@tangle-network/agent-eval 0.217.1 → 0.219.0

package/dist/{default-registry-DXgLmoDI.d.ts → default-registry-Cwg_QN3d.d.ts}

@@ -1,8 +1,8 @@
-import { W as TraceAnalysisStore, n as AnalystContext } from "./types-CW8ewATZ.js";
-import { l as ExactCapableAnalyst } from "./exact-types-CflbJ9f3.js";
-import { i as TraceAnalystLimits, r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-BvZ0R-36.js";
-import { t as TraceToolGroupName } from "./tool-groups-DyNz2JZ_.js";
-import { i as AnalystRegistryOptions, t as AnalystRegistry } from "./registry-ssNoHzwD.js";
+import { W as TraceAnalysisStore, n as AnalystContext } from "./types-8yQzNm8u.js";
+import { l as ExactCapableAnalyst } from "./exact-types-BYFmYq44.js";
+import { i as TraceAnalystLimits, r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-BpGGw6cd.js";
+import { t as TraceToolGroupName } from "./tool-groups-LISbBWCg.js";
+import { i as AnalystRegistryOptions, t as AnalystRegistry } from "./registry-B_yG-1jh.js";
 //#region src/analyst/behavioral-analyst.d.ts
 interface BehavioralAnalystOptions {
   /** Refuse larger unfiltered datasets instead of sampling them silently. Default 1,000. */
@@ -67,4 +67,4 @@ interface DefaultAnalystRegistryOptions {
 declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
 //#endregion
 export { createTraceAnalyst as a, behavioralAnalyst as c, TraceAnalystDefinition as i, buildDefaultAnalystRegistry as n, runTraceAnalyst as o, CreateTraceAnalystOptions as r, BehavioralAnalystOptions as s, DefaultAnalystRegistryOptions as t };
-//# sourceMappingURL=default-registry-DXgLmoDI.d.ts.map
+//# sourceMappingURL=default-registry-Cwg_QN3d.d.ts.map

package/dist/{default-registry-DXgLmoDI.d.ts.map → default-registry-Cwg_QN3d.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"default-registry-DXgLmoDI.d.ts","names":[],"sources":["../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;UA+BiB;;EAEf;;EAEA;;;iBAoGc,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgHI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;;;UCrKN;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
+{"version":3,"file":"default-registry-Cwg_QN3d.d.ts","names":[],"sources":["../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;UA+BiB;;EAEf;;EAEA;;;iBAoGc,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgHI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;;;UCrKN;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}

package/dist/{define-agent-eval-BIkH_gAR.d.ts → define-agent-eval-C2nFhHPu.d.ts}

@@ -1,12 +1,12 @@
-import { a as CostLedgerHandle, c as CostLedgerSummary, u as CostReceipt } from "./cost-ledger-Jk62OBfF.js";
-import { v as ProposalFinding } from "./types-CW8ewATZ.js";
-import { R as PowerPreflight, ct as EvidenceReceipt, nt as CampaignEvidenceContext } from "./statistical-heldout-C9AbgREN.js";
+import { d as CostReceipt, l as CostLedgerSummary, o as CostLedgerHandle } from "./cost-ledger-C1W3ICmP.js";
+import { v as ProposalFinding } from "./types-8yQzNm8u.js";
+import { R as PowerPreflight, ct as EvidenceReceipt, nt as CampaignEvidenceContext } from "./statistical-heldout-3z3I-kJN.js";
 import { nt as EvaluationClaim } from "./search-ledger-types-p2xKfiDJ.js";
-import { $n as CampaignComparisonUnits, Di as PremeasuredOptimizationBaseline, Jr as OptimizationMethod, Li as CampaignCellRetryPolicy, Ni as openAutoPr, Oi as RunOptimizationOptions, Pi as RunEvalOptions, Qr as OptimizationMethodResult, Ri as RunCampaignOptions, Ti as RunImprovementLoopResult, Xi as FinalEvidenceUse, Yi as FinalEvidencePolicy, Yn as LoopProvenanceRecord, Zr as OptimizationMethodProvenance, ti as ComparisonCost } from "./index-ZuWx52WG.js";
-import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-D8GhJoHk.js";
-import { At as SearchHistoryReceipt, Ot as SearchHistoryCoverageRow, Tt as SearchHistoryAdmissionOptions, Wt as CampaignStorage } from "./search-kernel-DxqjBxA-.js";
+import { $r as OptimizationMethodResult, Ei as RunImprovementLoopResult, Fi as RunEvalOptions, Oi as PremeasuredOptimizationBaseline, Pi as openAutoPr, Qr as OptimizationMethodProvenance, Ri as CampaignCellRetryPolicy, Xi as FinalEvidencePolicy, Xn as LoopProvenanceRecord, Yr as OptimizationMethod, Zi as FinalEvidenceUse, er as CampaignComparisonUnits, ki as RunOptimizationOptions, ni as ComparisonCost, zi as RunCampaignOptions } from "./index-Q6DvdcwI.js";
+import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-CHfn1zg5.js";
+import { At as SearchHistoryReceipt, Ot as SearchHistoryCoverageRow, Tt as SearchHistoryAdmissionOptions, Wt as CampaignStorage } from "./search-kernel-Dloxd6gs.js";
 import { n as HostedTenant } from "./client-hwo3cJjR.js";
-import { d as InsightReport } from "./analyze-runs-DyeNcenW.js";
+import { d as InsightReport } from "./analyze-runs-BWAHBDAw.js";
 //#region src/campaign/presets/run-final-comparison.d.ts
 interface FinalComparisonOptions<TScenario extends Scenario, TArtifact> extends Omit<RunCampaignOptions<TScenario, TArtifact>, 'dispatch'> {
   baselineSurface: MutableSurface;
@@ -512,4 +512,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
 declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
 //#endregion
 export { SelfImproveMethodResult as _, DefinedAgentEval as a, SelfImproveMethodOptions as c, SelfImproveProposerOptions as d, SelfImproveProposerResult as f, SelfImproveMethodProvenance as g, selfImprove as h, DefineAgentEvalOptions as i, SelfImproveOptions as l, SelfImproveRunError as m, AgentEvalEvaluateOptions as n, defineAgentEval as o, SelfImproveResult as p, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveProgressEvent as u };
-//# sourceMappingURL=define-agent-eval-BIkH_gAR.d.ts.map
+//# sourceMappingURL=define-agent-eval-C2nFhHPu.d.ts.map

package/dist/{define-agent-eval-BIkH_gAR.d.ts.map → define-agent-eval-C2nFhHPu.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"define-agent-eval-BIkH_gAR.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaU,uBAAuB,kBAAkB,UAAU,mBACnD,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}
+{"version":3,"file":"define-agent-eval-C2nFhHPu.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaU,uBAAuB,kBAAkB,UAAU,mBACnD,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}

package/dist/{define-agent-eval-BDxNlP3D.js → define-agent-eval-UQY3VbJ2.js}

@@ -1,12 +1,12 @@
 import { s as ValidationError } from "./errors-BcOs5MKW.js";
 import { i as modelHasSnapshot } from "./run-record-DhEJgoPz.js";
 import { t as analyzeRuns } from "./analyze-runs-DqPJaJ4X.js";
-import { B as defaultProductionGate, E as assertSearchHistoryAdmissionOptions, J as evaluationUnitMap, K as assertIndependentEvaluationSplit, L as runFinalComparison, R as openAutoPr, X as reserveFinalEvidence, Z as runEval, d as costFromLedgerSummary, et as resolveRunDir, p as runImprovementLoop, q as captureFinalEvidencePolicy, s as executeOptimizationMethod, u as combineComparisonCosts, z as captureJudge } from "./llm-judge-C8jMTNNG.js";
+import { B as defaultProductionGate, E as assertSearchHistoryAdmissionOptions, J as evaluationUnitMap, K as assertIndependentEvaluationSplit, L as runFinalComparison, R as openAutoPr, X as reserveFinalEvidence, Z as runEval, d as costFromLedgerSummary, et as resolveRunDir, p as runImprovementLoop, q as captureFinalEvidencePolicy, s as executeOptimizationMethod, u as combineComparisonCosts, z as captureJudge } from "./llm-judge-Cc8vDJh_.js";
 import { n as campaignCellToRunRecord } from "./run-record-CysKLB93.js";
-import { D as surfaceContentHash, O as surfaceDispatchRef, c as campaignMeasurementDigest, d as loopProvenanceArgsFromResult, h as pairedCampaignComposites, k as surfaceHash, l as canonicalDigest, t as createCampaignEvidenceReceipt, u as emitLoopProvenance, y as powerPreflight, z as campaignSplitDigest } from "./campaign-evidence-CQYw8_0H.js";
-import { M as fsCampaignStorage, N as inMemoryCampaignStorage, j as createRunCostLedger } from "./external-optimizer-subprocess-XIHBOSKt.js";
+import { D as surfaceContentHash, O as surfaceDispatchRef, c as campaignMeasurementDigest, d as loopProvenanceArgsFromResult, h as pairedCampaignComposites, k as surfaceHash, l as canonicalDigest, t as createCampaignEvidenceReceipt, u as emitLoopProvenance, y as powerPreflight, z as campaignSplitDigest } from "./campaign-evidence-DunzcJGv.js";
+import { M as fsCampaignStorage, N as inMemoryCampaignStorage, j as createRunCostLedger } from "./external-optimizer-subprocess-D670Oxf2.js";
 import { t as defineEvaluationClaim } from "./claim-BPVuyHuy.js";
-import { a as createHostedClient, i as startSearchShipper, n as shipBestEffort, r as shipSearchLedger } from "./search-shipper-Cw2z3-td.js";
+import { a as createHostedClient, i as startSearchShipper, n as shipBestEffort, r as shipSearchLedger } from "./search-shipper-Dn3NLsuH.js";
 import { fileURLToPath } from "node:url";
 //#region src/contract/self-improve-reporting.ts
 function pairedCompositeSummary(baseline, winner, independentUnitByScenarioId) {
@@ -68,16 +68,112 @@ function cellsToRunRecords(cells, candidateId, runId, surface, splitTag, fallbac
 }
 //#endregion
 //#region src/contract/self-improve-method.ts
+const METHOD_FINAL_PHASE = "selfImprove.method-final";
 /** Complete methods own search and selection; this path only measures their final choice. */
 async function runSelfImproveMethod(args) {
-	const { opts, train, selection, holdout, costLedger, storage, runDir, startedAt } = args;
+	const { opts, holdout, costLedger, startedAt } = args;
 	const budget = opts.budget ?? {};
+	assertAutoPrTarget(opts);
+	const evidenceContext = captureEvidenceContext(opts, budget);
+	const baselineSurface = structuredClone(opts.baselineSurface);
+	const { selected, cost: methodCost, history } = await searchWithMethod(args, baselineSurface);
+	const totalCost = () => methodTotalCost(opts.method.name, methodCost, costLedger);
+	const comparison = await measureMethodWinner(args, baselineSurface, selected, totalCost);
+	const evidence = evidenceContext ? methodEvidenceReceipts(comparison, baselineSurface, selected, evidenceContext) : void 0;
+	const deferred = comparison.holdout === "deferred";
+	const report = deferred ? null : pairedCompositeSummary(comparison.baselineOnHoldout, comparison.winnerOnHoldout, opts.claim ? evaluationUnitMap(opts.claim, holdout) : void 0);
+	const baseline = report?.baseline ?? null;
+	const winner = report?.winner ?? null;
+	const lift = baseline && winner ? winner.compositeMean - baseline.compositeMean : void 0;
+	const insight = deferred ? void 0 : await analyzeMethodHoldout(args, comparison, baselineSurface, selected);
+	const cost = totalCost();
+	const optimization = {
+		name: opts.method.name,
+		cost: methodCost,
+		...selected.durationMs === void 0 ? {} : { durationMs: selected.durationMs },
+		...selected.provenance === void 0 ? {} : { provenance: selected.provenance }
+	};
+	const durationMs = Date.now() - startedAt;
+	const receipts = costLedger.list();
+	const provenance = recordMethodProvenance({
+		args,
+		comparison,
+		baselineSurface,
+		selected,
+		report,
+		lift,
+		optimization,
+		cost,
+		durationMs,
+		receipts
+	});
+	opts.onProvenance?.(provenance);
+	opts.onProgress?.({
+		kind: "gate.decided",
+		decision: comparison.gateResult.decision,
+		...lift === void 0 ? {} : { lift }
+	});
+	const prResult = opts.autoOnPromote === "pr" && comparison.gateResult.decision === "ship" ? openAutoPr({
+		result: comparison.winnerOnHoldout,
+		gate: comparison.gateResult,
+		promotedDiff: comparison.promotedDiff,
+		ghOwner: opts.ghOwner,
+		ghRepo: opts.ghRepo
+	}) : void 0;
+	const result = {
+		...evidence ? { evidence } : {},
+		mode: "method",
+		...opts.claim ? { claim: opts.claim } : {},
+		...comparison.finalEvidence ? { finalEvidence: comparison.finalEvidence } : {},
+		searchHistoryCoverage: history,
+		baseline,
+		winner: {
+			surface: selected.winnerSurface,
+			compositeMean: winner?.compositeMean ?? null,
+			perScenario: winner?.perScenario ?? {},
+			label: opts.method.name
+		},
+		...lift === void 0 ? {} : { lift },
+		diff: comparison.promotedDiff,
+		gateDecision: comparison.gateResult.decision,
+		provenance,
+		optimization,
+		cost,
+		ledgerCost: costLedger.summary(),
+		totalCostUsd: cost.totalCostUsd,
+		durationMs,
+		receipts,
+		...insight ? { insight } : {},
+		...selected.searchHistory ? { searchHistory: selected.searchHistory } : {},
+		raw: {
+			...comparison,
+			kind: "method",
+			baselineSurface,
+			winnerSurface: selected.winnerSurface,
+			winnerSurfaceHash: surfaceHash(selected.winnerSurface),
+			method: selected,
+			cost,
+			...prResult ? { prResult } : {}
+		}
+	};
+	await shipMethodSearchLedger(opts, selected);
+	return result;
+}
+/** An auto-PR needs a target repository before any spend. */
+function assertAutoPrTarget(opts) {
 	if (opts.autoOnPromote === "pr" && (!opts.ghOwner || !opts.ghRepo)) throw new Error("selfImprove: autoOnPromote='pr' requires ghOwner + ghRepo");
+}
+/** Detach the caller's evidence context; receipts need a measured holdout. */
+function captureEvidenceContext(opts, budget) {
 	const evidenceContext = opts.evidence === void 0 ? void 0 : structuredClone(opts.evidence);
 	if (evidenceContext && budget.holdout === "deferred") throw new Error("selfImprove: evidence receipts require measured holdout");
-	const baselineSurface = structuredClone(opts.baselineSurface);
-	const judge = opts.judge;
-	const { selected, cost: methodCost, history } = await executeOptimizationMethod({
+	return evidenceContext;
+}
+/** Let the method search and select a winner on the train and selection splits. */
+function searchWithMethod(args, baselineSurface) {
+	const { opts, train, selection, costLedger, storage, runDir } = args;
+	const budget = opts.budget ?? {};
+	return executeOptimizationMethod({
 		method: opts.method,
 		storage,
 		searchHistoryPolicy: opts.searchHistoryPolicy,
@@ -87,7 +183,7 @@ async function runSelfImproveMethod(args) {
 			trainScenarios: train,
 			selectionScenarios: selection,
 			dispatchWithSurface: opts.agent,
-			judges: [judge],
+			judges: [opts.judge],
 			runDir: `${runDir}/optimization/${opts.method.name.replace(/[^a-zA-Z0-9._-]/g, "_")}`,
 			seed: 42,
 			runOptions: {
@@ -105,29 +201,39 @@ async function runSelfImproveMethod(args) {
 			costLedger
 		}
 	});
-	const finalPhase = "selfImprove.method-final";
-	const finalCost = () => combineComparisonCosts([
+}
+/** Spend on the final holdout comparison, read back from the shared ledger. */
+function methodFinalCost(costLedger) {
+	return combineComparisonCosts([
 		"holdout.baseline",
 		"holdout.winner",
 		"holdout.neutralized",
 		"promotion.gate"
 	].map((phase) => ({
 		label: phase,
-		cost: costFromLedgerSummary(costLedger.summary({ phase: `${finalPhase}.${phase}` }))
+		cost: costFromLedgerSummary(costLedger.summary({ phase: `${METHOD_FINAL_PHASE}.${phase}` }))
 	})));
-	const totalCost = () => combineComparisonCosts([{
-		label: opts.method.name,
+}
+/** The method's own spend plus the final comparison's. */
+function methodTotalCost(methodName, methodCost, costLedger) {
+	return combineComparisonCosts([{
+		label: methodName,
 		cost: methodCost
 	}, {
 		label: "final comparison",
-		cost: finalCost()
+		cost: methodFinalCost(costLedger)
 	}]);
+}
+/** Measure baseline and winner on the holdout; the gate holds when spend breaks the budget. */
+function measureMethodWinner(args, baselineSurface, selected, totalCost) {
+	const { opts, holdout, costLedger, storage, runDir } = args;
+	const budget = opts.budget ?? {};
 	const gate = opts.gate ?? defaultProductionGate({
 		holdoutScenarios: holdout,
 		deltaThreshold: opts.claim?.minimumEffect ?? .05,
 		independentUnitByScenarioId: opts.claim ? evaluationUnitMap(opts.claim, holdout) : void 0
 	});
-	const comparison = await runFinalComparison({
+	return runFinalComparison({
 		claim: opts.claim,
 		finalEvidence: opts.finalEvidence,
 		baselineSurface,
@@ -135,7 +241,7 @@ async function runSelfImproveMethod(args) {
 		scenarios: holdout,
 		dispatchWithSurface: opts.agent,
 		dispatchRef: opts.dispatchRef,
-		judges: [judge],
+		judges: [opts.judge],
 		holdout: budget.holdout,
 		neutralize: opts.neutralize,
 		gate: {
@@ -160,7 +266,7 @@ async function runSelfImproveMethod(args) {
 		runDir,
 		storage,
 		costLedger,
-		costPhase: finalPhase,
+		costPhase: METHOD_FINAL_PHASE,
 		reps: budget.reps,
 		maxConcurrency: budget.maxConcurrency ?? 2,
 		dispatchTimeoutMs: opts.dispatchTimeoutMs,
@@ -169,7 +275,10 @@ async function runSelfImproveMethod(args) {
 		expectUsage: opts.expectUsage ?? "assert",
 		label: "selfImprove"
 	});
-	const evidence = evidenceContext ? {
+}
+/** Evidence receipts for the baseline and winner holdout campaigns. */
+function methodEvidenceReceipts(comparison, baselineSurface, selected, evidenceContext) {
+	return {
 		baseline: createCampaignEvidenceReceipt({
 			campaign: comparison.baselineOnHoldout,
 			surface: baselineSurface,
@@ -180,28 +289,26 @@ async function runSelfImproveMethod(args) {
 			surface: selected.winnerSurface,
 			context: evidenceContext
 		})
-	} : void 0;
-	const deferred = comparison.holdout === "deferred";
-	const report = deferred ? null : pairedCompositeSummary(comparison.baselineOnHoldout, comparison.winnerOnHoldout, opts.claim ? evaluationUnitMap(opts.claim, holdout) : void 0);
-	const baseline = report?.baseline ?? null;
-	const winner = report?.winner ?? null;
-	const lift = baseline && winner ? winner.compositeMean - baseline.compositeMean : void 0;
-	const insight = deferred ? void 0 : await analyzeRuns({
+	};
+}
+/** Analyze the measured holdout runs; a winner that is the baseline adds no second candidate. */
+function analyzeMethodHoldout(args, comparison, baselineSurface, selected) {
+	const { opts, holdout, runDir } = args;
+	return analyzeRuns({
 		runs: [...cellsToRunRecords(comparison.baselineOnHoldout.cells, "baseline", runDir, baselineSurface, "holdout", opts.model), ...comparison.winnerOnHoldout === comparison.baselineOnHoldout ? [] : cellsToRunRecords(comparison.winnerOnHoldout.cells, "winner", runDir, selected.winnerSurface, "holdout", opts.model)],
 		baselineCandidateId: "baseline",
 		independentUnitByScenarioId: opts.claim ? evaluationUnitMap(opts.claim, holdout) : void 0,
 		decisionThreshold: opts.claim?.minimumEffect ?? .05,
 		...comparison.winnerOnHoldout === comparison.baselineOnHoldout ? {} : { candidateCandidateId: "winner" }
 	});
-	const cost = totalCost();
-	const optimization = {
-		name: opts.method.name,
-		cost: methodCost,
-		...selected.durationMs === void 0 ? {} : { durationMs: selected.durationMs },
-		...selected.provenance === void 0 ? {} : { provenance: selected.provenance }
-	};
-	const durationMs = Date.now() - startedAt;
-	const receipts = costLedger.list();
+}
+/** Build, persist and return the digest-sealed provenance record. */
+function recordMethodProvenance(input) {
+	const { args, comparison, baselineSurface, selected, report, lift, optimization } = input;
+	const { opts, train, selection, holdout, storage, runDir, startedAt } = args;
+	const budget = opts.budget ?? {};
+	const baseline = report?.baseline ?? null;
+	const winner = report?.winner ?? null;
 	const record = {
 		schema: "tangle.method-improvement",
 		runId: `${runDir}#${startedAt}`,
@@ -218,20 +325,20 @@ async function runSelfImproveMethod(args) {
 			holdoutSplitDigest: campaignSplitDigest(holdout, budget.reps ?? 1),
 			baselineCampaignDigest: campaignMeasurementDigest(comparison.baselineOnHoldout),
 			winnerCampaignDigest: campaignMeasurementDigest(comparison.winnerOnHoldout),
-			costReceiptsDigest: canonicalDigest(receipts),
+			costReceiptsDigest: canonicalDigest(input.receipts),
 			...report ? { holdoutObservations: report.observations } : {},
 			...comparison.neutralizedOnHoldout ? { neutralizedCampaignDigest: campaignMeasurementDigest(comparison.neutralizedOnHoldout) } : {}
 		},
 		gate: comparison.gateResult,
-		...deferred ? { holdout: "deferred" } : {},
+		...comparison.holdout === "deferred" ? { holdout: "deferred" } : {},
 		...baseline && winner ? {
 			baselineHoldoutComposite: baseline.compositeMean,
 			winnerHoldoutComposite: winner.compositeMean,
 			heldOutLift: lift
 		} : {},
-		cost,
-		totalCostUsd: cost.totalCostUsd,
-		totalDurationMs: durationMs
+		cost: input.cost,
+		totalCostUsd: input.cost.totalCostUsd,
+		totalDurationMs: input.durationMs
 	};
 	const serialized = JSON.parse(JSON.stringify(record));
 	const provenance = {
@@ -240,70 +347,23 @@ async function runSelfImproveMethod(args) {
 	};
 	storage.ensureDir(runDir);
 	storage.write(`${runDir}/method-provenance.json`, JSON.stringify(provenance, null, 2));
-	opts.onProvenance?.(provenance);
-	opts.onProgress?.({
-		kind: "gate.decided",
-		decision: comparison.gateResult.decision,
-		...lift === void 0 ? {} : { lift }
-	});
-	const prResult = opts.autoOnPromote === "pr" && comparison.gateResult.decision === "ship" ? openAutoPr({
-		result: comparison.winnerOnHoldout,
-		gate: comparison.gateResult,
-		promotedDiff: comparison.promotedDiff,
-		ghOwner: opts.ghOwner,
-		ghRepo: opts.ghRepo
-	}) : void 0;
-	const result = {
-		...evidence ? { evidence } : {},
-		mode: "method",
-		...opts.claim ? { claim: opts.claim } : {},
-		...comparison.finalEvidence ? { finalEvidence: comparison.finalEvidence } : {},
-		searchHistoryCoverage: history,
-		baseline,
-		winner: {
-			surface: selected.winnerSurface,
-			compositeMean: winner?.compositeMean ?? null,
-			perScenario: winner?.perScenario ?? {},
-			label: opts.method.name
-		},
-		...lift === void 0 ? {} : { lift },
-		diff: comparison.promotedDiff,
-		gateDecision: comparison.gateResult.decision,
-		provenance,
-		optimization,
-		cost,
-		ledgerCost: costLedger.summary(),
-		totalCostUsd: cost.totalCostUsd,
-		durationMs,
-		receipts,
-		...insight ? { insight } : {},
-		...selected.searchHistory ? { searchHistory: selected.searchHistory } : {},
-		raw: {
-			...comparison,
-			kind: "method",
-			baselineSurface,
-			winnerSurface: selected.winnerSurface,
-			winnerSurfaceHash: surfaceHash(selected.winnerSurface),
-			method: selected,
-			cost,
-			...prResult ? { prResult } : {}
-		}
-	};
-	if (opts.hostedTenant) {
-		const receipt = selected.searchHistory;
-		if (receipt) {
-			const path = fileURLToPath(receipt.ledger.uri);
-			await shipBestEffort(() => shipSearchLedger({
-				tenant: opts.hostedTenant,
-				ledger: {
-					path,
-					searchId: receipt.summary.searchId
-				},
-				runKind: "optimization"
-			}), path);
-		} else console.warn(`[agent-eval] hostedTenant: method ${opts.method.name} recorded no search ledger, so nothing was shipped`);
-	}
-	return result;
+	return provenance;
+}
+/** Ship the method's search ledger to the hosted tenant, when one is configured. */
+async function shipMethodSearchLedger(opts, selected) {
+	if (!opts.hostedTenant) return;
+	const receipt = selected.searchHistory;
+	if (receipt) {
+		const path = fileURLToPath(receipt.ledger.uri);
+		await shipBestEffort(() => shipSearchLedger({
+			tenant: opts.hostedTenant,
+			ledger: {
+				path,
+				searchId: receipt.summary.searchId
+			},
+			runKind: "optimization"
+		}), path);
+	} else console.warn(`[agent-eval] hostedTenant: method ${opts.method.name} recorded no search ledger, so nothing was shipped`);
 }
 //#endregion
 //#region src/contract/self-improve.ts
@@ -728,4 +788,4 @@ function requirePositiveInteger(value, field) {
 //#endregion
 export { SelfImproveRunError as n, selfImprove as r, defineAgentEval as t };
 
-//# sourceMappingURL=define-agent-eval-BDxNlP3D.js.map
+//# sourceMappingURL=define-agent-eval-UQY3VbJ2.js.map