@tangle-network/agent-eval 0.177.0 → 0.178.0

package/dist/cli.js

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import { i as runRolloutReleaseCli } from "./hf-dataset-D8_RNIis.js";
 import { t as createChatClient } from "./chat-client-CkmjYlfB.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-BrsZhMSk.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-DIxGSUP3.js";
 import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-D_cjseFN.js";
 import { t as runSupervisorRunCommand } from "./report-command-DKlXfU5r.js";
 import { writeFileSync } from "node:fs";

package/dist/contract/index.d.ts

@@ -3,12 +3,12 @@ import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin,
 import { S as createChatClient, _ as CreateChatClientOpts, p as ChatClient } from "../types-gvRsyJLh.js";
 import { $ as paretoSignificanceGate, G as BuildEvidenceVectorOptions, J as ParetoSignificanceGateOptions, K as EvidenceVector, Q as paretoPolicy, U as AxisEvidence, W as AxisVerdict, X as PromotionPolicy, Y as PromotionObjective, Z as buildEvidenceVector, q as ObjectiveSource } from "../statistical-heldout-Z9NROFFS.js";
 import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-BJz2CPTM.js";
-import { Ar as SkillOptOptimizationMethodConfig, Br as gepaOptimizationMethod, Ci as ReferenceEquivalenceScenario, Di as LlmJudgeOptions, Do as inMemoryCampaignStorage, Ei as LlmJudgeDimension, Eo as fsCampaignStorage, Fi as runImprovementLoop, Fr as GepaEngineOptions, Gr as ExternalTextOptimizerContext, Hr as OptimizerModelBudget, Ir as GepaEngineRun, Jr as ExternalTextEvaluationResponse, Kr as ExternalTextOptimizerResult, Lr as GepaOptimizationMethodConfig, Mi as transientDispatchFailure, Mr as SkillOptTrainerConfig, Ni as RunImprovementLoopOptions, Nr as skillOptOptimizationMethod, Oi as llmJudge, Pi as RunImprovementLoopResult, Pr as GepaAdaptiveEngineRun, Qr as OptimizationMethodComparison, Rr as GepaOptimizationRecipe, Si as ReferenceEquivalenceJudgeResult, Ti as runReferenceEquivalenceJudge, Ur as externalTextOptimizationMethod, Vr as OpenAICompatibleOptimizerModel, Wr as ExternalTextOptimizationMethodConfig, Xr as CompareOptimizationMethodsOptions, Zr as OptimizationMethod, bi as ReferenceEquivalenceJudgeInput, ci as compareOptimizationMethods, co as RunEvalOptions, do as CampaignCellRetryPolicy, dr as DefaultProductionGateCheck, ei as OptimizationMethodInput, fo as RunCampaignOptions, fr as DefaultProductionGateOptions, jo as campaignSplitDigest, jr as SkillOptRunnerCommand, kr as composeGate, lo as runEval, lr as HeldOutGateOptions, mr as defaultProductionGate, ni as OptimizationMethodProvenance, oi as OptimizationPackageSource, po as runCampaign, pr as DefaultProductionRewardHackingOptions, qr as ExternalOptimizationExample, ri as OptimizationMethodResult, si as OptimizationTokenUsage, ui as ComparisonCost, uo as CampaignCellFailureReceipt, ur as heldOutGate, vi as REFERENCE_EQUIVALENCE_INPUT_LIMITS, wi as createReferenceEquivalenceJudge, wo as CampaignStorage, xi as ReferenceEquivalenceJudgeOptions, yi as REFERENCE_EQUIVALENCE_JUDGE_VERSION, zr as GepaRunnerCommand } from "../index-CtGf6e6X.js";
+import { Ar as SkillOptOptimizationMethodConfig, Br as gepaOptimizationMethod, Ci as ReferenceEquivalenceScenario, Di as LlmJudgeOptions, Do as inMemoryCampaignStorage, Ei as LlmJudgeDimension, Eo as fsCampaignStorage, Fi as runImprovementLoop, Fr as GepaEngineOptions, Gr as ExternalTextOptimizerContext, Hr as OptimizerModelBudget, Ir as GepaEngineRun, Jr as ExternalTextEvaluationResponse, Kr as ExternalTextOptimizerResult, Lr as GepaOptimizationMethodConfig, Mi as transientDispatchFailure, Mr as SkillOptTrainerConfig, Ni as RunImprovementLoopOptions, Nr as skillOptOptimizationMethod, Oi as llmJudge, Pi as RunImprovementLoopResult, Pr as GepaAdaptiveEngineRun, Qr as OptimizationMethodComparison, Rr as GepaOptimizationRecipe, Si as ReferenceEquivalenceJudgeResult, Ti as runReferenceEquivalenceJudge, Ur as externalTextOptimizationMethod, Vr as OpenAICompatibleOptimizerModel, Wr as ExternalTextOptimizationMethodConfig, Xr as CompareOptimizationMethodsOptions, Zr as OptimizationMethod, bi as ReferenceEquivalenceJudgeInput, ci as compareOptimizationMethods, co as RunEvalOptions, do as CampaignCellRetryPolicy, dr as DefaultProductionGateCheck, ei as OptimizationMethodInput, fo as RunCampaignOptions, fr as DefaultProductionGateOptions, jo as campaignSplitDigest, jr as SkillOptRunnerCommand, kr as composeGate, lo as runEval, lr as HeldOutGateOptions, mr as defaultProductionGate, ni as OptimizationMethodProvenance, oi as OptimizationPackageSource, po as runCampaign, pr as DefaultProductionRewardHackingOptions, qr as ExternalOptimizationExample, ri as OptimizationMethodResult, si as OptimizationTokenUsage, ui as ComparisonCost, uo as CampaignCellFailureReceipt, ur as heldOutGate, vi as REFERENCE_EQUIVALENCE_INPUT_LIMITS, wi as createReferenceEquivalenceJudge, wo as CampaignStorage, xi as ReferenceEquivalenceJudgeOptions, yi as REFERENCE_EQUIVALENCE_JUDGE_VERSION, zr as GepaRunnerCommand } from "../index-CZ2tAJW2.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
 import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-DETqPc_A.js";
 import { n as HostedTenant } from "../client-vyYQg3bm.js";
-import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-CqmlXUfQ.js";
-import { $ as diffRunBaselineToWinner, A as CodeAgentSessionActionStatus, At as measuredComparisonFromCandidateExperiment, B as AgentTraceFile, C as fromOpenCodeSession, Ct as EvaluatePairedMeasurementsOptions, D as streamCodeAgentJsonlFile, Dt as RunCandidateExperimentOptions, E as parseCodeAgentJsonlFile, Et as PairedMeasurementEvaluation, F as CodeAgentSessionTerminalStatus, Ft as verifyCandidateExperiment, G as PartitionByAuthoringModelResult, H as AgentTraceRange, I as observeCodeAgentSession, It as verifyCandidateExperimentComparison, J as EvalCellScoreDelta, K as parseAgentTrace, L as AgentTraceContributor, M as CodeAgentSessionExecutionReceipt, Mt as sealCandidateBenchmarkSuite, N as CodeAgentSessionObservation, Nt as sealCandidateBenchmarkTask, O as CodeAgentSessionAction, Ot as SealCandidateBenchmarkSuiteOptions, P as CodeAgentSessionSource, Pt as sealCandidateExperiment, Q as diffGenerations, R as AgentTraceContributorType, S as fromKimiCodeSession, St as CompareCandidateExperimentOptions, T as parseCodeAgentJsonl, Tt as PairedMeasurementAdapter, U as AgentTraceRecord, V as AgentTraceIndex, W as AuthoringProvenance, X as EvalGenerationDiff, Y as EvalDimensionDelta, Z as EvalRunDiff, _ as CodeAgentSessionIntakeResult, _t as sealAgentProfileImprovementSuite, a as FromFeedbackTableOptions, at as FromRunRecordDirOptions, b as fromClaudeCodeSession, bt as CandidateExperimentExecutionInput, c as CodeAgentStoreObservation, ct as fromRunRecordDir, d as CodeAgentStoreScope, dt as CompareAgentProfileImprovementExperimentOptions, et as diffRuns, f as CodeAgentStoreSession, ft as RunAgentProfileImprovementExperimentOptions, g as CodeAgentSessionIntakeOptions, gt as sealAgentProfileImprovementExperiment, h as CodeAgentSessionDiagnostic, ht as runAgentProfileImprovementExperiment, i as FeedbackTableRow, it as evalReportingSuite, j as CodeAgentSessionActionSurface, jt as runCandidateExperiment, k as CodeAgentSessionActionKind, kt as evaluatePairedMeasurements, l as CodeAgentStoreRejection, lt as AgentProfileImprovementExperimentExecutionInput, m as CodeAgentJsonlLine, mt as measuredComparisonFromAgentProfileImprovementExperiment, n as fromOtelSpans, nt as EvalReportingSuiteOptions, o as FromFeedbackTableResult, ot as FromRunRecordDirResult, p as observeCodeAgentStore, pt as SealAgentProfileImprovementSuiteOptions, q as partitionRunsByAuthoringModel, r as FeedbackTableMeta, rt as EvalReportingSuiteResult, s as fromFeedbackTable, st as RunRecordRejection, t as FromOtelSpansOptions, tt as EvalReportingSuiteInput, u as CodeAgentStoreRejections, ut as AgentProfileImprovementExperimentRun, v as CodeAgentSessionMetrics, vt as sealAgentProfileImprovementTask, w as fromPiSession, wt as PairedMeasurement, x as fromCodexSession, xt as CandidateExperimentRun, y as ParsedCodeAgentJsonl, yt as verifyAgentProfileImprovementExperimentComparison, z as AgentTraceConversation } from "../index-DuaNwvse.js";
+import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-DMsxDp1u.js";
+import { $ as diffRunBaselineToWinner, A as CodeAgentSessionActionStatus, At as measuredComparisonFromCandidateExperiment, B as AgentTraceFile, C as fromOpenCodeSession, Ct as EvaluatePairedMeasurementsOptions, D as streamCodeAgentJsonlFile, Dt as RunCandidateExperimentOptions, E as parseCodeAgentJsonlFile, Et as PairedMeasurementEvaluation, F as CodeAgentSessionTerminalStatus, Ft as verifyCandidateExperiment, G as PartitionByAuthoringModelResult, H as AgentTraceRange, I as observeCodeAgentSession, It as verifyCandidateExperimentComparison, J as EvalCellScoreDelta, K as parseAgentTrace, L as AgentTraceContributor, M as CodeAgentSessionExecutionReceipt, Mt as sealCandidateBenchmarkSuite, N as CodeAgentSessionObservation, Nt as sealCandidateBenchmarkTask, O as CodeAgentSessionAction, Ot as SealCandidateBenchmarkSuiteOptions, P as CodeAgentSessionSource, Pt as sealCandidateExperiment, Q as diffGenerations, R as AgentTraceContributorType, S as fromKimiCodeSession, St as CompareCandidateExperimentOptions, T as parseCodeAgentJsonl, Tt as PairedMeasurementAdapter, U as AgentTraceRecord, V as AgentTraceIndex, W as AuthoringProvenance, X as EvalGenerationDiff, Y as EvalDimensionDelta, Z as EvalRunDiff, _ as CodeAgentSessionIntakeResult, _t as sealAgentProfileImprovementSuite, a as FromFeedbackTableOptions, at as FromRunRecordDirOptions, b as fromClaudeCodeSession, bt as CandidateExperimentExecutionInput, c as CodeAgentStoreObservation, ct as fromRunRecordDir, d as CodeAgentStoreScope, dt as CompareAgentProfileImprovementExperimentOptions, et as diffRuns, f as CodeAgentStoreSession, ft as RunAgentProfileImprovementExperimentOptions, g as CodeAgentSessionIntakeOptions, gt as sealAgentProfileImprovementExperiment, h as CodeAgentSessionDiagnostic, ht as runAgentProfileImprovementExperiment, i as FeedbackTableRow, it as evalReportingSuite, j as CodeAgentSessionActionSurface, jt as runCandidateExperiment, k as CodeAgentSessionActionKind, kt as evaluatePairedMeasurements, l as CodeAgentStoreRejection, lt as AgentProfileImprovementExperimentExecutionInput, m as CodeAgentJsonlLine, mt as measuredComparisonFromAgentProfileImprovementExperiment, n as fromOtelSpans, nt as EvalReportingSuiteOptions, o as FromFeedbackTableResult, ot as FromRunRecordDirResult, p as observeCodeAgentStore, pt as SealAgentProfileImprovementSuiteOptions, q as partitionRunsByAuthoringModel, r as FeedbackTableMeta, rt as EvalReportingSuiteResult, s as fromFeedbackTable, st as RunRecordRejection, t as FromOtelSpansOptions, tt as EvalReportingSuiteInput, u as CodeAgentStoreRejections, ut as AgentProfileImprovementExperimentRun, v as CodeAgentSessionMetrics, vt as sealAgentProfileImprovementTask, w as fromPiSession, wt as PairedMeasurement, x as fromCodexSession, xt as CandidateExperimentRun, y as ParsedCodeAgentJsonl, yt as verifyAgentProfileImprovementExperimentComparison, z as AgentTraceConversation } from "../index-nrWQbUgi.js";
 import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-FfNzaUHV.js";
 import { S as summarizeExecution, b as SummarizeExecutionOptions, v as AnalyzeRunsOptions, x as analyzeRuns, y as ExecutionReport } from "../engine-CvW_I72-.js";
 export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentProfileImprovementExperimentRun, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type CandidateExperimentRun, type ChatClient, type CodeAgentJsonlLine, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeAgentStoreObservation, type CodeAgentStoreRejection, type CodeAgentStoreRejections, type CodeAgentStoreScope, type CodeAgentStoreSession, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };

package/dist/contract/index.js

@@ -1,9 +1,9 @@
 import { s as ValidationError } from "../errors-Dngq5h35.js";
 import { a as hashCanonical } from "../canonical-DPyQ_rpt.js";
 import { r as pairedBootstrap } from "../paired-tests-C8iCsioC.js";
-import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-CJG7LD9M.js";
+import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-Dzidv34q.js";
 import { a as parseRunRecordSafe } from "../run-record-DQpSf7t-.js";
-import { G as runCampaign, R as defaultProductionGate, W as runEval, a as transientDispatchFailure, c as compareOptimizationMethods, h as runImprovementLoop, t as llmJudge } from "../llm-judge-CV80fkYA.js";
+import { G as runCampaign, R as defaultProductionGate, W as runEval, a as transientDispatchFailure, c as compareOptimizationMethods, h as runImprovementLoop, t as llmJudge } from "../llm-judge-v80Kmu9g.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-Qv-cpptD.js";
 import { i as CostLedger } from "../cost-ledger-B1qx30B4.js";
 import { M as heldoutSignificance, P as decidePairedPromotion, et as campaignSplitDigest } from "../campaign-evidence-D8DBLqLI.js";
@@ -17,7 +17,7 @@ import { t as extractUsage } from "../extract-usage-BrQ8mCLX.js";
 import { t as createChatClient } from "../chat-client-CkmjYlfB.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
 import { i as summarizeTraceErrors, n as recordAggregateMeasurements, r as summarizeExecutionMeasurements, t as readTaskFailureLabels } from "../task-failure-attributes-CUy9mkIY.js";
-import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-D0o2c2yM.js";
+import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-C3oYul8v.js";
 import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-DWOm70gx.js";
 import { createHash } from "node:crypto";
 import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";

package/dist/{define-agent-eval-CqmlXUfQ.d.ts → define-agent-eval-DMsxDp1u.d.ts}

@@ -2,7 +2,7 @@ import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from
 import { _ as ProposalFinding } from "./types-DN2WdT5S.js";
 import { B as PowerPreflight, ft as EvidenceReceipt, it as CampaignEvidenceContext } from "./statistical-heldout-Z9NROFFS.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-BJz2CPTM.js";
-import { $n as LoopProvenanceRecord, Ii as PremeasuredOptimizationBaseline, Li as RunOptimizationOptions, Pi as RunImprovementLoopResult, Zr as OptimizationMethod, aa as SearchHistoryReceipt, co as RunEvalOptions, do as CampaignCellRetryPolicy, ea as SearchHistoryAdmissionOptions, fo as RunCampaignOptions, ni as OptimizationMethodProvenance, ra as SearchHistoryCoverageRow, ri as OptimizationMethodResult, so as openAutoPr, ui as ComparisonCost, wo as CampaignStorage } from "./index-CtGf6e6X.js";
+import { $n as LoopProvenanceRecord, Ii as PremeasuredOptimizationBaseline, Li as RunOptimizationOptions, Pi as RunImprovementLoopResult, Zr as OptimizationMethod, aa as SearchHistoryReceipt, co as RunEvalOptions, do as CampaignCellRetryPolicy, ea as SearchHistoryAdmissionOptions, fo as RunCampaignOptions, ni as OptimizationMethodProvenance, ra as SearchHistoryCoverageRow, ri as OptimizationMethodResult, so as openAutoPr, ui as ComparisonCost, wo as CampaignStorage } from "./index-CZ2tAJW2.js";
 import { o as InsightReport } from "./insight-report-DETqPc_A.js";
 import { n as HostedTenant } from "./client-vyYQg3bm.js";
 //#region src/campaign/presets/run-final-comparison.d.ts
@@ -502,4 +502,4 @@ 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-CqmlXUfQ.d.ts.map
+//# sourceMappingURL=define-agent-eval-DMsxDp1u.d.ts.map

package/dist/{define-agent-eval-CqmlXUfQ.d.ts.map → define-agent-eval-DMsxDp1u.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"define-agent-eval-CqmlXUfQ.d.ts","names":[],"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":";;;;;;;;UAMiB,uBAAuB,kBAAkB,UAAU,mBAC1D,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;UCWnC;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D;IACE;IACA;IACA;IACA;IACA;IACA;IACA;;EAEF,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;;;;;UCxBhB;;;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;;;;;;;;;;;;;;EAeA,eAAe;;;EAIf,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;;;EAQX,mBAAmB,uBAAuB,WAAW;;;;;;EAOrD,eAAe,uBAAuB,WAAW;;;;EAKjD,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;UAGI,0BAA0B,kBAAkB,UAAU;EACrE;;;;EAIA;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;;;;;EAKT,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;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgMtB,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;;;KC1mB5B,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-DMsxDp1u.d.ts","names":[],"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":";;;;;;;;UAMiB,uBAAuB,kBAAkB,UAAU,mBAC1D,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;UCWnC;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D;IACE;IACA;IACA;IACA;IACA;IACA;IACA;;EAEF,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;;;;;UCxBhB;;;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;;;;;;;;;;;;;;EAeA,eAAe;;;EAIf,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;;;EAQX,mBAAmB,uBAAuB,WAAW;;;;;;EAOrD,eAAe,uBAAuB,WAAW;;;;EAKjD,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;UAGI,0BAA0B,kBAAkB,UAAU;EACrE;;;;EAIA;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;;;;;EAKT,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;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgMtB,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;;;KC1mB5B,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-CJG7LD9M.js → define-agent-eval-Dzidv34q.js}

@@ -10,7 +10,7 @@ import "./query-D1nLIKt7.js";
 import { r as observedSplitScore } from "./reward-nw2xZGZG.js";
 import { c as validateRunRecord, i as modelHasSnapshot } from "./run-record-DQpSf7t-.js";
 import { r as paretoChart } from "./summary-report-B16xy9Kd.js";
-import { C as assertSearchHistoryAdmissionOptions, I as runFinalComparison, L as openAutoPr, R as defaultProductionGate, W as runEval, d as combineComparisonCosts, f as costFromLedgerSummary, h as runImprovementLoop, q as resolveRunDir, u as executeOptimizationMethod } from "./llm-judge-CV80fkYA.js";
+import { C as assertSearchHistoryAdmissionOptions, I as runFinalComparison, L as openAutoPr, R as defaultProductionGate, W as runEval, d as combineComparisonCosts, f as costFromLedgerSummary, h as runImprovementLoop, q as resolveRunDir, u as executeOptimizationMethod } from "./llm-judge-v80Kmu9g.js";
 import { a as campaignCellToRunRecord, i as campaignCellTaskScore, n as campaignCellExecutionEvidence, r as campaignCellJudgeDimensions } from "./run-record-CR63CpHK.js";
 import { H as surfaceContentHash, O as powerPreflight, U as surfaceDispatchRef, W as surfaceHash, et as campaignSplitDigest, f as campaignMeasurementDigest, h as loopProvenanceArgsFromResult, m as emitLoopProvenance, p as canonicalDigest, t as createCampaignEvidenceReceipt } from "./campaign-evidence-D8DBLqLI.js";
 import { C as inMemoryCampaignStorage, S as fsCampaignStorage, x as createRunCostLedger } from "./external-optimizer-subprocess-q3VzlGAO.js";
@@ -1714,4 +1714,4 @@ function requirePositiveInteger(value, field) {
 //#endregion
 export { summarizeExecution as a, analyzeRuns as i, SelfImproveRunError as n, checkCanaries as o, selfImprove as r, defineAgentEval as t };
 
-//# sourceMappingURL=define-agent-eval-CJG7LD9M.js.map
+//# sourceMappingURL=define-agent-eval-Dzidv34q.js.map