@tangle-network/agent-eval 0.222.0 → 0.223.0

package/dist/judge-gate-C2BCOt-M.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"judge-gate-C2BCOt-M.js","names":["parseIso"],"sources":["../src/series-convergence.ts","../src/judge-retry.ts","../src/meta-eval/sentinel.ts","../src/meta-eval/judge-gate.ts"],"sourcesContent":["/**\n * Series convergence — detects whether a sequence of scalar measurements\n * is stabilizing, drifting, or noisy. It reads drift *across* runs, e.g.\n * \"are my nightly eval scores stabilizing?\".\n *\n * Three signals:\n *   - stabilized: last K values have low variance (< epsilon) — done\n *   - drifting:   recent trend is monotonic and beyond noise — regressing or improving\n *   - noisy:      neither — keep iterating, but flag as untrustworthy for gating\n */\n\nexport interface SeriesConvergenceOptions {\n  /** Window size for \"recent\" analysis (default 5). */\n  window?: number\n  /** Coefficient-of-variation threshold below which the window is stabilized (default 0.05 = 5%). */\n  stableCv?: number\n  /** Minimum monotone run length to call drift (default 3). */\n  driftRun?: number\n}\n\nexport interface SeriesConvergenceResult {\n  state: 'stabilized' | 'drifting-up' | 'drifting-down' | 'noisy' | 'insufficient-data'\n  windowMean: number\n  windowCv: number\n  /** Longest monotonic run at the tail of the series (positive for up, negative for down). */\n  tailRun: number\n  /** True when n ≥ window AND windowCv ≤ stableCv. */\n  stable: boolean\n}\n\nexport function analyzeSeries(\n  values: number[],\n  options: SeriesConvergenceOptions = {},\n): SeriesConvergenceResult {\n  const window = options.window ?? 5\n  const stableCv = options.stableCv ?? 0.05\n  const driftRun = options.driftRun ?? 3\n\n  if (values.length < Math.max(2, Math.min(window, 3))) {\n    return { state: 'insufficient-data', windowMean: 0, windowCv: 0, tailRun: 0, stable: false }\n  }\n\n  const tail = values.slice(-window)\n  const mean = tail.reduce((a, b) => a + b, 0) / tail.length\n  const variance = tail.reduce((acc, v) => acc + (v - mean) ** 2, 0) / tail.length\n  const stdDev = Math.sqrt(variance)\n  const refMean = Math.abs(mean) > 1e-9 ? Math.abs(mean) : 1\n  const cv = stdDev / refMean\n  const stable = tail.length >= window && cv <= stableCv\n\n  // Tail monotonic run: count how many consecutive strictly-increasing (or decreasing)\n  // steps end at the final value.\n  let tailRun = 0\n  let direction: 1 | -1 | 0 = 0\n  for (let i = values.length - 1; i > 0; i--) {\n    const delta = values[i]! - values[i - 1]!\n    if (delta === 0) break\n    const dir = delta > 0 ? 1 : -1\n    if (direction === 0) direction = dir\n    if (dir !== direction) break\n    tailRun += dir\n  }\n\n  let state: SeriesConvergenceResult['state']\n  if (stable) {\n    state = 'stabilized'\n  } else if (Math.abs(tailRun) >= driftRun) {\n    state = tailRun > 0 ? 'drifting-up' : 'drifting-down'\n  } else {\n    state = 'noisy'\n  }\n\n  return { state, windowMean: mean, windowCv: cv, tailRun, stable }\n}\n","/**\n * Wrap a single judge LLM call with retry, optional fallback-model\n * rotation, exponential backoff, and a typed `JudgeRetryOutcome`. Callers\n * MUST inspect `succeeded` before using `value`; on failure the library\n * returns `value: null` rather than substituting a default, so a judge\n * abort cannot silently corrupt a downstream composite.\n *\n * Reporting contract: callers ship `TrialResult.judgeSucceeded = succeeded`\n * and `TrialResult.judgeAttempts = attempts` so `aggregateTrialsByMode`\n * with `mode: 'exclude-failed'` drops the trial.\n */\n\nimport { backoffMs, isTransientLlmError } from './llm-client'\n\n/** Retry policy for judge LLM calls. */\nexport interface JudgeRetryPolicy {\n  /** Max attempts per model. Default 3 (one initial + two retries). */\n  maxAttempts?: number\n  /** Per-attempt timeout in ms. Default 300_000. */\n  timeoutMs?: number\n  /**\n   * Models to try, in order. The first model is the primary; subsequent\n   * models are fallbacks invoked only when ALL retries on the previous\n   * model have been exhausted. Example: `['claude-code/sonnet', 'kimi-code/k2p6']`\n   * runs claude-code up to maxAttempts times, then falls back to kimi.\n   * If omitted, the caller's judge function controls model selection and\n   * the retries apply to that single model.\n   */\n  models?: readonly string[]\n  /** Exponential backoff function, default `attempt → min(500 * 2^attempt, 16_000)`. */\n  backoffMs?: (attempt: number) => number\n  /**\n   * Predicate deciding whether an error should trigger a retry. Defaults to\n   * `isTransientLlmError` — the package-wide classifier shared with\n   * `callLlm` — which retries aborts/timeouts, network faults, HTTP/2\n   * transport faults, and any `LlmCallError` with status in {429,502,503,504}.\n   * JSON-parse and schema-rejection errors are NOT retriable (the model\n   * needs prompt adjustment, not another shot).\n   */\n  isRetryable?: (err: unknown) => boolean\n}\n\n/** Outcome of a wrapped judge invocation. */\nexport interface JudgeRetryOutcome<T> {\n  /** The judge's returned value when `succeeded === true`. */\n  value: T | null\n  /** True iff one of the attempts completed without throwing. */\n  succeeded: boolean\n  /** Total attempts made across all models. */\n  attempts: number\n  /** Which model the successful attempt used (when succeeded). */\n  modelUsed?: string\n  /** Last error captured when `succeeded === false`. */\n  error?: Error\n  /** Per-attempt error log for forensics. */\n  attemptErrors: Array<{ attempt: number; model: string; error: string }>\n}\n\nconst DEFAULT_MAX_ATTEMPTS = 3\nconst DEFAULT_TIMEOUT_MS = 300_000\n\nfunction sleep(ms: number): Promise<void> {\n  return new Promise((resolve) => setTimeout(resolve, ms))\n}\n\n/**\n * Wrap a judge call with retry + fallback-model + typed outcome semantics.\n *\n * The `judgeFn` signature is `(model: string, signal: AbortSignal) => Promise<T>`.\n * The signal will be aborted at `timeoutMs`. Callers should pass the signal\n * to their underlying fetch/SDK call so the abort actually fires.\n *\n * Returns a typed outcome — callers MUST inspect `succeeded` before using\n * `value`. The library refuses to default to a silent zero score because a\n * synthetic zero is indistinguishable from a real low score downstream.\n */\nexport async function withJudgeRetry<T>(\n  judgeFn: (model: string, signal: AbortSignal) => Promise<T>,\n  policy: JudgeRetryPolicy = {},\n): Promise<JudgeRetryOutcome<T>> {\n  const maxAttempts = policy.maxAttempts ?? DEFAULT_MAX_ATTEMPTS\n  const timeoutMs = policy.timeoutMs ?? DEFAULT_TIMEOUT_MS\n  const backoff = policy.backoffMs ?? backoffMs\n  const isRetryable = policy.isRetryable ?? isTransientLlmError\n  const models =\n    policy.models && policy.models.length > 0 ? policy.models : [undefined as unknown as string]\n\n  let totalAttempts = 0\n  const attemptErrors: JudgeRetryOutcome<T>['attemptErrors'] = []\n  let lastError: Error | undefined\n\n  for (const model of models) {\n    for (let attempt = 0; attempt < maxAttempts; attempt++) {\n      totalAttempts += 1\n      const controller = new AbortController()\n      const timer = setTimeout(() => controller.abort(new Error('TimeoutError')), timeoutMs)\n      try {\n        const value = await judgeFn(model as string, controller.signal)\n        clearTimeout(timer)\n        return {\n          value,\n          succeeded: true,\n          attempts: totalAttempts,\n          modelUsed: model,\n          attemptErrors,\n        }\n      } catch (err) {\n        clearTimeout(timer)\n        const errObj = err instanceof Error ? err : new Error(String(err))\n        lastError = errObj\n        attemptErrors.push({\n          attempt: totalAttempts,\n          model: model ?? '(default)',\n          error: errObj.message,\n        })\n        if (!isRetryable(errObj)) {\n          // Non-retriable (e.g., JSON parse, schema rejection): break out of\n          // attempts on this model AND skip the fallback rotation — a\n          // permanent error here will be permanent on the fallback too.\n          return {\n            value: null,\n            succeeded: false,\n            attempts: totalAttempts,\n            error: errObj,\n            attemptErrors,\n          }\n        }\n        if (attempt < maxAttempts - 1) {\n          await sleep(backoff(attempt))\n        }\n      }\n    }\n    // Exhausted attempts on this model — fall through to next model in the\n    // rotation (if any). The backoff between MODEL rotations is the same\n    // as the last per-attempt backoff (already slept above).\n  }\n\n  return {\n    value: null,\n    succeeded: false,\n    attempts: totalAttempts,\n    error: lastError,\n    attemptErrors,\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\nconst 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\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","/**\n * Judge calibration gate: a judge may gate only while its measured agreement\n * with its owner's verdicts holds.\n *\n * A judge (a model judge or a deterministic grader) registers a calibration\n * set: items its owner has decided, each verdict with its provenance. A\n * calibration run replays the judge on those items, once or several times, and\n * `measureJudgeAgreement` reports how often it reached the owner's verdict\n * (accuracy, and Cohen's κ for the pass/fail pair), how much its own verdict\n * moves between runs and between samples, and where it disagreed.\n *\n * `judgeGateDecision` turns that report into the one decision a consumer acts\n * on. A judge either decides (`decides: 'verdict'`, the default: it passes and\n * blocks, so it must reach the owner's verdict) or only vetoes\n * (`decides: 'veto'`: a necessary-condition check such as a copy or policy\n * check, which blocks but never passes work on its own; owners also reject\n * for reasons it does not cover, so its error is blocking what they approved).\n * It refuses when the judge has no measured agreement, too few decided\n * items of either verdict, a report measured on a different model or rubric\n * than the one the judge now runs, a report older than the policy allows,\n * agreement below the policy floors, or an alarmed drift history. A refused\n * judge's verdict is advisory: the consumer reports it but does not let it\n * pass or block anything.\n *\n * Drift is the judge sentinel (`./sentinel`): `snapshotFromJudgeAgreement`\n * records accuracy as `sentinelPassRate` and κ as `calibrationKappa`, so each\n * calibration run appends one snapshot to a `SentinelStore` and the gate reads\n * the alarms `judgeSentinelReport` raises over that history.\n *\n * Every timestamp is caller-supplied ISO-8601, so decisions are reproducible.\n */\n\nimport { JudgeError, ValidationError } from '../errors'\nimport { type ComputedInterval, computeInterval } from '../experiment/ast/interval'\nimport {\n  judgeSentinelReport,\n  type SentinelSnapshot,\n  type SentinelThresholds,\n  validateSentinelSnapshot,\n} from './sentinel'\n\nexport type OwnerVerdict = 'pass' | 'fail'\n\n/** One item the owner decided, with where that decision is recorded. */\nexport interface CalibrationExample {\n  /** Stable id of the decided item; a judge run names it in `exampleId`. */\n  id: string\n  verdict: OwnerVerdict\n  /** The person whose verdict this is. Never a model. */\n  labeledBy: string\n  /** ISO-8601 time the verdict was given. */\n  labeledAt: string\n  /** Where the verdict is recorded, e.g. `gtm:asset_decision:<id>`. */\n  source: string\n  /** What the owner judged (an image digest, a run id), so a replay judges the same thing. */\n  evidenceRef?: string\n  note?: string\n}\n\nexport interface JudgeCalibrationSet {\n  judgeId: string\n  /** Whose verdicts the judge stands in for. */\n  owner: string\n  examples: readonly CalibrationExample[]\n}\n\nfunction parseIso(value: unknown, label: string): number {\n  const ms = typeof value === 'string' ? Date.parse(value) : Number.NaN\n  if (!Number.isFinite(ms)) {\n    throw new ValidationError(\n      `judge-gate: ${label} is not a parseable ISO timestamp: ${JSON.stringify(value)}`,\n    )\n  }\n  return ms\n}\n\nfunction requireText(value: unknown, label: string): string {\n  if (typeof value !== 'string' || value.trim().length === 0) {\n    throw new ValidationError(`judge-gate: ${label} must be a non-empty string`)\n  }\n  return value\n}\n\nfunction requireVerdict(value: unknown, label: string): OwnerVerdict {\n  if (value !== 'pass' && value !== 'fail') {\n    throw new ValidationError(\n      `judge-gate: ${label} must be 'pass' or 'fail', got ${JSON.stringify(value)}`,\n    )\n  }\n  return value\n}\n\nfunction checkExample(example: CalibrationExample, label: string): CalibrationExample {\n  if (example === null || typeof example !== 'object') {\n    throw new ValidationError(`judge-gate: ${label} must be an object`)\n  }\n  requireText(example.id, `${label}.id`)\n  requireVerdict(example.verdict, `${label}.verdict`)\n  requireText(example.labeledBy, `${label}.labeledBy`)\n  parseIso(example.labeledAt, `${label}.labeledAt`)\n  requireText(example.source, `${label}.source`)\n  if (example.evidenceRef !== undefined) requireText(example.evidenceRef, `${label}.evidenceRef`)\n  const copy: CalibrationExample = {\n    id: example.id,\n    verdict: example.verdict,\n    labeledBy: example.labeledBy,\n    labeledAt: example.labeledAt,\n    source: example.source,\n  }\n  if (example.evidenceRef !== undefined) copy.evidenceRef = example.evidenceRef\n  if (example.note !== undefined) copy.note = example.note\n  return copy\n}\n\n/**\n * Validate a calibration set and return a normalized copy ordered by example\n * id. Every verdict needs its person, time and source; ids are unique.\n */\nexport function registerCalibrationSet(set: JudgeCalibrationSet): JudgeCalibrationSet {\n  if (set === null || typeof set !== 'object') {\n    throw new ValidationError('judge-gate: calibration set must be an object')\n  }\n  const judgeId = requireText(set.judgeId, 'calibration set judgeId')\n  const owner = requireText(set.owner, `calibration set \"${judgeId}\" owner`)\n  if (!Array.isArray(set.examples)) {\n    throw new ValidationError(`judge-gate: calibration set \"${judgeId}\" examples must be an array`)\n  }\n  const seen = new Set<string>()\n  const examples = set.examples.map((example, index) => {\n    const checked = checkExample(example, `calibration set \"${judgeId}\" example ${index}`)\n    if (seen.has(checked.id)) {\n      throw new ValidationError(\n        `judge-gate: calibration set \"${judgeId}\" has duplicate example id \"${checked.id}\"`,\n      )\n    }\n    seen.add(checked.id)\n    return checked\n  })\n  examples.sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0))\n  return { judgeId, owner, examples }\n}\n\nexport interface CalibrationMerge {\n  set: JudgeCalibrationSet\n  /** Ids that joined the set. */\n  added: string[]\n  /** Ids whose verdict a later decision replaced. */\n  superseded: string[]\n  /** Ids whose incoming verdict was already recorded, or older than the recorded one. */\n  unchanged: string[]\n}\n\n/**\n * Join new owner verdicts to a set. A verdict already recorded from the same\n * source is unchanged, so ingestion can rerun. A later verdict on an item\n * replaces an earlier one: the owner changed their mind, and the set keeps\n * what they decided last.\n */\nexport function addCalibrationVerdicts(\n  set: JudgeCalibrationSet,\n  verdicts: readonly CalibrationExample[],\n): CalibrationMerge {\n  const base = registerCalibrationSet(set)\n  const byId = new Map(base.examples.map((example) => [example.id, example]))\n  const added: string[] = []\n  const superseded: string[] = []\n  const unchanged: string[] = []\n  for (const [index, verdict] of verdicts.entries()) {\n    const incoming = checkExample(verdict, `incoming verdict ${index} for \"${base.judgeId}\"`)\n    const current = byId.get(incoming.id)\n    if (!current) {\n      byId.set(incoming.id, incoming)\n      added.push(incoming.id)\n      continue\n    }\n    const sameDecision = current.source === incoming.source && current.verdict === incoming.verdict\n    if (\n      sameDecision ||\n      parseIso(incoming.labeledAt, 'labeledAt') <= parseIso(current.labeledAt, 'labeledAt')\n    ) {\n      unchanged.push(incoming.id)\n      continue\n    }\n    byId.set(incoming.id, incoming)\n    superseded.push(incoming.id)\n  }\n  return {\n    set: registerCalibrationSet({ ...base, examples: [...byId.values()] }),\n    added,\n    superseded,\n    unchanged,\n  }\n}\n\n/** One gate decision the judge made on a calibration example. */\nexport interface JudgeRun {\n  exampleId: string\n  /** The judge's decision after its own sample aggregation (median, majority). */\n  verdict: OwnerVerdict\n  /** The per-sample verdicts behind that decision, when the judge samples. */\n  samples?: readonly OwnerVerdict[]\n}\n\nexport interface JudgeRunMeta {\n  judgeId: string\n  /** Exact model identity, or `deterministic` for a grader without one. */\n  judgeModel: string\n  rubricVersion: string\n  measuredAt: string\n}\n\nexport interface JudgeConfusion {\n  /** Judge pass, owner pass. */\n  truePass: number\n  /** Judge fail, owner fail. */\n  trueFail: number\n  /** Judge pass, owner fail: the judge let through what the owner rejected. */\n  falsePass: number\n  /** Judge fail, owner pass: the judge blocked what the owner accepted. */\n  falseFail: number\n}\n\nexport interface JudgeExampleAgreement {\n  id: string\n  verdict: OwnerVerdict\n  runs: number\n  /** Share of runs that passed. */\n  passRate: number\n  /** Share of runs that reached the owner's verdict. */\n  agreement: number\n}\n\nexport interface JudgeAgreement extends JudgeRunMeta {\n  owner: string\n  /** Examples in the calibration set. */\n  examples: number\n  /** Examples the judge ran on at least once. */\n  judged: number\n  /** Judged examples per owner verdict. */\n  judgedByVerdict: Record<OwnerVerdict, number>\n  runs: number\n  /** Runs that reached the owner's verdict, over all runs; null with no runs. */\n  accuracy: number | null\n  /**\n   * Exact (Clopper-Pearson) 95% interval for the share of judged examples on\n   * which every run reached the owner's verdict. Examples, not runs, are the\n   * independent units.\n   */\n  exampleAccuracy: { agreed: number; judged: number; interval: ComputedInterval } | null\n  /**\n   * Cohen's κ between the judge and the owner over runs. Null when undefined:\n   * no runs, or both sides used a single verdict.\n   */\n  kappa: number | null\n  /** Runs that failed what the owner passed, over runs on owner passes; null with none. A veto judge is held to this. */\n  falseFailRate: number | null\n  /** Runs that failed what the owner failed, over runs on owner fails; null with none. */\n  catchRate: number | null\n  confusion: JudgeConfusion\n  variance: {\n    /** Mean over judged examples of p(1 − p), p the example's run pass rate. 0 = no run changed its mind. */\n    betweenRuns: number | null\n    /** Mean over runs that reported samples of p(1 − p) across those samples; null when none did. */\n    withinRun: number | null\n    /** Examples whose runs disagree with each other. */\n    unstable: string[]\n  }\n  /** Examples with at least one run against the owner, least agreement first. */\n  disagreements: JudgeExampleAgreement[]\n  /** Examples without any run. */\n  unjudged: string[]\n}\n\nfunction bernoulliVariance(passes: number, n: number): number {\n  const p = passes / n\n  return p * (1 - p)\n}\n\n/** Cohen's κ for two raters over a binary pass/fail table. */\nfunction binaryKappa(confusion: JudgeConfusion): number | null {\n  const n = confusion.truePass + confusion.trueFail + confusion.falsePass + confusion.falseFail\n  if (n === 0) return null\n  const observed = (confusion.truePass + confusion.trueFail) / n\n  const judgePass = (confusion.truePass + confusion.falsePass) / n\n  const ownerPass = (confusion.truePass + confusion.falseFail) / n\n  const expected = judgePass * ownerPass + (1 - judgePass) * (1 - ownerPass)\n  if (expected === 1) return null\n  return (observed - expected) / (1 - expected)\n}\n\n/**\n * Agreement between a judge's runs and the owner's verdicts on one\n * calibration set. A run that names an example outside the set is an error:\n * a silently dropped run would overstate agreement.\n */\nexport function measureJudgeAgreement(\n  set: JudgeCalibrationSet,\n  runs: readonly JudgeRun[],\n  meta: JudgeRunMeta,\n): JudgeAgreement {\n  const registered = registerCalibrationSet(set)\n  if (meta.judgeId !== registered.judgeId) {\n    throw new ValidationError(\n      `judge-gate: runs are for judge \"${meta.judgeId}\" but the set belongs to \"${registered.judgeId}\"`,\n    )\n  }\n  requireText(meta.judgeModel, 'judgeModel')\n  requireText(meta.rubricVersion, 'rubricVersion')\n  parseIso(meta.measuredAt, 'measuredAt')\n  const byId = new Map(registered.examples.map((example) => [example.id, example]))\n  const runsById = new Map<string, JudgeRun[]>()\n  const confusion: JudgeConfusion = { truePass: 0, trueFail: 0, falsePass: 0, falseFail: 0 }\n  const withinRun: number[] = []\n  for (const [index, run] of runs.entries()) {\n    const example = byId.get(run?.exampleId)\n    if (!example) {\n      throw new ValidationError(\n        `judge-gate: run ${index} names example ${JSON.stringify(run?.exampleId)}, which is not in \"${registered.judgeId}\"'s calibration set`,\n      )\n    }\n    const verdict = requireVerdict(run.verdict, `run ${index} verdict`)\n    if (run.samples !== undefined) {\n      if (!Array.isArray(run.samples) || run.samples.length === 0) {\n        throw new ValidationError(`judge-gate: run ${index} samples must be a non-empty array`)\n      }\n      const samples = run.samples.map((sample, s) =>\n        requireVerdict(sample, `run ${index} sample ${s}`),\n      )\n      withinRun.push(\n        bernoulliVariance(samples.filter((sample) => sample === 'pass').length, samples.length),\n      )\n    }\n    const list = runsById.get(example.id) ?? []\n    list.push({ ...run, verdict })\n    runsById.set(example.id, list)\n    if (verdict === 'pass') {\n      if (example.verdict === 'pass') confusion.truePass++\n      else confusion.falsePass++\n    } else if (example.verdict === 'fail') confusion.trueFail++\n    else confusion.falseFail++\n  }\n\n  const perExample: JudgeExampleAgreement[] = []\n  const unjudged: string[] = []\n  const judgedByVerdict: Record<OwnerVerdict, number> = { pass: 0, fail: 0 }\n  for (const example of registered.examples) {\n    const list = runsById.get(example.id)\n    if (!list) {\n      unjudged.push(example.id)\n      continue\n    }\n    judgedByVerdict[example.verdict]++\n    const passes = list.filter((run) => run.verdict === 'pass').length\n    perExample.push({\n      id: example.id,\n      verdict: example.verdict,\n      runs: list.length,\n      passRate: passes / list.length,\n      agreement: list.filter((run) => run.verdict === example.verdict).length / list.length,\n    })\n  }\n\n  const total = runs.length\n  const agreedRuns = confusion.truePass + confusion.trueFail\n  const agreedExamples = perExample.filter((example) => example.agreement === 1).length\n  const mean = (values: readonly number[]) =>\n    values.length === 0 ? null : values.reduce((sum, value) => sum + value, 0) / values.length\n  return {\n    ...meta,\n    owner: registered.owner,\n    examples: registered.examples.length,\n    judged: perExample.length,\n    judgedByVerdict,\n    runs: total,\n    accuracy: total === 0 ? null : agreedRuns / total,\n    exampleAccuracy:\n      perExample.length === 0\n        ? null\n        : {\n            agreed: agreedExamples,\n            judged: perExample.length,\n            interval: computeInterval(\n              { kind: 'clopper-pearson', level: 0.95 },\n              { kind: 'binomial', successes: agreedExamples, trials: perExample.length },\n            ),\n          },\n    kappa: binaryKappa(confusion),\n    falseFailRate:\n      confusion.truePass + confusion.falseFail === 0\n        ? null\n        : confusion.falseFail / (confusion.truePass + confusion.falseFail),\n    catchRate:\n      confusion.trueFail + confusion.falsePass === 0\n        ? null\n        : confusion.trueFail / (confusion.trueFail + confusion.falsePass),\n    confusion,\n    variance: {\n      betweenRuns: mean(perExample.map((example) => example.passRate * (1 - example.passRate))),\n      withinRun: mean(withinRun),\n      unstable: perExample\n        .filter((example) => example.passRate > 0 && example.passRate < 1)\n        .map((example) => example.id),\n    },\n    disagreements: perExample\n      .filter((example) => example.agreement < 1)\n      .sort((a, b) => a.agreement - b.agreement || (a.id < b.id ? -1 : 1)),\n    unjudged,\n  }\n}\n\nexport interface JudgeGatePolicy {\n  /**\n   * `verdict`: the judge passes and blocks, so accuracy and κ must hold.\n   * `veto`: it only blocks, so its false-fail rate on owner passes must hold.\n   */\n  decides: 'verdict' | 'veto'\n  /** Judged examples required before the judge may gate; for a veto judge, owner passes. */\n  minExamples: number\n  /** Judged examples required of each owner verdict, so κ is defined. */\n  minPerVerdict: number\n  /** Floor on run accuracy. */\n  minAccuracy: number\n  /** Floor on Cohen's κ. */\n  minKappa: number\n  /** Ceiling on a veto judge's false-fail rate. */\n  maxFalseFailRate: number\n  /** A report older than this (days, vs `asOf`) no longer calibrates the judge. */\n  maxAgeDays: number\n  /** Drift thresholds forwarded to `judgeSentinelReport` over the history. */\n  drift?: SentinelThresholds\n}\n\nexport const DEFAULT_JUDGE_GATE_POLICY: Readonly<JudgeGatePolicy> = Object.freeze({\n  decides: 'verdict',\n  minExamples: 10,\n  minPerVerdict: 1,\n  minAccuracy: 0.9,\n  minKappa: 0.6,\n  maxFalseFailRate: 0.1,\n  maxAgeDays: 30,\n})\n\nexport type JudgeGateStatus = 'calibrated' | 'uncalibrated' | 'below-threshold' | 'drifted'\n\nexport interface JudgeGateDecision {\n  judgeId: string\n  judgeModel: string\n  rubricVersion: string\n  /** True only for `calibrated`. */\n  mayGate: boolean\n  status: JudgeGateStatus\n  /** Why the judge may not gate; empty when it may. */\n  reasons: string[]\n  agreement: JudgeAgreement | null\n  policy: JudgeGatePolicy\n}\n\nexport interface JudgeGateInput {\n  judgeId: string\n  /** The model the judge runs now. */\n  judgeModel: string\n  /** The rubric the judge runs now. */\n  rubricVersion: string\n  /** The latest calibration report; null when the judge was never measured. */\n  agreement: JudgeAgreement | null\n  asOf: string\n  /** Sentinel snapshots for this judge, oldest first or in any order. */\n  history?: readonly SentinelSnapshot[]\n  policy?: Partial<JudgeGatePolicy>\n}\n\nconst DAY_MS = 86_400_000\n\nfunction percent(value: number): string {\n  return `${(value * 100).toFixed(1)}%`\n}\n\n/** Whether a judge may gate now, and every reason it may not. */\nexport function judgeGateDecision(input: JudgeGateInput): JudgeGateDecision {\n  const judgeId = requireText(input.judgeId, 'judgeId')\n  const judgeModel = requireText(input.judgeModel, 'judgeModel')\n  const rubricVersion = requireText(input.rubricVersion, 'rubricVersion')\n  const asOf = parseIso(input.asOf, 'asOf')\n  const policy: JudgeGatePolicy = {\n    ...DEFAULT_JUDGE_GATE_POLICY,\n    ...Object.fromEntries(\n      Object.entries(input.policy ?? {}).filter(([, value]) => value !== undefined),\n    ),\n  }\n  if (policy.decides !== 'verdict' && policy.decides !== 'veto') {\n    throw new ValidationError(\n      `judge-gate: policy.decides must be 'verdict' or 'veto', got ${JSON.stringify(policy.decides)}`,\n    )\n  }\n  const agreement = input.agreement\n  const decide = (status: JudgeGateStatus, reasons: string[]): JudgeGateDecision => ({\n    judgeId,\n    judgeModel,\n    rubricVersion,\n    mayGate: status === 'calibrated',\n    status,\n    reasons,\n    agreement,\n    policy,\n  })\n\n  if (!agreement) return decide('uncalibrated', ['no measured agreement with owner verdicts'])\n  if (agreement.judgeId !== judgeId) {\n    throw new ValidationError(\n      `judge-gate: agreement belongs to \"${agreement.judgeId}\", not \"${judgeId}\"`,\n    )\n  }\n\n  const uncalibrated: string[] = []\n  if (agreement.judgeModel !== judgeModel) {\n    uncalibrated.push(\n      `agreement was measured on model ${agreement.judgeModel}; the judge now runs ${judgeModel}`,\n    )\n  }\n  if (agreement.rubricVersion !== rubricVersion) {\n    uncalibrated.push(\n      `agreement was measured on rubric ${agreement.rubricVersion}; the judge now runs ${rubricVersion}`,\n    )\n  }\n  const ageDays = (asOf - parseIso(agreement.measuredAt, 'agreement.measuredAt')) / DAY_MS\n  if (ageDays > policy.maxAgeDays) {\n    uncalibrated.push(`agreement is ${ageDays.toFixed(1)} days old (limit ${policy.maxAgeDays})`)\n  }\n  if (policy.decides === 'veto') {\n    // Only owner passes can show a veto judge blocking what it should not.\n    if (agreement.judgedByVerdict.pass < policy.minExamples) {\n      uncalibrated.push(\n        `${agreement.judgedByVerdict.pass} judged examples the owner marked pass; a veto judge needs at least ${policy.minExamples}`,\n      )\n    }\n  } else {\n    if (agreement.judged < policy.minExamples) {\n      uncalibrated.push(\n        `${agreement.judged} judged ${agreement.judged === 1 ? 'example' : 'examples'}; at least ${policy.minExamples} required`,\n      )\n    }\n    for (const verdict of ['pass', 'fail'] as const) {\n      if (agreement.judgedByVerdict[verdict] < policy.minPerVerdict) {\n        uncalibrated.push(\n          `${agreement.judgedByVerdict[verdict]} judged examples the owner marked ${verdict}; at least ${policy.minPerVerdict} required`,\n        )\n      }\n    }\n  }\n  if (uncalibrated.length > 0) return decide('uncalibrated', uncalibrated)\n\n  const below: string[] = []\n  if (policy.decides === 'veto') {\n    if (agreement.falseFailRate === null || agreement.falseFailRate > policy.maxFalseFailRate) {\n      below.push(\n        `it fails ${agreement.falseFailRate === null ? 'an unmeasured share' : percent(agreement.falseFailRate)} of what the owner passed (limit ${percent(policy.maxFalseFailRate)})`,\n      )\n    }\n  } else if (agreement.accuracy === null || agreement.accuracy < policy.minAccuracy) {\n    below.push(\n      `accuracy ${agreement.accuracy === null ? 'unmeasured' : percent(agreement.accuracy)} is below ${percent(policy.minAccuracy)}`,\n    )\n  }\n  if (\n    policy.decides === 'verdict' &&\n    (agreement.kappa === null || agreement.kappa < policy.minKappa)\n  ) {\n    below.push(\n      `kappa ${agreement.kappa === null ? 'undefined' : agreement.kappa.toFixed(2)} is below ${policy.minKappa}`,\n    )\n  }\n  if (below.length > 0) return decide('below-threshold', below)\n\n  const history = (input.history ?? []).filter((snapshot) => snapshot.judgeId === judgeId)\n  if (history.length > 0) {\n    const report = judgeSentinelReport([...history], {\n      asOf: input.asOf,\n      ...(policy.drift ? { thresholds: policy.drift } : {}),\n    })\n    if (!report.healthy) return decide('drifted', report.alarms)\n  }\n  return decide('calibrated', [])\n}\n\n/** Thrown by `assertJudgeMayGate`; carries the refused decision. */\nexport class JudgeGateRefusedError extends JudgeError {\n  readonly decision: JudgeGateDecision\n\n  constructor(decision: JudgeGateDecision) {\n    super(\n      `judge \"${decision.judgeId}\" may not gate (${decision.status}): ${decision.reasons.join('; ')}`,\n    )\n    this.decision = decision\n  }\n}\n\n/** Throw unless the decision lets the judge gate. */\nexport function assertJudgeMayGate(decision: JudgeGateDecision): void {\n  if (!decision.mayGate) throw new JudgeGateRefusedError(decision)\n}\n\n/**\n * The sentinel snapshot for one calibration run. A deciding judge records\n * accuracy as `sentinelPassRate` and κ as `calibrationKappa` when it is\n * defined; a veto judge records the share of owner passes it let through as\n * `sentinelPassRate`. Append it to the judge's `SentinelStore`; the gate reads\n * drift from that history.\n */\nexport function snapshotFromJudgeAgreement(\n  agreement: JudgeAgreement,\n  options: { decides?: JudgeGatePolicy['decides'] } = {},\n): SentinelSnapshot {\n  const veto = options.decides === 'veto'\n  const rate = veto\n    ? agreement.falseFailRate === null\n      ? null\n      : 1 - agreement.falseFailRate\n    : agreement.accuracy\n  if (rate === null) {\n    throw new ValidationError(\n      `snapshotFromJudgeAgreement: judge \"${agreement.judgeId}\" has no ${veto ? 'runs on owner passes' : 'runs'} to snapshot`,\n    )\n  }\n  const snapshot: SentinelSnapshot = {\n    at: agreement.measuredAt,\n    judgeId: agreement.judgeId,\n    judgeModel: agreement.judgeModel,\n    metrics: {\n      sentinelPassRate: rate,\n      ...(veto || agreement.kappa === null ? {} : { calibrationKappa: agreement.kappa }),\n    },\n  }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n"],"mappings":";;;;AA8BA,SAAgB,cACd,QACA,UAAoC,CAAC,GACZ;CACzB,MAAM,SAAS,QAAQ,UAAU;CACjC,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,WAAW,QAAQ,YAAY;CAErC,IAAI,OAAO,SAAS,KAAK,IAAI,GAAG,KAAK,IAAI,QAAQ,CAAC,CAAC,GACjD,OAAO;EAAE,OAAO;EAAqB,YAAY;EAAG,UAAU;EAAG,SAAS;EAAG,QAAQ;CAAM;CAG7F,MAAM,OAAO,OAAO,MAAM,CAAC,MAAM;CACjC,MAAM,OAAO,KAAK,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK;CACpD,MAAM,WAAW,KAAK,QAAQ,KAAK,MAAM,OAAO,IAAI,SAAS,GAAG,CAAC,IAAI,KAAK;CAG1E,MAAM,KAFS,KAAK,KAAK,QAET,KADA,KAAK,IAAI,IAAI,IAAI,OAAO,KAAK,IAAI,IAAI,IAAI;CAEzD,MAAM,SAAS,KAAK,UAAU,UAAU,MAAM;CAI9C,IAAI,UAAU;CACd,IAAI,YAAwB;CAC5B,KAAK,IAAI,IAAI,OAAO,SAAS,GAAG,IAAI,GAAG,KAAK;EAC1C,MAAM,QAAQ,OAAO,KAAM,OAAO,IAAI;EACtC,IAAI,UAAU,GAAG;EACjB,MAAM,MAAM,QAAQ,IAAI,IAAI;EAC5B,IAAI,cAAc,GAAG,YAAY;EACjC,IAAI,QAAQ,WAAW;EACvB,WAAW;CACb;CAEA,IAAI;CACJ,IAAI,QACF,QAAQ;MACH,IAAI,KAAK,IAAI,OAAO,KAAK,UAC9B,QAAQ,UAAU,IAAI,gBAAgB;MAEtC,QAAQ;CAGV,OAAO;EAAE;EAAO,YAAY;EAAM,UAAU;EAAI;EAAS;CAAO;AAClE;;;;;;;;;;;;;ACfA,MAAM,uBAAuB;AAC7B,MAAM,qBAAqB;AAE3B,SAAS,MAAM,IAA2B;CACxC,OAAO,IAAI,SAAS,YAAY,WAAW,SAAS,EAAE,CAAC;AACzD;;;;;;;;;;;;AAaA,eAAsB,eACpB,SACA,SAA2B,CAAC,GACG;CAC/B,MAAM,cAAc,OAAO,eAAe;CAC1C,MAAM,YAAY,OAAO,aAAa;CACtC,MAAM,UAAU,OAAO,aAAa;CACpC,MAAM,cAAc,OAAO,eAAe;CAC1C,MAAM,SACJ,OAAO,UAAU,OAAO,OAAO,SAAS,IAAI,OAAO,SAAS,CAAC,KAAA,CAA8B;CAE7F,IAAI,gBAAgB;CACpB,MAAM,gBAAuD,CAAC;CAC9D,IAAI;CAEJ,KAAK,MAAM,SAAS,QAClB,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;EACtD,iBAAiB;EACjB,MAAM,aAAa,IAAI,gBAAgB;EACvC,MAAM,QAAQ,iBAAiB,WAAW,sBAAM,IAAI,MAAM,cAAc,CAAC,GAAG,SAAS;EACrF,IAAI;GACF,MAAM,QAAQ,MAAM,QAAQ,OAAiB,WAAW,MAAM;GAC9D,aAAa,KAAK;GAClB,OAAO;IACL;IACA,WAAW;IACX,UAAU;IACV,WAAW;IACX;GACF;EACF,SAAS,KAAK;GACZ,aAAa,KAAK;GAClB,MAAM,SAAS,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC;GACjE,YAAY;GACZ,cAAc,KAAK;IACjB,SAAS;IACT,OAAO,SAAS;IAChB,OAAO,OAAO;GAChB,CAAC;GACD,IAAI,CAAC,YAAY,MAAM,GAIrB,OAAO;IACL,OAAO;IACP,WAAW;IACX,UAAU;IACV,OAAO;IACP;GACF;GAEF,IAAI,UAAU,cAAc,GAC1B,MAAM,MAAM,QAAQ,OAAO,CAAC;EAEhC;CACF;CAMF,OAAO;EACL,OAAO;EACP,WAAW;EACX,UAAU;EACV,OAAO;EACP;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC3FA,MAAM,wBAAwB;CAAC;CAAO;CAAoB;AAAkB;AAmC5E,SAASA,WAAS,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,WAAS,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;;;;;;;AAgBA,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,SAASA,WAAS,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,OAAOA,WAAS,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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC1YA,SAAS,SAAS,OAAgB,OAAuB;CACvD,MAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;CAC3D,IAAI,CAAC,OAAO,SAAS,EAAE,GACrB,MAAM,IAAI,gBACR,eAAe,MAAM,qCAAqC,KAAK,UAAU,KAAK,GAChF;CAEF,OAAO;AACT;AAEA,SAAS,YAAY,OAAgB,OAAuB;CAC1D,IAAI,OAAO,UAAU,YAAY,MAAM,KAAK,CAAC,CAAC,WAAW,GACvD,MAAM,IAAI,gBAAgB,eAAe,MAAM,4BAA4B;CAE7E,OAAO;AACT;AAEA,SAAS,eAAe,OAAgB,OAA6B;CACnE,IAAI,UAAU,UAAU,UAAU,QAChC,MAAM,IAAI,gBACR,eAAe,MAAM,iCAAiC,KAAK,UAAU,KAAK,GAC5E;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAA6B,OAAmC;CACpF,IAAI,YAAY,QAAQ,OAAO,YAAY,UACzC,MAAM,IAAI,gBAAgB,eAAe,MAAM,mBAAmB;CAEpE,YAAY,QAAQ,IAAI,GAAG,MAAM,IAAI;CACrC,eAAe,QAAQ,SAAS,GAAG,MAAM,SAAS;CAClD,YAAY,QAAQ,WAAW,GAAG,MAAM,WAAW;CACnD,SAAS,QAAQ,WAAW,GAAG,MAAM,WAAW;CAChD,YAAY,QAAQ,QAAQ,GAAG,MAAM,QAAQ;CAC7C,IAAI,QAAQ,gBAAgB,KAAA,GAAW,YAAY,QAAQ,aAAa,GAAG,MAAM,aAAa;CAC9F,MAAM,OAA2B;EAC/B,IAAI,QAAQ;EACZ,SAAS,QAAQ;EACjB,WAAW,QAAQ;EACnB,WAAW,QAAQ;EACnB,QAAQ,QAAQ;CAClB;CACA,IAAI,QAAQ,gBAAgB,KAAA,GAAW,KAAK,cAAc,QAAQ;CAClE,IAAI,QAAQ,SAAS,KAAA,GAAW,KAAK,OAAO,QAAQ;CACpD,OAAO;AACT;;;;;AAMA,SAAgB,uBAAuB,KAA+C;CACpF,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,gBAAgB,+CAA+C;CAE3E,MAAM,UAAU,YAAY,IAAI,SAAS,yBAAyB;CAClE,MAAM,QAAQ,YAAY,IAAI,OAAO,oBAAoB,QAAQ,QAAQ;CACzE,IAAI,CAAC,MAAM,QAAQ,IAAI,QAAQ,GAC7B,MAAM,IAAI,gBAAgB,gCAAgC,QAAQ,4BAA4B;CAEhG,MAAM,uBAAO,IAAI,IAAY;CAC7B,MAAM,WAAW,IAAI,SAAS,KAAK,SAAS,UAAU;EACpD,MAAM,UAAU,aAAa,SAAS,oBAAoB,QAAQ,YAAY,OAAO;EACrF,IAAI,KAAK,IAAI,QAAQ,EAAE,GACrB,MAAM,IAAI,gBACR,gCAAgC,QAAQ,8BAA8B,QAAQ,GAAG,EACnF;EAEF,KAAK,IAAI,QAAQ,EAAE;EACnB,OAAO;CACT,CAAC;CACD,SAAS,MAAM,GAAG,MAAO,EAAE,KAAK,EAAE,KAAK,KAAK,EAAE,KAAK,EAAE,KAAK,IAAI,CAAE;CAChE,OAAO;EAAE;EAAS;EAAO;CAAS;AACpC;;;;;;;AAkBA,SAAgB,uBACd,KACA,UACkB;CAClB,MAAM,OAAO,uBAAuB,GAAG;CACvC,MAAM,OAAO,IAAI,IAAI,KAAK,SAAS,KAAK,YAAY,CAAC,QAAQ,IAAI,OAAO,CAAC,CAAC;CAC1E,MAAM,QAAkB,CAAC;CACzB,MAAM,aAAuB,CAAC;CAC9B,MAAM,YAAsB,CAAC;CAC7B,KAAK,MAAM,CAAC,OAAO,YAAY,SAAS,QAAQ,GAAG;EACjD,MAAM,WAAW,aAAa,SAAS,oBAAoB,MAAM,QAAQ,KAAK,QAAQ,EAAE;EACxF,MAAM,UAAU,KAAK,IAAI,SAAS,EAAE;EACpC,IAAI,CAAC,SAAS;GACZ,KAAK,IAAI,SAAS,IAAI,QAAQ;GAC9B,MAAM,KAAK,SAAS,EAAE;GACtB;EACF;EAEA,IADqB,QAAQ,WAAW,SAAS,UAAU,QAAQ,YAAY,SAAS,WAGtF,SAAS,SAAS,WAAW,WAAW,KAAK,SAAS,QAAQ,WAAW,WAAW,GACpF;GACA,UAAU,KAAK,SAAS,EAAE;GAC1B;EACF;EACA,KAAK,IAAI,SAAS,IAAI,QAAQ;EAC9B,WAAW,KAAK,SAAS,EAAE;CAC7B;CACA,OAAO;EACL,KAAK,uBAAuB;GAAE,GAAG;GAAM,UAAU,CAAC,GAAG,KAAK,OAAO,CAAC;EAAE,CAAC;EACrE;EACA;EACA;CACF;AACF;AAiFA,SAAS,kBAAkB,QAAgB,GAAmB;CAC5D,MAAM,IAAI,SAAS;CACnB,OAAO,KAAK,IAAI;AAClB;;AAGA,SAAS,YAAY,WAA0C;CAC7D,MAAM,IAAI,UAAU,WAAW,UAAU,WAAW,UAAU,YAAY,UAAU;CACpF,IAAI,MAAM,GAAG,OAAO;CACpB,MAAM,YAAY,UAAU,WAAW,UAAU,YAAY;CAC7D,MAAM,aAAa,UAAU,WAAW,UAAU,aAAa;CAC/D,MAAM,aAAa,UAAU,WAAW,UAAU,aAAa;CAC/D,MAAM,WAAW,YAAY,aAAa,IAAI,cAAc,IAAI;CAChE,IAAI,aAAa,GAAG,OAAO;CAC3B,QAAQ,WAAW,aAAa,IAAI;AACtC;;;;;;AAOA,SAAgB,sBACd,KACA,MACA,MACgB;CAChB,MAAM,aAAa,uBAAuB,GAAG;CAC7C,IAAI,KAAK,YAAY,WAAW,SAC9B,MAAM,IAAI,gBACR,mCAAmC,KAAK,QAAQ,4BAA4B,WAAW,QAAQ,EACjG;CAEF,YAAY,KAAK,YAAY,YAAY;CACzC,YAAY,KAAK,eAAe,eAAe;CAC/C,SAAS,KAAK,YAAY,YAAY;CACtC,MAAM,OAAO,IAAI,IAAI,WAAW,SAAS,KAAK,YAAY,CAAC,QAAQ,IAAI,OAAO,CAAC,CAAC;CAChF,MAAM,2BAAW,IAAI,IAAwB;CAC7C,MAAM,YAA4B;EAAE,UAAU;EAAG,UAAU;EAAG,WAAW;EAAG,WAAW;CAAE;CACzF,MAAM,YAAsB,CAAC;CAC7B,KAAK,MAAM,CAAC,OAAO,QAAQ,KAAK,QAAQ,GAAG;EACzC,MAAM,UAAU,KAAK,IAAI,KAAK,SAAS;EACvC,IAAI,CAAC,SACH,MAAM,IAAI,gBACR,mBAAmB,MAAM,iBAAiB,KAAK,UAAU,KAAK,SAAS,EAAE,qBAAqB,WAAW,QAAQ,oBACnH;EAEF,MAAM,UAAU,eAAe,IAAI,SAAS,OAAO,MAAM,SAAS;EAClE,IAAI,IAAI,YAAY,KAAA,GAAW;GAC7B,IAAI,CAAC,MAAM,QAAQ,IAAI,OAAO,KAAK,IAAI,QAAQ,WAAW,GACxD,MAAM,IAAI,gBAAgB,mBAAmB,MAAM,mCAAmC;GAExF,MAAM,UAAU,IAAI,QAAQ,KAAK,QAAQ,MACvC,eAAe,QAAQ,OAAO,MAAM,UAAU,GAAG,CACnD;GACA,UAAU,KACR,kBAAkB,QAAQ,QAAQ,WAAW,WAAW,MAAM,CAAC,CAAC,QAAQ,QAAQ,MAAM,CACxF;EACF;EACA,MAAM,OAAO,SAAS,IAAI,QAAQ,EAAE,KAAK,CAAC;EAC1C,KAAK,KAAK;GAAE,GAAG;GAAK;EAAQ,CAAC;EAC7B,SAAS,IAAI,QAAQ,IAAI,IAAI;EAC7B,IAAI,YAAY,QAAQ;GACtB,IAAI,QAAQ,YAAY,QAAQ,UAAU;QACrC,UAAU;EACjB,OAAO,IAAI,QAAQ,YAAY,QAAQ,UAAU;OAC5C,UAAU;CACjB;CAEA,MAAM,aAAsC,CAAC;CAC7C,MAAM,WAAqB,CAAC;CAC5B,MAAM,kBAAgD;EAAE,MAAM;EAAG,MAAM;CAAE;CACzE,KAAK,MAAM,WAAW,WAAW,UAAU;EACzC,MAAM,OAAO,SAAS,IAAI,QAAQ,EAAE;EACpC,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,QAAQ,EAAE;GACxB;EACF;EACA,gBAAgB,QAAQ,QAAQ;EAChC,MAAM,SAAS,KAAK,QAAQ,QAAQ,IAAI,YAAY,MAAM,CAAC,CAAC;EAC5D,WAAW,KAAK;GACd,IAAI,QAAQ;GACZ,SAAS,QAAQ;GACjB,MAAM,KAAK;GACX,UAAU,SAAS,KAAK;GACxB,WAAW,KAAK,QAAQ,QAAQ,IAAI,YAAY,QAAQ,OAAO,CAAC,CAAC,SAAS,KAAK;EACjF,CAAC;CACH;CAEA,MAAM,QAAQ,KAAK;CACnB,MAAM,aAAa,UAAU,WAAW,UAAU;CAClD,MAAM,iBAAiB,WAAW,QAAQ,YAAY,QAAQ,cAAc,CAAC,CAAC,CAAC;CAC/E,MAAM,QAAQ,WACZ,OAAO,WAAW,IAAI,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;CACtF,OAAO;EACL,GAAG;EACH,OAAO,WAAW;EAClB,UAAU,WAAW,SAAS;EAC9B,QAAQ,WAAW;EACnB;EACA,MAAM;EACN,UAAU,UAAU,IAAI,OAAO,aAAa;EAC5C,iBACE,WAAW,WAAW,IAClB,OACA;GACE,QAAQ;GACR,QAAQ,WAAW;GACnB,UAAU,gBACR;IAAE,MAAM;IAAmB,OAAO;GAAK,GACvC;IAAE,MAAM;IAAY,WAAW;IAAgB,QAAQ,WAAW;GAAO,CAC3E;EACF;EACN,OAAO,YAAY,SAAS;EAC5B,eACE,UAAU,WAAW,UAAU,cAAc,IACzC,OACA,UAAU,aAAa,UAAU,WAAW,UAAU;EAC5D,WACE,UAAU,WAAW,UAAU,cAAc,IACzC,OACA,UAAU,YAAY,UAAU,WAAW,UAAU;EAC3D;EACA,UAAU;GACR,aAAa,KAAK,WAAW,KAAK,YAAY,QAAQ,YAAY,IAAI,QAAQ,SAAS,CAAC;GACxF,WAAW,KAAK,SAAS;GACzB,UAAU,WACP,QAAQ,YAAY,QAAQ,WAAW,KAAK,QAAQ,WAAW,CAAC,CAAC,CACjE,KAAK,YAAY,QAAQ,EAAE;EAChC;EACA,eAAe,WACZ,QAAQ,YAAY,QAAQ,YAAY,CAAC,CAAC,CAC1C,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,cAAc,EAAE,KAAK,EAAE,KAAK,KAAK,EAAE;EACrE;CACF;AACF;AAwBA,MAAa,4BAAuD,OAAO,OAAO;CAChF,SAAS;CACT,aAAa;CACb,eAAe;CACf,aAAa;CACb,UAAU;CACV,kBAAkB;CAClB,YAAY;AACd,CAAC;AA+BD,MAAM,SAAS;AAEf,SAAS,QAAQ,OAAuB;CACtC,OAAO,IAAI,QAAQ,IAAA,CAAK,QAAQ,CAAC,EAAE;AACrC;;AAGA,SAAgB,kBAAkB,OAA0C;CAC1E,MAAM,UAAU,YAAY,MAAM,SAAS,SAAS;CACpD,MAAM,aAAa,YAAY,MAAM,YAAY,YAAY;CAC7D,MAAM,gBAAgB,YAAY,MAAM,eAAe,eAAe;CACtE,MAAM,OAAO,SAAS,MAAM,MAAM,MAAM;CACxC,MAAM,SAA0B;EAC9B,GAAG;EACH,GAAG,OAAO,YACR,OAAO,QAAQ,MAAM,UAAU,CAAC,CAAC,CAAC,CAAC,QAAQ,GAAG,WAAW,UAAU,KAAA,CAAS,CAC9E;CACF;CACA,IAAI,OAAO,YAAY,aAAa,OAAO,YAAY,QACrD,MAAM,IAAI,gBACR,+DAA+D,KAAK,UAAU,OAAO,OAAO,GAC9F;CAEF,MAAM,YAAY,MAAM;CACxB,MAAM,UAAU,QAAyB,aAA0C;EACjF;EACA;EACA;EACA,SAAS,WAAW;EACpB;EACA;EACA;EACA;CACF;CAEA,IAAI,CAAC,WAAW,OAAO,OAAO,gBAAgB,CAAC,2CAA2C,CAAC;CAC3F,IAAI,UAAU,YAAY,SACxB,MAAM,IAAI,gBACR,qCAAqC,UAAU,QAAQ,UAAU,QAAQ,EAC3E;CAGF,MAAM,eAAyB,CAAC;CAChC,IAAI,UAAU,eAAe,YAC3B,aAAa,KACX,mCAAmC,UAAU,WAAW,uBAAuB,YACjF;CAEF,IAAI,UAAU,kBAAkB,eAC9B,aAAa,KACX,oCAAoC,UAAU,cAAc,uBAAuB,eACrF;CAEF,MAAM,WAAW,OAAO,SAAS,UAAU,YAAY,sBAAsB,KAAK;CAClF,IAAI,UAAU,OAAO,YACnB,aAAa,KAAK,gBAAgB,QAAQ,QAAQ,CAAC,EAAE,mBAAmB,OAAO,WAAW,EAAE;CAE9F,IAAI,OAAO,YAAY,QAEjB;MAAA,UAAU,gBAAgB,OAAO,OAAO,aAC1C,aAAa,KACX,GAAG,UAAU,gBAAgB,KAAK,sEAAsE,OAAO,aACjH;CAAA,OAEG;EACL,IAAI,UAAU,SAAS,OAAO,aAC5B,aAAa,KACX,GAAG,UAAU,OAAO,UAAU,UAAU,WAAW,IAAI,YAAY,WAAW,aAAa,OAAO,YAAY,UAChH;EAEF,KAAK,MAAM,WAAW,CAAC,QAAQ,MAAM,GACnC,IAAI,UAAU,gBAAgB,WAAW,OAAO,eAC9C,aAAa,KACX,GAAG,UAAU,gBAAgB,SAAS,oCAAoC,QAAQ,aAAa,OAAO,cAAc,UACtH;CAGN;CACA,IAAI,aAAa,SAAS,GAAG,OAAO,OAAO,gBAAgB,YAAY;CAEvE,MAAM,QAAkB,CAAC;CACzB,IAAI,OAAO,YAAY,QACjB;MAAA,UAAU,kBAAkB,QAAQ,UAAU,gBAAgB,OAAO,kBACvE,MAAM,KACJ,YAAY,UAAU,kBAAkB,OAAO,wBAAwB,QAAQ,UAAU,aAAa,EAAE,mCAAmC,QAAQ,OAAO,gBAAgB,EAAE,EAC9K;CAAA,OAEG,IAAI,UAAU,aAAa,QAAQ,UAAU,WAAW,OAAO,aACpE,MAAM,KACJ,YAAY,UAAU,aAAa,OAAO,eAAe,QAAQ,UAAU,QAAQ,EAAE,YAAY,QAAQ,OAAO,WAAW,GAC7H;CAEF,IACE,OAAO,YAAY,cAClB,UAAU,UAAU,QAAQ,UAAU,QAAQ,OAAO,WAEtD,MAAM,KACJ,SAAS,UAAU,UAAU,OAAO,cAAc,UAAU,MAAM,QAAQ,CAAC,EAAE,YAAY,OAAO,UAClG;CAEF,IAAI,MAAM,SAAS,GAAG,OAAO,OAAO,mBAAmB,KAAK;CAE5D,MAAM,WAAW,MAAM,WAAW,CAAC,EAAA,CAAG,QAAQ,aAAa,SAAS,YAAY,OAAO;CACvF,IAAI,QAAQ,SAAS,GAAG;EACtB,MAAM,SAAS,oBAAoB,CAAC,GAAG,OAAO,GAAG;GAC/C,MAAM,MAAM;GACZ,GAAI,OAAO,QAAQ,EAAE,YAAY,OAAO,MAAM,IAAI,CAAC;EACrD,CAAC;EACD,IAAI,CAAC,OAAO,SAAS,OAAO,OAAO,WAAW,OAAO,MAAM;CAC7D;CACA,OAAO,OAAO,cAAc,CAAC,CAAC;AAChC;;AAGA,IAAa,wBAAb,cAA2C,WAAW;CACpD;CAEA,YAAY,UAA6B;EACvC,MACE,UAAU,SAAS,QAAQ,kBAAkB,SAAS,OAAO,KAAK,SAAS,QAAQ,KAAK,IAAI,GAC9F;EACA,KAAK,WAAW;CAClB;AACF;;AAGA,SAAgB,mBAAmB,UAAmC;CACpE,IAAI,CAAC,SAAS,SAAS,MAAM,IAAI,sBAAsB,QAAQ;AACjE;;;;;;;;AASA,SAAgB,2BACd,WACA,UAAoD,CAAC,GACnC;CAClB,MAAM,OAAO,QAAQ,YAAY;CACjC,MAAM,OAAO,OACT,UAAU,kBAAkB,OAC1B,OACA,IAAI,UAAU,gBAChB,UAAU;CACd,IAAI,SAAS,MACX,MAAM,IAAI,gBACR,sCAAsC,UAAU,QAAQ,WAAW,OAAO,yBAAyB,OAAO,aAC5G;CAEF,MAAM,WAA6B;EACjC,IAAI,UAAU;EACd,SAAS,UAAU;EACnB,YAAY,UAAU;EACtB,SAAS;GACP,kBAAkB;GAClB,GAAI,QAAQ,UAAU,UAAU,OAAO,CAAC,IAAI,EAAE,kBAAkB,UAAU,MAAM;EAClF;CACF;CACA,yBAAyB,QAAQ;CACjC,OAAO;AACT"}

package/dist/mean-CdDYG-us.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"mean-CdDYG-us.js","names":[],"sources":["../src/math/student-t.ts","../src/statistics/paired-tests.ts","../src/statistics/mean.ts"],"sourcesContent":["import { regularizedIncompleteBeta } from './special-functions'\n\n/**\n * Student-t CDF via the regularized incomplete beta function.\n */\nexport function studentTCdf(t: number, degreesOfFreedom: number): number {\n  if (degreesOfFreedom <= 0) return 0.5\n\n  const x = degreesOfFreedom / (degreesOfFreedom + t * t)\n  const beta = regularizedIncompleteBeta(x, degreesOfFreedom / 2, 0.5)\n  return t >= 0 ? 1 - 0.5 * beta : 0.5 * beta\n}\n\n/**\n * Inverse Student-t CDF, solved against {@link studentTCdf}.\n *\n * The CDF is monotone, so bracket expansion followed by bisection is stable\n * across fractional degrees of freedom and does not need a separate\n * approximation with a different error profile.\n */\nexport function studentTQuantile(probability: number, degreesOfFreedom: number): number {\n  if (!Number.isFinite(probability) || probability < 0 || probability > 1) {\n    throw new RangeError(`studentTQuantile: probability must be in [0,1], got ${probability}`)\n  }\n  if (!Number.isFinite(degreesOfFreedom) || degreesOfFreedom <= 0) {\n    throw new RangeError(\n      `studentTQuantile: degreesOfFreedom must be positive and finite, got ${degreesOfFreedom}`,\n    )\n  }\n  if (probability === 0) return Number.NEGATIVE_INFINITY\n  if (probability === 1) return Number.POSITIVE_INFINITY\n  if (probability === 0.5) return 0\n  if (probability < 0.5) return -studentTQuantile(1 - probability, degreesOfFreedom)\n\n  let low = 0\n  let high = 1\n  while (studentTCdf(high, degreesOfFreedom) < probability) high *= 2\n  for (let iteration = 0; iteration < 64; iteration++) {\n    const middle = (low + high) / 2\n    if (studentTCdf(middle, degreesOfFreedom) < probability) low = middle\n    else high = middle\n  }\n  return (low + high) / 2\n}\n","/**\n * Paired significance tests on continuous scores: the paired t-test, the\n * promotion-gate paired bootstrap, the exact sign test, and the package-wide\n * paired-delta decision statistic.\n */\n\nimport { ValidationError } from '../errors'\nimport { studentTCdf } from '../math/student-t'\nimport { assertFiniteSample, binomialHalfUpperTail, makeRng, medianInPlace } from './internal'\n\nexport interface PairedTTestResult {\n  /** Null when the statistic is undefined — see {@link pairedTTest}. */\n  t: number | null\n  df: number\n  /** Null exactly when `t` is null. */\n  p: number | null\n}\n\n/**\n * Paired t-test — before/after measurements on the SAME items.\n * Pairing removes inter-item variance, giving tighter significance than\n * an unpaired test when comparing prompt v1 vs prompt v2 on identical\n * scenarios.\n *\n * Returns `t = p = null` where the statistic is undefined: fewer than two\n * pairs, or a non-zero constant delta whose observed variance is zero. A\n * constant shift carries no information about the variance it would have to\n * be compared against, so the honest answer is \"undefined\", not `p = 0` —\n * three observations cannot buy absolute certainty. This is the same contract\n * {@link pairedCohensDz} states for the same condition. An all-zero delta is\n * different: it is a measured null, and returns `t = 0, p = 1`.\n */\nexport function pairedTTest(before: number[], after: number[]): PairedTTestResult {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedTTest: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  assertFiniteSample('pairedTTest', 'before', before)\n  assertFiniteSample('pairedTTest', 'after', after)\n  const n = before.length\n  if (n < 2) return { t: null, df: 0, p: null }\n\n  const diffs = before.map((b, i) => after[i]! - b)\n  const mean = diffs.reduce((a, b) => a + b, 0) / n\n  const variance = diffs.reduce((acc, d) => acc + (d - mean) ** 2, 0) / (n - 1)\n  const se = Math.sqrt(variance / n)\n  if (se === 0) {\n    return mean === 0 ? { t: 0, df: n - 1, p: 1 } : { t: null, df: n - 1, p: null }\n  }\n\n  const t = mean / se\n  const df = n - 1\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\n// ── Paired bootstrap (promotion-gate effect size) ────────────────────\n\nexport interface PairedBootstrapResult {\n  /** Number of paired observations. */\n  n: number\n  /** Median of paired deltas (after − before). */\n  median: number\n  /** Mean of paired deltas. */\n  mean: number\n  /** Lower bound of the bootstrap CI on the chosen statistic. */\n  low: number\n  /** Upper bound of the bootstrap CI on the chosen statistic. */\n  high: number\n  /** Confidence level used (e.g. 0.95). */\n  confidence: number\n  /** Number of bootstrap resamples used. */\n  resamples: number\n  /** False below {@link BOOTSTRAP_GATE_MIN_N}. See {@link pairedBootstrap}. */\n  gateEligible: boolean\n}\n\n/**\n * Pairs below which a percentile bootstrap interval is descriptive spread only.\n *\n * `P(low > 0)` under a true null, against a nominal 2.5 %, measured over 4000\n * seeded trials: 13.53 % at n = 3, 3.52 % at n = 10, 3.10 % at n = 20 on the\n * median; 13.85 %, 4.90 %, 3.80 % on the mean. This is intrinsic to resampling\n * three points, not an implementation error — scipy's BCa gives 16.0 % on the\n * same n = 3 data — so no change to the estimator moves it. Below this floor\n * the decision belongs to the exact sign test or exact signed-rank test.\n */\nexport const BOOTSTRAP_GATE_MIN_N = 20\n\nexport interface PairedBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Statistic to bootstrap. Default 'median'. */\n  statistic?: 'median' | 'mean'\n  /** Deterministic seed. If omitted, derived from the deltas so the interval\n   *  is reproducible regardless. */\n  seed?: number\n}\n\n/**\n * Paired bootstrap on (after − before) deltas. Returns a CI on the chosen\n * statistic (median by default); pairs are resampled with replacement. Throws\n * on unequal sample sizes.\n *\n * `low > threshold` carries the stated confidence ONLY at `n ≥\n * {@link BOOTSTRAP_GATE_MIN_N}`, which `gateEligible` reports. Below it the\n * check fires under a true null several times more often than nominal, so the\n * interval is descriptive spread and a promotion must not turn on it.\n */\nexport function pairedBootstrap(\n  before: number[],\n  after: number[],\n  opts: PairedBootstrapOptions = {},\n): PairedBootstrapResult {\n  if (before.length !== after.length) {\n    throw new Error(`pairedBootstrap: unequal sample sizes (${before.length} vs ${after.length})`)\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const statistic = opts.statistic ?? 'median'\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedBootstrap: confidence must be in (0,1), got ${confidence}`)\n  }\n\n  const n = before.length\n  const deltas = before.map((b, i) => after[i]! - b)\n  const gateEligible = n >= BOOTSTRAP_GATE_MIN_N\n  if (n === 0) {\n    return { n: 0, median: 0, mean: 0, low: 0, high: 0, confidence, resamples, gateEligible }\n  }\n  if (n === 1) {\n    const d = deltas[0]!\n    return { n: 1, median: d, mean: d, low: d, high: d, confidence, resamples, gateEligible }\n  }\n\n  const rng = makeRng(opts.seed, deltas)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    if (statistic === 'mean') {\n      let sum = 0\n      for (let k = 0; k < n; k++) {\n        sum += deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = sum / n\n    } else {\n      const acc = new Array<number>(n)\n      for (let k = 0; k < n; k++) {\n        acc[k] = deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = medianInPlace(acc)\n    }\n  }\n  samples.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n\n  return {\n    n,\n    median: medianInPlace([...deltas]),\n    mean: deltas.reduce((s, x) => s + x, 0) / n,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n    gateEligible,\n  }\n}\n\n/** Pre-registered direction for a one-sided paired sign test. */\nexport type SignTestAlternative = 'greater' | 'less'\n\n/** Exact one-sided sign-test result for paired numeric differences. */\nexport interface PairedSignTestResult {\n  /** Total supplied differences, including zero ties. */\n  n: number\n  /** Strictly positive differences. */\n  positive: number\n  /** Strictly negative differences. */\n  negative: number\n  /** Zero differences excluded from the binomial test. */\n  ties: number\n  /** Non-zero differences used by the binomial test. */\n  nNonTies: number\n  /** Direction of the pre-registered alternative hypothesis. */\n  alternative: SignTestAlternative\n  /** Exact one-sided p-value under P(positive) = P(negative) = 0.5. */\n  pValue: number\n}\n\n/**\n * Exact one-sided sign test over paired differences.\n *\n * Pass `after[i] - before[i]` for each matched item. `alternative = 'greater'`\n * tests whether positive signs are more likely than negative signs and returns\n * `P(Binomial(nNonTies, 0.5) >= positive)`. `alternative = 'less'` treats\n * negative signs as successes instead. With a continuous difference\n * distribution this is the usual directional median test. Exact zero\n * differences are ties and do not enter the binomial denominator. All-tie and\n * empty inputs return p = 1. Every input difference must be finite, and the\n * direction must be chosen explicitly so a caller cannot select it after\n * seeing the signs.\n */\nexport function pairedSignTest(\n  differences: readonly number[],\n  alternative: SignTestAlternative,\n): PairedSignTestResult {\n  if (alternative !== 'greater' && alternative !== 'less') {\n    throw new ValidationError(\n      `pairedSignTest: alternative must be 'greater' or 'less', got ${alternative}`,\n    )\n  }\n\n  let positive = 0\n  let negative = 0\n  let ties = 0\n  for (let i = 0; i < differences.length; i++) {\n    const difference = differences[i]!\n    if (!Number.isFinite(difference)) {\n      throw new ValidationError(\n        `pairedSignTest: difference at index ${i} must be finite, got ${difference}`,\n      )\n    }\n    if (difference > 0) positive++\n    else if (difference < 0) negative++\n    else ties++\n  }\n\n  const nNonTies = positive + negative\n  const successes = alternative === 'greater' ? positive : negative\n  return {\n    n: differences.length,\n    positive,\n    negative,\n    ties,\n    nNonTies,\n    alternative,\n    pValue: binomialHalfUpperTail(successes, nNonTies),\n  }\n}\n\n/** Fraction of paired observations whose delta is an exact tie (|after − before|\n *  < 1e-9). Throws on unequal sample sizes; 0 pairs ⇒ 0. */\nexport function pairedDeltaTieFraction(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number {\n  if (before.length !== after.length) {\n    throw new Error(\n      `pairedDeltaTieFraction: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const n = before.length\n  if (n === 0) return 0\n  let ties = 0\n  for (let i = 0; i < n; i++) {\n    if (Math.abs(after[i]! - before[i]!) < 1e-9) ties++\n  }\n  return ties / n\n}\n\n/**\n * The paired-delta statistic a DECISION is computed on, package-wide.\n *\n * The mean paired delta is the estimator that answers the question a promotion\n * gate asks — \"by how much did the candidate move the score\" — in the caller's\n * own units, and it equals the aggregate lift everyone quotes. The MEDIAN\n * answers a different question and loses the answer to this one in every regime\n * eval data actually lands in:\n *   - TWO-POINT (pass/fail) outcomes on any encoding: the delta vector lives in\n *     {−s, 0, +s} dominated by zeros, so the median and its whole bootstrap CI\n *     are pinned at exactly 0 however large the shift. (Decide these on\n *     {@link pairedRiskDifferenceExact} instead — same estimand, exact interval.)\n *   - TIE-DOMINATED outcomes: at half the pairs tied the sample median is 0 by\n *     construction, and `ci.low > threshold` then answers \"no\" forever at a\n *     non-negative threshold and \"yes\" forever at a negative one.\n *   - LOW-CARDINALITY outcomes, even well below half ties: judge dimensions on\n *     integer 0-100, and block scores like {⅔, 1} from averaging pass/fail\n *     leaves, put the median on a coarse lattice whose bootstrap percentiles\n *     land on atoms. Measured: 26 blocks of 3 pass/fail leaves carrying a real\n *     +12.8pp lift, only 23% of pairs tied, gives a median CI of [0, 0.333] —\n *     lower bound exactly 0, so a gate at threshold 0 refuses a real lift.\n * That last case is why there is no tie-fraction threshold here: any cutoff on\n * ties leaves the lattice case open on the other side of it.\n *\n * `heldoutSignificance` has defaulted to the mean since #316 for the same\n * reason. The median remains available per call site for callers who\n * specifically want outlier robustness and accept the blindness.\n */\nexport const DECISION_PAIRED_DELTA_STATISTIC: 'mean' = 'mean'\n","/** Arithmetic mean. An empty input has no mean and returns NaN. */\nexport function mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;;AAKA,SAAgB,YAAY,GAAW,kBAAkC;CACvE,IAAI,oBAAoB,GAAG,OAAO;CAElC,MAAM,IAAI,oBAAoB,mBAAmB,IAAI;CACrD,MAAM,OAAO,0BAA0B,GAAG,mBAAmB,GAAG,EAAG;CACnE,OAAO,KAAK,IAAI,IAAI,KAAM,OAAO,KAAM;AACzC;;;;;;;;AASA,SAAgB,iBAAiB,aAAqB,kBAAkC;CACtF,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,KAAK,cAAc,GACpE,MAAM,IAAI,WAAW,uDAAuD,aAAa;CAE3F,IAAI,CAAC,OAAO,SAAS,gBAAgB,KAAK,oBAAoB,GAC5D,MAAM,IAAI,WACR,uEAAuE,kBACzE;CAEF,IAAI,gBAAgB,GAAG,OAAO,OAAO;CACrC,IAAI,gBAAgB,GAAG,OAAO,OAAO;CACrC,IAAI,gBAAgB,IAAK,OAAO;CAChC,IAAI,cAAc,IAAK,OAAO,CAAC,iBAAiB,IAAI,aAAa,gBAAgB;CAEjF,IAAI,MAAM;CACV,IAAI,OAAO;CACX,OAAO,YAAY,MAAM,gBAAgB,IAAI,aAAa,QAAQ;CAClE,KAAK,IAAI,YAAY,GAAG,YAAY,IAAI,aAAa;EACnD,MAAM,UAAU,MAAM,QAAQ;EAC9B,IAAI,YAAY,QAAQ,gBAAgB,IAAI,aAAa,MAAM;OAC1D,OAAO;CACd;CACA,QAAQ,MAAM,QAAQ;AACxB;;;;;;;;;;;;;;;;;;;;;ACXA,SAAgB,YAAY,QAAkB,OAAoC;CAChF,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,sCAAsC,OAAO,OAAO,MAAM,MAAM,OAAO,EACzE;CAEF,mBAAmB,eAAe,UAAU,MAAM;CAClD,mBAAmB,eAAe,SAAS,KAAK;CAChD,MAAM,IAAI,OAAO;CACjB,IAAI,IAAI,GAAG,OAAO;EAAE,GAAG;EAAM,IAAI;EAAG,GAAG;CAAK;CAE5C,MAAM,QAAQ,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CAChD,MAAM,OAAO,MAAM,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAChD,MAAM,WAAW,MAAM,QAAQ,KAAK,MAAM,OAAO,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI;CAC3E,MAAM,KAAK,KAAK,KAAK,WAAW,CAAC;CACjC,IAAI,OAAO,GACT,OAAO,SAAS,IAAI;EAAE,GAAG;EAAG,IAAI,IAAI;EAAG,GAAG;CAAE,IAAI;EAAE,GAAG;EAAM,IAAI,IAAI;EAAG,GAAG;CAAK;CAGhF,MAAM,IAAI,OAAO;CACjB,MAAM,KAAK,IAAI;CAEf,OAAO;EAAE;EAAG;EAAI,GADN,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;CAC5B;AACpB;;;;;;;;;;;AAiCA,MAAa,uBAAuB;;;;;;;;;;;AAwBpC,SAAgB,gBACd,QACA,OACA,OAA+B,CAAC,GACT;CACvB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MAAM,0CAA0C,OAAO,OAAO,MAAM,MAAM,OAAO,EAAE;CAE/F,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,qDAAqD,YAAY;CAGnF,MAAM,IAAI,OAAO;CACjB,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,eAAe,KAAA;CACrB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,QAAQ;EAAG,MAAM;EAAG,KAAK;EAAG,MAAM;EAAG;EAAY;EAAW;CAAa;CAE1F,IAAI,MAAM,GAAG;EACX,MAAM,IAAI,OAAO;EACjB,OAAO;GAAE,GAAG;GAAG,QAAQ;GAAG,MAAM;GAAG,KAAK;GAAG,MAAM;GAAG;GAAY;GAAW;EAAa;CAC1F;CAEA,MAAM,MAAM,QAAQ,KAAK,MAAM,MAAM;CACrC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAC7B,IAAI,cAAc,QAAQ;EACxB,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,OAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEpC,QAAQ,KAAK,MAAM;CACrB,OAAO;EACL,MAAM,MAAM,IAAI,MAAc,CAAC;EAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEtC,QAAQ,KAAK,cAAc,GAAG;CAChC;CAEF,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAE5B,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;CACjD,MAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;CAElF,OAAO;EACL;EACA,QAAQ,cAAc,CAAC,GAAG,MAAM,CAAC;EACjC,MAAM,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;EAC1C,KAAK,QAAQ;EACb,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM;EACtC;EACA;EACA;CACF;AACF;;;;;;;;;;;;;;AAoCA,SAAgB,eACd,aACA,aACsB;CACtB,IAAI,gBAAgB,aAAa,gBAAgB,QAC/C,MAAM,IAAI,gBACR,gEAAgE,aAClE;CAGF,IAAI,WAAW;CACf,IAAI,WAAW;CACf,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;EAC3C,MAAM,aAAa,YAAY;EAC/B,IAAI,CAAC,OAAO,SAAS,UAAU,GAC7B,MAAM,IAAI,gBACR,uCAAuC,EAAE,uBAAuB,YAClE;EAEF,IAAI,aAAa,GAAG;OACf,IAAI,aAAa,GAAG;OACpB;CACP;CAEA,MAAM,WAAW,WAAW;CAC5B,MAAM,YAAY,gBAAgB,YAAY,WAAW;CACzD,OAAO;EACL,GAAG,YAAY;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ,sBAAsB,WAAW,QAAQ;CACnD;AACF;;;AAIA,SAAgB,uBACd,QACA,OACQ;CACR,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,iDAAiD,OAAO,OAAO,MAAM,MAAM,OAAO,EACpF;CAEF,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO;CACpB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,IAAI,MAAM,KAAM,OAAO,EAAG,IAAI,MAAM;CAE/C,OAAO,OAAO;AAChB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,MAAa,kCAA0C;;;ACpSvD,SAAgB,KAAK,QAAmC;CACtD,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}