@tangle-network/agent-eval 0.106.3 → 0.107.0

package/dist/{chunk-J22CKVXN.js → chunk-V75NN2ZR.js}

@@ -6,7 +6,7 @@ import {
   pairHoldout,
   runCanaries,
   scoreRedTeamOutput
-} from "./chunk-3E5KUXYZ.js";
+} from "./chunk-G4DLZAV5.js";
 import {
   runCampaign
 } from "./chunk-3PFZBGMR.js";
@@ -418,4 +418,4 @@ export {
   runEval,
   evolutionaryProposer
 };
-//# sourceMappingURL=chunk-J22CKVXN.js.map
+//# sourceMappingURL=chunk-V75NN2ZR.js.map

package/dist/contract/index.d.ts

@@ -1,9 +1,9 @@
-import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, h as GateDecision } from '../types-Ctf7XIAL.js';
-export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, d as CampaignTraceWriter, m as CodeSurface, D as Dispatch, n as GateContext, G as GateResult, o as GenerationCandidate, e as GenerationRecord, c as JudgeDimension, J as JudgeScore, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-Ctf7XIAL.js';
-import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-Bsyjc67Z.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-Bsyjc67Z.js';
-import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../gepa-DQ3ruj18.js';
-export { G as GepaProposerOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../gepa-DQ3ruj18.js';
+import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, h as GateDecision } from '../types-CNZ0tHET.js';
+export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, d as CampaignTraceWriter, m as CodeSurface, D as Dispatch, n as GateContext, G as GateResult, o as GenerationCandidate, e as GenerationRecord, c as JudgeDimension, J as JudgeScore, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-CNZ0tHET.js';
+import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-CIjRcI6m.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-CIjRcI6m.js';
+import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../gepa-BmqTrdtg.js';
+export { G as GepaProposerOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../gepa-BmqTrdtg.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
 import { HostedTenant, EvalRunCellScore, EvalRunGenerationSnapshot, EvalRunEvent, TraceSpanEvent } from '../hosted/index.js';
 import { R as RunRecord, b as RunSplitTag } from '../run-record-I-Z3JNvO.js';

package/dist/contract/index.js

@@ -19,7 +19,7 @@ import {
   paretoSignificanceGate,
   powerPreflight,
   runEval
-} from "../chunk-J22CKVXN.js";
+} from "../chunk-V75NN2ZR.js";
 import {
   analyzeRuns
 } from "../chunk-GDZAWO2I.js";
@@ -29,7 +29,7 @@ import {
   heldOutGate,
   runImprovementLoop,
   surfaceContentHash
-} from "../chunk-3E5KUXYZ.js";
+} from "../chunk-G4DLZAV5.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,

package/dist/{gepa-DQ3ruj18.d.ts → gepa-BmqTrdtg.d.ts}

@@ -1,4 +1,4 @@
-import { S as Scenario, C as CampaignResult, G as GateResult, D as DispatchFn, a as JudgeConfig, L as LabeledScenarioStore, d as CampaignTraceWriter, e as GenerationRecord, M as MutableSurface, P as ParetoParent, f as SurfaceProposer, g as Gate } from './types-Ctf7XIAL.js';
+import { S as Scenario, C as CampaignResult, G as GateResult, D as DispatchFn, a as JudgeConfig, L as LabeledScenarioStore, d as CampaignTraceWriter, e as GenerationRecord, M as MutableSurface, P as ParetoParent, f as SurfaceProposer, g as Gate } from './types-CNZ0tHET.js';
 import { L as LlmClientOptions } from './llm-client-DyqEH4jH.js';
 
 /**
@@ -368,6 +368,14 @@ type RunImprovementLoopOptions<TScenario extends Scenario, TArtifact> = RunOptim
     /** Optional render override — substrate writes a diff-shaped surface; pass
      *  a function to format the promoted surface differently. */
     renderPromotedDiff?: (winnerSurface: MutableSurface, baselineSurface: MutableSurface) => string;
+    /** Placebo control. When supplied AND the winner differs from baseline, the
+     *  loop scores a THIRD holdout arm: the winner surface with its content
+     *  footprint-matched-blanked by this function (typically via `neutralizeText`).
+     *  Its scores are exposed to the gate as `ctx.neutralizedJudgeScores`, letting
+     *  a `neutralizationGate` reject a win whose lift survives blanking the content
+     *  (decorative — driven by footprint, not content). Costs one extra holdout
+     *  campaign; omit to skip. Return a byte/layout-matched blank of the winner. */
+    neutralize?: (winnerSurface: MutableSurface, baselineSurface: MutableSurface) => MutableSurface;
 };
 interface RunImprovementLoopResult<TArtifact, TScenario extends Scenario> extends RunOptimizationResult<TArtifact, TScenario> {
     baselineOnHoldout: CampaignResult<TArtifact, TScenario>;

package/dist/hosted/index.d.ts

@@ -1,4 +1,4 @@
-import { M as MutableSurface, h as GateDecision } from '../types-Ctf7XIAL.js';
+import { M as MutableSurface, h as GateDecision } from '../types-CNZ0tHET.js';
 import { I as InsightReport } from '../insight-report-B4xrdwEK.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';

package/dist/index.d.ts

@@ -18,8 +18,8 @@ import { A as AgentEvalError, J as JudgeError, a as ConfigError } from './errors
 export { b as AgentEvalErrorCode, C as CaptureIntegrityError, N as NotFoundError, R as ReplayError, V as ValidationError, c as VerificationError } from './errors-oeQrLqXC.js';
 import { b as FeedbackLabel, F as FeedbackTrajectoryStore, a as FeedbackTrajectory } from './feedback-trajectory-C9KCo8ag.js';
 export { c as FeedbackArtifactType, d as FeedbackAttempt, e as FeedbackLabelKind, f as FeedbackLabelSource, g as FeedbackOptimizerRow, h as FeedbackOutcome, i as FeedbackReplayAdapter, j as FeedbackReplayResult, k as FeedbackSeverity, l as FeedbackSplitPolicy, m as FeedbackTask, n as FeedbackTrajectoryFilter, o as FileSystemFeedbackTrajectoryStore, I as InMemoryFeedbackTrajectoryStore, P as PreferenceMemoryEntry, p as ProposedSideEffect, q as assignFeedbackSplit, r as controlRunToFeedbackTrajectory, s as createFeedbackTrajectory, t as feedbackTrajectoriesToDatasetScenarios, u as feedbackTrajectoriesToOptimizerRows, v as feedbackTrajectoryToDatasetScenario, w as feedbackTrajectoryToOptimizerRow, x as parseFeedbackTrajectoriesJsonl, y as renderPreferenceMemoryMarkdown, z as replayFeedbackTrajectories, A as replayFeedbackTrajectory, B as serializeFeedbackTrajectoriesJsonl, C as summarizePreferenceMemory, D as withAssignedFeedbackSplit } from './feedback-trajectory-C9KCo8ag.js';
-import { b as CorrectnessChecker } from './pre-registration-BkkTmtQG.js';
-export { A as ArtifactCheckArtifact, d as ArtifactEventLike, e as ArtifactValidator, f as BackendIntegrityError, B as BackendIntegrityReport, C as CompletionRequirement, a as CompletionVerdict, H as HypothesisManifest, g as HypothesisResult, h as LlmCorrectnessCheckerOpts, L as LlmJudgeDimension, c as LlmJudgeOptions, i as ProducedProposal, P as ProducedState, j as ProposalEventLike, k as RequirementCheck, R as RuntimeEventLike, m as SatisfiedBy, S as SignedManifest, n as SignedManifestAlgo, T as TaskGold, o as ToolCallEventLike, V as ValidationContext, p as ValidationIssue, q as ValidationResult, r as assertRealBackend, s as byteLengthRange, t as canonicalize, u as completionVerdict, v as composeValidators, w as containsAll, x as createLlmCorrectnessChecker, y as createTokenRecallChecker, z as evaluateHypothesis, D as extractProducedState, E as hashJson, F as jsonHasKeys, l as llmJudge, G as parseCorrectnessResponse, I as regexMatch, J as signManifest, K as summarizeBackendIntegrity, M as verifyCompletion, N as verifyManifest } from './pre-registration-BkkTmtQG.js';
+import { b as CorrectnessChecker } from './pre-registration-x3f0VpxT.js';
+export { A as ArtifactCheckArtifact, d as ArtifactEventLike, e as ArtifactValidator, f as BackendIntegrityError, B as BackendIntegrityReport, C as CompletionRequirement, a as CompletionVerdict, H as HypothesisManifest, g as HypothesisResult, h as LlmCorrectnessCheckerOpts, L as LlmJudgeDimension, c as LlmJudgeOptions, i as ProducedProposal, P as ProducedState, j as ProposalEventLike, k as RequirementCheck, R as RuntimeEventLike, m as SatisfiedBy, S as SignedManifest, n as SignedManifestAlgo, T as TaskGold, o as ToolCallEventLike, V as ValidationContext, p as ValidationIssue, q as ValidationResult, r as assertRealBackend, s as byteLengthRange, t as canonicalize, u as completionVerdict, v as composeValidators, w as containsAll, x as createLlmCorrectnessChecker, y as createTokenRecallChecker, z as evaluateHypothesis, D as extractProducedState, E as hashJson, F as jsonHasKeys, l as llmJudge, G as parseCorrectnessResponse, I as regexMatch, J as signManifest, K as summarizeBackendIntegrity, M as verifyCompletion, N as verifyManifest } from './pre-registration-x3f0VpxT.js';
 export { DataAcquisitionPlan, KnowledgeAcquisitionMode, KnowledgeBundle, KnowledgeFallbackPolicy, KnowledgeFreshness, KnowledgeImportance, KnowledgeReadinessReport, KnowledgeRecommendedAction, KnowledgeRequirement, KnowledgeRequirementCategory, KnowledgeResponsibleSurface, KnowledgeSensitivity, ScoreKnowledgeReadinessOptions, UserQuestion, acquisitionPlansForKnowledgeGaps, blockingKnowledgeEval, knowledgeReadinessTracePayload, scoreKnowledgeReadiness, userQuestionsForKnowledgeGaps } from './knowledge/index.js';
 import { h as ReleaseConfidenceThresholds, f as ReleaseConfidenceScorecard } from './release-report-DeJpsBiA.js';
 export { A as ActionableSideInfo, o as AsiSeverity, B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, g as ReleaseConfidenceStatus, i as ReleaseTraceEvidence, j as RenderReleaseReportOptions, V as Verdict, k as assertReleaseConfidence, l as bootstrapCi, m as evaluateReleaseConfidence, n as judgeReplayGate, r as renderReleaseReport } from './release-report-DeJpsBiA.js';
@@ -31,7 +31,7 @@ import { a as AnalyzeTracesInput, A as AnalyzeTracesOptions, b as AnalyzeTracesR
 export { c as AnalyzeTracesTurnSnapshot, d as analyzeTraces } from './analyst-C8HHvfJp.js';
 import { a as TraceAnalystSpan } from './store-C1YxJDEK.js';
 export { D as DEFAULT_TRACE_ANALYST_BUDGETS, b as DatasetOverview, E as ErrorCluster, Q as QueryTracesPage, S as SearchSpanResult, c as SearchTraceResult, d as SpanMatchRecord, e as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, T as TraceAnalysisStore, f as TraceAnalystByteBudgets, g as TraceAnalystFilters, h as TraceAnalystSpanKind, i as TraceAnalystSpanStatus, j as TraceAnalystTraceSummary, V as ViewSpansResult, k as ViewTraceOversized, l as ViewTraceResult } from './store-C1YxJDEK.js';
-import { a as JudgeConfig, J as JudgeScore, S as Scenario$1, g as Gate } from './types-Ctf7XIAL.js';
+import { a as JudgeConfig, J as JudgeScore, S as Scenario$1, g as Gate } from './types-CNZ0tHET.js';
 import { A as AnalyzeRunsOptions } from './analyze-runs-DJYpep3L.js';
 import { S as SteeringBundle } from './harness-optimizer-mOl9XX_O.js';
 export { D as DEFAULT_HARNESS_OBJECTIVES, H as HarnessAdapter, a as HarnessExperimentConfig, b as HarnessExperimentResult, c as HarnessIntervention, d as HarnessRunRequest, e as HarnessRunResult, f as HarnessScenario, g as HarnessSelection, h as HarnessVariant, i as HarnessVariantReport, M as MeasurementPolicy, j as SteeringDelta, k as SteeringRolePrompt, W as WorkflowTopology, m as mergeSteeringBundle, r as renderSteeringText, l as runHarnessExperiment, s as selectHarnessVariant, n as summarizeHarnessResults } from './harness-optimizer-mOl9XX_O.js';
@@ -76,7 +76,7 @@ export { C as CallbackResearcher, d as CallbackResearcherOptions, e as CampaignF
 export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, m as GateDecision, n as GateEvidence, H as HeldOutGate, o as HeldOutGateConfig, q as HeldOutGateRejectionCode, P as ParetoFigureSpec, c as ParetoPoint, R as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as ResearchReport, e as ResearchReportCandidate, f as ResearchReportDecision, g as ResearchReportMethodology, h as ResearchReportOptions, i as ResearchReportRecommendation, S as SummaryTable, j as SummaryTableOptions, k as SummaryTableRow, l as gainHistogram, p as paretoChart, r as researchReport, s as summaryTable } from './summary-report-QMZVe3P-.js';
 export { L as LockedJsonlAppender } from './testing-C21CHsq2.js';
 export { I as InterimReleaseConfidence, a as InterimReleaseConfidenceInput, P as PairedEvalueOptions, b as PairedEvalueSequence, c as PairedEvalueStep, S as SequentialDecision, e as evaluateInterimReleaseConfidence, p as pairedEvalueSequence } from './sequential-5iSVfzl2.js';
-import { j as GepaProposerConstraints, a as RunImprovementLoopResult } from './gepa-DQ3ruj18.js';
+import { j as GepaProposerConstraints, a as RunImprovementLoopResult } from './gepa-BmqTrdtg.js';
 export { IntegrityResult, IntegrityViolation, JourneySpec, PerfBaseline, PerfGateResult, PerfRegression, PerfScenario, PerfStat, ScenarioAxes, assertRecordIntegrity, checkRecordIntegrity, expandMatrix, gatePerf, scenarioKey, summarizeRecords } from './perf/index.js';
 export { AgentProfileRuntimeReceipt, ProductBenchmarkArm, ProductBenchmarkArtifactPaths, ProductBenchmarkBudgets, ProductBenchmarkExportOptions, ProductBenchmarkExportResult, ProductBenchmarkManifest, ProductBenchmarkProfileRef, ProductBenchmarkRecord, ProductBenchmarkRepoRef, ProductBenchmarkRunInput, ProductBenchmarkScenario, ProductBenchmarkSingleRunExportOptions, ProductBenchmarkSplit, ProductBenchmarkSubstrateVersions, ProductBenchmarkValidationReport, RuntimeResolution, assertProductBenchmarkRun, buildProductBenchmarkManifest, exportProductBenchmark, exportProductBenchmarkRuns, findProductBenchmarkArtifacts, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, readProductBenchmarkManifest, readProductBenchmarkRecords, runRecordToProductBenchmarkRecord, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun } from './product-benchmark/index.js';
 import '@ax-llm/ax';

package/dist/index.js

@@ -72,7 +72,7 @@ import {
   llmJudge,
   parseCorrectnessResponse,
   verifyCompletion
-} from "./chunk-2HLM4SJJ.js";
+} from "./chunk-6PF6LXRY.js";
 import {
   DEFAULT_MUTATION_PRIMITIVES,
   DEFAULT_RED_TEAM_CORPUS,
@@ -96,7 +96,7 @@ import {
   scoreRedTeamOutput,
   surfaceContentHash,
   toolNamesForRun
-} from "./chunk-3E5KUXYZ.js";
+} from "./chunk-G4DLZAV5.js";
 import {
   BackendIntegrityError,
   assertRealBackend,

package/dist/multishot/index.d.ts

@@ -1,4 +1,4 @@
-import { J as JudgeScore } from '../types-Ctf7XIAL.js';
+import { J as JudgeScore } from '../types-CNZ0tHET.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { M as MatrixResult } from '../types-BUxNaJ8c.js';
 import '../run-record-I-Z3JNvO.js';

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.106.3",
+    "version": "0.107.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/{pre-registration-BkkTmtQG.d.ts → pre-registration-x3f0VpxT.d.ts}

@@ -1,7 +1,7 @@
 import { A as AgentEvalError } from './errors-oeQrLqXC.js';
 import { R as RunRecord } from './run-record-I-Z3JNvO.js';
 import { C as ChatClient } from './types-D1ytG0Yg.js';
-import { c as JudgeDimension, S as Scenario, a as JudgeConfig } from './types-Ctf7XIAL.js';
+import { c as JudgeDimension, S as Scenario, a as JudgeConfig } from './types-CNZ0tHET.js';
 import { TCloud } from '@tangle-network/tcloud';
 import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
 import { D as DefaultVerdict } from './verdict-C9MlYujm.js';

package/dist/{provenance-Bsyjc67Z.d.ts → provenance-CIjRcI6m.d.ts}

@@ -1,9 +1,9 @@
-import { S as Scenario, g as Gate, G as GateResult, n as GateContext, C as CampaignResult, p as Mutator, f as SurfaceProposer, M as MutableSurface, h as GateDecision } from './types-Ctf7XIAL.js';
+import { S as Scenario, g as Gate, G as GateResult, n as GateContext, C as CampaignResult, p as Mutator, f as SurfaceProposer, M as MutableSurface, h as GateDecision } from './types-CNZ0tHET.js';
 import { R as RedTeamCase } from './red-team-KmmiqBlY.js';
 import { R as RunRecord } from './run-record-I-Z3JNvO.js';
 import { D as Direction } from './pareto-E-pembql.js';
 import { a as PairedBootstrapResult } from './statistics-D88peojY.js';
-import { b as RunCampaignOptions, C as CampaignStorage } from './gepa-DQ3ruj18.js';
+import { b as RunCampaignOptions, C as CampaignStorage } from './gepa-BmqTrdtg.js';
 import { HostedClient, TraceSpanEvent } from './hosted/index.js';
 
 /**

package/dist/rl.d.ts

@@ -10,7 +10,7 @@ import { R as Researcher, F as FailureMode, S as SteeringChange, E as Experiment
 export { r as runEvalCampaign } from './researcher-Wc7dx6GM.js';
 import { a as VerificationReport } from './multi-layer-verifier-CI4jdX-q.js';
 import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
-import { C as CampaignResult } from './types-Ctf7XIAL.js';
+import { C as CampaignResult } from './types-CNZ0tHET.js';
 import '@tangle-network/agent-interface';
 import './errors-oeQrLqXC.js';
 import './schema-m0gsnbt3.js';

package/dist/{types-Ctf7XIAL.d.ts → types-CNZ0tHET.d.ts}

@@ -289,6 +289,16 @@ interface GateContext<TArtifact, TScenario extends Scenario> {
      *  cannot represent both. A gate computing a holdout delta MUST read
      *  candidate from `judgeScores` and baseline from here. */
     baselineJudgeScores?: Map<string, Record<string, JudgeScore>>;
+    /** Neutralized-arm judge scores, keyed by cellId — the winner surface with its
+     *  content footprint-matched-blanked (via a `neutralize` fn). Same scenarios as
+     *  `judgeScores`. Present ONLY when `runImprovementLoop` was given a `neutralize`
+     *  function. A placebo gate (`neutralizationGate`) compares this arm's lift
+     *  against the candidate's to reject decorative wins (lift from footprint, not
+     *  content). Undefined otherwise. */
+    neutralizedJudgeScores?: Map<string, Record<string, JudgeScore>>;
+    /** Neutralized-arm artifacts, keyed by cellId. Present alongside
+     *  `neutralizedJudgeScores`. */
+    neutralizedArtifacts?: Map<string, TArtifact>;
     scenarios: TScenario[];
     cost: {
         candidate: number;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.106.3",
+  "version": "0.107.0",
   "description": "Evaluate and improve AI agents from runs, traces, judges, and feedback. Compare candidates, cluster failures, measure lift, and gate releases.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {