@tangle-network/agent-eval 1.0.5 → 1.0.7

package/dist/judge-gate-C9sLLJ-v.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"judge-gate-C9sLLJ-v.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 { meanOrNull } from '../statistics/mean'\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  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: meanOrNull(\n        perExample.map((example) => example.passRate * (1 - example.passRate)),\n      ),\n      withinRun: meanOrNull(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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACzYA,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,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,WACX,WAAW,KAAK,YAAY,QAAQ,YAAY,IAAI,QAAQ,SAAS,CACvE;GACA,WAAW,WAAW,SAAS;GAC/B,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/promotion-gate-CixshGI8.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"promotion-gate-CixshGI8.d.ts","names":[],"sources":["../src/statistics/rank-tests.ts","../src/release-confidence.ts","../src/promotion-gate.ts"],"mappings":";;;;;KAiCY;;;;;KAMA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;UAYe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;;KCxLE;;UAGY;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;;;;;;;;;;;;;;;;;;;;;;;;;KC3JS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;;EAGA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA0Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB"}

package/dist/sandbox-harness-BS6c6UZV.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"sandbox-harness-BS6c6UZV.d.ts","names":[],"sources":["../src/sandbox-harness.ts"],"mappings":"A,UAMiB;;EAEf;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA,MAAM;;EAEN,aAAa;;UAGE;EACf;EACA,MACE,gBACA,gBACA;IACG;IAAoB;;;UAGV;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA,KACE,OAAO,wBACP,iBACA,QAAQ,gBACP,QAAQ;;UAiDI;EACf;EACA,QAAQ;EACR,MAAM;EACN,OAAO;EACP;;EAEA;;;UA4De;EACf;EACA;EACA;EACA;EACA;EACA;;KAEU,qBACR,6BACU,cAAc,iBAAiB,UAAU,KAAK;UAC3C;EACf;EACA;EACA;EACA;;UAEe;EACf;EACA,YAAY;EACZ,SAAS;EACT;EACA,WAAW;;UAEI,gBAAgB;EAC/B;EACA,OAAO,QAAQ,cAAc,QAAQ;EACrC,OAAO,UAAU,IAAI;EACrB,SAAS,UAAU,GAAG,QAAQ,gBAAgB;EAC9C,YAAY,UAAU;;UAEP;EACf,QAAQ;EACR,SAAS,GAAG,OAAO,gBAAgB,KAAK;IAAU,WAAW;IAAG,QAAQ;;EACxE,SAAS,YAAY;EACrB,QAAQ,cAAc,WAAW,mBAAmB,QAAQ;;EAE5D,MAAM,cAAc;EACpB,KAAK,GACH,YAAY,QAAQ,IACpB,QAAQ,OAAO,eACf;IAAU;IAAmB;MAC5B,QAAQ;;UAEI,gBAAgB;EAC/B,MAAM,OAAO,eAAe,QAAQ;EACpC,QAAQ;IACN;;IAEA;IACA,QAAQ,OAAO;IACf,IAAI,OAAO,GAAG,OAAO,eAAe;IACpC,UAAU,OAAO,GAAG,QAAQ,kBAAkB;;EAEhD,gBAAgB,OAAO,MAAM;EAC7B,YAAY,OAAO,MAAM;;EAEzB;;;iBAIoB,mBAAmB,GACvC,MAAM,gBAAgB,IACtB;EACE,SAAS;EACT,UAAU;IAEX,QAAQ"}

package/dist/sandbox-harness-BzAUasgY.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"sandbox-harness-BzAUasgY.js","names":["delay"],"sources":["../src/sandbox-harness.ts"],"sourcesContent":["import { setTimeout as delay } from 'node:timers/promises'\nimport { runBoundedProcess } from './bounded-process'\n\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /** Captured stdout + stderr cap, default 16 MiB; excess is dropped. */\n  maxOutputBytes?: number\n  image?: string\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n  /** A timed-out phase always fails, even with exitCode 0. Absent/false means natural exit. */\n  killedByTimeout?: boolean\n  /** Captured output exceeded maxOutputBytes. */\n  outputTruncated?: boolean\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\ninterface SubprocessSandboxDriverOptions {\n  cwd?: string\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  constructor(private options: SubprocessSandboxDriverOptions = {}) {}\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    const result = await runBoundedProcess({\n      command,\n      shell: true,\n      cwd: config.cwd ?? this.options.cwd,\n      env: { ...process.env, ...this.options.env, ...config.env },\n      envMode: 'replace',\n      timeoutMs: config.timeoutMs ?? 10 * 60_000,\n      maxOutputBytes: config.maxOutputBytes ?? 16 * 1024 * 1024,\n    })\n    // A timed-out phase never reaches the parser: the output of a killed\n    // process is a fragment, and a fragment that happens to parse would\n    // report a pass for a run that never finished.\n    const parsed =\n      !result.killedByTimeout && phase === 'test' && config.testParser\n        ? config.testParser.parse(result.stdout, result.stderr, result.exitCode)\n        : undefined\n    return {\n      phase,\n      exitCode: result.exitCode,\n      stdout: result.stdout,\n      stderr: result.runnerError ? result.stderr + result.runnerError : result.stderr,\n      wallMs: result.wallMs,\n      testsTotal: parsed?.testsTotal,\n      testsPassed: parsed?.testsPassed,\n      killedByTimeout: result.killedByTimeout || undefined,\n      outputTruncated: result.outputTruncated || undefined,\n    }\n  }\n}\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  constructor(private driver: SandboxDriver = new SubprocessSandboxDriver()) {}\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      for (const phase of ['setup', 'run', 'test'] as const) {\n        const command = config[`${phase}Command`]\n        if (!command) continue\n        const measured = await this.driver.exec(phase, command, config)\n        result[phase] = measured\n        result.totalWallMs += measured.wallMs\n        const passed = !measured.killedByTimeout && measured.exitCode === 0\n        if (phase !== 'test' && !passed) {\n          await handle.fail(\n            measured.killedByTimeout\n              ? `${phase} timed out after ${measured.wallMs}ms`\n              : `${phase} failed (exit ${measured.exitCode})`,\n            {\n              exitCode: measured.exitCode,\n              wallMs: result.totalWallMs,\n            } as Partial<SandboxSpan>,\n          )\n          return result\n        }\n      }\n      const test = result.test\n      result.passed = !test || (!test.killedByTimeout && test.exitCode === 0)\n      result.score =\n        test?.testsTotal !== undefined && test.testsTotal > 0\n          ? (test.testsPassed ?? 0) / test.testsTotal\n          : result.passed\n            ? 1\n            : 0\n      await handle.end({\n        exitCode: test?.exitCode,\n        testsTotal: test?.testsTotal,\n        testsPassed: test?.testsPassed,\n        status: result.passed ? 'ok' : 'error',\n        wallMs: result.totalWallMs,\n      } as Partial<SandboxSpan>)\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n\n/** Portable integration report; products may retain extra timing and assertion fields. */\nexport interface HarnessAssertion {\n  name: string\n  status: 'pass' | 'fail' | 'skip'\n  expected?: string\n  actual?: string\n  message?: string\n  soft?: boolean\n}\nexport type HarnessExpectation =\n  | HarnessAssertion\n  | readonly [name: string, passed: boolean, details?: Omit<HarnessAssertion, 'name' | 'status'>]\nexport interface HarnessTiming {\n  name: string\n  durationMs: number\n  success: boolean\n  error?: string\n}\nexport interface HarnessReport {\n  pass: boolean\n  assertions: HarnessAssertion[]\n  timings: HarnessTiming[]\n  error?: string\n  metadata?: Record<string, unknown>\n}\nexport interface HarnessResource<T> {\n  name: string\n  create(signal: AbortSignal): Promise<T>\n  retire(resource: T): Promise<unknown>\n  ready?: (resource: T, signal: AbortSignal) => Promise<unknown>\n  onCreate?: (resource: T) => void\n}\nexport interface HarnessScope {\n  signal: AbortSignal\n  resource<T>(input: HarnessResource<T>): Promise<{ resource?: T; timing: HarnessTiming }>\n  check(...assertions: HarnessExpectation[]): void\n  measure(name: string, run: () => Promise<unknown>): Promise<HarnessTiming>\n  /** Register retirement immediately after allocation; cleanup ignores workload cancellation. */\n  defer(retire: () => Promise<unknown>): void\n  poll<T>(\n    read: () => Promise<T>,\n    ready: (value: T) => boolean,\n    budget: { timeoutMs: number; intervalMs: number },\n  ): Promise<T | undefined>\n}\nexport interface HarnessScenario<T> {\n  setup(scope: HarnessScope): Promise<T>\n  steps?: ReadonlyArray<{\n    name: string\n    /** Diagnostic measurements retain failures without gating the scenario. */\n    advisory?: boolean\n    when?: (state: T) => boolean\n    run(state: T, scope: HarnessScope): Promise<unknown>\n    expect?: (state: T, timing: HarnessTiming) => HarnessExpectation[]\n  }>\n  expectations?: (state: T) => HarnessExpectation[]\n  metadata?: (state: T) => Record<string, unknown>\n  /** Includes setup and steps. Transports must honor scope.signal. */\n  budgetMs: number\n}\n\n/** Campaigns own whole-scenario retries; polling only repeats the declared read. */\nexport async function runSandboxScenario<T>(\n  spec: HarnessScenario<T>,\n  options: {\n    signal?: AbortSignal\n    measure?: HarnessScope['measure']\n  } = {},\n): Promise<HarnessReport> {\n  if (!(spec.budgetMs > 0 && Number.isSafeInteger(spec.budgetMs)))\n    throw new Error('Scenario budgetMs must be positive and finite')\n  const signal = AbortSignal.any([\n    AbortSignal.timeout(spec.budgetMs),\n    ...(options.signal ? [options.signal] : []),\n  ])\n  const report: HarnessReport = { pass: false, assertions: [], timings: [] }\n  let measurementFailures = 0\n  const retirements: Array<() => Promise<unknown>> = []\n  const scope: HarnessScope = {\n    signal,\n    check(...assertions) {\n      for (const value of assertions)\n        report.assertions.push(\n          'status' in value\n            ? value\n            : {\n                ...value[2],\n                name: value[0],\n                status: value[1] ? 'pass' : 'fail',\n              },\n        )\n    },\n    defer: (retire) => retirements.push(retire),\n    async measure(name, run) {\n      const start = performance.now()\n      let measured: HarnessTiming = { name, durationMs: 0, success: true }\n      try {\n        if (options.measure) measured = await options.measure(name, run)\n        else {\n          const result = (await run()) as { success?: boolean; error?: string } | undefined\n          if (result?.success === false)\n            throw new Error(result.error ?? 'measured operation returned success=false')\n          measured.durationMs = Math.round(performance.now() - start)\n        }\n      } catch (error) {\n        measured = {\n          name,\n          durationMs: Math.round(performance.now() - start),\n          success: false,\n          error: error instanceof Error ? error.message : String(error),\n        }\n      }\n      if (!measured.success) measurementFailures++\n      report.timings.push(measured)\n      return measured\n    },\n    async resource<T>(input: HarnessResource<T>) {\n      let resource: T | undefined\n      const timing = await scope.measure(input.name, async () => {\n        const owned = await input.create(signal)\n        resource = owned\n        scope.defer(() => input.retire(owned))\n        input.onCreate?.(owned)\n        await input.ready?.(owned, signal)\n      })\n      return { resource, timing }\n    },\n    async poll(read, ready, budget) {\n      if (!(budget.timeoutMs > 0 && budget.intervalMs > 0))\n        throw new Error('Polling timeout and interval must be positive')\n      const deadline = performance.now() + budget.timeoutMs\n      while (performance.now() < deadline) {\n        signal.throwIfAborted()\n        const value = await read()\n        if (ready(value)) return value\n        await delay(\n          Math.min(budget.intervalMs, Math.max(0, deadline - performance.now())),\n          undefined,\n          { signal },\n        )\n      }\n      return undefined\n    },\n  }\n  try {\n    signal.throwIfAborted()\n    const state = await spec.setup(scope)\n    for (const step of spec.steps ?? []) {\n      signal.throwIfAborted()\n      if (step.when && !step.when(state)) continue\n      const failuresBefore = measurementFailures\n      const timing = await scope.measure(step.name, () => step.run(state, scope))\n      if (step.advisory) measurementFailures = failuresBefore\n      scope.check(...(step.expect?.(state, timing) ?? []))\n    }\n    signal.throwIfAborted()\n    scope.check(...(spec.expectations?.(state) ?? []))\n    report.metadata = spec.metadata?.(state)\n  } catch (error) {\n    report.error = error instanceof Error ? error.message : String(error)\n  } finally {\n    for (const retire of retirements.reverse()) {\n      const measured = await scope.measure('cleanup', retire)\n      if (!measured.success)\n        scope.check({ name: 'cleanup', status: 'fail', message: measured.error })\n    }\n  }\n  report.pass =\n    !report.error &&\n    measurementFailures === 0 &&\n    report.assertions.every((a) => a.soft || a.status === 'pass')\n  return report\n}\n"],"mappings":";;;AA6DA,IAAa,0BAAb,MAA8D;CAExC;CADpB,KAAK;CACL,YAAY,UAAkD,CAAC,GAAG;EAA9C,KAAA,UAAA;CAA+C;CAEnE,MAAM,KACJ,OACA,SACA,QACwB;EACxB,MAAM,SAAS,MAAM,kBAAkB;GACrC;GACA,OAAO;GACP,KAAK,OAAO,OAAO,KAAK,QAAQ;GAChC,KAAK;IAAE,GAAG,QAAQ;IAAK,GAAG,KAAK,QAAQ;IAAK,GAAG,OAAO;GAAI;GAC1D,SAAS;GACT,WAAW,OAAO,aAAa;GAC/B,gBAAgB,OAAO,kBAAkB;EAC3C,CAAC;EAID,MAAM,SACJ,CAAC,OAAO,mBAAmB,UAAU,UAAU,OAAO,aAClD,OAAO,WAAW,MAAM,OAAO,QAAQ,OAAO,QAAQ,OAAO,QAAQ,IACrE,KAAA;EACN,OAAO;GACL;GACA,UAAU,OAAO;GACjB,QAAQ,OAAO;GACf,QAAQ,OAAO,cAAc,OAAO,SAAS,OAAO,cAAc,OAAO;GACzE,QAAQ,OAAO;GACf,YAAY,QAAQ;GACpB,aAAa,QAAQ;GACrB,iBAAiB,OAAO,mBAAmB,KAAA;GAC3C,iBAAiB,OAAO,mBAAmB,KAAA;EAC7C;CACF;AACF;AAcA,IAAa,iBAAb,MAA4B;CACN;CAApB,YAAY,SAAgC,IAAI,wBAAwB,GAAG;EAAvD,KAAA,SAAA;CAAwD;CAE5E,MAAM,IAAI,QAAuB,SAAsD;EACrF,MAAM,SAAS,MAAM,QAAQ,QAAQ;GACnC,MAAM,WAAW,KAAK,OAAO,GAAG;GAChC,OAAO,OAAO;GACd,SAAS;IAAC,OAAO;IAAc,OAAO;IAAY,OAAO;GAAW,CAAC,CAClE,OAAO,OAAO,CAAC,CACf,KAAK,MAAM;EAChB,CAAC;EACD,MAAM,SAA+B;GAAE,QAAQ;GAAO,aAAa;GAAG,OAAO;EAAE;EAC/E,IAAI;GACF,KAAK,MAAM,SAAS;IAAC;IAAS;IAAO;GAAM,GAAY;IACrD,MAAM,UAAU,OAAO,GAAG,MAAM;IAChC,IAAI,CAAC,SAAS;IACd,MAAM,WAAW,MAAM,KAAK,OAAO,KAAK,OAAO,SAAS,MAAM;IAC9D,OAAO,SAAS;IAChB,OAAO,eAAe,SAAS;IAC/B,MAAM,SAAS,CAAC,SAAS,mBAAmB,SAAS,aAAa;IAClE,IAAI,UAAU,UAAU,CAAC,QAAQ;KAC/B,MAAM,OAAO,KACX,SAAS,kBACL,GAAG,MAAM,mBAAmB,SAAS,OAAO,MAC5C,GAAG,MAAM,gBAAgB,SAAS,SAAS,IAC/C;MACE,UAAU,SAAS;MACnB,QAAQ,OAAO;KACjB,CACF;KACA,OAAO;IACT;GACF;GACA,MAAM,OAAO,OAAO;GACpB,OAAO,SAAS,CAAC,QAAS,CAAC,KAAK,mBAAmB,KAAK,aAAa;GACrE,OAAO,QACL,MAAM,eAAe,KAAA,KAAa,KAAK,aAAa,KAC/C,KAAK,eAAe,KAAK,KAAK,aAC/B,OAAO,SACL,IACA;GACR,MAAM,OAAO,IAAI;IACf,UAAU,MAAM;IAChB,YAAY,MAAM;IAClB,aAAa,MAAM;IACnB,QAAQ,OAAO,SAAS,OAAO;IAC/B,QAAQ,OAAO;GACjB,CAAyB;EAC3B,SAAS,KAAK;GACZ,MAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;GAC1D,MAAM;EACR;EACA,OAAO;CACT;AACF;;AAgEA,eAAsB,mBACpB,MACA,UAGI,CAAC,GACmB;CACxB,IAAI,EAAE,KAAK,WAAW,KAAK,OAAO,cAAc,KAAK,QAAQ,IAC3D,MAAM,IAAI,MAAM,+CAA+C;CACjE,MAAM,SAAS,YAAY,IAAI,CAC7B,YAAY,QAAQ,KAAK,QAAQ,GACjC,GAAI,QAAQ,SAAS,CAAC,QAAQ,MAAM,IAAI,CAAC,CAC3C,CAAC;CACD,MAAM,SAAwB;EAAE,MAAM;EAAO,YAAY,CAAC;EAAG,SAAS,CAAC;CAAE;CACzE,IAAI,sBAAsB;CAC1B,MAAM,cAA6C,CAAC;CACpD,MAAM,QAAsB;EAC1B;EACA,MAAM,GAAG,YAAY;GACnB,KAAK,MAAM,SAAS,YAClB,OAAO,WAAW,KAChB,YAAY,QACR,QACA;IACE,GAAG,MAAM;IACT,MAAM,MAAM;IACZ,QAAQ,MAAM,KAAK,SAAS;GAC9B,CACN;EACJ;EACA,QAAQ,WAAW,YAAY,KAAK,MAAM;EAC1C,MAAM,QAAQ,MAAM,KAAK;GACvB,MAAM,QAAQ,YAAY,IAAI;GAC9B,IAAI,WAA0B;IAAE;IAAM,YAAY;IAAG,SAAS;GAAK;GACnE,IAAI;IACF,IAAI,QAAQ,SAAS,WAAW,MAAM,QAAQ,QAAQ,MAAM,GAAG;SAC1D;KACH,MAAM,SAAU,MAAM,IAAI;KAC1B,IAAI,QAAQ,YAAY,OACtB,MAAM,IAAI,MAAM,OAAO,SAAS,2CAA2C;KAC7E,SAAS,aAAa,KAAK,MAAM,YAAY,IAAI,IAAI,KAAK;IAC5D;GACF,SAAS,OAAO;IACd,WAAW;KACT;KACA,YAAY,KAAK,MAAM,YAAY,IAAI,IAAI,KAAK;KAChD,SAAS;KACT,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;IAC9D;GACF;GACA,IAAI,CAAC,SAAS,SAAS;GACvB,OAAO,QAAQ,KAAK,QAAQ;GAC5B,OAAO;EACT;EACA,MAAM,SAAY,OAA2B;GAC3C,IAAI;GACJ,MAAM,SAAS,MAAM,MAAM,QAAQ,MAAM,MAAM,YAAY;IACzD,MAAM,QAAQ,MAAM,MAAM,OAAO,MAAM;IACvC,WAAW;IACX,MAAM,YAAY,MAAM,OAAO,KAAK,CAAC;IACrC,MAAM,WAAW,KAAK;IACtB,MAAM,MAAM,QAAQ,OAAO,MAAM;GACnC,CAAC;GACD,OAAO;IAAE;IAAU;GAAO;EAC5B;EACA,MAAM,KAAK,MAAM,OAAO,QAAQ;GAC9B,IAAI,EAAE,OAAO,YAAY,KAAK,OAAO,aAAa,IAChD,MAAM,IAAI,MAAM,+CAA+C;GACjE,MAAM,WAAW,YAAY,IAAI,IAAI,OAAO;GAC5C,OAAO,YAAY,IAAI,IAAI,UAAU;IACnC,OAAO,eAAe;IACtB,MAAM,QAAQ,MAAM,KAAK;IACzB,IAAI,MAAM,KAAK,GAAG,OAAO;IACzB,MAAMA,WACJ,KAAK,IAAI,OAAO,YAAY,KAAK,IAAI,GAAG,WAAW,YAAY,IAAI,CAAC,CAAC,GACrE,KAAA,GACA,EAAE,OAAO,CACX;GACF;EAEF;CACF;CACA,IAAI;EACF,OAAO,eAAe;EACtB,MAAM,QAAQ,MAAM,KAAK,MAAM,KAAK;EACpC,KAAK,MAAM,QAAQ,KAAK,SAAS,CAAC,GAAG;GACnC,OAAO,eAAe;GACtB,IAAI,KAAK,QAAQ,CAAC,KAAK,KAAK,KAAK,GAAG;GACpC,MAAM,iBAAiB;GACvB,MAAM,SAAS,MAAM,MAAM,QAAQ,KAAK,YAAY,KAAK,IAAI,OAAO,KAAK,CAAC;GAC1E,IAAI,KAAK,UAAU,sBAAsB;GACzC,MAAM,MAAM,GAAI,KAAK,SAAS,OAAO,MAAM,KAAK,CAAC,CAAE;EACrD;EACA,OAAO,eAAe;EACtB,MAAM,MAAM,GAAI,KAAK,eAAe,KAAK,KAAK,CAAC,CAAE;EACjD,OAAO,WAAW,KAAK,WAAW,KAAK;CACzC,SAAS,OAAO;EACd,OAAO,QAAQ,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;CACtE,UAAU;EACR,KAAK,MAAM,UAAU,YAAY,QAAQ,GAAG;GAC1C,MAAM,WAAW,MAAM,MAAM,QAAQ,WAAW,MAAM;GACtD,IAAI,CAAC,SAAS,SACZ,MAAM,MAAM;IAAE,MAAM;IAAW,QAAQ;IAAQ,SAAS,SAAS;GAAM,CAAC;EAC5E;CACF;CACA,OAAO,OACL,CAAC,OAAO,SACR,wBAAwB,KACxB,OAAO,WAAW,OAAO,MAAM,EAAE,QAAQ,EAAE,WAAW,MAAM;CAC9D,OAAO;AACT"}

package/dist/summary-report-BstMpOfy.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"summary-report-BstMpOfy.js","names":["avg","hi"],"sources":["../src/research-report-render.ts","../src/research-report-analysis.ts","../src/statistics/effect-sizes.ts","../src/summary-report-evidence.ts","../src/summary-report-figures.ts","../src/research-report.ts"],"sourcesContent":["/** Markdown research reports and their shared numeric text formatting. */\n\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport type {\n  ResearchReportCandidate,\n  ResearchReportMethodology,\n  ResearchReportRecommendation,\n} from './research-report'\nimport type {\n  GainDistributionFigureSpec,\n  ParetoFigureSpec,\n  SummaryTable,\n} from './summary-report-figures'\n\nexport function renderResearchMarkdown(report: {\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  pareto: ParetoFigureSpec\n  gains: GainDistributionFigureSpec[]\n  rope: { low: number; high: number } | null\n  runFingerprint: string\n  preregistrationHash: string | null\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n}): string {\n  const lines: string[] = []\n  lines.push(`# ${report.title}`)\n  lines.push('')\n  lines.push(`**Generated:** ${report.generatedAt}`)\n  lines.push(`**Primary split:** ${report.split}`)\n  lines.push(`**Comparator:** ${report.comparator ?? 'not configured'}`)\n  lines.push(`**ROPE:** ${report.rope ? formatRope(report.rope) : 'not configured'}`)\n  lines.push(`**Run fingerprint:** \\`${report.runFingerprint}\\``)\n  lines.push(\n    `**Preregistration:** ${report.preregistrationHash ? `\\`${report.preregistrationHash}\\`` : 'none'}`,\n  )\n  lines.push('')\n  lines.push('## Executive Summary')\n  lines.push('')\n  for (const item of report.executiveSummary) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Recommendation')\n  lines.push('')\n  lines.push(`**Decision:** ${report.recommendation.decision}`)\n  lines.push(`**Candidate:** ${report.recommendation.candidateId ?? 'N/A'}`)\n  lines.push('')\n  lines.push('### Rationale')\n  lines.push('')\n  for (const item of report.recommendation.rationale) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Risks')\n  lines.push('')\n  for (const item of report.recommendation.risks.length\n    ? report.recommendation.risks\n    : ['No material report-level risks detected.']) {\n    lines.push(`- ${item}`)\n  }\n  lines.push('')\n  lines.push('### Next Actions')\n  lines.push('')\n  for (const item of report.recommendation.nextActions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Candidate Decision Table')\n  lines.push('')\n  lines.push(\n    '| Candidate | Decision | Mean | Δ̄ | Pr(Δ>0) | q | dz | Paired N | Median Gain CI | MDE | Pareto | Gate |',\n  )\n  lines.push('|---|---|---:|---:|---:|---:|---:|---:|---|---:|---|---|')\n  for (const c of report.candidates) {\n    const delta = c.meanDeltaVsComparator === null ? '-' : signed(c.meanDeltaVsComparator)\n    const prGt = c.prGreaterThanZero === null ? '-' : c.prGreaterThanZero.toFixed(3)\n    const q = c.qValue !== null ? c.qValue.toFixed(4) : '-'\n    const d = c.cohensD !== null ? c.cohensD.toFixed(3) : '-'\n    const gain = c.gainCi ? `[${fmt(c.gainCi.low)}, ${fmt(c.gainCi.high)}]` : '-'\n    const mde = c.mde === null || !Number.isFinite(c.mde) ? '-' : fmt(c.mde)\n    lines.push(\n      `| ${c.candidateId} | ${c.decision} | ${fmt(c.mean)} | ${delta} | ${prGt} | ${q} | ${d} | ${c.pairedN} | ${gain} | ${mde} | ${c.onParetoFrontier ? 'yes' : 'no'} | ${c.gate ?? '-'} |`,\n    )\n  }\n  lines.push('')\n  lines.push('## Statistical Summary')\n  lines.push('')\n  lines.push(report.summary.markdown)\n  lines.push('')\n  lines.push('## Methodology')\n  lines.push('')\n  lines.push('### Assumptions')\n  lines.push('')\n  for (const item of report.methodology.assumptions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Methods')\n  lines.push('')\n  for (const item of report.methodology.methods) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Alternatives Considered')\n  lines.push('')\n  for (const item of report.methodology.alternatives) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### When NOT To Apply')\n  lines.push('')\n  for (const item of report.methodology.whenNotToApply) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Citations')\n  lines.push('')\n  for (const item of report.methodology.citations) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Chart Specs')\n  lines.push('')\n  lines.push(\n    'The report carries JSON chart specs for Pareto cost/quality and paired gain histograms.',\n  )\n  lines.push('')\n  lines.push('```json')\n  lines.push(JSON.stringify({ pareto: report.pareto, gains: report.gains }, null, 2))\n  lines.push('```')\n  if (report.failureClusters) {\n    lines.push('')\n    lines.push('## Failure Clusters')\n    lines.push('')\n    lines.push('| Failure Class | Runs | Scenarios | Tool | Example |')\n    lines.push('|---|---:|---:|---|---|')\n    for (const c of report.failureClusters.clusters.slice(0, 10)) {\n      lines.push(\n        `| ${c.failureClass} | ${c.runCount} | ${c.scenarioIds.length} | ${c.toolName ?? '-'} | ${escapePipes(c.exampleError ?? c.exampleRunId)} |`,\n      )\n    }\n  }\n  return lines.join('\\n')\n}\n\nfunction escapePipes(s: string): string {\n  return s.replace(/\\|/g, '\\\\|')\n}\n\nfunction signed(x: number): string {\n  return `${x >= 0 ? '+' : ''}${fmt(x)}`\n}\n\nexport function formatRope(rope: { low: number; high: number }): string {\n  return `[${fmt(rope.low)}, ${fmt(rope.high)}]`\n}\n\nexport function fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/** Evidence-based report verdicts, recommendations, and methodology. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport type {\n  ResearchReportCandidate,\n  ResearchReportDecision,\n  ResearchReportMethodology,\n  ResearchReportRecommendation,\n} from './research-report'\nimport { fmt, formatRope } from './research-report-render'\nimport type { pairedPosterior } from './summary-report-evidence'\nimport type { ParetoPoint, SummaryTableRow } from './summary-report-figures'\n\n/**\n * Hard floor below which a paired comparison is treated as uninformative\n * regardless of `minPairs`. Mirrors the lower limit on Wilcoxon signed-rank\n * exact tables; below this the test has no power to separate effect sizes.\n */\nexport const RESEARCH_REPORT_HARD_PAIR_FLOOR = 6\n\nexport function buildMethodology(ctx: {\n  split: 'search' | 'holdout'\n  comparator: string | null\n  fdr: number\n  minPairs: number\n  rope: { low: number; high: number } | null\n  confidence: number\n  mdePower: number\n  mdeAlpha: number\n}): ResearchReportMethodology {\n  const assumptions: string[] = [\n    'Pairs are matched by (experimentId, scenarioId, seed); input order does not affect the comparison.',\n    'Paired deltas are exchangeable conditional on the matched scenario — no mid-run distribution shift.',\n    `Decisions are pre-specified at fdr=${ctx.fdr}, minPairs=${ctx.minPairs}, confidence=${ctx.confidence}; deviating from these post-hoc invalidates the false-discovery control.`,\n  ]\n  if (ctx.rope) {\n    assumptions.push(\n      `The Region of Practical Equivalence ${formatRope(ctx.rope)} is supplied by the domain owner; equivalent verdicts are only meaningful if that range is treated as the standing definition of \"no material difference.\"`,\n    )\n  }\n  if (ctx.comparator === null) {\n    assumptions.push('No comparator was configured; this run is descriptive, not causal.')\n  }\n  const methods: string[] = [\n    \"Marginal scores summarised with BH-FDR-adjusted Wilcoxon signed-rank q-values and paired Cohen's dz via summaryTable.\",\n    'Paired evidence summarised with a frequentist bootstrap CI on the median delta and a Dirichlet-weight Bayesian bootstrap for Pr(Δ>0) and Pr(Δ∈ROPE) on the mean delta.',\n    `Minimum detectable effect reported per candidate at α=${ctx.mdeAlpha} (two-sided), power=${ctx.mdePower}, standardised by the observed paired-delta SD.`,\n    'Pareto frontier flagged as a separate axis (cost vs quality); a candidate can be on-frontier without winning the paired test.',\n    'Held-out gate decisions, when supplied, override the statistical verdict in the reject direction.',\n  ]\n  const alternatives: string[] = [\n    'Paired t-test rejected: not robust to the heavy-tailed score distributions common in agent benchmarks.',\n    'Unpaired Mann–Whitney rejected: matched scenarios make pairing free; unpaired throws away that variance reduction.',\n    'Sequential / always-valid inference (e-values, mSPRT) is the right tool for iterative sweeps and is out of scope for this single-look report — preregister and run once, or wrap this report in an alpha-spending schedule.',\n    'Hierarchical Bayesian shrinkage across many candidates is future work; the current ranking uses raw paired statistics.',\n  ]\n  const whenNotToApply: string[] = [\n    `Paired N below ${RESEARCH_REPORT_HARD_PAIR_FLOOR} on any candidate — the bootstrap CI is degenerate.`,\n    'Comparator chosen post-hoc by inspecting the same data; q-values are no longer false-discovery-controlled.',\n    'Scenarios not drawn under a stable preregistered protocol; the report can describe the data but cannot anchor a launch decision.',\n    'Score distributions with mid-run shift (judge model swap, rubric change, infra outage) — pair exchangeability is violated.',\n  ]\n  const citations: string[] = [\n    'Benjamini, Y. & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. JRSS B, 57(1), 289–300.',\n    'Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1(6), 80–83.',\n    'Efron, B. (1979). Bootstrap methods: another look at the jackknife. Annals of Statistics, 7(1), 1–26.',\n    'Rubin, D. B. (1981). The Bayesian bootstrap. Annals of Statistics, 9(1), 130–134.',\n    'Kruschke, J. K. (2018). Rejecting or accepting parameter values in Bayesian estimation. Advances in Methods and Practices in Psychological Science, 1(2), 270–280. (ROPE.)',\n  ]\n  return { assumptions, methods, alternatives, whenNotToApply, citations }\n}\n\nexport function classifyCandidate(\n  row: SummaryTableRow,\n  ctx: {\n    comparator: string | null\n    posterior: ReturnType<typeof pairedPosterior> | null\n    point?: ParetoPoint\n    gateDecision?: HeldOutGateDecision\n    split: 'search' | 'holdout'\n    fdr: number\n    minPairs: number\n    rope: { low: number; high: number } | null\n  },\n): { decision: ResearchReportDecision; reason: string } {\n  if (ctx.comparator && row.candidateId === ctx.comparator) {\n    return { decision: 'hold', reason: 'Comparator baseline.' }\n  }\n  if (ctx.gateDecision && !ctx.gateDecision.promote) {\n    const code = ctx.gateDecision.rejectionCode ?? 'rejected'\n    return {\n      decision: 'reject',\n      reason: `Held-out gate returned ${code}: ${ctx.gateDecision.reason}`,\n    }\n  }\n  if (!ctx.comparator) {\n    return {\n      decision: ctx.point?.onFrontier ? 'hold' : 'needs_more_data',\n      reason:\n        'No comparator configured; report ranks candidates but cannot anchor a promotion call.',\n    }\n  }\n  if (!ctx.posterior || ctx.posterior.n < RESEARCH_REPORT_HARD_PAIR_FLOOR) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior?.n ?? 0} paired observations; below hard floor of ${RESEARCH_REPORT_HARD_PAIR_FLOOR} for any paired inference.`,\n    }\n  }\n  const ci = ctx.posterior.ci\n  if (ctx.split === 'search') {\n    return {\n      decision: 'hold',\n      reason:\n        'Search-split evidence is descriptive; release decisions require matched holdout runs.',\n    }\n  }\n  // A ZERO-WIDTH interval cannot support any directional or equivalence claim:\n  // every resample was identical, so it says nothing about how far the estimate\n  // could be wrong. It is the shape that sits fully inside any ROPE containing\n  // it (a fake \"practically equivalent\") and clears any threshold below it (a\n  // fake \"promote\"). Both are an absence of evidence, so it is disqualified from\n  // each claim rather than short-circuiting the whole classifier — the\n  // sample-size and MDE diagnostics below are still the more useful answer when\n  // they apply.\n  const degenerate = !Number.isFinite(ci.low) || !Number.isFinite(ci.high) || ci.low === ci.high\n  if (!degenerate && ctx.rope && ci.low >= ctx.rope.low && ci.high <= ctx.rope.high) {\n    return {\n      decision: 'equivalent',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] is fully inside ROPE ${formatRope(ctx.rope)}; candidate is practically equivalent to comparator.`,\n    }\n  }\n  const gainPositive = !degenerate && ci.low > 0\n  const gainNegative = !degenerate && ci.high < 0\n  if (gainNegative) {\n    return {\n      decision: 'reject',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] lies entirely below zero.`,\n    }\n  }\n  if (ctx.posterior.n < ctx.minPairs) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior.n} paired observations; minimum detectable effect at this N is ${fmt(ctx.posterior.mde)} score units (need ≥ ${ctx.minPairs} pairs to issue a directional verdict).`,\n    }\n  }\n  if (degenerate) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] has zero width — every resample was identical, so the interval carries no information about how far the estimate could be wrong and cannot support a promote or equivalence call.`,\n    }\n  }\n  if (row.qValue !== null && row.qValue <= ctx.fdr && gainPositive) {\n    return {\n      decision: 'promote',\n      reason: `BH-adjusted q=${fmt(row.qValue)} ≤ ${ctx.fdr} and paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] excludes zero; Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)}.`,\n    }\n  }\n  return {\n    decision: 'hold',\n    reason: `Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)} but CI [${fmt(ci.low)}, ${fmt(ci.high)}] crosses zero; effect not decisive at fdr=${ctx.fdr}.`,\n  }\n}\n\nexport function buildRecommendation(\n  candidates: ResearchReportCandidate[],\n  ctx: {\n    comparator: string | null\n    failureClusters?: FailureClusterReport\n    rope: { low: number; high: number } | null\n    minPairs: number\n    preregistrationHash: string | null\n  },\n): ResearchReportRecommendation {\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  const bestPromote = nonComparator.find((c) => c.decision === 'promote')\n  const bestEquivalent = nonComparator.find((c) => c.decision === 'equivalent')\n  const chosen = bestPromote ?? bestEquivalent ?? nonComparator[0] ?? null\n  const decision: ResearchReportDecision = bestPromote\n    ? 'promote'\n    : nonComparator.some((c) => c.decision === 'needs_more_data')\n      ? 'needs_more_data'\n      : bestEquivalent\n        ? 'equivalent'\n        : nonComparator.some((c) => c.decision === 'hold')\n          ? 'hold'\n          : 'reject'\n\n  const rationale: string[] = []\n  const risks: string[] = []\n  const nextActions: string[] = []\n\n  if (chosen) {\n    rationale.push(`${chosen.candidateId}: ${chosen.decisionReason}`)\n    if (chosen.gainCi) {\n      const probSummary =\n        chosen.prGreaterThanZero !== null ? `, Pr(Δ>0)=${fmt(chosen.prGreaterThanZero)}` : ''\n      rationale.push(\n        `Median paired gain CI: [${fmt(chosen.gainCi.low)}, ${fmt(chosen.gainCi.high)}]${probSummary}.`,\n      )\n    }\n    if (chosen.mde !== null && Number.isFinite(chosen.mde)) {\n      rationale.push(`MDE at current paired N=${chosen.pairedN}: ${fmt(chosen.mde)} score units.`)\n    }\n  }\n  if (!ctx.comparator) {\n    risks.push('No comparator was configured; verdict is descriptive, not causal.')\n    nextActions.push('Re-run with a stable comparator candidate for paired inference.')\n  }\n  if (!ctx.preregistrationHash) {\n    risks.push(\n      'No preregistration hash supplied; readers cannot verify the analysis was specified before data inspection.',\n    )\n    nextActions.push(\n      'Sign a HypothesisManifest before the next sweep and pass `preregistrationHash` so the report cites it.',\n    )\n  }\n  if (ctx.rope === null && nonComparator.length > 0) {\n    risks.push(\n      'No ROPE configured; the report cannot distinguish \"equivalent\" from \"inconclusive\".',\n    )\n    nextActions.push(\n      'Define a domain-specific Region of Practical Equivalence and pass it to lock in the equivalence threshold.',\n    )\n  }\n  const inconclusive = nonComparator.filter((c) => c.decision === 'needs_more_data')\n  if (inconclusive.length > 0) {\n    const worst = inconclusive.reduce((a, b) => (b.pairedN < a.pairedN ? b : a))\n    risks.push(\n      `${inconclusive.length} candidate(s) below soft floor (${ctx.minPairs} pairs); thinnest is ${worst.candidateId} with ${worst.pairedN}.`,\n    )\n    nextActions.push(\n      `Collect at least ${ctx.minPairs - worst.pairedN} more matched holdout runs for ${worst.candidateId}.`,\n    )\n  }\n  const rejected = nonComparator.filter((c) => c.decision === 'reject')\n  if (rejected.length > 0) {\n    risks.push(\n      `${rejected.length} candidate(s) failed the paired test or held-out gate; do not ship those variants.`,\n    )\n  }\n  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {\n    const top = ctx.failureClusters.clusters[0]!\n    risks.push(`Top failure cluster: ${top.failureClass} across ${top.runCount} run(s).`)\n    nextActions.push('Prioritize the largest failure cluster before broad rollout.')\n  }\n  if (decision === 'promote') {\n    nextActions.push('Ship behind the existing promotion gate and monitor canaries.')\n  } else if (decision === 'hold') {\n    nextActions.push('Keep current production candidate while expanding holdout evidence.')\n  } else if (decision === 'equivalent') {\n    nextActions.push(\n      'Either keep the comparator (no quality regression) or promote on cost/latency grounds — equivalence does not justify either; the choice is a product decision, not a stats one.',\n    )\n  } else if (decision === 'reject') {\n    nextActions.push(\n      'Do not promote this sweep; inspect failures and generate a revised candidate.',\n    )\n  }\n\n  return {\n    decision,\n    candidateId: chosen?.candidateId ?? null,\n    rationale,\n    risks,\n    nextActions,\n  }\n}\n\nexport function buildExecutiveSummary(\n  candidates: ResearchReportCandidate[],\n  recommendation: ResearchReportRecommendation,\n  ctx: {\n    comparator: string | null\n    split: 'search' | 'holdout'\n    failureClusters?: FailureClusterReport\n    preregistrationHash: string | null\n  },\n): string[] {\n  const lines: string[] = []\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  lines.push(\n    `Evaluated ${nonComparator.length} candidate(s) on the ${ctx.split} split${ctx.comparator ? ` against ${ctx.comparator}` : ''}.`,\n  )\n  lines.push(\n    `Recommendation: ${recommendation.decision}${recommendation.candidateId ? ` ${recommendation.candidateId}` : ''}.`,\n  )\n  const promoted = nonComparator.filter((c) => c.decision === 'promote').length\n  const held = nonComparator.filter((c) => c.decision === 'hold').length\n  const equivalent = nonComparator.filter((c) => c.decision === 'equivalent').length\n  const rejected = nonComparator.filter((c) => c.decision === 'reject').length\n  const more = nonComparator.filter((c) => c.decision === 'needs_more_data').length\n  lines.push(\n    `Decision mix: ${promoted} promote, ${equivalent} equivalent, ${held} hold, ${rejected} reject, ${more} need more data.`,\n  )\n  const frontier = nonComparator.filter((c) => c.onParetoFrontier).map((c) => c.candidateId)\n  if (frontier.length > 0) lines.push(`Pareto-frontier candidates: ${frontier.join(', ')}.`)\n  if (ctx.failureClusters) {\n    lines.push(\n      `Failure clustering found ${ctx.failureClusters.totalFailures}/${ctx.failureClusters.totalRuns} failed runs across ${ctx.failureClusters.clusters.length} reportable cluster(s).`,\n    )\n  }\n  lines.push(\n    ctx.preregistrationHash\n      ? `Preregistered analysis: ${ctx.preregistrationHash.slice(0, 12)}…`\n      : 'Analysis is post-hoc — no preregistration hash supplied.',\n  )\n  return lines\n}\n\nexport function decisionWeight(decision: ResearchReportDecision): number {\n  if (decision === 'promote') return 5\n  if (decision === 'equivalent') return 4\n  if (decision === 'hold') return 3\n  if (decision === 'needs_more_data') return 2\n  return 1\n}\n","/**\n * Standardized effect sizes: Cohen's d family and Cliff's delta, with the\n * shared degenerate-input contract (null where the effect is undefined).\n */\n\nimport { ValidationError } from '../errors'\n\n/**\n * Cohen's d — standardized effect size for two independent groups.\n * Positive d means group b has higher mean than group a.\n * Rule of thumb: |d| < 0.2 negligible, 0.2–0.5 small, 0.5–0.8 medium, > 0.8 large.\n *\n * Returns null where the standardized effect is undefined: fewer than two\n * observations in either group, or a zero pooled standard deviation with\n * unequal means. Null is NOT \"no effect\" — zero within-group spread across a\n * real mean gap is an unbounded effect, the opposite of negligible. Equal\n * means with zero spread is a genuine 0. Same contract as\n * {@link pairedCohensDz}.\n */\nexport function cohensD(a: number[], b: number[]): number | null {\n  if (a.length < 2 || b.length < 2) return null\n  const meanA = a.reduce((x, y) => x + y, 0) / a.length\n  const meanB = b.reduce((x, y) => x + y, 0) / b.length\n  const varA = a.reduce((acc, x) => acc + (x - meanA) ** 2, 0) / (a.length - 1)\n  const varB = b.reduce((acc, x) => acc + (x - meanB) ** 2, 0) / (b.length - 1)\n  const pooled = Math.sqrt(\n    ((a.length - 1) * varA + (b.length - 1) * varB) / (a.length + b.length - 2),\n  )\n  if (pooled === 0) return meanB === meanA ? 0 : null\n  return (meanB - meanA) / pooled\n}\n\n/**\n * Cohen's dz for paired observations: mean(after - before) divided by the\n * sample standard deviation of those within-pair deltas.\n *\n * Returns null when fewer than two pairs exist or a non-zero constant delta\n * has zero observed variance. In that case the standardized effect is\n * undefined, not an arbitrarily large finite number.\n */\nexport function pairedCohensDz(before: number[], after: number[]): number | null {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedCohensDz: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  if (before.length < 2) return null\n  const deltas = before.map((value, index) => after[index]! - value)\n  if (deltas.some((value) => !Number.isFinite(value))) {\n    throw new ValidationError('pairedCohensDz: all paired values must be finite')\n  }\n  const meanDelta = deltas.reduce((sum, value) => sum + value, 0) / deltas.length\n  const variance =\n    deltas.reduce((sum, value) => sum + (value - meanDelta) ** 2, 0) / (deltas.length - 1)\n  const standardDeviation = Math.sqrt(variance)\n  const scale = Math.max(1, Math.abs(meanDelta), ...deltas.map(Math.abs))\n  if (standardDeviation <= Number.EPSILON * scale) return meanDelta === 0 ? 0 : null\n  return meanDelta / standardDeviation\n}\n\nexport type CliffsMagnitude = 'negligible' | 'small' | 'medium' | 'large'\n\n/**\n * Cliff's delta — a non-parametric effect size for two independent samples.\n * `δ = (#(after > before) − #(after < before)) / (n_before · n_after)`,\n * ranging [-1, 1]. Positive ⇒ `after` tends to exceed `before` (improvement).\n *\n * Distribution-free counterpart to Cohen's d: no normality assumption, robust\n * to the bounded/skewed score distributions judges produce. Pairs with\n * `pairedBootstrap` / `wilcoxonSignedRank` for the non-parametric reporting\n * path. Returns 0 when either sample is empty.\n */\nexport function cliffsDelta(before: number[], after: number[]): number {\n  const n = before.length * after.length\n  if (n === 0) return 0\n  let dominance = 0\n  for (const a of after) {\n    for (const b of before) {\n      if (a > b) dominance += 1\n      else if (a < b) dominance -= 1\n    }\n  }\n  return dominance / n\n}\n\n/**\n * Map a Cliff's delta to a qualitative magnitude using the standard\n * Romano et al. thresholds (|δ|): <0.147 negligible, <0.33 small,\n * <0.474 medium, else large.\n */\nexport function interpretCliffs(delta: number): CliffsMagnitude {\n  const d = Math.abs(delta)\n  if (d < 0.147) return 'negligible'\n  if (d < 0.33) return 'small'\n  if (d < 0.474) return 'medium'\n  return 'large'\n}\n","/** Matched split scores and paired statistical evidence for report consumers. */\n\nimport { interpolatedQuantile } from './math/quantile'\nimport { pairRunRecords } from './paired-arms'\nimport { observedSplitScore, type ScorePreference } from './rollout/reward'\nimport type { RunRecord } from './run-record'\nimport { pairedBootstrap, pairedMde } from './statistics'\nimport { makeRng } from './statistics/internal'\n\nexport function pairScoresByKey(\n  candidate: RunRecord[],\n  baseline: RunRecord[],\n  split: ScorePreference,\n): {\n  before: number[]\n  after: number[]\n  unpairedCandidateN: number\n  unpairedComparatorN: number\n} {\n  // RAW, split-exact reads through `observedSplitScore` — the single allowed\n  // spelling of the split-score derivation (see rollout/score-derivation-guard).\n  const scoreOf = (run: RunRecord): number | undefined => {\n    const v = observedSplitScore(run, split)\n    return typeof v === 'number' && Number.isFinite(v) ? v : undefined\n  }\n  const scoredCandidate = candidate.filter((run) => scoreOf(run) !== undefined)\n  const scoredBaseline = baseline.filter((run) => scoreOf(run) !== undefined)\n  const pairing = pairRunRecords(scoredBaseline, scoredCandidate)\n  return {\n    before: pairing.pairs.map((pair) => scoreOf(pair.baseline) as number),\n    after: pairing.pairs.map((pair) => scoreOf(pair.treatment) as number),\n    unpairedCandidateN: pairing.unpairedTreatment.length,\n    unpairedComparatorN: pairing.unpairedBaseline.length,\n  }\n}\n\n/**\n * Internal: paired posterior summary on (candidate − comparator) deltas.\n *\n * Returns the bootstrap CI on the median (matching `gainHistogram`) plus\n * Bayesian-bootstrap posterior summaries Pr(Δ>0) and Pr(Δ∈ROPE) computed\n * from Dirichlet(1, ..., 1) weights on the observed paired deltas\n * (Rubin 1981), and the\n * minimum detectable paired effect at the configured power and α.\n *\n * `null` is returned when no paired observations exist; callers must\n * gate on `n` before consuming the bootstrap statistics.\n */\nexport function pairedPosterior(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: {\n    split: 'search' | 'holdout'\n    confidence: number\n    seed?: number\n    rope: { low: number; high: number } | null\n    mdePower: number\n    mdeAlpha: number\n  },\n): {\n  n: number\n  meanDelta: number\n  medianDelta: number\n  sdDelta: number\n  ci: { low: number; high: number }\n  prGreaterThanZero: number\n  prInRope: number | null\n  mde: number\n} | null {\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === opts.split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === opts.split)\n  const { before, after } = pairScoresByKey(candidate, baseline, opts.split)\n  const n = before.length\n  if (n === 0) return null\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const meanDelta = deltas.reduce((s, x) => s + x, 0) / n\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const medianDelta = interpolatedQuantile(sortedDeltas, 0.5)\n  const sdDelta = stdev(deltas, meanDelta)\n\n  const ci = pairedBootstrap(before, after, {\n    confidence: opts.confidence,\n    resamples: 2000,\n    statistic: 'median',\n    seed: opts.seed,\n  })\n\n  const meanSamples = bayesianBootstrapMeanSamples(deltas, 2000, opts.seed)\n  const prGreaterThanZero =\n    meanSamples.length === 0 ? 0 : meanSamples.filter((s) => s > 0).length / meanSamples.length\n  const prInRope =\n    opts.rope === null || meanSamples.length === 0\n      ? null\n      : meanSamples.filter((s) => s >= opts.rope!.low && s <= opts.rope!.high).length /\n        meanSamples.length\n\n  const dStandardised = pairedMde({ nPaired: n, alpha: opts.mdeAlpha, power: opts.mdePower })\n  const mde = sdDelta === 0 ? 0 : dStandardised * sdDelta\n\n  return {\n    n,\n    meanDelta,\n    medianDelta,\n    sdDelta,\n    ci: { low: ci.low, high: ci.high },\n    prGreaterThanZero,\n    prInRope,\n    mde,\n  }\n}\n\nfunction bayesianBootstrapMeanSamples(\n  deltas: number[],\n  resamples: number,\n  seed?: number,\n): number[] {\n  const n = deltas.length\n  if (n === 0) return []\n  if (n === 1) return new Array<number>(resamples).fill(deltas[0]!)\n  const rng = makeRng(seed, deltas)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let weightedSum = 0\n    let weightSum = 0\n    for (let k = 0; k < n; k++) {\n      const weight = -Math.log(Math.max(Number.MIN_VALUE, rng()))\n      weightedSum += deltas[k]! * weight\n      weightSum += weight\n    }\n    samples[b] = weightedSum / weightSum\n  }\n  return samples\n}\n\nfunction stdev(xs: number[], mean: number): number {\n  if (xs.length < 2) return 0\n  let sse = 0\n  for (const x of xs) sse += (x - mean) ** 2\n  return Math.sqrt(sse / (xs.length - 1))\n}\n","/** Tables and library-independent chart specifications from scored runs. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport { interpolatedQuantile } from './math/quantile'\nimport { fmt } from './research-report-render'\nimport { observedSplitScore } from './rollout/reward'\nimport type { RunRecord } from './run-record'\nimport {\n  benjaminiHochberg,\n  confidenceInterval,\n  pairedBootstrap,\n  pairedCohensDz,\n  wilcoxonSignedRank,\n} from './statistics'\nimport { mean as avg } from './statistics/mean'\nimport { pairScoresByKey } from './summary-report-evidence'\n\n// ── summaryTable ───────────────────────────────────────────────────────\n\nexport interface SummaryTableOptions {\n  /** Comparator candidate id. Wilcoxon + paired Cohen's dz are computed\n   *  versus this candidate. Required for paired stats columns. */\n  comparator?: string\n  /** Which split to read scores from. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the bootstrap CI on the mean. Default 0.95. */\n  confidence?: number\n  /** FDR for BH adjustment of the comparison p-values. Default 0.05. */\n  fdr?: number\n}\n\nexport interface SummaryTableRow {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  /** BH-adjusted q-value vs comparator, or null when unavailable. */\n  qValue: number | null\n  /** Paired Cohen's dz vs comparator, or null when the paired variance is zero. */\n  cohensD: number | null\n  /** Matched observations used for paired comparison, or null on the comparator row. */\n  pairedN: number | null\n  /** Candidate observations without a comparator match. */\n  unpairedCandidateN: number | null\n  /** Comparator observations without a candidate match. */\n  unpairedComparatorN: number | null\n}\n\nexport interface SummaryTable {\n  rows: SummaryTableRow[]\n  comparator: string | null\n  split: 'search' | 'holdout'\n  /** Pre-rendered markdown — drop into a paper or PR. */\n  markdown: string\n}\n\n/**\n * Table 1 helper. Buckets runs by `candidateId`, computes mean +\n * bootstrap CI on the chosen split, and (when a comparator is given)\n * BH-adjusted Wilcoxon p + paired Cohen's dz versus that comparator.\n */\nexport function summaryTable(runs: RunRecord[], opts: SummaryTableOptions = {}): SummaryTable {\n  const split = opts.split ?? 'holdout'\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const comparator = opts.comparator ?? null\n\n  const byCandidate = new Map<string, { runs: RunRecord[]; scores: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    // RAW (`observedSplitScore`): every figure in this module reports what the\n    // runs scored. Zeroing a gamed run here would erase it from the very table\n    // a reviewer reads to notice it, and the split is already filtered above,\n    // so the exact-split reader is the right one — no cross-split fallback.\n    const v = observedSplitScore(r, split)\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    const bucket = byCandidate.get(r.candidateId) ?? { runs: [], scores: [] }\n    bucket.runs.push(r)\n    bucket.scores.push(v)\n    byCandidate.set(r.candidateId, bucket)\n  }\n\n  const candidateIds = [...byCandidate.keys()].sort()\n  const compRuns = comparator ? byCandidate.get(comparator) : undefined\n\n  // First pass: per-candidate means + CIs + raw p-values.\n  const tentative: Array<SummaryTableRow & { rawP: number | null }> = []\n  for (const id of candidateIds) {\n    const bucket = byCandidate.get(id)!\n    const ci = confidenceInterval(bucket.scores, confidence)\n    let rawP: number | null = null\n    let d: number | null = null\n    let pairedN: number | null = null\n    let unpairedCandidateN: number | null = null\n    let unpairedComparatorN: number | null = null\n    if (comparator && compRuns && id !== comparator) {\n      const paired = pairScoresByKey(bucket.runs, compRuns.runs, split)\n      pairedN = paired.before.length\n      unpairedCandidateN = paired.unpairedCandidateN\n      unpairedComparatorN = paired.unpairedComparatorN\n      if (paired.before.length >= 6) {\n        rawP = wilcoxonSignedRank(paired.before, paired.after).p\n      }\n      d = pairedCohensDz(paired.before, paired.after)\n    }\n    tentative.push({\n      candidateId: id,\n      n: bucket.scores.length,\n      mean: ci.mean,\n      ciLow: ci.lower,\n      ciHigh: ci.upper,\n      qValue: rawP,\n      cohensD: d,\n      pairedN,\n      unpairedCandidateN,\n      unpairedComparatorN,\n      rawP,\n    })\n  }\n\n  // BH-adjust across the comparison set (skip NaN rows / the\n  // comparator itself). Adjustment is a no-op when there are 0 or 1\n  // comparators.\n  if (comparator) {\n    const idxs: number[] = []\n    const ps: number[] = []\n    for (let i = 0; i < tentative.length; i++) {\n      const r = tentative[i]!\n      if (r.candidateId === comparator) continue\n      if (r.rawP === null || !Number.isFinite(r.rawP)) continue\n      idxs.push(i)\n      ps.push(r.rawP)\n    }\n    if (ps.length > 0) {\n      const { qValues } = benjaminiHochberg(ps, fdr)\n      for (let k = 0; k < idxs.length; k++) {\n        tentative[idxs[k]!]!.qValue = qValues[k]!\n      }\n    }\n  }\n\n  const rows = tentative.map(({ rawP: _rawP, ...rest }) => rest)\n  const markdown = renderSummaryTableMarkdown(rows, comparator, split)\n  return { rows, comparator, split, markdown }\n}\n\nfunction renderSummaryTableMarkdown(\n  rows: SummaryTableRow[],\n  comparator: string | null,\n  split: 'search' | 'holdout',\n): string {\n  const lines: string[] = []\n  const cmpLabel = comparator ? ` (vs ${comparator})` : ''\n  lines.push(`Summary Table — ${split} split${cmpLabel}`)\n  lines.push('')\n  lines.push(\"| Candidate | N | Pairs | Unpaired | Mean | 95% CI | q (BH) | Cohen's dz |\")\n  lines.push('|---|---:|---:|---:|---:|---|---:|---:|')\n  for (const r of rows) {\n    const ci = `[${fmt(r.ciLow)}, ${fmt(r.ciHigh)}]`\n    const q = r.qValue !== null ? r.qValue.toFixed(4) : '—'\n    const d = r.cohensD !== null ? r.cohensD.toFixed(3) : '—'\n    const paired = r.pairedN ?? '—'\n    const unpaired =\n      r.unpairedCandidateN === null || r.unpairedComparatorN === null\n        ? '—'\n        : `${r.unpairedCandidateN}/${r.unpairedComparatorN}`\n    lines.push(\n      `| ${r.candidateId} | ${r.n} | ${paired} | ${unpaired} | ${fmt(r.mean)} | ${ci} | ${q} | ${d} |`,\n    )\n  }\n  return lines.join('\\n')\n}\n\n// ── paretoChart ─────────────────────────────────────────────────────\n\nexport interface ParetoPoint {\n  candidateId: string\n  /** Mean USD cost per run on the chosen split. */\n  cost: number\n  /** Mean score on the chosen split. */\n  quality: number\n  /** Number of runs that informed this point. */\n  n: number\n  /** Whether this candidate is on the Pareto frontier — high\n   *  quality, low cost, no dominator. */\n  onFrontier: boolean\n  /** Optional gate verdict for this candidate, if a `HeldOutGateDecision`\n   *  for it was passed in. */\n  gate?: 'promote' | 'reject'\n}\n\nexport interface ParetoFigureSpec {\n  kind: 'pareto-cost-quality'\n  split: 'search' | 'holdout'\n  points: ParetoPoint[]\n  axes: { x: 'costUsd'; y: 'score' }\n}\n\n/**\n * Cost vs quality scatter spec. `gateDecisions` is keyed by\n * candidate id; if present, every point picks up the gate verdict\n * for overlay.\n */\nexport function paretoChart(\n  runs: RunRecord[],\n  opts: {\n    split?: 'search' | 'holdout'\n    gateDecisions?: Record<string, HeldOutGateDecision>\n  } = {},\n): ParetoFigureSpec {\n  const split = opts.split ?? 'holdout'\n\n  const buckets = new Map<string, { cost: number[]; quality: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    const v = observedSplitScore(r, split)\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    if (r.costUsd === null) continue\n    const bucket = buckets.get(r.candidateId) ?? { cost: [], quality: [] }\n    bucket.cost.push(r.costUsd)\n    bucket.quality.push(v)\n    buckets.set(r.candidateId, bucket)\n  }\n\n  const points: ParetoPoint[] = []\n  for (const [candidateId, bucket] of buckets.entries()) {\n    points.push({\n      candidateId,\n      cost: avg(bucket.cost),\n      quality: avg(bucket.quality),\n      n: bucket.cost.length,\n      onFrontier: false,\n      gate: opts.gateDecisions?.[candidateId]\n        ? gateLabel(opts.gateDecisions[candidateId]!)\n        : undefined,\n    })\n  }\n\n  // Pareto: minimize cost, maximize quality. A point is dominated if\n  // some other point has lower-or-equal cost AND higher-or-equal\n  // quality with strict inequality somewhere.\n  for (const p of points) {\n    p.onFrontier = !points.some((q) => q !== p && dominates(q, p))\n  }\n\n  return {\n    kind: 'pareto-cost-quality',\n    split,\n    axes: { x: 'costUsd', y: 'score' },\n    points,\n  }\n}\n\nfunction dominates(a: ParetoPoint, b: ParetoPoint): boolean {\n  return a.cost <= b.cost && a.quality >= b.quality && (a.cost < b.cost || a.quality > b.quality)\n}\n\nexport function gateLabel(d: HeldOutGateDecision): ParetoPoint['gate'] {\n  return d.promote ? 'promote' : 'reject'\n}\n\n// ── gainHistogram ───────────────────────────────────────────\n\nexport interface GainDistributionBin {\n  /** Inclusive lower edge. */\n  lo: number\n  /** Exclusive upper edge (or inclusive if it's the last bin). */\n  hi: number\n  /** Number of pairs whose delta lands in this bin. */\n  count: number\n}\n\nexport interface GainDistributionFigureSpec {\n  kind: 'gain-distribution'\n  candidateId: string\n  comparator: string\n  split: 'search' | 'holdout'\n  /** Number of pairs used. */\n  n: number\n  /** Candidate rows without a comparator match. */\n  unpairedCandidateN: number\n  /** Comparator rows without a candidate match. */\n  unpairedComparatorN: number\n  bins: GainDistributionBin[]\n  median: number | null\n  ci: { low: number; high: number } | null\n}\n\nexport interface GainDistributionOptions {\n  /** Number of histogram bins. Default 11 (so the centre is exact at 0). */\n  bins?: number\n  /** Which split to use. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the CI. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Deterministic seed. */\n  seed?: number\n}\n\n/**\n * Held-out improvement distribution: per-pair delta (candidate −\n * comparator), histogrammed. Includes the bootstrap CI on the median\n * delta — same primitive the promotion gate uses.\n */\nexport function gainHistogram(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: GainDistributionOptions = {},\n): GainDistributionFigureSpec {\n  const split = opts.split ?? 'holdout'\n  const binCount = opts.bins ?? 11\n  if (binCount < 1) throw new Error('gainHistogram: bins must be ≥ 1')\n\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === split)\n  // pairScoresByKey returns before=baseline-score, after=candidate-score\n  // for each canonical work-identity pair where both sides recorded a\n  // valid score on this split. delta = after - before = candidate - baseline.\n  const { before, after, unpairedCandidateN, unpairedComparatorN } = pairScoresByKey(\n    candidate,\n    baseline,\n    split,\n  )\n  const n = before.length\n\n  if (n === 0) {\n    return {\n      kind: 'gain-distribution',\n      candidateId,\n      comparator,\n      split,\n      n: 0,\n      unpairedCandidateN,\n      unpairedComparatorN,\n      bins: [],\n      median: null,\n      ci: null,\n    }\n  }\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const median = interpolatedQuantile(sortedDeltas, 0.5)\n  const min = sortedDeltas[0]!\n  const max = sortedDeltas[sortedDeltas.length - 1]!\n\n  // Symmetric bins around the wider of (|min|, |max|) so the chart\n  // visually centres on zero without dropping outliers.\n  const bound = Math.max(Math.abs(min), Math.abs(max), 1e-6)\n  const lo = -bound\n  const hi = bound\n  const width = (hi - lo) / binCount\n  const bins: GainDistributionBin[] = []\n  for (let i = 0; i < binCount; i++) {\n    bins.push({ lo: lo + i * width, hi: lo + (i + 1) * width, count: 0 })\n  }\n  for (const d of deltas) {\n    let idx = Math.floor((d - lo) / width)\n    if (idx < 0) idx = 0\n    if (idx >= binCount) idx = binCount - 1\n    bins[idx]!.count += 1\n  }\n\n  const ci = pairedBootstrap(before, after, {\n    confidence: opts.confidence ?? 0.95,\n    resamples: opts.resamples ?? 2000,\n    statistic: 'median',\n    seed: opts.seed,\n  })\n\n  return {\n    kind: 'gain-distribution',\n    candidateId,\n    comparator,\n    split,\n    n,\n    unpairedCandidateN,\n    unpairedComparatorN,\n    bins,\n    median,\n    ci: { low: ci.low, high: ci.high },\n  }\n}\n","/** Assemble research report evidence, guidance, figures, and rendered documents. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport { hashJson } from './pre-registration'\nimport { renderResearchHtml } from './reporting-html'\nimport {\n  buildExecutiveSummary,\n  buildMethodology,\n  buildRecommendation,\n  classifyCandidate,\n  decisionWeight,\n  RESEARCH_REPORT_HARD_PAIR_FLOOR,\n} from './research-report-analysis'\nimport { renderResearchMarkdown } from './research-report-render'\nimport type { RunRecord } from './run-record'\nimport { pairedPosterior } from './summary-report-evidence'\nimport {\n  type GainDistributionFigureSpec,\n  gainHistogram,\n  gateLabel,\n  type ParetoFigureSpec,\n  type ParetoPoint,\n  paretoChart,\n  type SummaryTable,\n  summaryTable,\n} from './summary-report-figures'\n\nexport { RESEARCH_REPORT_HARD_PAIR_FLOOR } from './research-report-analysis'\n\nexport type ResearchReportDecision =\n  | 'promote'\n  | 'hold'\n  | 'reject'\n  | 'equivalent'\n  | 'needs_more_data'\n\nexport interface ResearchReportOptions {\n  /** Human-readable report title. */\n  title?: string\n  /** Comparator candidate id. Required for statistical decision guidance. */\n  comparator?: string\n  /** Which split to use for the primary decision. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level used by lower-level report helpers. Default 0.95. */\n  confidence?: number\n  /** FDR threshold for q-values. Default 0.05. */\n  fdr?: number\n  /**\n   * Soft floor on paired observations before issuing a directional\n   * promote / reject. Below this we report `needs_more_data` and surface the\n   * minimum detectable effect at the current N. Default 20 — chosen so the\n   * Wilcoxon signed-rank approximation is reasonable and so the paired\n   * bootstrap CI has non-degenerate coverage. Hard floor is enforced at\n   * `RESEARCH_REPORT_HARD_PAIR_FLOOR` (6) regardless of this value.\n   */\n  minPairs?: number\n  /**\n   * Region of Practical Equivalence on the paired delta. When a candidate's\n   * paired-delta CI is fully contained in `[low, high]`, the decision is\n   * `equivalent` rather than `hold`. Sourced from the domain owner — there is\n   * no statistically-defensible default.\n   */\n  rope?: { low: number; high: number }\n  /**\n   * Power for the minimum detectable effect (MDE) reported on each candidate.\n   * Default 0.8.\n   */\n  mdePower?: number\n  /**\n   * Two-sided alpha for the MDE. Default matches `fdr` so the reported MDE\n   * lines up with the test the report actually runs.\n   */\n  mdeAlpha?: number\n  /** Optional held-out gate decisions keyed by candidate id. */\n  gateDecisions?: Record<string, HeldOutGateDecision>\n  /** Optional failure clusters from failureClusterView. */\n  failureClusters?: FailureClusterReport\n  /** Build gain histograms for these candidates. Defaults to all non-comparator candidates. */\n  candidateIds?: string[]\n  /** Deterministic bootstrap seed passed to gainHistogram and the posterior helper. */\n  seed?: number\n  /** Report timestamp. Defaults to current time. */\n  generatedAt?: string\n  /**\n   * Hash of a preregistered protocol (e.g. `signManifest({...}).contentHash`).\n   * Embedded verbatim in the report so the analysis can be cited as the\n   * preregistered one rather than a post-hoc fishing expedition.\n   */\n  preregistrationHash?: string\n}\n\nexport interface ResearchReportRecommendation {\n  decision: ResearchReportDecision\n  candidateId: string | null\n  rationale: string[]\n  risks: string[]\n  nextActions: string[]\n}\n\nexport interface ResearchReportCandidate {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  qValue: number | null\n  cohensD: number | null\n  meanDeltaVsComparator: number | null\n  pairedN: number\n  medianGain: number | null\n  meanGain: number | null\n  gainCi: { low: number; high: number } | null\n  /**\n   * Bayesian-bootstrap posterior summaries on the paired mean delta.\n   * Dirichlet(1, ..., 1) weights represent uncertainty over the empirical\n   * distribution of matched deltas.\n   */\n  prGreaterThanZero: number | null\n  prInRope: number | null\n  /**\n   * Minimum detectable effect (in score units) at the candidate's paired N,\n   * the configured power, and the configured alpha. Standardised by the\n   * observed paired-delta SD and inverted via `requiredSampleSize`. Reported\n   * for every candidate so a `needs_more_data` verdict is actionable.\n   */\n  mde: number | null\n  onParetoFrontier: boolean\n  gate?: ParetoPoint['gate']\n  decision: ResearchReportDecision\n  decisionReason: string\n}\n\nexport interface ResearchReportMethodology {\n  /**\n   * Plain-language assumptions the report depends on. Read these first when\n   * deciding whether the verdict is load-bearing for a launch decision.\n   */\n  assumptions: string[]\n  /** Tests and estimators the verdict was computed from. */\n  methods: string[]\n  /** Alternatives the author considered and why this report didn't take them. */\n  alternatives: string[]\n  /** Failure modes — when this report should NOT drive a decision. */\n  whenNotToApply: string[]\n  /** Citations for the methodological choices above. */\n  citations: string[]\n}\n\nexport interface ResearchReport {\n  kind: 'agent-eval-research-report'\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  /**\n   * SHA-256 over the canonicalised set of `(runId, candidateId, split)` triples\n   * the report was computed from, plus the comparator and split. Stable across\n   * key insertion order; recomputable by the reader to verify provenance.\n   */\n  runFingerprint: string\n  preregistrationHash: string | null\n  rope: { low: number; high: number } | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  charts: {\n    pareto: ParetoFigureSpec\n    gains: GainDistributionFigureSpec[]\n  }\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n  markdown: string\n  html: string\n}\n\n/**\n * Executive research report for CPO / AI-lead / launch-review consumption.\n *\n * Composes:\n *   - `summaryTable`         marginal stats with BH-FDR-adjusted q-values\n *   - `paretoChart`           cost-vs-quality frontier with gate overlay\n *   - `gainHistogram`         per-candidate paired-delta distribution\n *   - paired posterior (this file): bootstrap CI on median, Bayesian-bootstrap Pr(Δ>0),\n *                              Pr(Δ∈ROPE), MDE at the configured power\n *\n * Decisions are made on paired evidence — never on marginal means alone —\n * and respect any held-out gate decision the caller passes through. The\n * report embeds a SHA-256 fingerprint of the input run set and, optionally,\n * the hash of a preregistered protocol so a downstream reader can verify\n * provenance and that the analysis was the preregistered one.\n *\n * Async because the fingerprint uses Web Crypto via `hashJson`; deterministic\n * for any fixed `runs`, `seed`, and ROPE.\n */\nexport async function researchReport(\n  runs: RunRecord[],\n  opts: ResearchReportOptions = {},\n): Promise<ResearchReport> {\n  const split = opts.split ?? 'holdout'\n  const comparator = opts.comparator ?? null\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const minPairs = Math.max(opts.minPairs ?? 20, RESEARCH_REPORT_HARD_PAIR_FLOOR)\n  const rope = opts.rope ?? null\n  const mdePower = opts.mdePower ?? 0.8\n  const mdeAlpha = opts.mdeAlpha ?? fdr\n  const title = opts.title ?? 'Agent Evaluation Research Report'\n  const generatedAt = opts.generatedAt ?? new Date().toISOString()\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  if (rope && !(Number.isFinite(rope.low) && Number.isFinite(rope.high) && rope.low <= rope.high)) {\n    throw new Error(\n      `researchReport: rope must satisfy low ≤ high with finite bounds, got ${JSON.stringify(rope)}`,\n    )\n  }\n\n  const summary = summaryTable(runs, {\n    comparator: comparator ?? undefined,\n    split,\n    confidence,\n    fdr,\n  })\n  const pareto = paretoChart(runs, { split, gateDecisions: opts.gateDecisions })\n  const candidateIds =\n    opts.candidateIds ?? summary.rows.map((r) => r.candidateId).filter((id) => id !== comparator)\n  const gains = comparator\n    ? candidateIds.map((id) =>\n        gainHistogram(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n        }),\n      )\n    : []\n\n  const gainByCandidate = new Map(gains.map((g) => [g.candidateId, g]))\n  const paretoByCandidate = new Map(pareto.points.map((p) => [p.candidateId, p]))\n  const posteriorByCandidate = new Map<string, ReturnType<typeof pairedPosterior>>()\n  if (comparator) {\n    for (const id of candidateIds) {\n      posteriorByCandidate.set(\n        id,\n        pairedPosterior(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n          rope,\n          mdePower,\n          mdeAlpha,\n        }),\n      )\n    }\n  }\n\n  const candidates = summary.rows\n    .map((row) => {\n      const gain = gainByCandidate.get(row.candidateId)\n      const point = paretoByCandidate.get(row.candidateId)\n      const posterior = posteriorByCandidate.get(row.candidateId) ?? null\n      const gateDecision = opts.gateDecisions?.[row.candidateId]\n      const classified = classifyCandidate(row, {\n        comparator,\n        posterior,\n        point,\n        gateDecision,\n        split,\n        fdr,\n        minPairs,\n        rope,\n      })\n      return {\n        candidateId: row.candidateId,\n        n: row.n,\n        mean: row.mean,\n        ciLow: row.ciLow,\n        ciHigh: row.ciHigh,\n        qValue: row.qValue,\n        cohensD: row.cohensD,\n        meanDeltaVsComparator: posterior ? posterior.meanDelta : null,\n        pairedN: posterior?.n ?? gain?.n ?? 0,\n        medianGain: posterior ? posterior.medianDelta : gain ? gain.median : null,\n        meanGain: posterior ? posterior.meanDelta : null,\n        gainCi: posterior ? posterior.ci : gain ? gain.ci : null,\n        prGreaterThanZero: posterior ? posterior.prGreaterThanZero : null,\n        prInRope: posterior ? posterior.prInRope : null,\n        mde: posterior ? posterior.mde : null,\n        onParetoFrontier: point?.onFrontier ?? false,\n        gate: gateDecision ? gateLabel(gateDecision) : undefined,\n        decision: classified.decision,\n        decisionReason: classified.reason,\n      } satisfies ResearchReportCandidate\n    })\n    .sort((a, b) => {\n      const decisionRank = decisionWeight(b.decision) - decisionWeight(a.decision)\n      if (decisionRank !== 0) return decisionRank\n      return b.mean - a.mean\n    })\n\n  const recommendation = buildRecommendation(candidates, {\n    comparator,\n    failureClusters: opts.failureClusters,\n    rope,\n    minPairs,\n    preregistrationHash,\n  })\n  const executiveSummary = buildExecutiveSummary(candidates, recommendation, {\n    comparator,\n    split,\n    failureClusters: opts.failureClusters,\n    preregistrationHash,\n  })\n  const methodology = buildMethodology({\n    split,\n    comparator,\n    fdr,\n    minPairs,\n    rope,\n    confidence,\n    mdePower,\n    mdeAlpha,\n  })\n\n  const runFingerprint = await hashJson({\n    triples: runs\n      .filter((r) => r.splitTag === split)\n      .map((r) => ({ runId: r.runId, candidateId: r.candidateId, splitTag: r.splitTag }))\n      .sort((a, b) => a.runId.localeCompare(b.runId)),\n    comparator,\n    split,\n  })\n\n  const markdown = renderResearchMarkdown({\n    title,\n    generatedAt,\n    split,\n    comparator,\n    rope,\n    runFingerprint,\n    preregistrationHash,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    pareto,\n    gains,\n    methodology,\n    failureClusters: opts.failureClusters,\n  })\n  const html = renderResearchHtml(markdown, title)\n\n  return {\n    kind: 'agent-eval-research-report',\n    title,\n    generatedAt,\n    split,\n    comparator,\n    runFingerprint,\n    preregistrationHash,\n    rope,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    charts: { pareto, gains },\n    methodology,\n    failureClusters: opts.failureClusters,\n    markdown,\n    html,\n  }\n}\n"],"mappings":";;;;;;;;;;;;AAcA,SAAgB,uBAAuB,QAgB5B;CACT,MAAM,QAAkB,CAAC;CACzB,MAAM,KAAK,KAAK,OAAO,OAAO;CAC9B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,kBAAkB,OAAO,aAAa;CACjD,MAAM,KAAK,sBAAsB,OAAO,OAAO;CAC/C,MAAM,KAAK,mBAAmB,OAAO,cAAc,kBAAkB;CACrE,MAAM,KAAK,aAAa,OAAO,OAAO,WAAW,OAAO,IAAI,IAAI,kBAAkB;CAClF,MAAM,KAAK,0BAA0B,OAAO,eAAe,GAAG;CAC9D,MAAM,KACJ,wBAAwB,OAAO,sBAAsB,KAAK,OAAO,oBAAoB,MAAM,QAC7F;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,sBAAsB;CACjC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,kBAAkB,MAAM,KAAK,KAAK,MAAM;CAClE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,mBAAmB;CAC9B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,iBAAiB,OAAO,eAAe,UAAU;CAC5D,MAAM,KAAK,kBAAkB,OAAO,eAAe,eAAe,OAAO;CACzE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,eAAe;CAC1B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,WAAW,MAAM,KAAK,KAAK,MAAM;CAC1E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,WAAW;CACtB,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,MAAM,SAC3C,OAAO,eAAe,QACtB,CAAC,0CAA0C,GAC7C,MAAM,KAAK,KAAK,MAAM;CAExB,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,kBAAkB;CAC7B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,aAAa,MAAM,KAAK,KAAK,MAAM;CAC5E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,6BAA6B;CACxC,MAAM,KAAK,EAAE;CACb,MAAM,KACJ,2GACF;CACA,MAAM,KAAK,0DAA0D;CACrE,KAAK,MAAM,KAAK,OAAO,YAAY;EACjC,MAAM,QAAQ,EAAE,0BAA0B,OAAO,MAAM,OAAO,EAAE,qBAAqB;EACrF,MAAM,OAAO,EAAE,sBAAsB,OAAO,MAAM,EAAE,kBAAkB,QAAQ,CAAC;EAC/E,MAAM,IAAI,EAAE,WAAW,OAAO,EAAE,OAAO,QAAQ,CAAC,IAAI;EACpD,MAAM,IAAI,EAAE,YAAY,OAAO,EAAE,QAAQ,QAAQ,CAAC,IAAI;EACtD,MAAM,OAAO,EAAE,SAAS,IAAI,IAAI,EAAE,OAAO,GAAG,EAAE,IAAI,IAAI,EAAE,OAAO,IAAI,EAAE,KAAK;EAC1E,MAAM,MAAM,EAAE,QAAQ,QAAQ,CAAC,OAAO,SAAS,EAAE,GAAG,IAAI,MAAM,IAAI,EAAE,GAAG;EACvE,MAAM,KACJ,KAAK,EAAE,YAAY,KAAK,EAAE,SAAS,KAAK,IAAI,EAAE,IAAI,EAAE,KAAK,MAAM,KAAK,KAAK,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,EAAE,mBAAmB,QAAQ,KAAK,KAAK,EAAE,QAAQ,IAAI,GACrL;CACF;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,wBAAwB;CACnC,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,OAAO,QAAQ,QAAQ;CAClC,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,gBAAgB;CAC3B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,iBAAiB;CAC5B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,aAAa,MAAM,KAAK,KAAK,MAAM;CACzE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,aAAa;CACxB,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,SAAS,MAAM,KAAK,KAAK,MAAM;CACrE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,6BAA6B;CACxC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,cAAc,MAAM,KAAK,KAAK,MAAM;CAC1E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,uBAAuB;CAClC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,gBAAgB,MAAM,KAAK,KAAK,MAAM;CAC5E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,eAAe;CAC1B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,WAAW,MAAM,KAAK,KAAK,MAAM;CACvE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,gBAAgB;CAC3B,MAAM,KAAK,EAAE;CACb,MAAM,KACJ,yFACF;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,SAAS;CACpB,MAAM,KAAK,KAAK,UAAU;EAAE,QAAQ,OAAO;EAAQ,OAAO,OAAO;CAAM,GAAG,MAAM,CAAC,CAAC;CAClF,MAAM,KAAK,KAAK;CAChB,IAAI,OAAO,iBAAiB;EAC1B,MAAM,KAAK,EAAE;EACb,MAAM,KAAK,qBAAqB;EAChC,MAAM,KAAK,EAAE;EACb,MAAM,KAAK,uDAAuD;EAClE,MAAM,KAAK,yBAAyB;EACpC,KAAK,MAAM,KAAK,OAAO,gBAAgB,SAAS,MAAM,GAAG,EAAE,GACzD,MAAM,KACJ,KAAK,EAAE,aAAa,KAAK,EAAE,SAAS,KAAK,EAAE,YAAY,OAAO,KAAK,EAAE,YAAY,IAAI,KAAK,YAAY,EAAE,gBAAgB,EAAE,YAAY,EAAE,GAC1I;CAEJ;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,YAAY,GAAmB;CACtC,OAAO,EAAE,QAAQ,OAAO,KAAK;AAC/B;AAEA,SAAS,OAAO,GAAmB;CACjC,OAAO,GAAG,KAAK,IAAI,MAAM,KAAK,IAAI,CAAC;AACrC;AAEA,SAAgB,WAAW,MAA6C;CACtE,OAAO,IAAI,IAAI,KAAK,GAAG,EAAE,IAAI,IAAI,KAAK,IAAI,EAAE;AAC9C;AAEA,SAAgB,IAAI,GAAmB;CACrC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO,OAAO,CAAC;CACxC,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;ACpIA,MAAa,kCAAkC;AAE/C,SAAgB,iBAAiB,KASH;CAC5B,MAAM,cAAwB;EAC5B;EACA;EACA,sCAAsC,IAAI,IAAI,aAAa,IAAI,SAAS,eAAe,IAAI,WAAW;CACxG;CACA,IAAI,IAAI,MACN,YAAY,KACV,uCAAuC,WAAW,IAAI,IAAI,EAAE,2JAC9D;CAEF,IAAI,IAAI,eAAe,MACrB,YAAY,KAAK,oEAAoE;CA4BvF,OAAO;EAAE;EAAa,SAAA;GAzBpB;GACA;GACA,yDAAyD,IAAI,SAAS,sBAAsB,IAAI,SAAS;GACzG;GACA;EAqB0B;EAAG,cAAA;GAlB7B;GACA;GACA;GACA;EAewC;EAAG,gBAAA;GAZ3C;GACA;GACA;GACA;EASwD;EAAG,WAAA;GAN3D;GACA;GACA;GACA;GACA;EAEmE;CAAE;AACzE;AAEA,SAAgB,kBACd,KACA,KAUsD;CACtD,IAAI,IAAI,cAAc,IAAI,gBAAgB,IAAI,YAC5C,OAAO;EAAE,UAAU;EAAQ,QAAQ;CAAuB;CAE5D,IAAI,IAAI,gBAAgB,CAAC,IAAI,aAAa,SAExC,OAAO;EACL,UAAU;EACV,QAAQ,0BAHG,IAAI,aAAa,iBAAiB,WAGN,IAAI,IAAI,aAAa;CAC9D;CAEF,IAAI,CAAC,IAAI,YACP,OAAO;EACL,UAAU,IAAI,OAAO,aAAa,SAAS;EAC3C,QACE;CACJ;CAEF,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,IAAA,GAClC,OAAO;EACL,UAAU;EACV,QAAQ,QAAQ,IAAI,WAAW,KAAK,EAAE;CACxC;CAEF,MAAM,KAAK,IAAI,UAAU;CACzB,IAAI,IAAI,UAAU,UAChB,OAAO;EACL,UAAU;EACV,QACE;CACJ;CAUF,MAAM,aAAa,CAAC,OAAO,SAAS,GAAG,GAAG,KAAK,CAAC,OAAO,SAAS,GAAG,IAAI,KAAK,GAAG,QAAQ,GAAG;CAC1F,IAAI,CAAC,cAAc,IAAI,QAAQ,GAAG,OAAO,IAAI,KAAK,OAAO,GAAG,QAAQ,IAAI,KAAK,MAC3E,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,yBAAyB,WAAW,IAAI,IAAI,EAAE;CACzG;CAEF,MAAM,eAAe,CAAC,cAAc,GAAG,MAAM;CAE7C,IADqB,CAAC,cAAc,GAAG,OAAO,GAE5C,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE;CAC3D;CAEF,IAAI,IAAI,UAAU,IAAI,IAAI,UACxB,OAAO;EACL,UAAU;EACV,QAAQ,QAAQ,IAAI,UAAU,EAAE,+DAA+D,IAAI,IAAI,UAAU,GAAG,EAAE,uBAAuB,IAAI,SAAS;CAC5J;CAEF,IAAI,YACF,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE;CAC3D;CAEF,IAAI,IAAI,WAAW,QAAQ,IAAI,UAAU,IAAI,OAAO,cAClD,OAAO;EACL,UAAU;EACV,QAAQ,iBAAiB,IAAI,IAAI,MAAM,EAAE,KAAK,IAAI,IAAI,wBAAwB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,2BAA2B,IAAI,IAAI,UAAU,iBAAiB,EAAE;CAC7K;CAEF,OAAO;EACL,UAAU;EACV,QAAQ,WAAW,IAAI,IAAI,UAAU,iBAAiB,EAAE,WAAW,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,6CAA6C,IAAI,IAAI;CACvJ;AACF;AAEA,SAAgB,oBACd,YACA,KAO8B;CAC9B,MAAM,gBAAgB,WAAW,QAAQ,MAAM,EAAE,gBAAgB,IAAI,UAAU;CAC/E,MAAM,cAAc,cAAc,MAAM,MAAM,EAAE,aAAa,SAAS;CACtE,MAAM,iBAAiB,cAAc,MAAM,MAAM,EAAE,aAAa,YAAY;CAC5E,MAAM,SAAS,eAAe,kBAAkB,cAAc,MAAM;CACpE,MAAM,WAAmC,cACrC,YACA,cAAc,MAAM,MAAM,EAAE,aAAa,iBAAiB,IACxD,oBACA,iBACE,eACA,cAAc,MAAM,MAAM,EAAE,aAAa,MAAM,IAC7C,SACA;CAEV,MAAM,YAAsB,CAAC;CAC7B,MAAM,QAAkB,CAAC;CACzB,MAAM,cAAwB,CAAC;CAE/B,IAAI,QAAQ;EACV,UAAU,KAAK,GAAG,OAAO,YAAY,IAAI,OAAO,gBAAgB;EAChE,IAAI,OAAO,QAAQ;GACjB,MAAM,cACJ,OAAO,sBAAsB,OAAO,aAAa,IAAI,OAAO,iBAAiB,MAAM;GACrF,UAAU,KACR,2BAA2B,IAAI,OAAO,OAAO,GAAG,EAAE,IAAI,IAAI,OAAO,OAAO,IAAI,EAAE,GAAG,YAAY,EAC/F;EACF;EACA,IAAI,OAAO,QAAQ,QAAQ,OAAO,SAAS,OAAO,GAAG,GACnD,UAAU,KAAK,2BAA2B,OAAO,QAAQ,IAAI,IAAI,OAAO,GAAG,EAAE,cAAc;CAE/F;CACA,IAAI,CAAC,IAAI,YAAY;EACnB,MAAM,KAAK,mEAAmE;EAC9E,YAAY,KAAK,iEAAiE;CACpF;CACA,IAAI,CAAC,IAAI,qBAAqB;EAC5B,MAAM,KACJ,4GACF;EACA,YAAY,KACV,wGACF;CACF;CACA,IAAI,IAAI,SAAS,QAAQ,cAAc,SAAS,GAAG;EACjD,MAAM,KACJ,yFACF;EACA,YAAY,KACV,4GACF;CACF;CACA,MAAM,eAAe,cAAc,QAAQ,MAAM,EAAE,aAAa,iBAAiB;CACjF,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,QAAQ,aAAa,QAAQ,GAAG,MAAO,EAAE,UAAU,EAAE,UAAU,IAAI,CAAE;EAC3E,MAAM,KACJ,GAAG,aAAa,OAAO,kCAAkC,IAAI,SAAS,uBAAuB,MAAM,YAAY,QAAQ,MAAM,QAAQ,EACvI;EACA,YAAY,KACV,oBAAoB,IAAI,WAAW,MAAM,QAAQ,iCAAiC,MAAM,YAAY,EACtG;CACF;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,QAAQ;CACpE,IAAI,SAAS,SAAS,GACpB,MAAM,KACJ,GAAG,SAAS,OAAO,mFACrB;CAEF,IAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,SAAS,GAAG;EAClE,MAAM,MAAM,IAAI,gBAAgB,SAAS;EACzC,MAAM,KAAK,wBAAwB,IAAI,aAAa,UAAU,IAAI,SAAS,SAAS;EACpF,YAAY,KAAK,8DAA8D;CACjF;CACA,IAAI,aAAa,WACf,YAAY,KAAK,+DAA+D;MAC3E,IAAI,aAAa,QACtB,YAAY,KAAK,qEAAqE;MACjF,IAAI,aAAa,cACtB,YAAY,KACV,iLACF;MACK,IAAI,aAAa,UACtB,YAAY,KACV,+EACF;CAGF,OAAO;EACL;EACA,aAAa,QAAQ,eAAe;EACpC;EACA;EACA;CACF;AACF;AAEA,SAAgB,sBACd,YACA,gBACA,KAMU;CACV,MAAM,QAAkB,CAAC;CACzB,MAAM,gBAAgB,WAAW,QAAQ,MAAM,EAAE,gBAAgB,IAAI,UAAU;CAC/E,MAAM,KACJ,aAAa,cAAc,OAAO,uBAAuB,IAAI,MAAM,QAAQ,IAAI,aAAa,YAAY,IAAI,eAAe,GAAG,EAChI;CACA,MAAM,KACJ,mBAAmB,eAAe,WAAW,eAAe,cAAc,IAAI,eAAe,gBAAgB,GAAG,EAClH;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACvE,MAAM,OAAO,cAAc,QAAQ,MAAM,EAAE,aAAa,MAAM,CAAC,CAAC;CAChE,MAAM,aAAa,cAAc,QAAQ,MAAM,EAAE,aAAa,YAAY,CAAC,CAAC;CAC5E,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,OAAO,cAAc,QAAQ,MAAM,EAAE,aAAa,iBAAiB,CAAC,CAAC;CAC3E,MAAM,KACJ,iBAAiB,SAAS,YAAY,WAAW,eAAe,KAAK,SAAS,SAAS,WAAW,KAAK,iBACzG;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,gBAAgB,CAAC,CAAC,KAAK,MAAM,EAAE,WAAW;CACzF,IAAI,SAAS,SAAS,GAAG,MAAM,KAAK,+BAA+B,SAAS,KAAK,IAAI,EAAE,EAAE;CACzF,IAAI,IAAI,iBACN,MAAM,KACJ,4BAA4B,IAAI,gBAAgB,cAAc,GAAG,IAAI,gBAAgB,UAAU,sBAAsB,IAAI,gBAAgB,SAAS,OAAO,wBAC3J;CAEF,MAAM,KACJ,IAAI,sBACA,2BAA2B,IAAI,oBAAoB,MAAM,GAAG,EAAE,EAAE,KAChE,0DACN;CACA,OAAO;AACT;AAEA,SAAgB,eAAe,UAA0C;CACvE,IAAI,aAAa,WAAW,OAAO;CACnC,IAAI,aAAa,cAAc,OAAO;CACtC,IAAI,aAAa,QAAQ,OAAO;CAChC,IAAI,aAAa,mBAAmB,OAAO;CAC3C,OAAO;AACT;;;;;;;;;;;;;;;;;;ACzSA,SAAgB,QAAQ,GAAa,GAA4B;CAC/D,IAAI,EAAE,SAAS,KAAK,EAAE,SAAS,GAAG,OAAO;CACzC,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;CAC/C,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;CAC/C,MAAM,OAAO,EAAE,QAAQ,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;CAC3E,MAAM,OAAO,EAAE,QAAQ,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;CAC3E,MAAM,SAAS,KAAK,OAChB,EAAE,SAAS,KAAK,QAAQ,EAAE,SAAS,KAAK,SAAS,EAAE,SAAS,EAAE,SAAS,EAC3E;CACA,IAAI,WAAW,GAAG,OAAO,UAAU,QAAQ,IAAI;CAC/C,QAAQ,QAAQ,SAAS;AAC3B;;;;;;;;;AAUA,SAAgB,eAAe,QAAkB,OAAgC;CAC/E,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,yCAAyC,OAAO,OAAO,MAAM,MAAM,OAAO,EAC5E;CAEF,IAAI,OAAO,SAAS,GAAG,OAAO;CAC9B,MAAM,SAAS,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,KAAK;CACjE,IAAI,OAAO,MAAM,UAAU,CAAC,OAAO,SAAS,KAAK,CAAC,GAChD,MAAM,IAAI,gBAAgB,kDAAkD;CAE9E,MAAM,YAAY,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;CACzE,MAAM,WACJ,OAAO,QAAQ,KAAK,UAAU,OAAO,QAAQ,cAAc,GAAG,CAAC,KAAK,OAAO,SAAS;CACtF,MAAM,oBAAoB,KAAK,KAAK,QAAQ;CAC5C,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,SAAS,GAAG,GAAG,OAAO,IAAI,KAAK,GAAG,CAAC;CACtE,IAAI,qBAAqB,OAAO,UAAU,OAAO,OAAO,cAAc,IAAI,IAAI;CAC9E,OAAO,YAAY;AACrB;;;;;;;;;;;AAcA,SAAgB,YAAY,QAAkB,OAAyB;CACrE,MAAM,IAAI,OAAO,SAAS,MAAM;CAChC,IAAI,MAAM,GAAG,OAAO;CACpB,IAAI,YAAY;CAChB,KAAK,MAAM,KAAK,OACd,KAAK,MAAM,KAAK,QACd,IAAI,IAAI,GAAG,aAAa;MACnB,IAAI,IAAI,GAAG,aAAa;CAGjC,OAAO,YAAY;AACrB;;;;;;AAOA,SAAgB,gBAAgB,OAAgC;CAC9D,MAAM,IAAI,KAAK,IAAI,KAAK;CACxB,IAAI,IAAI,MAAO,OAAO;CACtB,IAAI,IAAI,KAAM,OAAO;CACrB,IAAI,IAAI,MAAO,OAAO;CACtB,OAAO;AACT;;;ACvFA,SAAgB,gBACd,WACA,UACA,OAMA;CAGA,MAAM,WAAW,QAAuC;EACtD,MAAM,IAAI,mBAAmB,KAAK,KAAK;EACvC,OAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI,KAAA;CAC3D;CACA,MAAM,kBAAkB,UAAU,QAAQ,QAAQ,QAAQ,GAAG,MAAM,KAAA,CAAS;CAC5E,MAAM,iBAAiB,SAAS,QAAQ,QAAQ,QAAQ,GAAG,MAAM,KAAA,CAAS;CAC1E,MAAM,UAAU,eAAe,gBAAgB,eAAe;CAC9D,OAAO;EACL,QAAQ,QAAQ,MAAM,KAAK,SAAS,QAAQ,KAAK,QAAQ,CAAW;EACpE,OAAO,QAAQ,MAAM,KAAK,SAAS,QAAQ,KAAK,SAAS,CAAW;EACpE,oBAAoB,QAAQ,kBAAkB;EAC9C,qBAAqB,QAAQ,iBAAiB;CAChD;AACF;;;;;;;;;;;;;AAcA,SAAgB,gBACd,MACA,aACA,YACA,MAiBO;CAGP,MAAM,EAAE,QAAQ,UAAU,gBAFR,KAAK,QAAQ,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAAK,KAExC,GADjC,KAAK,QAAQ,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAAK,KAC5B,GAAG,KAAK,KAAK;CACzE,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO;CAEpB,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,YAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CACtD,MAAM,eAAe,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACrD,MAAM,cAAc,qBAAqB,cAAc,EAAG;CAC1D,MAAM,UAAU,MAAM,QAAQ,SAAS;CAEvC,MAAM,KAAK,gBAAgB,QAAQ,OAAO;EACxC,YAAY,KAAK;EACjB,WAAW;EACX,WAAW;EACX,MAAM,KAAK;CACb,CAAC;CAED,MAAM,cAAc,6BAA6B,QAAQ,KAAM,KAAK,IAAI;CACxE,MAAM,oBACJ,YAAY,WAAW,IAAI,IAAI,YAAY,QAAQ,MAAM,IAAI,CAAC,CAAC,CAAC,SAAS,YAAY;CACvF,MAAM,WACJ,KAAK,SAAS,QAAQ,YAAY,WAAW,IACzC,OACA,YAAY,QAAQ,MAAM,KAAK,KAAK,KAAM,OAAO,KAAK,KAAK,KAAM,IAAI,CAAC,CAAC,SACvE,YAAY;CAElB,MAAM,gBAAgB,UAAU;EAAE,SAAS;EAAG,OAAO,KAAK;EAAU,OAAO,KAAK;CAAS,CAAC;CAC1F,MAAM,MAAM,YAAY,IAAI,IAAI,gBAAgB;CAEhD,OAAO;EACL;EACA;EACA;EACA;EACA,IAAI;GAAE,KAAK,GAAG;GAAK,MAAM,GAAG;EAAK;EACjC;EACA;EACA;CACF;AACF;AAEA,SAAS,6BACP,QACA,WACA,MACU;CACV,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO,CAAC;CACrB,IAAI,MAAM,GAAG,OAAO,IAAI,MAAc,SAAS,CAAC,CAAC,KAAK,OAAO,EAAG;CAChE,MAAM,MAAM,QAAQ,MAAM,MAAM;CAChC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,IAAI,cAAc;EAClB,IAAI,YAAY;EAChB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,OAAO,WAAW,IAAI,CAAC,CAAC;GAC1D,eAAe,OAAO,KAAM;GAC5B,aAAa;EACf;EACA,QAAQ,KAAK,cAAc;CAC7B;CACA,OAAO;AACT;AAEA,SAAS,MAAM,IAAc,MAAsB;CACjD,IAAI,GAAG,SAAS,GAAG,OAAO;CAC1B,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,IAAI,QAAQ,IAAI,SAAS;CACzC,OAAO,KAAK,KAAK,OAAO,GAAG,SAAS,EAAE;AACxC;;;;;;;AC/EA,SAAgB,aAAa,MAAmB,OAA4B,CAAC,GAAiB;CAC5F,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,aAAa,KAAK,cAAc;CAEtC,MAAM,8BAAc,IAAI,IAAqD;CAC7E,KAAK,MAAM,KAAK,MAAM;EACpB,IAAI,EAAE,aAAa,OAAO;EAK1B,MAAM,IAAI,mBAAmB,GAAG,KAAK;EACrC,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,MAAM,SAAS,YAAY,IAAI,EAAE,WAAW,KAAK;GAAE,MAAM,CAAC;GAAG,QAAQ,CAAC;EAAE;EACxE,OAAO,KAAK,KAAK,CAAC;EAClB,OAAO,OAAO,KAAK,CAAC;EACpB,YAAY,IAAI,EAAE,aAAa,MAAM;CACvC;CAEA,MAAM,eAAe,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;CAClD,MAAM,WAAW,aAAa,YAAY,IAAI,UAAU,IAAI,KAAA;CAG5D,MAAM,YAA8D,CAAC;CACrE,KAAK,MAAM,MAAM,cAAc;EAC7B,MAAM,SAAS,YAAY,IAAI,EAAE;EACjC,MAAM,KAAK,mBAAmB,OAAO,QAAQ,UAAU;EACvD,IAAI,OAAsB;EAC1B,IAAI,IAAmB;EACvB,IAAI,UAAyB;EAC7B,IAAI,qBAAoC;EACxC,IAAI,sBAAqC;EACzC,IAAI,cAAc,YAAY,OAAO,YAAY;GAC/C,MAAM,SAAS,gBAAgB,OAAO,MAAM,SAAS,MAAM,KAAK;GAChE,UAAU,OAAO,OAAO;GACxB,qBAAqB,OAAO;GAC5B,sBAAsB,OAAO;GAC7B,IAAI,OAAO,OAAO,UAAU,GAC1B,OAAO,mBAAmB,OAAO,QAAQ,OAAO,KAAK,CAAC,CAAC;GAEzD,IAAI,eAAe,OAAO,QAAQ,OAAO,KAAK;EAChD;EACA,UAAU,KAAK;GACb,aAAa;GACb,GAAG,OAAO,OAAO;GACjB,MAAM,GAAG;GACT,OAAO,GAAG;GACV,QAAQ,GAAG;GACX,QAAQ;GACR,SAAS;GACT;GACA;GACA;GACA;EACF,CAAC;CACH;CAKA,IAAI,YAAY;EACd,MAAM,OAAiB,CAAC;EACxB,MAAM,KAAe,CAAC;EACtB,KAAK,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;GACzC,MAAM,IAAI,UAAU;GACpB,IAAI,EAAE,gBAAgB,YAAY;GAClC,IAAI,EAAE,SAAS,QAAQ,CAAC,OAAO,SAAS,EAAE,IAAI,GAAG;GACjD,KAAK,KAAK,CAAC;GACX,GAAG,KAAK,EAAE,IAAI;EAChB;EACA,IAAI,GAAG,SAAS,GAAG;GACjB,MAAM,EAAE,YAAY,kBAAkB,IAAI,GAAG;GAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAC/B,UAAU,KAAK,GAAI,CAAE,SAAS,QAAQ;EAE1C;CACF;CAEA,MAAM,OAAO,UAAU,KAAK,EAAE,MAAM,OAAO,GAAG,WAAW,IAAI;CAE7D,OAAO;EAAE;EAAM;EAAY;EAAO,UADjB,2BAA2B,MAAM,YAAY,KACrB;CAAE;AAC7C;AAEA,SAAS,2BACP,MACA,YACA,OACQ;CACR,MAAM,QAAkB,CAAC;CACzB,MAAM,WAAW,aAAa,QAAQ,WAAW,KAAK;CACtD,MAAM,KAAK,mBAAmB,MAAM,QAAQ,UAAU;CACtD,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,4EAA4E;CACvF,MAAM,KAAK,yCAAyC;CACpD,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,KAAK,IAAI,IAAI,EAAE,KAAK,EAAE,IAAI,IAAI,EAAE,MAAM,EAAE;EAC9C,MAAM,IAAI,EAAE,WAAW,OAAO,EAAE,OAAO,QAAQ,CAAC,IAAI;EACpD,MAAM,IAAI,EAAE,YAAY,OAAO,EAAE,QAAQ,QAAQ,CAAC,IAAI;EACtD,MAAM,SAAS,EAAE,WAAW;EAC5B,MAAM,WACJ,EAAE,uBAAuB,QAAQ,EAAE,wBAAwB,OACvD,MACA,GAAG,EAAE,mBAAmB,GAAG,EAAE;EACnC,MAAM,KACJ,KAAK,EAAE,YAAY,KAAK,EAAE,EAAE,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI,EAAE,IAAI,EAAE,KAAK,GAAG,KAAK,EAAE,KAAK,EAAE,GAC/F;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;;;;;;AAgCA,SAAgB,YACd,MACA,OAGI,CAAC,GACa;CAClB,MAAM,QAAQ,KAAK,SAAS;CAE5B,MAAM,0BAAU,IAAI,IAAmD;CACvE,KAAK,MAAM,KAAK,MAAM;EACpB,IAAI,EAAE,aAAa,OAAO;EAC1B,MAAM,IAAI,mBAAmB,GAAG,KAAK;EACrC,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,IAAI,EAAE,YAAY,MAAM;EACxB,MAAM,SAAS,QAAQ,IAAI,EAAE,WAAW,KAAK;GAAE,MAAM,CAAC;GAAG,SAAS,CAAC;EAAE;EACrE,OAAO,KAAK,KAAK,EAAE,OAAO;EAC1B,OAAO,QAAQ,KAAK,CAAC;EACrB,QAAQ,IAAI,EAAE,aAAa,MAAM;CACnC;CAEA,MAAM,SAAwB,CAAC;CAC/B,KAAK,MAAM,CAAC,aAAa,WAAW,QAAQ,QAAQ,GAClD,OAAO,KAAK;EACV;EACA,MAAMA,KAAI,OAAO,IAAI;EACrB,SAASA,KAAI,OAAO,OAAO;EAC3B,GAAG,OAAO,KAAK;EACf,YAAY;EACZ,MAAM,KAAK,gBAAgB,eACvB,UAAU,KAAK,cAAc,YAAa,IAC1C,KAAA;CACN,CAAC;CAMH,KAAK,MAAM,KAAK,QACd,EAAE,aAAa,CAAC,OAAO,MAAM,MAAM,MAAM,KAAK,UAAU,GAAG,CAAC,CAAC;CAG/D,OAAO;EACL,MAAM;EACN;EACA,MAAM;GAAE,GAAG;GAAW,GAAG;EAAQ;EACjC;CACF;AACF;AAEA,SAAS,UAAU,GAAgB,GAAyB;CAC1D,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,WAAW,EAAE,YAAY,EAAE,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE;AACzF;AAEA,SAAgB,UAAU,GAA6C;CACrE,OAAO,EAAE,UAAU,YAAY;AACjC;;;;;;AA+CA,SAAgB,cACd,MACA,aACA,YACA,OAAgC,CAAC,GACL;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,QAAQ;CAC9B,IAAI,WAAW,GAAG,MAAM,IAAI,MAAM,iCAAiC;CAOnE,MAAM,EAAE,QAAQ,OAAO,oBAAoB,wBAAwB,gBALjD,KAAK,QAAQ,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAMnF,GALe,KAAK,QAAQ,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAMjF,GACA,KACF;CACA,MAAM,IAAI,OAAO;CAEjB,IAAI,MAAM,GACR,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA,GAAG;EACH;EACA;EACA,MAAM,CAAC;EACP,QAAQ;EACR,IAAI;CACN;CAGF,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,eAAe,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACrD,MAAM,SAAS,qBAAqB,cAAc,EAAG;CACrD,MAAM,MAAM,aAAa;CACzB,MAAM,MAAM,aAAa,aAAa,SAAS;CAI/C,MAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,GAAG,GAAG,KAAK,IAAI,GAAG,GAAG,IAAI;CACzD,MAAM,KAAK,CAAC;CAEZ,MAAM,SAASC,QAAK,MAAM;CAC1B,MAAM,OAA8B,CAAC;CACrC,KAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAC5B,KAAK,KAAK;EAAE,IAAI,KAAK,IAAI;EAAO,IAAI,MAAM,IAAI,KAAK;EAAO,OAAO;CAAE,CAAC;CAEtE,KAAK,MAAM,KAAK,QAAQ;EACtB,IAAI,MAAM,KAAK,OAAO,IAAI,MAAM,KAAK;EACrC,IAAI,MAAM,GAAG,MAAM;EACnB,IAAI,OAAO,UAAU,MAAM,WAAW;EACtC,KAAK,IAAI,CAAE,SAAS;CACtB;CAEA,MAAM,KAAK,gBAAgB,QAAQ,OAAO;EACxC,YAAY,KAAK,cAAc;EAC/B,WAAW,KAAK,aAAa;EAC7B,WAAW;EACX,MAAM,KAAK;CACb,CAAC;CAED,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,IAAI;GAAE,KAAK,GAAG;GAAK,MAAM,GAAG;EAAK;CACnC;AACF;;;;;;;;;;;;;;;;;;;;;AC9LA,eAAsB,eACpB,MACA,OAA8B,CAAC,GACN;CACzB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,WAAW,KAAK,IAAI,KAAK,YAAY,IAAA,CAAmC;CAC9E,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,cAAc,KAAK,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;CAC/D,MAAM,sBAAsB,KAAK,uBAAuB;CAExD,IAAI,QAAQ,EAAE,OAAO,SAAS,KAAK,GAAG,KAAK,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,OAAO,KAAK,OACxF,MAAM,IAAI,MACR,wEAAwE,KAAK,UAAU,IAAI,GAC7F;CAGF,MAAM,UAAU,aAAa,MAAM;EACjC,YAAY,cAAc,KAAA;EAC1B;EACA;EACA;CACF,CAAC;CACD,MAAM,SAAS,YAAY,MAAM;EAAE;EAAO,eAAe,KAAK;CAAc,CAAC;CAC7E,MAAM,eACJ,KAAK,gBAAgB,QAAQ,KAAK,KAAK,MAAM,EAAE,WAAW,CAAC,CAAC,QAAQ,OAAO,OAAO,UAAU;CAC9F,MAAM,QAAQ,aACV,aAAa,KAAK,OAChB,cAAc,MAAM,IAAI,YAAY;EAClC;EACA;EACA,MAAM,KAAK;CACb,CAAC,CACH,IACA,CAAC;CAEL,MAAM,kBAAkB,IAAI,IAAI,MAAM,KAAK,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;CACpE,MAAM,oBAAoB,IAAI,IAAI,OAAO,OAAO,KAAK,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;CAC9E,MAAM,uCAAuB,IAAI,IAAgD;CACjF,IAAI,YACF,KAAK,MAAM,MAAM,cACf,qBAAqB,IACnB,IACA,gBAAgB,MAAM,IAAI,YAAY;EACpC;EACA;EACA,MAAM,KAAK;EACX;EACA;EACA;CACF,CAAC,CACH;CAIJ,MAAM,aAAa,QAAQ,KACxB,KAAK,QAAQ;EACZ,MAAM,OAAO,gBAAgB,IAAI,IAAI,WAAW;EAChD,MAAM,QAAQ,kBAAkB,IAAI,IAAI,WAAW;EACnD,MAAM,YAAY,qBAAqB,IAAI,IAAI,WAAW,KAAK;EAC/D,MAAM,eAAe,KAAK,gBAAgB,IAAI;EAC9C,MAAM,aAAa,kBAAkB,KAAK;GACxC;GACA;GACA;GACA;GACA;GACA;GACA;GACA;EACF,CAAC;EACD,OAAO;GACL,aAAa,IAAI;GACjB,GAAG,IAAI;GACP,MAAM,IAAI;GACV,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb,uBAAuB,YAAY,UAAU,YAAY;GACzD,SAAS,WAAW,KAAK,MAAM,KAAK;GACpC,YAAY,YAAY,UAAU,cAAc,OAAO,KAAK,SAAS;GACrE,UAAU,YAAY,UAAU,YAAY;GAC5C,QAAQ,YAAY,UAAU,KAAK,OAAO,KAAK,KAAK;GACpD,mBAAmB,YAAY,UAAU,oBAAoB;GAC7D,UAAU,YAAY,UAAU,WAAW;GAC3C,KAAK,YAAY,UAAU,MAAM;GACjC,kBAAkB,OAAO,cAAc;GACvC,MAAM,eAAe,UAAU,YAAY,IAAI,KAAA;GAC/C,UAAU,WAAW;GACrB,gBAAgB,WAAW;EAC7B;CACF,CAAC,CAAC,CACD,MAAM,GAAG,MAAM;EACd,MAAM,eAAe,eAAe,EAAE,QAAQ,IAAI,eAAe,EAAE,QAAQ;EAC3E,IAAI,iBAAiB,GAAG,OAAO;EAC/B,OAAO,EAAE,OAAO,EAAE;CACpB,CAAC;CAEH,MAAM,iBAAiB,oBAAoB,YAAY;EACrD;EACA,iBAAiB,KAAK;EACtB;EACA;EACA;CACF,CAAC;CACD,MAAM,mBAAmB,sBAAsB,YAAY,gBAAgB;EACzE;EACA;EACA,iBAAiB,KAAK;EACtB;CACF,CAAC;CACD,MAAM,cAAc,iBAAiB;EACnC;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF,CAAC;CAED,MAAM,iBAAiB,MAAM,SAAS;EACpC,SAAS,KACN,QAAQ,MAAM,EAAE,aAAa,KAAK,CAAC,CACnC,KAAK,OAAO;GAAE,OAAO,EAAE;GAAO,aAAa,EAAE;GAAa,UAAU,EAAE;EAAS,EAAE,CAAC,CAClF,MAAM,GAAG,MAAM,EAAE,MAAM,cAAc,EAAE,KAAK,CAAC;EAChD;EACA;CACF,CAAC;CAED,MAAM,WAAW,uBAAuB;EACtC;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,iBAAiB,KAAK;CACxB,CAAC;CACD,MAAM,OAAO,mBAAmB,UAAU,KAAK;CAE/C,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;GAAE;GAAQ;EAAM;EACxB;EACA,iBAAiB,KAAK;EACtB;EACA;CACF;AACF"}

package/dist/tool-waste-BFG1vwTV.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"tool-waste-BFG1vwTV.d.ts","names":[],"sources":["../src/baseline.ts","../src/tool-use-metrics.ts","../src/trajectory.ts","../src/pipelines/budget-breach.ts","../src/pipelines/judge-agreement.ts","../src/pipelines/tool-waste.ts"],"mappings":";;;UAsDiB;;EAEf;;EAEA;;EAEA;;KAGU;;;;;;;;;UAUF;EACR;EACA;EACA;;KAGU,mBACP;EACC;EACA;EACA;EACA;EACA;EACA;EACA;MAED;EACC;EACA;EACA;EACA;EACA;EACA;EACA;;;iBA6EU,IAAI;;;;;;;;;iBAcJ,YAAY,aAAa,cAAc;UC9KtC;EACf;EACA;;EAEA;EACA,QAAQ,eAAe;EACvB;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe;;EAEf,kBAAkB;;iBAGE,sBACpB,OAAO,YACP,eACA,UAAS,iBACR,QAAQ;UChCM;EACf;EACA,MAAM;;EAEN;;EAEA,QAAQ;;UAGO;EACf;EACA,OAAO;EACP;EACA;EACA;EACA;EACA;;iBAGoB,gBAAgB,OAAO,YAAY,gBAAgB,QAAQ;UCtBhE;EACf;EACA;EACA;EACA,iBAAiB;EACjB;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,aAAa;EACb,YAAY;EACZ,WAAW;EACX;EACA;;iBAGoB,iBACpB,OAAO,YACP;EAAW;EAAqB;IAC/B,QAAQ;UCnBM;EACf;EACA;EACA;;EAEA;EACA;EACA;;UAGe;EACf,OAAO;EACP;EACA;;iBAGoB,mBAAmB,OAAO,aAAa,QAAQ;UCbpD;EACf;EACA;EACA;EACA;;UAGe;EACf,OAAO;EACP;;UAGe;EACf;EACA,eAAe,MAAM,UAAU;IAAS,KAAK,QAAQ,kBAAkB;;;iBAGnD,cACpB,OAAO,YACP,UAAS,mBACR,QAAQ"}