"2":0.0002573446203141277,"3":0.9995779322084342},"predicted":"3","confidence":0.9995779322084342,"entropy_confidence":0.9970473932409759,"candidate_mass":0.9999974265742503,"accepted":true,"accepted_predicted":"3","input_tokens":381,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0004221568872466709,"brier_hard":2.587326973084699e-07,"score_mae":0.6726438679438971,"within_one":true,"case_id":"typed:218/r0","base_id":"typed:218","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.817802}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0016808880756289354,"1":0.0010972600707447849,"2":0.9273610541093841,"3":0.06986079774424203},"predicted":"2","confidence":0.9273610541093841,"entropy_confidence":0.802302070291331,"candidate_mass":0.9999892914468479,"accepted":true,"accepted_predicted":"2","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.6612506204839654,"brier_hard":1.7251614896162315,"score_mae":0.5445982384777608,"within_one":true,"case_id":"typed:219/r0","base_id":"typed:219","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.12550200000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.569848561450138,"hold":0.42612416312760276,"manual_review":0.00376212855273169,"reject":0.00026514686952625304},"predicted":"approve","confidence":0.569848561450138,"entropy_confidence":0.48989543921072365,"candidate_mass":0.9999789230053049,"accepted":false,"accepted_predicted":null,"input_tokens":391,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":null,"nll_hard":5.582770377307092,"brier_hard":1.4987991521966377,"soft_accuracy":0.11218669685122336,"brier_soft":0.8679718933776543,"tvd":0.7493057245777414,"kl":3.7181745715139365,"case_id":"typed:219/r0","base_id":"typed:219","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.12550200000004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004162878465224736,"true":0.9958371215347765},"predicted":"true","confidence":0.9958371215347765,"entropy_confidence":0.9610858665762304,"candidate_mass":0.9999459679738374,"accepted":true,"accepted_predicted":"true","input_tokens":339,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.4815485052322686,"brier_hard":1.983383145253336,"soft_accuracy":0.04052459793585599,"brier_soft":1.8400146440900733,"tvd":0.9591701215347759,"kl":5.123506497104011,"case_id":"typed:219/r0","base_id":"typed:219","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.12550200000004}
{"id":"matches_order","type":"binary","probabilities":{"false":3.80732502821238e-05,"true":0.9999619267497172},"predicted":"true","confidence":0.9999619267497172,"entropy_confidence":0.9993861247949195,"candidate_mass":0.9999939504886176,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.175998614746572,"brier_hard":1.9998477098980147,"soft_accuracy":0.05003426592525388,"brier_soft":1.8048553245480712,"tvd":0.9499619267497175,"kl":8.60003864339843,"case_id":"typed:219/r0","base_id":"typed:219","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.12550200000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.002412424413777265,"1":0.7386568172453025,"2":0.2536557770456779,"3":0.005274981295241153},"predicted":"1","confidence":0.7386568172453025,"entropy_confidence":0.5571499084495669,"candidate_mass":0.9999896239045838,"accepted":false,"accepted_predicted":null,"input_tokens":386,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3717771395706961,"brier_hard":1.1026772380194683,"score_mae":0.7782066847776183,"within_one":true,"case_id":"typed:219/r0","base_id":"typed:219","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.12550200000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999985473176007,"1":8.307421120578774e-07,"2":4.5949673158005685e-07,"3":1.6244355519956538e-07},"predicted":"0","confidence":0.9999985473176007,"entropy_confidence":0.9999838931580952,"candidate_mass":0.9999997356354887,"accepted":true,"accepted_predicted":"0","input_tokens":409,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.4526834544052055e-06,"brier_hard":3.0379437648280187e-12,"score_mae":0.013330762933759182,"within_one":true,"case_id":"typed:220/r0","base_id":"typed:220","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.654102}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999990343857136,"hold":5.032961220794838e-07,"manual_review":3.802390026911405e-07,"reject":8.207916107613947e-08},"predicted":"approve","confidence":0.9999990343857136,"entropy_confidence":0.9999890178465838,"candidate_mass":0.9999987403830108,"accepted":true,"accepted_predicted":"approve","input_tokens":413,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":9.656147526247229e-07,"brier_hard":1.3370366244871361e-12,"soft_accuracy":0.969999076607169,"brier_soft":0.00134991555842611,"tvd":0.029999116464874957,"kl":0.24676571654542823,"case_id":"typed:220/r0","base_id":"typed:220","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.654102}
{"id":"duplicate","type":"binary","probabilities":{"false":0.002306733430077775,"true":0.9976932665699226},"predicted":"true","confidence":0.9976932665699226,"entropy_confidence":0.9764690981023362,"candidate_mass":0.9999838933708709,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.071922854764193,"brier_hard":1.9907837083179245,"soft_accuracy":0.008942975446521121,"brier_soft":1.964266122063038,"tvd":0.9910262665699224,"kl":5.991406443684054,"case_id":"typed:220/r0","base_id":"typed:220","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.654102}
{"id":"matches_order","type":"binary","probabilities":{"false":1.1794174778295585e-09,"true":0.9999999988205825},"predicted":"true","confidence":0.9999999988205825,"entropy_confidence":0.999999963317789,"candidate_mass":0.999999973494868,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.1794174532858272e-09,"brier_hard":2.782051114484464e-18,"soft_accuracy":0.9933329988363089,"brier_soft":8.889774654729725e-05,"tvd":0.006666998820582514,"kl":0.05476061443144668,"case_id":"typed:220/r0","base_id":"typed:220","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.654102}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004588100918259868,"1":0.9151470486059036,"2":0.08390252168106846,"3":0.0004916196212026282},"predicted":"1","confidence":0.9151470486059036,"entropy_confidence":0.7862378617452095,"candidate_mass":0.9999901654482102,"accepted":true,"accepted_predicted":"1","input_tokens":408,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.4780996101672166,"brier_hard":1.6767291625509535,"score_mae":0.5255730491683517,"within_one":true,"case_id":"typed:220/r0","base_id":"typed:220","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.654102}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.99997944076193,"1":1.107704591020455e-05,"2":6.629588024782775e-06,"3":2.8526041333576026e-06},"predicted":"0","confidence":0.99997944076193,"entropy_confidence":0.9998106995504376,"candidate_mass":0.9999976675316115,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":2.055944941401545e-05,"brier_hard":5.974720038350393e-10,"score_mae":0.23363410596564016,"within_one":true,"case_id":"typed:221/r0","base_id":"typed:221","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.572102}
{"id":"disposition","type":"choice","probabilities":{"approve":0.999990369810242,"hold":5.658426812720581e-06,"manual_review":2.9849195888372005e-06,"reject":9.868433547762866e-07},"predicted":"approve","confidence":0.999990369810242,"entropy_confidence":0.999906500308424,"candidate_mass":0.9999968398264141,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":9.630236128544902e-06,"brier_hard":1.3464195352811196e-10,"soft_accuracy":0.7663272855084234,"brier_soft":0.07577145229523763,"tvd":0.23365660347647654,"kl":1.9482027897090686,"case_id":"typed:221/r0","base_id":"typed:221","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.572102}
{"id":"duplicate","type":"binary","probabilities":{"false":1.0973254297144938e-05,"true":0.9999890267457042},"predicted":"true","confidence":0.9999890267457042,"entropy_confidence":0.9998033775407819,"candidate_mass":0.9999986648633299,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.420049673418262,"brier_hard":1.999956107223639,"soft_accuracy":0.22667299870503393,"brier_soft":1.1960539141001654,"tvd":0.7733220267457035,"kl":6.7862268623295705,"case_id":"typed:221/r0","base_id":"typed:221","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.572102}
{"id":"matches_order","type":"binary","probabilities":{"false":9.408011301886332e-09,"true":0.9999999905919879},"predicted":"true","confidence":0.9999999905919879,"entropy_confidence":0.9999997355769458,"candidate_mass":0.9999999456049512,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.40801219136053e-09,"brier_hard":1.7702136921649962e-16,"soft_accuracy":0.9663329912254669,"brier_soft":0.0022669325110420543,"tvd":0.03366699059198828,"kl":0.2769907411854033,"case_id":"typed:221/r0","base_id":"typed:221","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.572102}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008436198085632449,"1":0.9313931708242841,"2":0.06754277208274344,"3":0.00022043728440825156},"predicted":"1","confidence":0.9313931708242841,"entropy_confidence":0.8152974863105295,"candidate_mass":0.9999899380866909,"accepted":true,"accepted_predicted":"1","input_tokens":406,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.07107378062277192,"brier_hard":0.00926968335714505,"score_mae":0.26680702684299584,"within_one":true,"case_id":"typed:221/r0","base_id":"typed:221","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.572102}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999021554268124,"1":4.3297141672726775e-05,"2":3.78237292991089e-05,"3":1.672370221712918e-05},"predicted":"0","confidence":0.9999021554268124,"entropy_confidence":0.9992051138763944,"candidate_mass":0.999996102496826,"accepted":true,"accepted_predicted":"0","input_tokens":421,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":9.784936028013643e-05,"brier_hard":1.3158519693235588e-08,"score_mae":0.45649788429307764,"within_one":true,"case_id":"typed:222/r0","base_id":"typed:222","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.813002}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999850836106717,"hold":9.058798140420936e-06,"manual_review":4.388453455455139e-06,"reject":1.4691377340517735e-06},"predicted":"approve","confidence":0.9999850836106717,"entropy_confidence":0.9998600765815814,"candidate_mass":0.9999958314181634,"accepted":true,"accepted_predicted":"approve","input_tokens":425,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.4916500578709378e-05,"brier_hard":3.2597738375366555e-10,"soft_accuracy":0.7766569612966668,"brier_soft":0.07443452050998708,"tvd":0.2233180836106709,"kl":1.8557427921264653,"case_id":"typed:222/r0","base_id":"typed:222","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.813002}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0029130977961609227,"true":0.9970869022038396},"predicted":"true","confidence":0.9970869022038396,"entropy_confidence":0.9712657224791026,"candidate_mass":0.9999856045355394,"accepted":true,"accepted_predicted":"true","input_tokens":373,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.838538229250862,"brier_hard":1.9883645810928974,"soft_accuracy":0.03273831192839128,"brier_soft":1.8705141528284366,"tvd":0.9670869022038393,"kl":5.528727434667286,"case_id":"typed:222/r0","base_id":"typed:222","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.813002}
{"id":"matches_order","type":"binary","probabilities":{"false":1.916786740805183e-09,"true":0.9999999980832133},"predicted":"true","confidence":0.9999999980832133,"entropy_confidence":0.99999994172708,"candidate_mass":0.9999999687626429,"accepted":true,"accepted_predicted":"true","input_tokens":393,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.916786731176073e-09,"brier_hard":7.3481427754968e-18,"soft_accuracy":0.8966669984793479,"brier_soft":0.021355416985730714,"tvd":0.10333299808321328,"kl":0.8539319851204201,"case_id":"typed:222/r0","base_id":"typed:222","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.813002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002813653934448778,"1":0.9216502098899464,"2":0.0778103430532104,"3":0.00025808166339974796},"predicted":"1","confidence":0.9216502098899464,"entropy_confidence":0.7992367593485061,"candidate_mass":0.9999907566106242,"accepted":true,"accepted_predicted":"1","input_tokens":420,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.5534809125054103,"brier_hard":1.699873018542449,"score_mae":0.7986218590134337,"within_one":true,"case_id":"typed:222/r0","base_id":"typed:222","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.813002}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999921006682904,"1":4.602308094729062e-06,"2":2.4719259959159945e-06,"3":8.250976181525082e-07},"predicted":"0","confidence":0.9999921006682904,"entropy_confidence":0.9999221460787293,"candidate_mass":0.999998862926322,"accepted":true,"accepted_predicted":"0","input_tokens":419,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":7.899362909522671e-06,"brier_hard":9.037188546646154e-11,"score_mae":0.14332097854705897,"within_one":true,"case_id":"typed:223/r0","base_id":"typed:223","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.78420200000005}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999974173515359,"hold":1.4383949319418848e-06,"manual_review":9.089425015827062e-07,"reject":2.3531102862207559e-07},"predicted":"approve","confidence":0.9999974173515359,"entropy_confidence":0.9999724679003149,"candidate_mass":0.9999981667678169,"accepted":true,"accepted_predicted":"approve","input_tokens":423,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.5826517991312414e-06,"brier_hard":9.620600820671742e-12,"soft_accuracy":0.5933319234284498,"brier_soft":0.23837506341179288,"tvd":0.4066644173515369,"kl":3.435823499300005,"case_id":"typed:223/r0","base_id":"typed:223","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.78420200000005}
{"id":"duplicate","type":"binary","probabilities":{"false":3.690264176498035e-05,"true":0.9999630973582353},"predicted":"true","confidence":0.9999630973582353,"entropy_confidence":0.999403336507406,"candidate_mass":0.9999981838239308,"accepted":true,"accepted_predicted":"true","input_tokens":371,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.690332268393873e-05,"brier_hard":2.7236099384484193e-09,"soft_accuracy":0.5833268495843077,"brier_soft":0.347161276049465,"tvd":0.4166300973582352,"kl":3.5232516396009927,"case_id":"typed:223/r0","base_id":"typed:223","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.78420200000005}
{"id":"matches_order","type":"binary","probabilities":{"false":1.2129186851099549e-09,"true":0.9999999987870822},"predicted":"true","confidence":0.9999999987870822,"entropy_confidence":0.9999999623248487,"candidate_mass":0.9999999785471376,"accepted":true,"accepted_predicted":"true","input_tokens":391,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.2129177669601487e-09,"brier_hard":2.9423412443120475e-18,"soft_accuracy":0.9633329988760304,"brier_soft":0.0026889376001037134,"tvd":0.03666699878708178,"kl":0.30172915633585023,"case_id":"typed:223/r0","base_id":"typed:223","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.78420200000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002546334772005244,"1":0.9703772305664157,"2":0.02917347320678668,"3":0.00019466274959766968},"predicted":"1","confidence":0.9703772305664157,"entropy_confidence":0.901850832440359,"candidate_mass":0.9999914809816982,"accepted":true,"accepted_predicted":"1","input_tokens":418,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.030070385594587617,"brier_hard":0.0017287027396561904,"score_mae":0.05264116522878215,"within_one":true,"case_id":"typed:223/r0","base_id":"typed:223","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.78420200000005}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999663315625025,"1":1.519788983725822e-05,"2":1.349435637085985e-05,"3":4.976191289554582e-06},"predicted":"0","confidence":0.9999663315625025,"entropy_confidence":0.9997011041815765,"candidate_mass":0.9999971944834476,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.3669004292088765e-05,"brier_hard":1.5713996726441825e-09,"score_mae":0.053275884823552354,"within_one":true,"case_id":"typed:224/r0","base_id":"typed:224","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.311001}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999968825159186,"hold":1.630291496326863e-06,"manual_review":1.2272025321083962e-06,"reject":2.599900515766575e-07},"predicted":"approve","confidence":0.9999968825159186,"entropy_confidence":0.9999671864237347,"candidate_mass":0.9999983849041354,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":3.1174889408021244e-06,"brier_hard":1.3950178242750386e-11,"soft_accuracy":0.9566640737410929,"brier_soft":0.002644013119601736,"tvd":0.04332988251591925,"kl":0.3567343992932135,"case_id":"typed:224/r0","base_id":"typed:224","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.311001}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004538750568095998,"true":0.9954612494319031},"predicted":"true","confidence":0.9954612494319031,"entropy_confidence":0.9581395340613433,"candidate_mass":0.9999689966688535,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.3951035100699976,"brier_hard":1.9818861982410527,"soft_accuracy":0.021054455856659073,"brier_soft":1.9160763654419268,"tvd":0.9787942494319035,"kl":5.220491704526277,"case_id":"typed:224/r0","base_id":"typed:224","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.311001}
{"id":"matches_order","type":"binary","probabilities":{"false":1.6776152189157025e-09,"true":0.9999999983223855},"predicted":"true","confidence":0.9999999983223855,"entropy_confidence":0.9999999486756517,"candidate_mass":0.9999999695238775,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6776144931757039e-09,"brier_hard":5.628783205729292e-18,"soft_accuracy":0.9829999983794243,"brier_soft":0.0005779998859221965,"tvd":0.01699999832238516,"kl":0.13972111383791433,"case_id":"typed:224/r0","base_id":"typed:224","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.311001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007428951635068185,"1":0.7225725940399507,"2":0.2761567608216818,"3":0.0005277499748616342},"predicted":"1","confidence":0.7225725940399507,"entropy_confidence":0.5675658185434762,"candidate_mass":0.9999915569013981,"accepted":false,"accepted_predicted":null,"input_tokens":406,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.286786600509565,"brier_hard":1.0460610189750434,"score_mae":0.34019763439210093,"within_one":true,"case_id":"typed:224/r0","base_id":"typed:224","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.311001}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0012707479728035504,"1":0.0015966134555756115,"2":0.975677166179251,"3":0.02145547239237093},"predicted":"2","confidence":0.975677166179251,"entropy_confidence":0.9096820704702875,"candidate_mass":0.9999922055733502,"accepted":true,"accepted_predicted":"2","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.8417755429590414,"brier_hard":1.9094994890893489,"score_mae":0.6626826370088095,"within_one":true,"case_id":"typed:225/r0","base_id":"typed:225","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.67760200000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.8138571568536389,"hold":0.18257894672033206,"manual_review":0.003258943099259982,"reject":0.0003049533267698665},"predicted":"approve","confidence":0.8138571568536389,"entropy_confidence":0.6398673866727331,"candidate_mass":0.9999837695813318,"accepted":true,"accepted_predicted":"approve","input_tokens":391,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":5.7263523388495585,"brier_hard":1.6891913710555302,"soft_accuracy":0.07199757483790936,"brier_soft":1.0807588231342313,"tvd":0.7871901568536385,"kl":3.7722985450026316,"case_id":"typed:225/r0","base_id":"typed:225","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.67760200000004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.017713725721081604,"true":0.9822862742789176},"predicted":"true","confidence":0.9822862742789176,"entropy_confidence":0.8715962194191662,"candidate_mass":0.9999234753737993,"accepted":true,"accepted_predicted":"true","input_tokens":339,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.033415475451687,"brier_hard":1.9297726492735152,"soft_accuracy":0.027359451206659963,"brier_soft":1.8906811983023586,"tvd":0.972286274278918,"kl":3.937258511256752,"case_id":"typed:225/r0","base_id":"typed:225","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.67760200000004}
{"id":"matches_order","type":"binary","probabilities":{"false":2.859427314861817e-05,"true":0.9999714057268524},"predicted":"true","confidence":0.9999714057268524,"entropy_confidence":0.9995271482670466,"candidate_mass":0.9999943286457257,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.46230409974789,"brier_hard":1.9998856245426724,"soft_accuracy":0.023360259892797886,"brier_soft":1.9076451510813737,"tvd":0.9766384057268518,"kl":8.907753398972378,"case_id":"typed:225/r0","base_id":"typed:225","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.67760200000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008493645118315045,"1":0.6335555253373163,"2":0.36441621039620625,"3":0.0011788997546450994},"predicted":"1","confidence":0.6335555253373163,"entropy_confidence":0.5159916343680331,"candidate_mass":0.9999874455618971,"accepted":false,"accepted_predicted":null,"input_tokens":386,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.009458629415061,"brier_hard":0.8053614685172678,"score_mae":0.5007423546063359,"within_one":true,"case_id":"typed:225/r0","base_id":"typed:225","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.67760200000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999871652263298,"1":6.9819498679990505e-06,"2":4.29590785145609e-06,"3":1.5569159504423379e-06},"predicted":"0","confidence":0.9999871652263298,"entropy_confidence":0.999877634805448,"candidate_mass":0.9999986190290336,"accepted":true,"accepted_predicted":"0","input_tokens":406,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.2834856036612936e-05,"brier_hard":2.3435785066947e-10,"score_mae":0.9799797554865778,"within_one":true,"case_id":"typed:226/r0","base_id":"typed:226","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.019502}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999980824882168,"hold":1.0301666458232554e-06,"manual_review":7.619352479884133e-07,"reject":1.2540989009862413e-07},"predicted":"approve","confidence":0.9999980824882168,"entropy_confidence":0.9999791921288378,"candidate_mass":0.999998423391513,"accepted":true,"accepted_predicted":"approve","input_tokens":410,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.91751362167291e-06,"brier_hard":5.3343677197112595e-12,"soft_accuracy":0.41666649298689545,"brier_soft":0.4704189167859771,"tvd":0.5833310824882164,"kl":5.00791768011706,"case_id":"typed:226/r0","base_id":"typed:226","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.019502}
{"id":"duplicate","type":"binary","probabilities":{"false":2.4204166563651124e-05,"true":0.9999757958334364},"predicted":"true","confidence":0.9999757958334364,"entropy_confidence":0.9995939251814688,"candidate_mass":0.9999984208103391,"accepted":true,"accepted_predicted":"true","input_tokens":358,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.4204459489134594e-05,"brier_hard":1.171683358077929e-09,"soft_accuracy":0.5433309023217006,"brier_soft":0.41704528597315504,"tvd":0.4566427958334365,"kl":3.874620658950168,"case_id":"typed:226/r0","base_id":"typed:226","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.019502}
{"id":"matches_order","type":"binary","probabilities":{"false":5.080655581165938e-09,"true":0.9999999949193441},"predicted":"true","confidence":0.9999999949193441,"entropy_confidence":0.9999998526862222,"candidate_mass":0.9999999482610309,"accepted":true,"accepted_predicted":"true","input_tokens":378,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.0806559089013556e-09,"brier_hard":5.162612546793956e-17,"soft_accuracy":0.9666669952580512,"brier_soft":0.0022221771005860535,"tvd":0.0333329949193442,"kl":0.274237102610099,"case_id":"typed:226/r0","base_id":"typed:226","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.019502}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000741815275223524,"1":0.46659496578003734,"2":0.5312677766348413,"3":0.0013954423098971435},"predicted":"2","confidence":0.5312677766348413,"entropy_confidence":0.49056728130486815,"candidate_mass":0.9999874245294391,"accepted":false,"accepted_predicted":null,"input_tokens":405,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":0.7622937086085244,"brier_hard":0.5667688785708699,"score_mae":0.08331684597941136,"within_one":true,"case_id":"typed:226/r0","base_id":"typed:226","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.019502}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9988887624783828,"1":0.0003282872484861117,"2":0.00033803828382756554,"3":0.0004449119893052035},"predicted":"0","confidence":0.9988887624783828,"entropy_confidence":0.992873515971324,"candidate_mass":0.9999906738822515,"accepted":true,"accepted_predicted":"0","input_tokens":415,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0011118554038169646,"brier_hard":1.6548379065291996e-06,"score_mae":0.6143279002159431,"within_one":true,"case_id":"typed:227/r0","base_id":"typed:227","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.66090199999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997748331760576,"hold":0.00014382113253840864,"manual_review":5.7140318748313705e-05,"reject":2.420537265408507e-05},"predicted":"approve","confidence":0.9997748331760576,"entropy_confidence":0.9983314822639596,"candidate_mass":0.9999867424260198,"accepted":true,"accepted_predicted":"approve","input_tokens":419,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00022519217779762928,"brier_hard":7.523553286090122e-08,"soft_accuracy":0.5665827449241406,"brier_soft":0.26694002951736695,"tvd":0.4331078331760584,"kl":2.965578625115602,"case_id":"typed:227/r0","base_id":"typed:227","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.66090199999996}
{"id":"duplicate","type":"binary","probabilities":{"false":5.227728857125212e-05,"true":0.999947722711429},"predicted":"true","confidence":0.999947722711429,"entropy_confidence":0.9991810180643598,"candidate_mass":0.9999966093637659,"accepted":true,"accepted_predicted":"true","input_tokens":367,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.85894853414824,"brier_hard":1.9997908963115452,"soft_accuracy":0.45000522772885726,"brier_soft":0.6048849954309732,"tvd":0.5499477227114289,"kl":4.706382266683689,"case_id":"typed:227/r0","base_id":"typed:227","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.66090199999996}
{"id":"matches_order","type":"binary","probabilities":{"false":3.2285712265478056e-09,"true":0.9999999967714288},"predicted":"true","confidence":0.9999999967714288,"entropy_confidence":0.9999999042756009,"candidate_mass":0.9999999660283999,"accepted":true,"accepted_predicted":"true","input_tokens":387,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.2285711933859367e-09,"brier_hard":2.0847344082000327e-17,"soft_accuracy":0.8499999977400002,"brier_soft":0.044999998062857316,"tvd":0.14999999677142883,"kl":1.2434099684676045,"case_id":"typed:227/r0","base_id":"typed:227","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.66090199999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.233415542768127e-05,"1":0.9445926355677888,"2":0.05522456260695345,"3":0.00010046766982966614},"predicted":"1","confidence":0.9445926355677888,"entropy_confidence":0.8445554249716212,"candidate_mass":0.9999912242056045,"accepted":true,"accepted_predicted":"1","input_tokens":414,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.05700151790051325,"brier_hard":0.006119745221119012,"score_mae":0.16132383620881519,"within_one":true,"case_id":"typed:227/r0","base_id":"typed:227","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.66090199999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9996759751033185,"1":9.216861135878166e-05,"2":0.00012935335613658606,"3":0.00010250292918448172},"predicted":"0","confidence":0.9996759751033185,"entropy_confidence":0.9976339500353397,"candidate_mass":0.9999927867699644,"accepted":true,"accepted_predicted":"0","input_tokens":413,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":9.18561918238456,"brier_hard":1.999147085074596,"score_mae":2.6926746158888144,"within_one":false,"case_id":"typed:228/r0","base_id":"typed:228","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.09000199999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999717660739359,"hold":1.562018441268573e-05,"manual_review":9.961106092831985e-06,"reject":2.652635559793248e-06},"predicted":"approve","confidence":0.9999717660739359,"entropy_confidence":0.9997476139985032,"candidate_mass":0.9999954475177717,"accepted":true,"accepted_predicted":"approve","input_tokens":417,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":11.066946607432559,"brier_hard":1.9999122929264512,"soft_accuracy":0.03334509948310075,"brier_soft":1.4311207567304218,"tvd":0.9666387327409018,"kl":8.789960204790082,"case_id":"typed:228/r0","base_id":"typed:228","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.09000199999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.002515088000249751,"true":0.9974849119997509},"predicted":"true","confidence":0.9974849119997509,"entropy_confidence":0.9746578358074486,"candidate_mass":0.9999804502384343,"accepted":true,"accepted_predicted":"true","input_tokens":365,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.985447485883629,"brier_hard":1.9899522993343002,"soft_accuracy":0.02904794829684445,"brier_soft":1.8849748365191108,"tvd":0.9708179119997505,"kl":5.70294251367536,"case_id":"typed:228/r0","base_id":"typed:228","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.09000199999997}
{"id":"matches_order","type":"binary","probabilities":{"false":1.3126302460413546e-08,"true":0.9999999868736983},"predicted":"true","confidence":0.9999999868736983,"entropy_confidence":0.9999996373772426,"candidate_mass":0.9999998760867476,"accepted":true,"accepted_predicted":"true","input_tokens":385,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.14864779831307,"brier_hard":1.999999947494792,"soft_accuracy":0.056667011638646146,"brier_soft":1.7797542482481044,"tvd":0.9433329868736979,"kl":8.525751807271842,"case_id":"typed:228/r0","base_id":"typed:228","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.09000199999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00040624937659910235,"1":0.7782015172368394,"2":0.22108967328880674,"3":0.00030256009775387564},"predicted":"1","confidence":0.7782015172368394,"entropy_confidence":0.6144837648417956,"candidate_mass":0.9999898739756212,"accepted":false,"accepted_predicted":null,"input_tokens":412,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.5091868982755017,"brier_hard":1.2122991550682254,"score_mae":0.5453784558922854,"within_one":true,"case_id":"typed:228/r0","base_id":"typed:228","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.09000199999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.004231198246975882,"1":0.0014436209300724985,"2":0.560679063976565,"3":0.43364611684638676},"predicted":"2","confidence":0.560679063976565,"entropy_confidence":0.48113312934281083,"candidate_mass":0.9999663859287568,"accepted":false,"accepted_predicted":null,"input_tokens":406,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":0.8355264764717562,"brier_hard":0.6351377208248086,"score_mae":0.19292690057763728,"within_one":true,"case_id":"typed:229/r0","base_id":"typed:229","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":315.446502}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999370307928629,"hold":3.599566801980006e-05,"manual_review":1.6519828628354157e-05,"reject":1.045371048986475e-05},"predicted":"approve","confidence":0.9999370307928629,"entropy_confidence":0.9994712040369456,"candidate_mass":0.9999940123357893,"accepted":true,"accepted_predicted":"approve","input_tokens":410,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":11.010949163511558,"brier_hard":1.9998410275714629,"soft_accuracy":0.013352732990667308,"brier_soft":1.567946996803385,"tvd":0.9866040307928624,"kl":8.890148585927639,"case_id":"typed:229/r0","base_id":"typed:229","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":315.446502}
{"id":"duplicate","type":"binary","probabilities":{"false":0.002285000176773613,"true":0.9977149998232251},"predicted":"true","confidence":0.9977149998232251,"entropy_confidence":0.9766595560480986,"candidate_mass":0.9999804673668727,"accepted":true,"accepted_predicted":"true","input_tokens":358,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.081389177273341,"brier_hard":1.9908704417445187,"soft_accuracy":0.032147900166167157,"brier_soft":1.8729446417657316,"tvd":0.9677149998232257,"kl":5.764273962218574,"case_id":"typed:229/r0","base_id":"typed:229","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":315.446502}
{"id":"matches_order","type":"binary","probabilities":{"false":2.3115851465891185e-08,"true":0.9999999768841479},"predicted":"true","confidence":0.9999999768841479,"entropy_confidence":0.9999993802816523,"candidate_mass":0.9999998503960784,"accepted":true,"accepted_predicted":"true","input_tokens":378,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.582747244108894,"brier_hard":1.999999907536594,"soft_accuracy":0.03333302157481009,"brier_soft":1.8688900883966766,"tvd":0.9666669768841482,"kl":8.78995965147301,"case_id":"typed:229/r0","base_id":"typed:229","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":315.446502}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00028174184496274307,"1":0.9026661573189353,"2":0.09679314147565825,"3":0.00025895936044521037},"predicted":"1","confidence":0.9026661573189353,"entropy_confidence":0.7670722675463697,"candidate_mass":0.9999903645536264,"accepted":true,"accepted_predicted":"1","input_tokens":405,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.335179139733956,"brier_hard":1.6305889672927607,"score_mae":0.5029706816484127,"within_one":true,"case_id":"typed:229/r0","base_id":"typed:229","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":315.446502}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999918831832709,"1":4.641863923416449e-06,"2":2.357317199448392e-06,"3":1.1176356073026038e-06},"predicted":"0","confidence":0.9999918831832709,"entropy_confidence":0.9999199425462737,"candidate_mass":0.9999991266384297,"accepted":true,"accepted_predicted":"0","input_tokens":377,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":8.116849670610876e-06,"brier_hard":9.423566822644441e-11,"score_mae":0.10665429059485577,"within_one":true,"case_id":"typed:230/r0","base_id":"typed:230","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.99270199999995}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999756985861165,"hold":1.3792406032890853e-05,"manual_review":8.08214743306992e-06,"reject":2.4268604161404423e-06},"predicted":"approve","confidence":0.9999756985861165,"entropy_confidence":0.9997801303296795,"candidate_mass":0.9999953432998541,"accepted":true,"accepted_predicted":"approve","input_tokens":381,"reused_prefix_tokens":0,"expected":"reject","correct":false,"accepted_correct":false,"nll_hard":12.928912145794861,"brier_hard":1.9999465443034004,"soft_accuracy":0.21666562332160183,"brier_soft":0.9788421291590291,"tvd":0.7833086985861171,"kl":6.8593886976729985,"case_id":"typed:230/r0","base_id":"typed:230","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.99270199999995}
{"id":"duplicate","type":"binary","probabilities":{"false":3.723372825590028e-06,"true":0.9999962766271726},"predicted":"true","confidence":0.9999962766271726,"entropy_confidence":0.9999274774441451,"candidate_mass":0.9999986786304785,"accepted":true,"accepted_predicted":"true","input_tokens":329,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.723379759160915e-06,"brier_hard":2.7727010410096687e-11,"soft_accuracy":0.7499981383135859,"brier_soft":0.12499627665490053,"tvd":0.2499962766271735,"kl":2.0868263311900295,"case_id":"typed:230/r0","base_id":"typed:230","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.99270199999995}
{"id":"matches_order","type":"binary","probabilities":{"false":2.7807668140909025e-09,"true":0.9999999972192342},"predicted":"true","confidence":0.9999999972192342,"entropy_confidence":0.9999999169536098,"candidate_mass":0.9999999574941727,"accepted":true,"accepted_predicted":"true","input_tokens":349,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.7807658436823423e-09,"brier_hard":1.5465322730236925e-17,"soft_accuracy":0.9499999974973108,"brier_soft":0.0049999994438467595,"tvd":0.04999999721923372,"kl":0.4117883915723641,"case_id":"typed:230/r0","base_id":"typed:230","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.99270199999995}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005756241965489345,"1":0.8487608484828852,"2":0.14990579898121742,"3":0.0007577283393475748},"predicted":"1","confidence":0.8487608484828852,"entropy_confidence":0.6873679715653209,"candidate_mass":0.9999904533623069,"accepted":true,"accepted_predicted":"1","input_tokens":376,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.1639778184732315,"brier_hard":0.04534593501526586,"score_mae":0.02248736853663713,"within_one":true,"case_id":"typed:230/r0","base_id":"typed:230","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.99270199999995}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9950099511131523,"1":0.00046489900887068404,"2":0.0011990126056653186,"3":0.0033261372723112134},"predicted":"0","confidence":0.9950099511131523,"entropy_confidence":0.9743281885559,"candidate_mass":0.999981128411809,"accepted":true,"accepted_predicted":"0","input_tokens":420,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.005002540754809138,"brier_hard":3.761753936438144e-05,"score_mae":0.009508336037134961,"within_one":true,"case_id":"typed:231/r0","base_id":"typed:231","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.219702}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9983909078084996,"hold":0.0008167210838935508,"manual_review":0.00026781998070140396,"reject":0.0005245511269071399},"predicted":"approve","confidence":0.9983909078084996,"entropy_confidence":0.9902043552324844,"candidate_mass":0.9999761007198668,"accepted":true,"accepted_predicted":"approve","input_tokens":424,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0016103881707607624,"brier_hard":3.603092436426284e-06,"soft_accuracy":0.9584787918048284,"brier_soft":0.0022500528777787197,"tvd":0.038915458935405906,"kl":0.10942591367458676,"case_id":"typed:231/r0","base_id":"typed:231","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.219702}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0016784213965068179,"true":0.9983215786034944},"predicted":"true","confidence":0.9983215786034944,"entropy_confidence":0.9821077544180192,"candidate_mass":0.9999834244507286,"accepted":true,"accepted_predicted":"true","input_tokens":372,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.389901572246433,"brier_hard":1.9932919486107434,"soft_accuracy":0.02161128454064657,"brier_soft":1.9142262223224638,"tvd":0.9783215786034938,"kl":6.164098024152248,"case_id":"typed:231/r0","base_id":"typed:231","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.219702}
{"id":"matches_order","type":"binary","probabilities":{"false":5.310140601233326e-09,"true":0.9999999946898601},"predicted":"true","confidence":0.9999999946898601,"entropy_confidence":0.9999998463707418,"candidate_mass":0.9999999559921778,"accepted":true,"accepted_predicted":"true","input_tokens":392,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.310139897462056e-09,"brier_hard":5.639517878575184e-17,"soft_accuracy":0.9966669947252575,"brier_soft":2.2217707205267093e-05,"tvd":0.003332994689859788,"kl":0.02737063037803052,"case_id":"typed:231/r0","base_id":"typed:231","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.219702}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.958566378223174e-05,"1":4.22775051980353e-05,"2":0.00015866521073313358,"3":0.9997294716202849},"predicted":"3","confidence":0.9997294716202849,"entropy_confidence":0.9980159075810943,"candidate_mass":0.9999982504307561,"accepted":true,"accepted_predicted":"3","input_tokens":419,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00027056497911813406,"brier_hard":1.0498980537804249e-07,"score_mae":0.41621502278751876,"within_one":true,"case_id":"typed:231/r0","base_id":"typed:231","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.219702}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9988507989588518,"1":0.00023071609670719467,"2":0.0005027216233254861,"3":0.0004157633211172695},"predicted":"0","confidence":0.9988507989588518,"entropy_confidence":0.9926884778540738,"candidate_mass":0.9999878641690776,"accepted":true,"accepted_predicted":"0","input_tokens":385,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.59547397381748,"brier_hard":1.9966979541521723,"score_mae":1.1841835506932898,"within_one":false,"case_id":"typed:232/r0","base_id":"typed:232","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.762802}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997745889167277,"hold":0.00013822595027821488,"manual_review":6.398965070965947e-05,"reject":2.31954822830158e-05},"predicted":"approve","confidence":0.9997745889167277,"entropy_confidence":0.9983270409622202,"candidate_mass":0.999989104075175,"accepted":true,"accepted_predicted":"approve","input_tokens":389,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00022543649216894284,"brier_hard":7.454927558861223e-08,"soft_accuracy":0.34998291710391305,"brier_soft":0.6229390159529048,"tvd":0.6497745889167283,"kl":4.9088186657553905,"case_id":"typed:232/r0","base_id":"typed:232","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.762802}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003120982999128081,"true":0.9968790170008731},"predicted":"true","confidence":0.9968790170008731,"entropy_confidence":0.9695260192512491,"candidate_mass":0.9999292025317325,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.769607262959827,"brier_hard":1.9875355490732516,"soft_accuracy":0.016370758866473347,"brier_soft":1.934725535116561,"tvd":0.9835460170008725,"kl":5.621913646244202,"case_id":"typed:232/r0","base_id":"typed:232","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.762802}
{"id":"matches_order","type":"binary","probabilities":{"false":4.653638278126733e-09,"true":0.9999999953463625},"predicted":"true","confidence":0.9999999953463625,"entropy_confidence":0.9999998644781953,"candidate_mass":0.9999999386061768,"accepted":true,"accepted_predicted":"true","input_tokens":357,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.653637486199568e-09,"brier_hard":4.3312690975827415e-17,"soft_accuracy":0.8499999967424539,"brier_soft":0.04499999720781733,"tvd":0.14999999534636216,"kl":1.2434099696789107,"case_id":"typed:232/r0","base_id":"typed:232","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.762802}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004669840890153376,"1":0.8351925509629864,"2":0.16375145876490588,"3":0.0005890061830934685},"predicted":"1","confidence":0.8351925509629864,"entropy_confidence":0.6720270944726028,"candidate_mass":0.9999885288009571,"accepted":true,"accepted_predicted":"1","input_tokens":384,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.18009298075969224,"brier_hard":0.05397660050814562,"score_mae":0.08112948704207845,"within_one":true,"case_id":"typed:232/r0","base_id":"typed:232","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.762802}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.006783388065042846,"1":0.0023149943901047926,"2":0.2065628759994425,"3":0.7843387415454101},"predicted":"3","confidence":0.7843387415454101,"entropy_confidence":0.5929960001428335,"candidate_mass":0.9999588957457025,"accepted":false,"accepted_predicted":null,"input_tokens":376,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.24291428364864134,"brier_hard":0.08922937369204569,"score_mae":0.21845697102522044,"within_one":true,"case_id":"typed:233/r0","base_id":"typed:233","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":319.716902}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997455785107352,"hold":0.0001587230074832532,"manual_review":7.538490271604813e-05,"reject":2.0313579065174812e-05},"predicted":"approve","confidence":0.9997455785107352,"entropy_confidence":0.998140331795944,"candidate_mass":0.9999891533753803,"accepted":true,"accepted_predicted":"approve","input_tokens":380,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.492903532229043,"brier_hard":1.9993404832348507,"soft_accuracy":0.050077422239365364,"brier_soft":1.446558486339552,"tvd":0.9497455785107353,"kl":8.059131837840965,"case_id":"typed:233/r0","base_id":"typed:233","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":319.716902}
{"id":"duplicate","type":"binary","probabilities":{"false":0.03797149801708545,"true":0.962028501982914},"predicted":"true","confidence":0.962028501982914,"entropy_confidence":0.7670869987325257,"candidate_mass":0.9999142831647927,"accepted":true,"accepted_predicted":"true","input_tokens":328,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.2709194528675845,"brier_hard":1.85099767725498,"soft_accuracy":0.07493377817571859,"brier_soft":1.7002731169377139,"tvd":0.9220285019829142,"kl":2.973686975055286,"case_id":"typed:233/r0","base_id":"typed:233","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":319.716902}
{"id":"matches_order","type":"binary","probabilities":{"false":5.645812064278644e-07,"true":0.9999994354187929},"predicted":"true","confidence":0.9999994354187929,"entropy_confidence":0.9999874668560432,"candidate_mass":0.999999155905429,"accepted":true,"accepted_predicted":"true","input_tokens":348,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.387181608206244,"brier_hard":1.9999977416758103,"soft_accuracy":0.050000508123085756,"brier_soft":1.8049978545920515,"tvd":0.9499994354187932,"kl":8.600036767928744,"case_id":"typed:233/r0","base_id":"typed:233","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":319.716902}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00042055026359663633,"1":0.9747878960363532,"2":0.024377332312443087,"3":0.00041422138760603244},"predicted":"1","confidence":0.9747878960363532,"entropy_confidence":0.9120481069462089,"candidate_mass":0.9999931075168182,"accepted":true,"accepted_predicted":"1","input_tokens":375,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.025535374176914885,"brier_hard":0.0012302529588271787,"score_mae":0.39188177517594247,"within_one":true,"case_id":"typed:233/r0","base_id":"typed:233","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":319.716902}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9997316174758343,"1":0.0001199898609520481,"2":9.275299362515429e-05,"3":5.5639669588930127e-05},"predicted":"0","confidence":0.9997316174758343,"entropy_confidence":0.9980105435852923,"candidate_mass":0.9999927681458429,"accepted":true,"accepted_predicted":"0","input_tokens":379,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0002684185452004494,"brier_hard":9.812563666724957e-08,"score_mae":0.03786058514303085,"within_one":true,"case_id":"typed:234/r0","base_id":"typed:234","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.15970100000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998191675630852,"hold":0.00010396740728190212,"manual_review":5.0649327925653993e-05,"reject":2.6215701706381334e-05},"predicted":"approve","confidence":0.9998191675630852,"entropy_confidence":0.9986208909556931,"candidate_mass":0.9999881421170654,"accepted":true,"accepted_predicted":"approve","input_tokens":383,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00018084878907128528,"brier_hard":4.6762209452745075e-08,"soft_accuracy":0.9165070441815817,"brier_soft":0.010156573126813897,"tvd":0.08315308422916894,"kl":0.4332171391097748,"case_id":"typed:234/r0","base_id":"typed:234","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.15970100000004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0019470248620628283,"true":0.998052975137937},"predicted":"true","confidence":0.998052975137937,"entropy_confidence":0.9796617625844077,"candidate_mass":0.9999742808626407,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.24145278325253,"brier_hard":1.9922194823633754,"soft_accuracy":0.0251891649998498,"brier_soft":1.9001580598658012,"tvd":0.9747199751379371,"kl":5.985125028061115,"case_id":"typed:234/r0","base_id":"typed:234","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.15970100000004}
{"id":"matches_order","type":"binary","probabilities":{"false":3.99638415040976e-09,"true":0.9999999960036163},"predicted":"true","confidence":0.9999999960036163,"entropy_confidence":0.9999998827406841,"candidate_mass":0.9999999483102289,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.996383676422006e-09,"brier_hard":3.1942168702992036e-17,"soft_accuracy":0.9833329961368318,"brier_soft":0.0005555775115691086,"tvd":0.016666996003616076,"kl":0.13698141950082396,"case_id":"typed:234/r0","base_id":"typed:234","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.15970100000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.113945040041066e-05,"1":6.33176054263163e-05,"2":0.00021710647072075497,"3":0.9996384364734511},"predicted":"3","confidence":0.9996384364734511,"entropy_confidence":0.997425358268146,"candidate_mass":0.9999979813188614,"accepted":true,"accepted_predicted":"3","input_tokens":378,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0003616289064005781,"brier_hard":1.8845613292752043e-07,"score_mae":0.4060798399672212,"within_one":true,"case_id":"typed:234/r0","base_id":"typed:234","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":322.15970100000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0034240095533954826,"1":0.0010171864446477427,"2":0.46313729894015543,"3":0.5324215050618007},"predicted":"3","confidence":0.5324215050618007,"entropy_confidence":0.48168703642000643,"candidate_mass":0.9999686584371382,"accepted":false,"accepted_predicted":null,"input_tokens":418,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.6303198005988797,"brier_hard":0.4331385651079394,"score_mae":0.242110700489639,"within_one":true,"case_id":"typed:235/r0","base_id":"typed:235","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.372702}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9763480833998391,"hold":0.022176579532612272,"manual_review":0.0009006955603179119,"reject":0.0005746415072307534},"predicted":"approve","confidence":0.9763480833998391,"entropy_confidence":0.9145649316797121,"candidate_mass":0.9999787054925312,"accepted":true,"accepted_predicted":"approve","input_tokens":422,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.012343248333043,"brier_hard":1.9519471309830239,"soft_accuracy":0.021042746107077497,"brier_soft":1.5105518076675049,"tvd":0.9596810833998392,"kl":5.586204634881126,"case_id":"typed:235/r0","base_id":"typed:235","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.372702}
{"id":"duplicate","type":"binary","probabilities":{"false":0.014956384067467755,"true":0.9850436159325335},"predicted":"true","confidence":0.9850436159325335,"entropy_confidence":0.8879026961466495,"candidate_mass":0.9999799957267819,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.202617042371653,"brier_hard":1.9406218505788786,"soft_accuracy":0.03759142944857533,"brier_soft":1.8497746175946634,"tvd":0.9617106159325328,"kl":3.994167580652528,"case_id":"typed:235/r0","base_id":"typed:235","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.372702}
{"id":"matches_order","type":"binary","probabilities":{"false":7.383478264548157e-08,"true":0.9999999261652174},"predicted":"true","confidence":0.9999999261652174,"entropy_confidence":0.9999981442499272,"candidate_mass":0.9999998033061492,"accepted":true,"accepted_predicted":"true","input_tokens":390,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.42143590681208,"brier_hard":1.9999997046608804,"soft_accuracy":0.033333068912513024,"brier_soft":1.8688898922834196,"tvd":0.9666669261652174,"kl":8.789959653163624,"case_id":"typed:235/r0","base_id":"typed:235","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.372702}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.1251573460014725e-05,"1":3.1639956413163395e-05,"2":0.0001107718325492509,"3":0.9998063366375766},"predicted":"3","confidence":0.9998063366375766,"entropy_confidence":0.9985308420995269,"candidate_mass":0.9999982124563533,"accepted":true,"accepted_predicted":"3","input_tokens":417,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.108038034839762,"brier_hard":1.999391183013762,"score_mae":0.7663391935342414,"within_one":true,"case_id":"typed:235/r0","base_id":"typed:235","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.372702}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999982081931792,"1":9.979554315878928e-07,"2":5.722366206744755e-07,"3":2.2161476985727927e-07},"predicted":"0","confidence":0.9999982081931792,"entropy_confidence":0.9999803779551164,"candidate_mass":0.9999996983108297,"accepted":true,"accepted_predicted":"0","input_tokens":413,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.7918084260933562e-06,"brier_hard":4.5830545827810765e-12,"score_mae":0.04333019272701749,"within_one":true,"case_id":"typed:236/r0","base_id":"typed:236","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.718402}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997920656158334,"hold":0.00011346340381635178,"manual_review":4.907914885398403e-05,"reject":4.539183149595194e-05},"predicted":"approve","confidence":0.9997920656158334,"entropy_confidence":0.9984278149834331,"candidate_mass":0.9999902442410322,"accepted":true,"accepted_predicted":"approve","input_tokens":417,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00020795600551787885,"brier_hard":6.057983334310441e-08,"soft_accuracy":0.9364768207483966,"brier_soft":0.005675585870707339,"tvd":0.06312506561583356,"kl":0.3138117014065261,"case_id":"typed:236/r0","base_id":"typed:236","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.718402}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0038559826102658873,"true":0.9961440173897326},"predicted":"true","confidence":0.9961440173897326,"entropy_confidence":0.963527819483711,"candidate_mass":0.9999774262524577,"accepted":true,"accepted_predicted":"true","input_tokens":365,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.558129412029825,"brier_hard":1.9846058067627148,"soft_accuracy":0.02039444728593526,"brier_soft":1.9187504551893761,"tvd":0.9794770173897334,"kl":5.3807890261180615,"case_id":"typed:236/r0","base_id":"typed:236","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.718402}
{"id":"matches_order","type":"binary","probabilities":{"false":1.1105098889889695e-08,"true":0.999999988894901},"predicted":"true","confidence":0.999999988894901,"entropy_confidence":0.9999996905353469,"candidate_mass":0.99999992011163,"accepted":true,"accepted_predicted":"true","input_tokens":385,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.1105099013186506e-08,"brier_hard":2.4664644407738864e-16,"soft_accuracy":0.9833329892650785,"brier_soft":0.000555577037645511,"tvd":0.016666988894901067,"kl":0.1369814264910583,"case_id":"typed:236/r0","base_id":"typed:236","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.718402}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00013075729971028652,"1":5.844639031450055e-05,"2":0.00020207116070429596,"3":0.9996087251492717},"predicted":"3","confidence":0.9996087251492717,"entropy_confidence":0.9972234254564885,"candidate_mass":0.9999974537229067,"accepted":true,"accepted_predicted":"3","input_tokens":412,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0003913514187061251,"brier_hard":2.1444221476916967e-07,"score_mae":0.6792887641595384,"within_one":true,"case_id":"typed:236/r0","base_id":"typed:236","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.718402}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9991053875397328,"1":0.00023731317306590564,"2":0.0003519964370473549,"3":0.00030530285015493964},"predicted":"0","confidence":0.9991053875397328,"entropy_confidence":0.9941245669011243,"candidate_mass":0.9999868308721522,"accepted":true,"accepted_predicted":"0","input_tokens":417,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0008950128648166738,"brier_hard":1.0737603181827567e-06,"score_mae":0.049809785402374566,"within_one":true,"case_id":"typed:237/r0","base_id":"typed:237","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.235902}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9968952065445074,"hold":0.0015497104790096726,"manual_review":0.0005853766241722579,"reject":0.0009697063523097643},"predicted":"approve","confidence":0.9968952065445074,"entropy_confidence":0.9825350639361615,"candidate_mass":0.9999512343635107,"accepted":true,"accepted_predicted":"approve","input_tokens":421,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00310962332645199,"brier_hard":1.3324341171859523e-05,"soft_accuracy":0.8840401947649633,"brier_soft":0.01716838345625988,"tvd":0.11022820654450784,"kl":0.3215243476474787,"case_id":"typed:237/r0","base_id":"typed:237","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.235902}
{"id":"duplicate","type":"binary","probabilities":{"false":0.001934334748153185,"true":0.9980656652518485},"predicted":"true","confidence":0.9980656652518485,"entropy_confidence":0.9797760551082969,"candidate_mass":0.9999819581091494,"accepted":true,"accepted_predicted":"true","input_tokens":369,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.247991811016995,"brier_hard":1.9922701443092263,"soft_accuracy":0.3439390928439784,"brier_soft":0.857349725895576,"tvd":0.6547326652518477,"kl":3.460293278661887,"case_id":"typed:237/r0","base_id":"typed:237","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.235902}
{"id":"matches_order","type":"binary","probabilities":{"false":2.163276127546105e-09,"true":0.9999999978367242},"predicted":"true","confidence":0.9999999978367242,"entropy_confidence":0.9999999346110207,"candidate_mass":0.9999999640519934,"accepted":true,"accepted_predicted":"true","input_tokens":389,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.1632757830709364e-09,"brier_hard":9.359525707508429e-18,"soft_accuracy":0.9766669979376756,"brier_soft":0.0010888575760971426,"tvd":0.023332997836724094,"kl":0.1918462307013351,"case_id":"typed:237/r0","base_id":"typed:237","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.235902}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.32523958378059e-05,"1":4.8471995248373214e-05,"2":0.00016272583076261854,"3":0.9997055497781502},"predicted":"3","confidence":0.9997055497781502,"entropy_confidence":0.9978521695587518,"candidate_mass":0.9999973355440913,"accepted":true,"accepted_predicted":"3","input_tokens":416,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.72344379320241,"brier_hard":1.9990857703559,"score_mae":0.6828235729912242,"within_one":true,"case_id":"typed:237/r0","base_id":"typed:237","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.235902}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999893920805931,"1":5.726447869396346e-06,"2":3.5205512010187556e-06,"3":1.360920334884586e-06},"predicted":"0","confidence":0.9999893920805931,"entropy_confidence":0.9998973392078158,"candidate_mass":0.9999987679624192,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.0607975671254743e-05,"brier_hard":1.5956654426066746e-10,"score_mae":0.056650149688723915,"within_one":true,"case_id":"typed:238/r0","base_id":"typed:238","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.39590200000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.999999532931805,"hold":2.567676218278223e-07,"manual_review":1.7444884461156734e-07,"reject":3.5851726967580944e-08},"predicted":"approve","confidence":0.999999532931805,"entropy_confidence":0.9999944507546114,"candidate_mass":0.9999994697810399,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":4.6706830408557373e-07,"brier_hard":3.158000561211274e-13,"soft_accuracy":0.3833328964440571,"brier_soft":0.5256443485318116,"tvd":0.6166665329318058,"kl":5.312153832079427,"case_id":"typed:238/r0","base_id":"typed:238","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.39590200000004}
{"id":"duplicate","type":"binary","probabilities":{"false":1.821069764772347e-05,"true":0.9999817893023507},"predicted":"true","confidence":0.9999817893023507,"entropy_confidence":0.9996870029751995,"candidate_mass":0.9999986291798723,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.8210863466057708e-05,"brier_hard":6.632590176906637e-10,"soft_accuracy":0.5399985431441874,"brier_soft":0.4231664929795857,"tvd":0.45998178930235145,"kl":3.9040258896864546,"case_id":"typed:238/r0","base_id":"typed:238","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.39590200000004}
{"id":"matches_order","type":"binary","probabilities":{"false":1.3311806337635207e-09,"true":0.9999999986688195},"predicted":"true","confidence":0.9999999986688195,"entropy_confidence":0.999999958830122,"candidate_mass":0.9999999785082674,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.3311805000504958e-09,"brier_hard":3.544083401062829e-18,"soft_accuracy":0.9733329987398167,"brier_soft":0.0014222576360056346,"tvd":0.026666998668819428,"kl":0.21930391588054382,"case_id":"typed:238/r0","base_id":"typed:238","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.39590200000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004958686040253151,"1":0.9936913472290916,"2":0.005750182915897928,"3":6.260125098557442e-05},"predicted":"1","confidence":0.9936913472290916,"entropy_confidence":0.9709078892557532,"candidate_mass":0.999995704684624,"accepted":true,"accepted_predicted":"1","input_tokens":406,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.0063286364117019316,"brier_hard":7.311350793925784e-05,"score_mae":0.46128748318615576,"within_one":true,"case_id":"typed:238/r0","base_id":"typed:238","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":329.39590200000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0009733984636074926,"1":0.0007690117609112009,"2":0.9409392439293668,"3":0.05731834584611442},"predicted":"2","confidence":0.9409392439293668,"entropy_confidence":0.8316183637869956,"candidate_mass":0.9999947032933868,"accepted":true,"accepted_predicted":"2","input_tokens":411,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0608767069122607,"brier_hard":0.0067751045618269935,"score_mae":0.33873046284201225,"within_one":true,"case_id":"typed:239/r0","base_id":"typed:239","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.453302}
{"id":"disposition","type":"choice","probabilities":{"approve":0.8231530044595124,"hold":0.16598258321722256,"manual_review":0.005045475118501451,"reject":0.005818937204762216},"predicted":"approve","confidence":0.8231530044595124,"entropy_confidence":0.6285675008324787,"candidate_mass":0.9999860323216744,"accepted":true,"accepted_predicted":"approve","input_tokens":415,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":5.289263453495604,"brier_hard":1.6950994532945458,"soft_accuracy":0.06965717120122142,"brier_soft":1.0892085811305428,"tvd":0.8064860211264966,"kl":3.3151071594020207,"case_id":"typed:239/r0","base_id":"typed:239","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.453302}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003083239263655236,"true":0.9969167607363462},"predicted":"true","confidence":0.9969167607363462,"entropy_confidence":0.9698403513621119,"candidate_mass":0.9999572617626858,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.781774525676113,"brier_hard":1.9876860556740956,"soft_accuracy":0.032898244907835966,"brier_soft":1.8698560443857342,"tvd":0.9669167607363454,"kl":5.47367176179322,"case_id":"typed:239/r0","base_id":"typed:239","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.453302}
{"id":"matches_order","type":"binary","probabilities":{"false":0.0007091646962666184,"true":0.9992908353037334},"predicted":"true","confidence":0.9992908353037334,"entropy_confidence":0.991558262998061,"candidate_mass":0.9999827973337615,"accepted":true,"accepted_predicted":"true","input_tokens":383,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.251422764661168,"brier_hard":1.9971643470440665,"soft_accuracy":0.02400907081655064,"brier_soft":1.9049873925814982,"tvd":0.9759578353037334,"kl":6.971500458488231,"case_id":"typed:239/r0","base_id":"typed:239","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.453302}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000631759634403515,"1":0.28873562707556666,"2":0.7097928951071087,"3":0.0008397181829210464},"predicted":"2","confidence":0.7097928951071087,"entropy_confidence":0.5581127989420526,"candidate_mass":0.9999879380606456,"accepted":false,"accepted_predicted":null,"input_tokens":410,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":1.2422437947466918,"brier_hard":1.0097040663829808,"score_mae":0.36084057183854723,"within_one":true,"case_id":"typed:239/r0","base_id":"typed:239","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.453302}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999839557544402,"1":9.211649879107521e-06,"2":4.814861232130679e-06,"3":2.0177344485054692e-06},"predicted":"0","confidence":0.9999839557544402,"entropy_confidence":0.9998497682513293,"candidate_mass":0.9999982233368998,"accepted":true,"accepted_predicted":"0","input_tokens":394,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.604437427007206e-05,"brier_hard":3.6952645006738515e-10,"score_mae":0.9566421054243112,"within_one":true,"case_id":"typed:240/r0","base_id":"typed:240","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":312.873702}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999716902877074,"hold":1.7318318034020132e-05,"manual_review":8.240133347346866e-06,"reject":2.7512609118540637e-06},"predicted":"approve","confidence":0.9999716902877074,"entropy_confidence":0.9997476209495916,"candidate_mass":0.9999927108093614,"accepted":true,"accepted_predicted":"approve","input_tokens":398,"reused_prefix_tokens":0,"expected":"reject","correct":false,"accepted_correct":false,"nll_hard":12.803451237969465,"brier_hard":1.999937879230424,"soft_accuracy":0.3733256822271892,"brier_soft":0.6048698928538696,"tvd":0.626638690287707,"kl":5.341172402560466,"case_id":"typed:240/r0","base_id":"typed:240","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":312.873702}
{"id":"duplicate","type":"binary","probabilities":{"false":7.957498512761304e-05,"true":0.9999204250148713},"predicted":"true","confidence":0.9999204250148713,"entropy_confidence":0.9988016036578776,"candidate_mass":0.9999961144695951,"accepted":true,"accepted_predicted":"true","input_tokens":346,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.957815138577214e-05,"brier_hard":1.2664356516288207e-08,"soft_accuracy":0.6166484324504196,"brier_soft":0.29376637557126006,"tvd":0.3832534250148719,"kl":2.9525844812550845,"case_id":"typed:240/r0","base_id":"typed:240","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":312.873702}
{"id":"matches_order","type":"binary","probabilities":{"false":1.0387167035641966e-08,"true":0.9999999896128333},"predicted":"true","confidence":0.9999999896128333,"entropy_confidence":0.9999997095403501,"candidate_mass":0.9999999208298322,"accepted":true,"accepted_predicted":"true","input_tokens":366,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.0387166736281808e-08,"brier_hard":2.1578647071294137e-16,"soft_accuracy":0.6899999960528768,"brier_soft":0.19219998711991337,"tvd":0.3099999896128332,"kl":2.5991715823200883,"case_id":"typed:240/r0","base_id":"typed:240","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":312.873702}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003200201662768831,"1":0.9432933207054546,"2":0.05598277722834765,"3":0.0004038818999207332},"predicted":"1","confidence":0.9432933207054546,"entropy_confidence":0.8397302269132477,"candidate_mass":0.9999904516244318,"accepted":true,"accepted_predicted":"1","input_tokens":393,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.05837799412585549,"brier_hard":0.0063499843563091345,"score_mae":0.19352947913808793,"within_one":true,"case_id":"typed:240/r0","base_id":"typed:240","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":312.873702}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999903014342172,"1":5.6026411161743025e-06,"2":2.8913605505546764e-06,"3":1.2045641161820123e-06},"predicted":"0","confidence":0.9999903014342172,"entropy_confidence":0.9999056907040587,"candidate_mass":0.9999986451381927,"accepted":true,"accepted_predicted":"0","input_tokens":398,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":9.698612814205132e-06,"brier_hard":1.352627062634742e-10,"score_mae":1.036652000945434,"within_one":false,"case_id":"typed:241/r0","base_id":"typed:241","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.406901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9968112601996036,"hold":0.001602924568236722,"manual_review":0.0004792945935920522,"reject":0.0011065206385675152},"predicted":"approve","confidence":0.9968112601996036,"entropy_confidence":0.9821864191528367,"candidate_mass":0.999960560128352,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"reject","correct":false,"accepted_correct":false,"nll_hard":6.806534746530069,"brier_hard":1.9914236706619894,"soft_accuracy":0.2765751156635742,"brier_soft":0.8053530175455373,"tvd":0.7201445368663271,"kl":3.8027419989278974,"case_id":"typed:241/r0","base_id":"typed:241","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.406901}
{"id":"duplicate","type":"binary","probabilities":{"false":1.5121157414017486e-05,"true":0.9999848788425854},"predicted":"true","confidence":0.9999848788425854,"entropy_confidence":0.9997360487651962,"candidate_mass":0.9999982080204395,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.5121271740496978e-05,"brier_hard":4.5729880309793025e-10,"soft_accuracy":0.6466625644504107,"brier_soft":0.2496670470196481,"tvd":0.3533178788425857,"kl":2.9724294370208164,"case_id":"typed:241/r0","base_id":"typed:241","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.406901}
{"id":"matches_order","type":"binary","probabilities":{"false":2.4943885102117672e-09,"true":0.9999999975056113},"predicted":"true","confidence":0.9999999975056113,"entropy_confidence":0.9999999251150601,"candidate_mass":0.999999971906695,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.4943886987062495e-09,"brier_hard":1.2443949004589913e-17,"soft_accuracy":0.6499999992516834,"brier_soft":0.24499999650785592,"tvd":0.3499999975056114,"kl":2.9436102363529213,"case_id":"typed:241/r0","base_id":"typed:241","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.406901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00020704933477755093,"1":0.00012879485708945643,"2":0.00044965356100556454,"3":0.999214502247127},"predicted":"3","confidence":0.999214502247127,"entropy_confidence":0.9948346872361299,"candidate_mass":0.999995652207376,"accepted":true,"accepted_predicted":"3","input_tokens":397,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0007858064178805369,"brier_hard":8.78652586937966e-07,"score_mae":0.615338608720482,"within_one":true,"case_id":"typed:241/r0","base_id":"typed:241","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":326.406901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9998959858702531,"1":4.3696797888647896e-05,"2":4.017794069621944e-05,"3":2.0139391162871176e-05},"predicted":"0","confidence":0.9998959858702531,"entropy_confidence":0.999158115166411,"candidate_mass":0.999994680926377,"accepted":true,"accepted_predicted":"0","input_tokens":409,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00010401953959165132,"brier_hard":1.4748211327730726e-08,"score_mae":0.1031485291472303,"within_one":true,"case_id":"typed:242/r0","base_id":"typed:242","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.4866}
{"id":"disposition","type":"choice","probabilities":{"approve":0.998159192066549,"hold":0.0008945396905663202,"manual_review":0.00034379473472826116,"reject":0.000602473508155024},"predicted":"approve","confidence":0.998159192066549,"entropy_confidence":0.988943899517221,"candidate_mass":0.9999637174295224,"accepted":true,"accepted_predicted":"approve","input_tokens":413,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0018425043024876272,"brier_hard":4.669944253510064e-06,"soft_accuracy":0.5659134033560418,"brier_soft":0.25420772292126786,"tvd":0.43149219206654965,"kl":2.0611156225890035,"case_id":"typed:242/r0","base_id":"typed:242","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.4866}
{"id":"duplicate","type":"binary","probabilities":{"false":1.1804023614948846e-05,"true":0.9999881959763848},"predicted":"true","confidence":0.9999881959763848,"entropy_confidence":0.9997897343573494,"candidate_mass":0.9999987382867951,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.347070100296897,"brier_hard":1.9999527841842095,"soft_accuracy":0.45333410171673993,"brier_soft":0.5976638065759596,"tvd":0.5466551959763849,"kl":4.6763499147109755,"case_id":"typed:242/r0","base_id":"typed:242","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.4866}
{"id":"matches_order","type":"binary","probabilities":{"false":2.8885894795245304e-09,"true":0.9999999971114111},"predicted":"true","confidence":0.9999999971114111,"entropy_confidence":0.999999913892065,"candidate_mass":0.9999999559826139,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.8885889277827418e-09,"brier_hard":1.6687895150826615e-17,"soft_accuracy":0.9866669971884381,"brier_soft":0.0003555376239457795,"tvd":0.013332997111410838,"kl":0.10955775316097156,"case_id":"typed:242/r0","base_id":"typed:242","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.4866}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001545463387128175,"1":8.910723398354399e-05,"2":0.0004577995746097101,"3":0.9992985468526945},"predicted":"3","confidence":0.9992985468526945,"entropy_confidence":0.9953773197600494,"candidate_mass":0.9999971481755935,"accepted":true,"accepted_predicted":"3","input_tokens":408,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0007016992806718657,"brier_hard":7.334416383353855e-07,"score_mae":0.5655673469412865,"within_one":true,"case_id":"typed:242/r0","base_id":"typed:242","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.4866}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999275546709286,"1":3.7423542592090525e-05,"2":2.4435293664809523e-05,"3":1.0586492814224186e-05},"predicted":"0","confidence":0.9999275546709286,"entropy_confidence":0.9993979077766196,"candidate_mass":0.9999945924687359,"accepted":true,"accepted_predicted":"0","input_tokens":402,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":7.244795336095572e-05,"brier_hard":7.358004650990436e-09,"score_mae":1.3032149463916356,"within_one":false,"case_id":"typed:243/r0","base_id":"typed:243","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.56829999999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999756672097244,"hold":1.3186566619146826e-05,"manual_review":7.646533725232463e-06,"reject":3.4996899308194254e-06},"predicted":"approve","confidence":0.9999756672097244,"entropy_confidence":0.9997788689780114,"candidate_mass":0.9999953485488966,"accepted":true,"accepted_predicted":"approve","input_tokens":406,"reused_prefix_tokens":0,"expected":"reject","correct":false,"accepted_correct":false,"nll_hard":12.562836184587788,"brier_hard":1.9999443358762747,"soft_accuracy":0.2733311558731767,"brier_soft":0.729044481287783,"tvd":0.7266426672097246,"kl":6.333231725314071,"case_id":"typed:243/r0","base_id":"typed:243","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.56829999999997}
{"id":"duplicate","type":"binary","probabilities":{"false":2.816856220407213e-05,"true":0.9999718314377964},"predicted":"true","confidence":0.9999718314377964,"entropy_confidence":0.9995335784888452,"candidate_mass":0.9999987170863194,"accepted":true,"accepted_predicted":"true","input_tokens":354,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.816895894495532e-05,"brier_hard":1.5869357932603166e-09,"soft_accuracy":0.6133266151446798,"brier_soft":0.29897917195116913,"tvd":0.3866388314377962,"kl":3.261525800723049,"case_id":"typed:243/r0","base_id":"typed:243","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.56829999999997}
{"id":"matches_order","type":"binary","probabilities":{"false":5.324397548720846e-09,"true":0.9999999946756031},"predicted":"true","confidence":0.9999999946756031,"entropy_confidence":0.9999998459788657,"candidate_mass":0.9999999541665449,"accepted":true,"accepted_predicted":"true","input_tokens":374,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.324396937530891e-09,"brier_hard":5.669841185427047e-17,"soft_accuracy":0.9599999951015548,"brier_soft":0.003199999148096505,"tvd":0.0399999946756028,"kl":0.32922450525102376,"case_id":"typed:243/r0","base_id":"typed:243","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.56829999999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001248010251232249,"1":7.860914806894985e-05,"2":0.00027078302054927476,"3":0.9995258068062602},"predicted":"3","confidence":0.9995258068062602,"entropy_confidence":0.9967084251307547,"candidate_mass":0.9999976563904838,"accepted":true,"accepted_predicted":"3","input_tokens":401,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00047430565888712954,"brier_hard":3.199373232388346e-07,"score_mae":0.5658645956079482,"within_one":true,"case_id":"typed:243/r0","base_id":"typed:243","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.56829999999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0008001054240889304,"1":0.00036037653977889176,"2":0.06895597423301647,"3":0.9298835438031143},"predicted":"3","confidence":0.9298835438031143,"entropy_confidence":0.8120389307585653,"candidate_mass":0.9999890486001889,"accepted":true,"accepted_predicted":"3","input_tokens":384,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.07269592239281059,"brier_hard":0.009672013851974297,"score_mae":0.13792295641515473,"within_one":true,"case_id":"typed:244/r0","base_id":"typed:244","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.96320000000003}
{"id":"disposition","type":"choice","probabilities":{"approve":0.5420932059760311,"hold":0.45329276105582433,"manual_review":0.004186147901092377,"reject":0.00042788506705197956},"predicted":"approve","confidence":0.5420932059760311,"entropy_confidence":0.48291772840005365,"candidate_mass":0.999982374220363,"accepted":false,"accepted_predicted":null,"input_tokens":388,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":null,"nll_hard":0.7912170905962803,"brier_hard":0.5927715559992159,"soft_accuracy":0.23183694886683107,"brier_soft":0.4743717781552026,"tvd":0.5220932059760311,"kl":2.28883888782488,"case_id":"typed:244/r0","base_id":"typed:244","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.96320000000003}
{"id":"duplicate","type":"binary","probabilities":{"false":0.005234080021129756,"true":0.9947659199788711},"predicted":"true","confidence":0.9947659199788711,"entropy_confidence":0.952805550613699,"candidate_mass":0.9999554019756719,"accepted":true,"accepted_predicted":"true","input_tokens":336,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.252564186276924,"brier_hard":1.9791184711028178,"soft_accuracy":0.025024716820284582,"brier_soft":1.900337197504508,"tvd":0.9747659199788706,"kl":5.049578745787719,"case_id":"typed:244/r0","base_id":"typed:244","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.96320000000003}
{"id":"matches_order","type":"binary","probabilities":{"false":8.01117594878151e-05,"true":0.9999198882405108},"predicted":"true","confidence":0.9999198882405108,"entropy_confidence":0.9987942968945484,"candidate_mass":0.9999846925606861,"accepted":true,"accepted_predicted":"true","input_tokens":356,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.432087904580424,"brier_hard":1.999679565797834,"soft_accuracy":0.030075305053918505,"brier_soft":1.8814891792089723,"tvd":0.9699198882405115,"kl":8.828835827346357,"case_id":"typed:244/r0","base_id":"typed:244","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.96320000000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006418220302963333,"1":0.5285557591034138,"2":0.4693931298747091,"3":0.0014092889915800428},"predicted":"1","confidence":0.5285557591034138,"entropy_confidence":0.4907362899341111,"candidate_mass":0.9999847848419393,"accepted":false,"accepted_predicted":null,"input_tokens":383,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.6376069747840418,"brier_hard":0.44259198067911426,"score_mae":0.2049028858275721,"within_one":true,"case_id":"typed:244/r0","base_id":"typed:244","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.96320000000003}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.011991635073140024,"1":0.014070272218695258,"2":0.8532969657616624,"3":0.12064112694650139},"predicted":"2","confidence":0.8532969657616624,"entropy_confidence":0.6367592912426658,"candidate_mass":0.9999688720354152,"accepted":true,"accepted_predicted":"2","input_tokens":404,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":4.423545949503772,"brier_hard":1.719028495014737,"score_mae":1.3559205845815243,"within_one":false,"case_id":"typed:245/r0","base_id":"typed:245","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.8409}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999628910059258,"hold":2.1292019582046e-05,"manual_review":1.2572395398080902e-05,"reject":3.2445790932619167e-06},"predicted":"approve","confidence":0.9999628910059258,"entropy_confidence":0.9996760976176686,"candidate_mass":0.9999929727746416,"accepted":true,"accepted_predicted":"approve","input_tokens":408,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":3.710968262991929e-05,"brier_hard":1.999019958616705e-09,"soft_accuracy":0.6666469164962394,"brier_soft":0.19232124879639284,"tvd":0.3332991355850195,"kl":2.7259930664958083,"case_id":"typed:245/r0","base_id":"typed:245","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.8409}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0545445129418892,"true":0.9454554870581103},"predicted":"true","confidence":0.9454554870581103,"entropy_confidence":0.6946033042812219,"candidate_mass":0.9999666710326931,"accepted":true,"accepted_predicted":"true","input_tokens":356,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":2.908738159445395,"brier_hard":1.7877721560165782,"soft_accuracy":0.06998667285624566,"brier_soft":1.722822701965865,"tvd":0.9281224870581105,"kl":2.771823479330604,"case_id":"typed:245/r0","base_id":"typed:245","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.8409}
{"id":"matches_order","type":"binary","probabilities":{"false":5.607786463750303e-07,"true":0.9999994392213525},"predicted":"true","confidence":0.9999994392213525,"entropy_confidence":0.9999875458016919,"candidate_mass":0.9999993832915718,"accepted":true,"accepted_predicted":"true","input_tokens":376,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.607788047545387e-07,"brier_hard":6.289453817424791e-13,"soft_accuracy":0.6599998205508324,"brier_soft":0.23119923734166908,"tvd":0.339999439221353,"kl":2.857275903571214,"case_id":"typed:245/r0","base_id":"typed:245","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.8409}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00027912250850544606,"1":0.8602512648967529,"2":0.13920940657372072,"3":0.0002602060210212012},"predicted":"1","confidence":0.8602512648967529,"entropy_confidence":0.705390065769856,"candidate_mass":0.999989493429385,"accepted":true,"accepted_predicted":"1","input_tokens":403,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.9717759574043328,"brier_hard":1.4809928301041775,"score_mae":0.4438823038927422,"within_one":true,"case_id":"typed:245/r0","base_id":"typed:245","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":334.8409}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9996893286411498,"1":0.00010248650840838287,"2":0.00012672181766757275,"3":8.146303277356186e-05},"predicted":"0","confidence":0.9996893286411498,"entropy_confidence":0.9977232932515355,"candidate_mass":0.9999877724301858,"accepted":true,"accepted_predicted":"0","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":9.415361226218991,"brier_hard":1.9992158609315749,"score_mae":2.506066680757936,"within_one":false,"case_id":"typed:246/r0","base_id":"typed:246","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":310.494601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999867812883276,"hold":7.3243170205018906e-06,"manual_review":4.795236739218847e-06,"reject":1.099157911347833e-06},"predicted":"approve","confidence":0.9999867812883276,"entropy_confidence":0.9998747476311851,"candidate_mass":0.9999975467983404,"accepted":true,"accepted_predicted":"approve","input_tokens":391,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":11.824310646930144,"brier_hard":1.9999589141951966,"soft_accuracy":0.03667275204094383,"brier_soft":1.45831651652535,"tvd":0.9633197812883283,"kl":8.751408673554964,"case_id":"typed:246/r0","base_id":"typed:246","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":310.494601}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0015708669516387864,"true":0.9984291330483611},"predicted":"true","confidence":0.9984291330483611,"entropy_confidence":0.9831040934368541,"candidate_mass":0.9999628292407612,"accepted":true,"accepted_predicted":"true","input_tokens":339,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.456127613531365,"brier_hard":1.9937214674394041,"soft_accuracy":0.05141378025647491,"brier_soft":1.7990356408297319,"tvd":0.9484291330483612,"kl":5.934884594611761,"case_id":"typed:246/r0","base_id":"typed:246","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":310.494601}
{"id":"matches_order","type":"binary","probabilities":{"false":1.3930893904942465e-08,"true":0.9999999860691063},"predicted":"true","confidence":0.9999999860691063,"entropy_confidence":0.9999996163455233,"candidate_mass":0.9999998661674551,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.089156880004502,"brier_hard":1.9999999442764254,"soft_accuracy":0.10000001114471516,"brier_soft":1.619999949848783,"tvd":0.8999999860691063,"kl":8.05904782687225,"case_id":"typed:246/r0","base_id":"typed:246","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":310.494601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.009352884270110978,"1":0.9872314551491862,"2":0.0012069153597804312,"3":0.0022087452209214996},"predicted":"1","confidence":0.9872314551491862,"entropy_confidence":0.9437339455559938,"candidate_mass":0.9999910609838091,"accepted":true,"accepted_predicted":"1","input_tokens":386,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.012850763343099173,"brier_hard":0.0002568473819139505,"score_mae":0.7037284784684884,"within_one":true,"case_id":"typed:246/r0","base_id":"typed:246","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":310.494601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.5936740809427615,"1":0.009675057248980901,"2":0.04222016144496857,"3":0.35443070036328844},"predicted":"0","confidence":0.5936740809427615,"entropy_confidence":0.38275545568687863,"candidate_mass":0.9999552232975847,"accepted":false,"accepted_predicted":null,"input_tokens":381,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.0372424376958058,"brier_hard":0.7710847837818771,"score_mae":1.5425925187712168,"within_one":false,"case_id":"typed:247/r0","base_id":"typed:247","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.6144}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997613989784387,"hold":0.00013607168682103674,"manual_review":7.723655695397894e-05,"reject":2.5292777785917392e-05},"predicted":"approve","confidence":0.9997613989784387,"entropy_confidence":0.9982334370526152,"candidate_mass":0.9999887630374394,"accepted":true,"accepted_predicted":"approve","input_tokens":385,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.468637677342876,"brier_hard":1.9993684068941313,"soft_accuracy":0.023422938234755507,"brier_soft":1.377765921316528,"tvd":0.9764283989784389,"kl":8.117469870991531,"case_id":"typed:247/r0","base_id":"typed:247","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.6144}
{"id":"duplicate","type":"binary","probabilities":{"false":0.007003871573136802,"true":0.9929961284268639},"predicted":"true","confidence":0.9929961284268639,"entropy_confidence":0.9397998843763758,"candidate_mass":0.9999560336574975,"accepted":true,"accepted_predicted":"true","input_tokens":333,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.9612922009435465,"brier_hard":1.9720826221414802,"soft_accuracy":0.036583639278748616,"brier_soft":1.8547230867302567,"tvd":0.9629961284268636,"kl":4.677922122149766,"case_id":"typed:247/r0","base_id":"typed:247","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.6144}
{"id":"matches_order","type":"binary","probabilities":{"false":1.1027028222135884e-07,"true":0.9999998897297172},"predicted":"true","confidence":0.9999998897297172,"entropy_confidence":0.9999972922969447,"candidate_mass":0.9999997377476013,"accepted":true,"accepted_predicted":"true","input_tokens":353,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.020331373812464,"brier_hard":1.9999995589188944,"soft_accuracy":0.08333309189197531,"brier_soft":1.6805563734535083,"tvd":0.9166668897297174,"kl":8.235740027767996,"case_id":"typed:247/r0","base_id":"typed:247","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.6144}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005789219339204549,"1":0.9702828459924194,"2":0.028703765374731686,"3":0.00043446669892837253},"predicted":"1","confidence":0.9702828459924194,"entropy_confidence":0.8998270091185498,"candidate_mass":0.9999917648888309,"accepted":true,"accepted_predicted":"1","input_tokens":380,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.03016765618330197,"brier_hard":0.0017075393009159596,"score_mae":0.17100622316133207,"within_one":true,"case_id":"typed:247/r0","base_id":"typed:247","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.6144}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.04906338699386429,"1":0.0012568181808252606,"2":0.009793100799602807,"3":0.9398866940257065},"predicted":"3","confidence":0.9398866940257065,"entropy_confidence":0.8125391859222306,"candidate_mass":0.9999786903954042,"accepted":true,"accepted_predicted":"3","input_tokens":429,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.626077141270535,"brier_hard":1.8663054963658867,"score_mae":1.3538361018571499,"within_one":false,"case_id":"typed:248/r0","base_id":"typed:248","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":347.6448}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9737201977494327,"hold":0.02359464485922925,"manual_review":0.0011860890960243323,"reject":0.0014990682953136072},"predicted":"approve","confidence":0.9737201977494327,"entropy_confidence":0.9047291669848644,"candidate_mass":0.9999796599561405,"accepted":true,"accepted_predicted":"approve","input_tokens":433,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":6.737093858104821,"brier_hard":1.9463192065922765,"soft_accuracy":0.2830989801140393,"brier_soft":0.6993600849562508,"tvd":0.6903869144161494,"kl":2.768746508334768,"case_id":"typed:248/r0","base_id":"typed:248","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":347.6448}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0008927835506684783,"true":0.9991072164493312},"predicted":"true","confidence":0.9991072164493312,"entropy_confidence":0.9896691973047751,"candidate_mass":0.9999906717009305,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.021166390937427,"brier_hard":1.9964304599222622,"soft_accuracy":0.024184120913492976,"brier_soft":1.904270642974613,"tvd":0.9757742164493313,"kl":6.746620944545756,"case_id":"typed:248/r0","base_id":"typed:248","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":347.6448}
{"id":"matches_order","type":"binary","probabilities":{"false":5.179232307234396e-09,"true":0.999999994820767},"predicted":"true","confidence":0.999999994820767,"entropy_confidence":0.9999998499715731,"candidate_mass":0.9999999340698111,"accepted":true,"accepted_predicted":"true","input_tokens":401,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":19.078608994894807,"brier_hard":1.9999999792830694,"soft_accuracy":0.4733330002762289,"brier_soft":0.5547562468670764,"tvd":0.5266669948207674,"kl":4.496750020385,"case_id":"typed:248/r0","base_id":"typed:248","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":347.6448}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.873786458615202e-05,"1":4.269345981847167e-05,"2":0.0001837160869467813,"3":0.9997048525886468},"predicted":"3","confidence":0.9997048525886468,"entropy_confidence":0.9978620145829554,"candidate_mass":0.9999976639236686,"accepted":true,"accepted_predicted":"3","input_tokens":428,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.602118997738618,"brier_hard":1.9990424004146206,"score_mae":0.7495246833996525,"within_one":true,"case_id":"typed:248/r0","base_id":"typed:248","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":347.6448}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9695327672015448,"1":0.0016440304276179873,"2":0.0032090899695757973,"3":0.02561411240125968},"predicted":"0","confidence":0.9695327672015448,"entropy_confidence":0.8897570518291492,"candidate_mass":0.9999811520615185,"accepted":true,"accepted_predicted":"0","input_tokens":421,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.030941006828751186,"brier_hard":0.001597336122979395,"score_mae":0.15176245242945136,"within_one":true,"case_id":"typed:249/r0","base_id":"typed:249","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":348.45300000000003}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9986205449656912,"hold":0.000936372792967677,"manual_review":0.00020979336348819457,"reject":0.00023328887785199832},"predicted":"approve","confidence":0.9986205449656912,"entropy_confidence":0.9916062801910163,"candidate_mass":0.9999840070152451,"accepted":true,"accepted_predicted":"approve","input_tokens":425,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0013804073582974782,"brier_hard":2.8781271549831355e-06,"soft_accuracy":0.8322866256660363,"brier_soft":0.0393145922824647,"tvd":0.16528754496569167,"kl":0.667144471788186,"case_id":"typed:249/r0","base_id":"typed:249","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":348.45300000000003}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0014565851625157305,"true":0.9985434148374858},"predicted":"true","confidence":0.9985434148374858,"entropy_confidence":0.9841744380936306,"candidate_mass":0.9999804669525598,"accepted":true,"accepted_predicted":"true","input_tokens":373,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.531660512614511,"brier_hard":1.9941779026306115,"soft_accuracy":0.028045899649458157,"brier_soft":1.8890875314347266,"tvd":0.971876414837485,"kl":6.234561394358649,"case_id":"typed:249/r0","base_id":"typed:249","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":348.45300000000003}
{"id":"matches_order","type":"binary","probabilities":{"false":2.289823452858085e-08,"true":0.9999999771017658},"predicted":"true","confidence":0.9999999771017658,"entropy_confidence":0.9999993858033254,"candidate_mass":0.999999859646466,"accepted":true,"accepted_predicted":"true","input_tokens":393,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.2898234459819256e-08,"brier_hard":1.0486582738965347e-15,"soft_accuracy":0.9366669800021935,"brier_soft":0.008022131977145533,"tvd":0.06333297710176561,"kl":0.5220347754318124,"case_id":"typed:249/r0","base_id":"typed:249","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":348.45300000000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.275651599985234e-05,"1":3.585470009846189e-05,"2":0.00011744380378734748,"3":0.9997839449801135},"predicted":"3","confidence":0.9997839449801135,"entropy_confidence":0.9983747297790552,"candidate_mass":0.9999983264778335,"accepted":true,"accepted_predicted":"3","input_tokens":420,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00021607836313468324,"brier_hard":6.569675848580006e-08,"score_mae":0.6496225772480133,"within_one":true,"case_id":"typed:249/r0","base_id":"typed:249","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":348.45300000000003}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9943126418707243,"1":0.0004985740277429389,"2":0.001237666790148271,"3":0.003951117311384393},"predicted":"0","confidence":0.9943126418707243,"entropy_confidence":0.9714257601712012,"candidate_mass":0.9999826574542121,"accepted":true,"accepted_predicted":"0","input_tokens":429,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.694527292685631,"brier_hard":1.9861996879267956,"score_mae":1.2885057404578075,"within_one":false,"case_id":"typed:250/r0","base_id":"typed:250","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":345.2772}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9995099893342222,"hold":0.00031541310617468125,"manual_review":9.532861210168662e-05,"reject":7.926894750196139e-05},"predicted":"approve","confidence":0.9995099893342222,"entropy_confidence":0.9966358421514967,"candidate_mass":0.999989746128028,"accepted":true,"accepted_predicted":"approve","input_tokens":433,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.258180560447737,"brier_hard":1.9988296764112314,"soft_accuracy":0.283306231910867,"brier_soft":0.8029632475917009,"tvd":0.7161769893342219,"kl":5.304913793821009,"case_id":"typed:250/r0","base_id":"typed:250","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":345.2772}
{"id":"duplicate","type":"binary","probabilities":{"false":0.005366021063729764,"true":0.9946339789362698},"predicted":"true","confidence":0.9946339789362698,"entropy_confidence":0.9518091122364769,"candidate_mass":0.9999828516968362,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.227668601580191,"brier_hard":1.9785935041091929,"soft_accuracy":0.031746829696316786,"brier_soft":1.8739201446220188,"tvd":0.96796697893627,"kl":4.965447645532887,"case_id":"typed:250/r0","base_id":"typed:250","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":345.2772}
{"id":"matches_order","type":"binary","probabilities":{"false":3.0372965352076143e-09,"true":0.9999999969627034},"predicted":"true","confidence":0.9999999969627034,"entropy_confidence":0.9999999096791227,"candidate_mass":0.9999999411882214,"accepted":true,"accepted_predicted":"true","input_tokens":401,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.037296641033746e-09,"brier_hard":1.8450341100399475e-17,"soft_accuracy":0.649999999088811,"brier_soft":0.24499999574778478,"tvd":0.3499999969627034,"kl":2.9436102367058115,"case_id":"typed:250/r0","base_id":"typed:250","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":345.2772}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.6751458338186774e-05,"1":3.397013837436635e-05,"2":0.00015779882092145983,"3":0.9997614795823649},"predicted":"3","confidence":0.9997614795823649,"entropy_confidence":0.9982430934789872,"candidate_mass":0.999998389664995,"accepted":true,"accepted_predicted":"3","input_tokens":428,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00023854886815398159,"brier_hard":8.5132126670926e-08,"score_mae":0.6663010065273123,"within_one":true,"case_id":"typed:250/r0","base_id":"typed:250","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":345.2772}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0051036655055965606,"1":0.0013007428705459928,"2":0.10319313230961791,"3":0.8904024593142407},"predicted":"3","confidence":0.8904024593142407,"entropy_confidence":0.7307165383235182,"candidate_mass":0.9999896790096297,"accepted":true,"accepted_predicted":"3","input_tokens":416,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.11608171699155184,"brier_hard":0.02268818281384527,"score_mae":0.038894385432503675,"within_one":true,"case_id":"typed:251/r0","base_id":"typed:251","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":344.1751}
{"id":"disposition","type":"choice","probabilities":{"approve":0.997088717526821,"hold":0.0015203155176860502,"manual_review":0.0006229587652192749,"reject":0.000768008190271901},"predicted":"approve","confidence":0.997088717526821,"entropy_confidence":0.9834969173896229,"candidate_mass":0.9999732812441918,"accepted":true,"accepted_predicted":"approve","input_tokens":420,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.381030228816304,"brier_hard":1.992943282362319,"soft_accuracy":0.0009038021730525247,"brier_soft":1.3907150588801809,"tvd":0.9970887175268219,"kl":6.086273012697465,"case_id":"typed:251/r0","base_id":"typed:251","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":344.1751}
{"id":"duplicate","type":"binary","probabilities":{"false":0.008793444370735152,"true":0.9912065556292645},"predicted":"true","confidence":0.9912065556292645,"entropy_confidence":0.9273161223157027,"candidate_mass":0.9999768828583505,"accepted":true,"accepted_predicted":"true","input_tokens":368,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.733748792987981,"brier_hard":1.9649808718448611,"soft_accuracy":0.02516732369608106,"brier_soft":1.8994546909721695,"tvd":0.9745395556292646,"kl":4.570231175489762,"case_id":"typed:251/r0","base_id":"typed:251","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":344.1751}
{"id":"matches_order","type":"binary","probabilities":{"false":2.271104252417491e-06,"true":0.9999977288957492},"predicted":"true","confidence":0.9999977288957492,"entropy_confidence":0.9999541444337386,"candidate_mass":0.9999985943659377,"accepted":true,"accepted_predicted":"true","input_tokens":388,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.995244389942279,"brier_hard":1.9999909155933093,"soft_accuracy":0.023335165120901415,"brier_soft":1.9077479853380113,"tvd":0.9766647288957484,"kl":8.9077527847644,"case_id":"typed:251/r0","base_id":"typed:251","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":344.1751}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.672051679099438e-05,"1":4.609897582558651e-05,"2":0.00020859870897914758,"3":0.9996685817984029},"predicted":"3","confidence":0.9996685817984029,"entropy_confidence":0.9976292940671679,"candidate_mass":0.9999978425790277,"accepted":true,"accepted_predicted":"3","input_tokens":415,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00033147313274639877,"brier_hard":1.6136259900646846e-07,"score_mae":0.6494690417889926,"within_one":true,"case_id":"typed:251/r0","base_id":"typed:251","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":344.1751}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9679761229140043,"1":0.0007825363196787512,"2":0.007641789457989461,"3":0.023599551308328554},"predicted":"0","confidence":0.9679761229140043,"entropy_confidence":0.8825888580126944,"candidate_mass":0.9999721662569235,"accepted":true,"accepted_predicted":"0","input_tokens":411,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.7465275794902695,"brier_hard":1.8903946200461366,"score_mae":2.4964682308393567,"within_one":false,"case_id":"typed:252/r0","base_id":"typed:252","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.4188}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999481043213297,"hold":2.9948386825759218e-05,"manual_review":1.710924646082616e-05,"reject":4.838045382607275e-06},"predicted":"approve","confidence":0.9999481043213297,"entropy_confidence":0.9995593721890651,"candidate_mass":0.9999935710431093,"accepted":true,"accepted_predicted":"approve","input_tokens":415,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":10.41603509709854,"brier_hard":1.9998363157752082,"soft_accuracy":0.06002021492697868,"brier_soft":1.3370783804729027,"tvd":0.9399481043213304,"kl":8.471903361795446,"case_id":"typed:252/r0","base_id":"typed:252","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.4188}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004054703784008156,"true":0.9959452962159919},"predicted":"true","confidence":0.9959452962159919,"entropy_confidence":0.9619427368340723,"candidate_mass":0.9999821910494499,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.507877643649433,"brier_hard":1.9838140661095196,"soft_accuracy":0.030505450212391864,"brier_soft":1.8790008310307522,"tvd":0.9692782962159918,"kl":5.23814921867856,"case_id":"typed:252/r0","base_id":"typed:252","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.4188}
{"id":"matches_order","type":"binary","probabilities":{"false":2.080755534928367e-08,"true":0.9999999791924454},"predicted":"true","confidence":0.9999999791924454,"entropy_confidence":0.9999994390072452,"candidate_mass":0.9999998907776344,"accepted":true,"accepted_predicted":"true","input_tokens":383,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.68794967824774,"brier_hard":1.999999916769781,"soft_accuracy":0.06666701803320083,"brier_soft":1.7422209000964899,"tvd":0.933332979192445,"kl":8.41577736236211,"case_id":"typed:252/r0","base_id":"typed:252","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.4188}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00026737037138592463,"1":0.751889292268071,"2":0.2472336890711224,"3":0.0006096482894198219},"predicted":"1","confidence":0.751889292268071,"entropy_confidence":0.591273212513215,"candidate_mass":0.9999881731425289,"accepted":false,"accepted_predicted":null,"input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3974212800340666,"brier_hard":1.1319950698548045,"score_mae":1.0051473847214247,"within_one":false,"case_id":"typed:252/r0","base_id":"typed:252","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.4188}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999940300694267,"1":3.48677160901429e-06,"2":1.8706296766024878e-06,"3":6.125292878491942e-07},"predicted":"0","confidence":0.9999940300694267,"entropy_confidence":0.999939968475219,"candidate_mass":0.9999990420209605,"accepted":true,"accepted_predicted":"0","input_tokens":423,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":5.9699483934416175e-06,"brier_hard":5.1672094819329054e-11,"score_mae":0.2699909343811743,"within_one":true,"case_id":"typed:253/r0","base_id":"typed:253","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.3165}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999972961203765,"hold":1.5767760888570035e-06,"manual_review":8.750536386806035e-07,"reject":2.5204989735856766e-07},"predicted":"approve","confidence":0.9999972961203765,"entropy_confidence":0.9999712864242893,"candidate_mass":0.9999983841580277,"accepted":true,"accepted_predicted":"approve","input_tokens":427,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.703883279022566e-06,"brier_hard":1.0626435874277302e-11,"soft_accuracy":0.3666662482534121,"brier_soft":0.5416621124112468,"tvd":0.633330662787009,"kl":5.465339427546418,"case_id":"typed:253/r0","base_id":"typed:253","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.3165}
{"id":"duplicate","type":"binary","probabilities":{"false":2.8115611575330923e-05,"true":0.9999718843884235},"predicted":"true","confidence":0.9999718843884235,"entropy_confidence":0.9995343789366546,"candidate_mass":0.9999982608706747,"accepted":true,"accepted_predicted":"true","input_tokens":375,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.8116006827667033e-05,"brier_hard":1.5809752285727686e-09,"soft_accuracy":0.5599966261266104,"brier_soft":0.38715051810460155,"tvd":0.43997188438842405,"kl":3.7278671512916963,"case_id":"typed:253/r0","base_id":"typed:253","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.3165}
{"id":"matches_order","type":"binary","probabilities":{"false":1.6712645017542434e-08,"true":0.9999999832873546},"predicted":"true","confidence":0.9999999832873546,"entropy_confidence":0.999999544126178,"candidate_mass":0.9999999216055637,"accepted":true,"accepted_predicted":"true","input_tokens":395,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6712645518381002e-08,"brier_hard":5.586250190373955e-16,"soft_accuracy":0.9666669844015199,"brier_soft":0.0022221755496701417,"tvd":0.03333298328735475,"kl":0.27423711385435945,"case_id":"typed:253/r0","base_id":"typed:253","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.3165}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005120500103048799,"1":0.10870877757703293,"2":0.8900998737778807,"3":0.0006792986347806633},"predicted":"2","confidence":0.8900998737778807,"entropy_confidence":0.744862566506709,"candidate_mass":0.9999885132469217,"accepted":true,"accepted_predicted":"2","input_tokens":422,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.219082737636434,"brier_hard":1.5866785521094744,"score_mae":0.8542794210371363,"within_one":true,"case_id":"typed:253/r0","base_id":"typed:253","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.3165}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9681618578596841,"1":0.002386609388637582,"2":0.018950861663234536,"3":0.010500671088442628},"predicted":"0","confidence":0.9681618578596841,"entropy_confidence":0.8782815378560969,"candidate_mass":0.9999362565344754,"accepted":true,"accepted_predicted":"0","input_tokens":392,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":4.556316110676008,"brier_hard":1.9168111359928905,"score_mae":2.5348766540195653,"within_one":false,"case_id":"typed:254/r0","base_id":"typed:254","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.56370000000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999280920839909,"hold":4.4244092771206175e-05,"manual_review":1.9211802841154966e-05,"reject":8.452020396800996e-06},"predicted":"approve","confidence":0.9999280920839909,"entropy_confidence":0.9994064347936674,"candidate_mass":0.9999942310212259,"accepted":true,"accepted_predicted":"approve","input_tokens":396,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.859985736490277,"brier_hard":1.9998177681311178,"soft_accuracy":0.0666865122759311,"brier_soft":1.2959271227409377,"tvd":0.9332610920839909,"kl":8.273686368184494,"case_id":"typed:254/r0","base_id":"typed:254","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.56370000000004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004958366528341553,"true":0.995041633471657},"predicted":"true","confidence":0.995041633471657,"entropy_confidence":0.954903498883469,"candidate_mass":0.9999613036902715,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.3066789214493895,"brier_hard":1.9802157046838895,"soft_accuracy":0.024760031867207863,"brier_soft":1.9014123740061568,"tvd":0.9750416334716576,"kl":5.102605643740946,"case_id":"typed:254/r0","base_id":"typed:254","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.56370000000004}
{"id":"matches_order","type":"binary","probabilities":{"false":2.6681775594818655e-08,"true":0.9999999733182235},"predicted":"true","confidence":0.9999999733182235,"entropy_confidence":0.999999290204331,"candidate_mass":0.9999998892860987,"accepted":true,"accepted_predicted":"true","input_tokens":364,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.43928506658351,"brier_hard":1.9999998932728973,"soft_accuracy":0.06666702312418768,"brier_soft":1.7422208781660733,"tvd":0.933332973318224,"kl":8.415777362753726,"case_id":"typed:254/r0","base_id":"typed:254","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.56370000000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.009577229154038443,"1":0.9749438500835527,"2":0.0017659549663408124,"3":0.01371296579606861},"predicted":"1","confidence":0.9749438500835527,"entropy_confidence":0.899535661134888,"candidate_mass":0.9999934268969841,"accepted":true,"accepted_predicted":"1","input_tokens":391,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.339063677498775,"brier_hard":1.9472664882291955,"score_mae":0.6503853425955599,"within_one":true,"case_id":"typed:254/r0","base_id":"typed:254","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.56370000000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999975329526356,"1":1.59890393455347e-06,"2":6.299289612699913e-07,"3":2.382144694239392e-07},"predicted":"0","confidence":0.9999975329526356,"entropy_confidence":0.9999737191259721,"candidate_mass":0.9999996486925276,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":2.467050407562534e-06,"brier_hard":9.096373119794325e-12,"score_mae":0.04332942659473464,"within_one":true,"case_id":"typed:255/r0","base_id":"typed:255","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":324.1983}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999993994115135,"hold":3.371801588362445e-07,"manual_review":2.197493003748629e-07,"reject":4.36590262976734e-08},"predicted":"approve","confidence":0.9999993994115135,"entropy_confidence":0.999992978203605,"candidate_mass":0.9999992635634183,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":6.00588666852358e-07,"brier_hard":5.245928552205222e-13,"soft_accuracy":0.9499994419944231,"brier_soft":0.0034887926127051836,"tvd":0.04999939941151402,"kl":0.4117889594898696,"case_id":"typed:255/r0","base_id":"typed:255","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":324.1983}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003992539856755072,"true":0.9960074601432466},"predicted":"true","confidence":0.9960074601432466,"entropy_confidence":0.9624370325153405,"candidate_mass":0.9999744810832778,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.523327695018217,"brier_hard":1.9840617213219987,"soft_accuracy":0.02052645253317003,"brier_soft":1.9182154737471686,"tvd":0.9793404601432458,"kl":5.346569634290323,"case_id":"typed:255/r0","base_id":"typed:255","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":324.1983}
{"id":"matches_order","type":"binary","probabilities":{"false":2.3984685574815846e-09,"true":0.999999997601531},"predicted":"true","confidence":0.999999997601531,"entropy_confidence":0.9999999278590196,"candidate_mass":0.9999999661875988,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.398468981857654e-09,"brier_hard":1.1505304864363525e-17,"soft_accuracy":0.9866669976654885,"brier_soft":0.00035553765008487727,"tvd":0.013332997601531272,"kl":0.10955775267738639,"case_id":"typed:255/r0","base_id":"typed:255","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":324.1983}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000645920999518623,"1":0.22325276747515646,"2":0.7753843647930345,"3":0.0007169467322904288},"predicted":"2","confidence":0.7753843647930345,"entropy_confidence":0.6090680509408757,"candidate_mass":0.9999898501887958,"accepted":false,"accepted_predicted":null,"input_tokens":406,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.25439641800195284,"brier_hard":0.10029491299129943,"score_mae":0.040494662741903165,"within_one":true,"case_id":"typed:255/r0","base_id":"typed:255","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":324.1983}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9998759029962797,"1":5.491809226286919e-05,"2":4.700171647166483e-05,"3":2.2177194987329744e-05},"predicted":"0","confidence":0.9998759029962797,"entropy_confidence":0.9990125721205488,"candidate_mass":0.9999944216936832,"accepted":true,"accepted_predicted":"0","input_tokens":414,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.96532643624736,"brier_hard":1.9996578236766687,"score_mae":1.599784546889832,"within_one":false,"case_id":"typed:256/r0","base_id":"typed:256","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.4692}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999286205556933,"hold":4.553843694269868e-05,"manual_review":1.916882322733159e-05,"reject":6.6721841368912365e-06},"predicted":"approve","confidence":0.9999286205556933,"entropy_confidence":0.9994125671820818,"candidate_mass":0.999991409219432,"accepted":true,"accepted_predicted":"approve","input_tokens":418,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.862225388949948,"brier_hard":1.9998189110456681,"soft_accuracy":0.08335119685596402,"brier_soft":1.2887323705361948,"tvd":0.9165956205556932,"kl":8.071090594542028,"case_id":"typed:256/r0","base_id":"typed:256","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.4692}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004600496991343559,"true":0.9953995030086576},"predicted":"true","confidence":0.9953995030086576,"entropy_confidence":0.9576599426649213,"candidate_mass":0.9999547985510034,"accepted":true,"accepted_predicted":"true","input_tokens":366,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.381590939726735,"brier_hard":1.981640341179763,"soft_accuracy":0.01781082013857241,"brier_soft":1.928909232663305,"tvd":0.982066503008657,"kl":5.239090547465991,"case_id":"typed:256/r0","base_id":"typed:256","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.4692}
{"id":"matches_order","type":"binary","probabilities":{"false":3.4851115624239482e-09,"true":0.9999999965148891},"predicted":"true","confidence":0.9999999965148891,"entropy_confidence":0.9999998970538375,"candidate_mass":0.9999999505869214,"accepted":true,"accepted_predicted":"true","input_tokens":386,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.4851108726364444e-09,"brier_hard":2.4292000354779703e-17,"soft_accuracy":0.6833329987221286,"brier_soft":0.2005559733635212,"tvd":0.31666699651488883,"kl":2.656416111941162,"case_id":"typed:256/r0","base_id":"typed:256","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.4692}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006067058198758502,"1":0.7273846209461948,"2":0.2703364608277858,"3":0.001672212406143274},"predicted":"1","confidence":0.7273846209461948,"entropy_confidence":0.5669491634162667,"candidate_mass":0.9999874012305459,"accepted":false,"accepted_predicted":null,"input_tokens":413,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3080879445717208,"brier_hard":1.061500431572644,"score_mae":0.42692582017980363,"within_one":true,"case_id":"typed:256/r0","base_id":"typed:256","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.4692}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9996619926129234,"1":0.00015202432184338336,"2":0.0001248395766449484,"3":6.114348858952648e-05},"predicted":"0","confidence":0.9996619926129234,"entropy_confidence":0.9975547636388543,"candidate_mass":0.9999914979804394,"accepted":true,"accepted_predicted":"0","input_tokens":403,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0003380645244490244,"brier_hard":1.5668383424405598e-07,"score_mae":0.21608186605909813,"within_one":true,"case_id":"typed:257/r0","base_id":"typed:257","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.770201}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999952723631742,"hold":2.6202645382078687e-06,"manual_review":1.7682611855220702e-06,"reject":3.3911110332709834e-07},"predicted":"approve","confidence":0.9999952723631742,"entropy_confidence":0.9999517583501583,"candidate_mass":0.9999975907049607,"accepted":true,"accepted_predicted":"approve","input_tokens":407,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":4.727648001106666e-06,"brier_hard":3.245808016744022e-11,"soft_accuracy":0.8233294823627891,"brier_soft":0.04528663781122831,"tvd":0.17666227236317358,"kl":1.4671156527460705,"case_id":"typed:257/r0","base_id":"typed:257","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.770201}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0039593806638792955,"true":0.9960406193361221},"predicted":"true","confidence":0.9960406193361221,"entropy_confidence":0.9627012694222862,"candidate_mass":0.9999765893517341,"accepted":true,"accepted_predicted":"true","input_tokens":355,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.531667663957471,"brier_hard":1.9841938307349687,"soft_accuracy":0.023801005437324153,"brier_soft":1.905310581188079,"tvd":0.9760406193361214,"kl":5.323074542193853,"case_id":"typed:257/r0","base_id":"typed:257","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.770201}
{"id":"matches_order","type":"binary","probabilities":{"false":9.532234346189073e-09,"true":0.9999999904677654},"predicted":"true","confidence":0.9999999904677654,"entropy_confidence":0.9999997322659081,"candidate_mass":0.9999999239419843,"accepted":true,"accepted_predicted":"true","input_tokens":375,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.53223460267303e-09,"brier_hard":1.8172698728493152e-16,"soft_accuracy":0.9633329911668023,"brier_soft":0.0026889363799264196,"tvd":0.03666699046776555,"kl":0.3017291643501228,"case_id":"typed:257/r0","base_id":"typed:257","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.770201}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00032678871616550825,"1":0.9365012204135027,"2":0.06288627867561553,"3":0.0002857121947169544},"predicted":"1","confidence":0.9365012204135027,"entropy_confidence":0.8266144570490472,"candidate_mass":0.9999908904984555,"accepted":true,"accepted_predicted":"1","input_tokens":402,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.7664272841345934,"brier_hard":1.7552168509527395,"score_mae":0.7002020856511155,"within_one":true,"case_id":"typed:257/r0","base_id":"typed:257","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.770201}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.027197224882254612,"1":0.001992460298456625,"2":0.15009403394677007,"3":0.8207162808725181},"predicted":"3","confidence":0.8207162808725181,"entropy_confidence":0.5980408161356929,"candidate_mass":0.9999659444469853,"accepted":true,"accepted_predicted":"3","input_tokens":409,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.19757780675535128,"brier_hard":0.05541452990993283,"score_mae":0.004329370809551314,"within_one":true,"case_id":"typed:258/r0","base_id":"typed:258","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":370.9746}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998856838328645,"hold":6.478041163255367e-05,"manual_review":2.9298092014115328e-05,"reject":2.02376634903418e-05},"predicted":"approve","confidence":0.9998856838328645,"entropy_confidence":0.9990884866013725,"candidate_mass":0.9999926237503745,"accepted":true,"accepted_predicted":"approve","input_tokens":413,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.437988163033758,"brier_hard":1.9997127900143297,"soft_accuracy":3.445679108778997e-05,"brier_soft":1.5969247503941824,"tvd":0.9998856838328637,"kl":8.910134634880977,"case_id":"typed:258/r0","base_id":"typed:258","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":370.9746}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0032771307405828043,"true":0.9967228692594182},"predicted":"true","confidence":0.9967228692594182,"entropy_confidence":0.9682325447286732,"candidate_mass":0.9999764712887275,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.720787013636887,"brier_hard":1.9869129562094525,"soft_accuracy":0.02645720015744279,"brier_soft":1.8949756751537326,"tvd":0.9733898692594176,"kl":5.476639069492072,"case_id":"typed:258/r0","base_id":"typed:258","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":370.9746}
{"id":"matches_order","type":"binary","probabilities":{"false":2.3248118120919183e-08,"true":0.9999999767518821},"predicted":"true","confidence":0.9999999767518821,"entropy_confidence":0.9999993769270508,"candidate_mass":0.9999998974873757,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.577041649224682,"brier_hard":1.9999999070075292,"soft_accuracy":0.016667022473165358,"brier_soft":1.9338874863354347,"tvd":0.983332976751882,"kl":8.988591521902045,"case_id":"typed:258/r0","base_id":"typed:258","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":370.9746}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.002196674969002453,"1":0.945769402885321,"2":0.051534262982385816,"3":0.0004996591632891794},"predicted":"1","confidence":0.945769402885321,"entropy_confidence":0.8392825366024915,"candidate_mass":0.9999900543784939,"accepted":true,"accepted_predicted":"1","input_tokens":408,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.055756499808104806,"brier_hard":0.005601812964751236,"score_mae":0.013669906339960125,"within_one":true,"case_id":"typed:258/r0","base_id":"typed:258","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":370.9746}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9030781159393836,"1":0.0002952427688345762,"2":0.0028792654436254443,"3":0.0937473758481574},"predicted":"0","confidence":0.9030781159393836,"entropy_confidence":0.7596299547167087,"candidate_mass":0.999972479741283,"accepted":true,"accepted_predicted":"0","input_tokens":411,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.3671516054763586,"brier_hard":1.636852279608515,"score_mae":2.3293710987994425,"within_one":false,"case_id":"typed:259/r0","base_id":"typed:259","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":351.9186}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9994292360723285,"hold":0.0003033929697240952,"manual_review":0.0001689467355444711,"reject":9.842422240196118e-05},"predicted":"approve","confidence":0.9994292360723285,"entropy_confidence":0.9961020016677358,"candidate_mass":0.9999825673173725,"accepted":true,"accepted_predicted":"approve","input_tokens":415,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":8.100481662959393,"brier_hard":1.998252142254291,"soft_accuracy":0.05018882597056187,"brier_soft":1.2857037540306153,"tvd":0.949429236072329,"kl":6.944432498471886,"case_id":"typed:259/r0","base_id":"typed:259","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":351.9186}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0004726777790402596,"true":0.9995273222209582},"predicted":"true","confidence":0.9995273222209582,"entropy_confidence":0.9940966282588244,"candidate_mass":0.9999914729090245,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.6570966298584295,"brier_hard":1.9981097357324016,"soft_accuracy":0.050425410001136164,"brier_soft":1.8032042712882095,"tvd":0.9495273222209589,"kl":7.0757501944959555,"case_id":"typed:259/r0","base_id":"typed:259","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":351.9186}
{"id":"matches_order","type":"binary","probabilities":{"false":4.102165249812543e-09,"true":0.9999999958978343},"predicted":"true","confidence":0.9999999958978343,"entropy_confidence":0.9999998797915343,"candidate_mass":0.9999999459849424,"accepted":true,"accepted_predicted":"true","input_tokens":383,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":19.31175089292663,"brier_hard":1.999999983591338,"soft_accuracy":0.10000000328173216,"brier_soft":1.6199999852322042,"tvd":0.8999999958978345,"kl":8.059047825889378,"case_id":"typed:259/r0","base_id":"typed:259","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":351.9186}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00023110631452950305,"1":0.00012214516137467143,"2":0.000595894332605949,"3":0.99905085419149},"predicted":"3","confidence":0.99905085419149,"entropy_confidence":0.9939341814766867,"candidate_mass":0.9999956376868163,"accepted":true,"accepted_predicted":"3","input_tokens":410,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0009495965326174285,"brier_hard":1.3242973905066854e-06,"score_mae":0.6451334964010562,"within_one":true,"case_id":"typed:259/r0","base_id":"typed:259","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":351.9186}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.02783357035791739,"1":0.016393312950494614,"2":0.7429923589524919,"3":0.21278075773909777},"predicted":"2","confidence":0.7429923589524919,"entropy_confidence":0.48274019274929625,"candidate_mass":0.9999669250278618,"accepted":false,"accepted_predicted":null,"input_tokens":393,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.2970695183650492,"brier_hard":0.11237202676919136,"score_mae":0.6673873040727718,"within_one":true,"case_id":"typed:260/r0","base_id":"typed:260","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":364.6604}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998569357294727,"hold":8.005585245407637e-05,"manual_review":4.38993378576934e-05,"reject":1.9109080215507565e-05},"predicted":"approve","confidence":0.9998569357294727,"entropy_confidence":0.9988845807324462,"candidate_mass":0.9999922242875928,"accepted":true,"accepted_predicted":"approve","input_tokens":397,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.033611320965157,"brier_hard":1.9996261019518637,"soft_accuracy":0.3333163860231137,"brier_soft":0.8269702063593518,"tvd":0.6665430448096882,"kl":5.635409281150591,"case_id":"typed:260/r0","base_id":"typed:260","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":364.6604}
{"id":"duplicate","type":"binary","probabilities":{"false":0.017800120576470888,"true":0.9821998794235303},"predicted":"true","confidence":0.9821998794235303,"entropy_confidence":0.8710960253459663,"candidate_mass":0.9999586647369343,"accepted":true,"accepted_predicted":"true","input_tokens":345,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.028550047748002,"brier_hard":1.9294332062791926,"soft_accuracy":0.037088115753412074,"brier_soft":1.85165721592531,"tvd":0.9621998794235297,"kl":3.850299142476082,"case_id":"typed:260/r0","base_id":"typed:260","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":364.6604}
{"id":"matches_order","type":"binary","probabilities":{"false":3.0361602030981864e-07,"true":0.9999996963839805},"predicted":"true","confidence":0.9999996963839805,"entropy_confidence":0.9999929883087613,"candidate_mass":0.9999995610341235,"accepted":true,"accepted_predicted":"true","input_tokens":365,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.007502024998207,"brier_hard":1.9999987855361046,"soft_accuracy":0.3266671052533515,"brier_soft":0.9067538400394421,"tvd":0.6733326963839801,"kl":5.836146607905055,"case_id":"typed:260/r0","base_id":"typed:260","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":364.6604}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003091601817387497,"1":0.9050655095064482,"2":0.09434749859874474,"3":0.0002778317130691821},"predicted":"1","confidence":0.9050655095064482,"entropy_confidence":0.7707667424681939,"candidate_mass":0.999989994336537,"accepted":true,"accepted_predicted":"1","input_tokens":392,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.09974795170699478,"brier_hard":0.017914180747589187,"score_mae":0.0612610018431452,"within_one":true,"case_id":"typed:260/r0","base_id":"typed:260","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":364.6604}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0012483964694837762,"1":0.0006568982614564118,"2":0.300004451183629,"3":0.6980902540854319},"predicted":"3","confidence":0.6980902540854319,"entropy_confidence":0.5489763257805538,"candidate_mass":0.9999779963172406,"accepted":false,"accepted_predicted":null,"input_tokens":398,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.3594068807314628,"brier_hard":0.18115415541726043,"score_mae":0.17826956288500995,"within_one":true,"case_id":"typed:261/r0","base_id":"typed:261","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.75899999999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.8629169230737654,"hold":0.12983990029670867,"manual_review":0.003536115342005744,"reject":0.0037070612875209482},"predicted":"approve","confidence":0.8629169230737654,"entropy_confidence":0.6876575566452137,"candidate_mass":0.9999827571899039,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":5.644726515393224,"brier_hard":1.7544380315672439,"soft_accuracy":0.11684575643116418,"brier_soft":1.0083744271669273,"tvd":0.7629169230737651,"kl":3.3577044517172356,"case_id":"typed:261/r0","base_id":"typed:261","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.75899999999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0009912140187862673,"true":0.9990087859812147},"predicted":"true","confidence":0.9990087859812147,"entropy_confidence":0.9886798446770637,"candidate_mass":0.9999838539605929,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.916580084502517,"brier_hard":1.996037108935319,"soft_accuracy":0.03758552432993263,"brier_soft":1.8522034260910263,"tvd":0.9623417859812142,"kl":6.505803177465644,"case_id":"typed:261/r0","base_id":"typed:261","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.75899999999996}
{"id":"matches_order","type":"binary","probabilities":{"false":7.236893072877356e-06,"true":0.999992763106927},"predicted":"true","confidence":0.999992763106927,"entropy_confidence":0.9998659807881903,"candidate_mass":0.999996365035065,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.83631857725129,"brier_hard":1.9999710525324534,"soft_accuracy":0.05000651320376558,"brier_soft":1.804972499911068,"tvd":0.9499927631069269,"kl":8.600037101545638,"case_id":"typed:261/r0","base_id":"typed:261","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.75899999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012809488198502143,"1":8.150542214069481e-05,"2":0.0002779207490996882,"3":0.9995124789467744},"predicted":"3","confidence":0.9995124789467744,"entropy_confidence":0.9966251702679538,"candidate_mass":0.9999966245491533,"accepted":true,"accepted_predicted":"3","input_tokens":397,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.188174560096268,"brier_hard":1.9984694543635022,"score_mae":0.8491747837606636,"within_one":true,"case_id":"typed:261/r0","base_id":"typed:261","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":337.75899999999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999682567230547,"1":1.6307319695518964e-05,"2":1.1195000276220436e-05,"3":4.240956974272567e-06},"predicted":"0","confidence":0.9999682567230547,"entropy_confidence":0.9997175201932138,"candidate_mass":0.9999972051016015,"accepted":true,"accepted_predicted":"0","input_tokens":402,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.1743780773806276e-05,"brier_hard":1.4168780541218929e-09,"score_mae":0.08328157980882922,"within_one":true,"case_id":"typed:262/r0","base_id":"typed:262","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":356.0811}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999983548501273,"hold":8.675799818468862e-07,"manual_review":6.462696887157571e-07,"reject":1.3130020236784164e-07},"predicted":"approve","confidence":0.9999983548501273,"entropy_confidence":0.9999819333696043,"candidate_mass":0.9999989756955602,"accepted":true,"accepted_predicted":"approve","input_tokens":406,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.6451512259249336e-06,"brier_hard":3.894117382123725e-12,"soft_accuracy":0.6666661069521522,"brier_soft":0.1516203935778444,"tvd":0.3333313548501272,"kl":2.7998016100862406,"case_id":"typed:262/r0","base_id":"typed:262","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":356.0811}
{"id":"duplicate","type":"binary","probabilities":{"false":2.0092591629953423e-05,"true":0.9999799074083712},"predicted":"true","confidence":0.9999799074083712,"entropy_confidence":0.9996575085675965,"candidate_mass":0.9999985478377545,"accepted":true,"accepted_predicted":"true","input_tokens":354,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.815159386464142,"brier_hard":1.999919630440907,"soft_accuracy":0.3600056259256568,"brier_soft":0.8191485637728533,"tvd":0.6399799074083706,"kl":5.5268306223589,"case_id":"typed:262/r0","base_id":"typed:262","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":356.0811}
{"id":"matches_order","type":"binary","probabilities":{"false":2.8384425826170984e-09,"true":0.9999999971615576},"predicted":"true","confidence":0.9999999971615576,"entropy_confidence":0.9999999153152125,"candidate_mass":0.9999999740909672,"accepted":true,"accepted_predicted":"true","input_tokens":374,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.838442374062877e-09,"brier_hard":1.611351138282109e-17,"soft_accuracy":0.963332997369712,"brier_soft":0.0026889373616913355,"tvd":0.036666997161557524,"kl":0.30172915790177174,"case_id":"typed:262/r0","base_id":"typed:262","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":356.0811}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005827054321223105,"1":0.5055200781601344,"2":0.4933512417591565,"3":0.0005459746485866433},"predicted":"1","confidence":0.5055200781601344,"entropy_confidence":0.4937146540490487,"candidate_mass":0.9999861079231925,"accepted":false,"accepted_predicted":null,"input_tokens":401,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.6821675219036115,"brier_hard":0.48790647848199875,"score_mae":0.5471934856242074,"within_one":true,"case_id":"typed:262/r0","base_id":"typed:262","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":356.0811}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9987187871703674,"1":0.00043421765680104035,"2":0.00038831218792102854,"3":0.0004586829849114243},"predicted":"0","confidence":0.9987187871703674,"entropy_confidence":0.9919081090870928,"candidate_mass":0.999987436909837,"accepted":true,"accepted_predicted":"0","input_tokens":377,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0012820342845039456,"brier_hard":2.1912277242281298e-06,"score_mae":0.6007461090126226,"within_one":true,"case_id":"typed:263/r0","base_id":"typed:263","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":343.9601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999397582251245,"hold":3.613389274488818e-05,"manual_review":1.842113274862869e-05,"reject":5.686749380808783e-06},"predicted":"approve","confidence":0.9999397582251245,"entropy_confidence":0.9994955357797694,"candidate_mass":0.9999925966530859,"accepted":true,"accepted_predicted":"approve","input_tokens":381,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":6.024358948411079e-05,"brier_hard":5.306406895313864e-09,"soft_accuracy":0.6732996171343016,"brier_soft":0.1648355452660525,"tvd":0.3266067582251251,"kl":2.6009183258497695,"case_id":"typed:263/r0","base_id":"typed:263","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":343.9601}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0038805301021534377,"true":0.9961194698978465},"predicted":"true","confidence":0.9961194698978465,"entropy_confidence":0.9633312306062182,"candidate_mass":0.9999430143103819,"accepted":true,"accepted_predicted":"true","input_tokens":329,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.55178351041285,"brier_hard":1.9845079966191337,"soft_accuracy":0.013802919500110368,"brier_soft":1.9448632178232197,"tvd":0.9861194698978466,"kl":5.440303021742817,"case_id":"typed:263/r0","base_id":"typed:263","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":343.9601}
{"id":"matches_order","type":"binary","probabilities":{"false":5.915286386298738e-09,"true":0.999999994084714},"predicted":"true","confidence":0.999999994084714,"entropy_confidence":0.9999998297840863,"candidate_mass":0.9999999223556463,"accepted":true,"accepted_predicted":"true","input_tokens":349,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.915286056266532e-09,"brier_hard":6.9981221952413e-17,"soft_accuracy":0.7166669974367056,"brier_soft":0.16055517107401684,"tvd":0.28333299408471374,"kl":2.3708400751255607,"case_id":"typed:263/r0","base_id":"typed:263","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":343.9601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002902481391434136,"1":0.9686243760579212,"2":0.030775566768879877,"3":0.0003098090340562091},"predicted":"1","confidence":0.9686243760579212,"entropy_confidence":0.8969366769285486,"candidate_mass":0.9999919524347358,"accepted":true,"accepted_predicted":"1","input_tokens":376,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.031878383050589355,"brier_hard":0.0019317455133203924,"score_mae":0.025562063302150495,"within_one":true,"case_id":"typed:263/r0","base_id":"typed:263","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":343.9601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999962902339692,"1":2.0803342950953185e-06,"2":1.225208854841841e-06,"3":4.04222881530747e-07},"predicted":"0","confidence":0.9999962902339692,"entropy_confidence":0.999961367906519,"candidate_mass":0.9999994158625151,"accepted":true,"accepted_predicted":"0","input_tokens":423,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.7097729120117645e-06,"brier_hard":1.9754687658657757e-11,"score_mae":0.04332725657935063,"within_one":true,"case_id":"typed:264/r0","base_id":"typed:264","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":362.88759999999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999998052315261,"hold":1.1093776283892959e-07,"manual_review":7.298212605936663e-08,"reject":1.0848584687463413e-08},"predicted":"approve","confidence":0.9999998052315261,"entropy_confidence":0.9999975693240211,"candidate_mass":0.9999993952293539,"accepted":true,"accepted_predicted":"approve","input_tokens":427,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.9476849287192023e-07,"brier_hard":5.568602816467585e-14,"soft_accuracy":0.6599998857378037,"brier_soft":0.16619983898750057,"tvd":0.33999980523152623,"kl":2.857275662004408,"case_id":"typed:264/r0","base_id":"typed:264","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":362.88759999999996}
{"id":"duplicate","type":"binary","probabilities":{"false":2.8957187269431733e-05,"true":0.999971042812729},"predicted":"true","confidence":0.999971042812729,"entropy_confidence":0.9995216737960878,"candidate_mass":0.9999978573064722,"accepted":true,"accepted_predicted":"true","input_tokens":375,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.449692119837984,"brier_hard":1.9998841729279564,"soft_accuracy":0.28334554813378776,"brier_soft":1.0271401688129198,"tvd":0.7166380428127297,"kl":6.243434783238857,"case_id":"typed:264/r0","base_id":"typed:264","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":362.88759999999996}
{"id":"matches_order","type":"binary","probabilities":{"false":1.757939552578693e-09,"true":0.9999999982420604},"predicted":"true","confidence":0.9999999982420604,"entropy_confidence":0.9999999463368516,"candidate_mass":0.9999999687723027,"accepted":true,"accepted_predicted":"true","input_tokens":395,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7579395732345245e-09,"brier_hard":6.180703008231907e-18,"soft_accuracy":0.9933329982655008,"brier_soft":8.889773111927336e-05,"tvd":0.006666998242060416,"kl":0.054760615006111776,"case_id":"typed:264/r0","base_id":"typed:264","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":362.88759999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007123327797388116,"1":0.7744532144436141,"2":0.22361160100687857,"3":0.001222851769766778},"predicted":"1","confidence":0.7744532144436141,"entropy_confidence":0.6059656858130387,"candidate_mass":0.9999879595255028,"accepted":false,"accepted_predicted":null,"input_tokens":422,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.497844656189823,"brier_hard":1.2025587302375889,"score_mae":0.6079880282333285,"within_one":true,"case_id":"typed:264/r0","base_id":"typed:264","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":362.88759999999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.970163547929824,"1":0.0024669747872718495,"2":0.009512110277890395,"3":0.01785736700501444},"predicted":"0","confidence":0.970163547929824,"entropy_confidence":0.8843224872551766,"candidate_mass":0.9999640988688429,"accepted":true,"accepted_predicted":"0","input_tokens":393,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.03029061558390117,"brier_hard":0.0013056656350274739,"score_mae":0.18160370364190404,"within_one":true,"case_id":"typed:265/r0","base_id":"typed:265","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":387.476901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9991514574056717,"hold":0.0005491602497688982,"manual_review":0.00017476967952026722,"reject":0.00012461266503925185},"predicted":"approve","confidence":0.9991514574056717,"entropy_confidence":0.9945154401932472,"candidate_mass":0.9999856174369269,"accepted":true,"accepted_predicted":"approve","input_tokens":397,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0008489028103823702,"brier_hard":1.0676742714834702e-06,"soft_accuracy":0.8593243379728148,"brier_soft":0.028544164317985194,"tvd":0.13915145740567164,"kl":0.6011761041318597,"case_id":"typed:265/r0","base_id":"typed:265","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":387.476901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0024495708005352673,"true":0.9975504291994659},"predicted":"true","confidence":0.9975504291994659,"entropy_confidence":0.9752245953504233,"candidate_mass":0.9999179979040775,"accepted":true,"accepted_predicted":"true","input_tokens":345,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.011842453226908,"brier_hard":1.9902137175920749,"soft_accuracy":0.02235158796851388,"brier_soft":1.9112096832561174,"tvd":0.9775504291994652,"kl":5.793615542400789,"case_id":"typed:265/r0","base_id":"typed:265","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":387.476901}
{"id":"matches_order","type":"binary","probabilities":{"false":1.529098054262215e-08,"true":0.9999999847090192},"predicted":"true","confidence":0.9999999847090192,"entropy_confidence":0.9999995809439671,"candidate_mass":0.9999999086793359,"accepted":true,"accepted_predicted":"true","input_tokens":365,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.5290980931781947e-08,"brier_hard":4.676281802357213e-16,"soft_accuracy":0.9166669872575057,"brier_soft":0.013888772681027294,"tvd":0.08333298470901931,"kl":0.6877651836570698,"case_id":"typed:265/r0","base_id":"typed:265","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":387.476901}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.510783734805683e-05,"1":2.412424724575897e-05,"2":8.226767899586878e-05,"3":0.9998585002364088},"predicted":"3","confidence":0.9998585002364088,"entropy_confidence":0.9988949956924742,"candidate_mass":0.9999985236008896,"accepted":true,"accepted_predicted":"3","input_tokens":392,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00014150977562726164,"brier_hard":2.8604693652174876e-08,"score_mae":0.5764311603144638,"within_one":true,"case_id":"typed:265/r0","base_id":"typed:265","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":387.476901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9995441381320901,"1":0.00013121966028587306,"2":0.00018092897610800154,"3":0.00014371323151684566},"predicted":"0","confidence":0.9995441381320901,"entropy_confidence":0.9967832577563926,"candidate_mass":0.9999890579335061,"accepted":true,"accepted_predicted":"0","input_tokens":411,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00045596580451957787,"brier_hard":2.7841742916835867e-07,"score_mae":1.0457427826929475,"within_one":false,"case_id":"typed:266/r0","base_id":"typed:266","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":358.951701}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9977259174136849,"hold":0.0016178364326894044,"manual_review":0.00032544639668742164,"reject":0.0003307997569372419},"predicted":"approve","confidence":0.9977259174136849,"entropy_confidence":0.9870638921922819,"candidate_mass":0.9999802152129671,"accepted":true,"accepted_predicted":"approve","input_tokens":415,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.002276672238920657,"brier_hard":8.0041901686254e-06,"soft_accuracy":0.45931884426497405,"brier_soft":0.4520884266216629,"tvd":0.5377263774132254,"kl":3.0271868674570492,"case_id":"typed:266/r0","base_id":"typed:266","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":358.951701}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0026125245762058684,"true":0.9973874754237959},"predicted":"true","confidence":0.9973874754237959,"entropy_confidence":0.9738195031695867,"candidate_mass":0.99998165681049,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.947438254575314,"brier_hard":1.9895635522645025,"soft_accuracy":0.02914018819045855,"brier_soft":1.8845964828139974,"tvd":0.970720475423795,"kl":5.665949479558282,"case_id":"typed:266/r0","base_id":"typed:266","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":358.951701}
{"id":"matches_order","type":"binary","probabilities":{"false":6.262898329585098e-09,"true":0.9999999937371022},"predicted":"true","confidence":0.9999999937371022,"entropy_confidence":0.9999998202972973,"candidate_mass":0.9999999246813414,"accepted":true,"accepted_predicted":"true","input_tokens":383,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.262897775571729e-09,"brier_hard":7.844778378832592e-17,"soft_accuracy":0.6233329984551562,"brier_soft":0.28375604834189194,"tvd":0.37666699373710194,"kl":3.174597733656726,"case_id":"typed:266/r0","base_id":"typed:266","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":358.951701}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004913848190125986,"1":0.00027480775742327,"2":0.001364623356139921,"3":0.9978691840674245},"predicted":"3","confidence":0.9978691840674245,"entropy_confidence":0.9876450667454834,"candidate_mass":0.9999968238294975,"accepted":true,"accepted_predicted":"3","input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.596876817988605,"brier_hard":1.9930158409743557,"score_mae":0.7499446066719768,"within_one":true,"case_id":"typed:266/r0","base_id":"typed:266","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":358.951701}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9998967432082376,"1":5.0684763560221816e-05,"2":3.512388215904579e-05,"3":1.7448146043028243e-05},"predicted":"0","confidence":0.9998967432082376,"entropy_confidence":0.9991661674836158,"candidate_mass":0.9999958047612765,"accepted":true,"accepted_predicted":"0","input_tokens":390,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00010326212311189828,"brier_hard":1.4769035200475213e-08,"score_mae":0.0598267230339926,"within_one":true,"case_id":"typed:267/r0","base_id":"typed:267","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.0521}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9982238406572967,"hold":0.0009652927985997882,"manual_review":0.0003884908049748497,"reject":0.0004223757391280834},"predicted":"approve","confidence":0.9982238406572967,"entropy_confidence":0.9893173523446787,"candidate_mass":0.9999667636631944,"accepted":true,"accepted_predicted":"approve","input_tokens":394,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.001777738583975062,"brier_hard":4.415858568254668e-06,"soft_accuracy":0.9383749822033431,"brier_soft":0.005057710544475331,"tvd":0.058223840657297035,"kl":0.16258289125370837,"case_id":"typed:267/r0","base_id":"typed:267","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.0521}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0008249728921524727,"true":0.9991750271078489},"predicted":"true","confidence":0.9991750271078489,"entropy_confidence":0.9903598077715137,"candidate_mass":0.9999505502506757,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.100160030166432,"brier_hard":1.9967014695919385,"soft_accuracy":0.0207919739764664,"brier_soft":1.9175674674233105,"tvd":0.9791750271078482,"kl":6.860134222550762,"case_id":"typed:267/r0","base_id":"typed:267","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.0521}
{"id":"matches_order","type":"binary","probabilities":{"false":3.2565412920592787e-09,"true":0.9999999967434583},"predicted":"true","confidence":0.9999999967434583,"entropy_confidence":0.9999999034868381,"candidate_mass":0.9999999566497406,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.2565417082698293e-09,"brier_hard":2.1210125050052262e-17,"soft_accuracy":0.976666996895428,"brier_soft":0.0010888574740604946,"tvd":0.02333299674345855,"kl":0.19184623176909185,"case_id":"typed:267/r0","base_id":"typed:267","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.0521}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.130428028687301e-05,"1":4.053349142500248e-05,"2":0.00011521375583021016,"3":0.9997829484724571},"predicted":"3","confidence":0.9997829484724571,"entropy_confidence":0.9983651157362818,"candidate_mass":0.9999984002025684,"accepted":true,"accepted_predicted":"3","input_tokens":389,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0002170750866348125,"brier_hard":6.57867538498113e-08,"score_mae":0.6062868064204565,"within_one":true,"case_id":"typed:267/r0","base_id":"typed:267","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.0521}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0407524440656124,"1":0.0036566694854745364,"2":0.18548622076040486,"3":0.7701046656885064},"predicted":"3","confidence":0.7701046656885064,"entropy_confidence":0.5205838302396223,"candidate_mass":0.9999663208122533,"accepted":false,"accepted_predicted":null,"input_tokens":391,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.6847746813087312,"brier_hard":1.258168025615421,"score_mae":0.8353601080718036,"within_one":true,"case_id":"typed:268/r0","base_id":"typed:268","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.019001}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9992295123944724,"hold":0.0006187047064844628,"manual_review":0.0001025014231492989,"reject":4.928147589518091e-05},"predicted":"approve","confidence":0.9992295123944724,"entropy_confidence":0.9951154524611566,"candidate_mass":0.9999885623373979,"accepted":true,"accepted_predicted":"approve","input_tokens":395,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.185633875098794,"brier_hard":1.998255011324516,"soft_accuracy":0.11679569952291548,"brier_soft":1.2215348684590204,"tvd":0.8825626290613551,"kl":6.713255095822373,"case_id":"typed:268/r0","base_id":"typed:268","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.019001}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0008207294264979271,"true":0.9991792705735023},"predicted":"true","confidence":0.9991792705735023,"entropy_confidence":0.990403285874504,"candidate_mass":0.9999749987826475,"accepted":true,"accepted_predicted":"true","input_tokens":343,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.105317068606553,"brier_hard":1.996718429487592,"soft_accuracy":0.03443246643129413,"brier_soft":1.8644278892559993,"tvd":0.9655122705735022,"kl":6.718863423366783,"case_id":"typed:268/r0","base_id":"typed:268","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.019001}
{"id":"matches_order","type":"binary","probabilities":{"false":3.713913696750976e-07,"true":0.9999996286086287},"predicted":"true","confidence":0.9999996286086287,"entropy_confidence":0.9999915310693213,"candidate_mass":0.9999994837827995,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.8060094255754,"brier_hard":1.9999985144347938,"soft_accuracy":0.11666728473313565,"brier_soft":1.560553065529262,"tvd":0.8833326286086296,"kl":7.885146570780843,"case_id":"typed:268/r0","base_id":"typed:268","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.019001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0019021318467628734,"1":0.9885551941726973,"2":0.009300947891407414,"3":0.00024172608913134884},"predicted":"1","confidence":0.9885551941726973,"entropy_confidence":0.9503604702636217,"candidate_mass":0.9999938300968856,"accepted":true,"accepted_predicted":"1","input_tokens":390,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.677638960209556,"brier_hard":1.9587296603117477,"score_mae":0.6921177317770937,"within_one":true,"case_id":"typed:268/r0","base_id":"typed:268","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":346.019001}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999646006996833,"1":1.709442591955376e-05,"2":1.2839479212405589e-05,"3":5.4653951862639765e-06},"predicted":"0","confidence":0.9999646006996833,"entropy_confidence":0.9996870249672022,"candidate_mass":0.9999974177420528,"accepted":true,"accepted_predicted":"0","input_tokens":393,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.539992688672572e-05,"brier_hard":1.740052631419389e-09,"score_mae":0.023273830430096842,"within_one":true,"case_id":"typed:269/r0","base_id":"typed:269","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":369.973901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998874859274172,"hold":6.142112331114533e-05,"manual_review":3.630526004684983e-05,"reject":1.4787689224604746e-05},"predicted":"approve","confidence":0.9998874859274172,"entropy_confidence":0.9991027966860314,"candidate_mass":0.9999912016446533,"accepted":true,"accepted_predicted":"approve","input_tokens":397,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00011252040276592879,"brier_hard":1.796871857765163e-08,"soft_accuracy":0.9432296128019647,"brier_soft":0.004626239775623294,"tvd":0.0565544859274173,"kl":0.29785748726590344,"case_id":"typed:269/r0","base_id":"typed:269","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":369.973901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.021251537547039,"true":0.9787484624529595},"predicted":"true","confidence":0.9787484624529595,"entropy_confidence":0.8515889418363332,"candidate_mass":0.9999421245008007,"accepted":true,"accepted_predicted":"true","input_tokens":345,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.851326031074443,"brier_hard":1.9158971055080676,"soft_accuracy":0.03401784404680963,"brier_soft":1.864054030286526,"tvd":0.9654154624529603,"kl":3.729453581761086,"case_id":"typed:269/r0","base_id":"typed:269","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":369.973901}
{"id":"matches_order","type":"binary","probabilities":{"false":4.4948090943002585e-09,"true":0.9999999955051901},"predicted":"true","confidence":0.9999999955051901,"entropy_confidence":0.9999998688783807,"candidate_mass":0.9999999628043319,"accepted":true,"accepted_predicted":"true","input_tokens":365,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.4948098677486e-09,"brier_hard":4.040662445060443e-17,"soft_accuracy":0.9899999955950862,"brier_soft":0.00019999982020766174,"tvd":0.009999995505190533,"kl":0.08215357567465728,"case_id":"typed:269/r0","base_id":"typed:269","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":369.973901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012573949811223785,"1":8.392087649784666e-05,"2":0.00031073018661674615,"3":0.9994796094387725},"predicted":"3","confidence":0.9994796094387725,"entropy_confidence":0.996431602862038,"candidate_mass":0.9999976455862877,"accepted":true,"accepted_predicted":"3","input_tokens":392,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0005205260113890156,"brier_hard":3.902127199872826e-07,"score_mae":0.5491442095660486,"within_one":true,"case_id":"typed:269/r0","base_id":"typed:269","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":369.973901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.00420616971708961,"1":0.00275285477152855,"2":0.9163345657813112,"3":0.07670640973006902},"predicted":"2","confidence":0.9163345657813112,"entropy_confidence":0.7718599972433045,"candidate_mass":0.9999866735952531,"accepted":true,"accepted_predicted":"2","input_tokens":415,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.5677700052600727,"brier_hard":1.6921653603522455,"score_mae":0.8011257844756421,"within_one":true,"case_id":"typed:270/r0","base_id":"typed:270","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.0082}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9994970395868521,"hold":0.0003425989463986573,"manual_review":0.00011236756364907586,"reject":4.799390309977479e-05},"predicted":"approve","confidence":0.9994970395868521,"entropy_confidence":0.9965840393507597,"candidate_mass":0.9999884520915149,"accepted":true,"accepted_predicted":"approve","input_tokens":419,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.093735241786355,"brier_hard":1.9987697293195053,"soft_accuracy":0.013497902773399788,"brier_soft":1.458465678500557,"tvd":0.986164026253839,"kl":7.825874037003431,"case_id":"typed:270/r0","base_id":"typed:270","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.0082}
{"id":"duplicate","type":"binary","probabilities":{"false":0.09519574595570322,"true":0.9048042540442971},"predicted":"true","confidence":0.9048042540442971,"entropy_confidence":0.5464212232283381,"candidate_mass":0.999945633339705,"accepted":true,"accepted_predicted":"true","input_tokens":367,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":2.351820023522489,"brier_hard":1.6373414762733134,"soft_accuracy":0.10329183103658916,"brier_soft":1.6013493061115414,"tvd":0.8948042540442969,"kl":2.2733006554579247,"case_id":"typed:270/r0","base_id":"typed:270","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.0082}
{"id":"matches_order","type":"binary","probabilities":{"false":1.6680267673786852e-05,"true":0.9999833197323262},"predicted":"true","confidence":0.9999833197323262,"entropy_confidence":0.9997111948256568,"candidate_mass":0.9999951200581325,"accepted":true,"accepted_predicted":"true","input_tokens":387,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.001284113573295,"brier_hard":1.9999332794857676,"soft_accuracy":0.02334890186630252,"brier_soft":1.9076916940665098,"tvd":0.9766503197323262,"kl":8.907753120976595,"case_id":"typed:270/r0","base_id":"typed:270","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.0082}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.266551536571002e-05,"1":0.9794433727666322,"2":0.020373306884271244,"3":0.00010065483372938244},"predicted":"1","confidence":0.9794433727666322,"entropy_confidence":0.92687593752753,"candidate_mass":0.9999933272417801,"accepted":true,"accepted_predicted":"1","input_tokens":414,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.02077085565604971,"brier_hard":0.0008376635215953173,"score_mae":0.27950804896363723,"within_one":true,"case_id":"typed:270/r0","base_id":"typed:270","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":360.0082}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9987830518576083,"1":0.0003208730768887783,"2":0.0006124136485533639,"3":0.00028366141694996977},"predicted":"0","confidence":0.9987830518576083,"entropy_confidence":0.9923212352854816,"candidate_mass":0.9999833831297625,"accepted":true,"accepted_predicted":"0","input_tokens":391,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.398102607437973,"brier_hard":1.996343315854699,"score_mae":1.5809363153751548,"within_one":false,"case_id":"typed:271/r0","base_id":"typed:271","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.785901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999853418163748,"hold":9.019766768392298e-06,"manual_review":4.594670433339037e-06,"reject":1.0437464236717261e-06},"predicted":"approve","confidence":0.9999853418163748,"entropy_confidence":0.9998627426890415,"candidate_mass":0.9999962962577079,"accepted":true,"accepted_predicted":"approve","input_tokens":395,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":11.61609208138676,"brier_hard":1.9999526444176317,"soft_accuracy":0.20000264070048185,"brier_soft":0.9477207603282047,"tvd":0.7999853418163747,"kl":7.046387646752338,"case_id":"typed:271/r0","base_id":"typed:271","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.785901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.03518490381128952,"true":0.9648150961887103},"predicted":"true","confidence":0.9648150961887103,"entropy_confidence":0.7802382188049694,"candidate_mass":0.9999004217802376,"accepted":true,"accepted_predicted":"true","input_tokens":343,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.347138157366153,"brier_hard":1.861736339667261,"soft_accuracy":0.04757966316625767,"brier_soft":1.8106363587353247,"tvd":0.9514820961887104,"kl":3.2321792162410437,"case_id":"typed:271/r0","base_id":"typed:271","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.785901}
{"id":"matches_order","type":"binary","probabilities":{"false":1.747909266037191e-09,"true":0.9999999982520915},"predicted":"true","confidence":0.9999999982520915,"entropy_confidence":0.9999999466286097,"candidate_mass":0.9999999673495138,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7479084861448475e-09,"brier_hard":6.1103708728956616e-18,"soft_accuracy":0.8499999987764641,"brier_soft":0.0449999989512547,"tvd":0.14999999825209115,"kl":1.243409967209041,"case_id":"typed:271/r0","base_id":"typed:271","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.785901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00019144409795068248,"1":0.9647939695075042,"2":0.03478742728628871,"3":0.00022715910825717628},"predicted":"1","confidence":0.9647939695075042,"entropy_confidence":0.8882219989864825,"candidate_mass":0.9999926650578811,"accepted":true,"accepted_predicted":"1","input_tokens":390,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.358499242357574,"brier_hard":1.8624628023747716,"score_mae":0.5016166985951469,"within_one":true,"case_id":"typed:271/r0","base_id":"typed:271","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.785901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999775392087475,"1":1.0265008023002075e-05,"2":8.511447452936128e-06,"3":3.6843357748222356e-06},"predicted":"0","confidence":0.9999775392087475,"entropy_confidence":0.9997938157435265,"candidate_mass":0.9999983649428055,"accepted":true,"accepted_predicted":"0","input_tokens":408,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":12.511420299636029,"brier_hard":1.9999477104418222,"score_mae":2.6532946590897466,"within_one":false,"case_id":"typed:272/r0","base_id":"typed:272","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.086601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999928577098856,"hold":3.801476076505898e-06,"manual_review":2.8199405487764137e-06,"reject":5.208734904530218e-07},"predicted":"approve","confidence":0.9999928577098856,"entropy_confidence":0.9999291950810081,"candidate_mass":0.9999982238590553,"accepted":true,"accepted_predicted":"approve","input_tokens":412,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":12.480121125469918,"brier_hard":1.999978112541305,"soft_accuracy":0.023336213257468878,"brier_soft":1.50706884675652,"tvd":0.9766598577098848,"kl":8.907752898424315,"case_id":"typed:272/r0","base_id":"typed:272","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.086601}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0011970606770728746,"true":0.9988029393229279},"predicted":"true","confidence":0.9988029393229279,"entropy_confidence":0.9866550248799374,"candidate_mass":0.9999706345896554,"accepted":true,"accepted_predicted":"true","input_tokens":360,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.727886162736262,"brier_hard":1.995214623200239,"soft_accuracy":0.014498139857058057,"brier_soft":1.9423020026182685,"tvd":0.9854699393229274,"kl":6.567390106283318,"case_id":"typed:272/r0","base_id":"typed:272","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.086601}
{"id":"matches_order","type":"binary","probabilities":{"false":1.0033623076745246e-08,"true":0.9999999899663763},"predicted":"true","confidence":0.9999999899663763,"entropy_confidence":0.9999997189253351,"candidate_mass":0.9999998942420503,"accepted":true,"accepted_predicted":"true","input_tokens":380,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.417324076195733,"brier_hard":1.9999999598655067,"soft_accuracy":0.06666700869579995,"brier_soft":1.7422209403191529,"tvd":0.9333329899663766,"kl":8.415777361643844,"case_id":"typed:272/r0","base_id":"typed:272","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.086601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012022941185087531,"1":0.8583834464978227,"2":0.14136318234468867,"3":0.00013314174563834393},"predicted":"1","confidence":0.8583834464978227,"entropy_confidence":0.704306341579416,"candidate_mass":0.9999894732197371,"accepted":true,"accepted_predicted":"1","input_tokens":407,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.1527043719873973,"brier_hard":0.04003882973028865,"score_mae":0.09182376357588495,"within_one":true,"case_id":"typed:272/r0","base_id":"typed:272","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.086601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.012576669029676186,"1":0.0026774668931218264,"2":0.7060371348530334,"3":0.2787087292241679},"predicted":"2","confidence":0.7060371348530334,"entropy_confidence":0.5147276536890986,"candidate_mass":0.9999730363198425,"accepted":false,"accepted_predicted":null,"input_tokens":410,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.2775880237641815,"brier_hard":1.0189148745217411,"score_mae":0.5791220757283075,"within_one":true,"case_id":"typed:273/r0","base_id":"typed:273","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.582101}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997805326045504,"hold":0.0001571395540854656,"manual_review":4.302192324562636e-05,"reject":1.9305918119246867e-05},"predicted":"approve","confidence":0.9997805326045504,"entropy_confidence":0.9983857453308546,"candidate_mass":0.9999921755216027,"accepted":true,"accepted_predicted":"approve","input_tokens":414,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.053800729274911,"brier_hard":1.9994750964449908,"soft_accuracy":6.61663499086148e-05,"brier_soft":1.5644283442590265,"tvd":0.9997805326045499,"kl":8.71015961020648,"case_id":"typed:273/r0","base_id":"typed:273","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.582101}
{"id":"duplicate","type":"binary","probabilities":{"false":0.008697082893011238,"true":0.9913029171069875},"predicted":"true","confidence":0.9913029171069875,"entropy_confidence":0.9279737525396305,"candidate_mass":0.9999634236722311,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.744767609200977,"brier_hard":1.9653629469296483,"soft_accuracy":0.03490023267399534,"brier_soft":1.8610449051456803,"tvd":0.9646359171069881,"kl":4.49551363245894,"case_id":"typed:273/r0","base_id":"typed:273","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.582101}
{"id":"matches_order","type":"binary","probabilities":{"false":3.563229799880741e-08,"true":0.9999999643677018},"predicted":"true","confidence":0.9999999643677018,"entropy_confidence":0.9999990669706196,"candidate_mass":0.9999998631816127,"accepted":true,"accepted_predicted":"true","input_tokens":382,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.15001336323087,"brier_hard":1.99999985747081,"soft_accuracy":0.023333033969481175,"brier_soft":1.9077567185744435,"tvd":0.9766669643677018,"kl":8.907752732604072,"case_id":"typed:273/r0","base_id":"typed:273","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.582101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00047449003179698485,"1":0.9876399734973378,"2":0.011743343865808725,"3":0.000142192605055236},"predicted":"1","confidence":0.9876399734973378,"entropy_confidence":0.9499621150548556,"candidate_mass":0.9999928496363375,"accepted":true,"accepted_predicted":"1","input_tokens":409,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.444468678407745,"brier_hard":1.9520841810028824,"score_mae":0.8384467609558792,"within_one":true,"case_id":"typed:273/r0","base_id":"typed:273","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.582101}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0022965108543580555,"1":0.0006492907546685812,"2":0.009458127511353949,"3":0.9875960708796181},"predicted":"3","confidence":0.9875960708796181,"entropy_confidence":0.9458052500565461,"candidate_mass":0.9999922592025215,"accepted":true,"accepted_predicted":"3","input_tokens":370,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.076364327731017,"brier_hard":1.9708481292247528,"score_mae":2.304020758416231,"within_one":false,"case_id":"typed:274/r0","base_id":"typed:274","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":314.0069}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9983393375684209,"hold":0.0010797344249472483,"manual_review":0.00033287337606759576,"reject":0.00024805463056379816},"predicted":"approve","confidence":0.9983393375684209,"entropy_confidence":0.9900743494118921,"candidate_mass":0.9999784337595065,"accepted":true,"accepted_predicted":"approve","input_tokens":374,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0016620428599303226,"brier_hard":4.095961924313291e-06,"soft_accuracy":0.6390951464245835,"brier_soft":0.21351485602759904,"tvd":0.3583393375684211,"kl":1.9486126158108261,"case_id":"typed:274/r0","base_id":"typed:274","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":314.0069}
{"id":"duplicate","type":"binary","probabilities":{"false":0.019119672246055475,"true":0.9808803277539432},"predicted":"true","confidence":0.9808803277539432,"entropy_confidence":0.8635311833928478,"candidate_mass":0.999934438849786,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.957037513448512,"brier_hard":1.9242524347493686,"soft_accuracy":0.044766943646484314,"brier_soft":1.8210461497265111,"tvd":0.9542133277539439,"kl":3.72907179719117,"case_id":"typed:274/r0","base_id":"typed:274","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":314.0069}
{"id":"matches_order","type":"binary","probabilities":{"false":4.2598149892509613e-07,"true":0.9999995740185011},"predicted":"true","confidence":0.9999995740185011,"entropy_confidence":0.9999903705175582,"candidate_mass":0.9999994318285326,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.259815895829011e-07,"brier_hard":3.6292047479131275e-13,"soft_accuracy":0.7099998210877704,"brier_soft":0.16819950586182425,"tvd":0.2899995740185012,"kl":2.4278308909678112,"case_id":"typed:274/r0","base_id":"typed:274","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":314.0069}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.404019725954263e-05,"1":3.415786489269576e-05,"2":0.00011866841925952263,"3":0.9997831335185868},"predicted":"3","confidence":0.9997831335185868,"entropy_confidence":0.9983703481749376,"candidate_mass":0.9999979822440288,"accepted":true,"accepted_predicted":"3","input_tokens":369,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.039177346848493,"brier_hard":1.9993289965798258,"score_mae":0.6996208952591725,"within_one":true,"case_id":"typed:274/r0","base_id":"typed:274","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":314.0069}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.999971492142838,"1":1.3463851101454561e-05,"2":1.0006763257843364e-05,"3":5.03724280208525e-06},"predicted":"0","confidence":0.999971492142838,"entropy_confidence":0.9997430854252077,"candidate_mass":0.99999727167816,"accepted":true,"accepted_predicted":"0","input_tokens":379,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":2.8508263518701887e-05,"brier_hard":1.1194823323978996e-09,"score_mae":0.0402844108939766,"within_one":true,"case_id":"typed:275/r0","base_id":"typed:275","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.313501}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999985187620118,"hold":8.501793244241405e-07,"manual_review":5.358080304415004e-07,"reject":9.525063300152536e-08},"predicted":"approve","confidence":0.9999985187620118,"entropy_confidence":0.999983667554267,"candidate_mass":0.9999989340890135,"accepted":true,"accepted_predicted":"approve","input_tokens":383,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.4812390851886545e-06,"brier_hard":3.2130337898033134e-12,"soft_accuracy":0.9568636083740175,"brier_soft":0.00254862596120173,"tvd":0.043133518762012005,"kl":0.3550984359528097,"case_id":"typed:275/r0","base_id":"typed:275","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.313501}
{"id":"duplicate","type":"binary","probabilities":{"false":0.005266544768117737,"true":0.9947334552318825},"predicted":"true","confidence":0.9947334552318825,"entropy_confidence":0.9525599291957444,"candidate_mass":0.9999223784663325,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.246380773179283,"brier_hard":1.9789892939151188,"soft_accuracy":0.015161213872755385,"brier_soft":1.9393999557058434,"tvd":0.9847334552318824,"kl":5.137968235711643,"case_id":"typed:275/r0","base_id":"typed:275","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.313501}
{"id":"matches_order","type":"binary","probabilities":{"false":9.907000782214023e-10,"true":0.9999999990092991},"predicted":"true","confidence":0.9999999990092991,"entropy_confidence":0.9999999689380566,"candidate_mass":0.999999983174238,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.907008551768548e-10,"brier_hard":1.9629748284636827e-18,"soft_accuracy":0.9833329990423232,"brier_soft":0.0005555777119519824,"tvd":0.01666699900929952,"kl":0.13698141654523685,"case_id":"typed:275/r0","base_id":"typed:275","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.313501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006573016151160021,"1":0.8489352464850314,"2":0.1491746868785993,"3":0.0012327650212544537},"predicted":"1","confidence":0.8489352464850314,"entropy_confidence":0.6855419556578061,"candidate_mass":0.9999899761397163,"accepted":true,"accepted_predicted":"1","input_tokens":378,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.16377236591517177,"brier_hard":0.045075598714877235,"score_mae":0.2823500846940066,"within_one":true,"case_id":"typed:275/r0","base_id":"typed:275","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":318.313501}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0017710978546986132,"1":0.0006848838234433048,"2":0.05740210091270897,"3":0.9401419174091475},"predicted":"3","confidence":0.9401419174091475,"entropy_confidence":0.8281184383794404,"candidate_mass":0.9999891927483225,"accepted":true,"accepted_predicted":"3","input_tokens":411,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.06172443914663359,"brier_hard":0.00688159709410867,"score_mae":0.39591483787630377,"within_one":true,"case_id":"typed:276/r0","base_id":"typed:276","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.875901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9979451552476106,"hold":0.0015776282700133806,"manual_review":0.00032838024215040484,"reject":0.00014883624022428096},"predicted":"approve","confidence":0.9979451552476106,"entropy_confidence":0.9883307478220853,"candidate_mass":0.999977588334158,"accepted":true,"accepted_predicted":"approve","input_tokens":415,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":8.02133834617155,"brier_hard":1.995240391294645,"soft_accuracy":0.033971551088430495,"brier_soft":1.4107324051580847,"tvd":0.9646121552476112,"kl":6.394403598438064,"case_id":"typed:276/r0","base_id":"typed:276","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.875901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.02911344576632977,"true":0.9708865542336685},"predicted":"true","confidence":0.9708865542336685,"entropy_confidence":0.810073839696656,"candidate_mass":0.9999510408486592,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.5365551576929235,"brier_hard":1.8852414023834556,"soft_accuracy":0.06050556879087157,"brier_soft":1.758013334112372,"tvd":0.9375535542336694,"kl":3.2735123862470097,"case_id":"typed:276/r0","base_id":"typed:276","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.875901}
{"id":"matches_order","type":"binary","probabilities":{"false":1.788915697747485e-07,"true":0.99999982110843},"predicted":"true","confidence":0.99999982110843,"entropy_confidence":0.999995732164826,"candidate_mass":0.9999996534762016,"accepted":true,"accepted_predicted":"true","input_tokens":383,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.536485970097946,"brier_hard":1.9999992844337844,"soft_accuracy":0.1000001431132558,"brier_soft":1.6199993559904124,"tvd":0.8999998211084301,"kl":8.059047843368319,"case_id":"typed:276/r0","base_id":"typed:276","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.875901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010551015737625038,"1":0.9364024201697714,"2":0.06105627105427647,"3":0.001486207202189387},"predicted":"1","confidence":0.9364024201697714,"entropy_confidence":0.8202754805634132,"candidate_mass":0.9999920646160353,"accepted":true,"accepted_predicted":"1","input_tokens":410,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":2.79595936372899,"brier_hard":1.7584681406774842,"score_mae":0.8503594161151076,"within_one":true,"case_id":"typed:276/r0","base_id":"typed:276","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.875901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.01529437290671846,"1":0.0020699709387740884,"2":0.08017118253724395,"3":0.9024644736172647},"predicted":"3","confidence":0.9024644736172647,"entropy_confidence":0.7319017773615268,"candidate_mass":0.9999710380971923,"accepted":true,"accepted_predicted":"3","input_tokens":384,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.10262595396290304,"brier_hard":0.01617880003847448,"score_mae":0.29313875686505586,"within_one":true,"case_id":"typed:277/r0","base_id":"typed:277","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":321.6164}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998761054608378,"hold":8.508486546052009e-05,"manual_review":2.871333580710194e-05,"reject":1.009633789357865e-05},"predicted":"approve","confidence":0.9998761054608378,"entropy_confidence":0.9990350387391139,"candidate_mass":0.9999931479803361,"accepted":true,"accepted_predicted":"approve","input_tokens":388,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":10.45814888085242,"brier_hard":1.9996948077657442,"soft_accuracy":0.05004327994423824,"brier_soft":1.2979988445491437,"tvd":0.9498761554607884,"kl":8.274895221798221,"case_id":"typed:277/r0","base_id":"typed:277","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":321.6164}
{"id":"duplicate","type":"binary","probabilities":{"false":0.029805673617222167,"true":0.9701943263827785},"predicted":"true","confidence":0.9701943263827785,"entropy_confidence":0.8065836173777059,"candidate_mass":0.9998932035314486,"accepted":true,"accepted_predicted":"true","input_tokens":336,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.5130565137529537,"brier_hard":1.8825540618906653,"soft_accuracy":0.06742121972784443,"brier_soft":1.7305229696694204,"tvd":0.9301943263827781,"kl":3.2058004611115543,"case_id":"typed:277/r0","base_id":"typed:277","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":321.6164}
{"id":"matches_order","type":"binary","probabilities":{"false":1.3407522823192082e-07,"true":0.9999998659247727},"predicted":"true","confidence":0.9999998659247727,"entropy_confidence":0.9999967455720229,"candidate_mass":0.9999997120389553,"accepted":true,"accepted_predicted":"true","input_tokens":356,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.824864789808345,"brier_hard":1.999999463699125,"soft_accuracy":0.1233331010034281,"brier_soft":1.5370895876207251,"tvd":0.8766668659247723,"kl":7.816281881055639,"case_id":"typed:277/r0","base_id":"typed:277","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":321.6164}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.005677651787835837,"1":0.9895279757299057,"2":0.0025211828840502057,"3":0.002273189598207603},"predicted":"1","confidence":0.9895279757299057,"entropy_confidence":0.9504451501625749,"candidate_mass":0.9999927120790777,"accepted":true,"accepted_predicted":"1","input_tokens":383,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.983027089159165,"brier_hard":1.9741670084679326,"score_mae":0.7486100897073711,"within_one":true,"case_id":"typed:277/r0","base_id":"typed:277","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":321.6164}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.005292660562933934,"1":0.001753820320698387,"2":0.5829088389607733,"3":0.41004468015559337},"predicted":"2","confidence":0.5829088389607733,"entropy_confidence":0.48132851775214336,"candidate_mass":0.9999633971214463,"accepted":false,"accepted_predicted":null,"input_tokens":390,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":0.8914891492319832,"brier_hard":0.6878610820928646,"score_mae":0.4022944612909747,"within_one":true,"case_id":"typed:278/r0","base_id":"typed:278","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.360601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.999440858178535,"hold":0.0004054418836364954,"manual_review":0.00011358277477837629,"reject":4.0117163049322323e-05},"predicted":"approve","confidence":0.999440858178535,"entropy_confidence":0.9962753170357832,"candidate_mass":0.9999861101284414,"accepted":true,"accepted_predicted":"approve","input_tokens":394,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.082978693646325,"brier_hard":1.9986550423406444,"soft_accuracy":0.016835022006144215,"brier_soft":1.5346569194339128,"tvd":0.9827738581785355,"kl":7.9094562955921255,"case_id":"typed:278/r0","base_id":"typed:278","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.360601}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0004525347318465026,"true":0.9995474652681519},"predicted":"true","confidence":0.9995474652681519,"entropy_confidence":0.9943197603268366,"candidate_mass":0.9999669315710583,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.700646042307959,"brier_hard":1.9981902706479777,"soft_accuracy":0.04041633195329872,"brier_soft":1.8414626762050732,"tvd":0.9595474652681526,"kl":7.224694158367731,"case_id":"typed:278/r0","base_id":"typed:278","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.360601}
{"id":"matches_order","type":"binary","probabilities":{"false":7.884153049209443e-07,"true":0.9999992115846953},"predicted":"true","confidence":0.9999992115846953,"entropy_confidence":0.9999828777992669,"candidate_mass":0.9999993066781828,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.05324084922189,"brier_hard":1.999996846340024,"soft_accuracy":0.016667762134269154,"brier_soft":1.9338844766800956,"tvd":0.9833322115846952,"kl":8.988591534655091,"case_id":"typed:278/r0","base_id":"typed:278","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.360601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.008092156673436272,"1":0.9855976277960011,"2":0.0018795879903507244,"3":0.004430627540212446},"predicted":"1","confidence":0.9855976277960011,"entropy_confidence":0.9357387011172025,"candidate_mass":0.9999939663573533,"accepted":true,"accepted_predicted":"1","input_tokens":389,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.014507093068601748,"brier_hard":0.00029607463614351814,"score_mae":0.5973513136026602,"within_one":true,"case_id":"typed:278/r0","base_id":"typed:278","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":342.360601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0016665888715107968,"1":0.0006450654046346473,"2":0.021555347212613606,"3":0.97613299851124},"predicted":"3","confidence":0.97613299851124,"entropy_confidence":0.9122188793386088,"candidate_mass":0.999990273318904,"accepted":true,"accepted_predicted":"3","input_tokens":419,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.024156432886382338,"brier_hard":0.0010374603813637002,"score_mae":0.1321547553635818,"within_one":true,"case_id":"typed:279/r0","base_id":"typed:279","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.917401}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9611946789781817,"hold":0.02442268285577025,"manual_review":0.0036398904848727125,"reject":0.010742747681174985},"predicted":"approve","confidence":0.9611946789781817,"entropy_confidence":0.857282254425675,"candidate_mass":0.9999731330688194,"accepted":true,"accepted_predicted":"approve","input_tokens":423,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":5.615801684359528,"brier_hard":1.9173405527945813,"soft_accuracy":0.029523023969749315,"brier_soft":1.332974285824828,"tvd":0.9411946789781819,"kl":3.9303489264107276,"case_id":"typed:279/r0","base_id":"typed:279","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.917401}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003253825756596954,"true":0.9967461742434044},"predicted":"true","confidence":0.9967461742434044,"entropy_confidence":0.9684248988053897,"candidate_mass":0.9999677128343496,"accepted":true,"accepted_predicted":"true","input_tokens":371,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.727923819069236,"brier_hard":1.9870058717377235,"soft_accuracy":0.03305859621120118,"brier_soft":1.8691963308285149,"tvd":0.9667461742434037,"kl":5.421441710246159,"case_id":"typed:279/r0","base_id":"typed:279","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.917401}
{"id":"matches_order","type":"binary","probabilities":{"false":8.515049171914018e-06,"true":0.9999914849508271},"predicted":"true","confidence":0.9999914849508271,"entropy_confidence":0.9998443087751303,"candidate_mass":0.9999967989615332,"accepted":true,"accepted_predicted":"true","input_tokens":391,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":11.673675469125953,"brier_hard":1.9999659399483227,"soft_accuracy":0.036674890606555904,"brier_soft":1.8559881266115545,"tvd":0.9633244849508276,"kl":8.751408501083898,"case_id":"typed:279/r0","base_id":"typed:279","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.917401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00013313565540902164,"1":7.418076695918812e-05,"2":0.00026760080071919386,"3":0.9995250827769133},"predicted":"3","confidence":0.9995250827769133,"entropy_confidence":0.9967037309096859,"candidate_mass":0.9999968528406632,"accepted":true,"accepted_predicted":"3","input_tokens":418,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00047503003198907923,"brier_hard":3.2038444625775964e-07,"score_mae":0.6825176306991376,"within_one":true,"case_id":"typed:279/r0","base_id":"typed:279","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.917401}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.13862171392776473,"1":0.010576477570388728,"2":0.11206781325995664,"3":0.7387339952418915},"predicted":"3","confidence":0.7387339952418915,"entropy_confidence":0.4294069392384746,"candidate_mass":0.999939925467703,"accepted":false,"accepted_predicted":null,"input_tokens":398,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.30281737514369667,"brier_hard":0.10014696146120045,"score_mae":0.15908591018402296,"within_one":true,"case_id":"typed:280/r0","base_id":"typed:280","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":335.96799999999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999766768031875,"hold":1.4204516981161561e-05,"manual_review":6.820336767915919e-06,"reject":2.298343063082363e-06},"predicted":"approve","confidence":0.9999766768031875,"entropy_confidence":0.9997887568563254,"candidate_mass":0.9999960096409025,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":11.161950546674369,"brier_hard":1.9999249453699517,"soft_accuracy":0.036676034896906184,"brier_soft":1.361556842388102,"tvd":0.9633096768031877,"kl":8.75140904406289,"case_id":"typed:280/r0","base_id":"typed:280","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":335.96799999999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.008155076036318631,"true":0.9918449239636817},"predicted":"true","confidence":0.9918449239636817,"entropy_confidence":0.9317021980759405,"candidate_mass":0.9999682969920319,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.809114719063157,"brier_hard":1.9675127063850426,"soft_accuracy":0.027828872994865896,"brier_soft":1.888965112467948,"tvd":0.9718449239636815,"kl":4.615057081613503,"case_id":"typed:280/r0","base_id":"typed:280","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":335.96799999999996}
{"id":"matches_order","type":"binary","probabilities":{"false":2.0914947454413806e-08,"true":0.9999999790850522},"predicted":"true","confidence":0.9999999790850522,"entropy_confidence":0.9999994362671769,"candidate_mass":0.9999998695062754,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.68280174439517,"brier_hard":1.9999999163402105,"soft_accuracy":0.05666701854457278,"brier_soft":1.7797542188589597,"tvd":0.9433329790850524,"kl":8.5257518077132,"case_id":"typed:280/r0","base_id":"typed:280","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":335.96799999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0013423638870882198,"1":0.8461643329448865,"2":0.15089711032957664,"3":0.001596192838449694},"predicted":"1","confidence":0.8461643329448865,"entropy_confidence":0.6783713101722775,"candidate_mass":0.9999905134830933,"accepted":true,"accepted_predicted":"1","input_tokens":397,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.891157062963746,"brier_hard":1.4369741453671108,"score_mae":0.8639198678806113,"within_one":true,"case_id":"typed:280/r0","base_id":"typed:280","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":335.96799999999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999984323977932,"1":9.394884598764896e-07,"2":4.778434098590602e-07,"3":1.5027033725300526e-07},"predicted":"0","confidence":0.9999984323977932,"entropy_confidence":0.9999827445357423,"candidate_mass":0.9999997312532566,"accepted":true,"accepted_predicted":"0","input_tokens":392,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.5676034355096859e-06,"brier_hard":3.5909307436720937e-12,"score_mae":0.07666465401370864,"within_one":true,"case_id":"typed:281/r0","base_id":"typed:281","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":308.385601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999951492603519,"hold":2.840465441085468e-06,"manual_review":1.6311156262573785e-06,"reject":3.7915857925424424e-07},"predicted":"approve","confidence":0.9999951492603519,"entropy_confidence":0.9999506089892907,"candidate_mass":0.9999984223329055,"accepted":true,"accepted_predicted":"approve","input_tokens":396,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":4.850751412941318e-06,"brier_hard":3.440221846976056e-11,"soft_accuracy":0.8333293258936378,"brier_soft":0.040420024545830484,"tvd":0.16666214926035267,"kl":1.3831289378463385,"case_id":"typed:281/r0","base_id":"typed:281","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":308.385601}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0007352949448932924,"true":0.9992647050551057},"predicted":"true","confidence":0.9992647050551057,"entropy_confidence":0.9912856123865199,"candidate_mass":0.9999722563581628,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.215238853675853,"brier_hard":1.9970599015377366,"soft_accuracy":0.05398986397430125,"brier_soft":1.7895735812569207,"tvd":0.9459317050551062,"kl":6.622253244588068,"case_id":"typed:281/r0","base_id":"typed:281","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":308.385601}
{"id":"matches_order","type":"binary","probabilities":{"false":2.3099659322239267e-09,"true":0.9999999976900344},"predicted":"true","confidence":0.9999999976900344,"entropy_confidence":0.9999999303956998,"candidate_mass":0.9999999759888997,"accepted":true,"accepted_predicted":"true","input_tokens":364,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.3099655548398593e-09,"brier_hard":1.0671883460255934e-17,"soft_accuracy":0.9633329978594334,"brier_soft":0.002688937439201956,"tvd":0.03666699769003426,"kl":0.30172915739267264,"case_id":"typed:281/r0","base_id":"typed:281","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":308.385601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00043111814911547174,"1":0.8513900663795877,"2":0.14763566369093004,"3":0.0005431517803673539},"predicted":"1","confidence":0.8513900663795877,"entropy_confidence":0.6921085052914229,"candidate_mass":0.9999889519867556,"accepted":true,"accepted_predicted":"1","input_tokens":391,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.9130077720798333,"brier_hard":1.4513904878181552,"score_mae":0.5917091508974501,"within_one":true,"case_id":"typed:281/r0","base_id":"typed:281","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":308.385601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9998992849567556,"1":4.958018874687718e-05,"2":3.4237272246587176e-05,"3":1.6897582249228853e-05},"predicted":"0","confidence":0.9998992849567556,"entropy_confidence":0.9991849815301093,"candidate_mass":0.999995787871441,"accepted":true,"accepted_predicted":"0","input_tokens":419,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00010072011534491012,"brier_hard":1.4059434148649989e-08,"score_mae":0.15983125252001226,"within_one":true,"case_id":"typed:282/r0","base_id":"typed:282","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.112701}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999764760460935,"hold":1.3451436906989217e-05,"manual_review":7.511010010005971e-06,"reject":2.561506988315927e-06},"predicted":"approve","confidence":0.9999764760460935,"entropy_confidence":0.9997864787153631,"candidate_mass":0.9999955934615762,"accepted":true,"accepted_predicted":"approve","input_tokens":423,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.3524230599022076e-05,"brier_hard":7.972941516784662e-10,"soft_accuracy":0.8399819267837106,"brier_soft":0.03738909932205986,"tvd":0.1599764760460941,"kl":1.2681672665094679,"case_id":"typed:282/r0","base_id":"typed:282","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.112701}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0021480660275281973,"true":0.9978519339724718},"predicted":"true","confidence":0.9978519339724718,"entropy_confidence":0.9778665631383237,"candidate_mass":0.9999857369006561,"accepted":true,"accepted_predicted":"true","input_tokens":371,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.143187363735237,"brier_hard":1.9914169642652046,"soft_accuracy":0.025380824378287568,"brier_soft":1.8993743053416856,"tvd":0.9745189339724718,"kl":5.8891571360954735,"case_id":"typed:282/r0","base_id":"typed:282","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.112701}
{"id":"matches_order","type":"binary","probabilities":{"false":6.884216647462181e-09,"true":0.9999999931157841},"predicted":"true","confidence":0.9999999931157841,"entropy_confidence":0.9999998034091103,"candidate_mass":0.9999999356379575,"accepted":true,"accepted_predicted":"true","input_tokens":391,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.884215887729863e-09,"brier_hard":9.4784866911808e-17,"soft_accuracy":0.9533329937583156,"brier_soft":0.004355616492937217,"tvd":0.04666699311578376,"kl":0.38425821044159014,"case_id":"typed:282/r0","base_id":"typed:282","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.112701}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006558395940879188,"1":0.8806483883906863,"2":0.11824944691265389,"3":0.0004463251025705251},"predicted":"1","confidence":0.8806483883906863,"entropy_confidence":0.7312002811205907,"candidate_mass":0.9999891431316466,"accepted":true,"accepted_predicted":"1","input_tokens":418,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.1270968378280059,"brier_hard":0.02822836822055938,"score_mae":0.004846742476294352,"within_one":true,"case_id":"typed:282/r0","base_id":"typed:282","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.112701}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0007286270213012988,"1":0.0003302830085456482,"2":0.6808097281421976,"3":0.3181313618279539},"predicted":"2","confidence":0.6808097281421976,"entropy_confidence":0.5426535753707019,"candidate_mass":0.9999728153169651,"accepted":false,"accepted_predicted":null,"input_tokens":385,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.1452908939898272,"brier_hard":0.9284473656398557,"score_mae":0.46365617522319713,"within_one":true,"case_id":"typed:283/r0","base_id":"typed:283","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.5301}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9927626117169481,"hold":0.006646881067211173,"manual_review":0.00048800748583177834,"reject":0.00010249973000731558},"predicted":"approve","confidence":0.9927626117169481,"entropy_confidence":0.9673960589900477,"candidate_mass":0.9999859679666094,"accepted":true,"accepted_predicted":"approve","input_tokens":389,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.625179812409563,"brier_hard":1.9846460179368148,"soft_accuracy":0.025229232568715246,"brier_soft":1.4713641468386698,"tvd":0.9694295883839256,"kl":6.008947629020858,"case_id":"typed:283/r0","base_id":"typed:283","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.5301}
{"id":"duplicate","type":"binary","probabilities":{"false":0.02927020087901418,"true":0.9707297991209866},"predicted":"true","confidence":0.9707297991209866,"entropy_confidence":0.8092813561105381,"candidate_mass":0.999914209768509,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":3.5311853153359145,"brier_hard":1.8846326858029403,"soft_accuracy":0.05751398882627335,"brier_soft":1.7699451099084218,"tvd":0.9407297991209862,"kl":3.291398801300435,"case_id":"typed:283/r0","base_id":"typed:283","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.5301}
{"id":"matches_order","type":"binary","probabilities":{"false":1.8196957044873764e-05,"true":0.9999818030429561},"predicted":"true","confidence":0.9999818030429561,"entropy_confidence":0.9996872193261127,"candidate_mass":0.9999957216614032,"accepted":true,"accepted_predicted":"true","input_tokens":357,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.91425617319683,"brier_hard":1.999927212834081,"soft_accuracy":0.0200174690787631,"brier_soft":1.9207286685906446,"tvd":0.9799818030429557,"kl":8.947893468370548,"case_id":"typed:283/r0","base_id":"typed:283","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.5301}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0033781507381238974,"1":0.37700771482126755,"2":0.6095884101801147,"3":0.010025724260495313},"predicted":"2","confidence":0.6095884101801147,"entropy_confidence":0.46990810554381934,"candidate_mass":0.9999918416791306,"accepted":false,"accepted_predicted":null,"input_tokens":384,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.4949712869621621,"brier_hard":0.2946679535498015,"score_mae":0.24373829203701725,"within_one":true,"case_id":"typed:283/r0","base_id":"typed:283","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":330.5301}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.001961436394805738,"1":0.00016723851065782272,"2":0.006957225806274316,"3":0.990914099288263},"predicted":"3","confidence":0.990914099288263,"entropy_confidence":0.9586740816836312,"candidate_mass":0.9999858876518719,"accepted":true,"accepted_predicted":"3","input_tokens":385,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.009127429248440719,"brier_hard":0.0001348317841133475,"score_mae":0.1768239879879956,"within_one":true,"case_id":"typed:284/r0","base_id":"typed:284","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.745401}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9997768782396214,"hold":0.00014130934134431064,"manual_review":6.309182171099075e-05,"reject":1.87205973221855e-05},"predicted":"approve","confidence":0.9997768782396214,"entropy_confidence":0.9983483391681416,"candidate_mass":0.9999908621005311,"accepted":true,"accepted_predicted":"approve","input_tokens":389,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":9.670919405200664,"brier_hard":1.9994276469185095,"soft_accuracy":0.023414613810851758,"brier_soft":1.3898135207182278,"tvd":0.976443878239622,"kl":8.271156733957048,"case_id":"typed:284/r0","base_id":"typed:284","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.745401}
{"id":"duplicate","type":"binary","probabilities":{"false":0.00026319239601864505,"true":0.999736807603981},"predicted":"true","confidence":0.999736807603981,"entropy_confidence":0.9964905663400673,"candidate_mass":0.999980339044126,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.242625249437118,"brier_hard":1.9989473689563992,"soft_accuracy":0.026916155292769375,"brier_soft":1.8937297009408978,"tvd":0.9730698076039812,"kl":7.89986798294689,"case_id":"typed:284/r0","base_id":"typed:284","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.745401}
{"id":"matches_order","type":"binary","probabilities":{"false":1.1087318204260019e-08,"true":0.9999999889126824},"predicted":"true","confidence":0.9999999889126824,"entropy_confidence":0.9999996910052088,"candidate_mass":0.9999998843512947,"accepted":true,"accepted_predicted":"true","input_tokens":357,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.31746388590421,"brier_hard":1.9999999556507286,"soft_accuracy":0.04000001020033277,"brier_soft":1.8431999574246993,"tvd":0.959999988912682,"kl":8.713171724545901,"case_id":"typed:284/r0","base_id":"typed:284","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.745401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012468160347077885,"1":0.8247776424541401,"2":0.17488297663765429,"3":0.00021469930473453258},"predicted":"1","confidence":0.8247776424541401,"entropy_confidence":0.6633086838621816,"candidate_mass":0.9999860237557111,"accepted":true,"accepted_predicted":"1","input_tokens":384,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.19264145327366938,"brier_hard":0.061286991742869196,"score_mae":0.5081453063563477,"within_one":true,"case_id":"typed:284/r0","base_id":"typed:284","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.745401}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.018542415564377027,"1":0.002794700511283483,"2":0.11020294749233617,"3":0.8684599364320039},"predicted":"3","confidence":0.8684599364320039,"entropy_confidence":0.6711364364314159,"candidate_mass":0.9999829585344147,"accepted":true,"accepted_predicted":"3","input_tokens":451,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.14103382397695086,"brier_hard":0.02979910948538087,"score_mae":0.06191340479196761,"within_one":true,"case_id":"typed:285/r0","base_id":"typed:285","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.42020099999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999433839176471,"hold":3.469220153361021e-05,"manual_review":1.5522312578188615e-05,"reject":6.401568242257783e-06},"predicted":"approve","confidence":0.9999433839176471,"entropy_confidence":0.9995229679349716,"candidate_mass":0.9999931465887157,"accepted":true,"accepted_predicted":"approve","input_tokens":455,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":11.073232047982955,"brier_hard":1.9998557279009896,"soft_accuracy":0.04001977297047123,"brier_soft":1.3062696643632037,"tvd":0.9599433839176466,"kl":8.713173988809812,"case_id":"typed:285/r0","base_id":"typed:285","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.42020099999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0017200639623410642,"true":0.9982799360376594},"predicted":"true","confidence":0.9982799360376594,"entropy_confidence":0.9817247063211627,"candidate_mass":0.9999807596354007,"accepted":true,"accepted_predicted":"true","input_tokens":403,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.365393801440614,"brier_hard":1.9931256613907058,"soft_accuracy":0.02165126140384743,"brier_soft":1.914063266507693,"tvd":0.9782799360376592,"kl":6.140081243031488,"case_id":"typed:285/r0","base_id":"typed:285","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.42020099999996}
{"id":"matches_order","type":"binary","probabilities":{"false":2.8442617322139116e-08,"true":0.9999999715573833},"predicted":"true","confidence":0.9999999715573833,"entropy_confidence":0.9999992459843513,"candidate_mass":0.9999998345695212,"accepted":true,"accepted_predicted":"true","input_tokens":423,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.375377206634756,"brier_hard":1.9999998862295334,"soft_accuracy":0.05666702521910177,"brier_soft":1.7797541904545646,"tvd":0.9433329715573829,"kl":8.525751808139772,"case_id":"typed:285/r0","base_id":"typed:285","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.42020099999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0037817366315005983,"1":0.1666800017448827,"2":0.8115873950057322,"3":0.017950866617882807},"predicted":"2","confidence":0.8115873950057322,"entropy_confidence":0.5950901733530775,"candidate_mass":0.9999895531898368,"accepted":true,"accepted_predicted":"2","input_tokens":450,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.20876320219380068,"brier_hard":0.06361806784668314,"score_mae":0.039625608390004396,"within_one":true,"case_id":"typed:285/r0","base_id":"typed:285","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":338.42020099999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999911325829857,"1":4.717934501292293e-06,"2":2.9448450204615648e-06,"3":1.2046374915951946e-06},"predicted":"0","confidence":0.9999911325829857,"entropy_confidence":0.999912968608036,"candidate_mass":0.9999990616368704,"accepted":true,"accepted_predicted":"0","input_tokens":383,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":8.867456330110138e-06,"brier_hard":1.110132541453043e-10,"score_mae":0.029985778462982997,"within_one":true,"case_id":"typed:286/r0","base_id":"typed:286","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.8657}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999884689171894,"hold":6.672166690319638e-06,"manual_review":3.925273891550271e-06,"reject":9.336422270680493e-07},"predicted":"approve","confidence":0.9999884689171894,"entropy_confidence":0.999889726195087,"candidate_mass":0.9999964963814797,"accepted":true,"accepted_predicted":"approve","input_tokens":387,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":1.1531149294047257e-05,"brier_hard":1.9376314206019448e-10,"soft_accuracy":0.9133229641314299,"brier_soft":0.011619327543282085,"tvd":0.08665640255941733,"kl":0.6981245680424443,"case_id":"typed:286/r0","base_id":"typed:286","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.8657}
{"id":"duplicate","type":"binary","probabilities":{"false":0.005218689719453935,"true":0.994781310280547},"predicted":"true","confidence":0.994781310280547,"entropy_confidence":0.9529220923124108,"candidate_mass":0.9999447424702123,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.255508920195137,"brier_hard":1.979179710566962,"soft_accuracy":0.025009942130675797,"brier_soft":1.9003972057445182,"tvd":0.9747813102805465,"kl":5.052464275604369,"case_id":"typed:286/r0","base_id":"typed:286","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.8657}
{"id":"matches_order","type":"binary","probabilities":{"false":1.218670234987387e-09,"true":0.9999999987813304},"predicted":"true","confidence":0.9999999987813304,"entropy_confidence":0.9999999621545135,"candidate_mass":0.9999999754765984,"accepted":true,"accepted_predicted":"true","input_tokens":355,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.2186696104131197e-09,"brier_hard":2.970312759178764e-18,"soft_accuracy":0.9866669988138275,"brier_soft":0.0003555377130059008,"tvd":0.013332998781330117,"kl":0.10955775151331722,"case_id":"typed:286/r0","base_id":"typed:286","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.8657}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007664609210697118,"1":0.7343335518410639,"2":0.2640619912630488,"3":0.0008379959748184177},"predicted":"1","confidence":0.7343335518410639,"entropy_confidence":0.5745428676260664,"candidate_mass":0.9999888428335718,"accepted":false,"accepted_predicted":null,"input_tokens":382,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3315713880103033,"brier_hard":1.0808518077628184,"score_mae":0.4183614777083833,"within_one":true,"case_id":"typed:286/r0","base_id":"typed:286","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.8657}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.021833164679665448,"1":0.002794708813253409,"2":0.16213715965558947,"3":0.8132349668514914},"predicted":"3","confidence":0.8132349668514914,"entropy_confidence":0.5938578579145354,"candidate_mass":0.9999207047620501,"accepted":true,"accepted_predicted":"3","input_tokens":378,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.20673519907282772,"brier_hard":0.06165413362542588,"score_mae":0.016773928678906813,"within_one":true,"case_id":"typed:287/r0","base_id":"typed:287","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.758201}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999635773559495,"hold":1.9654051624539945e-05,"manual_review":1.3013114381398429e-05,"reject":3.7554780445714536e-06},"predicted":"approve","confidence":0.9999635773559495,"entropy_confidence":0.9996806423863894,"candidate_mass":0.9999951093285775,"accepted":true,"accepted_predicted":"approve","input_tokens":382,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":10.837227051359129,"brier_hard":1.9998878485049856,"soft_accuracy":0.023347802272085248,"brier_soft":1.393432099131718,"tvd":0.9766305540229262,"kl":8.907753302383638,"case_id":"typed:287/r0","base_id":"typed:287","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.758201}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0013028527490598334,"true":0.9986971472509398},"predicted":"true","confidence":0.9986971472509398,"entropy_confidence":0.9856349213819959,"candidate_mass":0.9999656037084749,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.643198996389042,"brier_hard":1.9947919838543315,"soft_accuracy":0.01792642345552267,"brier_soft":1.928766420219406,"tvd":0.9820301472509401,"kl":6.447731092630219,"case_id":"typed:287/r0","base_id":"typed:287","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.758201}
{"id":"matches_order","type":"binary","probabilities":{"false":6.5429531795546e-08,"true":0.9999999345704683},"predicted":"true","confidence":0.9999999345704683,"entropy_confidence":0.9999983440977758,"candidate_mass":0.9999997269102003,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.542292123897397,"brier_hard":1.9999997382818817,"soft_accuracy":0.05000005888657862,"brier_soft":1.8049997513677878,"tvd":0.9499999345704682,"kl":8.600036742971152,"case_id":"typed:287/r0","base_id":"typed:287","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.758201}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007288150324823132,"1":0.9934014498612793,"2":0.005581928307336064,"3":0.0002878067989029977},"predicted":"1","confidence":0.9934014498612793,"entropy_confidence":0.9688748733596887,"candidate_mass":0.9999941588806315,"accepted":true,"accepted_predicted":"1","input_tokens":377,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.188220987494724,"brier_hard":1.9757143558995531,"score_mae":0.8779042731273394,"within_one":true,"case_id":"typed:287/r0","base_id":"typed:287","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":332.758201}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9991332335066273,"1":0.00019967089165749222,"2":0.0004280597484917254,"3":0.00023903585322191396},"predicted":"0","confidence":0.9991332335066273,"entropy_confidence":0.9943152131385495,"candidate_mass":0.9999864670073825,"accepted":true,"accepted_predicted":"0","input_tokens":398,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.756247772845818,"brier_hard":1.9974113790421777,"score_mae":1.3082271020516933,"within_one":false,"case_id":"typed:288/r0","base_id":"typed:288","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":317.641901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9967213881050284,"hold":0.002425150792617903,"manual_review":0.00040709548809678115,"reject":0.00044636561425554416},"predicted":"approve","confidence":0.9967213881050284,"entropy_confidence":0.9823280244989278,"candidate_mass":0.999974723022022,"accepted":true,"accepted_predicted":"approve","input_tokens":402,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.8064627855509485,"brier_hard":1.9926455808551862,"soft_accuracy":0.28316695367043077,"brier_soft":0.7891036200465181,"tvd":0.7133883881050291,"kl":4.095071255405257,"case_id":"typed:288/r0","base_id":"typed:288","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":317.641901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.00033117663074011885,"true":0.9996688233692608},"predicted":"true","confidence":0.9996688233692608,"entropy_confidence":0.9956938534623165,"candidate_mass":0.9999748693677548,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.012858697302143,"brier_hard":1.9986755128329627,"soft_accuracy":0.036973890123701454,"brier_soft":1.85474502362504,"tvd":0.9630018233692603,"kl":7.561860324765393,"case_id":"typed:288/r0","base_id":"typed:288","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":317.641901}
{"id":"matches_order","type":"binary","probabilities":{"false":4.57517454195829e-09,"true":0.999999995424826},"predicted":"true","confidence":0.999999995424826,"entropy_confidence":0.999999866650952,"candidate_mass":0.9999999575050387,"accepted":true,"accepted_predicted":"true","input_tokens":370,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.575174027884357e-09,"brier_hard":4.186443937904228e-17,"soft_accuracy":0.7099999980784271,"brier_soft":0.16819999469279795,"tvd":0.2899999954248258,"kl":2.427830591769256,"case_id":"typed:288/r0","base_id":"typed:288","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":317.641901}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.437261493319276e-05,"1":4.184776518435643e-05,"2":0.00015625956178640592,"3":0.9997275200580951},"predicted":"3","confidence":0.9997275200580951,"entropy_confidence":0.9980012867006002,"candidate_mass":0.9999974882989764,"accepted":true,"accepted_predicted":"3","input_tokens":397,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0002725170713091371,"brier_hard":1.0594489069312241e-07,"score_mae":0.4328699270630425,"within_one":true,"case_id":"typed:288/r0","base_id":"typed:288","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":317.641901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999690738562521,"1":1.6171789374863758e-05,"2":1.0510120675938812e-05,"3":4.244233698837579e-06},"predicted":"0","confidence":0.9999690738562521,"entropy_confidence":0.9997242165521292,"candidate_mass":0.9999976945717856,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":3.092662197093247e-05,"brier_hard":1.3464292950132048e-09,"score_mae":0.03328307526817675,"within_one":true,"case_id":"typed:289/r0","base_id":"typed:289","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.6856}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999940763414282,"hold":3.5231757372475915e-06,"manual_review":1.8707626737635601e-06,"reject":5.297201605777919e-07},"predicted":"approve","confidence":0.9999940763414282,"entropy_confidence":0.9999404961190149,"candidate_mass":0.9999975673282644,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":5.9236761167061865e-06,"brier_hard":5.128285458052117e-11,"soft_accuracy":0.9399946056335307,"brier_soft":0.005487839245077711,"tvd":0.059994076341428344,"kl":0.490993864553183,"case_id":"typed:289/r0","base_id":"typed:289","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.6856}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0017770866850530642,"true":0.9982229133149482},"predicted":"true","confidence":0.9982229133149482,"entropy_confidence":0.9812025410720143,"candidate_mass":0.9999728558355552,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.332779949315075,"brier_hard":1.9928979693339626,"soft_accuracy":0.03167046148394991,"brier_soft":1.8749112197361688,"tvd":0.9682229133149476,"kl":6.008107742683158,"case_id":"typed:289/r0","base_id":"typed:289","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.6856}
{"id":"matches_order","type":"binary","probabilities":{"false":1.901311630569295e-09,"true":0.9999999980986889},"predicted":"true","confidence":0.9999999980986889,"entropy_confidence":0.9999999421753096,"candidate_mass":0.9999999663129492,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.901311110361674e-09,"brier_hard":7.229969848049605e-18,"soft_accuracy":0.9833329981620673,"brier_soft":0.0005555776512433801,"tvd":0.01666699809868862,"kl":0.13698141744067002,"case_id":"typed:289/r0","base_id":"typed:289","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.6856}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.016436741220729724,"1":0.19496798001375312,"2":0.6651280518927699,"3":0.1234672268727482},"predicted":"2","confidence":0.6651280518927699,"entropy_confidence":0.33940990405319593,"candidate_mass":0.9999890272098789,"accepted":false,"accepted_predicted":null,"input_tokens":406,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.4077756976261852,"brier_hard":0.16566605743337856,"score_mae":0.007707235582462335,"within_one":true,"case_id":"typed:289/r0","base_id":"typed:289","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":320.6856}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.19806038453698177,"1":0.004444817977923079,"2":0.13907354943943603,"3":0.6584212480456593},"predicted":"3","confidence":0.6584212480456593,"entropy_confidence":0.35490756164958526,"candidate_mass":0.9999146794223926,"accepted":false,"accepted_predicted":null,"input_tokens":383,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.4179103578956962,"brier_hard":0.17526516827016225,"score_mae":0.5554773390062273,"within_one":true,"case_id":"typed:290/r0","base_id":"typed:290","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.28750099999996}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9978851381006505,"hold":0.0012252752232919583,"manual_review":0.0004828460443914593,"reject":0.0004067406316659336},"predicted":"approve","confidence":0.9978851381006505,"entropy_confidence":0.9875999827319839,"candidate_mass":0.9999620039509189,"accepted":true,"accepted_predicted":"approve","input_tokens":387,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":false,"nll_hard":7.635812703794784,"brier_hard":1.9948109566309884,"soft_accuracy":0.007314329149619387,"brier_soft":1.5269030881985324,"tvd":0.9912181381006506,"kl":6.57699379878401,"case_id":"typed:290/r0","base_id":"typed:290","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.28750099999996}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0030553887102550063,"true":0.9969446112897439},"predicted":"true","confidence":0.9969446112897439,"entropy_confidence":0.9700727207304862,"candidate_mass":0.9999626092062995,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.790848457053222,"brier_hard":1.9877971159593193,"soft_accuracy":0.026245805940702218,"brier_soft":1.8958391392764247,"tvd":0.9736116112897444,"kl":5.545060578908393,"case_id":"typed:290/r0","base_id":"typed:290","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.28750099999996}
{"id":"matches_order","type":"binary","probabilities":{"false":1.7628875356704901e-07,"true":0.9999998237112471},"predicted":"true","confidence":0.9999998237112471,"entropy_confidence":0.9999957905328705,"candidate_mass":0.9999996854094577,"accepted":true,"accepted_predicted":"true","input_tokens":355,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.55114254102914,"brier_hard":1.9999992948450491,"soft_accuracy":0.013333171587837672,"brier_soft":1.9470228420248807,"tvd":0.9866668237112467,"kl":9.029973503800138,"case_id":"typed:290/r0","base_id":"typed:290","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.28750099999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002435528026922979,"1":0.00015929562886327927,"2":0.0007672903459451797,"3":0.998829861222498},"predicted":"3","confidence":0.998829861222498,"entropy_confidence":0.9927199025746435,"candidate_mass":0.9999971247875531,"accepted":true,"accepted_predicted":"3","input_tokens":382,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.172645280711521,"brier_hard":1.9961271844054043,"score_mae":0.8648504599882476,"within_one":true,"case_id":"typed:290/r0","base_id":"typed:290","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.28750099999996}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999815655988193,"1":9.581345656562333e-06,"2":5.786579139810587e-06,"3":3.066476384809574e-06},"predicted":"0","confidence":0.9999815655988193,"entropy_confidence":0.9998284144050096,"candidate_mass":0.9999975241896687,"accepted":true,"accepted_predicted":"0","input_tokens":415,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.843457109631751e-05,"brier_hard":4.745171070397781e-10,"score_mae":0.10663664606690938,"within_one":true,"case_id":"typed:291/r0","base_id":"typed:291","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.110601}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999992912260669,"hold":3.784935027465451e-07,"manual_review":2.905935344322603e-07,"reject":3.9686895754097696e-08},"predicted":"approve","confidence":0.9999992912260669,"entropy_confidence":0.9999918085518226,"candidate_mass":0.9999992992839996,"accepted":true,"accepted_predicted":"approve","input_tokens":419,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":7.087741842805669e-07,"brier_hard":7.316374718121114e-13,"soft_accuracy":0.32166693592516965,"brier_soft":0.6216484941585262,"tvd":0.678332291226067,"kl":5.882830970456548,"case_id":"typed:291/r0","base_id":"typed:291","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.110601}
{"id":"duplicate","type":"binary","probabilities":{"false":8.564473062580206e-05,"true":0.9999143552693724},"predicted":"true","confidence":0.9999143552693724,"entropy_confidence":0.9987192764706408,"candidate_mass":0.9999972003055438,"accepted":true,"accepted_predicted":"true","input_tokens":367,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.56483983469118e-05,"brier_hard":1.4670039768232532e-08,"soft_accuracy":0.6166470161724231,"brier_soft":0.29375707064193857,"tvd":0.38324735526937337,"kl":2.92441021521264,"case_id":"typed:291/r0","base_id":"typed:291","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.110601}
{"id":"matches_order","type":"binary","probabilities":{"false":3.2248385523479358e-09,"true":0.9999999967751609},"predicted":"true","confidence":0.9999999967751609,"entropy_confidence":0.9999999043808887,"candidate_mass":0.9999999623650087,"accepted":true,"accepted_predicted":"true","input_tokens":387,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.2248390676543143e-09,"brier_hard":2.0799170667442083e-17,"soft_accuracy":0.9566669970546449,"brier_soft":0.0037554972190322533,"tvd":0.04333299677516115,"kl":0.3567314199795176,"case_id":"typed:291/r0","base_id":"typed:291","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.110601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.014397366772218688,"1":0.010290354590279778,"2":0.0012593483045746305,"3":0.9740529303329254},"predicted":"3","confidence":0.9740529303329254,"entropy_confidence":0.8974491053861331,"candidate_mass":0.9999875902576285,"accepted":true,"accepted_predicted":"3","input_tokens":414,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.026289633558519678,"brier_hard":0.000988011950027737,"score_mae":0.9683008421982053,"within_one":true,"case_id":"typed:291/r0","base_id":"typed:291","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":328.110601}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9998459383498225,"1":7.04479725453982e-05,"2":5.263493642163097e-05,"3":3.097874121101762e-05},"predicted":"0","confidence":0.9998459383498225,"entropy_confidence":0.9987969573447796,"candidate_mass":0.9999934074124354,"accepted":true,"accepted_predicted":"0","input_tokens":381,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.852130468172213,"brier_hard":1.9995866392548296,"score_mae":1.3697313459309783,"within_one":false,"case_id":"typed:292/r0","base_id":"typed:292","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.589901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9992331428320119,"hold":0.0004839088557105262,"manual_review":0.0001649862320729678,"reject":0.00011796208020467121},"predicted":"approve","confidence":0.9992331428320119,"entropy_confidence":0.9949761802757093,"candidate_mass":0.9999762588535248,"accepted":true,"accepted_predicted":"approve","input_tokens":385,"reused_prefix_tokens":0,"expected":"hold","correct":false,"accepted_correct":false,"nll_hard":7.633613983628745,"brier_hard":1.9974993313258087,"soft_accuracy":0.2934135596190331,"brier_soft":0.8611506253551636,"tvd":0.705900142832012,"kl":4.577228401763252,"case_id":"typed:292/r0","base_id":"typed:292","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.589901}
{"id":"duplicate","type":"binary","probabilities":{"false":3.8088917269810627e-06,"true":0.9999961911082728},"predicted":"true","confidence":0.9999961911082728,"entropy_confidence":0.9999259365209281,"candidate_mass":0.9999982099488561,"accepted":true,"accepted_predicted":"true","input_tokens":333,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.47817229644781,"brier_hard":1.9999847644621067,"soft_accuracy":0.04667045339262652,"brier_soft":1.8176730932383076,"tvd":0.9533291911082729,"kl":8.637500400577958,"case_id":"typed:292/r0","base_id":"typed:292","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.589901}
{"id":"matches_order","type":"binary","probabilities":{"false":8.189188881122826e-09,"true":0.999999991810812},"predicted":"true","confidence":0.999999991810812,"entropy_confidence":0.9999997681941403,"candidate_mass":0.9999999235152466,"accepted":true,"accepted_predicted":"true","input_tokens":353,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":18.620450981699932,"brier_hard":1.9999999672432465,"soft_accuracy":0.4200000013102706,"brier_soft":0.6727999810010832,"tvd":0.5799999918108116,"kl":4.9776471807496625,"case_id":"typed:292/r0","base_id":"typed:292","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.589901}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.7646370743664965e-05,"1":3.2811473291872485e-05,"2":0.00011313552962527519,"3":0.9997964066263402},"predicted":"3","confidence":0.9997964066263402,"entropy_confidence":0.9984612985933294,"candidate_mass":0.9999977458038997,"accepted":true,"accepted_predicted":"3","input_tokens":380,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.086924080715374,"brier_hard":1.9993666008430364,"score_mae":0.8996483024115629,"within_one":true,"case_id":"typed:292/r0","base_id":"typed:292","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":325.589901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9999914699771051,"1":4.442397060958682e-06,"2":2.97292992157588e-06,"3":1.1146959114337452e-06},"predicted":"0","confidence":0.9999914699771051,"entropy_confidence":0.9999160409581682,"candidate_mass":0.9999987556594153,"accepted":true,"accepted_predicted":"0","input_tokens":420,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":8.53005927576696e-06,"brier_hard":1.0257704152855292e-10,"score_mae":0.9899862676553616,"within_one":true,"case_id":"typed:293/r0","base_id":"typed:293","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.122101}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999990602477179,"hold":5.340124722828484e-07,"manual_review":3.425926369842708e-07,"reject":6.314717293701096e-08},"predicted":"approve","confidence":0.9999990602477179,"entropy_confidence":0.9999893245363304,"candidate_mass":0.9999989518895859,"accepted":true,"accepted_predicted":"approve","input_tokens":424,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":9.397527236705315e-07,"brier_hard":1.289660952637358e-12,"soft_accuracy":0.519999593178403,"brier_soft":0.3542879919179192,"tvd":0.4799990602477178,"kl":4.080922104168519,"case_id":"typed:293/r0","base_id":"typed:293","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.122101}
{"id":"duplicate","type":"binary","probabilities":{"false":7.661122844460378e-05,"true":0.9999233887715561},"predicted":"true","confidence":0.9999233887715561,"entropy_confidence":0.9988420423964327,"candidate_mass":0.9999969670687129,"accepted":true,"accepted_predicted":"true","input_tokens":372,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.661416323398936e-05,"brier_hard":1.1738560647480009e-08,"soft_accuracy":0.5166644462413114,"brier_soft":0.46707347457705006,"tvd":0.48325638877155574,"kl":3.8878822633923975,"case_id":"typed:293/r0","base_id":"typed:293","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.122101}
{"id":"matches_order","type":"binary","probabilities":{"false":2.1671447271017517e-09,"true":0.9999999978328553},"predicted":"true","confidence":0.9999999978328553,"entropy_confidence":0.9999999344996708,"candidate_mass":0.9999999658753472,"accepted":true,"accepted_predicted":"true","input_tokens":392,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.1671446882755274e-09,"brier_hard":9.393032357947727e-18,"soft_accuracy":0.9833329979050949,"brier_soft":0.0005555776335208053,"tvd":0.01666699783285528,"kl":0.13698141770207292,"case_id":"typed:293/r0","base_id":"typed:293","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.122101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006529732243210348,"1":0.9957997320971629,"2":0.003480320905636664,"3":6.697377288019482e-05},"predicted":"1","confidence":0.9957997320971629,"entropy_confidence":0.9788466037643483,"candidate_mass":0.9999960850989318,"accepted":true,"accepted_predicted":"1","input_tokens":419,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.004209113806865662,"brier_hard":3.018574357974935e-05,"score_mae":0.019628295227076764,"within_one":true,"case_id":"typed:293/r0","base_id":"typed:293","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.122101}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.958496325215449,"1":0.0011739558994194653,"2":0.006065840327444832,"3":0.03426387855768537},"predicted":"0","confidence":0.958496325215449,"entropy_confidence":0.8592558921837103,"candidate_mass":0.9999574786710018,"accepted":true,"accepted_predicted":"0","input_tokens":407,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.04238955039481591,"brier_hard":0.002934740985769421,"score_mae":0.7372357277726348,"within_one":true,"case_id":"typed:294/r0","base_id":"typed:294","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.37420099999997}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9995961838352189,"hold":0.00028823020965210817,"manual_review":7.775995018739311e-05,"reject":3.7826004940599985e-05},"predicted":"approve","confidence":0.9995961838352189,"entropy_confidence":0.9972053302015846,"candidate_mass":0.9999843392142248,"accepted":true,"accepted_predicted":"approve","input_tokens":411,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.0004038977204850171,"brier_hard":2.5362156519755474e-07,"soft_accuracy":0.46657104447639963,"brier_soft":0.4071395701172035,"tvd":0.5329291838352194,"kl":3.5729130595339473,"case_id":"typed:294/r0","base_id":"typed:294","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.37420099999997}
{"id":"duplicate","type":"binary","probabilities":{"false":0.006506257459707033,"true":0.9934937425402935},"predicted":"true","confidence":0.9934937425402935,"entropy_confidence":0.9433828863758852,"candidate_mass":0.999969484126222,"accepted":true,"accepted_predicted":"true","input_tokens":359,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.034990879056863,"brier_hard":1.974059632933437,"soft_accuracy":0.03940551129989822,"brier_soft":1.8438173030310545,"tvd":0.9601607425402933,"kl":4.721233486181506,"case_id":"typed:294/r0","base_id":"typed:294","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.37420099999997}
{"id":"matches_order","type":"binary","probabilities":{"false":2.0601042180358755e-09,"true":0.999999997939895},"predicted":"true","confidence":0.999999997939895,"entropy_confidence":0.9999999375843404,"candidate_mass":0.9999999696198381,"accepted":true,"accepted_predicted":"true","input_tokens":379,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.060104977620705e-09,"brier_hard":8.48806189924366e-18,"soft_accuracy":0.8733329984617896,"brier_soft":0.03208905673421092,"tvd":0.1266669979398954,"kl":1.0483634029272608,"case_id":"typed:294/r0","base_id":"typed:294","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.37420099999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.005647165988427778,"1":0.764627948944868,"2":0.21727357147653975,"3":0.012451313590163278},"predicted":"1","confidence":0.764627948944868,"entropy_confidence":0.5522356579221579,"candidate_mass":0.9999917945669515,"accepted":false,"accepted_predicted":null,"input_tokens":406,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.526598021405622,"brier_hard":1.1975034879105486,"score_mae":0.6334709673315628,"within_one":true,"case_id":"typed:294/r0","base_id":"typed:294","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.37420099999997}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.0007680632889394545,"1":0.0007024604830388131,"2":0.9244873249294231,"3":0.07404215129860017},"predicted":"2","confidence":0.9244873249294231,"entropy_confidence":0.8009539391105935,"candidate_mass":0.9999915603703698,"accepted":true,"accepted_predicted":"2","input_tokens":397,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.6031207358021398,"brier_hard":1.7120758348988314,"score_mae":0.8448634357623144,"within_one":true,"case_id":"typed:295/r0","base_id":"typed:295","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.0717}
{"id":"disposition","type":"choice","probabilities":{"approve":0.15421481014760946,"hold":0.5936029446911927,"manual_review":0.019633081555156216,"reject":0.23254916360604133},"predicted":"hold","confidence":0.5936029446911927,"entropy_confidence":0.26836734525669814,"candidate_mass":0.999988792936188,"accepted":false,"accepted_predicted":null,"input_tokens":401,"reused_prefix_tokens":0,"expected":"manual_review","correct":false,"accepted_correct":null,"nll_hard":3.9305393010907466,"brier_hard":1.3913450718898268,"soft_accuracy":0.08875532036496006,"brier_soft":0.9369899720482192,"tvd":0.7970339184448436,"kl":2.796540861405607,"case_id":"typed:295/r0","base_id":"typed:295","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.0717}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003448344785106645,"true":0.9965516552148919},"predicted":"true","confidence":0.9965516552148919,"entropy_confidence":0.9668266301408538,"candidate_mass":0.9999743348905457,"accepted":true,"accepted_predicted":"true","input_tokens":349,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.669860935499344,"brier_hard":1.9862304030230842,"soft_accuracy":0.02331041099370235,"brier_soft":1.907306270605893,"tvd":0.9765516552148925,"kl":5.458493689590217,"case_id":"typed:295/r0","base_id":"typed:295","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.0717}
{"id":"matches_order","type":"binary","probabilities":{"false":4.459257676037373e-05,"true":0.9999554074232383},"predicted":"true","confidence":0.9999554074232383,"entropy_confidence":0.9992911787058867,"candidate_mass":0.9999898063675202,"accepted":true,"accepted_predicted":"true","input_tokens":369,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.017943153141353,"brier_hard":1.9998216336699517,"soft_accuracy":4.459257676037373e-05,"brier_soft":1.9998216336699517,"tvd":0.999955407423239,"kl":9.210340371976184,"case_id":"typed:295/r0","base_id":"typed:295","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.0717}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00011696576227925247,"1":7.712826861245414e-05,"2":0.00028300385403019106,"3":0.999522902115077},"predicted":"3","confidence":0.999522902115077,"entropy_confidence":0.9966973015774319,"candidate_mass":0.9999976548307253,"accepted":true,"accepted_predicted":"3","input_tokens":396,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.170050041932527,"brier_hard":1.998480123865426,"score_mae":0.7492118423219036,"within_one":true,"case_id":"typed:295/r0","base_id":"typed:295","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":323.0717}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.13796610726502506,"1":0.001196355327152672,"2":0.03200130949798376,"3":0.8288362279098392},"predicted":"3","confidence":0.8288362279098392,"entropy_confidence":0.6053701686785161,"candidate_mass":0.9999538164600031,"accepted":true,"accepted_predicted":"3","input_tokens":419,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":1.9807472235182395,"brier_hard":1.4310974399952798,"score_mae":1.5683746580526376,"within_one":false,"case_id":"typed:296/r0","base_id":"typed:296","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.824901}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9961916524810106,"hold":0.0031079164655880326,"manual_review":0.00041023773191622965,"reject":0.00029019332148665},"predicted":"approve","confidence":0.9961916524810106,"entropy_confidence":0.9803010464137357,"candidate_mass":0.9999797935730571,"accepted":true,"accepted_predicted":"approve","input_tokens":423,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.003815617738620601,"brier_hard":2.4415162742989128e-05,"soft_accuracy":0.4330352177939755,"brier_soft":0.5024484400928132,"tvd":0.5628586524810099,"kl":2.4975857796700565,"case_id":"typed:296/r0","base_id":"typed:296","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.824901}
{"id":"duplicate","type":"binary","probabilities":{"false":0.004725665953498085,"true":0.9952743340465035},"predicted":"true","confidence":0.9952743340465035,"entropy_confidence":0.9566914101665176,"candidate_mass":0.9999781365578994,"accepted":true,"accepted_predicted":"true","input_tokens":371,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.35474678542862,"brier_hard":1.9811420000234188,"soft_accuracy":0.0443476126772183,"brier_soft":1.8250981065759784,"tvd":0.9552743340465026,"kl":4.972802240965931,"case_id":"typed:296/r0","base_id":"typed:296","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.824901}
{"id":"matches_order","type":"binary","probabilities":{"false":4.176567752764894e-09,"true":0.9999999958234316},"predicted":"true","confidence":0.9999999958234316,"entropy_confidence":0.9999998777195759,"candidate_mass":0.9999999605679236,"accepted":true,"accepted_predicted":"true","input_tokens":391,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.176568433162891e-09,"brier_hard":3.488744199747342e-17,"soft_accuracy":0.6766669985242763,"brier_soft":0.20908845237631085,"tvd":0.3233329958234319,"kl":2.7137170940938184,"case_id":"typed:296/r0","base_id":"typed:296","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.824901}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00012610742460548985,"1":7.755173262412227e-05,"2":0.00029217798503675487,"3":0.9995041628577351},"predicted":"3","confidence":0.9995041628577351,"entropy_confidence":0.9965810038565421,"candidate_mass":0.9999971323153019,"accepted":true,"accepted_predicted":"3","input_tokens":418,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.138147403591596,"brier_hard":1.998424322885197,"score_mae":0.9491743962759029,"within_one":true,"case_id":"typed:296/r0","base_id":"typed:296","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":331.824901}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.999987628535715,"1":7.843484265680391e-06,"2":3.2841150972268914e-06,"3":1.2438649236542768e-06},"predicted":"0","confidence":0.999987628535715,"entropy_confidence":0.9998824507819364,"candidate_mass":0.9999989596145417,"accepted":true,"accepted_predicted":"0","input_tokens":396,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":1.2371540812205795e-05,"brier_hard":2.269059859013935e-10,"score_mae":0.039981856690768904,"within_one":true,"case_id":"typed:297/r0","base_id":"typed:297","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.689501}
{"id":"disposition","type":"choice","probabilities":{"approve":0.999997889328064,"hold":1.2354124501699177e-06,"manual_review":7.307395246797299e-07,"reject":1.4451995941671126e-07},"predicted":"approve","confidence":0.999997889328064,"entropy_confidence":0.9999772643596905,"candidate_mass":0.9999989206215715,"accepted":true,"accepted_predicted":"approve","input_tokens":400,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.1106741634775923e-06,"brier_hard":6.53604621507901e-12,"soft_accuracy":0.9433310618465702,"brier_soft":0.004644170525523377,"tvd":0.05666488932806488,"kl":0.4668941407237207,"case_id":"typed:297/r0","base_id":"typed:297","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.689501}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0018462279069534346,"true":0.998153772093045},"predicted":"true","confidence":0.998153772093045,"entropy_confidence":0.9805729448571193,"candidate_mass":0.9999740555582691,"accepted":true,"accepted_predicted":"true","input_tokens":348,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.29461069063392,"brier_hard":1.992621905487152,"soft_accuracy":0.018451685745903026,"brier_soft":1.9266325675872529,"tvd":0.9814867720930458,"kl":6.104961778883306,"case_id":"typed:297/r0","base_id":"typed:297","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.689501}
{"id":"matches_order","type":"binary","probabilities":{"false":6.980697185205917e-09,"true":0.9999999930193031},"predicted":"true","confidence":0.9999999930193031,"entropy_confidence":0.9999998007941012,"candidate_mass":0.9999999603029622,"accepted":true,"accepted_predicted":"true","input_tokens":368,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.980696933773899e-09,"brier_hard":9.746026253257186e-17,"soft_accuracy":0.979999993298531,"brier_soft":0.0007999994415443352,"tvd":0.01999999301930297,"kl":0.16440816110943776,"case_id":"typed:297/r0","base_id":"typed:297","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.689501}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.764284951730904e-05,"1":0.9980579728508553,"2":0.001800080796845669,"3":4.4303502780632245e-05},"predicted":"1","confidence":0.9980579728508553,"entropy_confidence":0.989423397310832,"candidate_mass":0.9999961493876048,"accepted":true,"accepted_predicted":"1","input_tokens":395,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.001943915328862266,"brier_hard":7.02325724960816e-06,"score_mae":0.22820895504711136,"within_one":true,"case_id":"typed:297/r0","base_id":"typed:297","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":316.689501}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9996236846568671,"1":0.0001442251454632382,"2":0.00014716120732702334,"3":8.492899034326989e-05},"predicted":"0","confidence":0.9996236846568671,"entropy_confidence":0.9972975166201815,"candidate_mass":0.9999917856164319,"accepted":true,"accepted_predicted":"0","input_tokens":378,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.0003763861675203698,"brier_hard":1.9128348440377734e-07,"score_mae":0.3893066654688529,"within_one":true,"case_id":"typed:298/r0","base_id":"typed:298","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.19980100000004}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9998674179522814,"hold":8.562878686921752e-05,"manual_review":3.6627919690745835e-05,"reject":1.032534115887053e-05},"predicted":"approve","confidence":0.9998674179522814,"entropy_confidence":0.9989704811098327,"candidate_mass":0.9999918694873832,"accepted":true,"accepted_predicted":"approve","input_tokens":382,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":0.00013259083749520733,"brier_hard":2.6358505688869545e-08,"soft_accuracy":0.8432318301512142,"brier_soft":0.03560537329526033,"tvd":0.15653357461843798,"kl":0.9665547876929648,"case_id":"typed:298/r0","base_id":"typed:298","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.19980100000004}
{"id":"duplicate","type":"binary","probabilities":{"false":0.0021246780477632986,"true":0.9978753219522355},"predicted":"true","confidence":0.9978753219522355,"entropy_confidence":0.9780739570581314,"candidate_mass":0.9999577490430079,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.15413499501901,"brier_hard":1.9915103163225578,"soft_accuracy":0.028678360468763858,"brier_soft":1.8864912092585566,"tvd":0.971208321952236,"kl":5.867121197733684,"case_id":"typed:298/r0","base_id":"typed:298","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.19980100000004}
{"id":"matches_order","type":"binary","probabilities":{"false":1.3700904287993958e-09,"true":0.9999999986299102},"predicted":"true","confidence":0.9999999986299102,"entropy_confidence":0.9999999576836938,"candidate_mass":0.9999999724628953,"accepted":true,"accepted_predicted":"true","input_tokens":350,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.370089820313271e-09,"brier_hard":3.754293896241905e-18,"soft_accuracy":0.9033329988947952,"brier_soft":0.018689017248229974,"tvd":0.09666699862990982,"kl":0.7984995072853096,"case_id":"typed:298/r0","base_id":"typed:298","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.19980100000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000136360154308771,"1":0.9917027167150471,"2":0.008050722237285686,"3":0.00011020089335949155},"predicted":"1","confidence":0.9917027167150471,"entropy_confidence":0.9644366923607146,"candidate_mass":0.9999947367569841,"accepted":true,"accepted_predicted":"1","input_tokens":377,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.008331897341307112,"brier_hard":0.000133689776781266,"score_mae":0.11186523613030319,"within_one":true,"case_id":"typed:298/r0","base_id":"typed:298","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":336.19980100000004}
{"id":"discrepancy_severity","type":"ordinal","probabilities":{"0":0.9996081003877259,"1":0.00011166150444537,"2":0.0001833909334516492,"3":9.684717437619234e-05},"predicted":"0","confidence":0.9996081003877259,"entropy_confidence":0.9972005077388268,"candidate_mass":0.9999913438566223,"accepted":true,"accepted_predicted":"0","input_tokens":409,"reused_prefix_tokens":0,"expected":"0","correct":true,"accepted_correct":true,"nll_hard":0.00039197642499643907,"brier_hard":2.0906520733255166e-07,"score_mae":0.5858980151055228,"within_one":true,"case_id":"typed:299/r0","base_id":"typed:299","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.04900100000003}
{"id":"disposition","type":"choice","probabilities":{"approve":0.9999742926344503,"hold":1.5209436360046329e-05,"manual_review":8.308831262796884e-06,"reject":2.1890979276510818e-06},"predicted":"approve","confidence":0.9999742926344503,"entropy_confidence":0.9997690526384169,"candidate_mass":0.9999936654057017,"accepted":true,"accepted_predicted":"approve","input_tokens":413,"reused_prefix_tokens":0,"expected":"approve","correct":true,"accepted_correct":true,"nll_hard":2.5707695989727994e-05,"brier_hard":9.66024424588893e-10,"soft_accuracy":0.7333175830123183,"brier_soft":0.09998062021112165,"tvd":0.26664055930071656,"kl":2.186726225084333,"case_id":"typed:299/r0","base_id":"typed:299","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.04900100000003}
{"id":"duplicate","type":"binary","probabilities":{"false":0.003064599572545742,"true":0.9969354004274557},"predicted":"true","confidence":0.9969354004274557,"entropy_confidence":0.9699958299761899,"candidate_mass":0.9999796972207065,"accepted":true,"accepted_predicted":"true","input_tokens":361,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.787838363301464,"brier_hard":1.9877603852508998,"soft_accuracy":0.03288072359819304,"brier_soft":1.8699281371996053,"tvd":0.9669354004274548,"kl":5.47955312337487,"case_id":"typed:299/r0","base_id":"typed:299","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.04900100000003}
{"id":"matches_order","type":"binary","probabilities":{"false":6.4708493092413044e-09,"true":0.9999999935291513},"predicted":"true","confidence":0.9999999935291513,"entropy_confidence":0.999999814635451,"candidate_mass":0.9999999518654399,"accepted":true,"accepted_predicted":"true","input_tokens":381,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.470848766656956e-09,"brier_hard":8.374377427290788e-17,"soft_accuracy":0.899999994823321,"brier_soft":0.01999999741166046,"tvd":0.09999999352915095,"kl":0.8262095789293383,"case_id":"typed:299/r0","base_id":"typed:299","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.04900100000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0018830621487629243,"1":0.8299899417286989,"2":0.16501055693692038,"3":0.003116439185618726},"predicted":"1","confidence":0.8299899417286989,"entropy_confidence":0.6524764551043178,"candidate_mass":0.999992104755879,"accepted":true,"accepted_predicted":"1","input_tokens":408,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":1.8017458256924002,"brier_hard":1.3861039315138546,"score_mae":0.5806396268406042,"within_one":true,"case_id":"typed:299/r0","base_id":"typed:299","repeat":0,"workflow":"invoice_processing","case_elapsed_ms":339.04900100000003}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0014359293133621028,"true":0.9985640706866381},"predicted":"true","confidence":0.9985640706866381,"entropy_confidence":0.9843692513460487,"candidate_mass":0.9999676050901533,"accepted":true,"accepted_predicted":"true","input_tokens":358,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0014369612478333166,"brier_hard":4.123785985944521e-06,"soft_accuracy":0.6994256282746553,"brier_soft":0.1782810086099516,"tvd":0.29856407068663804,"kl":1.353924481092354,"case_id":"typed:300/r0","base_id":"typed:300","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.56510099999997}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.531142293929658e-06,"contain":4.234279464118872e-06,"investigate":0.9999746322616091,"monitor":1.460231663269053e-05},"predicted":"investigate","confidence":0.9999746322616091,"entropy_confidence":0.9997703827917118,"candidate_mass":0.9999952560346846,"accepted":true,"accepted_predicted":"investigate","input_tokens":410,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":12.372297382678077,"brier_hard":1.9999407968816245,"soft_accuracy":0.36666108862505864,"brier_soft":0.6782935315241375,"tvd":0.6333079989282425,"kl":5.4047254462169425,"case_id":"typed:300/r0","base_id":"typed:300","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.56510099999997}
{"id":"severity","type":"ordinal","probabilities":{"0":2.959304368126724e-05,"1":0.00026933970529915963,"2":0.9979465941852146,"3":0.0017540122764127739,"4":4.6078939108599227e-07},"predicted":"2","confidence":0.9979465941852146,"entropy_confidence":0.9902381074026765,"candidate_mass":0.9999978732321415,"accepted":true,"accepted_predicted":"2","input_tokens":430,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.345849385961557,"brier_hard":1.992392530271947,"score_mae":0.4985735919374692,"within_one":true,"case_id":"typed:300/r0","base_id":"typed:300","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.56510099999997}
{"id":"true_positive","type":"binary","probabilities":{"false":3.565964913967484e-07,"true":0.9999996434035087},"predicted":"true","confidence":0.9999996434035087,"entropy_confidence":0.9999918475270515,"candidate_mass":0.9999990524808963,"accepted":true,"accepted_predicted":"true","input_tokens":376,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.56596554916003e-07,"brier_hard":2.5432211530923146e-13,"soft_accuracy":0.8499997503824561,"brier_soft":0.04499978604235952,"tvd":0.14999964340350866,"kl":1.2434102688303907,"case_id":"typed:300/r0","base_id":"typed:300","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.56510099999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003270017617718451,"1":0.00027249939140040026,"2":0.05766408845678591,"3":0.9417364103900411},"predicted":"3","confidence":0.9417364103900411,"entropy_confidence":0.8370361120641002,"candidate_mass":0.9999879523789749,"accepted":true,"accepted_predicted":"3","input_tokens":405,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.060029862681300686,"brier_hard":0.006719974157860259,"score_mae":0.5241429074750954,"within_one":true,"case_id":"typed:300/r0","base_id":"typed:300","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.56510099999997}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.008341803399715662,"true":0.9916581966002843},"predicted":"true","confidence":0.9916581966002843,"entropy_confidence":0.9304119597012376,"candidate_mass":0.9999046895436347,"accepted":true,"accepted_predicted":"true","input_tokens":340,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.008376790950404013,"brier_hard":0.00013917136791901595,"soft_accuracy":0.8933265572802274,"brier_soft":0.016802450008032738,"tvd":0.0916581966002843,"kl":0.16110372356481506,"case_id":"typed:301/r0","base_id":"typed:301","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3284}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.00019135220952614377,"contain":0.0027098471976256208,"investigate":0.996717974638253,"monitor":0.0003808259545967879},"predicted":"investigate","confidence":0.996717974638253,"entropy_confidence":0.9827376360514507,"candidate_mass":0.9999919082849721,"accepted":true,"accepted_predicted":"investigate","input_tokens":392,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":5.9108630303201934,"brier_hard":1.9880345514876403,"soft_accuracy":0.26770074773674923,"brier_soft":1.0200751481873929,"tvd":0.7300509746382522,"kl":3.676862371074444,"case_id":"typed:301/r0","base_id":"typed:301","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3284}
{"id":"severity","type":"ordinal","probabilities":{"0":0.0001446402303155107,"1":0.0008003901137060534,"2":0.9966238240651402,"3":0.002429681999323641,"4":1.4635915145234335e-06},"predicted":"2","confidence":0.9966238240651402,"entropy_confidence":0.9844649372228781,"candidate_mass":0.9999951138516928,"accepted":true,"accepted_predicted":"2","input_tokens":412,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.0199948946690505,"brier_hard":1.9884062475974664,"score_mae":0.6753240613919846,"within_one":true,"case_id":"typed:301/r0","base_id":"typed:301","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3284}
{"id":"true_positive","type":"binary","probabilities":{"false":8.637836224263423e-07,"true":0.9999991362163776},"predicted":"true","confidence":0.9999991362163776,"entropy_confidence":0.9999813547803759,"candidate_mass":0.9999983577962209,"accepted":true,"accepted_predicted":"true","input_tokens":358,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.637839954680876e-07,"brier_hard":1.4922442927101863e-12,"soft_accuracy":0.9499992225947398,"brier_soft":0.00499982724476777,"tvd":0.049999136216377624,"kl":0.4117892095254322,"case_id":"typed:301/r0","base_id":"typed:301","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3284}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001450030261060483,"1":0.00012645507271814746,"2":0.0017438680376864423,"3":0.9979846738634884},"predicted":"3","confidence":0.9979846738634884,"entropy_confidence":0.9888145024046838,"candidate_mass":0.9999938742938678,"accepted":true,"accepted_predicted":"3","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0020173596388026968,"brier_hard":7.139631932367052e-06,"score_mae":0.13423521273855643,"within_one":true,"case_id":"typed:301/r0","base_id":"typed:301","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3284}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0008657301767613436,"true":0.9991342698232374},"predicted":"true","confidence":0.9991342698232374,"entropy_confidence":0.9899437948810761,"candidate_mass":0.9999618312370417,"accepted":true,"accepted_predicted":"true","input_tokens":320,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.051937272129745,"brier_hard":1.99653857827043,"soft_accuracy":0.20051943810605657,"brier_soft":1.2772311624118395,"tvd":0.799134269823238,"kl":5.14132061519312,"case_id":"typed:302/r0","base_id":"typed:302","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.451601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.3620362884137668e-05,"contain":1.1024584844827152e-05,"investigate":0.9999433269368645,"monitor":3.202811540535874e-05},"predicted":"investigate","confidence":0.9999433269368645,"entropy_confidence":0.999519165068602,"candidate_mass":0.9999948580157647,"accepted":true,"accepted_predicted":"investigate","input_tokens":372,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":5.6674669114223934e-05,"brier_hard":4.544692017676178e-09,"soft_accuracy":0.5166485643856384,"brier_soft":0.33492275964731577,"tvd":0.48327684360334844,"kl":4.022123063864216,"case_id":"typed:302/r0","base_id":"typed:302","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.451601}
{"id":"severity","type":"ordinal","probabilities":{"0":4.170115992352704e-05,"1":0.0002505821116040297,"2":0.9985936712179189,"3":0.0011132559904003213,"4":7.895201528650129e-07},"predicted":"2","confidence":0.9985936712179189,"entropy_confidence":0.9928637156820993,"candidate_mass":0.9999983000855746,"accepted":true,"accepted_predicted":"2","input_tokens":392,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.001407318590509014,"brier_hard":3.2816305482089807e-06,"score_mae":0.03255214940074547,"within_one":true,"case_id":"typed:302/r0","base_id":"typed:302","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.451601}
{"id":"true_positive","type":"binary","probabilities":{"false":4.437774333749711e-06,"true":0.9999955622256659},"predicted":"true","confidence":0.9999955622256659,"entropy_confidence":0.9999146863441365,"candidate_mass":0.9999957012484398,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.325357583383546,"brier_hard":1.9999822489420518,"soft_accuracy":0.4000008875548666,"brier_soft":0.7199893493809861,"tvd":0.5999955622256661,"kl":5.15968970554972,"case_id":"typed:302/r0","base_id":"typed:302","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.451601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00023396672300889386,"1":0.00022065878978822805,"2":0.01153605951762335,"3":0.9880093149695797},"predicted":"3","confidence":0.9880093149695797,"entropy_confidence":0.951518702663242,"candidate_mass":0.9999958777821174,"accepted":true,"accepted_predicted":"3","input_tokens":367,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":8.418892987156227,"brier_hard":1.9758542729870046,"score_mae":1.437320722733774,"within_one":false,"case_id":"typed:302/r0","base_id":"typed:302","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.451601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0038442635831052287,"true":0.9961557364168944},"predicted":"true","confidence":0.9961557364168944,"entropy_confidence":0.9636217511778415,"candidate_mass":0.9999262287343285,"accepted":true,"accepted_predicted":"true","input_tokens":336,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0038516717564292076,"brier_hard":2.9556724992781215e-05,"soft_accuracy":0.6653855762427889,"brier_soft":0.2171256548512036,"tvd":0.3294887364168946,"kl":1.2197763982243819,"case_id":"typed:303/r0","base_id":"typed:303","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.016501}
{"id":"disposition","type":"choice","probabilities":{"close_benign":3.463513212329266e-06,"contain":3.488824139321068e-06,"investigate":0.9999750287329097,"monitor":1.8018929739711302e-05},"predicted":"investigate","confidence":0.9999750287329097,"entropy_confidence":0.9997769598968541,"candidate_mass":0.9999983672381745,"accepted":true,"accepted_predicted":"investigate","input_tokens":388,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.4971578877615133e-05,"brier_hard":9.724138267085686e-10,"soft_accuracy":0.48332757479915156,"brier_soft":0.39607226439593,"tvd":0.5166420287329091,"kl":4.4072840203250925,"case_id":"typed:303/r0","base_id":"typed:303","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.016501}
{"id":"severity","type":"ordinal","probabilities":{"0":5.572505656606362e-06,"1":1.2298255542320706e-05,"2":0.9999294607379695,"3":5.26066125597823e-05,"4":6.188827029812637e-08},"predicted":"2","confidence":0.9999294607379695,"entropy_confidence":0.9995052070331277,"candidate_mass":0.9999994893373614,"accepted":true,"accepted_predicted":"2","input_tokens":408,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":7.054175004123522e-05,"brier_hard":7.925546911656163e-09,"score_mae":0.049970712877757784,"within_one":true,"case_id":"typed:303/r0","base_id":"typed:303","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.016501}
{"id":"true_positive","type":"binary","probabilities":{"false":2.578581278892958e-08,"true":0.9999999742141877},"predicted":"true","confidence":0.9999999742141877,"entropy_confidence":0.9999993127683215,"candidate_mass":0.9999997715735466,"accepted":true,"accepted_predicted":"true","input_tokens":354,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.5785812595655598e-08,"brier_hard":1.329816255258799e-15,"soft_accuracy":0.6166669939832934,"brier_soft":0.2938883382397899,"tvd":0.38333297421418755,"kl":3.2325144933301897,"case_id":"typed:303/r0","base_id":"typed:303","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.016501}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002418591901848246,"1":0.0002445356484411249,"2":0.8379341736576712,"3":0.1615794315037022},"predicted":"2","confidence":0.8379341736576712,"entropy_confidence":0.677754002815316,"candidate_mass":0.999981790158031,"accepted":true,"accepted_predicted":"2","input_tokens":383,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.17681573330655811,"brier_hard":0.0523733630466327,"score_mae":0.12751817747489014,"within_one":true,"case_id":"typed:303/r0","base_id":"typed:303","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.016501}
{"id":"credential_compromise","type":"binary","probabilities":{"false":3.536060052242079e-05,"true":0.9999646393994772},"predicted":"true","confidence":0.9999646393994772,"entropy_confidence":0.9994260915401041,"candidate_mass":0.9999903583849777,"accepted":true,"accepted_predicted":"true","input_tokens":312,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.249912337238907,"brier_hard":1.9998585600986536,"soft_accuracy":0.1400254596323761,"brier_soft":1.4790783620349464,"tvd":0.8599646393994773,"kl":7.645641870578923,"case_id":"typed:304/r0","base_id":"typed:304","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.451601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.025138516755700054,"contain":0.00030029789506533727,"investigate":0.9163846551285465,"monitor":0.05817653022068719},"predicted":"investigate","confidence":0.9163846551285465,"entropy_confidence":0.7543685431878459,"candidate_mass":0.9999792647326305,"accepted":true,"accepted_predicted":"investigate","input_tokens":364,"reused_prefix_tokens":0,"expected":"close_benign","correct":false,"accepted_correct":false,"nll_hard":3.6833540769360624,"brier_hard":1.793500346515686,"soft_accuracy":0.1682371793412583,"brier_soft":0.9555922191225693,"tvd":0.7630516551285469,"kl":2.052693145917227,"case_id":"typed:304/r0","base_id":"typed:304","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.451601}
{"id":"severity","type":"ordinal","probabilities":{"0":2.2114527256548472e-05,"1":0.00011524642164355564,"2":0.9998165624364496,"3":4.599710262997398e-05,"4":7.951202077309888e-08},"predicted":"2","confidence":0.9998165624364496,"entropy_confidence":0.9988031553030989,"candidate_mass":0.9999994061754927,"accepted":true,"accepted_predicted":"2","input_tokens":384,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00018345439027809846,"brier_hard":4.9535869511259276e-08,"score_mae":0.7132196806505158,"within_one":true,"case_id":"typed:304/r0","base_id":"typed:304","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.451601}
{"id":"true_positive","type":"binary","probabilities":{"false":5.459344070706222e-05,"true":0.999945406559293},"predicted":"true","confidence":0.999945406559293,"entropy_confidence":0.9991481475077778,"candidate_mass":0.9999920532712945,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.815596816000319,"brier_hard":1.9997816321980595,"soft_accuracy":0.1733686677509909,"brier_soft":1.3665761413554918,"tvd":0.826612406559293,"kl":7.310120773843886,"case_id":"typed:304/r0","base_id":"typed:304","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.451601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001095806435144173,"1":0.10855392926587314,"2":0.888213412919362,"3":0.002136851379621344},"predicted":"2","confidence":0.888213412919362,"entropy_confidence":0.735305747037097,"candidate_mass":0.9999930572051451,"accepted":true,"accepted_predicted":"2","input_tokens":359,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":6.8162647163236585,"brier_hard":1.7985211765041946,"score_mae":0.8980583092434611,"within_one":true,"case_id":"typed:304/r0","base_id":"typed:304","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.451601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00020520168161099263,"true":0.9997947983183906},"predicted":"true","confidence":0.9997947983183906,"entropy_confidence":0.9971901288788317,"candidate_mass":0.9999703830100358,"accepted":true,"accepted_predicted":"true","input_tokens":320,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00020522273835512711,"brier_hard":8.421546027131452e-08,"soft_accuracy":0.6932536554865792,"brier_soft":0.18783766765708287,"tvd":0.3064617983183898,"kl":1.9878025604309884,"case_id":"typed:305/r0","base_id":"typed:305","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.166901}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.943915247723698e-06,"contain":3.3198151319418857e-06,"investigate":0.9999838675422914,"monitor":9.868727327978846e-06},"predicted":"investigate","confidence":0.9999838675422914,"entropy_confidence":0.9998490543167392,"candidate_mass":0.9999979104903661,"accepted":true,"accepted_predicted":"investigate","input_tokens":372,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.6132587838090685e-05,"brier_hard":3.7733578028966224e-10,"soft_accuracy":0.5666600714465053,"brier_soft":0.2833366703469081,"tvd":0.4333168675422919,"kl":3.6692960889100945,"case_id":"typed:305/r0","base_id":"typed:305","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.166901}
{"id":"severity","type":"ordinal","probabilities":{"0":1.955835173628933e-05,"1":0.00012411580568406235,"2":0.7198771492670875,"3":0.2799772624426313,"4":1.914132861631464e-06},"predicted":"2","confidence":0.7198771492670875,"entropy_confidence":0.6306889637751949,"candidate_mass":0.9999947917434776,"accepted":false,"accepted_predicted":null,"input_tokens":392,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.2730468846722687,"brier_hard":1.036655868427442,"score_mae":0.6768491418008002,"within_one":true,"case_id":"typed:305/r0","base_id":"typed:305","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.166901}
{"id":"true_positive","type":"binary","probabilities":{"false":1.2300555779723407e-08,"true":0.9999999876994449},"predicted":"true","confidence":0.9999999876994449,"entropy_confidence":0.9999996590360207,"candidate_mass":0.9999996643061695,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.2300555205669983e-08,"brier_hard":3.0260732899670283e-16,"soft_accuracy":0.7599999936037115,"brier_soft":0.11519998819146707,"tvd":0.23999998769944453,"kl":2.0019096958893683,"case_id":"typed:305/r0","base_id":"typed:305","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.166901}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.5225277621128076e-05,"1":1.2722590852359751e-05,"2":0.0001526515120954828,"3":0.9998194006194326},"predicted":"3","confidence":0.9998194006194326,"entropy_confidence":0.9986768439969983,"candidate_mass":0.9999971598948257,"accepted":true,"accepted_predicted":"3","input_tokens":367,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00018061569059924022,"brier_hard":5.631229380298804e-08,"score_mae":0.5631092274733414,"within_one":true,"case_id":"typed:305/r0","base_id":"typed:305","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.166901}
{"id":"credential_compromise","type":"binary","probabilities":{"false":1.7221420134591552e-05,"true":0.9999827785798666},"predicted":"true","confidence":0.9999827785798666,"entropy_confidence":0.9997026184681882,"candidate_mass":0.9999941857515314,"accepted":true,"accepted_predicted":"true","input_tokens":304,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7221568423742456e-05,"brier_hard":5.931546228626472e-10,"soft_accuracy":0.7499913892899336,"brier_soft":0.12498277917302064,"tvd":0.249982778579866,"kl":2.086836454831528,"case_id":"typed:306/r0","base_id":"typed:306","repeat":0,"workflow":"security_incidents","case_elapsed_ms":336.208601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.887791183755561e-05,"contain":0.0002534286318704791,"investigate":0.9995495942717677,"monitor":0.00011809918452534495},"predicted":"investigate","confidence":0.9995495942717677,"entropy_confidence":0.9968534121934878,"candidate_mass":0.9999927692284536,"accepted":true,"accepted_predicted":"investigate","input_tokens":356,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":8.280428305653274,"brier_hard":1.9985926185403282,"soft_accuracy":0.3499854798260502,"brier_soft":0.6863507939299688,"tvd":0.6495495942717672,"kl":4.432388832270894,"case_id":"typed:306/r0","base_id":"typed:306","repeat":0,"workflow":"security_incidents","case_elapsed_ms":336.208601}
{"id":"severity","type":"ordinal","probabilities":{"0":1.0895112674350273e-05,"1":5.459846588970513e-05,"2":0.015428923292456214,"3":0.984503260204586,"4":2.3229243934472013e-06},"predicted":"3","confidence":0.984503260204586,"entropy_confidence":0.9500269531475173,"candidate_mass":0.9999987757001809,"accepted":true,"accepted_predicted":"3","input_tokens":376,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.015618069375086958,"brier_hard":0.0004782037233432011,"score_mae":0.09556848263786621,"within_one":true,"case_id":"typed:306/r0","base_id":"typed:306","repeat":0,"workflow":"security_incidents","case_elapsed_ms":336.208601}
{"id":"true_positive","type":"binary","probabilities":{"false":3.8958017677266666e-07,"true":0.9999996104198217},"predicted":"true","confidence":0.9999996104198217,"entropy_confidence":0.9999911431786227,"candidate_mass":0.999998608365049,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.8958025422572114e-07,"brier_hard":3.035454294891371e-13,"soft_accuracy":0.8599997195022714,"brier_soft":0.03919978183540412,"tvd":0.13999961041982245,"kl":1.1597403019439425,"case_id":"typed:306/r0","base_id":"typed:306","repeat":0,"workflow":"security_incidents","case_elapsed_ms":336.208601}
{"id":"urgency","type":"ordinal","probabilities":{"0":5.354087155892381e-06,"1":3.1976904648593675e-06,"2":2.6578096443894787e-05,"3":0.9999648701259337},"predicted":"3","confidence":0.9999648701259337,"entropy_confidence":0.9996966103035296,"candidate_mass":0.9999973329390993,"accepted":true,"accepted_predicted":"3","input_tokens":351,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.5130491134764115e-05,"brier_hard":1.9793947360760605e-09,"score_mae":0.33661796426115353,"within_one":true,"case_id":"typed:306/r0","base_id":"typed:306","repeat":0,"workflow":"security_incidents","case_elapsed_ms":336.208601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0005746021315229786,"true":0.9994253978684774},"predicted":"true","confidence":0.9994253978684774,"entropy_confidence":0.9929855849839682,"candidate_mass":0.9999699958561169,"accepted":true,"accepted_predicted":"true","input_tokens":314,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0005747672785929835,"brier_hard":6.603352191010723e-07,"soft_accuracy":0.5000000000000002,"brier_soft":0.4988514560721735,"tvd":0.49942539786847717,"kl":3.038056554003804,"case_id":"typed:307/r0","base_id":"typed:307","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.5598}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.450614070289812e-08,"contain":4.744265922664282e-08,"investigate":0.9999997527400634,"monitor":1.4531113729464686e-07},"predicted":"investigate","confidence":0.9999997527400634,"entropy_confidence":0.9999969366041476,"candidate_mass":0.999999790940923,"accepted":true,"accepted_predicted":"investigate","input_tokens":366,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.4725996720528896e-07,"brier_hard":8.747452817619183e-14,"soft_accuracy":0.5166664096226353,"brier_soft":0.3566664729019801,"tvd":0.48333326940654636,"kl":4.11048163233499,"case_id":"typed:307/r0","base_id":"typed:307","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.5598}
{"id":"severity","type":"ordinal","probabilities":{"0":6.777503048696705e-06,"1":2.4965293060101318e-05,"2":0.9997467477495606,"3":0.0002214129220229682,"4":9.653230602462708e-08},"predicted":"2","confidence":0.9997467477495606,"entropy_confidence":0.998469450911174,"candidate_mass":0.9999989592301526,"accepted":true,"accepted_predicted":"2","input_tokens":386,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0002532843242058327,"brier_hard":1.1382959411499386e-07,"score_mae":0.27981691431252553,"within_one":true,"case_id":"typed:307/r0","base_id":"typed:307","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.5598}
{"id":"true_positive","type":"binary","probabilities":{"false":3.1074267449504678e-09,"true":0.9999999968925728},"predicted":"true","confidence":0.9999999968925728,"entropy_confidence":0.9999999076959776,"candidate_mass":0.9999998975314756,"accepted":true,"accepted_predicted":"true","input_tokens":332,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.1074272092687344e-09,"brier_hard":1.9312204806131495e-17,"soft_accuracy":0.7499999984462863,"brier_soft":0.12499999689257305,"tvd":0.24999999689257302,"kl":2.0868235409857805,"case_id":"typed:307/r0","base_id":"typed:307","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.5598}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.9223551578006894e-05,"1":1.7236071343480774e-05,"2":0.00020894419087967862,"3":0.9997545961861971},"predicted":"3","confidence":0.9997545961861971,"entropy_confidence":0.9982589122226749,"candidate_mass":0.9999966986321308,"accepted":true,"accepted_predicted":"3","input_tokens":361,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.473443370941965,"brier_hard":1.9990914085379687,"score_mae":0.799698913011694,"within_one":true,"case_id":"typed:307/r0","base_id":"typed:307","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.5598}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0017623150448203542,"true":0.9982376849551811},"predicted":"true","confidence":0.9982376849551811,"entropy_confidence":0.9813375498392409,"candidate_mass":0.9999486586666188,"accepted":true,"accepted_predicted":"true","input_tokens":278,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.341126967893363,"brier_hard":1.9929569513293557,"soft_accuracy":0.46678448649477866,"brier_soft":0.5651347862074393,"tvd":0.5315706849551805,"kl":2.6918320551474904,"case_id":"typed:308/r0","base_id":"typed:308","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.51010099999996}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.87772314175408e-08,"contain":7.445990471230867e-08,"investigate":0.9999995289408216,"monitor":3.1782204293332297e-07},"predicted":"investigate","confidence":0.9999995289408216,"entropy_confidence":0.9999944190190162,"candidate_mass":0.9999996128467771,"accepted":true,"accepted_predicted":"investigate","input_tokens":330,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.710592893902224e-07,"brier_hard":3.3465773016816546e-13,"soft_accuracy":0.5333328671427854,"brier_soft":0.34731059915240703,"tvd":0.4666665289408213,"kl":3.962904086111531,"case_id":"typed:308/r0","base_id":"typed:308","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.51010099999996}
{"id":"severity","type":"ordinal","probabilities":{"0":2.4546741310493675e-05,"1":0.00014503580288096363,"2":0.9362260882474571,"3":0.06360304595974563,"4":1.2832486055730467e-06},"predicted":"2","confidence":0.9362260882474571,"entropy_confidence":0.8518191163569129,"candidate_mass":0.9999959208165871,"accepted":true,"accepted_predicted":"2","input_tokens":350,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0658982843994664,"brier_hard":0.008112480915151994,"score_mae":0.22325551682854572,"within_one":true,"case_id":"typed:308/r0","base_id":"typed:308","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.51010099999996}
{"id":"true_positive","type":"binary","probabilities":{"false":5.3010582375500005e-09,"true":0.9999999946989426},"predicted":"true","confidence":0.9999999946989426,"entropy_confidence":0.9999998466204147,"candidate_mass":0.9999998767227614,"accepted":true,"accepted_predicted":"true","input_tokens":296,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.301057384894014e-09,"brier_hard":5.620242768686996e-17,"soft_accuracy":0.6166669987630833,"brier_soft":0.29388836964971854,"tvd":0.3833329946989422,"kl":3.232514480697917,"case_id":"typed:308/r0","base_id":"typed:308","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.51010099999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.1338095188694899e-05,"1":1.0330918516386388e-05,"2":6.395155809383415e-05,"3":0.9999143794282008},"predicted":"3","confidence":0.9999143794282008,"entropy_confidence":0.9993140452272092,"candidate_mass":0.9999946766995264,"accepted":true,"accepted_predicted":"3","input_tokens":325,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":8.562423744956787e-05,"brier_hard":1.165596437774634e-08,"score_mae":0.9032143723193067,"within_one":true,"case_id":"typed:308/r0","base_id":"typed:308","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.51010099999996}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.004216112429311723,"true":0.995783887570689},"predicted":"true","confidence":0.995783887570689,"entropy_confidence":0.9606656929938977,"candidate_mass":0.9999199763550434,"accepted":true,"accepted_predicted":"true","input_tokens":264,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.46884180080459,"brier_hard":1.9831711014907878,"soft_accuracy":0.417369682594144,"brier_soft":0.6707527389391232,"tvd":0.5791168875706887,"kl":2.512722944653332,"case_id":"typed:309/r0","base_id":"typed:309","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.993601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0015942582422182776,"contain":0.0008068850890872108,"investigate":0.9932919859009367,"monitor":0.004306870767758872},"predicted":"investigate","confidence":0.9932919859009367,"entropy_confidence":0.9667001437838085,"candidate_mass":0.9999836372893974,"accepted":true,"accepted_predicted":"investigate","input_tokens":316,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":5.447543678524276,"brier_hard":1.9780369695782092,"soft_accuracy":0.3335219856139588,"brier_soft":0.7101624176638092,"tvd":0.6599589859009362,"kl":2.7296270124547997,"case_id":"typed:309/r0","base_id":"typed:309","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.993601}
{"id":"severity","type":"ordinal","probabilities":{"0":3.789182289893937e-05,"1":0.00031306492483433834,"2":0.9984319014239093,"3":0.001216689273141967,"4":4.525552158672499e-07},"predicted":"2","confidence":0.9984319014239093,"entropy_confidence":0.9921392692229729,"candidate_mass":0.9999977038061943,"accepted":true,"accepted_predicted":"2","input_tokens":336,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0015693293294595915,"brier_hard":4.038711573926919e-06,"score_mae":0.23416174581294236,"within_one":true,"case_id":"typed:309/r0","base_id":"typed:309","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.993601}
{"id":"true_positive","type":"binary","probabilities":{"false":4.9529899370572595e-06,"true":0.9999950470100621},"predicted":"true","confidence":0.9999950470100621,"entropy_confidence":0.9999055664902573,"candidate_mass":0.9999963013314013,"accepted":true,"accepted_predicted":"true","input_tokens":282,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.215519136046638,"brier_hard":1.9999801880893144,"soft_accuracy":0.38333415570095364,"brier_soft":0.7605441604452812,"tvd":0.6166620470100626,"kl":5.312155551685928,"case_id":"typed:309/r0","base_id":"typed:309","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.993601}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00015135039152068273,"1":0.00010849812166290352,"2":0.0014926142244601184,"3":0.9982475372623576},"predicted":"3","confidence":0.9982475372623576,"entropy_confidence":0.9900559180456499,"candidate_mass":0.9999929014317592,"accepted":true,"accepted_predicted":"3","input_tokens":311,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":9.128777696996252,"brier_hard":1.9962834119830426,"score_mae":1.547836338357656,"within_one":false,"case_id":"typed:309/r0","base_id":"typed:309","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.993601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0005620259533505687,"true":0.9994379740466484},"predicted":"true","confidence":0.9994379740466484,"entropy_confidence":0.9931211591188976,"candidate_mass":0.9999750950594252,"accepted":true,"accepted_predicted":"true","input_tokens":272,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.483962528797772,"brier_hard":1.9977525279329402,"soft_accuracy":0.3834641397637987,"brier_soft":0.759170678090044,"tvd":0.616104974046649,"kl":3.9496559828953903,"case_id":"typed:310/r0","base_id":"typed:310","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.645401}
{"id":"disposition","type":"choice","probabilities":{"close_benign":4.959375714523248e-05,"contain":4.365253826160757e-05,"investigate":0.9998067159154312,"monitor":0.0001000377891630495},"predicted":"investigate","confidence":0.9998067159154312,"entropy_confidence":0.9985252741376397,"candidate_mass":0.9999896568942901,"accepted":true,"accepted_predicted":"investigate","input_tokens":324,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00019330276634472923,"brier_hard":5.173138145267444e-08,"soft_accuracy":0.5332634934466823,"brier_soft":0.29919745222487926,"tvd":0.46647371591543074,"kl":3.3118139207305943,"case_id":"typed:310/r0","base_id":"typed:310","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.645401}
{"id":"severity","type":"ordinal","probabilities":{"0":5.35258969074989e-05,"1":0.0003844390262893733,"2":0.9964874267587162,"3":0.003073481770301594,"4":1.1265477869310604e-06},"predicted":"2","confidence":0.9964874267587162,"entropy_confidence":0.9845590093614114,"candidate_mass":0.9999970435273872,"accepted":true,"accepted_predicted":"2","input_tokens":344,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.784944232885298,"brier_hard":1.9868498250974524,"score_mae":0.5274157559542254,"within_one":true,"case_id":"typed:310/r0","base_id":"typed:310","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.645401}
{"id":"true_positive","type":"binary","probabilities":{"false":4.958825242781499e-07,"true":0.9999995041174742},"predicted":"true","confidence":0.9999995041174742,"entropy_confidence":0.9999888990781217,"candidate_mass":0.9999987775758694,"accepted":true,"accepted_predicted":"true","input_tokens":290,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.516926784487385,"brier_hard":1.9999980164703917,"soft_accuracy":0.45000004958825174,"brier_soft":0.6049989090589367,"tvd":0.549999504117475,"kl":4.706358964436096,"case_id":"typed:310/r0","base_id":"typed:310","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.645401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003521396808076836,"1":0.00032647204460026024,"2":0.005381538903584732,"3":0.9939398493710082},"predicted":"3","confidence":0.9939398493710082,"entropy_confidence":0.9714492163784957,"candidate_mass":0.9999914205466116,"accepted":true,"accepted_predicted":"3","input_tokens":319,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.224780904152031,"brier_hard":1.9771825379078145,"score_mae":0.8095760979647939,"within_one":true,"case_id":"typed:310/r0","base_id":"typed:310","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.645401}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.075049875619926e-05,"true":0.999949249501244},"predicted":"true","confidence":0.999949249501244,"entropy_confidence":0.9992027666621542,"candidate_mass":0.9999848139250523,"accepted":true,"accepted_predicted":"true","input_tokens":268,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.88858911238231,"brier_hard":1.9997970031562018,"soft_accuracy":0.4000101500997513,"brier_soft":0.7198782039542115,"tvd":0.5999492495012438,"kl":5.15970823115069,"case_id":"typed:311/r0","base_id":"typed:311","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.454001}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.5377912536057343e-05,"contain":2.494670334622076e-05,"investigate":0.9998734823384666,"monitor":7.619304565082689e-05},"predicted":"investigate","confidence":0.9998734823384666,"entropy_confidence":0.9990031449832792,"candidate_mass":0.9999928637506814,"accepted":true,"accepted_predicted":"investigate","input_tokens":320,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00012652566556782449,"brier_hard":2.307847533795507e-08,"soft_accuracy":0.5832807318661914,"brier_soft":0.2409630795790257,"tvd":0.41654048233846674,"kl":3.0693779928002853,"case_id":"typed:311/r0","base_id":"typed:311","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.454001}
{"id":"severity","type":"ordinal","probabilities":{"0":6.631818642890985e-05,"1":0.000335706387362931,"2":0.8235396171374819,"3":0.17605483246142262,"4":3.5258273030572283e-06},"predicted":"2","confidence":0.8235396171374819,"entropy_confidence":0.7085610756861218,"candidate_mass":0.9999945591776837,"accepted":true,"accepted_predicted":"2","input_tokens":340,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":1.7369597843887448,"brier_hard":1.3571032572144364,"score_mae":0.5810734586441932,"within_one":true,"case_id":"typed:311/r0","base_id":"typed:311","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.454001}
{"id":"true_positive","type":"binary","probabilities":{"false":3.489966089199527e-07,"true":0.9999996510033929},"predicted":"true","confidence":0.9999996510033929,"entropy_confidence":0.9999920104281608,"candidate_mass":0.9999992549317587,"accepted":true,"accepted_predicted":"true","input_tokens":286,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.86820363135293,"brier_hard":1.9999986040138116,"soft_accuracy":0.46666702326620885,"brier_soft":0.5688874332526118,"tvd":0.533332651003392,"kl":4.556513346137286,"case_id":"typed:311/r0","base_id":"typed:311","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.454001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0003657325984741284,"1":0.0003646932892690263,"2":0.011825808964237756,"3":0.9874437651480185},"predicted":"3","confidence":0.9874437651480185,"entropy_confidence":0.9489754117761211,"candidate_mass":0.9999943690659394,"accepted":true,"accepted_predicted":"3","input_tokens":315,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.012635730513935102,"brier_hard":0.00029777555284558644,"score_mae":0.8030146066617996,"within_one":true,"case_id":"typed:311/r0","base_id":"typed:311","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.454001}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0003688031638603893,"true":0.9996311968361383},"predicted":"true","confidence":0.9996311968361383,"entropy_confidence":0.9952618786828837,"candidate_mass":0.9999796095046759,"accepted":true,"accepted_predicted":"true","input_tokens":313,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.905247487521482,"brier_hard":1.9985250593761033,"soft_accuracy":0.45003688031638545,"brier_soft":0.6041889050710532,"tvd":0.5496311968361389,"kl":3.6599132964580403,"case_id":"typed:312/r0","base_id":"typed:312","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.39}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0002813132159501452,"contain":9.190034398953081e-05,"investigate":0.9913163124998358,"monitor":0.008310473940223866},"predicted":"investigate","confidence":0.9913163124998358,"entropy_confidence":0.9627717936534299,"candidate_mass":0.9999870461057954,"accepted":true,"accepted_predicted":"investigate","input_tokens":365,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.008721610414564663,"brier_hard":0.00014455798851034223,"soft_accuracy":0.5312322600627067,"brier_soft":0.31086794064076834,"tvd":0.45798331249983615,"kl":1.7634500847787422,"case_id":"typed:312/r0","base_id":"typed:312","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.39}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00036347557235029746,"1":0.6593186389693336,"2":0.3395474305569012,"3":0.0007695643896588597,"4":8.905117577467872e-07},"predicted":"1","confidence":0.6593186389693336,"entropy_confidence":0.5962528332131032,"candidate_mass":0.9999947688660464,"accepted":false,"accepted_predicted":null,"input_tokens":385,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.4165483424156377,"brier_hard":0.23135697169613542,"score_mae":0.04072575529914357,"within_one":true,"case_id":"typed:312/r0","base_id":"typed:312","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.39}
{"id":"true_positive","type":"binary","probabilities":{"false":2.5961687112222415e-06,"true":0.9999974038312892},"predicted":"true","confidence":0.9999974038312892,"entropy_confidence":0.9999480821358898,"candidate_mass":0.9999972882352421,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.596172080840758e-06,"brier_hard":1.3480183952008848e-11,"soft_accuracy":0.5333328269238169,"brier_soft":0.43555133160642484,"tvd":0.46666440383128904,"kl":3.9629052195043117,"case_id":"typed:312/r0","base_id":"typed:312","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.39}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.13502277020578554,"1":0.8559343607253612,"2":0.00667855219375658,"3":0.0023643168750959466},"predicted":"1","confidence":0.8559343607253612,"entropy_confidence":0.6744863433858618,"candidate_mass":0.9999872392760926,"accepted":true,"accepted_predicted":"1","input_tokens":360,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":true,"nll_hard":0.15556158717616952,"brier_hard":0.03903624994734531,"score_mae":0.5402825842618377,"within_one":true,"case_id":"typed:312/r0","base_id":"typed:312","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.39}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00034784634908969434,"true":0.9996521536509116},"predicted":"true","confidence":0.9996521536509116,"entropy_confidence":0.9955017531779152,"candidate_mass":0.9999740124115113,"accepted":true,"accepted_predicted":"true","input_tokens":296,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.9637497012952565,"brier_hard":1.998608856597809,"soft_accuracy":0.21686411269925354,"brier_soft":1.2261315016754812,"tvd":0.7829851536509109,"kl":5.7156863162024045,"case_id":"typed:313/r0","base_id":"typed:313","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.649401}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.02413616678321626,"contain":0.0002850237240564785,"investigate":0.7879291735831203,"monitor":0.18764963590960687},"predicted":"investigate","confidence":0.7879291735831203,"entropy_confidence":0.5715320082030018,"candidate_mass":0.999974324286433,"accepted":false,"accepted_predicted":null,"input_tokens":348,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":null,"nll_hard":1.673178693666209,"brier_hard":1.2813281324066836,"soft_accuracy":0.2982728116431055,"brier_soft":0.3995259364956865,"tvd":0.5212621735831203,"kl":0.8393959911803943,"case_id":"typed:313/r0","base_id":"typed:313","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.649401}
{"id":"severity","type":"ordinal","probabilities":{"0":3.4904974101634265e-05,"1":0.0002900018827945041,"2":0.9995519379633746,"3":0.0001230029142632455,"4":1.5226546570593358e-07},"predicted":"2","confidence":0.9995519379633746,"entropy_confidence":0.9973417701205635,"candidate_mass":0.9999990341263302,"accepted":true,"accepted_predicted":"2","input_tokens":368,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00044816244641405844,"brier_hard":3.012087780083143e-07,"score_mae":0.4164304956141962,"within_one":true,"case_id":"typed:313/r0","base_id":"typed:313","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.649401}
{"id":"true_positive","type":"binary","probabilities":{"false":1.066298115028297e-06,"true":0.9999989337018854},"predicted":"true","confidence":0.9999989337018854,"entropy_confidence":0.9999773074145898,"candidate_mass":0.999997814365835,"accepted":true,"accepted_predicted":"true","input_tokens":314,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.751317613692482,"brier_hard":1.9999957348098147,"soft_accuracy":0.26666749760507624,"brier_soft":1.0755514499738923,"tvd":0.7333319337018852,"kl":6.401778489102058,"case_id":"typed:313/r0","base_id":"typed:313","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.649401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008522384462157491,"1":0.0012588903114981993,"2":0.9569021724024772,"3":0.04098669883981048},"predicted":"2","confidence":0.9569021724024772,"entropy_confidence":0.8647347936482183,"candidate_mass":0.9999852553300754,"accepted":true,"accepted_predicted":"2","input_tokens":343,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":6.677524651227554,"brier_hard":1.9148262075225548,"score_mae":0.8213563316358841,"within_one":true,"case_id":"typed:313/r0","base_id":"typed:313","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.649401}
{"id":"credential_compromise","type":"binary","probabilities":{"false":3.896814283679254e-05,"true":0.9999610318571637},"predicted":"true","confidence":0.9999610318571637,"entropy_confidence":0.9993730021052183,"candidate_mass":0.9999858358210808,"accepted":true,"accepted_predicted":"true","input_tokens":306,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.896890211412965e-05,"brier_hard":3.037032312261005e-09,"soft_accuracy":0.7999766191142984,"brier_soft":0.07996882852276302,"tvd":0.1999610318571634,"kl":1.5301819707789648,"case_id":"typed:314/r0","base_id":"typed:314","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.43130199999996}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.2071750217253546e-05,"contain":1.8717109512651782e-05,"investigate":0.9999538782977687,"monitor":1.5332842503025652e-05},"predicted":"investigate","confidence":0.9999538782977687,"entropy_confidence":0.9995985288887237,"candidate_mass":0.9999937896696274,"accepted":true,"accepted_predicted":"investigate","input_tokens":358,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.6122765869696224e-05,"brier_hard":2.8583648177501047e-09,"soft_accuracy":0.44998909251553776,"brier_soft":0.5104851522008267,"tvd":0.5499538782977679,"kl":4.6079579995639435,"case_id":"typed:314/r0","base_id":"typed:314","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.43130199999996}
{"id":"severity","type":"ordinal","probabilities":{"0":8.956179735660639e-06,"1":0.00011724302337060051,"2":0.18868565234910262,"3":0.8111852851436296,"4":2.8633041618858075e-06},"predicted":"3","confidence":0.8111852851436296,"entropy_confidence":0.6982701670559941,"candidate_mass":0.9999961669047928,"accepted":true,"accepted_predicted":"3","input_tokens":378,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.20925878592558433,"brier_hard":0.07125328578303705,"score_mae":0.10227714363088802,"within_one":true,"case_id":"typed:314/r0","base_id":"typed:314","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.43130199999996}
{"id":"true_positive","type":"binary","probabilities":{"false":5.566437906338426e-09,"true":0.999999994433562},"predicted":"true","confidence":0.999999994433562,"entropy_confidence":0.999999839334269,"candidate_mass":0.999999714683998,"accepted":true,"accepted_predicted":"true","input_tokens":324,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.566437992715065e-09,"brier_hard":6.197046271938567e-17,"soft_accuracy":0.8699999958808359,"brier_soft":0.03379999710545234,"tvd":0.12999999443356208,"kl":1.0761862546195533,"case_id":"typed:314/r0","base_id":"typed:314","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.43130199999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.8863223465775654e-06,"1":2.442620869972905e-06,"2":1.4998226095585399e-05,"3":0.9999786728306862},"predicted":"3","confidence":0.9999786728306862,"entropy_confidence":0.9998067498999089,"candidate_mass":0.9999980032978545,"accepted":true,"accepted_predicted":"3","input_tokens":353,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.1327396741109693e-05,"brier_hard":7.008648350497535e-10,"score_mae":0.2899684575651196,"within_one":true,"case_id":"typed:314/r0","base_id":"typed:314","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.43130199999996}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.010877268365877116,"true":0.9891227316341223},"predicted":"true","confidence":0.9891227316341223,"entropy_confidence":0.9134456593505073,"candidate_mass":0.9999343964078907,"accepted":true,"accepted_predicted":"true","input_tokens":291,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.010936858360985048,"brier_hard":0.0002366299342066338,"soft_accuracy":0.5619552199251674,"brier_soft":0.36259379113011636,"tvd":0.42578973163412265,"kl":1.2952641416696,"case_id":"typed:315/r0","base_id":"typed:315","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.496801}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.076485886547867e-08,"contain":5.611075828886579e-08,"investigate":0.9999995393802013,"monitor":3.437441802532389e-07},"predicted":"investigate","confidence":0.9999995393802013,"entropy_confidence":0.9999945731605244,"candidate_mass":0.9999996933795722,"accepted":true,"accepted_predicted":"investigate","input_tokens":343,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.6061990483311765e-07,"brier_hard":3.371714457251148e-13,"soft_accuracy":0.6333334209617341,"brier_soft":0.1897992164377744,"tvd":0.36666590604656857,"kl":3.087842173071231,"case_id":"typed:315/r0","base_id":"typed:315","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.496801}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00016416107495660894,"1":0.0016729903984859297,"2":0.997301039848502,"3":0.0008613722131415737,"4":4.3646491243567426e-07},"predicted":"2","confidence":0.997301039848502,"entropy_confidence":0.9870099550422005,"candidate_mass":0.9999978838564311,"accepted":true,"accepted_predicted":"2","input_tokens":363,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":6.3931425960937815,"brier_hard":1.9912669510939436,"score_mae":0.4555279325945645,"within_one":true,"case_id":"typed:315/r0","base_id":"typed:315","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.496801}
{"id":"true_positive","type":"binary","probabilities":{"false":4.949418810624018e-08,"true":0.9999999505058111},"predicted":"true","confidence":0.9999999505058111,"entropy_confidence":0.999998727461826,"candidate_mass":0.9999995319549709,"accepted":true,"accepted_predicted":"true","input_tokens":309,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.949419008330509e-08,"brier_hard":4.899349387053553e-15,"soft_accuracy":0.7033329798723958,"brier_soft":0.17602255904483527,"tvd":0.29666695050581154,"kl":2.484883745955454,"case_id":"typed:315/r0","base_id":"typed:315","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.496801}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022046326223869463,"1":0.00012064616869446542,"2":0.0016945284504426433,"3":0.997964362118623},"predicted":"3","confidence":0.997964362118623,"entropy_confidence":0.988609886218407,"candidate_mass":0.9999913320477282,"accepted":true,"accepted_predicted":"3","input_tokens":338,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.380350777183438,"brier_hard":1.9925467457441624,"score_mae":1.1574027894254486,"within_one":false,"case_id":"typed:315/r0","base_id":"typed:315","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.496801}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.000510880855049213,"true":0.9994891191449504},"predicted":"true","confidence":0.9994891191449504,"entropy_confidence":0.99367680111046,"candidate_mass":0.9999662103064634,"accepted":true,"accepted_predicted":"true","input_tokens":345,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.5793741553200356,"brier_hard":1.9979569985782986,"soft_accuracy":0.42674192885148654,"brier_soft":0.656250360363424,"tvd":0.5728221191449505,"kl":3.6633705271274426,"case_id":"typed:316/r0","base_id":"typed:316","repeat":0,"workflow":"security_incidents","case_elapsed_ms":315.713801}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0053561860989206e-07,"contain":8.583232198729007e-08,"investigate":0.9999994840460696,"monitor":3.2958598902882203e-07},"predicted":"investigate","confidence":0.9999994840460696,"entropy_confidence":0.9999939034317675,"candidate_mass":0.9999996514028954,"accepted":true,"accepted_predicted":"investigate","input_tokens":397,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":5.159540634621346e-07,"brier_hard":3.923099805228335e-13,"soft_accuracy":0.5499998139311522,"brier_soft":0.2838881180082274,"tvd":0.44999948404607,"kl":3.815843100748426,"case_id":"typed:316/r0","base_id":"typed:316","repeat":0,"workflow":"security_incidents","case_elapsed_ms":315.713801}
{"id":"severity","type":"ordinal","probabilities":{"0":9.905333123898101e-06,"1":3.250337338222656e-05,"2":0.9998854148029986,"3":7.209318230269226e-05,"4":8.330819317422689e-08},"predicted":"2","confidence":0.9998854148029986,"entropy_confidence":0.9992211206958737,"candidate_mass":0.9999991691634028,"accepted":true,"accepted_predicted":"2","input_tokens":417,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001145917623865814,"brier_hard":1.9481786152144587e-08,"score_mae":0.0999800542409397,"within_one":true,"case_id":"typed:316/r0","base_id":"typed:316","repeat":0,"workflow":"security_incidents","case_elapsed_ms":315.713801}
{"id":"true_positive","type":"binary","probabilities":{"false":2.3334301211379226e-09,"true":0.9999999976665706},"predicted":"true","confidence":0.9999999976665706,"entropy_confidence":0.9999999297226961,"candidate_mass":0.9999998833367126,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.3334294523859732e-09,"brier_hard":1.0889789126790769e-17,"soft_accuracy":0.5666669996888748,"brier_soft":0.3755549737333914,"tvd":0.4333329976665702,"kl":3.6692869484272195,"case_id":"typed:316/r0","base_id":"typed:316","repeat":0,"workflow":"security_incidents","case_elapsed_ms":315.713801}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.483166394297578e-05,"1":7.352698134546114e-05,"2":0.006710829532019561,"3":0.9931508118226915},"predicted":"3","confidence":0.9931508118226915,"entropy_confidence":0.9698967780806738,"candidate_mass":0.9999948073076719,"accepted":true,"accepted_predicted":"3","input_tokens":392,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.004032709407216,"brier_hard":1.9729719208024015,"score_mae":1.276280621513459,"within_one":false,"case_id":"typed:316/r0","base_id":"typed:316","repeat":0,"workflow":"security_incidents","case_elapsed_ms":315.713801}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00010400316460037406,"true":0.9998959968353982},"predicted":"true","confidence":0.9998959968353982,"entropy_confidence":0.9984738882468968,"candidate_mass":0.9999767406312788,"accepted":true,"accepted_predicted":"true","input_tokens":326,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00010400857330593757,"brier_hard":2.1633316494080734e-08,"soft_accuracy":0.5999791993670791,"brier_soft":0.3198336165699548,"tvd":0.39989599683539895,"kl":2.995486430309167,"case_id":"typed:317/r0","base_id":"typed:317","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.205001}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.006486535627989268,"contain":0.0005681867031363599,"investigate":0.9890682577448372,"monitor":0.0038770199240376526},"predicted":"investigate","confidence":0.9890682577448372,"entropy_confidence":0.9499925189853364,"candidate_mass":0.9999732515598955,"accepted":true,"accepted_predicted":"investigate","input_tokens":378,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.01099193280990468,"brier_hard":0.0001769322528074919,"soft_accuracy":0.6605262476547169,"brier_soft":0.1511502302110482,"tvd":0.32240125774483697,"kl":1.0434604103627487,"case_id":"typed:317/r0","base_id":"typed:317","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.205001}
{"id":"severity","type":"ordinal","probabilities":{"0":1.3478530139159245e-05,"1":8.952537400501322e-05,"2":0.9985573693402953,"3":0.0013394253978111924,"4":2.0135775019658175e-07},"predicted":"2","confidence":0.9985573693402953,"entropy_confidence":0.9929843201447521,"candidate_mass":0.9999981732508788,"accepted":true,"accepted_predicted":"2","input_tokens":398,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.615514564587176,"brier_hard":1.9944397713250888,"score_mae":0.6987766543209704,"within_one":true,"case_id":"typed:317/r0","base_id":"typed:317","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.205001}
{"id":"true_positive","type":"binary","probabilities":{"false":2.948296954271021e-05,"true":0.9999705170304581},"predicted":"true","confidence":0.9999705170304581,"entropy_confidence":0.9995137541176672,"candidate_mass":0.9999899618671875,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.9483404173150144e-05,"brier_hard":1.7384909860626696e-09,"soft_accuracy":0.6499911551091377,"brier_soft":0.24495872558113177,"tvd":0.3499705170304577,"kl":2.943629398944281,"case_id":"typed:317/r0","base_id":"typed:317","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.205001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.001173900715953909,"1":0.0019742105058816995,"2":0.10614831415151015,"3":0.8907035746266541},"predicted":"3","confidence":0.8907035746266541,"entropy_confidence":0.7393118101666788,"candidate_mass":0.9999866432955002,"accepted":true,"accepted_predicted":"3","input_tokens":373,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.11574359527443438,"brier_hard":0.02321844874661152,"score_mae":0.6697145626888639,"within_one":true,"case_id":"typed:317/r0","base_id":"typed:317","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.205001}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.6206080505309535e-05,"true":0.9999437939194946},"predicted":"true","confidence":0.9999437939194946,"entropy_confidence":0.9991253452136306,"candidate_mass":0.9999892339607707,"accepted":true,"accepted_predicted":"true","input_tokens":289,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.786485612872529,"brier_hard":1.9997751819962257,"soft_accuracy":0.40001124121610104,"brier_soft":0.7198651117250341,"tvd":0.5999437939194947,"kl":5.159710413500099,"case_id":"typed:318/r0","base_id":"typed:318","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.179601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.753228619807239e-06,"contain":2.6723254703529617e-06,"investigate":0.9999885469527001,"monitor":6.027493211186869e-06},"predicted":"investigate","confidence":0.9999885469527001,"entropy_confidence":0.9998893163321636,"candidate_mass":0.9999987206582664,"accepted":true,"accepted_predicted":"investigate","input_tokens":341,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.1453112886527842e-05,"brier_hard":1.8222455811663438e-10,"soft_accuracy":0.5999949880634776,"brier_soft":0.22887589573866957,"tvd":0.3999885469526994,"kl":3.377647646398611,"case_id":"typed:318/r0","base_id":"typed:318","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.179601}
{"id":"severity","type":"ordinal","probabilities":{"0":8.677983648198886e-05,"1":0.0005143097273279744,"2":0.9668019355972252,"3":0.03259597487981329,"4":9.999591510465261e-07},"predicted":"2","confidence":0.9668019355972252,"entropy_confidence":0.9074488116782786,"candidate_mass":0.9999973839238592,"accepted":true,"accepted_predicted":"2","input_tokens":361,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.03376162808900318,"brier_hard":0.0021648811046917674,"score_mae":0.2514228946021775,"within_one":true,"case_id":"typed:318/r0","base_id":"typed:318","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.179601}
{"id":"true_positive","type":"binary","probabilities":{"false":2.1024684987875997e-07,"true":0.9999997897531515},"predicted":"true","confidence":0.9999997897531515,"entropy_confidence":0.9999950331058466,"candidate_mass":0.9999992095313954,"accepted":true,"accepted_predicted":"true","input_tokens":307,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.374983520946333,"brier_hard":1.9999991590126915,"soft_accuracy":0.4833330070083692,"brier_soft":0.533889143267653,"tvd":0.5166667897531507,"kl":4.407272052356209,"case_id":"typed:318/r0","base_id":"typed:318","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.179601}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.180173682867366e-05,"1":1.5511967612591256e-05,"2":0.0002014888470156404,"3":0.9997611974485446},"predicted":"3","confidence":0.9997611974485446,"entropy_confidence":0.9982982067014388,"candidate_mass":0.9999952538740019,"accepted":true,"accepted_predicted":"3","input_tokens":336,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.509776527907704,"brier_hard":1.999119515543409,"score_mae":1.183035082007278,"within_one":false,"case_id":"typed:318/r0","base_id":"typed:318","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.179601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0032756047088889485,"true":0.9967243952911096},"predicted":"true","confidence":0.9967243952911096,"entropy_confidence":0.9682451328915498,"candidate_mass":0.9999160715304166,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.721252783003958,"brier_hard":1.986919040336859,"soft_accuracy":0.2516378023544441,"brier_soft":1.1151946450457486,"tvd":0.7467243952911102,"kl":3.7294246879456967,"case_id":"typed:319/r0","base_id":"typed:319","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.279002}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.14445318658781e-05,"contain":0.00010737139922845517,"investigate":0.9997126955801925,"monitor":0.00010848848871453735},"predicted":"investigate","confidence":0.9997126955801925,"entropy_confidence":0.997878529815301,"candidate_mass":0.9999926903721733,"accepted":true,"accepted_predicted":"investigate","input_tokens":383,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00028734569962910356,"brier_hard":1.109465203303052e-07,"soft_accuracy":0.5832061759077454,"brier_soft":0.23690172806941456,"tvd":0.41637969558019183,"kl":2.7364225720893787,"case_id":"typed:319/r0","base_id":"typed:319","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.279002}
{"id":"severity","type":"ordinal","probabilities":{"0":8.43385298399081e-06,"1":5.6091125878594056e-05,"2":0.9388454900108961,"3":0.06108947968943968,"4":5.053208032583798e-07},"predicted":"2","confidence":0.9388454900108961,"entropy_confidence":0.8566762727322664,"candidate_mass":0.9999962423089039,"accepted":true,"accepted_predicted":"2","input_tokens":403,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.7954156094718137,"brier_hard":1.7629838224812464,"score_mae":0.6223154685007972,"within_one":true,"case_id":"typed:319/r0","base_id":"typed:319","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.279002}
{"id":"true_positive","type":"binary","probabilities":{"false":1.4821298887747636e-07,"true":0.99999985178701},"predicted":"true","confidence":0.99999985178701,"entropy_confidence":0.9999964238399747,"candidate_mass":0.9999986055363467,"accepted":true,"accepted_predicted":"true","input_tokens":349,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.724615483677479,"brier_hard":1.9999994071480862,"soft_accuracy":0.36000004149963644,"brier_soft":0.819199620574791,"tvd":0.6399998517870106,"kl":5.526823442309924,"case_id":"typed:319/r0","base_id":"typed:319","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.279002}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006133522060543869,"1":0.0043982814459102015,"2":0.9896407937795892,"3":0.005347572568447127},"predicted":"2","confidence":0.9896407937795892,"entropy_confidence":0.9518981835058377,"candidate_mass":0.9999928806843075,"accepted":true,"accepted_predicted":"2","input_tokens":378,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.231112546562695,"brier_hard":1.9687420731887824,"score_mae":0.00027741328956998323,"within_one":true,"case_id":"typed:319/r0","base_id":"typed:319","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.279002}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0008202759533326524,"true":0.9991797240466687},"predicted":"true","confidence":0.9991797240466687,"entropy_confidence":0.9904079339835705,"candidate_mass":0.9999504386367021,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.105869745864521,"brier_hard":1.9967202418919512,"soft_accuracy":0.4833603430786291,"brier_soft":0.5321956854193585,"tvd":0.515846724046668,"kl":2.9791735332715557,"case_id":"typed:320/r0","base_id":"typed:320","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.75840100000005}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.523533264325846e-05,"contain":4.5859673092249406e-05,"investigate":0.9992926458521384,"monitor":0.0005962591421242066},"predicted":"investigate","confidence":0.9992926458521384,"entropy_confidence":0.9955124461735256,"candidate_mass":0.9999916015592298,"accepted":true,"accepted_predicted":"investigate","input_tokens":396,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0007076044408443335,"brier_hard":8.622336133049171e-07,"soft_accuracy":0.6663217655515024,"brier_soft":0.15983190061288538,"tvd":0.33262564585213933,"kl":1.850745480903257,"case_id":"typed:320/r0","base_id":"typed:320","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.75840100000005}
{"id":"severity","type":"ordinal","probabilities":{"0":1.2218386088096202e-05,"1":6.452372488861757e-05,"2":0.996605904328402,"3":0.0033169526871531995,"4":4.008734683926033e-07},"predicted":"2","confidence":0.996605904328402,"entropy_confidence":0.985653048360185,"candidate_mass":0.9999979728556502,"accepted":true,"accepted_predicted":"2","input_tokens":416,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.708708782865347,"brier_hard":1.9866004296558153,"score_mae":0.663438206062974,"within_one":true,"case_id":"typed:320/r0","base_id":"typed:320","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.75840100000005}
{"id":"true_positive","type":"binary","probabilities":{"false":3.043401182298405e-06,"true":0.9999969565988194},"predicted":"true","confidence":0.9999969565988194,"entropy_confidence":0.9999398362897174,"candidate_mass":0.9999979977823805,"accepted":true,"accepted_predicted":"true","input_tokens":362,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.043405811779397e-06,"brier_hard":1.8524581502642703e-11,"soft_accuracy":0.6166662898710294,"brier_soft":0.2938837112521043,"tvd":0.3833299565988186,"kl":3.2325163541968616,"case_id":"typed:320/r0","base_id":"typed:320","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.75840100000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006323620494726428,"1":0.0007896998607309839,"2":0.785812534551643,"3":0.21276540353815357},"predicted":"2","confidence":0.785812534551643,"entropy_confidence":0.6184232184751136,"candidate_mass":0.9999852369400193,"accepted":false,"accepted_predicted":null,"input_tokens":391,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.24103702065941585,"brier_hard":0.0911464108055761,"score_mae":0.005956020421522545,"within_one":true,"case_id":"typed:320/r0","base_id":"typed:320","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.75840100000005}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0014646357415256592,"true":0.9985353642584737},"predicted":"true","confidence":0.9985353642584737,"entropy_confidence":0.9840986249385713,"candidate_mass":0.9999600056014476,"accepted":true,"accepted_predicted":"true","input_tokens":262,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0014657093688974076,"brier_hard":4.290315710710807e-06,"soft_accuracy":0.49999999999999967,"brier_soft":0.4970750188326587,"tvd":0.498535364258474,"kl":2.5706600281375653,"case_id":"typed:321/r0","base_id":"typed:321","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.268401}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0010221776166301855,"contain":0.00018762240416072827,"investigate":0.9981900384316652,"monitor":0.0006001615475445915},"predicted":"investigate","confidence":0.9981900384316652,"entropy_confidence":0.9892453996640733,"candidate_mass":0.9999843849020904,"accepted":true,"accepted_predicted":"investigate","input_tokens":314,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0018116015279155198,"brier_hard":4.716204008482887e-06,"soft_accuracy":0.5492616319946549,"brier_soft":0.2834171068560291,"tvd":0.44819003843166483,"kl":2.2974090344600175,"case_id":"typed:321/r0","base_id":"typed:321","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.268401}
{"id":"severity","type":"ordinal","probabilities":{"0":1.722043122724508e-05,"1":6.141038812646949e-05,"2":0.9995992773764365,"3":0.0003217520844248499,"4":3.3971978452508746e-07},"predicted":"2","confidence":0.9995992773764365,"entropy_confidence":0.9976528467529028,"candidate_mass":0.9999988052997639,"accepted":true,"accepted_predicted":"2","input_tokens":334,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.000400802934329627,"brier_hard":2.6817091929837704e-07,"score_mae":0.19977341972658813,"within_one":true,"case_id":"typed:321/r0","base_id":"typed:321","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.268401}
{"id":"true_positive","type":"binary","probabilities":{"false":4.1229811417193305e-05,"true":0.9999587701885839},"predicted":"true","confidence":0.9999587701885839,"entropy_confidence":0.9993399677024623,"candidate_mass":0.9999920443373884,"accepted":true,"accepted_predicted":"true","input_tokens":280,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.123066138811453e-05,"brier_hard":3.3997946989025933e-09,"soft_accuracy":0.6166573796831835,"brier_soft":0.29382516218859567,"tvd":0.3832917701885834,"kl":3.20461692961298,"case_id":"typed:321/r0","base_id":"typed:321","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.268401}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00039585102610277154,"1":0.0004178995017623524,"2":0.9233189589049879,"3":0.07586729056714707},"predicted":"2","confidence":0.9233189589049879,"entropy_confidence":0.801153484991858,"candidate_mass":0.9999771564924969,"accepted":true,"accepted_predicted":"2","input_tokens":309,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.5787696417875914,"brier_hard":1.706539495855125,"score_mae":0.1086753109868206,"within_one":true,"case_id":"typed:321/r0","base_id":"typed:321","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.268401}
{"id":"credential_compromise","type":"binary","probabilities":{"false":8.289808419956303e-05,"true":0.9999171019158001},"predicted":"true","confidence":0.9999171019158001,"entropy_confidence":0.9987564510652839,"candidate_mass":0.9999840755340685,"accepted":true,"accepted_predicted":"true","input_tokens":256,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.39789860586614,"brier_hard":1.9996684214073857,"soft_accuracy":0.4166808162921011,"brier_soft":0.6803613627695828,"tvd":0.5832501019158003,"kl":4.802945551634818,"case_id":"typed:322/r0","base_id":"typed:322","repeat":0,"workflow":"security_incidents","case_elapsed_ms":292.874001}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.671282146372198e-05,"contain":1.5383260605483497e-05,"investigate":0.9999312923429138,"monitor":3.6611575018302775e-05},"predicted":"investigate","confidence":0.9999312923429138,"entropy_confidence":0.9994250797951045,"candidate_mass":0.9999950366884995,"accepted":true,"accepted_predicted":"investigate","input_tokens":308,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":6.871001756541371e-05,"brier_hard":6.5771126757328695e-09,"soft_accuracy":0.5333105473849831,"brier_soft":0.3215751398268722,"tvd":0.46659775900937983,"kl":3.7941082930105656,"case_id":"typed:322/r0","base_id":"typed:322","repeat":0,"workflow":"security_incidents","case_elapsed_ms":292.874001}
{"id":"severity","type":"ordinal","probabilities":{"0":5.562766862666082e-06,"1":6.368265551068281e-05,"2":0.02840316739004493,"3":0.9715263321339931,"4":1.2550535875099703e-06},"predicted":"3","confidence":0.9715263321339931,"entropy_confidence":0.9192793110164498,"candidate_mass":0.9999988788724604,"accepted":true,"accepted_predicted":"3","input_tokens":328,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.561254612215258,"brier_hard":1.887863823255427,"score_mae":0.65478703405193,"within_one":true,"case_id":"typed:322/r0","base_id":"typed:322","repeat":0,"workflow":"security_incidents","case_elapsed_ms":292.874001}
{"id":"true_positive","type":"binary","probabilities":{"false":1.5205416561218525e-06,"true":0.9999984794583426},"predicted":"true","confidence":0.9999984794583426,"entropy_confidence":0.9999684188406724,"candidate_mass":0.999998216498773,"accepted":true,"accepted_predicted":"true","input_tokens":274,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.520542813387344e-06,"brier_hard":4.624093859777111e-12,"soft_accuracy":0.4999999999999994,"brier_soft":0.4999969589213106,"tvd":0.49999847945834325,"kl":4.258597355979526,"case_id":"typed:322/r0","base_id":"typed:322","repeat":0,"workflow":"security_incidents","case_elapsed_ms":292.874001}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.141036910619557e-05,"1":1.62112093387886e-05,"2":0.00012741525055968476,"3":0.9998249631709943},"predicted":"3","confidence":0.9998249631709943,"entropy_confidence":0.9986855822349858,"candidate_mass":0.9999963954441909,"accepted":true,"accepted_predicted":"3","input_tokens":303,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.968059115870442,"brier_hard":1.9993951439628215,"score_mae":0.9997459312234414,"within_one":true,"case_id":"typed:322/r0","base_id":"typed:322","repeat":0,"workflow":"security_incidents","case_elapsed_ms":292.874001}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.007273360748075955,"true":0.9927266392519231},"predicted":"true","confidence":0.9927266392519231,"entropy_confidence":0.9378811511635198,"candidate_mass":0.9999414186031358,"accepted":true,"accepted_predicted":"true","input_tokens":267,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.007299940598009877,"brier_hard":0.00010580355314331836,"soft_accuracy":0.4999999999999995,"brier_soft":0.48555908205699044,"tvd":0.49272663925192356,"kl":1.7722711987587676,"case_id":"typed:323/r0","base_id":"typed:323","repeat":0,"workflow":"security_incidents","case_elapsed_ms":303.752101}
{"id":"disposition","type":"choice","probabilities":{"close_benign":3.669297018973928e-08,"contain":3.153483950583283e-08,"investigate":0.9999998031960986,"monitor":1.2857609183562245e-07},"predicted":"investigate","confidence":0.9999998031960986,"entropy_confidence":0.9999975403695935,"candidate_mass":0.9999997992556798,"accepted":true,"accepted_predicted":"investigate","input_tokens":319,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.9680392076364774e-07,"brier_hard":5.760440716110436e-14,"soft_accuracy":0.5666663616120863,"brier_soft":0.2750000331683739,"tvd":0.4333333698625319,"kl":3.669292033008851,"case_id":"typed:323/r0","base_id":"typed:323","repeat":0,"workflow":"security_incidents","case_elapsed_ms":303.752101}
{"id":"severity","type":"ordinal","probabilities":{"0":5.145032132620505e-06,"1":1.77688474977334e-05,"2":0.9997366675594762,"3":0.0002403067140781959,"4":1.1184681593177956e-07},"predicted":"2","confidence":0.9997366675594762,"entropy_confidence":0.9984313051198325,"candidate_mass":0.9999991119195712,"accepted":true,"accepted_predicted":"2","input_tokens":339,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00026336711859894693,"brier_hard":1.2743350687002837e-07,"score_mae":0.3502124714959485,"within_one":true,"case_id":"typed:323/r0","base_id":"typed:323","repeat":0,"workflow":"security_incidents","case_elapsed_ms":303.752101}
{"id":"true_positive","type":"binary","probabilities":{"false":3.5351039507229394e-08,"true":0.9999999646489612},"predicted":"true","confidence":0.9999999646489612,"entropy_confidence":0.9999990739311915,"candidate_mass":0.9999997215285635,"accepted":true,"accepted_predicted":"true","input_tokens":285,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.5351039387802076e-08,"brier_hard":2.499391935861576e-15,"soft_accuracy":0.6766669875092763,"brier_soft":0.20908841205737233,"tvd":0.3233329646489609,"kl":2.713717115188554,"case_id":"typed:323/r0","base_id":"typed:323","repeat":0,"workflow":"security_incidents","case_elapsed_ms":303.752101}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.1967085799679262e-05,"1":1.032674612455525e-05,"2":6.483186429201315e-05,"3":0.999912874303783},"predicted":"3","confidence":0.999912874303783,"entropy_confidence":0.9993027974058191,"candidate_mass":0.9999953870106312,"accepted":true,"accepted_predicted":"3","input_tokens":314,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.643713342565611,"brier_hard":1.9996960969228923,"score_mae":1.0665456133860574,"within_one":false,"case_id":"typed:323/r0","base_id":"typed:323","repeat":0,"workflow":"security_incidents","case_elapsed_ms":303.752101}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.7289721957432367e-05,"true":0.9999727102780411},"predicted":"true","confidence":0.9999727102780411,"entropy_confidence":0.9995468825991268,"candidate_mass":0.9999907655116138,"accepted":true,"accepted_predicted":"true","input_tokens":323,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.729009433010439e-05,"brier_hard":1.48945784910626e-09,"soft_accuracy":0.7666524454634284,"brier_soft":0.10886310889668321,"tvd":0.23330571027804187,"kl":1.908845130678104,"case_id":"typed:324/r0","base_id":"typed:324","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.599301}
{"id":"disposition","type":"choice","probabilities":{"close_benign":3.7893147754744145e-06,"contain":1.777600501778493e-05,"investigate":0.9999734256433082,"monitor":5.009036900266616e-06},"predicted":"investigate","confidence":0.9999734256433082,"entropy_confidence":0.9997623615857899,"candidate_mass":0.9999983368289851,"accepted":true,"accepted_predicted":"investigate","input_tokens":375,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":10.93766104278841,"brier_hard":1.999911300338213,"soft_accuracy":0.2333395764666055,"brier_soft":1.0821564771930374,"tvd":0.7666442149580828,"kl":6.721703353539677,"case_id":"typed:324/r0","base_id":"typed:324","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.599301}
{"id":"severity","type":"ordinal","probabilities":{"0":8.695255022820689e-06,"1":2.084465389415875e-05,"2":0.0002637758057137467,"3":0.9996831986761646,"4":2.3485609204136488e-05},"predicted":"3","confidence":0.9996831986761646,"entropy_confidence":0.9980945496937674,"candidate_mass":0.9999991532501437,"accepted":true,"accepted_predicted":"3","input_tokens":395,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":10.659122698683944,"brier_hard":1.9993195971363562,"score_mae":0.5636410652693677,"within_one":true,"case_id":"typed:324/r0","base_id":"typed:324","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.599301}
{"id":"true_positive","type":"binary","probabilities":{"false":1.647547485669559e-08,"true":0.9999999835245249},"predicted":"true","confidence":0.9999999835245249,"entropy_confidence":0.9999995502557864,"candidate_mass":0.9999996872466898,"accepted":true,"accepted_predicted":"true","input_tokens":341,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.647547523103286e-08,"brier_hard":5.428825513698618e-16,"soft_accuracy":0.8833329873688134,"brier_soft":0.02722237008942357,"tvd":0.11666698352452498,"kl":0.9649626630898072,"case_id":"typed:324/r0","base_id":"typed:324","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.599301}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.513894965913729e-06,"1":1.2685095277114496e-06,"2":7.941719214931428e-06,"3":0.9999892758762914},"predicted":"3","confidence":0.9999892758762914,"entropy_confidence":0.9998979310994158,"candidate_mass":0.9999985853088292,"accepted":true,"accepted_predicted":"3","input_tokens":370,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":1.0724181212442195e-05,"brier_hard":1.8197872779623346e-10,"score_mae":0.2733179795768317,"within_one":true,"case_id":"typed:324/r0","base_id":"typed:324","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.599301}
{"id":"credential_compromise","type":"binary","probabilities":{"false":8.766265920805836e-06,"true":0.9999912337340802},"predicted":"true","confidence":0.9999912337340802,"entropy_confidence":0.9998400831923964,"candidate_mass":0.9999926165370759,"accepted":true,"accepted_predicted":"true","input_tokens":321,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.766304343768737e-06,"brier_hard":1.5369483637154398e-10,"soft_accuracy":0.7333289090817485,"brier_soft":0.14221322723635815,"tvd":0.2666582337340797,"kl":2.228653087222939,"case_id":"typed:325/r0","base_id":"typed:325","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.594101}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.948797665169305e-05,"contain":9.475587869117482e-05,"investigate":0.9996092844786073,"monitor":0.00022647166604808006},"predicted":"investigate","confidence":0.9996092844786073,"entropy_confidence":0.9972339699564614,"candidate_mass":0.9999892846805924,"accepted":true,"accepted_predicted":"investigate","input_tokens":373,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0003907918705898799,"brier_hard":2.177552896254397e-07,"soft_accuracy":0.3666211586229631,"brier_soft":0.559865547244578,"tvd":0.6329422844786081,"kl":4.349204118033545,"case_id":"typed:325/r0","base_id":"typed:325","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.594101}
{"id":"severity","type":"ordinal","probabilities":{"0":2.363276004862071e-06,"1":2.970269430286585e-05,"2":0.003940965514783351,"3":0.9960251151356188,"4":1.8533792910727003e-06},"predicted":"3","confidence":0.9960251151356188,"entropy_confidence":0.9837519880529906,"candidate_mass":0.9999993383700087,"accepted":true,"accepted_predicted":"3","input_tokens":393,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.003982805715839569,"brier_hard":3.133181014393583e-05,"score_mae":0.26266139264789023,"within_one":true,"case_id":"typed:325/r0","base_id":"typed:325","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.594101}
{"id":"true_positive","type":"binary","probabilities":{"false":7.189978697533902e-07,"true":0.9999992810021316},"predicted":"true","confidence":0.9999992810021316,"entropy_confidence":0.9999842897500951,"candidate_mass":0.9999981111858854,"accepted":true,"accepted_predicted":"true","input_tokens":339,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.189981268332566e-07,"brier_hard":1.0339158714077466e-12,"soft_accuracy":0.6266668178533945,"brier_soft":0.2787539840765082,"tvd":0.373332281002131,"kl":3.145657909472184,"case_id":"typed:325/r0","base_id":"typed:325","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.594101}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.0050673676122597e-05,"1":1.8389428042604257e-05,"2":0.00013076015878073214,"3":0.9998307997395018},"predicted":"3","confidence":0.9998307997395018,"entropy_confidence":0.998733408327081,"candidate_mass":0.9999941399130209,"accepted":true,"accepted_predicted":"3","input_tokens":368,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00016921457647710306,"brier_hard":4.646714785560764e-08,"score_mae":0.8331053089641096,"within_one":true,"case_id":"typed:325/r0","base_id":"typed:325","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.594101}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.1924805025631828e-05,"true":0.9999780751949746},"predicted":"true","confidence":0.9999780751949746,"entropy_confidence":0.9996290376554865,"candidate_mass":0.999993794433222,"accepted":true,"accepted_predicted":"true","input_tokens":313,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.1925045377437008e-05,"brier_hard":9.6139415081317e-10,"soft_accuracy":0.5933289073843452,"brier_soft":0.3307204343606529,"tvd":0.4066450751949745,"kl":3.4358349757804136,"case_id":"typed:326/r0","base_id":"typed:326","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.352502}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.891389323901745e-06,"contain":9.22367186749541e-06,"investigate":0.9999840404921516,"monitor":3.844446658254903e-06},"predicted":"investigate","confidence":0.9999840404921516,"entropy_confidence":0.9998501701463096,"candidate_mass":0.9999981739027437,"accepted":true,"accepted_predicted":"investigate","input_tokens":365,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":11.59373734940571,"brier_hard":1.9999496340034901,"soft_accuracy":0.18333640625040556,"brier_soft":1.0446957983140617,"tvd":0.8166508571589677,"kl":7.2035658975683114,"case_id":"typed:326/r0","base_id":"typed:326","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.352502}
{"id":"severity","type":"ordinal","probabilities":{"0":1.168316391966078e-05,"1":0.00010301525384494455,"2":0.8510703000602364,"3":0.14881375004306074,"4":1.2514789371692682e-06},"predicted":"2","confidence":0.8510703000602364,"entropy_confidence":0.7378969467274484,"candidate_mass":0.9999943527024918,"accepted":true,"accepted_predicted":"2","input_tokens":385,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":13.591184556284713,"brier_hard":1.7464636956388302,"score_mae":1.0513101285807513,"within_one":false,"case_id":"typed:326/r0","base_id":"typed:326","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.352502}
{"id":"true_positive","type":"binary","probabilities":{"false":7.367687410673889e-09,"true":0.9999999926323131},"predicted":"true","confidence":0.9999999926323131,"entropy_confidence":0.9999997903241927,"candidate_mass":0.9999997279026275,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.367686932154621e-09,"brier_hard":1.0856562811170573e-16,"soft_accuracy":0.7466669963652696,"brier_soft":0.12835521031208696,"tvd":0.2533329926323129,"kl":2.115154869884784,"case_id":"typed:326/r0","base_id":"typed:326","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.352502}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.2834696118993047e-05,"1":1.0305525488824462e-05,"2":9.445854028273607e-05,"3":0.9998824012381082},"predicted":"3","confidence":0.9998824012381082,"entropy_confidence":0.9990940808993625,"candidate_mass":0.999995916933597,"accepted":true,"accepted_predicted":"3","input_tokens":360,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00011760567716832496,"brier_hard":2.3022817910889826e-08,"score_mae":0.4665134263203794,"within_one":true,"case_id":"typed:326/r0","base_id":"typed:326","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.352502}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0007149503007215149,"true":0.9992850496992792},"predicted":"true","confidence":0.9992850496992792,"entropy_confidence":0.9914977762683435,"candidate_mass":0.9999624753122295,"accepted":true,"accepted_predicted":"true","input_tokens":326,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0007152059995689335,"brier_hard":1.0223078650025091e-06,"soft_accuracy":0.9127419748947043,"brier_soft":0.014775509695014623,"tvd":0.08595204969927892,"kl":0.3336501988250927,"case_id":"typed:327/r0","base_id":"typed:327","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.791301}
{"id":"disposition","type":"choice","probabilities":{"close_benign":3.913975352833043e-06,"contain":3.762945801313357e-06,"investigate":0.9999836229210268,"monitor":8.700157819030674e-06},"predicted":"investigate","confidence":0.9999836229210268,"entropy_confidence":0.9998460025835758,"candidate_mass":0.9999974166303369,"accepted":true,"accepted_predicted":"investigate","input_tokens":378,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":12.490308449509222,"brier_hard":1.9999597203238315,"soft_accuracy":0.2999975536121252,"brier_soft":0.935171535258856,"tvd":0.6999839229207269,"kl":6.042720073500876,"case_id":"typed:327/r0","base_id":"typed:327","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.791301}
{"id":"severity","type":"ordinal","probabilities":{"0":5.077173925335934e-06,"1":3.135977467118538e-05,"2":0.9998494098299263,"3":0.00011402977736415201,"4":1.2344411420999518e-07},"predicted":"2","confidence":0.9998494098299263,"entropy_confidence":0.9990214474094562,"candidate_mass":0.9999992690115966,"accepted":true,"accepted_predicted":"2","input_tokens":398,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001506015099118045,"brier_hard":3.668941784948392e-08,"score_mae":0.3399272374569269,"within_one":true,"case_id":"typed:327/r0","base_id":"typed:327","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.791301}
{"id":"true_positive","type":"binary","probabilities":{"false":1.8665349075678828e-08,"true":0.9999999813346501},"predicted":"true","confidence":0.9999999813346501,"entropy_confidence":0.9999994938375927,"candidate_mass":0.9999995015157416,"accepted":true,"accepted_predicted":"true","input_tokens":344,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.8665350106531798e-08,"brier_hard":6.967905442134311e-16,"soft_accuracy":0.9666669825789943,"brier_soft":0.002222175289312309,"tvd":0.033332981334650444,"kl":0.2742371157419746,"case_id":"typed:327/r0","base_id":"typed:327","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.791301}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.293029082956215e-05,"1":3.874190844994017e-05,"2":0.0009036893926659915,"3":0.9990246384080556},"predicted":"3","confidence":0.9990246384080556,"entropy_confidence":0.9941987012840229,"candidate_mass":0.9999950022235684,"accepted":true,"accepted_predicted":"3","input_tokens":373,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0009758375665853908,"brier_hard":1.7705700929818954e-06,"score_mae":0.20225303591794885,"within_one":true,"case_id":"typed:327/r0","base_id":"typed:327","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.791301}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00010126107832433644,"true":0.9998987389216769},"predicted":"true","confidence":0.9998987389216769,"entropy_confidence":0.998510221232506,"candidate_mass":0.9999860737365012,"accepted":true,"accepted_predicted":"true","input_tokens":318,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.197808442414068,"brier_hard":1.9995949761943173,"soft_accuracy":0.060089109748925486,"brier_soft":1.7668192788531147,"tvd":0.9398987389216762,"kl":8.418978489340954,"case_id":"typed:328/r0","base_id":"typed:328","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.263001}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.5472335124479067e-05,"contain":2.3080515203231683e-05,"investigate":0.999878397948998,"monitor":7.304920067366512e-05},"predicted":"investigate","confidence":0.999878397948998,"entropy_confidence":0.9990382946557552,"candidate_mass":0.9999955159317857,"accepted":true,"accepted_predicted":"investigate","input_tokens":370,"reused_prefix_tokens":0,"expected":"close_benign","correct":false,"accepted_correct":false,"nll_hard":10.57791759180597,"brier_hard":1.9997058725325416,"soft_accuracy":0.06669053015541726,"brier_soft":1.4863751835594827,"tvd":0.9332345451311349,"kl":8.169499036712832,"case_id":"typed:328/r0","base_id":"typed:328","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.263001}
{"id":"severity","type":"ordinal","probabilities":{"0":9.218804096321209e-05,"1":0.0006651251412228958,"2":0.9989621802039937,"3":0.0002803406308079009,"4":1.659830131146408e-07},"predicted":"2","confidence":0.9989621802039937,"entropy_confidence":0.9943736579473323,"candidate_mass":0.9999989021675497,"accepted":true,"accepted_predicted":"2","input_tokens":390,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0010383587038625063,"brier_hard":1.6065509141980118e-06,"score_mae":0.8094311713736866,"within_one":true,"case_id":"typed:328/r0","base_id":"typed:328","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.263001}
{"id":"true_positive","type":"binary","probabilities":{"false":2.4573964178950953e-06,"true":0.9999975426035834},"predicted":"true","confidence":0.9999975426035834,"entropy_confidence":0.9999506625296333,"candidate_mass":0.9999976825431846,"accepted":true,"accepted_predicted":"true","input_tokens":336,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.916408135194846,"brier_hard":1.9999901704264087,"soft_accuracy":0.0900020150650628,"brier_soft":1.656191055089119,"tvd":0.9099975426035828,"kl":8.164694854885608,"case_id":"typed:328/r0","base_id":"typed:328","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.263001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00035705580303507166,"1":0.00034735658461619194,"2":0.926361599829503,"3":0.07293398778284588},"predicted":"2","confidence":0.926361599829503,"entropy_confidence":0.8071010657016534,"candidate_mass":0.9999806830270693,"accepted":true,"accepted_predicted":"2","input_tokens":365,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":7.93761847737732,"brier_hard":1.8627513167519578,"score_mae":1.4352055195921598,"within_one":false,"case_id":"typed:328/r0","base_id":"typed:328","repeat":0,"workflow":"security_incidents","case_elapsed_ms":324.263001}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0012496665411945442,"true":0.9987503334588046},"predicted":"true","confidence":0.9987503334588046,"entropy_confidence":0.9861461538283847,"candidate_mass":0.9999436841141067,"accepted":true,"accepted_predicted":"true","input_tokens":320,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.684878530300949,"brier_hard":1.9950044571681484,"soft_accuracy":0.35037489996235804,"brier_soft":0.8417539903258215,"tvd":0.6487503334588051,"kl":3.69816206246993,"case_id":"typed:329/r0","base_id":"typed:329","repeat":0,"workflow":"security_incidents","case_elapsed_ms":318.96700100000004}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0916696347813469e-05,"contain":1.078393741678713e-05,"investigate":0.9999384758699819,"monitor":3.982349625293696e-05},"predicted":"investigate","confidence":0.9999384758699819,"entropy_confidence":0.999485648482223,"candidate_mass":0.9999962100490071,"accepted":true,"accepted_predicted":"investigate","input_tokens":372,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":6.152602270506064e-05,"brier_hard":5.606596993656518e-09,"soft_accuracy":0.5666450482702914,"brier_soft":0.27936481636197796,"tvd":0.43327147586998216,"kl":3.5183047129301563,"case_id":"typed:329/r0","base_id":"typed:329","repeat":0,"workflow":"security_incidents","case_elapsed_ms":318.96700100000004}
{"id":"severity","type":"ordinal","probabilities":{"0":5.7978725326182355e-05,"1":0.0016986467149710579,"2":0.9979503667449643,"3":0.000292737100311552,"4":2.707144263316e-07},"predicted":"2","confidence":0.9979503667449643,"entropy_confidence":0.990162497543562,"candidate_mass":0.9999985496249304,"accepted":true,"accepted_predicted":"2","input_tokens":392,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.002051736627862487,"brier_hard":7.1754537582058116e-06,"score_mae":0.49847867436353965,"within_one":true,"case_id":"typed:329/r0","base_id":"typed:329","repeat":0,"workflow":"security_incidents","case_elapsed_ms":318.96700100000004}
{"id":"true_positive","type":"binary","probabilities":{"false":4.164165488562214e-07,"true":0.9999995835834504},"predicted":"true","confidence":0.9999995835834504,"entropy_confidence":0.9999905730938734,"candidate_mass":0.999998672305875,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.1641663627708877e-07,"brier_hard":3.4680548492185164e-13,"soft_accuracy":0.5499999583583448,"brier_soft":0.4049992504505582,"tvd":0.4499995835834507,"kl":3.8158430460028407,"case_id":"typed:329/r0","base_id":"typed:329","repeat":0,"workflow":"security_incidents","case_elapsed_ms":318.96700100000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0010588415082304038,"1":0.7726033032508213,"2":0.2216456396863194,"3":0.004692215554627275},"predicted":"1","confidence":0.7726033032508213,"entropy_confidence":0.5919450704185041,"candidate_mass":0.9999820055832317,"accepted":false,"accepted_predicted":null,"input_tokens":367,"reused_prefix_tokens":0,"expected":"1","correct":true,"accepted_correct":null,"nll_hard":0.2579895532597157,"brier_hard":0.10085918531654631,"score_mae":0.28669577071265806,"within_one":true,"case_id":"typed:329/r0","base_id":"typed:329","repeat":0,"workflow":"security_incidents","case_elapsed_ms":318.96700100000004}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00010041049796676327,"true":0.9998995895020325},"predicted":"true","confidence":0.9998995895020325,"entropy_confidence":0.9985215131659755,"candidate_mass":0.9999815505947499,"accepted":true,"accepted_predicted":"true","input_tokens":315,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00010041553943902743,"brier_hard":2.0164536204013848e-08,"soft_accuracy":0.5666536118666636,"brier_soft":0.3753809532132738,"tvd":0.4332325895020328,"kl":3.305194468774518,"case_id":"typed:330/r0","base_id":"typed:330","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.53340099999997}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.1355153119232733e-07,"contain":4.0747920235158523e-07,"investigate":0.9999987745897249,"monitor":6.043795402975954e-07},"predicted":"investigate","confidence":0.9999987745897249,"entropy_confidence":0.9999861826364211,"candidate_mass":0.9999992161219015,"accepted":true,"accepted_predicted":"investigate","input_tokens":367,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.225411025956566e-06,"brier_hard":2.07854852797461e-12,"soft_accuracy":0.5833330443293464,"brier_soft":0.27086520225688154,"tvd":0.4166651912561422,"kl":3.5232257237163953,"case_id":"typed:330/r0","base_id":"typed:330","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.53340099999997}
{"id":"severity","type":"ordinal","probabilities":{"0":1.3786615872847955e-06,"1":4.301151869077808e-06,"2":5.778024172408279e-05,"3":0.999928246687549,"4":8.293257269772307e-06},"predicted":"3","confidence":0.999928246687549,"entropy_confidence":0.9995001957852383,"candidate_mass":0.9999996556642603,"accepted":true,"accepted_predicted":"3","input_tokens":387,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":7.175588684308926e-05,"brier_hard":8.576272912702501e-09,"score_mae":0.050062225272956784,"within_one":true,"case_id":"typed:330/r0","base_id":"typed:330","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.53340099999997}
{"id":"true_positive","type":"binary","probabilities":{"false":6.635426526045761e-08,"true":0.9999999336457344},"predicted":"true","confidence":0.9999999336457344,"entropy_confidence":0.999998322037948,"candidate_mass":0.9999992367473247,"accepted":true,"accepted_predicted":"true","input_tokens":333,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.63542678390581e-08,"brier_hard":8.805777086562175e-15,"soft_accuracy":0.7333329690347202,"brier_soft":0.1422225070000372,"tvd":0.26666693364573457,"kl":2.2286467072624503,"case_id":"typed:330/r0","base_id":"typed:330","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.53340099999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.359600524380691e-07,"1":7.684645497351169e-07,"2":5.602152288145782e-06,"3":0.9999926934231079},"predicted":"3","confidence":0.9999926934231079,"entropy_confidence":0.9999286864195389,"candidate_mass":0.9999988182789271,"accepted":true,"accepted_predicted":"3","input_tokens":362,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":7.306603585237331e-06,"brier_hard":8.623673512333162e-11,"score_mae":0.46665705303844973,"within_one":true,"case_id":"typed:330/r0","base_id":"typed:330","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.53340099999997}
{"id":"credential_compromise","type":"binary","probabilities":{"false":9.548985617762571e-07,"true":0.9999990451014374},"predicted":"true","confidence":0.9999990451014374,"entropy_confidence":0.9999795261693474,"candidate_mass":0.9999972912453879,"accepted":true,"accepted_predicted":"true","input_tokens":319,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.548990184710736e-07,"brier_hard":1.823662528052258e-12,"soft_accuracy":0.5633328790468183,"brier_soft":0.38135446988906196,"tvd":0.43666604510143786,"kl":3.6985642437808766,"case_id":"typed:331/r0","base_id":"typed:331","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.180002}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.083545727513627e-05,"contain":1.6383833059101084e-05,"investigate":0.9999466810381196,"monitor":2.6099671546636935e-05},"predicted":"investigate","confidence":0.9999466810381196,"entropy_confidence":0.9995432585533224,"candidate_mass":0.9999951035527602,"accepted":true,"accepted_predicted":"investigate","input_tokens":371,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":5.332038338681012e-05,"brier_hard":3.9099416709192e-09,"soft_accuracy":0.46665355778089307,"brier_soft":0.4252754482023946,"tvd":0.5332796810381193,"kl":4.544934106209835,"case_id":"typed:331/r0","base_id":"typed:331","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.180002}
{"id":"severity","type":"ordinal","probabilities":{"0":1.7538389289292108e-06,"1":7.5569092060919185e-06,"2":8.85519648030817e-05,"3":0.9999004412539125,"4":1.6960331488966705e-06},"predicted":"3","confidence":0.9999004412539125,"entropy_confidence":0.999340879731072,"candidate_mass":0.9999996858834233,"accepted":true,"accepted_predicted":"3","input_tokens":391,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":13.287218475445552,"brier_hard":1.9997975082579809,"score_mae":0.13010723126685475,"within_one":true,"case_id":"typed:331/r0","base_id":"typed:331","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.180002}
{"id":"true_positive","type":"binary","probabilities":{"false":1.5196119200920884e-07,"true":0.9999998480388088},"predicted":"true","confidence":0.9999998480388088,"entropy_confidence":0.9999963388767057,"candidate_mass":0.9999989311973528,"accepted":true,"accepted_predicted":"true","input_tokens":337,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.519612027320927e-07,"brier_hard":4.618440750352456e-14,"soft_accuracy":0.6599999513724192,"brier_soft":0.2311997933328256,"tvd":0.33999984803880834,"kl":2.8572756337515965,"case_id":"typed:331/r0","base_id":"typed:331","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.180002}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.590978468011431e-05,"1":1.3476421081420066e-05,"2":8.589608616325019e-05,"3":0.9998847177080739},"predicted":"3","confidence":0.9998847177080739,"entropy_confidence":0.999100927233188,"candidate_mass":0.99999553453218,"accepted":true,"accepted_predicted":"3","input_tokens":366,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00011528893744030511,"brier_hard":2.110287962364059e-08,"score_mae":0.5498394217176292,"within_one":true,"case_id":"typed:331/r0","base_id":"typed:331","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.180002}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0001958064657316648,"true":0.999804193534267},"predicted":"true","confidence":0.999804193534267,"entropy_confidence":0.9973055389923723,"candidate_mass":0.9999833852274488,"accepted":true,"accepted_predicted":"true","input_tokens":317,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00019582563832182514,"brier_hard":7.66803440451852e-08,"soft_accuracy":0.5833003657195817,"brier_soft":0.3468965100877149,"tvd":0.41647119353426765,"kl":2.878583618823879,"case_id":"typed:332/r0","base_id":"typed:332","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.779101}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.04437278583464667,"contain":0.00047591928333772144,"investigate":0.7133689229549702,"monitor":0.24178237192704696},"predicted":"investigate","confidence":0.7133689229549702,"entropy_confidence":0.47624723887028964,"candidate_mass":0.9999792680015959,"accepted":false,"accepted_predicted":null,"input_tokens":369,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":null,"nll_hard":0.3377565688859647,"brier_hard":0.14258526032455432,"soft_accuracy":0.47153955137145276,"brier_soft":0.035688159047589214,"tvd":0.1551512948820164,"kl":0.637364587919951,"case_id":"typed:332/r0","base_id":"typed:332","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.779101}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00010537214307086891,"1":0.000530048354076085,"2":0.9358052038017738,"3":0.06355699409087719,"4":2.3816102034206934e-06},"predicted":"2","confidence":0.9358052038017738,"entropy_confidence":0.849491542103298,"candidate_mass":0.999995407040316,"accepted":true,"accepted_predicted":"2","input_tokens":389,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.7558182308976917,"brier_hard":1.752657174838811,"score_mae":0.40717903532893107,"within_one":true,"case_id":"typed:332/r0","base_id":"typed:332","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.779101}
{"id":"true_positive","type":"binary","probabilities":{"false":7.356874712332219e-05,"true":0.9999264312528752},"predicted":"true","confidence":0.9999264312528752,"entropy_confidence":0.9988837274642874,"candidate_mass":0.9999922524293964,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.357145343784125e-05,"brier_hard":1.0824721106812555e-08,"soft_accuracy":0.5999852862505745,"brier_soft":0.31988230082932256,"tvd":0.39992643125287597,"kl":3.1339485774178146,"case_id":"typed:332/r0","base_id":"typed:332","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.779101}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0018529546442703818,"1":0.001110742141444983,"2":0.24583201675590072,"3":0.7512042864583852},"predicted":"3","confidence":0.7512042864583852,"entropy_confidence":0.5823078294632738,"candidate_mass":0.9999915630026243,"accepted":false,"accepted_predicted":null,"input_tokens":364,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.4031068350198304,"brier_hard":1.1330818941329421,"score_mae":0.6630546350284017,"within_one":true,"case_id":"typed:332/r0","base_id":"typed:332","repeat":0,"workflow":"security_incidents","case_elapsed_ms":321.779101}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00011524021431839651,"true":0.9998847597856818},"predicted":"true","confidence":0.9998847597856818,"entropy_confidence":0.998326057519032,"candidate_mass":0.9999845218023377,"accepted":true,"accepted_predicted":"true","input_tokens":345,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00011524685498187471,"brier_hard":2.65606139922525e-08,"soft_accuracy":0.516663158582696,"brier_soft":0.4669988067445856,"tvd":0.4832177597856817,"kl":3.6905693857653894,"case_id":"typed:333/r0","base_id":"typed:333","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.649701}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.2921542453536028e-06,"contain":1.1020242750432856e-06,"investigate":0.9999914522146902,"monitor":6.153606787755846e-06},"predicted":"investigate","confidence":0.9999914522146902,"entropy_confidence":0.9999170304827497,"candidate_mass":0.9999990081090483,"accepted":true,"accepted_predicted":"investigate","input_tokens":397,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":8.547821842312093e-06,"brier_hard":1.1381563029709629e-10,"soft_accuracy":0.6499958526399422,"brier_soft":0.1738800770413117,"tvd":0.349991452214691,"kl":2.943615790815767,"case_id":"typed:333/r0","base_id":"typed:333","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.649701}
{"id":"severity","type":"ordinal","probabilities":{"0":4.376533789788802e-05,"1":0.0009079721014771914,"2":0.9987908702689475,"3":0.0002571530242075974,"4":2.392674685485284e-07},"predicted":"2","confidence":0.9987908702689475,"entropy_confidence":0.993701778691509,"candidate_mass":0.9999987971678395,"accepted":true,"accepted_predicted":"2","input_tokens":417,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0012098613181876926,"brier_hard":2.354451183485281e-06,"score_mae":0.27592912878186926,"within_one":true,"case_id":"typed:333/r0","base_id":"typed:333","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.649701}
{"id":"true_positive","type":"binary","probabilities":{"false":8.863492514197585e-07,"true":0.9999991136507502},"predicted":"true","confidence":0.9999991136507502,"entropy_confidence":0.9999809006664374,"candidate_mass":0.9999990046996331,"accepted":true,"accepted_predicted":"true","input_tokens":363,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.863496425812166e-07,"brier_hard":1.5712299880663917e-12,"soft_accuracy":0.6333327636407916,"brier_soft":0.2688880777994885,"tvd":0.36666611365074936,"kl":3.087847931864805,"case_id":"typed:333/r0","base_id":"typed:333","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.649701}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002890796923792256,"1":0.00026611669503129147,"2":0.4398136636508247,"3":0.5596311399617654},"predicted":"3","confidence":0.5596311399617654,"entropy_confidence":0.5017910919134769,"candidate_mass":0.9999871362806153,"accepted":false,"accepted_predicted":null,"input_tokens":392,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":null,"nll_hard":8.2315756422595,"brier_hard":1.5060909925439672,"score_mae":1.008786863881977,"within_one":false,"case_id":"typed:333/r0","base_id":"typed:333","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.649701}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.014735313628703634,"true":0.9852646863712947},"predicted":"true","confidence":0.9852646863712947,"entropy_confidence":0.8892406655807165,"candidate_mass":0.9998925354476556,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.21750837841067,"brier_hard":1.9414930044206549,"soft_accuracy":0.4773546381457963,"brier_soft":0.5173432131644694,"tvd":0.5085976863712955,"kl":1.5221794859786208,"case_id":"typed:334/r0","base_id":"typed:334","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.622401}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.577427862126314e-06,"contain":7.777721719740688e-06,"investigate":0.9999779056085804,"monitor":8.739241836641418e-06},"predicted":"investigate","confidence":0.9999779056085804,"entropy_confidence":0.9997959639137936,"candidate_mass":0.9999981110747603,"accepted":true,"accepted_predicted":"investigate","input_tokens":374,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.2094635504308352e-05,"brier_hard":6.561371367908225e-10,"soft_accuracy":0.5999899088815924,"brier_soft":0.24373171188811307,"tvd":0.39997790560858093,"kl":3.357320625218271,"case_id":"typed:334/r0","base_id":"typed:334","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.622401}
{"id":"severity","type":"ordinal","probabilities":{"0":2.2777419784936523e-06,"1":1.7282721208809692e-05,"2":0.003309495170229673,"3":0.9966703731955602,"4":5.711710235859037e-07},"predicted":"3","confidence":0.9966703731955602,"entropy_confidence":0.986049916225621,"candidate_mass":0.9999996888802145,"accepted":true,"accepted_predicted":"3","input_tokens":394,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0033351823471186033,"brier_hard":2.2039477145414918e-05,"score_mae":0.09664967733244323,"within_one":true,"case_id":"typed:334/r0","base_id":"typed:334","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.622401}
{"id":"true_positive","type":"binary","probabilities":{"false":9.889285237111946e-09,"true":0.9999999901107142},"predicted":"true","confidence":0.9999999901107142,"entropy_confidence":0.9999997227619806,"candidate_mass":0.9999998202690642,"accepted":true,"accepted_predicted":"true","input_tokens":340,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.889285883573398e-09,"brier_hard":1.9559593682085225e-16,"soft_accuracy":0.7933329941982921,"brier_soft":0.08542248960284435,"tvd":0.20666699011071452,"kl":1.71980713655682,"case_id":"typed:334/r0","base_id":"typed:334","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.622401}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.30248719905103e-05,"1":7.31810632300389e-05,"2":0.004428453576364409,"3":0.9954153404884141},"predicted":"3","confidence":0.9954153404884141,"entropy_confidence":0.9783220285011975,"candidate_mass":0.9999965034786897,"accepted":true,"accepted_predicted":"3","input_tokens":369,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.004595201295671762,"brier_hard":4.064255251257459e-05,"score_mae":0.4285091096812015,"within_one":true,"case_id":"typed:334/r0","base_id":"typed:334","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.622401}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00011861388817976606,"true":0.9998813861118205},"predicted":"true","confidence":0.9998813861118205,"entropy_confidence":0.9982819906417034,"candidate_mass":0.9999823253724566,"accepted":true,"accepted_predicted":"true","input_tokens":317,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00011862092336301773,"brier_hard":2.8138508938172448e-08,"soft_accuracy":0.733277646931259,"brier_soft":0.14209608427763215,"tvd":0.2665483861118204,"kl":1.830744353830892,"case_id":"typed:335/r0","base_id":"typed:335","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.21210099999996}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.714787706949314e-05,"contain":0.0044279509451872765,"investigate":0.9954447497949689,"monitor":7.015138277593835e-05},"predicted":"investigate","confidence":0.9954447497949689,"entropy_confidence":0.9785234183366088,"candidate_mass":0.9999933157092761,"accepted":true,"accepted_predicted":"investigate","input_tokens":369,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0045656569728933895,"brier_hard":4.036524109977983e-05,"soft_accuracy":0.4991371618327884,"brier_soft":0.37001119894346074,"tvd":0.49544474979496805,"kl":2.391309858252214,"case_id":"typed:335/r0","base_id":"typed:335","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.21210099999996}
{"id":"severity","type":"ordinal","probabilities":{"0":7.006034850204144e-06,"1":6.949846562429077e-05,"2":0.05084978120039657,"3":0.9490721840061812,"4":1.5302929466522164e-06},"predicted":"3","confidence":0.9490721840061812,"entropy_confidence":0.8745817996536899,"candidate_mass":0.9999982778483144,"accepted":true,"accepted_predicted":"3","input_tokens":389,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.05227042003390988,"brier_hard":0.005179347571491515,"score_mae":0.2856587340567476,"within_one":true,"case_id":"typed:335/r0","base_id":"typed:335","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.21210099999996}
{"id":"true_positive","type":"binary","probabilities":{"false":2.4042013807894834e-07,"true":0.9999997595798615},"predicted":"true","confidence":0.9999997595798615,"entropy_confidence":0.999994366803789,"candidate_mass":0.9999995310775096,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.4042016736518513e-07,"brier_hard":1.1560368577307406e-13,"soft_accuracy":0.849999831705903,"brier_soft":0.044999855748032655,"tvd":0.14999975957986175,"kl":1.2434101700804612,"case_id":"typed:335/r0","base_id":"typed:335","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.21210099999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.2088901541557785e-06,"1":8.946879851882916e-07,"2":6.564544174465782e-06,"3":0.9999913318776861},"predicted":"3","confidence":0.9999913318776861,"entropy_confidence":0.9999163666368094,"candidate_mass":0.9999985803623361,"accepted":true,"accepted_predicted":"3","input_tokens":364,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":8.668159882315541e-06,"brier_hard":1.2049146666335333e-10,"score_mae":0.5399880194093925,"within_one":true,"case_id":"typed:335/r0","base_id":"typed:335","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.21210099999996}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.006608877985020994,"true":0.9933911220149799},"predicted":"true","confidence":0.9933911220149799,"entropy_confidence":0.9426395908793959,"candidate_mass":0.9999501609439104,"accepted":true,"accepted_predicted":"true","input_tokens":296,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.006630813317837446,"brier_hard":8.73545364417782e-05,"soft_accuracy":0.5164466996612477,"brier_soft":0.4545317770219061,"tvd":0.4767241220149795,"kl":1.7368477520233951,"case_id":"typed:336/r0","base_id":"typed:336","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.287101}
{"id":"disposition","type":"choice","probabilities":{"close_benign":3.614018096872623e-08,"contain":3.071939036128094e-08,"investigate":0.9999997795468042,"monitor":1.5359362515219973e-07},"predicted":"investigate","confidence":0.9999997795468042,"entropy_confidence":0.9999972726822188,"candidate_mass":0.9999998407344626,"accepted":true,"accepted_predicted":"investigate","input_tokens":348,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.204532201092429e-07,"brier_hard":7.444040685460794e-14,"soft_accuracy":0.5499999236544145,"brier_soft":0.28722188956285377,"tvd":0.4499997795468038,"kl":3.8158429382229624,"case_id":"typed:336/r0","base_id":"typed:336","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.287101}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00045515668311481815,"1":0.003367793637080231,"2":0.9935636752962435,"3":0.002612769319181971,"4":6.050643786494092e-07},"predicted":"2","confidence":0.9935636752962435,"entropy_confidence":0.9722635058879228,"candidate_mass":0.9999977207656596,"accepted":true,"accepted_predicted":"2","input_tokens":368,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00645712715051163,"brier_hard":5.9802041161690485e-05,"score_mae":0.2716688724446277,"within_one":true,"case_id":"typed:336/r0","base_id":"typed:336","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.287101}
{"id":"true_positive","type":"binary","probabilities":{"false":7.975371741933549e-09,"true":0.9999999920246285},"predicted":"true","confidence":0.9999999920246285,"entropy_confidence":0.9999997739421069,"candidate_mass":0.9999997748645293,"accepted":true,"accepted_predicted":"true","input_tokens":314,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.975371506147958e-09,"brier_hard":1.2721310457582307e-16,"soft_accuracy":0.6999999968098515,"brier_soft":0.17999999042955425,"tvd":0.29999999202462846,"kl":2.5134296564185026,"case_id":"typed:336/r0","base_id":"typed:336","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.287101}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.432884705775328e-05,"1":1.4311393757643828e-05,"2":0.00012809048067720025,"3":0.9998432692785073},"predicted":"3","confidence":0.9998432692785073,"entropy_confidence":0.9988283778163867,"candidate_mass":0.999996231348621,"accepted":true,"accepted_predicted":"3","input_tokens":343,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.962773663476227,"brier_hard":1.9994303989774824,"score_mae":1.383133300190634,"within_one":false,"case_id":"typed:336/r0","base_id":"typed:336","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.287101}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.2300510326137401,"true":0.7699489673862598},"predicted":"true","confidence":0.7699489673862598,"entropy_confidence":0.22189974575373128,"candidate_mass":0.999819338330139,"accepted":false,"accepted_predicted":null,"input_tokens":289,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":null,"nll_hard":1.4694541137415698,"brier_hard":1.1856428247583357,"soft_accuracy":0.32003420170847163,"brier_soft":0.7278982643468727,"tvd":0.6032819673862599,"kl":0.8175547871738267,"case_id":"typed:337/r0","base_id":"typed:337","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.195402}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.597731649158421e-08,"contain":1.4255577290403432e-08,"investigate":0.999999907096586,"monitor":6.267051929026273e-08},"predicted":"investigate","confidence":0.999999907096586,"entropy_confidence":0.9999987905262014,"candidate_mass":0.9999999210611494,"accepted":true,"accepted_predicted":"investigate","input_tokens":341,"reused_prefix_tokens":0,"expected":"close_benign","correct":false,"accepted_correct":false,"nll_hard":17.95209583989744,"brier_hard":1.9999997822385522,"soft_accuracy":0.33333299288480495,"brier_soft":0.6594447839795752,"tvd":0.6666669070965865,"kl":5.7740259523802475,"case_id":"typed:337/r0","base_id":"typed:337","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.195402}
{"id":"severity","type":"ordinal","probabilities":{"0":0.0001314327292534079,"1":0.001281845792847566,"2":0.9981802713673057,"3":0.0004060914518814472,"4":3.5865871059220805e-07},"predicted":"2","confidence":0.9981802713673057,"entropy_confidence":0.9908629311387716,"candidate_mass":0.9999988824239586,"accepted":true,"accepted_predicted":"2","input_tokens":361,"reused_prefix_tokens":0,"expected":"0","correct":false,"accepted_correct":false,"nll_hard":8.937015401853142,"brier_hard":1.996102814001996,"score_mae":1.3621950975179455,"within_one":false,"case_id":"typed:337/r0","base_id":"typed:337","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.195402}
{"id":"true_positive","type":"binary","probabilities":{"false":2.0101790380938737e-08,"true":0.9999999798982098},"predicted":"true","confidence":0.9999999798982098,"entropy_confidence":0.9999994570346453,"candidate_mass":0.9999997414388077,"accepted":true,"accepted_predicted":"true","input_tokens":307,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":17.72245695216966,"brier_hard":1.9999999195928397,"soft_accuracy":0.3333330067006102,"brier_soft":0.8888897241731999,"tvd":0.6666669798982097,"kl":5.774025928113063,"case_id":"typed:337/r0","base_id":"typed:337","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.195402}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.498562490279651e-05,"1":3.49630775046157e-05,"2":0.0004819871133269336,"3":0.9994480641842668},"predicted":"3","confidence":0.9994480641842668,"entropy_confidence":0.9964287950467808,"candidate_mass":0.9999943183936932,"accepted":true,"accepted_predicted":"3","input_tokens":336,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":10.261217981743062,"brier_hard":1.9988267416046572,"score_mae":2.1226761298569587,"within_one":false,"case_id":"typed:337/r0","base_id":"typed:337","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.195402}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.401874449496213e-05,"true":0.9999759812555055},"predicted":"true","confidence":0.9999759812555055,"entropy_confidence":0.9995967695356991,"candidate_mass":0.9999908637906233,"accepted":true,"accepted_predicted":"true","input_tokens":300,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.401903294919855e-05,"brier_hard":1.1538001742080824e-09,"soft_accuracy":0.5499975981255508,"brier_soft":0.40495676741370956,"tvd":0.4499759812555052,"kl":3.8158560274418125,"case_id":"typed:338/r0","base_id":"typed:338","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.828801}
{"id":"disposition","type":"choice","probabilities":{"close_benign":9.952989991858205e-05,"contain":6.096093580914284e-05,"investigate":0.9996997850921158,"monitor":0.00013972407215778025},"predicted":"investigate","confidence":0.9996997850921158,"entropy_confidence":0.9978004999465899,"candidate_mass":0.9999888907467611,"accepted":true,"accepted_predicted":"investigate","input_tokens":352,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":8.875840993318803,"brier_hard":1.9991202453141599,"soft_accuracy":0.43327549329976506,"brier_soft":0.5439598624149453,"tvd":0.5663667850921151,"kl":4.084430094587823,"case_id":"typed:338/r0","base_id":"typed:338","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.828801}
{"id":"severity","type":"ordinal","probabilities":{"0":6.992656682696782e-06,"1":4.512700112063046e-05,"2":0.9965330285838426,"3":0.0034145351464965738,"4":3.1661185776452875e-07},"predicted":"2","confidence":0.9965330285838426,"entropy_confidence":0.9854645992226708,"candidate_mass":0.999998094046977,"accepted":true,"accepted_predicted":"2","input_tokens":372,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.679713916309158,"brier_hard":1.986260667901203,"score_mae":0.5399769439442732,"within_one":true,"case_id":"typed:338/r0","base_id":"typed:338","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.828801}
{"id":"true_positive","type":"binary","probabilities":{"false":3.4310406717620223e-06,"true":0.9999965689593281},"predicted":"true","confidence":0.9999965689593281,"entropy_confidence":0.9999327666482641,"candidate_mass":0.9999957721397538,"accepted":true,"accepted_predicted":"true","input_tokens":318,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.431046557941979e-06,"brier_hard":2.354408018357531e-11,"soft_accuracy":0.49999999999999994,"brier_soft":0.4999931379422004,"tvd":0.4999965689593282,"kl":4.258598311231398,"case_id":"typed:338/r0","base_id":"typed:338","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.828801}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00045176322518475484,"1":0.00046545341908378894,"2":0.009050002757002826,"3":0.9900327805987302},"predicted":"3","confidence":0.9900327805987302,"entropy_confidence":0.957044919803845,"candidate_mass":0.9999912288744617,"accepted":true,"accepted_predicted":"3","input_tokens":347,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.704990216629152,"brier_hard":1.9621472244328466,"score_mae":0.8253308007292799,"within_one":true,"case_id":"typed:338/r0","base_id":"typed:338","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.828801}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0002317113827232295,"true":0.9997682886172768},"predicted":"true","confidence":0.9997682886172768,"entropy_confidence":0.9968677460706341,"candidate_mass":0.99998319158464,"accepted":true,"accepted_predicted":"true","input_tokens":296,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00023173823195319092,"brier_hard":1.0738032976698701e-07,"soft_accuracy":0.8764924439372077,"brier_soft":0.030307854518468184,"tvd":0.12310128861727684,"kl":0.6589890105380186,"case_id":"typed:339/r0","base_id":"typed:339","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.853501}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.242284057579786e-06,"contain":1.6259238463599215e-06,"investigate":0.9999947667489351,"monitor":2.365043160972339e-06},"predicted":"investigate","confidence":0.9999947667489351,"entropy_confidence":0.9999463046117606,"candidate_mass":0.9999982419040307,"accepted":true,"accepted_predicted":"investigate","input_tokens":348,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":13.329434382892444,"brier_hard":1.9999862816873446,"soft_accuracy":0.22333313099970162,"brier_soft":1.133745829313634,"tvd":0.7766630090329927,"kl":6.818572026463309,"case_id":"typed:339/r0","base_id":"typed:339","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.853501}
{"id":"severity","type":"ordinal","probabilities":{"0":1.7030132800756376e-06,"1":6.345016004780754e-06,"2":0.00022430073787331142,"3":0.9997646841344052,"4":2.9670984359599313e-06},"predicted":"3","confidence":0.9997646841344052,"entropy_confidence":0.9985980803011925,"candidate_mass":0.9999996535562975,"accepted":true,"accepted_predicted":"3","input_tokens":368,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":12.727926040273026,"brier_hard":1.9995235398082793,"score_mae":0.6835721327112889,"within_one":true,"case_id":"typed:339/r0","base_id":"typed:339","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.853501}
{"id":"true_positive","type":"binary","probabilities":{"false":1.5778422208148953e-08,"true":0.999999984221578},"predicted":"true","confidence":0.999999984221578,"entropy_confidence":0.9999995682997361,"candidate_mass":0.9999996859599807,"accepted":true,"accepted_predicted":"true","input_tokens":314,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.577842213786983e-08,"brier_hard":4.979172086112548e-16,"soft_accuracy":0.9066669871668729,"brier_soft":0.017422091887410614,"tvd":0.09333298422157789,"kl":0.7707934426701016,"case_id":"typed:339/r0","base_id":"typed:339","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.853501}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.051364291767066e-06,"1":5.205288681233759e-06,"2":3.639414204410408e-05,"3":0.999951349204982},"predicted":"3","confidence":0.999951349204982,"entropy_confidence":0.9995905560164432,"candidate_mass":0.9999960764910458,"accepted":true,"accepted_predicted":"3","input_tokens":343,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":4.8651978506348016e-05,"brier_hard":3.768250199643454e-09,"score_mae":0.19326504118771526,"within_one":true,"case_id":"typed:339/r0","base_id":"typed:339","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.853501}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.598248142299104e-05,"true":0.9999440175185765},"predicted":"true","confidence":0.9999440175185765,"entropy_confidence":0.9991285028143136,"candidate_mass":0.9999901020305058,"accepted":true,"accepted_predicted":"true","input_tokens":316,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.598404850106599e-05,"brier_hard":6.2680764526043226e-09,"soft_accuracy":0.5166651338799639,"brier_soft":0.46711335132330156,"tvd":0.48327701751857677,"kl":4.039495506592521,"case_id":"typed:340/r0","base_id":"typed:340","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.470801}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.756153637391798e-06,"contain":9.417958988493306e-06,"investigate":0.9999704648181781,"monitor":1.2361069196101969e-05},"predicted":"investigate","confidence":0.9999704648181781,"entropy_confidence":0.9997334717956325,"candidate_mass":0.9999980721542381,"accepted":true,"accepted_predicted":"investigate","input_tokens":368,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.9535617993986762e-05,"brier_hard":1.1739788676802165e-09,"soft_accuracy":0.6333182254839238,"brier_soft":0.18885965762048743,"tvd":0.36663746481817805,"kl":3.0777752551943887,"case_id":"typed:340/r0","base_id":"typed:340","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.470801}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00012547537765070197,"1":0.0010394098361238813,"2":0.9937826639544478,"3":0.005051305572903648,"4":1.1452588732381523e-06},"predicted":"2","confidence":0.9937826639544478,"entropy_confidence":0.9744056635718992,"candidate_mass":0.9999972825353193,"accepted":true,"accepted_predicted":"2","input_tokens":388,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.288108539819955,"brier_hard":1.9775279838367719,"score_mae":0.7562367645007764,"within_one":true,"case_id":"typed:340/r0","base_id":"typed:340","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.470801}
{"id":"true_positive","type":"binary","probabilities":{"false":4.82031035899968e-07,"true":0.9999995179689651},"predicted":"true","confidence":0.9999995179689651,"entropy_confidence":0.9999891894584736,"candidate_mass":0.999998692383811,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.820311510422123e-07,"brier_hard":4.647078381440271e-13,"soft_accuracy":0.5666669357288745,"brier_soft":0.37555414225864603,"tvd":0.43333251796896455,"kl":3.6692872202560882,"case_id":"typed:340/r0","base_id":"typed:340","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.470801}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.036272651288437e-05,"1":4.967860529040945e-05,"2":0.0007748293166507,"3":0.999095129351547},"predicted":"3","confidence":0.999095129351547,"entropy_confidence":0.9944423709352265,"candidate_mass":0.9999932979714059,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.162867789446306,"brier_hard":1.9966420281472845,"score_mae":0.7655517252932329,"within_one":true,"case_id":"typed:340/r0","base_id":"typed:340","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.470801}
{"id":"credential_compromise","type":"binary","probabilities":{"false":8.773167679096441e-05,"true":0.9999122683232083},"predicted":"true","confidence":0.9999122683232083,"entropy_confidence":0.9986911158147959,"candidate_mass":0.9999843629638792,"accepted":true,"accepted_predicted":"true","input_tokens":323,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.773552544031497e-05,"brier_hard":1.5393694225230197e-08,"soft_accuracy":0.5999824536646414,"brier_soft":0.3198596447108282,"tvd":0.3999122683232087,"kl":3.0635319858745826,"case_id":"typed:341/r0","base_id":"typed:341","repeat":0,"workflow":"security_incidents","case_elapsed_ms":368.237302}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.137185046215745e-06,"contain":5.903321895570071e-06,"investigate":0.9999731586119991,"monitor":1.5800881058741045e-05},"predicted":"investigate","confidence":0.9999731586119991,"entropy_confidence":0.9997582269839014,"candidate_mass":0.9999973251104207,"accepted":true,"accepted_predicted":"investigate","input_tokens":375,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":2.6841748237409937e-05,"brier_hard":1.0313678316495948e-09,"soft_accuracy":0.6166549690910047,"brier_soft":0.21274773562935662,"tvd":0.3833061586119993,"kl":3.23253102984929,"case_id":"typed:341/r0","base_id":"typed:341","repeat":0,"workflow":"security_incidents","case_elapsed_ms":368.237302}
{"id":"severity","type":"ordinal","probabilities":{"0":0.0001260509382365059,"1":0.0017165413014842681,"2":0.9975874124235482,"3":0.0005691694326109955,"4":8.259041216709271e-07},"predicted":"2","confidence":0.9975874124235482,"entropy_confidence":0.9883589999080641,"candidate_mass":0.9999981275360967,"accepted":true,"accepted_predicted":"2","input_tokens":395,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.002415502555230364,"brier_hard":9.106936217917504e-06,"score_mae":0.11526917806290049,"within_one":true,"case_id":"typed:341/r0","base_id":"typed:341","repeat":0,"workflow":"security_incidents","case_elapsed_ms":368.237302}
{"id":"true_positive","type":"binary","probabilities":{"false":3.3614942872117366e-08,"true":0.9999999663850568},"predicted":"true","confidence":0.9999999663850568,"entropy_confidence":0.9999991169684912,"candidate_mass":0.9999996747724196,"accepted":true,"accepted_predicted":"true","input_tokens":341,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.361494372732336e-08,"brier_hard":2.2599287881031404e-15,"soft_accuracy":0.6666669887949964,"brier_soft":0.22222173295812303,"tvd":0.33333296638505694,"kl":2.799800535728182,"case_id":"typed:341/r0","base_id":"typed:341","repeat":0,"workflow":"security_incidents","case_elapsed_ms":368.237302}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0002284155943705496,"1":0.0001994165701848862,"2":0.0038808012130195184,"3":0.9956913666224239},"predicted":"3","confidence":0.9956913666224239,"entropy_confidence":0.97875011061862,"candidate_mass":0.9999949853977733,"accepted":true,"accepted_predicted":"3","input_tokens":370,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.551713648463867,"brier_hard":1.9836548476990983,"score_mae":1.0783681188634957,"within_one":false,"case_id":"typed:341/r0","base_id":"typed:341","repeat":0,"workflow":"security_incidents","case_elapsed_ms":368.237302}
{"id":"credential_compromise","type":"binary","probabilities":{"false":1.626289004449877e-05,"true":0.9999837371099569},"predicted":"true","confidence":0.9999837371099569,"entropy_confidence":0.99971782682115,"candidate_mass":0.9999921133368191,"accepted":true,"accepted_predicted":"true","input_tokens":307,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.6263022285313938e-05,"brier_hard":5.289631851528946e-10,"soft_accuracy":0.6499951211329876,"brier_soft":0.24497723248290187,"tvd":0.34998373710995623,"kl":2.943620805696054,"case_id":"typed:342/r0","base_id":"typed:342","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.278601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.1384261791286544e-07,"contain":8.334775096869277e-08,"investigate":0.9999995242207825,"monitor":2.7858884772460625e-07},"predicted":"investigate","confidence":0.9999995242207825,"entropy_confidence":0.9999943306243639,"candidate_mass":0.9999996702862284,"accepted":true,"accepted_predicted":"investigate","input_tokens":359,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.7577933065107637e-07,"brier_hard":3.238845990958835e-13,"soft_accuracy":0.616666759642295,"brier_soft":0.221021426935299,"tvd":0.38333252422078296,"kl":3.232514770826342,"case_id":"typed:342/r0","base_id":"typed:342","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.278601}
{"id":"severity","type":"ordinal","probabilities":{"0":6.818978647142783e-06,"1":6.330818116635863e-05,"2":0.7912067443385733,"3":0.20872246293299113,"4":6.655686224080887e-07},"predicted":"2","confidence":0.7912067443385733,"entropy_confidence":0.681245277930143,"candidate_mass":0.9999963556473019,"accepted":false,"accepted_predicted":null,"input_tokens":379,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.5667498380460572,"brier_hard":1.2521282570085435,"score_mae":0.5280201520682244,"within_one":true,"case_id":"typed:342/r0","base_id":"typed:342","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.278601}
{"id":"true_positive","type":"binary","probabilities":{"false":9.246228025382655e-09,"true":0.9999999907537713},"predicted":"true","confidence":0.9999999907537713,"entropy_confidence":0.9999997398926669,"candidate_mass":0.9999998125230254,"accepted":true,"accepted_predicted":"true","input_tokens":325,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.246228714346338e-09,"brier_hard":1.709854773448889e-16,"soft_accuracy":0.7766669948837471,"brier_soft":0.09975524951804851,"tvd":0.22333299075377167,"kl":1.8606753461809067,"case_id":"typed:342/r0","base_id":"typed:342","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.278601}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.496112082234248e-07,"1":5.094202933892428e-07,"2":4.3818054845930155e-06,"3":0.9999944591630153},"predicted":"3","confidence":0.9999944591630153,"entropy_confidence":0.9999450043263248,"candidate_mass":0.9999988065300353,"accepted":true,"accepted_predicted":"3","input_tokens":354,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.540852335168284e-06,"brier_hard":5.0582597552710693e-11,"score_mae":0.5833256505203086,"within_one":true,"case_id":"typed:342/r0","base_id":"typed:342","repeat":0,"workflow":"security_incidents","case_elapsed_ms":329.278601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.003419501269393184,"true":0.9965804987306071},"predicted":"true","confidence":0.9965804987306071,"entropy_confidence":0.9670625980255506,"candidate_mass":0.9999455367914701,"accepted":true,"accepted_predicted":"true","input_tokens":313,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.678260566204322,"brier_hard":1.9863453809002904,"soft_accuracy":0.351025850380818,"brier_soft":0.8361326826774409,"tvd":0.646580498730607,"kl":3.0446216053923463,"case_id":"typed:343/r0","base_id":"typed:343","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.27030099999996}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.9986981985256093,"contain":0.00010393347407612106,"investigate":0.0006737455696020337,"monitor":0.0005241224307134959},"predicted":"close_benign","confidence":0.9986981985256093,"entropy_confidence":0.991968914627818,"candidate_mass":0.9999962275180605,"accepted":true,"accepted_predicted":"close_benign","input_tokens":365,"reused_prefix_tokens":0,"expected":"investigate","correct":false,"accepted_correct":false,"nll_hard":7.3026580114925075,"brier_hard":1.9960513400386752,"soft_accuracy":0.11702501256312121,"brier_soft":1.134215370985136,"tvd":0.8820313151924921,"kl":5.522633137896001,"case_id":"typed:343/r0","base_id":"typed:343","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.27030099999996}
{"id":"severity","type":"ordinal","probabilities":{"0":7.374473643666529e-05,"1":0.0008488813350285609,"2":0.9988945747782833,"3":0.00018243035432493243,"4":3.6879592747133245e-07},"predicted":"2","confidence":0.9988945747782833,"entropy_confidence":0.9941684849332276,"candidate_mass":0.9999985005585736,"accepted":true,"accepted_predicted":"2","input_tokens":385,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0011060366548143975,"brier_hard":1.981283698109066e-06,"score_mae":0.03251979713827979,"within_one":true,"case_id":"typed:343/r0","base_id":"typed:343","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.27030099999996}
{"id":"true_positive","type":"binary","probabilities":{"false":0.006155304864068523,"true":0.9938446951359329},"predicted":"true","confidence":0.9938446951359329,"entropy_confidence":0.9459428830989686,"candidate_mass":0.9998990805563585,"accepted":true,"accepted_predicted":"true","input_tokens":331,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.090440989448297,"brier_hard":1.975454556099668,"soft_accuracy":0.4835381809323395,"brier_soft":0.5212443817411259,"tvd":0.5105116951359321,"kl":1.9404556307676848,"case_id":"typed:343/r0","base_id":"typed:343","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.27030099999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006387612584911207,"1":0.9637969469827491,"2":0.0343584703502764,"3":0.0012058214084826355},"predicted":"1","confidence":0.9637969469827491,"entropy_confidence":0.8815825687901789,"candidate_mass":0.9999900721888806,"accepted":true,"accepted_predicted":"1","input_tokens":360,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.370906701184282,"brier_hard":1.8613699808187405,"score_mae":0.8305356480912502,"within_one":true,"case_id":"typed:343/r0","base_id":"typed:343","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.27030099999996}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0008100723447728031,"true":0.9991899276552275},"predicted":"true","confidence":0.9991899276552275,"entropy_confidence":0.9905126172624489,"candidate_mass":0.9999737447755529,"accepted":true,"accepted_predicted":"true","input_tokens":292,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0008104006306765606,"brier_hard":1.3124344075308964e-06,"soft_accuracy":0.5499189927655229,"brier_soft":0.4035431822138167,"tvd":0.4491899276552273,"kl":2.5155810565215804,"case_id":"typed:344/r0","base_id":"typed:344","repeat":0,"workflow":"security_incidents","case_elapsed_ms":311.870802}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.002600412567864335,"contain":0.00015130125882323835,"investigate":0.9665253499740714,"monitor":0.030722936199242157},"predicted":"investigate","confidence":0.9665253499740714,"entropy_confidence":0.8869526158101151,"candidate_mass":0.9999868010187872,"accepted":true,"accepted_predicted":"investigate","input_tokens":344,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0340477520444296,"brier_hard":0.002071236040655137,"soft_accuracy":0.6340216780377184,"brier_soft":0.13930075769136105,"tvd":0.3165253499740708,"kl":0.9930061418374065,"case_id":"typed:344/r0","base_id":"typed:344","repeat":0,"workflow":"security_incidents","case_elapsed_ms":311.870802}
{"id":"severity","type":"ordinal","probabilities":{"0":6.495087640427766e-05,"1":0.00031341961114725114,"2":0.9274373626900283,"3":0.07218327165390204,"4":9.951685196106132e-07},"predicted":"2","confidence":0.9274373626900283,"entropy_confidence":0.8367320332475405,"candidate_mass":0.9999955581053641,"accepted":true,"accepted_predicted":"2","input_tokens":364,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.628546954472977,"brier_hard":1.7209840455637513,"score_mae":0.9015910593730112,"within_one":true,"case_id":"typed:344/r0","base_id":"typed:344","repeat":0,"workflow":"security_incidents","case_elapsed_ms":311.870802}
{"id":"true_positive","type":"binary","probabilities":{"false":0.00017683195107918282,"true":0.9998231680489207},"predicted":"true","confidence":0.9998231680489207,"entropy_confidence":0.9975406387515531,"candidate_mass":0.9999891688924174,"accepted":true,"accepted_predicted":"true","input_tokens":310,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00017684758769217175,"brier_hard":6.25390778449877e-08,"soft_accuracy":0.599964633609784,"brier_soft":0.31971713141735103,"tvd":0.39982316804892076,"kl":2.783218723669256,"case_id":"typed:344/r0","base_id":"typed:344","repeat":0,"workflow":"security_incidents","case_elapsed_ms":311.870802}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.021987951692083712,"1":0.024679317364997705,"2":0.8185163117686295,"3":0.13481641917428824},"predicted":"2","confidence":0.8185163117686295,"entropy_confidence":0.5604398422072354,"candidate_mass":0.9999796824021662,"accepted":true,"accepted_predicted":"2","input_tokens":339,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.0038412839578683,"brier_hard":1.419604119886937,"score_mae":0.3838388015748788,"within_one":true,"case_id":"typed:344/r0","base_id":"typed:344","repeat":0,"workflow":"security_incidents","case_elapsed_ms":311.870802}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.77803878931294e-05,"true":0.999942219612107},"predicted":"true","confidence":0.999942219612107,"entropy_confidence":0.9991031493474514,"candidate_mass":0.9999918257676009,"accepted":true,"accepted_predicted":"true","input_tokens":281,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.75886114967199,"brier_hard":1.9997688851255742,"soft_accuracy":0.41667663002612865,"brier_soft":0.6804199636271031,"tvd":0.5832752196121069,"kl":5.0079409570288576,"case_id":"typed:345/r0","base_id":"typed:345","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.563801}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.00012099493401200619,"contain":5.685219506203442e-05,"investigate":0.9954120974436099,"monitor":0.0044100554273155046},"predicted":"investigate","confidence":0.9954120974436099,"entropy_confidence":0.9782557007251573,"candidate_mass":0.999991987823935,"accepted":true,"accepted_predicted":"investigate","input_tokens":333,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.004598459282519199,"brier_hard":4.0515310685065485e-05,"soft_accuracy":0.5324399071170827,"brier_soft":0.33974013374173967,"tvd":0.4620790974436102,"kl":1.9642658379219242,"case_id":"typed:345/r0","base_id":"typed:345","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.563801}
{"id":"severity","type":"ordinal","probabilities":{"0":1.7017145063182562e-05,"1":0.00010828971344764872,"2":0.9998414235426354,"3":3.320645231027431e-05,"4":6.314654469084092e-08},"predicted":"2","confidence":0.9998414235426354,"entropy_confidence":0.9989575912150025,"candidate_mass":0.9999997054085349,"accepted":true,"accepted_predicted":"2","input_tokens":353,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":9.130700390484577,"brier_hard":1.999466305923786,"score_mae":0.6465580087418281,"within_one":true,"case_id":"typed:345/r0","base_id":"typed:345","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.563801}
{"id":"true_positive","type":"binary","probabilities":{"false":3.026082655793257e-06,"true":0.9999969739173435},"predicted":"true","confidence":0.9999969739173435,"entropy_confidence":0.9999401537381718,"candidate_mass":0.9999980392945249,"accepted":true,"accepted_predicted":"true","input_tokens":299,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.708241627771805,"brier_hard":1.99998789568769,"soft_accuracy":0.4500003026082653,"brier_soft":0.604993342636471,"tvd":0.5499969739173439,"kl":4.70636010302816,"case_id":"typed:345/r0","base_id":"typed:345","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.563801}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0015133847029196478,"1":0.005336378613281787,"2":0.9737374684900947,"3":0.019412768193704532},"predicted":"2","confidence":0.9737374684900947,"entropy_confidence":0.8988744826418453,"candidate_mass":0.9999920589228781,"accepted":true,"accepted_predicted":"2","input_tokens":328,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.026613551209802473,"brier_hard":0.0010973434002145986,"score_mae":0.344382620174585,"within_one":true,"case_id":"typed:345/r0","base_id":"typed:345","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.563801}
{"id":"credential_compromise","type":"binary","probabilities":{"false":7.369970346979412e-05,"true":0.9999263002965286},"predicted":"true","confidence":0.9999263002965286,"entropy_confidence":0.9988819295439233,"candidate_mass":0.9999885253973579,"accepted":true,"accepted_predicted":"true","input_tokens":281,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.370241942796505e-05,"brier_hard":1.0863292583304172e-08,"soft_accuracy":0.8166203234720014,"brier_soft":0.06716794229034706,"tvd":0.18325930029652937,"kl":1.2681574911329565,"case_id":"typed:346/r0","base_id":"typed:346","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.511601}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.3045839991863295e-05,"contain":0.00016522508779121583,"investigate":0.9996963408029196,"monitor":8.538826929586998e-05},"predicted":"investigate","confidence":0.9996963408029196,"entropy_confidence":0.9977893788644334,"candidate_mass":0.9999932895630381,"accepted":true,"accepted_predicted":"investigate","input_tokens":333,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":8.708201845264483,"brier_hard":1.9990623610435119,"soft_accuracy":0.2667012766806741,"brier_soft":0.9848565467916497,"tvd":0.7330296074696537,"kl":5.603366170646043,"case_id":"typed:346/r0","base_id":"typed:346","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.511601}
{"id":"severity","type":"ordinal","probabilities":{"0":1.6196534843754232e-06,"1":3.0539976111986367e-06,"2":5.6402140162612155e-05,"3":0.9999374124870763,"4":1.5117216652595173e-06},"predicted":"3","confidence":0.9999374124870763,"entropy_confidence":0.9995681681230508,"candidate_mass":0.9999996327793673,"accepted":true,"accepted_predicted":"3","input_tokens":353,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":6.258947160376419e-05,"brier_hard":7.112633670102978e-09,"score_mae":0.20006585737417382,"within_one":true,"case_id":"typed:346/r0","base_id":"typed:346","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.511601}
{"id":"true_positive","type":"binary","probabilities":{"false":9.718636278874031e-08,"true":0.9999999028136374},"predicted":"true","confidence":0.9999999028136374,"entropy_confidence":0.9999975958653892,"candidate_mass":0.9999995349523745,"accepted":true,"accepted_predicted":"true","input_tokens":299,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.718636734683015e-08,"brier_hard":1.889037819223401e-14,"soft_accuracy":0.8833329254905202,"brier_soft":0.027222332424253363,"tvd":0.11666690281363726,"kl":0.9649627343844017,"case_id":"typed:346/r0","base_id":"typed:346","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.511601}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.055132366694082e-07,"1":5.974783286782112e-07,"2":3.756395657934837e-06,"3":0.9999947406127773},"predicted":"3","confidence":0.9999947406127773,"entropy_confidence":0.9999470915737585,"candidate_mass":0.9999984900209387,"accepted":true,"accepted_predicted":"3","input_tokens":328,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.259401053359031e-06,"brier_hard":4.294859687262833e-11,"score_mae":0.3099923321079765,"within_one":true,"case_id":"typed:346/r0","base_id":"typed:346","repeat":0,"workflow":"security_incidents","case_elapsed_ms":332.511601}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.017186951366486618,"true":0.9828130486335129},"predicted":"true","confidence":0.9828130486335129,"entropy_confidence":0.8746595482840497,"candidate_mass":0.9999085128482449,"accepted":true,"accepted_predicted":"true","input_tokens":266,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.063604824560141,"brier_hard":1.9318429771286005,"soft_accuracy":0.4839059098368502,"brier_soft":0.4989606379658733,"tvd":0.49948004863351314,"kl":1.4153182496205265,"case_id":"typed:347/r0","base_id":"typed:347","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.295202}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.00343411461116003,"contain":0.00024119180305928815,"investigate":0.9887341198611004,"monitor":0.00759057372468175},"predicted":"investigate","confidence":0.9887341198611004,"entropy_confidence":0.9496897046055387,"candidate_mass":0.9999796349287803,"accepted":true,"accepted_predicted":"investigate","input_tokens":318,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.011329820852409996,"brier_hard":0.00019638818142232769,"soft_accuracy":0.5627359121312918,"brier_soft":0.2649707591970395,"tvd":0.4220671198610996,"kl":1.350697128841214,"case_id":"typed:347/r0","base_id":"typed:347","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.295202}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00018739953073306878,"1":0.002757568832208766,"2":0.996245842008574,"3":0.0008082541474276231,"4":9.354810570153444e-07},"predicted":"2","confidence":0.996245842008574,"entropy_confidence":0.9829908592202599,"candidate_mass":0.9999966702550167,"accepted":true,"accepted_predicted":"2","input_tokens":338,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.893405845331742,"brier_hard":1.9869989326350455,"score_mae":0.5676777572158678,"within_one":true,"case_id":"typed:347/r0","base_id":"typed:347","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.295202}
{"id":"true_positive","type":"binary","probabilities":{"false":0.0017575835359028705,"true":0.9982424164640963},"predicted":"true","confidence":0.9982424164640963,"entropy_confidence":0.9813808323085536,"candidate_mass":0.9999604372355964,"accepted":true,"accepted_predicted":"true","input_tokens":284,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0017591298980192509,"brier_hard":6.1781997713566505e-06,"soft_accuracy":0.6328643122288143,"brier_soft":0.2663177644483351,"tvd":0.3649094164640967,"kl":1.6700239341902634,"case_id":"typed:347/r0","base_id":"typed:347","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.295202}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.002614091077232483,"1":0.037382489754196696,"2":0.9406121171875265,"3":0.01939130198104589},"predicted":"2","confidence":0.9406121171875265,"entropy_confidence":0.8034673498705178,"candidate_mass":0.9999745631068169,"accepted":true,"accepted_predicted":"2","input_tokens":313,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.06122442715194865,"brier_hard":0.005307227229850891,"score_mae":0.2934476300723874,"within_one":true,"case_id":"typed:347/r0","base_id":"typed:347","repeat":0,"workflow":"security_incidents","case_elapsed_ms":307.295202}
{"id":"credential_compromise","type":"binary","probabilities":{"false":1.6820664285614534e-05,"true":0.9999831793357145},"predicted":"true","confidence":0.9999831793357145,"entropy_confidence":0.9997089673745417,"candidate_mass":0.9999930745956823,"accepted":true,"accepted_predicted":"true","input_tokens":325,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.992902410405364,"brier_hard":1.9999327179087272,"soft_accuracy":0.40000336413285714,"brier_soft":0.719959630971584,"tvd":0.5999831793357144,"kl":5.159694658758349,"case_id":"typed:348/r0","base_id":"typed:348","repeat":0,"workflow":"security_incidents","case_elapsed_ms":352.386501}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.8378144772688085e-06,"contain":3.4865075824604244e-06,"investigate":0.9999888767918675,"monitor":4.798886072075628e-06},"predicted":"investigate","confidence":0.9999888767918675,"entropy_confidence":0.9998918301799417,"candidate_mass":0.9999984697948837,"accepted":true,"accepted_predicted":"investigate","input_tokens":377,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.112326999579581e-05,"brier_hard":1.6696399282068025e-10,"soft_accuracy":0.5999950317246018,"brier_soft":0.24007670807949533,"tvd":0.3999888767918679,"kl":3.3776474484928762,"case_id":"typed:348/r0","base_id":"typed:348","repeat":0,"workflow":"security_incidents","case_elapsed_ms":352.386501}
{"id":"severity","type":"ordinal","probabilities":{"0":9.536224267408777e-06,"1":6.590174691948108e-05,"2":0.9980473072940811,"3":0.0018770877364568922,"4":1.669982751784901e-07},"predicted":"2","confidence":0.9980473072940811,"entropy_confidence":0.9910015088476091,"candidate_mass":0.9999985183047925,"accepted":true,"accepted_predicted":"2","input_tokens":397,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.278033779552285,"brier_hard":1.9923477800164302,"score_mae":0.6615405524624469,"within_one":true,"case_id":"typed:348/r0","base_id":"typed:348","repeat":0,"workflow":"security_incidents","case_elapsed_ms":352.386501}
{"id":"true_positive","type":"binary","probabilities":{"false":1.0715166498319044e-07,"true":0.9999998928483351},"predicted":"true","confidence":0.9999998928483351,"entropy_confidence":0.9999973644400691,"candidate_mass":0.999999560990656,"accepted":true,"accepted_predicted":"true","input_tokens":343,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.04902057634036,"brier_hard":1.9999995713933632,"soft_accuracy":0.4500000107151666,"brier_soft":0.6049997642663603,"tvd":0.5499998928483352,"kl":4.706358789507156,"case_id":"typed:348/r0","base_id":"typed:348","repeat":0,"workflow":"security_incidents","case_elapsed_ms":352.386501}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.251815412671071e-05,"1":2.995156911611766e-05,"2":0.0007711789237180752,"3":0.9991563513530389},"predicted":"3","confidence":0.9991563513530389,"entropy_confidence":0.9948706922587663,"candidate_mass":0.9999946230801658,"accepted":true,"accepted_predicted":"3","input_tokens":372,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":7.167590144245128,"brier_hard":1.9967716540235037,"score_mae":1.0990413634756693,"within_one":false,"case_id":"typed:348/r0","base_id":"typed:348","repeat":0,"workflow":"security_incidents","case_elapsed_ms":352.386501}
{"id":"credential_compromise","type":"binary","probabilities":{"false":1.900893189204004e-05,"true":0.9999809910681089},"predicted":"true","confidence":0.9999809910681089,"entropy_confidence":0.9996744597932554,"candidate_mass":0.9999936701474964,"accepted":true,"accepted_predicted":"true","input_tokens":306,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.9009112563142272e-05,"brier_hard":7.226789833169565e-10,"soft_accuracy":0.5000000000000004,"brier_soft":0.4999619828588958,"tvd":0.49998099106810845,"kl":4.258606100264401,"case_id":"typed:349/r0","base_id":"typed:349","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.168301}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0016431390079470532,"contain":0.00023100033033817777,"investigate":0.992395181459574,"monitor":0.005730679202142014},"predicted":"investigate","confidence":0.992395181459574,"entropy_confidence":0.9642025989293258,"candidate_mass":0.9999846231081817,"accepted":true,"accepted_predicted":"investigate","input_tokens":358,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.007633882618062413,"brier_hard":9.34272161027244e-05,"soft_accuracy":0.7287046082800794,"brier_soft":0.09500847530932774,"tvd":0.25906144812584,"kl":0.8320953708650247,"case_id":"typed:349/r0","base_id":"typed:349","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.168301}
{"id":"severity","type":"ordinal","probabilities":{"0":7.564188256321453e-05,"1":0.0006037818329005092,"2":0.9972977035125852,"3":0.0020219308592488,"4":9.419127019599153e-07},"predicted":"2","confidence":0.9972977035125852,"entropy_confidence":0.9872947053382662,"candidate_mass":0.9999973281322124,"accepted":true,"accepted_predicted":"2","input_tokens":378,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.203702353207465,"brier_hard":1.9905633061924615,"score_mae":0.8287312509133749,"within_one":true,"case_id":"typed:349/r0","base_id":"typed:349","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.168301}
{"id":"true_positive","type":"binary","probabilities":{"false":5.689324430586214e-06,"true":0.9999943106755697},"predicted":"true","confidence":0.9999943106755697,"entropy_confidence":0.9998926651930383,"candidate_mass":0.9999953959568026,"accepted":true,"accepted_predicted":"true","input_tokens":324,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.689340614614487e-06,"brier_hard":6.473682495020683e-11,"soft_accuracy":0.5499994310675571,"brier_soft":0.4049897592807619,"tvd":0.4499943106755695,"kl":3.815845946111029,"case_id":"typed:349/r0","base_id":"typed:349","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.168301}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0018019722981646247,"1":0.0010018034085981678,"2":0.019510379491555065,"3":0.9776858448016817},"predicted":"3","confidence":0.9776858448016817,"entropy_confidence":0.9154748089994,"candidate_mass":0.9999881858068052,"accepted":true,"accepted_predicted":"3","input_tokens":353,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.02256688263435434,"brier_hard":0.0008828271443519577,"score_mae":0.4630800967967539,"within_one":true,"case_id":"typed:349/r0","base_id":"typed:349","repeat":0,"workflow":"security_incidents","case_elapsed_ms":334.168301}
{"id":"credential_compromise","type":"binary","probabilities":{"false":3.293373629418237e-05,"true":0.9999670662637058},"predicted":"true","confidence":0.9999670662637058,"entropy_confidence":0.999462101664599,"candidate_mass":0.9999901621603445,"accepted":true,"accepted_predicted":"true","input_tokens":309,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.321013006300106,"brier_hard":1.999868267224085,"soft_accuracy":0.4333373911867951,"brier_soft":0.6421483301010834,"tvd":0.5666340662637057,"kl":4.8568360206722545,"case_id":"typed:350/r0","base_id":"typed:350","repeat":0,"workflow":"security_incidents","case_elapsed_ms":328.055502}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.554291433970404e-05,"contain":9.952388355425969e-05,"investigate":0.9996536554889948,"monitor":0.00017127771311173167},"predicted":"investigate","confidence":0.9996536554889948,"entropy_confidence":0.9975000038888163,"candidate_mass":0.9999892952463741,"accepted":true,"accepted_predicted":"investigate","input_tokens":361,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00034640450211751643,"brier_hard":1.649023106168988e-07,"soft_accuracy":0.6498192734530635,"brier_soft":0.16966892897417316,"tvd":0.3496536554889945,"kl":2.1589347172508533,"case_id":"typed:350/r0","base_id":"typed:350","repeat":0,"workflow":"security_incidents","case_elapsed_ms":328.055502}
{"id":"severity","type":"ordinal","probabilities":{"0":1.2206314785200515e-05,"1":0.00011702834660138823,"2":0.46342922313436813,"3":0.5364401191635754,"4":1.423040669857808e-06},"predicted":"3","confidence":0.5364401191635754,"entropy_confidence":0.570200276024643,"candidate_mass":0.9999948972769435,"accepted":false,"accepted_predicted":null,"input_tokens":381,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.6228003370016709,"brier_hard":0.42965442182265723,"score_mae":0.07296852426874345,"within_one":true,"case_id":"typed:350/r0","base_id":"typed:350","repeat":0,"workflow":"security_incidents","case_elapsed_ms":328.055502}
{"id":"true_positive","type":"binary","probabilities":{"false":9.148644220800906e-07,"true":0.9999990851355768},"predicted":"true","confidence":0.9999990851355768,"entropy_confidence":0.9999803280060074,"candidate_mass":0.9999986921725988,"accepted":true,"accepted_predicted":"true","input_tokens":327,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":9.148648417069096e-07,"brier_hard":1.673953823658304e-12,"soft_accuracy":0.7666665120716974,"brier_soft":0.10888772390743265,"tvd":0.23333208513557735,"kl":1.945371535781382,"case_id":"typed:350/r0","base_id":"typed:350","repeat":0,"workflow":"security_incidents","case_elapsed_ms":328.055502}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.3889378100456856e-06,"1":2.8041370423394458e-06,"2":2.188781759176007e-05,"3":0.9999719191075574},"predicted":"3","confidence":0.9999719191075574,"entropy_confidence":0.9997536883164223,"candidate_mass":0.9999978780765371,"accepted":true,"accepted_predicted":"3","input_tokens":356,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.8081286718191464e-05,"brier_hard":1.2869611633328113e-09,"score_mae":0.6999623370948984,"within_one":true,"case_id":"typed:350/r0","base_id":"typed:350","repeat":0,"workflow":"security_incidents","case_elapsed_ms":328.055502}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0001473418225999334,"true":0.9998526581774013},"predicted":"true","confidence":0.9998526581774013,"entropy_confidence":0.9979119992679842,"candidate_mass":0.9999723209737965,"accepted":true,"accepted_predicted":"true","input_tokens":317,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.822755346758061,"brier_hard":1.9994106761288282,"soft_accuracy":0.1667652277835056,"brier_soft":1.3883966819850169,"tvd":0.8331856581774006,"kl":6.90175599486359,"case_id":"typed:351/r0","base_id":"typed:351","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.915501}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.000764951002260014,"contain":0.006322985190634768,"investigate":0.9915581282286814,"monitor":0.0013539355784241967},"predicted":"investigate","confidence":0.9915581282286814,"entropy_confidence":0.9604308946090406,"candidate_mass":0.9999880203365049,"accepted":true,"accepted_predicted":"investigate","input_tokens":369,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.008477706186361922,"brier_hard":0.00011366363231075406,"soft_accuracy":0.43057703262690883,"brier_soft":0.453186610391856,"tvd":0.5582251282286812,"kl":2.626193276120423,"case_id":"typed:351/r0","base_id":"typed:351","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.915501}
{"id":"severity","type":"ordinal","probabilities":{"0":0.0001317945816187079,"1":0.0007461596167612628,"2":0.9811840306915774,"3":0.017936733860298586,"4":1.2812497440563062e-06},"predicted":"2","confidence":0.9811840306915774,"entropy_confidence":0.9395271057208424,"candidate_mass":0.9999959008293396,"accepted":true,"accepted_predicted":"2","input_tokens":389,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":7.200571017160488,"brier_hard":1.9615520833978501,"score_mae":0.366929547579788,"within_one":true,"case_id":"typed:351/r0","base_id":"typed:351","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.915501}
{"id":"true_positive","type":"binary","probabilities":{"false":5.037363291531987e-07,"true":0.9999994962636717},"predicted":"true","confidence":0.9999994962636717,"entropy_confidence":0.9999887346812395,"candidate_mass":0.9999990131928487,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.501212862211187,"brier_hard":1.9999979850551926,"soft_accuracy":0.4333330671651801,"brier_soft":0.6422218359754907,"tvd":0.5666664962636712,"kl":4.856821967448126,"case_id":"typed:351/r0","base_id":"typed:351","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.915501}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.813391679313065e-06,"1":6.885730773512754e-06,"2":4.386462472200683e-05,"3":0.9999394362528249},"predicted":"3","confidence":0.9999394362528249,"entropy_confidence":0.9994981389705309,"candidate_mass":0.99999645092804,"accepted":true,"accepted_predicted":"3","input_tokens":364,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":11.886059291472346,"brier_hard":1.9998651067798914,"score_mae":1.6499129237386922,"within_one":false,"case_id":"typed:351/r0","base_id":"typed:351","repeat":0,"workflow":"security_incidents","case_elapsed_ms":316.915501}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.008938436359582722,"true":0.9910615636404189},"predicted":"true","confidence":0.9910615636404189,"entropy_confidence":0.9263294984008517,"candidate_mass":0.9999147807640814,"accepted":true,"accepted_predicted":"true","input_tokens":248,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.717394608943984,"brier_hard":1.9644060458507808,"soft_accuracy":0.22173210797813966,"brier_soft":1.1993738803916671,"tvd":0.7743945636404181,"kl":3.1745792342004053,"case_id":"typed:352/r0","base_id":"typed:352","repeat":0,"workflow":"security_incidents","case_elapsed_ms":305.491701}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.500474898137229e-06,"contain":6.809895486715694e-06,"investigate":0.9999690483211132,"monitor":1.664130850352882e-05},"predicted":"investigate","confidence":0.9999690483211132,"entropy_confidence":0.9997232955636881,"candidate_mass":0.9999966167902837,"accepted":true,"accepted_predicted":"investigate","input_tokens":300,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.095215789987856e-05,"brier_hard":1.337571374857628e-09,"soft_accuracy":0.4166609221493002,"brier_soft":0.4899495736812342,"tvd":0.5833024649876958,"kl":4.992486586053303,"case_id":"typed:352/r0","base_id":"typed:352","repeat":0,"workflow":"security_incidents","case_elapsed_ms":305.491701}
{"id":"severity","type":"ordinal","probabilities":{"0":3.974331052719234e-06,"1":3.538011882669593e-05,"2":0.8513199094177144,"3":0.14864035277190477,"4":3.83360501181009e-07},"predicted":"2","confidence":0.8513199094177144,"entropy_confidence":0.738545721788089,"candidate_mass":0.9999964855856344,"accepted":true,"accepted_predicted":"2","input_tokens":320,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":1.9062256305885263,"brier_hard":1.4495588383670273,"score_mae":0.3080692092880253,"within_one":true,"case_id":"typed:352/r0","base_id":"typed:352","repeat":0,"workflow":"security_incidents","case_elapsed_ms":305.491701}
{"id":"true_positive","type":"binary","probabilities":{"false":1.0486359489518894e-07,"true":0.9999998951364069},"predicted":"true","confidence":0.9999998951364069,"entropy_confidence":0.9999974174531829,"candidate_mass":0.9999994792682975,"accepted":true,"accepted_predicted":"true","input_tokens":266,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.07060542758222,"brier_hard":1.9999995805456459,"soft_accuracy":0.40000002097271975,"brier_soft":0.7199997483273963,"tvd":0.599999895136406,"kl":5.159687972381487,"case_id":"typed:352/r0","base_id":"typed:352","repeat":0,"workflow":"security_incidents","case_elapsed_ms":305.491701}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.983148747361846e-05,"1":7.079666114350217e-05,"2":0.0017542615807955568,"3":0.9980751102705876},"predicted":"3","confidence":0.9980751102705876,"entropy_confidence":0.9894313727895854,"candidate_mass":0.9999952951964035,"accepted":true,"accepted_predicted":"3","input_tokens":295,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.3457072622885775,"brier_hard":1.9926484949922416,"score_mae":1.1144716506344978,"within_one":false,"case_id":"typed:352/r0","base_id":"typed:352","repeat":0,"workflow":"security_incidents","case_elapsed_ms":305.491701}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0003579322886001768,"true":0.9996420677114011},"predicted":"true","confidence":0.9996420677114011,"entropy_confidence":0.9953860872552543,"candidate_mass":0.9999693937590394,"accepted":true,"accepted_predicted":"true","input_tokens":234,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.935166727440752,"brier_hard":1.9985685270766482,"soft_accuracy":0.31679824160053227,"brier_soft":0.932909886230784,"tvd":0.6829750677114004,"kl":4.798143088047236,"case_id":"typed:353/r0","base_id":"typed:353","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.386201}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.108874656668148e-05,"contain":6.732274113865358e-05,"investigate":0.9997252524795979,"monitor":0.00013633603269805732},"predicted":"investigate","confidence":0.9997252524795979,"entropy_confidence":0.9979702244519253,"candidate_mass":0.9999902395070377,"accepted":true,"accepted_predicted":"investigate","input_tokens":286,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":8.900387891044343,"brier_hard":1.999177936553475,"soft_accuracy":0.34997928179962773,"brier_soft":0.6533810227976153,"tvd":0.6497252524795972,"kl":4.7786358803703495,"case_id":"typed:353/r0","base_id":"typed:353","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.386201}
{"id":"severity","type":"ordinal","probabilities":{"0":4.97710574043914e-05,"1":0.000555754657708665,"2":0.9982413858784196,"3":0.001152784350653407,"4":3.040558127724108e-07},"predicted":"2","confidence":0.9982413858784196,"entropy_confidence":0.9911649295816015,"candidate_mass":0.9999978911266357,"accepted":true,"accepted_predicted":"2","input_tokens":306,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0017601622987584683,"brier_hard":4.7329758779033775e-06,"score_mae":0.3671650956897592,"within_one":true,"case_id":"typed:353/r0","base_id":"typed:353","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.386201}
{"id":"true_positive","type":"binary","probabilities":{"false":3.470552729014816e-06,"true":0.99999652944727},"predicted":"true","confidence":0.99999652944727,"entropy_confidence":0.9999320497161298,"candidate_mass":0.9999968385288712,"accepted":true,"accepted_predicted":"true","input_tokens":252,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.571196688789954,"brier_hard":1.9999861178131715,"soft_accuracy":0.38333380979795006,"brier_soft":0.7605478171007294,"tvd":0.6166635294472705,"kl":5.312154983416431,"case_id":"typed:353/r0","base_id":"typed:353","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.386201}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00041561468356255836,"1":0.0003854136715923923,"2":0.006351249983353221,"3":0.9928477216614928},"predicted":"3","confidence":0.9928477216614928,"entropy_confidence":0.9671614074229189,"candidate_mass":0.9999903410520944,"accepted":true,"accepted_predicted":"3","input_tokens":281,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":7.861193328839466,"brier_hard":1.9850164307208469,"score_mae":1.6249640786227775,"within_one":false,"case_id":"typed:353/r0","base_id":"typed:353","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.386201}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0009387016672702605,"true":0.9990612983327282},"predicted":"true","confidence":0.9990612983327282,"entropy_confidence":0.9892058100551048,"candidate_mass":0.9999667167769197,"accepted":true,"accepted_predicted":"true","input_tokens":246,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0009391425235918719,"brier_hard":1.76232164027489e-06,"soft_accuracy":0.5332704205146489,"brier_soft":0.43380569573579886,"tvd":0.465728298332729,"kl":2.5627191720346465,"case_id":"typed:354/r0","base_id":"typed:354","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.809701}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.8295765248588402e-06,"contain":1.827131050420288e-06,"investigate":0.9999906963165944,"monitor":5.646975831068271e-06},"predicted":"investigate","confidence":0.9999906963165944,"entropy_confidence":0.999909213605938,"candidate_mass":0.9999980708110697,"accepted":true,"accepted_predicted":"investigate","input_tokens":298,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":9.303726685101313e-06,"brier_hard":1.2513261908348005e-10,"soft_accuracy":0.6166628545432081,"brier_soft":0.2138781614499051,"tvd":0.3833236963165941,"kl":3.2325202147301537,"case_id":"typed:354/r0","base_id":"typed:354","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.809701}
{"id":"severity","type":"ordinal","probabilities":{"0":6.156480453767402e-06,"1":2.5506613066130355e-05,"2":0.0005817393449638705,"3":0.9993827779951315,"4":3.81956638560858e-06},"predicted":"3","confidence":0.9993827779951315,"entropy_confidence":0.9966808409899421,"candidate_mass":0.9999993346556775,"accepted":true,"accepted_predicted":"3","input_tokens":318,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0006174125647859983,"brier_hard":7.200867474219003e-07,"score_mae":0.25268559755393083,"within_one":true,"case_id":"typed:354/r0","base_id":"typed:354","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.809701}
{"id":"true_positive","type":"binary","probabilities":{"false":8.25616548888614e-08,"true":0.9999999174383446},"predicted":"true","confidence":0.9999999174383446,"entropy_confidence":0.9999979382168657,"candidate_mass":0.9999995033502626,"accepted":true,"accepted_predicted":"true","input_tokens":264,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.256165880942795e-08,"brier_hard":1.3632853800576309e-14,"soft_accuracy":0.6333329779836134,"brier_soft":0.268889256687476,"tvd":0.3666669174383448,"kl":3.0878474227993498,"case_id":"typed:354/r0","base_id":"typed:354","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.809701}
{"id":"urgency","type":"ordinal","probabilities":{"0":6.5474486480233295e-06,"1":4.817713354069518e-06,"2":4.3350573745391776e-05,"3":0.999945284264253},"predicted":"3","confidence":0.999945284264253,"entropy_confidence":0.9995474555321997,"candidate_mass":0.999997098508348,"accepted":true,"accepted_predicted":"3","input_tokens":293,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":5.471723270747612e-05,"brier_hard":4.9391634281508674e-09,"score_mae":0.5665943716536037,"within_one":true,"case_id":"typed:354/r0","base_id":"typed:354","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.809701}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.008259719108152134,"true":0.9917402808918485},"predicted":"true","confidence":0.9917402808918485,"entropy_confidence":0.9309783840843092,"candidate_mass":0.9999423556295882,"accepted":true,"accepted_predicted":"true","input_tokens":241,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.008294019593520796,"brier_hard":0.00013644591949113627,"soft_accuracy":0.5327823575659363,"brier_soft":0.4202744703493157,"tvd":0.45840728089184823,"kl":1.551805245191459,"case_id":"typed:355/r0","base_id":"typed:355","repeat":0,"workflow":"security_incidents","case_elapsed_ms":268.754901}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.7433034230845503e-06,"contain":1.447113269110985e-06,"investigate":0.999992076705316,"monitor":4.732877993005445e-06},"predicted":"investigate","confidence":0.999992076705316,"entropy_confidence":0.999921714678076,"candidate_mass":0.9999977160537737,"accepted":true,"accepted_predicted":"investigate","input_tokens":293,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":13.44593983479458,"brier_hard":1.9999812592744057,"soft_accuracy":0.2666659007223147,"brier_soft":0.9866519787235413,"tvd":0.7333270866754719,"kl":6.401782859535628,"case_id":"typed:355/r0","base_id":"typed:355","repeat":0,"workflow":"security_incidents","case_elapsed_ms":268.754901}
{"id":"severity","type":"ordinal","probabilities":{"0":4.278767485481508e-06,"1":1.4929809048132778e-05,"2":0.0003834723829965138,"3":0.99959327343241,"4":4.0456080596370175e-06},"predicted":"3","confidence":0.99959327343241,"entropy_confidence":0.9977059292772207,"candidate_mass":0.9999995290022569,"accepted":true,"accepted_predicted":"3","input_tokens":313,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":12.417878694984427,"brier_hard":1.999178768383844,"score_mae":0.18375512269549032,"within_one":true,"case_id":"typed:355/r0","base_id":"typed:355","repeat":0,"workflow":"security_incidents","case_elapsed_ms":268.754901}
{"id":"true_positive","type":"binary","probabilities":{"false":1.7783011807940699e-07,"true":0.9999998221698811},"predicted":"true","confidence":0.9999998221698811,"entropy_confidence":0.9999957559611918,"candidate_mass":0.9999992486712637,"accepted":true,"accepted_predicted":"true","input_tokens":259,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.7783013466441734e-07,"brier_hard":6.324710206727986e-14,"soft_accuracy":0.8666668695911275,"brier_soft":0.035555282935570524,"tvd":0.13333282216988157,"kl":1.1040220214140761,"case_id":"typed:355/r0","base_id":"typed:355","repeat":0,"workflow":"security_incidents","case_elapsed_ms":268.754901}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.774018036869428e-05,"1":1.150437766089177e-05,"2":7.110786074638197e-05,"3":0.9998996475812231},"predicted":"3","confidence":0.9998996475812231,"entropy_confidence":0.9992033225528839,"candidate_mass":0.9999943420551863,"accepted":true,"accepted_predicted":"3","input_tokens":288,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0001003574544178003,"brier_hard":1.5574000519184975e-08,"score_mae":0.31651966284282285,"within_one":true,"case_id":"typed:355/r0","base_id":"typed:355","repeat":0,"workflow":"security_incidents","case_elapsed_ms":268.754901}
{"id":"credential_compromise","type":"binary","probabilities":{"false":4.873158859642716e-05,"true":0.9999512684114035},"predicted":"true","confidence":0.9999512684114035,"entropy_confidence":0.9992316274492058,"candidate_mass":0.9999875742515127,"accepted":true,"accepted_predicted":"true","input_tokens":299,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.929183101672503,"brier_hard":1.9998050783951495,"soft_accuracy":0.1333687365307918,"brier_soft":1.5020544462887588,"tvd":0.8666182684114034,"kl":7.713651159075829,"case_id":"typed:356/r0","base_id":"typed:356","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.268702}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.8802287200106583,"contain":0.00044782342153869,"investigate":0.11767707659250023,"monitor":0.0016463799753043801},"predicted":"close_benign","confidence":0.8802287200106583,"entropy_confidence":0.7272543768065762,"candidate_mass":0.999984404637956,"accepted":true,"accepted_predicted":"close_benign","input_tokens":351,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":6.409176355674454,"brier_hard":1.7853606450491903,"soft_accuracy":0.32844873561516974,"brier_soft":0.48564491154745953,"tvd":0.5302287200106576,"kl":2.4036849578516035,"case_id":"typed:356/r0","base_id":"typed:356","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.268702}
{"id":"severity","type":"ordinal","probabilities":{"0":3.6053874642278697e-05,"1":0.0001922786187482568,"2":0.9996569223318938,"3":0.00011459285491135649,"4":1.5231980309797297e-07},"predicted":"2","confidence":0.9996569223318938,"entropy_confidence":0.9978878754879624,"candidate_mass":0.9999991887198676,"accepted":true,"accepted_predicted":"2","input_tokens":371,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":8.55656509850781,"brier_hard":1.9989294565310722,"score_mae":0.4831835111264824,"within_one":true,"case_id":"typed:356/r0","base_id":"typed:356","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.268702}
{"id":"true_positive","type":"binary","probabilities":{"false":0.0004106096976757947,"true":0.9995893903023242},"predicted":"true","confidence":0.9995893903023242,"entropy_confidence":0.9947884006781911,"candidate_mass":0.999982600856196,"accepted":true,"accepted_predicted":"true","input_tokens":317,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.797867435353929,"brier_hard":1.9983578984099444,"soft_accuracy":0.18359305308226778,"brier_soft":1.3325489894187605,"tvd":0.8162563903023242,"kl":5.89192627781845,"case_id":"typed:356/r0","base_id":"typed:356","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.268702}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.003573297899442996,"1":0.08757754200394233,"2":0.8933795090154777,"3":0.015469651081138206},"predicted":"2","confidence":0.8933795090154777,"entropy_confidence":0.7124575477761321,"candidate_mass":0.9999853708508255,"accepted":true,"accepted_predicted":"2","input_tokens":346,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.435230683726161,"brier_hard":1.630893767546754,"score_mae":0.8040785132783121,"within_one":true,"case_id":"typed:356/r0","base_id":"typed:356","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.268702}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0001570224807487232,"true":0.9998429775192512},"predicted":"true","confidence":0.9998429775192512,"entropy_confidence":0.9977892297715641,"candidate_mass":0.9999792986782308,"accepted":true,"accepted_predicted":"true","input_tokens":315,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00015703481006918156,"brier_hard":4.931211892098438e-08,"soft_accuracy":0.8932094757531593,"brier_soft":0.022688750622302793,"tvd":0.1065099775192512,"kl":0.5949598543951906,"case_id":"typed:357/r0","base_id":"typed:357","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.076201}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.317655574209308e-05,"contain":2.956082123938147e-05,"investigate":0.9998143065724918,"monitor":0.00010295605052631793},"predicted":"investigate","confidence":0.9998143065724918,"entropy_confidence":0.99858429294764,"candidate_mass":0.9999936729723016,"accepted":true,"accepted_predicted":"investigate","input_tokens":367,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":10.42906068015309,"brier_hard":1.9995695402860905,"soft_accuracy":0.349957725680586,"brier_soft":0.7844021283453974,"tvd":0.649814306572492,"kl":5.4580617224915,"case_id":"typed:357/r0","base_id":"typed:357","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.076201}
{"id":"severity","type":"ordinal","probabilities":{"0":2.241924175759336e-05,"1":8.771446390683896e-05,"2":0.9997122727368462,"3":0.0001772130984398322,"4":3.8045904873890347e-07},"predicted":"2","confidence":0.9997122727368462,"entropy_confidence":0.9982083846615217,"candidate_mass":0.999998152050177,"accepted":true,"accepted_predicted":"2","input_tokens":387,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0002877686645845658,"brier_hard":1.223880545491864e-07,"score_mae":0.3999545789308865,"within_one":true,"case_id":"typed:357/r0","base_id":"typed:357","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.076201}
{"id":"true_positive","type":"binary","probabilities":{"false":3.4094704682114056e-06,"true":0.9999965905295305},"predicted":"true","confidence":0.9999965905295305,"entropy_confidence":0.9999331583084116,"candidate_mass":0.9999976761551147,"accepted":true,"accepted_predicted":"true","input_tokens":333,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.4094762817641118e-06,"brier_hard":2.32489777560423e-11,"soft_accuracy":0.9499969314765774,"brier_soft":0.004999318129155215,"tvd":0.04999659052953119,"kl":0.4117916279331043,"case_id":"typed:357/r0","base_id":"typed:357","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.076201}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00013410288693439394,"1":0.00010768093835140495,"2":0.004631719354990451,"3":0.9951264968197233},"predicted":"3","confidence":0.9951264968197233,"entropy_confidence":0.9769631241557988,"candidate_mass":0.9999929882274805,"accepted":true,"accepted_predicted":"3","input_tokens":362,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.004885417422059435,"brier_hard":4.523343620032849e-05,"score_mae":0.13475061010750267,"within_one":true,"case_id":"typed:357/r0","base_id":"typed:357","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.076201}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00011366960215314173,"true":0.9998863303978486},"predicted":"true","confidence":0.9998863303978486,"entropy_confidence":0.9983466211973764,"candidate_mass":0.999984798881702,"accepted":true,"accepted_predicted":"true","input_tokens":304,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.082214544288128,"brier_hard":1.9995453474329479,"soft_accuracy":0.3933572495909064,"brier_soft":0.7358138852334414,"tvd":0.6065533303978478,"kl":4.839708909054349,"case_id":"typed:358/r0","base_id":"typed:358","repeat":0,"workflow":"security_incidents","case_elapsed_ms":581.989902}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.00016538351473985012,"contain":6.764953548127732e-05,"investigate":0.9992625980785692,"monitor":0.0005043688712082571},"predicted":"investigate","confidence":0.9992625980785692,"entropy_confidence":0.9951987470540866,"candidate_mass":0.9999913726427371,"accepted":true,"accepted_predicted":"investigate","input_tokens":356,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0007376739359585672,"brier_hard":8.300777185722533e-07,"soft_accuracy":0.44987984932006697,"brier_soft":0.45896597386207033,"tvd":0.5492630480781199,"kl":3.282550210384434,"case_id":"typed:358/r0","base_id":"typed:358","repeat":0,"workflow":"security_incidents","case_elapsed_ms":581.989902}
{"id":"severity","type":"ordinal","probabilities":{"0":9.795904638832768e-05,"1":0.000317895323802568,"2":0.9977944650467275,"3":0.0017890508582758663,"4":6.297248075041472e-07},"predicted":"2","confidence":0.9977944650467275,"entropy_confidence":0.9894408657209148,"candidate_mass":0.9999981226270834,"accepted":true,"accepted_predicted":"2","input_tokens":376,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.3260700465332675,"brier_hard":1.9920190041181152,"score_mae":0.772056503108685,"within_one":true,"case_id":"typed:358/r0","base_id":"typed:358","repeat":0,"workflow":"security_incidents","case_elapsed_ms":581.989902}
{"id":"true_positive","type":"binary","probabilities":{"false":9.547571464212867e-05,"true":0.999904524285359},"predicted":"true","confidence":0.999904524285359,"entropy_confidence":0.998587233168555,"candidate_mass":0.9999896440348608,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.2566386397576,"brier_hard":1.999618115372658,"soft_accuracy":0.4900019095142934,"brier_soft":0.5200052477733554,"tvd":0.5099045242853585,"kl":4.02798532438554,"case_id":"typed:358/r0","base_id":"typed:358","repeat":0,"workflow":"security_incidents","case_elapsed_ms":581.989902}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.004428439861207621,"1":0.003373139458606162,"2":0.7914351149166905,"3":0.2007633057634973},"predicted":"2","confidence":0.7914351149166905,"entropy_confidence":0.6027724654680431,"candidate_mass":0.9999849970932911,"accepted":false,"accepted_predicted":null,"input_tokens":351,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.23390738038502334,"brier_hard":0.08383620538031315,"score_mae":0.07479971341752067,"within_one":true,"case_id":"typed:358/r0","base_id":"typed:358","repeat":0,"workflow":"security_incidents","case_elapsed_ms":581.989902}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00021847836116951143,"true":0.9997815216388312},"predicted":"true","confidence":0.9997815216388312,"entropy_confidence":0.9970280914161321,"candidate_mass":0.999971164287385,"accepted":true,"accepted_predicted":"true","input_tokens":323,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00021850223104275245,"brier_hard":9.546558859833714e-08,"soft_accuracy":0.5832965870854577,"brier_soft":0.34685874235033554,"tvd":0.41644852163883084,"kl":2.8329467168052265,"case_id":"typed:359/r0","base_id":"typed:359","repeat":0,"workflow":"security_incidents","case_elapsed_ms":717.651403}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.717528277630692e-07,"contain":6.374709472876837e-07,"investigate":0.9999968506218831,"monitor":1.8401543419485634e-06},"predicted":"investigate","confidence":0.9999968506218831,"entropy_confidence":0.999966751860287,"candidate_mass":0.9999990973185422,"accepted":true,"accepted_predicted":"investigate","input_tokens":375,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.1493830761529054e-06,"brier_hard":1.4162371595337096e-11,"soft_accuracy":0.5666652470111871,"brier_soft":0.2765296941691706,"tvd":0.4333304172883164,"kl":3.6692937061363513,"case_id":"typed:359/r0","base_id":"typed:359","repeat":0,"workflow":"security_incidents","case_elapsed_ms":717.651403}
{"id":"severity","type":"ordinal","probabilities":{"0":4.066653870101989e-06,"1":2.2318144775046286e-05,"2":0.9995319765321481,"3":0.0004415696543772802,"4":6.901483127565693e-08},"predicted":"2","confidence":0.9995319765321481,"entropy_confidence":0.9974091808402265,"candidate_mass":0.9999990630598765,"accepted":true,"accepted_predicted":"2","input_tokens":395,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":7.7251747828052935,"brier_hard":1.9981812282999099,"score_mae":0.6295887437684717,"within_one":true,"case_id":"typed:359/r0","base_id":"typed:359","repeat":0,"workflow":"security_incidents","case_elapsed_ms":717.651403}
{"id":"true_positive","type":"binary","probabilities":{"false":1.0362001397480028e-07,"true":0.9999998963799843},"predicted":"true","confidence":0.9999998963799843,"entropy_confidence":0.9999974462962604,"candidate_mass":0.9999993670462799,"accepted":true,"accepted_predicted":"true","input_tokens":341,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.0362002104753879e-07,"brier_hard":2.1474214945450818e-14,"soft_accuracy":0.6666669654599251,"brier_soft":0.22222163961813984,"tvd":0.3333328963799852,"kl":2.799800582398257,"case_id":"typed:359/r0","base_id":"typed:359","repeat":0,"workflow":"security_incidents","case_elapsed_ms":717.651403}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.9821285476950404e-05,"1":6.328025073391843e-05,"2":0.0033056790217585457,"3":0.9965812194420292},"predicted":"3","confidence":0.9965812194420292,"entropy_confidence":0.9831199518422619,"candidate_mass":0.9999956023842625,"accepted":true,"accepted_predicted":"3","input_tokens":370,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.712113373904021,"brier_hard":1.9865737029013901,"score_mae":0.9797512966203388,"within_one":true,"case_id":"typed:359/r0","base_id":"typed:359","repeat":0,"workflow":"security_incidents","case_elapsed_ms":717.651403}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.671672125366513e-06,"true":0.9999943283278734},"predicted":"true","confidence":0.9999943283278734,"entropy_confidence":0.9998929727940481,"candidate_mass":0.9999959058173079,"accepted":true,"accepted_predicted":"true","input_tokens":258,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.671688210610183e-06,"brier_hard":6.433572940950382e-11,"soft_accuracy":0.5566663572067106,"brier_soft":0.3930782400846612,"tvd":0.443327328327874,"kl":3.757162101461566,"case_id":"typed:360/r0","base_id":"typed:360","repeat":0,"workflow":"security_incidents","case_elapsed_ms":683.7985960000001}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.86300588552355e-06,"contain":2.8333030211288565e-06,"investigate":0.999985715254817,"monitor":8.588436276689806e-06},"predicted":"investigate","confidence":0.999985715254817,"entropy_confidence":0.9998649603290432,"candidate_mass":0.9999972766307146,"accepted":true,"accepted_predicted":"investigate","input_tokens":310,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.4284847210890524e-05,"brier_hard":2.940395913305529e-10,"soft_accuracy":0.5866612306804638,"brier_soft":0.25013816722074594,"tvd":0.41331871525481684,"kl":3.4940777116098367,"case_id":"typed:360/r0","base_id":"typed:360","repeat":0,"workflow":"security_incidents","case_elapsed_ms":683.7985960000001}
{"id":"severity","type":"ordinal","probabilities":{"0":3.855447829920841e-06,"1":2.9430296271584605e-05,"2":0.9926950199015957,"3":0.0072715339339360285,"4":1.6042036608155057e-07},"predicted":"2","confidence":0.9926950199015957,"entropy_confidence":0.9730096095307339,"candidate_mass":0.999997974521436,"accepted":true,"accepted_predicted":"2","input_tokens":330,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":4.923788014653688,"brier_hard":1.9709532107563423,"score_mae":0.7360982864172643,"within_one":true,"case_id":"typed:360/r0","base_id":"typed:360","repeat":0,"workflow":"security_incidents","case_elapsed_ms":683.7985960000001}
{"id":"true_positive","type":"binary","probabilities":{"false":2.5723902869548516e-07,"true":0.999999742760971},"predicted":"true","confidence":0.999999742760971,"entropy_confidence":0.9999939978206566,"candidate_mass":0.9999991252082967,"accepted":true,"accepted_predicted":"true","input_tokens":276,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.572390620527583e-07,"brier_hard":1.323438359079755e-13,"soft_accuracy":0.6433329262583165,"brier_soft":0.25442233078344145,"tvd":0.3566667427609711,"kl":3.0012550781395086,"case_id":"typed:360/r0","base_id":"typed:360","repeat":0,"workflow":"security_incidents","case_elapsed_ms":683.7985960000001}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00015294478256846746,"1":0.0001390250611964656,"2":0.006235346332455886,"3":0.9934726838237802},"predicted":"3","confidence":0.9934726838237802,"entropy_confidence":0.9706091154749378,"candidate_mass":0.9999963261677235,"accepted":true,"accepted_predicted":"3","input_tokens":305,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.006548712261281494,"brier_hard":8.152812042416761e-05,"score_mae":0.7263607691974485,"within_one":true,"case_id":"typed:360/r0","base_id":"typed:360","repeat":0,"workflow":"security_incidents","case_elapsed_ms":683.7985960000001}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00038986301696638365,"true":0.9996101369830322},"predicted":"true","confidence":0.9996101369830322,"entropy_confidence":0.9950225569317463,"candidate_mass":0.9999729649070174,"accepted":true,"accepted_predicted":"true","input_tokens":256,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0003899390333116999,"brier_hard":3.039863439973275e-07,"soft_accuracy":0.5666150180044951,"brier_soft":0.3748795197214184,"tvd":0.4329431369830329,"kl":2.7175298915727812,"case_id":"typed:361/r0","base_id":"typed:361","repeat":0,"workflow":"security_incidents","case_elapsed_ms":897.1776729999999}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.135165982352575e-05,"contain":1.752571630804422e-05,"investigate":0.9998828069154567,"monitor":7.831570841102106e-05},"predicted":"investigate","confidence":0.9998828069154567,"entropy_confidence":0.9990772483722794,"candidate_mass":0.999993789776439,"accepted":true,"accepted_predicted":"investigate","input_tokens":308,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00011719995218937475,"brier_hard":2.0630613358017286e-08,"soft_accuracy":0.6499423077913511,"brier_soft":0.17488101887882468,"tvd":0.3498828069154571,"kl":2.541990229536593,"case_id":"typed:361/r0","base_id":"typed:361","repeat":0,"workflow":"security_incidents","case_elapsed_ms":897.1776729999999}
{"id":"severity","type":"ordinal","probabilities":{"0":1.472739613925611e-05,"1":7.931332831048645e-05,"2":0.999815928486616,"3":8.987428954594759e-05,"4":1.5649938938699184e-07},"predicted":"2","confidence":0.999815928486616,"entropy_confidence":0.9987967144549225,"candidate_mass":0.9999990884325525,"accepted":true,"accepted_predicted":"2","input_tokens":328,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00018408845662418454,"brier_hard":4.8467234697636567e-08,"score_mae":0.3833144191677378,"within_one":true,"case_id":"typed:361/r0","base_id":"typed:361","repeat":0,"workflow":"security_incidents","case_elapsed_ms":897.1776729999999}
{"id":"true_positive","type":"binary","probabilities":{"false":1.1526978927216947e-06,"true":0.9999988473021055},"predicted":"true","confidence":0.9999988473021055,"entropy_confidence":0.9999755982520593,"candidate_mass":0.9999982451789555,"accepted":true,"accepted_predicted":"true","input_tokens":274,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.15269855882543e-06,"brier_hard":2.6574248677976134e-12,"soft_accuracy":0.6499996541906311,"brier_soft":0.24499838622560632,"tvd":0.3499988473021064,"kl":2.943610983985632,"case_id":"typed:361/r0","base_id":"typed:361","repeat":0,"workflow":"security_incidents","case_elapsed_ms":897.1776729999999}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0006808172599588237,"1":0.00036336024749352987,"2":0.006129582361343234,"3":0.992826240131203},"predicted":"3","confidence":0.992826240131203,"entropy_confidence":0.966660446611834,"candidate_mass":0.999989923912007,"accepted":true,"accepted_predicted":"3","input_tokens":303,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.0946286616415755,"brier_hard":1.9734829456931102,"score_mae":1.1411012453637892,"within_one":false,"case_id":"typed:361/r0","base_id":"typed:361","repeat":0,"workflow":"security_incidents","case_elapsed_ms":897.1776729999999}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0003718839155626256,"true":0.9996281160844385},"predicted":"true","confidence":0.9996281160844385,"entropy_confidence":0.9952267632953782,"candidate_mass":0.9999778501869473,"accepted":true,"accepted_predicted":"true","input_tokens":254,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.89692880732958,"brier_hard":1.998512740933045,"soft_accuracy":0.4666917920131154,"brier_soft":0.5680951025159395,"tvd":0.532961116084438,"kl":3.520942956004148,"case_id":"typed:362/r0","base_id":"typed:362","repeat":0,"workflow":"security_incidents","case_elapsed_ms":372.173289}
{"id":"disposition","type":"choice","probabilities":{"close_benign":9.957857217223509e-05,"contain":0.00011109184466120691,"investigate":0.9997242405497826,"monitor":6.508903338354049e-05},"predicted":"investigate","confidence":0.9997242405497826,"entropy_confidence":0.9979569670335601,"candidate_mass":0.9999884860562632,"accepted":true,"accepted_predicted":"investigate","input_tokens":306,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00027579747884592736,"brier_hard":1.0253714663710012e-07,"soft_accuracy":0.616528223779614,"brier_soft":0.21361461385548386,"tvd":0.38305724054978285,"kl":2.6029614916775348,"case_id":"typed:362/r0","base_id":"typed:362","repeat":0,"workflow":"security_incidents","case_elapsed_ms":372.173289}
{"id":"severity","type":"ordinal","probabilities":{"0":6.803074304445934e-06,"1":6.813622312305684e-05,"2":0.0066541288683972626,"3":0.9932690282096182,"4":1.9036245564486575e-06},"predicted":"3","confidence":0.9932690282096182,"entropy_confidence":0.9746359126680716,"candidate_mass":0.9999991664538995,"accepted":true,"accepted_predicted":"3","input_tokens":326,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.006753726948034012,"brier_hard":8.958810469066194e-05,"score_mae":0.049858093086997624,"within_one":true,"case_id":"typed:362/r0","base_id":"typed:362","repeat":0,"workflow":"security_incidents","case_elapsed_ms":372.173289}
{"id":"true_positive","type":"binary","probabilities":{"false":7.984443048496162e-07,"true":0.9999992015556946},"predicted":"true","confidence":0.9999992015556946,"entropy_confidence":0.9999826745575451,"candidate_mass":0.9999984349649783,"accepted":true,"accepted_predicted":"true","input_tokens":272,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":7.984446242066655e-07,"brier_hard":1.2750266168520663e-12,"soft_accuracy":0.5166669733846572,"brier_soft":0.4672200341213497,"tvd":0.4833322015556949,"kl":4.110477333921915,"case_id":"typed:362/r0","base_id":"typed:362","repeat":0,"workflow":"security_incidents","case_elapsed_ms":372.173289}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.0252552992670073e-05,"1":1.5306218624044445e-05,"2":0.00012911624340805137,"3":0.9998353249849736},"predicted":"3","confidence":0.9998353249849736,"entropy_confidence":0.9987668930959083,"candidate_mass":0.999996513753354,"accepted":true,"accepted_predicted":"3","input_tokens":301,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0001646885754453876,"brier_hard":4.4433311117030984e-08,"score_mae":0.6164465136603612,"within_one":true,"case_id":"typed:362/r0","base_id":"typed:362","repeat":0,"workflow":"security_incidents","case_elapsed_ms":372.173289}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.14387258478113604,"true":0.8561274152188637},"predicted":"true","confidence":0.8561274152188637,"entropy_confidence":0.4057084045191066,"candidate_mass":0.9998398103018173,"accepted":true,"accepted_predicted":"true","input_tokens":255,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":1.938827199006198,"brier_hard":1.4659083021786659,"soft_accuracy":0.321936292390568,"brier_soft":0.7347808869598018,"tvd":0.6061274152188638,"kl":0.9306192707265009,"case_id":"typed:363/r0","base_id":"typed:363","repeat":0,"workflow":"security_incidents","case_elapsed_ms":430.869887}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.4304893326257732e-06,"contain":2.3478364816763256e-06,"investigate":0.9999874080067289,"monitor":7.813667456770043e-06},"predicted":"investigate","confidence":0.9999874080067289,"entropy_confidence":0.9998800179450502,"candidate_mass":0.9999979812031213,"accepted":true,"accepted_predicted":"investigate","input_tokens":307,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":11.759636119599406,"brier_hard":1.9999591889095756,"soft_accuracy":0.3999985692573232,"brier_soft":0.5421998555078428,"tvd":0.5999874080067289,"kl":5.159692967265069,"case_id":"typed:363/r0","base_id":"typed:363","repeat":0,"workflow":"security_incidents","case_elapsed_ms":430.869887}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00013174945817555792,"1":0.0037399353019383647,"2":0.9955742597369199,"3":0.0005537530558034536,"4":3.0244716424593093e-07},"predicted":"2","confidence":0.9955742597369199,"entropy_confidence":0.9809552089991271,"candidate_mass":0.9999987406992552,"accepted":true,"accepted_predicted":"2","input_tokens":327,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":5.588686966652286,"brier_hard":1.98370254716336,"score_mae":0.9465509237318455,"within_one":true,"case_id":"typed:363/r0","base_id":"typed:363","repeat":0,"workflow":"security_incidents","case_elapsed_ms":430.869887}
{"id":"true_positive","type":"binary","probabilities":{"false":9.693762057139248e-08,"true":0.9999999030623787},"predicted":"true","confidence":0.9999999030623787,"entropy_confidence":0.9999976016602156,"candidate_mass":0.9999995769033496,"accepted":true,"accepted_predicted":"true","input_tokens":273,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.14919815220496,"brier_hard":1.9999996122495352,"soft_accuracy":0.3666670258499673,"brier_soft":0.8022211322028414,"tvd":0.633332903062379,"kl":5.465335093331348,"case_id":"typed:363/r0","base_id":"typed:363","repeat":0,"workflow":"security_incidents","case_elapsed_ms":430.869887}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008764109878434202,"1":0.0006626951697522919,"2":0.04630379293536754,"3":0.9521571009070353},"predicted":"3","confidence":0.9521571009070353,"entropy_confidence":0.8557524042911226,"candidate_mass":0.9999898906628079,"accepted":true,"accepted_predicted":"3","input_tokens":302,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":7.319195447636218,"brier_hard":1.9074230029694947,"score_mae":1.6497415837615932,"within_one":false,"case_id":"typed:363/r0","base_id":"typed:363","repeat":0,"workflow":"security_incidents","case_elapsed_ms":430.869887}
{"id":"credential_compromise","type":"binary","probabilities":{"false":9.660159613928157e-05,"true":0.9999033984038601},"predicted":"true","confidence":0.9999033984038601,"entropy_confidence":0.998572207272343,"candidate_mass":0.999983138106488,"accepted":true,"accepted_predicted":"true","input_tokens":304,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.24491529370134,"brier_hard":1.9996136122791783,"soft_accuracy":0.27337679278798405,"brier_soft":1.0558090876734887,"tvd":0.7265703984038603,"kl":6.131455458555902,"case_id":"typed:364/r0","base_id":"typed:364","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.83259}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.054412776450400636,"contain":0.000577384969042815,"investigate":0.6814748573686676,"monitor":0.2635349812118889},"predicted":"investigate","confidence":0.6814748573686676,"entropy_confidence":0.44059888597449237,"candidate_mass":0.999968890675604,"accepted":false,"accepted_predicted":null,"input_tokens":356,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":null,"nll_hard":0.38349592160217233,"brier_hard":0.173870036425105,"soft_accuracy":0.41878013808376985,"brier_soft":0.0678815327729006,"tvd":0.22147485736866765,"kl":0.22578493918241468,"case_id":"typed:364/r0","base_id":"typed:364","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.83259}
{"id":"severity","type":"ordinal","probabilities":{"0":7.584896582251457e-05,"1":0.00038092891563062107,"2":0.99684103708843,"3":0.002701470094345711,"4":7.149357721869857e-07},"predicted":"2","confidence":0.99684103708843,"entropy_confidence":0.9857968768187286,"candidate_mass":0.9999976445465885,"accepted":true,"accepted_predicted":"2","input_tokens":376,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.913959174759672,"brier_hard":1.9882965618359314,"score_mae":0.5844967268813832,"within_one":true,"case_id":"typed:364/r0","base_id":"typed:364","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.83259}
{"id":"true_positive","type":"binary","probabilities":{"false":5.316798488168006e-05,"true":0.9999468320151185},"predicted":"true","confidence":0.9999468320151185,"entropy_confidence":0.9991683602374327,"candidate_mass":0.9999913725416384,"accepted":true,"accepted_predicted":"true","input_tokens":322,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":9.84205413070206,"brier_hard":1.999787333714143,"soft_accuracy":0.33335072269707255,"brier_soft":0.8887480020677611,"tvd":0.6666138320151185,"kl":5.7740436445275325,"case_id":"typed:364/r0","base_id":"typed:364","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.83259}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0025398546738922183,"1":0.001310330999915541,"2":0.23715923017956675,"3":0.7589905841466246},"predicted":"3","confidence":0.7589905841466246,"entropy_confidence":0.5856175870362785,"candidate_mass":0.9999853792615994,"accepted":false,"accepted_predicted":null,"input_tokens":351,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.4390235061443306,"brier_hard":1.1580009147525594,"score_mae":0.9359335437989225,"within_one":true,"case_id":"typed:364/r0","base_id":"typed:364","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.83259}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.004057385093313133,"true":0.9959426149066866},"predicted":"true","confidence":0.9959426149066866,"entropy_confidence":0.9619214476475793,"candidate_mass":0.9999407571144041,"accepted":true,"accepted_predicted":"true","input_tokens":320,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.004065638612914273,"brier_hard":3.292474759088171e-05,"soft_accuracy":0.5661260126159681,"brier_soft":0.3685551071070279,"tvd":0.42927561490668664,"kl":1.7045308691617986,"case_id":"typed:365/r0","base_id":"typed:365","repeat":0,"workflow":"security_incidents","case_elapsed_ms":341.75039000000004}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.10743264897567e-05,"contain":4.6876209416162305e-05,"investigate":0.9997846316584815,"monitor":9.741780561185811e-05},"predicted":"investigate","confidence":0.9997846316584815,"entropy_confidence":0.99836882148978,"candidate_mass":0.9999897450941386,"accepted":true,"accepted_predicted":"investigate","input_tokens":372,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":9.968000273128407,"brier_hard":1.9994755740208212,"soft_accuracy":0.2666468714663751,"brier_soft":0.9854751717742076,"tvd":0.7331178983252153,"kl":6.2034741790847026,"case_id":"typed:365/r0","base_id":"typed:365","repeat":0,"workflow":"security_incidents","case_elapsed_ms":341.75039000000004}
{"id":"severity","type":"ordinal","probabilities":{"0":1.4775477592605617e-05,"1":0.00010720738667259987,"2":0.007363402244923224,"3":0.9925101320623675,"4":4.4828284424903035e-06},"predicted":"3","confidence":0.9925101320623675,"entropy_confidence":0.9721488867505849,"candidate_mass":0.9999986966059428,"accepted":true,"accepted_predicted":"3","input_tokens":392,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":12.315256361984012,"brier_hard":1.985121628014028,"score_mae":0.2709506606226082,"within_one":true,"case_id":"typed:365/r0","base_id":"typed:365","repeat":0,"workflow":"security_incidents","case_elapsed_ms":341.75039000000004}
{"id":"true_positive","type":"binary","probabilities":{"false":5.444152607225306e-07,"true":0.9999994555847377},"predicted":"true","confidence":0.9999994555847377,"entropy_confidence":0.999987885952641,"candidate_mass":0.9999990367242141,"accepted":true,"accepted_predicted":"true","input_tokens":338,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.444154105097877e-07,"brier_hard":5.927759539499023e-13,"soft_accuracy":0.9166665463202517,"brier_soft":0.01388859630756482,"tvd":0.08333245558473848,"kl":0.6877656686879733,"case_id":"typed:365/r0","base_id":"typed:365","repeat":0,"workflow":"security_incidents","case_elapsed_ms":341.75039000000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.8544612171343548e-05,"1":2.3147721778056187e-05,"2":0.00021975963315606375,"3":0.9997285480328936},"predicted":"3","confidence":0.9997285480328936,"entropy_confidence":0.9980752953812589,"candidate_mass":0.999995310734671,"accepted":true,"accepted_predicted":"3","input_tokens":367,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00027148881686042795,"brier_hard":1.23331078718364e-07,"score_mae":0.33631531108677093,"within_one":true,"case_id":"typed:365/r0","base_id":"typed:365","repeat":0,"workflow":"security_incidents","case_elapsed_ms":341.75039000000004}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0012972609607475372,"true":0.9987027390392516},"predicted":"true","confidence":0.9987027390392516,"entropy_confidence":0.9856885205399182,"candidate_mass":0.9999411332495604,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":6.647500190534515,"brier_hard":1.9948143219290086,"soft_accuracy":0.41688320929528355,"brier_soft":0.6775312030379366,"tvd":0.582035739039252,"kl":3.199053727234133,"case_id":"typed:366/r0","base_id":"typed:366","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.781889}
{"id":"disposition","type":"choice","probabilities":{"close_benign":8.49287978704899e-06,"contain":9.002020063793337e-06,"investigate":0.9999604115295612,"monitor":2.2093570587138444e-05},"predicted":"investigate","confidence":0.9999604115295612,"entropy_confidence":0.9996536183005541,"candidate_mass":0.9999964816011343,"accepted":true,"accepted_predicted":"investigate","input_tokens":394,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.958925408301194e-05,"brier_hard":2.2085382252814934e-09,"soft_accuracy":0.5999830900649764,"brier_soft":0.2341546451377077,"tvd":0.39996041152956163,"kl":3.3567962094453083,"case_id":"typed:366/r0","base_id":"typed:366","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.781889}
{"id":"severity","type":"ordinal","probabilities":{"0":2.0032627779167535e-05,"1":0.00016656954476651766,"2":0.9965074703072923,"3":0.003305579403807402,"4":3.481163555050801e-07},"predicted":"2","confidence":0.9965074703072923,"entropy_confidence":0.9850634613267509,"candidate_mass":0.9999980275183324,"accepted":true,"accepted_predicted":"2","input_tokens":414,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0034986428121861423,"brier_hard":2.3152765689925426e-05,"score_mae":0.2602333591638044,"within_one":true,"case_id":"typed:366/r0","base_id":"typed:366","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.781889}
{"id":"true_positive","type":"binary","probabilities":{"false":1.4443258513137378e-06,"true":0.9999985556741485},"predicted":"true","confidence":0.9999985556741485,"entropy_confidence":0.9999698946650527,"candidate_mass":0.9999986138465978,"accepted":true,"accepted_predicted":"true","input_tokens":360,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4443268945136135e-06,"brier_hard":4.172154330009325e-12,"soft_accuracy":0.5499998555674148,"brier_soft":0.4049974002176396,"tvd":0.4499985556741486,"kl":3.8158436113534826,"case_id":"typed:366/r0","base_id":"typed:366","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.781889}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0019695640371883536,"1":0.0037936689448789065,"2":0.7164182664457946,"3":0.2778185005721373},"predicted":"2","confidence":0.7164182664457946,"entropy_confidence":0.5468759606034646,"candidate_mass":0.9999849437699351,"accepted":false,"accepted_predicted":null,"input_tokens":389,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":null,"nll_hard":0.3334911115692769,"brier_hard":0.1576199899723189,"score_mae":0.22008570355288049,"within_one":true,"case_id":"typed:366/r0","base_id":"typed:366","repeat":0,"workflow":"security_incidents","case_elapsed_ms":340.781889}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0026946727184155063,"true":0.997305327281586},"predicted":"true","confidence":0.997305327281586,"entropy_confidence":0.9731168031724413,"candidate_mass":0.9999364465930995,"accepted":true,"accepted_predicted":"true","input_tokens":316,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.002698309884394281,"brier_hard":1.4522522118737233e-05,"soft_accuracy":0.6160382412359221,"brier_soft":0.2897710723914465,"tvd":0.3806383272815853,"kl":1.6039731785347011,"case_id":"typed:367/r0","base_id":"typed:367","repeat":0,"workflow":"security_incidents","case_elapsed_ms":317.19989100000004}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.004032604077375445,"contain":0.00021493881915904511,"investigate":0.995397619595945,"monitor":0.0003548375075211649},"predicted":"investigate","confidence":0.995397619595945,"entropy_confidence":0.9773072216760981,"candidate_mass":0.9999855823342942,"accepted":true,"accepted_predicted":"investigate","input_tokens":368,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0046130039650584586,"brier_hard":3.761590938122011e-05,"soft_accuracy":0.5310031504674708,"brier_soft":0.3666268164334162,"tvd":0.46276368700678344,"kl":2.969768955473042,"case_id":"typed:367/r0","base_id":"typed:367","repeat":0,"workflow":"security_incidents","case_elapsed_ms":317.19989100000004}
{"id":"severity","type":"ordinal","probabilities":{"0":1.2218795672537433e-05,"1":4.5682804678639775e-05,"2":0.000974406180022177,"3":0.9989487732253676,"4":1.8918994258117143e-05},"predicted":"3","confidence":0.9989487732253676,"entropy_confidence":0.9946519108945627,"candidate_mass":0.9999986544678905,"accepted":true,"accepted_predicted":"3","input_tokens":388,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":10.87534415335826,"brier_hard":1.9978617656015003,"score_mae":0.38441650918214165,"within_one":true,"case_id":"typed:367/r0","base_id":"typed:367","repeat":0,"workflow":"security_incidents","case_elapsed_ms":317.19989100000004}
{"id":"true_positive","type":"binary","probabilities":{"false":0.00037481312182558607,"true":0.9996251868781749},"predicted":"true","confidence":0.9996251868781749,"entropy_confidence":0.9951934094361581,"candidate_mass":0.9999758049621194,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0003748833818200194,"brier_hard":2.8096975258490347e-07,"soft_accuracy":0.8164296181063021,"brier_soft":0.06694739629149815,"tvd":0.18295818687817464,"kl":0.9702253873835112,"case_id":"typed:367/r0","base_id":"typed:367","repeat":0,"workflow":"security_incidents","case_elapsed_ms":317.19989100000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.5290560552799972e-05,"1":1.2044662290219913e-05,"2":0.00012516258500729112,"3":0.9998475021921497},"predicted":"3","confidence":0.9998475021921497,"entropy_confidence":0.9988579923499863,"candidate_mass":0.9999964298691275,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00015250943682329238,"brier_hard":3.930012921656352e-08,"score_mae":0.3731378764087543,"within_one":true,"case_id":"typed:367/r0","base_id":"typed:367","repeat":0,"workflow":"security_incidents","case_elapsed_ms":317.19989100000004}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.9489296393858985e-05,"true":0.9999705107036062},"predicted":"true","confidence":0.9999705107036062,"entropy_confidence":0.9995136589013318,"candidate_mass":0.9999932976572486,"accepted":true,"accepted_predicted":"true","input_tokens":291,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.431483194565196,"brier_hard":1.9998820445536618,"soft_accuracy":0.25001474464819695,"brier_soft":1.1249115338500557,"tvd":0.7499705107036061,"kl":6.561189061134968,"case_id":"typed:368/r0","base_id":"typed:368","repeat":0,"workflow":"security_incidents","case_elapsed_ms":335.92469}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0299738065115405e-06,"contain":1.4352120565721806e-06,"investigate":0.9999961238606812,"monitor":1.410953454180933e-06},"predicted":"investigate","confidence":0.9999961238606812,"entropy_confidence":0.999959321587553,"candidate_mass":0.9999994182103209,"accepted":true,"accepted_predicted":"investigate","input_tokens":343,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.876146831095426e-06,"brier_hard":2.0135925358414594e-11,"soft_accuracy":0.46666589446181994,"brier_soft":0.4177716143738584,"tvd":0.533329123860682,"kl":4.556514992141871,"case_id":"typed:368/r0","base_id":"typed:368","repeat":0,"workflow":"security_incidents","case_elapsed_ms":335.92469}
{"id":"severity","type":"ordinal","probabilities":{"0":6.141951847902718e-05,"1":0.00035793843374914644,"2":0.0623940299053532,"3":0.9371801209043606,"4":6.491238057372916e-06},"predicted":"3","confidence":0.9371801209043606,"entropy_confidence":0.8524849129294076,"candidate_mass":0.9999982407761208,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0648797837297082,"brier_hard":0.007839484111836631,"score_mae":0.7967123259097666,"within_one":true,"case_id":"typed:368/r0","base_id":"typed:368","repeat":0,"workflow":"security_incidents","case_elapsed_ms":335.92469}
{"id":"true_positive","type":"binary","probabilities":{"false":7.897542606206973e-08,"true":0.9999999210245742},"predicted":"true","confidence":0.9999999210245742,"entropy_confidence":0.9999980227146928,"candidate_mass":0.9999995007682723,"accepted":true,"accepted_predicted":"true","input_tokens":309,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":16.354129095374812,"brier_hard":1.9999996840983087,"soft_accuracy":0.3166670289576036,"brier_soft":0.9338877619119534,"tvd":0.683332921024574,"kl":5.929592724476535,"case_id":"typed:368/r0","base_id":"typed:368","repeat":0,"workflow":"security_incidents","case_elapsed_ms":335.92469}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.6034094795717263e-06,"1":2.129745570996304e-06,"2":2.83104855751619e-05,"3":0.9999659563593758},"predicted":"3","confidence":0.9999659563593758,"entropy_confidence":0.9997089394667905,"candidate_mass":0.9999988221614347,"accepted":true,"accepted_predicted":"3","input_tokens":338,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":10.472278306915435,"brier_hard":1.9998752937255746,"score_mae":1.418289619794849,"within_one":false,"case_id":"typed:368/r0","base_id":"typed:368","repeat":0,"workflow":"security_incidents","case_elapsed_ms":335.92469}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00013333149358831393,"true":0.9998666685064107},"predicted":"true","confidence":0.9998666685064107,"entropy_confidence":0.9980913205171404,"candidate_mass":0.9999760611803954,"accepted":true,"accepted_predicted":"true","input_tokens":321,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.000133340383023056,"brier_hard":3.555457436524277e-08,"soft_accuracy":0.5166625555279922,"brier_soft":0.4669638392894114,"tvd":0.4831996685064112,"kl":3.620099265729511,"case_id":"typed:369/r0","base_id":"typed:369","repeat":0,"workflow":"security_incidents","case_elapsed_ms":327.46489}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.1972373706437091e-05,"contain":1.1904312619658608e-05,"investigate":0.9999397972785959,"monitor":3.632603507964643e-05},"predicted":"investigate","confidence":0.9999397972785959,"entropy_confidence":0.9994934544328421,"candidate_mass":0.9999965979351831,"accepted":true,"accepted_predicted":"investigate","input_tokens":373,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":6.0204533660624577e-05,"brier_hard":5.228998880175331e-09,"soft_accuracy":0.6033072664177384,"brier_soft":0.22715360472871401,"tvd":0.3966067972785951,"kl":3.1831558723107243,"case_id":"typed:369/r0","base_id":"typed:369","repeat":0,"workflow":"security_incidents","case_elapsed_ms":327.46489}
{"id":"severity","type":"ordinal","probabilities":{"0":2.9226908768504257e-06,"1":9.7438919289695e-06,"2":0.9998892996033774,"3":9.797382118851261e-05,"4":5.999262876174578e-08},"predicted":"2","confidence":0.9998892996033774,"entropy_confidence":0.9992756806577242,"candidate_mass":0.9999993775618866,"accepted":true,"accepted_predicted":"2","input_tokens":393,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.00011070652436378204,"brier_hard":2.1956936601689625e-08,"score_mae":0.023250495467235677,"within_one":true,"case_id":"typed:369/r0","base_id":"typed:369","repeat":0,"workflow":"security_incidents","case_elapsed_ms":327.46489}
{"id":"true_positive","type":"binary","probabilities":{"false":2.0983292089550745e-06,"true":0.9999979016707906},"predicted":"true","confidence":0.9999979016707906,"entropy_confidence":0.9999573933812311,"candidate_mass":0.9999980035784615,"accepted":true,"accepted_predicted":"true","input_tokens":339,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.098331410883721e-06,"brier_hard":8.805970940124503e-12,"soft_accuracy":0.573332692246448,"brier_soft":0.36408587663549125,"tvd":0.42666490167079085,"kl":3.6108108985321925,"case_id":"typed:369/r0","base_id":"typed:369","repeat":0,"workflow":"security_incidents","case_elapsed_ms":327.46489}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001732403435259056,"1":0.00019031556762554543,"2":0.04081144923183351,"3":0.9588249948570161},"predicted":"3","confidence":0.9588249948570161,"entropy_confidence":0.8744901460287772,"candidate_mass":0.9999918723271436,"accepted":true,"accepted_predicted":"3","input_tokens":368,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":3.1987926185107156,"brier_hard":1.8393881129195242,"score_mae":1.1216211986023408,"within_one":false,"case_id":"typed:369/r0","base_id":"typed:369","repeat":0,"workflow":"security_incidents","case_elapsed_ms":327.46489}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0002811797486394833,"true":0.9997188202513588},"predicted":"true","confidence":0.9997188202513588,"entropy_confidence":0.9962775423207246,"candidate_mass":0.9999775862800534,"accepted":true,"accepted_predicted":"true","input_tokens":312,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.176516418261688,"brier_hard":1.9988754391295407,"soft_accuracy":0.3334267267703324,"brier_soft":0.8881401228641549,"tvd":0.6663858202513597,"kl":4.814593473436382,"case_id":"typed:370/r0","base_id":"typed:370","repeat":0,"workflow":"security_incidents","case_elapsed_ms":295.749491}
{"id":"disposition","type":"choice","probabilities":{"close_benign":9.946046287920836e-06,"contain":1.6099869422732914e-05,"investigate":0.999954123333118,"monitor":1.9830751171934142e-05},"predicted":"investigate","confidence":0.999954123333118,"entropy_confidence":0.9996011959593687,"candidate_mass":0.9999972731509974,"accepted":true,"accepted_predicted":"investigate","input_tokens":364,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.58777192485181e-05,"brier_hard":2.8560568884402494e-09,"soft_accuracy":0.5999794767943143,"brier_soft":0.2305047737116642,"tvd":0.3999541233331177,"kl":3.305706205028751,"case_id":"typed:370/r0","base_id":"typed:370","repeat":0,"workflow":"security_incidents","case_elapsed_ms":295.749491}
{"id":"severity","type":"ordinal","probabilities":{"0":4.157859193626431e-06,"1":5.157825194698504e-05,"2":0.18148232695398483,"3":0.8184610629215263,"4":8.740133463760692e-07},"predicted":"3","confidence":0.8184610629215263,"entropy_confidence":0.7053306858959041,"candidate_mass":0.9999984747456999,"accepted":true,"accepted_predicted":"3","input_tokens":384,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.20032945456513934,"brier_hard":0.06589222335058284,"score_mae":0.10840291697788107,"within_one":true,"case_id":"typed:370/r0","base_id":"typed:370","repeat":0,"workflow":"security_incidents","case_elapsed_ms":295.749491}
{"id":"true_positive","type":"binary","probabilities":{"false":3.9207861609379495e-09,"true":0.9999999960792145},"predicted":"true","confidence":0.9999999960792145,"entropy_confidence":0.9999998848508049,"candidate_mass":0.9999999049737706,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.920785481751549e-09,"brier_hard":3.0745122853443764e-17,"soft_accuracy":0.5499999996079218,"brier_soft":0.4049999929425855,"tvd":0.44999999607921415,"kl":3.815842819130123,"case_id":"typed:370/r0","base_id":"typed:370","repeat":0,"workflow":"security_incidents","case_elapsed_ms":295.749491}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0005613233132989869,"1":0.0004413219572681169,"2":0.3838118871648439,"3":0.6151854675645886},"predicted":"3","confidence":0.6151854675645886,"entropy_confidence":0.5137923392173768,"candidate_mass":0.9999855030774994,"accepted":false,"accepted_predicted":null,"input_tokens":359,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.48583148336394544,"brier_hard":0.2953942989514552,"score_mae":0.2802884989807213,"within_one":true,"case_id":"typed:370/r0","base_id":"typed:370","repeat":0,"workflow":"security_incidents","case_elapsed_ms":295.749491}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.006311588783359824,"true":0.9936884112166386},"predicted":"true","confidence":0.9936884112166386,"entropy_confidence":0.9447993894768277,"candidate_mass":0.9999392733770017,"accepted":true,"accepted_predicted":"true","input_tokens":291,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.006331591068395461,"brier_hard":7.967230594048641e-05,"soft_accuracy":0.8620384973511422,"brier_soft":0.03226887781493323,"tvd":0.12702141121663946,"kl":0.2881942287476731,"case_id":"typed:371/r0","base_id":"typed:371","repeat":0,"workflow":"security_incidents","case_elapsed_ms":294.291992}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.08133254117956701,"contain":0.00034952576500715815,"investigate":0.9079502892670652,"monitor":0.010367643788361088},"predicted":"investigate","confidence":0.9079502892670652,"entropy_confidence":0.7533644616691417,"candidate_mass":0.9999767684761086,"accepted":true,"accepted_predicted":"investigate","input_tokens":343,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.09656564938277515,"brier_hard":0.015195741706725678,"soft_accuracy":0.45623708333812324,"brier_soft":0.3589774513426096,"tvd":0.469282830446632,"kl":2.830350576825437,"case_id":"typed:371/r0","base_id":"typed:371","repeat":0,"workflow":"security_incidents","case_elapsed_ms":294.291992}
{"id":"severity","type":"ordinal","probabilities":{"0":0.002414190673950047,"1":0.08700258684010836,"2":0.9100492642420133,"3":0.0005330435982028222,"4":9.146457246604893e-07},"predicted":"2","confidence":0.9100492642420133,"entropy_confidence":0.803160094183663,"candidate_mass":0.9999953775026541,"accepted":true,"accepted_predicted":"2","input_tokens":363,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.09425654440688545,"brier_hard":0.0156666974331981,"score_mae":0.008703904701642173,"within_one":true,"case_id":"typed:371/r0","base_id":"typed:371","repeat":0,"workflow":"security_incidents","case_elapsed_ms":294.291992}
{"id":"true_positive","type":"binary","probabilities":{"false":0.0029388781882154595,"true":0.9970611218117851},"predicted":"true","confidence":0.9970611218117851,"entropy_confidence":0.9710488413834037,"candidate_mass":0.9999660112682541,"accepted":true,"accepted_predicted":"true","input_tokens":309,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00294320517044613,"brier_hard":1.7274010010334004e-05,"soft_accuracy":0.9208447517598968,"brier_soft":0.010871671891786769,"tvd":0.07372812181178484,"kl":0.17911268080270884,"case_id":"typed:371/r0","base_id":"typed:371","repeat":0,"workflow":"security_incidents","case_elapsed_ms":294.291992}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.005473739098572477,"1":0.471804188249711,"2":0.4999935139606788,"3":0.022728558691037144},"predicted":"2","confidence":0.4999935139606788,"entropy_confidence":0.4117376753970984,"candidate_mass":0.9999717148201247,"accepted":false,"accepted_predicted":null,"input_tokens":338,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":3.784133053113351,"brier_hard":1.427682137870533,"score_mae":0.88669010775582,"within_one":true,"case_id":"typed:371/r0","base_id":"typed:371","repeat":0,"workflow":"security_incidents","case_elapsed_ms":294.291992}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0002550134730567442,"true":0.9997449865269435},"predicted":"true","confidence":0.9997449865269435,"entropy_confidence":0.996588008990099,"candidate_mass":0.9999716794108211,"accepted":true,"accepted_predicted":"true","input_tokens":260,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0002550459945212323,"brier_hard":1.3006374288077732e-07,"soft_accuracy":0.8498214905688605,"brier_soft":0.044847121979908935,"tvd":0.14974498652694343,"kl":0.8186368280920948,"case_id":"typed:372/r0","base_id":"typed:372","repeat":0,"workflow":"security_incidents","case_elapsed_ms":291.196991}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.028253946457986e-05,"contain":5.700712495475725e-05,"investigate":0.9997814325516005,"monitor":0.00010127778397933944},"predicted":"investigate","confidence":0.9997814325516005,"entropy_confidence":0.99834603196394,"candidate_mass":0.9999906168638228,"accepted":true,"accepted_predicted":"investigate","input_tokens":312,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":9.772334298735206,"brier_hard":1.9994489157660074,"soft_accuracy":0.28331373133486865,"brier_soft":0.9593579451241796,"tvd":0.716448432551601,"kl":6.073235179050148,"case_id":"typed:372/r0","base_id":"typed:372","repeat":0,"workflow":"security_incidents","case_elapsed_ms":291.196991}
{"id":"severity","type":"ordinal","probabilities":{"0":2.0868501476903247e-05,"1":8.131877688481132e-05,"2":0.0011897422349694115,"3":0.9986864215068744,"4":2.1648979793395797e-05},"predicted":"3","confidence":0.9986864215068744,"entropy_confidence":0.99344635857844,"candidate_mass":0.9999985369677278,"accepted":true,"accepted_predicted":"3","input_tokens":332,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":10.740552227332977,"brier_hard":1.9973326935461213,"score_mae":0.6047263363133801,"within_one":true,"case_id":"typed:372/r0","base_id":"typed:372","repeat":0,"workflow":"security_incidents","case_elapsed_ms":291.196991}
{"id":"true_positive","type":"binary","probabilities":{"false":1.0708134593221721e-05,"true":0.9999892918654064},"predicted":"true","confidence":0.9999892918654064,"entropy_confidence":0.9998077502140376,"candidate_mass":0.9999950863003698,"accepted":true,"accepted_predicted":"true","input_tokens":278,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.070819192606696e-05,"brier_hard":2.2932829294087152e-10,"soft_accuracy":0.8933245762745918,"brier_soft":0.022751129188957574,"tvd":0.10665629186540655,"kl":0.8812711286834195,"case_id":"typed:372/r0","base_id":"typed:372","repeat":0,"workflow":"security_incidents","case_elapsed_ms":291.196991}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.558787876821521e-05,"1":4.51886892191543e-05,"2":0.0002913993168985486,"3":0.9995878241151125},"predicted":"3","confidence":0.9995878241151125,"entropy_confidence":0.9971479635954039,"candidate_mass":0.9999951560856396,"accepted":true,"accepted_predicted":"3","input_tokens":307,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00041226085271614803,"brier_hard":2.625580670217674e-07,"score_mae":0.2827244596683536,"within_one":true,"case_id":"typed:372/r0","base_id":"typed:372","repeat":0,"workflow":"security_incidents","case_elapsed_ms":291.196991}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.10064099580669422,"true":0.8993590041933041},"predicted":"true","confidence":0.8993590041933041,"entropy_confidence":0.5289757846997936,"candidate_mass":0.9999143431725055,"accepted":true,"accepted_predicted":"true","input_tokens":258,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":2.296195591338357,"brier_hard":1.617693236847146,"soft_accuracy":0.3402563983226772,"brier_soft":0.7184624318151802,"tvd":0.5993590041933049,"kl":1.0282945079724282,"case_id":"typed:373/r0","base_id":"typed:373","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.500592}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0528724193757057e-05,"contain":9.112920225557996e-06,"investigate":0.9998984104033186,"monitor":8.194795226071267e-05},"predicted":"investigate","confidence":0.9998984104033186,"entropy_confidence":0.9992071637003137,"candidate_mass":0.9999955652408513,"accepted":true,"accepted_predicted":"investigate","input_tokens":310,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":9.409426240753348,"brier_hard":1.9996329421319283,"soft_accuracy":0.3666637251925876,"brier_soft":0.5903582654292743,"tvd":0.6332314104033192,"kl":5.144116414483382,"case_id":"typed:373/r0","base_id":"typed:373","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.500592}
{"id":"severity","type":"ordinal","probabilities":{"0":7.38819520303598e-05,"1":0.0013578872965676583,"2":0.9982681900127014,"3":0.0002996854176373351,"4":3.5532106296042077e-07},"predicted":"2","confidence":0.9982681900127014,"entropy_confidence":0.9914043009564215,"candidate_mass":0.9999982912214482,"accepted":true,"accepted_predicted":"2","input_tokens":330,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":6.6018252455078965,"brier_hard":1.9938255437260284,"score_mae":0.8221277448591344,"within_one":true,"case_id":"typed:373/r0","base_id":"typed:373","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.500592}
{"id":"true_positive","type":"binary","probabilities":{"false":3.5452624921119893e-06,"true":0.999996454737509},"predicted":"true","confidence":0.999996454737509,"entropy_confidence":0.9999306959020081,"candidate_mass":0.9999967933619804,"accepted":true,"accepted_predicted":"true","input_tokens":276,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.549898354991676,"brier_hard":1.9999858189751714,"soft_accuracy":0.3833338272302787,"brier_soft":0.7605476328175982,"tvd":0.6166634547375085,"kl":5.312155012055249,"case_id":"typed:373/r0","base_id":"typed:373","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.500592}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0008407086379010029,"1":0.00907848467972287,"2":0.9812654438369453,"3":0.00881536284543238},"predicted":"2","confidence":0.9812654438369453,"entropy_confidence":0.9214418925184171,"candidate_mass":0.9999901650990991,"accepted":true,"accepted_predicted":"2","input_tokens":305,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":4.701847985788945,"brier_hard":1.944885738206262,"score_mae":0.6813884608899106,"within_one":true,"case_id":"typed:373/r0","base_id":"typed:373","repeat":0,"workflow":"security_incidents","case_elapsed_ms":281.500592}
{"id":"credential_compromise","type":"binary","probabilities":{"false":4.628884967273959e-05,"true":0.9999537111503278},"predicted":"true","confidence":0.9999537111503278,"entropy_confidence":0.9992667088271788,"candidate_mass":0.9999886044629778,"accepted":true,"accepted_predicted":"true","input_tokens":259,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.628992103403491e-05,"brier_hard":4.285315207998366e-09,"soft_accuracy":0.766642312582649,"brier_soft":0.10884537919867274,"tvd":0.23328671115032756,"kl":1.785568649801454,"case_id":"typed:374/r0","base_id":"typed:374","repeat":0,"workflow":"security_incidents","case_elapsed_ms":296.172391}
{"id":"disposition","type":"choice","probabilities":{"close_benign":4.496779767898621e-05,"contain":8.919567975440072e-05,"investigate":0.9998154724325645,"monitor":5.036409000392242e-05},"predicted":"investigate","confidence":0.9998154724325645,"entropy_confidence":0.9985827293563815,"candidate_mass":0.9999898840296387,"accepted":true,"accepted_predicted":"investigate","input_tokens":311,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":9.324677952797867,"brier_hard":1.999452600070557,"soft_accuracy":0.3166638245485692,"brier_soft":0.7797694820672345,"tvd":0.6831487890992469,"kl":5.412260031408012,"case_id":"typed:374/r0","base_id":"typed:374","repeat":0,"workflow":"security_incidents","case_elapsed_ms":296.172391}
{"id":"severity","type":"ordinal","probabilities":{"0":2.5061790303683846e-06,"1":5.906144772302565e-06,"2":6.948989332413898e-05,"3":0.9999133598468122,"4":8.737936061758243e-06},"predicted":"3","confidence":0.9999133598468122,"entropy_confidence":0.9994052816451191,"candidate_mass":0.9999994072955987,"accepted":true,"accepted_predicted":"3","input_tokens":331,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":11.647836544724534,"brier_hard":1.9998092562743774,"score_mae":0.3967470827838957,"within_one":true,"case_id":"typed:374/r0","base_id":"typed:374","repeat":0,"workflow":"security_incidents","case_elapsed_ms":296.172391}
{"id":"true_positive","type":"binary","probabilities":{"false":1.3824158258017296e-07,"true":0.9999998617584172},"predicted":"true","confidence":0.9999998617584172,"entropy_confidence":0.9999966505446999,"candidate_mass":0.9999995059302668,"accepted":true,"accepted_predicted":"true","input_tokens":277,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.3824159240306643e-07,"brier_hard":3.822147038250771e-14,"soft_accuracy":0.8166669124469054,"brier_soft":0.06722187640106186,"tvd":0.18333286175841726,"kl":1.523164894447132,"case_id":"typed:374/r0","base_id":"typed:374","repeat":0,"workflow":"security_incidents","case_elapsed_ms":296.172391}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.19851633077684e-06,"1":4.0342947307510444e-06,"2":2.1112264334039724e-05,"3":0.9999666549246046},"predicted":"3","confidence":0.9999666549246046,"entropy_confidence":0.9997065861409935,"candidate_mass":0.9999967285533778,"accepted":true,"accepted_predicted":"3","input_tokens":306,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":3.3345631354815684e-05,"brier_hard":1.6411129624378624e-09,"score_mae":0.4966132235972127,"within_one":true,"case_id":"typed:374/r0","base_id":"typed:374","repeat":0,"workflow":"security_incidents","case_elapsed_ms":296.172391}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0005639803190344642,"true":0.999436019680967},"predicted":"true","confidence":0.999436019680967,"entropy_confidence":0.993100064169302,"candidate_mass":0.9999779273596748,"accepted":true,"accepted_predicted":"true","input_tokens":259,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.4804912024040355,"brier_hard":1.9977447148714655,"soft_accuracy":0.2169865884714663,"brier_soft":1.2254546763446017,"tvd":0.7827690196809662,"kl":5.337180836949245,"case_id":"typed:375/r0","base_id":"typed:375","repeat":0,"workflow":"security_incidents","case_elapsed_ms":285.229091}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.7955710469111882e-06,"contain":3.3798438351100377e-06,"investigate":0.9999880725075172,"monitor":5.752077600228715e-06},"predicted":"investigate","confidence":0.9999880725075172,"entropy_confidence":0.9998848309633678,"candidate_mass":0.9999973955820551,"accepted":true,"accepted_predicted":"investigate","input_tokens":311,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.192756361590354e-05,"brier_hard":1.945900354743498e-10,"soft_accuracy":0.533329343491464,"brier_soft":0.3183175948935848,"tvd":0.4666545391739842,"kl":3.96290548214173,"case_id":"typed:375/r0","base_id":"typed:375","repeat":0,"workflow":"security_incidents","case_elapsed_ms":285.229091}
{"id":"severity","type":"ordinal","probabilities":{"0":2.5314574001508534e-05,"1":0.000289369217889595,"2":0.9749950279408536,"3":0.02468927240832788,"4":1.0158589261208493e-06},"predicted":"2","confidence":0.9749950279408536,"entropy_confidence":0.926238892341104,"candidate_mass":0.9999969040852397,"accepted":true,"accepted_predicted":"2","input_tokens":331,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.7013864451427914,"brier_hard":1.9018464042411867,"score_mae":0.1589816942397153,"within_one":true,"case_id":"typed:375/r0","base_id":"typed:375","repeat":0,"workflow":"security_incidents","case_elapsed_ms":285.229091}
{"id":"true_positive","type":"binary","probabilities":{"false":3.522047167573419e-07,"true":0.9999996477952829},"predicted":"true","confidence":0.9999996477952829,"entropy_confidence":0.999991941634544,"candidate_mass":0.999999009002427,"accepted":true,"accepted_predicted":"true","input_tokens":277,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":14.859053248704413,"brier_hard":1.9999985911813805,"soft_accuracy":0.4666670234800795,"brier_soft":0.568887426408655,"tvd":0.5333326477952831,"kl":4.5565133476344055,"case_id":"typed:375/r0","base_id":"typed:375","repeat":0,"workflow":"security_incidents","case_elapsed_ms":285.229091}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001885088816221228,"1":0.00013782085131713853,"2":0.0019626897807097456,"3":0.997710980486351},"predicted":"3","confidence":0.997710980486351,"entropy_confidence":0.987475536769407,"candidate_mass":0.9999935507634975,"accepted":true,"accepted_predicted":"3","input_tokens":306,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.23343940943683,"brier_hard":1.991505727702977,"score_mae":1.3805291418717895,"within_one":false,"case_id":"typed:375/r0","base_id":"typed:375","repeat":0,"workflow":"security_incidents","case_elapsed_ms":285.229091}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0006690650620170597,"true":0.9993309349379816},"predicted":"true","confidence":0.9993309349379816,"entropy_confidence":0.9919793964523003,"candidate_mass":0.9999748215969302,"accepted":true,"accepted_predicted":"true","input_tokens":316,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.309629249923432,"brier_hard":1.9973246350480436,"soft_accuracy":0.4001338130124029,"brier_soft":0.7183951391472718,"tvd":0.5993309349379823,"kl":3.7130335985391762,"case_id":"typed:376/r0","base_id":"typed:376","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.90929100000005}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0003977864894857469,"contain":8.076622461098483e-05,"investigate":0.9976951789348102,"monitor":0.0018262683510927595},"predicted":"investigate","confidence":0.9976951789348102,"entropy_confidence":0.9872369771177463,"candidate_mass":0.9999923295536755,"accepted":true,"accepted_predicted":"investigate","input_tokens":368,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0023074812535525797,"brier_hard":8.812213507001066e-06,"soft_accuracy":0.6987176237379085,"brier_soft":0.12518620038531042,"tvd":0.2976951789348104,"kl":1.3241757114877963,"case_id":"typed:376/r0","base_id":"typed:376","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.90929100000005}
{"id":"severity","type":"ordinal","probabilities":{"0":1.3351581986309806e-05,"1":2.7499901141202324e-05,"2":0.0005645253626699831,"3":0.9993834830012114,"4":1.1140152990524858e-05},"predicted":"3","confidence":0.9993834830012114,"entropy_confidence":0.9966420552823606,"candidate_mass":0.9999994306146278,"accepted":true,"accepted_predicted":"3","input_tokens":388,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0006167071235409603,"brier_hard":6.998407072059671e-07,"score_mae":0.2760185602420777,"within_one":true,"case_id":"typed:376/r0","base_id":"typed:376","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.90929100000005}
{"id":"true_positive","type":"binary","probabilities":{"false":2.8523200429221176e-06,"true":0.9999971476799558},"predicted":"true","confidence":0.9999971476799558,"entropy_confidence":0.9999433468608117,"candidate_mass":0.9999967890307391,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":12.767377844694717,"brier_hard":1.9999885907360972,"soft_accuracy":0.4833330950792357,"brier_soft":0.5338836829957117,"tvd":0.5166641476799565,"kl":4.407273329359328,"case_id":"typed:376/r0","base_id":"typed:376","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.90929100000005}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.690408656736818e-05,"1":6.07488062986378e-05,"2":0.0007974751299644647,"3":0.9990448719771702},"predicted":"3","confidence":0.9990448719771702,"entropy_confidence":0.9941359712607971,"candidate_mass":0.9999969805860832,"accepted":true,"accepted_predicted":"3","input_tokens":363,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0009555844482527704,"brier_hard":1.5613169423668083e-06,"score_mae":0.7154573149977383,"within_one":true,"case_id":"typed:376/r0","base_id":"typed:376","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.90929100000005}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0026209522759303968,"true":0.9973790477240686},"predicted":"true","confidence":0.9973790477240686,"entropy_confidence":0.9737472420908974,"candidate_mass":0.9999516269490257,"accepted":true,"accepted_predicted":"true","input_tokens":306,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":5.944217563150634,"brier_hard":1.9895299296779418,"soft_accuracy":0.35078628568277875,"brier_soft":0.8381992628642452,"tvd":0.6473790477240691,"kl":3.217213314557896,"case_id":"typed:377/r0","base_id":"typed:377","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.20599100000004}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.271621658606226e-05,"contain":3.068953765253706e-05,"investigate":0.9996933740920908,"monitor":0.00022322015367130324},"predicted":"investigate","confidence":0.9996933740920908,"entropy_confidence":0.9978204715052493,"candidate_mass":0.9999927281815785,"accepted":true,"accepted_predicted":"investigate","input_tokens":358,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0003066729272447076,"brier_hard":1.4756753161865477e-07,"soft_accuracy":0.48325987242661156,"brier_soft":0.384733981298799,"tvd":0.5163598907586071,"kl":3.5438616507801233,"case_id":"typed:377/r0","base_id":"typed:377","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.20599100000004}
{"id":"severity","type":"ordinal","probabilities":{"0":9.926917831983128e-06,"1":3.575355723316707e-05,"2":0.0005344864440239564,"3":0.999413955322386,"4":5.877758524422305e-06},"predicted":"3","confidence":0.999413955322386,"entropy_confidence":0.9967914058977227,"candidate_mass":0.9999993387864007,"accepted":true,"accepted_predicted":"3","input_tokens":378,"reused_prefix_tokens":0,"expected":"4","correct":false,"accepted_correct":false,"nll_hard":12.044335072059738,"brier_hard":1.9988167856632546,"score_mae":0.11062989655346334,"within_one":true,"case_id":"typed:377/r0","base_id":"typed:377","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.20599100000004}
{"id":"true_positive","type":"binary","probabilities":{"false":1.6428622687360838e-06,"true":0.9999983571377301},"predicted":"true","confidence":0.9999983571377301,"entropy_confidence":0.9999660616676459,"candidate_mass":0.999997961583957,"accepted":true,"accepted_predicted":"true","input_tokens":324,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":13.319070551555026,"brier_hard":1.9999934285563206,"soft_accuracy":0.45000016428622636,"brier_soft":0.6049963857084054,"tvd":0.5499983571377307,"kl":4.706359480577533,"case_id":"typed:377/r0","base_id":"typed:377","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.20599100000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.8117667788186477e-05,"1":3.530406272001694e-05,"2":0.00039137728939455974,"3":0.999525200980097},"predicted":"3","confidence":0.999525200980097,"entropy_confidence":0.9968364088924536,"candidate_mass":0.9999950788626583,"accepted":true,"accepted_predicted":"3","input_tokens":353,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00047491177264898565,"brier_hard":3.821719787525897e-07,"score_mae":0.8560606615818003,"within_one":true,"case_id":"typed:377/r0","base_id":"typed:377","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.20599100000004}
{"id":"credential_compromise","type":"binary","probabilities":{"false":1.4645149021414711e-05,"true":0.9999853548509802},"predicted":"true","confidence":0.9999853548509802,"entropy_confidence":0.9997436820359172,"candidate_mass":0.999993756472265,"accepted":true,"accepted_predicted":"true","input_tokens":297,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4645256261073777e-05,"brier_hard":4.289607796725254e-10,"soft_accuracy":0.5000000000000008,"brier_soft":0.4999707101309195,"tvd":0.4999853548509794,"kl":4.25860391833625,"case_id":"typed:378/r0","base_id":"typed:378","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.73939}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.00023438730959504382,"contain":9.911974537092893e-05,"investigate":0.998151706949041,"monitor":0.0015147859959944645},"predicted":"investigate","confidence":0.998151706949041,"entropy_confidence":0.989501334338046,"candidate_mass":0.9999877801678518,"accepted":true,"accepted_predicted":"investigate","input_tokens":349,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.001850003252187054,"brier_hard":5.775525950705929e-06,"soft_accuracy":0.632465693042343,"brier_soft":0.18637808000644507,"tvd":0.3648180736154069,"kl":1.819872502670715,"case_id":"typed:378/r0","base_id":"typed:378","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.73939}
{"id":"severity","type":"ordinal","probabilities":{"0":9.729800273749222e-06,"1":6.261560597479719e-05,"2":0.9981176599524464,"3":0.0018096211025171267,"4":3.735387890877858e-07},"predicted":"2","confidence":0.9981176599524464,"entropy_confidence":0.9912817342218112,"candidate_mass":0.9999987725109991,"accepted":true,"accepted_predicted":"2","input_tokens":369,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.31463779124406,"brier_hard":1.9926228996479705,"score_mae":0.7816047070264247,"within_one":true,"case_id":"typed:378/r0","base_id":"typed:378","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.73939}
{"id":"true_positive","type":"binary","probabilities":{"false":2.4337845170603086e-07,"true":0.9999997566215496},"predicted":"true","confidence":0.9999997566215496,"entropy_confidence":0.9999943017827241,"candidate_mass":0.9999991591650659,"accepted":true,"accepted_predicted":"true","input_tokens":315,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.4337848000955685e-07,"brier_hard":1.1846614087053102e-13,"soft_accuracy":0.5999999513243105,"brier_soft":0.3199996105945968,"tvd":0.39999975662154896,"kl":3.377640920557967,"case_id":"typed:378/r0","base_id":"typed:378","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.73939}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00022524040319386618,"1":0.00016039771009427266,"2":0.001856927858526763,"3":0.9977574340281861},"predicted":"3","confidence":0.9977574340281861,"entropy_confidence":0.9875847908877434,"candidate_mass":0.999992776600608,"accepted":true,"accepted_predicted":"3","input_tokens":344,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":6.288831845763191,"brier_hard":1.991809566083193,"score_mae":0.7804795555117061,"within_one":true,"case_id":"typed:378/r0","base_id":"typed:378","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.73939}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00015411277187877574,"true":0.9998458872281222},"predicted":"true","confidence":0.9998458872281222,"entropy_confidence":0.9978260374930094,"candidate_mass":0.9999863558541631,"accepted":true,"accepted_predicted":"true","input_tokens":285,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.00015412464847129023,"brier_hard":4.750149291202563e-08,"soft_accuracy":0.6999383548912491,"brier_soft":0.179815112175239,"tvd":0.2998458872281218,"kl":2.0225913667830313,"case_id":"typed:379/r0","base_id":"typed:379","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.43949}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.9366658533185887e-06,"contain":2.569390927668812e-06,"investigate":0.9999911673647928,"monitor":4.3265784276624806e-06},"predicted":"investigate","confidence":0.9999911673647928,"entropy_confidence":0.999912848408149,"candidate_mass":0.9999979655131864,"accepted":true,"accepted_predicted":"investigate","input_tokens":337,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":8.832674215186365e-06,"brier_hard":1.0708716996138659e-10,"soft_accuracy":0.5833296001733016,"brier_soft":0.28912339275713034,"tvd":0.4166575840312087,"kl":3.523230161288491,"case_id":"typed:379/r0","base_id":"typed:379","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.43949}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00010754255561077185,"1":0.0007345990590937909,"2":0.9966723216415756,"3":0.002484467291122163,"4":1.0694525985570386e-06},"predicted":"2","confidence":0.9966723216415756,"entropy_confidence":0.9847638686612459,"candidate_mass":0.9999977002764322,"accepted":true,"accepted_predicted":"2","input_tokens":357,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.997697012183782,"brier_hard":1.9883935059242073,"score_mae":0.9251300779739946,"within_one":true,"case_id":"typed:379/r0","base_id":"typed:379","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.43949}
{"id":"true_positive","type":"binary","probabilities":{"false":1.0694735512231488e-07,"true":0.9999998930526446},"predicted":"true","confidence":0.9999998930526446,"entropy_confidence":0.9999973691709081,"candidate_mass":0.999999437928002,"accepted":true,"accepted_predicted":"true","input_tokens":303,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.0694736113492264e-07,"brier_hard":2.287547359814627e-14,"soft_accuracy":0.8266669301276565,"brier_soft":0.060088583627999126,"tvd":0.1733328930526447,"kl":1.4390972037683518,"case_id":"typed:379/r0","base_id":"typed:379","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.43949}
{"id":"urgency","type":"ordinal","probabilities":{"0":7.59194304838198e-05,"1":6.14128072244124e-05,"2":0.0005004412356204315,"3":0.999362226526671},"predicted":"3","confidence":0.999362226526671,"entropy_confidence":0.9958474419319158,"candidate_mass":0.9999907036303596,"accepted":true,"accepted_predicted":"3","input_tokens":332,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0006379769373445925,"brier_hard":6.66731726407676e-07,"score_mae":0.4158159748584782,"within_one":true,"case_id":"typed:379/r0","base_id":"typed:379","repeat":0,"workflow":"security_incidents","case_elapsed_ms":333.43949}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0009778235383276844,"true":0.9990221764616721},"predicted":"true","confidence":0.9990221764616721,"entropy_confidence":0.9888135738482984,"candidate_mass":0.9999671125889493,"accepted":true,"accepted_predicted":"true","input_tokens":278,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.000978301919637686,"brier_hard":1.9122777442157097e-06,"soft_accuracy":0.5998044352923343,"brier_soft":0.3184373946164198,"tvd":0.39902217646167226,"kl":2.0996478482892877,"case_id":"typed:380/r0","base_id":"typed:380","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.029491}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0008874641536386357,"contain":0.0001546741592715933,"investigate":0.9970395144783382,"monitor":0.0019183472087529105},"predicted":"investigate","confidence":0.9970395144783382,"entropy_confidence":0.983732664598737,"candidate_mass":0.9999870759522825,"accepted":true,"accepted_predicted":"investigate","input_tokens":330,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0029648764272067318,"brier_hard":1.325604725683936e-05,"soft_accuracy":0.6485314078871156,"brier_soft":0.17091834700770198,"tvd":0.34703951447833753,"kl":1.4633347087822408,"case_id":"typed:380/r0","base_id":"typed:380","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.029491}
{"id":"severity","type":"ordinal","probabilities":{"0":3.716825464228443e-05,"1":0.00010040560145556338,"2":0.999654421155522,"3":0.00020755555263157302,"4":4.4943574973332213e-07},"predicted":"2","confidence":0.999654421155522,"entropy_confidence":0.9978777288870947,"candidate_mass":0.9999984552434421,"accepted":true,"accepted_predicted":"2","input_tokens":350,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0003456385706072746,"brier_hard":1.7396701112820969e-07,"score_mae":0.3032992876866065,"within_one":true,"case_id":"typed:380/r0","base_id":"typed:380","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.029491}
{"id":"true_positive","type":"binary","probabilities":{"false":0.00025831408726427924,"true":0.9997416859127375},"predicted":"true","confidence":0.9997416859127375,"entropy_confidence":0.9965486410056827,"candidate_mass":0.9999850308294328,"accepted":true,"accepted_predicted":"true","input_tokens":296,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0002583474560929011,"brier_hard":1.3345233535744247e-07,"soft_accuracy":0.6832382850068803,"brier_soft":0.20022891304204965,"tvd":0.31640868591273663,"kl":1.9919369298109295,"case_id":"typed:380/r0","base_id":"typed:380","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.029491}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.00042394054237816385,"1":0.0018122411800061383,"2":0.9949444786002157,"3":0.00281933967740031},"predicted":"2","confidence":0.9949444786002157,"entropy_confidence":0.9737941025296205,"candidate_mass":0.9999880709373538,"accepted":true,"accepted_predicted":"2","input_tokens":325,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.005068343782237855,"brier_hard":3.6970916518222305e-05,"score_mae":0.016826217412638567,"within_one":true,"case_id":"typed:380/r0","base_id":"typed:380","repeat":0,"workflow":"security_incidents","case_elapsed_ms":322.029491}
{"id":"credential_compromise","type":"binary","probabilities":{"false":3.977467833534295e-05,"true":0.9999602253216655},"predicted":"true","confidence":0.9999602253216655,"entropy_confidence":0.9993612005063965,"candidate_mass":0.9999858362480718,"accepted":true,"accepted_predicted":"true","input_tokens":271,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.977546936798589e-05,"brier_hard":3.1640500732923155e-09,"soft_accuracy":0.786644195824572,"brier_soft":0.09098799993623718,"tvd":0.21329322532166506,"kl":1.6432416991406833,"case_id":"typed:381/r0","base_id":"typed:381","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.44338899999997}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0834831335733621e-05,"contain":1.643068733326262e-05,"investigate":0.9999489930347588,"monitor":2.3741446572672293e-05},"predicted":"investigate","confidence":0.9999489930347588,"entropy_confidence":0.9995609232286791,"candidate_mass":0.9999891492693405,"accepted":true,"accepted_predicted":"investigate","input_tokens":323,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":5.100826614067324e-05,"brier_hard":3.5527278447957255e-09,"soft_accuracy":0.5999767849168747,"brier_soft":0.22494441978849625,"tvd":0.3999489930347586,"kl":3.270129153669272,"case_id":"typed:381/r0","base_id":"typed:381","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.44338899999997}
{"id":"severity","type":"ordinal","probabilities":{"0":4.670606197933022e-06,"1":3.8759091185297194e-05,"2":0.9313798294781567,"3":0.06857625890719712,"4":4.819172618721944e-07},"predicted":"2","confidence":0.9313798294781567,"entropy_confidence":0.8443932809612273,"candidate_mass":0.9999950106555547,"accepted":true,"accepted_predicted":"2","input_tokens":343,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":2.679808884219838,"brier_hard":1.7350185737543868,"score_mae":0.6914708775618621,"within_one":true,"case_id":"typed:381/r0","base_id":"typed:381","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.44338899999997}
{"id":"true_positive","type":"binary","probabilities":{"false":1.4758882641108659e-08,"true":0.999999985241118},"predicted":"true","confidence":0.999999985241118,"entropy_confidence":0.9999995947722041,"candidate_mass":0.9999994251597428,"accepted":true,"accepted_predicted":"true","input_tokens":289,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.4758882114329052e-08,"brier_hard":4.356492148638329e-16,"soft_accuracy":0.8499999896687828,"brier_soft":0.044999991144671056,"tvd":0.14999998524111768,"kl":1.2434099782683687,"case_id":"typed:381/r0","base_id":"typed:381","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.44338899999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":2.550295023488666e-05,"1":2.5006124965596045e-05,"2":0.00017071286470309981,"3":0.9997787780600956},"predicted":"3","confidence":0.9997787780600956,"entropy_confidence":0.9983863933851025,"candidate_mass":0.9999941105048459,"accepted":true,"accepted_predicted":"3","input_tokens":318,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00022124641308712094,"brier_hard":7.935773562666976e-08,"score_mae":0.603035766034659,"within_one":true,"case_id":"typed:381/r0","base_id":"typed:381","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.44338899999997}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.7914410868786172e-05,"true":0.9999720855891309},"predicted":"true","confidence":0.9999720855891309,"entropy_confidence":0.9995374217760422,"candidate_mass":0.9999902592508809,"accepted":true,"accepted_predicted":"true","input_tokens":279,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.48636748389228,"brier_hard":1.9998883439149528,"soft_accuracy":0.40000558288217364,"brier_soft":0.7199330069723431,"tvd":0.5999720855891311,"kl":5.159699096356241,"case_id":"typed:382/r0","base_id":"typed:382","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.05798999999996}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.825275738512506e-05,"contain":4.669994903313039e-05,"investigate":0.9997672714317264,"monitor":0.00012777586185400494},"predicted":"investigate","confidence":0.9997672714317264,"entropy_confidence":0.9982601608841635,"candidate_mass":0.9999901217470226,"accepted":true,"accepted_predicted":"investigate","input_tokens":331,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00023275565376931615,"brier_hard":7.606352634588543e-08,"soft_accuracy":0.5165973164627086,"brier_soft":0.35578454155756206,"tvd":0.48310027143172707,"kl":3.3708718447207984,"case_id":"typed:382/r0","base_id":"typed:382","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.05798999999996}
{"id":"severity","type":"ordinal","probabilities":{"0":7.721234993272772e-06,"1":3.6732640753127294e-05,"2":0.9998184582735836,"3":0.0001369624425589919,"4":1.2540811202477647e-07},"predicted":"2","confidence":0.9998184582735836,"entropy_confidence":0.9988394034823421,"candidate_mass":0.9999988793791763,"accepted":true,"accepted_predicted":"2","input_tokens":351,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0001815582071102465,"brier_hard":5.312502919567868e-08,"score_mae":0.10675203814804557,"within_one":true,"case_id":"typed:382/r0","base_id":"typed:382","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.05798999999996}
{"id":"true_positive","type":"binary","probabilities":{"false":3.2116557207896223e-06,"true":0.9999967883442789},"predicted":"true","confidence":0.9999967883442789,"entropy_confidence":0.9999367594669107,"candidate_mass":0.9999953746959807,"accepted":true,"accepted_predicted":"true","input_tokens":297,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.211660878524289e-06,"brier_hard":2.0629464940057037e-11,"soft_accuracy":0.516666892942668,"brier_soft":0.46721536860185114,"tvd":0.48332978834427903,"kl":4.110478580751118,"case_id":"typed:382/r0","base_id":"typed:382","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.05798999999996}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.000260035975455175,"1":0.0002306640217525243,"2":0.004729789381753602,"3":0.9947795106210388},"predicted":"3","confidence":0.9947795106210388,"entropy_confidence":0.975035776960643,"candidate_mass":0.9999893711053741,"accepted":true,"accepted_predicted":"3","input_tokens":326,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":5.35387460564019,"brier_hard":1.9801491877201214,"score_mae":1.1773617746483762,"within_one":false,"case_id":"typed:382/r0","base_id":"typed:382","repeat":0,"workflow":"security_incidents","case_elapsed_ms":344.05798999999996}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0003446444888966697,"true":0.9996553555111019},"predicted":"true","confidence":0.9996553555111019,"entropy_confidence":0.9955385599149712,"candidate_mass":0.9999760479729373,"accepted":true,"accepted_predicted":"true","input_tokens":282,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0003447038924591452,"brier_hard":2.3755964745471318e-07,"soft_accuracy":0.7598207848657726,"brier_soft":0.11486937885030593,"tvd":0.23965535551110256,"kl":1.3627013602781426,"case_id":"typed:383/r0","base_id":"typed:383","repeat":0,"workflow":"security_incidents","case_elapsed_ms":354.195198}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.0003166328517567137,"contain":0.00022099434591908357,"investigate":0.998662153351787,"monitor":0.000800219450537104},"predicted":"investigate","confidence":0.998662153351787,"entropy_confidence":0.9917372803318499,"candidate_mass":0.9999877359932828,"accepted":true,"accepted_predicted":"investigate","input_tokens":334,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":8.417373440850344,"brier_hard":1.9968848972914228,"soft_accuracy":0.34971525971073386,"brier_soft":0.746229236562605,"tvd":0.6486625033514372,"kl":4.441914456376332,"case_id":"typed:383/r0","base_id":"typed:383","repeat":0,"workflow":"security_incidents","case_elapsed_ms":354.195198}
{"id":"severity","type":"ordinal","probabilities":{"0":3.0073704121726994e-05,"1":0.00013151008867613462,"2":0.015422669066781365,"3":0.9844120657191774,"4":3.681421244690216e-06},"predicted":"3","confidence":0.9844120657191774,"entropy_confidence":0.9494591779592548,"candidate_mass":0.999997705951254,"accepted":true,"accepted_predicted":"3","input_tokens":354,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.015710703613109403,"brier_hard":0.00048086062917066517,"score_mae":0.08243922893525024,"within_one":true,"case_id":"typed:383/r0","base_id":"typed:383","repeat":0,"workflow":"security_incidents","case_elapsed_ms":354.195198}
{"id":"true_positive","type":"binary","probabilities":{"false":1.424057504277225e-06,"true":0.9999985759424966},"predicted":"true","confidence":0.9999985759424966,"entropy_confidence":0.9999702881008431,"candidate_mass":0.9999980662927159,"accepted":true,"accepted_predicted":"true","input_tokens":300,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.424058517333901e-06,"brier_hard":4.055879548372888e-12,"soft_accuracy":0.8599989746785978,"brier_soft":0.039199202531853744,"tvd":0.1399985759424962,"kl":1.1597411915952487,"case_id":"typed:383/r0","base_id":"typed:383","repeat":0,"workflow":"security_incidents","case_elapsed_ms":354.195198}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.2450384134320406e-05,"1":8.324648447372409e-06,"2":5.525138153194557e-05,"3":0.9999239735858869},"predicted":"3","confidence":0.9999239735858869,"entropy_confidence":0.9993827674105515,"candidate_mass":0.9999949828028736,"accepted":true,"accepted_predicted":"3","input_tokens":329,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":7.602930426739173e-05,"brier_hard":9.057042640947227e-09,"score_mae":0.28989074816917215,"within_one":true,"case_id":"typed:383/r0","base_id":"typed:383","repeat":0,"workflow":"security_incidents","case_elapsed_ms":354.195198}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0034317819642950605,"true":0.9965682180357065},"predicted":"true","confidence":0.9965682180357065,"entropy_confidence":0.9669620871670359,"candidate_mass":0.9999552517768914,"accepted":true,"accepted_predicted":"true","input_tokens":334,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0034376840349672386,"brier_hard":2.3554254900911026e-05,"soft_accuracy":0.5662094267835736,"brier_soft":0.3696301145371667,"tvd":0.4299012180357057,"kl":1.776740560988118,"case_id":"typed:384/r0","base_id":"typed:384","repeat":0,"workflow":"security_incidents","case_elapsed_ms":363.7864}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.451013015870141e-08,"contain":4.8246753439937894e-08,"investigate":0.9999996435027749,"monitor":2.337403415668927e-07},"predicted":"investigate","confidence":0.9999996435027749,"entropy_confidence":0.9999956999131012,"candidate_mass":0.9999996380396021,"accepted":true,"accepted_predicted":"investigate","input_tokens":386,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.564972886187329e-07,"brier_hard":1.8960432747473269e-13,"soft_accuracy":0.6166662142171612,"brier_soft":0.20499994857179882,"tvd":0.38333326016915825,"kl":3.2325200584226654,"case_id":"typed:384/r0","base_id":"typed:384","repeat":0,"workflow":"security_incidents","case_elapsed_ms":363.7864}
{"id":"severity","type":"ordinal","probabilities":{"0":1.5803357253723492e-05,"1":5.186150987148655e-05,"2":0.9990421750486996,"3":0.0008899831340514572,"4":1.7695012360344676e-07},"predicted":"2","confidence":0.9990421750486996,"entropy_confidence":0.9950926608733642,"candidate_mass":0.9999982734781147,"accepted":true,"accepted_predicted":"2","input_tokens":406,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":7.024308045921632,"brier_hard":1.9963060962673063,"score_mae":0.5325261311900804,"within_one":true,"case_id":"typed:384/r0","base_id":"typed:384","repeat":0,"workflow":"security_incidents","case_elapsed_ms":363.7864}
{"id":"true_positive","type":"binary","probabilities":{"false":3.717789425795569e-08,"true":0.9999999628221061},"predicted":"true","confidence":0.9999999628221061,"entropy_confidence":0.9999990287767844,"candidate_mass":0.9999996678921886,"accepted":true,"accepted_predicted":"true","input_tokens":352,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.7177894565682725e-08,"brier_hard":2.7643916144056253e-15,"soft_accuracy":0.7666669801717653,"brier_soft":0.10888854307868455,"tvd":0.23333296282210594,"kl":1.9453708628877633,"case_id":"typed:384/r0","base_id":"typed:384","repeat":0,"workflow":"security_incidents","case_elapsed_ms":363.7864}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.355126253281456e-06,"1":6.52976818776753e-06,"2":6.783008235464698e-05,"3":0.9999172850232059},"predicted":"3","confidence":0.9999172850232059,"entropy_confidence":0.9993439833522395,"candidate_mass":0.9999960432407824,"accepted":true,"accepted_predicted":"3","input_tokens":381,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":8.271839786647293e-05,"brier_hard":1.1555133465585797e-08,"score_mae":0.6665610450025148,"within_one":true,"case_id":"typed:384/r0","base_id":"typed:384","repeat":0,"workflow":"security_incidents","case_elapsed_ms":363.7864}
{"id":"credential_compromise","type":"binary","probabilities":{"false":8.172511960990712e-05,"true":0.9999182748803904},"predicted":"true","confidence":0.9999182748803904,"entropy_confidence":0.9987723663627505,"candidate_mass":0.999983891672868,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":8.172845928912294e-05,"brier_hard":1.3357990350453287e-08,"soft_accuracy":0.5499918274880392,"brier_soft":0.4048529081426927,"tvd":0.4499182748803902,"kl":3.5473732507442355,"case_id":"typed:385/r0","base_id":"typed:385","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.783}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.4348892251750282e-05,"contain":3.8951298149750806e-05,"investigate":0.9998882109584845,"monitor":4.8488851113096336e-05},"predicted":"investigate","confidence":0.9998882109584845,"entropy_confidence":0.9991000333169647,"candidate_mass":0.9999930834755257,"accepted":true,"accepted_predicted":"investigate","input_tokens":382,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00011179529037615061,"brier_hard":1.695803066666973e-08,"soft_accuracy":0.5332916251159997,"brier_soft":0.35743750419900006,"tvd":0.46655521095848496,"kl":3.742468097685721,"case_id":"typed:385/r0","base_id":"typed:385","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.783}
{"id":"severity","type":"ordinal","probabilities":{"0":2.856360050264972e-06,"1":2.7442678531888244e-05,"2":0.013198314652810643,"3":0.9867703765048175,"4":1.009803789035876e-06},"predicted":"3","confidence":0.9867703765048175,"entropy_confidence":0.9561348285735242,"candidate_mass":0.9999993750565891,"accepted":true,"accepted_predicted":"3","input_tokens":402,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.013317914533498528,"brier_hard":0.0003492192097779832,"score_mae":0.04659375928623799,"within_one":true,"case_id":"typed:385/r0","base_id":"typed:385","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.783}
{"id":"true_positive","type":"binary","probabilities":{"false":1.795406354464538e-07,"true":0.9999998204593635},"predicted":"true","confidence":0.9999998204593635,"entropy_confidence":0.9999957176180968,"candidate_mass":0.999999154809876,"accepted":true,"accepted_predicted":"true","input_tokens":348,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":1.795406525864997e-07,"brier_hard":6.446967992023793e-14,"soft_accuracy":0.7433329126236764,"brier_soft":0.13175571344943882,"tvd":0.256666820459364,"kl":2.143509708540391,"case_id":"typed:385/r0","base_id":"typed:385","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.783}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0001756981456847952,"1":0.00018063826140046993,"2":0.008866115436755705,"3":0.9907775481561596},"predicted":"3","confidence":0.9907775481561596,"entropy_confidence":0.9609369204809882,"candidate_mass":0.9999941704777155,"accepted":true,"accepted_predicted":"3","input_tokens":377,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.009265241942454582,"brier_hard":0.00016372512096971199,"score_mae":0.4469125136033907,"within_one":true,"case_id":"typed:385/r0","base_id":"typed:385","repeat":0,"workflow":"security_incidents","case_elapsed_ms":331.783}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0005437207015761593,"true":0.9994562792984254},"predicted":"true","confidence":0.9994562792984254,"entropy_confidence":0.9933192225189966,"candidate_mass":0.9999459115150643,"accepted":true,"accepted_predicted":"true","input_tokens":330,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0005438685712775703,"brier_hard":5.912644026432477e-07,"soft_accuracy":0.6831336361052369,"brier_soft":0.1998678554287796,"tvd":0.31612327929842465,"kl":1.7564496248229033,"case_id":"typed:386/r0","base_id":"typed:386","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.59700000000004}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.103819414080902e-05,"contain":2.2983742661090533e-05,"investigate":0.9999120233536503,"monitor":4.395470954722455e-05},"predicted":"investigate","confidence":0.9999120233536503,"entropy_confidence":0.9992779392470934,"candidate_mass":0.9999942203619918,"accepted":true,"accepted_predicted":"investigate","input_tokens":382,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":8.79805165218841e-05,"brier_hard":1.0642764833745893e-08,"soft_accuracy":0.4999707347668657,"brier_soft":0.3798830144838073,"tvd":0.4999115233531506,"kl":4.038171852567326,"case_id":"typed:386/r0","base_id":"typed:386","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.59700000000004}
{"id":"severity","type":"ordinal","probabilities":{"0":7.61458810257832e-05,"1":0.000541870275380242,"2":0.9972505700265669,"3":0.002130483676236469,"4":9.301407899362971e-07},"predicted":"2","confidence":0.9972505700265669,"entropy_confidence":0.9871624482173949,"candidate_mass":0.9999969935968951,"accepted":true,"accepted_predicted":"2","input_tokens":402,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0027532165983217174,"brier_hard":1.2397748329222038e-05,"score_mae":0.24856181807961697,"within_one":true,"case_id":"typed:386/r0","base_id":"typed:386","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.59700000000004}
{"id":"true_positive","type":"binary","probabilities":{"false":5.650422980208036e-09,"true":0.9999999943495776},"predicted":"true","confidence":0.9999999943495776,"entropy_confidence":0.9999998370322585,"candidate_mass":0.9999998083830844,"accepted":true,"accepted_predicted":"true","input_tokens":348,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.65042236819609e-09,"brier_hard":6.38545526138712e-17,"soft_accuracy":0.7833329967980979,"brier_soft":0.09388917288095959,"tvd":0.21666699434957737,"kl":1.8042889507990845,"case_id":"typed:386/r0","base_id":"typed:386","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.59700000000004}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.017811828617755637,"1":0.7740046875738262,"2":0.122623329817007,"3":0.08556015399141216},"predicted":"1","confidence":0.7740046875738262,"entropy_confidence":0.4678465971189544,"candidate_mass":0.9999595216050375,"accepted":false,"accepted_predicted":null,"input_tokens":377,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":2.4585355949634256,"brier_hard":1.450637230608585,"score_mae":1.107401190817923,"within_one":false,"case_id":"typed:386/r0","base_id":"typed:386","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.59700000000004}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0005350446341738479,"true":0.9994649553658275},"predicted":"true","confidence":0.9994649553658275,"entropy_confidence":0.9934134067867535,"candidate_mass":0.9999604258237332,"accepted":true,"accepted_predicted":"true","input_tokens":324,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.533160386185887,"brier_hard":1.9978603940088284,"soft_accuracy":0.3168631826758344,"brier_soft":0.9324260957035071,"tvd":0.6827979553658268,"kl":4.523494999546584,"case_id":"typed:387/r0","base_id":"typed:387","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.69780000000003}
{"id":"disposition","type":"choice","probabilities":{"close_benign":6.8229158343277145e-06,"contain":7.737314240250465e-06,"investigate":0.9999589746111407,"monitor":2.6465158784869925e-05},"predicted":"investigate","confidence":0.9999589746111407,"entropy_confidence":0.999644965028387,"candidate_mass":0.999996369729554,"accepted":true,"accepted_predicted":"investigate","input_tokens":376,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":4.1026230423588565e-05,"brier_hard":2.4899053727003456e-09,"soft_accuracy":0.649979117150972,"brier_soft":0.17051529518824282,"tvd":0.3499589746111406,"kl":2.851024899098787,"case_id":"typed:387/r0","base_id":"typed:387","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.69780000000003}
{"id":"severity","type":"ordinal","probabilities":{"0":2.0304521849347288e-05,"1":0.00012369819979894762,"2":0.9961902446309666,"3":0.0036653905475876264,"4":3.620997956492222e-07},"predicted":"2","confidence":0.9961902446309666,"entropy_confidence":0.9840324655327847,"candidate_mass":0.999997149081636,"accepted":true,"accepted_predicted":"2","input_tokens":396,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.608820387905336,"brier_hard":1.9850776732042454,"score_mae":1.0798311924963224,"within_one":false,"case_id":"typed:387/r0","base_id":"typed:387","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.69780000000003}
{"id":"true_positive","type":"binary","probabilities":{"false":3.545553655181315e-08,"true":0.9999999645444635},"predicted":"true","confidence":0.9999999645444635,"entropy_confidence":0.9999990713447275,"candidate_mass":0.9999996643683703,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.545553713529304e-08,"brier_hard":2.5141901411581203e-15,"soft_accuracy":0.5166669988181252,"brier_soft":0.4672215092306792,"tvd":0.4833329645444635,"kl":4.110476939710632,"case_id":"typed:387/r0","base_id":"typed:387","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.69780000000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.049915974364328e-06,"1":9.98747710559055e-06,"2":0.00012331306610118358,"3":0.9998576495408192},"predicted":"3","confidence":0.9998576495408192,"entropy_confidence":0.9989379263773721,"candidate_mass":0.9999963556925278,"accepted":true,"accepted_predicted":"3","input_tokens":371,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":9.00078418340691,"brier_hard":1.9994687086008522,"score_mae":0.7498295622317657,"within_one":true,"case_id":"typed:387/r0","base_id":"typed:387","repeat":0,"workflow":"security_incidents","case_elapsed_ms":310.69780000000003}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.008743511881851513,"true":0.9912564881181478},"predicted":"true","confidence":0.9912564881181478,"entropy_confidence":0.9276566995259217,"candidate_mass":0.9999378246238185,"accepted":true,"accepted_predicted":"true","input_tokens":314,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":4.739443352876101,"brier_hard":1.9651788504726488,"soft_accuracy":0.28712186097781006,"brier_soft":1.0023113300587316,"tvd":0.7079234881181481,"kl":2.8030171461399336,"case_id":"typed:388/r0","base_id":"typed:388","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.0458}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.5735037293123634,"contain":0.000598000188012249,"investigate":0.4232535827282565,"monitor":0.002644687771368767},"predicted":"close_benign","confidence":0.5735037293123634,"entropy_confidence":0.49296161284576556,"candidate_mass":0.9999643518221341,"accepted":false,"accepted_predicted":null,"input_tokens":366,"reused_prefix_tokens":0,"expected":"investigate","correct":false,"accepted_correct":null,"nll_hard":0.8597837932008048,"brier_hard":0.6615503093486135,"soft_accuracy":0.29035859651646195,"brier_soft":0.3745627595502026,"tvd":0.5235037293123629,"kl":1.6673454130243408,"case_id":"typed:388/r0","base_id":"typed:388","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.0458}
{"id":"severity","type":"ordinal","probabilities":{"0":1.9056290949757703e-05,"1":7.007386600344752e-05,"2":0.9994672784413122,"3":0.0004432793609251336,"4":3.120408080902947e-07},"predicted":"2","confidence":0.9994672784413122,"entropy_confidence":0.9969943659745809,"candidate_mass":0.9999980071457022,"accepted":true,"accepted_predicted":"2","input_tokens":386,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":7.721310375101911,"brier_hard":1.9980484837232113,"score_mae":0.49299728300536394,"within_one":true,"case_id":"typed:388/r0","base_id":"typed:388","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.0458}
{"id":"true_positive","type":"binary","probabilities":{"false":0.0006665511028548653,"true":0.9993334488971456},"predicted":"true","confidence":0.9993334488971456,"entropy_confidence":0.9920059119737759,"candidate_mass":0.9999842859774111,"accepted":true,"accepted_predicted":"true","input_tokens":332,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0006667733468042192,"brier_hard":8.885807454334634e-07,"soft_accuracy":0.6830885983733209,"brier_soft":0.19971256740639481,"tvd":0.3160004488971454,"kl":1.692034523227886,"case_id":"typed:388/r0","base_id":"typed:388","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.0458}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0007538365246894655,"1":0.0007417333212529851,"2":0.8130324385330397,"3":0.185471991621018},"predicted":"2","confidence":0.8130324385330397,"entropy_confidence":0.6454264419973998,"candidate_mass":0.9999789419609707,"accepted":true,"accepted_predicted":"2","input_tokens":361,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.20698427043566525,"brier_hard":0.06935784715459434,"score_mae":0.09988958525038649,"within_one":true,"case_id":"typed:388/r0","base_id":"typed:388","repeat":0,"workflow":"security_incidents","case_elapsed_ms":312.0458}
{"id":"credential_compromise","type":"binary","probabilities":{"false":2.94798769095153e-05,"true":0.9999705201230894},"predicted":"true","confidence":0.9999705201230894,"entropy_confidence":0.9995138006611499,"candidate_mass":0.9999931469468162,"accepted":true,"accepted_predicted":"true","input_tokens":318,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":10.431802666041431,"brier_hard":1.9998820822304861,"soft_accuracy":0.28334577463298044,"brier_soft":1.0271386704963443,"tvd":0.7166375201230899,"kl":6.243434931338407,"case_id":"typed:389/r0","base_id":"typed:389","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.9294}
{"id":"disposition","type":"choice","probabilities":{"close_benign":8.306488272411293e-06,"contain":1.031046662276622e-05,"investigate":0.9999650160434344,"monitor":1.636700167119125e-05},"predicted":"investigate","confidence":0.9999650160434344,"entropy_confidence":0.9996891617475473,"candidate_mass":0.9999980695203856,"accepted":true,"accepted_predicted":"investigate","input_tokens":370,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":3.498456851852935e-05,"brier_hard":1.6670594300908372e-09,"soft_accuracy":0.5499862298378393,"brier_soft":0.2916237574124912,"tvd":0.449965566042884,"kl":3.7867977485926843,"case_id":"typed:389/r0","base_id":"typed:389","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.9294}
{"id":"severity","type":"ordinal","probabilities":{"0":0.0002807924968946948,"1":0.12857733551139475,"2":0.870436402476887,"3":0.0007051869115370862,"4":2.826032857774155e-07},"predicted":"2","confidence":0.870436402476887,"entropy_confidence":0.7564731169352832,"candidate_mass":0.9999963352864722,"accepted":true,"accepted_predicted":"2","input_tokens":390,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":2.0512247229020772,"brier_hard":1.517037567074412,"score_mae":0.7348998316129232,"within_one":true,"case_id":"typed:389/r0","base_id":"typed:389","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.9294}
{"id":"true_positive","type":"binary","probabilities":{"false":2.95324279655301e-07,"true":0.9999997046757204},"predicted":"true","confidence":0.9999997046757204,"entropy_confidence":0.9999931680001508,"candidate_mass":0.9999992444656817,"accepted":true,"accepted_predicted":"true","input_tokens":336,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.03519183127403,"brier_hard":1.999998818703056,"soft_accuracy":0.33333309844162345,"brier_soft":0.8888889902463684,"tvd":0.6666667046757204,"kl":5.774026019853816,"case_id":"typed:389/r0","base_id":"typed:389","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.9294}
{"id":"urgency","type":"ordinal","probabilities":{"0":8.08253039086617e-05,"1":7.124520756886222e-05,"2":0.0032171341603448836,"3":0.9966307953281782},"predicted":"3","confidence":0.9966307953281782,"entropy_confidence":0.9832146123311369,"candidate_mass":0.9999959552218112,"accepted":true,"accepted_predicted":"3","input_tokens":365,"reused_prefix_tokens":0,"expected":"1","correct":false,"accepted_correct":false,"nll_hard":9.549383003269124,"brier_hard":1.9931408133421544,"score_mae":1.7797308995127936,"within_one":false,"case_id":"typed:389/r0","base_id":"typed:389","repeat":0,"workflow":"security_incidents","case_elapsed_ms":349.9294}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0004793029457804617,"true":0.9995206970542191},"predicted":"true","confidence":0.9995206970542191,"entropy_confidence":0.9940235122244304,"candidate_mass":0.9999625636877821,"accepted":true,"accepted_predicted":"true","input_tokens":317,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0004794178481546609,"brier_hard":4.5946262766808883e-07,"soft_accuracy":0.6165551623264489,"brier_soft":0.2931539066961679,"tvd":0.38285369705421934,"kl":2.264505639157687,"case_id":"typed:390/r0","base_id":"typed:390","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.2859}
{"id":"disposition","type":"choice","probabilities":{"close_benign":4.544080325255053e-06,"contain":5.24450930007011e-06,"investigate":0.9999820141128454,"monitor":8.197297528281143e-06},"predicted":"investigate","confidence":0.9999820141128454,"entropy_confidence":0.9998314539931914,"candidate_mass":0.999996727623618,"accepted":true,"accepted_predicted":"investigate","input_tokens":369,"reused_prefix_tokens":0,"expected":"contain","correct":false,"accepted_correct":false,"nll_hard":12.158328876246657,"brier_hard":1.999953539645932,"soft_accuracy":0.28333211889861315,"brier_soft":0.9191001091344567,"tvd":0.7166487307795626,"kl":6.210676634222575,"case_id":"typed:390/r0","base_id":"typed:390","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.2859}
{"id":"severity","type":"ordinal","probabilities":{"0":5.234257987155987e-06,"1":6.084432921688354e-05,"2":0.7173728850285226,"3":0.28256036260796574,"4":6.737763061092543e-07},"predicted":"2","confidence":0.7173728850285226,"entropy_confidence":0.629644807128527,"candidate_mass":0.9999948756470993,"accepted":false,"accepted_predicted":null,"input_tokens":389,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.2638630782719034,"brier_hard":1.0293434932052432,"score_mae":0.5075096026846162,"within_one":true,"case_id":"typed:390/r0","base_id":"typed:390","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.2859}
{"id":"true_positive","type":"binary","probabilities":{"false":6.50134381902612e-08,"true":0.9999999349865604},"predicted":"true","confidence":0.9999999349865604,"entropy_confidence":0.9999983540299611,"candidate_mass":0.9999994867745239,"accepted":true,"accepted_predicted":"true","input_tokens":335,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.501344174520734e-08,"brier_hard":8.453494478080985e-15,"soft_accuracy":0.906666947122359,"brier_soft":0.017422073506411277,"tvd":0.09333293498656109,"kl":0.7707934873098692,"case_id":"typed:390/r0","base_id":"typed:390","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.2859}
{"id":"urgency","type":"ordinal","probabilities":{"0":4.237218197833337e-06,"1":2.8019232233470674e-06,"2":2.0240737669144074e-05,"3":0.9999727201209104},"predicted":"3","confidence":0.9999727201209104,"entropy_confidence":0.9997588665331095,"candidate_mass":0.999997488211356,"accepted":true,"accepted_predicted":"3","input_tokens":364,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":2.7280251192238945e-05,"brier_hard":1.1796840563382525e-09,"score_mae":0.3766284437612928,"within_one":true,"case_id":"typed:390/r0","base_id":"typed:390","repeat":0,"workflow":"security_incidents","case_elapsed_ms":314.2859}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0001342412026961276,"true":0.9998657587973026},"predicted":"true","confidence":0.9998657587973026,"entropy_confidence":0.9980796147522065,"candidate_mass":0.9999756002740197,"accepted":true,"accepted_predicted":"true","input_tokens":342,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.91587235598778,"brier_hard":1.9994630712306138,"soft_accuracy":0.17008859919377922,"brier_soft":1.3773543552484475,"tvd":0.8298657587973032,"kl":6.944310664978854,"case_id":"typed:391/r0","base_id":"typed:391","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.2681}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.0589375788195918e-05,"contain":1.1434001216734805e-05,"investigate":0.9999595155555902,"monitor":1.846106740617924e-05},"predicted":"investigate","confidence":0.9999595155555902,"entropy_confidence":0.9996442881857451,"candidate_mass":0.9999958689271431,"accepted":true,"accepted_predicted":"investigate","input_tokens":394,"reused_prefix_tokens":0,"expected":"monitor","correct":false,"accepted_correct":false,"nll_hard":10.89984650790797,"brier_hard":1.9998821111990404,"soft_accuracy":0.35999553424975933,"brier_soft":0.6229279648343342,"tvd":0.6399595155555895,"kl":5.486698587930635,"case_id":"typed:391/r0","base_id":"typed:391","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.2681}
{"id":"severity","type":"ordinal","probabilities":{"0":5.8397720286130225e-06,"1":2.7506413612747778e-05,"2":0.9964478103475086,"3":0.0035186743842861006,"4":1.6908256408194674e-07},"predicted":"2","confidence":0.9964478103475086,"entropy_confidence":0.9852202540330464,"candidate_mass":0.9999979098779221,"accepted":true,"accepted_predicted":"2","input_tokens":414,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.649670955675244,"brier_hard":1.9858832718379291,"score_mae":0.586520173408255,"within_one":true,"case_id":"typed:391/r0","base_id":"typed:391","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.2681}
{"id":"true_positive","type":"binary","probabilities":{"false":1.1804638193832747e-07,"true":0.9999998819536166},"predicted":"true","confidence":0.9999998819536166,"entropy_confidence":0.9999971129587824,"candidate_mass":0.9999995032669976,"accepted":true,"accepted_predicted":"true","input_tokens":360,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":15.952188222437488,"brier_hard":1.9999995278144973,"soft_accuracy":0.2766670527273048,"brier_soft":1.0464209162306515,"tvd":0.7233328819536173,"kl":6.306642486797278,"case_id":"typed:391/r0","base_id":"typed:391","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.2681}
{"id":"urgency","type":"ordinal","probabilities":{"0":9.949138985332891e-05,"1":0.00011918807931171041,"2":0.007297107437768864,"3":0.9924842130930672},"predicted":"3","confidence":0.9924842130930672,"entropy_confidence":0.9672616699449468,"candidate_mass":0.9999931045299126,"accepted":true,"accepted_predicted":"3","input_tokens":389,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":4.920277250753685,"brier_hard":1.9704839702447203,"score_mae":1.105499042234051,"within_one":false,"case_id":"typed:391/r0","base_id":"typed:391","repeat":0,"workflow":"security_incidents","case_elapsed_ms":326.2681}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.002634268518229966,"true":0.9973657314817709},"predicted":"true","confidence":0.9973657314817709,"entropy_confidence":0.9736331453493604,"candidate_mass":0.9999648314240774,"accepted":true,"accepted_predicted":"true","input_tokens":369,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.002637744308994856,"brier_hard":1.387874125227055e-05,"soft_accuracy":0.5828939570031412,"brier_soft":0.3428462054765117,"tvd":0.41403273148177044,"kl":1.7969930089177015,"case_id":"typed:392/r0","base_id":"typed:392","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.8614}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.3886294322499762e-08,"contain":1.9039483155732967e-08,"investigate":0.9999998498600629,"monitor":1.0721416004199881e-07},"predicted":"investigate","confidence":0.9999998498600629,"entropy_confidence":0.9999981039933284,"candidate_mass":0.9999998695812199,"accepted":true,"accepted_predicted":"investigate","input_tokens":421,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":1.5013994842051971e-07,"brier_hard":3.496993381607114e-14,"soft_accuracy":0.6333329191242656,"brier_soft":0.20468901924962957,"tvd":0.36666684986006265,"kl":3.087847465598911,"case_id":"typed:392/r0","base_id":"typed:392","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.8614}
{"id":"severity","type":"ordinal","probabilities":{"0":2.464727101327835e-06,"1":2.2691659536099586e-05,"2":0.9838927140310614,"3":0.016082014777895184,"4":1.1480440443879935e-07},"predicted":"2","confidence":0.9838927140310614,"entropy_confidence":0.9486325243644183,"candidate_mass":0.9999973352051846,"accepted":true,"accepted_predicted":"2","input_tokens":441,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":4.130053725947562,"brier_hard":1.9361394748879335,"score_mae":0.7939453767270379,"within_one":true,"case_id":"typed:392/r0","base_id":"typed:392","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.8614}
{"id":"true_positive","type":"binary","probabilities":{"false":4.2244433926751466e-08,"true":0.9999999577555672},"predicted":"true","confidence":0.9999999577555672,"entropy_confidence":0.9999989042064852,"candidate_mass":0.9999996917164835,"accepted":true,"accepted_predicted":"true","input_tokens":387,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.2244433733028394e-08,"brier_hard":3.569184303826813e-15,"soft_accuracy":0.7833329760615163,"brier_soft":0.09388914116610503,"tvd":0.21666695775556666,"kl":1.8042889794643813,"case_id":"typed:392/r0","base_id":"typed:392","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.8614}
{"id":"urgency","type":"ordinal","probabilities":{"0":1.5414464350404538e-05,"1":1.573205951152997e-05,"2":0.00024466312647441216,"3":0.9997241903496645},"predicted":"3","confidence":0.9997241903496645,"entropy_confidence":0.9980847586901764,"candidate_mass":0.9999968651620466,"accepted":true,"accepted_predicted":"3","input_tokens":416,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.0002758476928122259,"brier_hard":1.3641611208209147e-07,"score_mae":0.49967762936145377,"within_one":true,"case_id":"typed:392/r0","base_id":"typed:392","repeat":0,"workflow":"security_incidents","case_elapsed_ms":338.8614}
{"id":"credential_compromise","type":"binary","probabilities":{"false":5.4381442370243884e-05,"true":0.9999456185576281},"predicted":"true","confidence":0.9999456185576281,"entropy_confidence":0.9991511501756916,"candidate_mass":0.9999903151290319,"accepted":true,"accepted_predicted":"true","input_tokens":333,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":5.4382921096156686e-05,"brier_hard":5.914682548717429e-09,"soft_accuracy":0.5666597491047621,"brier_soft":0.37546072259841456,"tvd":0.4332786185576289,"kl":3.570907159389398,"case_id":"typed:393/r0","base_id":"typed:393","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.52579999999995}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.844422656866156e-05,"contain":4.187683681420177e-05,"investigate":0.9995360856364065,"monitor":0.0003435933002090527},"predicted":"investigate","confidence":0.9995360856364065,"entropy_confidence":0.9968491405400097,"candidate_mass":0.9999936728758424,"accepted":true,"accepted_predicted":"investigate","input_tokens":385,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0004640220051541373,"brier_hard":3.411800588404276e-07,"soft_accuracy":0.6331217429489485,"brier_soft":0.19338430433297485,"tvd":0.3662030856364073,"kl":2.2084802513691164,"case_id":"typed:393/r0","base_id":"typed:393","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.52579999999995}
{"id":"severity","type":"ordinal","probabilities":{"0":6.261210656621821e-05,"1":0.0002824958723001004,"2":0.9976526767253545,"3":0.0020015506291151167,"4":6.64666664074779e-07},"predicted":"2","confidence":0.9976526767253545,"entropy_confidence":0.9889987496858577,"candidate_mass":0.9999974541933594,"accepted":true,"accepted_predicted":"2","input_tokens":405,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.002350082556720383,"brier_hard":9.599856112140551e-06,"score_mae":0.19840484012298942,"within_one":true,"case_id":"typed:393/r0","base_id":"typed:393","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.52579999999995}
{"id":"true_positive","type":"binary","probabilities":{"false":3.343646661168658e-05,"true":0.9999665635333899},"predicted":"true","confidence":0.9999665635333899,"entropy_confidence":0.9994546215046597,"candidate_mass":0.9999946322492337,"accepted":true,"accepted_predicted":"true","input_tokens":351,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.343702562124784e-05,"brier_hard":2.2359945988452526e-09,"soft_accuracy":0.6333240836311955,"brier_soft":0.2688403398183833,"tvd":0.36663356353338905,"kl":3.0878685472820746,"case_id":"typed:393/r0","base_id":"typed:393","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.52579999999995}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0014995495394495948,"1":0.000992908980363394,"2":0.26726678025488604,"3":0.7302407612253008},"predicted":"3","confidence":0.7302407612253008,"entropy_confidence":0.5680201379014495,"candidate_mass":0.9999871933162997,"accepted":false,"accepted_predicted":null,"input_tokens":380,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":null,"nll_hard":1.3195079424423568,"brier_hard":1.0701527751900128,"score_mae":0.5762487531660381,"within_one":true,"case_id":"typed:393/r0","base_id":"typed:393","repeat":0,"workflow":"security_incidents","case_elapsed_ms":323.52579999999995}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.07122165016730911,"true":0.9287783498326916},"predicted":"true","confidence":0.9287783498326916,"entropy_confidence":0.629533839086959,"candidate_mass":0.9998780442251862,"accepted":true,"accepted_predicted":"true","input_tokens":341,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":2.6419584314022018,"brier_hard":1.7252584462358742,"soft_accuracy":0.41996166055013123,"brier_soft":0.5452005232482298,"tvd":0.5221113498326913,"kl":0.9219853085990679,"case_id":"typed:394/r0","base_id":"typed:394","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3336}
{"id":"disposition","type":"choice","probabilities":{"close_benign":0.624371471696737,"contain":0.0004332260877494981,"investigate":0.37034367534118845,"monitor":0.004851626874323708},"predicted":"close_benign","confidence":0.624371471696737,"entropy_confidence":0.5014313587107943,"candidate_mass":0.9999752347857472,"accepted":false,"accepted_predicted":null,"input_tokens":393,"reused_prefix_tokens":0,"expected":"investigate","correct":false,"accepted_correct":null,"nll_hard":0.993323852186716,"brier_hard":0.7863305478197627,"soft_accuracy":0.26759100734198216,"brier_soft":0.41312466159617517,"tvd":0.5477044716967376,"kl":1.5469106221492293,"case_id":"typed:394/r0","base_id":"typed:394","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3336}
{"id":"severity","type":"ordinal","probabilities":{"0":0.00012011569001517481,"1":0.0008808412931231297,"2":0.9968640683433518,"3":0.00213357327136571,"4":1.4014021456102374e-06},"predicted":"2","confidence":0.9968640683433518,"entropy_confidence":0.98536633949503,"candidate_mass":0.9999965021068248,"accepted":true,"accepted_predicted":"2","input_tokens":413,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":6.1499571126563275,"brier_hard":1.9894761666573582,"score_mae":0.26565169659749355,"within_one":true,"case_id":"typed:394/r0","base_id":"typed:394","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3336}
{"id":"true_positive","type":"binary","probabilities":{"false":0.8668245252145707,"true":0.13317547478543068},"predicted":"false","confidence":0.8668245252145707,"entropy_confidence":0.43391693654880725,"candidate_mass":0.9999419814307886,"accepted":true,"accepted_predicted":"false","input_tokens":359,"reused_prefix_tokens":0,"expected":"true","correct":false,"accepted_correct":false,"nll_hard":2.016087661061114,"brier_hard":1.5027695150269293,"soft_accuracy":0.44375332825074976,"brier_soft":0.3933694658742912,"tvd":0.44349152521457,"kl":0.5417684130751894,"case_id":"typed:394/r0","base_id":"typed:394","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3336}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.005377911767863334,"1":0.7962461218864297,"2":0.1890312652076576,"3":0.009344701138049917},"predicted":"1","confidence":0.7962461218864297,"entropy_confidence":0.5902111692382033,"candidate_mass":0.9999747539940176,"accepted":false,"accepted_predicted":null,"input_tokens":388,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":4.6729458195330515,"brier_hard":1.6511675489434294,"score_mae":0.8143242442841052,"within_one":true,"case_id":"typed:394/r0","base_id":"typed:394","repeat":0,"workflow":"security_incidents","case_elapsed_ms":319.3336}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0017270212218641935,"true":0.9982729787781345},"predicted":"true","confidence":0.9982729787781345,"entropy_confidence":0.9816608531524708,"candidate_mass":0.999973593332842,"accepted":true,"accepted_predicted":"true","input_tokens":376,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0017285142422487877,"brier_hard":5.965204601543169e-06,"soft_accuracy":0.6328724621588495,"brier_soft":0.2663623762211714,"tvd":0.3649399787781351,"kl":1.6764365388084643,"case_id":"typed:395/r0","base_id":"typed:395","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.61740000000003}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.11501531412837e-05,"contain":2.5517212246391767e-05,"investigate":0.9997618362230063,"monitor":0.00014149641160698172},"predicted":"investigate","confidence":0.9997618362230063,"entropy_confidence":0.9982387141536189,"candidate_mass":0.9999909835376484,"accepted":true,"accepted_predicted":"investigate","input_tokens":428,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0002381921424898862,"brier_hard":8.24566915824152e-08,"soft_accuracy":0.49990755628460715,"brier_soft":0.38470864233348984,"tvd":0.4997618362230058,"kl":3.860979037145809,"case_id":"typed:395/r0","base_id":"typed:395","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.61740000000003}
{"id":"severity","type":"ordinal","probabilities":{"0":2.0261413373384914e-05,"1":0.00024254837337264753,"2":0.9993697793549793,"3":0.0003671515985179678,"4":2.5925975707392536e-07},"predicted":"2","confidence":0.9993697793549793,"entropy_confidence":0.9964111479801785,"candidate_mass":0.9999983851327023,"accepted":true,"accepted_predicted":"2","input_tokens":448,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":7.909735720220848,"brier_hard":1.9980058467315858,"score_mae":0.7499154010820868,"within_one":true,"case_id":"typed:395/r0","base_id":"typed:395","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.61740000000003}
{"id":"true_positive","type":"binary","probabilities":{"false":3.2398229400270196e-07,"true":0.9999996760177076},"predicted":"true","confidence":0.9999996760177076,"entropy_confidence":0.9999925483176587,"candidate_mass":0.9999991067944133,"accepted":true,"accepted_predicted":"true","input_tokens":394,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":3.23982344920626e-07,"brier_hard":2.0992905264086286e-13,"soft_accuracy":0.7666668272092284,"brier_soft":0.10888827539516825,"tvd":0.2333326760177068,"kl":1.945371082771271,"case_id":"typed:395/r0","base_id":"typed:395","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.61740000000003}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0004740134360811275,"1":0.0004519118333213663,"2":0.6374775166872142,"3":0.36159655804338425},"predicted":"2","confidence":0.6374775166872142,"entropy_confidence":0.5225041200276744,"candidate_mass":0.999975535605915,"accepted":false,"accepted_predicted":null,"input_tokens":423,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":null,"nll_hard":1.0172261691947497,"brier_hard":0.8139369678967943,"score_mae":0.05647038066209742,"within_one":true,"case_id":"typed:395/r0","base_id":"typed:395","repeat":0,"workflow":"security_incidents","case_elapsed_ms":330.61740000000003}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.00030834894857258177,"true":0.9996916510514262},"predicted":"true","confidence":0.9996916510514262,"entropy_confidence":0.9959588955307239,"candidate_mass":0.9999758305654636,"accepted":true,"accepted_predicted":"true","input_tokens":257,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":8.08427846638314,"brier_hard":1.9987667943638554,"soft_accuracy":0.49000616697897087,"brier_soft":0.5195711583030589,"tvd":0.5096916510514269,"kl":3.430185964914887,"case_id":"typed:396/r0","base_id":"typed:396","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.9782}
{"id":"disposition","type":"choice","probabilities":{"close_benign":2.577559555157282e-05,"contain":4.900929995574205e-05,"investigate":0.9998965867092086,"monitor":2.8628395285275924e-05},"predicted":"investigate","confidence":0.9998965867092086,"entropy_confidence":0.9991620945090971,"candidate_mass":0.9999956085238151,"accepted":true,"accepted_predicted":"investigate","input_tokens":309,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00010341863831447242,"brier_hard":1.4580186537115742e-08,"soft_accuracy":0.6332822752676028,"brier_soft":0.193783915241398,"tvd":0.36656358670920797,"kl":2.7238810846001007,"case_id":"typed:396/r0","base_id":"typed:396","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.9782}
{"id":"severity","type":"ordinal","probabilities":{"0":1.4932896884306068e-05,"1":8.683380515458336e-05,"2":0.002268877202031035,"3":0.9975941817798876,"4":3.517431604407036e-05},"predicted":"3","confidence":0.9975941817798876,"entropy_confidence":0.9890920963436015,"candidate_mass":0.9999987686929592,"accepted":true,"accepted_predicted":"3","input_tokens":329,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.002408716850752078,"brier_hard":1.0944765399757279e-05,"score_mae":0.03754783081305568,"within_one":true,"case_id":"typed:396/r0","base_id":"typed:396","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.9782}
{"id":"true_positive","type":"binary","probabilities":{"false":4.4516209918694106e-07,"true":0.9999995548379021},"predicted":"true","confidence":0.9999995548379021,"entropy_confidence":0.9999899652180104,"candidate_mass":0.9999990915327969,"accepted":true,"accepted_predicted":"true","input_tokens":275,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":4.4516219693540254e-07,"brier_hard":3.963385879153837e-13,"soft_accuracy":0.5499999554837909,"brier_soft":0.4049991987086189,"tvd":0.44999955483790144,"kl":3.815843061812899,"case_id":"typed:396/r0","base_id":"typed:396","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.9782}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.066991784760256e-05,"1":2.1835743169930113e-05,"2":0.00013166842711671706,"3":0.9998158259118674},"predicted":"3","confidence":0.9998158259118674,"entropy_confidence":0.9986195477721498,"candidate_mass":0.9999957129866088,"accepted":true,"accepted_predicted":"3","input_tokens":304,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":true,"nll_hard":0.00018419105026270751,"brier_hard":5.2674112979444924e-08,"score_mae":0.4163996503330054,"within_one":true,"case_id":"typed:396/r0","base_id":"typed:396","repeat":0,"workflow":"security_incidents","case_elapsed_ms":320.9782}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.001315606390946824,"true":0.9986843936090545},"predicted":"true","confidence":0.9986843936090545,"entropy_confidence":0.9855128029418064,"candidate_mass":0.9999683724355288,"accepted":true,"accepted_predicted":"true","input_tokens":250,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0013164725608092232,"brier_hard":3.4616403517968314e-06,"soft_accuracy":0.6329821725061526,"brier_soft":0.2669632816241555,"tvd":0.3653513936090538,"kl":1.775945814465968,"case_id":"typed:397/r0","base_id":"typed:397","repeat":0,"workflow":"security_incidents","case_elapsed_ms":298.41339999999997}
{"id":"disposition","type":"choice","probabilities":{"close_benign":1.7845869506634523e-07,"contain":2.1948073183579076e-07,"investigate":0.9999991036501508,"monitor":4.984104238418957e-07},"predicted":"investigate","confidence":0.9999991036501508,"entropy_confidence":0.9999897082932777,"candidate_mass":0.9999994123197261,"accepted":true,"accepted_predicted":"investigate","input_tokens":302,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":8.963502509549129e-07,"brier_hard":1.1318753003065e-12,"soft_accuracy":0.6666665242690009,"brier_soft":0.15642103654143166,"tvd":0.33333210365015,"kl":2.799801110885341,"case_id":"typed:397/r0","base_id":"typed:397","repeat":0,"workflow":"security_incidents","case_elapsed_ms":298.41339999999997}
{"id":"severity","type":"ordinal","probabilities":{"0":6.108014563698547e-05,"1":0.0006772872283325567,"2":0.9982007544826782,"3":0.0010602473703813545,"4":6.307729702181662e-07},"predicted":"2","confidence":0.9982007544826782,"entropy_confidence":0.9909262358625347,"candidate_mass":0.9999980565292239,"accepted":true,"accepted_predicted":"2","input_tokens":322,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0018008661037178827,"brier_hard":4.823858089731146e-06,"score_mae":0.3769290613967138,"within_one":true,"case_id":"typed:397/r0","base_id":"typed:397","repeat":0,"workflow":"security_incidents","case_elapsed_ms":298.41339999999997}
{"id":"true_positive","type":"binary","probabilities":{"false":6.705375118290455e-09,"true":0.9999999932946242},"predicted":"true","confidence":0.9999999932946242,"entropy_confidence":0.999999808261609,"candidate_mass":0.9999998390643198,"accepted":true,"accepted_predicted":"true","input_tokens":268,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":6.705375832637544e-09,"brier_hard":8.992412023242082e-17,"soft_accuracy":0.7799999962449894,"brier_soft":0.09679999409926966,"tvd":0.2199999932946245,"kl":1.8324750268121237,"case_id":"typed:397/r0","base_id":"typed:397","repeat":0,"workflow":"security_incidents","case_elapsed_ms":298.41339999999997}
{"id":"urgency","type":"ordinal","probabilities":{"0":3.4644968262120116e-05,"1":3.111495781740756e-05,"2":0.0002641303029900843,"3":0.9996701097709291},"predicted":"3","confidence":0.9996701097709291,"entropy_confidence":0.9977026920497101,"candidate_mass":0.9999948380317466,"accepted":true,"accepted_predicted":"3","input_tokens":297,"reused_prefix_tokens":0,"expected":"2","correct":false,"accepted_correct":false,"nll_hard":8.239068004652381,"brier_hard":1.9988121396966725,"score_mae":1.1995697048765843,"within_one":false,"case_id":"typed:397/r0","base_id":"typed:397","repeat":0,"workflow":"security_incidents","case_elapsed_ms":298.41339999999997}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0004900504665286019,"true":0.9995099495334709},"predicted":"true","confidence":0.9995099495334709,"entropy_confidence":0.9939051818083958,"candidate_mass":0.9999733100847029,"accepted":true,"accepted_predicted":"true","input_tokens":279,"reused_prefix_tokens":0,"expected":"false","correct":false,"accepted_correct":false,"nll_hard":7.621002179247494,"brier_hard":1.9980402784328042,"soft_accuracy":0.44338853937957334,"brier_soft":0.6186655983847147,"tvd":0.5561769495334712,"kl":3.5558666658916875,"case_id":"typed:398/r0","base_id":"typed:398","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.128799}
{"id":"disposition","type":"choice","probabilities":{"close_benign":7.841612115177366e-05,"contain":8.772662997306382e-05,"investigate":0.9996297386385726,"monitor":0.00020411861030287796},"predicted":"investigate","confidence":0.9996297386385726,"entropy_confidence":0.9973560200572537,"candidate_mass":0.9999944609417781,"accepted":true,"accepted_predicted":"investigate","input_tokens":331,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.00037032992509009293,"brier_hard":1.9260293250094882e-07,"soft_accuracy":0.7131153828035428,"brier_soft":0.12191744205099218,"tvd":0.28629673863857247,"kl":1.6731186059822452,"case_id":"typed:398/r0","base_id":"typed:398","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.128799}
{"id":"severity","type":"ordinal","probabilities":{"0":8.503923662596774e-06,"1":2.68587116720015e-05,"2":0.9995872752009048,"3":0.0003771647054356495,"4":1.9745832498741222e-07},"predicted":"2","confidence":0.9995872752009048,"entropy_confidence":0.9976570816549598,"candidate_mass":0.9999990509904171,"accepted":true,"accepted_predicted":"2","input_tokens":351,"reused_prefix_tokens":0,"expected":"2","correct":true,"accepted_correct":true,"nll_hard":0.0004128099934170611,"brier_hard":3.1338872091462484e-07,"score_mae":0.3729993069369115,"within_one":true,"case_id":"typed:398/r0","base_id":"typed:398","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.128799}
{"id":"true_positive","type":"binary","probabilities":{"false":2.0701688715100402e-07,"true":0.9999997929831111},"predicted":"true","confidence":0.9999997929831111,"entropy_confidence":0.9999951047869554,"candidate_mass":0.9999992835082034,"accepted":true,"accepted_predicted":"true","input_tokens":297,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.0701691028149422e-07,"brier_hard":8.571198383627189e-14,"soft_accuracy":0.5599999751579726,"brier_soft":0.38719963565036275,"tvd":0.439999792983112,"kl":3.7278515222573425,"case_id":"typed:398/r0","base_id":"typed:398","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.128799}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0015956799397366535,"1":0.005336681088594682,"2":0.9503552326782616,"3":0.04271240629340645},"predicted":"2","confidence":0.9503552326782616,"entropy_confidence":0.8403804448904448,"candidate_mass":0.9999872598935168,"accepted":true,"accepted_predicted":"2","input_tokens":326,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":3.1532658554347,"brier_hard":1.8196056317030243,"score_mae":0.2224826346746629,"within_one":true,"case_id":"typed:398/r0","base_id":"typed:398","repeat":0,"workflow":"security_incidents","case_elapsed_ms":304.128799}
{"id":"credential_compromise","type":"binary","probabilities":{"false":0.0009087139198052442,"true":0.9990912860801938},"predicted":"true","confidence":0.9990912860801938,"entropy_confidence":0.989508056740371,"candidate_mass":0.9999760207213276,"accepted":true,"accepted_predicted":"true","input_tokens":285,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":0.0009091270505976349,"brier_hard":1.6515219760973265e-06,"soft_accuracy":0.5332724196778218,"brier_soft":0.4338615621047203,"tvd":0.4657582860801943,"kl":2.5778546235301887,"case_id":"typed:399/r0","base_id":"typed:399","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.8347}
{"id":"disposition","type":"choice","probabilities":{"close_benign":5.181070844993859e-05,"contain":3.2556412607910645e-05,"investigate":0.9997268293870885,"monitor":0.00018880349185442153},"predicted":"investigate","confidence":0.9997268293870885,"entropy_confidence":0.9980236952640258,"candidate_mass":0.999991787306284,"accepted":true,"accepted_predicted":"investigate","input_tokens":337,"reused_prefix_tokens":0,"expected":"investigate","correct":true,"accepted_correct":true,"nll_hard":0.0002732079307996783,"brier_hard":1.1401321180687234e-07,"soft_accuracy":0.6931821960478518,"brier_soft":0.13382314882570062,"tvd":0.30639313605339474,"kl":1.933590103399051,"case_id":"typed:399/r0","base_id":"typed:399","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.8347}
{"id":"severity","type":"ordinal","probabilities":{"0":7.756591396386e-06,"1":3.615710484296576e-05,"2":0.9956684828353043,"3":0.0042874776272397675,"4":1.258412172031545e-07},"predicted":"2","confidence":0.9956684828353043,"entropy_confidence":0.9825027320951558,"candidate_mass":0.9999981095362286,"accepted":true,"accepted_predicted":"2","input_tokens":357,"reused_prefix_tokens":0,"expected":"3","correct":false,"accepted_correct":false,"nll_hard":5.4520566846401906,"brier_hard":1.982799156288998,"score_mae":0.9057639409779603,"within_one":true,"case_id":"typed:399/r0","base_id":"typed:399","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.8347}
{"id":"true_positive","type":"binary","probabilities":{"false":2.1528146476699216e-06,"true":0.9999978471853526},"predicted":"true","confidence":0.9999978471853526,"entropy_confidence":0.9999563666710326,"candidate_mass":0.9999978755120339,"accepted":true,"accepted_predicted":"true","input_tokens":303,"reused_prefix_tokens":0,"expected":"true","correct":true,"accepted_correct":true,"nll_hard":2.152816964746885e-06,"brier_hard":9.269221813446225e-12,"soft_accuracy":0.5833326411989941,"brier_soft":0.3472191897599862,"tvd":0.41666484718535246,"kl":3.5232313684842405,"case_id":"typed:399/r0","base_id":"typed:399","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.8347}
{"id":"urgency","type":"ordinal","probabilities":{"0":0.0009423655544231814,"1":0.0007506632849621129,"2":0.24295319637983048,"3":0.7553537747807846},"predicted":"3","confidence":0.7553537747807846,"entropy_confidence":0.590529714322241,"candidate_mass":0.9999849976117096,"accepted":false,"accepted_predicted":null,"input_tokens":332,"reused_prefix_tokens":0,"expected":"3","correct":true,"accepted_correct":null,"nll_hard":0.2805690636119138,"brier_hard":0.1188794826933931,"score_mae":0.2627183803869766,"within_one":true,"case_id":"typed:399/r0","base_id":"typed:399","repeat":0,"workflow":"security_incidents","case_elapsed_ms":313.8347}
