@remnic/bench 9.21.0 → 9.22.1

package/dist/index.d.ts

@@ -125,6 +125,15 @@ interface BenchRecallSupportRequest {
 }
 interface BenchResponder {
     respond(question: string, recalledText: string, control?: BenchPhaseControl): Promise<BenchResponse>;
+    /**
+     * Deterministic non-secret fingerprint of the responder's identity
+     * (e.g. model name + provider base URL hash). Used by the published
+     * harness to pin paired-run replay keys to a concrete responder so a
+     * pair of systems with no `systemProvider` cannot cross-replay through
+     * a shared cache. Optional: responders that do not declare an identity
+     * disable the replay cache and always invoke `respond` directly.
+     */
+    identity?(): string;
 }
 interface BenchJudgeResult {
     score: number;
@@ -590,6 +599,13 @@ interface BenchmarkDefinition {
     runnerAvailable: boolean;
     meta: BenchmarkMeta;
 }
+interface PairedAnswerReplayEntry {
+    sourceRuntimeProfile: BenchRuntimeProfile | null;
+    finalAnswer: string;
+    answeredText: string;
+    model?: string;
+}
+type PairedAnswerReplayCache = Map<string, PairedAnswerReplayEntry>;
 interface RunBenchmarkOptions {
     mode?: BenchmarkMode;
     datasetDir?: string;
@@ -625,6 +641,12 @@ interface RunBenchmarkOptions {
      * `noJudgeCache` is true. The directory is created on demand.
      */
     judgeCacheDir?: string;
+    /**
+     * Ephemeral cross-profile answer cache for a paired benchmark matrix. A
+     * cached answer may only be reused when the responder-facing input is
+     * identical and it originated from a different runtime profile.
+     */
+    pairedAnswerReplayCache?: PairedAnswerReplayCache;
     /** Called after each task completes for progress logging and partial result tracking. */
     onTaskComplete?: (task: TaskResult, completedCount: number, totalCount?: number) => void;
 }
@@ -2617,7 +2639,7 @@ interface ProviderResponderOptions {
     contextBudgetChars?: number;
     promptBudgetChars?: number;
 }
-declare function createResponderFromProvider(provider: LlmProvider, options?: ProviderResponderOptions): BenchResponder;
+declare function createResponderFromProvider(provider: LlmProvider, options?: ProviderResponderOptions, responderIdentity?: string): BenchResponder;
 declare function createProviderBackedResponder(config: ProviderFactoryConfig, providerInstance?: LlmProvider): BenchResponder;
 declare function createProviderBackedJudge(config: ProviderFactoryConfig, providerInstance?: LlmProvider): BenchJudge;
 declare function createProviderBackedAmaBenchRecommendedJudge(config: ProviderFactoryConfig, providerInstance?: LlmProvider): BenchJudge;
@@ -5781,4 +5803,4 @@ declare function checkCodingGraphRegression(report: CodingGraphBenchReport, base
  */
 declare function buildBaselineFromReport(report: CodingGraphBenchReport, note: string): CodingGraphBaseline;
 
-export { AMA_BENCH_DIAGNOSTIC_VARIANTS, ASSISTANT_AGENT_CONFIG_KEY, ASSISTANT_JUDGE_CONFIG_KEY, ASSISTANT_MEETING_PREP_SCENARIOS, ASSISTANT_MEETING_PREP_SMOKE_SCENARIOS, ASSISTANT_MORNING_BRIEF_SCENARIOS, ASSISTANT_MORNING_BRIEF_SMOKE_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SMOKE_SCENARIOS, ASSISTANT_RUBRIC_DIMENSIONS, ASSISTANT_RUBRIC_ID_KEY, ASSISTANT_SEEDS_CONFIG_KEY, ASSISTANT_SPOT_CHECK_DIR_KEY, ASSISTANT_SYNTHESIS_SCENARIOS, ASSISTANT_SYNTHESIS_SMOKE_SCENARIOS, type AblationConfigOverrides, type AbstentionRetrievalCase, type AggregateMetrics, type AmaBenchDiagnosticAdapterOptions, type AmaBenchDiagnosticAnswererMode, type AmaBenchDiagnosticBreakdown, type AmaBenchDiagnosticMatrixArtifact, type AmaBenchDiagnosticRecallMode, type AmaBenchDiagnosticRunContext, type AmaBenchDiagnosticTaskEvidence, type AmaBenchDiagnosticTaskRow, type AmaBenchDiagnosticVariant, type AmaBenchDiagnosticVariantSummary, type AnthropicProviderConfig, type AssistantAgent, type AssistantMemoryFact, type AssistantMemoryGraph, type AssistantRubricDimension, type AssistantRubricRequest, type AssistantRubricScores, type AssistantRunnerOptions, type AssistantScenario, type AssistantStance, type AttackRecallOptions, type AttackRetrievalHit, type AttackerMode, BENCHMARK_ARTIFACT_SCHEMA_VERSION, BENCHMARK_INTEGRITY_META_SCHEMA, BENCHMARK_REPRO_MANIFEST_FILENAME, BENCHMARK_REPRO_MANIFEST_SCHEMA_VERSION, BENCHMARK_RESULT_SCHEMA, BENCHMARK_SPLIT_TYPES, BUILD_WEEK_EVIDENCE_RECEIPT_SCHEMA_VERSION, BUILD_WEEK_LIMITATIONS, type BaselineRow, type BaselineScenario, type BeamDatasetPreview, type BenchConfig, type BenchJudge, type BenchJudgeResult, type BenchMemoryAdapter, type BenchModelSource, type BenchReasoningEffort, type BenchRecallLineageStatus, type BenchRecallOptions, type BenchRecallSupportAssessment, type BenchRecallSupportRequest, type BenchRecallSupportStatus, type BenchRecallTrace, type BenchRecallTraceCoreCapture, type BenchRecallTraceLcmCandidate, type BenchRecallTraceRange, type BenchRecallTraceSection, type BenchRecallTraceSelection, type BenchRecallWithTraceResult, type BenchResponder, type BenchResponse, type BenchRuntimeProfile, type BenchTier, type BenchmarkArtifact, type BenchmarkArtifactEnvironment, type BenchmarkArtifactHardware, type BenchmarkArtifactJudgeCalibration, type BenchmarkArtifactPerTaskScore, type BenchmarkArtifactSystem, type BenchmarkArtifactTier, type BenchmarkCategory, type BenchmarkDefinition, type BenchmarkIntegrityMeta, type BenchmarkMeta, type BenchmarkMode, type BenchmarkReport, type BenchmarkReproManifest, type BenchmarkReproManifestDataset, type BenchmarkReproManifestFile, type BenchmarkReproManifestResult, type BenchmarkResult, type BenchmarkSplitType, type BenchmarkStatus, type BenchmarkSuiteResult, type BenchmarkTier, type BootstrapKappaOptions, type BootstrapKappaResult, type BuildBenchmarkArtifactInput, type BuildBenchmarkPublishFeedOptions, type BuildBenchmarkReproManifestOptions, type BuildBuildWeekEvidenceReceiptOptions, type BuildWeekEvidenceReceipt, type BuildWeekEvidenceReceiptProvider, type BuildWeekLimitationCode, type BuiltInProvider, CALIBRATION_SLICE_SIZE, CANARY_FIXED_RECALL, CANARY_SCORE_FLOOR, DEFAULT_10K_FIXTURE as CODING_GRAPH_10K_FIXTURE, CODING_GRAPH_BENCH_SCHEMA_VERSION, DEFAULT_TOLERANCE_PERCENT as CODING_GRAPH_DEFAULT_TOLERANCE, MIN_ITERATIONS as CODING_GRAPH_MIN_ITERATIONS, DEFAULT_SMOKE_FIXTURE as CODING_GRAPH_SMOKE_FIXTURE, type CalibrationAnswer, type CalibrationVerdictPair, type CanaryAdapterOptions, type CanaryFloorCheck, type CaptureLoCoMoRetrievalTraceOptions, type ClaudeCliProviderConfig, type CodexCliNativeUsage, type CodexCliProviderConfig, CodexCreditAccountingError, type CodexCreditBudgetConfig, CodexCreditDispatchError, type CodexCreditReceipt, type CodexCreditReceiptScope, type CodexCreditReconciliationReceipt, type CodingGraphBaseline, type CodingGraphBenchConfig, type CodingGraphBenchReport, type MachineFingerprint as CodingGraphMachineFingerprint, type CodingGraphMetricKey, type RegressionMetricDetail as CodingGraphRegressionDetail, type RegressionMetricKey as CodingGraphRegressionKey, type RegressionGateResult as CodingGraphRegressionResult, type CohenKappaResult, type ComparisonMetricDelta, type ComparisonResult, type CompletionOpts, type CompletionResult, type ConfidenceInterval, type ContaminationCheckResult, type ContaminationEntry, type ContaminationManifest, type CustomBenchmarkScoring, type CustomBenchmarkSpec, type CustomBenchmarkTask, DEFAULT_ABLATION_BENCHMARK, DEFAULT_ABLATION_BOOTSTRAP_SEED, DEFAULT_ASSISTANT_RUBRIC_ID, DEFAULT_BASELINE_SCENARIOS, DEFAULT_JUDGE_BINARIZATION_THRESHOLD, DEFAULT_KAPPA_BOOTSTRAP_SAMPLES, DEFAULT_KAPPA_CONFIDENCE_LEVEL, DEFAULT_OPENAI_RESPONSES_JUDGE_MODEL, type DatasetSource, type DiagnoseLoComoProfileDeltaOptions, type DiagnoseLoComoRecallDeltaOptions, type DiscoveredModel, EMPTY_CONTAMINATION_MANIFEST, type EffectSizeInterpretation, type EffectSizeSummary, type ExplainResult, type ExtractedEntity, type ExtractedLink, type ExtractedPage, type ExtractionAttackOptions, type ExtractionAttackResult, type ExtractionAttackTarget, type FixtureGenerator, type FixtureOutput, type FixtureVariant, GENERAL_ANSWER_JUDGE_RUBRIC, type GeneratedFile, type GeneratedRepo, type GoldEntity, type GoldEntityType, type GoldGraph, type GoldLink, type GoldPage, type HarnessRng, INTEGRITY_CIPHER_ALGORITHM, INTEGRITY_HASH_ALGORITHM, INTEGRITY_META_FIELDS, type IngestionBenchAdapter, type IngestionLog, JUDGE_CALIBRATION_KAPPA_THRESHOLD, JUDGE_CALIBRATION_PROTOCOL_VERSION, type JudgeCalibrationCheckpointProvenance, type JudgeCalibrationIdentities, type JudgeCalibrationResult, type JudgeCategory, type KappaConfidenceInterval, LOCAL_LAB_PROVIDER_KINDS, LOCOMO_DATASET_FILENAMES, LOCOMO_FULL_TASK_COUNT, LOCOMO_RECALL_DIFF_LINE_LIMIT, LOCOMO_RECALL_EXCERPT_CHARS, LOCOMO_RETRIEVAL_TRACE_BUDGET_VERSION, LOCOMO_RETRIEVAL_TRACE_DELTA_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SELECTION_VERSION, LONG_MEM_EVAL_DATASET_FILENAMES, type LeaderboardArtifactWrite, type LettaAdapterConfig, LettaMemCorrectAdapter, type LlmJudge, type LlmProvider, type LoCoMoCategory, type LoCoMoRetrievalMechanism, type LoCoMoRetrievalMechanismSummary, type LoCoMoRetrievalSessionReceipt, type LoCoMoRetrievalStructuralTrace, type LoCoMoRetrievalTaskDelta, type LoCoMoRetrievalTaskReceipt, type LoCoMoRetrievalTraceCoreCaptureReceipt, type LoCoMoRetrievalTraceDeltaReport, type LoCoMoRetrievalTraceProfile, type LoCoMoRetrievalTraceReceipt, type LoCoMoRetrievalTraceSelectionManifest, type LoCoMoRetrievalTraceSelector, type LoCoMoStructuralMultisetDelta, type LoComoCategoryDelta, type LoComoFinalContextRegression, type LoComoMetricDelta, type LoComoProfileArtifactEvidence, type LoComoProfileDeltaReport, type LoComoRawResultEvidence, type LoComoRecallCategoryDelta, type LoComoRecallContextSummary, type LoComoRecallDeltaReport, type LoComoRecallLineDelta, type LoComoRecallLineEvidence, type LoComoRecallMetricDelta, type LoComoRecallResultProvenance, type LoComoRecallTextDigest, type LoComoTaskRegression, type LoadDatasetOptions, type LoadSealedQrelsOptions, type LoadedDataset, type LoadedJudgeCalibrationState, type LocalLabManifest, type LocalLabManifestNotes, type LocalLabPhase, type LocalLabPhaseDescriptor, type LocalLabPhaseExecute, type LocalLabPhaseName, type LocalLabPhaseOutcome, LocalLabPreflightError, type LocalLabPreflightFailure, type LocalLabPreflightInput, type LocalLabPreflightOptions, type LocalLabPreflightResult, type LocalLabPreflightSuccess, type LocalLabProviderKind, type LocalLabRoleConfig, type LocalLlmProviderConfig, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC_VERSION, MEMCORRECT_STALE_HARM_RUBRIC, MEMCORRECT_STALE_HARM_RUBRIC_VERSION, MEMORY_EVAL_DIMENSIONS, MEMORY_EVAL_PUBLIC_LINE, MIN_CALIBRATION_SOURCE_TASKS, MITIGATED_BASELINE_SCENARIOS, type McpArgumentSemantic, type McpBackendErrorCode, type McpBackendResult, type McpBenchMemoryAdapter, type McpConformanceResult, type McpHttpTransportConfig, type McpListedTool, type McpMemCorrectAdapter, type McpMemoryAdapterOptions, McpMemoryBackendError, type McpMemoryToolMapping, type McpMemoryTransportConfig, type McpStdioTransportConfig, type McpToolCallResult, type McpToolClient, type McpToolMappingEntry, type McpToolMappingValue, type McpToolOperation, type Mem0AdapterConfig, Mem0MemCorrectAdapter, type MemCorrectGeneratorOptions, type MemCorrectJudgeRequest, type MemCorrectJudgeResult, type MemCorrectSystemAdapter, type MemoryEvalCategory, type MemoryEvalDimension, type MemoryEvalDimensionId, type MemoryEvalMetric, type MemoryGraph, type MemoryStats, type MemorySystem, type Message, type MetricAggregate, type MicroMetric, MissingCredentialError, type MitigatedBaselineConfig, type MitigatedTargetConfig, type MultipleChoiceQuestion, OPENAI_RESPONSES_JUDGE_RUBRIC_VERSION, OTHER_NAMESPACE_MEMORIES, type OllamaProviderConfig, type OpenAiCompatibleProviderConfig, OpenAiResponsesJudgeError, type OpenAiResponsesJudgeErrorCode, type OpenAiResponsesJudgeTelemetry, OpenAiResponsesProvider, type OpenAiResponsesProviderConfig, type OpenAiResponsesVerdict, type OpenAiResponsesVerdictResult, PROCEDURAL_REAL_SCENARIOS, PROCEDURAL_REAL_SCENARIOS_SMOKE, PUBLISHED_BENCHMARK_ARTIFACT_IDS, type PersonalizationRetrievalCase, type PreflightDiscoveredModel, type ProceduralAblationArtifact, type ProceduralAblationPerCase, type ProceduralAblationScenario, type ProceduralRealScenario, type ProceduralRealScenarioCategory, type ProviderBaseConfig, type ProviderConfig, type ProviderDiscoveryResult, type ProviderFactoryConfig, type PublishSkipReason, type PublishSkipRecord, type PublishedBenchmarkFeed, type PublishedBenchmarkFeedEntry, type PublishedBenchmarkId, REQUIRED_FRONTMATTER_FIELDS, type RecallMetrics, type RecoveredMemory, type RegressionDetail, type RegressionGateResult$1 as RegressionGateResult, type RemnicAdapterOptions, type ReportCardProvenanceContext, type ResolveBenchRuntimeProfileOptions, type ResolvedBenchRuntimeProfile, type ResolvedLocalLabProfile, type ResolvedLocalLabRole, type ResolvedRunBenchmarkOptions, type RotatedChoices, type RunBenchmarkOptions, type RunJudgeCalibrationOptions, type RunProceduralAblationCliArgs, type RunProceduralAblationOptions, type RunSequentialPhasesOptions, SCHEMA_TIER_FIXTURE, SCHEMA_TIER_SMOKE_FIXTURE, SEALED_PROMPT_REGISTRY, SINGLE_FLAG_ABLATION_MATRIX, SYNTHETIC_MEMORIES, type SanitizedDiagnosticProvider, type SavedBaseline, type SchemaTierCorpus, type SchemaTierFixture, type SchemaTierName, type SchemaTierPage, type SchemaTierPageFrontmatter, type SealedArtifact, type SealedJudgeDecision, type SealedJudgeInput, type SealedQrelsArtifact, type SealedQrelsHandle, type SealedRubric, type SearchResult, type SeededMemory, type SeededRng, type SequentialPhaseHooks, type SingleFlagAblationCell, type SingleFlagAblationId, type SpotCheckLogger, type StatisticalReport, type StructuredJudge, StructuredJudgeError, type StructuredJudgeErrorCode, type StructuredJudgeProvider, type StructuredJudgeTelemetry, type StructuredJudgeVerdict, type StructuredJudgeVerdictResult, type StructuredVerdictRequest, type SyntheticEdge, type SyntheticEmailIngestionAdapterOptions, type SyntheticFileIR, type SyntheticRepoConfig, type SyntheticSymbol, type SyntheticTargetOptions, type TaskResult, type TaskTokenUsage, type TemporalRetrievalCase, type ThirdPartyAdapterConfig, type TierDetail, type TimelineEntry, type TokenUsage, type WallMetric, type WriteBenchmarkArtifactResult, type ZepAdapterConfig, ZepMemCorrectAdapter, addContaminationEntry, aggregateTaskScores, answerBenchmarkQuestion, assertCanaryUnderFloor, assertIntegrityMetaPresent, assertPublishableIntegrity, assertSha256Hex, assistantMeetingPrepDefinition, assistantMorningBriefDefinition, assistantNextBestActionDefinition, assistantSynthesisDefinition, backlinkF1, binarizeJudgeScore, bootstrapCohensKappaConfidenceInterval, bootstrapMeanConfidenceInterval, buildAmaBenchDiagnosticMatrixArtifact, buildAmaBenchDiagnosticVariantSummary, buildAmaBenchLeaderboardRows, buildBaselineFromReport, buildBenchmarkArtifact, buildBenchmarkArtifactFilename, buildBenchmarkPublishFeed, buildBenchmarkReproManifest, buildBenchmarkRunSeeds, buildBuildWeekEvidenceReceipt, buildCodexCreditReceipt, buildJudgePayload, buildOracleTrajectoryRecall, buildProviderFreeLoCoMoRetrievalConfig, buildSchemaTierFixture, buildSchemaTierSmokeFixture, calculateCodexBudgetUnits, calendarFixture, canonicalJsonStringify, captureLoCoMoRetrievalTrace, captureMachineFingerprint, chatFixture, checkCodingGraphRegression, checkDatasetContamination, checkRegression, clampScore, cohensD, compareResults, computeCohensKappa, computeSealHash, containsAnswer, createSeededRng$1 as createAdamSeededRng, createAmaBenchDiagnosticAdapter, createAnthropicProvider, createCanaryAdapter, createClaudeCliProvider, createCodexCliProvider, createSeededRng as createCodingGraphSeededRng, createDeterministicSpotCheckLogger, createGatewayResponder, createLightweightAdapter, createLiteLlmProvider, createLocalLlmProvider, createMcpDemoMemCorrectAdapter, createMcpDemoMemoryAdapter, createMcpMemCorrectAdapter, createMcpMemoryAdapter, createMitigatedTarget, createOllamaProvider, createOpenAiCompatibleProvider, createOpenAiResponsesBenchJudge, createOpenAiResponsesProvider, createSeededRandom as createProceduralAblationSeededRandom, createProvider, createProviderBackedAmaBenchRecommendedJudge, createProviderBackedJudge, createProviderBackedResponder, createProviderBackedStructuredJudge, createRemnicAdapter, createResponderFromProvider, createSeededRng$2 as createSeededRng, createSpotCheckFileLogger, createStructuredBenchJudge, createStructuredJudgeFromProvider, createSyntheticEmailIngestionAdapter, createSyntheticTarget, createTimeoutGuardedAdapter, defaultBenchmarkBaselineDir, defaultBenchmarkPublishPath, deleteBenchmarkResults, diagnoseLoCoMoRetrievalTraceDelta, diagnoseLoComoProfileDelta, diagnoseLoComoRecallDelta, discoverAllProviders, discoveryEndpointFor, emailFixture, entityRecall, exactMatch, extractMetrics as extractCodingGraphMetrics, extractMarkdownSectionsByTitle, f1Score, fixtureToAblationScenarios, formatHandoffNote, formatMissingDatasetError, generateReport, generateSyntheticRepo, getAblationCell, getBenchmark, getBenchmarkLowerIsBetter, getGitSha, getMemoryEvalDimension, getProviderBackedJudgePromptIdentity, getRemnicVersion, hashBenchmarkArtifact, hashBytes, hashCanonicalJson, hashOrderedQuestionIds, hashString, integrityMetaIsComplete, interpretEffectSize, isAmaBenchUnknownLikeAnswer, isContaminationEntry, isContaminationManifest, isSealedQrelsArtifact, isSha256Hex, isStructuredJudgeProvider, judgeMemCorrectCorrectionAcceptance, judgeMemCorrectStaleMemoryHarm, linkMatches, listBenchmarkBaselines, listBenchmarkResults, listBenchmarks, listMemoryEvalBenchmarkIds, listMemoryEvalDimensions, llmJudgeScore, llmJudgeScoreDetailed, loadAblationFixture, loadBaseline, loadBeamDatasetPreview, loadBenchmarkArtifact, loadBenchmarkBaseline, loadBenchmarkReportCardProvenance, loadBenchmarkResult, loadCustomBenchmarkFile, loadJudgeCalibrationState, loadLoCoMo10, loadLocalLabManifest, loadLongMemEvalS, loadSealKeyFromEnv, loadSealedQrels, loadSealedRubric, matchEntity, mergeContaminationManifests, openSeal, orchestrateBenchmarkRuns, pairedDeltaConfidenceInterval, parseBenchmarkArtifact, parseCodexJsonlUsage, parseCustomBenchmark, parseLocalLabManifest, parseRubricResponse, parseSealedQrels, pickStableQualifiedName, precisionAtK, preflightLoCoMoRetrievalTraceCapture, preflightLocalLabRole, projectFolderFixture, recallAtK, reconcileCodexCreditLedger, redactBenchmarkResultSecrets, renderBaselineMarkdown, renderBenchmarkResultExport, renderLoComoProfileDeltaMarkdown, renderLoComoRecallDeltaMarkdown, renderMemorySummaryForJudge, renderMemoryViewForAgent, resolveAssistantAgent, resolveAssistantRubricId, resolveAssistantSeeds, resolveAssistantSpotCheckDir, resolveBenchRuntimeProfile, resolveBenchmarkPhaseTimeoutMs, resolveBenchmarkProgressLogging, resolveBenchmarkResultReference, resolveBenchmarkRunCount, resolveCodexCreditBudgetConfig, resolveLocalLabJudgeProviderConfig, resolveLocalLabProfile, resolveLocalLabRole, resolveStructuredJudge, rotateDistractors, rougeL, runAssistantBenchmark, runAssistantMeetingPrepBenchmark, runAssistantMorningBriefBenchmark, runAssistantNextBestActionBenchmark, runAssistantSynthesisBenchmark, runBaseline, runBenchSuite, runBenchmark, runCodingGraphBenchmark, runCustomBenchmarkFile, runExplain, runExtractionAttack, runJudgeCalibration, runMitigatedBaseline, runProceduralAblation, runProceduralAblationCli, runSealedJudge, runSequentialPhases, runWithinCodexCreditBudget, safeHexEqual, sanitizeLoComoResultReference, saveBaseline, saveBenchmarkBaseline, schemaCompleteness, sealPayload, selectAmaBenchDiagnosticVariants, selectCalibrationSlice, selectFixtureVariant, serializeBenchmarkArtifact, serializeBuildWeekEvidenceReceipt, serializeJsonl, serializeLoCoMoRetrievalTraceDelta, serializeLoCoMoRetrievalTraceReceipt, serializeSealedQrels, shuffleTasks, timed, verifyRubricDigest, writeBenchmarkArtifact, writeBenchmarkPublishFeed, writeBenchmarkReproManifest, writeBenchmarkResult, writeBuildWeekEvidenceReceipt, writeJudgeCalibrationState, writeLeaderboardArtifactsForResult, zeroScores };
+export { AMA_BENCH_DIAGNOSTIC_VARIANTS, ASSISTANT_AGENT_CONFIG_KEY, ASSISTANT_JUDGE_CONFIG_KEY, ASSISTANT_MEETING_PREP_SCENARIOS, ASSISTANT_MEETING_PREP_SMOKE_SCENARIOS, ASSISTANT_MORNING_BRIEF_SCENARIOS, ASSISTANT_MORNING_BRIEF_SMOKE_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SMOKE_SCENARIOS, ASSISTANT_RUBRIC_DIMENSIONS, ASSISTANT_RUBRIC_ID_KEY, ASSISTANT_SEEDS_CONFIG_KEY, ASSISTANT_SPOT_CHECK_DIR_KEY, ASSISTANT_SYNTHESIS_SCENARIOS, ASSISTANT_SYNTHESIS_SMOKE_SCENARIOS, type AblationConfigOverrides, type AbstentionRetrievalCase, type AggregateMetrics, type AmaBenchDiagnosticAdapterOptions, type AmaBenchDiagnosticAnswererMode, type AmaBenchDiagnosticBreakdown, type AmaBenchDiagnosticMatrixArtifact, type AmaBenchDiagnosticRecallMode, type AmaBenchDiagnosticRunContext, type AmaBenchDiagnosticTaskEvidence, type AmaBenchDiagnosticTaskRow, type AmaBenchDiagnosticVariant, type AmaBenchDiagnosticVariantSummary, type AnthropicProviderConfig, type AssistantAgent, type AssistantMemoryFact, type AssistantMemoryGraph, type AssistantRubricDimension, type AssistantRubricRequest, type AssistantRubricScores, type AssistantRunnerOptions, type AssistantScenario, type AssistantStance, type AttackRecallOptions, type AttackRetrievalHit, type AttackerMode, BENCHMARK_ARTIFACT_SCHEMA_VERSION, BENCHMARK_INTEGRITY_META_SCHEMA, BENCHMARK_REPRO_MANIFEST_FILENAME, BENCHMARK_REPRO_MANIFEST_SCHEMA_VERSION, BENCHMARK_RESULT_SCHEMA, BENCHMARK_SPLIT_TYPES, BUILD_WEEK_EVIDENCE_RECEIPT_SCHEMA_VERSION, BUILD_WEEK_LIMITATIONS, type BaselineRow, type BaselineScenario, type BeamDatasetPreview, type BenchConfig, type BenchJudge, type BenchJudgeResult, type BenchMemoryAdapter, type BenchModelSource, type BenchReasoningEffort, type BenchRecallLineageStatus, type BenchRecallOptions, type BenchRecallSupportAssessment, type BenchRecallSupportRequest, type BenchRecallSupportStatus, type BenchRecallTrace, type BenchRecallTraceCoreCapture, type BenchRecallTraceLcmCandidate, type BenchRecallTraceRange, type BenchRecallTraceSection, type BenchRecallTraceSelection, type BenchRecallWithTraceResult, type BenchResponder, type BenchResponse, type BenchRuntimeProfile, type BenchTier, type BenchmarkArtifact, type BenchmarkArtifactEnvironment, type BenchmarkArtifactHardware, type BenchmarkArtifactJudgeCalibration, type BenchmarkArtifactPerTaskScore, type BenchmarkArtifactSystem, type BenchmarkArtifactTier, type BenchmarkCategory, type BenchmarkDefinition, type BenchmarkIntegrityMeta, type BenchmarkMeta, type BenchmarkMode, type BenchmarkReport, type BenchmarkReproManifest, type BenchmarkReproManifestDataset, type BenchmarkReproManifestFile, type BenchmarkReproManifestResult, type BenchmarkResult, type BenchmarkSplitType, type BenchmarkStatus, type BenchmarkSuiteResult, type BenchmarkTier, type BootstrapKappaOptions, type BootstrapKappaResult, type BuildBenchmarkArtifactInput, type BuildBenchmarkPublishFeedOptions, type BuildBenchmarkReproManifestOptions, type BuildBuildWeekEvidenceReceiptOptions, type BuildWeekEvidenceReceipt, type BuildWeekEvidenceReceiptProvider, type BuildWeekLimitationCode, type BuiltInProvider, CALIBRATION_SLICE_SIZE, CANARY_FIXED_RECALL, CANARY_SCORE_FLOOR, DEFAULT_10K_FIXTURE as CODING_GRAPH_10K_FIXTURE, CODING_GRAPH_BENCH_SCHEMA_VERSION, DEFAULT_TOLERANCE_PERCENT as CODING_GRAPH_DEFAULT_TOLERANCE, MIN_ITERATIONS as CODING_GRAPH_MIN_ITERATIONS, DEFAULT_SMOKE_FIXTURE as CODING_GRAPH_SMOKE_FIXTURE, type CalibrationAnswer, type CalibrationVerdictPair, type CanaryAdapterOptions, type CanaryFloorCheck, type CaptureLoCoMoRetrievalTraceOptions, type ClaudeCliProviderConfig, type CodexCliNativeUsage, type CodexCliProviderConfig, CodexCreditAccountingError, type CodexCreditBudgetConfig, CodexCreditDispatchError, type CodexCreditReceipt, type CodexCreditReceiptScope, type CodexCreditReconciliationReceipt, type CodingGraphBaseline, type CodingGraphBenchConfig, type CodingGraphBenchReport, type MachineFingerprint as CodingGraphMachineFingerprint, type CodingGraphMetricKey, type RegressionMetricDetail as CodingGraphRegressionDetail, type RegressionMetricKey as CodingGraphRegressionKey, type RegressionGateResult as CodingGraphRegressionResult, type CohenKappaResult, type ComparisonMetricDelta, type ComparisonResult, type CompletionOpts, type CompletionResult, type ConfidenceInterval, type ContaminationCheckResult, type ContaminationEntry, type ContaminationManifest, type CustomBenchmarkScoring, type CustomBenchmarkSpec, type CustomBenchmarkTask, DEFAULT_ABLATION_BENCHMARK, DEFAULT_ABLATION_BOOTSTRAP_SEED, DEFAULT_ASSISTANT_RUBRIC_ID, DEFAULT_BASELINE_SCENARIOS, DEFAULT_JUDGE_BINARIZATION_THRESHOLD, DEFAULT_KAPPA_BOOTSTRAP_SAMPLES, DEFAULT_KAPPA_CONFIDENCE_LEVEL, DEFAULT_OPENAI_RESPONSES_JUDGE_MODEL, type DatasetSource, type DiagnoseLoComoProfileDeltaOptions, type DiagnoseLoComoRecallDeltaOptions, type DiscoveredModel, EMPTY_CONTAMINATION_MANIFEST, type EffectSizeInterpretation, type EffectSizeSummary, type ExplainResult, type ExtractedEntity, type ExtractedLink, type ExtractedPage, type ExtractionAttackOptions, type ExtractionAttackResult, type ExtractionAttackTarget, type FixtureGenerator, type FixtureOutput, type FixtureVariant, GENERAL_ANSWER_JUDGE_RUBRIC, type GeneratedFile, type GeneratedRepo, type GoldEntity, type GoldEntityType, type GoldGraph, type GoldLink, type GoldPage, type HarnessRng, INTEGRITY_CIPHER_ALGORITHM, INTEGRITY_HASH_ALGORITHM, INTEGRITY_META_FIELDS, type IngestionBenchAdapter, type IngestionLog, JUDGE_CALIBRATION_KAPPA_THRESHOLD, JUDGE_CALIBRATION_PROTOCOL_VERSION, type JudgeCalibrationCheckpointProvenance, type JudgeCalibrationIdentities, type JudgeCalibrationResult, type JudgeCategory, type KappaConfidenceInterval, LOCAL_LAB_PROVIDER_KINDS, LOCOMO_DATASET_FILENAMES, LOCOMO_FULL_TASK_COUNT, LOCOMO_RECALL_DIFF_LINE_LIMIT, LOCOMO_RECALL_EXCERPT_CHARS, LOCOMO_RETRIEVAL_TRACE_BUDGET_VERSION, LOCOMO_RETRIEVAL_TRACE_DELTA_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SELECTION_VERSION, LONG_MEM_EVAL_DATASET_FILENAMES, type LeaderboardArtifactWrite, type LettaAdapterConfig, LettaMemCorrectAdapter, type LlmJudge, type LlmProvider, type LoCoMoCategory, type LoCoMoRetrievalMechanism, type LoCoMoRetrievalMechanismSummary, type LoCoMoRetrievalSessionReceipt, type LoCoMoRetrievalStructuralTrace, type LoCoMoRetrievalTaskDelta, type LoCoMoRetrievalTaskReceipt, type LoCoMoRetrievalTraceCoreCaptureReceipt, type LoCoMoRetrievalTraceDeltaReport, type LoCoMoRetrievalTraceProfile, type LoCoMoRetrievalTraceReceipt, type LoCoMoRetrievalTraceSelectionManifest, type LoCoMoRetrievalTraceSelector, type LoCoMoStructuralMultisetDelta, type LoComoCategoryDelta, type LoComoFinalContextRegression, type LoComoMetricDelta, type LoComoProfileArtifactEvidence, type LoComoProfileDeltaReport, type LoComoRawResultEvidence, type LoComoRecallCategoryDelta, type LoComoRecallContextSummary, type LoComoRecallDeltaReport, type LoComoRecallLineDelta, type LoComoRecallLineEvidence, type LoComoRecallMetricDelta, type LoComoRecallResultProvenance, type LoComoRecallTextDigest, type LoComoTaskRegression, type LoadDatasetOptions, type LoadSealedQrelsOptions, type LoadedDataset, type LoadedJudgeCalibrationState, type LocalLabManifest, type LocalLabManifestNotes, type LocalLabPhase, type LocalLabPhaseDescriptor, type LocalLabPhaseExecute, type LocalLabPhaseName, type LocalLabPhaseOutcome, LocalLabPreflightError, type LocalLabPreflightFailure, type LocalLabPreflightInput, type LocalLabPreflightOptions, type LocalLabPreflightResult, type LocalLabPreflightSuccess, type LocalLabProviderKind, type LocalLabRoleConfig, type LocalLlmProviderConfig, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC_VERSION, MEMCORRECT_STALE_HARM_RUBRIC, MEMCORRECT_STALE_HARM_RUBRIC_VERSION, MEMORY_EVAL_DIMENSIONS, MEMORY_EVAL_PUBLIC_LINE, MIN_CALIBRATION_SOURCE_TASKS, MITIGATED_BASELINE_SCENARIOS, type McpArgumentSemantic, type McpBackendErrorCode, type McpBackendResult, type McpBenchMemoryAdapter, type McpConformanceResult, type McpHttpTransportConfig, type McpListedTool, type McpMemCorrectAdapter, type McpMemoryAdapterOptions, McpMemoryBackendError, type McpMemoryToolMapping, type McpMemoryTransportConfig, type McpStdioTransportConfig, type McpToolCallResult, type McpToolClient, type McpToolMappingEntry, type McpToolMappingValue, type McpToolOperation, type Mem0AdapterConfig, Mem0MemCorrectAdapter, type MemCorrectGeneratorOptions, type MemCorrectJudgeRequest, type MemCorrectJudgeResult, type MemCorrectSystemAdapter, type MemoryEvalCategory, type MemoryEvalDimension, type MemoryEvalDimensionId, type MemoryEvalMetric, type MemoryGraph, type MemoryStats, type MemorySystem, type Message, type MetricAggregate, type MicroMetric, MissingCredentialError, type MitigatedBaselineConfig, type MitigatedTargetConfig, type MultipleChoiceQuestion, OPENAI_RESPONSES_JUDGE_RUBRIC_VERSION, OTHER_NAMESPACE_MEMORIES, type OllamaProviderConfig, type OpenAiCompatibleProviderConfig, OpenAiResponsesJudgeError, type OpenAiResponsesJudgeErrorCode, type OpenAiResponsesJudgeTelemetry, OpenAiResponsesProvider, type OpenAiResponsesProviderConfig, type OpenAiResponsesVerdict, type OpenAiResponsesVerdictResult, PROCEDURAL_REAL_SCENARIOS, PROCEDURAL_REAL_SCENARIOS_SMOKE, PUBLISHED_BENCHMARK_ARTIFACT_IDS, type PairedAnswerReplayCache, type PairedAnswerReplayEntry, type PersonalizationRetrievalCase, type PreflightDiscoveredModel, type ProceduralAblationArtifact, type ProceduralAblationPerCase, type ProceduralAblationScenario, type ProceduralRealScenario, type ProceduralRealScenarioCategory, type ProviderBaseConfig, type ProviderConfig, type ProviderDiscoveryResult, type ProviderFactoryConfig, type PublishSkipReason, type PublishSkipRecord, type PublishedBenchmarkFeed, type PublishedBenchmarkFeedEntry, type PublishedBenchmarkId, REQUIRED_FRONTMATTER_FIELDS, type RecallMetrics, type RecoveredMemory, type RegressionDetail, type RegressionGateResult$1 as RegressionGateResult, type RemnicAdapterOptions, type ReportCardProvenanceContext, type ResolveBenchRuntimeProfileOptions, type ResolvedBenchRuntimeProfile, type ResolvedLocalLabProfile, type ResolvedLocalLabRole, type ResolvedRunBenchmarkOptions, type RotatedChoices, type RunBenchmarkOptions, type RunJudgeCalibrationOptions, type RunProceduralAblationCliArgs, type RunProceduralAblationOptions, type RunSequentialPhasesOptions, SCHEMA_TIER_FIXTURE, SCHEMA_TIER_SMOKE_FIXTURE, SEALED_PROMPT_REGISTRY, SINGLE_FLAG_ABLATION_MATRIX, SYNTHETIC_MEMORIES, type SanitizedDiagnosticProvider, type SavedBaseline, type SchemaTierCorpus, type SchemaTierFixture, type SchemaTierName, type SchemaTierPage, type SchemaTierPageFrontmatter, type SealedArtifact, type SealedJudgeDecision, type SealedJudgeInput, type SealedQrelsArtifact, type SealedQrelsHandle, type SealedRubric, type SearchResult, type SeededMemory, type SeededRng, type SequentialPhaseHooks, type SingleFlagAblationCell, type SingleFlagAblationId, type SpotCheckLogger, type StatisticalReport, type StructuredJudge, StructuredJudgeError, type StructuredJudgeErrorCode, type StructuredJudgeProvider, type StructuredJudgeTelemetry, type StructuredJudgeVerdict, type StructuredJudgeVerdictResult, type StructuredVerdictRequest, type SyntheticEdge, type SyntheticEmailIngestionAdapterOptions, type SyntheticFileIR, type SyntheticRepoConfig, type SyntheticSymbol, type SyntheticTargetOptions, type TaskResult, type TaskTokenUsage, type TemporalRetrievalCase, type ThirdPartyAdapterConfig, type TierDetail, type TimelineEntry, type TokenUsage, type WallMetric, type WriteBenchmarkArtifactResult, type ZepAdapterConfig, ZepMemCorrectAdapter, addContaminationEntry, aggregateTaskScores, answerBenchmarkQuestion, assertCanaryUnderFloor, assertIntegrityMetaPresent, assertPublishableIntegrity, assertSha256Hex, assistantMeetingPrepDefinition, assistantMorningBriefDefinition, assistantNextBestActionDefinition, assistantSynthesisDefinition, backlinkF1, binarizeJudgeScore, bootstrapCohensKappaConfidenceInterval, bootstrapMeanConfidenceInterval, buildAmaBenchDiagnosticMatrixArtifact, buildAmaBenchDiagnosticVariantSummary, buildAmaBenchLeaderboardRows, buildBaselineFromReport, buildBenchmarkArtifact, buildBenchmarkArtifactFilename, buildBenchmarkPublishFeed, buildBenchmarkReproManifest, buildBenchmarkRunSeeds, buildBuildWeekEvidenceReceipt, buildCodexCreditReceipt, buildJudgePayload, buildOracleTrajectoryRecall, buildProviderFreeLoCoMoRetrievalConfig, buildSchemaTierFixture, buildSchemaTierSmokeFixture, calculateCodexBudgetUnits, calendarFixture, canonicalJsonStringify, captureLoCoMoRetrievalTrace, captureMachineFingerprint, chatFixture, checkCodingGraphRegression, checkDatasetContamination, checkRegression, clampScore, cohensD, compareResults, computeCohensKappa, computeSealHash, containsAnswer, createSeededRng$1 as createAdamSeededRng, createAmaBenchDiagnosticAdapter, createAnthropicProvider, createCanaryAdapter, createClaudeCliProvider, createCodexCliProvider, createSeededRng as createCodingGraphSeededRng, createDeterministicSpotCheckLogger, createGatewayResponder, createLightweightAdapter, createLiteLlmProvider, createLocalLlmProvider, createMcpDemoMemCorrectAdapter, createMcpDemoMemoryAdapter, createMcpMemCorrectAdapter, createMcpMemoryAdapter, createMitigatedTarget, createOllamaProvider, createOpenAiCompatibleProvider, createOpenAiResponsesBenchJudge, createOpenAiResponsesProvider, createSeededRandom as createProceduralAblationSeededRandom, createProvider, createProviderBackedAmaBenchRecommendedJudge, createProviderBackedJudge, createProviderBackedResponder, createProviderBackedStructuredJudge, createRemnicAdapter, createResponderFromProvider, createSeededRng$2 as createSeededRng, createSpotCheckFileLogger, createStructuredBenchJudge, createStructuredJudgeFromProvider, createSyntheticEmailIngestionAdapter, createSyntheticTarget, createTimeoutGuardedAdapter, defaultBenchmarkBaselineDir, defaultBenchmarkPublishPath, deleteBenchmarkResults, diagnoseLoCoMoRetrievalTraceDelta, diagnoseLoComoProfileDelta, diagnoseLoComoRecallDelta, discoverAllProviders, discoveryEndpointFor, emailFixture, entityRecall, exactMatch, extractMetrics as extractCodingGraphMetrics, extractMarkdownSectionsByTitle, f1Score, fixtureToAblationScenarios, formatHandoffNote, formatMissingDatasetError, generateReport, generateSyntheticRepo, getAblationCell, getBenchmark, getBenchmarkLowerIsBetter, getGitSha, getMemoryEvalDimension, getProviderBackedJudgePromptIdentity, getRemnicVersion, hashBenchmarkArtifact, hashBytes, hashCanonicalJson, hashOrderedQuestionIds, hashString, integrityMetaIsComplete, interpretEffectSize, isAmaBenchUnknownLikeAnswer, isContaminationEntry, isContaminationManifest, isSealedQrelsArtifact, isSha256Hex, isStructuredJudgeProvider, judgeMemCorrectCorrectionAcceptance, judgeMemCorrectStaleMemoryHarm, linkMatches, listBenchmarkBaselines, listBenchmarkResults, listBenchmarks, listMemoryEvalBenchmarkIds, listMemoryEvalDimensions, llmJudgeScore, llmJudgeScoreDetailed, loadAblationFixture, loadBaseline, loadBeamDatasetPreview, loadBenchmarkArtifact, loadBenchmarkBaseline, loadBenchmarkReportCardProvenance, loadBenchmarkResult, loadCustomBenchmarkFile, loadJudgeCalibrationState, loadLoCoMo10, loadLocalLabManifest, loadLongMemEvalS, loadSealKeyFromEnv, loadSealedQrels, loadSealedRubric, matchEntity, mergeContaminationManifests, openSeal, orchestrateBenchmarkRuns, pairedDeltaConfidenceInterval, parseBenchmarkArtifact, parseCodexJsonlUsage, parseCustomBenchmark, parseLocalLabManifest, parseRubricResponse, parseSealedQrels, pickStableQualifiedName, precisionAtK, preflightLoCoMoRetrievalTraceCapture, preflightLocalLabRole, projectFolderFixture, recallAtK, reconcileCodexCreditLedger, redactBenchmarkResultSecrets, renderBaselineMarkdown, renderBenchmarkResultExport, renderLoComoProfileDeltaMarkdown, renderLoComoRecallDeltaMarkdown, renderMemorySummaryForJudge, renderMemoryViewForAgent, resolveAssistantAgent, resolveAssistantRubricId, resolveAssistantSeeds, resolveAssistantSpotCheckDir, resolveBenchRuntimeProfile, resolveBenchmarkPhaseTimeoutMs, resolveBenchmarkProgressLogging, resolveBenchmarkResultReference, resolveBenchmarkRunCount, resolveCodexCreditBudgetConfig, resolveLocalLabJudgeProviderConfig, resolveLocalLabProfile, resolveLocalLabRole, resolveStructuredJudge, rotateDistractors, rougeL, runAssistantBenchmark, runAssistantMeetingPrepBenchmark, runAssistantMorningBriefBenchmark, runAssistantNextBestActionBenchmark, runAssistantSynthesisBenchmark, runBaseline, runBenchSuite, runBenchmark, runCodingGraphBenchmark, runCustomBenchmarkFile, runExplain, runExtractionAttack, runJudgeCalibration, runMitigatedBaseline, runProceduralAblation, runProceduralAblationCli, runSealedJudge, runSequentialPhases, runWithinCodexCreditBudget, safeHexEqual, sanitizeLoComoResultReference, saveBaseline, saveBenchmarkBaseline, schemaCompleteness, sealPayload, selectAmaBenchDiagnosticVariants, selectCalibrationSlice, selectFixtureVariant, serializeBenchmarkArtifact, serializeBuildWeekEvidenceReceipt, serializeJsonl, serializeLoCoMoRetrievalTraceDelta, serializeLoCoMoRetrievalTraceReceipt, serializeSealedQrels, shuffleTasks, timed, verifyRubricDigest, writeBenchmarkArtifact, writeBenchmarkPublishFeed, writeBenchmarkReproManifest, writeBenchmarkResult, writeBuildWeekEvidenceReceipt, writeJudgeCalibrationState, writeLeaderboardArtifactsForResult, zeroScores };

package/dist/index.js

@@ -9777,7 +9777,7 @@ async function runWithBenchmarkPhaseTimeout(label, timeoutMs, fn, options = {})
   }
 }
 function wrapResponder(responder, run) {
-  return {
+  const wrapped = {
     respond(question, recalledText, control) {
       return run("respond", async (signal) => {
         const merged = mergeBenchPhaseControl(signal, control);
@@ -9789,6 +9789,11 @@ function wrapResponder(responder, run) {
       });
     }
   };
+  if (typeof responder.identity === "function") {
+    const inner = responder.identity.bind(responder);
+    wrapped.identity = () => inner();
+  }
+  return wrapped;
 }
 function wrapJudge(judge, run) {
   const wrapped = {
@@ -15030,6 +15035,24 @@ function getProviderBackedJudgePromptIdentity(config) {
   };
   return `sha256:${createHash9("sha256").update(JSON.stringify(contract)).digest("hex")}`;
 }
+function getProviderBackedResponderIdentity(config) {
+  const sanitized = {};
+  for (const [key, value] of Object.entries(config)) {
+    if (key === "apiKey" || key === "authToken" || key === "bearerToken") continue;
+    sanitized[key] = value;
+  }
+  return `responder:sha256:${createHash9("sha256").update(stableStringifyForIdentity(sanitized)).digest("hex")}`;
+}
+function stableStringifyForIdentity(value) {
+  if (value === null || typeof value !== "object") {
+    return JSON.stringify(value);
+  }
+  if (Array.isArray(value)) {
+    return `[${value.map((entry) => stableStringifyForIdentity(entry)).join(",")}]`;
+  }
+  const keys = Object.keys(value).sort();
+  return `{${keys.map((key) => `${JSON.stringify(key)}:${stableStringifyForIdentity(value[key])}`).join(",")}}`;
+}
 var AMA_BENCH_RECOMMENDED_JUDGE_SYSTEM_PROMPT = [
   "You are evaluating an AMA-Bench long-horizon memory question.",
   "Decide whether the predicted answer correctly answers the question using the reference answer as ground truth.",
@@ -15043,8 +15066,8 @@ var CONTEXT_COMPACTION_MARKER = "[...omitted unrelated recalled context...]";
 var COMPACTED_CONTEXT_PREFIX = "[Remnic memory context compacted for the responder prompt; full recalled text is preserved in the benchmark artifact.]";
 var TRAJECTORY_ANALYSIS_HEADING = "## Trajectory analysis";
 var TRAJECTORY_LABELS = Object.freeze(["action", "observation", "step", "turn"]);
-function createResponderFromProvider(provider, options = {}) {
-  return {
+function createResponderFromProvider(provider, options = {}, responderIdentity) {
+  const responder = {
     async respond(question, recalledText, control) {
       const responderQuestion = options.promptBudgetChars === void 0 ? question : compactResponderQuestion(question, options.promptBudgetChars);
       const responderContext = options.contextBudgetChars === void 0 ? recalledText : compactResponderContext(recalledText, question, options.contextBudgetChars);
@@ -15072,13 +15095,21 @@ function createResponderFromProvider(provider, options = {}) {
       };
     }
   };
+  if (responderIdentity && responderIdentity.trim().length > 0) {
+    responder.identity = () => responderIdentity;
+  }
+  return responder;
 }
 function createProviderBackedResponder(config, providerInstance) {
   validateProviderConfig(config, "responder");
-  return createResponderFromProvider(providerInstance ?? createProvider(config), {
-    contextBudgetChars: config.responderContextBudgetChars,
-    promptBudgetChars: config.responderPromptBudgetChars
-  });
+  return createResponderFromProvider(
+    providerInstance ?? createProvider(config),
+    {
+      contextBudgetChars: config.responderContextBudgetChars,
+      promptBudgetChars: config.responderPromptBudgetChars
+    },
+    getProviderBackedResponderIdentity(config)
+  );
 }
 function compactResponderQuestion(question, maxChars) {
   if (!Number.isInteger(maxChars) || maxChars <= 0) {
@@ -15576,7 +15607,11 @@ function createGatewayResponder(options) {
     ...options.workspaceDir ? { workspaceDir: options.workspaceDir } : {}
   };
   const llm = options.llmFactory?.(options.gatewayConfig, runtimeContext) ?? new FallbackLlmClient(options.gatewayConfig, runtimeContext);
-  return {
+  const responderIdentity = getGatewayResponderIdentity(
+    options.gatewayConfig,
+    options.agentId
+  );
+  const responder = {
     async respond(question, recalledText, control) {
       const startedAt = performance.now();
       const response = await llm.chatCompletion(
@@ -15614,6 +15649,25 @@ function createGatewayResponder(options) {
       };
     }
   };
+  if (responderIdentity) {
+    responder.identity = () => responderIdentity;
+  }
+  return responder;
+}
+function getGatewayResponderIdentity(config, agentId) {
+  const sanitized = sanitizeGatewayConfigForIdentity(config);
+  return `gateway:sha256:${createHash9("sha256").update(stableStringifyForIdentity({ agentId: agentId ?? null, config: sanitized })).digest("hex")}`;
+}
+function sanitizeGatewayConfigForIdentity(value) {
+  if (value === null || typeof value !== "object") return value;
+  if (Array.isArray(value)) return value.map(sanitizeGatewayConfigForIdentity);
+  const record = value;
+  const out = {};
+  for (const [key, child] of Object.entries(record)) {
+    if (key === "apiKey" || key === "headers" || key === "authHeader") continue;
+    out[key] = sanitizeGatewayConfigForIdentity(child);
+  }
+  return out;
 }
 function validateProviderConfig(config, kind) {
   if (typeof config.model !== "string" || config.model.trim().length === 0) {
@@ -17365,7 +17419,7 @@ async function resolveLocalLabRuntimeProfile(options) {
 // src/benchmark.ts
 import fs2 from "fs";
 import path35 from "path";
-import { createHash as createHash15 } from "crypto";
+import { createHash as createHash16 } from "crypto";
 import { expandTildePath as expandTildePath3 } from "@remnic/core";
 
 // src/judges/judge-cache.ts
@@ -21503,7 +21557,7 @@ function formatMissingDatasetError(benchmark, datasetDir, filenames, errors) {
 }
 
 // src/benchmarks/published/harness.ts
-import { randomUUID as randomUUID5 } from "crypto";
+import { createHash as createHash11, randomUUID as randomUUID5 } from "crypto";
 
 // src/benchmarks/published/category-aggregates.ts
 function computeCategoryAggregates(tasks) {
@@ -21534,34 +21588,45 @@ async function runPublishedHarness(ctx) {
   validateContext(ctx);
   const executionProvenance = captureBenchmarkExecutionProvenance();
   const answerSupportGate = resolveAnswerSupportGate(ctx.options);
-  const trialConcurrency = resolveTrialConcurrency(
-    ctx.options.benchmarkOptions?.trialConcurrency
-  );
+  const trialConcurrency = resolveTrialConcurrency(ctx.options.benchmarkOptions?.trialConcurrency);
   const tasks = [];
-  for await (const plan of toAsyncIterable(ctx.plans)) {
-    await ctx.options.system.reset();
-    for (const session of plan.ingestSessions) {
-      if (session.messages.length > 0) {
-        await ctx.options.system.store(session.sessionId, session.messages);
+  const pendingPairedAnswerReplays = /* @__PURE__ */ new Map();
+  try {
+    for await (const plan of toAsyncIterable(ctx.plans)) {
+      await ctx.options.system.reset();
+      for (const session of plan.ingestSessions) {
+        if (session.messages.length > 0) {
+          await ctx.options.system.store(session.sessionId, session.messages);
+        }
+      }
+      try {
+        await ctx.options.system.drain?.();
+      } catch (drainErr) {
+        throw new Error(
+          `PublishedBenchmarkHarness: drain failed before scoring; public benchmark evidence would be incomplete: ${drainErr instanceof Error ? drainErr.message : String(drainErr)}`,
+          { cause: drainErr }
+        );
       }
+      const planIndex = tasks.length;
+      await executePlanTrials(ctx, plan.trials, {
+        planIndex,
+        tasks,
+        trialConcurrency,
+        answerSupportGate,
+        pendingPairedAnswerReplays
+      });
     }
-    try {
-      await ctx.options.system.drain?.();
-    } catch (drainErr) {
-      throw new Error(
-        `PublishedBenchmarkHarness: drain failed before scoring; public benchmark evidence would be incomplete: ${drainErr instanceof Error ? drainErr.message : String(drainErr)}`,
-        { cause: drainErr }
-      );
+  } catch (error) {
+    if (ctx.options.runtimeProfile === "baseline") {
+      ctx.options.pairedAnswerReplayCache?.clear();
     }
-    const planIndex = tasks.length;
-    await executePlanTrials(ctx, plan.trials, {
-      planIndex,
-      tasks,
-      trialConcurrency,
-      answerSupportGate
-    });
+    throw error;
+  }
+  const result = await buildBenchmarkResult(ctx, tasks, executionProvenance);
+  if (ctx.options.runtimeProfile === "baseline" && result.meta.status === "partial") {
+    ctx.options.pairedAnswerReplayCache?.clear();
   }
-  return buildBenchmarkResult(ctx, tasks, executionProvenance);
+  return result;
 }
 async function executePlanTrials(ctx, trials, options) {
   if (options.trialConcurrency === 1 || trials.length <= 1) {
@@ -21569,28 +21634,28 @@ async function executePlanTrials(ctx, trials, options) {
       appendCompletedTask(
         ctx,
         options.tasks,
+        options.pendingPairedAnswerReplays,
         await executeTrialWithFailure(
           ctx,
           trial,
           options.planIndex,
-          options.answerSupportGate
+          options.answerSupportGate,
+          options.pendingPairedAnswerReplays
         )
       );
     }
     return;
   }
   for (let batchStart = 0; batchStart < trials.length; batchStart += options.trialConcurrency) {
-    const batch = trials.slice(
-      batchStart,
-      batchStart + options.trialConcurrency
-    );
+    const batch = trials.slice(batchStart, batchStart + options.trialConcurrency);
     const settled = await Promise.allSettled(
       batch.map(
         (trial) => executeTrialWithFailure(
           ctx,
           trial,
           options.planIndex,
-          options.answerSupportGate
+          options.answerSupportGate,
+          options.pendingPairedAnswerReplays
         )
       )
     );
@@ -21611,7 +21676,7 @@ async function executePlanTrials(ctx, trials, options) {
           `PublishedBenchmarkHarness: concurrent trial ${batchStart + offset} did not settle before canonical emission.`
         );
       }
-      appendCompletedTask(ctx, options.tasks, result.value);
+      appendCompletedTask(ctx, options.tasks, options.pendingPairedAnswerReplays, result.value);
     }
     if (terminalOffset >= 0) {
       const terminalResult = settled[terminalOffset];
@@ -21625,14 +21690,18 @@ async function executePlanTrials(ctx, trials, options) {
     }
   }
 }
-function appendCompletedTask(ctx, tasks, task) {
+function appendCompletedTask(ctx, tasks, pendingPairedAnswerReplays, task) {
+  const pendingReplay = pendingPairedAnswerReplays.get(task);
+  if (pendingReplay) {
+    ctx.options.pairedAnswerReplayCache?.set(pendingReplay.key, pendingReplay.entry);
+  }
   tasks.push(task);
   ctx.options.onTaskComplete?.(task, tasks.length, ctx.totalCount);
 }
-async function executeTrialWithFailure(ctx, trial, planIndex, answerSupportGate) {
+async function executeTrialWithFailure(ctx, trial, planIndex, answerSupportGate, pendingPairedAnswerReplays) {
   const trialId = trial.taskId ?? trial.question.slice(0, 60);
   try {
-    return await executeTrial(ctx, trial, answerSupportGate);
+    return await executeTrial(ctx, trial, answerSupportGate, pendingPairedAnswerReplays);
   } catch (err) {
     const blocked = findBenchmarkRunBlockedError(err);
     if (blocked) {
@@ -21678,13 +21747,7 @@ function validateContext(ctx) {
       "PublishedBenchmarkHarness requires metricsSpec.metrics: one of f1, contains_answer, rouge_l, llm_judge, judge_accuracy."
     );
   }
-  const allowed = [
-    "f1",
-    "contains_answer",
-    "rouge_l",
-    "llm_judge",
-    "judge_accuracy"
-  ];
+  const allowed = ["f1", "contains_answer", "rouge_l", "llm_judge", "judge_accuracy"];
   for (const metric of ctx.metricsSpec.metrics) {
     if (!allowed.includes(metric)) {
       throw new Error(
@@ -21713,9 +21776,7 @@ function resolveTrialConcurrency(raw) {
   }
   const parsed = typeof raw === "number" ? raw : Number(raw);
   if (!Number.isInteger(parsed) || parsed <= 0 || parsed > 64) {
-    throw new Error(
-      "PublishedBenchmarkHarness: benchmarkOptions.trialConcurrency must be an integer from 1 to 64."
-    );
+    throw new Error("PublishedBenchmarkHarness: benchmarkOptions.trialConcurrency must be an integer from 1 to 64.");
   }
   return parsed;
 }
@@ -21740,15 +21801,77 @@ function resolveAnswerSupportGate(options) {
     "PublishedBenchmarkHarness: answerSupportGate must be a boolean or one of true/false, 1/0, yes/no, on/off."
   );
 }
-async function executeTrial(ctx, trial, answerSupportGate) {
+function stableStringify3(value) {
+  if (value === null || typeof value !== "object") {
+    return JSON.stringify(value);
+  }
+  if (Array.isArray(value)) {
+    return `[${value.map((entry) => stableStringify3(entry)).join(",")}]`;
+  }
+  const keys = Object.keys(value).sort();
+  return `{${keys.map((key) => `${JSON.stringify(key)}:${stableStringify3(value[key])}`).join(",")}}`;
+}
+function pairedAnswerReplayKey(trial, recalledText, recallSupport, systemProvider, responderIdentity) {
+  return createHash11("sha256").update(
+    stableStringify3({
+      responder: {
+        baseUrl: systemProvider?.baseUrl ?? null,
+        disableThinking: systemProvider?.disableThinking ?? null,
+        model: systemProvider?.model ?? null,
+        provider: systemProvider?.provider ?? null,
+        providerRequestTimeoutMs: systemProvider?.providerRequestTimeoutMs ?? null,
+        reasoningEffort: systemProvider?.reasoningEffort ?? null,
+        responderContextBudgetChars: systemProvider?.responderContextBudgetChars ?? null,
+        responderPromptBudgetChars: systemProvider?.responderPromptBudgetChars ?? null,
+        retryOptions: systemProvider?.retryOptions ?? null,
+        seed: systemProvider?.seed ?? null,
+        temperature: systemProvider?.temperature ?? null
+      },
+      responderIdentity,
+      responderPrompt: buildStrictBenchmarkQuestion(
+        trial.question,
+        trial.answerFormat ?? "auto"
+      ),
+      answerMode: "strict",
+      question: trial.question,
+      recalledText,
+      recallSupport: recallSupport ? {
+        evidenceCount: recallSupport.evidenceCount ?? null,
+        maxScore: recallSupport.maxScore ?? null,
+        reason: recallSupport.reason ?? null,
+        status: recallSupport.status,
+        supportThreshold: recallSupport.supportThreshold ?? null
+      } : null,
+      taskId: trial.taskId
+    })
+  ).digest("hex");
+}
+function pairedAnswerReplayEntry(sourceRuntimeProfile, answer) {
+  return {
+    sourceRuntimeProfile,
+    finalAnswer: answer.finalAnswer,
+    answeredText: answer.answeredText,
+    ...answer.model === void 0 ? {} : { model: answer.model }
+  };
+}
+function resolveResponderIdentity(responder) {
+  if (!responder || typeof responder.identity !== "function") {
+    return null;
+  }
+  let raw;
+  try {
+    raw = responder.identity();
+  } catch {
+    return null;
+  }
+  const trimmed = typeof raw === "string" ? raw.trim() : "";
+  return trimmed.length > 0 ? trimmed : null;
+}
+async function executeTrial(ctx, trial, answerSupportGate, pendingPairedAnswerReplays) {
   const { result: recallResult, durationMs } = await timed(async () => {
-    const recallBudget = benchmarkRecallBudgetForSessionCount(
-      trial.recallSessionIds.length
-    );
+    const recallBudget = benchmarkRecallBudgetForSessionCount(trial.recallSessionIds.length);
     const recalledSessions = await Promise.all(
-      trial.recallSessionIds.map(
-        (sessionId) => ctx.options.system.recall(sessionId, trial.question, recallBudget)
-      )
+      trial.recallSessionIds.map((sessionId) => ctx.options.system.recall(sessionId, trial.question, recallBudget))
     );
     const rawRecalledText = recalledSessions.filter(Boolean).join("\n\n");
     const recalledText2 = trial.recallTextTransform ? trial.recallTextTransform({
@@ -21759,17 +21882,39 @@ async function executeTrial(ctx, trial, answerSupportGate) {
     return { recalledText: recalledText2, recallSupport: recallSupport2 };
   });
   const { recalledText, recallSupport } = recallResult;
-  let answered = await answerBenchmarkQuestion({
-    question: trial.question,
+  const responderIdentity = resolveResponderIdentity(ctx.options.system.responder);
+  const answerReplayKey = ctx.options.pairedAnswerReplayCache && responderIdentity !== null ? pairedAnswerReplayKey(
+    trial,
     recalledText,
-    responder: ctx.options.system.responder,
-    answerMode: "strict",
-    answerFormat: trial.answerFormat,
-    recallSupport
-  }).catch(
-    (error) => answerWithTrialFallback(trial, recalledText, error)
-  );
-  answered = refineTrialAnswer(trial, recalledText, answered);
+    recallSupport,
+    ctx.options.systemProvider,
+    responderIdentity
+  ) : void 0;
+  const cachedAnswer = answerReplayKey ? ctx.options.pairedAnswerReplayCache?.get(answerReplayKey) : void 0;
+  const currentProfile = ctx.options.runtimeProfile ?? null;
+  const pairedAnswerReusedFrom = cachedAnswer?.sourceRuntimeProfile === "baseline" && currentProfile === "real" ? "baseline" : void 0;
+  let answered;
+  if (pairedAnswerReusedFrom) {
+    const reusedAnswer = cachedAnswer;
+    answered = {
+      finalAnswer: reusedAnswer.finalAnswer,
+      recalledText,
+      answeredText: reusedAnswer.answeredText,
+      latencyMs: 0,
+      tokens: { input: 0, output: 0 },
+      model: reusedAnswer.model
+    };
+  } else {
+    answered = await answerBenchmarkQuestion({
+      question: trial.question,
+      recalledText,
+      responder: ctx.options.system.responder,
+      answerMode: "strict",
+      answerFormat: trial.answerFormat,
+      recallSupport
+    }).catch((error) => answerWithTrialFallback(trial, recalledText, error));
+    answered = refineTrialAnswer(trial, recalledText, answered);
+  }
   const hookResult = trial.postAnswerHook ? await trial.postAnswerHook({
     question: trial.question,
     recalledText,
@@ -21789,10 +21934,7 @@ async function executeTrial(ctx, trial, answerSupportGate) {
         scores.f1 = f1Score(answered.finalAnswer, trial.expected);
         break;
       case "contains_answer":
-        scores.contains_answer = containsAnswer(
-          answered.finalAnswer,
-          trial.expected
-        );
+        scores.contains_answer = containsAnswer(answered.finalAnswer, trial.expected);
         break;
       case "rouge_l":
         scores.rouge_l = rougeL(answered.finalAnswer, trial.expected);
@@ -21811,9 +21953,7 @@ async function executeTrial(ctx, trial, answerSupportGate) {
         break;
       default: {
         const exhaustive = metric;
-        throw new Error(
-          `PublishedBenchmarkHarness: metric ${String(exhaustive)} not handled.`
-        );
+        throw new Error(`PublishedBenchmarkHarness: metric ${String(exhaustive)} not handled.`);
       }
     }
   }
@@ -21834,6 +21974,7 @@ async function executeTrial(ctx, trial, answerSupportGate) {
     answeredText: answered.finalAnswer,
     ...trial.answerFormat ? { answerFormat: trial.answerFormat } : {},
     ...answerSupportGate ? { answerSupportGate: true, recallSupport } : {},
+    ...pairedAnswerReusedFrom ? { pairedAnswerReusedFrom } : {},
     responderModel: answered.model,
     judgeModel: judgeResult.model,
     ...answered.fallbackReason ? { answerFallbackReason: answered.fallbackReason } : {},
@@ -21849,7 +21990,7 @@ async function executeTrial(ctx, trial, answerSupportGate) {
   if (hookResult.extraDetails) {
     Object.assign(details, hookResult.extraDetails);
   }
-  return {
+  const task = {
     taskId: trial.taskId,
     question: trial.question,
     expected: trial.expected,
@@ -21862,6 +22003,13 @@ async function executeTrial(ctx, trial, answerSupportGate) {
     },
     details
   };
+  if (answerReplayKey && currentProfile === "baseline" && answered.fallbackReason === void 0) {
+    pendingPairedAnswerReplays.set(task, {
+      key: answerReplayKey,
+      entry: pairedAnswerReplayEntry(currentProfile, answered)
+    });
+  }
+  return task;
 }
 async function assessRecallSupport(ctx, trial, recalledText) {
   if (recalledText.trim().length === 0) {
@@ -21898,13 +22046,7 @@ async function assessRecallSupport(ctx, trial, recalledText) {
   }
 }
 function validateRecallSupportAssessment(assessment) {
-  const allowed = [
-    "supported",
-    "weak",
-    "empty",
-    "unavailable",
-    "backend_failure"
-  ];
+  const allowed = ["supported", "weak", "empty", "unavailable", "backend_failure"];
   if (!assessment || !allowed.includes(assessment.status)) {
     throw new Error("adapter returned an invalid recall support status");
   }
@@ -21919,21 +22061,11 @@ function validateRecallSupportAssessment(assessment) {
 }
 async function scoreTrialJudge(ctx, trial, answeredText) {
   if (!trial.binaryJudgePrompt) {
-    return llmJudgeScoreDetailed(
-      ctx.options.system.judge,
-      trial.question,
-      answeredText,
-      trial.expected
-    );
+    return llmJudgeScoreDetailed(ctx.options.system.judge, trial.question, answeredText, trial.expected);
   }
   const judge = ctx.options.system.judge;
   if (!judge?.scoreBinaryPrompt) {
-    return llmJudgeScoreDetailed(
-      judge,
-      trial.question,
-      answeredText,
-      trial.expected
-    );
+    return llmJudgeScoreDetailed(judge, trial.question, answeredText, trial.expected);
   }
   const prompt = trial.binaryJudgePrompt({
     question: trial.question,
@@ -21941,21 +22073,15 @@ async function scoreTrialJudge(ctx, trial, answeredText) {
     answeredText
   });
   if (typeof prompt !== "string" || prompt.trim().length === 0) {
-    throw new Error(
-      "PublishedBenchmarkHarness: binaryJudgePrompt returned an empty prompt."
-    );
+    throw new Error("PublishedBenchmarkHarness: binaryJudgePrompt returned an empty prompt.");
   }
   const binaryJudge = {
     scoreBinaryPrompt: judge.scoreBinaryPrompt.bind(judge)
   };
-  return llmBinaryJudgeScoreDetailed(
-    binaryJudge,
-    prompt,
-    {
-      predicted: answeredText,
-      expected: trial.expected
-    }
-  );
+  return llmBinaryJudgeScoreDetailed(binaryJudge, prompt, {
+    predicted: answeredText,
+    expected: trial.expected
+  });
 }
 function answerWithTrialFallback(trial, recalledText, error) {
   if (isBenchmarkRunBlockedError(error)) {
@@ -22000,14 +22126,8 @@ function refineTrialAnswer(trial, recalledText, answered) {
 async function buildBenchmarkResult(ctx, tasks, executionProvenance) {
   const remnicVersion = await getRemnicVersion();
   const totalLatencyMs = tasks.reduce((sum, task) => sum + task.latencyMs, 0);
-  const totalInputTokens = tasks.reduce(
-    (sum, task) => sum + task.tokens.input,
-    0
-  );
-  const totalOutputTokens = tasks.reduce(
-    (sum, task) => sum + task.tokens.output,
-    0
-  );
+  const totalInputTokens = tasks.reduce((sum, task) => sum + task.tokens.input, 0);
+  const totalOutputTokens = tasks.reduce((sum, task) => sum + task.tokens.output, 0);
   const mode = ctx.options.mode;
   const failedTasks = tasks.flatMap((task) => {
     const marker = task.details?.benchmarkFailure;
@@ -22015,10 +22135,12 @@ async function buildBenchmarkResult(ctx, tasks, executionProvenance) {
       return [];
     }
     const message = marker.message;
-    return [{
-      taskId: task.taskId,
-      message: typeof message === "string" ? message : "unknown trial failure"
-    }];
+    return [
+      {
+        taskId: task.taskId,
+        message: typeof message === "string" ? message : "unknown trial failure"
+      }
+    ];
   });
   const failureReason = failedTasks.length > 0 ? `trial_execution_failure: ${failedTasks.length}/${tasks.length} scored trial(s) failed (${failedTasks.slice(0, 3).map((failure) => `${failure.taskId}: ${failure.message.slice(0, 240)}`).join("; ")}${failedTasks.length > 3 ? `; and ${failedTasks.length - 3} more` : ""})` : void 0;
   const categoryAggregates = computeCategoryAggregates(tasks);
@@ -25283,7 +25405,7 @@ var StructuredLiteralParser = class {
 };
 
 // src/benchmarks/published/personamem/runner.ts
-import { createHash as createHash11, randomUUID as randomUUID7 } from "crypto";
+import { createHash as createHash12, randomUUID as randomUUID7 } from "crypto";
 import { readFile as readFile17, realpath as realpath5 } from "fs/promises";
 import path19 from "path";
 
@@ -25888,7 +26010,7 @@ function buildMcqPrompt(sample, seed) {
 function deterministicShuffle(values, seedMaterial) {
   return values.map((value, index) => ({
     value,
-    key: createHash11("sha256").update(`${seedMaterial}:${index}:${value}`).digest("hex"),
+    key: createHash12("sha256").update(`${seedMaterial}:${index}:${value}`).digest("hex"),
     index
   })).sort((left, right) => {
     const byKey = left.key.localeCompare(right.key);
@@ -34354,7 +34476,7 @@ function pairedDeltaConfidenceInterval(candidateValues, baselineValues, options
 }
 
 // src/judges/sealed-rubric.ts
-import { createHash as createHash12 } from "crypto";
+import { createHash as createHash13 } from "crypto";
 import { appendFileSync, mkdirSync } from "fs";
 import path31 from "path";
 
@@ -34463,7 +34585,7 @@ function loadSealedRubric(id = DEFAULT_ASSISTANT_RUBRIC_ID, options = {}) {
   if (typeof prompt !== "string" || prompt.length === 0) {
     throw new Error(`sealed rubric not found in registry: ${id}`);
   }
-  const sha2564 = createHash12("sha256").update(prompt, "utf8").digest("hex");
+  const sha2564 = createHash13("sha256").update(prompt, "utf8").digest("hex");
   const version = parseVersionFromId(id);
   return { id, version, prompt, sha256: sha2564 };
 }
@@ -36666,7 +36788,7 @@ async function runRetentionAgedDatasetBenchmark(options) {
 import { randomUUID as randomUUID31 } from "crypto";
 
 // src/benchmarks/remnic/memcorrect/generator.ts
-import { createHash as createHash13 } from "crypto";
+import { createHash as createHash14 } from "crypto";
 
 // src/benchmarks/remnic/memcorrect/token-pools.ts
 var PERSONAS = [
@@ -36990,7 +37112,7 @@ function corpusHash(corpus) {
     uptakeLatencyCap: corpus.options.uptakeLatencyCap,
     scenarios: corpus.scenarios
   });
-  return createHash13("sha256").update(canonical).digest("hex");
+  return createHash14("sha256").update(canonical).digest("hex");
 }
 
 // src/benchmarks/remnic/memcorrect/schema.ts
@@ -37947,7 +38069,7 @@ import { mkdir as mkdir17, writeFile as writeFile16 } from "fs/promises";
 import path34 from "path";
 
 // src/benchmarks/remnic/bounded-memory-contracts/fixture.ts
-import { createHash as createHash14 } from "crypto";
+import { createHash as createHash15 } from "crypto";
 var SCOPE_ACME = "project:acme";
 var SCOPE_BETA = "project:beta";
 var SCOPE_ALICE = "user:alice";
@@ -38430,7 +38552,7 @@ var BOUNDED_MEMORY_SMOKE_FIXTURE = [
 function fixtureHash(tasks) {
   const source = tasks ?? BOUNDED_MEMORY_FIXTURE;
   const payload = JSON.stringify(source);
-  return createHash14("sha256").update(payload, "utf8").digest("hex");
+  return createHash15("sha256").update(payload, "utf8").digest("hex");
 }
 
 // src/benchmarks/remnic/bounded-memory-contracts/agent.ts
@@ -39690,13 +39812,13 @@ function wrapJudgeWithCache(args) {
       // differentiator is part of the prompt hash. Bumping
       // JUDGE_CACHE_PROTOCOL_VERSION invalidates verdicts when judge
       // prompt/parse semantics change (PR #1591, High).
-      judgePromptHash: createHash15("sha256").update(JUDGE_CACHE_PROTOCOL_VERSION).update("
").update(args.amaBenchJudgeProtocol).update("
").update(args.role).digest("hex"),
+      judgePromptHash: createHash16("sha256").update(JUDGE_CACHE_PROTOCOL_VERSION).update("
").update(args.amaBenchJudgeProtocol).update("
").update(args.role).digest("hex"),
       judgeModelId: args.provider?.model !== void 0 && args.provider.model.length > 0 ? `${args.provider.model}${crossJudgeIdSuffix}` : `unknown-${args.role}-judge`,
       // Full judge configuration, deterministically serialized (sorted
       // keys) so provider/base-url/retry changes produce fresh cache
       // keys. `role` is included so primary and cross judges never
       // share a paramsHash.
-      judgeParamsHash: createHash15("sha256").update(
+      judgeParamsHash: createHash16("sha256").update(
         stableStringify2({
           role: args.role,
           provider: args.provider
@@ -40402,7 +40524,7 @@ function formatSignedScore(value) {
 }
 
 // src/stats/locomo-recall-delta.ts
-import { createHash as createHash16 } from "crypto";
+import { createHash as createHash17 } from "crypto";
 import { basename as basename2 } from "path";
 var LOCOMO_FULL_TASK_COUNT = 1986;
 var LOCOMO_RECALL_EXCERPT_CHARS = 240;
@@ -40914,7 +41036,7 @@ function normalizeText3(value) {
   return value.replaceAll("\r\n", "\n").replaceAll("\r", "\n");
 }
 function sha2563(value) {
-  return createHash16("sha256").update(value).digest("hex");
+  return createHash17("sha256").update(value).digest("hex");
 }
 function stableJson(value) {
   return JSON.stringify(value);
@@ -43430,7 +43552,7 @@ var chatFixture = {
 };
 
 // src/judges/calibration-slice.ts
-import { createHash as createHash17, randomBytes as randomBytes3 } from "crypto";
+import { createHash as createHash18, randomBytes as randomBytes3 } from "crypto";
 import { chmod as chmod2, lstat as lstat5, mkdir as mkdir18, open as open3, readFile as readFile23, rename as rename4, unlink as unlink5, writeFile as writeFile17 } from "fs/promises";
 import path37 from "path";
 
@@ -43579,7 +43701,7 @@ function selectCalibrationSlice(questionIds, size = CALIBRATION_SLICE_SIZE) {
       unique.push(id);
     }
   }
-  return unique.map((id) => ({ id, digest: createHash17("sha256").update(id, "utf8").digest("hex") })).sort((a, b) => a.digest < b.digest ? -1 : a.digest > b.digest ? 1 : 0).slice(0, Math.min(size, unique.length)).map((entry) => entry.id);
+  return unique.map((id) => ({ id, digest: createHash18("sha256").update(id, "utf8").digest("hex") })).sort((a, b) => a.digest < b.digest ? -1 : a.digest > b.digest ? 1 : 0).slice(0, Math.min(size, unique.length)).map((entry) => entry.id);
 }
 async function runJudgeCalibration(options) {
   const binScore = options.binScore ?? ((score) => binarizeJudgeScore(score));
@@ -43696,7 +43818,7 @@ function hashOrderedQuestionIds(questionIds) {
   if (questionIds.some((id) => typeof id !== "string" || id.length === 0)) {
     throw new Error("hashOrderedQuestionIds: question ids must be non-empty strings.");
   }
-  return createHash17("sha256").update(JSON.stringify(questionIds), "utf8").digest("hex");
+  return createHash18("sha256").update(JSON.stringify(questionIds), "utf8").digest("hex");
 }
 function validatePinnedQuestionIds(ids, availableIds) {
   if (ids.length === 0 || ids.length > CALIBRATION_SLICE_SIZE || ids.some((id) => typeof id !== "string" || id.length === 0) || new Set(ids).size !== ids.length) {
@@ -43709,7 +43831,7 @@ function validatePinnedQuestionIds(ids, availableIds) {
   return [...ids];
 }
 function hashCalibrationAnswerSet(answers) {
-  return createHash17("sha256").update(JSON.stringify(answers.map((answer) => [
+  return createHash18("sha256").update(JSON.stringify(answers.map((answer) => [
     answer.questionId,
     answer.question,
     answer.predicted,
@@ -43775,7 +43897,7 @@ async function loadOrInitializeCheckpoint(benchmarkId, provenance, sliceQuestion
       frontierJudgeConfigHash: provenance.frontierJudgeConfigHash,
       binningIdentity: provenance.binningIdentity
     };
-    const contractHash = createHash17("sha256").update(stableJson2(contract)).digest("hex");
+    const contractHash = createHash18("sha256").update(stableJson2(contract)).digest("hex");
     let raw;
     try {
       const info = await lstat5(checkpointPath);
@@ -45303,7 +45425,7 @@ function createMitigatedTarget(config) {
 }
 
 // src/coding-graph/generator.ts
-import { createHash as createHash18 } from "crypto";
+import { createHash as createHash19 } from "crypto";
 function createSeededRng3(seed) {
   let state = seed >>> 0;
   return function rng() {
@@ -45332,7 +45454,7 @@ var EDGE_TYPE_WEIGHTS = [
 var PROVENANCE_VALUES = ["heuristic", "heuristic", "heuristic", "trace"];
 var AVG_BYTES_PER_LINE = 40;
 function hashContent(input) {
-  return createHash18("sha256").update(input).digest("hex").slice(0, 16);
+  return createHash19("sha256").update(input).digest("hex").slice(0, 16);
 }
 function generateSyntheticRepo(config) {
   const rng = createSeededRng3(config.seed);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/bench",
-  "version": "9.21.0",
+  "version": "9.22.1",
   "description": "Retrieval latency ladder benchmarks + CI regression gates for @remnic/core",
   "type": "module",
   "main": "./dist/index.js",
@@ -40,8 +40,8 @@
     "hyparquet": "^1.25.7",
     "yaml": "^2.4.2",
     "zod": "^3.24.0",
-    "@remnic/coding-graph": "^9.21.0",
-    "@remnic/core": "^9.21.0"
+    "@remnic/coding-graph": "^9.22.1",
+    "@remnic/core": "^9.22.1"
   },
   "devDependencies": {
     "tsup": "^8.5.1",