@tangle-network/agent-eval 0.129.0 → 0.130.0

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

package/dist/metrics-C9YY1OcL.js

@@ -0,0 +1,239 @@
+//#region src/metrics.ts
+/** Per-1K token pricing for exact model ids. */
+const MODEL_PRICING = {
+	"gpt-4o": {
+		input: .0025,
+		output: .01
+	},
+	"gpt-4o-mini": {
+		input: 15e-5,
+		output: 6e-4
+	},
+	"gpt-4-turbo": {
+		input: .01,
+		output: .03
+	},
+	"claude-sonnet-4-20250514": {
+		input: .003,
+		output: .015
+	},
+	"claude-opus-4-20250514": {
+		input: .015,
+		output: .075
+	},
+	"claude-3-haiku-20240307": {
+		input: 25e-5,
+		output: .00125
+	}
+};
+/** Family-level pricing fallbacks (per-1K), matched against a normalized id
+*  after exact lookup misses. Ordered — first match wins. Covers the model
+*  ids actually used through the Tangle router + cli-bridge harnesses
+*  (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,
+*  `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),
+*  none of which appear in the exact table above — without this they priced
+*  to a silent $0, blanking every cost/Pareto axis downstream. */
+const FAMILY_PRICING = [
+	[/claude.*opus/, {
+		input: .015,
+		output: .075
+	}],
+	[/claude.*haiku/, {
+		input: 8e-4,
+		output: .004
+	}],
+	[/claude.*sonnet|claude-code|claude-sonnet/, {
+		input: .003,
+		output: .015
+	}],
+	[/gpt-4o-mini/, {
+		input: 15e-5,
+		output: 6e-4
+	}],
+	[/gpt-5|gpt-4\.1|o[134]\b/, {
+		input: .00125,
+		output: .01
+	}],
+	[/gpt-4o|gpt-4/, {
+		input: .0025,
+		output: .01
+	}],
+	[/deepseek/, {
+		input: 3e-4,
+		output: .0011
+	}],
+	[/glm|zhipu|zai/, {
+		input: 6e-4,
+		output: .0022
+	}],
+	[/kimi|moonshot/, {
+		input: 6e-4,
+		output: .0025
+	}],
+	[/qwen/, {
+		input: 4e-4,
+		output: .0012
+	}],
+	[/gemini.*flash/, {
+		input: 1e-4,
+		output: 4e-4
+	}],
+	[/gemini/, {
+		input: .00125,
+		output: .005
+	}],
+	[/llama/, {
+		input: 2e-4,
+		output: 6e-4
+	}]
+];
+/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,
+*  and keep the final harness/provider-prefixed segment so family regexes
+*  match (`opencode/zai-coding-plan/glm-5.1` → `glm-5.1`). */
+function normalizeModelId(model) {
+	return (model.split("@")[0] ?? model).trim().toLowerCase();
+}
+/** Resolve pricing for a model id: exact table, then family fallback.
+*  Returns null when the id matches nothing (caller decides — never a
+*  silent-zero masquerading as a real $0 cost). */
+function resolveModelPricing(model) {
+	if (MODEL_PRICING[model]) return MODEL_PRICING[model];
+	const id = normalizeModelId(model);
+	if (MODEL_PRICING[id]) return MODEL_PRICING[id];
+	for (const [pattern, price] of FAMILY_PRICING) if (pattern.test(id)) return price;
+	return null;
+}
+/** True when `model` has known pricing (exact or family). Lets cost-aware
+*  callers distinguish a real $0 from an unpriced model. */
+function isModelPriced(model) {
+	return resolveModelPricing(model) !== null;
+}
+const warnedUnpricedModels = /* @__PURE__ */ new Set();
+/** Estimate token count from string length (chars / 4 approximation) */
+function estimateTokens(text) {
+	return Math.ceil(text.length / 4);
+}
+/** Calculate cost in USD from token counts and model. Unknown models warn
+*  once (not a silent zero) and return 0 so callers that ignore pricing keep
+*  working; cost-sensitive callers should gate on {@link isModelPriced}. */
+function estimateCost(inputTokens, outputTokens, model) {
+	const pricing = resolveModelPricing(model);
+	if (!pricing) {
+		if (!warnedUnpricedModels.has(model)) {
+			warnedUnpricedModels.add(model);
+			console.warn(`estimateCost: no pricing for model "${model}" — returning 0; add it to MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)`);
+		}
+		return 0;
+	}
+	return inputTokens / 1e3 * pricing.input + outputTokens / 1e3 * pricing.output;
+}
+/**
+* TokenCounter — accumulates token usage and cost across turns.
+*/
+var TokenCounter = class {
+	totalInput = 0;
+	totalOutput = 0;
+	totalCost = 0;
+	model;
+	constructor(model = "gpt-4o") {
+		this.model = model;
+	}
+	/** Record tokens for a turn, returns per-turn cost */
+	record(inputTokens, outputTokens) {
+		this.totalInput += inputTokens;
+		this.totalOutput += outputTokens;
+		const cost = estimateCost(inputTokens, outputTokens, this.model);
+		this.totalCost += cost;
+		return cost;
+	}
+	/** Estimate and record from raw text */
+	recordFromText(inputText, outputText) {
+		const inputTokens = estimateTokens(inputText);
+		const outputTokens = estimateTokens(outputText);
+		return {
+			inputTokens,
+			outputTokens,
+			cost: this.record(inputTokens, outputTokens)
+		};
+	}
+	getTotalInput() {
+		return this.totalInput;
+	}
+	getTotalOutput() {
+		return this.totalOutput;
+	}
+	getTotalCost() {
+		return this.totalCost;
+	}
+};
+/**
+* MetricsCollector — collects per-turn metrics from the product.
+*
+* After each turn, queries the product's APIs to measure state changes.
+*/
+var MetricsCollector = class {
+	client;
+	workspaceId;
+	metrics = [];
+	constructor(client, workspaceId) {
+		this.client = client;
+		this.workspaceId = workspaceId;
+	}
+	/** Collect metrics after a turn completes */
+	async collect(turn, responseLatencyMs, responseChars, codeBlocksProduced, blocksExtracted, completionCriteriaMet, completionCriteriaTotal, qualityScore, inputTokens = 0, outputTokens = 0, estimatedCostUsd = 0) {
+		const state = await this.getState();
+		const m = {
+			turn,
+			timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+			tasks: state.tasks,
+			events: state.events,
+			proposals: state.proposals,
+			vaultFiles: state.vaultFiles.length,
+			responseLatencyMs,
+			responseChars,
+			codeBlocksProduced,
+			blocksExtracted,
+			qualityScore,
+			inputTokens,
+			outputTokens,
+			estimatedCostUsd,
+			totalCostUsd: estimatedCostUsd,
+			completionPercent: completionCriteriaTotal > 0 ? completionCriteriaMet / completionCriteriaTotal * 100 : 0
+		};
+		this.metrics.push(m);
+		return m;
+	}
+	/** Get current product state */
+	async getState() {
+		const [tasks, events, approvals, vaultFiles] = await Promise.all([
+			this.client.getTasks(this.workspaceId),
+			this.client.getEvents(this.workspaceId),
+			this.client.getApprovals(this.workspaceId),
+			this.client.getVaultTree(this.workspaceId)
+		]);
+		return {
+			tasks: tasks.length,
+			events: events.length,
+			proposals: {
+				pending: approvals.filter((a) => a.status === "pending").length,
+				approved: approvals.filter((a) => a.status === "approved").length,
+				rejected: approvals.filter((a) => a.status === "rejected").length
+			},
+			vaultFiles,
+			codeBlocks: 0,
+			generations: 0
+		};
+	}
+	/** Get all collected metrics */
+	getMetrics() {
+		return [...this.metrics];
+	}
+	/** Get convergence curve (completion% over turns) */
+	getConvergenceCurve() {
+		return this.metrics.map((m) => m.completionPercent);
+	}
+};
+//#endregion
+export { estimateTokens as a, estimateCost as i, MetricsCollector as n, isModelPriced as o, TokenCounter as r, resolveModelPricing as s, MODEL_PRICING as t };
+
+//# sourceMappingURL=metrics-C9YY1OcL.js.map

package/dist/metrics-C9YY1OcL.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"metrics-C9YY1OcL.js","names":[],"sources":["../src/metrics.ts"],"sourcesContent":["import type { ProductClient } from './client'\nimport type { DriverState, TurnMetrics } from './types'\n\ninterface TokenPrice {\n  input: number\n  output: number\n}\n\n/** Per-1K token pricing for exact model ids. */\nexport const MODEL_PRICING: Record<string, TokenPrice> = {\n  'gpt-4o': { input: 0.0025, output: 0.01 },\n  'gpt-4o-mini': { input: 0.00015, output: 0.0006 },\n  'gpt-4-turbo': { input: 0.01, output: 0.03 },\n  'claude-sonnet-4-20250514': { input: 0.003, output: 0.015 },\n  'claude-opus-4-20250514': { input: 0.015, output: 0.075 },\n  'claude-3-haiku-20240307': { input: 0.00025, output: 0.00125 },\n}\n\n/** Family-level pricing fallbacks (per-1K), matched against a normalized id\n *  after exact lookup misses. Ordered — first match wins. Covers the model\n *  ids actually used through the Tangle router + cli-bridge harnesses\n *  (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,\n *  `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),\n *  none of which appear in the exact table above — without this they priced\n *  to a silent $0, blanking every cost/Pareto axis downstream. */\nconst FAMILY_PRICING: Array<[RegExp, TokenPrice]> = [\n  [/claude.*opus/, { input: 0.015, output: 0.075 }],\n  [/claude.*haiku/, { input: 0.0008, output: 0.004 }],\n  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 0.003, output: 0.015 }],\n  [/gpt-4o-mini/, { input: 0.00015, output: 0.0006 }],\n  [/gpt-5|gpt-4\\.1|o[134]\\b/, { input: 0.00125, output: 0.01 }],\n  [/gpt-4o|gpt-4/, { input: 0.0025, output: 0.01 }],\n  [/deepseek/, { input: 0.0003, output: 0.0011 }],\n  [/glm|zhipu|zai/, { input: 0.0006, output: 0.0022 }],\n  [/kimi|moonshot/, { input: 0.0006, output: 0.0025 }],\n  [/qwen/, { input: 0.0004, output: 0.0012 }],\n  [/gemini.*flash/, { input: 0.0001, output: 0.0004 }],\n  [/gemini/, { input: 0.00125, output: 0.005 }],\n  [/llama/, { input: 0.0002, output: 0.0006 }],\n]\n\n/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,\n *  and keep the final harness/provider-prefixed segment so family regexes\n *  match (`opencode/zai-coding-plan/glm-5.1` → `glm-5.1`). */\nfunction normalizeModelId(model: string): string {\n  return (model.split('@')[0] ?? model).trim().toLowerCase()\n}\n\n/** Resolve pricing for a model id: exact table, then family fallback.\n *  Returns null when the id matches nothing (caller decides — never a\n *  silent-zero masquerading as a real $0 cost). */\nexport function resolveModelPricing(model: string): TokenPrice | null {\n  if (MODEL_PRICING[model]) return MODEL_PRICING[model]\n  const id = normalizeModelId(model)\n  if (MODEL_PRICING[id]) return MODEL_PRICING[id]\n  for (const [pattern, price] of FAMILY_PRICING) {\n    if (pattern.test(id)) return price\n  }\n  return null\n}\n\n/** True when `model` has known pricing (exact or family). Lets cost-aware\n *  callers distinguish a real $0 from an unpriced model. */\nexport function isModelPriced(model: string): boolean {\n  return resolveModelPricing(model) !== null\n}\n\nconst warnedUnpricedModels = new Set<string>()\n\n/** Estimate token count from string length (chars / 4 approximation) */\nexport function estimateTokens(text: string): number {\n  return Math.ceil(text.length / 4)\n}\n\n/** Calculate cost in USD from token counts and model. Unknown models warn\n *  once (not a silent zero) and return 0 so callers that ignore pricing keep\n *  working; cost-sensitive callers should gate on {@link isModelPriced}. */\nexport function estimateCost(inputTokens: number, outputTokens: number, model: string): number {\n  const pricing = resolveModelPricing(model)\n  if (!pricing) {\n    if (!warnedUnpricedModels.has(model)) {\n      warnedUnpricedModels.add(model)\n      console.warn(\n        `estimateCost: no pricing for model \"${model}\" — returning 0; add it to ` +\n          'MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)',\n      )\n    }\n    return 0\n  }\n  return (inputTokens / 1000) * pricing.input + (outputTokens / 1000) * pricing.output\n}\n\n/**\n * TokenCounter — accumulates token usage and cost across turns.\n */\nexport class TokenCounter {\n  private totalInput = 0\n  private totalOutput = 0\n  private totalCost = 0\n  private model: string\n\n  constructor(model = 'gpt-4o') {\n    this.model = model\n  }\n\n  /** Record tokens for a turn, returns per-turn cost */\n  record(inputTokens: number, outputTokens: number): number {\n    this.totalInput += inputTokens\n    this.totalOutput += outputTokens\n    const cost = estimateCost(inputTokens, outputTokens, this.model)\n    this.totalCost += cost\n    return cost\n  }\n\n  /** Estimate and record from raw text */\n  recordFromText(\n    inputText: string,\n    outputText: string,\n  ): { inputTokens: number; outputTokens: number; cost: number } {\n    const inputTokens = estimateTokens(inputText)\n    const outputTokens = estimateTokens(outputText)\n    const cost = this.record(inputTokens, outputTokens)\n    return { inputTokens, outputTokens, cost }\n  }\n\n  getTotalInput(): number {\n    return this.totalInput\n  }\n  getTotalOutput(): number {\n    return this.totalOutput\n  }\n  getTotalCost(): number {\n    return this.totalCost\n  }\n}\n\n/**\n * MetricsCollector — collects per-turn metrics from the product.\n *\n * After each turn, queries the product's APIs to measure state changes.\n */\nexport class MetricsCollector {\n  private client: ProductClient\n  private workspaceId: string\n  private metrics: TurnMetrics[] = []\n  constructor(client: ProductClient, workspaceId: string) {\n    this.client = client\n    this.workspaceId = workspaceId\n  }\n\n  /** Collect metrics after a turn completes */\n  async collect(\n    turn: number,\n    responseLatencyMs: number,\n    responseChars: number,\n    codeBlocksProduced: number,\n    blocksExtracted: number,\n    completionCriteriaMet: number,\n    completionCriteriaTotal: number,\n    qualityScore?: number,\n    inputTokens = 0,\n    outputTokens = 0,\n    estimatedCostUsd = 0,\n  ): Promise<TurnMetrics> {\n    const state = await this.getState()\n\n    const m: TurnMetrics = {\n      turn,\n      timestamp: new Date().toISOString(),\n      tasks: state.tasks,\n      events: state.events,\n      proposals: state.proposals,\n      vaultFiles: state.vaultFiles.length,\n      responseLatencyMs,\n      responseChars,\n      codeBlocksProduced,\n      blocksExtracted,\n      qualityScore,\n      inputTokens,\n      outputTokens,\n      estimatedCostUsd,\n      totalCostUsd: estimatedCostUsd,\n      completionPercent:\n        completionCriteriaTotal > 0 ? (completionCriteriaMet / completionCriteriaTotal) * 100 : 0,\n    }\n\n    this.metrics.push(m)\n    return m\n  }\n\n  /** Get current product state */\n  async getState(): Promise<DriverState> {\n    const [tasks, events, approvals, vaultFiles] = await Promise.all([\n      this.client.getTasks(this.workspaceId),\n      this.client.getEvents(this.workspaceId),\n      this.client.getApprovals(this.workspaceId),\n      this.client.getVaultTree(this.workspaceId),\n    ])\n\n    return {\n      tasks: tasks.length,\n      events: events.length,\n      proposals: {\n        pending: approvals.filter((a) => a.status === 'pending').length,\n        approved: approvals.filter((a) => a.status === 'approved').length,\n        rejected: approvals.filter((a) => a.status === 'rejected').length,\n      },\n      vaultFiles,\n      codeBlocks: 0,\n      generations: 0,\n    }\n  }\n\n  /** Get all collected metrics */\n  getMetrics(): TurnMetrics[] {\n    return [...this.metrics]\n  }\n\n  /** Get convergence curve (completion% over turns) */\n  getConvergenceCurve(): number[] {\n    return this.metrics.map((m) => m.completionPercent)\n  }\n}\n"],"mappings":";;AASA,MAAa,gBAA4C;CACvD,UAAU;EAAE,OAAO;EAAQ,QAAQ;CAAK;CACxC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO;CAChD,eAAe;EAAE,OAAO;EAAM,QAAQ;CAAK;CAC3C,4BAA4B;EAAE,OAAO;EAAO,QAAQ;CAAM;CAC1D,0BAA0B;EAAE,OAAO;EAAO,QAAQ;CAAM;CACxD,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAQ;AAC/D;;;;;;;;AASA,MAAM,iBAA8C;CAClD,CAAC,gBAAgB;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAChD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAM,CAAC;CAClD,CAAC,4CAA4C;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAC5E,CAAC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO,CAAC;CAClD,CAAC,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAK,CAAC;CAC5D,CAAC,gBAAgB;EAAE,OAAO;EAAQ,QAAQ;CAAK,CAAC;CAChD,CAAC,YAAY;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC9C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,QAAQ;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC1C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,UAAU;EAAE,OAAO;EAAS,QAAQ;CAAM,CAAC;CAC5C,CAAC,SAAS;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;AAC7C;;;;AAKA,SAAS,iBAAiB,OAAuB;CAC/C,QAAQ,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,MAAA,CAAO,KAAK,CAAC,CAAC,YAAY;AAC3D;;;;AAKA,SAAgB,oBAAoB,OAAkC;CACpE,IAAI,cAAc,QAAQ,OAAO,cAAc;CAC/C,MAAM,KAAK,iBAAiB,KAAK;CACjC,IAAI,cAAc,KAAK,OAAO,cAAc;CAC5C,KAAK,MAAM,CAAC,SAAS,UAAU,gBAC7B,IAAI,QAAQ,KAAK,EAAE,GAAG,OAAO;CAE/B,OAAO;AACT;;;AAIA,SAAgB,cAAc,OAAwB;CACpD,OAAO,oBAAoB,KAAK,MAAM;AACxC;AAEA,MAAM,uCAAuB,IAAI,IAAY;;AAG7C,SAAgB,eAAe,MAAsB;CACnD,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;AAClC;;;;AAKA,SAAgB,aAAa,aAAqB,cAAsB,OAAuB;CAC7F,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SAAS;EACZ,IAAI,CAAC,qBAAqB,IAAI,KAAK,GAAG;GACpC,qBAAqB,IAAI,KAAK;GAC9B,QAAQ,KACN,uCAAuC,MAAM,oGAE/C;EACF;EACA,OAAO;CACT;CACA,OAAQ,cAAc,MAAQ,QAAQ,QAAS,eAAe,MAAQ,QAAQ;AAChF;;;;AAKA,IAAa,eAAb,MAA0B;CACxB,aAAqB;CACrB,cAAsB;CACtB,YAAoB;CACpB;CAEA,YAAY,QAAQ,UAAU;EAC5B,KAAK,QAAQ;CACf;;CAGA,OAAO,aAAqB,cAA8B;EACxD,KAAK,cAAc;EACnB,KAAK,eAAe;EACpB,MAAM,OAAO,aAAa,aAAa,cAAc,KAAK,KAAK;EAC/D,KAAK,aAAa;EAClB,OAAO;CACT;;CAGA,eACE,WACA,YAC6D;EAC7D,MAAM,cAAc,eAAe,SAAS;EAC5C,MAAM,eAAe,eAAe,UAAU;EAE9C,OAAO;GAAE;GAAa;GAAc,MADvB,KAAK,OAAO,aAAa,YACC;EAAE;CAC3C;CAEA,gBAAwB;EACtB,OAAO,KAAK;CACd;CACA,iBAAyB;EACvB,OAAO,KAAK;CACd;CACA,eAAuB;EACrB,OAAO,KAAK;CACd;AACF;;;;;;AAOA,IAAa,mBAAb,MAA8B;CAC5B;CACA;CACA,UAAiC,CAAC;CAClC,YAAY,QAAuB,aAAqB;EACtD,KAAK,SAAS;EACd,KAAK,cAAc;CACrB;;CAGA,MAAM,QACJ,MACA,mBACA,eACA,oBACA,iBACA,uBACA,yBACA,cACA,cAAc,GACd,eAAe,GACf,mBAAmB,GACG;EACtB,MAAM,QAAQ,MAAM,KAAK,SAAS;EAElC,MAAM,IAAiB;GACrB;GACA,4BAAW,IAAI,KAAK,EAAA,CAAE,YAAY;GAClC,OAAO,MAAM;GACb,QAAQ,MAAM;GACd,WAAW,MAAM;GACjB,YAAY,MAAM,WAAW;GAC7B;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA,cAAc;GACd,mBACE,0BAA0B,IAAK,wBAAwB,0BAA2B,MAAM;EAC5F;EAEA,KAAK,QAAQ,KAAK,CAAC;EACnB,OAAO;CACT;;CAGA,MAAM,WAAiC;EACrC,MAAM,CAAC,OAAO,QAAQ,WAAW,cAAc,MAAM,QAAQ,IAAI;GAC/D,KAAK,OAAO,SAAS,KAAK,WAAW;GACrC,KAAK,OAAO,UAAU,KAAK,WAAW;GACtC,KAAK,OAAO,aAAa,KAAK,WAAW;GACzC,KAAK,OAAO,aAAa,KAAK,WAAW;EAC3C,CAAC;EAED,OAAO;GACL,OAAO,MAAM;GACb,QAAQ,OAAO;GACf,WAAW;IACT,SAAS,UAAU,QAAQ,MAAM,EAAE,WAAW,SAAS,CAAC,CAAC;IACzD,UAAU,UAAU,QAAQ,MAAM,EAAE,WAAW,UAAU,CAAC,CAAC;IAC3D,UAAU,UAAU,QAAQ,MAAM,EAAE,WAAW,UAAU,CAAC,CAAC;GAC7D;GACA;GACA,YAAY;GACZ,aAAa;EACf;CACF;;CAGA,aAA4B;EAC1B,OAAO,CAAC,GAAG,KAAK,OAAO;CACzB;;CAGA,sBAAgC;EAC9B,OAAO,KAAK,QAAQ,KAAK,MAAM,EAAE,iBAAiB;CACpD;AACF"}

package/dist/mint-yN2M2eh0.js

@@ -0,0 +1,201 @@
+import { s as ValidationError } from "./errors-8YnH8WlF.js";
+import { o as runTaskScore } from "./run-record-BuoE80Dq.js";
+import { a as scoreOrigin, i as rolloutRewardFields } from "./reward-nw2xZGZG.js";
+import { i as ROLLOUT_SCHEMA, s as assertMinted } from "./schema-C6DW4ZHR.js";
+import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
+//#region src/rollout/mint.ts
+/**
+* Rollout minting — `tangle.rollout.v1` lines joined from the records the
+* substrate ALREADY keeps. There is no separate rollout store: a rollout
+* is the JOIN of a RunRecord (identity, provenance, cost, outcome) with
+* its trace (spans share `runId`), projected into the canonical line.
+*
+* Composition, not duplication:
+*   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)
+*   - step structure          → `buildTrajectory` over the shared TraceStore
+*   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)
+*   - PRM / reward-model      → `reward-model-export.ts`
+*
+* Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is
+* never exported with a positive reward OR with any of the numbers that reward
+* was computed from. The gate travels into the training data (`reward` forced
+* to 0, `realness_gated: true`) and the whole outcome is transformed by
+* `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and
+* `verdict` to `provenance.gated_evidence`. Mint returns
+* `MintedRolloutLine[]`: the brand the training exporters require, which only
+* this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.
+*
+* A record carrying NEITHER split score is REJECTED (`ValidationError`), never
+* minted at 0 — "nobody graded this" is not the same claim as "graded a total
+* failure", and a trainer reading 0 learns the second. Lines that already
+* carry `reward: null` (interchange imports, existing ledgers) remain valid on
+* the wire; only the RunRecord→line door refuses.
+*
+* Records without spans become labeled GAP LINES (messages: [],
+* provenance.gap) — present in the output AND surfaced in
+* `missingTraces`; a capture gap is a finding, never a silent omission.
+*/
+const asText = (v, scrub) => {
+	return scrub((typeof v === "string" ? v : JSON.stringify(v)) ?? "");
+};
+function projectStep(span, scrub) {
+	const base = {
+		kind: span.kind,
+		name: scrub(span.name),
+		status: span.status,
+		durationMs: span.endedAt !== void 0 ? span.endedAt - span.startedAt : void 0
+	};
+	if (span.kind === "llm") {
+		const llm = span;
+		const last = llm.messages[llm.messages.length - 1];
+		if (last) base.input = scrub(last.content);
+		if (llm.output !== void 0) base.output = scrub(llm.output);
+	} else if (span.kind === "tool") {
+		const tool = span;
+		base.input = asText(tool.args, scrub);
+		if (tool.result !== void 0) base.output = asText(tool.result, scrub);
+	}
+	return base;
+}
+/** The final llm span's history + output is the completed conversation. */
+function finalConversation(spans, scrub) {
+	const llms = spans.filter((s) => s.kind === "llm");
+	const last = llms[llms.length - 1];
+	if (!last) return [];
+	const messages = last.messages.map((m) => ({
+		role: m.role,
+		content: scrub(m.content)
+	}));
+	if (last.output !== void 0 && last.output !== "") messages.push({
+		role: "assistant",
+		content: scrub(last.output)
+	});
+	return messages;
+}
+const REWARD_SOURCE = {
+	holdout: "run-record/holdout-score",
+	search: "run-record/search-score",
+	unscored: "run-record/unscored"
+};
+/**
+* The mint door refuses an execution-only record: a missing training label is
+* not a zero reward, and not a mintable line either. Lines that already carry
+* `reward: null` — interchange imports, existing ledgers — stay valid on the
+* wire and keep their labeled gap; this guard is only about the
+* RunRecord→line door, where the producer can still be told to go score the
+* run instead of shipping an unlabeled row.
+*/
+function requireTaskScore(record) {
+	if (runTaskScore(record) === void 0) throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`);
+}
+const SPLIT_FROM_TAG = {
+	search: "search",
+	dev: "dev",
+	holdout: "holdout"
+};
+function mintLine(record, steps, messages, options, capturedAt, gap) {
+	requireTaskScore(record);
+	const rewardFields = rolloutRewardFields(record);
+	const uncaptured = record.costProvenance.kind === "uncaptured";
+	const terminalOutcome = record.terminalOutcome;
+	const isCompleted = terminalOutcome === "succeeded" || terminalOutcome === "failed";
+	const isTruncated = terminalOutcome === "cancelled" || terminalOutcome === "incomplete";
+	const terminalError = terminalOutcome === "failed" || terminalOutcome === "cancelled" || terminalOutcome === "incomplete" ? record.terminalFailureReason ?? `run ended ${terminalOutcome}` : null;
+	return assertMinted({
+		schema: ROLLOUT_SCHEMA,
+		rollout_id: record.runId,
+		parent_rollout_id: null,
+		run_id: record.runId,
+		experiment_id: record.experimentId,
+		candidate_id: record.candidateId,
+		generation: null,
+		candidate_index: null,
+		role: options.role ?? "agent",
+		task: {
+			suite: options.suite ?? record.experimentId,
+			instance_id: record.scenarioId,
+			split: SPLIT_FROM_TAG[record.splitTag],
+			seed: record.seed,
+			rep: 0
+		},
+		policy: {
+			harness: null,
+			harness_version: null,
+			model: record.model,
+			provider: null,
+			profile_commit: record.commitSha,
+			prompt_hash: record.promptHash,
+			config_hash: record.configHash,
+			agent_profile_cell_id: record.agentProfile?.cellId ?? null,
+			sampling: null
+		},
+		messages,
+		tool_defs: [],
+		...steps.length > 0 ? { steps } : {},
+		outcome: {
+			...rewardFields,
+			reward_source: REWARD_SOURCE[scoreOrigin(record)],
+			verdict: null,
+			metrics: { ...record.outcome.raw },
+			is_completed: isCompleted,
+			is_truncated: isTruncated,
+			error: terminalError
+		},
+		cost: {
+			usd: uncaptured ? null : record.costUsd,
+			tokens_in: record.tokenUsage.input,
+			tokens_out: record.tokenUsage.output,
+			tokens_reasoning: record.tokenUsage.reasoning ?? null,
+			cache_read: record.tokenUsage.cached ?? null,
+			cache_write: record.tokenUsage.cacheWrite ?? null,
+			wall_s: Math.round(record.wallMs / 1e3)
+		},
+		artifacts: {
+			patch_path: null,
+			run_dir: null,
+			transcript_ref: null
+		},
+		provenance: {
+			captured_at: capturedAt,
+			capture: "mint",
+			...gap !== void 0 ? { gap } : {}
+		}
+	}, `minted rollout line for run ${record.runId}`);
+}
+/**
+* Join RunRecords with their traces into canonical rollout lines. Records
+* without spans are emitted as labeled gap lines and reported in
+* `missingTraces`. Execution-only records without a task score are rejected
+* because a missing training label is not a zero reward.
+*/
+async function mintRolloutRows(records, store, options = {}) {
+	const scrub = options.scrub ?? ((t) => t);
+	const capturedAt = (options.now?.() ?? /* @__PURE__ */ new Date()).toISOString();
+	const rows = [];
+	const missingTraces = [];
+	for (const record of records) {
+		const trajectory = await buildTrajectory(store, record.runId);
+		if (trajectory.steps.length === 0) {
+			missingTraces.push(record.runId);
+			rows.push(mintLine(record, [], [], options, capturedAt, "no trace spans recorded for this runId"));
+			continue;
+		}
+		let steps = trajectory.steps.map((s) => projectStep(s.span, scrub));
+		if (options.maxSteps !== void 0 && steps.length > options.maxSteps) {
+			const head = Math.ceil(options.maxSteps / 2);
+			const tail = options.maxSteps - head;
+			steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)];
+		}
+		const conversation = finalConversation(trajectory.steps.map((s) => s.span), scrub);
+		const gap = conversation.length === 0 ? "trace has no llm spans — no conversation to inline" : void 0;
+		rows.push(mintLine(record, steps, conversation, options, capturedAt, gap));
+	}
+	return {
+		rows,
+		missingTraces
+	};
+}
+//#endregion
+export { mintRolloutRows as t };
+
+//# sourceMappingURL=mint-yN2M2eh0.js.map