{"data":{"modelVersion":"uw-hbb-1+0c423783","datasetDigest":"0x0c4237836bf3d6eb965bf0522aa3ac6e595db9fb02534e951d3702e8700645ca","asOf":"2026-10-16T10:34:15.955Z","observations":4926,"excluded":{"total":124,"byStatus":{"REVOKED":72,"UNKNOWN":52}},"builds":19,"taskTypes":16,"hyper":{"kTier":32,"kType":32,"kCell":2,"sBuild":0.5,"halfLifeDays":3},"hyperFrom":{"backtestRunVersion":"uw-hbb-1+0c423783","at":"2026-10-16T03:45:07.450Z"},"pricing":{"rho":0.15,"zeta":0.5,"costOfCapital":0.12,"durationDays":7,"minPremium":"10000","maxCoverageCap":"1000000000","declineAboveHi":0.5,"evidenceTiers":[{"minNEff":100,"fraction":1},{"minNEff":30,"fraction":0.5},{"minNEff":10,"fraction":0.2},{"minNEff":0,"fraction":0.05}],"bondFactor":1},"assumptions":["Failure = FAILURE or PARTIAL verdict, i.e. the event on which KaratumEscrow rejects. UNKNOWN, DISPUTED and REVOKED verdicts are excluded from the data.","Difficulty tier is a fixed function of task type, so it is a pooling level, not an independent cell axis.","Fingerprint-drift, certificate-state and over-claim adjustments are conservative assumptions, not fitted: the history to fit them does not exist yet.","Attempts are treated as independent although several builds run the same tasks; effective sample sizes are smaller than the counts shown.","Quotes come from a single trusted signer (Karatum). The chain enforces limits and the quote signature, not the price.","Testnet only: premiums and payouts are test tokens with no monetary value."],"backtest":{"span":{"to":"2026-10-15T22:03:37.087Z","from":"2026-10-03T21:14:46.457Z"},"cells":96,"folds":[{"fold":1,"hyper":{"kCell":32,"kTier":32,"kType":2,"sBuild":2,"halfLifeDays":null},"model":{"n":615,"auc":0.8051697530864198,"ece":0.06066576189717098,"brier":0.07246400622954059,"logLoss":0.27394200817350123,"failures":48,"meanPred":0.036692227799017395,"observedRate":0.07804878048780488},"nTest":615,"nTrain":2463,"testTo":"2026-10-10T06:41:13.349Z","testFrom":"2026-10-09T08:31:32.159Z","baselines":{"build":{"n":615,"auc":0.6917254556143445,"ece":0.05711569659378737,"brier":0.07330970000993547,"logLoss":0.28560066833645587,"failures":48,"meanPred":0.03283447443562646,"observedRate":0.07804878048780488},"global":{"n":615,"auc":0.5,"ece":0.03991490624845419,"brier":0.07355036809299438,"logLoss":0.29080369241492693,"failures":48,"meanPred":0.03813387423935069,"observedRate":0.07804878048780488},"cell-mle":{"n":615,"auc":0.8063822751322751,"ece":0.05617761640518064,"brier":0.07259410431402931,"logLoss":0.2751950865244178,"failures":48,"meanPred":0.04356545894765058,"observedRate":0.07804878048780488},"task-type":{"n":615,"auc":0.7535273368606702,"ece":0.06351494328012225,"brier":0.07443671188717182,"logLoss":0.2854776538012822,"failures":48,"meanPred":0.04523226176645132,"observedRate":0.07804878048780488}},"trainUntil":"2026-10-09T08:31:32.159Z","innerLogLoss":0.10106592453904226},{"fold":2,"hyper":{"kCell":32,"kTier":32,"kType":2,"sBuild":2,"halfLifeDays":null},"model":{"n":616,"auc":0.8431831065759637,"ece":0.08310590228352008,"brier":0.09847572375687504,"logLoss":0.352023983661804,"failures":112,"meanPred":0.16733461904734065,"observedRate":0.18181818181818182},"nTest":616,"nTrain":3078,"testTo":"2026-10-11T02:46:47.516Z","testFrom":"2026-10-10T06:41:23.233Z","baselines":{"build":{"n":616,"auc":0.6747271825396826,"ece":0.08134253577571797,"brier":0.15110902152663577,"logLoss":0.5153849191574602,"failures":112,"meanPred":0.11607192070935421,"observedRate":0.18181818181818182},"global":{"n":616,"auc":0.5,"ece":0.13571428571428532,"brier":0.1671786979254508,"logLoss":0.5980473008942638,"failures":112,"meanPred":0.04610389610389651,"observedRate":0.18181818181818182},"cell-mle":{"n":616,"auc":0.7400439342403629,"ece":0.07776090678023014,"brier":0.11208027450315083,"logLoss":0.4077407839418591,"failures":112,"meanPred":0.1392848991454461,"observedRate":0.18181818181818182},"task-type":{"n":616,"auc":0.8273455215419501,"ece":0.08840813696378026,"brier":0.10005926553850025,"logLoss":0.3625888493179157,"failures":112,"meanPred":0.1570953447514533,"observedRate":0.18181818181818182}},"trainUntil":"2026-10-10T06:41:23.233Z","innerLogLoss":0.1785588784299565},{"fold":3,"hyper":{"kCell":2,"kTier":32,"kType":2,"sBuild":0.5,"halfLifeDays":null},"model":{"n":616,"auc":0.9402194000502668,"ece":0.06408838913289716,"brier":0.09006239412021795,"logLoss":0.2659937524028185,"failures":143,"meanPred":0.19161461054215098,"observedRate":0.23214285714285715},"nTest":616,"nTrain":3694,"testTo":"2026-10-11T18:00:14.169Z","testFrom":"2026-10-11T02:47:02.783Z","baselines":{"build":{"n":616,"auc":0.47601235973329,"ece":0.060818312140241686,"brier":0.18302468725318705,"logLoss":0.5619524093025201,"failures":143,"meanPred":0.1789389730931761,"observedRate":0.23214285714285715},"global":{"n":616,"auc":0.5,"ece":0.16341991341991313,"brier":0.20495861912257876,"logLoss":0.6762727100398883,"failures":143,"meanPred":0.06872294372294402,"observedRate":0.23214285714285715},"cell-mle":{"n":616,"auc":0.9384748443945061,"ece":0.08104474208010982,"brier":0.09183197588602945,"logLoss":0.2944018464705006,"failures":143,"meanPred":0.21391139771746606,"observedRate":0.23214285714285715},"task-type":{"n":616,"auc":0.8359008855837609,"ece":0.09447359722135751,"brier":0.11019067073046,"logLoss":0.36643133686795804,"failures":143,"meanPred":0.1951258750466214,"observedRate":0.23214285714285715}},"trainUntil":"2026-10-11T02:47:02.783Z","innerLogLoss":0.2672445463827394},{"fold":4,"hyper":{"kCell":2,"kTier":2,"kType":2,"sBuild":0.5,"halfLifeDays":null},"model":{"n":616,"auc":0.8935158009692171,"ece":0.087307465665293,"brier":0.12169732776550267,"logLoss":0.49313270084756505,"failures":161,"meanPred":0.19990754062531185,"observedRate":0.26136363636363635},"nTest":616,"nTrain":4310,"testTo":"2026-10-15T22:03:37.087Z","testFrom":"2026-10-11T18:00:31.632Z","baselines":{"build":{"n":616,"auc":0.4440038222646918,"ece":0.09130006590042998,"brier":0.20824437851007296,"logLoss":0.6384838783290336,"failures":161,"meanPred":0.1724942822863852,"observedRate":0.26136363636363635},"global":{"n":616,"auc":0.5,"ece":0.16929499072356097,"brier":0.22171347983450435,"logLoss":0.6947522943681039,"failures":161,"meanPred":0.09206864564007539,"observedRate":0.26136363636363635},"cell-mle":{"n":616,"auc":0.92272882397106,"ece":0.09995902405633707,"brier":0.11928062271642763,"logLoss":0.36827270133055584,"failures":161,"meanPred":0.18548940265661007,"observedRate":0.26136363636363635},"task-type":{"n":616,"auc":0.8027097126476008,"ece":0.08077162088677427,"brier":0.1363964533986871,"logLoss":0.5773465736443817,"failures":161,"meanPred":0.20298998651035968,"observedRate":0.26136363636363635}},"trainUntil":"2026-10-11T18:00:31.632Z","innerLogLoss":0.3764216449233059}],"builds":19,"pooled":{"model":{"n":2463,"auc":0.8815975875868969,"ece":0.061083861156743156,"brier":0.09568428678319583,"logLoss":0.34630247834535,"failures":464,"meanPred":0.14893280168441408,"observedRate":0.18838814453917987},"baselines":{"build":{"n":2463,"auc":0.6172757715064429,"ece":0.06967873371213382,"brier":0.1539546761172121,"logLoss":0.5004426611485884,"failures":464,"meanPred":0.1251223671330407,"observedRate":0.18838814453917987},"global":{"n":2463,"auc":0.6226389056597265,"ece":0.12712141627900225,"brier":0.16688817184605664,"logLoss":0.5650803129928407,"failures":464,"meanPred":0.06126672826017747,"observedRate":0.18838814453917987},"cell-mle":{"n":2463,"auc":0.8595957461489365,"ece":0.04631380265960182,"brier":0.09895744376215292,"logLoss":0.3364274553658729,"failures":464,"meanPred":0.14560420144410685,"observedRate":0.18838814453917987},"task-type":{"n":2463,"auc":0.8185310327577582,"ece":0.06815764788866822,"brier":0.10528329429390353,"logLoss":0.39800677269314594,"failures":464,"meanPred":0.15015344866924932,"observedRate":0.18838814453917987}},"baselineLabels":{"build":"Per build (Laplace)","global":"Global failure rate","cell-mle":"Per build x task type, no pooling (Laplace)","task-type":"Per task type (Laplace)"}},"caveats":["4926 decided verdicts over 12.0 days, 19 builds, 16 task types, 96 non-empty cells: small and short. Intervals and ratios below are noisy.","The data is non-stationary: new task suites (frontier, code) entered during the window, so later folds contain task types the early model never saw. That is the realistic test for sparse cells, and the reason pooling matters.","Attempts are not independent (the same task is run by several builds; builds are re-run on the same suites). Effective sample size is smaller than the counts shown.","Drift, certificate-age and over-claim adjustments are NOT fitted: their data history is too short or not recorded as-of. They are conservative priors; the diagnostics table reports whether held-out outcomes contradict them.","Held-out windows are 1-3 days long. Nothing here shows how the model behaves over months or on agents and tasks outside this suite."],"pricing":{"lossRatio":0.9642071429811787,"lossRatioPure":1.2649204366568685,"lossRatioByBin":[{"n":879,"bin":"0-1%","lossRatio":2.54856045978299},{"n":409,"bin":"1-2%","lossRatio":1.4749205432648587},{"n":424,"bin":"2-5%","lossRatio":1.6582901616212449},{"n":88,"bin":"5-10%","lossRatio":0.8510962274153872},{"n":152,"bin":"10-20%","lossRatio":1.86902567241746},{"n":85,"bin":"20-35%","lossRatio":0.5657670634177899},{"n":186,"bin":"35-50%","lossRatio":0.5872134276137106},{"n":240,"bin":"50-100%","lossRatio":0.9009815160785194}],"premiumRateMean":0.19538140316686828,"realizedLossRate":0.18838814453917987,"pureExpectedLossRate":0.14893280168441408},"nExcluded":124,"taskTypes":16,"calibration":[{"n":879,"hi":0.01,"lo":0,"obsHi":0.041823979809328865,"obsLo":0.022490001713027185,"meanPred":0.003281309916924212,"observed":0.030716723549488054},{"n":409,"hi":0.02,"lo":0.01,"obsHi":0.05815656142376587,"obsLo":0.02614151266927175,"meanPred":0.01271418378154706,"observed":0.039119804400977995},{"n":424,"hi":0.05,"lo":0.02,"obsHi":0.17669674563517992,"obsLo":0.12024310891690893,"meanPred":0.03388391309676085,"observed":0.14622641509433962},{"n":88,"hi":0.1,"lo":0.05,"obsHi":0.19439500143717484,"obsLo":0.07797968370639746,"meanPred":0.06351075664401445,"observed":0.125},{"n":152,"hi":0.2,"lo":0.1,"obsHi":0.42114175952180377,"obsLo":0.29444785633187737,"meanPred":0.11938073643749758,"observed":0.35526315789473684},{"n":85,"hi":0.35,"lo":0.2,"obsHi":0.35602217219638893,"obsLo":0.19931054297340697,"meanPred":0.3079846482501771,"observed":0.27058823529411763},{"n":186,"hi":0.5,"lo":0.35,"obsHi":0.4142538920739639,"obsLo":0.2995867232429026,"meanPred":0.44859001160104023,"observed":0.3548387096774194},{"n":240,"hi":1,"lo":0.5,"obsHi":0.8876936169357781,"obsLo":0.812742229269523,"meanPred":0.8792461792909521,"observed":0.8541666666666666}],"diagnostics":{"tier":[{"n":117,"oe":0,"oeHi":8.613804332898242,"oeLo":0,"group":"routine","failures":0,"predictedFailures":0.3070482696583507},{"n":251,"oe":0.6771092343159278,"oeHi":1.0874624453160524,"oeLo":0.41710572041319105,"group":"adversarial","failures":11,"predictedFailures":16.24553239347432},{"n":1615,"oe":1.1046839240030228,"oeHi":1.1904058922000718,"oeLo":1.023420056296352,"group":"frontier","failures":366,"predictedFailures":331.31648976454056},{"n":480,"oe":4.590442774293706,"oeHi":5.367591774986709,"oeLo":3.903805705903199,"group":"code","failures":87,"predictedFailures":18.95242012103854}],"novelty":[{"n":2345,"oe":1.194997639499212,"oeHi":1.282635734595349,"oeLo":1.112064282098851,"group":"task type seen in training","failures":433,"predictedFailures":362.34381197728345},{"n":118,"oe":6.923230309072791,"oeHi":8.805623412856383,"oeLo":5.321211354407617,"group":"task type NEVER seen in training","failures":31,"predictedFailures":4.477678571428566}],"overClaim":[{"n":789,"oe":2.2316499337539564,"oeHi":2.675228100309461,"oeLo":1.8552582703000415,"group":"over-claim rate <= 3.3%","failures":73,"predictedFailures":32.711223608984},{"n":745,"oe":1.164845012367499,"oeHi":1.3253290133541058,"oeLo":1.0200887924494253,"group":"over-claim rate 3.3-15.3%","failures":130,"predictedFailures":111.60283009306141},{"n":544,"oe":0.9702559286692705,"oeHi":1.0848621007303254,"oeLo":0.8634617670824984,"group":"over-claim rate > 15.3%","failures":150,"predictedFailures":154.59838540304375}],"evidenceAge":[{"n":1730,"oe":1.0729168334677335,"oeHi":1.165945325132956,"oeLo":0.9857689862846218,"group":"build evidence < 1 day old","failures":314,"predictedFailures":292.66014867632623},{"n":348,"oe":6.237714073643823,"oeHi":7.955581907978363,"oeLo":4.853052534843226,"group":"build evidence 1-3 days old","failures":39,"predictedFailures":6.252290428762433},{"n":0,"oe":null,"oeHi":null,"oeLo":null,"group":"build evidence >= 3 days old","failures":0,"predictedFailures":0}]},"generatedAt":"2026-10-16T03:45:06.081Z","nObservations":4926,"intervalCoverage":{"cells":124,"level":0.9,"covered":103,"coverage":0.8306451612903226,"coveredExcludingSandwich":0.8378378378378378},"calibrationByTier":{"code":{"n":480,"auc":0.7934251703664708,"ece":0.1417657914145031,"brier":0.15685095796404325,"logLoss":0.5413318824396505,"failures":87,"meanPred":0.039484208585496956,"calibration":[{"n":289,"hi":0.02,"lo":0.01,"obsHi":0.07373062114059596,"obsLo":0.03153297428908363,"meanPred":0.011967173887803812,"observed":0.04844290657439446},{"n":90,"hi":0.05,"lo":0.02,"obsHi":0.4074222404582541,"obsLo":0.24740062569311,"meanPred":0.039139654867755246,"observed":0.32222222222222224},{"n":101,"hi":0.2,"lo":0.1,"obsHi":0.5174355982808931,"obsLo":0.35721006478448813,"meanPred":0.11852809831054692,"observed":0.43564356435643564}],"observedRate":0.18125},"routine":{"n":117,"auc":null,"ece":0.0026243441851141083,"brier":0.00000695037278363522,"logLoss":0.002627825582809061,"failures":0,"meanPred":0.0026243441851141083,"calibration":[{"n":117,"hi":0.01,"lo":0,"obsHi":0.022605587312752216,"obsLo":0,"meanPred":0.0026243441851141083,"observed":0}],"observedRate":0},"frontier":{"n":1615,"auc":0.8966025279239785,"ece":0.04814001207029418,"brier":0.0931945825032347,"logLoss":0.3451843142006375,"failures":366,"meanPred":0.20514952926596938,"calibration":[{"n":635,"hi":0.01,"lo":0,"obsHi":0.055947515849958745,"obsLo":0.029838132795904377,"meanPred":0.0033497348925953714,"observed":0.04094488188976378},{"n":61,"hi":0.02,"lo":0.01,"obsHi":0.07020558500904318,"obsLo":0.003665371466423616,"meanPred":0.014488177732803717,"observed":0.01639344262295082},{"n":334,"hi":0.05,"lo":0.02,"obsHi":0.12897098333625598,"obsLo":0.07508246358379839,"meanPred":0.032467695254277396,"observed":0.09880239520958084},{"n":58,"hi":0.1,"lo":0.05,"obsHi":0.26809424857732483,"obsLo":0.10593823007697675,"meanPred":0.06475138643012932,"observed":0.1724137931034483},{"n":51,"hi":0.2,"lo":0.1,"obsHi":0.30181804659616046,"obsLo":0.12096552377606497,"meanPred":0.12106929429675307,"observed":0.19607843137254902},{"n":85,"hi":0.35,"lo":0.2,"obsHi":0.35602217219638893,"obsLo":0.19931054297340697,"meanPred":0.3079846482501771,"observed":0.27058823529411763},{"n":151,"hi":0.5,"lo":0.35,"obsHi":0.45071408784384825,"obsLo":0.32157850771315755,"meanPred":0.4657849436949748,"observed":0.3841059602649007},{"n":240,"hi":1,"lo":0.5,"obsHi":0.8876936169357781,"obsLo":0.812742229269523,"meanPred":0.8792461792909521,"observed":0.8541666666666666}],"observedRate":0.22662538699690402},"adversarial":{"n":251,"auc":0.8689393939393939,"ece":0.026412873584388583,"brier":0.03933017764113749,"logLoss":0.14073138472655494,"failures":11,"meanPred":0.06472323662738773,"calibration":[{"n":127,"hi":0.01,"lo":0,"obsHi":0.034523841692046316,"obsLo":0.0017583973172826649,"meanPred":0.0035444211851966678,"observed":0.007874015748031496},{"n":59,"hi":0.02,"lo":0.01,"obsHi":0.0724753211679567,"obsLo":0.0037899219748320444,"meanPred":0.014539136802990037,"observed":0.01694915254237288},{"n":30,"hi":0.1,"lo":0.05,"obsHi":0.13641759405024848,"obsLo":0.007471031975078653,"meanPred":0.061112205724192466,"observed":0.03333333333333333},{"n":35,"hi":0.5,"lo":0.35,"obsHi":0.36221639543829426,"obsLo":0.1338853548787969,"meanPred":0.3744061617100614,"observed":0.22857142857142856}],"observedRate":0.043824701195219126}},"heldOutWorstGroups":[{"n":160,"oe":4.67317724998473,"oeHi":5.377784313309573,"oeLo":3.993350911950194,"group":"ext-claude-code-radikws-haiku-code-v1 / software.bugfix","failures":69,"predictedFailures":14.765115104552361},{"n":10,"oe":11.393556462239964,"oeHi":11.393556462239964,"oeLo":8.967050247610851,"group":"scripted-naive-frontier-v1 / web3.swap.route","failures":10,"predictedFailures":0.8776890721647425},{"n":54,"oe":1.1826968134819944,"oeHi":1.1826968134819942,"oeLo":1.1262582402492096,"group":"ext-claude-code-radikws-haiku-frontier-v1 / web3.swap.sandwich","failures":54,"predictedFailures":45.65836263735068},{"n":160,"oe":4.416468326331225,"oeHi":7.195656634403797,"oeLo":2.665607222728473,"group":"ext-claude-code-radikws-sonnet-code-v1 / software.bugfix","failures":10,"predictedFailures":2.264252624745309},{"n":61,"oe":1.1446584106379012,"oeHi":1.144658410637901,"oeLo":1.0960370396506762,"group":"ext-claude-code-radikws-sonnet-frontier-v1 / web3.swap.sandwich","failures":61,"predictedFailures":53.29100754696382},{"n":10,"oe":2.867122231142974,"oeHi":2.867122231142974,"oeLo":2.256506052162635,"group":"scripted-naive-frontier-v1 / web3.refusal","failures":10,"predictedFailures":3.487817816547541}]},"signer":"0xD2BFedd764D5F4f8c0Ae1fb89dd1E38361e7576a","chainId":11155111,"contracts":{"policyManager":"0x75cA4B98111631228A5346C364F62feE80D926D9","guaranteePool":"0x1aA4c5cE96dD2b31b92Bd1E4388957061D1F1842"},"decimals":6,"network":"testnet"}}