@tangle-network/agent-eval 0.132.0 → 0.133.1

package/dist/meta-eval/index.d.ts

@@ -1,4 +1,217 @@
+import { f as Run } from "../schema-BtVldJ3T.js";
+import { s as TraceStore } from "../store-CT9YIIve.js";
+import { a as ContinuousCalibrationResult, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, t as CalibrationResult } from "../judge-calibration-DFtEMlde.js";
+import { i as CorpusAgreementReport } from "../statistics-DbvkkDPa.js";
 import { a as OutcomeFilter, i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
-import { A as EvalMetricSpec, C as CalibrationPair, D as CorrelationResult, E as calibrationFromPairs, M as correlationStudy, O as CorrelationStudyOptions, S as CalibrationOptions, T as calibrationCurve, _ as snapshotFromAgreement, a as SentinelMetrics, b as validateSentinelSnapshot, c as SentinelSnapshot, d as SentinelTrend, f as SnapshotMeta, g as judgeSentinelReport, h as inMemorySentinelStore, i as SentinelMetricName, j as OutcomePair, k as CorrelationStudyResult, l as SentinelStore, m as fileSentinelStore, n as JudgeSentinelOptions, o as SentinelReport, p as evalHealthStamp, r as SENTINEL_METRIC_NAMES, s as SentinelSetOptions, t as EvalHealthStamp, u as SentinelThresholds, v as snapshotFromCalibration, w as CalibrationReport, x as CalibrationBin, y as snapshotFromSentinelSet } from "../index-CXs7QlR5.js";
+import { n as SeriesConvergenceResult, t as SeriesConvergenceOptions } from "../series-convergence-ofsqPWhs.js";
 import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "../rubric-predictive-validity-C1dCLcvb.js";
-export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomePair, OutcomeStore, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SENTINEL_METRIC_NAMES, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, calibrationCurve, calibrationFromPairs, correlationStudy, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, rubricPredictiveValidity, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot };
+//#region src/meta-eval/correlation-study.d.ts
+interface EvalMetricSpec {
+  id: string;
+  /** Extract a scalar from a run (defaults cover score/pass/durationMs/costUsd/tokens). */
+  extract?: (run: Run, store: TraceStore) => Promise<number | null>;
+}
+interface OutcomePair {
+  evalMetric: string;
+  outcomeMetric: string;
+}
+interface CorrelationResult {
+  evalMetric: string;
+  outcomeMetric: string;
+  n: number;
+  pearson: number;
+  spearman: number;
+  /** 95% bootstrap CI for Pearson. */
+  pearsonCi95: {
+    lower: number;
+    upper: number;
+  };
+  /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */
+  verdict: 'strong' | 'moderate' | 'weak';
+}
+interface CorrelationStudyResult {
+  pairs: CorrelationResult[];
+  joinedSamples: number;
+  skippedRuns: number;
+}
+interface CorrelationStudyOptions {
+  /** Only join outcomes captured within this window after run.startedAt. */
+  maxCaptureLagMs?: number;
+  /** Restrict to a subset of outcomes (cohort, region, source). */
+  outcomeFilter?: OutcomeFilter;
+  /** Which outcome per run to use when multiple exist. Default 'latest'. */
+  reduction?: 'latest' | 'mean' | 'max';
+  /** Bootstrap iterations for the CI. Default 500. */
+  bootstrapIterations?: number;
+}
+declare function correlationStudy(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetrics: EvalMetricSpec[], outcomeMetricNames: string[], options?: CorrelationStudyOptions): Promise<CorrelationStudyResult>;
+//#endregion
+//#region src/meta-eval/calibration.d.ts
+interface CalibrationBin {
+  lower: number;
+  upper: number;
+  n: number;
+  evalMean: number;
+  outcomeMean: number;
+  /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */
+  gap: number;
+}
+interface CalibrationReport {
+  evalMetric: string;
+  outcomeMetric: string;
+  n: number;
+  bins: CalibrationBin[];
+  /** Expected Calibration Error — Σ (n_i/N) × |outcomeMean_i − evalMean_i|. */
+  ece: number;
+  /** Max bin gap — upper bound on miscalibration. */
+  maxGap: number;
+}
+interface CalibrationOptions {
+  bins?: number;
+  /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */
+  binning?: 'equal-width' | 'equal-frequency';
+  /** Clip eval values to [lo, hi] before binning. */
+  range?: {
+    lo: number;
+    hi: number;
+  };
+}
+interface CalibrationPair {
+  evalScore: number;
+  outcome: number;
+}
+declare function calibrationCurve(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetric: EvalMetricSpec, outcomeMetric: string, options?: CalibrationOptions): Promise<CalibrationReport | null>;
+declare function calibrationFromPairs(inputPairs: CalibrationPair[], evalMetric: string, outcomeMetric: string, options?: CalibrationOptions): CalibrationReport | null;
+//#endregion
+//#region src/meta-eval/sentinel.d.ts
+declare const SENTINEL_METRIC_NAMES: readonly ['irr', 'calibrationKappa', 'sentinelPassRate'];
+type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number];
+interface SentinelMetrics {
+  /** Inter-rater reliability — ICC(2,1) from agreement instruments. */
+  irr?: number;
+  /** Weighted κ vs the human golden set (`calibrateJudge*`). */
+  calibrationKappa?: number;
+  /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */
+  sentinelPassRate?: number;
+}
+interface SentinelSnapshot {
+  /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */
+  at: string;
+  judgeId: string;
+  /** Exact model identity behind the judge (pin the full version string). */
+  judgeModel: string;
+  /**
+   * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a
+   * model-change marker — it records the new `judgeModel` immediately and
+   * the silent-upgrade alarm stays raised until a golden-grounded snapshot
+   * (calibrationKappa or sentinelPassRate) follows.
+   */
+  metrics: SentinelMetrics;
+}
+/** Identity + timestamp the caller supplies alongside an instrument output. */
+interface SnapshotMeta {
+  at: string;
+  judgeId: string;
+  judgeModel: string;
+}
+/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */
+declare function validateSentinelSnapshot(snapshot: SentinelSnapshot, source?: string): void;
+/**
+ * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the
+ * un-rounded continuous κ_w when the report carries one (integer κ
+ * discards information for [0,1] judges).
+ */
+declare function snapshotFromCalibration(report: CalibrationResult | ContinuousCalibrationResult, meta: SnapshotMeta): SentinelSnapshot;
+/**
+ * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`
+ * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability
+ * coefficient both instruments compute. For ensemble-level agreement,
+ * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its
+ * composition (e.g. a joined list of member model strings).
+ */
+declare function snapshotFromAgreement(report: CorpusAgreementReport | ContinuousAgreement, meta: SnapshotMeta): SentinelSnapshot;
+interface SentinelSetOptions {
+  /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */
+  tolerance?: number;
+}
+/**
+ * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the
+ * judge was originally calibrated on, re-scored periodically. Pass rate =
+ * fraction of joined items where the judge stays within `tolerance` of the
+ * human score. Items the judge didn't score are excluded from the
+ * denominator; zero joined items is an error, not a 0% pass rate.
+ */
+declare function snapshotFromSentinelSet(scores: CandidateScore[], golden: GoldenItem[], meta: SnapshotMeta, options?: SentinelSetOptions): SentinelSnapshot;
+interface SentinelStore {
+  append(snapshot: SentinelSnapshot): Promise<void>;
+  /** All snapshots (optionally for one judge), in append order. */
+  history(judgeId?: string): Promise<SentinelSnapshot[]>;
+}
+declare function inMemorySentinelStore(initial?: SentinelSnapshot[]): SentinelStore;
+/**
+ * JSONL store — one snapshot per line, appended atomically per call.
+ * A corrupt or shape-invalid line is a loud error naming the file and
+ * line number: a sentinel history that silently drops records would
+ * defeat the drift detection it exists to provide.
+ */
+declare function fileSentinelStore(path: string): SentinelStore;
+interface SentinelThresholds {
+  /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */
+  minIrr?: number;
+  /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */
+  maxKappaDrop?: number;
+  /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */
+  maxSentinelDrop?: number;
+  /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */
+  staleAfterDays?: number;
+}
+interface JudgeSentinelOptions {
+  /** Caller-supplied ISO timestamp staleness is measured against. */
+  asOf: string;
+  thresholds?: SentinelThresholds;
+  /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */
+  convergence?: SeriesConvergenceOptions;
+}
+interface SentinelTrend {
+  judgeId: string;
+  metric: SentinelMetricName;
+  /** Verbatim `analyzeSeries` state for this judge×metric history. */
+  state: SeriesConvergenceResult['state'];
+  /** Newest value in the series. */
+  current: number;
+  /** Oldest value in the series — the reference the drop thresholds compare against. */
+  baseline: number;
+  /** current − baseline (signed; negative = decayed). */
+  drift: number;
+  alarmed: boolean;
+  reason?: string;
+}
+interface SentinelReport {
+  perJudge: SentinelTrend[];
+  alarms: string[];
+  healthy: boolean;
+  /**
+   * `judgeId:metric` pairs with too few snapshots for the convergence
+   * machine. Named, not counted — a blind spot you can't see is a blind
+   * spot you won't fix. Floor/drop checks still apply to thin series, so
+   * insufficient history alone never masks an alarm.
+   */
+  insufficientHistory: string[];
+}
+declare function judgeSentinelReport(history: SentinelSnapshot[], opts: JudgeSentinelOptions): SentinelReport;
+interface EvalHealthStamp {
+  healthy: boolean;
+  alarms: string[];
+}
+/**
+ * The object a campaign attaches to its verdicts. Compute it from the
+ * sentinel report in a post-campaign hook (or a nightly job feeding the
+ * next day's campaigns) and store it alongside the verdict payload.
+ * `healthy: false` means the judges that produced those verdicts have an
+ * unresolved drift / decay / silent-upgrade / staleness alarm — treat the
+ * verdicts as carrying an integrity warning until recalibration clears it.
+ */
+declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
+//#endregion
+export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomePair, OutcomeStore, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SENTINEL_METRIC_NAMES, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, calibrationCurve, calibrationFromPairs, correlationStudy, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, rubricPredictiveValidity, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot };
+//# sourceMappingURL=index.d.ts.map

package/dist/{index-CXs7QlR5.d.ts.map → meta-eval/index.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"index-CXs7QlR5.d.ts","names":[],"sources":["../src/meta-eval/correlation-study.ts","../src/meta-eval/calibration.ts","../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;UAiBiB;EACf;;EAEA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;;UAGe;EACf,OAAO;EACP;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UCjDM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;EACf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;iBAgCK,qBACd,YAAY,mBACZ,oBACA,uBACA,UAAS,qBACR;;;cCtCU;KAED,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;UAiBiB;EACf;;EAEA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;;UAGe;EACf,OAAO;EACP;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UCjDM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;EACf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;iBAgCK,qBACd,YAAY,mBACZ,oBACA,uBACA,UAAS,qBACR;;;cCtCU;KAED,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}

package/dist/meta-eval/index.js

@@ -1,10 +1,100 @@
 import { s as ValidationError } from "../errors-8YnH8WlF.js";
-import { A as spearmanR, E as pearsonR } from "../statistics-CnnxdpOg.js";
+import { A as spearmanR, E as pearsonR } from "../statistics-DWM_AyLe.js";
 import { o as llmSpans, t as aggregateLlm } from "../query-Di7eEQ79.js";
 import { t as analyzeSeries } from "../series-convergence-CjO2QdRW.js";
-import { t as rubricPredictiveValidity } from "../rubric-predictive-validity-B3xmbmS1.js";
+import { t as rubricPredictiveValidity } from "../rubric-predictive-validity-QG7ydk0s.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
-import { n as calibrationFromPairs, t as calibrationCurve } from "../calibration-CNWWA6K8.js";
+//#region src/meta-eval/calibration.ts
+async function calibrationCurve(traceStore, outcomeStore, evalMetric, outcomeMetric, options = {}) {
+	const runs = await traceStore.listRuns();
+	const outcomes = await outcomeStore.list();
+	const byRun = /* @__PURE__ */ new Map();
+	for (const o of outcomes) {
+		const arr = byRun.get(o.runId) ?? [];
+		arr.push(o);
+		byRun.set(o.runId, arr);
+	}
+	const extract = evalMetric.extract ?? defaultExtract$1(evalMetric.id);
+	const pairs = [];
+	for (const run of runs) {
+		const os = byRun.get(run.runId);
+		if (!os?.length) continue;
+		const x = await extract(run, traceStore);
+		if (x === null || !Number.isFinite(x)) continue;
+		const y = [...os].sort((a, b) => b.capturedAt - a.capturedAt)[0].metrics[outcomeMetric];
+		if (typeof y !== "number" || !Number.isFinite(y)) continue;
+		pairs.push({
+			x,
+			y
+		});
+	}
+	if (pairs.length < 2) return null;
+	return calibrationFromPairs(pairs.map((p) => ({
+		evalScore: p.x,
+		outcome: p.y
+	})), evalMetric.id, outcomeMetric, options);
+}
+function calibrationFromPairs(inputPairs, evalMetric, outcomeMetric, options = {}) {
+	const pairs = inputPairs.filter((pair) => Number.isFinite(pair.evalScore) && Number.isFinite(pair.outcome));
+	if (pairs.length < 2) return null;
+	const numBins = options.bins ?? 10;
+	const binning = options.binning ?? "equal-width";
+	const xs = pairs.map((p) => p.evalScore);
+	const lo = options.range?.lo ?? Math.min(...xs);
+	const hi = options.range?.hi ?? Math.max(...xs);
+	const bins = [];
+	if (binning === "equal-frequency") {
+		const sorted = [...pairs].sort((a, b) => a.evalScore - b.evalScore);
+		const perBin = Math.max(1, Math.floor(sorted.length / numBins));
+		for (let i = 0; i < sorted.length; i += perBin) {
+			const chunk = sorted.slice(i, i + perBin);
+			if (chunk.length === 0) continue;
+			bins.push(toBin(chunk));
+		}
+	} else {
+		const width = (hi - lo) / numBins;
+		if (width === 0) return null;
+		for (let i = 0; i < numBins; i++) {
+			const binLo = lo + i * width;
+			const binHi = i === numBins - 1 ? hi + 1e-9 : lo + (i + 1) * width;
+			const chunk = pairs.filter((p) => p.evalScore >= binLo && p.evalScore < binHi);
+			if (chunk.length === 0) continue;
+			bins.push(toBin(chunk, binLo, binHi));
+		}
+	}
+	const total = bins.reduce((a, b) => a + b.n, 0);
+	const ece = bins.reduce((a, b) => a + b.n / total * b.gap, 0);
+	const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0);
+	return {
+		evalMetric,
+		outcomeMetric,
+		n: pairs.length,
+		bins,
+		ece,
+		maxGap
+	};
+}
+function toBin(chunk, lower, upper) {
+	const xs = chunk.map((c) => c.evalScore);
+	const ys = chunk.map((c) => c.outcome);
+	const evalMean = mean(xs);
+	const outcomeMean = mean(ys);
+	return {
+		lower: lower ?? Math.min(...xs),
+		upper: upper ?? Math.max(...xs),
+		n: chunk.length,
+		evalMean,
+		outcomeMean,
+		gap: Math.abs(outcomeMean - evalMean)
+	};
+}
+function mean(xs) {
+	return xs.reduce((a, b) => a + b, 0) / xs.length;
+}
+function defaultExtract$1(metric) {
+	return async (run) => run.outcome?.score ?? (metric === "pass" ? run.outcome?.pass === true ? 1 : 0 : null);
+}
+//#endregion
 //#region src/meta-eval/correlation-study.ts
 /**
 * Correlation study — "does our eval score predict real-world outcomes?"

package/dist/meta-eval/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/sentinel.ts"],"sourcesContent":["/**\n * Correlation study — \"does our eval score predict real-world outcomes?\"\n *\n * This is the load-bearing signal. Takes a TraceStore + OutcomeStore,\n * joins on runId, computes Pearson + Spearman + bootstrap CI for every\n * (evalMetric, outcomeMetric) pair the caller declares.\n *\n * Without this number the framework is ornamental. With it and r > 0.6\n * the framework is a moat — no other agent-eval tool publishes one.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport { aggregateLlm, llmSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { DeploymentOutcome, OutcomeFilter, OutcomeStore } from './outcome-store'\n\nexport interface EvalMetricSpec {\n  id: string\n  /** Extract a scalar from a run (defaults cover score/pass/durationMs/costUsd/tokens). */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface OutcomePair {\n  evalMetric: string\n  outcomeMetric: string\n}\n\nexport interface CorrelationResult {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  pearson: number\n  spearman: number\n  /** 95% bootstrap CI for Pearson. */\n  pearsonCi95: { lower: number; upper: number }\n  /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */\n  verdict: 'strong' | 'moderate' | 'weak'\n}\n\nexport interface CorrelationStudyResult {\n  pairs: CorrelationResult[]\n  joinedSamples: number\n  skippedRuns: number\n}\n\nexport interface CorrelationStudyOptions {\n  /** Only join outcomes captured within this window after run.startedAt. */\n  maxCaptureLagMs?: number\n  /** Restrict to a subset of outcomes (cohort, region, source). */\n  outcomeFilter?: OutcomeFilter\n  /** Which outcome per run to use when multiple exist. Default 'latest'. */\n  reduction?: 'latest' | 'mean' | 'max'\n  /** Bootstrap iterations for the CI. Default 500. */\n  bootstrapIterations?: number\n}\n\nexport async function correlationStudy(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetrics: EvalMetricSpec[],\n  outcomeMetricNames: string[],\n  options: CorrelationStudyOptions = {},\n): Promise<CorrelationStudyResult> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list(options.outcomeFilter)\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  const reduction = options.reduction ?? 'latest'\n  const maxLag = options.maxCaptureLagMs ?? Infinity\n\n  const pairs: Array<{ evalMetric: string; outcomeMetric: string; xs: number[]; ys: number[] }> = []\n  for (const em of evalMetrics) {\n    for (const om of outcomeMetricNames) {\n      pairs.push({ evalMetric: em.id, outcomeMetric: om, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    const eligible = os.filter((o) => o.capturedAt - run.startedAt <= maxLag)\n    if (eligible.length === 0) {\n      skipped++\n      continue\n    }\n\n    for (const em of evalMetrics) {\n      const extract = em.extract ?? defaultExtract(em.id)\n      const x = await extract(run, traceStore)\n      if (x === null || !Number.isFinite(x)) continue\n\n      for (const om of outcomeMetricNames) {\n        const values = eligible\n          .map((o) => o.metrics[om])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, reduction, eligible)\n        if (y === null) continue\n        const pair = pairs.find((p) => p.evalMetric === em.id && p.outcomeMetric === om)!\n        pair.xs.push(x)\n        pair.ys.push(y)\n      }\n    }\n    joined++\n  }\n\n  const results: CorrelationResult[] = pairs\n    .filter((p) => p.xs.length >= 3)\n    .map((p) => {\n      const pearson = pearsonR(p.xs, p.ys)\n      const spearman = spearmanR(p.xs, p.ys)\n      const pearsonCi95 = bootstrapPearsonCi(p.xs, p.ys, options.bootstrapIterations ?? 500)\n      const verdict: CorrelationResult['verdict'] =\n        Math.abs(pearson) >= 0.7 ? 'strong' : Math.abs(pearson) >= 0.4 ? 'moderate' : 'weak'\n      return {\n        evalMetric: p.evalMetric,\n        outcomeMetric: p.outcomeMetric,\n        n: p.xs.length,\n        pearson,\n        spearman,\n        pearsonCi95,\n        verdict,\n      }\n    })\n\n  return { pairs: results, joinedSamples: joined, skippedRuns: skipped }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  kind: 'latest' | 'mean' | 'max',\n  outcomes: DeploymentOutcome[],\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((a, b) => a + b, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest': pick the outcome captured last, then lookup its metric\n  const latest = [...outcomes].sort((a, b) => b.capturedAt - a.capturedAt)[0]\n  if (!latest) return null\n  const latestKey = Object.keys(latest.metrics)[0]\n  const v = latestKey !== undefined ? latest.metrics[latestKey] : undefined\n  // For 'latest' we already have `values` aligned; use the last-captured one\n  const paired = outcomes\n    .map((o) => {\n      const k = Object.keys(o.metrics)[0]\n      return {\n        at: o.capturedAt,\n        v: k !== undefined ? values.find((x) => o.metrics[k] === x) : undefined,\n      }\n    })\n    .filter((p) => p.v !== undefined)\n  if (paired.length === 0) return v ?? null\n  return paired.sort((a, b) => b.at - a.at)[0]?.v ?? null\n}\n\nfunction bootstrapPearsonCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n): { lower: number; upper: number } {\n  const n = xs.length\n  if (n < 3) return { lower: NaN, upper: NaN }\n  const rs: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx: number[] = new Array(n)\n    const ry: number[] = new Array(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(Math.random() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) rs.push(r)\n  }\n  rs.sort((a, b) => a - b)\n  if (rs.length === 0) return { lower: NaN, upper: NaN }\n  return {\n    lower: rs[Math.floor(0.025 * rs.length)]!,\n    upper: rs[Math.min(rs.length - 1, Math.floor(0.975 * rs.length))]!,\n  }\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * Judge sentinel — eval trustworthiness as a continuously measured,\n * alarmed trend.\n *\n * Judges are models; models change underneath us; calibration decays.\n * This module composes three existing instruments into one loop:\n *\n *   snapshot — adapters turn real instrument outputs (`calibrateJudge` /\n *              `calibrateJudgeContinuous`, `corpusInterRaterAgreement` /\n *              `continuousAgreement`, a rerun sentinel golden set) into\n *              `SentinelSnapshot`s\n *   series   — `analyzeSeries` (src/series-convergence) classifies each\n *              judge×metric history as stabilized / drifting / noisy\n *   alarm    — `judgeSentinelReport` turns trends + thresholds into named\n *              alarms; `evalHealthStamp` is the tiny object a campaign\n *              attaches to its verdicts\n *\n * Alarm conditions:\n *   - convergence state `drifting-down` on any tracked metric\n *   - irr below `minIrr`\n *   - calibrationKappa / sentinelPassRate dropped vs the series baseline\n *     beyond `maxKappaDrop` / `maxSentinelDrop`\n *   - judge model changed with no post-change golden-grounded snapshot\n *     (the silent-upgrade trap: judges agreeing with each other after a\n *     model swap proves nothing — only re-measuring against gold does)\n *   - newest snapshot older than `staleAfterDays` relative to `asOf`\n *\n * Wire-in: run the snapshot adapters from a post-campaign hook or a\n * nightly job, append to a `SentinelStore`, then compute\n * `judgeSentinelReport(await store.history(), { asOf })` with `asOf` =\n * the campaign timestamp. Attach `evalHealthStamp(report)` to campaign\n * verdicts: an alarmed sentinel means every downstream verdict carries\n * an integrity warning until the judge is recalibrated.\n *\n * The module is clock-free — every timestamp (`at`, `asOf`) is\n * caller-supplied ISO-8601, so reports are reproducible.\n */\n\nimport { ValidationError } from '../errors'\nimport type {\n  CalibrationResult,\n  CandidateScore,\n  ContinuousAgreement,\n  ContinuousCalibrationResult,\n  GoldenItem,\n} from '../judge-calibration'\nimport {\n  analyzeSeries,\n  type SeriesConvergenceOptions,\n  type SeriesConvergenceResult,\n} from '../series-convergence'\nimport type { CorpusAgreementReport } from '../statistics'\n\nexport const SENTINEL_METRIC_NAMES = ['irr', 'calibrationKappa', 'sentinelPassRate'] as const\n\nexport type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number]\n\nexport interface SentinelMetrics {\n  /** Inter-rater reliability — ICC(2,1) from agreement instruments. */\n  irr?: number\n  /** Weighted κ vs the human golden set (`calibrateJudge*`). */\n  calibrationKappa?: number\n  /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */\n  sentinelPassRate?: number\n}\n\nexport interface SentinelSnapshot {\n  /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */\n  at: string\n  judgeId: string\n  /** Exact model identity behind the judge (pin the full version string). */\n  judgeModel: string\n  /**\n   * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a\n   * model-change marker — it records the new `judgeModel` immediately and\n   * the silent-upgrade alarm stays raised until a golden-grounded snapshot\n   * (calibrationKappa or sentinelPassRate) follows.\n   */\n  metrics: SentinelMetrics\n}\n\n/** Identity + timestamp the caller supplies alongside an instrument output. */\nexport interface SnapshotMeta {\n  at: string\n  judgeId: string\n  judgeModel: string\n}\n\nfunction parseIso(value: string, label: string): number {\n  const ms = typeof value === 'string' ? Date.parse(value) : NaN\n  if (!Number.isFinite(ms)) {\n    throw new ValidationError(\n      `judge-sentinel: ${label} is not a parseable ISO timestamp: ${JSON.stringify(value)}`,\n    )\n  }\n  return ms\n}\n\n/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */\nexport function validateSentinelSnapshot(snapshot: SentinelSnapshot, source = 'snapshot'): void {\n  parseIso(snapshot.at, `${source}.at`)\n  if (typeof snapshot.judgeId !== 'string' || snapshot.judgeId.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeId must be a non-empty string`)\n  }\n  if (typeof snapshot.judgeModel !== 'string' || snapshot.judgeModel.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeModel must be a non-empty string`)\n  }\n  if (snapshot.metrics === null || typeof snapshot.metrics !== 'object') {\n    throw new ValidationError(`judge-sentinel: ${source}.metrics must be an object (may be empty)`)\n  }\n  for (const [key, value] of Object.entries(snapshot.metrics)) {\n    if (!(SENTINEL_METRIC_NAMES as readonly string[]).includes(key)) {\n      // A typo'd key would silently never trend — reject instead.\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics carries unknown key \"${key}\" — known: ${SENTINEL_METRIC_NAMES.join(', ')}`,\n      )\n    }\n    if (value !== undefined && !Number.isFinite(value)) {\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics.${key} is not a finite number: ${String(value)}`,\n      )\n    }\n  }\n}\n\n// ── Snapshot adapters — real instrument output → SentinelSnapshot ───\n\n/**\n * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the\n * un-rounded continuous κ_w when the report carries one (integer κ\n * discards information for [0,1] judges).\n */\nexport function snapshotFromCalibration(\n  report: CalibrationResult | ContinuousCalibrationResult,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const kappa = 'weightedKappaContinuous' in report ? report.weightedKappaContinuous : report.kappa\n  if (!Number.isFinite(kappa)) {\n    throw new ValidationError(\n      `snapshotFromCalibration: κ is not finite (n=${report.n}) — calibrate against ≥2 joined golden items before snapshotting`,\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { calibrationKappa: kappa } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n/**\n * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`\n * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability\n * coefficient both instruments compute. For ensemble-level agreement,\n * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its\n * composition (e.g. a joined list of member model strings).\n */\nexport function snapshotFromAgreement(\n  report: CorpusAgreementReport | ContinuousAgreement,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const irr = 'overallIcc' in report ? report.overallIcc : report.icc\n  if (!Number.isFinite(irr)) {\n    throw new ValidationError(\n      'snapshotFromAgreement: ICC is not finite — agreement needs ≥2 raters and ≥2 complete items',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { irr } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\nexport interface SentinelSetOptions {\n  /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */\n  tolerance?: number\n}\n\n/**\n * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the\n * judge was originally calibrated on, re-scored periodically. Pass rate =\n * fraction of joined items where the judge stays within `tolerance` of the\n * human score. Items the judge didn't score are excluded from the\n * denominator; zero joined items is an error, not a 0% pass rate.\n */\nexport function snapshotFromSentinelSet(\n  scores: CandidateScore[],\n  golden: GoldenItem[],\n  meta: SnapshotMeta,\n  options: SentinelSetOptions = {},\n): SentinelSnapshot {\n  const tolerance = options.tolerance ?? 0.1\n  if (!Number.isFinite(tolerance) || tolerance < 0) {\n    throw new ValidationError(\n      `snapshotFromSentinelSet: tolerance must be a finite number ≥ 0, got ${tolerance}`,\n    )\n  }\n  const goldenById = new Map<string, number>()\n  for (const item of golden) {\n    if (!Number.isFinite(item.humanScore)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: golden item \"${item.itemId}\" has a non-finite humanScore`,\n      )\n    }\n    if (goldenById.has(item.itemId)) {\n      throw new ValidationError(`snapshotFromSentinelSet: duplicate golden itemId \"${item.itemId}\"`)\n    }\n    goldenById.set(item.itemId, item.humanScore)\n  }\n  let joined = 0\n  let passed = 0\n  for (const s of scores) {\n    const human = goldenById.get(s.itemId)\n    if (human === undefined) continue\n    if (!Number.isFinite(s.score)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: judge score for item \"${s.itemId}\" is not finite`,\n      )\n    }\n    joined++\n    if (Math.abs(s.score - human) <= tolerance) passed++\n  }\n  if (joined === 0) {\n    throw new ValidationError(\n      'snapshotFromSentinelSet: no judge scores joined the golden set — itemIds mismatch or empty sentinel run',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { sentinelPassRate: passed / joined } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n// ── Persistence ──────────────────────────────────────────────────────\n\nexport interface SentinelStore {\n  append(snapshot: SentinelSnapshot): Promise<void>\n  /** All snapshots (optionally for one judge), in append order. */\n  history(judgeId?: string): Promise<SentinelSnapshot[]>\n}\n\nexport function inMemorySentinelStore(initial: SentinelSnapshot[] = []): SentinelStore {\n  for (const s of initial) validateSentinelSnapshot(s)\n  const items = initial.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      items.push({ ...snapshot, metrics: { ...snapshot.metrics } })\n    },\n    async history(judgeId) {\n      const all = items.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n      return judgeId === undefined ? all : all.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n/**\n * JSONL store — one snapshot per line, appended atomically per call.\n * A corrupt or shape-invalid line is a loud error naming the file and\n * line number: a sentinel history that silently drops records would\n * defeat the drift detection it exists to provide.\n */\nexport function fileSentinelStore(path: string): SentinelStore {\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      const fs = await import('node:fs/promises')\n      const pathMod = await import('node:path')\n      await fs.mkdir(pathMod.dirname(path), { recursive: true })\n      await fs.appendFile(path, `${JSON.stringify(snapshot)}\\n`, 'utf8')\n    },\n    async history(judgeId) {\n      const fs = await import('node:fs/promises')\n      let raw: string\n      try {\n        raw = await fs.readFile(path, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') return []\n        throw err\n      }\n      const lines = raw.split('\\n')\n      const snapshots: SentinelSnapshot[] = []\n      for (let i = 0; i < lines.length; i++) {\n        const line = lines[i]!\n        if (line.trim().length === 0) continue\n        let parsed: unknown\n        try {\n          parsed = JSON.parse(line)\n        } catch (err) {\n          throw new ValidationError(\n            `judge-sentinel: store at ${path} has a corrupt JSONL line ${i + 1}: ${(err as Error).message}`,\n          )\n        }\n        const snapshot = parsed as SentinelSnapshot\n        validateSentinelSnapshot(snapshot, `${path}:${i + 1}`)\n        snapshots.push(snapshot)\n      }\n      return judgeId === undefined ? snapshots : snapshots.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n// ── Trend report ─────────────────────────────────────────────────────\n\nexport interface SentinelThresholds {\n  /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */\n  minIrr?: number\n  /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */\n  maxKappaDrop?: number\n  /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */\n  maxSentinelDrop?: number\n  /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */\n  staleAfterDays?: number\n}\n\nexport interface JudgeSentinelOptions {\n  /** Caller-supplied ISO timestamp staleness is measured against. */\n  asOf: string\n  thresholds?: SentinelThresholds\n  /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */\n  convergence?: SeriesConvergenceOptions\n}\n\nexport interface SentinelTrend {\n  judgeId: string\n  metric: SentinelMetricName\n  /** Verbatim `analyzeSeries` state for this judge×metric history. */\n  state: SeriesConvergenceResult['state']\n  /** Newest value in the series. */\n  current: number\n  /** Oldest value in the series — the reference the drop thresholds compare against. */\n  baseline: number\n  /** current − baseline (signed; negative = decayed). */\n  drift: number\n  alarmed: boolean\n  reason?: string\n}\n\nexport interface SentinelReport {\n  perJudge: SentinelTrend[]\n  alarms: string[]\n  healthy: boolean\n  /**\n   * `judgeId:metric` pairs with too few snapshots for the convergence\n   * machine. Named, not counted — a blind spot you can't see is a blind\n   * spot you won't fix. Floor/drop checks still apply to thin series, so\n   * insufficient history alone never masks an alarm.\n   */\n  insufficientHistory: string[]\n}\n\nconst MS_PER_DAY = 86_400_000\n\nexport function judgeSentinelReport(\n  history: SentinelSnapshot[],\n  opts: JudgeSentinelOptions,\n): SentinelReport {\n  const asOfMs = parseIso(opts.asOf, 'asOf')\n  const minIrr = opts.thresholds?.minIrr ?? 0.6\n  const maxKappaDrop = opts.thresholds?.maxKappaDrop ?? 0.15\n  const maxSentinelDrop = opts.thresholds?.maxSentinelDrop ?? 0.1\n  const staleAfterDays = opts.thresholds?.staleAfterDays ?? 30\n\n  const byJudge = new Map<string, Array<SentinelSnapshot & { atMs: number }>>()\n  for (const snapshot of history) {\n    validateSentinelSnapshot(snapshot)\n    const atMs = parseIso(snapshot.at, `snapshot.at (judge \"${snapshot.judgeId}\")`)\n    const arr = byJudge.get(snapshot.judgeId) ?? []\n    arr.push({ ...snapshot, atMs })\n    byJudge.set(snapshot.judgeId, arr)\n  }\n\n  const perJudge: SentinelTrend[] = []\n  const alarms: string[] = []\n  const insufficientHistory: string[] = []\n\n  for (const judgeId of [...byJudge.keys()].sort()) {\n    const snaps = byJudge.get(judgeId)!.sort((a, b) => a.atMs - b.atMs)\n    const newest = snaps[snaps.length - 1]!\n\n    const ageDays = (asOfMs - newest.atMs) / MS_PER_DAY\n    if (ageDays > staleAfterDays) {\n      alarms.push(\n        `judge \"${judgeId}\": stale — newest snapshot ${newest.at} is ${ageDays.toFixed(1)}d old vs asOf (limit ${staleAfterDays}d)`,\n      )\n    }\n\n    // Silent-upgrade trap: only the latest model change matters for current\n    // trust, and only golden-grounded metrics clear it — irr alone proves\n    // judges still agree with each other, not that they're still right.\n    let changeIdx = -1\n    for (let i = 1; i < snaps.length; i++) {\n      if (snaps[i]!.judgeModel !== snaps[i - 1]!.judgeModel) changeIdx = i\n    }\n    if (changeIdx >= 0) {\n      const recalibrated = snaps\n        .slice(changeIdx)\n        .some(\n          (s) =>\n            Number.isFinite(s.metrics.calibrationKappa) ||\n            Number.isFinite(s.metrics.sentinelPassRate),\n        )\n      if (!recalibrated) {\n        alarms.push(\n          `judge \"${judgeId}\": model changed \"${snaps[changeIdx - 1]!.judgeModel}\" → \"${snaps[changeIdx]!.judgeModel}\" at ${snaps[changeIdx]!.at} with no post-change calibration snapshot — silent-upgrade trap; rerun calibrateJudge or the sentinel set against the new model`,\n        )\n      }\n    }\n\n    for (const metric of SENTINEL_METRIC_NAMES) {\n      const values = snaps\n        .filter((s) => typeof s.metrics[metric] === 'number')\n        .map((s) => s.metrics[metric]!)\n      if (values.length === 0) continue\n\n      const convergence = analyzeSeries(values, opts.convergence)\n      if (convergence.state === 'insufficient-data')\n        insufficientHistory.push(`${judgeId}:${metric}`)\n\n      const current = values[values.length - 1]!\n      const baseline = values[0]!\n      const drift = current - baseline\n      const reasons: string[] = []\n      if (convergence.state === 'drifting-down') {\n        reasons.push(\n          `convergence state drifting-down (tail run ${convergence.tailRun}, window mean ${convergence.windowMean.toFixed(3)})`,\n        )\n      }\n      if (metric === 'irr' && current < minIrr) {\n        reasons.push(`irr ${current.toFixed(3)} below floor ${minIrr}`)\n      }\n      if (metric === 'calibrationKappa' && baseline - current > maxKappaDrop) {\n        reasons.push(\n          `calibrationKappa dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxKappaDrop})`,\n        )\n      }\n      if (metric === 'sentinelPassRate' && baseline - current > maxSentinelDrop) {\n        reasons.push(\n          `sentinelPassRate dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxSentinelDrop})`,\n        )\n      }\n\n      const alarmed = reasons.length > 0\n      const trend: SentinelTrend = {\n        judgeId,\n        metric,\n        state: convergence.state,\n        current,\n        baseline,\n        drift,\n        alarmed,\n      }\n      if (alarmed) {\n        trend.reason = reasons.join('; ')\n        alarms.push(`judge \"${judgeId}\" ${metric}: ${trend.reason}`)\n      }\n      perJudge.push(trend)\n    }\n  }\n\n  return { perJudge, alarms, healthy: alarms.length === 0, insufficientHistory }\n}\n\n// ── Verdict stamp ────────────────────────────────────────────────────\n\nexport interface EvalHealthStamp {\n  healthy: boolean\n  alarms: string[]\n}\n\n/**\n * The object a campaign attaches to its verdicts. Compute it from the\n * sentinel report in a post-campaign hook (or a nightly job feeding the\n * next day's campaigns) and store it alongside the verdict payload.\n * `healthy: false` means the judges that produced those verdicts have an\n * unresolved drift / decay / silent-upgrade / staleness alarm — treat the\n * verdicts as carrying an integrity warning until recalibration clears it.\n */\nexport function evalHealthStamp(report: SentinelReport): EvalHealthStamp {\n  return { healthy: report.healthy, alarms: [...report.alarms] }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAyDA,eAAsB,iBACpB,YACA,cACA,aACA,oBACA,UAAmC,CAAC,GACH;CACjC,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,MAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;CAC9D,MAAM,gCAAgB,IAAI,IAAiC;CAC3D,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;EAC3C,IAAI,KAAK,CAAC;EACV,cAAc,IAAI,EAAE,OAAO,GAAG;CAChC;CAEA,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,SAAS,QAAQ,mBAAmB;CAE1C,MAAM,QAA0F,CAAC;CACjG,KAAK,MAAM,MAAM,aACf,KAAK,MAAM,MAAM,oBACf,MAAM,KAAK;EAAE,YAAY,GAAG;EAAI,eAAe;EAAI,IAAI,CAAC;EAAG,IAAI,CAAC;CAAE,CAAC;CAIvE,IAAI,SAAS;CACb,IAAI,UAAU;CACd,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,cAAc,IAAI,IAAI,KAAK;EACtC,IAAI,CAAC,MAAM,GAAG,WAAW,GAAG;GAC1B;GACA;EACF;EACA,MAAM,WAAW,GAAG,QAAQ,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;EACxE,IAAI,SAAS,WAAW,GAAG;GACzB;GACA;EACF;EAEA,KAAK,MAAM,MAAM,aAAa;GAE5B,MAAM,IAAI,OADM,GAAG,WAAW,eAAe,GAAG,EAAE,EAAA,CAC1B,KAAK,UAAU;GACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;GAEvC,KAAK,MAAM,MAAM,oBAAoB;IACnC,MAAM,SAAS,SACZ,KAAK,MAAM,EAAE,QAAQ,GAAG,CAAC,CACzB,QAAQ,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;IACzE,IAAI,OAAO,WAAW,GAAG;IACzB,MAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;IAC5C,IAAI,MAAM,MAAM;IAChB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;IAC/E,KAAK,GAAG,KAAK,CAAC;IACd,KAAK,GAAG,KAAK,CAAC;GAChB;EACF;EACA;CACF;CAqBA,OAAO;EAAE,OAnB4B,MAClC,QAAQ,MAAM,EAAE,GAAG,UAAU,CAAC,CAAC,CAC/B,KAAK,MAAM;GACV,MAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;GACnC,MAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;GACrC,MAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;GACrF,MAAM,UACJ,KAAK,IAAI,OAAO,KAAK,KAAM,WAAW,KAAK,IAAI,OAAO,KAAK,KAAM,aAAa;GAChF,OAAO;IACL,YAAY,EAAE;IACd,eAAe,EAAE;IACjB,GAAG,EAAE,GAAG;IACR;IACA;IACA;IACA;GACF;EACF,CAEoB;EAAG,eAAe;EAAQ,aAAa;CAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;CACf,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,SAAS,QAAQ,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;CACvE,IAAI,SAAS,OAAO,OAAO,KAAK,IAAI,GAAG,MAAM;CAE7C,MAAM,SAAS,CAAC,GAAG,QAAQ,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,CAAC,CAAC;CACzE,IAAI,CAAC,QAAQ,OAAO;CACpB,MAAM,YAAY,OAAO,KAAK,OAAO,OAAO,CAAC,CAAC;CAC9C,MAAM,IAAI,cAAc,KAAA,IAAY,OAAO,QAAQ,aAAa,KAAA;CAEhE,MAAM,SAAS,SACZ,KAAK,MAAM;EACV,MAAM,IAAI,OAAO,KAAK,EAAE,OAAO,CAAC,CAAC;EACjC,OAAO;GACL,IAAI,EAAE;GACN,GAAG,MAAM,KAAA,IAAY,OAAO,MAAM,MAAM,EAAE,QAAQ,OAAO,CAAC,IAAI,KAAA;EAChE;CACF,CAAC,CAAC,CACD,QAAQ,MAAM,EAAE,MAAM,KAAA,CAAS;CAClC,IAAI,OAAO,WAAW,GAAG,OAAO,KAAK;CACrC,OAAO,OAAO,MAAM,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,KAAK;AACrD;AAEA,SAAS,mBACP,IACA,IACA,YACkC;CAClC,MAAM,IAAI,GAAG;CACb,IAAI,IAAI,GAAG,OAAO;EAAE,OAAO;EAAK,OAAO;CAAI;CAC3C,MAAM,KAAe,CAAC;CACtB,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,KAAe,IAAI,MAAM,CAAC;EAChC,MAAM,KAAe,IAAI,MAAM,CAAC;EAChC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;GACxC,GAAG,KAAK,GAAG;GACX,GAAG,KAAK,GAAG;EACb;EACA,MAAM,IAAI,SAAS,IAAI,EAAE;EACzB,IAAI,OAAO,SAAS,CAAC,GAAG,GAAG,KAAK,CAAC;CACnC;CACA,GAAG,MAAM,GAAG,MAAM,IAAI,CAAC;CACvB,IAAI,GAAG,WAAW,GAAG,OAAO;EAAE,OAAO;EAAK,OAAO;CAAI;CACrD,OAAO;EACL,OAAO,GAAG,KAAK,MAAM,OAAQ,GAAG,MAAM;EACtC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,OAAQ,GAAG,MAAM,CAAC;CACjE;AACF;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACpKA,MAAa,wBAAwB;CAAC;CAAO;CAAoB;AAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;CACtD,MAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;CAC3D,IAAI,CAAC,OAAO,SAAS,EAAE,GACrB,MAAM,IAAI,gBACR,mBAAmB,MAAM,qCAAqC,KAAK,UAAU,KAAK,GACpF;CAEF,OAAO;AACT;;AAGA,SAAgB,yBAAyB,UAA4B,SAAS,YAAkB;CAC9F,SAAS,SAAS,IAAI,GAAG,OAAO,IAAI;CACpC,IAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GACtE,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,oCAAoC;CAE1F,IAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAC5E,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,uCAAuC;CAE7F,IAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAC3D,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,0CAA0C;CAEhG,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,SAAS,OAAO,GAAG;EAC3D,IAAI,CAAE,sBAA4C,SAAS,GAAG,GAE5D,MAAM,IAAI,gBACR,mBAAmB,OAAO,gCAAgC,IAAI,aAAa,sBAAsB,KAAK,IAAI,GAC5G;EAEF,IAAI,UAAU,KAAA,KAAa,CAAC,OAAO,SAAS,KAAK,GAC/C,MAAM,IAAI,gBACR,mBAAmB,OAAO,WAAW,IAAI,2BAA2B,OAAO,KAAK,GAClF;CAEJ;AACF;;;;;;AASA,SAAgB,wBACd,QACA,MACkB;CAClB,MAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;CAC5F,IAAI,CAAC,OAAO,SAAS,KAAK,GACxB,MAAM,IAAI,gBACR,+CAA+C,OAAO,EAAE,iEAC1D;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,kBAAkB,MAAM;CAAE;CACnF,yBAAyB,QAAQ;CACjC,OAAO;AACT;;;;;;;;AASA,SAAgB,sBACd,QACA,MACkB;CAClB,MAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;CAChE,IAAI,CAAC,OAAO,SAAS,GAAG,GACtB,MAAM,IAAI,gBACR,4FACF;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,IAAI;CAAE;CAC/D,yBAAyB,QAAQ;CACjC,OAAO;AACT;;;;;;;;AAcA,SAAgB,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;CAClB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAC7C,MAAM,IAAI,gBACR,uEAAuE,WACzE;CAEF,MAAM,6BAAa,IAAI,IAAoB;CAC3C,KAAK,MAAM,QAAQ,QAAQ;EACzB,IAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAClC,MAAM,IAAI,gBACR,yCAAyC,KAAK,OAAO,8BACvD;EAEF,IAAI,WAAW,IAAI,KAAK,MAAM,GAC5B,MAAM,IAAI,gBAAgB,qDAAqD,KAAK,OAAO,EAAE;EAE/F,WAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;CAC7C;CACA,IAAI,SAAS;CACb,IAAI,SAAS;CACb,KAAK,MAAM,KAAK,QAAQ;EACtB,MAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;EACrC,IAAI,UAAU,KAAA,GAAW;EACzB,IAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAC1B,MAAM,IAAI,gBACR,kDAAkD,EAAE,OAAO,gBAC7D;EAEF;EACA,IAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,WAAW;CAC9C;CACA,IAAI,WAAW,GACb,MAAM,IAAI,gBACR,yGACF;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO;CAAE;CAC7F,yBAAyB,QAAQ;CACjC,OAAO;AACT;AAUA,SAAgB,sBAAsB,UAA8B,CAAC,GAAkB;CACrF,KAAK,MAAM,KAAK,SAAS,yBAAyB,CAAC;CACnD,MAAM,QAAQ,QAAQ,KAAK,OAAO;EAAE,GAAG;EAAG,SAAS,EAAE,GAAG,EAAE,QAAQ;CAAE,EAAE;CACtE,OAAO;EACL,MAAM,OAAO,UAAU;GACrB,yBAAyB,QAAQ;GACjC,MAAM,KAAK;IAAE,GAAG;IAAU,SAAS,EAAE,GAAG,SAAS,QAAQ;GAAE,CAAC;EAC9D;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,MAAM,MAAM,KAAK,OAAO;IAAE,GAAG;IAAG,SAAS,EAAE,GAAG,EAAE,QAAQ;GAAE,EAAE;GAClE,OAAO,YAAY,KAAA,IAAY,MAAM,IAAI,QAAQ,MAAM,EAAE,YAAY,OAAO;EAC9E;CACF;AACF;;;;;;;AAQA,SAAgB,kBAAkB,MAA6B;CAC7D,OAAO;EACL,MAAM,OAAO,UAAU;GACrB,yBAAyB,QAAQ;GACjC,MAAM,KAAK,MAAM,OAAO;GACxB,MAAM,UAAU,MAAM,OAAO;GAC7B,MAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;GACzD,MAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,EAAE,KAAK,MAAM;EACnE;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,KAAK,MAAM,OAAO;GACxB,IAAI;GACJ,IAAI;IACF,MAAM,MAAM,GAAG,SAAS,MAAM,MAAM;GACtC,SAAS,KAAK;IACZ,IAAK,IAA8B,SAAS,UAAU,OAAO,CAAC;IAC9D,MAAM;GACR;GACA,MAAM,QAAQ,IAAI,MAAM,IAAI;GAC5B,MAAM,YAAgC,CAAC;GACvC,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;IACrC,MAAM,OAAO,MAAM;IACnB,IAAI,KAAK,KAAK,CAAC,CAAC,WAAW,GAAG;IAC9B,IAAI;IACJ,IAAI;KACF,SAAS,KAAK,MAAM,IAAI;IAC1B,SAAS,KAAK;KACZ,MAAM,IAAI,gBACR,4BAA4B,KAAK,4BAA4B,IAAI,EAAE,IAAK,IAAc,SACxF;IACF;IACA,MAAM,WAAW;IACjB,yBAAyB,UAAU,GAAG,KAAK,GAAG,IAAI,GAAG;IACrD,UAAU,KAAK,QAAQ;GACzB;GACA,OAAO,YAAY,KAAA,IAAY,YAAY,UAAU,QAAQ,MAAM,EAAE,YAAY,OAAO;EAC1F;CACF;AACF;AAmDA,MAAM,aAAa;AAEnB,SAAgB,oBACd,SACA,MACgB;CAChB,MAAM,SAAS,SAAS,KAAK,MAAM,MAAM;CACzC,MAAM,SAAS,KAAK,YAAY,UAAU;CAC1C,MAAM,eAAe,KAAK,YAAY,gBAAgB;CACtD,MAAM,kBAAkB,KAAK,YAAY,mBAAmB;CAC5D,MAAM,iBAAiB,KAAK,YAAY,kBAAkB;CAE1D,MAAM,0BAAU,IAAI,IAAwD;CAC5E,KAAK,MAAM,YAAY,SAAS;EAC9B,yBAAyB,QAAQ;EACjC,MAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,QAAQ,GAAG;EAC9E,MAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;EAC9C,IAAI,KAAK;GAAE,GAAG;GAAU;EAAK,CAAC;EAC9B,QAAQ,IAAI,SAAS,SAAS,GAAG;CACnC;CAEA,MAAM,WAA4B,CAAC;CACnC,MAAM,SAAmB,CAAC;CAC1B,MAAM,sBAAgC,CAAC;CAEvC,KAAK,MAAM,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,KAAK,GAAG;EAChD,MAAM,QAAQ,QAAQ,IAAI,OAAO,CAAC,CAAE,MAAM,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;EAClE,MAAM,SAAS,MAAM,MAAM,SAAS;EAEpC,MAAM,WAAW,SAAS,OAAO,QAAQ;EACzC,IAAI,UAAU,gBACZ,OAAO,KACL,UAAU,QAAQ,6BAA6B,OAAO,GAAG,MAAM,QAAQ,QAAQ,CAAC,EAAE,uBAAuB,eAAe,GAC1H;EAMF,IAAI,YAAY;EAChB,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAChC,IAAI,MAAM,EAAE,CAAE,eAAe,MAAM,IAAI,EAAE,CAAE,YAAY,YAAY;EAErE,IAAI,aAAa,GAQX;OAAA,CAPiB,MAClB,MAAM,SAAS,CAAC,CAChB,MACE,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB,CAEhC,GACd,OAAO,KACL,UAAU,QAAQ,oBAAoB,MAAM,YAAY,EAAE,CAAE,WAAW,OAAO,MAAM,UAAU,CAAE,WAAW,OAAO,MAAM,UAAU,CAAE,GAAG,gIACzI;EAAA;EAIJ,KAAK,MAAM,UAAU,uBAAuB;GAC1C,MAAM,SAAS,MACZ,QAAQ,MAAM,OAAO,EAAE,QAAQ,YAAY,QAAQ,CAAC,CACpD,KAAK,MAAM,EAAE,QAAQ,OAAQ;GAChC,IAAI,OAAO,WAAW,GAAG;GAEzB,MAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;GAC1D,IAAI,YAAY,UAAU,qBACxB,oBAAoB,KAAK,GAAG,QAAQ,GAAG,QAAQ;GAEjD,MAAM,UAAU,OAAO,OAAO,SAAS;GACvC,MAAM,WAAW,OAAO;GACxB,MAAM,QAAQ,UAAU;GACxB,MAAM,UAAoB,CAAC;GAC3B,IAAI,YAAY,UAAU,iBACxB,QAAQ,KACN,6CAA6C,YAAY,QAAQ,gBAAgB,YAAY,WAAW,QAAQ,CAAC,EAAE,EACrH;GAEF,IAAI,WAAW,SAAS,UAAU,QAChC,QAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,EAAE,eAAe,QAAQ;GAEhE,IAAI,WAAW,sBAAsB,WAAW,UAAU,cACxD,QAAQ,KACN,6BAA6B,WAAW,QAAA,CAAS,QAAQ,CAAC,EAAE,iBAAiB,SAAS,QAAQ,CAAC,EAAE,UAAU,aAAa,EAC1H;GAEF,IAAI,WAAW,sBAAsB,WAAW,UAAU,iBACxD,QAAQ,KACN,6BAA6B,WAAW,QAAA,CAAS,QAAQ,CAAC,EAAE,iBAAiB,SAAS,QAAQ,CAAC,EAAE,UAAU,gBAAgB,EAC7H;GAGF,MAAM,UAAU,QAAQ,SAAS;GACjC,MAAM,QAAuB;IAC3B;IACA;IACA,OAAO,YAAY;IACnB;IACA;IACA;IACA;GACF;GACA,IAAI,SAAS;IACX,MAAM,SAAS,QAAQ,KAAK,IAAI;IAChC,OAAO,KAAK,UAAU,QAAQ,IAAI,OAAO,IAAI,MAAM,QAAQ;GAC7D;GACA,SAAS,KAAK,KAAK;EACrB;CACF;CAEA,OAAO;EAAE;EAAU;EAAQ,SAAS,OAAO,WAAW;EAAG;CAAoB;AAC/E;;;;;;;;;AAiBA,SAAgB,gBAAgB,QAAyC;CACvE,OAAO;EAAE,SAAS,OAAO;EAAS,QAAQ,CAAC,GAAG,OAAO,MAAM;CAAE;AAC/D"}
+{"version":3,"file":"index.js","names":["defaultExtract"],"sources":["../../src/meta-eval/calibration.ts","../../src/meta-eval/correlation-study.ts","../../src/meta-eval/sentinel.ts"],"sourcesContent":["/**\n * Calibration curve — binned \"if eval says X, what does reality show?\"\n *\n * Companion to correlationStudy. Raw correlation is a single number;\n * the calibration curve shows *where* the eval is well-calibrated vs\n * overconfident / underconfident. Buckets the eval metric, computes\n * mean outcome per bucket, reports expected-calibration-error (ECE).\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { EvalMetricSpec } from './correlation-study'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface CalibrationBin {\n  lower: number\n  upper: number\n  n: number\n  evalMean: number\n  outcomeMean: number\n  /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */\n  gap: number\n}\n\nexport interface CalibrationReport {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  bins: CalibrationBin[]\n  /** Expected Calibration Error — Σ (n_i/N) × |outcomeMean_i − evalMean_i|. */\n  ece: number\n  /** Max bin gap — upper bound on miscalibration. */\n  maxGap: number\n}\n\nexport interface CalibrationOptions {\n  bins?: number\n  /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */\n  binning?: 'equal-width' | 'equal-frequency'\n  /** Clip eval values to [lo, hi] before binning. */\n  range?: { lo: number; hi: number }\n}\n\nexport interface CalibrationPair {\n  evalScore: number\n  outcome: number\n}\n\nexport async function calibrationCurve(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetric: EvalMetricSpec,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): Promise<CalibrationReport | null> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list()\n  const byRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = byRun.get(o.runId) ?? []\n    arr.push(o)\n    byRun.set(o.runId, arr)\n  }\n\n  const extract = evalMetric.extract ?? defaultExtract(evalMetric.id)\n  const pairs: Array<{ x: number; y: number }> = []\n  for (const run of runs) {\n    const os = byRun.get(run.runId)\n    if (!os?.length) continue\n    const x = await extract(run, traceStore)\n    if (x === null || !Number.isFinite(x)) continue\n    const latest = [...os].sort((a, b) => b.capturedAt - a.capturedAt)[0]!\n    const y = latest.metrics[outcomeMetric]\n    if (typeof y !== 'number' || !Number.isFinite(y)) continue\n    pairs.push({ x, y })\n  }\n  if (pairs.length < 2) return null\n\n  return calibrationFromPairs(\n    pairs.map((p) => ({ evalScore: p.x, outcome: p.y })),\n    evalMetric.id,\n    outcomeMetric,\n    options,\n  )\n}\n\nexport function calibrationFromPairs(\n  inputPairs: CalibrationPair[],\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): CalibrationReport | null {\n  const pairs = inputPairs.filter(\n    (pair) => Number.isFinite(pair.evalScore) && Number.isFinite(pair.outcome),\n  )\n  if (pairs.length < 2) return null\n\n  const numBins = options.bins ?? 10\n  const binning = options.binning ?? 'equal-width'\n  const xs = pairs.map((p) => p.evalScore)\n  const lo = options.range?.lo ?? Math.min(...xs)\n  const hi = options.range?.hi ?? Math.max(...xs)\n\n  const bins: CalibrationBin[] = []\n  if (binning === 'equal-frequency') {\n    const sorted = [...pairs].sort((a, b) => a.evalScore - b.evalScore)\n    const perBin = Math.max(1, Math.floor(sorted.length / numBins))\n    for (let i = 0; i < sorted.length; i += perBin) {\n      const chunk = sorted.slice(i, i + perBin)\n      if (chunk.length === 0) continue\n      bins.push(toBin(chunk))\n    }\n  } else {\n    const width = (hi - lo) / numBins\n    if (width === 0) return null\n    for (let i = 0; i < numBins; i++) {\n      const binLo = lo + i * width\n      const binHi = i === numBins - 1 ? hi + 1e-9 : lo + (i + 1) * width\n      const chunk = pairs.filter((p) => p.evalScore >= binLo && p.evalScore < binHi)\n      if (chunk.length === 0) continue\n      bins.push(toBin(chunk, binLo, binHi))\n    }\n  }\n\n  const total = bins.reduce((a, b) => a + b.n, 0)\n  const ece = bins.reduce((a, b) => a + (b.n / total) * b.gap, 0)\n  const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0)\n\n  return { evalMetric, outcomeMetric, n: pairs.length, bins, ece, maxGap }\n}\n\nfunction toBin(chunk: CalibrationPair[], lower?: number, upper?: number): CalibrationBin {\n  const xs = chunk.map((c) => c.evalScore)\n  const ys = chunk.map((c) => c.outcome)\n  const evalMean = mean(xs)\n  const outcomeMean = mean(ys)\n  return {\n    lower: lower ?? Math.min(...xs),\n    upper: upper ?? Math.max(...xs),\n    n: chunk.length,\n    evalMean,\n    outcomeMean,\n    gap: Math.abs(outcomeMean - evalMean),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((a, b) => a + b, 0) / xs.length\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run) =>\n    run.outcome?.score ?? (metric === 'pass' ? (run.outcome?.pass === true ? 1 : 0) : null)\n}\n","/**\n * Correlation study — \"does our eval score predict real-world outcomes?\"\n *\n * This is the load-bearing signal. Takes a TraceStore + OutcomeStore,\n * joins on runId, computes Pearson + Spearman + bootstrap CI for every\n * (evalMetric, outcomeMetric) pair the caller declares.\n *\n * Without this number the framework is ornamental. With it and r > 0.6\n * the framework is a moat — no other agent-eval tool publishes one.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport { aggregateLlm, llmSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { DeploymentOutcome, OutcomeFilter, OutcomeStore } from './outcome-store'\n\nexport interface EvalMetricSpec {\n  id: string\n  /** Extract a scalar from a run (defaults cover score/pass/durationMs/costUsd/tokens). */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface OutcomePair {\n  evalMetric: string\n  outcomeMetric: string\n}\n\nexport interface CorrelationResult {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  pearson: number\n  spearman: number\n  /** 95% bootstrap CI for Pearson. */\n  pearsonCi95: { lower: number; upper: number }\n  /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */\n  verdict: 'strong' | 'moderate' | 'weak'\n}\n\nexport interface CorrelationStudyResult {\n  pairs: CorrelationResult[]\n  joinedSamples: number\n  skippedRuns: number\n}\n\nexport interface CorrelationStudyOptions {\n  /** Only join outcomes captured within this window after run.startedAt. */\n  maxCaptureLagMs?: number\n  /** Restrict to a subset of outcomes (cohort, region, source). */\n  outcomeFilter?: OutcomeFilter\n  /** Which outcome per run to use when multiple exist. Default 'latest'. */\n  reduction?: 'latest' | 'mean' | 'max'\n  /** Bootstrap iterations for the CI. Default 500. */\n  bootstrapIterations?: number\n}\n\nexport async function correlationStudy(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetrics: EvalMetricSpec[],\n  outcomeMetricNames: string[],\n  options: CorrelationStudyOptions = {},\n): Promise<CorrelationStudyResult> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list(options.outcomeFilter)\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  const reduction = options.reduction ?? 'latest'\n  const maxLag = options.maxCaptureLagMs ?? Infinity\n\n  const pairs: Array<{ evalMetric: string; outcomeMetric: string; xs: number[]; ys: number[] }> = []\n  for (const em of evalMetrics) {\n    for (const om of outcomeMetricNames) {\n      pairs.push({ evalMetric: em.id, outcomeMetric: om, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    const eligible = os.filter((o) => o.capturedAt - run.startedAt <= maxLag)\n    if (eligible.length === 0) {\n      skipped++\n      continue\n    }\n\n    for (const em of evalMetrics) {\n      const extract = em.extract ?? defaultExtract(em.id)\n      const x = await extract(run, traceStore)\n      if (x === null || !Number.isFinite(x)) continue\n\n      for (const om of outcomeMetricNames) {\n        const values = eligible\n          .map((o) => o.metrics[om])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, reduction, eligible)\n        if (y === null) continue\n        const pair = pairs.find((p) => p.evalMetric === em.id && p.outcomeMetric === om)!\n        pair.xs.push(x)\n        pair.ys.push(y)\n      }\n    }\n    joined++\n  }\n\n  const results: CorrelationResult[] = pairs\n    .filter((p) => p.xs.length >= 3)\n    .map((p) => {\n      const pearson = pearsonR(p.xs, p.ys)\n      const spearman = spearmanR(p.xs, p.ys)\n      const pearsonCi95 = bootstrapPearsonCi(p.xs, p.ys, options.bootstrapIterations ?? 500)\n      const verdict: CorrelationResult['verdict'] =\n        Math.abs(pearson) >= 0.7 ? 'strong' : Math.abs(pearson) >= 0.4 ? 'moderate' : 'weak'\n      return {\n        evalMetric: p.evalMetric,\n        outcomeMetric: p.outcomeMetric,\n        n: p.xs.length,\n        pearson,\n        spearman,\n        pearsonCi95,\n        verdict,\n      }\n    })\n\n  return { pairs: results, joinedSamples: joined, skippedRuns: skipped }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  kind: 'latest' | 'mean' | 'max',\n  outcomes: DeploymentOutcome[],\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((a, b) => a + b, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest': pick the outcome captured last, then lookup its metric\n  const latest = [...outcomes].sort((a, b) => b.capturedAt - a.capturedAt)[0]\n  if (!latest) return null\n  const latestKey = Object.keys(latest.metrics)[0]\n  const v = latestKey !== undefined ? latest.metrics[latestKey] : undefined\n  // For 'latest' we already have `values` aligned; use the last-captured one\n  const paired = outcomes\n    .map((o) => {\n      const k = Object.keys(o.metrics)[0]\n      return {\n        at: o.capturedAt,\n        v: k !== undefined ? values.find((x) => o.metrics[k] === x) : undefined,\n      }\n    })\n    .filter((p) => p.v !== undefined)\n  if (paired.length === 0) return v ?? null\n  return paired.sort((a, b) => b.at - a.at)[0]?.v ?? null\n}\n\nfunction bootstrapPearsonCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n): { lower: number; upper: number } {\n  const n = xs.length\n  if (n < 3) return { lower: NaN, upper: NaN }\n  const rs: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx: number[] = new Array(n)\n    const ry: number[] = new Array(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(Math.random() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) rs.push(r)\n  }\n  rs.sort((a, b) => a - b)\n  if (rs.length === 0) return { lower: NaN, upper: NaN }\n  return {\n    lower: rs[Math.floor(0.025 * rs.length)]!,\n    upper: rs[Math.min(rs.length - 1, Math.floor(0.975 * rs.length))]!,\n  }\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * Judge sentinel — eval trustworthiness as a continuously measured,\n * alarmed trend.\n *\n * Judges are models; models change underneath us; calibration decays.\n * This module composes three existing instruments into one loop:\n *\n *   snapshot — adapters turn real instrument outputs (`calibrateJudge` /\n *              `calibrateJudgeContinuous`, `corpusInterRaterAgreement` /\n *              `continuousAgreement`, a rerun sentinel golden set) into\n *              `SentinelSnapshot`s\n *   series   — `analyzeSeries` (src/series-convergence) classifies each\n *              judge×metric history as stabilized / drifting / noisy\n *   alarm    — `judgeSentinelReport` turns trends + thresholds into named\n *              alarms; `evalHealthStamp` is the tiny object a campaign\n *              attaches to its verdicts\n *\n * Alarm conditions:\n *   - convergence state `drifting-down` on any tracked metric\n *   - irr below `minIrr`\n *   - calibrationKappa / sentinelPassRate dropped vs the series baseline\n *     beyond `maxKappaDrop` / `maxSentinelDrop`\n *   - judge model changed with no post-change golden-grounded snapshot\n *     (the silent-upgrade trap: judges agreeing with each other after a\n *     model swap proves nothing — only re-measuring against gold does)\n *   - newest snapshot older than `staleAfterDays` relative to `asOf`\n *\n * Wire-in: run the snapshot adapters from a post-campaign hook or a\n * nightly job, append to a `SentinelStore`, then compute\n * `judgeSentinelReport(await store.history(), { asOf })` with `asOf` =\n * the campaign timestamp. Attach `evalHealthStamp(report)` to campaign\n * verdicts: an alarmed sentinel means every downstream verdict carries\n * an integrity warning until the judge is recalibrated.\n *\n * The module is clock-free — every timestamp (`at`, `asOf`) is\n * caller-supplied ISO-8601, so reports are reproducible.\n */\n\nimport { ValidationError } from '../errors'\nimport type {\n  CalibrationResult,\n  CandidateScore,\n  ContinuousAgreement,\n  ContinuousCalibrationResult,\n  GoldenItem,\n} from '../judge-calibration'\nimport {\n  analyzeSeries,\n  type SeriesConvergenceOptions,\n  type SeriesConvergenceResult,\n} from '../series-convergence'\nimport type { CorpusAgreementReport } from '../statistics'\n\nexport const SENTINEL_METRIC_NAMES = ['irr', 'calibrationKappa', 'sentinelPassRate'] as const\n\nexport type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number]\n\nexport interface SentinelMetrics {\n  /** Inter-rater reliability — ICC(2,1) from agreement instruments. */\n  irr?: number\n  /** Weighted κ vs the human golden set (`calibrateJudge*`). */\n  calibrationKappa?: number\n  /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */\n  sentinelPassRate?: number\n}\n\nexport interface SentinelSnapshot {\n  /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */\n  at: string\n  judgeId: string\n  /** Exact model identity behind the judge (pin the full version string). */\n  judgeModel: string\n  /**\n   * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a\n   * model-change marker — it records the new `judgeModel` immediately and\n   * the silent-upgrade alarm stays raised until a golden-grounded snapshot\n   * (calibrationKappa or sentinelPassRate) follows.\n   */\n  metrics: SentinelMetrics\n}\n\n/** Identity + timestamp the caller supplies alongside an instrument output. */\nexport interface SnapshotMeta {\n  at: string\n  judgeId: string\n  judgeModel: string\n}\n\nfunction parseIso(value: string, label: string): number {\n  const ms = typeof value === 'string' ? Date.parse(value) : NaN\n  if (!Number.isFinite(ms)) {\n    throw new ValidationError(\n      `judge-sentinel: ${label} is not a parseable ISO timestamp: ${JSON.stringify(value)}`,\n    )\n  }\n  return ms\n}\n\n/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */\nexport function validateSentinelSnapshot(snapshot: SentinelSnapshot, source = 'snapshot'): void {\n  parseIso(snapshot.at, `${source}.at`)\n  if (typeof snapshot.judgeId !== 'string' || snapshot.judgeId.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeId must be a non-empty string`)\n  }\n  if (typeof snapshot.judgeModel !== 'string' || snapshot.judgeModel.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeModel must be a non-empty string`)\n  }\n  if (snapshot.metrics === null || typeof snapshot.metrics !== 'object') {\n    throw new ValidationError(`judge-sentinel: ${source}.metrics must be an object (may be empty)`)\n  }\n  for (const [key, value] of Object.entries(snapshot.metrics)) {\n    if (!(SENTINEL_METRIC_NAMES as readonly string[]).includes(key)) {\n      // A typo'd key would silently never trend — reject instead.\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics carries unknown key \"${key}\" — known: ${SENTINEL_METRIC_NAMES.join(', ')}`,\n      )\n    }\n    if (value !== undefined && !Number.isFinite(value)) {\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics.${key} is not a finite number: ${String(value)}`,\n      )\n    }\n  }\n}\n\n// ── Snapshot adapters — real instrument output → SentinelSnapshot ───\n\n/**\n * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the\n * un-rounded continuous κ_w when the report carries one (integer κ\n * discards information for [0,1] judges).\n */\nexport function snapshotFromCalibration(\n  report: CalibrationResult | ContinuousCalibrationResult,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const kappa = 'weightedKappaContinuous' in report ? report.weightedKappaContinuous : report.kappa\n  if (!Number.isFinite(kappa)) {\n    throw new ValidationError(\n      `snapshotFromCalibration: κ is not finite (n=${report.n}) — calibrate against ≥2 joined golden items before snapshotting`,\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { calibrationKappa: kappa } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n/**\n * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`\n * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability\n * coefficient both instruments compute. For ensemble-level agreement,\n * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its\n * composition (e.g. a joined list of member model strings).\n */\nexport function snapshotFromAgreement(\n  report: CorpusAgreementReport | ContinuousAgreement,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const irr = 'overallIcc' in report ? report.overallIcc : report.icc\n  if (!Number.isFinite(irr)) {\n    throw new ValidationError(\n      'snapshotFromAgreement: ICC is not finite — agreement needs ≥2 raters and ≥2 complete items',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { irr } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\nexport interface SentinelSetOptions {\n  /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */\n  tolerance?: number\n}\n\n/**\n * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the\n * judge was originally calibrated on, re-scored periodically. Pass rate =\n * fraction of joined items where the judge stays within `tolerance` of the\n * human score. Items the judge didn't score are excluded from the\n * denominator; zero joined items is an error, not a 0% pass rate.\n */\nexport function snapshotFromSentinelSet(\n  scores: CandidateScore[],\n  golden: GoldenItem[],\n  meta: SnapshotMeta,\n  options: SentinelSetOptions = {},\n): SentinelSnapshot {\n  const tolerance = options.tolerance ?? 0.1\n  if (!Number.isFinite(tolerance) || tolerance < 0) {\n    throw new ValidationError(\n      `snapshotFromSentinelSet: tolerance must be a finite number ≥ 0, got ${tolerance}`,\n    )\n  }\n  const goldenById = new Map<string, number>()\n  for (const item of golden) {\n    if (!Number.isFinite(item.humanScore)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: golden item \"${item.itemId}\" has a non-finite humanScore`,\n      )\n    }\n    if (goldenById.has(item.itemId)) {\n      throw new ValidationError(`snapshotFromSentinelSet: duplicate golden itemId \"${item.itemId}\"`)\n    }\n    goldenById.set(item.itemId, item.humanScore)\n  }\n  let joined = 0\n  let passed = 0\n  for (const s of scores) {\n    const human = goldenById.get(s.itemId)\n    if (human === undefined) continue\n    if (!Number.isFinite(s.score)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: judge score for item \"${s.itemId}\" is not finite`,\n      )\n    }\n    joined++\n    if (Math.abs(s.score - human) <= tolerance) passed++\n  }\n  if (joined === 0) {\n    throw new ValidationError(\n      'snapshotFromSentinelSet: no judge scores joined the golden set — itemIds mismatch or empty sentinel run',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { sentinelPassRate: passed / joined } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n// ── Persistence ──────────────────────────────────────────────────────\n\nexport interface SentinelStore {\n  append(snapshot: SentinelSnapshot): Promise<void>\n  /** All snapshots (optionally for one judge), in append order. */\n  history(judgeId?: string): Promise<SentinelSnapshot[]>\n}\n\nexport function inMemorySentinelStore(initial: SentinelSnapshot[] = []): SentinelStore {\n  for (const s of initial) validateSentinelSnapshot(s)\n  const items = initial.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      items.push({ ...snapshot, metrics: { ...snapshot.metrics } })\n    },\n    async history(judgeId) {\n      const all = items.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n      return judgeId === undefined ? all : all.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n/**\n * JSONL store — one snapshot per line, appended atomically per call.\n * A corrupt or shape-invalid line is a loud error naming the file and\n * line number: a sentinel history that silently drops records would\n * defeat the drift detection it exists to provide.\n */\nexport function fileSentinelStore(path: string): SentinelStore {\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      const fs = await import('node:fs/promises')\n      const pathMod = await import('node:path')\n      await fs.mkdir(pathMod.dirname(path), { recursive: true })\n      await fs.appendFile(path, `${JSON.stringify(snapshot)}\\n`, 'utf8')\n    },\n    async history(judgeId) {\n      const fs = await import('node:fs/promises')\n      let raw: string\n      try {\n        raw = await fs.readFile(path, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') return []\n        throw err\n      }\n      const lines = raw.split('\\n')\n      const snapshots: SentinelSnapshot[] = []\n      for (let i = 0; i < lines.length; i++) {\n        const line = lines[i]!\n        if (line.trim().length === 0) continue\n        let parsed: unknown\n        try {\n          parsed = JSON.parse(line)\n        } catch (err) {\n          throw new ValidationError(\n            `judge-sentinel: store at ${path} has a corrupt JSONL line ${i + 1}: ${(err as Error).message}`,\n          )\n        }\n        const snapshot = parsed as SentinelSnapshot\n        validateSentinelSnapshot(snapshot, `${path}:${i + 1}`)\n        snapshots.push(snapshot)\n      }\n      return judgeId === undefined ? snapshots : snapshots.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n// ── Trend report ─────────────────────────────────────────────────────\n\nexport interface SentinelThresholds {\n  /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */\n  minIrr?: number\n  /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */\n  maxKappaDrop?: number\n  /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */\n  maxSentinelDrop?: number\n  /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */\n  staleAfterDays?: number\n}\n\nexport interface JudgeSentinelOptions {\n  /** Caller-supplied ISO timestamp staleness is measured against. */\n  asOf: string\n  thresholds?: SentinelThresholds\n  /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */\n  convergence?: SeriesConvergenceOptions\n}\n\nexport interface SentinelTrend {\n  judgeId: string\n  metric: SentinelMetricName\n  /** Verbatim `analyzeSeries` state for this judge×metric history. */\n  state: SeriesConvergenceResult['state']\n  /** Newest value in the series. */\n  current: number\n  /** Oldest value in the series — the reference the drop thresholds compare against. */\n  baseline: number\n  /** current − baseline (signed; negative = decayed). */\n  drift: number\n  alarmed: boolean\n  reason?: string\n}\n\nexport interface SentinelReport {\n  perJudge: SentinelTrend[]\n  alarms: string[]\n  healthy: boolean\n  /**\n   * `judgeId:metric` pairs with too few snapshots for the convergence\n   * machine. Named, not counted — a blind spot you can't see is a blind\n   * spot you won't fix. Floor/drop checks still apply to thin series, so\n   * insufficient history alone never masks an alarm.\n   */\n  insufficientHistory: string[]\n}\n\nconst MS_PER_DAY = 86_400_000\n\nexport function judgeSentinelReport(\n  history: SentinelSnapshot[],\n  opts: JudgeSentinelOptions,\n): SentinelReport {\n  const asOfMs = parseIso(opts.asOf, 'asOf')\n  const minIrr = opts.thresholds?.minIrr ?? 0.6\n  const maxKappaDrop = opts.thresholds?.maxKappaDrop ?? 0.15\n  const maxSentinelDrop = opts.thresholds?.maxSentinelDrop ?? 0.1\n  const staleAfterDays = opts.thresholds?.staleAfterDays ?? 30\n\n  const byJudge = new Map<string, Array<SentinelSnapshot & { atMs: number }>>()\n  for (const snapshot of history) {\n    validateSentinelSnapshot(snapshot)\n    const atMs = parseIso(snapshot.at, `snapshot.at (judge \"${snapshot.judgeId}\")`)\n    const arr = byJudge.get(snapshot.judgeId) ?? []\n    arr.push({ ...snapshot, atMs })\n    byJudge.set(snapshot.judgeId, arr)\n  }\n\n  const perJudge: SentinelTrend[] = []\n  const alarms: string[] = []\n  const insufficientHistory: string[] = []\n\n  for (const judgeId of [...byJudge.keys()].sort()) {\n    const snaps = byJudge.get(judgeId)!.sort((a, b) => a.atMs - b.atMs)\n    const newest = snaps[snaps.length - 1]!\n\n    const ageDays = (asOfMs - newest.atMs) / MS_PER_DAY\n    if (ageDays > staleAfterDays) {\n      alarms.push(\n        `judge \"${judgeId}\": stale — newest snapshot ${newest.at} is ${ageDays.toFixed(1)}d old vs asOf (limit ${staleAfterDays}d)`,\n      )\n    }\n\n    // Silent-upgrade trap: only the latest model change matters for current\n    // trust, and only golden-grounded metrics clear it — irr alone proves\n    // judges still agree with each other, not that they're still right.\n    let changeIdx = -1\n    for (let i = 1; i < snaps.length; i++) {\n      if (snaps[i]!.judgeModel !== snaps[i - 1]!.judgeModel) changeIdx = i\n    }\n    if (changeIdx >= 0) {\n      const recalibrated = snaps\n        .slice(changeIdx)\n        .some(\n          (s) =>\n            Number.isFinite(s.metrics.calibrationKappa) ||\n            Number.isFinite(s.metrics.sentinelPassRate),\n        )\n      if (!recalibrated) {\n        alarms.push(\n          `judge \"${judgeId}\": model changed \"${snaps[changeIdx - 1]!.judgeModel}\" → \"${snaps[changeIdx]!.judgeModel}\" at ${snaps[changeIdx]!.at} with no post-change calibration snapshot — silent-upgrade trap; rerun calibrateJudge or the sentinel set against the new model`,\n        )\n      }\n    }\n\n    for (const metric of SENTINEL_METRIC_NAMES) {\n      const values = snaps\n        .filter((s) => typeof s.metrics[metric] === 'number')\n        .map((s) => s.metrics[metric]!)\n      if (values.length === 0) continue\n\n      const convergence = analyzeSeries(values, opts.convergence)\n      if (convergence.state === 'insufficient-data')\n        insufficientHistory.push(`${judgeId}:${metric}`)\n\n      const current = values[values.length - 1]!\n      const baseline = values[0]!\n      const drift = current - baseline\n      const reasons: string[] = []\n      if (convergence.state === 'drifting-down') {\n        reasons.push(\n          `convergence state drifting-down (tail run ${convergence.tailRun}, window mean ${convergence.windowMean.toFixed(3)})`,\n        )\n      }\n      if (metric === 'irr' && current < minIrr) {\n        reasons.push(`irr ${current.toFixed(3)} below floor ${minIrr}`)\n      }\n      if (metric === 'calibrationKappa' && baseline - current > maxKappaDrop) {\n        reasons.push(\n          `calibrationKappa dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxKappaDrop})`,\n        )\n      }\n      if (metric === 'sentinelPassRate' && baseline - current > maxSentinelDrop) {\n        reasons.push(\n          `sentinelPassRate dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxSentinelDrop})`,\n        )\n      }\n\n      const alarmed = reasons.length > 0\n      const trend: SentinelTrend = {\n        judgeId,\n        metric,\n        state: convergence.state,\n        current,\n        baseline,\n        drift,\n        alarmed,\n      }\n      if (alarmed) {\n        trend.reason = reasons.join('; ')\n        alarms.push(`judge \"${judgeId}\" ${metric}: ${trend.reason}`)\n      }\n      perJudge.push(trend)\n    }\n  }\n\n  return { perJudge, alarms, healthy: alarms.length === 0, insufficientHistory }\n}\n\n// ── Verdict stamp ────────────────────────────────────────────────────\n\nexport interface EvalHealthStamp {\n  healthy: boolean\n  alarms: string[]\n}\n\n/**\n * The object a campaign attaches to its verdicts. Compute it from the\n * sentinel report in a post-campaign hook (or a nightly job feeding the\n * next day's campaigns) and store it alongside the verdict payload.\n * `healthy: false` means the judges that produced those verdicts have an\n * unresolved drift / decay / silent-upgrade / staleness alarm — treat the\n * verdicts as carrying an integrity warning until recalibration clears it.\n */\nexport function evalHealthStamp(report: SentinelReport): EvalHealthStamp {\n  return { healthy: report.healthy, alarms: [...report.alarms] }\n}\n"],"mappings":";;;;;;;AAgDA,eAAsB,iBACpB,YACA,cACA,YACA,eACA,UAA8B,CAAC,GACI;CACnC,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,MAAM,WAAW,MAAM,aAAa,KAAK;CACzC,MAAM,wBAAQ,IAAI,IAAiC;CACnD,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;EACnC,IAAI,KAAK,CAAC;EACV,MAAM,IAAI,EAAE,OAAO,GAAG;CACxB;CAEA,MAAM,UAAU,WAAW,WAAWA,iBAAe,WAAW,EAAE;CAClE,MAAM,QAAyC,CAAC;CAChD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,MAAM,IAAI,IAAI,KAAK;EAC9B,IAAI,CAAC,IAAI,QAAQ;EACjB,MAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;EACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;EAEvC,MAAM,IADS,CAAC,GAAG,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,CAAC,CAAC,EACnD,CAAC,QAAQ;EACzB,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,MAAM,KAAK;GAAE;GAAG;EAAE,CAAC;CACrB;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,OAAO,qBACL,MAAM,KAAK,OAAO;EAAE,WAAW,EAAE;EAAG,SAAS,EAAE;CAAE,EAAE,GACnD,WAAW,IACX,eACA,OACF;AACF;AAEA,SAAgB,qBACd,YACA,YACA,eACA,UAA8B,CAAC,GACL;CAC1B,MAAM,QAAQ,WAAW,QACtB,SAAS,OAAO,SAAS,KAAK,SAAS,KAAK,OAAO,SAAS,KAAK,OAAO,CAC3E;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,MAAM,UAAU,QAAQ,QAAQ;CAChC,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAE9C,MAAM,OAAyB,CAAC;CAChC,IAAI,YAAY,mBAAmB;EACjC,MAAM,SAAS,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EAClE,MAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,OAAO,SAAS,OAAO,CAAC;EAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK,QAAQ;GAC9C,MAAM,QAAQ,OAAO,MAAM,GAAG,IAAI,MAAM;GACxC,IAAI,MAAM,WAAW,GAAG;GACxB,KAAK,KAAK,MAAM,KAAK,CAAC;EACxB;CACF,OAAO;EACL,MAAM,SAAS,KAAK,MAAM;EAC1B,IAAI,UAAU,GAAG,OAAO;EACxB,KAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;GAChC,MAAM,QAAQ,KAAK,IAAI;GACvB,MAAM,QAAQ,MAAM,UAAU,IAAI,KAAK,OAAO,MAAM,IAAI,KAAK;GAC7D,MAAM,QAAQ,MAAM,QAAQ,MAAM,EAAE,aAAa,SAAS,EAAE,YAAY,KAAK;GAC7E,IAAI,MAAM,WAAW,GAAG;GACxB,KAAK,KAAK,MAAM,OAAO,OAAO,KAAK,CAAC;EACtC;CACF;CAEA,MAAM,QAAQ,KAAK,QAAQ,GAAG,MAAM,IAAI,EAAE,GAAG,CAAC;CAC9C,MAAM,MAAM,KAAK,QAAQ,GAAG,MAAM,IAAK,EAAE,IAAI,QAAS,EAAE,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,GAAG,GAAG,CAAC;CAE1D,OAAO;EAAE;EAAY;EAAe,GAAG,MAAM;EAAQ;EAAM;EAAK;CAAO;AACzE;AAEA,SAAS,MAAM,OAA0B,OAAgB,OAAgC;CACvF,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,OAAO;CACrC,MAAM,WAAW,KAAK,EAAE;CACxB,MAAM,cAAc,KAAK,EAAE;CAC3B,OAAO;EACL,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,GAAG,MAAM;EACT;EACA;EACA,KAAK,KAAK,IAAI,cAAc,QAAQ;CACtC;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,OAAO,GAAG,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAASA,iBAAe,QAAyE;CAC/F,OAAO,OAAO,QACZ,IAAI,SAAS,UAAU,WAAW,SAAU,IAAI,SAAS,SAAS,OAAO,IAAI,IAAK;AACtF;;;;;;;;;;;;;AChGA,eAAsB,iBACpB,YACA,cACA,aACA,oBACA,UAAmC,CAAC,GACH;CACjC,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,MAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;CAC9D,MAAM,gCAAgB,IAAI,IAAiC;CAC3D,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;EAC3C,IAAI,KAAK,CAAC;EACV,cAAc,IAAI,EAAE,OAAO,GAAG;CAChC;CAEA,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,SAAS,QAAQ,mBAAmB;CAE1C,MAAM,QAA0F,CAAC;CACjG,KAAK,MAAM,MAAM,aACf,KAAK,MAAM,MAAM,oBACf,MAAM,KAAK;EAAE,YAAY,GAAG;EAAI,eAAe;EAAI,IAAI,CAAC;EAAG,IAAI,CAAC;CAAE,CAAC;CAIvE,IAAI,SAAS;CACb,IAAI,UAAU;CACd,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,cAAc,IAAI,IAAI,KAAK;EACtC,IAAI,CAAC,MAAM,GAAG,WAAW,GAAG;GAC1B;GACA;EACF;EACA,MAAM,WAAW,GAAG,QAAQ,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;EACxE,IAAI,SAAS,WAAW,GAAG;GACzB;GACA;EACF;EAEA,KAAK,MAAM,MAAM,aAAa;GAE5B,MAAM,IAAI,OADM,GAAG,WAAW,eAAe,GAAG,EAAE,EAAA,CAC1B,KAAK,UAAU;GACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;GAEvC,KAAK,MAAM,MAAM,oBAAoB;IACnC,MAAM,SAAS,SACZ,KAAK,MAAM,EAAE,QAAQ,GAAG,CAAC,CACzB,QAAQ,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;IACzE,IAAI,OAAO,WAAW,GAAG;IACzB,MAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;IAC5C,IAAI,MAAM,MAAM;IAChB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;IAC/E,KAAK,GAAG,KAAK,CAAC;IACd,KAAK,GAAG,KAAK,CAAC;GAChB;EACF;EACA;CACF;CAqBA,OAAO;EAAE,OAnB4B,MAClC,QAAQ,MAAM,EAAE,GAAG,UAAU,CAAC,CAAC,CAC/B,KAAK,MAAM;GACV,MAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;GACnC,MAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;GACrC,MAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;GACrF,MAAM,UACJ,KAAK,IAAI,OAAO,KAAK,KAAM,WAAW,KAAK,IAAI,OAAO,KAAK,KAAM,aAAa;GAChF,OAAO;IACL,YAAY,EAAE;IACd,eAAe,EAAE;IACjB,GAAG,EAAE,GAAG;IACR;IACA;IACA;IACA;GACF;EACF,CAEoB;EAAG,eAAe;EAAQ,aAAa;CAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;CACf,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,SAAS,QAAQ,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;CACvE,IAAI,SAAS,OAAO,OAAO,KAAK,IAAI,GAAG,MAAM;CAE7C,MAAM,SAAS,CAAC,GAAG,QAAQ,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,CAAC,CAAC;CACzE,IAAI,CAAC,QAAQ,OAAO;CACpB,MAAM,YAAY,OAAO,KAAK,OAAO,OAAO,CAAC,CAAC;CAC9C,MAAM,IAAI,cAAc,KAAA,IAAY,OAAO,QAAQ,aAAa,KAAA;CAEhE,MAAM,SAAS,SACZ,KAAK,MAAM;EACV,MAAM,IAAI,OAAO,KAAK,EAAE,OAAO,CAAC,CAAC;EACjC,OAAO;GACL,IAAI,EAAE;GACN,GAAG,MAAM,KAAA,IAAY,OAAO,MAAM,MAAM,EAAE,QAAQ,OAAO,CAAC,IAAI,KAAA;EAChE;CACF,CAAC,CAAC,CACD,QAAQ,MAAM,EAAE,MAAM,KAAA,CAAS;CAClC,IAAI,OAAO,WAAW,GAAG,OAAO,KAAK;CACrC,OAAO,OAAO,MAAM,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,KAAK;AACrD;AAEA,SAAS,mBACP,IACA,IACA,YACkC;CAClC,MAAM,IAAI,GAAG;CACb,IAAI,IAAI,GAAG,OAAO;EAAE,OAAO;EAAK,OAAO;CAAI;CAC3C,MAAM,KAAe,CAAC;CACtB,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,KAAe,IAAI,MAAM,CAAC;EAChC,MAAM,KAAe,IAAI,MAAM,CAAC;EAChC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;GACxC,GAAG,KAAK,GAAG;GACX,GAAG,KAAK,GAAG;EACb;EACA,MAAM,IAAI,SAAS,IAAI,EAAE;EACzB,IAAI,OAAO,SAAS,CAAC,GAAG,GAAG,KAAK,CAAC;CACnC;CACA,GAAG,MAAM,GAAG,MAAM,IAAI,CAAC;CACvB,IAAI,GAAG,WAAW,GAAG,OAAO;EAAE,OAAO;EAAK,OAAO;CAAI;CACrD,OAAO;EACL,OAAO,GAAG,KAAK,MAAM,OAAQ,GAAG,MAAM;EACtC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,OAAQ,GAAG,MAAM,CAAC;CACjE;AACF;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACpKA,MAAa,wBAAwB;CAAC;CAAO;CAAoB;AAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;CACtD,MAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;CAC3D,IAAI,CAAC,OAAO,SAAS,EAAE,GACrB,MAAM,IAAI,gBACR,mBAAmB,MAAM,qCAAqC,KAAK,UAAU,KAAK,GACpF;CAEF,OAAO;AACT;;AAGA,SAAgB,yBAAyB,UAA4B,SAAS,YAAkB;CAC9F,SAAS,SAAS,IAAI,GAAG,OAAO,IAAI;CACpC,IAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GACtE,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,oCAAoC;CAE1F,IAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAC5E,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,uCAAuC;CAE7F,IAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAC3D,MAAM,IAAI,gBAAgB,mBAAmB,OAAO,0CAA0C;CAEhG,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,SAAS,OAAO,GAAG;EAC3D,IAAI,CAAE,sBAA4C,SAAS,GAAG,GAE5D,MAAM,IAAI,gBACR,mBAAmB,OAAO,gCAAgC,IAAI,aAAa,sBAAsB,KAAK,IAAI,GAC5G;EAEF,IAAI,UAAU,KAAA,KAAa,CAAC,OAAO,SAAS,KAAK,GAC/C,MAAM,IAAI,gBACR,mBAAmB,OAAO,WAAW,IAAI,2BAA2B,OAAO,KAAK,GAClF;CAEJ;AACF;;;;;;AASA,SAAgB,wBACd,QACA,MACkB;CAClB,MAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;CAC5F,IAAI,CAAC,OAAO,SAAS,KAAK,GACxB,MAAM,IAAI,gBACR,+CAA+C,OAAO,EAAE,iEAC1D;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,kBAAkB,MAAM;CAAE;CACnF,yBAAyB,QAAQ;CACjC,OAAO;AACT;;;;;;;;AASA,SAAgB,sBACd,QACA,MACkB;CAClB,MAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;CAChE,IAAI,CAAC,OAAO,SAAS,GAAG,GACtB,MAAM,IAAI,gBACR,4FACF;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,IAAI;CAAE;CAC/D,yBAAyB,QAAQ;CACjC,OAAO;AACT;;;;;;;;AAcA,SAAgB,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;CAClB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAC7C,MAAM,IAAI,gBACR,uEAAuE,WACzE;CAEF,MAAM,6BAAa,IAAI,IAAoB;CAC3C,KAAK,MAAM,QAAQ,QAAQ;EACzB,IAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAClC,MAAM,IAAI,gBACR,yCAAyC,KAAK,OAAO,8BACvD;EAEF,IAAI,WAAW,IAAI,KAAK,MAAM,GAC5B,MAAM,IAAI,gBAAgB,qDAAqD,KAAK,OAAO,EAAE;EAE/F,WAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;CAC7C;CACA,IAAI,SAAS;CACb,IAAI,SAAS;CACb,KAAK,MAAM,KAAK,QAAQ;EACtB,MAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;EACrC,IAAI,UAAU,KAAA,GAAW;EACzB,IAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAC1B,MAAM,IAAI,gBACR,kDAAkD,EAAE,OAAO,gBAC7D;EAEF;EACA,IAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,WAAW;CAC9C;CACA,IAAI,WAAW,GACb,MAAM,IAAI,gBACR,yGACF;CAEF,MAAM,WAA6B;EAAE,GAAG;EAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO;CAAE;CAC7F,yBAAyB,QAAQ;CACjC,OAAO;AACT;AAUA,SAAgB,sBAAsB,UAA8B,CAAC,GAAkB;CACrF,KAAK,MAAM,KAAK,SAAS,yBAAyB,CAAC;CACnD,MAAM,QAAQ,QAAQ,KAAK,OAAO;EAAE,GAAG;EAAG,SAAS,EAAE,GAAG,EAAE,QAAQ;CAAE,EAAE;CACtE,OAAO;EACL,MAAM,OAAO,UAAU;GACrB,yBAAyB,QAAQ;GACjC,MAAM,KAAK;IAAE,GAAG;IAAU,SAAS,EAAE,GAAG,SAAS,QAAQ;GAAE,CAAC;EAC9D;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,MAAM,MAAM,KAAK,OAAO;IAAE,GAAG;IAAG,SAAS,EAAE,GAAG,EAAE,QAAQ;GAAE,EAAE;GAClE,OAAO,YAAY,KAAA,IAAY,MAAM,IAAI,QAAQ,MAAM,EAAE,YAAY,OAAO;EAC9E;CACF;AACF;;;;;;;AAQA,SAAgB,kBAAkB,MAA6B;CAC7D,OAAO;EACL,MAAM,OAAO,UAAU;GACrB,yBAAyB,QAAQ;GACjC,MAAM,KAAK,MAAM,OAAO;GACxB,MAAM,UAAU,MAAM,OAAO;GAC7B,MAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;GACzD,MAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,EAAE,KAAK,MAAM;EACnE;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,KAAK,MAAM,OAAO;GACxB,IAAI;GACJ,IAAI;IACF,MAAM,MAAM,GAAG,SAAS,MAAM,MAAM;GACtC,SAAS,KAAK;IACZ,IAAK,IAA8B,SAAS,UAAU,OAAO,CAAC;IAC9D,MAAM;GACR;GACA,MAAM,QAAQ,IAAI,MAAM,IAAI;GAC5B,MAAM,YAAgC,CAAC;GACvC,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;IACrC,MAAM,OAAO,MAAM;IACnB,IAAI,KAAK,KAAK,CAAC,CAAC,WAAW,GAAG;IAC9B,IAAI;IACJ,IAAI;KACF,SAAS,KAAK,MAAM,IAAI;IAC1B,SAAS,KAAK;KACZ,MAAM,IAAI,gBACR,4BAA4B,KAAK,4BAA4B,IAAI,EAAE,IAAK,IAAc,SACxF;IACF;IACA,MAAM,WAAW;IACjB,yBAAyB,UAAU,GAAG,KAAK,GAAG,IAAI,GAAG;IACrD,UAAU,KAAK,QAAQ;GACzB;GACA,OAAO,YAAY,KAAA,IAAY,YAAY,UAAU,QAAQ,MAAM,EAAE,YAAY,OAAO;EAC1F;CACF;AACF;AAmDA,MAAM,aAAa;AAEnB,SAAgB,oBACd,SACA,MACgB;CAChB,MAAM,SAAS,SAAS,KAAK,MAAM,MAAM;CACzC,MAAM,SAAS,KAAK,YAAY,UAAU;CAC1C,MAAM,eAAe,KAAK,YAAY,gBAAgB;CACtD,MAAM,kBAAkB,KAAK,YAAY,mBAAmB;CAC5D,MAAM,iBAAiB,KAAK,YAAY,kBAAkB;CAE1D,MAAM,0BAAU,IAAI,IAAwD;CAC5E,KAAK,MAAM,YAAY,SAAS;EAC9B,yBAAyB,QAAQ;EACjC,MAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,QAAQ,GAAG;EAC9E,MAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;EAC9C,IAAI,KAAK;GAAE,GAAG;GAAU;EAAK,CAAC;EAC9B,QAAQ,IAAI,SAAS,SAAS,GAAG;CACnC;CAEA,MAAM,WAA4B,CAAC;CACnC,MAAM,SAAmB,CAAC;CAC1B,MAAM,sBAAgC,CAAC;CAEvC,KAAK,MAAM,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,KAAK,GAAG;EAChD,MAAM,QAAQ,QAAQ,IAAI,OAAO,CAAC,CAAE,MAAM,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;EAClE,MAAM,SAAS,MAAM,MAAM,SAAS;EAEpC,MAAM,WAAW,SAAS,OAAO,QAAQ;EACzC,IAAI,UAAU,gBACZ,OAAO,KACL,UAAU,QAAQ,6BAA6B,OAAO,GAAG,MAAM,QAAQ,QAAQ,CAAC,EAAE,uBAAuB,eAAe,GAC1H;EAMF,IAAI,YAAY;EAChB,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAChC,IAAI,MAAM,EAAE,CAAE,eAAe,MAAM,IAAI,EAAE,CAAE,YAAY,YAAY;EAErE,IAAI,aAAa,GAQX;OAAA,CAPiB,MAClB,MAAM,SAAS,CAAC,CAChB,MACE,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB,CAEhC,GACd,OAAO,KACL,UAAU,QAAQ,oBAAoB,MAAM,YAAY,EAAE,CAAE,WAAW,OAAO,MAAM,UAAU,CAAE,WAAW,OAAO,MAAM,UAAU,CAAE,GAAG,gIACzI;EAAA;EAIJ,KAAK,MAAM,UAAU,uBAAuB;GAC1C,MAAM,SAAS,MACZ,QAAQ,MAAM,OAAO,EAAE,QAAQ,YAAY,QAAQ,CAAC,CACpD,KAAK,MAAM,EAAE,QAAQ,OAAQ;GAChC,IAAI,OAAO,WAAW,GAAG;GAEzB,MAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;GAC1D,IAAI,YAAY,UAAU,qBACxB,oBAAoB,KAAK,GAAG,QAAQ,GAAG,QAAQ;GAEjD,MAAM,UAAU,OAAO,OAAO,SAAS;GACvC,MAAM,WAAW,OAAO;GACxB,MAAM,QAAQ,UAAU;GACxB,MAAM,UAAoB,CAAC;GAC3B,IAAI,YAAY,UAAU,iBACxB,QAAQ,KACN,6CAA6C,YAAY,QAAQ,gBAAgB,YAAY,WAAW,QAAQ,CAAC,EAAE,EACrH;GAEF,IAAI,WAAW,SAAS,UAAU,QAChC,QAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,EAAE,eAAe,QAAQ;GAEhE,IAAI,WAAW,sBAAsB,WAAW,UAAU,cACxD,QAAQ,KACN,6BAA6B,WAAW,QAAA,CAAS,QAAQ,CAAC,EAAE,iBAAiB,SAAS,QAAQ,CAAC,EAAE,UAAU,aAAa,EAC1H;GAEF,IAAI,WAAW,sBAAsB,WAAW,UAAU,iBACxD,QAAQ,KACN,6BAA6B,WAAW,QAAA,CAAS,QAAQ,CAAC,EAAE,iBAAiB,SAAS,QAAQ,CAAC,EAAE,UAAU,gBAAgB,EAC7H;GAGF,MAAM,UAAU,QAAQ,SAAS;GACjC,MAAM,QAAuB;IAC3B;IACA;IACA,OAAO,YAAY;IACnB;IACA;IACA;IACA;GACF;GACA,IAAI,SAAS;IACX,MAAM,SAAS,QAAQ,KAAK,IAAI;IAChC,OAAO,KAAK,UAAU,QAAQ,IAAI,OAAO,IAAI,MAAM,QAAQ;GAC7D;GACA,SAAS,KAAK,KAAK;EACrB;CACF;CAEA,OAAO;EAAE;EAAU;EAAQ,SAAS,OAAO,WAAW;EAAG;CAAoB;AAC/E;;;;;;;;;AAiBA,SAAgB,gBAAgB,QAAyC;CACvE,OAAO;EAAE,SAAS,OAAO;EAAS,QAAQ,CAAC,GAAG,OAAO,MAAM;CAAE;AAC/D"}

package/dist/{mint-D5_87M5L.js → mint-BvkwcYZU.js}

@@ -1,6 +1,6 @@
 import { s as ValidationError } from "./errors-8YnH8WlF.js";
 import { a as scoreOrigin, i as rolloutRewardFields } from "./reward-nw2xZGZG.js";
-import { o as runTaskScore } from "./run-record-CN8Zd21B.js";
+import { o as runTaskScore } from "./run-record-BIwU2wdV.js";
 import { i as ROLLOUT_SCHEMA, s as assertMinted } from "./schema-C6DW4ZHR.js";
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
 //#region src/rollout/mint.ts
@@ -198,4 +198,4 @@ async function mintRolloutRows(records, store, options = {}) {
 //#endregion
 export { mintRolloutRows as t };
 
-//# sourceMappingURL=mint-D5_87M5L.js.map
+//# sourceMappingURL=mint-BvkwcYZU.js.map

package/dist/{mint-D5_87M5L.js.map → mint-BvkwcYZU.js.map}

@@ -1 +1 @@
-{"version":3,"file":"mint-D5_87M5L.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, runTaskScore } from '../run-record'\nimport type { LlmSpan, Message, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor applied to every exported string (secrets, PII). Identity by default. */\nexport type RolloutScrubber = (text: string) => string\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** The final llm span's history + output is the completed conversation. */\nfunction finalConversation(spans: Span[], scrub: RolloutScrubber): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = last.messages.map((m: Message) => ({\n    role: m.role,\n    content: scrub(m.content),\n  }))\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: scrub(last.output) })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  const uncaptured = record.costProvenance.kind === 'uncaptured'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: null,\n      candidate_index: null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: uncaptured ? null : record.costUsd,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: { patch_path: null, run_dir: null, transcript_ref: null },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    if (trajectory.steps.length === 0) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const conversation = finalConversation(\n      trajectory.steps.map((s) => s.span),\n      scrub,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EACzC,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,kBAAkB,OAAe,OAAuC;CAC/E,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,KAAK,SAAS,KAAK,OAAgB;EACjE,MAAM,EAAE;EACR,SAAS,MAAM,EAAE,OAAO;CAC1B,EAAE;CACF,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,MAAM,KAAK,MAAM;CAAE,CAAC;CAElE,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACmB;CAInB,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAC/C,MAAM,aAAa,OAAO,eAAe,SAAS;CAClD,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB;EACnB,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY;EACZ,iBAAiB;EACjB,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK;EACP;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,OAAO;GAChC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GAAE,YAAY;GAAM,SAAS;GAAM,gBAAgB;EAAK;EACnE,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,IAAI,WAAW,MAAM,WAAW,GAAG;GACjC,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC,CACxF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,eAAe,kBACnB,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI,GAClC,KACF;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CAC3E;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}
+{"version":3,"file":"mint-BvkwcYZU.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, runTaskScore } from '../run-record'\nimport type { LlmSpan, Message, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor applied to every exported string (secrets, PII). Identity by default. */\nexport type RolloutScrubber = (text: string) => string\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** The final llm span's history + output is the completed conversation. */\nfunction finalConversation(spans: Span[], scrub: RolloutScrubber): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = last.messages.map((m: Message) => ({\n    role: m.role,\n    content: scrub(m.content),\n  }))\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: scrub(last.output) })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  const uncaptured = record.costProvenance.kind === 'uncaptured'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: null,\n      candidate_index: null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: uncaptured ? null : record.costUsd,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: { patch_path: null, run_dir: null, transcript_ref: null },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    if (trajectory.steps.length === 0) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const conversation = finalConversation(\n      trajectory.steps.map((s) => s.span),\n      scrub,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EACzC,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,kBAAkB,OAAe,OAAuC;CAC/E,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,KAAK,SAAS,KAAK,OAAgB;EACjE,MAAM,EAAE;EACR,SAAS,MAAM,EAAE,OAAO;CAC1B,EAAE;CACF,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,MAAM,KAAK,MAAM;CAAE,CAAC;CAElE,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACmB;CAInB,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAC/C,MAAM,aAAa,OAAO,eAAe,SAAS;CAClD,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB;EACnB,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY;EACZ,iBAAiB;EACjB,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK;EACP;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,OAAO;GAChC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GAAE,YAAY;GAAM,SAAS;GAAM,gBAAgB;EAAK;EACnE,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,IAAI,WAAW,MAAM,WAAW,GAAG;GACjC,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC,CACxF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,eAAe,kBACnB,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI,GAClC,KACF;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CAC3E;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.132.0",
+    "version": "0.133.1",
     "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",