@tangle-network/agent-eval 0.133.0 → 0.133.1

package/dist/index.js

@@ -1,8 +1,7 @@
 import { t as __exportAll } from "./rolldown-runtime-8H4AJuhK.js";
-import { _ as observeAll, a as jsonlReviewStore, c as scoreFromEvals, d as runAgentControlLoop, f as stopOnNoProgress, g as noProgressDetector, h as errorStreakDetector, i as inMemoryReviewStore, l as allCriticalPassed, m as subjectiveEval, n as runProposeReviewAsControlLoop, o as runProposeReview, p as stopOnRepeatedAction, r as createLlmReviewer, s as controlRunToRunRecord, t as controlFailureClassFromVerification, u as objectiveEval, v as repeatedActionDetector, y as evaluateActionPolicy } from "./propose-review-control-Bqb7daEJ.js";
+import { _ as observeAll, a as jsonlReviewStore, c as scoreFromEvals, d as runAgentControlLoop, f as stopOnNoProgress, g as noProgressDetector, h as errorStreakDetector, i as inMemoryReviewStore, l as allCriticalPassed, m as subjectiveEval, n as runProposeReviewAsControlLoop, o as runProposeReview, p as stopOnRepeatedAction, r as createLlmReviewer, s as controlRunToRunRecord, t as controlFailureClassFromVerification, u as objectiveEval, v as repeatedActionDetector, y as evaluateActionPolicy } from "./propose-review-control-SQ-n9-We.js";
 import { a as NotFoundError, c as VerificationError, i as JudgeError, n as CaptureIntegrityError, o as ReplayError, r as ConfigError, s as ValidationError, t as AgentEvalError } from "./errors-8YnH8WlF.js";
-import { a as verifyManifest, i as signManifest, n as evaluateHypothesis, r as hashJson, t as canonicalize } from "./pre-registration-DakwTRXk.js";
-import { AGENT_PROFILE_KINDS, AgentProfileCellValidationError, agentProfileCellHashMaterial, agentProfileCellKey, assertRunAgentProfileCell, buildAgentInterfaceProfileCell, buildAgentProfileCell, groupRunsByAgentProfileCell, requireAgentProfileCell, toAgentProfileJson, validateAgentProfileCell, verifyAgentProfileCell } from "./profile-cell.js";
+import { _ as verifyManifest, a as assertRunAgentProfileCell, c as groupRunsByAgentProfileCell, d as validateAgentProfileCell, f as verifyAgentProfileCell, g as signManifest, h as hashJson, i as agentProfileCellKey, l as requireAgentProfileCell, m as evaluateHypothesis, n as AgentProfileCellValidationError, o as buildAgentInterfaceProfileCell, p as canonicalize, r as agentProfileCellHashMaterial, s as buildAgentProfileCell, t as AGENT_PROFILE_KINDS, u as toAgentProfileJson } from "./agent-profile-cell-OhuTee9n.js";
 import { F as makeFinding, I as computeTraceMetrics, L as createChatClient, P as computeFindingId, R as createAnalystAi, a as KNOWLEDGE_GAP_KIND_SPEC, d as renderUpstreamFindings, i as KNOWLEDGE_POISONING_KIND_SPEC, l as createTraceAnalystKind, n as AnalystRegistry, o as IMPROVEMENT_KIND_SPEC, r as DEFAULT_TRACE_ANALYST_KINDS, s as FAILURE_MODE_KIND_SPEC, t as buildDefaultAnalystRegistry, u as renderPriorFindings } from "./default-registry-D3T9XbuY.js";
 import { _ as resolveModelPricing, a as CostLedgerPersistenceError, c as costForTokenPricing, d as MODEL_PRICING, f as MetricsCollector, g as isModelPriced, h as estimateTokens, i as CostLedger, l as costForUsage, m as estimateCost, n as CostCallConflictError, o as CostReceiptCaptureError, p as TokenCounter, r as CostCeilingReachedError, s as CostReservationExceededError, t as CostAccountingIncompleteError, u as modelPriceKey } from "./cost-ledger-BrJxbrMy.js";
 import { a as providerFromBaseUrl, i as defaultProviderRedactor, n as InMemoryRawProviderSink, r as NoopRawProviderSink, t as FileSystemRawProviderSink } from "./raw-provider-sink-BQd7mzyT.js";
@@ -12,34 +11,34 @@ import { S as traceSpanKindToOpenInferenceKind, a as SpanNotFoundError, b as cla
 import { n as aggregateRunScore, r as clamp01, t as DEFAULT_RUN_SCORE_WEIGHTS } from "./run-score-iEEAWiBY.js";
 import { n as mapConcurrent, t as Mutex } from "./concurrency-MUjT7VjM.js";
 import { a as RunCritic, d as defaultIsMaterial, f as diffFindings, i as runSemanticConceptJudge, n as SEMANTIC_CONCEPT_JUDGE_VERSION, o as SKILL_USAGE_ANALYST, p as LockedJsonlAppender, r as createSemanticConceptJudge, s as SkillUsageAnalyst, t as DEFAULT_COMPLEXITY_WEIGHTS, u as FindingsStore } from "./semantic-concept-judge-BypLt6Fw.js";
-import { At as dominates, B as surfaceContentHash, Bt as JudgeParseError, Ct as llmJudge, D as runCanaries, Dt as fileVerdictCache, Et as contentHash, Ft as BackendIntegrityError, G as DEFAULT_MUTATION_PRIMITIVES, It as assertRealAgentReceipts, K as buildReflectionPrompt, Lt as assertRealBackend, Mt as paretoFrontierWithCrowding, Nt as scalarScore, Ot as inMemoryVerdictCache, Rt as summarizeAgentReceiptIntegrity, St as hashScenarios, Tt as canonicalJson, _t as redTeamReport, bt as Dataset, dt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, ft as REFERENCE_EQUIVALENCE_JUDGE_VERSION, gt as redTeamDataset, ht as DEFAULT_RED_TEAM_CORPUS, jt as paretoFrontier, kt as crowdingDistance, mt as runReferenceEquivalenceJudge, pt as createReferenceEquivalenceJudge, q as parseReflectionResponse, vt as scoreRedTeamOutput, wt as cachedJudge, xt as HoldoutLockedError, yt as toolNamesForRun, zt as summarizeBackendIntegrity } from "./skillopt-optimization-method-DgN7U9iR.js";
-import { A as spearmanR, B as verbosityBias, C as pairedTTest, D as ranks, E as pearsonR, F as calibrateJudge, I as calibrateJudgeContinuous, L as continuousAgreement, M as weightedMean, N as wilcoxonSignedRank, O as requiredPairedSampleSize, P as wilson, R as positionalBias, S as pairedSignTest, T as passAtK, _ as normalizeScores, a as confidenceInterval, b as pairedMde, c as eProcess, d as interpretCliffs, f as mannWhitneyU, g as mulberry32, h as mcnemarRequiredN, i as cohensD, j as weightedComposite, k as requiredSampleSize, l as holm, m as mcnemarPower, n as bonferroni, o as corpusInterRaterAgreement, p as mcnemar, r as cliffsDelta, s as corpusInterRaterAgreementFromJudgeScores, t as benjaminiHochberg, u as interRaterReliability, v as pairedBootstrap, w as partialCredit, x as pairedRiskDifference, y as pairedCohensDz, z as selfPreference } from "./statistics-CnnxdpOg.js";
-import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-D9D0wXj2.js";
+import { At as dominates, B as surfaceContentHash, Bt as JudgeParseError, Ct as llmJudge, D as runCanaries, Dt as fileVerdictCache, Et as contentHash, Ft as BackendIntegrityError, G as DEFAULT_MUTATION_PRIMITIVES, It as assertRealAgentReceipts, K as buildReflectionPrompt, Lt as assertRealBackend, Mt as paretoFrontierWithCrowding, Nt as scalarScore, Ot as inMemoryVerdictCache, Rt as summarizeAgentReceiptIntegrity, St as hashScenarios, Tt as canonicalJson, _t as redTeamReport, bt as Dataset, dt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, ft as REFERENCE_EQUIVALENCE_JUDGE_VERSION, gt as redTeamDataset, ht as DEFAULT_RED_TEAM_CORPUS, jt as paretoFrontier, kt as crowdingDistance, mt as runReferenceEquivalenceJudge, pt as createReferenceEquivalenceJudge, q as parseReflectionResponse, vt as scoreRedTeamOutput, wt as cachedJudge, xt as HoldoutLockedError, yt as toolNamesForRun, zt as summarizeBackendIntegrity } from "./skillopt-optimization-method-BoIzh7Dl.js";
+import { A as spearmanR, B as positionalBias, C as pairedTTest, D as ranks, E as pearsonR, H as verbosityBias, L as calibrateJudge, M as weightedMean, N as wilcoxonSignedRank, O as requiredPairedSampleSize, P as wilson, R as calibrateJudgeContinuous, S as pairedSignTest, T as passAtK, V as selfPreference, _ as normalizeScores, a as confidenceInterval, b as pairedMde, c as eProcess, d as interpretCliffs, f as mannWhitneyU, g as mulberry32, h as mcnemarRequiredN, i as cohensD, j as weightedComposite, k as requiredSampleSize, l as holm, m as mcnemarPower, n as bonferroni, o as corpusInterRaterAgreement, p as mcnemar, r as cliffsDelta, s as corpusInterRaterAgreementFromJudgeScores, t as benjaminiHochberg, u as interRaterReliability, v as pairedBootstrap, w as partialCredit, x as pairedRiskDifference, y as pairedCohensDz, z as continuousAgreement } from "./statistics-DWM_AyLe.js";
+import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-6XItKzd1.js";
 import { n as llmSpanFromProvider, t as TraceEmitter } from "./emitter-CPBAhxum.js";
 import { a as scoreOrigin, n as observedScore, o as trainingReward, r as observedSplitScore, s as trainingScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
 import { a as isRetrievalSpan, i as isLlmSpan, n as TRACE_SCHEMA_VERSION, o as isSandboxSpan, r as isJudgeSpan, s as isToolSpan, t as FAILURE_CLASSES } from "./schema-CRhEY1SO.js";
-import { a as roundTripRunRecord, i as parseRunRecordSafe, n as isRunRecord, o as runTaskScore, r as modelHasSnapshot, s as validateRunRecord, t as RunRecordValidationError } from "./run-record-CN8Zd21B.js";
-import { a as evaluateReleaseConfidence, i as assertReleaseConfidence, n as bootstrapCi, r as judgeReplayGate, t as renderReleaseReport } from "./release-report-oWt9f2k-.js";
+import { a as roundTripRunRecord, i as parseRunRecordSafe, n as isRunRecord, o as runTaskScore, r as modelHasSnapshot, s as validateRunRecord, t as RunRecordValidationError } from "./run-record-BIwU2wdV.js";
+import { a as evaluateReleaseConfidence, i as assertReleaseConfidence, n as bootstrapCi, r as judgeReplayGate, t as renderReleaseReport } from "./release-report-wuilQkvK.js";
 import { c as assertMintedLines, f as isRolloutLine, i as ROLLOUT_SCHEMA, m as validateRolloutLine, p as isTrainableSplit, s as assertMinted, u as assertRolloutLine } from "./schema-C6DW4ZHR.js";
 import { i as toRewardRows, r as toJsonl, s as toSftRows } from "./exporters-q9iL-2Jf.js";
-import { a as toHarborTrajectories, i as relabelImportedSplit, n as HARBOR_IMPORT_GAP, o as toHarborTrajectory, r as fromHarborTrajectory, t as ATIF_SCHEMA_VERSION } from "./rollout-VEo41J0N.js";
+import { a as toHarborTrajectories, i as relabelImportedSplit, n as HARBOR_IMPORT_GAP, o as toHarborTrajectory, r as fromHarborTrajectory, t as ATIF_SCHEMA_VERSION } from "./rollout-CeTlDrf6.js";
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
-import { t as mintRolloutRows } from "./mint-D5_87M5L.js";
+import { t as mintRolloutRows } from "./mint-BvkwcYZU.js";
 import { D as showMeasured, E as isUnavailable, S as rollupSupervisorRuns, T as SUPERVISOR_RUN_SCHEMA, _ as claudeCodeSupervisorRunReader, f as renderSupervisorRunHeadline, l as writeSupervisorRunReport, n as analyzeSupervisorRun, p as renderSupervisorRunMarkdown, t as supervisorRunRolloutLines, v as readClaudeCodeSupervisorRun, y as analyzeSupervisorRunSources } from "./supervisor-run-_lnTLM3z.js";
 import { n as TRACE_ANALYST_ACTOR_DESCRIPTION, r as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, t as analyzeTraces } from "./analyst-LsnNpSkm.js";
-import { C as planTraceInsightQuestions, E as traceAnalystOnRunComplete, S as inferDomainKeywords, T as tokenizeDomainWords, _ as buildTraceInsightContext, a as convertTraceStoresToOtlp, b as describeTraceInsightScope, c as otelRunCompleteHook, d as captureFetchToRawSink, f as otlpRowsToRunRecords, g as flattenOtlpExportToNdjson, h as otlpToTraceRunRecords, i as iterateRawCalls, l as OTEL_AGENT_EVAL_SCOPE, m as otlpToRunRecords, n as ReplayCacheMissError, o as createOtelExporter, p as otlpRowsToTraceRunRecords, r as createReplayFetch, s as createOtelTracingStore, t as ReplayCache, u as exportRunAsOtlp, v as buildTraceInsightPrompt, w as scoreTraceInsightReadiness, x as domainEvidencePattern, y as defaultTraceInsightPanel } from "./replay-BMR4TEYY.js";
+import { C as planTraceInsightQuestions, E as traceAnalystOnRunComplete, S as inferDomainKeywords, T as tokenizeDomainWords, _ as buildTraceInsightContext, a as convertTraceStoresToOtlp, b as describeTraceInsightScope, c as otelRunCompleteHook, d as captureFetchToRawSink, f as otlpRowsToRunRecords, g as flattenOtlpExportToNdjson, h as otlpToTraceRunRecords, i as iterateRawCalls, l as OTEL_AGENT_EVAL_SCOPE, m as otlpToRunRecords, n as ReplayCacheMissError, o as createOtelExporter, p as otlpRowsToTraceRunRecords, r as createReplayFetch, s as createOtelTracingStore, t as ReplayCache, u as exportRunAsOtlp, v as buildTraceInsightPrompt, w as scoreTraceInsightReadiness, x as domainEvidencePattern, y as defaultTraceInsightPanel } from "./replay-CJfGLdx4.js";
 import { n as extractUsageFromResponse, r as extractUsageFromSse, t as extractUsage } from "./extract-usage-2j25whHw.js";
-import { B as harnessAxisOf, F as HARNESS_NATIVE_MODEL, G as parseCorrectnessResponse, H as completionVerdict, I as agentProfileHash, K as verifyCompletion, L as agentProfileId, P as CODING_HARNESSES, R as agentProfileModelId, U as createLlmCorrectnessChecker, V as extractProducedState, W as createTokenRecallChecker, z as expandProfileAxes } from "./campaign-CyQMCCEU.js";
+import { B as harnessAxisOf, F as HARNESS_NATIVE_MODEL, G as parseCorrectnessResponse, H as completionVerdict, I as agentProfileHash, K as verifyCompletion, L as agentProfileId, P as CODING_HARNESSES, R as agentProfileModelId, U as createLlmCorrectnessChecker, V as extractProducedState, W as createTokenRecallChecker, z as expandProfileAxes } from "./campaign-CnzHQndg.js";
 import { a as judgeSpans, c as runsForScenario, i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, r as groupBy, s as runFailureClass, t as aggregateLlm } from "./query-Di7eEQ79.js";
-import { a as checkBehavioralCanary, i as canaryLeakView, o as checkCanaries, r as HoldoutAuditor, s as runBehavioralCanaries, t as analyzeRuns } from "./analyze-runs-BjPn_fOS.js";
-import { a as summaryTable, i as researchReport, n as gainHistogram, r as paretoChart, t as RESEARCH_REPORT_HARD_PAIR_FLOOR } from "./summary-report-BNs5nmXI.js";
+import { a as checkBehavioralCanary, i as canaryLeakView, o as checkCanaries, r as HoldoutAuditor, s as runBehavioralCanaries, t as analyzeRuns } from "./analyze-runs-B-afTpCv.js";
+import { a as summaryTable, i as researchReport, n as gainHistogram, r as paretoChart, t as RESEARCH_REPORT_HARD_PAIR_FLOOR } from "./summary-report-Ci17nIdU.js";
 import { a as composeParsers, c as vitestTestParser, i as SubprocessSandboxDriver, n as DockerSandboxDriver, o as jestTestParser, r as SandboxHarness, s as pytestTestParser, t as runTestGradedScenario } from "./test-graded-scenario-BsqWLmPt.js";
 import { a as InMemoryTraceStore, i as FileSystemTraceStore, n as assertRunCaptured, r as throwIfRunIncomplete, t as RunIntegrityError } from "./integrity-BzRbCHzi.js";
 import { i as redactValue, n as REDACTION_VERSION, r as redactString, t as DEFAULT_REDACTION_RULES } from "./redact-7Aq1ukl-.js";
-import { a as DEFAULT_RULES, i as computeToolUseMetrics, n as iqr, o as classifyFailure, r as welchsTTest, t as compareToBaseline } from "./baseline-HsBvw_dk.js";
+import { a as DEFAULT_RULES, i as computeToolUseMetrics, n as iqr, o as classifyFailure, r as welchsTTest, t as compareToBaseline } from "./baseline-DcX5hQDv.js";
 import { t as analyzeSeries } from "./series-convergence-CjO2QdRW.js";
-import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-DncrVrSr.js";
-import { t as runEvalCampaign } from "./eval-campaign-CYkUME2T.js";
+import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-BU7P6PCW.js";
+import { t as runEvalCampaign } from "./eval-campaign-DXhpZghy.js";
 import { n as pairedEvalueSequence, t as evaluateInterimReleaseConfidence } from "./sequential-Br0mAPHA.js";
 import { AxGEPA, AxSignature, ax } from "@ax-llm/ax";
 import { accessSync, appendFileSync, constants, cpSync, existsSync, mkdirSync, promises, readFileSync, readdirSync, statSync, writeFileSync } from "node:fs";