@tangle-network/agent-eval 0.106.3 → 0.107.0

package/dist/adapters/http.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-Ctf7XIAL.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-CNZ0tHET.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/adapters/langchain.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-Ctf7XIAL.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-CNZ0tHET.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/adapters/otel.d.ts

@@ -1,5 +1,5 @@
 import { TraceSpanEvent, HostedClient } from '../hosted/index.js';
-import '../types-Ctf7XIAL.js';
+import '../types-CNZ0tHET.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/campaign/index.d.ts

@@ -1,11 +1,11 @@
-import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-BkkTmtQG.js';
-export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-BkkTmtQG.js';
+import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-x3f0VpxT.js';
+export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-x3f0VpxT.js';
 import { A as AnalyzeTracesOptions, a as AnalyzeTracesInput, b as AnalyzeTracesResult } from '../analyst-C8HHvfJp.js';
-import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, e as GenerationRecord, g as Gate, 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, y as LabeledScenarioSource, C as CampaignResult, m as CodeSurface } from '../types-Ctf7XIAL.js';
-export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, n as GateContext, h as GateDecision, G as GateResult, o as GenerationCandidate, A as JudgeAggregate, c 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-Ctf7XIAL.js';
-import { e as CampaignRunPlan, P as PlanCampaignRunOptions, C as CampaignStorage, b as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-DQ3ruj18.js';
-export { h as CampaignRunPlanCell, j as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, k as OpenAutoPrResult, a as RunImprovementLoopResult, R as RunOptimizationOptions, l as RunOptimizationResult, m as countSentenceEdits, n as defaultRenderDiff, o as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, p as openAutoPr, q as planCampaignRun, r as runCampaign, d as runImprovementLoop, s as runOptimization, t as surfaceHash } from '../gepa-DQ3ruj18.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, z as surfaceContentHash } from '../provenance-Bsyjc67Z.js';
+import { S as Scenario, M as MutableSurface, b as DispatchContext, a 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, y as LabeledScenarioSource, C as CampaignResult, m as CodeSurface } from '../types-CNZ0tHET.js';
+export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, n as GateContext, h as GateDecision, G as GateResult, o as GenerationCandidate, A as JudgeAggregate, c 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-CNZ0tHET.js';
+import { e as CampaignRunPlan, P as PlanCampaignRunOptions, C as CampaignStorage, b as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-BmqTrdtg.js';
+export { h as CampaignRunPlanCell, j as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, k as OpenAutoPrResult, a as RunImprovementLoopResult, R as RunOptimizationOptions, l as RunOptimizationResult, m as countSentenceEdits, n as defaultRenderDiff, o as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, p as openAutoPr, q as planCampaignRun, r as runCampaign, d as runImprovementLoop, s as runOptimization, t as surfaceHash } from '../gepa-BmqTrdtg.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, z as surfaceContentHash } from '../provenance-CIjRcI6m.js';
 import { E as EProcessState, a as PairedBootstrapResult } from '../statistics-D88peojY.js';
 import { L as LlmClientOptions } from '../llm-client-DyqEH4jH.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
@@ -435,6 +435,43 @@ declare function loadEvalFixtureScenarios(evalsDir: string, options?: LoadEvalFi
  */
 declare function planEvalFixtureRun<TArtifact = unknown>(options: PlanEvalFixtureRunOptions<TArtifact>): EvalFixtureRunPlan;
 
+/**
+ * @module
+ * Composable placebo / neutralization promotion gate.
+ *
+ * A held-out gate proves a candidate beat baseline. It CANNOT prove the lift came
+ * from the candidate's CONTENT rather than from the prompt/mount FOOTPRINT the
+ * content happened to add (more bytes, a longer prompt). This gate closes that
+ * hole: it compares the candidate's held-out lift against the lift of a
+ * FOOTPRINT-MATCHED neutralized variant (same layout + length, zero content, via
+ * `neutralizeText`). If the neutralized variant reproduces more than
+ * `maxDecorativeFraction` of the candidate's lift, the lift is decorative — it
+ * survives blanking the content — and the candidate is HELD regardless of how
+ * large or significant its raw lift is.
+ *
+ * Compose it AFTER the significance gate — significance says the lift is real,
+ * this says the lift is CAUSED BY THE CONTENT:
+ *   composeGate(heldOutGate({ ... }), neutralizationGate({ ... }))
+ *
+ * Requires `ctx.neutralizedJudgeScores`, populated by `runImprovementLoop` when it
+ * is given a `neutralize` function. A gate composed without that wiring fails
+ * loud rather than silently passing an unproven candidate.
+ */
+
+interface NeutralizationGateOptions<TScenario extends Scenario = Scenario> {
+    scenarios: TScenario[];
+    /** Reject when the neutralized (content-blanked, footprint-matched) variant
+     *  reproduces at least this fraction of the candidate's held-out lift. Default
+     *  0.5 — if blanking the content keeps half the lift, the content is decorative.
+     *  Equality rejects: a neutralized lift == threshold·candidateLift is decorative. */
+    maxDecorativeFraction?: number;
+}
+/**
+ * Composable placebo gate: ships only when the candidate's held-out lift is NOT
+ * mostly reproduced by a footprint-matched neutralized variant.
+ */
+declare function neutralizationGate<TArtifact, TScenario extends Scenario>(options: NeutralizationGateOptions<TScenario>): Gate<TArtifact, TScenario>;
+
 /**
  * Anytime-valid sequential promotion gate — an e-process (betting
  * test-martingale, see `eProcess` in `statistics.ts`) over paired
@@ -751,6 +788,34 @@ declare class FsLabeledScenarioStore implements LabeledScenarioStore {
     private pathForSource;
 }
 
+/**
+ * @module
+ * Footprint-matched neutralization — the placebo control for content-vs-footprint
+ * attribution in a promotion gate.
+ *
+ * A promoted surface can raise a held-out score two different ways:
+ *  1. its CONTENT is informative (the thing we want to promote), or
+ *  2. it merely added prompt/mount FOOTPRINT — more bytes, more lines, a longer
+ *     more authoritative-looking prompt — that the model spends attention on
+ *     regardless of what the bytes say.
+ *
+ * A held-out gate proves the candidate beat baseline; it cannot separate (1) from
+ * (2). `neutralizeText` produces a variant that keeps the input's layout and
+ * length while carrying ZERO information, so scoring it isolates the footprint
+ * contribution (2). Feed the neutralized variant's scores to `neutralizationGate`:
+ * any lift it still holds over baseline is decorative, and a candidate whose lift
+ * survives neutralization is rejected however large its raw lift.
+ */
+/**
+ * Blank every non-whitespace character to a 1-byte filler while preserving all
+ * whitespace. Line count, indentation, and word/line lengths are unchanged — so
+ * the neutralized variant has the same layout and (for ASCII) the same byte
+ * footprint as the input, but no readable content. Whitespace is preserved
+ * deliberately: collapsing it would change the token structure and stop the
+ * variant from being a true footprint match.
+ */
+declare function neutralizeText(content: string): string;
+
 /**
  * FAPO (Fully Autonomous Prompt Optimization) is an orchestration policy, not
  * a new prompt mutation primitive. The paper's loop evaluates an inspectable
@@ -2007,4 +2072,4 @@ declare function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAd
  *  as a ref under the adapter's worktree dir. */
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 
-export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type ApplySkillPatchResult, type BuildAnalystSurfaceDispatchOptions, type CampaignBreakdown, CampaignResult, CampaignRunPlan, CampaignStorage, CodeSurface, type CompareProposersOptions, type DimensionRegression, type DiscriminationScore, DispatchContext, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoEntryConfig, type FapoFailureCluster, type FapoOptimizationLevel, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, Gate, GenerationRecord, type GitWorktreeAdapterOptions, type Governor, type GovernorContext, type GovernorOp, type HaloProposerOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type HeuristicGovernorOptions, type JsonPrimitive, type JsonValue, JudgeConfig, JudgeScore, LabelTrust, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioSource, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, Lineage, type LineageEdge, type LineageGraph, type LineageNode, type LineageNodeInput, type LineageStore, type LoadEvalFixtureScenariosOptions, type MemoryCurationProposerOptions, MutableSurface, type OptimizerEntryConfig, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditProposerOptions, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, ProposeContext, type ProposePatchesArgs, ProposedCandidate, type ProposerComparison, type ProposerEntry, type ProposerPairwise, type ProposerScore, type RejectedEdit, RunCampaignOptions, RunImprovementLoopOptions, type RunLineageLoopOptions, type RunLineageLoopResult, type RunLineageLoopSeed, type RunLineageOptions, type RunLineageResult, type RunLineageSeed, type RunLineageStepResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, Scenario, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, SurfaceProposer, type SurfaceScore, type TraceAnalystProposerOptions, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, applySkillPatch, buildAnalystSurfaceDispatch, callbackGovernor, campaignBreakdown, campaignMeanComposite, compareProposers, detectScale, dimensionRegressions, discoverEvalFixtures, extractFapoAttributionSignals, failureModeRecallJudge, fapoEscalationEntry, fapoProposer, fsLineageStore, gepaParetoEntry, gepaReflectionEntry, gitWorktreeAdapter, haloProposer, heldoutSignificance, heuristicGovernor, lineageNodeId, loadEvalFixture, loadEvalFixtureScenarios, makePlaybackDispatch, memLineageStore, memoryCurationProposer, pairHoldout, parameterSweepProposer, parseSkillPatchResponse, patchEditCount, planEvalFixtureRun, policyEditProposer, renderScoreboardMarkdown, resolveRunDir, resolveWorktreePath, runLineage, runLineageLoop, runProfileMatrix, runSkillOpt, scoreDiscrimination, scoreUserStory, scoreboardSummary, selectDiscriminative, sequentialDecide, sequentialPairedGate, skillOptEntry, skillOptProposer, tangleTracesRoot, traceAnalystProposer, userStoryScoreboard };
+export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type ApplySkillPatchResult, type BuildAnalystSurfaceDispatchOptions, type CampaignBreakdown, CampaignResult, CampaignRunPlan, CampaignStorage, CodeSurface, type CompareProposersOptions, type DimensionRegression, type DiscriminationScore, DispatchContext, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoEntryConfig, type FapoFailureCluster, type FapoOptimizationLevel, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, Gate, GenerationRecord, type GitWorktreeAdapterOptions, type Governor, type GovernorContext, type GovernorOp, type HaloProposerOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type HeuristicGovernorOptions, type JsonPrimitive, type JsonValue, JudgeConfig, JudgeScore, LabelTrust, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioSource, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, Lineage, type LineageEdge, type LineageGraph, type LineageNode, type LineageNodeInput, type LineageStore, type LoadEvalFixtureScenariosOptions, type MemoryCurationProposerOptions, MutableSurface, type NeutralizationGateOptions, type OptimizerEntryConfig, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditProposerOptions, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, ProposeContext, type ProposePatchesArgs, ProposedCandidate, type ProposerComparison, type ProposerEntry, type ProposerPairwise, type ProposerScore, type RejectedEdit, RunCampaignOptions, RunImprovementLoopOptions, type RunLineageLoopOptions, type RunLineageLoopResult, type RunLineageLoopSeed, type RunLineageOptions, type RunLineageResult, type RunLineageSeed, type RunLineageStepResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, Scenario, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, SurfaceProposer, type SurfaceScore, type TraceAnalystProposerOptions, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, applySkillPatch, buildAnalystSurfaceDispatch, callbackGovernor, campaignBreakdown, campaignMeanComposite, compareProposers, detectScale, dimensionRegressions, discoverEvalFixtures, extractFapoAttributionSignals, failureModeRecallJudge, fapoEscalationEntry, fapoProposer, fsLineageStore, gepaParetoEntry, gepaReflectionEntry, gitWorktreeAdapter, haloProposer, heldoutSignificance, heuristicGovernor, lineageNodeId, loadEvalFixture, loadEvalFixtureScenarios, makePlaybackDispatch, memLineageStore, memoryCurationProposer, neutralizationGate, neutralizeText, pairHoldout, parameterSweepProposer, parseSkillPatchResponse, patchEditCount, planEvalFixtureRun, policyEditProposer, renderScoreboardMarkdown, resolveRunDir, resolveWorktreePath, runLineage, runLineageLoop, runProfileMatrix, runSkillOpt, scoreDiscrimination, scoreUserStory, scoreboardSummary, selectDiscriminative, sequentialDecide, sequentialPairedGate, skillOptEntry, skillOptProposer, tangleTracesRoot, traceAnalystProposer, userStoryScoreboard };

package/dist/campaign/index.js

@@ -7,7 +7,7 @@ import {
   paretoSignificanceGate,
   powerPreflight,
   runEval
-} from "../chunk-J22CKVXN.js";
+} from "../chunk-V75NN2ZR.js";
 import {
   HARNESS_NATIVE_MODEL,
   agentProfileHash,
@@ -17,7 +17,7 @@ import {
   harnessAxisOf,
   llmJudge,
   verifyCompletion
-} from "../chunk-2HLM4SJJ.js";
+} from "../chunk-6PF6LXRY.js";
 import {
   buildLoopProvenanceRecord,
   campaignBreakdown,
@@ -43,7 +43,7 @@ import {
   runOptimization,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-3E5KUXYZ.js";
+} from "../chunk-G4DLZAV5.js";
 import {
   assertRealBackend,
   contentHash,
@@ -737,6 +737,86 @@ function assertInside(root, target, label) {
   throw new Error(`loadEvalFixture: fixture path escapes evalsDir: ${label}`);
 }
 
+// src/campaign/gates/neutralization-gate.ts
+function mean(xs) {
+  return xs.length === 0 ? 0 : xs.reduce((a, b) => a + b, 0) / xs.length;
+}
+function pairedLift(arm, baseline, scenarioIds) {
+  const paired = pairHoldout(arm, baseline, scenarioIds, (s) => s.composite);
+  const deltas = paired.after.map((a, i) => a - (paired.before[i] ?? 0));
+  return { lift: mean(deltas), n: deltas.length };
+}
+function neutralizationGate(options) {
+  const maxDecorativeFraction = options.maxDecorativeFraction ?? 0.5;
+  return {
+    name: "neutralizationGate",
+    async decide(ctx) {
+      if (!ctx.baselineJudgeScores) {
+        throw new Error(
+          "neutralizationGate: ctx.baselineJudgeScores is required \u2014 the placebo control measures lift OVER baseline."
+        );
+      }
+      if (!ctx.neutralizedJudgeScores) {
+        throw new Error(
+          "neutralizationGate: ctx.neutralizedJudgeScores is required. It is populated by runImprovementLoop only when a `neutralize` function is supplied \u2014 composing this gate without that wiring would pass an unproven candidate."
+        );
+      }
+      const scenarioIds = new Set(options.scenarios.map((s) => s.id));
+      const cand = pairedLift(
+        ctx.judgeScores,
+        ctx.baselineJudgeScores,
+        scenarioIds
+      );
+      const neut = pairedLift(
+        ctx.neutralizedJudgeScores,
+        ctx.baselineJudgeScores,
+        scenarioIds
+      );
+      if (cand.lift <= 0) {
+        return {
+          decision: "hold",
+          reasons: [
+            `neutralization: candidate held-out lift ${cand.lift.toFixed(3)} \u2264 0 \u2014 no positive lift to attribute to content`
+          ],
+          contributingGates: [
+            {
+              name: "neutralizationGate",
+              passed: false,
+              detail: { candidateLift: cand.lift, neutralizedLift: neut.lift, n: cand.n }
+            }
+          ],
+          delta: cand.lift
+        };
+      }
+      const decorativeFraction = neut.lift / cand.lift;
+      const passed = decorativeFraction < maxDecorativeFraction;
+      const pct = (decorativeFraction * 100).toFixed(0);
+      return {
+        decision: passed ? "ship" : "hold",
+        reasons: passed ? [
+          `neutralization: content is causal \u2014 blanked variant reproduces ${pct}% of the lift (< ${(maxDecorativeFraction * 100).toFixed(0)}%); candidate \u0394 ${cand.lift.toFixed(3)}, neutralized \u0394 ${neut.lift.toFixed(3)}`
+        ] : [
+          `neutralization: lift is DECORATIVE \u2014 blanking the content (footprint-matched) reproduces ${pct}% of the lift (\u2265 ${(maxDecorativeFraction * 100).toFixed(0)}%); candidate \u0394 ${cand.lift.toFixed(3)}, neutralized \u0394 ${neut.lift.toFixed(3)}`
+        ],
+        contributingGates: [
+          {
+            name: "neutralizationGate",
+            passed,
+            detail: {
+              candidateLift: cand.lift,
+              neutralizedLift: neut.lift,
+              decorativeFraction,
+              maxDecorativeFraction,
+              n: cand.n
+            }
+          }
+        ],
+        delta: cand.lift
+      };
+    }
+  };
+}
+
 // src/campaign/gates/sequential.ts
 import { createHash as createHash3 } from "crypto";
 function verifyManifestSync(m) {
@@ -1214,6 +1294,12 @@ function appendLine(path, line) {
   }
 }
 
+// src/campaign/neutralize.ts
+var FILLER = "#";
+function neutralizeText(content) {
+  return content.replace(/\S/g, FILLER);
+}
+
 // src/campaign/proposers/fapo.ts
 var FAPO_LEVELS = ["prompt", "parameter", "structural"];
 var MAX_FINDING_DEPTH = 16;
@@ -2155,8 +2241,8 @@ async function compareProposerEntries(opts) {
     });
     const score = {
       name: w.name,
-      baselineComposite: mean(baselineArr),
-      winnerComposite: mean(w.arr),
+      baselineComposite: mean2(baselineArr),
+      winnerComposite: mean2(w.arr),
       lift: boot.mean,
       liftCi: { low: boot.low, high: boot.high },
       costUsd: w.costUsd,
@@ -2193,7 +2279,7 @@ async function compareProposerEntries(opts) {
   });
   return { scores, best, pairwise, holdoutScenarioIds: scenarioIds };
 }
-function mean(xs) {
+function mean2(xs) {
   return xs.length === 0 ? 0 : xs.reduce((a, b) => a + b, 0) / xs.length;
 }
 function slug(name) {
@@ -2614,12 +2700,12 @@ function sanitize(id) {
 function sha(input) {
   return createHash5("sha256").update(JSON.stringify(input)).digest("hex");
 }
-function mean2(xs) {
+function mean3(xs) {
   return xs.length === 0 ? 0 : xs.reduce((a, b) => a + b, 0) / xs.length;
 }
 function cellComposite(cell) {
   const composites = Object.values(cell.judgeScores).map((s) => s.composite);
-  return composites.length === 0 ? 0 : mean2(composites);
+  return composites.length === 0 ? 0 : mean3(composites);
 }
 function requireResolvedModel(cell, profileId) {
   const resolved = cell.resolvedModel?.trim();
@@ -2655,7 +2741,7 @@ function buildRunRecord(args) {
     if (js.notes) notes.push(`${judgeName}: ${js.notes}`);
   }
   const perDimMean = {};
-  for (const [dim, values] of Object.entries(dimAccum)) perDimMean[dim] = mean2(values);
+  for (const [dim, values] of Object.entries(dimAccum)) perDimMean[dim] = mean3(values);
   let costUsd = cell.costUsd;
   let costEstimated = false;
   if (costUsd === 0 && cell.tokenUsage.output > 0 && isModelPriced(model)) {
@@ -2819,7 +2905,7 @@ async function runProfileMatrix(opts) {
       // one harness, so the first record's model is representative).
       model: declaredModel === HARNESS_NATIVE_MODEL ? profileRecords[0]?.model ?? declaredModel : declaredModel,
       records: profileRecords.length,
-      meanComposite: mean2(profileRecords.map(compositeOf)),
+      meanComposite: mean3(profileRecords.map(compositeOf)),
       totalCostUsd: pricedTotalCostUsd,
       integrity: summarizeBackendIntegrity(profileRecords)
     };
@@ -2849,7 +2935,7 @@ function rollup(records, keyOf) {
     groups.set(key, arr);
   }
   const out = {};
-  for (const [key, xs] of groups) out[key] = { meanComposite: mean2(xs), n: xs.length };
+  for (const [key, xs] of groups) out[key] = { meanComposite: mean3(xs), n: xs.length };
   return out;
 }
 function rollupByPersona(records, scenarios, personaOf) {
@@ -3409,6 +3495,8 @@ export {
   makePlaybackDispatch,
   memLineageStore,
   memoryCurationProposer,
+  neutralizationGate,
+  neutralizeText,
   openAutoPr,
   pairHoldout,
   parameterSweepProposer,