@tangle-network/agent-eval 0.171.0 → 0.172.0

package/dist/cli.js

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import { i as runRolloutReleaseCli } from "./hf-dataset-D8_RNIis.js";
 import { n as LlmClient } from "./llm-client-BFMRpmqb.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-D8k3Gf0J.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-DoFcisuM.js";
 import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-CCEnywOR.js";
 import { writeFileSync } from "node:fs";
 //#region src/cli-config.ts

package/dist/{client-CDtcZ3p9.d.ts → client-Df7wdslk.d.ts}

@@ -1,5 +1,5 @@
 import { l as RunTerminalOutcome } from "./run-record-DQjRcYwA.js";
-import { _ as GateDecision, j as MutableSurface } from "./types-JHMOqZI4.js";
+import { _ as GateDecision, j as MutableSurface } from "./types-nokrtr7M.js";
 import { o as InsightReport } from "./insight-report-08F022xN.js";
 //#region src/hosted/types.d.ts
 declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
@@ -199,4 +199,4 @@ declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
 }): HostedClient | undefined;
 //#endregion
 export { UnixNanoTimestamp as _, hostedTenantFromEnv as a, EvalRunGenerationSnapshot as c, HostedIngestHeaders as d, HostedWireVersion as f, TraceSpanEvent as g, IngestTracesRequest as h, hostedClientFromEnv as i, EvalRunStatus as l, IngestResponse as m, HostedTenant as n, EvalRunCellScore as o, IngestEvalRunsRequest as p, createHostedClient as r, EvalRunEvent as s, HostedClient as t, HOSTED_WIRE_VERSION as u };
-//# sourceMappingURL=client-CDtcZ3p9.d.ts.map
+//# sourceMappingURL=client-Df7wdslk.d.ts.map

package/dist/{client-CDtcZ3p9.d.ts.map → client-Df7wdslk.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"client-CDtcZ3p9.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
+{"version":3,"file":"client-Df7wdslk.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}

package/dist/contract/index.d.ts

@@ -1,15 +1,15 @@
 import { c as CostLedgerHandle } from "../cost-ledger-DbQdN3nO.js";
 import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-DMoNFDWi.js";
-import { An as runEval, B as OptimizationMethodResult, F as OptimizationMethod, Gn as fsCampaignStorage, I as OptimizationMethodComparison, K as compareOptimizationMethods, Kn as inMemoryCampaignStorage, L as OptimizationMethodInput, Mn as CampaignCellRetryPolicy, N as CompareOptimizationMethodsOptions, Nn as RunCampaignOptions, P as ComparisonCost, Pn as runCampaign, U as OptimizationPackageSource, Un as CampaignStorage, W as OptimizationTokenUsage, at as transientDispatchFailure, ct as runImprovementLoop, et as LlmJudgeDimension, jn as CampaignCellFailureReceipt, kn as RunEvalOptions, nt as llmJudge, ot as RunImprovementLoopOptions, st as RunImprovementLoopResult, tt as LlmJudgeOptions, z as OptimizationMethodProvenance } from "../provenance-CafMdZKM.js";
+import { An as runEval, B as OptimizationMethodResult, F as OptimizationMethod, Gn as CampaignStorage, I as OptimizationMethodComparison, Jn as inMemoryCampaignStorage, K as compareOptimizationMethods, L as OptimizationMethodInput, Mn as CampaignCellRetryPolicy, N as CompareOptimizationMethodsOptions, Nn as RunCampaignOptions, P as ComparisonCost, Pn as runCampaign, U as OptimizationPackageSource, W as OptimizationTokenUsage, at as transientDispatchFailure, ct as runImprovementLoop, et as LlmJudgeDimension, jn as CampaignCellFailureReceipt, kn as RunEvalOptions, nt as llmJudge, ot as RunImprovementLoopOptions, qn as fsCampaignStorage, st as RunImprovementLoopResult, tt as LlmJudgeOptions, z as OptimizationMethodProvenance } from "../provenance-Dp-vvyrU.js";
 import { _ as CreateChatClientOpts, p as ChatClient, x as createChatClient } from "../types-Bfk0uxRj.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-JHMOqZI4.js";
-import { A as ReferenceEquivalenceJudgeInput, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as runReferenceEquivalenceJudge, M as ReferenceEquivalenceJudgeResult, N as ReferenceEquivalenceScenario, O as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as createReferenceEquivalenceJudge, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as GepaRunnerCommand, a as DefaultProductionRewardHackingOptions, b as OptimizerModelBudget, c as SkillOptOptimizationMethodConfig, d as skillOptOptimizationMethod, f as GepaAdaptiveEngineRun, g as GepaOptimizationRecipe, h as GepaOptimizationMethodConfig, i as DefaultProductionGateOptions, j as ReferenceEquivalenceJudgeOptions, k as REFERENCE_EQUIVALENCE_JUDGE_VERSION, l as SkillOptRunnerCommand, m as GepaEngineRun, n as heldOutGate, o as defaultProductionGate, p as GepaEngineOptions, r as DefaultProductionGateCheck, s as composeGate, t as HeldOutGateOptions, u as SkillOptTrainerConfig, v as gepaOptimizationMethod, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as OpenAICompatibleOptimizerModel, z as campaignSplitDigest } from "../heldout-gate-Dh2b62w8.js";
-import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-WSXtBgBb.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-nokrtr7M.js";
+import { A as ReferenceEquivalenceJudgeInput, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as runReferenceEquivalenceJudge, M as ReferenceEquivalenceJudgeResult, N as ReferenceEquivalenceScenario, O as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as createReferenceEquivalenceJudge, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as GepaRunnerCommand, a as DefaultProductionRewardHackingOptions, b as OptimizerModelBudget, c as SkillOptOptimizationMethodConfig, d as skillOptOptimizationMethod, f as GepaAdaptiveEngineRun, g as GepaOptimizationRecipe, h as GepaOptimizationMethodConfig, i as DefaultProductionGateOptions, j as ReferenceEquivalenceJudgeOptions, k as REFERENCE_EQUIVALENCE_JUDGE_VERSION, l as SkillOptRunnerCommand, m as GepaEngineRun, n as heldOutGate, o as defaultProductionGate, p as GepaEngineOptions, r as DefaultProductionGateCheck, s as composeGate, t as HeldOutGateOptions, u as SkillOptTrainerConfig, v as gepaOptimizationMethod, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as OpenAICompatibleOptimizerModel, z as campaignSplitDigest } from "../heldout-gate-JgNRDZwZ.js";
+import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-BBBcz5_3.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-08F022xN.js";
-import { n as HostedTenant } from "../client-CDtcZ3p9.js";
-import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-Cjy2yhqP.js";
-import { $ as FromRunRecordDirOptions, A as observeCodeAgentSession, At as verifyCandidateExperimentComparison, B as parseAgentTrace, C as CodeAgentSessionActionKind, Ct as evaluatePairedMeasurements, D as CodeAgentSessionObservation, Dt as sealCandidateBenchmarkTask, E as CodeAgentSessionExecutionReceipt, Et as sealCandidateBenchmarkSuite, F as AgentTraceIndex, G as EvalRunDiff, H as EvalCellScoreDelta, I as AgentTraceRange, J as diffRuns, K as diffGenerations, L as AgentTraceRecord, M as AgentTraceContributorType, N as AgentTraceConversation, O as CodeAgentSessionSource, Ot as sealCandidateExperiment, P as AgentTraceFile, Q as evalReportingSuite, R as AuthoringProvenance, S as CodeAgentSessionAction, St as SealCandidateBenchmarkSuiteOptions, T as CodeAgentSessionActionSurface, Tt as runCandidateExperiment, U as EvalDimensionDelta, V as partitionRunsByAuthoringModel, W as EvalGenerationDiff, X as EvalReportingSuiteOptions, Y as EvalReportingSuiteInput, Z as EvalReportingSuiteResult, _ as fromOpenCodeSession, _t as EvaluatePairedMeasurementsOptions, a as FromFeedbackTableOptions, at as CompareAgentProfileImprovementExperimentOptions, b as parseCodeAgentJsonlFile, bt as PairedMeasurementEvaluation, c as CodeAgentJsonlLine, ct as measuredComparisonFromAgentProfileImprovementExperiment, d as CodeAgentSessionIntakeResult, dt as sealAgentProfileImprovementSuite, et as FromRunRecordDirResult, f as CodeAgentSessionMetrics, ft as sealAgentProfileImprovementTask, g as fromKimiCodeSession, gt as CompareCandidateExperimentOptions, h as fromCodexSession, ht as CandidateExperimentRun, i as FeedbackTableRow, it as AgentProfileImprovementExperimentRun, j as AgentTraceContributor, k as CodeAgentSessionTerminalStatus, kt as verifyCandidateExperiment, l as CodeAgentSessionDiagnostic, lt as runAgentProfileImprovementExperiment, m as fromClaudeCodeSession, mt as CandidateExperimentExecutionInput, n as fromOtelSpans, nt as fromRunRecordDir, o as FromFeedbackTableResult, ot as RunAgentProfileImprovementExperimentOptions, p as ParsedCodeAgentJsonl, pt as verifyAgentProfileImprovementExperimentComparison, q as diffRunBaselineToWinner, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentExecutionInput, s as fromFeedbackTable, st as SealAgentProfileImprovementSuiteOptions, t as FromOtelSpansOptions, tt as RunRecordRejection, u as CodeAgentSessionIntakeOptions, ut as sealAgentProfileImprovementExperiment, v as fromPiSession, vt as PairedMeasurement, w as CodeAgentSessionActionStatus, wt as measuredComparisonFromCandidateExperiment, x as streamCodeAgentJsonlFile, xt as RunCandidateExperimentOptions, y as parseCodeAgentJsonl, yt as PairedMeasurementAdapter, z as PartitionByAuthoringModelResult } from "../index-fNXZMCzX.js";
-import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-B0s2zU-s.js";
-import { _ as AnalyzeRunsOptions, b as analyzeRuns, v as ExecutionReport, x as summarizeExecution, y as SummarizeExecutionOptions } from "../engine-CAmTUk52.js";
+import { n as HostedTenant } from "../client-Df7wdslk.js";
+import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-0wW7gFhr.js";
+import { $ as FromRunRecordDirOptions, A as observeCodeAgentSession, At as verifyCandidateExperimentComparison, B as parseAgentTrace, C as CodeAgentSessionActionKind, Ct as evaluatePairedMeasurements, D as CodeAgentSessionObservation, Dt as sealCandidateBenchmarkTask, E as CodeAgentSessionExecutionReceipt, Et as sealCandidateBenchmarkSuite, F as AgentTraceIndex, G as EvalRunDiff, H as EvalCellScoreDelta, I as AgentTraceRange, J as diffRuns, K as diffGenerations, L as AgentTraceRecord, M as AgentTraceContributorType, N as AgentTraceConversation, O as CodeAgentSessionSource, Ot as sealCandidateExperiment, P as AgentTraceFile, Q as evalReportingSuite, R as AuthoringProvenance, S as CodeAgentSessionAction, St as SealCandidateBenchmarkSuiteOptions, T as CodeAgentSessionActionSurface, Tt as runCandidateExperiment, U as EvalDimensionDelta, V as partitionRunsByAuthoringModel, W as EvalGenerationDiff, X as EvalReportingSuiteOptions, Y as EvalReportingSuiteInput, Z as EvalReportingSuiteResult, _ as fromOpenCodeSession, _t as EvaluatePairedMeasurementsOptions, a as FromFeedbackTableOptions, at as CompareAgentProfileImprovementExperimentOptions, b as parseCodeAgentJsonlFile, bt as PairedMeasurementEvaluation, c as CodeAgentJsonlLine, ct as measuredComparisonFromAgentProfileImprovementExperiment, d as CodeAgentSessionIntakeResult, dt as sealAgentProfileImprovementSuite, et as FromRunRecordDirResult, f as CodeAgentSessionMetrics, ft as sealAgentProfileImprovementTask, g as fromKimiCodeSession, gt as CompareCandidateExperimentOptions, h as fromCodexSession, ht as CandidateExperimentRun, i as FeedbackTableRow, it as AgentProfileImprovementExperimentRun, j as AgentTraceContributor, k as CodeAgentSessionTerminalStatus, kt as verifyCandidateExperiment, l as CodeAgentSessionDiagnostic, lt as runAgentProfileImprovementExperiment, m as fromClaudeCodeSession, mt as CandidateExperimentExecutionInput, n as fromOtelSpans, nt as fromRunRecordDir, o as FromFeedbackTableResult, ot as RunAgentProfileImprovementExperimentOptions, p as ParsedCodeAgentJsonl, pt as verifyAgentProfileImprovementExperimentComparison, q as diffRunBaselineToWinner, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentExecutionInput, s as fromFeedbackTable, st as SealAgentProfileImprovementSuiteOptions, t as FromOtelSpansOptions, tt as RunRecordRejection, u as CodeAgentSessionIntakeOptions, ut as sealAgentProfileImprovementExperiment, v as fromPiSession, vt as PairedMeasurement, w as CodeAgentSessionActionStatus, wt as measuredComparisonFromCandidateExperiment, x as streamCodeAgentJsonlFile, xt as RunCandidateExperimentOptions, y as parseCodeAgentJsonl, yt as PairedMeasurementAdapter, z as PartitionByAuthoringModelResult } from "../index-DnglhM0A.js";
+import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-XxedTLwu.js";
+import { _ as AnalyzeRunsOptions, b as analyzeRuns, v as ExecutionReport, x as summarizeExecution, y as SummarizeExecutionOptions } from "../engine-BfRay1qD.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 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 SelfImproveOptions, type SelfImproveProgressEvent, 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, 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-Dy8QgxAI.js";
+import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-jS8xj_Q_.js";
 import { a as parseRunRecordSafe } from "../run-record-DLORoL7t.js";
-import { O as compareOptimizationMethods, Z as defaultProductionGate, g as runImprovementLoop, i as transientDispatchFailure, it as runCampaign, mt as campaignSplitDigest, rt as runEval, t as llmJudge } from "../llm-judge-BtJ2Sfk_.js";
+import { O as compareOptimizationMethods, Z as defaultProductionGate, g as runImprovementLoop, i as transientDispatchFailure, it as runCampaign, mt as campaignSplitDigest, rt as runEval, t as llmJudge } from "../llm-judge-aQHIk5_-.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-Qv-cpptD.js";
 import { i as CostLedger } from "../cost-ledger-B1qx30B4.js";
 import { d as mapConcurrentRange } from "../ledger-core-PIfjCbKn.js";
@@ -11,12 +11,12 @@ import { C as inMemoryCampaignStorage, S as fsCampaignStorage } from "../externa
 import { a as heldoutSignificance, s as decidePairedPromotion } from "../power-preflight-CFXm0Vjo.js";
 import { i as makeProposalFinding } from "../types-CiWITkGo.js";
 import { G as isOtlpModelCall, W as classifyOtlpSpanRole } from "../kind-factory-DMeEoMQZ.js";
-import { n as buildDefaultAnalystRegistry, t as createChatClient } from "../chat-client-DEtybj5i.js";
+import { n as buildDefaultAnalystRegistry, t as createChatClient } from "../chat-client-DI79OPye.js";
 import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
 import { t as extractUsage } from "../extract-usage-BrQ8mCLX.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-CBGtLS_H.js";
-import { c as externalTextOptimizationMethod, d as REFERENCE_EQUIVALENCE_INPUT_LIMITS, f as REFERENCE_EQUIVALENCE_JUDGE_VERSION, m as runReferenceEquivalenceJudge, n as gepaOptimizationMethod, o as heldOutGate, p as createReferenceEquivalenceJudge, s as composeGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-DzlF2RM7.js";
+import { c as externalTextOptimizationMethod, d as REFERENCE_EQUIVALENCE_INPUT_LIMITS, f as REFERENCE_EQUIVALENCE_JUDGE_VERSION, m as runReferenceEquivalenceJudge, n as gepaOptimizationMethod, o as heldOutGate, p as createReferenceEquivalenceJudge, s as composeGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-LHi02MzH.js";
 import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-LY9mVQ7W.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/{default-registry-B0s2zU-s.d.ts → default-registry-XxedTLwu.d.ts}

@@ -1,8 +1,8 @@
 import { i as AnalystFinding, n as AnalystContext, w as TraceAnalysisStore } from "./types-DMoNFDWi.js";
 import { l as ExactCapableAnalyst } from "./exact-types-BEecmnWm.js";
 import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-xEb_xfns.js";
-import { c as RawAnalystFinding, i as TraceAnalystLimits, r as TraceAnalysisEngineResult, t as TraceAnalysisEngine } from "./engine-CAmTUk52.js";
-import { t as TraceToolGroupName } from "./tool-groups-RGYfVWpc.js";
+import { c as RawAnalystFinding, i as TraceAnalystLimits, r as TraceAnalysisEngineResult, t as TraceAnalysisEngine } from "./engine-BfRay1qD.js";
+import { t as TraceToolGroupName } from "./tool-groups-DjwlMBvW.js";
 //#region src/trace-analyst/behavioral-metrics.d.ts
 type SuboptimalCode = 'monotonic-input-growth' | 'output-length-decay' | 'single-tool-dependency' | 'no-self-verification';
 interface SuboptimalSignal {
@@ -110,4 +110,4 @@ interface DefaultAnalystRegistryOptions {
 declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
 //#endregion
 export { createTraceAnalyst as a, BehavioralAnalystOptions as c, TraceAnalystDefinition as i, behavioralAnalyst as l, buildDefaultAnalystRegistry as n, renderPriorFindings as o, CreateTraceAnalystOptions as r, runTraceAnalyst as s, DefaultAnalystRegistryOptions as t, deriveEfficiencyFindings as u };
-//# sourceMappingURL=default-registry-B0s2zU-s.d.ts.map
+//# sourceMappingURL=default-registry-XxedTLwu.d.ts.map

package/dist/{default-registry-B0s2zU-s.d.ts.map → default-registry-XxedTLwu.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"default-registry-B0s2zU-s.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;UC1BR;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,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;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;;;UCpZ1B;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
+{"version":3,"file":"default-registry-XxedTLwu.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;UC1BR;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,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;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;;;UCpZ1B;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}

package/dist/{define-agent-eval-Cjy2yhqP.d.ts → define-agent-eval-0wW7gFhr.d.ts}

@@ -1,10 +1,10 @@
 import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "./cost-ledger-DbQdN3nO.js";
 import { _ as ProposalFinding } from "./types-DMoNFDWi.js";
-import { B as OptimizationMethodResult, F as OptimizationMethod, Mn as CampaignCellRetryPolicy, Un as CampaignStorage, jt as SearchHistoryReceipt, kn as RunEvalOptions, l as LoopProvenanceRecord, lt as PremeasuredOptimizationBaseline, st as RunImprovementLoopResult, ut as RunOptimizationOptions, z as OptimizationMethodProvenance } from "./provenance-CafMdZKM.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 } from "./types-JHMOqZI4.js";
+import { B as OptimizationMethodResult, F as OptimizationMethod, Gn as CampaignStorage, Mn as CampaignCellRetryPolicy, jt as SearchHistoryReceipt, kn as RunEvalOptions, l as LoopProvenanceRecord, lt as PremeasuredOptimizationBaseline, st as RunImprovementLoopResult, ut as RunOptimizationOptions, z as OptimizationMethodProvenance } from "./provenance-Dp-vvyrU.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 } from "./types-nokrtr7M.js";
 import { t as PowerPreflight } from "./power-preflight-Ptse_Kq7.js";
 import { o as InsightReport } from "./insight-report-08F022xN.js";
-import { n as HostedTenant } from "./client-CDtcZ3p9.js";
+import { n as HostedTenant } from "./client-Df7wdslk.js";
 //#region src/contract/self-improve.d.ts
 interface SelfImproveBudget {
   /** Hard spend cap across the full run. Each paid call reserves its enforced
@@ -404,4 +404,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
 declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
 //#endregion
 export { DefinedAgentEval as a, SelfImproveOptions as c, SelfImproveRunError as d, selfImprove as f, DefineAgentEvalOptions as i, SelfImproveProgressEvent as l, AgentEvalEvaluateOptions as n, defineAgentEval as o, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveResult as u };
-//# sourceMappingURL=define-agent-eval-Cjy2yhqP.d.ts.map
+//# sourceMappingURL=define-agent-eval-0wW7gFhr.d.ts.map

package/dist/{define-agent-eval-Cjy2yhqP.d.ts.map → define-agent-eval-0wW7gFhr.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"define-agent-eval-Cjy2yhqP.d.ts","names":[],"sources":["../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;UAiEiB;;;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;;;;;;;;;;;;;;EAc9D,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;EAaT,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;;;;;EAMxB,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;;UAGlC,kBAAkB,kBAAkB,UAAU;;;;EAI7D;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;;;cAI9B,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAuMhB,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KCvkB5B,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-0wW7gFhr.d.ts","names":[],"sources":["../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;UAiEiB;;;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;;;;;;;;;;;;;;EAc9D,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;EAaT,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;;;;;EAMxB,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;;UAGlC,kBAAkB,kBAAkB,UAAU;;;;EAI7D;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;;;cAI9B,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAuMhB,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KCvkB5B,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-Dy8QgxAI.js → define-agent-eval-jS8xj_Q_.js}

@@ -10,7 +10,7 @@ import "./query-BPGMVlbM.js";
 import { r as observedSplitScore } from "./reward-nw2xZGZG.js";
 import { c as validateRunRecord, i as modelHasSnapshot } from "./run-record-DLORoL7t.js";
 import { r as paretoChart } from "./summary-report-Bgh8CpNK.js";
-import { E as assertOptimizationResult, F as surfaceContentHash, I as surfaceHash, Z as defaultProductionGate, d as loopProvenanceArgsFromResult, g as runImprovementLoop, ot as resolveRunDir, rt as runEval, u as emitLoopProvenance } from "./llm-judge-BtJ2Sfk_.js";
+import { E as assertOptimizationResult, F as surfaceContentHash, I as surfaceHash, Z as defaultProductionGate, d as loopProvenanceArgsFromResult, g as runImprovementLoop, ot as resolveRunDir, rt as runEval, u as emitLoopProvenance } from "./llm-judge-aQHIk5_-.js";
 import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "./reward-hacking-SkxYgT0x.js";
 import { C as inMemoryCampaignStorage, S as fsCampaignStorage, x as createRunCostLedger } from "./external-optimizer-subprocess-Cex8Da2i.js";
 import { t as powerPreflight } from "./power-preflight-CFXm0Vjo.js";
@@ -1509,4 +1509,4 @@ 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-Dy8QgxAI.js.map
+//# sourceMappingURL=define-agent-eval-jS8xj_Q_.js.map