@tangle-network/agent-eval 0.216.0 → 0.216.1

package/dist/index.d.ts

@@ -1,52 +1,143 @@
+import { _ as PendingCostCallView, a as CostLedgerHandle, c as CostLedgerSummary, d as CostReceiptInput, f as CostUsage, g as PendingCostCall, h as PaidCallResult, i as CostLedgerFilter, l as CostProvenance, m as MaximumCharge, n as ChannelRollup, o as CostLedgerOptions, p as CustomTokenPricing, r as CostChannel, s as CostLedgerPersistence, t as CostLedger, u as CostReceipt, v as RunPaidCallInput } from "./cost-ledger-Jk62OBfF.js";
 import { a as JudgeError, c as ValidationError, i as ConfigError, n as AgentEvalErrorCode, s as NotFoundError, t as AgentEvalError } from "./errors-DEE6u6ot.js";
-import { C as PendingCostCall, D as costForUsage, E as costForTokenPricing, O as modelPriceKey, S as PaidCallResult, T as RunPaidCallInput, _ as CostReservationExceededError, a as CostChannel, b as CustomTokenPricing, c as CostLedgerHandle, d as CostLedgerPersistenceError, f as CostLedgerSummary, g as CostReceiptInput, h as CostReceiptCaptureError, i as CostCeilingReachedError, l as CostLedgerOptions, m as CostReceipt, n as CostAccountingIncompleteError, o as CostLedger, p as CostProvenance, r as CostCallConflictError, s as CostLedgerFilter, t as ChannelRollup, u as CostLedgerPersistence, v as CostResult, w as PendingCostCallView, x as MaximumCharge, y as CostUsage } from "./cost-ledger-CO7UyEcQ.js";
 import { S as verifyAgentProfileCell, _ as buildAgentProfileCell, b as toAgentProfileJson, i as AgentProfileCellInput, l as AgentProfileJson, m as agentProfileCellKey, p as AgentProfileSourceInput, r as AgentProfileCell, s as AgentProfileDimensionValue, t as AGENT_PROFILE_KINDS, v as groupRunsByAgentProfileCell } from "./agent-profile-cell-BNp-Js9F.js";
 import { C as ToolSpan, D as isToolSpan, E as isLlmSpan, T as isJudgeSpan, a as FAILURE_CLASSES, c as JudgeSpan, f as Run, g as RunStatus, l as LlmSpan, n as BudgetLedgerEntry, o as FailureClass, p as RunCaptureReport, r as BudgetSpec, s as GenericSpan, t as Artifact, v as Span, w as TraceEvent } from "./schema-2g7trUwH.js";
-import { _ as runTaskScore, a as RunRecord, c as RunSplitTag, d as RunTokenUsage, f as RunTraceRef, g as parseRunRecordSafe, h as modelHasSnapshot, i as RunOutcome, l as RunTaskFailure, m as isRunRecord, n as RunCostProvenance, o as RunRecordValidationError, p as UNKNOWN_MODEL, r as RunJudgeMetadata, s as RunSearchCoordinates, t as JudgeScoresRecord, u as RunTerminalOutcome, v as searchCellRunId, y as validateRunRecord } from "./run-record-D1Hw-Jml.js";
-import { $ as DatasetOverview, C as makeFinding, Q as DEFAULT_TRACE_ANALYST_BUDGETS, S as computeFindingId, W as createBoundedTraceAnalysisStore, X as TraceAnalysisStore, a as AnalystFinding, at as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, b as ProposalFinding, ct as TraceAnalystSpan, d as AnalystRunResult, dt as TraceAnalystTraceSummary, et as ErrorCluster, f as AnalystRunSummary, ft as ViewSpansResult, h as EvidenceRef, it as SpanMatchRecord, l as AnalystRunEvent, lt as TraceAnalystSpanKind, m as AnalystUsageReceipt, mt as ViewTraceResult, n as Analyst, nt as SearchSpanResult, ot as TraceAnalystByteBudgets, p as AnalystSeverity, pt as ViewTraceOversized, r as AnalystContext, rt as SearchTraceResult, st as TraceAnalystFilters, tt as QueryTracesPage, u as AnalystRunInputs, ut as TraceAnalystSpanStatus, w as makeProposalFinding } from "./types-oKRhQAde.js";
-import { $ as NoopRawProviderSink, A as createChatClient, B as LlmToolCall, C as ChatResponse, Ct as judgeFamily, D as MockTransportOpts, E as CustomTransportOpts, F as LlmChargeBounds, G as costReceiptFromLlmError, H as LlmToolDefinition, I as LlmMessage, J as stripFencedJson, K as isTransientLlmError, L as LlmResponseError, M as LlmCallMetadata, N as LlmCallRequest, O as OpenAiCompatibleTransportOpts, P as LlmCallResult, S as ChatRequest, St as assertCrossFamily, T as CreateChatClientOpts, U as LlmUsage, V as LlmToolChoice, W as costReceiptFromLlm, Y as FileSystemRawProviderSink, Z as InMemoryRawProviderSink, _ as ScenarioResult, _t as checkServedModel, a as CompletionCriterion, b as ChatCallOpts, bt as CrossFamilyError, c as JudgeInput, ct as AssertServedModelOptions, d as PersonaConfig, dt as ServedModelCheck, f as ProductClientConfig, ft as ServedModelPolicy, g as ScenarioFile, gt as assertServedModels, h as Scenario, ht as assertServedModel, i as CollectedArtifacts, j as LlmCallError, k as SandboxSdkTransportOpts, l as JudgeRubric, lt as ModelSubstitutionError, m as RubricDimension, mt as assertCrossFamilyServed, n as ArtifactResult, nt as RawProviderEvent, o as DriverState, p as RouteMap, pt as ServedModelVerdict, q as maximumChargeForLlmRequest, r as CheckResult, rt as RawProviderSink, s as JudgeFn, st as AssertCrossFamilyServedOptions, t as ArtifactCheck, u as JudgeScore, ut as ServedCrossFamilyError, v as Turn, vt as servedModelAcceptable, w as ChatTransport, x as ChatClient, xt as JudgeFamily, y as TurnResult, yt as AssertCrossFamilyOptions, z as LlmTokenLogprob } from "./types-DBZ7keha.js";
+import { _ as runTaskScore, a as RunRecord, c as RunSplitTag, d as RunTokenUsage, f as RunTraceRef, g as parseRunRecordSafe, h as modelHasSnapshot, i as RunOutcome, l as RunTaskFailure, m as isRunRecord, n as RunCostProvenance, o as RunRecordValidationError, p as UNKNOWN_MODEL, r as RunJudgeMetadata, s as RunSearchCoordinates, t as JudgeScoresRecord, u as RunTerminalOutcome, v as searchCellRunId, y as validateRunRecord } from "./run-record-C1E24G2A.js";
+import { $ as DatasetOverview, C as makeFinding, Q as DEFAULT_TRACE_ANALYST_BUDGETS, S as computeFindingId, W as createBoundedTraceAnalysisStore, X as TraceAnalysisStore, a as AnalystFinding, at as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, b as ProposalFinding, ct as TraceAnalystSpan, d as AnalystRunResult, dt as TraceAnalystTraceSummary, et as ErrorCluster, f as AnalystRunSummary, ft as ViewSpansResult, h as EvidenceRef, it as SpanMatchRecord, l as AnalystRunEvent, lt as TraceAnalystSpanKind, m as AnalystUsageReceipt, mt as ViewTraceResult, n as Analyst, nt as SearchSpanResult, ot as TraceAnalystByteBudgets, p as AnalystSeverity, pt as ViewTraceOversized, r as AnalystContext, rt as SearchTraceResult, st as TraceAnalystFilters, tt as QueryTracesPage, u as AnalystRunInputs, ut as TraceAnalystSpanStatus, w as makeProposalFinding } from "./types-C1tnkEgM.js";
+import { A as createChatClient, B as LlmToolDefinition, C as ChatResponse, D as MockTransportOpts, E as CustomTransportOpts, F as LlmMessage, H as FileSystemRawProviderSink, K as NoopRawProviderSink, L as LlmTokenLogprob, M as LlmCallRequest, N as LlmCallResult, O as OpenAiCompatibleTransportOpts, P as LlmChargeBounds, R as LlmToolCall, S as ChatRequest, T as CreateChatClientOpts, V as LlmUsage, W as InMemoryRawProviderSink, X as RawProviderSink, Y as RawProviderEvent, _ as ScenarioResult, a as CompletionCriterion, at as ServedModelPolicy, b as ChatCallOpts, c as JudgeInput, ct as assertServedModel, d as PersonaConfig, dt as servedModelAcceptable, et as AssertCrossFamilyServedOptions, f as ProductClientConfig, ft as AssertCrossFamilyOptions, g as ScenarioFile, gt as judgeFamily, h as Scenario, ht as assertCrossFamily, i as CollectedArtifacts, it as ServedModelCheck, j as LlmCallMetadata, k as SandboxSdkTransportOpts, l as JudgeRubric, lt as assertServedModels, m as RubricDimension, mt as JudgeFamily, n as ArtifactResult, nt as ModelSubstitutionError, o as DriverState, ot as ServedModelVerdict, p as RouteMap, pt as CrossFamilyError, r as CheckResult, rt as ServedCrossFamilyError, s as JudgeFn, st as assertCrossFamilyServed, t as ArtifactCheck, tt as AssertServedModelOptions, u as JudgeScore, ut as checkServedModel, v as Turn, w as ChatTransport, x as ChatClient, y as TurnResult, z as LlmToolChoice } from "./types-C02qtnG8.js";
 import { a as ContinuousCalibrationResult, d as calibrateJudgeContinuous, f as continuousAgreement, h as verbosityBias, i as ContinuousAgreementOptions, l as VerbosityBiasResult, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, t as CalibrationResult, u as calibrateJudge } from "./judge-calibration-DFtEMlde.js";
-import { G as CorpusAgreementPerDimension, H as SeriesConvergenceResult, J as corpusInterRaterAgreement, K as CorpusAgreementReport, U as analyzeSeries, V as SeriesConvergenceOptions, W as CorpusAgreementOptions, X as interRaterReliability, Y as corpusInterRaterAgreementFromJudgeScores, _ as assertJudgeMayGate, a as JudgeCalibrationSet, b as registerCalibrationSet, c as JudgeGateDecision, d as JudgeGateRefusedError, f as JudgeGateStatus, g as addCalibrationVerdicts, h as OwnerVerdict, i as JudgeAgreement, l as JudgeGateInput, m as JudgeRunMeta, n as CalibrationMerge, o as JudgeConfusion, p as JudgeRun, q as CorpusScoreRecord, r as DEFAULT_JUDGE_GATE_POLICY, s as JudgeExampleAgreement, t as CalibrationExample, u as JudgeGatePolicy, v as judgeGateDecision, x as snapshotFromJudgeAgreement, y as measureJudgeAgreement } from "./judge-gate-DNcUrRFA.js";
+import { G as CorpusAgreementPerDimension, H as SeriesConvergenceResult, J as corpusInterRaterAgreement, K as CorpusAgreementReport, U as analyzeSeries, V as SeriesConvergenceOptions, W as CorpusAgreementOptions, X as interRaterReliability, Y as corpusInterRaterAgreementFromJudgeScores, _ as assertJudgeMayGate, a as JudgeCalibrationSet, b as registerCalibrationSet, c as JudgeGateDecision, d as JudgeGateRefusedError, f as JudgeGateStatus, g as addCalibrationVerdicts, h as OwnerVerdict, i as JudgeAgreement, l as JudgeGateInput, m as JudgeRunMeta, n as CalibrationMerge, o as JudgeConfusion, p as JudgeRun, q as CorpusScoreRecord, r as DEFAULT_JUDGE_GATE_POLICY, s as JudgeExampleAgreement, t as CalibrationExample, u as JudgeGatePolicy, v as judgeGateDecision, x as snapshotFromJudgeAgreement, y as measureJudgeAgreement } from "./judge-gate-CE7CQAgt.js";
 import { a as partialCredit, c as spearmanR, d as weightedMean, i as confidenceInterval, l as summarizeNumberSeries, n as WeightedCompositeInput, o as pearsonR, r as WeightedCompositeResult, s as ranks, t as SeriesDistribution, u as weightedComposite } from "./descriptive-B2iPaT9J.js";
 import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-DIWHvysC.js";
-import { A as PairArmsOptions, At as EProcess, Bt as isBinaryOutcomeVector, C as hashJson, D as ComparePairedArmsOptions, F as PairedCorrectness, Ft as ExactRiskDifferenceResult, Gt as pairedRiskDifferenceScore, Ht as pairedBinaryScale, I as PairedMetricDelta, It as McNemarResult, Kt as passAtK, L as comparePairedArms, Lt as ProportionInterval, M as PairRunRecordsResult, Mt as EProcessState, N as PairedArmRow, Nt as EProcessStep, O as MatchedPair, P as PairedArmsComparison, Pt as eProcess, R as pairArms, Rt as RiskDifferenceResult, Ut as pairedRiskDifference, Vt as mcnemar, Wt as pairedRiskDifferenceExact, _t as AttestationProvenance, bt as attest, et as Objective, gt as ATTESTATION_ALGORITHM, j as PairArmsResult, jt as EProcessOptions, k as MatchedRunRecordPair, nt as dominates, qt as wilson, rt as paretoFrontier, tt as ParetoResult, vt as AttestationVerification, w as manifestContentDigest, xt as verifyAttestation, yt as AttestedReport, z as pairRunRecords, zt as ScoreRiskDifferenceResult } from "./statistical-heldout-BeR4bOdg.js";
+import { A as PairArmsOptions, At as EProcess, Bt as isBinaryOutcomeVector, C as hashJson, D as ComparePairedArmsOptions, F as PairedCorrectness, Ft as ExactRiskDifferenceResult, Gt as pairedRiskDifferenceScore, Ht as pairedBinaryScale, I as PairedMetricDelta, It as McNemarResult, Kt as passAtK, L as comparePairedArms, Lt as ProportionInterval, M as PairRunRecordsResult, Mt as EProcessState, N as PairedArmRow, Nt as EProcessStep, O as MatchedPair, P as PairedArmsComparison, Pt as eProcess, R as pairArms, Rt as RiskDifferenceResult, Ut as pairedRiskDifference, Vt as mcnemar, Wt as pairedRiskDifferenceExact, _t as AttestationProvenance, bt as attest, et as Objective, gt as ATTESTATION_ALGORITHM, j as PairArmsResult, jt as EProcessOptions, k as MatchedRunRecordPair, nt as dominates, qt as wilson, rt as paretoFrontier, tt as ParetoResult, vt as AttestationVerification, w as manifestContentDigest, xt as verifyAttestation, yt as AttestedReport, z as pairRunRecords, zt as ScoreRiskDifferenceResult } from "./statistical-heldout-CYpGaoDH.js";
 import { _ as pairedSignTest, a as PairedPromotionDecision, c as BOOTSTRAP_GATE_MIN_N, d as PairedBootstrapResult, f as PairedSignTestResult, g as pairedDeltaTieFraction, h as pairedBootstrap, i as PairedMcNemarEvidence, l as DECISION_PAIRED_DELTA_STATISTIC, m as SignTestAlternative, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, p as PairedTTestResult, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod, u as PairedBootstrapOptions, v as pairedTTest } from "./paired-promotion-decision-DPKz8ZnF.js";
 import { a as requiredPairedSampleSize, i as pairedMde, n as mcnemarPower, o as requiredSampleSize, r as mcnemarRequiredN, t as mulberry32 } from "./random-DGEyr3zG.js";
-import { C as WilcoxonSignedRankResult, S as RankTestOptions, T as wilcoxonSignedRank, _ as bootstrapCi, a as ReleaseConfidenceIssue, b as RankTestMethod, f as evaluateReleaseConfidence, g as Verdict, i as ReleaseConfidenceInput, m as BootstrapResult, o as ReleaseConfidenceMetrics, p as BootstrapOptions, s as ReleaseConfidenceScorecard, t as ActionableSideInfo, u as ReleaseTraceEvidence, w as mannWhitneyU, x as RankTestMethodRequest, y as MannWhitneyResult } from "./release-confidence-BK2QCB5M.js";
+import { C as WilcoxonSignedRankResult, S as RankTestOptions, T as wilcoxonSignedRank, a as bootstrapCi, b as RankTestMethod, d as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseTraceEvidence, i as Verdict, n as BootstrapResult, p as ReleaseConfidenceScorecard, s as ActionableSideInfo, t as BootstrapOptions, u as ReleaseConfidenceInput, v as evaluateReleaseConfidence, w as mannWhitneyU, x as RankTestMethodRequest, y as MannWhitneyResult } from "./promotion-gate-BawrdPYr.js";
+import { Ar as DEFAULT_RED_TEAM_CORPUS, At as RequirementCheck, Br as CanaryKind, Ct as extractProducedState, Dt as LlmCorrectnessCheckerOpts, Et as CorrectnessChecker, Fr as redTeamDataset, Ft as verifyCompletion, Hr as CanaryReport, Ir as redTeamReport, Ji as llmJudge, Lr as scoreRedTeamOutput, Mr as RedTeamCategory, Mt as TaskGold, Nr as RedTeamFinding, Nt as completionVerdict, Ot as ProducedProposal, Pr as RedTeamReport, Pt as createLlmCorrectnessChecker, Qi as transientDispatchFailure, Rr as CanaryAlert, St as ToolCallEventLike, Tt as CompletionVerdict, Ur as runCanaries, Vr as CanaryOptions, Zi as quotaExhaustedUntil, _t as assertRealBackend, bt as ProposalEventLike, da as CampaignCellFailureReceipt, fa as CampaignCellRetryPolicy, gt as BackendIntegrityReport, ht as BackendIntegrityError, jr as RedTeamCase, jt as SatisfiedBy, kt as ProducedState, ma as runCampaign, pa as RunCampaignOptions, qi as LlmJudgeOptions, vt as summarizeBackendIntegrity, wt as CompletionRequirement, xt as RuntimeEventLike, yt as ArtifactEventLike, zr as CanaryEvaluation } from "./index-CyszSg4O.js";
+import { R as Scenario$1, S as JudgeConfig, _ as GateDecision, a as CampaignResult, w as JudgeScore$1 } from "./types-Cq8mSbq8.js";
 import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-C4avFjIi.js";
 import { a as MintedRolloutLine } from "./schema-C-ziyJD7.js";
-import { a as mintRolloutRows, c as ScorePreference, i as RolloutSearchLineage, n as MintRolloutResult, t as MintRolloutOptions } from "./mint-DCGoBYhB.js";
-import { R as Scenario$1, S as JudgeConfig, _ as GateDecision, a as CampaignResult, w as JudgeScore$1 } from "./types-B4C5nFw6.js";
-import { Ar as DEFAULT_RED_TEAM_CORPUS, At as RequirementCheck, Br as CanaryKind, Ct as extractProducedState, Dt as LlmCorrectnessCheckerOpts, Et as CorrectnessChecker, Fr as redTeamDataset, Ft as verifyCompletion, Hr as CanaryReport, Ir as redTeamReport, Ji as llmJudge, Lr as scoreRedTeamOutput, Mr as RedTeamCategory, Mt as TaskGold, Nr as RedTeamFinding, Nt as completionVerdict, Ot as ProducedProposal, Pr as RedTeamReport, Pt as createLlmCorrectnessChecker, Qi as transientDispatchFailure, Rr as CanaryAlert, St as ToolCallEventLike, Tt as CompletionVerdict, Ur as runCanaries, Vr as CanaryOptions, Zi as quotaExhaustedUntil, _a as runCampaign, _t as assertRealBackend, bt as ProposalEventLike, ga as RunCampaignOptions, gt as BackendIntegrityReport, ha as CampaignCellRetryPolicy, ht as BackendIntegrityError, jr as RedTeamCase, jt as SatisfiedBy, kt as ProducedState, ma as CampaignCellFailureReceipt, qi as LlmJudgeOptions, vt as summarizeBackendIntegrity, wt as CompletionRequirement, xt as RuntimeEventLike, yt as ArtifactEventLike, zr as CanaryEvaluation } from "./index-NvGBiO8H.js";
-import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-CIA2Atnt.js";
+import { n as ScorePreference } from "./reward-CpCmAM8L.js";
+import { a as mintRolloutRows, i as RolloutSearchLineage, n as MintRolloutResult, t as MintRolloutOptions } from "./mint-DNU8QPG4.js";
+import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-VFcQncbW.js";
 import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
 import { C as defineEquivalenceCheck, S as buildEquivalenceRecord, _ as StrategyChecker, a as CheckerIdentity, b as VerificationStrategyProfile, c as EquivalenceCheckDefinition, d as EquivalenceCheckerInput, f as EquivalenceCheckerResult, g as EquivalenceRecord, h as EquivalenceProtocolError, i as equivalenceVerdict, l as EquivalenceCheckSpec, m as EquivalenceObligationStatus, n as VerdictCertification, o as CheckerOutcome, p as EquivalenceObligation, r as certificationEvidenceDigest, s as EquivalenceArm, t as DefaultVerdict, u as EquivalenceChecker, v as VERIFICATION_STRATEGIES, w as runEquivalenceCheck, x as VerificationStrategySource, y as VERIFICATION_STRATEGY_SOURCES } from "./verdict-E4eRNf7-.js";
 import { a as MultiLayerVerifier, c as VerifyOptions, i as LayerStatus, l as gradeSemanticStatus, n as Layer, o as Severity, r as LayerResult, s as VerificationReport, t as Finding } from "./multi-layer-verifier-BUaQ4C17.js";
-import { C as HeldOutGateConfig, S as HeldOutGate, T as SplitCoverage, _ as paretoChart, a as ParetoPoint, b as GateDecision$1, d as ResearchReportOptions, g as gainHistogram, h as SummaryTableRow, m as SummaryTableOptions, n as GainDistributionFigureSpec, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, w as HeldOutGateRejectionCode, x as GateEvidence, y as summaryTable } from "./summary-report-C-CtItCN.js";
+import { C as HeldOutGateConfig, S as HeldOutGate, T as SplitCoverage, _ as gainHistogram, a as ResearchReportOptions, b as GateDecision$1, d as GainDistributionOptions, g as SummaryTableRow, h as SummaryTableOptions, l as GainDistributionBin, m as SummaryTable, p as ParetoPoint, t as ResearchReport, u as GainDistributionFigureSpec, v as paretoChart, w as HeldOutGateRejectionCode, x as GateEvidence, y as summaryTable } from "./summary-report-bdFDXq0b.js";
 import { a as FAILURE_BLAME, c as FailureContext, d as INFRA_FAILURE_BLAMES, f as classifyFailure, i as DEFAULT_FAILURE_REASON_RULES, l as FailureReasonRule, m as failureBlame, o as FailureBlame, p as classifyFailureReason, r as failureClusterView, s as FailureClassification, u as FailureRule } from "./failure-cluster-CbgEDcFJ.js";
-import { d as InsightReport, i as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-CSDAh8A7.js";
-import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-CD0lf5Qd.js";
-import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-JYfPxcVz.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, 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-CEWkX5wZ.js";
-import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-BrWh62zf.js";
-import { i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine, t as DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS } from "./engine-C_zNWfM8.js";
-import { A as SemanticConceptJudgeInput, C as DspyRlmTraceEngineOptions, D as ConceptFinding, E as createChatTraceEngine, M as SemanticConceptJudgeResult, N as runSemanticConceptJudge, O as ConceptSpec, T as ChatTraceEngineOptions, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as SemanticConceptJudgeOptions, k as SEMANTIC_CONCEPT_JUDGE_VERSION, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-oQ4aHziK.js";
-import { i as BenchmarkEvaluation } from "./types-CfZMIzB_.js";
+import { d as InsightReport, n as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-D68fd-su.js";
+import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-ClugMJfI.js";
+import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-D7_hm2fK.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, 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-WSck0CZJ.js";
+import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-DzHiGruW.js";
+import { i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine, t as DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS } from "./engine-3A_4jOlI.js";
+import { A as SemanticConceptJudgeInput, C as DspyRlmTraceEngineOptions, D as ConceptFinding, E as createChatTraceEngine, M as SemanticConceptJudgeResult, N as runSemanticConceptJudge, O as ConceptSpec, T as ChatTraceEngineOptions, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as SemanticConceptJudgeOptions, k as SEMANTIC_CONCEPT_JUDGE_VERSION, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-DV-jSzxF.js";
+import { i as BenchmarkEvaluation } from "./types-Cbe2fWBa.js";
 import { a as SpanHandle, n as CaptureWrite, o as TraceEmitter, s as TraceEmitterOptions, t as CaptureErrorHandler } from "./emitter-HcCmFQl_.js";
 import { n as SandboxDriver } from "./sandbox-harness-BlSOu4LX.js";
-import { S as shareAllowed, b as redactForShare, d as ShareSafetyVerdict, f as assessShareSafety, i as RedactOptions, o as RedactionProfile, r as REDACTION_VERSION, s as RedactionReport, u as ShareSafetyStatus, x as redactText, y as redact } from "./redact-Cveoq6fJ.js";
+import { _ as redactText, d as ShareSafetyVerdict, f as assessShareSafety, g as redactForShare, h as redact, i as RedactOptions, o as RedactionProfile, r as REDACTION_VERSION, s as RedactionReport, u as ShareSafetyStatus, v as shareAllowed } from "./redact-mtCCcT8z.js";
 import { a as PairedEvalueStep, c as SequentialDecision, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, n as InterimReleaseConfidenceInput, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence } from "./sequential-BhsrMupG.js";
+import { _ as WelchTestResult, b as welchsTTest, d as ToolStats, f as ToolUseMetrics, i as toolWasteView, m as computeToolUseMetrics, p as ToolUseOptions, s as judgeAgreementView, u as budgetBreachView, v as WelchTestStatus, y as iqr } from "./tool-waste-BKgvd-TP.js";
+import { A as AnalystFindingDigest, B as ControlRunResult, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlActionOutcome, G as StopDecision, H as ControlRuntimeError, I as ControlBudget, L as ControlContext, M as AnalystRunDigest, N as analystFindingDigest, O as summarizePreferenceMemory, P as analystRunDigest, R as ControlDecision, S as controlRunToFeedbackTrajectory, T as feedbackTrajectoriesToOptimizerRows, U as ControlSeverity, V as ControlRuntimeConfig, W as ControlStep, _ as PreferenceMemoryEntry, a as FeedbackLabel, b as analystRunToReviewRequests, c as FeedbackOptimizerRow, d as FeedbackTask, f as FeedbackTrajectory, g as InMemoryFeedbackTrajectoryStore, h as FileSystemFeedbackTrajectoryStore, i as FeedbackAttempt, j as AnalystReviewDecision, k as withAssignedFeedbackSplit, l as FeedbackOutcome, m as FeedbackTrajectoryStore, n as AnalystReviewRequest, o as FeedbackLabelKind, p as FeedbackTrajectoryFilter, r as FeedbackArtifactType, s as FeedbackLabelSource, t as AnalystFeedbackTrajectoryOptions, u as FeedbackSplitPolicy, v as ProposedSideEffect, w as feedbackTrajectoriesToDatasetScenarios, x as assignFeedbackSplit, y as analystRunToFeedbackTrajectory, z as ControlEvalResult } from "./feedback-trajectory-C4EG0nNH.js";
+import { B as inferDomainKeywords, C as exportRunAsOtlp, F as buildTraceInsightContext, G as scoreTraceInsightReadiness, I as buildTraceInsightPrompt, K as tokenizeDomainWords, M as TraceInsightReadiness, N as TraceInsightSuite, R as describeTraceInsightScope, S as OtlpSpan, X as analyzeTraces, Y as AnalyzeTracesResult, _ as spanRecordsToTraceAnalysisStore, a as captureFetchToRawSink, b as OtlpExport, c as ExtractedUsage, f as extractUsageFromSse, g as otlpTextToTraceAnalysisStore, h as SpanRecordsToTraceAnalysisStoreOptions, n as toolSpansToTraceAnalysisStore, o as OtlpFlatLine, p as OtlpFileTraceStore, u as extractUsage, z as domainEvidencePattern } from "./store-tool-spans-DdCSPa7C.js";
+import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-DAMEDFaC.js";
+import { a as SteeringChange, c as CampaignRunContext, d as CampaignVariant, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, l as CampaignRunOutcome, m as FailedRun, n as ExperimentResult, o as CampaignFactoryParams, p as EvalCampaignResult, r as FailureMode, s as CampaignIntegrityPolicy, t as ExperimentPlan, u as CampaignScenario } from "./researcher-CEcfku-T.js";
+import "./index-BFRO6chs.js";
 import { n as TrajectoryStep, r as buildTrajectory, t as Trajectory } from "./trajectory-Y1fmdlw7.js";
 import { a as runCounterfactual, i as CounterfactualRunner, n as CounterfactualMutation, r as CounterfactualResult, t as CounterfactualContext } from "./counterfactual-CPaWN99-.js";
-import { A as AnalystFindingDigest, B as ControlRunResult, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlActionOutcome, G as StopDecision, H as ControlRuntimeError, I as ControlBudget, J as subjectiveEval, K as objectiveEval, L as ControlContext, M as AnalystRunDigest, N as analystFindingDigest, O as summarizePreferenceMemory, P as analystRunDigest, R as ControlDecision, S as controlRunToFeedbackTrajectory, T as feedbackTrajectoriesToOptimizerRows, U as ControlSeverity, V as ControlRuntimeConfig, W as ControlStep, _ as PreferenceMemoryEntry, a as FeedbackLabel, b as analystRunToReviewRequests, c as FeedbackOptimizerRow, d as FeedbackTask, f as FeedbackTrajectory, g as InMemoryFeedbackTrajectoryStore, h as FileSystemFeedbackTrajectoryStore, i as FeedbackAttempt, j as AnalystReviewDecision, k as withAssignedFeedbackSplit, l as FeedbackOutcome, m as FeedbackTrajectoryStore, n as AnalystReviewRequest, o as FeedbackLabelKind, p as FeedbackTrajectoryFilter, q as runAgentControlLoop, r as FeedbackArtifactType, s as FeedbackLabelSource, t as AnalystFeedbackTrajectoryOptions, u as FeedbackSplitPolicy, v as ProposedSideEffect, w as feedbackTrajectoriesToDatasetScenarios, x as assignFeedbackSplit, y as analystRunToFeedbackTrajectory, z as ControlEvalResult } from "./feedback-trajectory-DkaDj0dj.js";
 import { y as OUTPUT_VALUE } from "./attribute-vocabulary-CcXqHODT.js";
-import { A as scoreTraceInsightReadiness, B as OtlpSpan, C as describeTraceInsightScope, F as analyzeTraces, I as OtlpFlatLine, J as extractUsage, K as ExtractedUsage, P as AnalyzeTracesResult, R as OtlpExport, T as inferDomainKeywords, V as exportRunAsOtlp, W as captureFetchToRawSink, X as extractUsageFromSse, _ as TraceInsightReadiness, a as SpanRecordsToTraceAnalysisStoreOptions, b as buildTraceInsightContext, c as spanRecordsToTraceAnalysisStore, j as tokenizeDomainWords, n as toolSpansToTraceAnalysisStore, r as OtlpFileTraceStore, s as otlpTextToTraceAnalysisStore, v as TraceInsightSuite, w as domainEvidencePattern, x as buildTraceInsightPrompt } from "./store-tool-spans-DR1SrCLK.js";
-import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-CCAxGVK6.js";
 import { d as runsForScenario, i as argHash, s as judgeSpans } from "./query-DK1nWcmp.js";
-import { a as SteeringChange, c as CampaignRunContext, d as CampaignVariant, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, l as CampaignRunOutcome, m as FailedRun, n as ExperimentResult, o as CampaignFactoryParams, p as EvalCampaignResult, r as FailureMode, s as CampaignIntegrityPolicy, t as ExperimentPlan, u as CampaignScenario } from "./researcher-f6mZhh58.js";
-import "./index-CPJJr98o.js";
-import { _ as WelchTestResult, a as ToolStats, b as welchsTTest, c as computeToolUseMetrics, d as judgeAgreementView, i as toolWasteView, m as budgetBreachView, o as ToolUseMetrics, s as ToolUseOptions, v as WelchTestStatus, y as iqr } from "./tool-waste-Wi4g6Ekb.js";
 import { SpanKind } from "@tangle-network/agent-trace-contract";
+//#region src/cost-ledger/errors.d.ts
+export declare class CostCeilingReachedError extends ValidationError {
+  constructor(ceilingUsd: number, committedUsd: number, requestedUsd: number, phase: string, actor: string);
+}
+export declare class CostAccountingIncompleteError extends ValidationError {}
+export declare class CostReservationExceededError extends ValidationError {
+  constructor(actor: string, actualUsd: number, maximumUsd: number);
+}
+export declare class CostCallConflictError extends ValidationError {
+  readonly callId?: string;
+  readonly receipt?: CostReceipt;
+  constructor(message: string, options?: {
+    callId?: string;
+    receipt?: CostReceipt;
+    cause?: unknown;
+  });
+}
+export declare class CostLedgerPersistenceError extends ValidationError {
+  readonly callId?: string;
+  readonly receipt?: CostReceipt;
+  constructor(cause: unknown, callId?: string, receipt?: CostReceipt);
+}
+export declare class CostReceiptCaptureError extends ValidationError {
+  readonly callId: string;
+  readonly receipt?: CostReceipt;
+  readonly receiptError: Error;
+  constructor(callId: string, cause: unknown, receiptError: unknown, receipt?: CostReceipt);
+}
+//#endregion
+//#region src/cost-ledger/pricing.d.ts
+/** Return the canonical pricing-table key, or null when the model is unpriced. */
+export declare function modelPriceKey(model: string): string | null;
+interface CostResult {
+  costUsd: number;
+  costUnknown: boolean;
+}
+export declare function costForUsage(model: string, usage: CostUsage): CostResult;
+/** Price token counts with caller-supplied per-million rates. */
+export declare function costForTokenPricing(pricing: CustomTokenPricing, usage: Pick<CostUsage, 'inputTokens' | 'outputTokens' | 'cachedTokens' | 'cacheWriteTokens'>): number;
+//#endregion
+//#region src/llm-client/accounting.d.ts
+export declare function maximumChargeForLlmRequest(request: Pick<LlmCallRequest, 'model' | 'messages' | 'jsonSchema' | 'tools' | 'toolChoice' | 'maxTokens' | 'thinking'>, options?: LlmChargeBounds): MaximumCharge | undefined;
+/** Convert a provider result into the canonical paid-call receipt input.
+ * The receipt is JSON-clean: absent optional fields are omitted, never carried
+ * as explicit-undefined keys. Receipts cross JSON boundaries (the external
+ * optimizer's loopback proxy validates them with `assertJsonValue`), where an
+ * explicit-undefined value is rejected as non-serializable. */
+export declare function costReceiptFromLlm(result: LlmCallResult, customTokenPricing?: CustomTokenPricing): CostReceiptInput;
+/** Structured-response failures retain their completed provider receipt. */
+export declare function costReceiptFromLlmError(error: Error, customTokenPricing?: CustomTokenPricing): CostReceiptInput | undefined;
+//#endregion
+//#region src/llm-client/errors.d.ts
+export declare class LlmCallError extends AgentEvalError {
+  readonly status: number;
+  readonly body: string;
+  readonly model: string;
+  constructor(message: string, status: number, body: string, model: string);
+}
+/** A provider response completed and incurred measurable usage, but its content
+ *  could not satisfy the caller's response contract. The response envelope is
+ *  retained so accounting can commit the receipt before the error propagates. */
+export declare class LlmResponseError extends AgentEvalError {
+  readonly result: LlmCallResult;
+  constructor(message: string, result: LlmCallResult, options?: {
+    cause?: unknown;
+  });
+}
+//#endregion
+//#region src/llm-client/json.d.ts
+/**
+ * Strip a ```json / ``` code fence if the model emitted one.
+ * Idempotent for naked JSON. Some models (claude-code via router, certain
+ * deepseek models) wrap output even under json_object.
+ */
+export declare function stripFencedJson(raw: string): string;
+//#endregion
+//#region src/llm-client/resilience.d.ts
+/**
+ * True when an error is a transient transport/network fault worth retrying,
+ * as opposed to a deterministic failure (4xx schema reject, JSON parse) that
+ * a retry cannot fix. Inspects `LlmCallError.status`, then the error's
+ * name/message/code, then recurses into `error.cause` — undici nests the
+ * real socket fault one or more levels under `.cause`.
+ *
+ * This is the retry classifier for the package: `callLlm` and
+ * `withJudgeRetry` both route through it, so connection failures are treated
+ * consistently across transports.
+ */
+export declare function isTransientLlmError(err: unknown): boolean;
+//#endregion
 //#region src/statistics/effect-sizes.d.ts
 /**
  * Standardized effect sizes: Cohen's d family and Cliff's delta, with the
@@ -151,694 +242,824 @@ export declare function sampleMedian(values: readonly number[]): number | null;
 /** Descriptive facts, not stability policy or a significance test. */
 export declare function summarizeSample(values: readonly number[], options?: SampleSummaryOptions): SampleSummary | null;
 //#endregion
-//#region src/profile/baselines.d.ts
+//#region src/analyst/policy-edit.d.ts
+type PolicyEditSchemaVersion = 'policy-edit/v1';
+declare const POLICY_EDIT_AXES: readonly ['carrier', 'representation', 'budget', 'sampling', 'output_contract', 'tool_contract', 'routing', 'memory', 'agent_profile', 'deployment_target'];
+type PolicyEditAxis = (typeof POLICY_EDIT_AXES)[number];
+declare const POLICY_EDIT_TARGET_SURFACES: readonly ['prompt', 'tool-contract', 'runtime-config', 'memory', 'agent-profile', 'code', 'deployment'];
+type PolicyEditTargetSurface = (typeof POLICY_EDIT_TARGET_SURFACES)[number];
+type PolicyEditRisk = 'low' | 'medium' | 'high' | 'unknown';
+type PolicyEditGainDirection = 'increase' | 'decrease';
+type PolicyEditGainUnit = 'absolute' | 'relative' | 'percent' | 'score';
+interface PolicyEditTarget {
+  surface: PolicyEditTargetSurface;
+  /** Stable path inside the target surface, for example `system-prompt:tools`
+   * or `budget.maxTurns`. */
+  path?: string;
+  /** Optional canonical deployment identity. Store the existing cell, not a
+   * local profile shape. */
+  agentProfileCell?: AgentProfileCell;
+  /** Human label when the path is not enough for a readable audit trail. */
+  label?: string;
+}
+type PolicyEditChange = {
+  kind: 'text';
+  mode: 'append' | 'prepend' | 'replace';
+  value: string;
+  /** Required when `mode === 'replace'`; exact match only. */
+  find?: string;
+} | {
+  kind: 'json';
+  mode: 'set' | 'merge' | 'remove';
+  path: string;
+  value?: AgentProfileJson;
+};
+interface PolicyEditExpectedGain {
+  /** Metric this edit is expected to move, e.g. `holdout.composite`. */
+  metric: string;
+  direction: PolicyEditGainDirection;
+  /** Positive magnitude in the metric's native units. */
+  amount: number;
+  unit?: PolicyEditGainUnit;
+  rationale?: string;
+}
+interface PolicyEditSource {
+  findingIds: string[];
+  analystIds: string[];
+  evidenceRefs: EvidenceRef[];
+  /** Mirrors `AnalystFinding.derived_from_judge`; admission rejects it. */
+  derivedFromJudge?: boolean;
+}
+interface PolicyEdit {
+  schemaVersion: PolicyEditSchemaVersion;
+  editId: string;
+  axis: PolicyEditAxis;
+  target: PolicyEditTarget;
+  change: PolicyEditChange;
+  claim: string;
+  expectedGain: PolicyEditExpectedGain;
+  confidence: number;
+  risk: PolicyEditRisk;
+  source: PolicyEditSource;
+  rationale?: string;
+  validationPlan?: string;
+  metadata?: Record<string, unknown>;
+}
+export declare const POLICY_EDIT_CANDIDATE_RECORD_SCHEMA: 'tangle.policy-edit-candidate.v1';
+/** JSON-safe attribution carried with a measured candidate and its scores. */
+interface PolicyEditCandidateRecord {
+  schema: typeof POLICY_EDIT_CANDIDATE_RECORD_SCHEMA;
+  policyEdit: PolicyEdit;
+}
+type PolicyEditInit = Omit<PolicyEdit, 'schemaVersion' | 'editId'> & {
+  schemaVersion?: PolicyEditSchemaVersion;
+  editId?: string;
+};
+export declare class PolicyEditValidationError extends ValidationError {
+  readonly path: string;
+  constructor(message: string, path?: string);
+}
+interface FindingToPolicyEditOptions {
+  expectedGain?: PolicyEditExpectedGain | ((finding: AnalystFinding) => PolicyEditExpectedGain | null | undefined);
+  risk?: PolicyEditRisk | ((finding: AnalystFinding) => PolicyEditRisk);
+  defaultAxis?: PolicyEditAxis;
+  defaultTargetSurface?: PolicyEditTargetSurface;
+}
+interface PolicyEditAdmissionOptions {
+  minScore?: number;
+  minExpectedGain?: number;
+  allowHighRisk?: boolean;
+  requireEvidence?: boolean;
+}
+interface PolicyEditAdmission {
+  edit: PolicyEdit;
+  decision: 'admit' | 'reject';
+  score: number;
+  reasons: string[];
+}
+export declare function makePolicyEdit(init: PolicyEditInit): PolicyEdit;
+export declare function computePolicyEditId(edit: Omit<PolicyEdit, 'editId'> | PolicyEdit): string;
+export declare function validatePolicyEdit(input: unknown): PolicyEdit;
+export declare function makePolicyEditCandidateRecord(edit: PolicyEdit): PolicyEditCandidateRecord;
+export declare function validatePolicyEditCandidateRecord(input: unknown): PolicyEditCandidateRecord;
+export declare function policyEditsFromFindings(findings: ReadonlyArray<AnalystFinding>, opts?: FindingToPolicyEditOptions): PolicyEdit[];
+export declare function admitPolicyEdit(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): PolicyEditAdmission;
+export declare function applyPolicyEditToSurface(surface: unknown, edit: PolicyEdit): unknown;
+//#endregion
+//#region src/capability-headroom.d.ts
 /**
- * Strong, generically-useful baseline ROLES — the top zone of an `AgentProfile`
- * before any domain layer. A product composes one of these with its own
- * environment description (its sandbox) and its domain guidance, which lives in
- * the product repo (not here): a profile is `<baseline role> + <environment> +
- * <domain sections>`. Domain strength is NOT generalized into the substrate —
- * it is lifted here only once ≥2 products genuinely reuse it.
+ * Capability-headroom gate — "can this task set even SEE the capability
+ * you're about to A/B?"
  *
- * Three roles cover the common shapes:
- *   - `engineerRole`   — builds + verifies real artifacts in a sandbox. Distilled
- *                        from agent-runtime's `coderProfile` doctrine (minimal
- *                        correct change, run the checks, fix the root cause,
- *                        never weaken a test or hide an error).
- *   - `researcherRole` — gathers from real sources and grounds every claim
- *                        (cite or mark inferred; never fabricate).
- *   - `generalistRole` — a strong default: helpful, grounded, verify-before-assert.
+ * A capability A/B (grant one arm a tool, a skill, extra context — anything)
+ * can only detect the capability on tasks the capability-ABSENT baseline
+ * FAILS. A task the baseline already passes is saturated: the treatment arm
+ * has no headroom to demonstrate anything, and averaging saturated tasks into
+ * the comparison dilutes a real effect toward a false null. Calibrate before
+ * measuring: run the baseline arm first, classify every task's headroom, and
+ * refuse to run the comparison when no task has a gap.
  *
- * All three are verification-first and domain-agnostic. They describe HOW a
- * capable IC operates, not WHAT domain it works in.
+ * Fail-closed on unknown outcomes: a baseline run whose outcome could not be
+ * determined (infra loss, unverified, missing) NEVER counts as a failure —
+ * treating "couldn't tell" as "failed" would manufacture headroom out of
+ * telemetry gaps and launder infra noise into a treatment effect.
+ *
+ * Domain-free by construction: the capability under test never appears here.
+ * Callers decide what the baseline arm lacks; this module only classifies
+ * baseline outcomes per task.
  */
-/** Senior-IC engineer operating in a sandbox — produces real artifacts and
- *  verifies them before declaring done. The shared operator foundation for any
- *  file-producing, tool-using product agent. */
-declare const engineerRole: string;
-/** Researcher who grounds every claim in real sources — gathers, cites, and
- *  distinguishes verified fact from inference. */
-declare const researcherRole: string;
-/** Strong general-purpose default — helpful, grounded, verifies before asserting. */
-declare const generalistRole: string;
-/** The named baseline roles, for selection by key (e.g. from a product config). */
-declare const BASELINE_ROLES: {
-  readonly engineer: string;
-  readonly researcher: string;
-  readonly generalist: string;
-};
-type BaselineRoleKey = keyof typeof BASELINE_ROLES;
-declare namespace index_d_exports {
-  export { AgentProfileSection, BASELINE_ROLES, BaselineRoleKey, ProfileSkill, PromptProfile, applyDomainPatch, baselineProfile, baselineProfileFromRole, engineerRole, generalistRole, prodProfile, profileToSurface, renderProfile, researcherRole, sectionHash };
+/** One baseline-arm (capability-absent) observation of one task. */
+interface HeadroomInput {
+  /** Stable task identity; multiple rows per task are multiple baseline reps. */
+  taskId: string;
+  /** Baseline outcome. 'unknown' = the outcome could not be determined —
+   *  distinct from 'fail' and never counted as one. */
+  baselineOutcome: 'pass' | 'fail' | 'unknown';
 }
-/** A named, addressable region of the system prompt. `evolvable` marks whether
- *  the self-improvement loop is allowed to patch its body; fixed scaffolding
- *  (e.g. a compliance preamble) sets `evolvable: false`. */
-interface AgentProfileSection {
-  id: string;
-  title: string;
-  body: string;
-  evolvable: boolean;
+/** Headroom classification for one task.
+ *  - 'gap': the baseline fails it — the comparison can see the capability here.
+ *  - 'saturated': the baseline already passes it — no headroom.
+ *  - 'unknown': every baseline outcome is unknown — fail-closed, not a gap. */
+type HeadroomClass = 'gap' | 'saturated' | 'unknown';
+interface TaskHeadroom {
+  taskId: string;
+  /** Total baseline observations for this task (known + unknown). */
+  n: number;
+  /** Observations with a known ('pass' | 'fail') outcome — the evidence the
+   *  classification actually rests on. A large `n` with a small `nKnown`
+   *  means the headroom verdict is thin, not well-replicated. */
+  nKnown: number;
+  /** Pass rate over KNOWN outcomes only; NaN when every outcome is unknown
+   *  (no data ≠ measured zero). */
+  baselinePassRate: number;
+  headroom: HeadroomClass;
 }
-/** A skill the agent can invoke. `triggers` are the phrases/conditions that
- *  should activate it; `scriptsDir` points at its executable scripts when the
- *  skill ships code (mirrors Harvey's `skills/<name>/scripts/`). */
-interface ProfileSkill {
-  name: string;
-  description: string;
-  triggers: string[];
-  scriptsDir?: string;
+interface CapabilityHeadroomOptions {
+  /** A task counts as 'gap' when its baseline pass rate is ≤ this. Default 0
+   *  (the baseline must fail every known rep). Must be in [0, 1) — a
+   *  threshold of 1 would classify a fully-saturated task as a gap. */
+  keepThreshold?: number;
 }
-/** The structured system prompt. The first four fields are fixed scaffolding;
- *  `domain` is the evolvable surface the loop optimizes. */
-interface PromptProfile {
-  role: string;
-  environment: string;
-  toolConventions: string;
-  skills: ProfileSkill[];
-  domain: AgentProfileSection[];
+interface CapabilityHeadroomResult {
+  /** Per-task classification, sorted by `taskId`. */
+  tasks: TaskHeadroom[];
+  summary: {
+    tasksWithGap: number;
+    tasksSaturated: number;
+    tasksUnknown: number;
+    /** Total unknown-outcome reps across ALL tasks (including tasks that
+     *  still classified via their known reps) — the size of the telemetry
+     *  hole behind the classifications. */
+    repsUnknown: number;
+  };
 }
 /**
- * Emit the sectioned system prompt in fixed order: role text, then
- * `## Environment`, `## Tool conventions`, `## Skills` (each skill as
- * `### <name>` + description + triggers), `## Domain guidance` (each section as
- * `### <title>` + body). The order is load-bearing — the loop diffs rendered
- * text, and reordering would make every candidate look like a full rewrite.
+ * Classify every task's baseline headroom from capability-absent baseline
+ * outcomes.
+ *
+ * A task is 'gap' when it has ≥ 1 known outcome and its pass rate over known
+ * outcomes is ≤ `keepThreshold`; 'saturated' when the rate exceeds the
+ * threshold; 'unknown' when every outcome is unknown. Unknown reps never
+ * count toward the rate in either direction (fail-closed — see the module
+ * doc), and evidence strength stays visible: each task reports `nKnown`
+ * alongside `n`, and the summary totals the unknown reps in `repsUnknown`.
+ * Throws on empty input, an out-of-range `keepThreshold`, or an
+ * unrecognized outcome value (rows are often projected from untyped records).
  */
-declare function renderProfile(p: PromptProfile): string;
-/** The string `MutableSurface` the self-improvement loop scores — a profile
- *  renders to exactly the text a candidate is graded on. */
-declare function profileToSurface(p: PromptProfile): string;
-/**
- * The fixed-scaffolding baseline: a `role`/`environment`/`toolConventions`
- * profile with the stock skill roster and NO domain guidance. The
- * production profile is this plus shipped domain sections.
- */
-declare function baselineProfile(args: {
-  role: string;
-  environment?: string;
-  toolConventions?: string;
-  skills?: ProfileSkill[];
-}): PromptProfile;
-/**
- * Baseline from one of the strong generic roles (`'engineer' | 'researcher' |
- * 'generalist'`) — the common case: pick a role foundation, optionally override
- * the environment to describe THIS product's sandbox, then layer domain via
- * `prodProfile`. Domain guidance stays in the product repo; this only supplies
- * the generically-useful role + stock scaffolding.
- */
-declare function baselineProfileFromRole(role: BaselineRoleKey, args?: {
-  environment?: string;
-  toolConventions?: string;
-  skills?: ProfileSkill[];
-}): PromptProfile;
-/** The production profile: the baseline scaffolding plus the domain sections
- *  shipped after self-improvement. Differs from the baseline ONLY in `domain`
- *  (and any skills the caller layered into the baseline) — the role,
- *  environment, and tool conventions are carried through unchanged. */
-declare function prodProfile(baseline: PromptProfile, shipped: AgentProfileSection[]): PromptProfile;
+export declare function capabilityHeadroom(rows: readonly HeadroomInput[], opts?: CapabilityHeadroomOptions): CapabilityHeadroomResult;
+interface AssertCapabilityHeadroomOptions {
+  /** Minimum tasks classified 'gap' for the comparison to proceed. Default 1. */
+  minTasksWithGap?: number;
+}
 /**
- * Section-scoped edit — replace the body of ONE domain section by id, leaving
- * every other section byte-identical. This is how the loop targets a single
- * evolvable surface. Throws on an unknown section id (a patch that hits no
- * section is a silent no-op otherwise — fail loud). Throws when the targeted
- * section is `evolvable: false`.
+ * The go/no-go guard: throws when fewer than `minTasksWithGap` tasks have
+ * baseline headroom — i.e. the benchmark cannot see the capability, so any
+ * comparison run on it would measure noise. Call between the baseline
+ * calibration run and the treatment run.
  */
-declare function applyDomainPatch(p: PromptProfile, sectionId: string, newBody: string): PromptProfile;
+export declare function assertCapabilityHeadroom(result: CapabilityHeadroomResult, opts?: AssertCapabilityHeadroomOptions): void;
+//#endregion
+//#region src/cost-tracker.d.ts
 /**
- * Content hash of a single section — its loop identity. Delegates to the
- * campaign provenance `surfaceContentHash` (the same helper the loop's
- * provenance record uses for full surfaces), so a section hash is byte-for-byte
- * comparable with the `sha256:`-prefixed content hashes that record carries.
- * Two sections that render identically (same title + body) share a hash; the
- * `id` and `evolvable` flag are excluded — they are routing metadata, not
- * content.
+ * Cost tracker — token + USD accounting per scenario and per run.
+ *
+ * Adapted from generic usage-event accounting. Every
+ * optimizer needs to know "is the quality gain worth the cost delta?",
+ * and every dashboard needs dollars-per-completed-task. MODEL_PRICING
+ * from metrics.ts stays authoritative for estimate math; this module
+ * adds the aggregation + per-scenario roll-up that was duplicated
+ * across 4 verticals.
  */
-declare function sectionHash(section: AgentProfileSection): string;
-//#endregion
-//#region src/budget-guard.d.ts
-export declare class BudgetBreachError extends AgentEvalError {
-  dimension: keyof BudgetSpec;
-  limit: number;
-  attempted: number;
-  constructor(dimension: keyof BudgetSpec, limit: number, attempted: number);
-}
-export declare class BudgetGuard {
-  private consumed;
-  private emitter;
-  private budget;
-  private startedAt;
-  constructor(emitter: TraceEmitter, budget: BudgetSpec, now?: () => number);
-  /** Record consumption. Throws `BudgetBreachError` if any dimension exceeds its cap. */
-  charge(delta: Partial<Record<keyof BudgetSpec, number>>, spanId?: string): Promise<void>;
-  /** Convenience: advance wall-clock budget based on elapsed wall time. */
-  tickWall(nowMs: number, spanId?: string): Promise<void>;
-  get state(): Record<keyof BudgetSpec, number>;
-}
-//#endregion
-//#region src/run-score.d.ts
-interface RunScore {
-  success: number;
-  goalProgress: number;
-  repoGroundedness: number;
-  driftPenalty: number;
-  toolUseQuality: number;
-  patchQuality: number;
-  testReality: number;
-  finalGate: number;
-  reviewerBlockers: number;
-  costUsd: number;
-  wallSeconds: number;
-  notes?: string[];
-}
-interface RunScoreWeights {
-  success: number;
-  goalProgress: number;
-  repoGroundedness: number;
-  driftPenalty: number;
-  toolUseQuality: number;
-  patchQuality: number;
-  testReality: number;
-  finalGate: number;
-  reviewerBlockers: number;
-  costUsd: number;
-  wallSeconds: number;
-}
-export declare function aggregateRunScore(score: RunScore, weights?: Partial<RunScoreWeights>): number;
-export declare function clamp01(value: number): number;
-//#endregion
-//#region src/scorecard.d.ts
-/** One commit's measurement of one (scenario, profile) cell. */
-interface ScorecardEntry {
-  commitSha: string;
-  /** ISO timestamp the entry was recorded. */
-  timestamp: string;
-  /** Per-seed (or per-rep) scores for this cell at this commit. */
-  scores: number[];
-  /** Median of `scores` — the cell's headline score for the commit. */
-  composite: number;
-  /** Per-dimension means, when the runs carried a judge breakdown. */
-  perDimension?: Record<string, number>;
-  /** RunRecord ids folded into this entry — provenance. */
-  runIds: string[];
-}
-/** A (scenario, profile) cell and its commit-ordered score timeline. */
-interface ScorecardCell {
-  scenarioId: string;
-  profileHash: string;
-  /** Model id — denormalised from the profile for readable filtering. */
-  model: string;
-  timeline: ScorecardEntry[];
-}
-/** The folded scorecard: every cell, plus the profile definitions by hash. */
-interface Scorecard {
-  cells: ScorecardCell[];
-  /** Profile definitions seen — keeps the scorecard self-describing. */
-  profiles: Record<string, AgentProfile>;
+interface TokenSpec {
+  inputTokens: number;
+  outputTokens: number;
+  cachedTokens?: number;
+  reasoningTokens?: number;
 }
-/** One append-only log line — a single cell's entry for a single commit. */
-interface ScorecardLogLine {
+interface CostEntry extends TokenSpec {
   scenarioId: string;
-  profileHash: string;
   model: string;
-  profile: AgentProfile;
-  entry: ScorecardEntry;
-}
-interface RecordRunsOptions {
-  /** The profile that produced these runs — keys the cell. */
-  profile: AgentProfile;
-  commitSha: string;
-  /** Defaults to `new Date().toISOString()`. */
-  timestamp?: string;
+  /** Override estimate with an observed cost (e.g. from provider response). */
+  actualCostUsd?: number;
+  timestamp: number;
+  /** Free-form tags (variant id, round #, etc.). */
+  tags?: Record<string, string>;
 }
-/**
- * Fold a benchmark's `RunRecord[]` into per-cell scorecard log lines — one
- * line per scenario the runs cover. All runs are attributed to the single
- * `profile` in `opts` (the harness ran them under it); the cell key is
- * `(scenarioId, agentProfileHash(profile))`.
- */
-export declare function recordRuns(runs: RunRecord[], opts: RecordRunsOptions): ScorecardLogLine[];
-/** Record runs and append them to the log in one call. Returns the lines. */
-export declare function recordRunsToScorecard(logPath: string, runs: RunRecord[], opts: RecordRunsOptions): ScorecardLogLine[];
-/**
- * Fold the JSONL log into a queryable `Scorecard`. A missing file yields an
- * empty scorecard; a malformed line is skipped — a corrupt append never
- * breaks the read. Each cell's timeline is sorted chronologically.
- */
-export declare function loadScorecard(logPath: string): Scorecard;
-type CellVerdict = 'improved' | 'regressed' | 'flat' | 'new';
-interface ScorecardCellDiff {
+interface ScenarioCost {
   scenarioId: string;
-  profileHash: string;
-  model: string;
-  verdict: CellVerdict;
-  /** Composite of the latest entry. */
-  current: number;
-  /** Composite of the comparison entry — `null` when `verdict === 'new'`. */
-  baseline: number | null;
-  /** `current − baseline` — `null` when new. */
-  delta: number | null;
-  /** Cohen's d of current vs baseline samples — `null` when new or n < 2. */
-  cohensD: number | null;
-  /** Welch's t-test p-value — `null` when new or n < 2. */
-  pValue: number | null;
-  currentCommit: string;
-  baselineCommit: string | null;
-}
-interface ScorecardDiff {
-  cells: ScorecardCellDiff[];
-  summary: {
-    improved: number;
-    regressed: number;
-    flat: number;
-    new: number;
-  };
+  entries: CostEntry[];
+  totalInputTokens: number;
+  totalOutputTokens: number;
+  totalCachedTokens: number;
+  totalCostUsd: number;
+  /** Pass flag — set by consumer via markOutcome; used for cost-per-completed-task. */
+  completed?: boolean;
 }
-interface DiffScorecardOptions {
-  /** Compare each cell against this commit instead of its immediate predecessor. */
-  baselineCommit?: string;
-  /** |Cohen's d| at/above which a move counts as real. Default 0.5. */
-  minEffect?: number;
-  /** p-value at/below which a move is significant. Default 0.05. */
-  maxP?: number;
+export declare class CostTracker {
+  private byScenario;
+  record(entry: Omit<CostEntry, 'timestamp'> & {
+    timestamp?: number;
+  }): CostEntry;
+  markOutcome(scenarioId: string, completed: boolean): void;
   /**
-   * |delta| at/above which a move counts when statistics are unavailable
-   * (a cell with fewer than 2 samples on either side). Default 0.05.
+   * Convenience: record + markOutcome in one call from a
+   * `{ usage, verdict }`-shaped response (starter-foundry's
+   * `invokeMetaJudge` returns this shape; consumers that wrap any
+   * judge/critic can follow the same convention).
+   *
+   * `usage.model` must be present in `MODEL_PRICING` for cost math to
+   * populate; otherwise totalCostUsd stays at 0 for the entry but
+   * tokens still aggregate.
    */
-  minDelta?: number;
+  recordVerdict(verdict: {
+    usage?: {
+      inputTokens: number;
+      outputTokens: number;
+      model: string;
+      cachedTokens?: number;
+      reasoningTokens?: number;
+    };
+    verdict?: 'pass' | 'fail' | 'borderline' | string;
+  }, scenarioId: string, tags?: Record<string, string>): CostEntry | null;
+  get(scenarioId: string): ScenarioCost | undefined;
+  list(): ScenarioCost[];
+  summary(): CostSummary;
+}
+interface CostSummary {
+  scenarioCount: number;
+  completedCount: number;
+  totalInputTokens: number;
+  totalOutputTokens: number;
+  totalCostUsd: number;
+  avgCostPerScenarioUsd: number;
+  /** Total USD / completed scenarios — null when nothing completed. */
+  costPerCompletedTaskUsd: number | null;
 }
-/**
- * Compare the latest entry of every cell against its predecessor (or against
- * `baselineCommit`) and classify the move. A move is `improved`/`regressed`
- * only when it clears both the effect-size and significance gates; otherwise
- * `flat`. Cells with no prior entry are `new`.
- */
-export declare function diffScorecard(scorecard: Scorecard, opts?: DiffScorecardOptions): ScorecardDiff;
-/**
- * Render a scorecard diff as a human-readable report — the block a feature
- * PR prints. Regressions are listed first; flat cells are summarised, not
- * enumerated.
- */
-export declare function formatScorecardDiff(diff: ScorecardDiff): string;
 //#endregion
-//#region src/trace-contracts/types.d.ts
-/**
- * Minimal structural span the checker reads. The eval-side `Span`
- * (trace/schema) and the output of {@link contractSpansFromOtlp} satisfy it.
- */
-interface ContractSpan {
-  spanId?: string;
-  parentSpanId?: string | null;
-  name?: string;
-  /** Eval kinds (`tool`, `llm`, ...) or contract kinds (`TOOL`, `LLM`, ...). */
-  kind?: string;
-  startedAt?: number;
-  endedAt?: number;
-  status?: string;
-  error?: string;
-  /** Typed field on eval-side ToolSpans; OTLP flattenings drop it (see
-   *  {@link contractSpanToolName}). */
-  toolName?: string;
-  /** Typed field on eval-side LlmSpans. */
-  model?: string;
-  inputTokens?: number;
-  outputTokens?: number;
-  /** Typed field on eval-side ToolSpans. */
-  args?: unknown;
-  /** False when the source observed the call but did not capture its arguments. */
-  argsCaptured?: boolean;
-  attributes?: Record<string, unknown>;
+//#region src/detectors/index.d.ts
+type DetectorSeverity = 'info' | 'warn' | 'critical';
+/** A normalized step a detector folds over. Fields are optional; each detector reads only what it
+ *  needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;
+ *  error-streak → `status`). */
+interface DetectorEvent {
+  /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */
+  readonly actionFingerprint?: string;
+  /** Fingerprint of observable state AFTER this step. */
+  readonly stateFingerprint?: string;
+  /** Score after this step (the score-flat half of no-progress). */
+  readonly score?: number;
+  /** Whether this step errored. */
+  readonly status?: 'ok' | 'error';
+  /** Free-form label carried into the signal evidence (e.g. the tool name). */
+  readonly label?: string;
 }
-/**
- * The run a `run` rule reads. The eval-side `Run` satisfies it. When no run is
- * passed, the checker derives it from the trace's single root span; a trace
- * with no unique root leaves the run unknown and every `run` rule fails.
- */
-interface ContractRun {
-  status?: string;
-  startedAt?: number;
-  endedAt?: number;
+interface DetectorSignal {
+  readonly detector: string;
+  readonly severity: DetectorSeverity;
+  readonly failureClass?: FailureClass;
+  readonly reason: string;
+  /** Consecutive matching steps at the moment the signal fired. */
+  readonly streak: number;
+  readonly evidence?: Record<string, unknown>;
 }
-/** JSON-safe RegExp form — what the builder normalizes RegExp matchers to. */
-interface SerializedRegex {
-  $regex: string;
-  flags: string;
+interface StreamingDetector {
+  readonly id: string;
+  /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */
+  readonly streak: number;
+  /** Fold one event; return a signal when the threshold trips, else null. */
+  observe(event: DetectorEvent): DetectorSignal | null;
+  /** Clear all state. */
+  reset(): void;
 }
-/** Matches when the value equals one of the listed strings. */
-interface OneOfMatcher {
-  oneOf: string[];
+interface RepeatedActionOptions {
+  /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).
+   *  `<= 0` disables the signal (the streak is still tracked, for telemetry). */
+  readonly maxRepeated?: number;
+  readonly severity?: DetectorSeverity;
+  /** Failure class to stamp on the signal (default `tool_recovery_failure`). */
+  readonly failureClass?: FailureClass;
 }
-type TextMatcher = string | RegExp | SerializedRegex | OneOfMatcher;
+/** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */
+export declare function repeatedActionDetector(opts?: RepeatedActionOptions): StreamingDetector;
+interface ErrorStreakOptions {
+  /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */
+  readonly maxErrors?: number;
+  readonly severity?: DetectorSeverity;
+  readonly failureClass?: FailureClass;
+}
+/** N consecutive tool errors = the worker is hammering a broken approach. */
+export declare function errorStreakDetector(opts?: ErrorStreakOptions): StreamingDetector;
+/** Fold one event through many detectors at once; returns every signal that fired this step. The
+ *  natural shape for an online pipe watching with a whole panel of detectors. */
+export declare function observeAll(detectors: ReadonlyArray<StreamingDetector>, event: DetectorEvent): DetectorSignal[];
+//#endregion
+//#region src/error-count-extractor.d.ts
 /**
- * Proposition over one span. All specified fields must match (AND); `not`
- * must NOT match. At least one field is required — an empty predicate would
- * match every span and is rejected.
+ * Toolchain error-count extractor.
  *
- * `kind` compares against the span's contract kind (`TOOL`, `LLM`, `AGENT`,
- * ...), resolved by agent-trace-contract's classifier, so eval kinds
- * (`tool`, `llm`) and OTLP spans that declare no kind are read the same way.
+ * Given stderr/stdout from a compiler or test runner, count the number
+ * of reported errors/failures. Patterns are deliberately narrow —
+ * unknown stderr returns `null` rather than zero so callers can
+ * distinguish "no errors" from "different toolchain, couldn't parse".
  *
- * `tool` resolves through {@link contractSpanToolName}; `model` through the
- * typed `model` field and the gen_ai model attributes.
+ * All patterns are anchored to the start of a line and use bounded
+ * character classes to avoid catastrophic backtracking on pathological
+ * inputs.
  *
- * `attr` values match by strict equality, or regex-test when the value is a
- * RegExp/SerializedRegex and the attribute is a string. Structured attribute
- * values need `custom`, or an `argument` rule for tool arguments.
+ * Add new toolchains by appending to {@link ERROR_COUNT_PATTERNS};
+ * order matters only in the sense that the first matching pattern wins.
  */
-interface SpanPredicate {
-  name?: TextMatcher;
-  tool?: TextMatcher;
-  model?: TextMatcher;
-  kind?: SpanKind;
-  attr?: Record<string, unknown>;
-  not?: SpanPredicate;
-  custom?: (span: ContractSpan) => boolean;
-  /** Stamped by the builder when `custom` is present. Survives JSON while
-   *  the function does not, so evaluation of a deserialized contract throws
-   *  instead of silently dropping the check. */
-  requiresCustom?: true;
+interface ErrorCountPattern {
+  /** Stable identifier for logging + tests. */
+  name: string;
+  /** Must be global (`g` flag) — the extractor counts matches. */
+  regex: RegExp;
+  /** Optional post-processing to extract a count from a single captured match. */
+  transform?: (match: RegExpMatchArray) => number;
 }
-declare const ORDER_MODES: readonly ['start-order', 'finish-before-start', 'all-occurrences'];
-type OrderMode = (typeof ORDER_MODES)[number];
-declare const ARGUMENT_OCCURRENCES: readonly ['first', 'last', 'any', 'all'];
-type ArgumentOccurrence = (typeof ARGUMENT_OCCURRENCES)[number];
-declare const ARGUMENT_TYPES: readonly ['string', 'number', 'boolean', 'object', 'array', 'null'];
-type ArgumentType = (typeof ARGUMENT_TYPES)[number];
-type ArgumentCheck = {
-  op: 'exists';
-} | {
-  op: 'equals';
-  value: unknown;
-} | {
-  op: 'oneOf';
-  values: unknown[];
-} | {
-  op: 'type';
-  type: ArgumentType;
-};
-/** A tool whose calls change state outside the run. */
-interface RetryWrite {
-  tool: string;
-  /** RFC 6901 JSON Pointer to the idempotency key in the call's arguments.
-   *  Without it, a repeated call of this tool always fails. */
-  idempotencyKey?: string;
+export declare const ERROR_COUNT_PATTERNS: ErrorCountPattern[];
+interface ExtractOptions {
+  /** Restrict to named patterns — default: all patterns. */
+  only?: string[];
+  /** Additional patterns to consider BEFORE the built-in list. */
+  extra?: ErrorCountPattern[];
 }
-interface RunChecks {
-  /** The run's status is `completed` and its end time is recorded. A failed
-   *  or aborted run fails this check. */
-  requireCompleted?: boolean;
-  /** The run's status must be one of these. */
-  allowedStatuses?: RunStatus[];
-  /** End minus start, in milliseconds. An unknown duration fails. */
-  maxDurationMs?: number;
+interface ExtractResult {
+  /** Total count of matched errors, or null when no pattern matched. */
+  count: number | null;
+  /** Name of the pattern that matched, or null. */
+  matched: string | null;
+  /** Original matches for callers that want to surface specifics. */
+  samples: string[];
 }
-type ContractRule = {
-  kind: 'always' | 'never' | 'eventually';
-  label: string;
-  p: SpanPredicate;
-} | {
-  kind: 'precedes';
-  label: string;
-  a: SpanPredicate;
-  b: SpanPredicate;
-  order?: OrderMode;
-} | {
-  kind: 'neverUnless';
-  label: string;
-  p: SpanPredicate;
-  prior: SpanPredicate;
-  order?: OrderMode;
-} | {
-  kind: 'atMost';
-  label: string;
-  p: SpanPredicate;
-  max: number;
-} | {
-  kind: 'tokensAtMost';
-  label: string;
-  p: SpanPredicate;
-  max: number;
-} | ({
-  kind: 'run';
-  label: string;
-} & RunChecks) | {
-  kind: 'argument';
-  label: string;
-  p: SpanPredicate;
-  /** RFC 6901 JSON Pointer into the call's arguments; `''` is the whole value. */
-  pointer: string;
-  check: ArgumentCheck;
-  /** Which matching calls must satisfy the check. Default `all`. */
-  occurrence?: ArgumentOccurrence;
-} | {
-  /**
-   * The tools the harness offered the model (the OTel GenAI
-   * `gen_ai.tool.definitions` attribute) are recorded, every tool call is
-   * one of them, and, when `declared` is set, every offered tool is
-   * declared. A trace that records no offered tools fails: what the
-   * harness enforced is unknown.
-   */
-  kind: 'toolsOffered';
-  label: string;
-  declared?: string[];
-} | {
-  /**
-   * A tool called again with the same arguments repeats its side effect.
-   * A repeated `reads` call passes; a repeated `writes` call passes only
-   * when every call carries the same idempotency key; a repeated call of
-   * any other tool fails, because its side effect is unknown.
-   */
-  kind: 'retrySafe';
-  label: string;
-  reads?: string[];
-  writes?: RetryWrite[];
-};
-/** One legitimate alternate path. The contract passes only when the base
- *  rules pass and at least one alternative passes completely. */
-interface ContractAlternative {
+/**
+ * Try each pattern in order; return the first with matches.
+ *
+ * Returning `null` (instead of zero) on no-match is deliberate — a
+ * callsite that greps for "typescript errors" on cargo output should
+ * NOT treat that as "zero TS errors" because the toolchain is wrong.
+ */
+export declare function extractErrorCount(text: string, opts?: ExtractOptions): ExtractResult;
+//#endregion
+//#region src/knowledge/types.d.ts
+type KnowledgeRequirementCategory = 'user_specific' | 'company_specific' | 'domain_specific' | 'codebase_specific' | 'market_specific' | 'regulatory' | 'tool_api' | 'credential_or_secret' | 'runtime_environment' | 'preference' | 'historical_context';
+type KnowledgeAcquisitionMode = 'ask_user' | 'search_web' | 'query_connector' | 'inspect_repo' | 'run_command' | 'infer_low_confidence' | 'not_available';
+type KnowledgeImportance = 'blocking' | 'high' | 'medium' | 'low';
+type KnowledgeFreshness = 'static' | 'monthly' | 'weekly' | 'daily' | 'realtime';
+type KnowledgeSensitivity = 'public' | 'private' | 'secret';
+type KnowledgeFallbackPolicy = 'block' | 'ask' | 'continue_with_caveat' | 'use_default';
+interface KnowledgeRequirement {
   id: string;
-  rules: ContractRule[];
-}
-/** Serializable plain object — `traceContract(name)....build()` output. */
-interface TraceContract {
-  name: string;
-  /** What the contract protects, for reports and `explainTraceContract`. */
-  description?: string;
-  /** Evaluate only the span matching `root` and its descendants. Exactly one
-   *  span must match; zero or several make every rule an error. */
-  scope?: {
-    root: SpanPredicate;
-  };
-  rules: ContractRule[];
-  alternatives?: ContractAlternative[];
-}
-type ContractStatus = 'pass' | 'fail' | 'error';
-interface ContractViolation {
-  rule: string;
-  spanId?: string;
-  detail: string;
+  description: string;
+  requiredFor: string[];
+  category: KnowledgeRequirementCategory;
+  acquisitionMode: KnowledgeAcquisitionMode;
+  importance: KnowledgeImportance;
+  freshness: KnowledgeFreshness;
+  sensitivity: KnowledgeSensitivity;
+  confidenceNeeded: number;
+  currentConfidence: number;
+  evidenceIds: string[];
+  fallbackPolicy: KnowledgeFallbackPolicy;
+  /**
+   * ISO timestamp after which this requirement must be treated as stale.
+   * Stale requirements score as missing even when they still have evidence.
+   */
+  validUntil?: string;
+  /** ISO timestamp for the last source-grounding or human verification pass. */
+  lastVerifiedAt?: string;
+  metadata?: Record<string, unknown>;
 }
-/** What happened to one rule. `error` means it could not be evaluated. */
-interface ContractRuleExecution {
-  rule: string;
-  /** Set for rules inside `alternatives`. */
-  alternative?: string;
-  status: ContractStatus;
-  violations: number;
-  error?: string;
+interface KnowledgeBundle {
+  taskId: string;
+  requirements: KnowledgeRequirement[];
+  evidenceIds: string[];
+  claimIds: string[];
+  wikiPageIds: string[];
+  userAnswers: Record<string, string>;
+  missing: KnowledgeRequirement[];
+  readinessScore: number;
+  metadata?: Record<string, unknown>;
 }
-interface ContractVerdict extends DefaultVerdict {
-  /** Contract name — keys this verdict in multi-contract reports. */
-  contract: string;
-  status: ContractStatus;
-  /** `status === 'pass'`. */
-  valid: boolean;
-  /** Fraction of rules that passed, in [0, 1]. An errored rule is not a pass. */
-  score: number;
-  /** 0|1 per evaluated base rule, keyed by label, plus `anyOf` when the
-   *  contract has alternatives. Errored rules are absent, never 0. */
-  scores: Record<string, number>;
-  violations: ContractViolation[];
-  ruleExecutions: ContractRuleExecution[];
-  /** Why rules errored, one line per errored rule or scope failure. */
-  errors: string[];
+type KnowledgeRecommendedAction = 'run_agent' | 'ask_user' | 'collect_web_data' | 'query_connectors' | 'inspect_repo' | 'build_domain_wiki' | 'continue_with_caveat' | 'abort_or_rescope';
+interface KnowledgeReadinessReport {
+  taskId: string;
+  readinessScore: number;
+  blockingMissingRequirements: KnowledgeRequirement[];
+  nonBlockingGaps: KnowledgeRequirement[];
+  recommendedAction: KnowledgeRecommendedAction;
+  bundle: KnowledgeBundle;
+  severity: ControlSeverity;
+  reason: string;
 }
-interface ContractCheckResult {
-  verdicts: ContractVerdict[];
-  /** `error` if any verdict errored, else `fail` if any failed, else `pass`. */
-  status: ContractStatus;
-  allValid: boolean;
+interface UserQuestion {
+  id: string;
+  question: string;
+  reason: string;
+  requirementId: string;
+  importance: KnowledgeImportance;
+  answerType: 'free_text' | 'select_one' | 'multi_select' | 'file_upload' | 'credential' | 'url';
+  defaultIfSkipped?: string;
+  impactIfUnknown: string;
+  options?: string[];
+  metadata?: Record<string, unknown>;
 }
-interface EvaluateTraceContractOptions {
-  /** The run `run` rules read. Ignored under `scope`, where the scope root is the run. */
-  run?: ContractRun;
+interface DataAcquisitionPlan {
+  id: string;
+  requirementIds: string[];
+  mode: Exclude<KnowledgeAcquisitionMode, 'not_available' | 'infer_low_confidence'> | 'build_domain_wiki';
+  description: string;
+  priority: KnowledgeImportance;
+  expectedEvidenceIds?: string[];
+  questions?: UserQuestion[];
+  metadata?: Record<string, unknown>;
 }
 //#endregion
-//#region src/trace-contracts/builder.d.ts
-interface OrderOptions {
-  order?: OrderMode;
-}
-interface ArgumentRuleOptions {
-  pointer: string;
-  check: ArgumentCheck;
-  occurrence?: ArgumentOccurrence;
+//#region src/leaderboard.d.ts
+interface LeaderboardRow {
+  /** Group key — the agent-profile cellId by default, else a caller dimension. */
+  key: string;
+  /** Display label: `harness · model` when the profile carries them, else the key. */
+  label: string;
+  harness?: string;
+  model?: string;
+  rank: number;
+  /** Measured runs in this group. */
+  n: number;
+  passRate: number;
+  passRateCi95: [number, number];
+  /** Observability rollup — means over the group; null when no run reported it. */
+  meanCostUsd: number | null;
+  meanTokensIn: number | null;
+  meanTokensOut: number | null;
+  meanWallMs: number | null;
 }
-declare class TraceContractBuilder {
-  private readonly name;
-  private readonly rules;
-  private readonly alternativeList;
-  private scopeRoot;
-  constructor(name: string);
-  /** Every span in the trace must satisfy `p`. */
-  always(p: SpanPredicate, label?: string): this;
-  /** No span in the trace may satisfy `p`. */
-  never(p: SpanPredicate, label?: string): this;
-  /** At least one span in the trace must satisfy `p`. */
-  eventually(p: SpanPredicate, label?: string): this;
-  /** Both `a` and `b` must occur, with `a` first under `options.order`
-   *  (default `start-order`). For "every `b` needs an earlier `a`, and a run
-   *  without `b` passes", use {@link neverUnless}. */
-  precedes(a: SpanPredicate, b: SpanPredicate, label?: string, options?: OrderOptions): this;
-  /** Every `p`-match needs an earlier `prior`-match; a run without `p` passes. */
-  neverUnless(p: SpanPredicate, prior: SpanPredicate, label?: string, options?: OrderOptions): this;
-  /** At most `max` spans may satisfy `p`. */
-  atMost(p: SpanPredicate, max: number, label?: string): this;
-  /** Spans satisfying `p` may use at most `max` input+output tokens in total. */
-  tokensAtMost(p: SpanPredicate, max: number, label?: string): this;
-  /** Terminal status and duration of the run. */
-  run(checks: RunChecks, label?: string): this;
-  /** A JSON Pointer check on the arguments of calls matching `p`. */
-  argument(p: SpanPredicate, options: ArgumentRuleOptions, label?: string): this;
-  /** The trace records the tools offered to the model, every call is one of
-   *  them, and, with `declared`, every offered tool is declared. */
-  toolsOffered(declared?: string[], label?: string): this;
-  /** No call repeats a side effect that is not proven safe to repeat. */
-  retrySafe(options: {
-    reads?: string[];
-    writes?: RetryWrite[];
-  }, label?: string): this;
-  /** Evaluate only the subtree rooted at the one span matching `root`. */
-  scope(root: SpanPredicate): this;
-  /** A named alternate path; at least one alternative must pass completely. */
-  alternative(id: string, define: (branch: TraceContractBuilder) => TraceContractBuilder): this;
-  build(): TraceContract;
-  private where;
-  private add;
+interface LeaderboardOptions {
+  /** Per-run pass predicate — the benchmark defines what "passed" means (real-build
+   *  hit-rate, test-pass, etc.). The leaderboard owns only the grouping + stats. */
+  passed: (record: RunRecord) => boolean;
+  /** Group key. Default: the agent-profile cellId (harness × model × dimensions),
+   *  so persona/harness/model sweeps separate without parsing labels. */
+  groupBy?: (record: RunRecord) => string;
 }
-export declare function traceContract(name: string): TraceContractBuilder;
-//#endregion
-//#region src/trace-contracts/evaluate.d.ts
-/**
- * Evaluate one contract over a span sequence. Pure and synchronous — works on
- * `Span[]` from a TraceStore, `contractSpansFromOtlp` output, or any array
- * satisfying `ContractSpan`. Throws
- * `ValidationError` for a malformed contract; a rule that cannot be evaluated
- * is reported as `error` in the verdict instead.
- */
-export declare function evaluateTraceContract(contract: TraceContract, spans: readonly ContractSpan[], options?: EvaluateTraceContractOptions): ContractVerdict;
 /**
- * Evaluate many contracts over one span sequence. Throws on an empty
- * contract list — `allValid: true` over zero contracts is a silent pass.
+ * Group `records`, compute pass-rate + Wilson CI + cost/token/latency means per
+ * group, and rank by pass-rate (cost as the tiebreaker — cheaper wins a tie).
+ * Pure projection: no I/O, deterministic, safe to call on any RunRecord[].
  */
-export declare function checkTraceContracts(spans: readonly ContractSpan[], contracts: readonly TraceContract[], options?: EvaluateTraceContractOptions): ContractCheckResult;
-//#endregion
-//#region src/trace-contracts/explain.d.ts
-/** Plain-language statement of everything a contract checks, one rule per line. */
-export declare function explainTraceContract(contract: TraceContract): string;
+export declare function leaderboard(records: RunRecord[], opts: LeaderboardOptions): LeaderboardRow[];
 //#endregion
-//#region src/trace-contracts/spec.d.ts
-interface ToolArgumentSpec {
-  tool: string;
-  /** RFC 6901 JSON Pointer into the call's arguments; `''` is the whole value. */
-  pointer: string;
-  /** Which calls must satisfy the check. Default `all`. */
-  occurrence?: ArgumentOccurrence;
-  /** Exactly one of `exists`, `equals`, `oneOf`, `type`. */
-  exists?: true;
-  equals?: unknown;
-  oneOf?: unknown[];
-  type?: ArgumentType;
+//#region src/paired-delta-test.d.ts
+interface PairedDeltaTestOptions extends PairedBootstrapOptions {
+  /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. */
+  threshold?: number;
+  /** Caller-required paired observations. The exact test may impose a higher minimum. */
+  minPairs?: number;
 }
-interface ToolsSpec {
-  /** Each must be called at least once. */
-  required?: string[];
-  /** None may be called. */
-  forbidden?: string[];
-  /** Every tool call must name one of these; a call without a name fails. */
-  allowed?: string[];
-  /** Ceiling on tool calls of any name. */
-  maxCalls?: number;
-  /** Ceiling per tool name. */
-  maxCallsPerTool?: Record<string, number>;
-  /** Each tool is called, and each comes before the next under `orderMode`. */
-  requiredOrder?: string[];
-  /** Default `start-order`. */
-  orderMode?: OrderMode;
-  arguments?: ToolArgumentSpec[];
-  /** The trace records the tools the harness offered the model
-   *  (`gen_ai.tool.definitions`), every call is one of them, and, with
-   *  `allowed`, every offered tool is allowed. */
-  enforced?: true;
+interface PairedDeltaTestResult {
+  bootstrap: PairedBootstrapResult;
+  method: 'bootstrap-ci' | 'exact-sign';
+  /** Exact one-sided p-value below the bootstrap minimum; otherwise null. */
+  pValue: number | null;
+  /** Effective observation minimum after accounting for confidence. */
+  minimumPairs: number;
+  sufficient: boolean;
+  /**
+   * The bootstrap interval has zero width (or is non-finite), so it carries no
+   * information about how far the estimate could be wrong and cannot support a
+   * decision in either direction. See {@link pairedDeltaTest}.
+   */
+  indeterminate: boolean;
+  significant: boolean;
 }
-interface RetriesSpec {
-  /** Tools that change nothing outside the run: repeating a call is safe. */
-  reads?: string[];
-  /** Tools that change state outside the run: a repeated call must carry one
-   *  idempotency key. Any other tool called twice with the same arguments fails. */
-  writes?: RetryWrite[];
+/** Smallest all-positive sample that can clear a one-sided exact sign test. */
+export declare function minimumPairsForPairedDeltaTest(confidence?: number): number;
+/**
+ * Tests whether a paired candidate-minus-baseline delta clears a threshold.
+ *
+ * At 20 or more pairs, the percentile bootstrap lower bound carries the
+ * decision. Below that point the interval is descriptive only, so the function
+ * switches to a pre-registered one-sided exact sign test. The exact path is
+ * deliberately conservative: it requires both a point estimate above the
+ * threshold and enough consistently positive paired differences.
+ *
+ * ## A zero-width interval is never significant
+ *
+ * When every paired delta is identical the resample distribution is a point
+ * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n
+ * identical deltas of g. Neither says the effect is certain — both say the
+ * sample carries no information about how far the estimate could be wrong, and
+ * `low > threshold` then answers on the point estimate alone. It fails in both
+ * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a
+ * tie-dominated pass/fail comparison laundered a regression into a
+ * noninferiority pass, and `[g, g]` clears every threshold below g with no
+ * spread behind it. Under a bounded asymmetric null whose true mean paired
+ * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —
+ * every sample that misses the drop is exactly that shape, and deciding on
+ * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.
+ *
+ * So `indeterminate` is reported and `significant` is false whenever the
+ * interval has zero width, on BOTH paths: at small n the exact sign test is a
+ * test of the MEDIAN and a zero-spread sample is precisely where it stops
+ * saying anything about the mean the caller is thresholding.
+ *
+ * `threshold` may be negative — that is a noninferiority margin, and it is the
+ * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome
+ * the percentile bootstrap is not a valid interval at a nonzero margin at all;
+ * use {@link decidePairedPromotion}, which routes those to Tango's score
+ * interval, rather than thresholding this function's bootstrap directly.
+ */
+export declare function pairedDeltaTest(before: number[], after: number[], options?: PairedDeltaTestOptions): PairedDeltaTestResult;
+//#endregion
+//#region src/partition-held-out.d.ts
+interface PartitionHeldOutOptions {
+  /** Partition seed. Bumping it reshuffles every assignment — do that only when
+   *  the corpus/policy changes meaningfully so old and new records can't pair.
+   *  Default 'held-out-v1'. */
+  seed?: string;
+  /** Fraction routed to the held-out bucket, in (0, 1). Default 0.5. */
+  holdoutFraction?: number;
+  /**
+   * Minimum held-out ids required for the split to be admissible. Below this,
+   * a paired comparison has too few observations to be significant, so the
+   * partition throws rather than hand back an underpowered holdout. Default 1
+   * (only the empty-holdout degenerate case is rejected). Raise it to enforce
+   * real statistical power. */
+  minHoldout?: number;
+  /** Same floor for the search side. Default 1. */
+  minSearch?: number;
 }
-interface LlmSpec {
-  maxCalls?: number;
-  /** Input plus output tokens across LLM calls; an unrecorded count fails. */
-  maxTotalTokens?: number;
-  /** Every LLM call must name one of these models; a call without one fails. */
-  allowedModels?: string[];
+interface HeldOutPartition {
+  /** Ids the optimizer is allowed to read/optimize against. */
+  search: string[];
+  /** Ids reserved for gated, held-out evaluation. */
+  holdout: string[];
+  seed: string;
+  holdoutFraction: number;
 }
-/** The checks one path carries: the base contract or one alternative. */
-interface PathSpec {
-  run?: RunChecks;
-  tools?: ToolsSpec;
-  llm?: LlmSpec;
-  retries?: RetriesSpec;
-  /** Low-level rules in their serialized form. */
-  rules?: ContractRule[];
+/**
+ * Partition a list of stable ids into disjoint `search` / `holdout` sets by a
+ * deterministic hash. Fail-loud:
+ *   - empty input throws,
+ *   - duplicate ids throw (a dup would be observed on both sides, double-counting),
+ *   - a holdout (or search) set below its floor throws (underpowered comparison).
+ *
+ * Order within each bucket follows the input order, so the result is stable and
+ * reproducible.
+ */
+export declare function partitionHeldOut(ids: string[], options?: PartitionHeldOutOptions): HeldOutPartition;
+interface TrainSelectionTestOptions<T> {
+  /** Partition seed. Bumping it reshuffles every unit. Default 'train-selection-test-v1'. */
+  seed?: string;
+  /** Share of units routed to selection, in (0, 1). Default 0.25. */
+  selectionFraction?: number;
+  /** Share of units routed to test, in (0, 1). Default 0.25. */
+  testFraction?: number;
+  /**
+   * The independent unit an item samples, such as the task a fresh run repeats.
+   * Every item of one unit lands in the same partition, so a final test never
+   * scores a unit the optimizer read. Default: the item id.
+   */
+  unitOf?: (item: T) => string;
+  /** Minimum items per partition. Defaults 1, 1 and 1. */
+  minTrain?: number;
+  minSelection?: number;
+  minTest?: number;
 }
-interface AlternativeSpec extends PathSpec {
+interface TrainSelectionTestPartition<T> {
+  /** Items an optimizer may learn from. */
+  train: T[];
+  /** Items an optimizer may use to accept candidates and stop. */
+  selection: T[];
+  /** Items reserved for the final comparison. */
+  test: T[];
+  seed: string;
+  selectionFraction: number;
+  testFraction: number;
+}
+/**
+ * Partition items into disjoint train, selection and test sets, the three
+ * partitions `compareOptimizationMethods` and Runtime `improve()` take.
+ *
+ * Assignment is per unit: `hashToUnit(unit, seed)` below `testFraction` is
+ * test, below `testFraction + selectionFraction` is selection, and the rest is
+ * train. A unit therefore keeps its partition as the corpus grows, and items of
+ * one unit never straddle two partitions. Fail-loud like `partitionHeldOut`:
+ * empty input, a duplicate or empty id, an empty unit, fractions outside
+ * (0, 1) or summing to 1 or more, and a partition below its floor all throw.
+ * Order within each partition follows the input order.
+ */
+export declare function partitionTrainSelectionTest<T extends {
   id: string;
+}>(items: readonly T[], options?: TrainSelectionTestOptions<T>): TrainSelectionTestPartition<T>;
+//#endregion
+//#region src/product-benchmark/export.d.ts
+interface ProductBenchmarkExportOptions {
+  /** Source eval run directories, each containing a `records.jsonl` of RunRecord rows. */
+  readonly runDirs: readonly string[];
+  /** Destination directory for the bundle (manifest + records + materialized source runs). */
+  readonly outDir: string;
+  readonly projectId: string;
+  readonly benchmarkId: string;
+  /** Repo-relative path of the product's canonical agent profile source. */
+  readonly agentProfilePath: string;
+  /** Pass threshold applied when a row carries no explicit `outcome.raw.pass`. Default 0.7. */
+  readonly passThreshold?: number;
+  /**
+   * First scenario tag. Defaults to `projectId` with a trailing `-agent`
+   * stripped (`tax-agent` → `tax`), matching the product exporters.
+   */
+  readonly scenarioTagPrefix?: string;
+  /** Profile id used when a row has no `agentProfile.profileId`. Defaults to the row's arm id. */
+  readonly fallbackProfileId?: string;
+  /** Arm mutable surfaces recorded in the manifest. Defaults to the full superset. */
+  readonly mutableSurfaces?: readonly string[];
+  /**
+   * Copy each run dir into `<outDir>/source-runs/` and record
+   * bundle-relative artifact paths (portable, self-contained). When false,
+   * artifacts keep absolute paths into the original run dirs. Default true.
+   */
+  readonly materializeSourceRuns?: boolean;
+  /**
+   * Override split classification for a row. Return undefined to fall back
+   * to the default (`outcome.raw.safety === 1` → safety, then splitTag).
+   */
+  readonly classifySplit?: (record: RunRecord) => ProductBenchmarkSplit | undefined;
+  /** Recovers a tool-call count when the row's raw bag carries none (e.g. from turn artifacts). */
+  readonly toolCallFallback?: (record: RunRecord, runDir: string) => number;
+  /** Backend version recorded per row. Defaults to the cwd package.json's `@tangle-network/sandbox` range. */
+  readonly backendVersion?: string;
+  /**
+   * Explicit substrate versions for the manifest, merged over what the cwd
+   * package.json / node_modules resolve. Use when a substrate package is not
+   * installed where the export runs — the validator refuses an `'unknown'`
+   * version, so provide the real one rather than shipping the sentinel.
+   */
+  readonly substrate?: Partial<ProductBenchmarkManifest['substrate']>;
 }
-interface TraceContractSpec extends PathSpec {
-  name: string;
-  description?: string;
-  /** Check only the one span matching `root` and its descendants: one
-   *  sub-agent (`{ "kind": "AGENT", "name": "researcher" }`) or one search
-   *  node (`{ "attr": { "agent.branch.id": "n7" } }`). */
-  scope?: {
-    root: SpanPredicate;
-  };
-  /** At least one alternative must pass completely, as well as the base checks. */
-  alternatives?: {
-    anyOf: AlternativeSpec[];
-  };
+interface ProductBenchmarkSingleRunExportOptions extends Omit<ProductBenchmarkExportOptions, 'runDirs'> {
+  readonly runDir: string;
 }
+interface ProductBenchmarkExportResult {
+  readonly manifestPath: string;
+  readonly recordsPath: string;
+  readonly records: number;
+}
+/** Repo identity from the exporting process's cwd. `'unknown'` values are flagged by `validateProductBenchmarkRun`. */
+export declare function productBenchmarkRepoIdentity(): ProductBenchmarkManifest['repo'];
+/** Single-run convenience wrapper over `exportProductBenchmarkRuns`. */
+export declare function exportProductBenchmark(options: ProductBenchmarkSingleRunExportOptions): ProductBenchmarkExportResult;
 /**
- * Parse and compile a declarative contract. Throws `ValidationError` naming
- * the path of the first problem: an unknown key, a wrong type, an unknown
- * status, or a contract that checks nothing.
+ * Export one or more product eval run dirs into a validated product
+ * benchmark bundle at `outDir`. Both the manifest and every record are
+ * run through the contract validators before anything is written.
  */
-export declare function compileTraceContractSpec(input: unknown): TraceContract;
-interface ContractLintFinding {
-  /** `error`: the contract can never pass, or contradicts itself.
-   *  `warning`: the contract works but likely does not say what was meant. */
-  level: 'error' | 'warning';
-  code: string;
-  path: string;
-  message: string;
+export declare function exportProductBenchmarkRuns(options: ProductBenchmarkExportOptions): ProductBenchmarkExportResult;
+//#endregion
+//#region src/product-benchmark/index.d.ts
+declare const productBenchmarkSplits: readonly ['practice', 'dev', 'holdout', 'safety', 'sentinel'];
+type ProductBenchmarkSplit = (typeof productBenchmarkSplits)[number];
+interface ProductBenchmarkRepoRef {
+  readonly url: string;
+  readonly commit: string;
+  readonly branch: string;
+}
+interface ProductBenchmarkSubstrateVersions {
+  readonly agentEval: string;
+  readonly agentRuntime: string;
+  readonly agentInterface: string;
+  readonly sandbox: string;
+  readonly agentBench?: string;
+}
+interface ProductBenchmarkProfileRef {
+  readonly id: string;
+  readonly profileHash: string;
+  readonly agentProfilePath: string;
+}
+interface ProductBenchmarkArm {
+  readonly id: string;
+  readonly profileId: string;
+  readonly mutableSurfaces: readonly string[];
+  readonly policyAxes: Record<string, unknown>;
+}
+interface ProductBenchmarkScenario {
+  readonly id: string;
+  readonly split: ProductBenchmarkSplit;
+  readonly tags: readonly string[];
+  readonly sourceAllowedForSynthesis: boolean;
+}
+interface ProductBenchmarkBudgets {
+  readonly maxUsd: number;
+  readonly maxCells: number;
+  readonly maxWallMs: number;
+}
+interface ProductBenchmarkManifest {
+  readonly schemaVersion: 1;
+  readonly projectId: string;
+  readonly benchmarkId: string;
+  readonly repo: ProductBenchmarkRepoRef;
+  readonly substrate: ProductBenchmarkSubstrateVersions;
+  readonly profiles: readonly ProductBenchmarkProfileRef[];
+  readonly arms: readonly ProductBenchmarkArm[];
+  readonly scenarios: readonly ProductBenchmarkScenario[];
+  readonly budgets: ProductBenchmarkBudgets;
+  readonly expectedArtifactDir: string;
 }
+interface ProductBenchmarkValidationReport {
+  readonly manifestPath: string;
+  readonly recordsPath: string;
+  readonly records: number;
+  /** Manifest repo fields that are empty or the `'unknown'` export sentinel. */
+  readonly repoFailures: readonly string[];
+  /** Manifest substrate versions that are empty or the `'unknown'` export
+   *  sentinel — a bundle without substrate identity is not reproducible. */
+  readonly substrateFailures: readonly string[];
+  readonly projects: readonly string[];
+  readonly benchmarks: readonly string[];
+  readonly arms: readonly string[];
+  readonly scenarios: readonly string[];
+  readonly passed: number;
+  readonly failed: number;
+  readonly inputTokens: number;
+  readonly outputTokens: number;
+  readonly costUsd: number;
+  readonly wallMs: number;
+  readonly integrityFailures: readonly string[];
+  readonly missingArtifacts: readonly string[];
+}
+export declare function readProductBenchmarkManifest(path: string): ProductBenchmarkManifest;
 /**
- * Find contradictions and likely mistakes in a declarative contract. Parses
- * strictly first, so a malformed spec throws instead of linting.
+ * Fail-loud gate over a bundle directory: locates the manifest + records,
+ * runs `validateProductBenchmarkRun`, and throws with every repo,
+ * integrity, and artifact failure listed. Returns the report when clean.
  */
-export declare function lintTraceContractSpec(input: unknown): ContractLintFinding[];
+export declare function assertProductBenchmarkRun(runDir: string): ProductBenchmarkValidationReport;
+//#endregion
+//#region src/runtime-trajectory.d.ts
+interface RuntimeTrajectoryHookEvent {
+  id: string;
+  runId: string;
+  scenarioId?: string;
+  target: string;
+  phase: string;
+  timestamp: number;
+  stepIndex?: number;
+  parentId?: string;
+  payload?: unknown;
+  metadata?: Record<string, unknown>;
+}
+interface RuntimeTrajectoryRecord {
+  id?: string;
+  scenarioId?: string;
+  splitTag?: RunSplitTag;
+  runtimeEvents?: unknown;
+  [key: string]: unknown;
+}
+interface RuntimeTrajectoryRunRecord {
+  runId: string;
+  scenarioId?: string;
+  splitTag: RunSplitTag;
+}
+interface RuntimeTrajectoryEvidenceSummary {
+  recordCount: number;
+  recordWithRuntimeEventsCount: number;
+  runtimeRunCount: number;
+  lifecycleEventCount: number;
+  defaultedSplitCount: number;
+}
+interface RuntimeTrajectoryEvidenceProjection {
+  runs: RuntimeTrajectoryRunRecord[];
+  events: RuntimeTrajectoryHookEvent[];
+  summary: RuntimeTrajectoryEvidenceSummary;
+  diagnostics: string[];
+}
+interface ProjectRuntimeTrajectoryEvidenceOptions<TRecord extends RuntimeTrajectoryRecord = RuntimeTrajectoryRecord> {
+  records: TRecord[];
+  defaultSplitTag?: RunSplitTag;
+  recordIdOf?: (record: TRecord, index: number) => string | undefined;
+  scenarioIdOf?: (record: TRecord, index: number) => string | undefined;
+}
+export declare function projectRuntimeTrajectoryEvidence<TRecord extends RuntimeTrajectoryRecord>(options: ProjectRuntimeTrajectoryEvidenceOptions<TRecord>): RuntimeTrajectoryEvidenceProjection;
 //#endregion
 //#region src/active-learning.d.ts
 type SynthesisReason = 'high-variance' | 'undersampled' | 'failure-cluster' | 'difficulty-gap';
@@ -1061,85 +1282,6 @@ interface BoundedProcessResult {
 }
 export declare function runBoundedProcess(input: BoundedProcessInput): Promise<BoundedProcessResult>;
 //#endregion
-//#region src/client.d.ts
-/**
- * ProductClient — configurable HTTP client for exercising any agent's APIs.
- *
- * Routes are config, not hardcoded. Each agent provides its own RouteMap.
- *
- * @deprecated A one-product REST client does not belong in the generic
- * substrate. Moves to the product repo in the next major
- * (tangle-network/agent-runtime#694); vendor it if you construct it directly.
- */
-export declare class ProductClient {
-  private baseUrl;
-  private routes;
-  private cookies;
-  private timeoutMs;
-  constructor(config: ProductClientConfig);
-  private route;
-  /**
-   * Single HTTP boundary for every JSON request. Aborts after `timeoutMs`,
-   * checks `res.ok` BEFORE parsing, and throws `${method} ${path} failed:
-   * HTTP ${status} — ${body}` on a non-ok response so a 4xx/5xx error body
-   * can never be parsed as a success shape. Callers inspect the resolved
-   * value only after a successful return.
-   */
-  private request;
-  /**
-   * Read a required collection field. A missing/non-array field is a contract
-   * violation (wrong route, drift, partial body) — surface it loud instead of
-   * masking it as a healthy empty set. A genuinely empty `[]` passes through.
-   */
-  private requireArray;
-  signup(name: string, email: string, password: string): Promise<{
-    userId: string;
-  }>;
-  login(email: string, password: string): Promise<void>;
-  createWorkspace(name: string, type?: string): Promise<string>;
-  createThread(workspaceId: string): Promise<string>;
-  chat(workspaceId: string, threadId: string, content: string, _opts?: {
-    blockPatterns?: RegExp[];
-  }): Promise<{
-    text: string;
-    blocks: {
-      type: string;
-      title: string;
-    }[];
-  }>;
-  getTasks(workspaceId: string): Promise<{
-    id: string;
-    title: string;
-    status: string;
-    priority: string;
-  }[]>;
-  getEvents(workspaceId: string): Promise<{
-    id: string;
-    title: string;
-    type: string;
-  }[]>;
-  getApprovals(workspaceId: string): Promise<{
-    id: string;
-    title: string;
-    status: string;
-    type: string;
-  }[]>;
-  getVaultTree(workspaceId: string): Promise<string[]>;
-  approveAction(workspaceId: string, id: string): Promise<void>;
-  rejectAction(workspaceId: string, id: string, reason: string): Promise<void>;
-  getGenerations(workspaceId: string): Promise<{
-    id: string;
-    type: string;
-    prompt: string;
-  }[]>;
-  /** Generic GET for custom routes */
-  get(path: string): Promise<Record<string, unknown>>;
-  /** Generic POST for custom routes */
-  post(path: string, body: Record<string, unknown>): Promise<Record<string, unknown>>;
-  /** Generic PATCH for custom routes */
-  patch(path: string, body: Record<string, unknown>): Promise<Record<string, unknown>>;
-}
-//#endregion
 //#region src/command-runner.d.ts
 /**
  * CommandRunner — abstract subprocess execution surface.
@@ -1567,891 +1709,875 @@ interface ReflectionProposal {
  */
 export declare function parseReflectionResponse(raw: string, maxProposals?: number): ReflectionProposal[];
 //#endregion
-//#region src/analyst/policy-edit.d.ts
-type PolicyEditSchemaVersion = 'policy-edit/v1';
-declare const POLICY_EDIT_AXES: readonly ['carrier', 'representation', 'budget', 'sampling', 'output_contract', 'tool_contract', 'routing', 'memory', 'agent_profile', 'deployment_target'];
-type PolicyEditAxis = (typeof POLICY_EDIT_AXES)[number];
-declare const POLICY_EDIT_TARGET_SURFACES: readonly ['prompt', 'tool-contract', 'runtime-config', 'memory', 'agent-profile', 'code', 'deployment'];
-type PolicyEditTargetSurface = (typeof POLICY_EDIT_TARGET_SURFACES)[number];
-type PolicyEditRisk = 'low' | 'medium' | 'high' | 'unknown';
-type PolicyEditGainDirection = 'increase' | 'decrease';
-type PolicyEditGainUnit = 'absolute' | 'relative' | 'percent' | 'score';
-interface PolicyEditTarget {
-  surface: PolicyEditTargetSurface;
-  /** Stable path inside the target surface, for example `system-prompt:tools`
-   * or `budget.maxTurns`. */
-  path?: string;
-  /** Optional canonical deployment identity. Store the existing cell, not a
-   * local profile shape. */
-  agentProfileCell?: AgentProfileCell;
-  /** Human label when the path is not enough for a readable audit trail. */
-  label?: string;
-}
-type PolicyEditChange = {
-  kind: 'text';
-  mode: 'append' | 'prepend' | 'replace';
-  value: string;
-  /** Required when `mode === 'replace'`; exact match only. */
-  find?: string;
-} | {
-  kind: 'json';
-  mode: 'set' | 'merge' | 'remove';
-  path: string;
-  value?: AgentProfileJson;
-};
-interface PolicyEditExpectedGain {
-  /** Metric this edit is expected to move, e.g. `holdout.composite`. */
-  metric: string;
-  direction: PolicyEditGainDirection;
-  /** Positive magnitude in the metric's native units. */
-  amount: number;
-  unit?: PolicyEditGainUnit;
-  rationale?: string;
-}
-interface PolicyEditSource {
-  findingIds: string[];
-  analystIds: string[];
-  evidenceRefs: EvidenceRef[];
-  /** Mirrors `AnalystFinding.derived_from_judge`; admission rejects it. */
-  derivedFromJudge?: boolean;
-}
-interface PolicyEdit {
-  schemaVersion: PolicyEditSchemaVersion;
-  editId: string;
-  axis: PolicyEditAxis;
-  target: PolicyEditTarget;
-  change: PolicyEditChange;
-  claim: string;
-  expectedGain: PolicyEditExpectedGain;
-  confidence: number;
-  risk: PolicyEditRisk;
-  source: PolicyEditSource;
-  rationale?: string;
-  validationPlan?: string;
-  metadata?: Record<string, unknown>;
+//#region src/eval-trace-store.d.ts
+/**
+ * Backing persistence for `EvalTraceStore`. The in-memory store is the default;
+ * the JSONL file store appends one validated `RunRecord` per line. Both keep an
+ * append order so `getBest` / `compareRuns` are reproducible.
+ */
+interface RunRecordBackend {
+  append(record: RunRecord): Promise<void>;
+  load(): Promise<RunRecord[]>;
 }
-export declare const POLICY_EDIT_CANDIDATE_RECORD_SCHEMA: 'tangle.policy-edit-candidate.v1';
-/** JSON-safe attribution carried with a measured candidate and its scores. */
-interface PolicyEditCandidateRecord {
-  schema: typeof POLICY_EDIT_CANDIDATE_RECORD_SCHEMA;
-  policyEdit: PolicyEdit;
+/**
+ * JSONL-backed store at `path`, one `RunRecord` per line. Malformed lines fail
+ * loud on load (a corrupt corpus must not silently shrink the analysis set);
+ * pass `skipInvalid` only for forensics on a known-bad file.
+ */
+export declare function jsonlRunRecordBackend(path: string, opts?: {
+  skipInvalid?: boolean;
+}): RunRecordBackend;
+//#endregion
+//#region src/integrity/preflight.d.ts
+interface ModelPreflight {
+  /** The model id as supplied by the caller. */
+  model: string;
+  /** Membership in the `{baseUrl}/models` served set. */
+  listed: boolean;
+  /**
+   * The router reached a provider for this model. `null` when `probe` was not
+   * requested. True on a 2xx, and also when the provider answered by
+   * exhausting its reasoning budget — that proves reachability while leaving
+   * identity unproven (see `budgetExhausted`).
+   */
+  served: boolean | null;
+  /** HTTP status of the probe. `null` when not probed. */
+  status: number | null;
+  /** Probe body's `error.message` when present, else `null`. */
+  detail: string | null;
+  /**
+   * The probe died on the token budget rather than on the model. Identity is
+   * unproven, not refuted: raise `probeMaxTokens` to resolve it.
+   */
+  budgetExhausted: boolean;
+  /**
+   * Identity of the model that answered the probe, compared against `model`.
+   * `null` when the model was not probed or the probe failed.
+   */
+  substitution: ServedModelCheck | null;
 }
-type PolicyEditInit = Omit<PolicyEdit, 'schemaVersion' | 'editId'> & {
-  schemaVersion?: PolicyEditSchemaVersion;
-  editId?: string;
+/** One check Agent Eval asks the caller's endpoint to perform. */
+type ModelEndpointCheck = {
+  readonly kind: 'list-models';
+} | {
+  readonly kind: 'probe';
+  readonly model: string;
+  /** Output-token budget the probe completion may bill. */
+  readonly maxOutputTokens: number;
 };
-export declare class PolicyEditValidationError extends ValidationError {
-  readonly path: string;
-  constructor(message: string, path?: string);
-}
-interface FindingToPolicyEditOptions {
-  expectedGain?: PolicyEditExpectedGain | ((finding: AnalystFinding) => PolicyEditExpectedGain | null | undefined);
-  risk?: PolicyEditRisk | ((finding: AnalystFinding) => PolicyEditRisk);
-  defaultAxis?: PolicyEditAxis;
-  defaultTargetSurface?: PolicyEditTargetSurface;
-}
-interface PolicyEditAdmissionOptions {
-  minScore?: number;
-  minExpectedGain?: number;
-  allowHighRisk?: boolean;
-  requireEvidence?: boolean;
+/**
+ * Caller-owned request into the model endpoint. The caller binds the base URL
+ * and the credential and returns the raw `Response`; a `list-models` request
+ * answers with an OpenAI-compatible `{ data: [{ id }] }` body, and a `probe`
+ * request answers with one minimal chat completion for `model`.
+ */
+type ModelEndpointRequest = (check: ModelEndpointCheck) => Promise<Response>;
+interface PreflightModelsOptions {
+  /** Caller-owned endpoint request. Agent Eval issues no provider HTTP itself. */
+  request: ModelEndpointRequest;
+  /** Model ids to check. */
+  models: string[];
+  /** When true, additionally spend a small chat probe per model. Default false. */
+  probe?: boolean;
+  /**
+   * Output-token budget per probe. Default `PROBE_MAX_TOKENS`. Lowering it
+   * below the reasoning floor makes healthy reasoning models report
+   * `budgetExhausted` instead of proving their identity.
+   */
+  probeMaxTokens?: number;
 }
-interface PolicyEditAdmission {
-  edit: PolicyEdit;
-  decision: 'admit' | 'reject';
-  score: number;
-  reasons: string[];
+interface PreflightOutcome {
+  succeeded: boolean;
+  value: ModelPreflight[] | null;
+  error: string | null;
 }
-export declare function makePolicyEdit(init: PolicyEditInit): PolicyEdit;
-export declare function computePolicyEditId(edit: Omit<PolicyEdit, 'editId'> | PolicyEdit): string;
-export declare function validatePolicyEdit(input: unknown): PolicyEdit;
-export declare function makePolicyEditCandidateRecord(edit: PolicyEdit): PolicyEditCandidateRecord;
-export declare function validatePolicyEditCandidateRecord(input: unknown): PolicyEditCandidateRecord;
-export declare function policyEditsFromFindings(findings: ReadonlyArray<AnalystFinding>, opts?: FindingToPolicyEditOptions): PolicyEdit[];
-export declare function admitPolicyEdit(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): PolicyEditAdmission;
-export declare function applyPolicyEditToSurface(surface: unknown, edit: PolicyEdit): unknown;
-//#endregion
-//#region src/analyst/steer-firewall.d.ts
 /**
- * THE steer firewall. Fail-loud guard for any path that admits analyst findings
- * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any
- * finding whose provenance is a judge verdict, rather than let `J` leak into the
- * loop. Returns the findings unchanged for chaining.
- *
- * Call this at the chokepoint where a detector that ALSO scores/gates has its
- * findings turned into a steer (the judge-and-steer dual-role case). It keys on
- * provenance, so it correctly admits evidence-less trace-analyst observations and
- * correctly rejects an artifact-citing judge verdict — the cases an evidence
- * check gets backwards.
+ * Check that `models` are reachable on the router. Returns a typed outcome —
+ * a network failure yields `{ succeeded: false, error }`, never a throw and
+ * never a partial result silently reported as success. No retries, no
+ * fallbacks.
  *
- * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge
- * whose output is laundered through a hand-built finding with no provenance flag
- * is out of its reach — provenance must be honestly set at every judge→finding
- * lift (today: createJudgeAdapter). That is why the integrity rule lives at the
- * lift site, and why ProposeContext.judgeScores?: never is the complementary
- * compile-time tripwire on the obvious direct channel.
+ * The membership check (one GET) always runs. When `probe` is true, each model
+ * additionally gets a small chat probe so a model that is listed but
+ * unconfigured (a 401 `model_not_found` from the router) is caught.
  */
-export declare function assertNoJudgeVerdict(findings: ReadonlyArray<AnalystFinding>, context?: string): ReadonlyArray<AnalystFinding>;
+export declare function preflightModels(opts: PreflightModelsOptions): Promise<PreflightOutcome>;
 //#endregion
-//#region src/capability-headroom.d.ts
+//#region src/model-seats.d.ts
+interface ModelSeats {
+  /** The model under evaluation — what the agent itself dispatches with. */
+  worker?: string;
+  /** Judge-panel model ids — thread into `ensembleJudge({ models })`. */
+  judges?: string[];
+  /** Analyst model — `analyzeRuns` / analyst-registry LLM calls. */
+  analyst?: string;
+  /** Reflection or candidate-generation model. */
+  reflection?: string;
+  /** Verifier model — completion/objective checking. */
+  verifier?: string;
+}
+type SeatName = keyof ModelSeats;
 /**
- * Capability-headroom gate — "can this task set even SEE the capability
- * you're about to A/B?"
+ * Tier presets — plain data, swap or spread freely.
  *
- * A capability A/B (grant one arm a tool, a skill, extra context — anything)
- * can only detect the capability on tasks the capability-ABSENT baseline
- * FAILS. A task the baseline already passes is saturated: the treatment arm
- * has no headroom to demonstrate anything, and averaging saturated tasks into
- * the comparison dilutes a real effect toward a false null. Calibrate before
- * measuring: run the baseline arm first, classify every task's headroom, and
- * refuse to run the comparison when no task has a gap.
+ * A preset names REQUESTED ids, and a requested id is NOT a guarantee of
+ * family, of provider, or of liveness. A routing gateway can accept any id
+ * below and answer from a different model on HTTP 200, with only the response
+ * body's `model` field betraying the swap. The served assertion is the
+ * guarantee: gate a run on `assertModelsServed({ probe: true })` and assert
+ * `assertServedModel` per call. `assertCrossFamily` over these ids proves the
+ * configuration is diverse; only `assertCrossFamilyServed` over the ids that
+ * ANSWERED proves the run was.
  *
- * Fail-closed on unknown outcomes: a baseline run whose outcome could not be
- * determined (infra loss, unverified, missing) NEVER counts as a failure —
- * treating "couldn't tell" as "failed" would manufacture headroom out of
- * telemetry gaps and launder infra noise into a treatment effect.
+ * `economy` names ids a live router probe answered from the provider their
+ * name implies, so the judge trio spanned three provider families (deepseek /
+ * zhipu / google) as configured and, at that probe, as served. A preset is
+ * only as good as its last probe: ids go dead and start resolving elsewhere
+ * without notice, so re-probe rather than trusting this list.
  *
- * Domain-free by construction: the capability under test never appears here.
- * Callers decide what the baseline arm lacks; this module only classifies
- * baseline outcomes per task.
+ * `frontier` is deliberately EMPTY: entitled frontier ids vary per router
+ * account, and a hardcoded claude/gpt-5 id 401s on keys that lack it. Supply
+ * your own: `{ ...seatPresets.frontier, worker: '<your-frontier-id>', … }` —
+ * `resolveSeat` throws on every seat you haven't filled.
  */
-/** One baseline-arm (capability-absent) observation of one task. */
-interface HeadroomInput {
-  /** Stable task identity; multiple rows per task are multiple baseline reps. */
-  taskId: string;
-  /** Baseline outcome. 'unknown' = the outcome could not be determined —
-   *  distinct from 'fail' and never counted as one. */
-  baselineOutcome: 'pass' | 'fail' | 'unknown';
-}
-/** Headroom classification for one task.
- *  - 'gap': the baseline fails it — the comparison can see the capability here.
- *  - 'saturated': the baseline already passes it — no headroom.
- *  - 'unknown': every baseline outcome is unknown — fail-closed, not a gap. */
-type HeadroomClass = 'gap' | 'saturated' | 'unknown';
-interface TaskHeadroom {
-  taskId: string;
-  /** Total baseline observations for this task (known + unknown). */
-  n: number;
-  /** Observations with a known ('pass' | 'fail') outcome — the evidence the
-   *  classification actually rests on. A large `n` with a small `nKnown`
-   *  means the headroom verdict is thin, not well-replicated. */
-  nKnown: number;
-  /** Pass rate over KNOWN outcomes only; NaN when every outcome is unknown
-   *  (no data ≠ measured zero). */
-  baselinePassRate: number;
-  headroom: HeadroomClass;
-}
-interface CapabilityHeadroomOptions {
-  /** A task counts as 'gap' when its baseline pass rate is ≤ this. Default 0
-   *  (the baseline must fail every known rep). Must be in [0, 1) — a
-   *  threshold of 1 would classify a fully-saturated task as a gap. */
-  keepThreshold?: number;
-}
-interface CapabilityHeadroomResult {
-  /** Per-task classification, sorted by `taskId`. */
-  tasks: TaskHeadroom[];
-  summary: {
-    tasksWithGap: number;
-    tasksSaturated: number;
-    tasksUnknown: number;
-    /** Total unknown-outcome reps across ALL tasks (including tasks that
-     *  still classified via their known reps) — the size of the telemetry
-     *  hole behind the classifications. */
-    repsUnknown: number;
-  };
-}
+export declare const seatPresets: Record<'economy' | 'frontier', ModelSeats>;
 /**
- * Classify every task's baseline headroom from capability-absent baseline
- * outcomes.
- *
- * A task is 'gap' when it has ≥ 1 known outcome and its pass rate over known
- * outcomes is ≤ `keepThreshold`; 'saturated' when the rate exceeds the
- * threshold; 'unknown' when every outcome is unknown. Unknown reps never
- * count toward the rate in either direction (fail-closed — see the module
- * doc), and evidence strength stays visible: each task reports `nKnown`
- * alongside `n`, and the summary totals the unknown reps in `repsUnknown`.
- * Throws on empty input, an out-of-range `keepThreshold`, or an
- * unrecognized outcome value (rows are often projected from untyped records).
+ * Read one seat. Blank strings and empty arrays count as unset (env-var
+ * plumbing produces them); malformed values (non-string seat, non-array or
+ * blank-entry `judges`) throw `ValidationError`. When the seat is unset, an
+ * explicit `fallback` is returned (`[fallback]` for `judges` — a one-model
+ * panel); without one, `SeatUnsetError`.
  */
-export declare function capabilityHeadroom(rows: readonly HeadroomInput[], opts?: CapabilityHeadroomOptions): CapabilityHeadroomResult;
-interface AssertCapabilityHeadroomOptions {
-  /** Minimum tasks classified 'gap' for the comparison to proceed. Default 1. */
-  minTasksWithGap?: number;
-}
+export declare function resolveSeat(seats: ModelSeats, seat: 'judges', fallback?: string): string[];
+export declare function resolveSeat(seats: ModelSeats, seat: Exclude<SeatName, 'judges'>, fallback?: string): string;
+export declare function resolveSeat(seats: ModelSeats, seat: SeatName, fallback?: string): string | string[];
+//#endregion
+//#region src/prompt-registry.d.ts
 /**
- * The go/no-go guard: throws when fewer than `minTasksWithGap` tasks have
- * baseline headroom — i.e. the benchmark cannot see the capability, so any
- * comparison run on it would measure noise. Call between the baseline
- * calibration run and the treatment run.
+ * Versioned prompt registry.
+ *
+ * Every prompt used in an eval run is registered with an explicit version.
+ * Reports include the content hash so A/B compares are rigorous: if the
+ * hash changes between two reports, the prompt actually changed; if it
+ * matches, the variance is elsewhere.
+ *
+ * Hash is SHA-256(content), truncated to 12 hex chars for readability.
+ * Uses the Web Crypto API (works in Workers, Node 22+, browsers).
  */
-export declare function assertCapabilityHeadroom(result: CapabilityHeadroomResult, opts?: AssertCapabilityHeadroomOptions): void;
+interface PromptHandle {
+  /** Stable human-readable id, e.g. 'browser.system' */
+  id: string;
+  /** Caller-chosen version string, e.g. 'v3' or '2026-04-20' */
+  version: string;
+  /** SHA-256 of content, 12-hex-char prefix */
+  hash: string;
+  /** Full prompt body */
+  content: string;
+}
+export declare class PromptRegistry {
+  private readonly entries;
+  /**
+   * Register a prompt. Re-registering the same id+version with DIFFERENT
+   * content throws — versions are immutable. Re-registering with the SAME
+   * content is a no-op (idempotent).
+   */
+  register(id: string, version: string, content: string): Promise<PromptHandle>;
+  /** Look up a registered prompt. Throws if unknown — no implicit defaults. */
+  get(id: string, version: string): PromptHandle;
+  /** Return all versions of an id, newest-first (lex-descending on version). */
+  listVersions(id: string): PromptHandle[];
+  /** Snapshot the whole registry — useful for including in reports. */
+  list(): PromptHandle[];
+  /** Verify a hash against registered content. Returns null if not found. */
+  verifyHash(id: string, version: string, expectedHash: string): boolean | null;
+}
+/** SHA-256(content) → first 12 hex chars. Stable across runtimes. */
+export declare function hashContent(content: string): Promise<string>;
 //#endregion
-//#region src/cost-tracker.d.ts
+//#region src/reference-replay.d.ts
 /**
- * Cost tracker — token + USD accounting per scenario and per run.
+ * Reference replay — score an agent against withheld historical outcomes.
  *
- * Adapted from generic usage-event accounting. Every
- * optimizer needs to know "is the quality gain worth the cost delta?",
- * and every dashboard needs dollars-per-completed-task. MODEL_PRICING
- * from metrics.ts stays authoritative for estimate math; this module
- * adds the aggregation + per-scenario roll-up that was duplicated
- * across 4 verticals.
+ * This is the generic version of the public-audit replay pattern:
+ * run a candidate system on an old task, keep the reference answers hidden
+ * until after execution, then score recall/precision and gate promotion
+ * across train/dev/test/holdout splits.
  */
-interface TokenSpec {
-  inputTokens: number;
-  outputTokens: number;
-  cachedTokens?: number;
-  reasoningTokens?: number;
+type ReferenceReplaySplit = 'train' | 'dev' | 'test' | 'holdout';
+interface ReferenceReplayItem {
+  id: string;
+  title: string;
+  description?: string;
+  severity?: string;
+  tags?: string[];
+  weight?: number;
+}
+interface ReferenceReplayCandidate {
+  id: string;
+  title: string;
+  description?: string;
+  severity?: string;
+  tags?: string[];
+  metadata?: Record<string, unknown>;
+}
+interface ReferenceReplayMatch {
+  scenarioId: string;
+  referenceId: string;
+  candidateId: string | null;
+  score: number;
+  matched: boolean;
+  weight: number;
+  reason: string;
+}
+interface ReferenceReplayScenarioScore {
+  scenarioId: string;
+  split: ReferenceReplaySplit;
+  matched: number;
+  total: number;
+  falsePositives: number;
+  matchedWeight: number;
+  totalWeight: number;
+  precision: number;
+  recall: number;
+  f1: number;
+  matches: ReferenceReplayMatch[];
+}
+interface ReferenceReplayAggregate {
+  matched: number;
+  total: number;
+  falsePositives: number;
+  matchedWeight: number;
+  totalWeight: number;
+  precision: number;
+  recall: number;
+  f1: number;
+  weightedRecall: number;
 }
-interface CostEntry extends TokenSpec {
-  scenarioId: string;
-  model: string;
-  /** Override estimate with an observed cost (e.g. from provider response). */
-  actualCostUsd?: number;
-  timestamp: number;
-  /** Free-form tags (variant id, round #, etc.). */
-  tags?: Record<string, string>;
+interface ReferenceReplayScore {
+  scenarios: ReferenceReplayScenarioScore[];
+  aggregate: ReferenceReplayAggregate;
+  bySplit: Partial<Record<ReferenceReplaySplit, ReferenceReplayAggregate>>;
 }
-interface ScenarioCost {
-  scenarioId: string;
-  entries: CostEntry[];
-  totalInputTokens: number;
-  totalOutputTokens: number;
-  totalCachedTokens: number;
-  totalCostUsd: number;
-  /** Pass flag — set by consumer via markOutcome; used for cost-per-completed-task. */
-  completed?: boolean;
+interface ReferenceReplayCaseRun<Input = unknown> {
+  caseId: string;
+  split: ReferenceReplaySplit;
+  input: Input;
+  metadata?: Record<string, unknown>;
+  references: ReferenceReplayItem[];
+  candidates: ReferenceReplayCandidate[];
+  score: ReferenceReplayScenarioScore;
+  durationMs: number;
+  error?: string;
 }
-export declare class CostTracker {
-  private byScenario;
-  record(entry: Omit<CostEntry, 'timestamp'> & {
-    timestamp?: number;
-  }): CostEntry;
-  markOutcome(scenarioId: string, completed: boolean): void;
-  /**
-   * Convenience: record + markOutcome in one call from a
-   * `{ usage, verdict }`-shaped response (starter-foundry's
-   * `invokeMetaJudge` returns this shape; consumers that wrap any
-   * judge/critic can follow the same convention).
-   *
-   * `usage.model` must be present in `MODEL_PRICING` for cost math to
-   * populate; otherwise totalCostUsd stays at 0 for the entry but
-   * tokens still aggregate.
-   */
-  recordVerdict(verdict: {
-    usage?: {
-      inputTokens: number;
-      outputTokens: number;
-      model: string;
-      cachedTokens?: number;
-      reasoningTokens?: number;
-    };
-    verdict?: 'pass' | 'fail' | 'borderline' | string;
-  }, scenarioId: string, tags?: Record<string, string>): CostEntry | null;
-  get(scenarioId: string): ScenarioCost | undefined;
-  list(): ScenarioCost[];
-  summary(): CostSummary;
+interface ReferenceReplayRun<Input = unknown> {
+  id: string;
+  variantId?: string;
+  startedAt: number;
+  completedAt: number;
+  durationMs: number;
+  cases: ReferenceReplayCaseRun<Input>[];
+  score: ReferenceReplayScore;
+  metadata?: Record<string, unknown>;
 }
-interface CostSummary {
-  scenarioCount: number;
-  completedCount: number;
-  totalInputTokens: number;
-  totalOutputTokens: number;
-  totalCostUsd: number;
-  avgCostPerScenarioUsd: number;
-  /** Total USD / completed scenarios — null when nothing completed. */
-  costPerCompletedTaskUsd: number | null;
+//#endregion
+//#region src/steering.d.ts
+interface SteeringRolePrompt {
+  system?: string;
+  append?: string;
+}
+interface SteeringBundle {
+  id: string;
+  coderPrompt?: string;
+  continuePrompt?: string;
+  reviewerPrompts?: Record<string, string>;
+  skills?: string[];
+  rolePrompts?: Record<string, SteeringRolePrompt>;
+  metadata?: Record<string, unknown>;
 }
 //#endregion
-//#region src/detectors/index.d.ts
-type DetectorSeverity = 'info' | 'warn' | 'critical';
-/** A normalized step a detector folds over. Fields are optional; each detector reads only what it
- *  needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;
- *  error-streak → `status`). */
-interface DetectorEvent {
-  /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */
-  readonly actionFingerprint?: string;
-  /** Fingerprint of observable state AFTER this step. */
-  readonly stateFingerprint?: string;
-  /** Score after this step (the score-flat half of no-progress). */
-  readonly score?: number;
-  /** Whether this step errored. */
-  readonly status?: 'ok' | 'error';
-  /** Free-form label carried into the signal evidence (e.g. the tool name). */
-  readonly label?: string;
+//#region src/run-score.d.ts
+interface RunScore {
+  success: number;
+  goalProgress: number;
+  repoGroundedness: number;
+  driftPenalty: number;
+  toolUseQuality: number;
+  patchQuality: number;
+  testReality: number;
+  finalGate: number;
+  reviewerBlockers: number;
+  costUsd: number;
+  wallSeconds: number;
+  notes?: string[];
 }
-interface DetectorSignal {
-  readonly detector: string;
-  readonly severity: DetectorSeverity;
-  readonly failureClass?: FailureClass;
-  readonly reason: string;
-  /** Consecutive matching steps at the moment the signal fired. */
-  readonly streak: number;
-  readonly evidence?: Record<string, unknown>;
+interface RunScoreWeights {
+  success: number;
+  goalProgress: number;
+  repoGroundedness: number;
+  driftPenalty: number;
+  toolUseQuality: number;
+  patchQuality: number;
+  testReality: number;
+  finalGate: number;
+  reviewerBlockers: number;
+  costUsd: number;
+  wallSeconds: number;
 }
-interface StreamingDetector {
-  readonly id: string;
-  /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */
-  readonly streak: number;
-  /** Fold one event; return a signal when the threshold trips, else null. */
-  observe(event: DetectorEvent): DetectorSignal | null;
-  /** Clear all state. */
-  reset(): void;
+export declare function aggregateRunScore(score: RunScore, weights?: Partial<RunScoreWeights>): number;
+export declare function clamp01(value: number): number;
+//#endregion
+//#region src/steering-optimizer.d.ts
+type SteeringOptimizerBackend = 'pairwise';
+interface SteeringOptimizationRow {
+  variantId: string;
+  scenarioId: string;
+  bundle: SteeringBundle;
+  score: RunScore;
+  metadata?: Record<string, unknown>;
 }
-interface RepeatedActionOptions {
-  /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).
-   *  `<= 0` disables the signal (the streak is still tracked, for telemetry). */
-  readonly maxRepeated?: number;
-  readonly severity?: DetectorSeverity;
-  /** Failure class to stamp on the signal (default `tool_recovery_failure`). */
-  readonly failureClass?: FailureClass;
+interface SteeringOptimizationSelector {
+  backend: SteeringOptimizerBackend;
+  signature?: string;
+  labels?: string[];
+  rationale?: string;
 }
-/** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */
-export declare function repeatedActionDetector(opts?: RepeatedActionOptions): StreamingDetector;
-interface ErrorStreakOptions {
-  /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */
-  readonly maxErrors?: number;
-  readonly severity?: DetectorSeverity;
-  readonly failureClass?: FailureClass;
+interface SteeringOptimizationResult {
+  backend: SteeringOptimizerBackend;
+  recommendedVariantId: string;
+  rationale: string;
+  rankings: Array<{
+    variantId: string;
+    mean: number;
+    runs: number;
+  }>;
+  selector?: SteeringOptimizationSelector;
+  skipped?: boolean;
+}
+interface SteeringOptimizerConfig {
+  weights?: Partial<RunScoreWeights>;
 }
-/** N consecutive tool errors = the worker is hammering a broken approach. */
-export declare function errorStreakDetector(opts?: ErrorStreakOptions): StreamingDetector;
-/** Fold one event through many detectors at once; returns every signal that fired this step. The
- *  natural shape for an online pipe watching with a whole panel of detectors. */
-export declare function observeAll(detectors: ReadonlyArray<StreamingDetector>, event: DetectorEvent): DetectorSignal[];
-//#endregion
-//#region src/error-count-extractor.d.ts
 /**
- * Toolchain error-count extractor.
- *
- * Given stderr/stdout from a compiler or test runner, count the number
- * of reported errors/failures. Patterns are deliberately narrow —
- * unknown stderr returns `null` rather than zero so callers can
- * distinguish "no errors" from "different toolchain, couldn't parse".
- *
- * All patterns are anchored to the start of a line and use bounded
- * character classes to avoid catastrophic backtracking on pathological
- * inputs.
+ * Rank already-evaluated steering variants.
  *
- * Add new toolchains by appending to {@link ERROR_COUNT_PATTERNS};
- * order matters only in the sense that the first matching pattern wins.
+ * Model-backed candidate generation belongs to the official GEPA campaign
+ * integration. This class only performs the deterministic selection it owns.
  */
-interface ErrorCountPattern {
-  /** Stable identifier for logging + tests. */
-  name: string;
-  /** Must be global (`g` flag) — the extractor counts matches. */
-  regex: RegExp;
-  /** Optional post-processing to extract a count from a single captured match. */
-  transform?: (match: RegExpMatchArray) => number;
+export declare class PairwiseSteeringOptimizer {
+  optimize(rows: SteeringOptimizationRow[], config?: SteeringOptimizerConfig): SteeringOptimizationResult;
 }
-export declare const ERROR_COUNT_PATTERNS: ErrorCountPattern[];
-interface ExtractOptions {
-  /** Restrict to named patterns — default: all patterns. */
-  only?: string[];
-  /** Additional patterns to consider BEFORE the built-in list. */
-  extra?: ErrorCountPattern[];
+//#endregion
+//#region src/scorecard.d.ts
+/** One commit's measurement of one (scenario, profile) cell. */
+interface ScorecardEntry {
+  commitSha: string;
+  /** ISO timestamp the entry was recorded. */
+  timestamp: string;
+  /** Per-seed (or per-rep) scores for this cell at this commit. */
+  scores: number[];
+  /** Median of `scores` — the cell's headline score for the commit. */
+  composite: number;
+  /** Per-dimension means, when the runs carried a judge breakdown. */
+  perDimension?: Record<string, number>;
+  /** RunRecord ids folded into this entry — provenance. */
+  runIds: string[];
 }
-interface ExtractResult {
-  /** Total count of matched errors, or null when no pattern matched. */
-  count: number | null;
-  /** Name of the pattern that matched, or null. */
-  matched: string | null;
-  /** Original matches for callers that want to surface specifics. */
-  samples: string[];
+/** A (scenario, profile) cell and its commit-ordered score timeline. */
+interface ScorecardCell {
+  scenarioId: string;
+  profileHash: string;
+  /** Model id — denormalised from the profile for readable filtering. */
+  model: string;
+  timeline: ScorecardEntry[];
 }
-/**
- * Try each pattern in order; return the first with matches.
- *
- * Returning `null` (instead of zero) on no-match is deliberate — a
- * callsite that greps for "typescript errors" on cargo output should
- * NOT treat that as "zero TS errors" because the toolchain is wrong.
- */
-export declare function extractErrorCount(text: string, opts?: ExtractOptions): ExtractResult;
-//#endregion
-//#region src/knowledge/types.d.ts
-type KnowledgeRequirementCategory = 'user_specific' | 'company_specific' | 'domain_specific' | 'codebase_specific' | 'market_specific' | 'regulatory' | 'tool_api' | 'credential_or_secret' | 'runtime_environment' | 'preference' | 'historical_context';
-type KnowledgeAcquisitionMode = 'ask_user' | 'search_web' | 'query_connector' | 'inspect_repo' | 'run_command' | 'infer_low_confidence' | 'not_available';
-type KnowledgeImportance = 'blocking' | 'high' | 'medium' | 'low';
-type KnowledgeFreshness = 'static' | 'monthly' | 'weekly' | 'daily' | 'realtime';
-type KnowledgeSensitivity = 'public' | 'private' | 'secret';
-type KnowledgeFallbackPolicy = 'block' | 'ask' | 'continue_with_caveat' | 'use_default';
-interface KnowledgeRequirement {
-  id: string;
-  description: string;
-  requiredFor: string[];
-  category: KnowledgeRequirementCategory;
-  acquisitionMode: KnowledgeAcquisitionMode;
-  importance: KnowledgeImportance;
-  freshness: KnowledgeFreshness;
-  sensitivity: KnowledgeSensitivity;
-  confidenceNeeded: number;
-  currentConfidence: number;
-  evidenceIds: string[];
-  fallbackPolicy: KnowledgeFallbackPolicy;
-  /**
-   * ISO timestamp after which this requirement must be treated as stale.
-   * Stale requirements score as missing even when they still have evidence.
-   */
-  validUntil?: string;
-  /** ISO timestamp for the last source-grounding or human verification pass. */
-  lastVerifiedAt?: string;
-  metadata?: Record<string, unknown>;
+/** The folded scorecard: every cell, plus the profile definitions by hash. */
+interface Scorecard {
+  cells: ScorecardCell[];
+  /** Profile definitions seen — keeps the scorecard self-describing. */
+  profiles: Record<string, AgentProfile>;
 }
-interface KnowledgeBundle {
-  taskId: string;
-  requirements: KnowledgeRequirement[];
-  evidenceIds: string[];
-  claimIds: string[];
-  wikiPageIds: string[];
-  userAnswers: Record<string, string>;
-  missing: KnowledgeRequirement[];
-  readinessScore: number;
-  metadata?: Record<string, unknown>;
+/** One append-only log line — a single cell's entry for a single commit. */
+interface ScorecardLogLine {
+  scenarioId: string;
+  profileHash: string;
+  model: string;
+  profile: AgentProfile;
+  entry: ScorecardEntry;
 }
-type KnowledgeRecommendedAction = 'run_agent' | 'ask_user' | 'collect_web_data' | 'query_connectors' | 'inspect_repo' | 'build_domain_wiki' | 'continue_with_caveat' | 'abort_or_rescope';
-interface KnowledgeReadinessReport {
-  taskId: string;
-  readinessScore: number;
-  blockingMissingRequirements: KnowledgeRequirement[];
-  nonBlockingGaps: KnowledgeRequirement[];
-  recommendedAction: KnowledgeRecommendedAction;
-  bundle: KnowledgeBundle;
-  severity: ControlSeverity;
-  reason: string;
+interface RecordRunsOptions {
+  /** The profile that produced these runs — keys the cell. */
+  profile: AgentProfile;
+  commitSha: string;
+  /** Defaults to `new Date().toISOString()`. */
+  timestamp?: string;
 }
-interface UserQuestion {
-  id: string;
-  question: string;
-  reason: string;
-  requirementId: string;
-  importance: KnowledgeImportance;
-  answerType: 'free_text' | 'select_one' | 'multi_select' | 'file_upload' | 'credential' | 'url';
-  defaultIfSkipped?: string;
-  impactIfUnknown: string;
-  options?: string[];
-  metadata?: Record<string, unknown>;
+/**
+ * Fold a benchmark's `RunRecord[]` into per-cell scorecard log lines — one
+ * line per scenario the runs cover. All runs are attributed to the single
+ * `profile` in `opts` (the harness ran them under it); the cell key is
+ * `(scenarioId, agentProfileHash(profile))`.
+ */
+export declare function recordRuns(runs: RunRecord[], opts: RecordRunsOptions): ScorecardLogLine[];
+/** Record runs and append them to the log in one call. Returns the lines. */
+export declare function recordRunsToScorecard(logPath: string, runs: RunRecord[], opts: RecordRunsOptions): ScorecardLogLine[];
+/**
+ * Fold the JSONL log into a queryable `Scorecard`. A missing file yields an
+ * empty scorecard; a malformed line is skipped — a corrupt append never
+ * breaks the read. Each cell's timeline is sorted chronologically.
+ */
+export declare function loadScorecard(logPath: string): Scorecard;
+type CellVerdict = 'improved' | 'regressed' | 'flat' | 'new';
+interface ScorecardCellDiff {
+  scenarioId: string;
+  profileHash: string;
+  model: string;
+  verdict: CellVerdict;
+  /** Composite of the latest entry. */
+  current: number;
+  /** Composite of the comparison entry — `null` when `verdict === 'new'`. */
+  baseline: number | null;
+  /** `current − baseline` — `null` when new. */
+  delta: number | null;
+  /** Cohen's d of current vs baseline samples — `null` when new or n < 2. */
+  cohensD: number | null;
+  /** Welch's t-test p-value — `null` when new or n < 2. */
+  pValue: number | null;
+  currentCommit: string;
+  baselineCommit: string | null;
 }
-interface DataAcquisitionPlan {
-  id: string;
-  requirementIds: string[];
-  mode: Exclude<KnowledgeAcquisitionMode, 'not_available' | 'infer_low_confidence'> | 'build_domain_wiki';
-  description: string;
-  priority: KnowledgeImportance;
-  expectedEvidenceIds?: string[];
-  questions?: UserQuestion[];
-  metadata?: Record<string, unknown>;
+interface ScorecardDiff {
+  cells: ScorecardCellDiff[];
+  summary: {
+    improved: number;
+    regressed: number;
+    flat: number;
+    new: number;
+  };
 }
-//#endregion
-//#region src/knowledge/readiness.d.ts
-interface ScoreKnowledgeReadinessOptions {
-  taskId: string;
-  requirements: KnowledgeRequirement[];
-  evidenceIds?: string[];
-  claimIds?: string[];
-  wikiPageIds?: string[];
-  userAnswers?: Record<string, string>;
-  metadata?: Record<string, unknown>;
-  now?: Date;
+interface DiffScorecardOptions {
+  /** Compare each cell against this commit instead of its immediate predecessor. */
+  baselineCommit?: string;
+  /** |Cohen's d| at/above which a move counts as real. Default 0.5. */
+  minEffect?: number;
+  /** p-value at/below which a move is significant. Default 0.05. */
+  maxP?: number;
+  /**
+   * |delta| at/above which a move counts when statistics are unavailable
+   * (a cell with fewer than 2 samples on either side). Default 0.05.
+   */
+  minDelta?: number;
 }
-export declare function scoreKnowledgeReadiness(options: ScoreKnowledgeReadinessOptions): KnowledgeReadinessReport;
-export declare function blockingKnowledgeEval(report: KnowledgeReadinessReport, options?: {
-  id?: string;
-  minimumScore?: number;
-  emitter?: TraceEmitter;
-}): ControlEvalResult;
-export declare function knowledgeReadinessTracePayload(report: KnowledgeReadinessReport, options?: {
-  passed?: boolean;
-  minimumScore?: number;
-}): Record<string, unknown>;
-export declare function userQuestionsForKnowledgeGaps(gaps: KnowledgeRequirement[]): UserQuestion[];
-export declare function acquisitionPlansForKnowledgeGaps(gaps: KnowledgeRequirement[]): DataAcquisitionPlan[];
+/**
+ * Compare the latest entry of every cell against its predecessor (or against
+ * `baselineCommit`) and classify the move. A move is `improved`/`regressed`
+ * only when it clears both the effect-size and significance gates; otherwise
+ * `flat`. Cells with no prior entry are `new`.
+ */
+export declare function diffScorecard(scorecard: Scorecard, opts?: DiffScorecardOptions): ScorecardDiff;
+/**
+ * Render a scorecard diff as a human-readable report — the block a feature
+ * PR prints. Regressions are listed first; flat cells are summarised, not
+ * enumerated.
+ */
+export declare function formatScorecardDiff(diff: ScorecardDiff): string;
 //#endregion
-//#region src/leaderboard.d.ts
-interface LeaderboardRow {
-  /** Group key — the agent-profile cellId by default, else a caller dimension. */
-  key: string;
-  /** Display label: `harness · model` when the profile carries them, else the key. */
-  label: string;
-  harness?: string;
+//#region src/trace-contracts/types.d.ts
+/**
+ * Minimal structural span the checker reads. The eval-side `Span`
+ * (trace/schema) and the output of {@link contractSpansFromOtlp} satisfy it.
+ */
+interface ContractSpan {
+  spanId?: string;
+  parentSpanId?: string | null;
+  name?: string;
+  /** Eval kinds (`tool`, `llm`, ...) or contract kinds (`TOOL`, `LLM`, ...). */
+  kind?: string;
+  startedAt?: number;
+  endedAt?: number;
+  status?: string;
+  error?: string;
+  /** Typed field on eval-side ToolSpans; OTLP flattenings drop it (see
+   *  {@link contractSpanToolName}). */
+  toolName?: string;
+  /** Typed field on eval-side LlmSpans. */
   model?: string;
-  rank: number;
-  /** Measured runs in this group. */
-  n: number;
-  passRate: number;
-  passRateCi95: [number, number];
-  /** Observability rollup — means over the group; null when no run reported it. */
-  meanCostUsd: number | null;
-  meanTokensIn: number | null;
-  meanTokensOut: number | null;
-  meanWallMs: number | null;
-}
-interface LeaderboardOptions {
-  /** Per-run pass predicate — the benchmark defines what "passed" means (real-build
-   *  hit-rate, test-pass, etc.). The leaderboard owns only the grouping + stats. */
-  passed: (record: RunRecord) => boolean;
-  /** Group key. Default: the agent-profile cellId (harness × model × dimensions),
-   *  so persona/harness/model sweeps separate without parsing labels. */
-  groupBy?: (record: RunRecord) => string;
+  inputTokens?: number;
+  outputTokens?: number;
+  /** Typed field on eval-side ToolSpans. */
+  args?: unknown;
+  /** False when the source observed the call but did not capture its arguments. */
+  argsCaptured?: boolean;
+  attributes?: Record<string, unknown>;
 }
 /**
- * Group `records`, compute pass-rate + Wilson CI + cost/token/latency means per
- * group, and rank by pass-rate (cost as the tiebreaker — cheaper wins a tie).
- * Pure projection: no I/O, deterministic, safe to call on any RunRecord[].
+ * The run a `run` rule reads. The eval-side `Run` satisfies it. When no run is
+ * passed, the checker derives it from the trace's single root span; a trace
+ * with no unique root leaves the run unknown and every `run` rule fails.
  */
-export declare function leaderboard(records: RunRecord[], opts: LeaderboardOptions): LeaderboardRow[];
-//#endregion
-//#region src/metrics.d.ts
-interface TokenPrice {
-  input: number;
-  output: number;
+interface ContractRun {
+  status?: string;
+  startedAt?: number;
+  endedAt?: number;
 }
-/** Per-1K token pricing for exact model ids. */
-export declare const MODEL_PRICING: Record<string, TokenPrice>;
-/** Resolve pricing for a model id: exact table, then family fallback.
- *  Returns null when the id matches nothing (caller decides — never a
- *  silent-zero masquerading as a real $0 cost). */
-export declare function resolveModelPricing(model: string): TokenPrice | null;
-/** True when `model` has known pricing (exact or family). Lets cost-aware
- *  callers distinguish a real $0 from an unpriced model. */
-export declare function isModelPriced(model: string): boolean;
-/** Estimate token count from string length (chars / 4 approximation) */
-export declare function estimateTokens(text: string): number;
-/** Calculate cost in USD from token counts and model. Unknown models warn
- *  once (not a silent zero) and return 0 so callers that ignore pricing keep
- *  working; cost-sensitive callers should gate on {@link isModelPriced}. */
-export declare function estimateCost(inputTokens: number, outputTokens: number, model: string): number;
-//#endregion
-//#region src/paired-delta-test.d.ts
-interface PairedDeltaTestOptions extends PairedBootstrapOptions {
-  /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. */
-  threshold?: number;
-  /** Caller-required paired observations. The exact test may impose a higher minimum. */
-  minPairs?: number;
+/** JSON-safe RegExp form — what the builder normalizes RegExp matchers to. */
+interface SerializedRegex {
+  $regex: string;
+  flags: string;
 }
-interface PairedDeltaTestResult {
-  bootstrap: PairedBootstrapResult;
-  method: 'bootstrap-ci' | 'exact-sign';
-  /** Exact one-sided p-value below the bootstrap minimum; otherwise null. */
-  pValue: number | null;
-  /** Effective observation minimum after accounting for confidence. */
-  minimumPairs: number;
-  sufficient: boolean;
-  /**
-   * The bootstrap interval has zero width (or is non-finite), so it carries no
-   * information about how far the estimate could be wrong and cannot support a
-   * decision in either direction. See {@link pairedDeltaTest}.
-   */
-  indeterminate: boolean;
-  significant: boolean;
+/** Matches when the value equals one of the listed strings. */
+interface OneOfMatcher {
+  oneOf: string[];
 }
-/** Smallest all-positive sample that can clear a one-sided exact sign test. */
-export declare function minimumPairsForPairedDeltaTest(confidence?: number): number;
+type TextMatcher = string | RegExp | SerializedRegex | OneOfMatcher;
 /**
- * Tests whether a paired candidate-minus-baseline delta clears a threshold.
- *
- * At 20 or more pairs, the percentile bootstrap lower bound carries the
- * decision. Below that point the interval is descriptive only, so the function
- * switches to a pre-registered one-sided exact sign test. The exact path is
- * deliberately conservative: it requires both a point estimate above the
- * threshold and enough consistently positive paired differences.
- *
- * ## A zero-width interval is never significant
+ * Proposition over one span. All specified fields must match (AND); `not`
+ * must NOT match. At least one field is required — an empty predicate would
+ * match every span and is rejected.
  *
- * When every paired delta is identical the resample distribution is a point
- * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n
- * identical deltas of g. Neither says the effect is certain — both say the
- * sample carries no information about how far the estimate could be wrong, and
- * `low > threshold` then answers on the point estimate alone. It fails in both
- * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a
- * tie-dominated pass/fail comparison laundered a regression into a
- * noninferiority pass, and `[g, g]` clears every threshold below g with no
- * spread behind it. Under a bounded asymmetric null whose true mean paired
- * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —
- * every sample that misses the drop is exactly that shape, and deciding on
- * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.
+ * `kind` compares against the span's contract kind (`TOOL`, `LLM`, `AGENT`,
+ * ...), resolved by agent-trace-contract's classifier, so eval kinds
+ * (`tool`, `llm`) and OTLP spans that declare no kind are read the same way.
  *
- * So `indeterminate` is reported and `significant` is false whenever the
- * interval has zero width, on BOTH paths: at small n the exact sign test is a
- * test of the MEDIAN and a zero-spread sample is precisely where it stops
- * saying anything about the mean the caller is thresholding.
+ * `tool` resolves through {@link contractSpanToolName}; `model` through the
+ * typed `model` field and the gen_ai model attributes.
  *
- * `threshold` may be negative — that is a noninferiority margin, and it is the
- * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome
- * the percentile bootstrap is not a valid interval at a nonzero margin at all;
- * use {@link decidePairedPromotion}, which routes those to Tango's score
- * interval, rather than thresholding this function's bootstrap directly.
+ * `attr` values match by strict equality, or regex-test when the value is a
+ * RegExp/SerializedRegex and the attribute is a string. Structured attribute
+ * values need `custom`, or an `argument` rule for tool arguments.
  */
-export declare function pairedDeltaTest(before: number[], after: number[], options?: PairedDeltaTestOptions): PairedDeltaTestResult;
-//#endregion
-//#region src/partition-held-out.d.ts
-interface PartitionHeldOutOptions {
-  /** Partition seed. Bumping it reshuffles every assignment — do that only when
-   *  the corpus/policy changes meaningfully so old and new records can't pair.
-   *  Default 'held-out-v1'. */
-  seed?: string;
-  /** Fraction routed to the held-out bucket, in (0, 1). Default 0.5. */
-  holdoutFraction?: number;
-  /**
-   * Minimum held-out ids required for the split to be admissible. Below this,
-   * a paired comparison has too few observations to be significant, so the
-   * partition throws rather than hand back an underpowered holdout. Default 1
-   * (only the empty-holdout degenerate case is rejected). Raise it to enforce
-   * real statistical power. */
-  minHoldout?: number;
-  /** Same floor for the search side. Default 1. */
-  minSearch?: number;
-}
-interface HeldOutPartition {
-  /** Ids the optimizer is allowed to read/optimize against. */
-  search: string[];
-  /** Ids reserved for gated, held-out evaluation. */
-  holdout: string[];
-  seed: string;
-  holdoutFraction: number;
+interface SpanPredicate {
+  name?: TextMatcher;
+  tool?: TextMatcher;
+  model?: TextMatcher;
+  kind?: SpanKind;
+  attr?: Record<string, unknown>;
+  not?: SpanPredicate;
+  custom?: (span: ContractSpan) => boolean;
+  /** Stamped by the builder when `custom` is present. Survives JSON while
+   *  the function does not, so evaluation of a deserialized contract throws
+   *  instead of silently dropping the check. */
+  requiresCustom?: true;
 }
-/**
- * Partition a list of stable ids into disjoint `search` / `holdout` sets by a
- * deterministic hash. Fail-loud:
- *   - empty input throws,
- *   - duplicate ids throw (a dup would be observed on both sides, double-counting),
- *   - a holdout (or search) set below its floor throws (underpowered comparison).
- *
- * Order within each bucket follows the input order, so the result is stable and
- * reproducible.
- */
-export declare function partitionHeldOut(ids: string[], options?: PartitionHeldOutOptions): HeldOutPartition;
-interface TrainSelectionTestOptions<T> {
-  /** Partition seed. Bumping it reshuffles every unit. Default 'train-selection-test-v1'. */
-  seed?: string;
-  /** Share of units routed to selection, in (0, 1). Default 0.25. */
-  selectionFraction?: number;
-  /** Share of units routed to test, in (0, 1). Default 0.25. */
-  testFraction?: number;
-  /**
-   * The independent unit an item samples, such as the task a fresh run repeats.
-   * Every item of one unit lands in the same partition, so a final test never
-   * scores a unit the optimizer read. Default: the item id.
-   */
-  unitOf?: (item: T) => string;
-  /** Minimum items per partition. Defaults 1, 1 and 1. */
-  minTrain?: number;
-  minSelection?: number;
-  minTest?: number;
+declare const ORDER_MODES: readonly ['start-order', 'finish-before-start', 'all-occurrences'];
+type OrderMode = (typeof ORDER_MODES)[number];
+declare const ARGUMENT_OCCURRENCES: readonly ['first', 'last', 'any', 'all'];
+type ArgumentOccurrence = (typeof ARGUMENT_OCCURRENCES)[number];
+declare const ARGUMENT_TYPES: readonly ['string', 'number', 'boolean', 'object', 'array', 'null'];
+type ArgumentType = (typeof ARGUMENT_TYPES)[number];
+type ArgumentCheck = {
+  op: 'exists';
+} | {
+  op: 'equals';
+  value: unknown;
+} | {
+  op: 'oneOf';
+  values: unknown[];
+} | {
+  op: 'type';
+  type: ArgumentType;
+};
+/** A tool whose calls change state outside the run. */
+interface RetryWrite {
+  tool: string;
+  /** RFC 6901 JSON Pointer to the idempotency key in the call's arguments.
+   *  Without it, a repeated call of this tool always fails. */
+  idempotencyKey?: string;
 }
-interface TrainSelectionTestPartition<T> {
-  /** Items an optimizer may learn from. */
-  train: T[];
-  /** Items an optimizer may use to accept candidates and stop. */
-  selection: T[];
-  /** Items reserved for the final comparison. */
-  test: T[];
-  seed: string;
-  selectionFraction: number;
-  testFraction: number;
+interface RunChecks {
+  /** The run's status is `completed` and its end time is recorded. A failed
+   *  or aborted run fails this check. */
+  requireCompleted?: boolean;
+  /** The run's status must be one of these. */
+  allowedStatuses?: RunStatus[];
+  /** End minus start, in milliseconds. An unknown duration fails. */
+  maxDurationMs?: number;
 }
-/**
- * Partition items into disjoint train, selection and test sets, the three
- * partitions `compareOptimizationMethods` and Runtime `improve()` take.
- *
- * Assignment is per unit: `hashToUnit(unit, seed)` below `testFraction` is
- * test, below `testFraction + selectionFraction` is selection, and the rest is
- * train. A unit therefore keeps its partition as the corpus grows, and items of
- * one unit never straddle two partitions. Fail-loud like `partitionHeldOut`:
- * empty input, a duplicate or empty id, an empty unit, fractions outside
- * (0, 1) or summing to 1 or more, and a partition below its floor all throw.
- * Order within each partition follows the input order.
- */
-export declare function partitionTrainSelectionTest<T extends {
-  id: string;
-}>(items: readonly T[], options?: TrainSelectionTestOptions<T>): TrainSelectionTestPartition<T>;
-//#endregion
-//#region src/product-benchmark/export.d.ts
-interface ProductBenchmarkExportOptions {
-  /** Source eval run directories, each containing a `records.jsonl` of RunRecord rows. */
-  readonly runDirs: readonly string[];
-  /** Destination directory for the bundle (manifest + records + materialized source runs). */
-  readonly outDir: string;
-  readonly projectId: string;
-  readonly benchmarkId: string;
-  /** Repo-relative path of the product's canonical agent profile source. */
-  readonly agentProfilePath: string;
-  /** Pass threshold applied when a row carries no explicit `outcome.raw.pass`. Default 0.7. */
-  readonly passThreshold?: number;
-  /**
-   * First scenario tag. Defaults to `projectId` with a trailing `-agent`
-   * stripped (`tax-agent` → `tax`), matching the product exporters.
-   */
-  readonly scenarioTagPrefix?: string;
-  /** Profile id used when a row has no `agentProfile.profileId`. Defaults to the row's arm id. */
-  readonly fallbackProfileId?: string;
-  /** Arm mutable surfaces recorded in the manifest. Defaults to the full superset. */
-  readonly mutableSurfaces?: readonly string[];
-  /**
-   * Copy each run dir into `<outDir>/source-runs/` and record
-   * bundle-relative artifact paths (portable, self-contained). When false,
-   * artifacts keep absolute paths into the original run dirs. Default true.
-   */
-  readonly materializeSourceRuns?: boolean;
+type ContractRule = {
+  kind: 'always' | 'never' | 'eventually';
+  label: string;
+  p: SpanPredicate;
+} | {
+  kind: 'precedes';
+  label: string;
+  a: SpanPredicate;
+  b: SpanPredicate;
+  order?: OrderMode;
+} | {
+  kind: 'neverUnless';
+  label: string;
+  p: SpanPredicate;
+  prior: SpanPredicate;
+  order?: OrderMode;
+} | {
+  kind: 'atMost';
+  label: string;
+  p: SpanPredicate;
+  max: number;
+} | {
+  kind: 'tokensAtMost';
+  label: string;
+  p: SpanPredicate;
+  max: number;
+} | ({
+  kind: 'run';
+  label: string;
+} & RunChecks) | {
+  kind: 'argument';
+  label: string;
+  p: SpanPredicate;
+  /** RFC 6901 JSON Pointer into the call's arguments; `''` is the whole value. */
+  pointer: string;
+  check: ArgumentCheck;
+  /** Which matching calls must satisfy the check. Default `all`. */
+  occurrence?: ArgumentOccurrence;
+} | {
   /**
-   * Override split classification for a row. Return undefined to fall back
-   * to the default (`outcome.raw.safety === 1` → safety, then splitTag).
+   * The tools the harness offered the model (the OTel GenAI
+   * `gen_ai.tool.definitions` attribute) are recorded, every tool call is
+   * one of them, and, when `declared` is set, every offered tool is
+   * declared. A trace that records no offered tools fails: what the
+   * harness enforced is unknown.
    */
-  readonly classifySplit?: (record: RunRecord) => ProductBenchmarkSplit | undefined;
-  /** Recovers a tool-call count when the row's raw bag carries none (e.g. from turn artifacts). */
-  readonly toolCallFallback?: (record: RunRecord, runDir: string) => number;
-  /** Backend version recorded per row. Defaults to the cwd package.json's `@tangle-network/sandbox` range. */
-  readonly backendVersion?: string;
+  kind: 'toolsOffered';
+  label: string;
+  declared?: string[];
+} | {
   /**
-   * Explicit substrate versions for the manifest, merged over what the cwd
-   * package.json / node_modules resolve. Use when a substrate package is not
-   * installed where the export runs — the validator refuses an `'unknown'`
-   * version, so provide the real one rather than shipping the sentinel.
+   * A tool called again with the same arguments repeats its side effect.
+   * A repeated `reads` call passes; a repeated `writes` call passes only
+   * when every call carries the same idempotency key; a repeated call of
+   * any other tool fails, because its side effect is unknown.
    */
-  readonly substrate?: Partial<ProductBenchmarkManifest['substrate']>;
-}
-interface ProductBenchmarkSingleRunExportOptions extends Omit<ProductBenchmarkExportOptions, 'runDirs'> {
-  readonly runDir: string;
+  kind: 'retrySafe';
+  label: string;
+  reads?: string[];
+  writes?: RetryWrite[];
+};
+/** One legitimate alternate path. The contract passes only when the base
+ *  rules pass and at least one alternative passes completely. */
+interface ContractAlternative {
+  id: string;
+  rules: ContractRule[];
 }
-interface ProductBenchmarkExportResult {
-  readonly manifestPath: string;
-  readonly recordsPath: string;
-  readonly records: number;
+/** Serializable plain object — `traceContract(name)....build()` output. */
+interface TraceContract {
+  name: string;
+  /** What the contract protects, for reports and `explainTraceContract`. */
+  description?: string;
+  /** Evaluate only the span matching `root` and its descendants. Exactly one
+   *  span must match; zero or several make every rule an error. */
+  scope?: {
+    root: SpanPredicate;
+  };
+  rules: ContractRule[];
+  alternatives?: ContractAlternative[];
 }
-/** Repo identity from the exporting process's cwd. `'unknown'` values are flagged by `validateProductBenchmarkRun`. */
-export declare function productBenchmarkRepoIdentity(): ProductBenchmarkManifest['repo'];
-/** Single-run convenience wrapper over `exportProductBenchmarkRuns`. */
-export declare function exportProductBenchmark(options: ProductBenchmarkSingleRunExportOptions): ProductBenchmarkExportResult;
-/**
- * Export one or more product eval run dirs into a validated product
- * benchmark bundle at `outDir`. Both the manifest and every record are
- * run through the contract validators before anything is written.
- */
-export declare function exportProductBenchmarkRuns(options: ProductBenchmarkExportOptions): ProductBenchmarkExportResult;
-//#endregion
-//#region src/product-benchmark/index.d.ts
-declare const productBenchmarkSplits: readonly ['practice', 'dev', 'holdout', 'safety', 'sentinel'];
-type ProductBenchmarkSplit = (typeof productBenchmarkSplits)[number];
-interface ProductBenchmarkRepoRef {
-  readonly url: string;
-  readonly commit: string;
-  readonly branch: string;
+type ContractStatus = 'pass' | 'fail' | 'error';
+interface ContractViolation {
+  rule: string;
+  spanId?: string;
+  detail: string;
 }
-interface ProductBenchmarkSubstrateVersions {
-  readonly agentEval: string;
-  readonly agentRuntime: string;
-  readonly agentInterface: string;
-  readonly sandbox: string;
-  readonly agentBench?: string;
+/** What happened to one rule. `error` means it could not be evaluated. */
+interface ContractRuleExecution {
+  rule: string;
+  /** Set for rules inside `alternatives`. */
+  alternative?: string;
+  status: ContractStatus;
+  violations: number;
+  error?: string;
 }
-interface ProductBenchmarkProfileRef {
-  readonly id: string;
-  readonly profileHash: string;
-  readonly agentProfilePath: string;
+interface ContractVerdict extends DefaultVerdict {
+  /** Contract name — keys this verdict in multi-contract reports. */
+  contract: string;
+  status: ContractStatus;
+  /** `status === 'pass'`. */
+  valid: boolean;
+  /** Fraction of rules that passed, in [0, 1]. An errored rule is not a pass. */
+  score: number;
+  /** 0|1 per evaluated base rule, keyed by label, plus `anyOf` when the
+   *  contract has alternatives. Errored rules are absent, never 0. */
+  scores: Record<string, number>;
+  violations: ContractViolation[];
+  ruleExecutions: ContractRuleExecution[];
+  /** Why rules errored, one line per errored rule or scope failure. */
+  errors: string[];
 }
-interface ProductBenchmarkArm {
-  readonly id: string;
-  readonly profileId: string;
-  readonly mutableSurfaces: readonly string[];
-  readonly policyAxes: Record<string, unknown>;
+interface ContractCheckResult {
+  verdicts: ContractVerdict[];
+  /** `error` if any verdict errored, else `fail` if any failed, else `pass`. */
+  status: ContractStatus;
+  allValid: boolean;
 }
-interface ProductBenchmarkScenario {
-  readonly id: string;
-  readonly split: ProductBenchmarkSplit;
-  readonly tags: readonly string[];
-  readonly sourceAllowedForSynthesis: boolean;
+interface EvaluateTraceContractOptions {
+  /** The run `run` rules read. Ignored under `scope`, where the scope root is the run. */
+  run?: ContractRun;
 }
-interface ProductBenchmarkBudgets {
-  readonly maxUsd: number;
-  readonly maxCells: number;
-  readonly maxWallMs: number;
+//#endregion
+//#region src/trace-contracts/builder.d.ts
+interface OrderOptions {
+  order?: OrderMode;
 }
-interface ProductBenchmarkManifest {
-  readonly schemaVersion: 1;
-  readonly projectId: string;
-  readonly benchmarkId: string;
-  readonly repo: ProductBenchmarkRepoRef;
-  readonly substrate: ProductBenchmarkSubstrateVersions;
-  readonly profiles: readonly ProductBenchmarkProfileRef[];
-  readonly arms: readonly ProductBenchmarkArm[];
-  readonly scenarios: readonly ProductBenchmarkScenario[];
-  readonly budgets: ProductBenchmarkBudgets;
-  readonly expectedArtifactDir: string;
+interface ArgumentRuleOptions {
+  pointer: string;
+  check: ArgumentCheck;
+  occurrence?: ArgumentOccurrence;
 }
-interface ProductBenchmarkValidationReport {
-  readonly manifestPath: string;
-  readonly recordsPath: string;
-  readonly records: number;
-  /** Manifest repo fields that are empty or the `'unknown'` export sentinel. */
-  readonly repoFailures: readonly string[];
-  /** Manifest substrate versions that are empty or the `'unknown'` export
-   *  sentinel — a bundle without substrate identity is not reproducible. */
-  readonly substrateFailures: readonly string[];
-  readonly projects: readonly string[];
-  readonly benchmarks: readonly string[];
-  readonly arms: readonly string[];
-  readonly scenarios: readonly string[];
-  readonly passed: number;
-  readonly failed: number;
-  readonly inputTokens: number;
-  readonly outputTokens: number;
-  readonly costUsd: number;
-  readonly wallMs: number;
-  readonly integrityFailures: readonly string[];
-  readonly missingArtifacts: readonly string[];
+declare class TraceContractBuilder {
+  private readonly name;
+  private readonly rules;
+  private readonly alternativeList;
+  private scopeRoot;
+  constructor(name: string);
+  /** Every span in the trace must satisfy `p`. */
+  always(p: SpanPredicate, label?: string): this;
+  /** No span in the trace may satisfy `p`. */
+  never(p: SpanPredicate, label?: string): this;
+  /** At least one span in the trace must satisfy `p`. */
+  eventually(p: SpanPredicate, label?: string): this;
+  /** Both `a` and `b` must occur, with `a` first under `options.order`
+   *  (default `start-order`). For "every `b` needs an earlier `a`, and a run
+   *  without `b` passes", use {@link neverUnless}. */
+  precedes(a: SpanPredicate, b: SpanPredicate, label?: string, options?: OrderOptions): this;
+  /** Every `p`-match needs an earlier `prior`-match; a run without `p` passes. */
+  neverUnless(p: SpanPredicate, prior: SpanPredicate, label?: string, options?: OrderOptions): this;
+  /** At most `max` spans may satisfy `p`. */
+  atMost(p: SpanPredicate, max: number, label?: string): this;
+  /** Spans satisfying `p` may use at most `max` input+output tokens in total. */
+  tokensAtMost(p: SpanPredicate, max: number, label?: string): this;
+  /** Terminal status and duration of the run. */
+  run(checks: RunChecks, label?: string): this;
+  /** A JSON Pointer check on the arguments of calls matching `p`. */
+  argument(p: SpanPredicate, options: ArgumentRuleOptions, label?: string): this;
+  /** The trace records the tools offered to the model, every call is one of
+   *  them, and, with `declared`, every offered tool is declared. */
+  toolsOffered(declared?: string[], label?: string): this;
+  /** No call repeats a side effect that is not proven safe to repeat. */
+  retrySafe(options: {
+    reads?: string[];
+    writes?: RetryWrite[];
+  }, label?: string): this;
+  /** Evaluate only the subtree rooted at the one span matching `root`. */
+  scope(root: SpanPredicate): this;
+  /** A named alternate path; at least one alternative must pass completely. */
+  alternative(id: string, define: (branch: TraceContractBuilder) => TraceContractBuilder): this;
+  build(): TraceContract;
+  private where;
+  private add;
 }
-export declare function readProductBenchmarkManifest(path: string): ProductBenchmarkManifest;
+export declare function traceContract(name: string): TraceContractBuilder;
+//#endregion
+//#region src/trace-contracts/evaluate.d.ts
 /**
- * Fail-loud gate over a bundle directory: locates the manifest + records,
- * runs `validateProductBenchmarkRun`, and throws with every repo,
- * integrity, and artifact failure listed. Returns the report when clean.
+ * Evaluate one contract over a span sequence. Pure and synchronous — works on
+ * `Span[]` from a TraceStore, `contractSpansFromOtlp` output, or any array
+ * satisfying `ContractSpan`. Throws
+ * `ValidationError` for a malformed contract; a rule that cannot be evaluated
+ * is reported as `error` in the verdict instead.
  */
-export declare function assertProductBenchmarkRun(runDir: string): ProductBenchmarkValidationReport;
+export declare function evaluateTraceContract(contract: TraceContract, spans: readonly ContractSpan[], options?: EvaluateTraceContractOptions): ContractVerdict;
+/**
+ * Evaluate many contracts over one span sequence. Throws on an empty
+ * contract list — `allValid: true` over zero contracts is a silent pass.
+ */
+export declare function checkTraceContracts(spans: readonly ContractSpan[], contracts: readonly TraceContract[], options?: EvaluateTraceContractOptions): ContractCheckResult;
+//#endregion
+//#region src/trace-contracts/explain.d.ts
+/** Plain-language statement of everything a contract checks, one rule per line. */
+export declare function explainTraceContract(contract: TraceContract): string;
 //#endregion
-//#region src/runtime-trajectory.d.ts
-interface RuntimeTrajectoryHookEvent {
-  id: string;
-  runId: string;
-  scenarioId?: string;
-  target: string;
-  phase: string;
-  timestamp: number;
-  stepIndex?: number;
-  parentId?: string;
-  payload?: unknown;
-  metadata?: Record<string, unknown>;
+//#region src/trace-contracts/spec.d.ts
+interface ToolArgumentSpec {
+  tool: string;
+  /** RFC 6901 JSON Pointer into the call's arguments; `''` is the whole value. */
+  pointer: string;
+  /** Which calls must satisfy the check. Default `all`. */
+  occurrence?: ArgumentOccurrence;
+  /** Exactly one of `exists`, `equals`, `oneOf`, `type`. */
+  exists?: true;
+  equals?: unknown;
+  oneOf?: unknown[];
+  type?: ArgumentType;
 }
-interface RuntimeTrajectoryRecord {
-  id?: string;
-  scenarioId?: string;
-  splitTag?: RunSplitTag;
-  runtimeEvents?: unknown;
-  [key: string]: unknown;
+interface ToolsSpec {
+  /** Each must be called at least once. */
+  required?: string[];
+  /** None may be called. */
+  forbidden?: string[];
+  /** Every tool call must name one of these; a call without a name fails. */
+  allowed?: string[];
+  /** Ceiling on tool calls of any name. */
+  maxCalls?: number;
+  /** Ceiling per tool name. */
+  maxCallsPerTool?: Record<string, number>;
+  /** Each tool is called, and each comes before the next under `orderMode`. */
+  requiredOrder?: string[];
+  /** Default `start-order`. */
+  orderMode?: OrderMode;
+  arguments?: ToolArgumentSpec[];
+  /** The trace records the tools the harness offered the model
+   *  (`gen_ai.tool.definitions`), every call is one of them, and, with
+   *  `allowed`, every offered tool is allowed. */
+  enforced?: true;
 }
-interface RuntimeTrajectoryRunRecord {
-  runId: string;
-  scenarioId?: string;
-  splitTag: RunSplitTag;
+interface RetriesSpec {
+  /** Tools that change nothing outside the run: repeating a call is safe. */
+  reads?: string[];
+  /** Tools that change state outside the run: a repeated call must carry one
+   *  idempotency key. Any other tool called twice with the same arguments fails. */
+  writes?: RetryWrite[];
 }
-interface RuntimeTrajectoryEvidenceSummary {
-  recordCount: number;
-  recordWithRuntimeEventsCount: number;
-  runtimeRunCount: number;
-  lifecycleEventCount: number;
-  defaultedSplitCount: number;
+interface LlmSpec {
+  maxCalls?: number;
+  /** Input plus output tokens across LLM calls; an unrecorded count fails. */
+  maxTotalTokens?: number;
+  /** Every LLM call must name one of these models; a call without one fails. */
+  allowedModels?: string[];
 }
-interface RuntimeTrajectoryEvidenceProjection {
-  runs: RuntimeTrajectoryRunRecord[];
-  events: RuntimeTrajectoryHookEvent[];
-  summary: RuntimeTrajectoryEvidenceSummary;
-  diagnostics: string[];
+/** The checks one path carries: the base contract or one alternative. */
+interface PathSpec {
+  run?: RunChecks;
+  tools?: ToolsSpec;
+  llm?: LlmSpec;
+  retries?: RetriesSpec;
+  /** Low-level rules in their serialized form. */
+  rules?: ContractRule[];
 }
-interface ProjectRuntimeTrajectoryEvidenceOptions<TRecord extends RuntimeTrajectoryRecord = RuntimeTrajectoryRecord> {
-  records: TRecord[];
-  defaultSplitTag?: RunSplitTag;
-  recordIdOf?: (record: TRecord, index: number) => string | undefined;
-  scenarioIdOf?: (record: TRecord, index: number) => string | undefined;
+interface AlternativeSpec extends PathSpec {
+  id: string;
 }
-export declare function projectRuntimeTrajectoryEvidence<TRecord extends RuntimeTrajectoryRecord>(options: ProjectRuntimeTrajectoryEvidenceOptions<TRecord>): RuntimeTrajectoryEvidenceProjection;
+interface TraceContractSpec extends PathSpec {
+  name: string;
+  description?: string;
+  /** Check only the one span matching `root` and its descendants: one
+   *  sub-agent (`{ "kind": "AGENT", "name": "researcher" }`) or one search
+   *  node (`{ "attr": { "agent.branch.id": "n7" } }`). */
+  scope?: {
+    root: SpanPredicate;
+  };
+  /** At least one alternative must pass completely, as well as the base checks. */
+  alternatives?: {
+    anyOf: AlternativeSpec[];
+  };
+}
+/**
+ * Parse and compile a declarative contract. Throws `ValidationError` naming
+ * the path of the first problem: an unknown key, a wrong type, an unknown
+ * status, or a contract that checks nothing.
+ */
+export declare function compileTraceContractSpec(input: unknown): TraceContract;
+interface ContractLintFinding {
+  /** `error`: the contract can never pass, or contradicts itself.
+   *  `warning`: the contract works but likely does not say what was meant. */
+  level: 'error' | 'warning';
+  code: string;
+  path: string;
+  message: string;
+}
+/**
+ * Find contradictions and likely mistakes in a declarative contract. Parses
+ * strictly first, so a malformed spec throws instead of linting.
+ */
+export declare function lintTraceContractSpec(input: unknown): ContractLintFinding[];
 //#endregion
 //#region src/action-policy.d.ts
 interface ActionExecutionPolicy {
@@ -3374,375 +3500,347 @@ interface VerdictCacheStore {
  */
 export declare function fileVerdictCache(path: string): VerdictCacheStore;
 //#endregion
-//#region src/anti-slop.d.ts
-interface AntiSlopConfig {
-  /** Domain label — appears in the JudgeScore output */
-  domain?: string;
-  /** Case-insensitive substrings that must not appear. Each occurrence = penalty. */
-  bannedPhrases?: string[];
-  /** Regexes matching opening formulas to penalize (e.g. /^great question/i). */
-  bannedOpenings?: RegExp[];
-  /** Regexes matching hedges (e.g. /i could be wrong/i). Ratio of hedged sentences drives score. */
-  hedgingPatterns?: RegExp[];
-  /** Regexes matching apology padding. */
-  apologyPatterns?: RegExp[];
-  /** Fraction of sentences that can be duplicates before penalty (default 0.15 = 15%). */
-  repetitionThreshold?: number;
-  /** Min output length in chars; below this the turn is deemed too terse. */
-  minLength?: number;
-  /** Max output length in chars; above this the turn is deemed too verbose. */
-  maxLength?: number;
-  /** How heavily each violation class reduces the score (default 1). */
-  penaltyWeights?: Partial<Record<SlopCategory, number>>;
-}
-type SlopCategory = 'banned_phrase' | 'banned_opening' | 'hedging' | 'apology' | 'repetition' | 'length';
-/** Create a reusable Judge function from an anti-slop config. */
-export declare function createAntiSlopJudge(config?: AntiSlopConfig): JudgeFn;
-interface AntiSlopIssue {
-  category: SlopCategory;
-  detail: string;
-  example?: string;
-}
-interface AntiSlopReport {
-  /** 0–10 score; 10 is clean, lower values mean more slop. */
-  score: number;
-  issues: AntiSlopIssue[];
-  /** Count of each category for programmatic aggregation. */
-  counts: Record<SlopCategory, number>;
-}
-/**
- * Pure function — analyze one or more outputs against the config. Exposed
- * separately so consumers can build their own reporters on top.
- */
-export declare function analyzeAntiSlop(outputs: string[], config: Omit<Required<AntiSlopConfig>, 'domain'> & {
-  penaltyWeights: Record<SlopCategory, number>;
-}): AntiSlopReport;
-//#endregion
-//#region src/eval-trace-store.d.ts
-/**
- * Backing persistence for `EvalTraceStore`. The in-memory store is the default;
- * the JSONL file store appends one validated `RunRecord` per line. Both keep an
- * append order so `getBest` / `compareRuns` are reproducible.
- */
-interface RunRecordBackend {
-  append(record: RunRecord): Promise<void>;
-  load(): Promise<RunRecord[]>;
-}
+//#region src/profile/baselines.d.ts
 /**
- * JSONL-backed store at `path`, one `RunRecord` per line. Malformed lines fail
- * loud on load (a corrupt corpus must not silently shrink the analysis set);
- * pass `skipInvalid` only for forensics on a known-bad file.
+ * Strong, generically-useful baseline ROLES — the top zone of an `AgentProfile`
+ * before any domain layer. A product composes one of these with its own
+ * environment description (its sandbox) and its domain guidance, which lives in
+ * the product repo (not here): a profile is `<baseline role> + <environment> +
+ * <domain sections>`. Domain strength is NOT generalized into the substrate —
+ * it is lifted here only once ≥2 products genuinely reuse it.
+ *
+ * Three roles cover the common shapes:
+ *   - `engineerRole`   — builds + verifies real artifacts in a sandbox. Distilled
+ *                        from agent-runtime's `coderProfile` doctrine (minimal
+ *                        correct change, run the checks, fix the root cause,
+ *                        never weaken a test or hide an error).
+ *   - `researcherRole` — gathers from real sources and grounds every claim
+ *                        (cite or mark inferred; never fabricate).
+ *   - `generalistRole` — a strong default: helpful, grounded, verify-before-assert.
+ *
+ * All three are verification-first and domain-agnostic. They describe HOW a
+ * capable IC operates, not WHAT domain it works in.
  */
-export declare function jsonlRunRecordBackend(path: string, opts?: {
-  skipInvalid?: boolean;
-}): RunRecordBackend;
-//#endregion
-//#region src/integrity/preflight.d.ts
-interface ModelPreflight {
-  /** The model id as supplied by the caller. */
-  model: string;
-  /** Membership in the `{baseUrl}/models` served set. */
-  listed: boolean;
-  /**
-   * The router reached a provider for this model. `null` when `probe` was not
-   * requested. True on a 2xx, and also when the provider answered by
-   * exhausting its reasoning budget — that proves reachability while leaving
-   * identity unproven (see `budgetExhausted`).
-   */
-  served: boolean | null;
-  /** HTTP status of the probe. `null` when not probed. */
-  status: number | null;
-  /** Probe body's `error.message` when present, else `null`. */
-  detail: string | null;
-  /**
-   * The probe died on the token budget rather than on the model. Identity is
-   * unproven, not refuted: raise `probeMaxTokens` to resolve it.
-   */
-  budgetExhausted: boolean;
-  /**
-   * Identity of the model that answered the probe, compared against `model`.
-   * `null` when the model was not probed or the probe failed.
-   */
-  substitution: ServedModelCheck | null;
-}
-/** One check Agent Eval asks the caller's endpoint to perform. */
-type ModelEndpointCheck = {
-  readonly kind: 'list-models';
-} | {
-  readonly kind: 'probe';
-  readonly model: string;
-  /** Output-token budget the probe completion may bill. */
-  readonly maxOutputTokens: number;
+/** Senior-IC engineer operating in a sandbox — produces real artifacts and
+ *  verifies them before declaring done. The shared operator foundation for any
+ *  file-producing, tool-using product agent. */
+declare const engineerRole: string;
+/** Researcher who grounds every claim in real sources — gathers, cites, and
+ *  distinguishes verified fact from inference. */
+declare const researcherRole: string;
+/** Strong general-purpose default — helpful, grounded, verifies before asserting. */
+declare const generalistRole: string;
+/** The named baseline roles, for selection by key (e.g. from a product config). */
+declare const BASELINE_ROLES: {
+  readonly engineer: string;
+  readonly researcher: string;
+  readonly generalist: string;
 };
-/**
- * Caller-owned request into the model endpoint. The caller binds the base URL
- * and the credential and returns the raw `Response`; a `list-models` request
- * answers with an OpenAI-compatible `{ data: [{ id }] }` body, and a `probe`
- * request answers with one minimal chat completion for `model`.
- */
-type ModelEndpointRequest = (check: ModelEndpointCheck) => Promise<Response>;
-interface PreflightModelsOptions {
-  /** Caller-owned endpoint request. Agent Eval issues no provider HTTP itself. */
-  request: ModelEndpointRequest;
-  /** Model ids to check. */
-  models: string[];
-  /** When true, additionally spend a small chat probe per model. Default false. */
-  probe?: boolean;
-  /**
-   * Output-token budget per probe. Default `PROBE_MAX_TOKENS`. Lowering it
-   * below the reasoning floor makes healthy reasoning models report
-   * `budgetExhausted` instead of proving their identity.
-   */
-  probeMaxTokens?: number;
+type BaselineRoleKey = keyof typeof BASELINE_ROLES;
+declare namespace index_d_exports {
+  export { AgentProfileSection, BASELINE_ROLES, BaselineRoleKey, ProfileSkill, PromptProfile, applyDomainPatch, baselineProfile, baselineProfileFromRole, engineerRole, generalistRole, prodProfile, profileToSurface, renderProfile, researcherRole, sectionHash };
 }
-interface PreflightOutcome {
-  succeeded: boolean;
-  value: ModelPreflight[] | null;
-  error: string | null;
+/** A named, addressable region of the system prompt. `evolvable` marks whether
+ *  the self-improvement loop is allowed to patch its body; fixed scaffolding
+ *  (e.g. a compliance preamble) sets `evolvable: false`. */
+interface AgentProfileSection {
+  id: string;
+  title: string;
+  body: string;
+  evolvable: boolean;
+}
+/** A skill the agent can invoke. `triggers` are the phrases/conditions that
+ *  should activate it; `scriptsDir` points at its executable scripts when the
+ *  skill ships code (mirrors Harvey's `skills/<name>/scripts/`). */
+interface ProfileSkill {
+  name: string;
+  description: string;
+  triggers: string[];
+  scriptsDir?: string;
+}
+/** The structured system prompt. The first four fields are fixed scaffolding;
+ *  `domain` is the evolvable surface the loop optimizes. */
+interface PromptProfile {
+  role: string;
+  environment: string;
+  toolConventions: string;
+  skills: ProfileSkill[];
+  domain: AgentProfileSection[];
 }
 /**
- * Check that `models` are reachable on the router. Returns a typed outcome —
- * a network failure yields `{ succeeded: false, error }`, never a throw and
- * never a partial result silently reported as success. No retries, no
- * fallbacks.
- *
- * The membership check (one GET) always runs. When `probe` is true, each model
- * additionally gets a small chat probe so a model that is listed but
- * unconfigured (a 401 `model_not_found` from the router) is caught.
+ * Emit the sectioned system prompt in fixed order: role text, then
+ * `## Environment`, `## Tool conventions`, `## Skills` (each skill as
+ * `### <name>` + description + triggers), `## Domain guidance` (each section as
+ * `### <title>` + body). The order is load-bearing — the loop diffs rendered
+ * text, and reordering would make every candidate look like a full rewrite.
  */
-export declare function preflightModels(opts: PreflightModelsOptions): Promise<PreflightOutcome>;
-//#endregion
-//#region src/model-seats.d.ts
-interface ModelSeats {
-  /** The model under evaluation — what the agent itself dispatches with. */
-  worker?: string;
-  /** Judge-panel model ids — thread into `ensembleJudge({ models })`. */
-  judges?: string[];
-  /** Analyst model — `analyzeRuns` / analyst-registry LLM calls. */
-  analyst?: string;
-  /** Reflection or candidate-generation model. */
-  reflection?: string;
-  /** Verifier model — completion/objective checking. */
-  verifier?: string;
-}
-type SeatName = keyof ModelSeats;
+declare function renderProfile(p: PromptProfile): string;
+/** The string `MutableSurface` the self-improvement loop scores — a profile
+ *  renders to exactly the text a candidate is graded on. */
+declare function profileToSurface(p: PromptProfile): string;
 /**
- * Tier presets — plain data, swap or spread freely.
- *
- * A preset names REQUESTED ids, and a requested id is NOT a guarantee of
- * family, of provider, or of liveness. A routing gateway can accept any id
- * below and answer from a different model on HTTP 200, with only the response
- * body's `model` field betraying the swap. The served assertion is the
- * guarantee: gate a run on `assertModelsServed({ probe: true })` and assert
- * `assertServedModel` per call. `assertCrossFamily` over these ids proves the
- * configuration is diverse; only `assertCrossFamilyServed` over the ids that
- * ANSWERED proves the run was.
- *
- * `economy` names ids a live router probe answered from the provider their
- * name implies, so the judge trio spanned three provider families (deepseek /
- * zhipu / google) as configured and, at that probe, as served. A preset is
- * only as good as its last probe: ids go dead and start resolving elsewhere
- * without notice, so re-probe rather than trusting this list.
- *
- * `frontier` is deliberately EMPTY: entitled frontier ids vary per router
- * account, and a hardcoded claude/gpt-5 id 401s on keys that lack it. Supply
- * your own: `{ ...seatPresets.frontier, worker: '<your-frontier-id>', … }` —
- * `resolveSeat` throws on every seat you haven't filled.
+ * The fixed-scaffolding baseline: a `role`/`environment`/`toolConventions`
+ * profile with the stock skill roster and NO domain guidance. The
+ * production profile is this plus shipped domain sections.
  */
-export declare const seatPresets: Record<'economy' | 'frontier', ModelSeats>;
+declare function baselineProfile(args: {
+  role: string;
+  environment?: string;
+  toolConventions?: string;
+  skills?: ProfileSkill[];
+}): PromptProfile;
 /**
- * Read one seat. Blank strings and empty arrays count as unset (env-var
- * plumbing produces them); malformed values (non-string seat, non-array or
- * blank-entry `judges`) throw `ValidationError`. When the seat is unset, an
- * explicit `fallback` is returned (`[fallback]` for `judges` — a one-model
- * panel); without one, `SeatUnsetError`.
+ * Baseline from one of the strong generic roles (`'engineer' | 'researcher' |
+ * 'generalist'`) — the common case: pick a role foundation, optionally override
+ * the environment to describe THIS product's sandbox, then layer domain via
+ * `prodProfile`. Domain guidance stays in the product repo; this only supplies
+ * the generically-useful role + stock scaffolding.
  */
-export declare function resolveSeat(seats: ModelSeats, seat: 'judges', fallback?: string): string[];
-export declare function resolveSeat(seats: ModelSeats, seat: Exclude<SeatName, 'judges'>, fallback?: string): string;
-export declare function resolveSeat(seats: ModelSeats, seat: SeatName, fallback?: string): string | string[];
+declare function baselineProfileFromRole(role: BaselineRoleKey, args?: {
+  environment?: string;
+  toolConventions?: string;
+  skills?: ProfileSkill[];
+}): PromptProfile;
+/** The production profile: the baseline scaffolding plus the domain sections
+ *  shipped after self-improvement. Differs from the baseline ONLY in `domain`
+ *  (and any skills the caller layered into the baseline) — the role,
+ *  environment, and tool conventions are carried through unchanged. */
+declare function prodProfile(baseline: PromptProfile, shipped: AgentProfileSection[]): PromptProfile;
+/**
+ * Section-scoped edit — replace the body of ONE domain section by id, leaving
+ * every other section byte-identical. This is how the loop targets a single
+ * evolvable surface. Throws on an unknown section id (a patch that hits no
+ * section is a silent no-op otherwise — fail loud). Throws when the targeted
+ * section is `evolvable: false`.
+ */
+declare function applyDomainPatch(p: PromptProfile, sectionId: string, newBody: string): PromptProfile;
+/**
+ * Content hash of a single section — its loop identity. Delegates to the
+ * campaign provenance `surfaceContentHash` (the same helper the loop's
+ * provenance record uses for full surfaces), so a section hash is byte-for-byte
+ * comparable with the `sha256:`-prefixed content hashes that record carries.
+ * Two sections that render identically (same title + body) share a hash; the
+ * `id` and `evolvable` flag are excluded — they are routing metadata, not
+ * content.
+ */
+declare function sectionHash(section: AgentProfileSection): string;
 //#endregion
-//#region src/prompt-registry.d.ts
+//#region src/budget-guard.d.ts
+export declare class BudgetBreachError extends AgentEvalError {
+  dimension: keyof BudgetSpec;
+  limit: number;
+  attempted: number;
+  constructor(dimension: keyof BudgetSpec, limit: number, attempted: number);
+}
+export declare class BudgetGuard {
+  private consumed;
+  private emitter;
+  private budget;
+  private startedAt;
+  constructor(emitter: TraceEmitter, budget: BudgetSpec, now?: () => number);
+  /** Record consumption. Throws `BudgetBreachError` if any dimension exceeds its cap. */
+  charge(delta: Partial<Record<keyof BudgetSpec, number>>, spanId?: string): Promise<void>;
+  /** Convenience: advance wall-clock budget based on elapsed wall time. */
+  tickWall(nowMs: number, spanId?: string): Promise<void>;
+  get state(): Record<keyof BudgetSpec, number>;
+}
+//#endregion
+//#region src/client.d.ts
 /**
- * Versioned prompt registry.
+ * ProductClient — configurable HTTP client for exercising any agent's APIs.
  *
- * Every prompt used in an eval run is registered with an explicit version.
- * Reports include the content hash so A/B compares are rigorous: if the
- * hash changes between two reports, the prompt actually changed; if it
- * matches, the variance is elsewhere.
+ * Routes are config, not hardcoded. Each agent provides its own RouteMap.
  *
- * Hash is SHA-256(content), truncated to 12 hex chars for readability.
- * Uses the Web Crypto API (works in Workers, Node 22+, browsers).
+ * @deprecated A one-product REST client does not belong in the generic
+ * substrate. Moves to the product repo in the next major
+ * (tangle-network/agent-runtime#694); vendor it if you construct it directly.
  */
-interface PromptHandle {
-  /** Stable human-readable id, e.g. 'browser.system' */
-  id: string;
-  /** Caller-chosen version string, e.g. 'v3' or '2026-04-20' */
-  version: string;
-  /** SHA-256 of content, 12-hex-char prefix */
-  hash: string;
-  /** Full prompt body */
-  content: string;
-}
-export declare class PromptRegistry {
-  private readonly entries;
+export declare class ProductClient {
+  private baseUrl;
+  private routes;
+  private cookies;
+  private timeoutMs;
+  constructor(config: ProductClientConfig);
+  private route;
   /**
-   * Register a prompt. Re-registering the same id+version with DIFFERENT
-   * content throws — versions are immutable. Re-registering with the SAME
-   * content is a no-op (idempotent).
+   * Single HTTP boundary for every JSON request. Aborts after `timeoutMs`,
+   * checks `res.ok` BEFORE parsing, and throws `${method} ${path} failed:
+   * HTTP ${status} — ${body}` on a non-ok response so a 4xx/5xx error body
+   * can never be parsed as a success shape. Callers inspect the resolved
+   * value only after a successful return.
    */
-  register(id: string, version: string, content: string): Promise<PromptHandle>;
-  /** Look up a registered prompt. Throws if unknown — no implicit defaults. */
-  get(id: string, version: string): PromptHandle;
-  /** Return all versions of an id, newest-first (lex-descending on version). */
-  listVersions(id: string): PromptHandle[];
-  /** Snapshot the whole registry — useful for including in reports. */
-  list(): PromptHandle[];
-  /** Verify a hash against registered content. Returns null if not found. */
-  verifyHash(id: string, version: string, expectedHash: string): boolean | null;
+  private request;
+  /**
+   * Read a required collection field. A missing/non-array field is a contract
+   * violation (wrong route, drift, partial body) — surface it loud instead of
+   * masking it as a healthy empty set. A genuinely empty `[]` passes through.
+   */
+  private requireArray;
+  signup(name: string, email: string, password: string): Promise<{
+    userId: string;
+  }>;
+  login(email: string, password: string): Promise<void>;
+  createWorkspace(name: string, type?: string): Promise<string>;
+  createThread(workspaceId: string): Promise<string>;
+  chat(workspaceId: string, threadId: string, content: string, _opts?: {
+    blockPatterns?: RegExp[];
+  }): Promise<{
+    text: string;
+    blocks: {
+      type: string;
+      title: string;
+    }[];
+  }>;
+  getTasks(workspaceId: string): Promise<{
+    id: string;
+    title: string;
+    status: string;
+    priority: string;
+  }[]>;
+  getEvents(workspaceId: string): Promise<{
+    id: string;
+    title: string;
+    type: string;
+  }[]>;
+  getApprovals(workspaceId: string): Promise<{
+    id: string;
+    title: string;
+    status: string;
+    type: string;
+  }[]>;
+  getVaultTree(workspaceId: string): Promise<string[]>;
+  approveAction(workspaceId: string, id: string): Promise<void>;
+  rejectAction(workspaceId: string, id: string, reason: string): Promise<void>;
+  getGenerations(workspaceId: string): Promise<{
+    id: string;
+    type: string;
+    prompt: string;
+  }[]>;
+  /** Generic GET for custom routes */
+  get(path: string): Promise<Record<string, unknown>>;
+  /** Generic POST for custom routes */
+  post(path: string, body: Record<string, unknown>): Promise<Record<string, unknown>>;
+  /** Generic PATCH for custom routes */
+  patch(path: string, body: Record<string, unknown>): Promise<Record<string, unknown>>;
 }
-/** SHA-256(content) → first 12 hex chars. Stable across runtimes. */
-export declare function hashContent(content: string): Promise<string>;
 //#endregion
-//#region src/reference-replay.d.ts
+//#region src/control-runtime.d.ts
+export declare function runAgentControlLoop<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult>(config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>>;
+//#endregion
+//#region src/control-runtime/policy.d.ts
+export declare function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
+export declare function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
+//#endregion
+//#region src/analyst/steer-firewall.d.ts
 /**
- * Reference replay — score an agent against withheld historical outcomes.
+ * THE steer firewall. Fail-loud guard for any path that admits analyst findings
+ * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any
+ * finding whose provenance is a judge verdict, rather than let `J` leak into the
+ * loop. Returns the findings unchanged for chaining.
  *
- * This is the generic version of the public-audit replay pattern:
- * run a candidate system on an old task, keep the reference answers hidden
- * until after execution, then score recall/precision and gate promotion
- * across train/dev/test/holdout splits.
+ * Call this at the chokepoint where a detector that ALSO scores/gates has its
+ * findings turned into a steer (the judge-and-steer dual-role case). It keys on
+ * provenance, so it correctly admits evidence-less trace-analyst observations and
+ * correctly rejects an artifact-citing judge verdict — the cases an evidence
+ * check gets backwards.
+ *
+ * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge
+ * whose output is laundered through a hand-built finding with no provenance flag
+ * is out of its reach — provenance must be honestly set at every judge→finding
+ * lift (today: createJudgeAdapter). That is why the integrity rule lives at the
+ * lift site, and why ProposeContext.judgeScores?: never is the complementary
+ * compile-time tripwire on the obvious direct channel.
  */
-type ReferenceReplaySplit = 'train' | 'dev' | 'test' | 'holdout';
-interface ReferenceReplayItem {
-  id: string;
-  title: string;
-  description?: string;
-  severity?: string;
-  tags?: string[];
-  weight?: number;
-}
-interface ReferenceReplayCandidate {
-  id: string;
-  title: string;
-  description?: string;
-  severity?: string;
-  tags?: string[];
-  metadata?: Record<string, unknown>;
-}
-interface ReferenceReplayMatch {
-  scenarioId: string;
-  referenceId: string;
-  candidateId: string | null;
-  score: number;
-  matched: boolean;
-  weight: number;
-  reason: string;
-}
-interface ReferenceReplayScenarioScore {
-  scenarioId: string;
-  split: ReferenceReplaySplit;
-  matched: number;
-  total: number;
-  falsePositives: number;
-  matchedWeight: number;
-  totalWeight: number;
-  precision: number;
-  recall: number;
-  f1: number;
-  matches: ReferenceReplayMatch[];
-}
-interface ReferenceReplayAggregate {
-  matched: number;
-  total: number;
-  falsePositives: number;
-  matchedWeight: number;
-  totalWeight: number;
-  precision: number;
-  recall: number;
-  f1: number;
-  weightedRecall: number;
-}
-interface ReferenceReplayScore {
-  scenarios: ReferenceReplayScenarioScore[];
-  aggregate: ReferenceReplayAggregate;
-  bySplit: Partial<Record<ReferenceReplaySplit, ReferenceReplayAggregate>>;
-}
-interface ReferenceReplayCaseRun<Input = unknown> {
-  caseId: string;
-  split: ReferenceReplaySplit;
-  input: Input;
-  metadata?: Record<string, unknown>;
-  references: ReferenceReplayItem[];
-  candidates: ReferenceReplayCandidate[];
-  score: ReferenceReplayScenarioScore;
-  durationMs: number;
-  error?: string;
-}
-interface ReferenceReplayRun<Input = unknown> {
-  id: string;
-  variantId?: string;
-  startedAt: number;
-  completedAt: number;
-  durationMs: number;
-  cases: ReferenceReplayCaseRun<Input>[];
-  score: ReferenceReplayScore;
-  metadata?: Record<string, unknown>;
-}
+export declare function assertNoJudgeVerdict(findings: ReadonlyArray<AnalystFinding>, context?: string): ReadonlyArray<AnalystFinding>;
 //#endregion
-//#region src/steering.d.ts
-interface SteeringRolePrompt {
-  system?: string;
-  append?: string;
-}
-interface SteeringBundle {
-  id: string;
-  coderPrompt?: string;
-  continuePrompt?: string;
-  reviewerPrompts?: Record<string, string>;
-  skills?: string[];
-  rolePrompts?: Record<string, SteeringRolePrompt>;
+//#region src/knowledge/readiness.d.ts
+interface ScoreKnowledgeReadinessOptions {
+  taskId: string;
+  requirements: KnowledgeRequirement[];
+  evidenceIds?: string[];
+  claimIds?: string[];
+  wikiPageIds?: string[];
+  userAnswers?: Record<string, string>;
   metadata?: Record<string, unknown>;
+  now?: Date;
 }
+export declare function scoreKnowledgeReadiness(options: ScoreKnowledgeReadinessOptions): KnowledgeReadinessReport;
+export declare function blockingKnowledgeEval(report: KnowledgeReadinessReport, options?: {
+  id?: string;
+  minimumScore?: number;
+  emitter?: TraceEmitter;
+}): ControlEvalResult;
+export declare function knowledgeReadinessTracePayload(report: KnowledgeReadinessReport, options?: {
+  passed?: boolean;
+  minimumScore?: number;
+}): Record<string, unknown>;
+export declare function userQuestionsForKnowledgeGaps(gaps: KnowledgeRequirement[]): UserQuestion[];
+export declare function acquisitionPlansForKnowledgeGaps(gaps: KnowledgeRequirement[]): DataAcquisitionPlan[];
 //#endregion
-//#region src/steering-optimizer.d.ts
-type SteeringOptimizerBackend = 'pairwise';
-interface SteeringOptimizationRow {
-  variantId: string;
-  scenarioId: string;
-  bundle: SteeringBundle;
-  score: RunScore;
-  metadata?: Record<string, unknown>;
+//#region src/metrics.d.ts
+interface TokenPrice {
+  input: number;
+  output: number;
 }
-interface SteeringOptimizationSelector {
-  backend: SteeringOptimizerBackend;
-  signature?: string;
-  labels?: string[];
-  rationale?: string;
+/** Per-1K token pricing for exact model ids. */
+export declare const MODEL_PRICING: Record<string, TokenPrice>;
+/** Resolve pricing for a model id: exact table, then family fallback.
+ *  Returns null when the id matches nothing (caller decides — never a
+ *  silent-zero masquerading as a real $0 cost). */
+export declare function resolveModelPricing(model: string): TokenPrice | null;
+/** True when `model` has known pricing (exact or family). Lets cost-aware
+ *  callers distinguish a real $0 from an unpriced model. */
+export declare function isModelPriced(model: string): boolean;
+/** Estimate token count from string length (chars / 4 approximation) */
+export declare function estimateTokens(text: string): number;
+/** Calculate cost in USD from token counts and model. Unknown models warn
+ *  once (not a silent zero) and return 0 so callers that ignore pricing keep
+ *  working; cost-sensitive callers should gate on {@link isModelPriced}. */
+export declare function estimateCost(inputTokens: number, outputTokens: number, model: string): number;
+//#endregion
+//#region src/anti-slop.d.ts
+interface AntiSlopConfig {
+  /** Domain label — appears in the JudgeScore output */
+  domain?: string;
+  /** Case-insensitive substrings that must not appear. Each occurrence = penalty. */
+  bannedPhrases?: string[];
+  /** Regexes matching opening formulas to penalize (e.g. /^great question/i). */
+  bannedOpenings?: RegExp[];
+  /** Regexes matching hedges (e.g. /i could be wrong/i). Ratio of hedged sentences drives score. */
+  hedgingPatterns?: RegExp[];
+  /** Regexes matching apology padding. */
+  apologyPatterns?: RegExp[];
+  /** Fraction of sentences that can be duplicates before penalty (default 0.15 = 15%). */
+  repetitionThreshold?: number;
+  /** Min output length in chars; below this the turn is deemed too terse. */
+  minLength?: number;
+  /** Max output length in chars; above this the turn is deemed too verbose. */
+  maxLength?: number;
+  /** How heavily each violation class reduces the score (default 1). */
+  penaltyWeights?: Partial<Record<SlopCategory, number>>;
 }
-interface SteeringOptimizationResult {
-  backend: SteeringOptimizerBackend;
-  recommendedVariantId: string;
-  rationale: string;
-  rankings: Array<{
-    variantId: string;
-    mean: number;
-    runs: number;
-  }>;
-  selector?: SteeringOptimizationSelector;
-  skipped?: boolean;
+type SlopCategory = 'banned_phrase' | 'banned_opening' | 'hedging' | 'apology' | 'repetition' | 'length';
+/** Create a reusable Judge function from an anti-slop config. */
+export declare function createAntiSlopJudge(config?: AntiSlopConfig): JudgeFn;
+interface AntiSlopIssue {
+  category: SlopCategory;
+  detail: string;
+  example?: string;
 }
-interface SteeringOptimizerConfig {
-  weights?: Partial<RunScoreWeights>;
+interface AntiSlopReport {
+  /** 0–10 score; 10 is clean, lower values mean more slop. */
+  score: number;
+  issues: AntiSlopIssue[];
+  /** Count of each category for programmatic aggregation. */
+  counts: Record<SlopCategory, number>;
 }
 /**
- * Rank already-evaluated steering variants.
- *
- * Model-backed candidate generation belongs to the official GEPA campaign
- * integration. This class only performs the deterministic selection it owns.
+ * Pure function — analyze one or more outputs against the config. Exposed
+ * separately so consumers can build their own reporters on top.
  */
-export declare class PairwiseSteeringOptimizer {
-  optimize(rows: SteeringOptimizationRow[], config?: SteeringOptimizerConfig): SteeringOptimizationResult;
-}
+export declare function analyzeAntiSlop(outputs: string[], config: Omit<Required<AntiSlopConfig>, 'domain'> & {
+  penaltyWeights: Record<SlopCategory, number>;
+}): AntiSlopReport;
 //#endregion
-export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AgentEvalAgent, AgentEvalError, type AgentEvalErrorCode, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type ArgumentCheck, type Artifact, type ArtifactCheck, type ArtifactEventLike, type ArtifactResult, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, type BoundedProcessInput, type BoundedProcessResult, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationExample, type CalibrationMerge, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureErrorHandler, type CaptureWrite, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTraceEngineOptions, type ChatTransport, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CollectedArtifacts, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractLintFinding, type ContractRule, type ContractRuleExecution, type ContractRun, type ContractSpan, type ContractStatus, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostProvenance, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, CostReservationExceededError, type CostResult, type CostSummary, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, type CustomTransportOpts, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_FAILURE_REASON_RULES, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type JudgeScore as DimensionJudgeScore, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureBlame, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureReasonRule, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type GateDecision$1 as HeldOutGateDecision, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, INFRA_FAILURE_BLAMES, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeAgreement, type JudgeCalibrationSet, type JudgeConfusion, JudgeError, type JudgeExampleAgreement, type JudgeFamily, type JudgeFn, type JudgeGateDecision, type JudgeGateInput, type JudgeGatePolicy, JudgeGateRefusedError, type JudgeGateStatus, type JudgeInput, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeRun, type JudgeRunMeta, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, type LinearRegressionResult, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MockTransportOpts, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type OpenAiCompatibleTransportOpts, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrderMode, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, type OwnerVerdict, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionAlternative, type PairedPromotionAlternativeCell, type PairedPromotionCallRates, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedPromotionPowerCall, type PairedPromotionPowerOptions, type PairedPromotionPowerResult, type PairedSignTestResult, type PairedTTestResult, type PairedValueLaw, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, type ProductClientConfig, type Scenario as ProductScenario, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactOptions, type RedactionProfile, type RedactionReport, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequiredPairsForPairedPromotionOptions, type RequiredPairsForPairedPromotionResult, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type RetryWrite, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RolloutSearchLineage, type RouteMap, type RoutedField, type RubricDimension, type Run, type RunCampaignOptions, type RunCaptureReport, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSearchCoordinates, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RunTraceRef, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SampleSummary, type SampleSummaryOptions, type SandboxDriver, type SandboxSdkTransportOpts, type SatisfiedBy, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, type ScenarioResult, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type ShareSafetyStatus, type ShareSafetyVerdict, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SpanPredicate, type SpanRecordsToTraceAnalysisStoreOptions, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, type TraceContractSpec, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type TrainSelectionTestOptions, type TrainSelectionTestPartition, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnResult, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, type WeightedCompositeInput, type WeightedCompositeResult, type WelchTestResult, type WelchTestStatus, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, addCalibrationVerdicts, agentProfileCellKey, agentProfileHash, agentProfileId, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCrossFamily, assertCrossFamilyServed, assertJudgeMayGate, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, captureFetchToRawSink, certificationEvidenceDigest, checkServedModel, classifyFailure, classifyFailureReason, comparePairedArms, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, describeTraceInsightScope, diffFindings, domainEvidencePattern, dominates, eProcess, equivalenceVerdict, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, expandProfileAxes, exportRunAsOtlp, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, gainHistogram, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashJson, holm, inferDomainKeywords, interRaterReliability, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isRunRecord, isToolSpan, isTransientLlmError, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, llmJudge, makeFinding, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, measureJudgeAgreement, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, objectiveEval, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseRunRecordSafe, partialCredit, passAtK, pearsonR, index_d_exports as profile, quotaExhaustedUntil, ranks, redTeamDataset, redTeamReport, redact, redactForShare, redactText, registerCalibrationSet, renderPreferenceMemoryMarkdown, requiredPairedSampleSize, requiredSampleSize, runAgentControlLoop, runCampaign, runCanaries, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, selfImprove, servedModelAcceptable, shareAllowed, snapshotFromJudgeAgreement, spanRecordsToTraceAnalysisStore, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, transientDispatchFailure, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, weightedComposite, weightedMean, welchsTTest, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit };
+export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AgentEvalAgent, AgentEvalError, type AgentEvalErrorCode, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type ArgumentCheck, type Artifact, type ArtifactCheck, type ArtifactEventLike, type ArtifactResult, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, type BoundedProcessInput, type BoundedProcessResult, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationExample, type CalibrationMerge, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureErrorHandler, type CaptureWrite, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTraceEngineOptions, type ChatTransport, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CollectedArtifacts, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractLintFinding, type ContractRule, type ContractRuleExecution, type ContractRun, type ContractSpan, type ContractStatus, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, type CostLedgerSummary, type CostProvenance, type CostReceipt, type CostReceiptInput, type CostResult, type CostSummary, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, type CustomTransportOpts, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_FAILURE_REASON_RULES, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type JudgeScore as DimensionJudgeScore, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureBlame, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureReasonRule, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type GateDecision$1 as HeldOutGateDecision, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, INFRA_FAILURE_BLAMES, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeAgreement, type JudgeCalibrationSet, type JudgeConfusion, JudgeError, type JudgeExampleAgreement, type JudgeFamily, type JudgeFn, type JudgeGateDecision, type JudgeGateInput, type JudgeGatePolicy, JudgeGateRefusedError, type JudgeGateStatus, type JudgeInput, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeRun, type JudgeRunMeta, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, type LinearRegressionResult, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MockTransportOpts, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type OpenAiCompatibleTransportOpts, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrderMode, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, type OwnerVerdict, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionAlternative, type PairedPromotionAlternativeCell, type PairedPromotionCallRates, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedPromotionPowerCall, type PairedPromotionPowerOptions, type PairedPromotionPowerResult, type PairedSignTestResult, type PairedTTestResult, type PairedValueLaw, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, type ProductClientConfig, type Scenario as ProductScenario, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactOptions, type RedactionProfile, type RedactionReport, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequiredPairsForPairedPromotionOptions, type RequiredPairsForPairedPromotionResult, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type RetryWrite, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RolloutSearchLineage, type RouteMap, type RoutedField, type RubricDimension, type Run, type RunCampaignOptions, type RunCaptureReport, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSearchCoordinates, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RunTraceRef, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SampleSummary, type SampleSummaryOptions, type SandboxDriver, type SandboxSdkTransportOpts, type SatisfiedBy, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, type ScenarioResult, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type ShareSafetyStatus, type ShareSafetyVerdict, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SpanPredicate, type SpanRecordsToTraceAnalysisStoreOptions, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, type TraceContractSpec, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type TrainSelectionTestOptions, type TrainSelectionTestPartition, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnResult, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, type WeightedCompositeInput, type WeightedCompositeResult, type WelchTestResult, type WelchTestStatus, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, addCalibrationVerdicts, agentProfileCellKey, agentProfileHash, agentProfileId, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCrossFamily, assertCrossFamilyServed, assertJudgeMayGate, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, captureFetchToRawSink, certificationEvidenceDigest, checkServedModel, classifyFailure, classifyFailureReason, comparePairedArms, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, describeTraceInsightScope, diffFindings, domainEvidencePattern, dominates, eProcess, equivalenceVerdict, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, expandProfileAxes, exportRunAsOtlp, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, gainHistogram, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashJson, holm, inferDomainKeywords, interRaterReliability, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isRunRecord, isToolSpan, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, llmJudge, makeFinding, makeProposalFinding, manifestContentDigest, mannWhitneyU, mcnemar, mcnemarPower, mcnemarRequiredN, measureJudgeAgreement, mintRolloutRows, modelHasSnapshot, mulberry32, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseRunRecordSafe, partialCredit, passAtK, pearsonR, index_d_exports as profile, quotaExhaustedUntil, ranks, redTeamDataset, redTeamReport, redact, redactForShare, redactText, registerCalibrationSet, renderPreferenceMemoryMarkdown, requiredPairedSampleSize, requiredSampleSize, runCampaign, runCanaries, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, selfImprove, servedModelAcceptable, shareAllowed, snapshotFromJudgeAgreement, spanRecordsToTraceAnalysisStore, spearmanR, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, transientDispatchFailure, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, weightedComposite, weightedMean, welchsTTest, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit };
 //# sourceMappingURL=index.d.ts.map