@tangle-network/agent-eval 0.95.0 → 0.96.0

package/dist/release-report-pidWUMZ2.d.ts

@@ -0,0 +1,233 @@
+import { D as DatasetSplit, c as DatasetManifest, a as DatasetScenario } from './dataset-BbGkaN2I.js';
+import { m as GateDecision } from './summary-report-CInXwsza.js';
+import { R as RunRecord, a as RunSplitTag } from './run-record-CP2ObebC.js';
+
+/**
+ * Release confidence gate.
+ *
+ * This is the production-facing composition layer over the lower-level
+ * primitives:
+ *   - Dataset manifests prove corpus/version coverage.
+ *   - RunRecord rows prove reproducible search/holdout outcomes.
+ *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.
+ *   - HeldOutGate decisions remain the paired promotion authority.
+ *
+ * The gate is intentionally pure and conservative. Missing declared evidence
+ * fails closed instead of being treated as a neutral zero.
+ */
+
+/** Severity of an actionable finding attached to a run/trace. */
+type AsiSeverity = 'info' | 'warning' | 'error' | 'critical';
+/** Actionable side-info — a diagnosed finding the loop can act on. */
+interface ActionableSideInfo {
+    /** Stable expectation/check id when available. */
+    expectationId?: string;
+    /** Human-readable diagnosis of what happened. */
+    message: string;
+    severity?: AsiSeverity;
+    /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */
+    evidence?: string;
+    /** Prompt/tool/context surface likely responsible. */
+    responsibleSurface?: string;
+    /** Suggested fix in natural language. */
+    suggestion?: string;
+    /** Whether this expectation was satisfied. Defaults to false for ASI rows. */
+    matched?: boolean;
+    metadata?: Record<string, unknown>;
+}
+type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail';
+type ReleaseConfidenceAxisName = 'corpus' | 'quality' | 'generalization' | 'diagnostics' | 'efficiency';
+interface ReleaseTraceEvidence {
+    scenarioId: string;
+    candidateId?: string;
+    split?: RunSplitTag;
+    score?: number;
+    ok?: boolean;
+    turnCount?: number;
+    costUsd?: number;
+    durationMs?: number;
+    failureMode?: string;
+    asi?: ActionableSideInfo[];
+    metadata?: Record<string, unknown>;
+}
+interface ReleaseConfidenceThresholds {
+    /** Require a Dataset manifest or explicit scenarios. Default true. */
+    requireCorpus?: boolean;
+    minScenarioCount?: number;
+    minSearchRuns?: number;
+    minHoldoutRuns?: number;
+    /** Require at least one holdout scenario/run. Default true. */
+    requireHoldout?: boolean;
+    minPassRate?: number;
+    minMeanScore?: number;
+    /** Search mean may exceed holdout mean by at most this much. */
+    maxOverfitGap?: number;
+    maxMeanCostUsd?: number;
+    maxP95WallMs?: number;
+    /** Low-score/failed rows must carry ASI. Default true. */
+    requireAsiForFailures?: boolean;
+    /** Score below this is considered a failure for ASI coverage. Default 0.5. */
+    failureScoreThreshold?: number;
+}
+interface ReleaseConfidenceInput {
+    target: string;
+    candidateId?: string;
+    baselineId?: string;
+    dataset?: DatasetManifest;
+    scenarios?: readonly DatasetScenario[];
+    runs?: readonly RunRecord[];
+    traces?: readonly ReleaseTraceEvidence[];
+    gateDecision?: GateDecision | null;
+    thresholds?: ReleaseConfidenceThresholds;
+}
+interface ReleaseConfidenceAxis {
+    name: ReleaseConfidenceAxisName;
+    status: ReleaseConfidenceStatus;
+    score: number;
+    detail: string;
+}
+interface ReleaseConfidenceIssue {
+    axis: ReleaseConfidenceAxisName;
+    severity: 'critical' | 'warning';
+    code: string;
+    detail: string;
+}
+interface ReleaseConfidenceMetrics {
+    scenarioCount: number;
+    searchRuns: number;
+    holdoutRuns: number;
+    passRate: number;
+    meanScore: number;
+    searchMeanScore: number;
+    holdoutMeanScore: number;
+    overfitGap: number;
+    meanCostUsd: number;
+    p95WallMs: number;
+    failedRows: number;
+    failuresWithAsi: number;
+    singleShotTraces: number;
+    multiShotTraces: number;
+    splitCounts: Record<DatasetSplit, number>;
+    domainCounts: Record<string, number>;
+    failureModeCounts: Record<string, number>;
+    responsibleSurfaceCounts: Record<string, number>;
+}
+interface ReleaseConfidenceScorecard {
+    target: string;
+    candidateId: string | null;
+    baselineId: string | null;
+    status: ReleaseConfidenceStatus;
+    promote: boolean;
+    axes: ReleaseConfidenceAxis[];
+    issues: ReleaseConfidenceIssue[];
+    metrics: ReleaseConfidenceMetrics;
+    dataset: DatasetManifest | null;
+    gateDecision: GateDecision | null;
+    summary: string;
+}
+declare function evaluateReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard;
+declare function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard;
+
+/**
+ * Bootstrap-CI promotion gate.
+ *
+ * In any iterative-improvement loop (GEPA, prompt evolution, dataset
+ * curation), the question is "did this generation actually improve, or are
+ * we celebrating noise?". With small N and noisy outcomes, point-estimate
+ * deltas lie. Bootstrap confidence intervals tell the operator whether the
+ * delta is real before code or prompts get promoted.
+ *
+ * This module is pure functions — no I/O, no model calls. Easy to unit-test
+ * and to compose into any verdict gate.
+ *
+ * Default gate:
+ *   - Bootstrap mean baseline vs candidate (1k resamples).
+ *   - Compute the delta distribution; pass if the lower CI bound > 0.
+ *   - Tunable confidence (default 95%) and resample count.
+ *
+ * Verdict semantics intentionally match the existing `experiments.jsonl`
+ * vocabulary:
+ *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)
+ *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)
+ *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)
+ *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal
+ */
+type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE';
+interface BootstrapResult {
+    baselineMean: number;
+    candidateMean: number;
+    /** candidateMean - baselineMean, point estimate. */
+    delta: number;
+    /** Lower bound of the (1 - alpha) CI on the delta. */
+    ciLower: number;
+    /** Upper bound of the (1 - alpha) CI on the delta. */
+    ciUpper: number;
+    /** Number of bootstrap resamples used. */
+    iterations: number;
+    alpha: number;
+    verdict: Verdict;
+}
+interface BootstrapOptions {
+    /** Confidence level alpha (default 0.05 → 95% CI). */
+    alpha?: number;
+    /** Number of resamples (default 1000). */
+    iterations?: number;
+    /**
+     * Minimum total samples (baseline + candidate) below which we always
+     * return INCONCLUSIVE — bootstrap with too few samples is meaningless.
+     * Default 6 (combined).
+     */
+    minTotalSamples?: number;
+    /** RNG seed for reproducibility. Default: Math.random. */
+    seed?: number;
+}
+/**
+ * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.
+ *
+ * Uses simple percentile bootstrap on the difference of resampled means.
+ * That's the standard non-parametric primitive — no distributional
+ * assumptions, robust to skew, easy to reason about.
+ */
+declare function bootstrapCi(baseline: number[], candidate: number[], options?: BootstrapOptions): BootstrapResult;
+/**
+ * Judge-replay promotion gate.
+ *
+ * The cheap inner-loop judge that drives an evolution run is by definition
+ * fast and noisy. When you're about to promote a winning variant to the
+ * canonical default, you want a STRONGER judge (a more expensive model, a
+ * human grader, a separately-trained reward model) to confirm the win
+ * generalises beyond the inner loop.
+ *
+ * This helper takes raw winner + baseline outputs, scores both through the
+ * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger
+ * judge agrees the winner is real with the configured confidence. Doesn't
+ * matter what shape your "output" is — pass a string, an object, anything
+ * the judge can read.
+ */
+interface JudgeReplayGateArgs<TOutput> {
+    baselineOutputs: TOutput[];
+    candidateOutputs: TOutput[];
+    /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */
+    judge: (output: TOutput) => Promise<number> | number;
+    alpha?: number;
+    iterations?: number;
+    /** RNG seed for reproducibility. */
+    seed?: number;
+    /** Maximum concurrent judge calls. Default 4. */
+    judgeConcurrency?: number;
+}
+declare function judgeReplayGate<TOutput>(args: JudgeReplayGateArgs<TOutput>): Promise<BootstrapResult & {
+    baselineSamples: number;
+    candidateSamples: number;
+}>;
+
+interface RenderReleaseReportOptions {
+    title?: string;
+    runs?: readonly RunRecord[];
+    comparator?: string;
+    traceAnalystFindings?: readonly string[];
+    nextActions?: readonly string[];
+}
+declare function renderReleaseReport(scorecard: ReleaseConfidenceScorecard, options?: RenderReleaseReportOptions): string;
+
+export { type ActionableSideInfo as A, type BootstrapOptions as B, type JudgeReplayGateArgs as J, type ReleaseConfidenceAxis as R, type Verdict as V, type BootstrapResult as a, type ReleaseConfidenceAxisName as b, type ReleaseConfidenceInput as c, type ReleaseConfidenceIssue as d, type ReleaseConfidenceMetrics as e, type ReleaseConfidenceScorecard as f, type ReleaseConfidenceStatus as g, type ReleaseConfidenceThresholds as h, type ReleaseTraceEvidence as i, type RenderReleaseReportOptions as j, assertReleaseConfidence as k, bootstrapCi as l, evaluateReleaseConfidence as m, judgeReplayGate as n, type AsiSeverity as o, renderReleaseReport as r };

package/dist/reporting.d.ts

@@ -1,15 +1,16 @@
-export type { RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, } from './meta-eval/rubric-predictive-validity';
-export { rubricPredictiveValidity } from './meta-eval/rubric-predictive-validity';
-export type { BootstrapOptions, BootstrapResult, JudgeReplayGateArgs, Verdict, } from './promotion-gate';
-export { bootstrapCi, judgeReplayGate, } from './promotion-gate';
-export type { ReleaseConfidenceAxis, ReleaseConfidenceAxisName, ReleaseConfidenceInput, ReleaseConfidenceIssue, ReleaseConfidenceMetrics, ReleaseConfidenceScorecard, ReleaseConfidenceStatus, ReleaseConfidenceThresholds, ReleaseTraceEvidence, } from './release-confidence';
-export { assertReleaseConfidence, evaluateReleaseConfidence } from './release-confidence';
-export type { RenderReleaseReportOptions } from './release-report';
-export { renderReleaseReport } from './release-report';
-export type { InterimReleaseConfidence, InterimReleaseConfidenceInput, PairedEvalueOptions, PairedEvalueSequence, PairedEvalueStep, SequentialDecision, } from './sequential';
-export { evaluateInterimReleaseConfidence, pairedEvalueSequence, } from './sequential';
-export type { PairedBootstrapOptions, PairedBootstrapResult, } from './statistics';
-export { benjaminiHochberg, pairedBootstrap, wilcoxonSignedRank, } from './statistics';
-export type { GainDistributionBin, GainDistributionFigureSpec, GainDistributionOptions, ParetoFigureSpec, ParetoPoint, ResearchReport, ResearchReportCandidate, ResearchReportDecision, ResearchReportMethodology, ResearchReportOptions, ResearchReportRecommendation, SummaryTable, SummaryTableOptions, SummaryTableRow, } from './summary-report';
-export { gainHistogram, paretoChart, RESEARCH_REPORT_HARD_PAIR_FLOOR, researchReport, summaryTable, } from './summary-report';
-//# sourceMappingURL=reporting.d.ts.map
+export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-C2hDKM8Z.js';
+export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, f as ReleaseConfidenceScorecard, g as ReleaseConfidenceStatus, h as ReleaseConfidenceThresholds, 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-pidWUMZ2.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 { P as PairedBootstrapOptions, a as PairedBootstrapResult, b as benjaminiHochberg, p as pairedBootstrap, w as wilcoxonSignedRank } from './statistics-CCJpTGOS.js';
+export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, 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-CInXwsza.js';
+import './run-record-CP2ObebC.js';
+import '@tangle-network/agent-interface';
+import './errors-CzMUYo7b.js';
+import './schema-m0gsnbt3.js';
+import './outcome-store-rnXLEqSn.js';
+import './dataset-BbGkaN2I.js';
+import './judge-calibration-0p2QcWNE.js';
+import './types-C7DGg5ex.js';
+import '@tangle-network/tcloud';
+import './failure-cluster-DH9Flgcf.js';
+import './store-BcFXE6LG.js';

package/dist/{eval-campaign.d.ts → researcher-Jr8ME1dZ.d.ts}

@@ -1,3 +1,12 @@
+import { a as RunSplitTag, b as RunTokenUsage, c as RunJudgeMetadata, J as JudgeScoresRecord, A as AgentProfileCell, d as AgentProfileCellInput, R as RunRecord } from './run-record-CP2ObebC.js';
+import { L as LlmClientOptions, a as LlmRouteRequirements } from './llm-client-Bj7g0rqu.js';
+import { h as ResearchReportOptions, d as ResearchReport, m as GateDecision } from './summary-report-CInXwsza.js';
+import { T as TraceEmitter, R as RunCompleteHook } from './emitter-C2rqGH_l.js';
+import { R as RunIntegrityExpectations, a as RunIntegrityReport } from './integrity-D2t12mMw.js';
+import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
+import { F as FailureClass } from './schema-m0gsnbt3.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
+
 /**
  * EvalCampaign — opinionated matrix runner that wires the four
  * capture-integrity directives by construction.
@@ -37,26 +46,17 @@
  *   - Resume from partial state across crashes
  *   - LLM-call retry beyond what `LlmClient` already does
  */
-import { type AgentProfileCell, type AgentProfileCellInput } from './agent-profile-cell';
-import { type LlmClientOptions, type LlmRouteRequirements } from './llm-client';
-import type { JudgeScoresRecord, RunJudgeMetadata, RunRecord, RunSplitTag, RunTokenUsage } from './run-record';
-import { type ResearchReport, type ResearchReportOptions } from './summary-report';
-import type { RunCompleteHook } from './trace/emitter';
-import { TraceEmitter } from './trace/emitter';
-import { type RunIntegrityExpectations, type RunIntegrityReport } from './trace/integrity';
-import { type RawProviderSink } from './trace/raw-provider-sink';
-import type { FailureClass } from './trace/schema';
-import type { TraceStore } from './trace/store';
-export interface CampaignVariant<V> {
+
+interface CampaignVariant<V> {
     id: string;
     payload: V;
 }
-export interface CampaignScenario {
+interface CampaignScenario {
     scenarioId: string;
     /** Free-form metadata propagated to runs and reports. */
     tags?: Record<string, string>;
 }
-export interface CampaignRunContext<V> {
+interface CampaignRunContext<V> {
     /** Stable run id. The campaign generates this; the runner does not. */
     runId: string;
     /** Logical experiment id (campaignId by default; overridable per-run via opts). */
@@ -83,7 +83,7 @@ export interface CampaignRunContext<V> {
      */
     llmOpts: LlmClientOptions;
 }
-export interface CampaignRunOutcome {
+interface CampaignRunOutcome {
     /** Did the run pass? Mirrors `RunOutcome.pass` semantics. */
     pass: boolean;
     /** Score for the run on its split. Maps to `searchScore` or `holdoutScore`. */
@@ -120,9 +120,9 @@ export interface CampaignRunOutcome {
      */
     agentProfile?: AgentProfileCell | AgentProfileCellInput;
 }
-export type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>;
-export type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log';
-export interface EvalCampaignOptions<V> {
+type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>;
+type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log';
+interface EvalCampaignOptions<V> {
     /**
      * Stable id for the campaign. Used as the default `experimentId` on
      * every run, and folded into the campaign fingerprint.
@@ -215,14 +215,14 @@ export interface EvalCampaignOptions<V> {
         scenarioTags: Record<string, string>;
     }) => AgentProfileCell | AgentProfileCellInput | Promise<AgentProfileCell | AgentProfileCellInput>);
 }
-export interface CampaignFactoryParams {
+interface CampaignFactoryParams {
     campaignId: string;
     runId: string;
     variantId: string;
     scenarioId: string;
     seed: number;
 }
-export interface FailedRun {
+interface FailedRun {
     runId: string;
     variantId: string;
     scenarioId: string;
@@ -230,7 +230,7 @@ export interface FailedRun {
     reason: string;
     error?: string;
 }
-export interface EvalCampaignResult {
+interface EvalCampaignResult {
     campaignId: string;
     /** SHA-256 over canonicalised `(variantIds, scenarioIds, seeds, comparator, splitTag, baseUrl, provider, preregistrationHash)`. */
     campaignFingerprint: string;
@@ -245,21 +245,143 @@ export interface EvalCampaignResult {
     startedAt: string;
     endedAt: string;
 }
-export declare function runEvalCampaign<V>(opts: EvalCampaignOptions<V>): Promise<EvalCampaignResult>;
+declare function runEvalCampaign<V>(opts: EvalCampaignOptions<V>): Promise<EvalCampaignResult>;
+
 /**
- * Abort a run whose owning work threw or was orphaned by a sibling's genuine
- * error, finalizing the emitter so hooks fire and the store records a terminal
- * status. Safe to call unconditionally: it only aborts runs that are still
- * `running`. Two no-op cases are intentional and benign:
+ * Researcher interface — stable hook for an external autonomous-research
+ * agent to drive the meta-loop.
  *
- *   - the run was never started (absent from the store) — nothing to finalize
- *   - the run is already terminal (`completed` / `failed` / `aborted`) —
- *     finalizing again would overwrite the real outcome (e.g. flip a passed run
- *     to `aborted` with `{ pass: false, notes: reason }`), so we leave it alone
+ * Implementations live downstream (typically in a private repo that
+ * runs the actual LLM). This package ships only the contract + a
+ * `NoopResearcher` so consumers can wire the surface without being
+ * forced to implement every method up front.
  *
- * Any OTHER failure of the store read or the abort write (disk full, FS fault,
- * backend down) is a genuine diagnostic and propagates rather than being
- * swallowed.
+ * The four methods mirror the four stages of the paper "Two Loops,
+ * Three Roles":
+ *
+ *   inspectFailures   — given the observed runs, what failure modes
+ *                       are present? (data → diagnosis)
+ *   proposeChange     — given diagnosed failure modes, what
+ *                       structural changes should we try?
+ *                       (diagnosis → plan delta)
+ *   applyChange       — fold the proposed deltas into a concrete
+ *                       experiment plan against an existing baseline.
+ *                       (plan delta → executable plan)
+ *   evaluateChange    — run the plan, return runs + the gate verdict.
+ *                       (executable plan → verdict)
+ *
+ * Composition is the discipline: a Researcher implementation MUST
+ * keep these four steps separate and inspectable. Conflating
+ * "diagnose + propose + run" into a single LLM call defeats the
+ * point of the framework — you can't audit which step lied.
+ *
+ * THIS INTERFACE IS STABLE. Breaking changes require a new module
+ * (e.g. `Researcher2`) so existing implementations keep working.
  */
-export declare function finalizeAbort(emitter: TraceEmitter, runId: string, reason: string): Promise<void>;
-//# sourceMappingURL=eval-campaign.d.ts.map
+
+/** A diagnosed failure mode with the run-IDs that exhibit it. */
+interface FailureMode {
+    /** Short machine-readable code. Must be stable across runs of the
+     *  same researcher to enable longitudinal tracking. */
+    code: string;
+    /** Human-readable description for the paper / dashboard. */
+    description: string;
+    evidence: {
+        /** Run IDs (from `RunRecord.runId`) where this failure mode was
+         *  observed. */
+        runIds: string[];
+        /** Number of run samples that informed the diagnosis. */
+        samples: number;
+    };
+}
+/** A single steering change the researcher wants to try. */
+interface SteeringChange {
+    kind: 'reviewer_prompt' | 'skill_add' | 'skill_remove' | 'threshold' | 'budget';
+    /** Implementation-specific payload. Researcher implementations
+     *  define the schema — keep this `unknown` here to avoid coupling
+     *  the public interface to any one researcher's internal model. */
+    payload: unknown;
+    /** Why the researcher proposed this change. Goes into the audit
+     *  trail next to the failure-mode evidence. */
+    rationale: string;
+    /** Optional self-reported expected delta on the headline metric. */
+    expectedDelta?: number;
+}
+/** A single experiment plan, mapped onto the search/holdout splits. */
+interface ExperimentPlan {
+    baselineCandidateId: string;
+    proposedCandidateId: string;
+    changes: SteeringChange[];
+    /** USD ceiling for the entire experiment. The runner must stop
+     *  before exceeding this and report a partial result. */
+    evaluationBudgetUsd: number;
+    /** Item IDs (your dataset keys) for the search vs holdout splits. */
+    splits: {
+        search: string[];
+        holdout: string[];
+    };
+}
+/** Result of running a plan: every run, plus the gate verdict. */
+interface ExperimentResult {
+    plan: ExperimentPlan;
+    runs: RunRecord[];
+    gateDecision: GateDecision;
+}
+/**
+ * The researcher loop. Stable, four-step, inspectable.
+ *
+ *   ┌──────────┐  inspectFailures  ┌──────────┐  proposeChange ┌──────────┐
+ *   │   runs   │ ─────────────────▶│ failures │ ──────────────▶│ changes  │
+ *   └──────────┘                   └──────────┘                └────┬─────┘
+ *                                                                   │
+ *                                                                   ▼
+ *                              ┌────────────────┐  applyChange ┌────────┐
+ *                              │ ExperimentPlan │ ◀────────────│  base  │
+ *                              └────────┬───────┘              └────────┘
+ *                                       │
+ *                       evaluateChange  ▼
+ *                              ┌────────────────┐
+ *                              │ ExperimentResult│
+ *                              └────────────────┘
+ */
+interface Researcher {
+    inspectFailures(runs: RunRecord[]): Promise<FailureMode[]>;
+    proposeChange(failures: FailureMode[]): Promise<SteeringChange[]>;
+    applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan>;
+    evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult>;
+}
+interface CallbackResearcherOptions {
+    inspectFailures: Researcher['inspectFailures'];
+    proposeChange: Researcher['proposeChange'];
+    applyChange: Researcher['applyChange'];
+    evaluateChange: Researcher['evaluateChange'];
+}
+/**
+ * Minimal concrete researcher for tests, scripts, and small integrations.
+ * Larger autonomous researchers can still implement `Researcher` directly.
+ */
+declare class CallbackResearcher implements Researcher {
+    private readonly callbacks;
+    constructor(callbacks: CallbackResearcherOptions);
+    inspectFailures(runs: RunRecord[]): Promise<FailureMode[]>;
+    proposeChange(failures: FailureMode[]): Promise<SteeringChange[]>;
+    applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan>;
+    evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult>;
+}
+/**
+ * No-op researcher — fails loud on every method. Use as a placeholder
+ * in code paths that wire the interface but don't have an implementation
+ * yet. Importantly, this does NOT silently succeed: a no-op researcher
+ * that returned empty arrays would muffle the loop's signal that
+ * nobody implemented the brain.
+ */
+declare class NoopResearcher implements Researcher {
+    private readonly hint;
+    constructor(hint?: string);
+    inspectFailures(_runs: RunRecord[]): Promise<FailureMode[]>;
+    proposeChange(_failures: FailureMode[]): Promise<SteeringChange[]>;
+    applyChange(_changes: SteeringChange[], _baseline: ExperimentPlan): Promise<ExperimentPlan>;
+    evaluateChange(_plan: ExperimentPlan): Promise<ExperimentResult>;
+}
+
+export { CallbackResearcher as C, type ExperimentPlan as E, type FailureMode as F, NoopResearcher as N, type Researcher as R, type SteeringChange as S, type ExperimentResult as a, type EvalCampaignResult as b, type EvalCampaignOptions as c, type CallbackResearcherOptions as d, type CampaignFactoryParams as e, type CampaignIntegrityPolicy as f, type CampaignRunContext as g, type CampaignRunOutcome as h, type CampaignRunner as i, type CampaignScenario as j, type CampaignVariant as k, type FailedRun as l, runEvalCampaign as r };