@tangle-network/agent-eval 0.115.1 → 0.115.3

package/CHANGELOG.md

@@ -4,6 +4,26 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ---
 
+## [0.115.3] — 2026-07-12 — fail-closed structured output parsing
+
+### Fixed
+
+- `callLlmJson()` now rejects responses terminated with finish reason `length`, even when the returned prefix happens to parse as JSON.
+- JSON extraction no longer descends into a valid nested object or array when a response declares an incomplete top-level JSON root.
+
+This patch prevents truncated structured responses from being silently accepted under the wrong response shape.
+Consumers of `callLlmJson()` should update.
+
+## [0.115.2] — 2026-07-12 — truthful code-agent session accounting
+
+### Fixed
+
+- fix(contract): ingest direct Codex 0.144.x exec JSONL lifecycle, tool, patch, terminal, and token events without double-counting transitions or reasoning tokens.
+- fix(contract): preserve observed, estimated, and uncaptured USD provenance through code-agent session intake and analyzeRuns.
+
+This patch corrects imported trace and cost semantics while retaining backward-compatible serialized RunRecords.
+Consumers importing code-agent execution traces should update.
+
 ## [0.115.1] — 2026-07-11 — fair cross-surface baseline selection
 
 ### Fixed

package/dist/analyst/index.d.ts

@@ -1,20 +1,20 @@
 import { M as MultiLayerVerifier, V as VerifyOptions, S as Severity } from '../multi-layer-verifier-BsqKuLyN.js';
-import { R as RunCritic, S as SemanticConceptJudgeOptions, a as RunTrace, b as SemanticConceptJudgeInput, B as BehavioralMetrics } from '../semantic-concept-judge-CpzbtwD0.js';
-export { C as CreateAnalystAiConfig, D as DEFAULT_TRACE_ANALYST_KINDS, c as DiffPolicy, F as FAILURE_MODE_KIND_SPEC, d as FINDING_SUBJECT_GRAMMAR_PROMPT, e as FINDING_SUBJECT_KINDS, f as FINDING_SUBJECT_SYNTAX, g as FindingSubject, h as FindingSubjectKind, i as FindingSubjectStringSchema, j as FindingsDiff, k as FindingsStore, I as IMPROVEMENT_KIND_SPEC, K as KIND_EXPECTED_SUBJECTS, l as KNOWLEDGE_GAP_KIND_SPEC, m as KNOWLEDGE_POISONING_KIND_SPEC, P as PersistedFinding, n as SKILL_USAGE_ANALYST, o as SkillUsageAnalyst, p as SkillUsageRecord, q as SkillUsageReport, r as SkillUsageScanConfig, s as buildSkillUsageReport, t as createAnalystAi, u as defaultIsMaterial, v as diffFindings, w as emitSkillUsageFindings, x as findingSubjectGrammarPromptFor, y as parseFindingSubject, z as renderFindingSubject } from '../semantic-concept-judge-CpzbtwD0.js';
+import { R as RunCritic, S as SemanticConceptJudgeOptions, a as RunTrace, b as SemanticConceptJudgeInput, B as BehavioralMetrics } from '../semantic-concept-judge-DxJmRkyJ.js';
+export { C as CreateAnalystAiConfig, D as DEFAULT_TRACE_ANALYST_KINDS, c as DiffPolicy, F as FAILURE_MODE_KIND_SPEC, d as FINDING_SUBJECT_GRAMMAR_PROMPT, e as FINDING_SUBJECT_KINDS, f as FINDING_SUBJECT_SYNTAX, g as FindingSubject, h as FindingSubjectKind, i as FindingSubjectStringSchema, j as FindingsDiff, k as FindingsStore, I as IMPROVEMENT_KIND_SPEC, K as KIND_EXPECTED_SUBJECTS, l as KNOWLEDGE_GAP_KIND_SPEC, m as KNOWLEDGE_POISONING_KIND_SPEC, P as PersistedFinding, n as SKILL_USAGE_ANALYST, o as SkillUsageAnalyst, p as SkillUsageRecord, q as SkillUsageReport, r as SkillUsageScanConfig, s as buildSkillUsageReport, t as createAnalystAi, u as defaultIsMaterial, v as diffFindings, w as emitSkillUsageFindings, x as findingSubjectGrammarPromptFor, y as parseFindingSubject, z as renderFindingSubject } from '../semantic-concept-judge-DxJmRkyJ.js';
 import { b as JudgeFn, a as JudgeInput } from '../types-C7DGg5ex.js';
-import { a as Analyst, g as AnalystSeverity, A as AnalystFinding } from '../kind-factory-20hcaYpf.js';
-export { h as ANALYST_SEVERITIES, b as AnalystContext, i as AnalystCost, j as AnalystInputKind, k as AnalystRequirements, f as AnalystRunEvent, e as AnalystRunInputs, d as AnalystRunResult, c as AnalystRunSummary, l as ChatCallOpts, C as ChatClient, m as ChatRequest, n as ChatResponse, o as ChatTransport, p as CliBridgeTransportOpts, q as CreateChatClientOpts, r as CreateTraceAnalystKindOpts, D as DirectProviderTransportOpts, E as EvidenceRef, M as MockTransportOpts, R as RAW_FINDING_SCHEMA_PROMPT, s as RawAnalystFinding, t as RawAnalystFindingSchema, u as RouterTransportOpts, S as SandboxSdkTransportOpts, v as TraceAnalystGolden, T as TraceAnalystKindSpec, w as computeFindingId, x as createChatClient, y as createTraceAnalystKind, z as makeFinding, B as parseRawFinding, F as renderPriorFindings } from '../kind-factory-20hcaYpf.js';
+import { a as Analyst, g as AnalystSeverity, A as AnalystFinding } from '../kind-factory-DcNg13sZ.js';
+export { h as ANALYST_SEVERITIES, b as AnalystContext, i as AnalystCost, j as AnalystInputKind, k as AnalystRequirements, f as AnalystRunEvent, e as AnalystRunInputs, d as AnalystRunResult, c as AnalystRunSummary, l as ChatCallOpts, C as ChatClient, m as ChatRequest, n as ChatResponse, o as ChatTransport, p as CliBridgeTransportOpts, q as CreateChatClientOpts, r as CreateTraceAnalystKindOpts, D as DirectProviderTransportOpts, E as EvidenceRef, M as MockTransportOpts, R as RAW_FINDING_SCHEMA_PROMPT, s as RawAnalystFinding, t as RawAnalystFindingSchema, u as RouterTransportOpts, S as SandboxSdkTransportOpts, v as TraceAnalystGolden, T as TraceAnalystKindSpec, w as computeFindingId, x as createChatClient, y as createTraceAnalystKind, z as makeFinding, B as parseRawFinding, F as renderPriorFindings } from '../kind-factory-DcNg13sZ.js';
 import { TCloud } from '@tangle-network/tcloud';
 import { T as TraceAnalysisStore } from '../store-C1YxJDEK.js';
-export { a as AnalystHooks, A as AnalystRegistry, b as AnalystRegistryOptions, B as BudgetPolicy, D as DefaultAnalystRegistryOptions, R as RegistryRunOpts, c as buildDefaultAnalystRegistry } from '../default-registry-Bcf1uKVI.js';
-export { F as FindingToPolicyEditOptions, P as POLICY_EDIT_AXES, a as POLICY_EDIT_TARGET_SURFACES, b as PolicyEdit, c as PolicyEditAdmission, d as PolicyEditAdmissionOptions, e as PolicyEditAxis, f as PolicyEditChange, g as PolicyEditExpectedGain, h as PolicyEditGainDirection, i as PolicyEditGainUnit, j as PolicyEditInit, k as PolicyEditRisk, l as PolicyEditSchemaVersion, m as PolicyEditSource, n as PolicyEditTarget, o as PolicyEditTargetSurface, p as PolicyEditValidationError, q as admitPolicyEdit, r as applyPolicyEditToSurface, s as computePolicyEditId, t as isPolicyEdit, u as makePolicyEdit, v as policyEditFromFinding, w as policyEditsFromFindings, x as scorePolicyEditReadiness, y as validatePolicyEdit } from '../policy-edit-az2qRmvN.js';
+export { a as AnalystHooks, A as AnalystRegistry, b as AnalystRegistryOptions, B as BudgetPolicy, D as DefaultAnalystRegistryOptions, R as RegistryRunOpts, c as buildDefaultAnalystRegistry } from '../default-registry-DltpYR5u.js';
+export { F as FindingToPolicyEditOptions, P as POLICY_EDIT_AXES, a as POLICY_EDIT_TARGET_SURFACES, b as PolicyEdit, c as PolicyEditAdmission, d as PolicyEditAdmissionOptions, e as PolicyEditAxis, f as PolicyEditChange, g as PolicyEditExpectedGain, h as PolicyEditGainDirection, i as PolicyEditGainUnit, j as PolicyEditInit, k as PolicyEditRisk, l as PolicyEditSchemaVersion, m as PolicyEditSource, n as PolicyEditTarget, o as PolicyEditTargetSurface, p as PolicyEditValidationError, q as admitPolicyEdit, r as applyPolicyEditToSurface, s as computePolicyEditId, t as isPolicyEdit, u as makePolicyEdit, v as policyEditFromFinding, w as policyEditsFromFindings, x as scorePolicyEditReadiness, y as validatePolicyEdit } from '../policy-edit-RLn8GWof.js';
 import { L as LlmClientOptions } from '../llm-client-DyqEH4jH.js';
 import { AxFunction } from '@ax-llm/ax';
 import '../verdict-C9MlYujm.js';
 import 'zod';
 import '../schema-SGWcK9wa.js';
 import '../store-BsVi7ncX.js';
-import '../run-record-DksGsfgv.js';
+import '../run-record-B7RTi_ix.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';
 import '../raw-provider-sink-C46HDghv.js';

package/dist/analyst/index.js

@@ -11,12 +11,12 @@ import {
   diffFindings,
   emitSkillUsageFindings,
   runSemanticConceptJudge
-} from "../chunk-I2HNIE6N.js";
+} from "../chunk-WSBUZMBU.js";
 import {
   behavioralAnalyst,
   buildDefaultAnalystRegistry,
   deriveEfficiencyFindings
-} from "../chunk-LVTGFSHF.js";
+} from "../chunk-I6LVHOV3.js";
 import {
   POLICY_EDIT_AXES,
   POLICY_EDIT_TARGET_SURFACES,
@@ -33,7 +33,7 @@ import {
   policyEditsFromFindings,
   scorePolicyEditReadiness,
   validatePolicyEdit
-} from "../chunk-AN5UYSVD.js";
+} from "../chunk-QMXXSNC4.js";
 import {
   ANALYST_SEVERITIES,
   AnalystRegistry,
@@ -62,11 +62,11 @@ import {
   renderPriorFindings,
   stripCodeFences,
   structureFindings
-} from "../chunk-DWLIGZBX.js";
+} from "../chunk-5S5NJ63F.js";
 import "../chunk-LNQEP766.js";
 import "../chunk-XJYR7XFV.js";
 import "../chunk-VSMTAMNK.js";
-import "../chunk-FUCQVFMU.js";
+import "../chunk-GY4SYVPJ.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";

package/dist/{analyze-runs-Dmz6LA9e.d.ts → analyze-runs-BYHg6Irm.d.ts}

@@ -1,7 +1,7 @@
-import { A as AnalystRegistry } from './default-registry-Bcf1uKVI.js';
+import { A as AnalystRegistry } from './default-registry-DltpYR5u.js';
 import { D as DatasetScenario } from './dataset-NENEzRgk.js';
-import { R as RunRecord } from './run-record-DksGsfgv.js';
-import { I as InsightReport } from './insight-report-D4cXFsLt.js';
+import { R as RunRecord } from './run-record-B7RTi_ix.js';
+import { I as InsightReport } from './insight-report-IwwvqZZv.js';
 
 /**
  * # `analyzeRuns()` — turn a set of agent runs into an actionable decision packet.

package/dist/belief-state/index.d.ts

@@ -1,9 +1,9 @@
 import { c as CalibrationReport } from '../calibration-Dz8TQV4y.js';
 import { O as OffPolicyEstimate, a as OffPolicyOptions, b as OffPolicyTrajectory } from '../off-policy-DiwuKKg7.js';
-import { d as CodeAgentSessionSource, a as CodeAgentSessionIntakeOptions, c as CodeAgentSessionMetrics, C as CodeAgentSessionDiagnostic } from '../code-agent-session-yitf9I-F.js';
-import { R as RunRecord, b as RunSplitTag } from '../run-record-DksGsfgv.js';
+import { d as CodeAgentSessionSource, a as CodeAgentSessionIntakeOptions, c as CodeAgentSessionMetrics, C as CodeAgentSessionDiagnostic } from '../code-agent-session-D-g04tcy.js';
+import { R as RunRecord, b as RunSplitTag } from '../run-record-B7RTi_ix.js';
 import { T as TraceStore } from '../store-BsVi7ncX.js';
-import { R as RuntimeTrajectoryRecord, P as ProjectRuntimeTrajectoryEvidenceOptions, a as RuntimeTrajectoryEvidenceProjection } from '../runtime-trajectory-h5i0SZUj.js';
+import { R as RuntimeTrajectoryRecord, P as ProjectRuntimeTrajectoryEvidenceOptions, a as RuntimeTrajectoryEvidenceProjection } from '../runtime-trajectory-Dws7Kpgi.js';
 import '../schema-SGWcK9wa.js';
 import '../outcome-store-rnXLEqSn.js';
 import '@tangle-network/agent-interface';

package/dist/belief-state/index.js

@@ -4,7 +4,7 @@ import {
   fromKimiCodeSession,
   fromOpenCodeSession,
   fromPiSession
-} from "../chunk-S42AWHMP.js";
+} from "../chunk-E4BUPP7Z.js";
 import {
   calibrationFromPairs
 } from "../chunk-NPCTHQIO.js";

package/dist/benchmarks/index.d.ts

@@ -1,6 +1,6 @@
-export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, k as BenchmarkDistribution, c as BenchmarkEvaluation, d as BenchmarkFamily, l as BenchmarkMetricCalibrationOptions, m as BenchmarkMetricCalibrationResult, n as BenchmarkReport, e as BenchmarkResponder, o as BenchmarkRunOptions, p as BenchmarkRunResult, f as BenchmarkScenario, q as BenchmarkSliceSummary, g as BenchmarkSource, h as BenchmarkTaskKind, r as BuildStandardRetrievalItemsOptions, R as RetrievalIdAdapterOptions, S as StandardRetrievalArtifact, s as StandardRetrievalDocument, t as StandardRetrievalEvaluationOptions, u as StandardRetrievalPayload, v as StandardRetrievalQrel, w as StandardRetrievalQuery, x as StandardRetrievalResult, y as buildStandardRetrievalItems, z as calibrateBenchmarkMetric, A as createRetrievalIdBenchmarkAdapter, i as deterministicSplit, C as evaluateStandardRetrieval, D as normalizeRetrievedDocumentIds, E as parseBeirCorpusJsonl, F as parseBeirQueriesJsonl, G as parseJsonlRows, H as parseQrels, I as parseTsvRows, J as renderBenchmarkReportMarkdown, K as retrievalMetricsAtCutoff, L as routing, M as runBenchmarkAdapter, N as summarizeBenchmarkCampaign } from '../index-CWr5SIG-.js';
-import '../types-CgSlO6wT.js';
-import '../run-record-DksGsfgv.js';
+export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, k as BenchmarkDistribution, c as BenchmarkEvaluation, d as BenchmarkFamily, l as BenchmarkMetricCalibrationOptions, m as BenchmarkMetricCalibrationResult, n as BenchmarkReport, e as BenchmarkResponder, o as BenchmarkRunOptions, p as BenchmarkRunResult, f as BenchmarkScenario, q as BenchmarkSliceSummary, g as BenchmarkSource, h as BenchmarkTaskKind, r as BuildStandardRetrievalItemsOptions, R as RetrievalIdAdapterOptions, S as StandardRetrievalArtifact, s as StandardRetrievalDocument, t as StandardRetrievalEvaluationOptions, u as StandardRetrievalPayload, v as StandardRetrievalQrel, w as StandardRetrievalQuery, x as StandardRetrievalResult, y as buildStandardRetrievalItems, z as calibrateBenchmarkMetric, A as createRetrievalIdBenchmarkAdapter, i as deterministicSplit, C as evaluateStandardRetrieval, D as normalizeRetrievedDocumentIds, E as parseBeirCorpusJsonl, F as parseBeirQueriesJsonl, G as parseJsonlRows, H as parseQrels, I as parseTsvRows, J as renderBenchmarkReportMarkdown, K as retrievalMetricsAtCutoff, L as routing, M as runBenchmarkAdapter, N as summarizeBenchmarkCampaign } from '../index-BTEpx9He.js';
+import '../types-C5gJrOVT.js';
+import '../run-record-B7RTi_ix.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';
 import '../schema-SGWcK9wa.js';

package/dist/benchmarks/index.js

@@ -17,21 +17,21 @@ import {
   runBenchmarkAdapter,
   summarizeBenchmarkCampaign
 } from "../chunk-3LXTCTWL.js";
-import "../chunk-KDCMEZDI.js";
-import "../chunk-N6MTC3GK.js";
+import "../chunk-KG4TD7EQ.js";
+import "../chunk-ADYLPOSX.js";
 import "../chunk-VI2UW6B6.js";
 import "../chunk-FAOEFFRT.js";
-import "../chunk-AN5UYSVD.js";
-import "../chunk-DWLIGZBX.js";
+import "../chunk-QMXXSNC4.js";
+import "../chunk-5S5NJ63F.js";
 import "../chunk-ARU2PZFM.js";
 import "../chunk-PJQFMIOX.js";
 import "../chunk-RPDDVKI7.js";
 import "../chunk-GGE4NNQT.js";
 import "../chunk-LNQEP766.js";
-import "../chunk-VK6HBGAE.js";
+import "../chunk-5UF54T55.js";
 import "../chunk-XJYR7XFV.js";
 import "../chunk-VSMTAMNK.js";
-import "../chunk-FUCQVFMU.js";
+import "../chunk-GY4SYVPJ.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";

package/dist/campaign/index.d.ts

@@ -1,20 +1,20 @@
-import { P as PairedArmsComparison, S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, b as ProducedState, c as CorrectnessChecker } from '../pre-registration-oNItiRBb.js';
-export { L as LlmJudgeDimension, d as LlmJudgeOptions, l as llmJudge } from '../pre-registration-oNItiRBb.js';
+import { P as PairedArmsComparison, S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, b as ProducedState, c as CorrectnessChecker } from '../pre-registration-D8h7ZxNL.js';
+export { L as LlmJudgeDimension, d as LlmJudgeOptions, l as llmJudge } from '../pre-registration-D8h7ZxNL.js';
 import { A as AnalyzeTracesOptions, a as AnalyzeTracesInput, b as AnalyzeTracesResult } from '../analyst-C8HHvfJp.js';
-import { S as Scenario, M as MutableSurface, D as DispatchContext, b as JudgeConfig, g as Gate, e as GenerationRecord, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, m as CodeSurface, y as LabeledScenarioSource, C as CampaignResult } from '../types-CgSlO6wT.js';
-export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, c as DispatchFn, n as GateContext, h as GateDecision, G as GateResult, o as GenerationCandidate, A as JudgeAggregate, a as JudgeDimension, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-CgSlO6wT.js';
-import { C as CampaignRunPlan, P as PlanCampaignRunOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-DolL_Fko.js';
-export { f as CampaignRunPlanCell, h as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, i as OpenAutoPrResult, a as RunImprovementLoopResult, R as RunOptimizationOptions, j as RunOptimizationResult, k as countSentenceEdits, l as defaultRenderDiff, m as extractH2Sections, g as gepaProposer, o as openAutoPr, p as planCampaignRun, r as runCampaign, d as runImprovementLoop, n as runOptimization } from '../gepa-DolL_Fko.js';
+import { S as Scenario, M as MutableSurface, D as DispatchContext, b as JudgeConfig, g as Gate, e as GenerationRecord, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, m as CodeSurface, y as LabeledScenarioSource, C as CampaignResult } from '../types-C5gJrOVT.js';
+export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, c as DispatchFn, n as GateContext, h as GateDecision, G as GateResult, o as GenerationCandidate, A as JudgeAggregate, a as JudgeDimension, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-C5gJrOVT.js';
+import { C as CampaignRunPlan, P as PlanCampaignRunOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-dne9JDPL.js';
+export { f as CampaignRunPlanCell, h as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, i as OpenAutoPrResult, a as RunImprovementLoopResult, R as RunOptimizationOptions, j as RunOptimizationResult, k as countSentenceEdits, l as defaultRenderDiff, m as extractH2Sections, g as gepaProposer, o as openAutoPr, p as planCampaignRun, r as runCampaign, d as runImprovementLoop, n as runOptimization } from '../gepa-dne9JDPL.js';
 import { a as PairedBootstrapResult, E as EProcessState } from '../statistics-oUbOJe-S.js';
 import { C as CampaignStorage } from '../storage-Dw_f7WMt.js';
 export { f as fsCampaignStorage, i as inMemoryCampaignStorage } from '../storage-Dw_f7WMt.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, l as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, m as EmitLoopProvenanceArgs, n as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, o as LoopProvenanceBackend, q as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, c as ParetoSignificanceGateOptions, P as PowerPreflight, s as PowerPreflightOptions, d as PromotionObjective, e as PromotionPolicy, R as RunEvalOptions, f as buildEvidenceVector, t as buildLoopProvenanceRecord, g as composeGate, h as defaultProductionGate, u as emitLoopProvenance, i as evolutionaryProposer, j as heldOutGate, v as loopProvenanceSpans, p as paretoPolicy, k as paretoSignificanceGate, w as powerPreflight, x as provenanceRecordPath, y as provenanceSpansPath, r as runEval } from '../provenance-BZmpWmn4.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, l as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, m as EmitLoopProvenanceArgs, n as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, o as LoopProvenanceBackend, q as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, c as ParetoSignificanceGateOptions, P as PowerPreflight, s as PowerPreflightOptions, d as PromotionObjective, e as PromotionPolicy, R as RunEvalOptions, f as buildEvidenceVector, t as buildLoopProvenanceRecord, g as composeGate, h as defaultProductionGate, u as emitLoopProvenance, i as evolutionaryProposer, j as heldOutGate, v as loopProvenanceSpans, p as paretoPolicy, k as paretoSignificanceGate, w as powerPreflight, x as provenanceRecordPath, y as provenanceSpansPath, r as runEval } from '../provenance-Bibyg1U9.js';
 import { L as LlmClientOptions } from '../llm-client-DyqEH4jH.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { A as AgentEvalError, V as ValidationError } from '../errors-oeQrLqXC.js';
-import { b as RunSplitTag, R as RunRecord } from '../run-record-DksGsfgv.js';
-import { b as PolicyEdit, F as FindingToPolicyEditOptions, d as PolicyEditAdmissionOptions, c as PolicyEditAdmission } from '../policy-edit-az2qRmvN.js';
-import { T as TraceAnalystKindSpec, A as AnalystFinding } from '../kind-factory-20hcaYpf.js';
+import { b as RunSplitTag, R as RunRecord } from '../run-record-B7RTi_ix.js';
+import { b as PolicyEdit, F as FindingToPolicyEditOptions, d as PolicyEditAdmissionOptions, c as PolicyEditAdmission } from '../policy-edit-RLn8GWof.js';
+import { T as TraceAnalystKindSpec, A as AnalystFinding } from '../kind-factory-DcNg13sZ.js';
 import '@tangle-network/tcloud';
 import '../raw-provider-sink-C46HDghv.js';
 import '../verdict-C9MlYujm.js';
@@ -26,8 +26,8 @@ import '../schema-SGWcK9wa.js';
 import '../judge-calibration-7C-IDmKr.js';
 import '../types-C7DGg5ex.js';
 import '../hosted/index.js';
-import '../insight-report-D4cXFsLt.js';
-import '../summary-report-Bz-0-t8v.js';
+import '../insight-report-IwwvqZZv.js';
+import '../summary-report-BJ5aNwZ1.js';
 import '../failure-cluster-C48PiReX.js';
 import 'zod';
 

package/dist/campaign/index.js

@@ -65,7 +65,7 @@ import {
   userStoryScoreboard,
   validateSearchLedgerEvent,
   verifyCodeSurface
-} from "../chunk-KDCMEZDI.js";
+} from "../chunk-KG4TD7EQ.js";
 import {
   assertCodeSurfaceIdentity,
   buildEvidenceVector,
@@ -100,7 +100,7 @@ import {
   runOptimization,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-N6MTC3GK.js";
+} from "../chunk-ADYLPOSX.js";
 import "../chunk-VI2UW6B6.js";
 import {
   fsCampaignStorage,
@@ -110,17 +110,17 @@ import {
   runCampaign,
   tangleTracesRoot
 } from "../chunk-FAOEFFRT.js";
-import "../chunk-AN5UYSVD.js";
-import "../chunk-DWLIGZBX.js";
+import "../chunk-QMXXSNC4.js";
+import "../chunk-5S5NJ63F.js";
 import "../chunk-ARU2PZFM.js";
 import "../chunk-PJQFMIOX.js";
 import "../chunk-RPDDVKI7.js";
 import "../chunk-GGE4NNQT.js";
 import "../chunk-LNQEP766.js";
-import "../chunk-VK6HBGAE.js";
+import "../chunk-5UF54T55.js";
 import "../chunk-XJYR7XFV.js";
 import "../chunk-VSMTAMNK.js";
-import "../chunk-FUCQVFMU.js";
+import "../chunk-GY4SYVPJ.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";

package/dist/{chunk-DRPIZQIT.js → chunk-4D5RVB3W.js}

@@ -1,6 +1,6 @@
 import {
   callLlmJson
-} from "./chunk-FUCQVFMU.js";
+} from "./chunk-GY4SYVPJ.js";
 
 // src/wire/schemas.ts
 import { extendZodWithOpenApi } from "@asteasolutions/zod-to-openapi";
@@ -1002,4 +1002,4 @@ export {
   startServer,
   startServerAsync
 };
-//# sourceMappingURL=chunk-DRPIZQIT.js.map
+//# sourceMappingURL=chunk-4D5RVB3W.js.map

package/dist/{chunk-DWLIGZBX.js → chunk-5S5NJ63F.js}

@@ -4,7 +4,7 @@ import {
 } from "./chunk-LNQEP766.js";
 import {
   callLlm
-} from "./chunk-FUCQVFMU.js";
+} from "./chunk-GY4SYVPJ.js";
 
 // src/analyst/types.ts
 import { createHash } from "crypto";
@@ -1148,4 +1148,4 @@ export {
   DEFAULT_TRACE_ANALYST_KINDS,
   AnalystRegistry
 };
-//# sourceMappingURL=chunk-DWLIGZBX.js.map
+//# sourceMappingURL=chunk-5S5NJ63F.js.map

package/dist/{chunk-VK6HBGAE.js → chunk-5UF54T55.js}

@@ -50,6 +50,9 @@ function validateRunRecord(input) {
   expectFiniteNumber(obj.wallMs, "wallMs");
   if (obj.queueMs !== void 0) expectFiniteNumber(obj.queueMs, "queueMs");
   expectFiniteNumber(obj.costUsd, "costUsd");
+  if (obj.costProvenance !== void 0) {
+    validateCostProvenance(obj.costProvenance, obj.costUsd);
+  }
   if (!modelHasSnapshot(obj.model)) {
     throw new RunRecordValidationError(
       `model "${obj.model}" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,
@@ -151,6 +154,54 @@ function validateRunRecord(input) {
   }
   return input;
 }
+function resolveRunCostProvenance(run) {
+  if (run.costProvenance) return run.costProvenance;
+  if (run.outcome.raw.cost_estimated === 1) {
+    return { kind: "estimated", usd: run.costUsd };
+  }
+  if (run.costUsd > 0) return { kind: "observed", usd: run.costUsd };
+  return { kind: "uncaptured", usd: null };
+}
+function validateCostProvenance(input, costUsd) {
+  if (input === null || typeof input !== "object") {
+    throw new RunRecordValidationError("costProvenance must be an object", "costProvenance");
+  }
+  const value = input;
+  if (value.kind !== "observed" && value.kind !== "estimated" && value.kind !== "uncaptured") {
+    throw new RunRecordValidationError(
+      "costProvenance.kind must be observed, estimated, or uncaptured",
+      "costProvenance.kind"
+    );
+  }
+  if (value.kind === "uncaptured") {
+    if (value.usd !== null) {
+      throw new RunRecordValidationError(
+        "uncaptured costProvenance.usd must be null",
+        "costProvenance.usd"
+      );
+    }
+    if (costUsd !== 0) {
+      throw new RunRecordValidationError(
+        "uncaptured costProvenance requires the compatibility costUsd sentinel 0",
+        "costUsd"
+      );
+    }
+    return;
+  }
+  expectFiniteNumber(value.usd, "costProvenance.usd");
+  if (value.usd < 0) {
+    throw new RunRecordValidationError(
+      "costProvenance.usd must be non-negative",
+      "costProvenance.usd"
+    );
+  }
+  if (value.usd !== costUsd) {
+    throw new RunRecordValidationError(
+      "costProvenance.usd must equal costUsd",
+      "costProvenance.usd"
+    );
+  }
+}
 function isRunRecord(input) {
   try {
     validateRunRecord(input);
@@ -241,9 +292,10 @@ function modelHasSnapshot(model) {
 export {
   RunRecordValidationError,
   validateRunRecord,
+  resolveRunCostProvenance,
   isRunRecord,
   parseRunRecordSafe,
   roundTripRunRecord,
   modelHasSnapshot
 };
-//# sourceMappingURL=chunk-VK6HBGAE.js.map
+//# sourceMappingURL=chunk-5UF54T55.js.map

package/dist/chunk-5UF54T55.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag and either a `searchScore` or a `holdoutScore`.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n *   const rec = validateRunRecord(rawJson)         // throws on missing\n *   const ok  = isRunRecord(rawJson)               // boolean check\n *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport { ValidationError } from './errors'\nimport type { FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n *  combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\nexport interface RunTokenUsage {\n  input: number\n  output: number\n  cached?: number\n}\n\n/**\n * How a run's USD amount was obtained.\n *\n * `costUsd` remains mandatory for wire compatibility. New producers should\n * always populate this discriminated union so a missing bill is never\n * mistaken for an observed zero-dollar run. For `uncaptured`, `costUsd` uses\n * the legacy `0` sentinel while this field carries the truthful null.\n */\nexport type RunCostProvenance =\n  | { kind: 'observed'; usd: number }\n  | { kind: 'estimated'; usd: number }\n  | { kind: 'uncaptured'; usd: null }\n\nexport interface RunJudgeMetadata {\n  model: string\n  promptVersion: string\n  /** [0,1] confidence the judge declared. Constant judge confidence\n   *  across many runs is a fallback signal (see `canary.ts`). */\n  confidence: number\n  /** True if the judge degraded to a fallback path (rules-only,\n   *  prior-call cache, etc.). The canary uses this to alert. */\n  fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n  /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n  perJudge: Record<string, Record<string, number>>\n  /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n  perDimMean: Record<string, number>\n  /** Composite mean across all dims and judges. Mirrors the score\n   *  the gate sees on `outcome.searchScore` / `holdoutScore`. */\n  composite: number\n  /** Judges that errored or returned an unparseable verdict. Recorded\n   *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n   *  not inferred from missing keys in `perJudge`. */\n  failedJudges?: string[]\n  /** Free-form notes the judges emitted (joined across judges or\n   *  first-judge only — consumer's choice). */\n  notes?: string\n}\n\nexport interface RunOutcome {\n  /** Score on the search/optimization split. Optional because a\n   *  holdout-only evaluation only fills `holdoutScore`. */\n  searchScore?: number\n  /** Score on the held-out split. Optional because a search-only run\n   *  only fills `searchScore`. At least one must be present. */\n  holdoutScore?: number\n  /** Bag of any other metric the run produced — judge dimensions,\n   *  pass/fail counters, latency stats, etc. Numeric only — keeps\n   *  reporters honest. */\n  raw: Record<string, number>\n  /** Per-judge / per-dim breakdown. Consumers writing ensemble\n   *  judgements populate this; substrate primitives like\n   *  `interRaterReliability` and `corpusInterRaterAgreement` accept\n   *  these records as input. Optional — single-judge or scalar-only\n   *  runs leave it unset. */\n  judgeScores?: JudgeScoresRecord\n  /** Authenticity / realness verdict — did the run build the REAL thing on the\n   *  intended infra, or fake it (see `./authenticity`)? Optional: only domains\n   *  with an authenticity config populate it. Carried in the corpus so the\n   *  flywheel / off-policy learning can optimize for real completion, not gamed\n   *  pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n   *  must not count as a real success regardless of `score`. */\n  realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n *     model (after any steering bundle merge).\n *   - `configHash` is the sha256 of the effective run config (model,\n *     temperature, tools, judges, splits). The pair (promptHash,\n *     configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n *   - `model` MUST encode a snapshot version. Bare aliases like\n *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n */\nexport interface RunRecord {\n  /** UUID for the run. */\n  runId: string\n  /** Logical experiment grouping (a treatment vs a baseline within\n   *  the same sweep should share `experimentId`). */\n  experimentId: string\n  /** Stable identifier for the candidate (variant) being run. The\n   *  promotion gate compares two `candidateId`s on matched items. */\n  candidateId: string\n  /** RNG seed for the run. Always recorded — silent re-seeding is\n   *  the most common cause of non-reproducible numbers. */\n  seed: number\n  /** Model identifier WITH snapshot version. */\n  model: string\n  /** sha256 of the effective prompt (post-steering). */\n  promptHash: string\n  /** sha256 of the effective config. */\n  configHash: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** End-to-end wall-clock duration in milliseconds. */\n  wallMs: number\n  /** Time spent queued before execution started, if known. */\n  queueMs?: number\n  /** Total USD cost. Mandatory — runs without a cost number are\n   *  unbounded by definition and must not be admitted into the gate.\n   *  `0` is retained as the compatibility sentinel for an uncaptured amount;\n   *  inspect `costProvenance` before treating it as observed. */\n  costUsd: number\n  /** Observed, model-priced estimate, or genuinely uncaptured USD amount.\n   *  Optional only so existing serialized RunRecords remain valid. */\n  costProvenance?: RunCostProvenance\n  /** Token usage breakdown. */\n  tokenUsage: RunTokenUsage\n  /** Judge-side metadata, if a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /** Per-split scores + raw bag. */\n  outcome: RunOutcome\n  /** Canonical, cross-agent failure class drawn from the shared\n   *  `FAILURE_CLASSES` taxonomy. This is the aggregation key that makes\n   *  \"which failure dominates across the whole fleet\" answerable in ONE\n   *  vocabulary — every agent classifies against the same enum. Producers\n   *  set it via the substrate classifier; leave unset only when the failure\n   *  genuinely can't be classified. */\n  failureClass?: FailureClass\n  /** Free-form domain-specific failure detail, scoped UNDER `failureClass`\n   *  (e.g. failureClass='tool_recovery_failure', failureMode='forge_build_unsatisfied').\n   *  The within-agent drill-down; `failureClass` is the cross-agent key. */\n  failureMode?: string\n  /** Which split this run was drawn from. */\n  splitTag: RunSplitTag\n  /**\n   * Stable scenario identifier the run was scored against. Optional for\n   * backwards compatibility, but **strongly recommended**: every primitive\n   * that pairs runs by scenario (preferences, paired stats, BT tournament)\n   * keys on this. The campaign artifact populates it canonically; legacy\n   * runs without it fall back to inference from `outcome.raw.scenario_id`\n   * or `experimentId`.\n   */\n  scenarioId?: string\n  /**\n   * Canonical identity for the agent profile cell that produced this row:\n   * profile artifact hash plus optional harness/model/prompt/reporting\n   * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n   * longitudinal reports by the complete source profile, not by a loose\n   * candidate label or opaque config hash.\n   */\n  agentProfile?: AgentProfileCell\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n  'runId',\n  'experimentId',\n  'candidateId',\n  'seed',\n  'model',\n  'promptHash',\n  'configHash',\n  'commitSha',\n  'wallMs',\n  'costUsd',\n  'tokenUsage',\n  'outcome',\n  'splitTag',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\n\nexport class RunRecordValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n\n  for (const key of MANDATORY_TOP_LEVEL) {\n    if (!(key in obj)) {\n      throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n    }\n  }\n\n  expectString(obj.runId, 'runId')\n  expectString(obj.experimentId, 'experimentId')\n  expectString(obj.candidateId, 'candidateId')\n  expectFiniteNumber(obj.seed, 'seed')\n  expectString(obj.model, 'model')\n  expectString(obj.promptHash, 'promptHash')\n  expectString(obj.configHash, 'configHash')\n  expectString(obj.commitSha, 'commitSha')\n  expectFiniteNumber(obj.wallMs, 'wallMs')\n  if (obj.queueMs !== undefined) expectFiniteNumber(obj.queueMs, 'queueMs')\n  expectFiniteNumber(obj.costUsd, 'costUsd')\n  if (obj.costProvenance !== undefined) {\n    validateCostProvenance(obj.costProvenance, obj.costUsd as number)\n  }\n\n  // Snapshot discipline: bare model aliases are not paper-grade.\n  if (!modelHasSnapshot(obj.model as string)) {\n    throw new RunRecordValidationError(\n      `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,\n      'model',\n    )\n  }\n\n  // Token usage.\n  const tu = obj.tokenUsage\n  if (tu === null || typeof tu !== 'object') {\n    throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n  }\n  const tuRec = tu as Record<string, unknown>\n  expectFiniteNumber(tuRec.input, 'tokenUsage.input')\n  expectFiniteNumber(tuRec.output, 'tokenUsage.output')\n  if (tuRec.cached !== undefined) expectFiniteNumber(tuRec.cached, 'tokenUsage.cached')\n\n  // Judge metadata, optional.\n  if (obj.judgeMetadata !== undefined) {\n    const jm = obj.judgeMetadata\n    if (jm === null || typeof jm !== 'object') {\n      throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n    }\n    const jmRec = jm as Record<string, unknown>\n    expectString(jmRec.model, 'judgeMetadata.model')\n    expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n    expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n    if (typeof jmRec.fallback !== 'boolean') {\n      throw new RunRecordValidationError(\n        'judgeMetadata.fallback must be boolean',\n        'judgeMetadata.fallback',\n      )\n    }\n  }\n\n  // Outcome.\n  const out = obj.outcome\n  if (out === null || typeof out !== 'object') {\n    throw new RunRecordValidationError('outcome must be an object', 'outcome')\n  }\n  const outRec = out as Record<string, unknown>\n  if (outRec.searchScore !== undefined)\n    expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n  if (outRec.holdoutScore !== undefined)\n    expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n  if (outRec.searchScore === undefined && outRec.holdoutScore === undefined) {\n    throw new RunRecordValidationError(\n      'outcome must define searchScore or holdoutScore (or both)',\n      'outcome',\n    )\n  }\n  const raw = outRec.raw\n  if (raw === null || typeof raw !== 'object') {\n    throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n  }\n  for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n    expectFiniteNumber(v, `outcome.raw.${k}`)\n  }\n  // Realness verdict, optional.\n  if (outRec.realness !== undefined) {\n    const r = outRec.realness\n    if (r === null || typeof r !== 'object') {\n      throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n    }\n    const rr = r as Record<string, unknown>\n    expectFiniteNumber(rr.score, 'outcome.realness.score')\n    if (typeof rr.gated !== 'boolean') {\n      throw new RunRecordValidationError(\n        'outcome.realness.gated must be a boolean',\n        'outcome.realness.gated',\n      )\n    }\n  }\n\n  // Per-judge / per-dim breakdown, optional.\n  if (outRec.judgeScores !== undefined) {\n    validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n  }\n\n  // Failure mode optional.\n  if (obj.failureMode !== undefined) expectString(obj.failureMode, 'failureMode')\n\n  if (obj.agentProfile !== undefined) {\n    try {\n      const profile = validateAgentProfileCell(obj.agentProfile)\n      if (profile.model !== undefined && profile.model !== obj.model) {\n        throw new RunRecordValidationError(\n          `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n          'agentProfile.model',\n        )\n      }\n      if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n        throw new RunRecordValidationError(\n          `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n          'agentProfile.promptHash',\n        )\n      }\n    } catch (error) {\n      if (error instanceof RunRecordValidationError) throw error\n      if (error instanceof Error) {\n        throw new RunRecordValidationError(error.message, 'agentProfile')\n      }\n      throw error\n    }\n  }\n\n  // Split tag.\n  if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n    throw new RunRecordValidationError(\n      `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n      'splitTag',\n    )\n  }\n\n  return input as RunRecord\n}\n\n/**\n * Resolve provenance for both new and legacy records.\n *\n * Legacy producers sometimes set `outcome.raw.cost_estimated = 1`. A positive\n * unlabeled amount is treated as observed, matching the historical contract.\n * Zero without an explicit label is conservatively uncaptured: claiming an\n * observed $0 would be stronger than the serialized evidence supports.\n */\nexport function resolveRunCostProvenance(\n  run: Pick<RunRecord, 'costUsd' | 'costProvenance' | 'outcome'>,\n): RunCostProvenance {\n  if (run.costProvenance) return run.costProvenance\n  if (run.outcome.raw.cost_estimated === 1) {\n    return { kind: 'estimated', usd: run.costUsd }\n  }\n  if (run.costUsd > 0) return { kind: 'observed', usd: run.costUsd }\n  return { kind: 'uncaptured', usd: null }\n}\n\nfunction validateCostProvenance(input: unknown, costUsd: number): void {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n  }\n  const value = input as Record<string, unknown>\n  if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n    throw new RunRecordValidationError(\n      'costProvenance.kind must be observed, estimated, or uncaptured',\n      'costProvenance.kind',\n    )\n  }\n  if (value.kind === 'uncaptured') {\n    if (value.usd !== null) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance.usd must be null',\n        'costProvenance.usd',\n      )\n    }\n    if (costUsd !== 0) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance requires the compatibility costUsd sentinel 0',\n        'costUsd',\n      )\n    }\n    return\n  }\n  expectFiniteNumber(value.usd, 'costProvenance.usd')\n  if ((value.usd as number) < 0) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must be non-negative',\n      'costProvenance.usd',\n    )\n  }\n  if (value.usd !== costUsd) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must equal costUsd',\n      'costProvenance.usd',\n    )\n  }\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n  try {\n    validateRunRecord(input)\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n  input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n  try {\n    return { ok: true, value: validateRunRecord(input) }\n  } catch (e) {\n    if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n    throw e\n  }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n  const json = JSON.stringify(record)\n  return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new RunRecordValidationError(`expected non-empty string`, path)\n  }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value)) {\n    throw new RunRecordValidationError(`expected finite number`, path)\n  }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object') {\n    throw new RunRecordValidationError('judgeScores must be an object', path)\n  }\n  const rec = value as Record<string, unknown>\n\n  const perJudge = rec.perJudge\n  if (perJudge === null || typeof perJudge !== 'object') {\n    throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n  }\n  for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n    if (dims === null || typeof dims !== 'object') {\n      throw new RunRecordValidationError(\n        'per-judge entry must be an object of dimension scores',\n        `${path}.perJudge.${judgeId}`,\n      )\n    }\n    for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n    }\n  }\n\n  const perDimMean = rec.perDimMean\n  if (perDimMean === null || typeof perDimMean !== 'object') {\n    throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n  }\n  for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n  }\n\n  expectFiniteNumber(rec.composite, `${path}.composite`)\n\n  if (rec.failedJudges !== undefined) {\n    if (!Array.isArray(rec.failedJudges)) {\n      throw new RunRecordValidationError(\n        'failedJudges must be an array of strings',\n        `${path}.failedJudges`,\n      )\n    }\n    for (let i = 0; i < rec.failedJudges.length; i++) {\n      const id = rec.failedJudges[i]\n      if (typeof id !== 'string' || id.length === 0) {\n        throw new RunRecordValidationError(\n          'failedJudges entry must be a non-empty string',\n          `${path}.failedJudges[${i}]`,\n        )\n      }\n    }\n  }\n\n  if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n    throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n  }\n}\n\n/**\n * Heuristic snapshot check. Accepts:\n *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)\n *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)\n *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)\n *   - explicit `:date-...` Vertex-style tags\n *\n * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap\n * silently as providers ship new snapshots.\n */\nexport function modelHasSnapshot(model: string): boolean {\n  if (model.includes('@')) return true\n  if (/-\\d{8}$/.test(model)) return true\n  if (/-\\d{4}-\\d{2}-\\d{2}$/.test(model)) return true\n  if (/:date-/.test(model)) return true\n  return false\n}\n"],"mappings":";;;;;;;;AAsNA,IAAM,sBAAsB;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,aAAyC,CAAC,UAAU,OAAO,SAAS;AAEnE,IAAM,2BAAN,cAAuC,gBAAgB;AAAA,EACnD;AAAA,EACT,YAAY,SAAiB,OAAO,IAAI;AACtC,UAAM,OAAO,GAAG,OAAO,QAAQ,IAAI,MAAM,OAAO;AAChD,SAAK,OAAO;AAAA,EACd;AACF;AAOO,SAAS,kBAAkB,OAA2B;AAC3D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iBAAiB;AAAA,EACtD;AACA,QAAM,MAAM;AAEZ,aAAW,OAAO,qBAAqB;AACrC,QAAI,EAAE,OAAO,MAAM;AACjB,YAAM,IAAI,yBAAyB,4BAA4B,GAAG,GAAG;AAAA,IACvE;AAAA,EACF;AAEA,eAAa,IAAI,OAAO,OAAO;AAC/B,eAAa,IAAI,cAAc,cAAc;AAC7C,eAAa,IAAI,aAAa,aAAa;AAC3C,qBAAmB,IAAI,MAAM,MAAM;AACnC,eAAa,IAAI,OAAO,OAAO;AAC/B,eAAa,IAAI,YAAY,YAAY;AACzC,eAAa,IAAI,YAAY,YAAY;AACzC,eAAa,IAAI,WAAW,WAAW;AACvC,qBAAmB,IAAI,QAAQ,QAAQ;AACvC,MAAI,IAAI,YAAY,OAAW,oBAAmB,IAAI,SAAS,SAAS;AACxE,qBAAmB,IAAI,SAAS,SAAS;AACzC,MAAI,IAAI,mBAAmB,QAAW;AACpC,2BAAuB,IAAI,gBAAgB,IAAI,OAAiB;AAAA,EAClE;AAGA,MAAI,CAAC,iBAAiB,IAAI,KAAe,GAAG;AAC1C,UAAM,IAAI;AAAA,MACR,UAAU,IAAI,KAAK;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AAGA,QAAM,KAAK,IAAI;AACf,MAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,UAAM,IAAI,yBAAyB,gCAAgC,YAAY;AAAA,EACjF;AACA,QAAM,QAAQ;AACd,qBAAmB,MAAM,OAAO,kBAAkB;AAClD,qBAAmB,MAAM,QAAQ,mBAAmB;AACpD,MAAI,MAAM,WAAW,OAAW,oBAAmB,MAAM,QAAQ,mBAAmB;AAGpF,MAAI,IAAI,kBAAkB,QAAW;AACnC,UAAM,KAAK,IAAI;AACf,QAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,YAAM,IAAI,yBAAyB,mCAAmC,eAAe;AAAA,IACvF;AACA,UAAM,QAAQ;AACd,iBAAa,MAAM,OAAO,qBAAqB;AAC/C,iBAAa,MAAM,eAAe,6BAA6B;AAC/D,uBAAmB,MAAM,YAAY,0BAA0B;AAC/D,QAAI,OAAO,MAAM,aAAa,WAAW;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAGA,QAAM,MAAM,IAAI;AAChB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,6BAA6B,SAAS;AAAA,EAC3E;AACA,QAAM,SAAS;AACf,MAAI,OAAO,gBAAgB;AACzB,uBAAmB,OAAO,aAAa,qBAAqB;AAC9D,MAAI,OAAO,iBAAiB;AAC1B,uBAAmB,OAAO,cAAc,sBAAsB;AAChE,MAAI,OAAO,gBAAgB,UAAa,OAAO,iBAAiB,QAAW;AACzE,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,QAAM,MAAM,OAAO;AACnB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,iCAAiC,aAAa;AAAA,EACnF;AACA,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,GAA8B,GAAG;AACnE,uBAAmB,GAAG,eAAe,CAAC,EAAE;AAAA,EAC1C;AAEA,MAAI,OAAO,aAAa,QAAW;AACjC,UAAM,IAAI,OAAO;AACjB,QAAI,MAAM,QAAQ,OAAO,MAAM,UAAU;AACvC,YAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;AAAA,IAC7F;AACA,UAAM,KAAK;AACX,uBAAmB,GAAG,OAAO,wBAAwB;AACrD,QAAI,OAAO,GAAG,UAAU,WAAW;AACjC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAGA,MAAI,OAAO,gBAAgB,QAAW;AACpC,wBAAoB,OAAO,aAAa,qBAAqB;AAAA,EAC/D;AAGA,MAAI,IAAI,gBAAgB,OAAW,cAAa,IAAI,aAAa,aAAa;AAE9E,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI;AACF,YAAM,UAAU,yBAAyB,IAAI,YAAY;AACzD,UAAI,QAAQ,UAAU,UAAa,QAAQ,UAAU,IAAI,OAAO;AAC9D,cAAM,IAAI;AAAA,UACR,uBAAuB,QAAQ,KAAK,2BAA2B,IAAI,KAAK;AAAA,UACxE;AAAA,QACF;AAAA,MACF;AACA,UAAI,QAAQ,eAAe,UAAa,QAAQ,eAAe,IAAI,YAAY;AAC7E,cAAM,IAAI;AAAA,UACR,4BAA4B,QAAQ,UAAU,gCAAgC,IAAI,UAAU;AAAA,UAC5F;AAAA,QACF;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,yBAA0B,OAAM;AACrD,UAAI,iBAAiB,OAAO;AAC1B,cAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;AAAA,MAClE;AACA,YAAM;AAAA,IACR;AAAA,EACF;AAGA,MAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GAAG;AACzF,UAAM,IAAI;AAAA,MACR,2BAA2B,WAAW,KAAK,IAAI,CAAC,SAAS,OAAO,IAAI,QAAQ,CAAC;AAAA,MAC7E;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAUO,SAAS,yBACd,KACmB;AACnB,MAAI,IAAI,eAAgB,QAAO,IAAI;AACnC,MAAI,IAAI,QAAQ,IAAI,mBAAmB,GAAG;AACxC,WAAO,EAAE,MAAM,aAAa,KAAK,IAAI,QAAQ;AAAA,EAC/C;AACA,MAAI,IAAI,UAAU,EAAG,QAAO,EAAE,MAAM,YAAY,KAAK,IAAI,QAAQ;AACjE,SAAO,EAAE,MAAM,cAAc,KAAK,KAAK;AACzC;AAEA,SAAS,uBAAuB,OAAgB,SAAuB;AACrE,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;AAAA,EACzF;AACA,QAAM,QAAQ;AACd,MAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAAc;AAC1F,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,SAAS,cAAc;AAC/B,QAAI,MAAM,QAAQ,MAAM;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,YAAY,GAAG;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA;AAAA,EACF;AACA,qBAAmB,MAAM,KAAK,oBAAoB;AAClD,MAAK,MAAM,MAAiB,GAAG;AAC7B,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,QAAQ,SAAS;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAGO,SAAS,YAAY,OAAoC;AAC9D,MAAI;AACF,sBAAkB,KAAK;AACvB,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGO,SAAS,mBACd,OACiF;AACjF,MAAI;AACF,WAAO,EAAE,IAAI,MAAM,OAAO,kBAAkB,KAAK,EAAE;AAAA,EACrD,SAAS,GAAG;AACV,QAAI,aAAa,yBAA0B,QAAO,EAAE,IAAI,OAAO,OAAO,EAAE;AACxE,UAAM;AAAA,EACR;AACF;AAGO,SAAS,mBAAmB,QAA8B;AAC/D,QAAM,OAAO,KAAK,UAAU,MAAM;AAClC,SAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;AACxD,MAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAAG;AACnD,UAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAAA,EACtE;AACF;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;AAC9D,MAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;AACxD,UAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAAA,EACnE;AACF;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;AAC/D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iCAAiC,IAAI;AAAA,EAC1E;AACA,QAAM,MAAM;AAEZ,QAAM,WAAW,IAAI;AACrB,MAAI,aAAa,QAAQ,OAAO,aAAa,UAAU;AACrD,UAAM,IAAI,yBAAyB,8BAA8B,GAAG,IAAI,WAAW;AAAA,EACrF;AACA,aAAW,CAAC,SAAS,IAAI,KAAK,OAAO,QAAQ,QAAmC,GAAG;AACjF,QAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI,aAAa,OAAO;AAAA,MAC7B;AAAA,IACF;AACA,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,IAA+B,GAAG;AAC1E,yBAAmB,OAAO,GAAG,IAAI,aAAa,OAAO,IAAI,GAAG,EAAE;AAAA,IAChE;AAAA,EACF;AAEA,QAAM,aAAa,IAAI;AACvB,MAAI,eAAe,QAAQ,OAAO,eAAe,UAAU;AACzD,UAAM,IAAI,yBAAyB,gCAAgC,GAAG,IAAI,aAAa;AAAA,EACzF;AACA,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,UAAqC,GAAG;AAC/E,uBAAmB,MAAM,GAAG,IAAI,eAAe,GAAG,EAAE;AAAA,EACtD;AAEA,qBAAmB,IAAI,WAAW,GAAG,IAAI,YAAY;AAErD,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GAAG;AACpC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI;AAAA,MACT;AAAA,IACF;AACA,aAAS,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;AAChD,YAAM,KAAK,IAAI,aAAa,CAAC;AAC7B,UAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAAG;AAC7C,cAAM,IAAI;AAAA,UACR;AAAA,UACA,GAAG,IAAI,iBAAiB,CAAC;AAAA,QAC3B;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,MAAI,IAAI,UAAU,UAAa,OAAO,IAAI,UAAU,UAAU;AAC5D,UAAM,IAAI,yBAAyB,0BAA0B,GAAG,IAAI,QAAQ;AAAA,EAC9E;AACF;AAYO,SAAS,iBAAiB,OAAwB;AACvD,MAAI,MAAM,SAAS,GAAG,EAAG,QAAO;AAChC,MAAI,UAAU,KAAK,KAAK,EAAG,QAAO;AAClC,MAAI,sBAAsB,KAAK,KAAK,EAAG,QAAO;AAC9C,MAAI,SAAS,KAAK,KAAK,EAAG,QAAO;AACjC,SAAO;AACT;","names":[]}

package/dist/{chunk-N6MTC3GK.js → chunk-ADYLPOSX.js}

@@ -13,7 +13,7 @@ import {
 } from "./chunk-GGE4NNQT.js";
 import {
   callLlm
-} from "./chunk-FUCQVFMU.js";
+} from "./chunk-GY4SYVPJ.js";
 import {
   ValidationError
 } from "./chunk-ONWEPEDO.js";
@@ -2633,4 +2633,4 @@ export {
   provenanceSpansPath,
   emitLoopProvenance
 };
-//# sourceMappingURL=chunk-N6MTC3GK.js.map
+//# sourceMappingURL=chunk-ADYLPOSX.js.map

package/dist/{chunk-IMWDSFUM.js → chunk-DXZRATT5.js}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-TVVP3ZZQ.js";
 import {
   validateRunRecord
-} from "./chunk-VK6HBGAE.js";
+} from "./chunk-5UF54T55.js";
 
 // src/action-policy.ts
 function evaluateActionPolicy(action, policy = {}, options = {}) {
@@ -1523,4 +1523,4 @@ export {
   runProposeReviewAsControlLoop,
   controlFailureClassFromVerification
 };
-//# sourceMappingURL=chunk-IMWDSFUM.js.map
+//# sourceMappingURL=chunk-DXZRATT5.js.map