@tangle-network/agent-eval 0.95.0 → 0.95.1

package/dist/release-confidence.d.ts

@@ -1,128 +0,0 @@
-/**
- * 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.
- */
-import type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset';
-import type { GateDecision } from './held-out-gate';
-import type { RunRecord, RunSplitTag } from './run-record';
-/** Severity of an actionable finding attached to a run/trace. */
-export type AsiSeverity = 'info' | 'warning' | 'error' | 'critical';
-/** Actionable side-info — a diagnosed finding the loop can act on. */
-export 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>;
-}
-export type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail';
-export type ReleaseConfidenceAxisName = 'corpus' | 'quality' | 'generalization' | 'diagnostics' | 'efficiency';
-export 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>;
-}
-export 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;
-}
-export interface ReleaseConfidenceInput {
-    target: string;
-    candidateId?: string;
-    baselineId?: string;
-    dataset?: DatasetManifest;
-    scenarios?: readonly DatasetScenario[];
-    runs?: readonly RunRecord[];
-    traces?: readonly ReleaseTraceEvidence[];
-    gateDecision?: GateDecision | null;
-    thresholds?: ReleaseConfidenceThresholds;
-}
-export interface ReleaseConfidenceAxis {
-    name: ReleaseConfidenceAxisName;
-    status: ReleaseConfidenceStatus;
-    score: number;
-    detail: string;
-}
-export interface ReleaseConfidenceIssue {
-    axis: ReleaseConfidenceAxisName;
-    severity: 'critical' | 'warning';
-    code: string;
-    detail: string;
-}
-export 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>;
-}
-export 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;
-}
-export declare function evaluateReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard;
-export declare function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard;
-//# sourceMappingURL=release-confidence.d.ts.map

package/dist/release-confidence.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"release-confidence.d.ts","sourceRoot":"","sources":["../src/release-confidence.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,eAAe,EAAE,YAAY,EAAE,MAAM,WAAW,CAAA;AAE/E,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAA;AACnD,OAAO,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,MAAM,cAAc,CAAA;AAE1D,iEAAiE;AACjE,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,GAAG,UAAU,CAAA;AAEnE,sEAAsE;AACtE,MAAM,WAAW,kBAAkB;IACjC,kDAAkD;IAClD,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,iDAAiD;IACjD,OAAO,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,WAAW,CAAA;IACtB,wEAAwE;IACxE,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,sDAAsD;IACtD,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B,yCAAyC;IACzC,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,8EAA8E;IAC9E,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,MAAM,uBAAuB,GAAG,MAAM,GAAG,MAAM,GAAG,MAAM,CAAA;AAC9D,MAAM,MAAM,yBAAyB,GACjC,QAAQ,GACR,SAAS,GACT,gBAAgB,GAChB,aAAa,GACb,YAAY,CAAA;AAEhB,MAAM,WAAW,oBAAoB;IACnC,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,KAAK,CAAC,EAAE,WAAW,CAAA;IACnB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,EAAE,CAAC,EAAE,OAAO,CAAA;IACZ,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,GAAG,CAAC,EAAE,kBAAkB,EAAE,CAAA;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,2BAA2B;IAC1C,sEAAsE;IACtE,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,+DAA+D;IAC/D,cAAc,CAAC,EAAE,OAAO,CAAA;IACxB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,gEAAgE;IAChE,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,0DAA0D;IAC1D,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B,8EAA8E;IAC9E,qBAAqB,CAAC,EAAE,MAAM,CAAA;CAC/B;AAED,MAAM,WAAW,sBAAsB;IACrC,MAAM,EAAE,MAAM,CAAA;IACd,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,OAAO,CAAC,EAAE,eAAe,CAAA;IACzB,SAAS,CAAC,EAAE,SAAS,eAAe,EAAE,CAAA;IACtC,IAAI,CAAC,EAAE,SAAS,SAAS,EAAE,CAAA;IAC3B,MAAM,CAAC,EAAE,SAAS,oBAAoB,EAAE,CAAA;IACxC,YAAY,CAAC,EAAE,YAAY,GAAG,IAAI,CAAA;IAClC,UAAU,CAAC,EAAE,2BAA2B,CAAA;CACzC;AAED,MAAM,WAAW,qBAAqB;IACpC,IAAI,EAAE,yBAAyB,CAAA;IAC/B,MAAM,EAAE,uBAAuB,CAAA;IAC/B,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,sBAAsB;IACrC,IAAI,EAAE,yBAAyB,CAAA;IAC/B,QAAQ,EAAE,UAAU,GAAG,SAAS,CAAA;IAChC,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,wBAAwB;IACvC,aAAa,EAAE,MAAM,CAAA;IACrB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE,MAAM,CAAA;IAChB,SAAS,EAAE,MAAM,CAAA;IACjB,eAAe,EAAE,MAAM,CAAA;IACvB,gBAAgB,EAAE,MAAM,CAAA;IACxB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE,MAAM,CAAA;IACvB,gBAAgB,EAAE,MAAM,CAAA;IACxB,eAAe,EAAE,MAAM,CAAA;IACvB,WAAW,EAAE,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAA;IACzC,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACpC,iBAAiB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACzC,wBAAwB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;CACjD;AAED,MAAM,WAAW,0BAA0B;IACzC,MAAM,EAAE,MAAM,CAAA;IACd,WAAW,EAAE,MAAM,GAAG,IAAI,CAAA;IAC1B,UAAU,EAAE,MAAM,GAAG,IAAI,CAAA;IACzB,MAAM,EAAE,uBAAuB,CAAA;IAC/B,OAAO,EAAE,OAAO,CAAA;IAChB,IAAI,EAAE,qBAAqB,EAAE,CAAA;IAC7B,MAAM,EAAE,sBAAsB,EAAE,CAAA;IAChC,OAAO,EAAE,wBAAwB,CAAA;IACjC,OAAO,EAAE,eAAe,GAAG,IAAI,CAAA;IAC/B,YAAY,EAAE,YAAY,GAAG,IAAI,CAAA;IACjC,OAAO,EAAE,MAAM,CAAA;CAChB;AAiBD,wBAAgB,yBAAyB,CACvC,KAAK,EAAE,sBAAsB,GAC5B,0BAA0B,CAyE5B;AAED,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,sBAAsB,GAAG,0BAA0B,CAMjG"}

package/dist/release-report.d.ts

@@ -1,11 +0,0 @@
-import type { ReleaseConfidenceScorecard } from './release-confidence';
-import type { RunRecord } from './run-record';
-export interface RenderReleaseReportOptions {
-    title?: string;
-    runs?: readonly RunRecord[];
-    comparator?: string;
-    traceAnalystFindings?: readonly string[];
-    nextActions?: readonly string[];
-}
-export declare function renderReleaseReport(scorecard: ReleaseConfidenceScorecard, options?: RenderReleaseReportOptions): string;
-//# sourceMappingURL=release-report.d.ts.map

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

@@ -1 +0,0 @@
-{"version":3,"file":"release-report.d.ts","sourceRoot":"","sources":["../src/release-report.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,0BAA0B,EAAE,MAAM,sBAAsB,CAAA;AACtE,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,cAAc,CAAA;AAG7C,MAAM,WAAW,0BAA0B;IACzC,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,IAAI,CAAC,EAAE,SAAS,SAAS,EAAE,CAAA;IAC3B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,oBAAoB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IACxC,WAAW,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;CAChC;AAED,wBAAgB,mBAAmB,CACjC,SAAS,EAAE,0BAA0B,EACrC,OAAO,GAAE,0BAA+B,GACvC,MAAM,CAkFR"}

package/dist/replay.d.ts

@@ -1,120 +0,0 @@
-/**
- * Replay-from-raw-events — turn every captured campaign run into a
- * re-runnable artifact.
- *
- * `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:
- *
- *   1. **Post-hoc judging** — apply a new judge / rubric / scoring callback
- *      to last week's runs without re-calling any LLM. The cost of trying
- *      a new rubric drops from "another full sweep" to a CPU-bound replay.
- *   2. **Determinism audits** — replay the same campaign and verify the
- *      raw responses match byte-for-byte. Any drift is a non-determinism
- *      bug (in the harness, the prompt builder, the sandbox, …).
- *   3. **Free judge calibration** — run two judges on identical responses
- *      and measure inter-judge agreement without doubling LLM spend.
- *
- * The interface is deliberately fetch-shaped. Inject `createReplayFetch`
- * into `LlmClientOptions.fetch` and every `callLlm` transparently reads
- * from the cache instead of calling the network. No new code path through
- * the LLM client is needed; the cache hit is invisible to the runner.
- */
-import { ReplayError } from './errors';
-import type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink';
-export declare class ReplayCacheMissError extends ReplayError {
-    readonly url: string;
-    readonly requestKey: string;
-    constructor(url: string, requestKey: string, message?: string);
-}
-export interface ReplayCacheEntry {
-    request: RawProviderEvent;
-    response: RawProviderEvent;
-}
-export interface ReplayCacheStats {
-    total: number;
-    byProvider: Record<string, number>;
-    byModel: Record<string, number>;
-    /** Spans for which we have a request but no response (run aborted mid-call). */
-    orphanRequests: number;
-}
-/**
- * In-memory deterministic cache of (request → response) keyed on a stable
- * hash of the request body. Built from a `RawProviderSink` containing
- * paired `request` and `response` events from a previous run.
- *
- * The cache is the source of truth for replay; `createReplayFetch` is a
- * thin wrapper that reads from it.
- */
-export declare class ReplayCache {
-    private byKey;
-    private orphans;
-    private byProvider;
-    private byModel;
-    /**
-     * Build a cache from a sink's events. The sink must implement `list()`.
-     * Filter by `runId` / `spanId` to scope to a specific replay.
-     */
-    static fromSink(sink: RawProviderSink, filter?: {
-        runId?: string;
-        spanId?: string;
-    }): Promise<ReplayCache>;
-    /** Build a cache from an in-memory event list. */
-    static fromEvents(events: RawProviderEvent[]): Promise<ReplayCache>;
-    /** Number of cacheable (request, response) pairs in the cache. */
-    size(): number;
-    stats(): ReplayCacheStats;
-    /** Iterate every cached `(request, response)` pair in insertion order. */
-    entries(): IterableIterator<ReplayCacheEntry>;
-    /**
-     * Look up a cached response by hashing the (model, messages, temperature,
-     * maxTokens, response_format) shape. Returns `undefined` on miss; the
-     * caller decides whether to throw, fall back to the network, or skip.
-     */
-    lookup(requestBody: unknown): Promise<ReplayCacheEntry | undefined>;
-}
-export interface ReplayFetchOptions {
-    /**
-     * Behaviour on cache miss. Default `'throw'`. `'fallback'` calls the
-     * `fallbackFetch` (typically `globalThis.fetch`) so a partial replay can
-     * still complete; `'fail-closed'` returns a synthetic 599 response so the
-     * call site sees a non-retriable failure.
-     */
-    onMiss?: 'throw' | 'fallback' | 'fail-closed';
-    fallbackFetch?: typeof fetch;
-    /** Optional callback fired once per replayed call (for telemetry / counters). */
-    onHit?: (info: {
-        url: string;
-        provider: string;
-        model: string;
-    }) => void;
-    /** Optional callback fired on cache miss before the `onMiss` policy applies. */
-    onMissNotify?: (info: {
-        url: string;
-        requestBody: unknown;
-    }) => void;
-}
-/**
- * Build a `fetch`-shaped function that serves cached responses out of a
- * `ReplayCache` for any URL ending in `/chat/completions`. Pass through
- * `LlmClientOptions.fetch` and `callLlm` becomes free.
- *
- * Non-`/chat/completions` URLs are passed straight to the fallback fetch
- * (default: `globalThis.fetch`). This matters because non-LLM HTTP work
- * (judge HTTP servers, sandbox callbacks) sometimes flows through the same
- * `fetch` and shouldn't be intercepted.
- */
-export declare function createReplayFetch(cache: ReplayCache, opts?: ReplayFetchOptions): typeof fetch;
-/**
- * Convenience iterator over `(request, response)` pairs in a sink — for
- * post-hoc scoring that doesn't need a `fetch` shim. The judge or scorer
- * runs purely in-process over cached LLM outputs.
- */
-export declare function iterateRawCalls(sink: RawProviderSink, filter?: {
-    runId?: string;
-    spanId?: string;
-}): AsyncGenerator<ReplayCacheEntry>;
-//# sourceMappingURL=replay.d.ts.map

package/dist/replay.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"replay.d.ts","sourceRoot":"","sources":["../src/replay.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,UAAU,CAAA;AAEtC,OAAO,KAAK,EAAE,gBAAgB,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAA;AAElF,qBAAa,oBAAqB,SAAQ,WAAW;aAEjC,GAAG,EAAE,MAAM;aACX,UAAU,EAAE,MAAM;gBADlB,GAAG,EAAE,MAAM,EACX,UAAU,EAAE,MAAM,EAClC,OAAO,CAAC,EAAE,MAAM;CAInB;AAED,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,gBAAgB,CAAA;IACzB,QAAQ,EAAE,gBAAgB,CAAA;CAC3B;AAED,MAAM,WAAW,gBAAgB;IAC/B,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAClC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC/B,gFAAgF;IAChF,cAAc,EAAE,MAAM,CAAA;CACvB;AAED;;;;;;;GAOG;AACH,qBAAa,WAAW;IACtB,OAAO,CAAC,KAAK,CAAsC;IACnD,OAAO,CAAC,OAAO,CAAI;IACnB,OAAO,CAAC,UAAU,CAA6B;IAC/C,OAAO,CAAC,OAAO,CAA6B;IAE5C;;;OAGG;WACU,QAAQ,CACnB,IAAI,EAAE,eAAe,EACrB,MAAM,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAO,GAC/C,OAAO,CAAC,WAAW,CAAC;IAQvB,kDAAkD;WACrC,UAAU,CAAC,MAAM,EAAE,gBAAgB,EAAE,GAAG,OAAO,CAAC,WAAW,CAAC;IA2BzE,kEAAkE;IAClE,IAAI,IAAI,MAAM;IAId,KAAK,IAAI,gBAAgB;IASzB,0EAA0E;IACzE,OAAO,IAAI,gBAAgB,CAAC,gBAAgB,CAAC;IAI9C;;;;OAIG;IACG,MAAM,CAAC,WAAW,EAAE,OAAO,GAAG,OAAO,CAAC,gBAAgB,GAAG,SAAS,CAAC;CAI1E;AAED,MAAM,WAAW,kBAAkB;IACjC;;;;;OAKG;IACH,MAAM,CAAC,EAAE,OAAO,GAAG,UAAU,GAAG,aAAa,CAAA;IAC7C,aAAa,CAAC,EAAE,OAAO,KAAK,CAAA;IAC5B,iFAAiF;IACjF,KAAK,CAAC,EAAE,CAAC,IAAI,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,KAAK,IAAI,CAAA;IACxE,gFAAgF;IAChF,YAAY,CAAC,EAAE,CAAC,IAAI,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,OAAO,CAAA;KAAE,KAAK,IAAI,CAAA;CACrE;AAED;;;;;;;;;GASG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,WAAW,EAAE,IAAI,GAAE,kBAAuB,GAAG,OAAO,KAAK,CA+CjG;AAED;;;;GAIG;AACH,wBAAuB,eAAe,CACpC,IAAI,EAAE,eAAe,EACrB,MAAM,GAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAA;CAAO,GAC/C,cAAc,CAAC,gBAAgB,CAAC,CAOlC"}

package/dist/reporter.d.ts

@@ -1,14 +0,0 @@
-import type { BenchmarkReport, DriverResult } from './types';
-/**
- * Report generation utilities.
- *
- * Outputs convergence curves, cost curves, quality curves,
- * and per-persona summaries in markdown format.
- */
-/** Generate a markdown report from benchmark results */
-export declare function formatBenchmarkReport(report: BenchmarkReport): string;
-/** Generate a markdown report from agent driver results */
-export declare function formatDriverReport(results: DriverResult[]): string;
-/** Print a compact summary to console */
-export declare function printDriverSummary(results: DriverResult[]): void;
-//# sourceMappingURL=reporter.d.ts.map

package/dist/reporter.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"reporter.d.ts","sourceRoot":"","sources":["../src/reporter.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAE5D;;;;;GAKG;AAEH,wDAAwD;AACxD,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,eAAe,GAAG,MAAM,CAwDrE;AAED,2DAA2D;AAC3D,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,MAAM,CA6ClE;AAED,yCAAyC;AACzC,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,IAAI,CAgBhE"}

package/dist/reporting.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"reporting.d.ts","sourceRoot":"","sources":["../src/reporting.ts"],"names":[],"mappings":"AAAA,YAAY,EACV,iBAAiB,EACjB,6BAA6B,EAC7B,8BAA8B,EAC9B,aAAa,GACd,MAAM,wCAAwC,CAAA;AAC/C,OAAO,EAAE,wBAAwB,EAAE,MAAM,wCAAwC,CAAA;AACjF,YAAY,EACV,gBAAgB,EAChB,eAAe,EACf,mBAAmB,EACnB,OAAO,GACR,MAAM,kBAAkB,CAAA;AACzB,OAAO,EACL,WAAW,EACX,eAAe,GAChB,MAAM,kBAAkB,CAAA;AACzB,YAAY,EACV,qBAAqB,EACrB,yBAAyB,EACzB,sBAAsB,EACtB,sBAAsB,EACtB,wBAAwB,EACxB,0BAA0B,EAC1B,uBAAuB,EACvB,2BAA2B,EAC3B,oBAAoB,GACrB,MAAM,sBAAsB,CAAA;AAC7B,OAAO,EAAE,uBAAuB,EAAE,yBAAyB,EAAE,MAAM,sBAAsB,CAAA;AACzF,YAAY,EAAE,0BAA0B,EAAE,MAAM,kBAAkB,CAAA;AAClE,OAAO,EAAE,mBAAmB,EAAE,MAAM,kBAAkB,CAAA;AACtD,YAAY,EACV,wBAAwB,EACxB,6BAA6B,EAC7B,mBAAmB,EACnB,oBAAoB,EACpB,gBAAgB,EAChB,kBAAkB,GACnB,MAAM,cAAc,CAAA;AACrB,OAAO,EACL,gCAAgC,EAChC,oBAAoB,GACrB,MAAM,cAAc,CAAA;AACrB,YAAY,EACV,sBAAsB,EACtB,qBAAqB,GACtB,MAAM,cAAc,CAAA;AACrB,OAAO,EACL,iBAAiB,EACjB,eAAe,EACf,kBAAkB,GACnB,MAAM,cAAc,CAAA;AACrB,YAAY,EACV,mBAAmB,EACnB,0BAA0B,EAC1B,uBAAuB,EACvB,gBAAgB,EAChB,WAAW,EACX,cAAc,EACd,uBAAuB,EACvB,sBAAsB,EACtB,yBAAyB,EACzB,qBAAqB,EACrB,4BAA4B,EAC5B,YAAY,EACZ,mBAAmB,EACnB,eAAe,GAChB,MAAM,kBAAkB,CAAA;AACzB,OAAO,EACL,aAAa,EACb,WAAW,EACX,+BAA+B,EAC/B,cAAc,EACd,YAAY,GACb,MAAM,kBAAkB,CAAA"}

package/dist/researcher.d.ts

@@ -1,140 +0,0 @@
-/**
- * Researcher interface — stable hook for an external autonomous-research
- * agent to drive the meta-loop.
- *
- * 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.
- *
- * 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.
- */
-import type { GateDecision } from './held-out-gate';
-import type { RunRecord, RunSplitTag } from './run-record';
-/** A diagnosed failure mode with the run-IDs that exhibit it. */
-export 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. */
-export 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. */
-export 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. */
-export 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│
- *                              └────────────────┘
- */
-export interface Researcher {
-    inspectFailures(runs: RunRecord[]): Promise<FailureMode[]>;
-    proposeChange(failures: FailureMode[]): Promise<SteeringChange[]>;
-    applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan>;
-    evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult>;
-}
-export 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.
- */
-export 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.
- */
-export 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>;
-}
-/** Re-export the split alias so callers don't have to import twice. */
-export type { RunSplitTag };
-//# sourceMappingURL=researcher.d.ts.map

package/dist/researcher.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"researcher.d.ts","sourceRoot":"","sources":["../src/researcher.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAA;AACnD,OAAO,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,MAAM,cAAc,CAAA;AAE1D,iEAAiE;AACjE,MAAM,WAAW,WAAW;IAC1B;2DACuD;IACvD,IAAI,EAAE,MAAM,CAAA;IACZ,4DAA4D;IAC5D,WAAW,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE;QACR;wBACgB;QAChB,MAAM,EAAE,MAAM,EAAE,CAAA;QAChB,yDAAyD;QACzD,OAAO,EAAE,MAAM,CAAA;KAChB,CAAA;CACF;AAED,4DAA4D;AAC5D,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,iBAAiB,GAAG,WAAW,GAAG,cAAc,GAAG,WAAW,GAAG,QAAQ,CAAA;IAC/E;;uEAEmE;IACnE,OAAO,EAAE,OAAO,CAAA;IAChB;mDAC+C;IAC/C,SAAS,EAAE,MAAM,CAAA;IACjB,oEAAoE;IACpE,aAAa,CAAC,EAAE,MAAM,CAAA;CACvB;AAED,uEAAuE;AACvE,MAAM,WAAW,cAAc;IAC7B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,OAAO,EAAE,cAAc,EAAE,CAAA;IACzB;6DACyD;IACzD,mBAAmB,EAAE,MAAM,CAAA;IAC3B,qEAAqE;IACrE,MAAM,EAAE;QAAE,MAAM,EAAE,MAAM,EAAE,CAAC;QAAC,OAAO,EAAE,MAAM,EAAE,CAAA;KAAE,CAAA;CAChD;AAED,kEAAkE;AAClE,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,cAAc,CAAA;IACpB,IAAI,EAAE,SAAS,EAAE,CAAA;IACjB,YAAY,EAAE,YAAY,CAAA;CAC3B;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,WAAW,UAAU;IACzB,eAAe,CAAC,IAAI,EAAE,SAAS,EAAE,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,CAAA;IAC1D,aAAa,CAAC,QAAQ,EAAE,WAAW,EAAE,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC,CAAA;IACjE,WAAW,CAAC,OAAO,EAAE,cAAc,EAAE,EAAE,QAAQ,EAAE,cAAc,GAAG,OAAO,CAAC,cAAc,CAAC,CAAA;IACzF,cAAc,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAAA;CAChE;AAED,MAAM,WAAW,yBAAyB;IACxC,eAAe,EAAE,UAAU,CAAC,iBAAiB,CAAC,CAAA;IAC9C,aAAa,EAAE,UAAU,CAAC,eAAe,CAAC,CAAA;IAC1C,WAAW,EAAE,UAAU,CAAC,aAAa,CAAC,CAAA;IACtC,cAAc,EAAE,UAAU,CAAC,gBAAgB,CAAC,CAAA;CAC7C;AAED;;;GAGG;AACH,qBAAa,kBAAmB,YAAW,UAAU;IACvC,OAAO,CAAC,QAAQ,CAAC,SAAS;gBAAT,SAAS,EAAE,yBAAyB;IAEjE,eAAe,CAAC,IAAI,EAAE,SAAS,EAAE,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC;IAI1D,aAAa,CAAC,QAAQ,EAAE,WAAW,EAAE,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC;IAIjE,WAAW,CAAC,OAAO,EAAE,cAAc,EAAE,EAAE,QAAQ,EAAE,cAAc,GAAG,OAAO,CAAC,cAAc,CAAC;IAIzF,cAAc,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,gBAAgB,CAAC;CAGhE;AAED;;;;;;GAMG;AACH,qBAAa,cAAe,YAAW,UAAU;IAC/C,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAQ;gBAEjB,IAAI,SAA4C;IAItD,eAAe,CAAC,KAAK,EAAE,SAAS,EAAE,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC;IAI3D,aAAa,CAAC,SAAS,EAAE,WAAW,EAAE,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC;IAIlE,WAAW,CACf,QAAQ,EAAE,cAAc,EAAE,EAC1B,SAAS,EAAE,cAAc,GACxB,OAAO,CAAC,cAAc,CAAC;IAIpB,cAAc,CAAC,KAAK,EAAE,cAAc,GAAG,OAAO,CAAC,gBAAgB,CAAC;CAGvE;AAED,uEAAuE;AACvE,YAAY,EAAE,WAAW,EAAE,CAAA"}

package/dist/reviewer.d.ts

@@ -1,118 +0,0 @@
-/**
- * Reviewer primitives — prompt builder + default ReviewFn factory.
- *
- * `buildReviewerPrompt` is the pure, LLM-agnostic piece: takes
- * `ReviewerPromptInput` (user request, trace summary, verification
- * summary, memory, optional extra context) and emits the system +
- * user message pair. No LLM dependency — callers that want to drive
- * their own transport get full control.
- *
- * `createDefaultReviewer` is the convenience factory: wires the prompt
- * builder to `callLlmJson` with a default schema + soft-fail policy.
- * Returns a function that maps `ReviewerPromptInput` to `ReviewerOutput`.
- *
- * Same pattern as `runSemanticConceptJudge` / `createSemanticConceptJudge`:
- * low-level pure builder + high-level factory built on top.
- */
-import { type LlmClientOptions } from './llm-client';
-export interface ReviewerMemoryEntry {
-    shot: number;
-    ts?: string;
-    observations?: string;
-    diagnosis?: string;
-    nextShotInstruction?: string;
-    shouldContinue?: boolean;
-    confidence?: number;
-}
-export interface ReviewerVerificationSummary {
-    blendedScore: number;
-    allPass: boolean;
-    failCount: number;
-    failingLayers?: string[];
-}
-export interface ReviewerPromptInput {
-    shot: number;
-    userRequest: string;
-    /**
-     * Compact trace summary — tool-call counts, errors, recent activity
-     * lines. Built by the caller from whatever trace format they have;
-     * agent-eval does not prescribe.
-     */
-    traceSummary: string;
-    verification: ReviewerVerificationSummary;
-    memory: ReviewerMemoryEntry[];
-    /**
-     * Optional extra context injected into the prompt between the trace
-     * and the verification blocks. Use for workdir file-tree snapshots,
-     * scaffold descriptions, or any environmental fact the reviewer
-     * needs to direct the next shot accurately.
-     */
-    extraContext?: string;
-    /**
-     * Optional extra section appended at the end of the prompt (e.g.
-     * leaf metadata, scenario id). Free-form — no agent-eval-shaped
-     * schema.
-     */
-    trailingContext?: string;
-}
-export interface ReviewerOutput {
-    shot: number;
-    observations: string;
-    diagnosis: string;
-    nextShotInstruction: string;
-    shouldContinue: boolean;
-    /** 0..1 self-assessed confidence in the directive. */
-    confidence: number;
-    /** LLM cost in USD if the transport reports it, else null. */
-    costUsd: number | null;
-    durationMs: number;
-    /** False when the LLM errored or returned malformed JSON; caller soft-fails to defaults. */
-    available: boolean;
-    error?: string;
-}
-export interface ReviewerSoftFailDefaults {
-    observations?: string;
-    diagnosis?: string;
-    nextShotInstruction?: string;
-    shouldContinue?: boolean;
-    confidence?: number;
-}
-export interface CreateDefaultReviewerOptions {
-    /** Model id to call. */
-    model: string;
-    /** Per-call timeout. Default 300s. */
-    timeoutMs?: number;
-    /** LlmClient transport config (baseUrl, apiKey, authHeader, etc.). */
-    llm?: LlmClientOptions;
-    /**
-     * Override the prompt builder. Default: `buildReviewerPrompt`.
-     * Consumers with different reviewer voices pass their own.
-     */
-    promptBuilder?: (input: ReviewerPromptInput) => {
-        system: string;
-        user: string;
-    };
-    /**
-     * Soft-fail values when the LLM throws or returns unparseable JSON.
-     * Matches VerticalBench's shipped policy: continue with generic
-     * instruction at confidence 0.3 so the worker keeps trying.
-     */
-    softFailDefaults?: ReviewerSoftFailDefaults;
-}
-/**
- * Build the reviewer's system + user messages. Pure function, no LLM
- * call. Callers that want their own transport or a different structured
- * output can use this and skip `createDefaultReviewer` entirely.
- */
-export declare function buildReviewerPrompt(input: ReviewerPromptInput): {
-    system: string;
-    user: string;
-};
-/**
- * Factory: returns a function that invokes the default reviewer against
- * an LLM and parses the structured output. Soft-fails to the provided
- * defaults on LLM throw or JSON-parse error so the shot loop keeps
- * moving rather than crashing.
- */
-export declare function createDefaultReviewer(options: CreateDefaultReviewerOptions): (input: ReviewerPromptInput) => Promise<ReviewerOutput>;
-//# sourceMappingURL=reviewer.d.ts.map

package/dist/reviewer.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"reviewer.d.ts","sourceRoot":"","sources":["../src/reviewer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,EAAe,KAAK,gBAAgB,EAAE,MAAM,cAAc,CAAA;AAIjE,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,MAAM,CAAA;IACZ,EAAE,CAAC,EAAE,MAAM,CAAA;IACX,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAC5B,cAAc,CAAC,EAAE,OAAO,CAAA;IACxB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,2BAA2B;IAC1C,YAAY,EAAE,MAAM,CAAA;IACpB,OAAO,EAAE,OAAO,CAAA;IAChB,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;CACzB;AAED,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB;;;;OAIG;IACH,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,2BAA2B,CAAA;IACzC,MAAM,EAAE,mBAAmB,EAAE,CAAA;IAC7B;;;;;OAKG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;OAIG;IACH,eAAe,CAAC,EAAE,MAAM,CAAA;CACzB;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,YAAY,EAAE,MAAM,CAAA;IACpB,SAAS,EAAE,MAAM,CAAA;IACjB,mBAAmB,EAAE,MAAM,CAAA;IAC3B,cAAc,EAAE,OAAO,CAAA;IACvB,sDAAsD;IACtD,UAAU,EAAE,MAAM,CAAA;IAClB,8DAA8D;IAC9D,OAAO,EAAE,MAAM,GAAG,IAAI,CAAA;IACtB,UAAU,EAAE,MAAM,CAAA;IAClB,4FAA4F;IAC5F,SAAS,EAAE,OAAO,CAAA;IAClB,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,wBAAwB;IACvC,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAC5B,cAAc,CAAC,EAAE,OAAO,CAAA;IACxB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,4BAA4B;IAC3C,wBAAwB;IACxB,KAAK,EAAE,MAAM,CAAA;IACb,sCAAsC;IACtC,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,sEAAsE;IACtE,GAAG,CAAC,EAAE,gBAAgB,CAAA;IACtB;;;OAGG;IACH,aAAa,CAAC,EAAE,CAAC,KAAK,EAAE,mBAAmB,KAAK;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAA;IAChF;;;;OAIG;IACH,gBAAgB,CAAC,EAAE,wBAAwB,CAAA;CAC5C;AAkCD;;;;GAIG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,mBAAmB,GAAG;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CA6ChG;AAcD;;;;;GAKG;AACH,wBAAgB,qBAAqB,CACnC,OAAO,EAAE,4BAA4B,GACpC,CAAC,KAAK,EAAE,mBAAmB,KAAK,OAAO,CAAC,cAAc,CAAC,CA0DzD"}

package/dist/reviewer.test.d.ts

@@ -1,2 +0,0 @@
-export {};
-//# sourceMappingURL=reviewer.test.d.ts.map

package/dist/reviewer.test.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"reviewer.test.d.ts","sourceRoot":"","sources":["../src/reviewer.test.ts"],"names":[],"mappings":""}

package/dist/reward-model-export.d.ts

@@ -1,60 +0,0 @@
-/**
- * Reward-model export — the productizable wrapper around PRM training
- * data. Takes a TraceStore + PrmGrader, produces an embeddable
- * inference scorer that customers plug into their own agent stack.
- *
- * Two export forms:
- *   - `exportRewardModel(store, graders)` — serializes the (step-context,
- *     score) corpus to a framework-agnostic payload. Customer fine-tunes
- *     their own model; we ship the scaffolding.
- *   - `loadScorerFromTraces(store, grader)` — a zero-deps "reward model"
- *     that literally replays the trained rubric at inference time. Works
- *     as a reference baseline + deterministic fallback.
- */
-import type { PrmGradedTrace, PrmGrader } from './prm/rubric';
-import { type PrmTrainingSample } from './prm/training-export';
-import type { TraceStore } from './trace/store';
-import type { Trajectory } from './trajectory';
-export interface ExportedRewardModel {
-    /** Version of the export format. Bump when payload shape changes. */
-    version: '1.0';
-    /** Metadata about the training corpus. */
-    metadata: {
-        nTraces: number;
-        nSamples: number;
-        rubrics: string[];
-        exportedAt: string;
-        /** Mean reward across training corpus — use as sanity check at load. */
-        meanReward: number;
-    };
-    /** NDJSON training payload suitable for most fine-tuning frameworks. */
-    trainingNdjson: string;
-}
-export declare function exportRewardModel(store: TraceStore, grader: PrmGrader, runIds: string[]): Promise<ExportedRewardModel>;
-/**
- * Zero-deps inference scorer — apply a grader to a trajectory and return
- * its aggregate score. This is the "reward model" customers embed when
- * they don't want (or can't) fine-tune one. Deterministic + portable.
- */
-export interface InferenceScorer {
-    /** Score a completed trajectory. Higher is better. */
-    score(trajectory: Trajectory, store: TraceStore): Promise<number>;
-    metadata: {
-        rubrics: string[];
-        deterministic: true;
-    };
-}
-export declare function loadScorerFromGrader(grader: PrmGrader): InferenceScorer;
-/**
- * Replay a trace corpus through a scorer — produces the canonical
- * "what would this reward model have said about every run?" table.
- * Callers use this to validate a trained model against the training
- * corpus (expect high agreement; drift indicates overfitting).
- */
-export declare function replayScorerOverCorpus(store: TraceStore, scorer: InferenceScorer, runIds: string[]): Promise<Array<{
-    runId: string;
-    score: number;
-    outcomeScore: number | null;
-}>>;
-export type { PrmGradedTrace, PrmTrainingSample };
-//# sourceMappingURL=reward-model-export.d.ts.map

package/dist/reward-model-export.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"reward-model-export.d.ts","sourceRoot":"","sources":["../src/reward-model-export.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,SAAS,EAAE,MAAM,cAAc,CAAA;AAC7D,OAAO,EAAsB,KAAK,iBAAiB,EAAY,MAAM,uBAAuB,CAAA;AAC5F,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAC/C,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,cAAc,CAAA;AAG9C,MAAM,WAAW,mBAAmB;IAClC,qEAAqE;IACrE,OAAO,EAAE,KAAK,CAAA;IACd,0CAA0C;IAC1C,QAAQ,EAAE;QACR,OAAO,EAAE,MAAM,CAAA;QACf,QAAQ,EAAE,MAAM,CAAA;QAChB,OAAO,EAAE,MAAM,EAAE,CAAA;QACjB,UAAU,EAAE,MAAM,CAAA;QAClB,wEAAwE;QACxE,UAAU,EAAE,MAAM,CAAA;KACnB,CAAA;IACD,wEAAwE;IACxE,cAAc,EAAE,MAAM,CAAA;CACvB;AAED,wBAAsB,iBAAiB,CACrC,KAAK,EAAE,UAAU,EACjB,MAAM,EAAE,SAAS,EACjB,MAAM,EAAE,MAAM,EAAE,GACf,OAAO,CAAC,mBAAmB,CAAC,CAiB9B;AAED;;;;GAIG;AACH,MAAM,WAAW,eAAe;IAC9B,sDAAsD;IACtD,KAAK,CAAC,UAAU,EAAE,UAAU,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CAAA;IACjE,QAAQ,EAAE;QAAE,OAAO,EAAE,MAAM,EAAE,CAAC;QAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAA;CACrD;AAED,wBAAgB,oBAAoB,CAAC,MAAM,EAAE,SAAS,GAAG,eAAe,CAWvE;AAED;;;;;GAKG;AACH,wBAAsB,sBAAsB,CAC1C,KAAK,EAAE,UAAU,EACjB,MAAM,EAAE,eAAe,EACvB,MAAM,EAAE,MAAM,EAAE,GACf,OAAO,CAAC,KAAK,CAAC;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAA;CAAE,CAAC,CAAC,CAc/E;AAGD,YAAY,EAAE,cAAc,EAAE,iBAAiB,EAAE,CAAA"}