@tangle-network/agent-eval 0.25.0 → 0.27.2

package/dist/knowledge/index.d.ts

@@ -1,4 +1,4 @@
-import { j as ControlSeverity, C as ControlEvalResult } from '../control-runtime-BuJHoLg0.js';
+import { j as ControlSeverity, C as ControlEvalResult } from '../control-runtime-BZ_lVLYW.js';
 import { T as TraceEmitter } from '../emitter-DP_cSSiw.js';
 import '../store-Db2Bv8Cf.js';
 

package/dist/knowledge/index.js

@@ -4,8 +4,8 @@ import {
   knowledgeReadinessTracePayload,
   scoreKnowledgeReadiness,
   userQuestionsForKnowledgeGaps
-} from "../chunk-WWYCWKUM.js";
-import "../chunk-LSH4MMOZ.js";
+} from "../chunk-3CKU6VGU.js";
+import "../chunk-NCRFYPS3.js";
 import "../chunk-TVVP3ZZQ.js";
 import "../chunk-PZ5AY32C.js";
 export {

package/dist/{multi-layer-verifier-LkP3LVKj.d.ts → multi-layer-verifier-U-c8ge1k.d.ts}

@@ -51,7 +51,7 @@ interface LayerResult {
      * diagnostic name; null = "diagnostic not applicable / not measured."
      * Renderers that know the keys can display them; ones that don't,
      * ignore. Free-form on purpose — consumers type the value shape in
-     * their own namespace. Added in 0.10.
+     * their own namespace.
      */
     diagnostics?: Record<string, number | null>;
     /** Any rich per-layer detail — rendered as-is by consumers that know the layer. */

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.25.0",
+    "version": "0.27.2",
     "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/optimization.d.ts

@@ -1,11 +1,11 @@
-export { C as CallbackResearcher, a as CallbackResearcherOptions, b as CampaignFactoryParams, c as CampaignIntegrityPolicy, d as CampaignRunContext, e as CampaignRunOutcome, f as CampaignRunner, g as CampaignScenario, h as CampaignVariant, E as EvalCampaignOptions, i as EvalCampaignResult, j as ExperimentPlan, k as ExperimentResult, F as FailedRun, l as FailureMode, N as NoopResearcher, R as Researcher, S as SteeringChange, r as runEvalCampaign } from './researcher-BPT8x_NT.js';
-export { F as FeedbackArtifactType, a as FeedbackAttempt, b as FeedbackLabel, c as FeedbackLabelKind, d as FeedbackLabelSource, e as FeedbackOptimizerRow, f as FeedbackOutcome, g as FeedbackReplayAdapter, h as FeedbackReplayResult, i as FeedbackSeverity, j as FeedbackSplitPolicy, k as FeedbackTask, l as FeedbackTrajectory, m as FeedbackTrajectoryFilter, n as FeedbackTrajectoryStore, 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-DfFdrraJ.js';
-export { A as ActionableSideInfo, a as AsiSeverity, D as DEFAULT_MUTATION_PRIMITIVES, E as EvolvableVariant, G as GenerationReport, I as InMemoryTrialCache, M as MultiShotGateConfig, b as MultiShotGateResult, c as MultiShotMutateAdapter, d as MultiShotOptimizationConfig, e as MultiShotOptimizationResult, f as MultiShotRun, g as MultiShotRunInput, h as MultiShotRunner, i as MultiShotScore, j as MultiShotScorer, k as MultiShotSplit, l as MultiShotTrace, m as MultiShotTrialResult, n as MultiShotVariant, o as MutateAdapter, P as PromptEvolutionConfig, p as PromptEvolutionEvent, q as PromptEvolutionResult, R as ReflectionContext, r as ReflectionProposal, S as ScenarioAggregate, s as ScoreAdapter, T as TrialCache, t as TrialResult, u as TrialTrace, V as VariantAggregate, v as buildReflectionPrompt, w as defaultMultiShotObjectives, x as parseReflectionResponse, y as runMultiShotOptimization, z as runPromptEvolution, B as trialTraceFromMultiShotTrial } from './summary-report-C7VPYEj2.js';
+export { C as CallbackResearcher, a as CallbackResearcherOptions, b as CampaignFactoryParams, c as CampaignIntegrityPolicy, d as CampaignRunContext, e as CampaignRunOutcome, f as CampaignRunner, g as CampaignScenario, h as CampaignVariant, E as EvalCampaignOptions, i as EvalCampaignResult, j as ExperimentPlan, k as ExperimentResult, F as FailedRun, l as FailureMode, N as NoopResearcher, R as Researcher, S as SteeringChange, r as runEvalCampaign } from './researcher-G81CWc0q.js';
+export { F as FeedbackArtifactType, a as FeedbackAttempt, b as FeedbackLabel, c as FeedbackLabelKind, d as FeedbackLabelSource, e as FeedbackOptimizerRow, f as FeedbackOutcome, g as FeedbackReplayAdapter, h as FeedbackReplayResult, i as FeedbackSeverity, j as FeedbackSplitPolicy, k as FeedbackTask, l as FeedbackTrajectory, m as FeedbackTrajectoryFilter, n as FeedbackTrajectoryStore, 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-D1aGKusy.js';
+export { A as ActionableSideInfo, a as AsiSeverity, D as DEFAULT_MUTATION_PRIMITIVES, E as EvolvableVariant, G as GenerationReport, I as InMemoryTrialCache, M as MultiShotGateConfig, b as MultiShotGateResult, c as MultiShotMutateAdapter, d as MultiShotOptimizationConfig, e as MultiShotOptimizationResult, f as MultiShotRun, g as MultiShotRunInput, h as MultiShotRunner, i as MultiShotScore, j as MultiShotScorer, k as MultiShotSplit, l as MultiShotTrace, m as MultiShotTrialResult, n as MultiShotVariant, o as MutateAdapter, P as PromptEvolutionConfig, p as PromptEvolutionEvent, q as PromptEvolutionResult, R as ReflectionContext, r as ReflectionProposal, S as ScenarioAggregate, s as ScoreAdapter, T as TrialCache, t as TrialResult, u as TrialTrace, V as VariantAggregate, v as buildReflectionPrompt, w as defaultMultiShotObjectives, x as parseReflectionResponse, y as runMultiShotOptimization, z as runPromptEvolution, B as trialTraceFromMultiShotTrial } from './summary-report-Dl4akLKX.js';
 import './errors-BZ9sTdz7.js';
 import './integrity-DK2EBVZC.js';
 import './store-Db2Bv8Cf.js';
 import './run-record-CqzahIbx.js';
 import './emitter-DP_cSSiw.js';
-import './control-runtime-BuJHoLg0.js';
+import './control-runtime-BZ_lVLYW.js';
 import './dataset-CiK_3LDr.js';
-import './failure-cluster-C2EGSDiT.js';
+import './failure-cluster-Cw65_5FY.js';

package/dist/optimization.js

@@ -25,18 +25,18 @@ import {
   summarizePreferenceMemory,
   trialTraceFromMultiShotTrial,
   withAssignedFeedbackSplit
-} from "./chunk-EDUKQ5AM.js";
+} from "./chunk-SZSBQUIJ.js";
 import "./chunk-NLMNWKVM.js";
 import {
   runEvalCampaign
-} from "./chunk-SESZDQPX.js";
+} from "./chunk-RUI6SIHY.js";
 import "./chunk-4S4BM3QQ.js";
-import "./chunk-2A5XJB43.js";
-import "./chunk-I4MBDTY5.js";
+import "./chunk-5AKPEK5L.js";
+import "./chunk-R5UQJNKC.js";
 import "./chunk-KTGTIOFD.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-TVVP3ZZQ.js";
-import "./chunk-4F5DQN55.js";
+import "./chunk-VSMTAMNK.js";
 import "./chunk-NG236HPC.js";
 import "./chunk-PZ5AY32C.js";
 export {

package/dist/pipelines/index.d.ts

@@ -1,5 +1,5 @@
 import { g as BudgetSpec, T as TraceStore, h as RunFilter, R as Run, a as ToolSpan } from '../store-Db2Bv8Cf.js';
-export { a as FailureCluster, F as FailureClusterReport, f as failureClusterView } from '../failure-cluster-C2EGSDiT.js';
+export { a as FailureCluster, F as FailureClusterReport, f as failureClusterView } from '../failure-cluster-Cw65_5FY.js';
 import { a as TrajectoryStep } from '../trajectory-CnoBo-JY.js';
 import { B as BaselineOptions, a as BaselineReport } from '../baseline-4R5deP0N.js';
 export { c as computeToolUseMetrics } from '../baseline-4R5deP0N.js';

package/dist/pipelines/index.js

@@ -2,13 +2,13 @@ import {
   compareToBaseline,
   computeToolUseMetrics,
   failureClusterView
-} from "../chunk-JLZQWFV3.js";
+} from "../chunk-K33INZHH.js";
 import {
   buildTrajectory
 } from "../chunk-RZTMDUO7.js";
 import {
   interRaterReliability
-} from "../chunk-I4MBDTY5.js";
+} from "../chunk-R5UQJNKC.js";
 import {
   aggregateLlm,
   argHash,

package/dist/{release-report-BNgMdqPF.d.ts → release-report-CCQqnK46.d.ts}

@@ -1,5 +1,5 @@
 import { D as DatasetSplit, b as DatasetManifest, a as DatasetScenario } from './dataset-CiK_3LDr.js';
-import { a3 as GateDecision, A as ActionableSideInfo, m as MultiShotTrialResult } from './summary-report-C7VPYEj2.js';
+import { a3 as GateDecision, A as ActionableSideInfo, m as MultiShotTrialResult } from './summary-report-Dl4akLKX.js';
 import { R as RunRecord, a as RunSplitTag } from './run-record-CqzahIbx.js';
 
 /**

package/dist/{replay-BL96gCEP.d.ts → replay-D7z0J43-.d.ts}

@@ -107,11 +107,10 @@ declare function redactValue(value: unknown, rules?: RedactionRule[], report?: R
  * Replay-from-raw-events — turn every captured campaign run into a
  * re-runnable artifact.
  *
- * The premise: 0.21 made `RawProviderSink` capture every provider HTTP
- * envelope. 0.22's `runEvalCampaign` makes capture the default. Together
- * they mean every past run is a complete fingerprint of what happened on
- * the wire — and that fingerprint is enough to replay the run without
- * burning new LLM cost.
+ * `RawProviderSink` captures every provider HTTP envelope; `runEvalCampaign`
+ * makes that capture the default. Together they make every past run a
+ * complete fingerprint of what happened on the wire — enough to replay
+ * the run without burning new LLM cost.
  *
  * Three use cases this primitive enables:
  *

package/dist/reporting.d.ts

@@ -1,10 +1,10 @@
 export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-C0uDYwG6.js';
-export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, P as PairedBootstrapOptions, b as PairedBootstrapResult, R as ReleaseConfidenceAxis, c as ReleaseConfidenceAxisName, d as ReleaseConfidenceInput, e as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceStatus, i as ReleaseConfidenceThresholds, j as ReleaseTraceEvidence, k as RenderReleaseReportOptions, V as Verdict, l as assertReleaseConfidence, m as bhAdjust, n as bootstrapCi, o as evaluateReleaseConfidence, p as judgeReplayGate, q as pairedBootstrap, r as pairedWilcoxon, s as releaseTraceEvidenceFromMultiShotTrials, t as renderReleaseReport } from './release-report-BNgMdqPF.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-Dgz1n51-.js';
-export { C as GainDistributionBin, F as GainDistributionFigureSpec, H as GainDistributionOptions, J as ParetoFigureSpec, K as ParetoPoint, L as RESEARCH_REPORT_HARD_PAIR_FLOOR, N as ResearchReport, O as ResearchReportCandidate, Q as ResearchReportDecision, U as ResearchReportMethodology, W as ResearchReportOptions, X as ResearchReportRecommendation, Y as SummaryTable, Z as SummaryTableOptions, _ as SummaryTableRow, $ as gainHistogram, a0 as paretoChart, a1 as researchReport, a2 as summaryTable } from './summary-report-C7VPYEj2.js';
+export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, P as PairedBootstrapOptions, b as PairedBootstrapResult, R as ReleaseConfidenceAxis, c as ReleaseConfidenceAxisName, d as ReleaseConfidenceInput, e as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceStatus, i as ReleaseConfidenceThresholds, j as ReleaseTraceEvidence, k as RenderReleaseReportOptions, V as Verdict, l as assertReleaseConfidence, m as bhAdjust, n as bootstrapCi, o as evaluateReleaseConfidence, p as judgeReplayGate, q as pairedBootstrap, r as pairedWilcoxon, s as releaseTraceEvidenceFromMultiShotTrials, t as renderReleaseReport } from './release-report-CCQqnK46.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';
+export { C as GainDistributionBin, F as GainDistributionFigureSpec, H as GainDistributionOptions, J as ParetoFigureSpec, K as ParetoPoint, L as RESEARCH_REPORT_HARD_PAIR_FLOOR, N as ResearchReport, O as ResearchReportCandidate, Q as ResearchReportDecision, U as ResearchReportMethodology, W as ResearchReportOptions, X as ResearchReportRecommendation, Y as SummaryTable, Z as SummaryTableOptions, _ as SummaryTableRow, $ as gainHistogram, a0 as paretoChart, a1 as researchReport, a2 as summaryTable } from './summary-report-Dl4akLKX.js';
 import './run-record-CqzahIbx.js';
 import './errors-BZ9sTdz7.js';
 import './outcome-store-D6KWmYvj.js';
 import './dataset-CiK_3LDr.js';
-import './failure-cluster-C2EGSDiT.js';
+import './failure-cluster-Cw65_5FY.js';
 import './store-Db2Bv8Cf.js';

package/dist/reporting.js

@@ -5,14 +5,14 @@ import {
   judgeReplayGate,
   releaseTraceEvidenceFromMultiShotTrials,
   renderReleaseReport
-} from "./chunk-RAF443UI.js";
+} from "./chunk-DBIGN5MJ.js";
 import {
   rubricPredictiveValidity
 } from "./chunk-YRZ4M5GS.js";
 import {
   evaluateInterimReleaseConfidence,
   pairedEvalueSequence
-} from "./chunk-NU65VQ7M.js";
+} from "./chunk-MAZ26DC7.js";
 import {
   RESEARCH_REPORT_HARD_PAIR_FLOOR,
   bhAdjust,
@@ -22,9 +22,9 @@ import {
   paretoChart,
   researchReport,
   summaryTable
-} from "./chunk-2A5XJB43.js";
-import "./chunk-I4MBDTY5.js";
-import "./chunk-4F5DQN55.js";
+} from "./chunk-5AKPEK5L.js";
+import "./chunk-R5UQJNKC.js";
+import "./chunk-VSMTAMNK.js";
 import "./chunk-NG236HPC.js";
 import "./chunk-PZ5AY32C.js";
 export {

package/dist/{researcher-BPT8x_NT.d.ts → researcher-G81CWc0q.d.ts}

@@ -1,7 +1,7 @@
 import { A as AgentEvalError, C as CaptureIntegrityError } from './errors-BZ9sTdz7.js';
 import { R as RawProviderSink, P as ProviderRedactor, a as RunIntegrityExpectations, b as RunIntegrityReport } from './integrity-DK2EBVZC.js';
 import { a as RunSplitTag, b as RunTokenUsage, c as RunJudgeMetadata, R as RunRecord } from './run-record-CqzahIbx.js';
-import { W as ResearchReportOptions, N as ResearchReport, a3 as GateDecision } from './summary-report-C7VPYEj2.js';
+import { W as ResearchReportOptions, N as ResearchReport, a3 as GateDecision } from './summary-report-Dl4akLKX.js';
 import { T as TraceEmitter, R as RunCompleteHook } from './emitter-DP_cSSiw.js';
 import { T as TraceStore } from './store-Db2Bv8Cf.js';
 
@@ -224,16 +224,15 @@ declare class LlmClient {
  * EvalCampaign — opinionated matrix runner that wires the four
  * capture-integrity directives by construction.
  *
- * Every consumer that ran a launch-grade benchmark before 0.22 reinvented
- * the same shape: matrix runner → for each (variant, scenario, seed) →
- * start a TraceEmitter → call LLMs → end the run → maybe analyze.
- * The bug class blueprint-agent reported (raw events not captured, route
- * silently wrong, integrity not asserted, analyst never ran) lives at the
- * integration boundary — not the agent-eval API surface. The four
- * directives in `SKILL.md § Capture integrity` are mitigations.
+ * The canonical benchmark shape — matrix runner → for each
+ * (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the
+ * run → analyze — has a bug class at the integration boundary: raw
+ * events not captured, route silently wrong, integrity not asserted,
+ * analyst never run. The directives in `SKILL.md § Capture integrity`
+ * are the mitigations.
  *
- * `EvalCampaign` is the structural fix. Consumers don't wire the integrity
- * surface anymore; the campaign owns it. Specifically, the campaign:
+ * `EvalCampaign` is the structural fix — consumers don't wire the
+ * integrity surface themselves; the campaign owns it. Specifically:
  *
  *   - calls `assertLlmRoute` once at preflight before any work runs
  *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories

package/dist/rl.d.ts

@@ -1,14 +1,14 @@
 import { R as RunRecord, a as RunSplitTag } from './run-record-CqzahIbx.js';
-import { V as VerificationReport } from './multi-layer-verifier-LkP3LVKj.js';
-import { t as TrialResult, V as VariantAggregate, q as PromptEvolutionResult, e as MultiShotOptimizationResult } from './summary-report-C7VPYEj2.js';
+import { V as VerificationReport } from './multi-layer-verifier-U-c8ge1k.js';
+import { t as TrialResult, V as VariantAggregate, q as PromptEvolutionResult, e as MultiShotOptimizationResult } from './summary-report-Dl4akLKX.js';
 import { O as OutcomeStore } from './outcome-store-D6KWmYvj.js';
 import { b as RubricPredictiveValidityReport } from './rubric-predictive-validity-C0uDYwG6.js';
-import { I as InterimReleaseConfidence } from './sequential-Dgz1n51-.js';
+import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
 import { S as Span, T as TraceStore } from './store-Db2Bv8Cf.js';
-import { R as Researcher, l as FailureMode, S as SteeringChange, j as ExperimentPlan, k as ExperimentResult, i as EvalCampaignResult, E as EvalCampaignOptions } from './researcher-BPT8x_NT.js';
-export { r as runEvalCampaign } from './researcher-BPT8x_NT.js';
+import { R as Researcher, l as FailureMode, S as SteeringChange, j as ExperimentPlan, k as ExperimentResult, i as EvalCampaignResult, E as EvalCampaignOptions } from './researcher-G81CWc0q.js';
+export { r as runEvalCampaign } from './researcher-G81CWc0q.js';
 import './errors-BZ9sTdz7.js';
-import './failure-cluster-C2EGSDiT.js';
+import './failure-cluster-Cw65_5FY.js';
 import './integrity-DK2EBVZC.js';
 import './emitter-DP_cSSiw.js';
 
@@ -529,17 +529,12 @@ declare function toAnthropicFormat(triples: PreferenceTriple[]): Array<{
 }>;
 
 /**
- * Adapters: convert legacy optimization outputs into the canonical
- * `RunRecord[]` artifact that 0.22+ primitives consume.
+ * Adapters: convert `TrialResult[]` (from `runMultiShotOptimization`,
+ * `runPromptEvolution`) into the canonical `RunRecord[]` artifact that
+ * `replayCache`, `pairedEvalueSequence`, and `rubricPredictiveValidity`
+ * consume.
  *
- * The 0.22 release standardized the campaign artifact: every cell of an
- * eval matrix produces one `RunRecord`. The pre-0.22 optimization
- * primitives (`runMultiShotOptimization`, `runPromptEvolution`) produce
- * `TrialResult[]` with a different shape. This file bridges the two so
- * the new primitives (`replayCache`, `pairedEvalueSequence`,
- * `rubricPredictiveValidity`) compose cleanly with the existing RL stack.
- *
- * The adapters are thin and explicit — every mandatory `RunRecord` field
+ * Adapters are thin and explicit — every mandatory `RunRecord` field
  * comes from a caller-supplied context (`commitSha`, `model`,
  * `promptHash`, `configHash`) plus the trial's runtime data. Defaults
  * exist for fields the trial doesn't carry (`tokenUsage`, `costUsd`),
@@ -1505,18 +1500,16 @@ interface DetectRewardHackingInput {
 declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport;
 
 /**
- * `analyzeOptimizationResult` — unifies the pre-0.22 auto-research stack
+ * `analyzeOptimizationResult` — unifies the auto-research stack
  * (`runPromptEvolution`, `runMultiShotOptimization`, reflective-mutation,
- * Ax/AxRLM trace analyst) with the 0.23 RL bridge in a single call.
+ * Ax/AxRLM trace analyst) with the RL bridge in a single call.
  *
- * What this fixes: until 0.23 the optimization stack and the RL bridge
- * lived in parallel namespaces. The optimization primitives produced
- * `TrialResult[]`; the RL bridge consumed `RunRecord[]`. Trace-analyst
- * was decoupled from both. `analyzeOptimizationResult` does the wiring
- * once so consumers don't have to:
+ * The optimization primitives produce `TrialResult[]`; the RL bridge
+ * consumes `RunRecord[]`. Trace-analyst is independent of both. This
+ * function does the wiring once so consumers don't have to:
  *
- *    Optimization (existing primitives)           RL bridge (0.23)
- *    ──────────────────────────────────           ────────────────
+ *    Optimization (existing primitives)           RL bridge
+ *    ──────────────────────────────────           ────────
  *    runPromptEvolution → TrialResult[]    →
  *    runMultiShotOptimization → MSTrial[]  → analyzeOptimizationResult →
  *    reflective-mutation → mutations.jsonl →                             ↓
@@ -1527,10 +1520,10 @@ declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHac
  *    ↓                                                                   │
  *    TraceAnalyst.analyze(progressLog)         ←─────────────────────────┘
  *
- * The output of this function is the canonical RL artifact set:
- * `RunRecord[]` (so every other 0.22+ primitive composes), preference
- * triples, verifiable reward signals, reward-hacking diagnosis,
- * sequential interim verdict, and (when wired) trace-analyst summary.
+ * The output is the canonical RL artifact set: `RunRecord[]` (so every
+ * other RL primitive composes), preference triples, verifiable reward
+ * signals, reward-hacking diagnosis, sequential interim verdict, and
+ * (when wired) trace-analyst summary.
  *
  * What this primitive does NOT do: it does not modify the optimization
  * primitives' internals. They keep producing `TrialResult` and emitting
@@ -1609,11 +1602,7 @@ declare function analyzeOptimizationResult(opts: AnalyzeOptimizationResultOption
  * `PredictiveValidityResearcher` — concrete `Researcher` implementation
  * that drives selection from outcome-anchored predictive validity.
  *
- * `Researcher` was a placeholder interface plus `NoopResearcher` until
- * 0.23. The 0.23 panel critique called this out: shipping the interface
- * without a default implementation that drives the loop is incomplete.
- *
- * This researcher answers each method:
+ * Each method:
  *
  *   - `inspectFailures(runs)` — synthesizes failure modes from the
  *     bottom-quartile of `RunRecord`s on the configured proxy reward.
@@ -1676,14 +1665,10 @@ declare class PredictiveValidityResearcher implements Researcher {
 }
 
 /**
- * `runRLCampaign` — the missing top-level orchestrator.
+ * `runRLCampaign` — top-level orchestrator that runs the matrix and
+ * produces every RL-ready artifact in one call.
  *
- * `runEvalCampaign` runs the matrix and produces `RunRecord[]`. The 0.23
- * RL primitives consume that artifact in different ways. Until 0.24 they
- * had to be wired together by hand at every consumer; that defeats the
- * cohesion the package is supposed to provide.
- *
- * `runRLCampaign` wires:
+ * Wires:
  *   1. `runEvalCampaign` for the matrix run (capture, integrity, hooks)
  *   2. `extractVerifiableReward` over each run, separating deterministic
  *      from probabilistic reward sources for the trainer
@@ -1697,9 +1682,6 @@ declare class PredictiveValidityResearcher implements Researcher {
  * stage's output is in there. The consumer's downstream fits in a single
  * line: pass `result.preferences` to their DPO trainer, `result.grpoRows`
  * to GRPO, `result.runs` plus `result.rewardSignals` to a custom RL loop.
- *
- * This is what the 0.23 panel critique called the "missing top-level
- * primitive." Now shipped.
  */
 
 interface RunRLCampaignOptions<V> extends EvalCampaignOptions<V> {

package/dist/rl.js

@@ -1,23 +1,23 @@
 import {
   runEvalCampaign
-} from "./chunk-SESZDQPX.js";
+} from "./chunk-RUI6SIHY.js";
 import "./chunk-4S4BM3QQ.js";
 import {
   rubricPredictiveValidity
 } from "./chunk-YRZ4M5GS.js";
 import {
   evaluateInterimReleaseConfidence
-} from "./chunk-NU65VQ7M.js";
+} from "./chunk-MAZ26DC7.js";
 import {
   benjaminiHochberg
-} from "./chunk-2A5XJB43.js";
+} from "./chunk-5AKPEK5L.js";
 import {
   wilcoxonSignedRank
-} from "./chunk-I4MBDTY5.js";
+} from "./chunk-R5UQJNKC.js";
 import "./chunk-KTGTIOFD.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-TVVP3ZZQ.js";
-import "./chunk-4F5DQN55.js";
+import "./chunk-VSMTAMNK.js";
 import {
   ValidationError
 } from "./chunk-NG236HPC.js";