@tangle-network/agent-eval 0.208.2 → 0.210.0

package/dist/reporting.d.ts

@@ -1,7 +1,7 @@
 import { t as benjaminiHochberg } from "./multiplicity-DIWHvysC.js";
 import { d as PairedBootstrapResult, h as pairedBootstrap, u as PairedBootstrapOptions } from "./paired-promotion-decision-DPKz8ZnF.js";
-import { _ as bootstrapCi, a as ReleaseConfidenceIssue, c as ReleaseConfidenceStatus, d as assertReleaseConfidence, f as evaluateReleaseConfidence, g as Verdict, h as JudgeReplayGateArgs, i as ReleaseConfidenceInput, k as wilcoxonSignedRank, l as ReleaseConfidenceThresholds, m as BootstrapResult, n as ReleaseConfidenceAxis, o as ReleaseConfidenceMetrics, p as BootstrapOptions, r as ReleaseConfidenceAxisName, s as ReleaseConfidenceScorecard, u as ReleaseTraceEvidence, v as judgeReplayGate } from "./release-confidence-CkoXFx9I.js";
-import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-DvvesgR_.js";
+import { T as wilcoxonSignedRank, _ as bootstrapCi, a as ReleaseConfidenceIssue, c as ReleaseConfidenceStatus, d as assertReleaseConfidence, f as evaluateReleaseConfidence, g as Verdict, h as JudgeReplayGateArgs, i as ReleaseConfidenceInput, l as ReleaseConfidenceThresholds, m as BootstrapResult, n as ReleaseConfidenceAxis, o as ReleaseConfidenceMetrics, p as BootstrapOptions, r as ReleaseConfidenceAxisName, s as ReleaseConfidenceScorecard, u as ReleaseTraceEvidence, v as judgeReplayGate } from "./release-confidence-C4kI7lQi.js";
+import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-CK0ueoJd.js";
 import { a as PairedEvalueStep, c as SequentialDecision, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, n as InterimReleaseConfidenceInput, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence } from "./sequential-BhsrMupG.js";
-import { a as RubricPredictiveValidityReport, i as RubricPredictiveValidityInput, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity } from "./rubric-predictive-validity-DIUwJJ9Y.js";
+import { a as RubricPredictiveValidityReport, i as RubricPredictiveValidityInput, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity } from "./rubric-predictive-validity-BiNNEjpG.js";
 export { type BootstrapOptions, type BootstrapResult, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeReplayGateArgs, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type ParetoFigureSpec, type ParetoPoint, RESEARCH_REPORT_HARD_PAIR_FLOOR, type ReleaseConfidenceAxis, type ReleaseConfidenceAxisName, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseConfidenceStatus, type ReleaseConfidenceThresholds, type ReleaseTraceEvidence, type ResearchReport, type ResearchReportCandidate, type ResearchReportDecision, type ResearchReportMethodology, type ResearchReportOptions, type ResearchReportRecommendation, type RubricOutcomePair, type RubricPredictiveValidityInput, type RubricPredictiveValidityReport, type RubricRanking, type SequentialDecision, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type Verdict, assertReleaseConfidence, benjaminiHochberg, bootstrapCi, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, gainHistogram, judgeReplayGate, pairedBootstrap, pairedEvalueSequence, paretoChart, researchReport, rubricPredictiveValidity, summaryTable, wilcoxonSignedRank };

package/dist/reporting.js

@@ -1,8 +1,8 @@
 import { t as benjaminiHochberg } from "./multiplicity-D9cjmStq.js";
 import { r as pairedBootstrap } from "./paired-tests-jbwPdSz5.js";
-import { l as wilcoxonSignedRank } from "./paired-arms-CgWT7sHy.js";
-import { a as summaryTable, i as researchReport, n as gainHistogram, r as paretoChart, t as RESEARCH_REPORT_HARD_PAIR_FLOOR } from "./summary-report-fXurE2TR.js";
+import { a as wilcoxonSignedRank } from "./paired-arms-DzKn0TfY.js";
+import { a as summaryTable, i as researchReport, n as gainHistogram, r as paretoChart, t as RESEARCH_REPORT_HARD_PAIR_FLOOR } from "./summary-report-DtdRucM0.js";
 import { n as pairedEvalueSequence, t as evaluateInterimReleaseConfidence } from "./sequential-CzK5DarL.js";
-import { i as judgeReplayGate, n as evaluateReleaseConfidence, r as bootstrapCi, t as assertReleaseConfidence } from "./release-confidence-CiFg1EPq.js";
+import { i as judgeReplayGate, n as evaluateReleaseConfidence, r as bootstrapCi, t as assertReleaseConfidence } from "./release-confidence-Dbe9AS_X.js";
 import { t as rubricPredictiveValidity } from "./rubric-predictive-validity-BsgkBfT0.js";
 export { RESEARCH_REPORT_HARD_PAIR_FLOOR, assertReleaseConfidence, benjaminiHochberg, bootstrapCi, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, gainHistogram, judgeReplayGate, pairedBootstrap, pairedEvalueSequence, paretoChart, researchReport, rubricPredictiveValidity, summaryTable, wilcoxonSignedRank };

package/dist/{researcher-CZFZjj-u.d.ts → researcher-DuIJ9IaA.d.ts}

@@ -1,8 +1,8 @@
-import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-BHcB-mZX.js";
-import { a as RunRecord, c as RunSplitTag, d as RunTokenUsage, l as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, t as JudgeScoresRecord } from "./run-record-DF-fuu31.js";
+import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-BNp-Js9F.js";
+import { a as RunRecord, c as RunSplitTag, d as RunTokenUsage, l as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, t as JudgeScoresRecord } from "./run-record-p2AYTrCn.js";
 import { rt as RawProviderSink, x as ChatClient } from "./types-Lub7o6bN.js";
 import { s as TraceStore } from "./store-C4avFjIi.js";
-import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-DvvesgR_.js";
+import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-CK0ueoJd.js";
 import { o as TraceEmitter, r as RunCompleteHook } from "./emitter-HcCmFQl_.js";
 import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-zloavsM_.js";
 //#region src/eval-campaign.d.ts
@@ -290,4 +290,4 @@ interface Researcher {
 }
 //#endregion
 export { SteeringChange as a, CampaignRunContext as c, CampaignVariant as d, EvalCampaignOptions as f, runEvalCampaign as h, Researcher as i, CampaignRunOutcome as l, FailedRun as m, ExperimentResult as n, CampaignFactoryParams as o, EvalCampaignResult as p, FailureMode as r, CampaignIntegrityPolicy as s, ExperimentPlan as t, CampaignScenario as u };
-//# sourceMappingURL=researcher-CZFZjj-u.d.ts.map
+//# sourceMappingURL=researcher-DuIJ9IaA.d.ts.map

package/dist/{researcher-CZFZjj-u.d.ts.map → researcher-DuIJ9IaA.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"researcher-CZFZjj-u.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;UAuEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;EAKT,MAAM;;;UAIS;;;;;;EAMf,SAAS;EACT;;UAGQ;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;;;EAQA,cAAc,QAAQ,uBAAuB;;;;;;;EAO7C;;;;;;EAMA,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAWoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UClRM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ"}
+{"version":3,"file":"researcher-DuIJ9IaA.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;UAuEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;EAKT,MAAM;;;UAIS;;;;;;EAMf,SAAS;EACT;;UAGQ;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;;;EAQA,cAAc,QAAQ,uBAAuB;;;;;;;EAO7C;;;;;;EAMA,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAWoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UClRM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ"}

package/dist/{reward-hacking-Cyk-QHk9.d.ts → reward-hacking-C6Kp54Av.d.ts}

@@ -1,4 +1,4 @@
-import { a as RunRecord } from "./run-record-DF-fuu31.js";
+import { a as RunRecord } from "./run-record-p2AYTrCn.js";
 import { x as VerificationStrategySource } from "./verdict-E4eRNf7-.js";
 import { s as VerificationReport } from "./multi-layer-verifier-BUaQ4C17.js";
 //#region src/rl/verifiable-reward.d.ts
@@ -251,4 +251,4 @@ interface DetectRewardHackingInput {
 declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport;
 //#endregion
 export { detectRewardHacking as a, VerifiableRewardSource as c, filterDeterministicallyRewarded as d, RewardHackingSignal as i, extractVerifiableReward as l, RewardHackingFinding as n, VerifiableReward as o, RewardHackingReport as r, VerifiableRewardExtractionOptions as s, DetectRewardHackingInput as t, extractVerifiableRewardsFromRecords as u };
-//# sourceMappingURL=reward-hacking-Cyk-QHk9.d.ts.map
+//# sourceMappingURL=reward-hacking-C6Kp54Av.d.ts.map

package/dist/{reward-hacking-Cyk-QHk9.d.ts.map → reward-hacking-C6Kp54Av.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"reward-hacking-Cyk-QHk9.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KA+DY,yBAAyB;UAEpB;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA+Dc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KC1YvB;UAMK;EACf,QAAQ;;EAER;EACA;;EAEA,QAAQ;;UAGO;EACf,UAAU;;EAEV,kBAAkB;;;;;;EAMlB;;EAEA;;EAEA;;UAGe;;;;;EAKf,MAAM;;;;;EAKN,WAAW,KAAK;;;;;;EAMhB,WAAW,KAAK;;;;;;EAMhB,qBAAqB,KAAK;;;;;EAK1B;;;;;EAKA;IAAe;IAAkB;;;;;EAIjC,0BAA0B;;iBAgBZ,oBAAoB,OAAO,2BAA2B"}
+{"version":3,"file":"reward-hacking-C6Kp54Av.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KA+DY,yBAAyB;UAEpB;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA+Dc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KC1YvB;UAMK;EACf,QAAQ;;EAER;EACA;;EAEA,QAAQ;;UAGO;EACf,UAAU;;EAEV,kBAAkB;;;;;;EAMlB;;EAEA;;EAEA;;UAGe;;;;;EAKf,MAAM;;;;;EAKN,WAAW,KAAK;;;;;;EAMhB,WAAW,KAAK;;;;;;EAMhB,qBAAqB,KAAK;;;;;EAK1B;;;;;EAKA;IAAe;IAAkB;;;;;EAIjC,0BAA0B;;iBAgBZ,oBAAoB,OAAO,2BAA2B"}

package/dist/reward-nw2xZGZG.js.map

@@ -1 +1 @@
-{"version":3,"file":"reward-nw2xZGZG.js","names":[],"sources":["../src/rollout/reward.ts"],"sourcesContent":["/**\n * The two named score derivations every consumer must choose between.\n *\n * The anti-Goodhart gate (`outcome.realness.gated`) only holds if it is\n * impossible to read a run's score WITHOUT deciding whether the gate applies.\n * A bare `outcome.holdoutScore ?? outcome.searchScore` makes that decision\n * invisible — and silently answers \"no gate\", which is the wrong default on\n * every path that produces training data. So the expression lives here, once,\n * behind two names that force the caller to state the intent:\n *\n *   - `trainingScore` / `trainingReward` — GATED. Anything that becomes\n *     training data, or a reward a trainer consumes, uses these.\n *   - `observedScore` — RAW. Analysis, reporting, and reward-hack DETECTION\n *     need the ungated number; that is how a gamed run is visible at all.\n *\n * A leaf module on purpose: it imports only the `RunRecord` type, so gate and\n * reporting code can depend on it without pulling in the trace store that\n * `mint.ts` needs.\n */\n\nimport type { RunRecord } from '../run-record'\n\n/**\n * Which split's score wins when a record carries both. `'holdout'` is the\n * canonical \"real signal\" default; `'search'` exists because some callers\n * deliberately score on the search split when both are present.\n */\nexport type ScorePreference = 'holdout' | 'search'\n\n/** Only the outcome is read, so every accessor here accepts anything carrying one. */\ntype Scored = Pick<RunRecord, 'outcome'>\n\n/** True when the authenticity gate flagged the run as gamed (`realness.gated`). */\nexport function isRealnessGated(record: Scored): boolean {\n  return record.outcome.realness?.gated === true\n}\n\n/**\n * The RAW score recorded on ONE split, with no cross-split fallback and no\n * anti-Goodhart gate.\n *\n * The narrowest of the three raw readers, and the one every split-scoped\n * consumer wants: a per-split report, a promotion gate, or a paired comparison\n * asks \"what did this run score on the split I am summarising\", and answering\n * it with the other split's number silently mixes populations. `undefined` =\n * that split was never scored.\n *\n * Same warning as `observedScore`: this INCLUDES runs flagged as gamed. Never\n * feed it into training data.\n */\nexport function observedSplitScore(record: Scored, split: ScorePreference): number | undefined {\n  return split === 'search' ? record.outcome.searchScore : record.outcome.holdoutScore\n}\n\n/**\n * The RAW split score the run carries, with NO anti-Goodhart gate applied.\n *\n * INCLUDES RUNS FLAGGED AS GAMED (`outcome.realness.gated === true`); NEVER\n * feed this into training data — a fine-tune that sees it learns from gamed\n * successes. It is exported anyway because analysis, reporting, and\n * reward-hacking detection legitimately need the ungated number: forcing a\n * gamed run to 0 collapses the proxy signal toward ground truth and makes a\n * detector report \"clean\" on exactly the population that is being gamed.\n *\n * Returns `undefined` when the record carries neither score — an unscored run\n * is a labeled gap, not a measured zero, and each caller picks its own\n * sentinel (`?? 0`, `?? null`, skip, throw). Non-finite values are returned\n * as-is; callers that care keep their own `Number.isFinite` guard.\n */\nexport function observedScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  return observedSplitScore(record, prefer) ?? observedSplitScore(record, other)\n}\n\n/** Which split actually carried the score, or that none did. */\nexport type ScoreOrigin = 'holdout' | 'search' | 'unscored'\n\n/**\n * Where `observedScore` / `trainingScore` read their number from — the\n * provenance label a rollout line's `reward_source` is built from, and the\n * only supported way to ask \"was this run scored at all\" without respelling\n * the field access.\n */\nexport function scoreOrigin(record: Scored, prefer: ScorePreference = 'holdout'): ScoreOrigin {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  if (observedSplitScore(record, prefer) !== undefined) return prefer\n  if (observedSplitScore(record, other) !== undefined) return other\n  return 'unscored'\n}\n\n/**\n * The GATED score — the only derivation allowed to reach training data.\n *\n * A realness-gated run scores 0 no matter what it claims, so a fine-tune\n * cannot learn from a gamed success. An unscored run stays `undefined` (a\n * labeled gap), keeping \"we never measured this\" distinct from \"we measured\n * zero\"; callers that need a number apply their own sentinel.\n */\nexport function trainingScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  if (isRealnessGated(record)) return 0\n  return observedScore(record, prefer)\n}\n\n/**\n * `{reward, gated}` as written onto a minted `RolloutLine` — `trainingScore`\n * plus the flag itself, so the gate travels into the exported row and a\n * downstream filter can drop or down-weight the line.\n *\n * An unscored record yields `reward: null`, matching the schema's \"no verdict\n * exists — a labeled gap, never 0\" rule. It previously collapsed to 0, which\n * made a run nobody graded indistinguishable from one graded as a total\n * failure, and taught any trainer reading the row that the trajectory was bad.\n * A gated run still yields 0, because that IS a verdict: the gate decided.\n */\nexport function trainingReward(record: Scored): { reward: number | null; gated: boolean } {\n  return { reward: trainingScore(record) ?? null, gated: isRealnessGated(record) }\n}\n\n/**\n * A reward the CALLER computed, with the gate applied on top.\n *\n * Several published exporters accept a `rewardOf` hook so a consumer can drive\n * training off a verifiable signal instead of the headline judge score. That is\n * a real feature and the gate does not remove it — but a reward derived from a\n * gamed run is a gamed reward whatever its source, so it falls to 0 exactly\n * like the score it replaced.\n *\n * It lives here because the same hook exists under the same name in two\n * modules, and for a while only one of them gated it: `rl/exporters.toGrpoRows`\n * forced a gated run to 0 while `rl/preferences.extractPreferences` handed the\n * caller's number straight through, which put a gamed run on the CHOSEN side of\n * every DPO pair against its honest sibling. Two same-named hooks with\n * different gating behaviour was the defect; one implementation is the fix.\n *\n * `null` means \"no reward\" and is preserved: a non-finite or absent value is a\n * gap, and a gap is not a zero.\n */\nexport function trainingRewardOverride(record: Scored, value: number | null): number | null {\n  if (value === null || !Number.isFinite(value)) return null\n  return isRealnessGated(record) ? 0 : value\n}\n\n/**\n * The two fields a rollout line's `outcome` has to state TOGETHER — the reward\n * and whether the authenticity gate fired on it.\n *\n * Stating them together is the point. `realness_gated` used to be optional on\n * the wire, so a producer that wrote `reward` and simply never thought about\n * the flag emitted a row on which `isLineRealnessGated` was structurally false\n * — which is how a second minting door existed for every supervisor and worker\n * row without anyone noticing it had never consulted the gate. The wire schema\n * now requires the flag, and this pair is how a producer states it.\n */\nexport interface RolloutRewardFields {\n  reward: number | null\n  realness_gated: boolean\n  /**\n   * Whether a screen RAN, as distinct from its verdict. Omitted when the\n   * producer cannot tell — see `unscreenedRewardFields` and the field's own doc\n   * on `RolloutOutcome`.\n   */\n  realness_screened?: boolean\n}\n\n/**\n * `trainingReward` in the line's own field names — what `mintRolloutRows`\n * writes onto every row it produces.\n *\n * `realness_screened: true` is written only when the record actually carries an\n * `outcome.realness` verdict, i.e. a screen genuinely ran and reported. A\n * record without one gets the field OMITTED rather than `false`: the record\n * cannot distinguish \"the screen ran elsewhere and this pipeline did not record\n * it\" from \"no screen exists\", and `false` is a load-bearing claim that\n * `assertMinted` refuses on. Absent = unknown, which is the truth here.\n */\nexport function rolloutRewardFields(record: Scored): RolloutRewardFields {\n  const { reward, gated } = trainingReward(record)\n  const screened = record.outcome.realness !== undefined\n  return { reward, realness_gated: gated, ...(screened ? { realness_screened: true } : {}) }\n}\n\n/**\n * The same fields for a producer that has a graded score but NO `RunRecord`\n * behind it — a supervision journal, a harness session store, any second\n * intake. There is no `outcome.realness` to consult, so no authenticity screen\n * has run on this score.\n *\n * `realness_gated: false` alone was the WRONG way to say that. `false` is the\n * gate's VERDICT, and a verdict is a claim that somebody looked; on this path\n * nobody did, so the row asserted \"screened and clean\" about a score that had\n * never been screened at all — indistinguishable on the wire from a genuinely\n * clean one, which is exactly what an adversary needs. The two claims are now\n * carried by two fields:\n *\n *   - `realness_screened: false` — no screen ran. Explicit, not inferable from\n *     absence, and impossible to read as \"clean\" because it is not the gate's\n *     verdict field.\n *   - `realness_gated: false` — no gate fired, which is trivially true when no\n *     gate ran, and is what a consumer filtering on the flag expects to see.\n *\n * `assertMinted` REFUSES a line carrying `realness_screened: false` with a\n * reward above zero, so these rows can be written, read, reported and analysed,\n * but a positive one cannot become training data until something screens it.\n *\n * A separate function from `rolloutRewardFields` on purpose: picking this one\n * is a producer declaring that its rewards were never screened.\n */\nexport function unscreenedRewardFields(score: number | null): RolloutRewardFields {\n  return { reward: score, realness_gated: false, realness_screened: false }\n}\n"],"mappings":";;AAiCA,SAAgB,gBAAgB,QAAyB;CACvD,OAAO,OAAO,QAAQ,UAAU,UAAU;AAC5C;;;;;;;;;;;;;;AAeA,SAAgB,mBAAmB,QAAgB,OAA4C;CAC7F,OAAO,UAAU,WAAW,OAAO,QAAQ,cAAc,OAAO,QAAQ;AAC1E;;;;;;;;;;;;;;;;AAiBA,SAAgB,cACd,QACA,SAA0B,WACN;CACpB,MAAM,QAAyB,WAAW,WAAW,YAAY;CACjE,OAAO,mBAAmB,QAAQ,MAAM,KAAK,mBAAmB,QAAQ,KAAK;AAC/E;;;;;;;AAWA,SAAgB,YAAY,QAAgB,SAA0B,WAAwB;CAC5F,MAAM,QAAyB,WAAW,WAAW,YAAY;CACjE,IAAI,mBAAmB,QAAQ,MAAM,MAAM,KAAA,GAAW,OAAO;CAC7D,IAAI,mBAAmB,QAAQ,KAAK,MAAM,KAAA,GAAW,OAAO;CAC5D,OAAO;AACT;;;;;;;;;AAUA,SAAgB,cACd,QACA,SAA0B,WACN;CACpB,IAAI,gBAAgB,MAAM,GAAG,OAAO;CACpC,OAAO,cAAc,QAAQ,MAAM;AACrC;;;;;;;;;;;;AAaA,SAAgB,eAAe,QAA2D;CACxF,OAAO;EAAE,QAAQ,cAAc,MAAM,KAAK;EAAM,OAAO,gBAAgB,MAAM;CAAE;AACjF;;;;;;;;;;;;AA2DA,SAAgB,oBAAoB,QAAqC;CACvE,MAAM,EAAE,QAAQ,UAAU,eAAe,MAAM;CAE/C,OAAO;EAAE;EAAQ,gBAAgB;EAAO,GADvB,OAAO,QAAQ,aAAa,KAAA,IACU,EAAE,mBAAmB,KAAK,IAAI,CAAC;CAAG;AAC3F;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAgB,uBAAuB,OAA2C;CAChF,OAAO;EAAE,QAAQ;EAAO,gBAAgB;EAAO,mBAAmB;CAAM;AAC1E"}
+{"version":3,"file":"reward-nw2xZGZG.js","names":[],"sources":["../src/rollout/reward.ts"],"sourcesContent":["/**\n * The two named score derivations every consumer must choose between.\n *\n * The anti-Goodhart gate (`outcome.realness.gated`) only holds if it is\n * impossible to read a run's score WITHOUT deciding whether the gate applies.\n * A bare `outcome.holdoutScore ?? outcome.searchScore` makes that decision\n * invisible — and silently answers \"no gate\", which is the wrong default on\n * every path that produces training data. So the expression lives here, once,\n * behind two names that force the caller to state the intent:\n *\n *   - `trainingScore` / `trainingReward` — GATED. Anything that becomes\n *     training data, or a reward a trainer consumes, uses these.\n *   - `observedScore` — RAW. Analysis, reporting, and reward-hack DETECTION\n *     need the ungated number; that is how a gamed run is visible at all.\n *\n * A leaf module on purpose: it imports only the `RunRecord` type, so gate and\n * reporting code can depend on it without pulling in the trace store that\n * `mint.ts` needs.\n */\n\nimport type { RunRecord } from '../run-record'\n\n/**\n * Which split's score wins when a record carries both. `'holdout'` is the\n * canonical \"real signal\" default; `'search'` exists because some callers\n * deliberately score on the search split when both are present.\n */\nexport type ScorePreference = 'holdout' | 'search'\n\n/** Only the outcome is read, so every accessor here accepts anything carrying one. */\ntype Scored = Pick<RunRecord, 'outcome'>\n\n/** True when the authenticity gate flagged the run as gamed (`realness.gated`). */\nexport function isRealnessGated(record: Scored): boolean {\n  return record.outcome.realness?.gated === true\n}\n\n/**\n * The RAW score recorded on ONE split, with no cross-split fallback and no\n * anti-Goodhart gate.\n *\n * The narrowest of the three raw readers, and the one every split-scoped\n * consumer wants: a per-split report, a promotion gate, or a paired comparison\n * asks \"what did this run score on the split I am summarising\", and answering\n * it with the other split's number silently mixes populations. `undefined` =\n * that split was never scored.\n *\n * Same warning as `observedScore`: this INCLUDES runs flagged as gamed. Never\n * feed it into training data.\n */\nexport function observedSplitScore(record: Scored, split: ScorePreference): number | undefined {\n  return split === 'search' ? record.outcome.searchScore : record.outcome.holdoutScore\n}\n\n/**\n * The RAW split score the run carries, with NO anti-Goodhart gate applied.\n *\n * INCLUDES RUNS FLAGGED AS GAMED (`outcome.realness.gated === true`); NEVER\n * feed this into training data — a fine-tune that sees it learns from gamed\n * successes. It is exported anyway because analysis, reporting, and\n * reward-hacking detection legitimately need the ungated number: forcing a\n * gamed run to 0 collapses the proxy signal toward ground truth and makes a\n * detector report \"clean\" on exactly the population that is being gamed.\n *\n * Returns `undefined` when the record carries neither score — an unscored run\n * is a labeled gap, not a measured zero, and each caller picks its own\n * sentinel (`?? 0`, `?? null`, skip, throw). Non-finite values are returned\n * as-is; callers that care keep their own `Number.isFinite` guard.\n */\nexport function observedScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  return observedSplitScore(record, prefer) ?? observedSplitScore(record, other)\n}\n\n/** Which split actually carried the score, or that none did. */\nexport type ScoreOrigin = 'holdout' | 'search' | 'unscored'\n\n/**\n * Where `observedScore` / `trainingScore` read their number from — the\n * provenance label a rollout line's `reward_source` is built from, and the\n * only supported way to ask \"was this run scored at all\" without respelling\n * the field access.\n */\nexport function scoreOrigin(record: Scored, prefer: ScorePreference = 'holdout'): ScoreOrigin {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  if (observedSplitScore(record, prefer) !== undefined) return prefer\n  if (observedSplitScore(record, other) !== undefined) return other\n  return 'unscored'\n}\n\n/**\n * The GATED score — the only derivation allowed to reach training data.\n *\n * A realness-gated run scores 0 no matter what it claims, so a fine-tune\n * cannot learn from a gamed success. An unscored run stays `undefined` (a\n * labeled gap), keeping \"we never measured this\" distinct from \"we measured\n * zero\"; callers that need a number apply their own sentinel.\n */\nexport function trainingScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  if (isRealnessGated(record)) return 0\n  return observedScore(record, prefer)\n}\n\n/**\n * `{reward, gated}` as written onto a minted `RolloutLine` — `trainingScore`\n * plus the flag itself, so the gate travels into the exported row and a\n * downstream filter can drop or down-weight the line.\n *\n * An unscored record yields `reward: null`, matching the schema's \"no verdict\n * exists — a labeled gap, never 0\" rule. It previously collapsed to 0, which\n * made a run nobody graded indistinguishable from one graded as a total\n * failure, and taught any trainer reading the row that the trajectory was bad.\n * A gated run still yields 0, because that IS a verdict: the gate decided.\n */\nexport function trainingReward(record: Scored): { reward: number | null; gated: boolean } {\n  return { reward: trainingScore(record) ?? null, gated: isRealnessGated(record) }\n}\n\n/**\n * The two fields a rollout line's `outcome` has to state TOGETHER — the reward\n * and whether the authenticity gate fired on it.\n *\n * Stating them together is the point. `realness_gated` used to be optional on\n * the wire, so a producer that wrote `reward` and simply never thought about\n * the flag emitted a row on which `isLineRealnessGated` was structurally false\n * — which is how a second minting door existed for every supervisor and worker\n * row without anyone noticing it had never consulted the gate. The wire schema\n * now requires the flag, and this pair is how a producer states it.\n */\nexport interface RolloutRewardFields {\n  reward: number | null\n  realness_gated: boolean\n  /**\n   * Whether a screen RAN, as distinct from its verdict. Omitted when the\n   * producer cannot tell — see `unscreenedRewardFields` and the field's own doc\n   * on `RolloutOutcome`.\n   */\n  realness_screened?: boolean\n}\n\n/**\n * `trainingReward` in the line's own field names — what `mintRolloutRows`\n * writes onto every row it produces.\n *\n * `realness_screened: true` is written only when the record actually carries an\n * `outcome.realness` verdict, i.e. a screen genuinely ran and reported. A\n * record without one gets the field OMITTED rather than `false`: the record\n * cannot distinguish \"the screen ran elsewhere and this pipeline did not record\n * it\" from \"no screen exists\", and `false` is a load-bearing claim that\n * `assertMinted` refuses on. Absent = unknown, which is the truth here.\n */\nexport function rolloutRewardFields(record: Scored): RolloutRewardFields {\n  const { reward, gated } = trainingReward(record)\n  const screened = record.outcome.realness !== undefined\n  return { reward, realness_gated: gated, ...(screened ? { realness_screened: true } : {}) }\n}\n\n/**\n * The same fields for a producer that has a graded score but NO `RunRecord`\n * behind it — a supervision journal, a harness session store, any second\n * intake. There is no `outcome.realness` to consult, so no authenticity screen\n * has run on this score.\n *\n * `realness_gated: false` alone was the WRONG way to say that. `false` is the\n * gate's VERDICT, and a verdict is a claim that somebody looked; on this path\n * nobody did, so the row asserted \"screened and clean\" about a score that had\n * never been screened at all — indistinguishable on the wire from a genuinely\n * clean one, which is exactly what an adversary needs. The two claims are now\n * carried by two fields:\n *\n *   - `realness_screened: false` — no screen ran. Explicit, not inferable from\n *     absence, and impossible to read as \"clean\" because it is not the gate's\n *     verdict field.\n *   - `realness_gated: false` — no gate fired, which is trivially true when no\n *     gate ran, and is what a consumer filtering on the flag expects to see.\n *\n * `assertMinted` REFUSES a line carrying `realness_screened: false` with a\n * reward above zero, so these rows can be written, read, reported and analysed,\n * but a positive one cannot become training data until something screens it.\n *\n * A separate function from `rolloutRewardFields` on purpose: picking this one\n * is a producer declaring that its rewards were never screened.\n */\nexport function unscreenedRewardFields(score: number | null): RolloutRewardFields {\n  return { reward: score, realness_gated: false, realness_screened: false }\n}\n"],"mappings":";;AAiCA,SAAgB,gBAAgB,QAAyB;CACvD,OAAO,OAAO,QAAQ,UAAU,UAAU;AAC5C;;;;;;;;;;;;;;AAeA,SAAgB,mBAAmB,QAAgB,OAA4C;CAC7F,OAAO,UAAU,WAAW,OAAO,QAAQ,cAAc,OAAO,QAAQ;AAC1E;;;;;;;;;;;;;;;;AAiBA,SAAgB,cACd,QACA,SAA0B,WACN;CACpB,MAAM,QAAyB,WAAW,WAAW,YAAY;CACjE,OAAO,mBAAmB,QAAQ,MAAM,KAAK,mBAAmB,QAAQ,KAAK;AAC/E;;;;;;;AAWA,SAAgB,YAAY,QAAgB,SAA0B,WAAwB;CAC5F,MAAM,QAAyB,WAAW,WAAW,YAAY;CACjE,IAAI,mBAAmB,QAAQ,MAAM,MAAM,KAAA,GAAW,OAAO;CAC7D,IAAI,mBAAmB,QAAQ,KAAK,MAAM,KAAA,GAAW,OAAO;CAC5D,OAAO;AACT;;;;;;;;;AAUA,SAAgB,cACd,QACA,SAA0B,WACN;CACpB,IAAI,gBAAgB,MAAM,GAAG,OAAO;CACpC,OAAO,cAAc,QAAQ,MAAM;AACrC;;;;;;;;;;;;AAaA,SAAgB,eAAe,QAA2D;CACxF,OAAO;EAAE,QAAQ,cAAc,MAAM,KAAK;EAAM,OAAO,gBAAgB,MAAM;CAAE;AACjF;;;;;;;;;;;;AAmCA,SAAgB,oBAAoB,QAAqC;CACvE,MAAM,EAAE,QAAQ,UAAU,eAAe,MAAM;CAE/C,OAAO;EAAE;EAAQ,gBAAgB;EAAO,GADvB,OAAO,QAAQ,aAAa,KAAA,IACU,EAAE,mBAAmB,KAAK,IAAI,CAAC;CAAG;AAC3F;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAgB,uBAAuB,OAA2C;CAChF,OAAO;EAAE,QAAQ;EAAO,gBAAgB;EAAO,mBAAmB;CAAM;AAC1E"}

package/dist/rl.d.ts

@@ -1,20 +1,20 @@
 import { v as Span } from "./schema-2g7trUwH.js";
-import { a as RunRecord, c as RunSplitTag } from "./run-record-DF-fuu31.js";
+import { a as RunRecord, c as RunSplitTag } from "./run-record-p2AYTrCn.js";
 import { a as PairedPromotionDecision, d as PairedBootstrapResult } from "./paired-promotion-decision-DPKz8ZnF.js";
 import { s as TraceStore } from "./store-C4avFjIi.js";
 import { a as MintedRolloutLine, b as RolloutSplit, t as ChatMessage } from "./schema-C-ziyJD7.js";
-import { t as MintRolloutOptions } from "./mint-C1W8w5wo.js";
-import { a as CampaignResult } from "./types-CFCUZk0B.js";
+import { t as MintRolloutOptions } from "./mint-CDy-brUR.js";
+import { a as CampaignResult } from "./types-Dv9sYwBB.js";
 import { s as VerificationReport } from "./multi-layer-verifier-BUaQ4C17.js";
-import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-Cyk-QHk9.js";
+import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-C6Kp54Av.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-CNt4iZ67.js";
-import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-dRvmsbTk.js";
+import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-C26UvI3D.js";
 import { t as InterimReleaseConfidence } from "./sequential-BhsrMupG.js";
 import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
-import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-CZFZjj-u.js";
+import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-DuIJ9IaA.js";
 import { o as TrainingExportOptions } from "./exporters-ChTnR90l.js";
-import { a as RubricPredictiveValidityReport, t as OutcomeMetricSpec } from "./rubric-predictive-validity-DIUwJJ9Y.js";
-import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-Bt4JcxMc.js";
+import { a as RubricPredictiveValidityReport, t as OutcomeMetricSpec } from "./rubric-predictive-validity-BiNNEjpG.js";
+import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-BReh3mvm.js";
 //#region src/rl/active-curriculum.d.ts
 /**
  * Adaptive curriculum: active scenario selection by variance.
@@ -310,26 +310,7 @@ export declare function checkerReadSignal(events: readonly TrajectoryEvent[], op
   checkers: readonly string[];
   firstClaimAt: string | number | null;
 }): IntegritySignal;
-/** What the referee and the System One model read about a claim: the source and the claim's parameters, never its prose. */
-export declare function claimBrief(claim: IntegrityClaim, scope: SourceScope): {
-  statement: string;
-  source: string;
-  excerpts: {
-    cite: string;
-    text: string;
-  }[];
-  claim: {
-    field: string | null;
-    instance: string | null;
-    asserts: string | null;
-    novelty: "known" | "new" | null;
-    parameters: {
-      [k: string]: unknown;
-    };
-    parameterMeaning: Readonly<Record<string, string>>;
-  };
-};
-export declare const REFEREE_VERDICTS: readonly ['holds', 'excluded', 'trivial', 'known', 'unclear'];
+declare const REFEREE_VERDICTS: readonly ['holds', 'excluded', 'trivial', 'known', 'unclear'];
 export type RefereeVerdict = (typeof REFEREE_VERDICTS)[number];
 /** The referee's messages for one claim, for any chat model. */
 export declare function sourceScopeRefereeMessages(claim: IntegrityClaim, scope: SourceScope): ({
@@ -361,7 +342,7 @@ export declare function refereeSignal(claim: IntegrityClaim, answer: SourceScope
  * System One typed questions over the same brief the referee reads. Each names a way the claim can be true of
  * the checker and not of the statement; `true` supports that hypothesis, except `inScope`.
  */
-export declare const claimIntegritySystemOneQuestions: {
+declare const claimIntegritySystemOneQuestions: {
   readonly excludedBySource: {
     readonly type: 'noul';
     readonly instructions: "Using only the source excerpts, does the source exclude the claim's case from its statement, by a hypothesis, a definition, or an assumption it states (for example a proof that calls a case easy and then assumes it away)?";

package/dist/rl.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"rl.d.ts","names":[],"sources":["../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/claim-integrity.ts","../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/rollout-input.ts","../src/rl/preferences.ts","../src/rl/process-reward.ts","../src/rl/exporters.ts","../src/rl/dataset.ts","../src/rl/corpus.ts","../src/rl/off-policy.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/rl-campaign.ts","../src/rl/run-record-adapters.ts","../src/rl/sim-fidelity.ts","../src/rl/tournament.ts","../src/rl/verified-findings-dataset.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgBiB;EACf;EACA;;EAEA;;iBAGe;EACf;EACA;;EAEA;;EAEA;;iBAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;wBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;;iBC9Cc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;iBAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;iBAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;iBAG7D;EACf,QAAQ;;;;;EAKR;;;;;EAKA;;wBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;iBAmEM;EACf,MAAM;IACJ;;IAEA,OAAO;;;EAGT,WAAW;;EAEX,cAAc;EACd,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;wBAUc,wBACd,GAAG,iBACH,GAAG,iBACH;EACE;EACA;EACA;;EAEA;IAED;;wBA0Ea,WAAW,OAAO,iBAAiB;;;YC5LvC;YACA;;iBAGK;WACN;WACA,IAAI;WACJ,OAAO,sBAAsB;;iBAGvB;;WAEN;WACA,eAAe;;WAEf;;WAEA;;WAEA;;iBAGM;WACN;WACA;;;;;;;iBAQM;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA,SAAS,SAAS;;;;;WAKlB,aAAa,SAAS;;WAEtB,mBAAmB;;WAEnB,oBAAoB;;WAEpB,kBAAkB;;WAElB,mBAAmB;;;iBAIb;;WAEN;;WAEA;WACA;WACA;WACA;;WAEA,YAAY,SAAS;;WAErB;;WAEA;;WAEA;;YAGC;YACA;iBAYK;WACN;WACA;WACA;;iBAGM;WACN,IAAI;WACJ,MAAM;;WAEN;WACA;WACA,mBAAmB;;WAEnB;;;wBA2BK,UAAU,MAAM,WAAW,OAAO,KAAK;;;;;;wBAyDvC,aACd,OAAO,gBACP,OAAO,0BACN;iBAsCc;;WAEN;;;;;;;;;wBAUK,cACd,OAAO,gBACP,qBAAqB,kBACrB,OAAO,yBACP,UAAS,iBACR;;;;;;;;;;;qBAkFU;WAEL;WAEF;;WAGE;WAEF;;WAGE;WACK;;WAGL;WACK;;;wBAKG,sBAAsB,kCAAkC;;iBAqCvD;WACN;WACA;;WAEA;;;;;;;wBAQK,kBACd,iBAAiB,mBACjB;EAAW;EAA6B;IACvC;;wBAuCa,WAAW,OAAO,gBAAgB,OAAO;;;;;;;;;;;;;;;;;;qBAsB5C;YACD,yBAAyB;;wBAkBrB,2BAA2B,OAAO,gBAAgB,OAAO;EAa7D;;;EACA;;;iBAIK;WACN,SAAS;WACT;WACA;WACA;;;wBAyBK,wBAAwB,eAAe;;;;;wBA8BvC,cACd,OAAO,gBACP,QAAQ;WAAsC;UAC9C;EAAW;EAAgB;IAC1B;;;;;qBA8CU;WACX;aACE;aACA;;WAGF;aACE;aACA;;WAGF;aACE;aACA;;WAEF;aACE;aACA;;;YAKQ,0CAA0C;iBAErC;WACN;;;;;WAKA;;WAEA;WACA;;;wBAIK,8BACd,OAAO,gBACP,OAAO,aACP,SAAS,iCACR,wBAAwB;;wBA+BX,gBACd,QAAQ;WAAmC;UAC3C;EAAW;IACV;;;;;;wBAqCa,6BAA6B,OAAO,iBAAiB;;wBAmBrD,cACd,mBAAmB;WAAoC;UACvD;EAAW;IACV;qBAsCU;iBAEI;;WAEN,sBAAsB;;iBAGhB;WACN;WACA;;WAEA;WACA,kBAAkB;;WAElB;;WAEA;;;wBAIK,qBAAqB,SAAQ;;wBAS7B,sBACd,kBAAkB,mBAClB,SAAQ,uBACP;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBC5zBc;;EAEf;;EAEA;;EAEA,OAAO;;iBAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;iBAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;iBAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;wBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;iBAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;iBAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;wBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;iBAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;iBAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;wBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;iBAkChB;EACf;EACA;EACA;EACA;;wBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;YC7LhD;iBAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;iBAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;iBAGJ;;EAEf;;EAEA;;;;;EAKA;;iBAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;;EAGF;IAAc;IAAW;;;EAEzB;EACA;EACA;EACA;;EAEA;;EAEA;;wBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;wBAyGK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;wBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;wBAgBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UClMP;;;;;;;;;;;EAWf,OAAO;;;;YCKG;iBAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;iBAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;iBAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;wBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;wBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;wBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;iBChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;iBAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;iBAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;wBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;iBAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;wBAGc,yBAAyB,aAAa,eAAe;iBA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;wBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;iBCrJc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;iBAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;YAIG,iBAAiB;;;;;;;;;;;;;;;;wBAwBP,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;wBAwCK,WAAW,MAAM;;YAOrB,+BAA+B;iBAE1B,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;iBAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;wBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;wBAyDK,YAAY,MAAM;iBAMjB,mBAAmB;;;;;EAKlC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;iBAGF;;EAEf,UAAU;EACV,OAAO;;;;;;;;;;;;;;;;;;;wBAoBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;wBAwDK,WAAW,MAAM;iBAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;iBAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;iBAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;wBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;wBA8FK,WAAW,MAAM;iBAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;;;;;;wBAoBc,mBAAmB,aAAa,cAAc,SAAS;;;YCpf3D;cACN;YACM,wBAAwB;wBAIpB,uBAAuB,iBAAiB;;;iBA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;iBAIa;EACf;EACA;EACA;EACA;EACA;EACA;;iBAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;iBAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;iBAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;;wBA0Da,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;wBA2DK,oBAAoB,GAAG;;;;;YC1O3B,eAAe;EAAc;EAAiB;;iBAEzC;EACf;;;EAGA;EACA;;;;;;;;wBASc,eAAe,SAAS,gBAAgB,qBAAqB;;wBAqB7D,WAAW,qBAAqB;iBAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;wBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCnEM;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;iBAGe;;EAEf;;EAEA;;iBAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;iBAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;wBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;wBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;wBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;wBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI;;;;iBCvQ1C;EACf,UAAU;;EAEV,eAAe;;EAEf;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;qBAOnD,wCAAwC;mBAClC;UACT;EAER,YAAY,MAAM;EAaZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EA+DhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EAiD9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAUlB,iBAAiB;;;;iBCxMF,qBAAqB,WAAW,oBAAoB;;EAEnE,UAAU;;EAEV,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf,0BAA0B;;EAE1B;IACE;IACA;IACA;MAAS;MAAa;;;;;;;;;;;;;IAYtB;;;EAGF;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;;;;;;;iBAYO;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;iBAGe;EACf,UAAU;;EAEV,cAAc;;EAEd,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;;;;EAKb,mBAAmB;;;;EAInB,eAAe;;EAEf,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;wBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;iBC5I9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;;wBAYc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;wBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;YCpCS,oBAAoB,QAAQ,cAAc;;;;;qBAMzC;;;;;;;;;;;;;;;;;;;;;;qBAuCA,yBAAyB;;;;;;;;;wBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;wBAqC3C,cAAc,kBAAkB;;;wBAsBhC,YAAY,eAAe;iBAU1B;;EAEf;;;EAGA;;EAEA;;iBAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;YAGU;iBAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;iBAGM;;EAEf,WAAW;;EAEX;;;;;;;;wBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;iBAuHc;;;EAGf;;;EAGA;;iBAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;wBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;iBAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;iBAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;wBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;iBA+Gc;;EAEf;;EAEA;;wBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;iBA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;wBAGc,0BACd,OAAO,iCACN;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBC/MU;;iBAKI;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA;EACA,KAAK;EACL;;iBAGe;EACf;EACA;EACA;EACA;EACA;EACA;;iBAGe;EACf;EACA,OAAO;;;iBAIQ;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,kBAAkB;IAAQ;IAAkB;;;;iBAI7B;EACf;EACA;EACA;;iBAGe;EACf;;;EAGA;;EAEA;EACA;;;iBAIe;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB;EACA;IAAU;IAAyB;;;iBAKpB;EACf;EACA;EACA;;EAEA;EACA;;YAGU;iBAEK;EACf,SAAS;EACT;EACA;EACA;EACA;;iBAGe;EACf,eAAe;;EAEf;EACA;EACA;EACA;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;;IAEE;;IAEA;;IAEA;;IAEA;IACA;;;EAGF;IACE;MAAU;MAAe;;IACzB,OAAO;;EAET;IACE;;IAEA;IACA;IACA;IACA;IACA;IACA;IACA;;;IAGA;IACA,wBAAwB;IACxB;IACA;IACA;IACA;;EAEF,KAAK;EACL;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;;;iBAIa;EACf;EACA;;;;EAIA;EACA,KAAK,OAAO;EACZ,UAAU;IAAiB;IAAc;IAAoB;;;iBAO9C;EACf,WAAW;EACX,OAAO;EACP,OAAO;EACP;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS;EACT;;;;;;wBAiEc,wBAAwB,MAAM,8BAA8B;wBA8I5D,0BAA0B,MAAM,uBAAuB;wBAwBvD,wBAAwB,MAAM;iBAM7B;EACf;;EAEA;;iBAGe;EACf;;EAEA;;EAEA,SAAS,eAAe;;EAExB;EACA;;iBAGe;EACf,MAAM;EACN,SAAS;EACT;IACE;IACA;IACA;IACA;IACA,SAAS;MAAiB;MAAoB;MAAsB;;;;;;;;;wBAuExD,4BACd,QAAQ,yBACP"}
+{"version":3,"file":"rl.d.ts","names":[],"sources":["../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/claim-integrity.ts","../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/rollout-input.ts","../src/rl/preferences.ts","../src/rl/process-reward.ts","../src/rl/exporters.ts","../src/rl/dataset.ts","../src/rl/corpus.ts","../src/rl/off-policy.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/rl-campaign.ts","../src/rl/run-record-adapters.ts","../src/rl/sim-fidelity.ts","../src/rl/tournament.ts","../src/rl/verified-findings-dataset.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgBiB;EACf;EACA;;EAEA;;iBAGe;EACf;EACA;;EAEA;;EAEA;;iBAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;wBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;;iBC9Cc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;iBAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;iBAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;iBAG7D;EACf,QAAQ;;;;;EAKR;;;;;EAKA;;wBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;iBAmEM;EACf,MAAM;IACJ;;IAEA,OAAO;;;EAGT,WAAW;;EAEX,cAAc;EACd,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;wBAUc,wBACd,GAAG,iBACH,GAAG,iBACH;EACE;EACA;EACA;;EAEA;IAED;;wBA0Ea,WAAW,OAAO,iBAAiB;;;YC5LvC;YACA;;iBAGK;WACN;WACA,IAAI;WACJ,OAAO,sBAAsB;;iBAGvB;;WAEN;WACA,eAAe;;WAEf;;WAEA;;WAEA;;iBAGM;WACN;WACA;;;;;;;iBAQM;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA,SAAS,SAAS;;;;;WAKlB,aAAa,SAAS;;WAEtB,mBAAmB;;WAEnB,oBAAoB;;WAEpB,kBAAkB;;WAElB,mBAAmB;;;iBAIb;;WAEN;;WAEA;WACA;WACA;WACA;;WAEA,YAAY,SAAS;;WAErB;;WAEA;;WAEA;;YAGC;YACA;iBAYK;WACN;WACA;WACA;;iBAGM;WACN,IAAI;WACJ,MAAM;;WAEN;WACA;WACA,mBAAmB;;WAEnB;;;wBA2BK,UAAU,MAAM,WAAW,OAAO,KAAK;;;;;;wBAyDvC,aACd,OAAO,gBACP,OAAO,0BACN;iBAsCc;;WAEN;;;;;;;;;wBAUK,cACd,OAAO,gBACP,qBAAqB,kBACrB,OAAO,yBACP,UAAS,iBACR;;;;;;;;;;;qBAkFU;WAEL;WAEF;;WAGE;WAEF;;WAGE;WACK;;WAGL;WACK;;;wBAKG,sBAAsB,kCAAkC;;iBAqCvD;WACN;WACA;;WAEA;;;;;;;wBAQK,kBACd,iBAAiB,mBACjB;EAAW;EAA6B;IACvC;cA6DG;YACM,yBAAyB;;wBAkBrB,2BAA2B,OAAO,gBAAgB,OAAO;EAa7D;;;EACA;;;iBAIK;WACN,SAAS;WACT;WACA;WACA;;;wBAyBK,wBAAwB,eAAe;;;;;wBA8BvC,cACd,OAAO,gBACP,QAAQ;WAAsC;UAC9C;EAAW;EAAgB;IAC1B;;;;;cA8CG;WACJ;aACE;aACA;;WAGF;aACE;aACA;;WAGF;aACE;aACA;;WAEF;aACE;aACA;;;YAKQ,0CAA0C;iBAErC;WACN;;;;;WAKA;;WAEA;WACA;;;wBAIK,8BACd,OAAO,gBACP,OAAO,aACP,SAAS,iCACR,wBAAwB;;wBA+BX,gBACd,QAAQ;WAAmC;UAC3C;EAAW;IACV;;;;;;wBAqCa,6BAA6B,OAAO,iBAAiB;;wBAmBrD,cACd,mBAAmB;WAAoC;UACvD;EAAW;IACV;qBAsCU;iBAEI;;WAEN,sBAAsB;;iBAGhB;WACN;WACA;;WAEA;WACA,kBAAkB;;WAElB;;WAEA;;;wBAIK,qBAAqB,SAAQ;;wBAS7B,sBACd,kBAAkB,mBAClB,SAAQ,uBACP;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBC5zBc;;EAEf;;EAEA;;EAEA,OAAO;;iBAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;iBAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;iBAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;wBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;iBAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;iBAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;wBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;iBAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;iBAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;wBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;iBAkChB;EACf;EACA;EACA;EACA;;wBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;YC7LhD;iBAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;iBAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;iBAGJ;;EAEf;;EAEA;;;;;EAKA;;iBAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;;EAGF;IAAc;IAAW;;;EAEzB;EACA;EACA;EACA;;EAEA;;EAEA;;wBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;wBAyGK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;wBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;wBAgBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UClMP;;;;;;;;;;;EAWf,OAAO;;;;YCKG;iBAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;iBAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;iBAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;wBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;wBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;wBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;iBChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;iBAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;iBAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;wBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;iBAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;wBAGc,yBAAyB,aAAa,eAAe;iBA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;wBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;iBCrJc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;iBAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;YAIG,iBAAiB;;;;;;;;;;;;;;;;wBAwBP,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;wBAwCK,WAAW,MAAM;;YAOrB,+BAA+B;iBAE1B,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;iBAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;wBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;wBAyDK,YAAY,MAAM;iBAMjB,mBAAmB;;;;;EAKlC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;iBAGF;;EAEf,UAAU;EACV,OAAO;;;;;;;;;;;;;;;;;;;wBAoBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;wBAwDK,WAAW,MAAM;iBAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;iBAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;iBAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;wBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;wBA8FK,WAAW,MAAM;iBAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;;;;;;wBAoBc,mBAAmB,aAAa,cAAc,SAAS;;;YCpf3D;cACN;YACM,wBAAwB;wBAIpB,uBAAuB,iBAAiB;;;iBA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;iBAIa;EACf;EACA;EACA;EACA;EACA;EACA;;iBAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;iBAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;iBAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;;wBA0Da,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;wBA2DK,oBAAoB,GAAG;;;;;YC1O3B,eAAe;EAAc;EAAiB;;iBAEzC;EACf;;;EAGA;EACA;;;;;;;;wBASc,eAAe,SAAS,gBAAgB,qBAAqB;;wBAqB7D,WAAW,qBAAqB;iBAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;wBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCnEM;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;iBAGe;;EAEf;;EAEA;;iBAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;iBAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;wBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;wBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;wBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;wBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI;;;;iBCvQ1C;EACf,UAAU;;EAEV,eAAe;;EAEf;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;qBAOnD,wCAAwC;mBAClC;UACT;EAER,YAAY,MAAM;EAaZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EA+DhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EAiD9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAUlB,iBAAiB;;;;iBCxMF,qBAAqB,WAAW,oBAAoB;;EAEnE,UAAU;;EAEV,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf,0BAA0B;;EAE1B;IACE;IACA;IACA;MAAS;MAAa;;;;;;;;;;;;;IAYtB;;;EAGF;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;;;;;;;iBAYO;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;iBAGe;EACf,UAAU;;EAEV,cAAc;;EAEd,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;;;;EAKb,mBAAmB;;;;EAInB,eAAe;;EAEf,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;wBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;iBC5I9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;;wBAYc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;wBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;YCpCS,oBAAoB,QAAQ,cAAc;;;;;qBAMzC;;;;;;;;;;;;;;;;;;;;;;qBAuCA,yBAAyB;;;;;;;;;wBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;wBAqC3C,cAAc,kBAAkB;;;wBAsBhC,YAAY,eAAe;iBAU1B;;EAEf;;;EAGA;;EAEA;;iBAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;YAGU;iBAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;iBAGM;;EAEf,WAAW;;EAEX;;;;;;;;wBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;iBAuHc;;;EAGf;;;EAGA;;iBAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;wBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;iBAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;iBAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;wBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;iBA+Gc;;EAEf;;EAEA;;wBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;iBA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;wBAGc,0BACd,OAAO,iCACN;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBC/MU;;iBAKI;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA;EACA,KAAK;EACL;;iBAGe;EACf;EACA;EACA;EACA;EACA;EACA;;iBAGe;EACf;EACA,OAAO;;;iBAIQ;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,kBAAkB;IAAQ;IAAkB;;;;iBAI7B;EACf;EACA;EACA;;iBAGe;EACf;;;EAGA;;EAEA;EACA;;;iBAIe;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB;EACA;IAAU;IAAyB;;;iBAKpB;EACf;EACA;EACA;;EAEA;EACA;;YAGU;iBAEK;EACf,SAAS;EACT;EACA;EACA;EACA;;iBAGe;EACf,eAAe;;EAEf;EACA;EACA;EACA;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;;IAEE;;IAEA;;IAEA;;IAEA;IACA;;;EAGF;IACE;MAAU;MAAe;;IACzB,OAAO;;EAET;IACE;;IAEA;IACA;IACA;IACA;IACA;IACA;IACA;;;IAGA;IACA,wBAAwB;IACxB;IACA;IACA;IACA;;EAEF,KAAK;EACL;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;IACA;;;iBAIa;EACf;EACA;;;;EAIA;EACA,KAAK,OAAO;EACZ,UAAU;IAAiB;IAAc;IAAoB;;;iBAO9C;EACf,WAAW;EACX,OAAO;EACP,OAAO;EACP;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS;EACT;;;;;;wBAiEc,wBAAwB,MAAM,8BAA8B;wBA8I5D,0BAA0B,MAAM,uBAAuB;wBAwBvD,wBAAwB,MAAM;iBAM7B;EACf;;EAEA;;iBAGe;EACf;;EAEA;;EAEA,SAAS,eAAe;;EAExB;EACA;;iBAGe;EACf,MAAM;EACN,SAAS;EACT;IACE;IACA;IACA;IACA;IACA,SAAS;MAAiB;MAAoB;MAAsB;;;;;;;;;wBAuExD,4BACd,QAAQ,yBACP"}

package/dist/rl.js

@@ -3,17 +3,17 @@ import { a as hashCanonical, i as compareCodeUnits } from "./canonical-BsZVgBhW.
 import { a as medianInPlace } from "./internal-CRHwS9iz.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { r as pairedBootstrap } from "./paired-tests-jbwPdSz5.js";
-import { l as wilcoxonSignedRank } from "./paired-arms-CgWT7sHy.js";
-import { t as decidePairedPromotion } from "./paired-promotion-decision-2deJ7sDX.js";
+import { a as wilcoxonSignedRank } from "./paired-arms-DzKn0TfY.js";
+import { t as decidePairedPromotion } from "./paired-promotion-decision-BN6IQTiu.js";
 import { r as observedSplitScore, s as trainingScore } from "./reward-nw2xZGZG.js";
-import { c as runTaskScore } from "./run-record-BRlM99rX.js";
-import { n as campaignCellToRunRecord } from "./run-record-BB290Y9A.js";
+import { o as runTaskScore } from "./run-record-DlTZ_lqe.js";
+import { n as campaignCellToRunRecord } from "./run-record-ES07wfz9.js";
 import { i as filterDeterministicallyRewarded, n as extractVerifiableReward, r as extractVerifiableRewardsFromRecords, t as detectRewardHacking } from "./reward-hacking-BFcQgkux.js";
 import { n as InMemoryTraceStore } from "./store-BhNuLdJw.js";
-import { t as runEvalCampaign } from "./eval-campaign-DU_p8qVt.js";
+import { t as runEvalCampaign } from "./eval-campaign-4byTtqrf.js";
 import { s as assertRewardGate } from "./schema-BuiujZ-g.js";
 import { n as isTrainingLineEligible, s as toSftRows$1, t as isSplitEligible } from "./exporters-X4xwla8A.js";
-import { t as mintRolloutRows } from "./mint-CWG-s2mI.js";
+import { t as mintRolloutRows } from "./mint-Csor1BLJ.js";
 import { t as evaluateInterimReleaseConfidence } from "./sequential-CzK5DarL.js";
 import { n as prepareSystemOneReview } from "./systemone-review-DTClwgmE.js";
 import { n as assertUniqueObservationIds, s as validateOutcomeMetricSpecifications, t as rubricPredictiveValidity } from "./rubric-predictive-validity-BsgkBfT0.js";
@@ -3711,6 +3711,6 @@ function loadVerifiedFindingsDataset(source) {
 	};
 }
 //#endregion
-export { ABSENT_CATEGORY, CLAIM_INTEGRITY_VERSION, FileSystemOutcomeStore, GRADER_REFERENCE_PHRASES, InMemoryOutcomeStore, PredictiveValidityResearcher, REFEREE_VERDICTS, VERIFIED_FINDING_SCHEMA, appendToCorpus, applyEloUpdate, bestOfN, bucketLabel, buildDatasetFromCorpus, buildPairwiseFromCampaign, buildRlDataset, buildVerifiedFindingRow, campaignToRunRecords, checkerReadSignal, claimBrief, claimIntegrityDigest, claimIntegritySystemOneQuestions, claimIntegritySystemOneReview, claimIntegrityTraceQuestions, claimIntegrityVerdict, compareAdaptationCurves, datasheetToMarkdown, defaultBehaviorFeatures, detectRewardHacking, doublyRobust, easyModeCheck, extractPreferences, extractStepRewards, extractVerifiableReward, extractVerifiableRewardsFromRecords, farmingSignal, filterDeterministicallyRewarded, firstPassK, fitBradleyTerry, graderReferenceSignal, injectIrrelevantClause, inverseProbabilityWeighting, jsDivergence, loadVerifiedFindingsDataset, monitorSignal, offPolicyEstimateAll, paretoFrontier, parseSourceScopeReferee, prmTrainingPairs, quantileEdges, readCorpus, refereeSignal, renameVariables, ruleHolds, runAdaptationCurve, runComputeCurve, runContaminationProbe, runEvalCampaign, runRLCampaign, runwiseStepRewardSummary, scopeSignals, selfConsistency, selfNormalizedImportanceWeighting, shuffleOrder, simFidelityReport, sourceScopeRefereeMessages, stepRewardsToJsonl, summarizeVerifiedFindings, systemOneSignal, toAnthropicFormat, toDpoJsonl, toDpoRows, toGrpoJsonl, toGrpoRows, toPrmJsonl, toPrmRows, toSftJsonl, toSftRows, toTRLFormat, validateDatasetFormats, varianceBasedCurriculum, verificationReportToRunRecord, verifiedFindingsToJsonl };
+export { ABSENT_CATEGORY, CLAIM_INTEGRITY_VERSION, FileSystemOutcomeStore, GRADER_REFERENCE_PHRASES, InMemoryOutcomeStore, PredictiveValidityResearcher, VERIFIED_FINDING_SCHEMA, appendToCorpus, applyEloUpdate, bestOfN, bucketLabel, buildDatasetFromCorpus, buildPairwiseFromCampaign, buildRlDataset, buildVerifiedFindingRow, campaignToRunRecords, checkerReadSignal, claimIntegrityDigest, claimIntegritySystemOneReview, claimIntegrityTraceQuestions, claimIntegrityVerdict, compareAdaptationCurves, datasheetToMarkdown, defaultBehaviorFeatures, detectRewardHacking, doublyRobust, easyModeCheck, extractPreferences, extractStepRewards, extractVerifiableReward, extractVerifiableRewardsFromRecords, farmingSignal, filterDeterministicallyRewarded, firstPassK, fitBradleyTerry, graderReferenceSignal, injectIrrelevantClause, inverseProbabilityWeighting, jsDivergence, loadVerifiedFindingsDataset, monitorSignal, offPolicyEstimateAll, paretoFrontier, parseSourceScopeReferee, prmTrainingPairs, quantileEdges, readCorpus, refereeSignal, renameVariables, ruleHolds, runAdaptationCurve, runComputeCurve, runContaminationProbe, runEvalCampaign, runRLCampaign, runwiseStepRewardSummary, scopeSignals, selfConsistency, selfNormalizedImportanceWeighting, shuffleOrder, simFidelityReport, sourceScopeRefereeMessages, stepRewardsToJsonl, summarizeVerifiedFindings, systemOneSignal, toAnthropicFormat, toDpoJsonl, toDpoRows, toGrpoJsonl, toGrpoRows, toPrmJsonl, toPrmRows, toSftJsonl, toSftRows, toTRLFormat, validateDatasetFormats, varianceBasedCurriculum, verificationReportToRunRecord, verifiedFindingsToJsonl };
 
 //# sourceMappingURL=rl.js.map