@tangle-network/agent-eval 0.131.1 → 0.133.0

package/dist/reporting.d.ts

@@ -1,6 +1,6 @@
-import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "./rubric-predictive-validity-lXLmashy.js";
-import { I as pairedBootstrap, Z as wilcoxonSignedRank, d as PairedBootstrapOptions, f as PairedBootstrapResult, y as benjaminiHochberg } from "./statistics-Cmj6nynr.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-CWwB_LiV.js";
-import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-Crg9oFJ0.js";
+import { I as pairedBootstrap, Z as wilcoxonSignedRank, d as PairedBootstrapOptions, f as PairedBootstrapResult, y as benjaminiHochberg } from "./statistics-DbvkkDPa.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-DnUcjVpV.js";
+import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-DfmKSIEE.js";
 import { a as PairedEvalueStep, c as pairedEvalueSequence, i as PairedEvalueSequence, n as InterimReleaseConfidenceInput, o as SequentialDecision, r as PairedEvalueOptions, s as evaluateInterimReleaseConfidence, t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
+import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "./rubric-predictive-validity-C1dCLcvb.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 RenderReleaseReportOptions, 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, renderReleaseReport, researchReport, rubricPredictiveValidity, summaryTable, wilcoxonSignedRank };

package/dist/{researcher-Q5rpPqZY.d.ts → researcher-DMimgHtN.d.ts}

@@ -1,11 +1,11 @@
 import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-CCm3l2v2.js";
-import { a as RunRecord, c as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-BJnYdTxO.js";
+import { a as RunRecord, c as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-DcObtIGh.js";
 import { l as RawProviderSink } from "./raw-provider-sink-BU29Sh8h.js";
 import { s as TraceStore } from "./store-CT9YIIve.js";
 import { i as TraceEmitter, t as RunCompleteHook } from "./emitter-DGQGoLyj.js";
 import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-rmVhXWA7.js";
-import { o as LlmClientOptions, u as LlmRouteRequirements } from "./llm-client-B_nIBlYo.js";
-import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-CWwB_LiV.js";
+import { o as LlmClientOptions, u as LlmRouteRequirements } from "./llm-client-BiK4HW0u.js";
+import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-DnUcjVpV.js";
 //#region src/eval-campaign.d.ts
 interface CampaignVariant<V> {
   id: string;
@@ -312,4 +312,4 @@ declare class NoopResearcher implements Researcher {
 }
 //#endregion
 export { EvalCampaignResult as _, FailureMode as a, SteeringChange as c, CampaignRunContext as d, CampaignRunOutcome as f, EvalCampaignOptions as g, CampaignVariant as h, ExperimentResult as i, CampaignFactoryParams as l, CampaignScenario as m, CallbackResearcherOptions as n, NoopResearcher as o, CampaignRunner as p, ExperimentPlan as r, Researcher as s, CallbackResearcher as t, CampaignIntegrityPolicy as u, FailedRun as v, runEvalCampaign as y };
-//# sourceMappingURL=researcher-Q5rpPqZY.d.ts.map
+//# sourceMappingURL=researcher-DMimgHtN.d.ts.map

package/dist/{researcher-Q5rpPqZY.d.ts.map → researcher-DMimgHtN.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"researcher-Q5rpPqZY.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;;UAyEiB,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;;;;;;EAMT,SAAS;;UAGD;;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;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,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;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;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;;UAG/B;EACf,iBAAiB;EACjB,eAAe;EACf,aAAa;EACb,gBAAgB;;;;;;cAOL,8BAA8B;mBACZ;EAA7B,YAA6B,WAAW;EAExC,gBAAgB,MAAM,cAAc,QAAQ;EAI5C,cAAc,UAAU,gBAAgB,QAAQ;EAIhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAI1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;;;;cAYnC,0BAA0B;mBACpB;EAEjB,YAAY;EAIN,gBAAgB,OAAO,cAAc,QAAQ;EAI7C,cAAc,WAAW,gBAAgB,QAAQ;EAIjD,YACJ,UAAU,kBACV,WAAW,iBACV,QAAQ;EAIL,eAAe,OAAO,iBAAiB,QAAQ"}
+{"version":3,"file":"researcher-DMimgHtN.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;;UAyEiB,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;;;;;;EAMT,SAAS;;UAGD;;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;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,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;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;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;;UAG/B;EACf,iBAAiB;EACjB,eAAe;EACf,aAAa;EACb,gBAAgB;;;;;;cAOL,8BAA8B;mBACZ;EAA7B,YAA6B,WAAW;EAExC,gBAAgB,MAAM,cAAc,QAAQ;EAI5C,cAAc,UAAU,gBAAgB,QAAQ;EAIhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAI1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;;;;cAYnC,0BAA0B;mBACpB;EAEjB,YAAY;EAIN,gBAAgB,OAAO,cAAc,QAAQ;EAI7C,cAAc,WAAW,gBAAgB,QAAQ;EAIjD,YACJ,UAAU,kBACV,WAAW,iBACV,QAAQ;EAIL,eAAe,OAAO,iBAAiB,QAAQ"}

package/dist/{reward-hacking-CW-3HN0n.d.ts → reward-hacking-D-QqXvg-.d.ts}

@@ -1,5 +1,5 @@
 import { s as VerificationReport } from "./multi-layer-verifier-BHY1gWAc.js";
-import { a as RunRecord } from "./run-record-BJnYdTxO.js";
+import { a as RunRecord } from "./run-record-DcObtIGh.js";
 //#region src/rl/verifiable-reward.d.ts
 type VerifiableRewardSource = 'compile' | 'test' | 'schema' | 'sandbox' | 'judge' | 'composite';
 interface VerifiableReward {
@@ -221,4 +221,4 @@ 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-CW-3HN0n.d.ts.map
+//# sourceMappingURL=reward-hacking-D-QqXvg-.d.ts.map

package/dist/{reward-hacking-CW-3HN0n.d.ts.map → reward-hacking-D-QqXvg-.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"reward-hacking-CW-3HN0n.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;KAiCY;UAQK;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA6Cc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KChWvB;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-D-QqXvg-.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;KAiCY;UAQK;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA6Cc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KChWvB;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/rl.d.ts

@@ -1,16 +1,15 @@
 import { s as VerificationReport } from "./multi-layer-verifier-BHY1gWAc.js";
 import { _ as Span } from "./schema-BtVldJ3T.js";
-import { a as RunRecord, s as RunSplitTag } from "./run-record-BJnYdTxO.js";
+import { a as RunRecord, s as RunSplitTag } from "./run-record-DcObtIGh.js";
 import { s as TraceStore } from "./store-CT9YIIve.js";
+import { a as CampaignResult } from "./types-BokuXvOG.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-D-QqXvg-.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-BYHIuO0e.js";
-import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-lXLmashy.js";
-import { a as doublyRobust, c as selfNormalizedImportanceWeighting, i as OffPolicyTrajectory, n as OffPolicyEstimate, o as inverseProbabilityWeighting, r as OffPolicyOptions, s as offPolicyEstimateAll, t as OffPolicyContributionCounts } from "./off-policy-mskQw8Mb.js";
-import { a as CampaignResult } from "./types-CsD5nTfV.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-CW-3HN0n.js";
 import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
 import { b as RolloutSplit, o as MintedRolloutLine } from "./schema-Cef2cFmb.js";
-import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-Q5rpPqZY.js";
+import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-DMimgHtN.js";
 import { t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
+import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-C1dCLcvb.js";
 //#region src/rl/active-curriculum.d.ts
 interface CellObservation {
   variantId: string;
@@ -1035,6 +1034,157 @@ interface HarvestOptions {
  */
 declare function buildDatasetFromCorpus(corpusPath: string, config: RlDatasetConfig, opts?: HarvestOptions): Promise<RlDatasetBundle>;
 //#endregion
+//#region src/rl/off-policy.d.ts
+/**
+ * Off-policy evaluation primitives.
+ *
+ * Standard inverse-probability-weighted (IPS), self-normalized
+ * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the
+ * value of a *target* policy given trajectories collected under a
+ * *behavior* policy. This is the canonical RL eval task: "we have last
+ * week's runs, we changed the policy — how would the new one do without
+ * re-running?"
+ *
+ * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan
+ * & Joachims 2015 for SNIPS) but the *application* to LLM-agent
+ * evaluation needs care:
+ *
+ *   - The "policy" is the (prompt, tool config, model snapshot) triple.
+ *     Two policies have the same probability over an action *iff* their
+ *     LLM call would emit the same token with the same probability —
+ *     which is generally unknowable without the model log-probs.
+ *   - For LLM agents, propensity scores must be supplied by the caller
+ *     (logged in the trace, recovered from token log-probs, or estimated
+ *     via a learned propensity model). We do NOT estimate propensity here.
+ *   - Doubly-robust requires two outputs from a Q-function: its prediction
+ *     for the logged action and its expectation under the target policy.
+ *     Consumers compute these with a tabular estimate, regression fit, or
+ *     learned reward model before constructing the trajectories.
+ *
+ * Bias / variance tradeoffs:
+ *   - IPS: unbiased; high variance for small overlap, infinite variance
+ *     when target has support outside behavior.
+ *   - SNIPS: lower variance, slight bias; usually preferred in practice.
+ *   - DR: doubly-robust — unbiased if either propensity OR Q-function is
+ *     correct. Lowest practical variance when Q is decent. Use this.
+ *
+ * Caveat the panel will land: on the LLM-agent setting, propensity scores
+ * recovered from token log-probs are noisy, the action space is enormous,
+ * and overlap is often poor. These estimators are useful but not magic;
+ * complement with `replayCampaign` (exact replay where the request hashes
+ * match) for high-confidence answers and OPE for the gap.
+ */
+interface OffPolicyTrajectory {
+  /** Stable id, for traceability through the dataset. */
+  runId: string;
+  /** Reward observed under the behavior policy (the realized outcome). */
+  reward: number;
+  /**
+   * Behavior-policy probability of the action that was taken. For LLM
+   * agents this is typically `exp(sum(token_log_probs))` over the chosen
+   * trajectory. Must be in (0, 1].
+   */
+  behaviorProb: number;
+  /**
+   * Target-policy probability of the same action. For replay-style
+   * counterfactual evaluation this is what the *new* policy would have
+   * assigned to the *old* trajectory. Must be in [0, 1].
+   */
+  targetProb: number;
+  /**
+   * Model-based reward prediction for the action selected by the behavior
+   * policy: `Q_hat(context, loggedAction)`. Supply this together with
+   * `vHatTarget` for contextual-bandit doubly-robust estimation.
+   */
+  qHatChosen?: number | null;
+  /**
+   * Expected model-based reward under the target policy:
+   * `sum_action targetPolicy(action | context) * Q_hat(context, action)`.
+   * Supply this together with `qHatChosen`. For an honest evaluation, both
+   * values must come from a model cross-fitted or trained outside this row.
+   */
+  vHatTarget?: number | null;
+}
+interface OffPolicyContributionCounts {
+  /** Contributions using the contextual-bandit doubly-robust formula. */
+  dr: number;
+  /** Contributions using exact IPS because no reward-model estimate was supplied. */
+  ipsFallback: number;
+}
+interface OffPolicyEstimate {
+  /** Estimated value of the target policy. */
+  value: number;
+  /** Standard error of the estimate. */
+  standardError: number;
+  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */
+  effectiveSampleSize: number;
+  /** Number of trajectories used. */
+  n: number;
+  /**
+   * Diagnostic: maximum importance weight observed. Large values (>>10x
+   * mean) are a red flag — variance is dominated by a few outliers.
+   */
+  maxImportanceWeight: number;
+  /** Populated by `doublyRobust` to expose which formula each row used. */
+  contributionCounts?: OffPolicyContributionCounts;
+}
+interface OffPolicyOptions {
+  /**
+   * Cap importance weights at this value (Ionides 2008 truncated IS) to
+   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).
+   * Set e.g. `10` for stable estimates when the policies are close.
+   */
+  weightCap?: number;
+  /** Reward clipping range. Default `[0, 1]`. */
+  rewardClip?: {
+    low: number;
+    high: number;
+  };
+}
+/**
+ * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator
+ * of E[reward under target policy]. Variance scales with the spread of
+ * target/behavior ratios.
+ */
+declare function inverseProbabilityWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at
+ * the cost of small bias (vanishing as N grows). The right default for
+ * LLM-agent evaluation where overlap is often poor.
+ */
+declare function selfNormalizedImportanceWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).
+ *
+ *     V_DR = (1/N) * sum_i [ v_hat_target_i
+ *              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]
+ *
+ * Unbiased if EITHER:
+ *   - the importance ratios are correct (IPS-style validity), OR
+ *   - the Q-hat function is correct (model-based validity).
+ *
+ * In practice both are imperfect, but the residual bias is the *product*
+ * of both errors — much smaller than either alone. This is why DR is the
+ * default in production OPE pipelines.
+ *
+ * `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither
+ * use the exact IPS contribution. `contributionCounts` makes the mix explicit
+ * in the result.
+ * Callers must cross-fit the Q-function or train it on independent rows;
+ * fitting and evaluating Q on the same outcomes leaks the answer.
+ */
+declare function doublyRobust(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Convenience: run all three estimators and return them side-by-side.
+ * The recommended diagnostic — agreement across estimators is a much
+ * stronger signal than any single one.
+ */
+declare function offPolicyEstimateAll(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): {
+  ips: OffPolicyEstimate;
+  snips: OffPolicyEstimate;
+  dr: OffPolicyEstimate;
+};
+//#endregion
 //#region src/rl/predictive-validity-researcher.d.ts
 interface PredictiveValidityResearcherOptions {
   outcomes: OutcomeStore;

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/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/predictive-validity-researcher.ts","../src/rl/rl-campaign.ts","../src/rl/run-record-adapters.ts","../src/rl/sim-fidelity.ts","../src/rl/tournament.ts"],"mappings":";;;;;;;;;;;;;;UAoCiB;EACf;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;iBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;UAoDc;EACf;;;;;;EAMA;;EAEA;EACA;;EAEA;;;;;;;;;;;;;iBAcc,mBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;iBAsDa,2BACd,MAAM,aACN;EAAQ;EAAwB;IAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UChMc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;UAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;UAG7D;EACf,QAAQ;;;;;EAKR;;;;;;EAMA;;iBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;UA0DM;EACf,MAAM;IACJ;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;;EAEA;;EAEA;;;;;;;iBAQc,wBACd,GAAG,iBACH,GAAG,iBACH;EAAQ;EAAqB;EAA6B;IACzD;;iBA8Ca,WAAW,OAAO,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC5KlC;;EAEf;;EAEA;;EAEA,OAAO;;UAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;UAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;iBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;UAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;UAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;iBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;UAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;UAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;iBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;UAkChB;EACf;EACA;EACA;EACA;;iBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3KhD;UAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;UAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;UAGJ;;EAEf;;;;;EAKA;;;;;EAKA;;UAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;IAEA;;;EAGF;IAAc;IAAW;;EACzB;EACA;EACA;EACA;;EAEA;;iBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;iBA0FK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;iBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;iBAuBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UC9MP;;;;;;;;;;;EAWf,OAAO;;;UAIQ;;EAEf;;EAEA;;EAEA;;;;KCLU;UAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;UAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;UAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;iBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;iBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;iBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;UChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;UAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;iBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;UAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;iBAGc,yBAAyB,aAAa,eAAe;UA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;iBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;UC3Jc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;KAIG,iBAAiB;cAEhB,yBAAyB;;;;;;;;;;;;;;;;iBAsBhB,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;iBAwCK,WAAW,MAAM;UAMhB;;EAEf;;EAEA;;;;;EAKA,cAAc;;UAGC,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;iBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;iBAyDK,YAAY,MAAM;UAMjB,mBAAmB;;EAElC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;EAEjB;;UAGe;EACf,UAAU;IAAQ;IAAuC;;EACzD,OAAO;;;;;;;;;;;;;;;;;;;;iBAqBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;iBAkDK,WAAW,MAAM;UAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;UAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;iBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;;;;;;;;iBA2CK,uBAAuB,MAAM,mBAAmB;cAwBnD,yBAAyB;iBA2BtB,WAAW,MAAM;UAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;cAGW,iCAAiC;;;;;;;;;;;iBAiB9B,mBAAmB,aAAa,cAAc,SAAS;;;KChf3D;cACN;KACM,wBAAwB;iBAIpB,uBAAuB,iBAAiB;;;UA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;UAIa;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;UAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;UAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;iBAyDa,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;iBA2DK,oBAAoB,GAAG;;;;;KCzO3B,eAAe;EAAc;EAAiB;;UAEzC;EACf;;;EAGA;EACA;;;;;;;;iBASc,eAAe,SAAS,gBAAgB,qBAAqB;;iBAqB7D,WAAW,qBAAqB;UAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;iBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;UCzEM;EACf,UAAU;EACV;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;cAOnD,wCAAwC;UAC3C;UACA;EAER,YAAY,MAAM;EAIZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EAmDhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EAwC9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAIlB,iBAAiB;;;;UC5KF,qBAAqB,WAAW,oBAAoB;;EAEnE,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf;;EAEA;IAAe;IAAgB;IAAgB;MAAS;MAAa;;;;EAErE;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;UAIO;EACf,UAAU;;EAEV,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;EAEb,mBAAmB;;EAEnB,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;iBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;UCrF9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;iBAWc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;iBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;KCnCS,oBAAoB,QAAQ,cAAc;;;;;cAMzC;;;cAIA;;;cAIA;;;;;cAMA;;;;;;;;;;;;;;;;;;;;;;cAyBA,yBAAyB;;;;;;;;;iBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;iBAqC3C,cAAc,kBAAkB;;;iBAsBhC,YAAY,eAAe;UAU1B;;EAEf;;;EAGA;;EAEA;;UAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;KAGU;UAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;UAGM;;EAEf,WAAW;;EAEX;;;;;;;;iBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;UAuHc;;;EAGf;;;EAGA;;UAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;iBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;iBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;UA+Gc;;EAEf;;EAEA;;iBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;UA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;iBAGc,0BACd,OAAO,iCACN"}
+{"version":3,"file":"rl.d.ts","names":[],"sources":["../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.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"],"mappings":";;;;;;;;;;;;;UAoCiB;EACf;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;iBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;UAoDc;EACf;;;;;;EAMA;;EAEA;EACA;;EAEA;;;;;;;;;;;;;iBAcc,mBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;iBAsDa,2BACd,MAAM,aACN;EAAQ;EAAwB;IAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UChMc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;UAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;UAG7D;EACf,QAAQ;;;;;EAKR;;;;;;EAMA;;iBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;UA0DM;EACf,MAAM;IACJ;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;;EAEA;;EAEA;;;;;;;iBAQc,wBACd,GAAG,iBACH,GAAG,iBACH;EAAQ;EAAqB;EAA6B;IACzD;;iBA8Ca,WAAW,OAAO,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC5KlC;;EAEf;;EAEA;;EAEA,OAAO;;UAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;UAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;iBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;UAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;UAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;iBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;UAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;UAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;iBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;UAkChB;EACf;EACA;EACA;EACA;;iBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3KhD;UAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;UAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;UAGJ;;EAEf;;;;;EAKA;;;;;EAKA;;UAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;IAEA;;;EAGF;IAAc;IAAW;;EACzB;EACA;EACA;EACA;;EAEA;;iBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;iBA0FK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;iBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;iBAuBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UC9MP;;;;;;;;;;;EAWf,OAAO;;;UAIQ;;EAEf;;EAEA;;EAEA;;;;KCLU;UAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;UAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;UAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;iBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;iBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;iBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;UChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;UAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;iBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;UAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;iBAGc,yBAAyB,aAAa,eAAe;UA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;iBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;UC3Jc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;KAIG,iBAAiB;cAEhB,yBAAyB;;;;;;;;;;;;;;;;iBAsBhB,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;iBAwCK,WAAW,MAAM;UAMhB;;EAEf;;EAEA;;;;;EAKA,cAAc;;UAGC,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;iBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;iBAyDK,YAAY,MAAM;UAMjB,mBAAmB;;EAElC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;EAEjB;;UAGe;EACf,UAAU;IAAQ;IAAuC;;EACzD,OAAO;;;;;;;;;;;;;;;;;;;;iBAqBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;iBAkDK,WAAW,MAAM;UAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;UAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;iBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;;;;;;;;iBA2CK,uBAAuB,MAAM,mBAAmB;cAwBnD,yBAAyB;iBA2BtB,WAAW,MAAM;UAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;cAGW,iCAAiC;;;;;;;;;;;iBAiB9B,mBAAmB,aAAa,cAAc,SAAS;;;KChf3D;cACN;KACM,wBAAwB;iBAIpB,uBAAuB,iBAAiB;;;UA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;UAIa;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;UAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;UAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;iBAyDa,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;iBA2DK,oBAAoB,GAAG;;;;;KCzO3B,eAAe;EAAc;EAAiB;;UAEzC;EACf;;;EAGA;EACA;;;;;;;;iBASc,eAAe,SAAS,gBAAgB,qBAAqB;;iBAqB7D,WAAW,qBAAqB;UAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;iBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UCnEM;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;UAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;iBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;iBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI;;;;UCtQ1C;EACf,UAAU;EACV;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;cAOnD,wCAAwC;UAC3C;UACA;EAER,YAAY,MAAM;EAIZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EAmDhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EAwC9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAIlB,iBAAiB;;;;UC5KF,qBAAqB,WAAW,oBAAoB;;EAEnE,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf;;EAEA;IAAe;IAAgB;IAAgB;MAAS;MAAa;;;;EAErE;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;UAIO;EACf,UAAU;;EAEV,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;EAEb,mBAAmB;;EAEnB,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;iBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;UCrF9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;iBAWc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;iBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;KCnCS,oBAAoB,QAAQ,cAAc;;;;;cAMzC;;;cAIA;;;cAIA;;;;;cAMA;;;;;;;;;;;;;;;;;;;;;;cAyBA,yBAAyB;;;;;;;;;iBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;iBAqC3C,cAAc,kBAAkB;;;iBAsBhC,YAAY,eAAe;UAU1B;;EAEf;;;EAGA;;EAEA;;UAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;KAGU;UAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;UAGM;;EAEf,WAAW;;EAEX;;;;;;;;iBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;UAuHc;;;EAGf;;;EAGA;;UAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;iBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;iBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;UA+Gc;;EAEf;;EAEA;;iBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;UA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;iBAGc,0BACd,OAAO,iCACN"}

package/dist/rl.js

@@ -7,11 +7,10 @@ import { t as isSplitEligible } from "./exporters-q9iL-2Jf.js";
 import { t as mintRolloutRows } from "./mint-D5_87M5L.js";
 import { a as InMemoryTraceStore } from "./integrity-BzRbCHzi.js";
 import { c as campaignCellToRunRecord, i as filterDeterministicallyRewarded, n as extractVerifiableReward, r as extractVerifiableRewardsFromRecords, t as detectRewardHacking } from "./reward-hacking-BEvjdUtD.js";
-import { t as runEvalCampaign } from "./eval-campaign-C2k-m4aY.js";
+import { t as runEvalCampaign } from "./eval-campaign-CYkUME2T.js";
 import { t as evaluateInterimReleaseConfidence } from "./sequential-Br0mAPHA.js";
 import { t as rubricPredictiveValidity } from "./rubric-predictive-validity-B3xmbmS1.js";
 import { n as thompsonCurriculum, r as varianceBasedCurriculum, t as observationsFromRunRecords } from "./active-curriculum-C4mk67HP.js";
-import { i as selfNormalizedImportanceWeighting, n as inverseProbabilityWeighting, r as offPolicyEstimateAll, t as doublyRobust } from "./off-policy-DvgzvtIx.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "./outcome-store-ChBKlTd_.js";
 import { appendFileSync, existsSync, mkdirSync, readFileSync } from "node:fs";
 import { dirname } from "node:path";
@@ -1242,6 +1241,222 @@ async function buildDatasetFromCorpus(corpusPath, config, opts = {}) {
 	return buildRlDataset(rows, lookups, config);
 }
 //#endregion
+//#region src/rl/off-policy.ts
+/**
+* Off-policy evaluation primitives.
+*
+* Standard inverse-probability-weighted (IPS), self-normalized
+* importance-weighted (SNIPS), and doubly-robust (DR) estimators for the
+* value of a *target* policy given trajectories collected under a
+* *behavior* policy. This is the canonical RL eval task: "we have last
+* week's runs, we changed the policy — how would the new one do without
+* re-running?"
+*
+* The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan
+* & Joachims 2015 for SNIPS) but the *application* to LLM-agent
+* evaluation needs care:
+*
+*   - The "policy" is the (prompt, tool config, model snapshot) triple.
+*     Two policies have the same probability over an action *iff* their
+*     LLM call would emit the same token with the same probability —
+*     which is generally unknowable without the model log-probs.
+*   - For LLM agents, propensity scores must be supplied by the caller
+*     (logged in the trace, recovered from token log-probs, or estimated
+*     via a learned propensity model). We do NOT estimate propensity here.
+*   - Doubly-robust requires two outputs from a Q-function: its prediction
+*     for the logged action and its expectation under the target policy.
+*     Consumers compute these with a tabular estimate, regression fit, or
+*     learned reward model before constructing the trajectories.
+*
+* Bias / variance tradeoffs:
+*   - IPS: unbiased; high variance for small overlap, infinite variance
+*     when target has support outside behavior.
+*   - SNIPS: lower variance, slight bias; usually preferred in practice.
+*   - DR: doubly-robust — unbiased if either propensity OR Q-function is
+*     correct. Lowest practical variance when Q is decent. Use this.
+*
+* Caveat the panel will land: on the LLM-agent setting, propensity scores
+* recovered from token log-probs are noisy, the action space is enormous,
+* and overlap is often poor. These estimators are useful but not magic;
+* complement with `replayCampaign` (exact replay where the request hashes
+* match) for high-confidence answers and OPE for the gap.
+*/
+/**
+* Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator
+* of E[reward under target policy]. Variance scales with the spread of
+* target/behavior ratios.
+*/
+function inverseProbabilityWeighting(trajectories, opts = {}) {
+	const cap = opts.weightCap ?? Infinity;
+	const clip = opts.rewardClip ?? {
+		low: 0,
+		high: 1
+	};
+	if (trajectories.length === 0) return zeroEstimate();
+	const weights = [];
+	const weightedRewards = [];
+	let maxW = 0;
+	for (const t of trajectories) {
+		if (t.behaviorProb <= 0) throw new ValidationError(`inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`);
+		const w = Math.min(cap, t.targetProb / t.behaviorProb);
+		const r = clamp(t.reward, clip.low, clip.high);
+		weights.push(w);
+		weightedRewards.push(w * r);
+		if (w > maxW) maxW = w;
+	}
+	const n = weights.length;
+	const value = weightedRewards.reduce((s, x) => s + x, 0) / n;
+	const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+	const sumW = weights.reduce((s, w) => s + w, 0);
+	const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+	return {
+		value,
+		standardError: Math.sqrt(variance / n),
+		effectiveSampleSize: effN,
+		n,
+		maxImportanceWeight: maxW
+	};
+}
+/**
+* Self-Normalized Importance Sampling. Lower variance than vanilla IPS at
+* the cost of small bias (vanishing as N grows). The right default for
+* LLM-agent evaluation where overlap is often poor.
+*/
+function selfNormalizedImportanceWeighting(trajectories, opts = {}) {
+	const cap = opts.weightCap ?? Infinity;
+	const clip = opts.rewardClip ?? {
+		low: 0,
+		high: 1
+	};
+	if (trajectories.length === 0) return zeroEstimate();
+	const weights = [];
+	const rewards = [];
+	let maxW = 0;
+	for (const t of trajectories) {
+		if (t.behaviorProb <= 0) throw new ValidationError(`selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`);
+		const w = Math.min(cap, t.targetProb / t.behaviorProb);
+		weights.push(w);
+		rewards.push(clamp(t.reward, clip.low, clip.high));
+		if (w > maxW) maxW = w;
+	}
+	const sumW = weights.reduce((s, w) => s + w, 0);
+	const sumWR = weights.reduce((s, w, i) => s + w * rewards[i], 0);
+	const value = sumW === 0 ? 0 : sumWR / sumW;
+	const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+	const variance = weights.map((w, i) => w * (rewards[i] - value)).reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW);
+	return {
+		value,
+		standardError: Math.sqrt(variance),
+		effectiveSampleSize: effN,
+		n: trajectories.length,
+		maxImportanceWeight: maxW
+	};
+}
+/**
+* Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).
+*
+*     V_DR = (1/N) * sum_i [ v_hat_target_i
+*              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]
+*
+* Unbiased if EITHER:
+*   - the importance ratios are correct (IPS-style validity), OR
+*   - the Q-hat function is correct (model-based validity).
+*
+* In practice both are imperfect, but the residual bias is the *product*
+* of both errors — much smaller than either alone. This is why DR is the
+* default in production OPE pipelines.
+*
+* `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither
+* use the exact IPS contribution. `contributionCounts` makes the mix explicit
+* in the result.
+* Callers must cross-fit the Q-function or train it on independent rows;
+* fitting and evaluating Q on the same outcomes leaks the answer.
+*/
+function doublyRobust(trajectories, opts = {}) {
+	const cap = opts.weightCap ?? Infinity;
+	const clip = opts.rewardClip ?? {
+		low: 0,
+		high: 1
+	};
+	if (trajectories.length === 0) return {
+		...zeroEstimate(),
+		contributionCounts: {
+			dr: 0,
+			ipsFallback: 0
+		}
+	};
+	const contributions = [];
+	const contributionCounts = {
+		dr: 0,
+		ipsFallback: 0
+	};
+	let maxW = 0;
+	let sumW = 0;
+	let sumW2 = 0;
+	for (const t of trajectories) {
+		if (t.behaviorProb <= 0) throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`);
+		const w = Math.min(cap, t.targetProb / t.behaviorProb);
+		const r = clamp(t.reward, clip.low, clip.high);
+		const rawQHatChosen = t.qHatChosen;
+		const rawVHatTarget = t.vHatTarget;
+		const hasQHatChosen = rawQHatChosen !== null && rawQHatChosen !== void 0;
+		const hasVHatTarget = rawVHatTarget !== null && rawVHatTarget !== void 0;
+		if (hasQHatChosen !== hasVHatTarget) throw new ValidationError(`doublyRobust: qHatChosen and vHatTarget must be supplied together (runId=${t.runId})`);
+		if (hasQHatChosen && hasVHatTarget) {
+			if (!Number.isFinite(rawQHatChosen) || !Number.isFinite(rawVHatTarget)) throw new ValidationError(`doublyRobust: qHatChosen and vHatTarget must be finite (runId=${t.runId})`);
+			const qHatChosen = clamp(rawQHatChosen, clip.low, clip.high);
+			const vHatTarget = clamp(rawVHatTarget, clip.low, clip.high);
+			contributions.push(vHatTarget + w * (r - qHatChosen));
+			contributionCounts.dr += 1;
+		} else {
+			contributions.push(w * r);
+			contributionCounts.ipsFallback += 1;
+		}
+		if (w > maxW) maxW = w;
+		sumW += w;
+		sumW2 += w * w;
+	}
+	const n = contributions.length;
+	const value = contributions.reduce((s, x) => s + x, 0) / n;
+	const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+	return {
+		value,
+		standardError: Math.sqrt(variance / n),
+		effectiveSampleSize: effN,
+		n,
+		maxImportanceWeight: maxW,
+		contributionCounts
+	};
+}
+/**
+* Convenience: run all three estimators and return them side-by-side.
+* The recommended diagnostic — agreement across estimators is a much
+* stronger signal than any single one.
+*/
+function offPolicyEstimateAll(trajectories, opts = {}) {
+	return {
+		ips: inverseProbabilityWeighting(trajectories, opts),
+		snips: selfNormalizedImportanceWeighting(trajectories, opts),
+		dr: doublyRobust(trajectories, opts)
+	};
+}
+function zeroEstimate() {
+	return {
+		value: 0,
+		standardError: 0,
+		effectiveSampleSize: 0,
+		n: 0,
+		maxImportanceWeight: 0
+	};
+}
+function clamp(x, lo, hi) {
+	if (!Number.isFinite(x)) return lo;
+	return Math.max(lo, Math.min(hi, x));
+}
+//#endregion
 //#region src/rl/predictive-validity-researcher.ts
 /**
 * Concrete `Researcher` driven by `rubricPredictiveValidity`. The brain: