@tangle-network/agent-eval 0.95.1 → 0.96.1

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

@@ -3,7 +3,7 @@ export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOut
 export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from '../rubric-predictive-validity-C2hDKM8Z.js';
 import { C as ContinuousAgreement, a as CalibrationResult, b as ContinuousCalibrationResult, c as CandidateScore, G as GoldenItem } from '../judge-calibration-0p2QcWNE.js';
 import { S as SeriesConvergenceOptions, a as SeriesConvergenceResult } from '../series-convergence-D5OWMBg6.js';
-import { C as CorpusAgreementReport } from '../statistics-CCJpTGOS.js';
+import { C as CorpusAgreementReport } from '../statistics-xP-cWc5k.js';
 import '../store-BcFXE6LG.js';
 import '../schema-m0gsnbt3.js';
 import '../run-record-CP2ObebC.js';

package/dist/meta-eval/index.js

@@ -11,7 +11,11 @@ import {
 } from "../chunk-3RF76KTD.js";
 import {
   rubricPredictiveValidity
-} from "../chunk-YRZ4M5GS.js";
+} from "../chunk-NF7OZ4J7.js";
+import {
+  pearsonR,
+  spearmanR
+} from "../chunk-HRGUJTER.js";
 import {
   aggregateLlm,
   llmSpans
@@ -71,7 +75,7 @@ async function correlationStudy(traceStore, outcomeStore, evalMetrics, outcomeMe
   }
   const results = pairs.filter((p) => p.xs.length >= 3).map((p) => {
     const pearson = pearsonR(p.xs, p.ys);
-    const spearman = pearsonR(ranks(p.xs), ranks(p.ys));
+    const spearman = spearmanR(p.xs, p.ys);
     const pearsonCi95 = bootstrapPearsonCi(p.xs, p.ys, options.bootstrapIterations ?? 500);
     const verdict = Math.abs(pearson) >= 0.7 ? "strong" : Math.abs(pearson) >= 0.4 ? "moderate" : "weak";
     return {
@@ -104,32 +108,6 @@ function reduce(values, kind, outcomes) {
   if (paired.length === 0) return v ?? null;
   return paired.sort((a, b) => b.at - a.at)[0]?.v ?? null;
 }
-function pearsonR(a, b) {
-  if (a.length !== b.length || a.length < 2) return NaN;
-  const mA = a.reduce((s, v) => s + v, 0) / a.length;
-  const mB = b.reduce((s, v) => s + v, 0) / b.length;
-  let num = 0, dA = 0, dB = 0;
-  for (let i = 0; i < a.length; i++) {
-    const da = a[i] - mA, db = b[i] - mB;
-    num += da * db;
-    dA += da * da;
-    dB += db * db;
-  }
-  if (dA === 0 || dB === 0) return dA === 0 && dB === 0 ? 1 : 0;
-  return num / Math.sqrt(dA * dB);
-}
-function ranks(xs) {
-  const indexed = xs.map((v, i) => ({ v, i })).sort((x, y) => x.v - y.v);
-  const r = new Array(xs.length);
-  for (let i = 0; i < indexed.length; i++) {
-    let j = i;
-    while (j + 1 < indexed.length && indexed[j + 1].v === indexed[i].v) j++;
-    const avg = (i + j + 2) / 2;
-    for (let k = i; k <= j; k++) r[indexed[k].i] = avg;
-    i = j;
-  }
-  return r;
-}
 function bootstrapPearsonCi(xs, ys, iterations) {
   const n = xs.length;
   if (n < 3) return { lower: NaN, upper: NaN };

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

@@ -1 +1 @@
-{"version":3,"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 { 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 = pearsonR(ranks(p.xs), ranks(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 pearsonR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return NaN\n  const mA = a.reduce((s, v) => s + v, 0) / a.length\n  const mB = b.reduce((s, v) => s + v, 0) / b.length\n  let num = 0,\n    dA = 0,\n    dB = 0\n  for (let i = 0; i < a.length; i++) {\n    const da = a[i]! - mA,\n      db = b[i]! - mB\n    num += da * db\n    dA += da * da\n    dB += db * db\n  }\n  if (dA === 0 || dB === 0) return dA === 0 && dB === 0 ? 1 : 0\n  return num / Math.sqrt(dA * dB)\n}\n\nfunction ranks(xs: number[]): number[] {\n  const indexed = xs.map((v, i) => ({ v, i })).sort((x, y) => x.v - y.v)\n  const r = new Array<number>(xs.length)\n  for (let i = 0; i < indexed.length; i++) {\n    let j = i\n    while (j + 1 < indexed.length && indexed[j + 1]!.v === indexed[i]!.v) j++\n    const avg = (i + j + 2) / 2\n    for (let k = i; k <= j; k++) r[indexed[k]!.i] = avg\n    i = j\n  }\n  return r\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":";;;;;;;;;;;;;;;;;;;;;;;;;AAwDA,eAAsB,iBACpB,YACA,cACA,aACA,oBACA,UAAmC,CAAC,GACH;AACjC,QAAM,OAAO,MAAM,WAAW,SAAS;AACvC,QAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;AAC9D,QAAM,gBAAgB,oBAAI,IAAiC;AAC3D,aAAW,KAAK,UAAU;AACxB,UAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;AAC3C,QAAI,KAAK,CAAC;AACV,kBAAc,IAAI,EAAE,OAAO,GAAG;AAAA,EAChC;AAEA,QAAM,YAAY,QAAQ,aAAa;AACvC,QAAM,SAAS,QAAQ,mBAAmB;AAE1C,QAAM,QAA0F,CAAC;AACjG,aAAW,MAAM,aAAa;AAC5B,eAAW,MAAM,oBAAoB;AACnC,YAAM,KAAK,EAAE,YAAY,GAAG,IAAI,eAAe,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AAEA,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,OAAO,MAAM;AACtB,UAAM,KAAK,cAAc,IAAI,IAAI,KAAK;AACtC,QAAI,CAAC,MAAM,GAAG,WAAW,GAAG;AAC1B;AACA;AAAA,IACF;AACA,UAAM,WAAW,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;AACxE,QAAI,SAAS,WAAW,GAAG;AACzB;AACA;AAAA,IACF;AAEA,eAAW,MAAM,aAAa;AAC5B,YAAM,UAAU,GAAG,WAAW,eAAe,GAAG,EAAE;AAClD,YAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;AACvC,UAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,EAAG;AAEvC,iBAAW,MAAM,oBAAoB;AACnC,cAAM,SAAS,SACZ,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,EACxB,OAAO,CAAC,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;AACzE,YAAI,OAAO,WAAW,EAAG;AACzB,cAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;AAC5C,YAAI,MAAM,KAAM;AAChB,cAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;AAC/E,aAAK,GAAG,KAAK,CAAC;AACd,aAAK,GAAG,KAAK,CAAC;AAAA,MAChB;AAAA,IACF;AACA;AAAA,EACF;AAEA,QAAM,UAA+B,MAClC,OAAO,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC,EAC9B,IAAI,CAAC,MAAM;AACV,UAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;AACnC,UAAM,WAAW,SAAS,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,EAAE,CAAC;AAClD,UAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;AACrF,UAAM,UACJ,KAAK,IAAI,OAAO,KAAK,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,MAAM,aAAa;AAChF,WAAO;AAAA,MACL,YAAY,EAAE;AAAA,MACd,eAAe,EAAE;AAAA,MACjB,GAAG,EAAE,GAAG;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAEH,SAAO,EAAE,OAAO,SAAS,eAAe,QAAQ,aAAa,QAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;AACf,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,SAAS,OAAQ,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACvE,MAAI,SAAS,MAAO,QAAO,KAAK,IAAI,GAAG,MAAM;AAE7C,QAAM,SAAS,CAAC,GAAG,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,EAAE,CAAC;AAC1E,MAAI,CAAC,OAAQ,QAAO;AACpB,QAAM,YAAY,OAAO,KAAK,OAAO,OAAO,EAAE,CAAC;AAC/C,QAAM,IAAI,cAAc,SAAY,OAAO,QAAQ,SAAS,IAAI;AAEhE,QAAM,SAAS,SACZ,IAAI,CAAC,MAAM;AACV,UAAM,IAAI,OAAO,KAAK,EAAE,OAAO,EAAE,CAAC;AAClC,WAAO;AAAA,MACL,IAAI,EAAE;AAAA,MACN,GAAG,MAAM,SAAY,OAAO,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,IAAI;AAAA,IAChE;AAAA,EACF,CAAC,EACA,OAAO,CAAC,MAAM,EAAE,MAAM,MAAS;AAClC,MAAI,OAAO,WAAW,EAAG,QAAO,KAAK;AACrC,SAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,GAAG,KAAK;AACrD;AAEA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,EAAG,QAAO;AAClD,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,MAAI,MAAM,GACR,KAAK,GACL,KAAK;AACP,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC,IAAK,IACjB,KAAK,EAAE,CAAC,IAAK;AACf,WAAO,KAAK;AACZ,UAAM,KAAK;AACX,UAAM,KAAK;AAAA,EACb;AACA,MAAI,OAAO,KAAK,OAAO,EAAG,QAAO,OAAO,KAAK,OAAO,IAAI,IAAI;AAC5D,SAAO,MAAM,KAAK,KAAK,KAAK,EAAE;AAChC;AAEA,SAAS,MAAM,IAAwB;AACrC,QAAM,UAAU,GAAG,IAAI,CAAC,GAAG,OAAO,EAAE,GAAG,EAAE,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AACrE,QAAM,IAAI,IAAI,MAAc,GAAG,MAAM;AACrC,WAAS,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,QAAI,IAAI;AACR,WAAO,IAAI,IAAI,QAAQ,UAAU,QAAQ,IAAI,CAAC,EAAG,MAAM,QAAQ,CAAC,EAAG,EAAG;AACtE,UAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,aAAS,IAAI,GAAG,KAAK,GAAG,IAAK,GAAE,QAAQ,CAAC,EAAG,CAAC,IAAI;AAChD,QAAI;AAAA,EACN;AACA,SAAO;AACT;AAEA,SAAS,mBACP,IACA,IACA,YACkC;AAClC,QAAM,IAAI,GAAG;AACb,MAAI,IAAI,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AAC3C,QAAM,KAAe,CAAC;AACtB,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;AACxC,SAAG,CAAC,IAAI,GAAG,GAAG;AACd,SAAG,CAAC,IAAI,GAAG,GAAG;AAAA,IAChB;AACA,UAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAI,OAAO,SAAS,CAAC,EAAG,IAAG,KAAK,CAAC;AAAA,EACnC;AACA,KAAG,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,MAAI,GAAG,WAAW,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AACrD,SAAO;AAAA,IACL,OAAO,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC;AAAA,IACvC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC;AAAA,EAClE;AACF;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;AClMO,IAAM,wBAAwB,CAAC,OAAO,oBAAoB,kBAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;AACtD,QAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;AAC3D,MAAI,CAAC,OAAO,SAAS,EAAE,GAAG;AACxB,UAAM,IAAI;AAAA,MACR,mBAAmB,KAAK,sCAAsC,KAAK,UAAU,KAAK,CAAC;AAAA,IACrF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,yBAAyB,UAA4B,SAAS,YAAkB;AAC9F,WAAS,SAAS,IAAI,GAAG,MAAM,KAAK;AACpC,MAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GAAG;AACzE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,qCAAqC;AAAA,EAC1F;AACA,MAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAAG;AAC/E,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,wCAAwC;AAAA,EAC7F;AACA,MAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAAU;AACrE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,2CAA2C;AAAA,EAChG;AACA,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,SAAS,OAAO,GAAG;AAC3D,QAAI,CAAE,sBAA4C,SAAS,GAAG,GAAG;AAE/D,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,iCAAiC,GAAG,mBAAc,sBAAsB,KAAK,IAAI,CAAC;AAAA,MAC7G;AAAA,IACF;AACA,QAAI,UAAU,UAAa,CAAC,OAAO,SAAS,KAAK,GAAG;AAClD,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,YAAY,GAAG,4BAA4B,OAAO,KAAK,CAAC;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACF;AASO,SAAS,wBACd,QACA,MACkB;AAClB,QAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;AAC5F,MAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,UAAM,IAAI;AAAA,MACR,oDAA+C,OAAO,CAAC;AAAA,IACzD;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,MAAM,EAAE;AACnF,2BAAyB,QAAQ;AACjC,SAAO;AACT;AASO,SAAS,sBACd,QACA,MACkB;AAClB,QAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;AAChE,MAAI,CAAC,OAAO,SAAS,GAAG,GAAG;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,IAAI,EAAE;AAC/D,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAcO,SAAS,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;AAClB,QAAM,YAAY,QAAQ,aAAa;AACvC,MAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAAG;AAChD,UAAM,IAAI;AAAA,MACR,4EAAuE,SAAS;AAAA,IAClF;AAAA,EACF;AACA,QAAM,aAAa,oBAAI,IAAoB;AAC3C,aAAW,QAAQ,QAAQ;AACzB,QAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAAG;AACrC,YAAM,IAAI;AAAA,QACR,yCAAyC,KAAK,MAAM;AAAA,MACtD;AAAA,IACF;AACA,QAAI,WAAW,IAAI,KAAK,MAAM,GAAG;AAC/B,YAAM,IAAI,gBAAgB,qDAAqD,KAAK,MAAM,GAAG;AAAA,IAC/F;AACA,eAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;AAAA,EAC7C;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,aAAW,KAAK,QAAQ;AACtB,UAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;AACrC,QAAI,UAAU,OAAW;AACzB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAAG;AAC7B,YAAM,IAAI;AAAA,QACR,kDAAkD,EAAE,MAAM;AAAA,MAC5D;AAAA,IACF;AACA;AACA,QAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,UAAW;AAAA,EAC9C;AACA,MAAI,WAAW,GAAG;AAChB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO,EAAE;AAC7F,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAUO,SAAS,sBAAsB,UAA8B,CAAC,GAAkB;AACrF,aAAW,KAAK,QAAS,0BAAyB,CAAC;AACnD,QAAM,QAAQ,QAAQ,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AACtE,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,EAAE,GAAG,UAAU,SAAS,EAAE,GAAG,SAAS,QAAQ,EAAE,CAAC;AAAA,IAC9D;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,MAAM,MAAM,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AAClE,aAAO,YAAY,SAAY,MAAM,IAAI,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC9E;AAAA,EACF;AACF;AAQO,SAAS,kBAAkB,MAA6B;AAC7D,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,YAAM,UAAU,MAAM,OAAO,MAAW;AACxC,YAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACzD,YAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,CAAC;AAAA,GAAM,MAAM;AAAA,IACnE;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,UAAI;AACJ,UAAI;AACF,cAAM,MAAM,GAAG,SAAS,MAAM,MAAM;AAAA,MACtC,SAAS,KAAK;AACZ,YAAK,IAA8B,SAAS,SAAU,QAAO,CAAC;AAC9D,cAAM;AAAA,MACR;AACA,YAAM,QAAQ,IAAI,MAAM,IAAI;AAC5B,YAAM,YAAgC,CAAC;AACvC,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,cAAM,OAAO,MAAM,CAAC;AACpB,YAAI,KAAK,KAAK,EAAE,WAAW,EAAG;AAC9B,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,SAAS,KAAK;AACZ,gBAAM,IAAI;AAAA,YACR,4BAA4B,IAAI,6BAA6B,IAAI,CAAC,KAAM,IAAc,OAAO;AAAA,UAC/F;AAAA,QACF;AACA,cAAM,WAAW;AACjB,iCAAyB,UAAU,GAAG,IAAI,IAAI,IAAI,CAAC,EAAE;AACrD,kBAAU,KAAK,QAAQ;AAAA,MACzB;AACA,aAAO,YAAY,SAAY,YAAY,UAAU,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC1F;AAAA,EACF;AACF;AAmDA,IAAM,aAAa;AAEZ,SAAS,oBACd,SACA,MACgB;AAChB,QAAM,SAAS,SAAS,KAAK,MAAM,MAAM;AACzC,QAAM,SAAS,KAAK,YAAY,UAAU;AAC1C,QAAM,eAAe,KAAK,YAAY,gBAAgB;AACtD,QAAM,kBAAkB,KAAK,YAAY,mBAAmB;AAC5D,QAAM,iBAAiB,KAAK,YAAY,kBAAkB;AAE1D,QAAM,UAAU,oBAAI,IAAwD;AAC5E,aAAW,YAAY,SAAS;AAC9B,6BAAyB,QAAQ;AACjC,UAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,OAAO,IAAI;AAC9E,UAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;AAC9C,QAAI,KAAK,EAAE,GAAG,UAAU,KAAK,CAAC;AAC9B,YAAQ,IAAI,SAAS,SAAS,GAAG;AAAA,EACnC;AAEA,QAAM,WAA4B,CAAC;AACnC,QAAM,SAAmB,CAAC;AAC1B,QAAM,sBAAgC,CAAC;AAEvC,aAAW,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,EAAE,KAAK,GAAG;AAChD,UAAM,QAAQ,QAAQ,IAAI,OAAO,EAAG,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClE,UAAM,SAAS,MAAM,MAAM,SAAS,CAAC;AAErC,UAAM,WAAW,SAAS,OAAO,QAAQ;AACzC,QAAI,UAAU,gBAAgB;AAC5B,aAAO;AAAA,QACL,UAAU,OAAO,mCAA8B,OAAO,EAAE,OAAO,QAAQ,QAAQ,CAAC,CAAC,wBAAwB,cAAc;AAAA,MACzH;AAAA,IACF;AAKA,QAAI,YAAY;AAChB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAI,MAAM,CAAC,EAAG,eAAe,MAAM,IAAI,CAAC,EAAG,WAAY,aAAY;AAAA,IACrE;AACA,QAAI,aAAa,GAAG;AAClB,YAAM,eAAe,MAClB,MAAM,SAAS,EACf;AAAA,QACC,CAAC,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB;AAAA,MAC9C;AACF,UAAI,CAAC,cAAc;AACjB,eAAO;AAAA,UACL,UAAU,OAAO,qBAAqB,MAAM,YAAY,CAAC,EAAG,UAAU,aAAQ,MAAM,SAAS,EAAG,UAAU,QAAQ,MAAM,SAAS,EAAG,EAAE;AAAA,QACxI;AAAA,MACF;AAAA,IACF;AAEA,eAAW,UAAU,uBAAuB;AAC1C,YAAM,SAAS,MACZ,OAAO,CAAC,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,QAAQ,EACnD,IAAI,CAAC,MAAM,EAAE,QAAQ,MAAM,CAAE;AAChC,UAAI,OAAO,WAAW,EAAG;AAEzB,YAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;AAC1D,UAAI,YAAY,UAAU;AACxB,4BAAoB,KAAK,GAAG,OAAO,IAAI,MAAM,EAAE;AAEjD,YAAM,UAAU,OAAO,OAAO,SAAS,CAAC;AACxC,YAAM,WAAW,OAAO,CAAC;AACzB,YAAM,QAAQ,UAAU;AACxB,YAAM,UAAoB,CAAC;AAC3B,UAAI,YAAY,UAAU,iBAAiB;AACzC,gBAAQ;AAAA,UACN,6CAA6C,YAAY,OAAO,iBAAiB,YAAY,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpH;AAAA,MACF;AACA,UAAI,WAAW,SAAS,UAAU,QAAQ;AACxC,gBAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,CAAC,gBAAgB,MAAM,EAAE;AAAA,MAChE;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,cAAc;AACtE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,YAAY;AAAA,QACzH;AAAA,MACF;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,iBAAiB;AACzE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,eAAe;AAAA,QAC5H;AAAA,MACF;AAEA,YAAM,UAAU,QAAQ,SAAS;AACjC,YAAM,QAAuB;AAAA,QAC3B;AAAA,QACA;AAAA,QACA,OAAO,YAAY;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,SAAS;AACX,cAAM,SAAS,QAAQ,KAAK,IAAI;AAChC,eAAO,KAAK,UAAU,OAAO,KAAK,MAAM,KAAK,MAAM,MAAM,EAAE;AAAA,MAC7D;AACA,eAAS,KAAK,KAAK;AAAA,IACrB;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,QAAQ,SAAS,OAAO,WAAW,GAAG,oBAAoB;AAC/E;AAiBO,SAAS,gBAAgB,QAAyC;AACvE,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,CAAC,GAAG,OAAO,MAAM,EAAE;AAC/D;","names":[]}
+{"version":3,"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;AACjC,QAAM,OAAO,MAAM,WAAW,SAAS;AACvC,QAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;AAC9D,QAAM,gBAAgB,oBAAI,IAAiC;AAC3D,aAAW,KAAK,UAAU;AACxB,UAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;AAC3C,QAAI,KAAK,CAAC;AACV,kBAAc,IAAI,EAAE,OAAO,GAAG;AAAA,EAChC;AAEA,QAAM,YAAY,QAAQ,aAAa;AACvC,QAAM,SAAS,QAAQ,mBAAmB;AAE1C,QAAM,QAA0F,CAAC;AACjG,aAAW,MAAM,aAAa;AAC5B,eAAW,MAAM,oBAAoB;AACnC,YAAM,KAAK,EAAE,YAAY,GAAG,IAAI,eAAe,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AAEA,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,OAAO,MAAM;AACtB,UAAM,KAAK,cAAc,IAAI,IAAI,KAAK;AACtC,QAAI,CAAC,MAAM,GAAG,WAAW,GAAG;AAC1B;AACA;AAAA,IACF;AACA,UAAM,WAAW,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;AACxE,QAAI,SAAS,WAAW,GAAG;AACzB;AACA;AAAA,IACF;AAEA,eAAW,MAAM,aAAa;AAC5B,YAAM,UAAU,GAAG,WAAW,eAAe,GAAG,EAAE;AAClD,YAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;AACvC,UAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,EAAG;AAEvC,iBAAW,MAAM,oBAAoB;AACnC,cAAM,SAAS,SACZ,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,EACxB,OAAO,CAAC,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;AACzE,YAAI,OAAO,WAAW,EAAG;AACzB,cAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;AAC5C,YAAI,MAAM,KAAM;AAChB,cAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;AAC/E,aAAK,GAAG,KAAK,CAAC;AACd,aAAK,GAAG,KAAK,CAAC;AAAA,MAChB;AAAA,IACF;AACA;AAAA,EACF;AAEA,QAAM,UAA+B,MAClC,OAAO,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC,EAC9B,IAAI,CAAC,MAAM;AACV,UAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;AACnC,UAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;AACrC,UAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;AACrF,UAAM,UACJ,KAAK,IAAI,OAAO,KAAK,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,MAAM,aAAa;AAChF,WAAO;AAAA,MACL,YAAY,EAAE;AAAA,MACd,eAAe,EAAE;AAAA,MACjB,GAAG,EAAE,GAAG;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAEH,SAAO,EAAE,OAAO,SAAS,eAAe,QAAQ,aAAa,QAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;AACf,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,SAAS,OAAQ,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACvE,MAAI,SAAS,MAAO,QAAO,KAAK,IAAI,GAAG,MAAM;AAE7C,QAAM,SAAS,CAAC,GAAG,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,EAAE,CAAC;AAC1E,MAAI,CAAC,OAAQ,QAAO;AACpB,QAAM,YAAY,OAAO,KAAK,OAAO,OAAO,EAAE,CAAC;AAC/C,QAAM,IAAI,cAAc,SAAY,OAAO,QAAQ,SAAS,IAAI;AAEhE,QAAM,SAAS,SACZ,IAAI,CAAC,MAAM;AACV,UAAM,IAAI,OAAO,KAAK,EAAE,OAAO,EAAE,CAAC;AAClC,WAAO;AAAA,MACL,IAAI,EAAE;AAAA,MACN,GAAG,MAAM,SAAY,OAAO,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,IAAI;AAAA,IAChE;AAAA,EACF,CAAC,EACA,OAAO,CAAC,MAAM,EAAE,MAAM,MAAS;AAClC,MAAI,OAAO,WAAW,EAAG,QAAO,KAAK;AACrC,SAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,GAAG,KAAK;AACrD;AAEA,SAAS,mBACP,IACA,IACA,YACkC;AAClC,QAAM,IAAI,GAAG;AACb,MAAI,IAAI,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AAC3C,QAAM,KAAe,CAAC;AACtB,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;AACxC,SAAG,CAAC,IAAI,GAAG,GAAG;AACd,SAAG,CAAC,IAAI,GAAG,GAAG;AAAA,IAChB;AACA,UAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAI,OAAO,SAAS,CAAC,EAAG,IAAG,KAAK,CAAC;AAAA,EACnC;AACA,KAAG,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,MAAI,GAAG,WAAW,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AACrD,SAAO;AAAA,IACL,OAAO,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC;AAAA,IACvC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC;AAAA,EAClE;AACF;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;ACpKO,IAAM,wBAAwB,CAAC,OAAO,oBAAoB,kBAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;AACtD,QAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;AAC3D,MAAI,CAAC,OAAO,SAAS,EAAE,GAAG;AACxB,UAAM,IAAI;AAAA,MACR,mBAAmB,KAAK,sCAAsC,KAAK,UAAU,KAAK,CAAC;AAAA,IACrF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,yBAAyB,UAA4B,SAAS,YAAkB;AAC9F,WAAS,SAAS,IAAI,GAAG,MAAM,KAAK;AACpC,MAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GAAG;AACzE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,qCAAqC;AAAA,EAC1F;AACA,MAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAAG;AAC/E,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,wCAAwC;AAAA,EAC7F;AACA,MAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAAU;AACrE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,2CAA2C;AAAA,EAChG;AACA,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,SAAS,OAAO,GAAG;AAC3D,QAAI,CAAE,sBAA4C,SAAS,GAAG,GAAG;AAE/D,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,iCAAiC,GAAG,mBAAc,sBAAsB,KAAK,IAAI,CAAC;AAAA,MAC7G;AAAA,IACF;AACA,QAAI,UAAU,UAAa,CAAC,OAAO,SAAS,KAAK,GAAG;AAClD,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,YAAY,GAAG,4BAA4B,OAAO,KAAK,CAAC;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACF;AASO,SAAS,wBACd,QACA,MACkB;AAClB,QAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;AAC5F,MAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,UAAM,IAAI;AAAA,MACR,oDAA+C,OAAO,CAAC;AAAA,IACzD;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,MAAM,EAAE;AACnF,2BAAyB,QAAQ;AACjC,SAAO;AACT;AASO,SAAS,sBACd,QACA,MACkB;AAClB,QAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;AAChE,MAAI,CAAC,OAAO,SAAS,GAAG,GAAG;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,IAAI,EAAE;AAC/D,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAcO,SAAS,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;AAClB,QAAM,YAAY,QAAQ,aAAa;AACvC,MAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAAG;AAChD,UAAM,IAAI;AAAA,MACR,4EAAuE,SAAS;AAAA,IAClF;AAAA,EACF;AACA,QAAM,aAAa,oBAAI,IAAoB;AAC3C,aAAW,QAAQ,QAAQ;AACzB,QAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAAG;AACrC,YAAM,IAAI;AAAA,QACR,yCAAyC,KAAK,MAAM;AAAA,MACtD;AAAA,IACF;AACA,QAAI,WAAW,IAAI,KAAK,MAAM,GAAG;AAC/B,YAAM,IAAI,gBAAgB,qDAAqD,KAAK,MAAM,GAAG;AAAA,IAC/F;AACA,eAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;AAAA,EAC7C;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,aAAW,KAAK,QAAQ;AACtB,UAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;AACrC,QAAI,UAAU,OAAW;AACzB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAAG;AAC7B,YAAM,IAAI;AAAA,QACR,kDAAkD,EAAE,MAAM;AAAA,MAC5D;AAAA,IACF;AACA;AACA,QAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,UAAW;AAAA,EAC9C;AACA,MAAI,WAAW,GAAG;AAChB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO,EAAE;AAC7F,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAUO,SAAS,sBAAsB,UAA8B,CAAC,GAAkB;AACrF,aAAW,KAAK,QAAS,0BAAyB,CAAC;AACnD,QAAM,QAAQ,QAAQ,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AACtE,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,EAAE,GAAG,UAAU,SAAS,EAAE,GAAG,SAAS,QAAQ,EAAE,CAAC;AAAA,IAC9D;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,MAAM,MAAM,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AAClE,aAAO,YAAY,SAAY,MAAM,IAAI,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC9E;AAAA,EACF;AACF;AAQO,SAAS,kBAAkB,MAA6B;AAC7D,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,YAAM,UAAU,MAAM,OAAO,MAAW;AACxC,YAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACzD,YAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,CAAC;AAAA,GAAM,MAAM;AAAA,IACnE;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,UAAI;AACJ,UAAI;AACF,cAAM,MAAM,GAAG,SAAS,MAAM,MAAM;AAAA,MACtC,SAAS,KAAK;AACZ,YAAK,IAA8B,SAAS,SAAU,QAAO,CAAC;AAC9D,cAAM;AAAA,MACR;AACA,YAAM,QAAQ,IAAI,MAAM,IAAI;AAC5B,YAAM,YAAgC,CAAC;AACvC,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,cAAM,OAAO,MAAM,CAAC;AACpB,YAAI,KAAK,KAAK,EAAE,WAAW,EAAG;AAC9B,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,SAAS,KAAK;AACZ,gBAAM,IAAI;AAAA,YACR,4BAA4B,IAAI,6BAA6B,IAAI,CAAC,KAAM,IAAc,OAAO;AAAA,UAC/F;AAAA,QACF;AACA,cAAM,WAAW;AACjB,iCAAyB,UAAU,GAAG,IAAI,IAAI,IAAI,CAAC,EAAE;AACrD,kBAAU,KAAK,QAAQ;AAAA,MACzB;AACA,aAAO,YAAY,SAAY,YAAY,UAAU,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC1F;AAAA,EACF;AACF;AAmDA,IAAM,aAAa;AAEZ,SAAS,oBACd,SACA,MACgB;AAChB,QAAM,SAAS,SAAS,KAAK,MAAM,MAAM;AACzC,QAAM,SAAS,KAAK,YAAY,UAAU;AAC1C,QAAM,eAAe,KAAK,YAAY,gBAAgB;AACtD,QAAM,kBAAkB,KAAK,YAAY,mBAAmB;AAC5D,QAAM,iBAAiB,KAAK,YAAY,kBAAkB;AAE1D,QAAM,UAAU,oBAAI,IAAwD;AAC5E,aAAW,YAAY,SAAS;AAC9B,6BAAyB,QAAQ;AACjC,UAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,OAAO,IAAI;AAC9E,UAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;AAC9C,QAAI,KAAK,EAAE,GAAG,UAAU,KAAK,CAAC;AAC9B,YAAQ,IAAI,SAAS,SAAS,GAAG;AAAA,EACnC;AAEA,QAAM,WAA4B,CAAC;AACnC,QAAM,SAAmB,CAAC;AAC1B,QAAM,sBAAgC,CAAC;AAEvC,aAAW,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,EAAE,KAAK,GAAG;AAChD,UAAM,QAAQ,QAAQ,IAAI,OAAO,EAAG,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClE,UAAM,SAAS,MAAM,MAAM,SAAS,CAAC;AAErC,UAAM,WAAW,SAAS,OAAO,QAAQ;AACzC,QAAI,UAAU,gBAAgB;AAC5B,aAAO;AAAA,QACL,UAAU,OAAO,mCAA8B,OAAO,EAAE,OAAO,QAAQ,QAAQ,CAAC,CAAC,wBAAwB,cAAc;AAAA,MACzH;AAAA,IACF;AAKA,QAAI,YAAY;AAChB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAI,MAAM,CAAC,EAAG,eAAe,MAAM,IAAI,CAAC,EAAG,WAAY,aAAY;AAAA,IACrE;AACA,QAAI,aAAa,GAAG;AAClB,YAAM,eAAe,MAClB,MAAM,SAAS,EACf;AAAA,QACC,CAAC,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB;AAAA,MAC9C;AACF,UAAI,CAAC,cAAc;AACjB,eAAO;AAAA,UACL,UAAU,OAAO,qBAAqB,MAAM,YAAY,CAAC,EAAG,UAAU,aAAQ,MAAM,SAAS,EAAG,UAAU,QAAQ,MAAM,SAAS,EAAG,EAAE;AAAA,QACxI;AAAA,MACF;AAAA,IACF;AAEA,eAAW,UAAU,uBAAuB;AAC1C,YAAM,SAAS,MACZ,OAAO,CAAC,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,QAAQ,EACnD,IAAI,CAAC,MAAM,EAAE,QAAQ,MAAM,CAAE;AAChC,UAAI,OAAO,WAAW,EAAG;AAEzB,YAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;AAC1D,UAAI,YAAY,UAAU;AACxB,4BAAoB,KAAK,GAAG,OAAO,IAAI,MAAM,EAAE;AAEjD,YAAM,UAAU,OAAO,OAAO,SAAS,CAAC;AACxC,YAAM,WAAW,OAAO,CAAC;AACzB,YAAM,QAAQ,UAAU;AACxB,YAAM,UAAoB,CAAC;AAC3B,UAAI,YAAY,UAAU,iBAAiB;AACzC,gBAAQ;AAAA,UACN,6CAA6C,YAAY,OAAO,iBAAiB,YAAY,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpH;AAAA,MACF;AACA,UAAI,WAAW,SAAS,UAAU,QAAQ;AACxC,gBAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,CAAC,gBAAgB,MAAM,EAAE;AAAA,MAChE;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,cAAc;AACtE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,YAAY;AAAA,QACzH;AAAA,MACF;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,iBAAiB;AACzE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,eAAe;AAAA,QAC5H;AAAA,MACF;AAEA,YAAM,UAAU,QAAQ,SAAS;AACjC,YAAM,QAAuB;AAAA,QAC3B;AAAA,QACA;AAAA,QACA,OAAO,YAAY;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,SAAS;AACX,cAAM,SAAS,QAAQ,KAAK,IAAI;AAChC,eAAO,KAAK,UAAU,OAAO,KAAK,MAAM,KAAK,MAAM,MAAM,EAAE;AAAA,MAC7D;AACA,eAAS,KAAK,KAAK;AAAA,IACrB;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,QAAQ,SAAS,OAAO,WAAW,GAAG,oBAAoB;AAC/E;AAiBO,SAAS,gBAAgB,QAAyC;AACvE,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,CAAC,GAAG,OAAO,MAAM,EAAE;AAC/D;","names":[]}

package/dist/multishot/index.d.ts

@@ -1,4 +1,4 @@
-import { J as JudgeScore } from '../types-DQRY8ZT-.js';
+import { J as JudgeScore } from '../types-BMahhhio.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { M as MatrixResult } from '../types-BUxNaJ8c.js';
 import '../run-record-CP2ObebC.js';

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.95.1",
+    "version": "0.96.1",
     "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",

package/dist/pipelines/index.js

@@ -3,13 +3,14 @@ import {
   classifyFailure,
   compareToBaseline,
   computeToolUseMetrics
-} from "../chunk-NACM5RS7.js";
+} from "../chunk-IH7LYRHL.js";
 import {
   buildTrajectory
 } from "../chunk-RZTMDUO7.js";
 import {
-  interRaterReliability
-} from "../chunk-56GC6VYO.js";
+  interRaterReliability,
+  pearsonR
+} from "../chunk-HRGUJTER.js";
 import {
   aggregateLlm,
   argHash,
@@ -228,7 +229,7 @@ async function judgeAgreementView(store) {
           judgeB: judgeJ,
           dimension: dim,
           commonItems: common.length,
-          pearson: pearson(
+          pearson: pearsonR(
             common.map((c) => c[0]),
             common.map((c) => c[1])
           ),
@@ -243,21 +244,6 @@ async function judgeAgreementView(store) {
     judgeIds
   };
 }
-function pearson(a, b) {
-  if (a.length !== b.length || a.length < 2) return NaN;
-  const mA = a.reduce((s, v) => s + v, 0) / a.length;
-  const mB = b.reduce((s, v) => s + v, 0) / b.length;
-  let num = 0, denA = 0, denB = 0;
-  for (let i = 0; i < a.length; i++) {
-    const dA = a[i] - mA;
-    const dB = b[i] - mB;
-    num += dA * dB;
-    denA += dA * dA;
-    denB += dB * dB;
-  }
-  if (denA === 0 || denB === 0) return denA === 0 && denB === 0 ? 1 : 0;
-  return num / Math.sqrt(denA * denB);
-}
 
 // src/pipelines/regression.ts
 async function regressionView(store, metrics, options) {

package/dist/pipelines/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/pipelines/budget-breach.ts","../../src/pipelines/failure-cluster.ts","../../src/pipelines/first-divergence.ts","../../src/pipelines/judge-agreement.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts","../../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearson(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n\nfunction pearson(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return NaN\n  const mA = a.reduce((s, v) => s + v, 0) / a.length\n  const mB = b.reduce((s, v) => s + v, 0) / b.length\n  let num = 0,\n    denA = 0,\n    denB = 0\n  for (let i = 0; i < a.length; i++) {\n    const dA = a[i]! - mA\n    const dB = b[i]! - mB\n    num += dA * dB\n    denA += dA * dA\n    denB += dB * dB\n  }\n  if (denA === 0 || denB === 0) return denA === 0 && denB === 0 ? 1 : 0\n  return num / Math.sqrt(denA * denB)\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\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      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { argHash, toolSpans } from '../trace/query'\nimport type { TraceStore } from '../trace/store'\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  occurrences: number\n  spanIds: string[]\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const tools = await toolSpans(store, runId)\n    const byKey = new Map<string, { spans: typeof tools; argHash: string; toolName: string }>()\n    for (const t of tools) {\n      const h = argHash(t.args)\n      // NUL delimiter never appears in a tool name or a JSON argHash, so the\n      // key cannot collide; toolName is carried in the bucket rather than\n      // re-derived by splitting the key (which mislabeled any tool whose name\n      // contained the delimiter, e.g. \"shell|grep\").\n      const key = `${t.toolName}\\u0000${h}`\n      const bucket = byKey.get(key) ?? { spans: [], argHash: h, toolName: t.toolName }\n      bucket.spans.push(t)\n      byKey.set(key, bucket)\n    }\n    for (const { spans, argHash: h, toolName } of byKey.values()) {\n      if (spans.length < minOccurrences) continue\n      const sorted = [...spans].sort((a, b) => a.startedAt - b.startedAt)\n      const first = sorted[0]!.startedAt\n      const last = sorted[sorted.length - 1]!.startedAt\n      findings.push({\n        runId,\n        toolName,\n        argHash: h,\n        occurrences: sorted.length,\n        spanIds: sorted.map((s) => s.spanId),\n        windowMs: last - first,\n      })\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;AAC7B,QAAM,OAAO,MAAM,MAAM,SAAS;AAAA,IAChC,YAAY,QAAQ;AAAA,IACpB,WAAW,QAAQ;AAAA,EACrB,CAAC;AACD,QAAM,WAAkC,CAAC;AACzC,QAAM,cAAsC,CAAC;AAC7C,QAAM,aAAqC,CAAC;AAC5C,QAAM,YAAoC,CAAC;AAE3C,aAAW,OAAO,MAAM;AACtB,UAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;AAC5C,eAAW,KAAK,SAAS;AACvB,UAAI,CAAC,EAAE,SAAU;AACjB,YAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;AACzD,eAAS,KAAK;AAAA,QACZ,OAAO,IAAI;AAAA,QACX,YAAY,IAAI;AAAA,QAChB,WAAW,IAAI;AAAA,QACf,WAAW,EAAE;AAAA,QACb,OAAO,EAAE;AAAA,QACT,UAAU,EAAE;AAAA,QACZ;AAAA,QACA,WAAW,EAAE;AAAA,MACf,CAAC;AACD,kBAAY,EAAE,SAAS,KAAK,YAAY,EAAE,SAAS,KAAK,KAAK;AAC7D,iBAAW,IAAI,UAAU,KAAK,WAAW,IAAI,UAAU,KAAK,KAAK;AACjE,UAAI,IAAI,UAAW,WAAU,IAAI,SAAS,KAAK,UAAU,IAAI,SAAS,KAAK,KAAK;AAAA,IAClF;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACzD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,WAAW,KAAK;AAAA,IAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;AAAA,EACxE;AACF;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;AAC/B,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,UAAU,QAAQ,kBAAkB;AAC1C,QAAM,OAAO,MAAM,MAAM,SAAS;AAGlC,QAAM,WAAW,oBAAI,IAAyB;AAC9C,MAAI,gBAAgB;AAEpB,aAAW,OAAO,MAAM;AACtB,QAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,MAAO;AAC/D;AACA,UAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACpD,UAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;AACtD,UAAM,MAAM,gBAAgB,EAAE,KAAK,OAAO,OAAO,GAAG,KAAK;AAEzD,QAAI;AACJ,QAAI;AACJ,QAAI;AACJ,QAAI,IAAI,eAAe;AACrB,YAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,IAAI,aAAa;AAC7D,UAAI,MAAM,SAAS,QAAQ;AACzB,mBAAW,KAAK;AAChB,oBAAY,QAAQ,KAAK,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC5C,WAAW,MAAM,SAAS,SAAS;AACjC,oBAAY,KAAK;AAAA,MACnB;AAAA,IACF;AAEA,QAAI,CAAC,UAAU;AACb,YAAM,KAAK,MAAM,UAAU,OAAO,IAAI,KAAK;AAC3C,YAAM,UAAU,GAAG,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO,EAAE,IAAI;AAC3D,UAAI,SAAS;AACX,mBAAW,QAAQ;AACnB,oBAAY,QAAQ,QAAQ,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC/C;AAAA,IACF;AAIA,QAAI,CAAC,WAAW;AACd,YAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;AACrF,UAAI,OAAO,SAAS,QAAS,aAAY,MAAM;AAAA,IACjD;AAEA,UAAM,MAAM,GAAG,IAAI,YAAY,IAAI,YAAY,EAAE,IAAI,aAAa,EAAE,IAAI,aAAa,EAAE;AACvF,QAAI,UAAU,SAAS,IAAI,GAAG;AAC9B,QAAI,CAAC,SAAS;AACZ,gBAAU;AAAA,QACR,cAAc,IAAI;AAAA,QAClB;AAAA,QACA;AAAA,QACA;AAAA,QACA,UAAU;AAAA,QACV,aAAa,CAAC;AAAA,QACd,cAAc,IAAI;AAAA,QAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;AAAA,MAChD;AACA,eAAS,IAAI,KAAK,OAAO;AAAA,IAC3B;AACA,YAAQ;AACR,QAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,EAAG,SAAQ,YAAY,KAAK,IAAI,UAAU;AAAA,EAC5F;AAEA,QAAM,MAAM,CAAC,GAAG,SAAS,OAAO,CAAC,EAC9B,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AAEzC,SAAO,EAAE,UAAU,KAAK,eAAe,WAAW,KAAK,OAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;AAC5D,QAAM,UAAU,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,OAAO;AACtD,SAAO,SAAS;AAClB;;;ACvFA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;AAC3B,QAAM,CAAC,GAAG,CAAC,IAAI,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;AAC7F,QAAM,KAAK,QAAQ,cAAc;AACjC,QAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;AACtD,WAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,UAAM,QAAQ,EAAE,MAAM,CAAC;AACvB,UAAM,QAAQ,EAAE,MAAM,CAAC;AACvB,QAAI,CAAC,GAAG,OAAO,KAAK,GAAG;AACrB,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA,sBAAsB;AAAA,QACtB;AAAA,QACA;AAAA,QACA,QAAQ,mBAAmB,OAAO,KAAK;AAAA,QACvC,iBAAiB;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AACA,MAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAAQ;AACrC,WAAO,EAAE,MAAM,MAAM,sBAAsB,MAAM,iBAAiB,OAAO;AAAA,EAC3E;AACA,QAAM,SAAqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI;AACjE,QAAM,QAAQ,OAAO,MAAM,SAAS;AAGpC,QAAM,SACJ,WAAW,IACP,0CAA0C,KAAK,iCAC/C,sBAAsB,KAAK,6BAA6B,SAAS,CAAC;AACxE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,sBAAsB;AAAA,IACtB,OAAO,EAAE,MAAM,MAAM;AAAA,IACrB,OAAO,EAAE,MAAM,MAAM;AAAA,IACrB;AAAA,IACA,iBAAiB;AAAA,EACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;AACxE,MAAI,EAAE,KAAK,SAAS,EAAE,KAAK,KAAM,QAAO;AACxC,MAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,OAAQ,QAAO,EAAE,KAAK,aAAa,EAAE,KAAK;AACxF,MAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,MAAO,QAAO,EAAE,KAAK,UAAU,EAAE,KAAK;AACnF,MAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS;AAC7C,WAAO,EAAE,KAAK,cAAc,EAAE,KAAK;AACrC,SAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;AACxE,MAAI,EAAE,KAAK,SAAS,EAAE,KAAK,KAAM,QAAO,QAAQ,EAAE,KAAK,IAAI,OAAO,EAAE,KAAK,IAAI;AAC7E,MAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UAAU;AAC3F,WAAO,QAAQ,EAAE,KAAK,QAAQ,OAAO,EAAE,KAAK,QAAQ;AAAA,EACtD;AACA,MAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAAO;AACnF,WAAO,SAAS,EAAE,KAAK,KAAK,OAAO,EAAE,KAAK,KAAK;AAAA,EACjD;AACA,SAAO,SAAS,EAAE,KAAK,IAAI,SAAS,EAAE,KAAK,IAAI;AACjD;;;AC9DA,eAAsB,mBAAmB,OAAkD;AACzF,QAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,GAAG;AAAA,IACjD,CAAC,MAAsB,EAAE,SAAS;AAAA,EACpC;AACA,MAAI,IAAI,WAAW,EAAG,QAAO,EAAE,OAAO,CAAC,GAAG,YAAY,CAAC,GAAG,UAAU,CAAC,EAAE;AAEvE,QAAM,cAAc,oBAAI,IAAyB;AACjD,aAAW,KAAK,KAAK;AACnB,UAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;AAC7C,QAAI,KAAK,CAAC;AACV,gBAAY,IAAI,EAAE,WAAW,GAAG;AAAA,EAClC;AAEA,QAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,IAAI,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,EAAE,KAAK;AAC9D,QAAM,QAAqB,CAAC;AAC5B,aAAW,CAAC,KAAK,KAAK,KAAK,aAAa;AACtC,UAAM,UAAU,oBAAI,IAAiC;AACrD,eAAW,KAAK,OAAO;AACrB,YAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,KAAK,oBAAI,IAAoB;AAC5D,QAAE,IAAI,EAAE,cAAc,EAAE,KAAK;AAC7B,cAAQ,IAAI,EAAE,SAAS,CAAC;AAAA,IAC1B;AACA,UAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;AACrC,aAAS,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC1C,eAAS,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC9C,cAAM,SAAS,WAAW,CAAC;AAC3B,cAAM,SAAS,WAAW,CAAC;AAC3B,cAAM,IAAI,QAAQ,IAAI,MAAM;AAC5B,cAAM,IAAI,QAAQ,IAAI,MAAM;AAC5B,cAAM,SAAkC,CAAC;AACzC,mBAAW,CAAC,QAAQ,MAAM,KAAK,GAAG;AAChC,gBAAM,SAAS,EAAE,IAAI,MAAM;AAC3B,cAAI,WAAW,OAAW,QAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;AAAA,QACxD;AACA,YAAI,OAAO,SAAS,EAAG;AACvB,cAAM,cAAc,OAAO;AAAA,UACzB,CAAC,CAAC,QAAQ,MAAM,MACd;AAAA,YACE,EAAE,WAAW,QAAQ,WAAW,KAAK,OAAO,QAAQ,WAAW,GAAG;AAAA,YAClE,EAAE,WAAW,QAAQ,WAAW,KAAK,OAAO,QAAQ,WAAW,GAAG;AAAA,UACpE;AAAA,QACJ;AACA,cAAM,IAAI;AAAA,UACR,YAAY,CAAC,EAAG,IAAI,CAAC,GAAG,OAAO,YAAY,IAAI,CAAC,SAAS,KAAK,EAAE,CAAE,CAAC;AAAA,QACrE;AACA,cAAM,KAAK;AAAA,UACT,QAAQ;AAAA,UACR,QAAQ;AAAA,UACR,WAAW;AAAA,UACX,aAAa,OAAO;AAAA,UACpB,SAAS;AAAA,YACP,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAAA,YACtB,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAAA,UACxB;AAAA,UACA,cAAc;AAAA,QAChB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AAAA,IACL,OAAO,MAAM,KAAK,CAAC,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;AAAA,IACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,EAAE,KAAK;AAAA,IACzC;AAAA,EACF;AACF;AAEA,SAAS,QAAQ,GAAa,GAAqB;AACjD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,EAAG,QAAO;AAClD,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,MAAI,MAAM,GACR,OAAO,GACP,OAAO;AACT,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,WAAO,KAAK;AACZ,YAAQ,KAAK;AACb,YAAQ,KAAK;AAAA,EACf;AACA,MAAI,SAAS,KAAK,SAAS,EAAG,QAAO,SAAS,KAAK,SAAS,IAAI,IAAI;AACpE,SAAO,MAAM,KAAK,KAAK,OAAO,IAAI;AACpC;;;ACvFA,eAAsB,eACpB,OACA,SACA,SACyB;AACzB,QAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;AAC1D,QAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;AAC5D,QAAM,UAAU,MAAM,QAAQ;AAAA,IAC5B,QAAQ,IAAI,OAAO,MAAM;AACvB,YAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;AACpD,YAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;AAC9D,YAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;AAChE,aAAO,EAAE,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,gBAAgB,UAAU,UAAU;AAAA,IACnF,CAAC;AAAA,EACH;AACA,SAAO,kBAAkB,SAAS,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;AACnB,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,MAAM;AACpB,UAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;AAChC,QAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,EAAG,KAAI,KAAK,CAAC;AAAA,EAClD;AACA,SAAO;AACT;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,gBAAgB;AACnB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,gBAAgB;AACnB,eAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;AAAA,MAClD;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;AClDA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;AAC1B,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;AAE5D,QAAM,WAA+B,CAAC;AACtC,aAAW,EAAE,MAAM,KAAK,MAAM;AAC5B,UAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,UAAM,QAAQ,oBAAI,IAAwE;AAC1F,eAAW,KAAK,OAAO;AACrB,YAAM,IAAI,QAAQ,EAAE,IAAI;AAKxB,YAAM,MAAM,GAAG,EAAE,QAAQ,KAAS,CAAC;AACnC,YAAM,SAAS,MAAM,IAAI,GAAG,KAAK,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,UAAU,EAAE,SAAS;AAC/E,aAAO,MAAM,KAAK,CAAC;AACnB,YAAM,IAAI,KAAK,MAAM;AAAA,IACvB;AACA,eAAW,EAAE,OAAO,SAAS,GAAG,SAAS,KAAK,MAAM,OAAO,GAAG;AAC5D,UAAI,MAAM,SAAS,eAAgB;AACnC,YAAM,SAAS,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAClE,YAAM,QAAQ,OAAO,CAAC,EAAG;AACzB,YAAM,OAAO,OAAO,OAAO,SAAS,CAAC,EAAG;AACxC,eAAS,KAAK;AAAA,QACZ;AAAA,QACA;AAAA,QACA,SAAS;AAAA,QACT,aAAa,OAAO;AAAA,QACpB,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,MAAM;AAAA,QACnC,UAAU,OAAO;AAAA,MACnB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACzD,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;AAAA,IACtE,WAAW,KAAK;AAAA,EAClB;AACF;;;AChDA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;AAC1B,QAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,EAAE,KAAK;AAE1F,QAAM,QAA4B,CAAC;AACnC,MAAI,aAAa;AACjB,MAAI,cAAc;AAClB,aAAW,SAAS,MAAM;AACxB,UAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,QAAI,MAAM,WAAW,GAAG;AACtB,YAAM,KAAK,EAAE,OAAO,aAAa,GAAG,YAAY,GAAG,WAAW,EAAE,CAAC;AACjE;AAAA,IACF;AACA,UAAM,OAAO,MAAM,SAAS,OAAO,KAAK;AAOxC,UAAM,YAAY,CAAC,GAAG,IAAI,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACpE,UAAM,aAAa,UAAU,IAAI,CAAC,MAAM,EAAE,SAAS;AACnD,UAAM,aAAa,gBAAgB,SAAS;AAC5C,QAAI,SAAS;AACb,eAAW,KAAK,OAAO;AACrB,UAAI,EAAE,WAAW,SAAS;AACxB;AACA;AAAA,MACF;AAEA,YAAM,SAAS,WAAW,YAAY,EAAE,SAAS;AACjD,UAAI,QAAQ,aAAa;AACvB,YAAI,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,EAAG;AAAA,MACjE,OAAO;AAKL,cAAM,YAAY,UAAU,EAAE,MAAM;AACpC,YAAI,cAAc,GAAI;AACtB,cAAM,WAAW,WAAW,MAAM,KAAK;AACvC,cAAM,OAAO,SAAS,SAAS,UAAU,MAAM,GAAG,GAAG,CAAC;AACtD,YAAI,CAAC,KAAM;AAAA,MACb;AAAA,IACF;AACA,UAAM,YAAY,SAAS,MAAM;AACjC,UAAM,KAAK,EAAE,OAAO,aAAa,QAAQ,YAAY,MAAM,QAAQ,UAAU,CAAC;AAC9E,kBAAc,MAAM;AACpB,mBAAe;AAAA,EACjB;AACA,SAAO,EAAE,OAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa,EAAE;AAClF;AAOA,SAAS,gBAAgB,OAA4B;AACnD,QAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;AACjD,SAAO,MAAM,MAAM,IAAI;AACvB,WAAS,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAAK;AAC1C,UAAM,MAAM,MAAM,CAAC,EAAG,SAAS;AAAA,MAAI,CAAC,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU;AAAA,IAC9C,EAAE,KAAK,IAAI;AACX,WAAO,CAAC,IAAI,GAAG,GAAG;AAAA,EAAK,OAAO,IAAI,CAAC,CAAC;AAAA,EACtC;AACA,SAAO;AACT;AAGA,SAAS,WAAW,QAAkB,QAAwB;AAC5D,MAAI,KAAK;AACT,MAAI,KAAK,OAAO;AAChB,SAAO,KAAK,IAAI;AACd,UAAM,MAAO,KAAK,OAAQ;AAC1B,QAAI,OAAO,GAAG,KAAM,OAAQ,MAAK,MAAM;AAAA,QAClC,MAAK;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,UAAU,GAAoB;AACrC,MAAI,MAAM,QAAQ,MAAM,OAAW,QAAO;AAC1C,MAAI,OAAO,MAAM,SAAU,QAAO;AAClC,MAAI;AACF,WAAO,KAAK,UAAU,CAAC;AAAA,EACzB,QAAQ;AACN,WAAO,OAAO,CAAC;AAAA,EACjB;AACF;","names":[]}
+{"version":3,"sources":["../../src/pipelines/budget-breach.ts","../../src/pipelines/failure-cluster.ts","../../src/pipelines/first-divergence.ts","../../src/pipelines/judge-agreement.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts","../../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\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      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { argHash, toolSpans } from '../trace/query'\nimport type { TraceStore } from '../trace/store'\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  occurrences: number\n  spanIds: string[]\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const tools = await toolSpans(store, runId)\n    const byKey = new Map<string, { spans: typeof tools; argHash: string; toolName: string }>()\n    for (const t of tools) {\n      const h = argHash(t.args)\n      // NUL delimiter never appears in a tool name or a JSON argHash, so the\n      // key cannot collide; toolName is carried in the bucket rather than\n      // re-derived by splitting the key (which mislabeled any tool whose name\n      // contained the delimiter, e.g. \"shell|grep\").\n      const key = `${t.toolName}\\u0000${h}`\n      const bucket = byKey.get(key) ?? { spans: [], argHash: h, toolName: t.toolName }\n      bucket.spans.push(t)\n      byKey.set(key, bucket)\n    }\n    for (const { spans, argHash: h, toolName } of byKey.values()) {\n      if (spans.length < minOccurrences) continue\n      const sorted = [...spans].sort((a, b) => a.startedAt - b.startedAt)\n      const first = sorted[0]!.startedAt\n      const last = sorted[sorted.length - 1]!.startedAt\n      findings.push({\n        runId,\n        toolName,\n        argHash: h,\n        occurrences: sorted.length,\n        spanIds: sorted.map((s) => s.spanId),\n        windowMs: last - first,\n      })\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;AAC7B,QAAM,OAAO,MAAM,MAAM,SAAS;AAAA,IAChC,YAAY,QAAQ;AAAA,IACpB,WAAW,QAAQ;AAAA,EACrB,CAAC;AACD,QAAM,WAAkC,CAAC;AACzC,QAAM,cAAsC,CAAC;AAC7C,QAAM,aAAqC,CAAC;AAC5C,QAAM,YAAoC,CAAC;AAE3C,aAAW,OAAO,MAAM;AACtB,UAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;AAC5C,eAAW,KAAK,SAAS;AACvB,UAAI,CAAC,EAAE,SAAU;AACjB,YAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;AACzD,eAAS,KAAK;AAAA,QACZ,OAAO,IAAI;AAAA,QACX,YAAY,IAAI;AAAA,QAChB,WAAW,IAAI;AAAA,QACf,WAAW,EAAE;AAAA,QACb,OAAO,EAAE;AAAA,QACT,UAAU,EAAE;AAAA,QACZ;AAAA,QACA,WAAW,EAAE;AAAA,MACf,CAAC;AACD,kBAAY,EAAE,SAAS,KAAK,YAAY,EAAE,SAAS,KAAK,KAAK;AAC7D,iBAAW,IAAI,UAAU,KAAK,WAAW,IAAI,UAAU,KAAK,KAAK;AACjE,UAAI,IAAI,UAAW,WAAU,IAAI,SAAS,KAAK,UAAU,IAAI,SAAS,KAAK,KAAK;AAAA,IAClF;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACzD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,WAAW,KAAK;AAAA,IAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;AAAA,EACxE;AACF;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;AAC/B,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,UAAU,QAAQ,kBAAkB;AAC1C,QAAM,OAAO,MAAM,MAAM,SAAS;AAGlC,QAAM,WAAW,oBAAI,IAAyB;AAC9C,MAAI,gBAAgB;AAEpB,aAAW,OAAO,MAAM;AACtB,QAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,MAAO;AAC/D;AACA,UAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACpD,UAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;AACtD,UAAM,MAAM,gBAAgB,EAAE,KAAK,OAAO,OAAO,GAAG,KAAK;AAEzD,QAAI;AACJ,QAAI;AACJ,QAAI;AACJ,QAAI,IAAI,eAAe;AACrB,YAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,IAAI,aAAa;AAC7D,UAAI,MAAM,SAAS,QAAQ;AACzB,mBAAW,KAAK;AAChB,oBAAY,QAAQ,KAAK,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC5C,WAAW,MAAM,SAAS,SAAS;AACjC,oBAAY,KAAK;AAAA,MACnB;AAAA,IACF;AAEA,QAAI,CAAC,UAAU;AACb,YAAM,KAAK,MAAM,UAAU,OAAO,IAAI,KAAK;AAC3C,YAAM,UAAU,GAAG,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO,EAAE,IAAI;AAC3D,UAAI,SAAS;AACX,mBAAW,QAAQ;AACnB,oBAAY,QAAQ,QAAQ,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC/C;AAAA,IACF;AAIA,QAAI,CAAC,WAAW;AACd,YAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;AACrF,UAAI,OAAO,SAAS,QAAS,aAAY,MAAM;AAAA,IACjD;AAEA,UAAM,MAAM,GAAG,IAAI,YAAY,IAAI,YAAY,EAAE,IAAI,aAAa,EAAE,IAAI,aAAa,EAAE;AACvF,QAAI,UAAU,SAAS,IAAI,GAAG;AAC9B,QAAI,CAAC,SAAS;AACZ,gBAAU;AAAA,QACR,cAAc,IAAI;AAAA,QAClB;AAAA,QACA;AAAA,QACA;AAAA,QACA,UAAU;AAAA,QACV,aAAa,CAAC;AAAA,QACd,cAAc,IAAI;AAAA,QAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;AAAA,MAChD;AACA,eAAS,IAAI,KAAK,OAAO;AAAA,IAC3B;AACA,YAAQ;AACR,QAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,EAAG,SAAQ,YAAY,KAAK,IAAI,UAAU;AAAA,EAC5F;AAEA,QAAM,MAAM,CAAC,GAAG,SAAS,OAAO,CAAC,EAC9B,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AAEzC,SAAO,EAAE,UAAU,KAAK,eAAe,WAAW,KAAK,OAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;AAC5D,QAAM,UAAU,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,OAAO;AACtD,SAAO,SAAS;AAClB;;;ACvFA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;AAC3B,QAAM,CAAC,GAAG,CAAC,IAAI,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;AAC7F,QAAM,KAAK,QAAQ,cAAc;AACjC,QAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;AACtD,WAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,UAAM,QAAQ,EAAE,MAAM,CAAC;AACvB,UAAM,QAAQ,EAAE,MAAM,CAAC;AACvB,QAAI,CAAC,GAAG,OAAO,KAAK,GAAG;AACrB,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA,sBAAsB;AAAA,QACtB;AAAA,QACA;AAAA,QACA,QAAQ,mBAAmB,OAAO,KAAK;AAAA,QACvC,iBAAiB;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AACA,MAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAAQ;AACrC,WAAO,EAAE,MAAM,MAAM,sBAAsB,MAAM,iBAAiB,OAAO;AAAA,EAC3E;AACA,QAAM,SAAqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI;AACjE,QAAM,QAAQ,OAAO,MAAM,SAAS;AAGpC,QAAM,SACJ,WAAW,IACP,0CAA0C,KAAK,iCAC/C,sBAAsB,KAAK,6BAA6B,SAAS,CAAC;AACxE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,sBAAsB;AAAA,IACtB,OAAO,EAAE,MAAM,MAAM;AAAA,IACrB,OAAO,EAAE,MAAM,MAAM;AAAA,IACrB;AAAA,IACA,iBAAiB;AAAA,EACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;AACxE,MAAI,EAAE,KAAK,SAAS,EAAE,KAAK,KAAM,QAAO;AACxC,MAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,OAAQ,QAAO,EAAE,KAAK,aAAa,EAAE,KAAK;AACxF,MAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,MAAO,QAAO,EAAE,KAAK,UAAU,EAAE,KAAK;AACnF,MAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS;AAC7C,WAAO,EAAE,KAAK,cAAc,EAAE,KAAK;AACrC,SAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;AACxE,MAAI,EAAE,KAAK,SAAS,EAAE,KAAK,KAAM,QAAO,QAAQ,EAAE,KAAK,IAAI,OAAO,EAAE,KAAK,IAAI;AAC7E,MAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UAAU;AAC3F,WAAO,QAAQ,EAAE,KAAK,QAAQ,OAAO,EAAE,KAAK,QAAQ;AAAA,EACtD;AACA,MAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAAO;AACnF,WAAO,SAAS,EAAE,KAAK,KAAK,OAAO,EAAE,KAAK,KAAK;AAAA,EACjD;AACA,SAAO,SAAS,EAAE,KAAK,IAAI,SAAS,EAAE,KAAK,IAAI;AACjD;;;AC9DA,eAAsB,mBAAmB,OAAkD;AACzF,QAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,GAAG;AAAA,IACjD,CAAC,MAAsB,EAAE,SAAS;AAAA,EACpC;AACA,MAAI,IAAI,WAAW,EAAG,QAAO,EAAE,OAAO,CAAC,GAAG,YAAY,CAAC,GAAG,UAAU,CAAC,EAAE;AAEvE,QAAM,cAAc,oBAAI,IAAyB;AACjD,aAAW,KAAK,KAAK;AACnB,UAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;AAC7C,QAAI,KAAK,CAAC;AACV,gBAAY,IAAI,EAAE,WAAW,GAAG;AAAA,EAClC;AAEA,QAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,IAAI,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,EAAE,KAAK;AAC9D,QAAM,QAAqB,CAAC;AAC5B,aAAW,CAAC,KAAK,KAAK,KAAK,aAAa;AACtC,UAAM,UAAU,oBAAI,IAAiC;AACrD,eAAW,KAAK,OAAO;AACrB,YAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,KAAK,oBAAI,IAAoB;AAC5D,QAAE,IAAI,EAAE,cAAc,EAAE,KAAK;AAC7B,cAAQ,IAAI,EAAE,SAAS,CAAC;AAAA,IAC1B;AACA,UAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;AACrC,aAAS,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC1C,eAAS,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC9C,cAAM,SAAS,WAAW,CAAC;AAC3B,cAAM,SAAS,WAAW,CAAC;AAC3B,cAAM,IAAI,QAAQ,IAAI,MAAM;AAC5B,cAAM,IAAI,QAAQ,IAAI,MAAM;AAC5B,cAAM,SAAkC,CAAC;AACzC,mBAAW,CAAC,QAAQ,MAAM,KAAK,GAAG;AAChC,gBAAM,SAAS,EAAE,IAAI,MAAM;AAC3B,cAAI,WAAW,OAAW,QAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;AAAA,QACxD;AACA,YAAI,OAAO,SAAS,EAAG;AACvB,cAAM,cAAc,OAAO;AAAA,UACzB,CAAC,CAAC,QAAQ,MAAM,MACd;AAAA,YACE,EAAE,WAAW,QAAQ,WAAW,KAAK,OAAO,QAAQ,WAAW,GAAG;AAAA,YAClE,EAAE,WAAW,QAAQ,WAAW,KAAK,OAAO,QAAQ,WAAW,GAAG;AAAA,UACpE;AAAA,QACJ;AACA,cAAM,IAAI;AAAA,UACR,YAAY,CAAC,EAAG,IAAI,CAAC,GAAG,OAAO,YAAY,IAAI,CAAC,SAAS,KAAK,EAAE,CAAE,CAAC;AAAA,QACrE;AACA,cAAM,KAAK;AAAA,UACT,QAAQ;AAAA,UACR,QAAQ;AAAA,UACR,WAAW;AAAA,UACX,aAAa,OAAO;AAAA,UACpB,SAAS;AAAA,YACP,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAAA,YACtB,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAAA,UACxB;AAAA,UACA,cAAc;AAAA,QAChB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AAAA,IACL,OAAO,MAAM,KAAK,CAAC,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;AAAA,IACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,EAAE,KAAK;AAAA,IACzC;AAAA,EACF;AACF;;;ACrEA,eAAsB,eACpB,OACA,SACA,SACyB;AACzB,QAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;AAC1D,QAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;AAC5D,QAAM,UAAU,MAAM,QAAQ;AAAA,IAC5B,QAAQ,IAAI,OAAO,MAAM;AACvB,YAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;AACpD,YAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;AAC9D,YAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;AAChE,aAAO,EAAE,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,gBAAgB,UAAU,UAAU;AAAA,IACnF,CAAC;AAAA,EACH;AACA,SAAO,kBAAkB,SAAS,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;AACnB,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,MAAM;AACpB,UAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;AAChC,QAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,EAAG,KAAI,KAAK,CAAC;AAAA,EAClD;AACA,SAAO;AACT;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,gBAAgB;AACnB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,gBAAgB;AACnB,eAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;AAAA,MAClD;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;AClDA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;AAC1B,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;AAE5D,QAAM,WAA+B,CAAC;AACtC,aAAW,EAAE,MAAM,KAAK,MAAM;AAC5B,UAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,UAAM,QAAQ,oBAAI,IAAwE;AAC1F,eAAW,KAAK,OAAO;AACrB,YAAM,IAAI,QAAQ,EAAE,IAAI;AAKxB,YAAM,MAAM,GAAG,EAAE,QAAQ,KAAS,CAAC;AACnC,YAAM,SAAS,MAAM,IAAI,GAAG,KAAK,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,UAAU,EAAE,SAAS;AAC/E,aAAO,MAAM,KAAK,CAAC;AACnB,YAAM,IAAI,KAAK,MAAM;AAAA,IACvB;AACA,eAAW,EAAE,OAAO,SAAS,GAAG,SAAS,KAAK,MAAM,OAAO,GAAG;AAC5D,UAAI,MAAM,SAAS,eAAgB;AACnC,YAAM,SAAS,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAClE,YAAM,QAAQ,OAAO,CAAC,EAAG;AACzB,YAAM,OAAO,OAAO,OAAO,SAAS,CAAC,EAAG;AACxC,eAAS,KAAK;AAAA,QACZ;AAAA,QACA;AAAA,QACA,SAAS;AAAA,QACT,aAAa,OAAO;AAAA,QACpB,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,MAAM;AAAA,QACnC,UAAU,OAAO;AAAA,MACnB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACzD,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;AAAA,IACtE,WAAW,KAAK;AAAA,EAClB;AACF;;;AChDA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;AAC1B,QAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,EAAE,KAAK;AAE1F,QAAM,QAA4B,CAAC;AACnC,MAAI,aAAa;AACjB,MAAI,cAAc;AAClB,aAAW,SAAS,MAAM;AACxB,UAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,QAAI,MAAM,WAAW,GAAG;AACtB,YAAM,KAAK,EAAE,OAAO,aAAa,GAAG,YAAY,GAAG,WAAW,EAAE,CAAC;AACjE;AAAA,IACF;AACA,UAAM,OAAO,MAAM,SAAS,OAAO,KAAK;AAOxC,UAAM,YAAY,CAAC,GAAG,IAAI,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACpE,UAAM,aAAa,UAAU,IAAI,CAAC,MAAM,EAAE,SAAS;AACnD,UAAM,aAAa,gBAAgB,SAAS;AAC5C,QAAI,SAAS;AACb,eAAW,KAAK,OAAO;AACrB,UAAI,EAAE,WAAW,SAAS;AACxB;AACA;AAAA,MACF;AAEA,YAAM,SAAS,WAAW,YAAY,EAAE,SAAS;AACjD,UAAI,QAAQ,aAAa;AACvB,YAAI,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,EAAG;AAAA,MACjE,OAAO;AAKL,cAAM,YAAY,UAAU,EAAE,MAAM;AACpC,YAAI,cAAc,GAAI;AACtB,cAAM,WAAW,WAAW,MAAM,KAAK;AACvC,cAAM,OAAO,SAAS,SAAS,UAAU,MAAM,GAAG,GAAG,CAAC;AACtD,YAAI,CAAC,KAAM;AAAA,MACb;AAAA,IACF;AACA,UAAM,YAAY,SAAS,MAAM;AACjC,UAAM,KAAK,EAAE,OAAO,aAAa,QAAQ,YAAY,MAAM,QAAQ,UAAU,CAAC;AAC9E,kBAAc,MAAM;AACpB,mBAAe;AAAA,EACjB;AACA,SAAO,EAAE,OAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa,EAAE;AAClF;AAOA,SAAS,gBAAgB,OAA4B;AACnD,QAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;AACjD,SAAO,MAAM,MAAM,IAAI;AACvB,WAAS,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAAK;AAC1C,UAAM,MAAM,MAAM,CAAC,EAAG,SAAS;AAAA,MAAI,CAAC,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU;AAAA,IAC9C,EAAE,KAAK,IAAI;AACX,WAAO,CAAC,IAAI,GAAG,GAAG;AAAA,EAAK,OAAO,IAAI,CAAC,CAAC;AAAA,EACtC;AACA,SAAO;AACT;AAGA,SAAS,WAAW,QAAkB,QAAwB;AAC5D,MAAI,KAAK;AACT,MAAI,KAAK,OAAO;AAChB,SAAO,KAAK,IAAI;AACd,UAAM,MAAO,KAAK,OAAQ;AAC1B,QAAI,OAAO,GAAG,KAAM,OAAQ,MAAK,MAAM;AAAA,QAClC,MAAK;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,UAAU,GAAoB;AACrC,MAAI,MAAM,QAAQ,MAAM,OAAW,QAAO;AAC1C,MAAI,OAAO,MAAM,SAAU,QAAO;AAClC,MAAI;AACF,WAAO,KAAK,UAAU,CAAC;AAAA,EACzB,QAAQ;AACN,WAAO,OAAO,CAAC;AAAA,EACjB;AACF;","names":[]}

package/dist/{provenance-CncDq9qE.d.ts → provenance-DJ--Jcbg.d.ts}

@@ -1,9 +1,9 @@
-import { S as Scenario, G as Gate, l as GateResult, k as GateContext, C as CampaignResult, p as Mutator, c as SurfaceProposer, M as MutableSurface, d as GateDecision } from './types-DQRY8ZT-.js';
+import { S as Scenario, f as Gate, G as GateResult, g as GateContext, C as CampaignResult, h as Mutator, e as SurfaceProposer, M as MutableSurface, i as GateDecision } from './types-BMahhhio.js';
 import { R as RedTeamCase } from './red-team-BWdoyleI.js';
 import { R as RunRecord } from './run-record-CP2ObebC.js';
 import { D as Direction } from './pareto-E-pembql.js';
-import { a as PairedBootstrapResult } from './statistics-CCJpTGOS.js';
-import { b as RunCampaignOptions, C as CampaignStorage } from './gepa-C1NCIZ9o.js';
+import { a as PairedBootstrapResult } from './statistics-xP-cWc5k.js';
+import { R as RunCampaignOptions, C as CampaignStorage } from './gepa-Dprxvz8r.js';
 import { HostedClient, TraceSpanEvent } from './hosted/index.js';
 
 /**

package/dist/reporting.d.ts

@@ -1,7 +1,7 @@
 export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-C2hDKM8Z.js';
 export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, f as ReleaseConfidenceScorecard, g as ReleaseConfidenceStatus, h as ReleaseConfidenceThresholds, i as ReleaseTraceEvidence, j as RenderReleaseReportOptions, V as Verdict, k as assertReleaseConfidence, l as bootstrapCi, m as evaluateReleaseConfidence, n as judgeReplayGate, r as renderReleaseReport } from './release-report-pidWUMZ2.js';
 export { I as InterimReleaseConfidence, a as InterimReleaseConfidenceInput, P as PairedEvalueOptions, b as PairedEvalueSequence, c as PairedEvalueStep, S as SequentialDecision, e as evaluateInterimReleaseConfidence, p as pairedEvalueSequence } from './sequential-5iSVfzl2.js';
-export { P as PairedBootstrapOptions, a as PairedBootstrapResult, b as benjaminiHochberg, p as pairedBootstrap, w as wilcoxonSignedRank } from './statistics-CCJpTGOS.js';
+export { P as PairedBootstrapOptions, a as PairedBootstrapResult, b as benjaminiHochberg, p as pairedBootstrap, w as wilcoxonSignedRank } from './statistics-xP-cWc5k.js';
 export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, P as ParetoFigureSpec, c as ParetoPoint, R as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as ResearchReport, e as ResearchReportCandidate, f as ResearchReportDecision, g as ResearchReportMethodology, h as ResearchReportOptions, i as ResearchReportRecommendation, S as SummaryTable, j as SummaryTableOptions, k as SummaryTableRow, l as gainHistogram, p as paretoChart, r as researchReport, s as summaryTable } from './summary-report-CInXwsza.js';
 import './run-record-CP2ObebC.js';
 import '@tangle-network/agent-interface';

package/dist/reporting.js

@@ -4,10 +4,10 @@ import {
   evaluateReleaseConfidence,
   judgeReplayGate,
   renderReleaseReport
-} from "./chunk-QTMYW64F.js";
+} from "./chunk-VDGPPGE3.js";
 import {
   rubricPredictiveValidity
-} from "./chunk-YRZ4M5GS.js";
+} from "./chunk-NF7OZ4J7.js";
 import {
   evaluateInterimReleaseConfidence,
   pairedEvalueSequence
@@ -18,12 +18,12 @@ import {
   paretoChart,
   researchReport,
   summaryTable
-} from "./chunk-URWVKRCS.js";
+} from "./chunk-6GT4NI4V.js";
 import {
   benjaminiHochberg,
   pairedBootstrap,
   wilcoxonSignedRank
-} from "./chunk-56GC6VYO.js";
+} from "./chunk-HRGUJTER.js";
 import "./chunk-VSMTAMNK.js";
 import "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";

package/dist/rl.d.ts

@@ -10,7 +10,7 @@ import { R as Researcher, F as FailureMode, S as SteeringChange, E as Experiment
 export { r as runEvalCampaign } from './researcher-Jr8ME1dZ.js';
 import { a as VerificationReport } from './multi-layer-verifier-DUZXrPDA.js';
 import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
-import { C as CampaignResult } from './types-DQRY8ZT-.js';
+import { C as CampaignResult } from './types-BMahhhio.js';
 import '@tangle-network/agent-interface';
 import './errors-CzMUYo7b.js';
 import './schema-m0gsnbt3.js';

package/dist/rl.js

@@ -9,26 +9,26 @@ import {
   extractVerifiableReward,
   extractVerifiableRewardsFromRecords,
   filterDeterministicallyRewarded
-} from "./chunk-5HRORJQY.js";
+} from "./chunk-AIGWQEME.js";
 import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
 } from "./chunk-3RF76KTD.js";
 import {
   runEvalCampaign
-} from "./chunk-X74V6ESX.js";
+} from "./chunk-OKQ2LAT7.js";
 import "./chunk-CWNP4DV4.js";
 import {
   rubricPredictiveValidity
-} from "./chunk-YRZ4M5GS.js";
+} from "./chunk-NF7OZ4J7.js";
 import {
   evaluateInterimReleaseConfidence
 } from "./chunk-MAZ26DC7.js";
-import "./chunk-URWVKRCS.js";
+import "./chunk-6GT4NI4V.js";
 import {
   benjaminiHochberg,
   wilcoxonSignedRank
-} from "./chunk-56GC6VYO.js";
+} from "./chunk-HRGUJTER.js";
 import {
   observationsFromRunRecords,
   thompsonCurriculum,

package/dist/{run-campaign-WXY7KI67.js → run-campaign-OWCFOEQG.js}

@@ -1,10 +1,10 @@
 import {
   runCampaign
-} from "./chunk-2T4EZACH.js";
-import "./chunk-56GC6VYO.js";
+} from "./chunk-HRGTA6U5.js";
+import "./chunk-HRGUJTER.js";
 import "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";
 export {
   runCampaign
 };
-//# sourceMappingURL=run-campaign-WXY7KI67.js.map
+//# sourceMappingURL=run-campaign-OWCFOEQG.js.map