@tangle-network/agent-eval 0.37.0 → 0.40.1

package/dist/{chunk-KHZRNY3F.js → chunk-WP7SY7AI.js}

@@ -343,17 +343,18 @@ function weightedMean(scores) {
   }
   return totalWeight > 0 ? weightedSum / totalWeight : 0;
 }
-function confidenceInterval(scores, confidence = 0.95) {
+function confidenceInterval(scores, confidence = 0.95, opts = {}) {
   if (scores.length === 0) return { mean: 0, lower: 0, upper: 0 };
   if (scores.length === 1) return { mean: scores[0], lower: scores[0], upper: scores[0] };
   const n = scores.length;
   const mean = scores.reduce((a, b) => a + b, 0) / n;
-  const B = 1e3;
+  const B = opts.resamples ?? 1e3;
+  const rng = makeRng(opts.seed);
   const bootstrapMeans = [];
   for (let i = 0; i < B; i++) {
     let sum = 0;
     for (let j = 0; j < n; j++) {
-      sum += scores[Math.floor(Math.random() * n)];
+      sum += scores[Math.floor(rng() * n)];
     }
     bootstrapMeans.push(sum / n);
   }
@@ -880,4 +881,4 @@ export {
   benjaminiHochberg,
   pairedBootstrap
 };
-//# sourceMappingURL=chunk-KHZRNY3F.js.map
+//# sourceMappingURL=chunk-WP7SY7AI.js.map

package/dist/chunk-WP7SY7AI.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/judge-calibration.ts","../src/statistics.ts"],"sourcesContent":["/**\n * Judge calibration — measure judge quality against human gold + bias.\n *\n * Workflow:\n *   1. Build a golden set: {itemId, humanScore}[].\n *   2. Run candidate judges; each produces {itemId, score}.\n *   3. `calibrateJudge(golden, candidate)` reports κ + Pearson + MAE.\n *   4. `calibrateJudgeContinuous(golden, candidate)` adds quadratic-weighted\n *      κ over the un-rounded [0,1] scores plus ICC(2,1), Pearson, Spearman,\n *      and bootstrap CIs — use this for fine-grained judges where rounding\n *      to int discards information (e.g. 0.78 vs 0.81 both round to 1 and\n *      look \"perfectly agreed\" to integer κ).\n *   5. Run bias probes (positional, verbosity, self-preference) to\n *      detect systematic score inflation.\n *   6. For N≥2 judges on the same items, `continuousAgreement(scores)`\n *      reports ICC(2,1) + κ_w + Pearson + Spearman with bootstrap CIs.\n *\n * Returns actionable diagnostics, not a single number. Consumers then\n * decide whether to trust the judge, retrain it, or add a tie-breaker.\n */\n\nexport interface GoldenItem {\n  itemId: string\n  humanScore: number\n  /** Optional group used for per-group bias audits (e.g. model-of-output family). */\n  group?: string\n}\n\nexport interface CandidateScore {\n  itemId: string\n  score: number\n  /** Optional — enables positional-bias analysis (did order matter?). */\n  positionOfAInput?: 'first' | 'second'\n}\n\nexport interface CalibrationResult {\n  n: number\n  pearson: number\n  /** Cohen's κ with quadratic weights over integer-rounded scores. */\n  kappa: number\n  /** Mean absolute error vs human. */\n  mae: number\n  /** Worst-5 miscalibrations (largest |judge - human|). */\n  worstItems: Array<{ itemId: string; judge: number; human: number; delta: number }>\n}\n\nexport function calibrateJudge(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n): CalibrationResult {\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const common = [...map.values()].filter((v) => Number.isFinite(v.j))\n  const n = common.length\n  if (n < 2) {\n    return { n, pearson: NaN, kappa: NaN, mae: NaN, worstItems: [] }\n  }\n  const humans = common.map((c) => c.h)\n  const judges = common.map((c) => c.j)\n  const pearson = pearsonR(humans, judges)\n  const kappa = weightedKappa(humans.map(Math.round), judges.map(Math.round))\n  const absDiffs = common.map((c) => Math.abs(c.j - c.h))\n  const mae = absDiffs.reduce((a, b) => a + b, 0) / n\n  const worst = [...map.entries()]\n    .filter(([, v]) => Number.isFinite(v.j))\n    .map(([itemId, v]) => ({ itemId, judge: v.j, human: v.h, delta: Math.abs(v.j - v.h) }))\n    .sort((a, b) => b.delta - a.delta)\n    .slice(0, 5)\n  return { n, pearson, kappa, mae, worstItems: worst }\n}\n\nexport interface PositionalBiasResult {\n  /**\n   * Score delta (first-position - second-position) averaged across items\n   * presented in both positions. Non-zero = positional bias.\n   */\n  avgDelta: number\n  n: number\n}\n\n/**\n * Feed the same items to the judge twice with A/B swapped and pass all\n * results here. Items that don't appear in both positions are ignored.\n */\nexport function positionalBias(scores: CandidateScore[]): PositionalBiasResult {\n  const pairs = new Map<string, { first?: number; second?: number }>()\n  for (const s of scores) {\n    const slot = pairs.get(s.itemId) ?? {}\n    if (s.positionOfAInput === 'first') slot.first = s.score\n    else if (s.positionOfAInput === 'second') slot.second = s.score\n    pairs.set(s.itemId, slot)\n  }\n  const deltas: number[] = []\n  for (const { first, second } of pairs.values()) {\n    if (first !== undefined && second !== undefined) deltas.push(first - second)\n  }\n  if (deltas.length === 0) return { avgDelta: 0, n: 0 }\n  return { avgDelta: deltas.reduce((a, b) => a + b, 0) / deltas.length, n: deltas.length }\n}\n\nexport interface VerbosityBiasResult {\n  /** Pearson correlation between output length and score. Strong positive = verbosity bias. */\n  pearson: number\n  n: number\n}\n\nexport function verbosityBias(\n  samples: Array<{ outputLen: number; score: number }>,\n): VerbosityBiasResult {\n  const n = samples.length\n  if (n < 3) return { pearson: NaN, n }\n  return {\n    pearson: pearsonR(\n      samples.map((s) => s.outputLen),\n      samples.map((s) => s.score),\n    ),\n    n,\n  }\n}\n\nexport interface SelfPreferenceResult {\n  /** Mean judge score when judge's family matches output's family. */\n  inFamilyMean: number\n  outOfFamilyMean: number\n  deltaMean: number\n  n: number\n}\n\n/**\n * Pass the same scenarios scored with judge-model X grading outputs from\n * model X (in-family) and model Y (out-of-family). Non-zero delta\n * indicates self-preference.\n */\nexport function selfPreference(\n  samples: Array<{ score: number; inFamily: boolean }>,\n): SelfPreferenceResult {\n  const inF = samples.filter((s) => s.inFamily).map((s) => s.score)\n  const outF = samples.filter((s) => !s.inFamily).map((s) => s.score)\n  if (inF.length === 0 || outF.length === 0)\n    return { inFamilyMean: 0, outOfFamilyMean: 0, deltaMean: 0, n: 0 }\n  const inMean = inF.reduce((a, b) => a + b, 0) / inF.length\n  const outMean = outF.reduce((a, b) => a + b, 0) / outF.length\n  return {\n    inFamilyMean: inMean,\n    outOfFamilyMean: outMean,\n    deltaMean: inMean - outMean,\n    n: samples.length,\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction pearsonR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return NaN\n  const mA = a.reduce((s, v) => s + v, 0) / a.length\n  const mB = b.reduce((s, v) => s + v, 0) / b.length\n  let num = 0,\n    dA = 0,\n    dB = 0\n  for (let i = 0; i < a.length; i++) {\n    const da = a[i]! - mA\n    const db = b[i]! - mB\n    num += da * db\n    dA += da * da\n    dB += db * db\n  }\n  if (dA === 0 || dB === 0) return dA === 0 && dB === 0 ? 1 : 0\n  return num / Math.sqrt(dA * dB)\n}\n\n/** Quadratic weighted Cohen's κ over bounded integer scores. */\nfunction weightedKappa(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return NaN\n  const min = Math.min(...a, ...b)\n  const max = Math.max(...a, ...b)\n  const K = max - min + 1\n  if (K < 2) return 1\n  const observed: number[][] = Array.from({ length: K }, () => new Array(K).fill(0))\n  const rowMarg = new Array(K).fill(0)\n  const colMarg = new Array(K).fill(0)\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]! - min\n    const bi = b[i]! - min\n    const row = observed[ai]!\n    row[bi] = (row[bi] ?? 0) + 1\n    rowMarg[ai]++\n    colMarg[bi]++\n  }\n  let num = 0\n  let den = 0\n  for (let i = 0; i < K; i++) {\n    for (let j = 0; j < K; j++) {\n      const w = (i - j) ** 2 / (K - 1) ** 2\n      const expected = (rowMarg[i] * colMarg[j]) / a.length\n      num += w * observed[i]![j]!\n      den += w * expected\n    }\n  }\n  if (den === 0) return 1\n  return 1 - num / den\n}\n\n// ── Continuous-value inter-rater agreement ──────────────────────────\n//\n// `weightedKappa` above quantises to integers. For [0,1] judges that is\n// lossy: 0.78 and 0.81 both round to 1 and look identical to integer κ.\n// `continuousAgreement` keeps the original magnitudes:\n//\n//   - κ_w   : Cohen's weighted κ generalised to continuous scores. For two\n//             raters, weighted κ = 1 − Σ w_ij·o_ij / Σ w_ij·e_ij collapses\n//             into the closed form below when each \"category\" is the score\n//             itself (Janson & Olsson 2001; Cohen 1968 quadratic weights).\n//   - ICC   : Intraclass Correlation Coefficient, ICC(2,1) per Shrout &\n//             Fleiss (1979) — two-way random effects, absolute agreement,\n//             single rater. The right reliability coefficient when judges\n//             are a random sample of the judge population.\n//   - r, ρ  : Pearson + Spearman for sanity-check. Pearson catches linear\n//             association regardless of scale; Spearman catches monotone\n//             association regardless of distribution.\n//   - CIs   : Bootstrap percentile CIs over items (rows). Bias is preserved\n//             because we resample the full row, i.e. a judge's vector\n//             stays internally consistent within each bootstrap replicate.\n//\n// References:\n//   - Shrout, P. E. & Fleiss, J. L. (1979). Intraclass correlations: uses\n//     in assessing rater reliability. Psychological Bulletin, 86(2), 420.\n//   - McGraw, K. O. & Wong, S. P. (1996). Forming inferences about some\n//     intraclass correlation coefficients. Psychological Methods, 1, 30.\n//   - Janson, H. & Olsson, U. (2001). A measure of agreement for interval\n//     or nominal multivariate observations. Educ. Psychol. Meas., 61, 277.\n\nexport interface ContinuousAgreement {\n  /** Cohen's κ_w with quadratic weights, computed on raw [0,1] scores. */\n  weightedKappa: number\n  /** ICC(2,1): two-way random effects, absolute agreement, single rater. */\n  icc: number\n  /** Pearson product-moment correlation (averaged over rater pairs if N>2). */\n  pearson: number\n  /** Spearman rank correlation (averaged over rater pairs if N>2). */\n  spearman: number\n  /** 95% bootstrap percentile CIs over items. */\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n  /** Number of complete items (no NaN across raters). */\n  n: number\n  /** Number of raters. */\n  raters: number\n}\n\nexport interface ContinuousAgreementOptions {\n  /** Bootstrap iterations. Default 1000. Set to 0 to skip CIs (CI = [NaN, NaN]). */\n  bootstrap?: number\n  /** κ weighting scheme. Default 'quadratic'. */\n  weights?: 'linear' | 'quadratic'\n  /** PRNG seed for reproducible bootstrap. Default 0xC0FFEE. */\n  seed?: number\n  /** Confidence level for percentile CI. Default 0.95. */\n  ciLevel?: number\n}\n\n/**\n * Inter-rater agreement on continuous (typically [0,1]) scores.\n *\n * `scores` has shape [n_items][n_raters]. Rows with any non-finite entry\n * are dropped. Returns NaN metrics if fewer than 2 raters or 2 complete\n * items remain.\n */\nexport function continuousAgreement(\n  scores: number[][],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousAgreement {\n  const bootstrap = opts.bootstrap ?? 1000\n  const weights = opts.weights ?? 'quadratic'\n  const seed = opts.seed ?? 0xc0ffee\n  const ciLevel = opts.ciLevel ?? 0.95\n\n  const matrix = scores.filter((row) => row.length >= 2 && row.every((v) => Number.isFinite(v)))\n  const raters = matrix[0]?.length ?? 0\n  // All rows must have the same rater count, else drop the offenders.\n  const clean = matrix.filter((row) => row.length === raters)\n  const nClean = clean.length\n\n  if (nClean < 2 || raters < 2) {\n    return {\n      weightedKappa: NaN,\n      icc: NaN,\n      pearson: NaN,\n      spearman: NaN,\n      ci: { icc: [NaN, NaN], weightedKappa: [NaN, NaN] },\n      n: nClean,\n      raters,\n    }\n  }\n\n  const kappa = continuousWeightedKappa(clean, weights)\n  const icc = icc21(clean)\n  const pearson = avgPairwise(clean, pearsonR)\n  const spearman = avgPairwise(clean, spearmanR)\n\n  const ciIcc: [number, number] = [NaN, NaN]\n  const ciKappa: [number, number] = [NaN, NaN]\n  if (bootstrap > 0) {\n    const rng = mulberry32(seed)\n    const iccs: number[] = []\n    const kappas: number[] = []\n    for (let b = 0; b < bootstrap; b++) {\n      const sample: number[][] = new Array(nClean)\n      for (let i = 0; i < nClean; i++) {\n        sample[i] = clean[Math.floor(rng() * nClean)]!\n      }\n      const iccB = icc21(sample)\n      const kB = continuousWeightedKappa(sample, weights)\n      if (Number.isFinite(iccB)) iccs.push(iccB)\n      if (Number.isFinite(kB)) kappas.push(kB)\n    }\n    const [lo, hi] = percentileBounds(ciLevel)\n    if (iccs.length > 0) {\n      iccs.sort((a, b) => a - b)\n      ciIcc[0] = quantile(iccs, lo)\n      ciIcc[1] = quantile(iccs, hi)\n    }\n    if (kappas.length > 0) {\n      kappas.sort((a, b) => a - b)\n      ciKappa[0] = quantile(kappas, lo)\n      ciKappa[1] = quantile(kappas, hi)\n    }\n  }\n\n  return {\n    weightedKappa: kappa,\n    icc,\n    pearson,\n    spearman,\n    ci: { icc: ciIcc, weightedKappa: ciKappa },\n    n: nClean,\n    raters,\n  }\n}\n\nexport interface ContinuousCalibrationResult extends CalibrationResult {\n  /** Cohen's κ_w computed on raw (un-rounded) scores. */\n  weightedKappaContinuous: number\n  /** ICC(2,1) treating golden + candidate as two raters. */\n  icc: number\n  spearman: number\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n}\n\n/**\n * Drop-in superset of `calibrateJudge` that adds continuous-value\n * agreement metrics. The old fields (n, pearson, kappa, mae, worstItems)\n * are preserved unchanged so existing callers continue to work.\n */\nexport function calibrateJudgeContinuous(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousCalibrationResult {\n  const base = calibrateJudge(golden, candidate)\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const rows: number[][] = []\n  for (const v of map.values()) {\n    if (Number.isFinite(v.j)) rows.push([v.h, v.j])\n  }\n  const agreement = continuousAgreement(rows, opts)\n  return {\n    ...base,\n    weightedKappaContinuous: agreement.weightedKappa,\n    icc: agreement.icc,\n    spearman: agreement.spearman,\n    ci: agreement.ci,\n  }\n}\n\n// ── Continuous-agreement internals ──────────────────────────────────\n\n/**\n * Quadratic-weighted κ on continuous scores. With weights w(x,y) = (x-y)^2\n * (or |x-y| for linear) the formula collapses to:\n *\n *   κ_w = 1 − E_obs[w] / E_exp[w]\n *\n * where E_obs averages w over paired (a_i, b_i) and E_exp averages w over\n * the independent product distribution (sum_{i,j} w(a_i, b_j) / n^2).\n * The normalisation by (max-min)^2 in the integer version cancels in the\n * ratio, so we don't need it here. Generalises to N raters by averaging κ_w\n * over all rater pairs (mean pairwise weighted agreement).\n */\nfunction continuousWeightedKappa(rows: number[][], scheme: 'linear' | 'quadratic'): number {\n  if (rows.length === 0) return NaN\n  const raters = rows[0]!.length\n  if (raters < 2) return NaN\n  const wFn =\n    scheme === 'linear'\n      ? (x: number, y: number) => Math.abs(x - y)\n      : (x: number, y: number) => (x - y) ** 2\n  let sum = 0\n  let pairs = 0\n  for (let r1 = 0; r1 < raters; r1++) {\n    for (let r2 = r1 + 1; r2 < raters; r2++) {\n      const a = rows.map((row) => row[r1]!)\n      const b = rows.map((row) => row[r2]!)\n      const n = a.length\n      let obs = 0\n      for (let i = 0; i < n; i++) obs += wFn(a[i]!, b[i]!)\n      obs /= n\n      // Expected under independence: average of w over all i,j cross pairs.\n      let exp = 0\n      for (let i = 0; i < n; i++) {\n        for (let j = 0; j < n; j++) exp += wFn(a[i]!, b[j]!)\n      }\n      exp /= n * n\n      if (exp === 0) {\n        // Degenerate: at least one rater is constant. Convention: perfect\n        // agreement iff observed disagreement is also zero, else 0.\n        sum += obs === 0 ? 1 : 0\n      } else {\n        sum += 1 - obs / exp\n      }\n      pairs++\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\n/**\n * ICC(2,1) — two-way random effects, absolute agreement, single rater.\n *\n *   ICC(2,1) = (MSR − MSE) / (MSR + (k−1)·MSE + k·(MSC − MSE)/n)\n *\n * where MSR = between-rows MS, MSC = between-columns MS, MSE = residual MS,\n * n = rows (items), k = columns (raters).\n */\nfunction icc21(rows: number[][]): number {\n  const n = rows.length\n  if (n < 2) return NaN\n  const k = rows[0]!.length\n  if (k < 2) return NaN\n\n  // Row means.\n  const rowMeans = rows.map((row) => row.reduce((s, v) => s + v, 0) / k)\n  // Column means.\n  const colMeans = new Array(k).fill(0)\n  for (let j = 0; j < k; j++) {\n    let s = 0\n    for (let i = 0; i < n; i++) s += rows[i]![j]!\n    colMeans[j] = s / n\n  }\n  // Grand mean.\n  let grand = 0\n  for (let i = 0; i < n; i++) grand += rowMeans[i]!\n  grand /= n\n\n  let ssR = 0\n  for (let i = 0; i < n; i++) ssR += (rowMeans[i]! - grand) ** 2\n  ssR *= k\n  let ssC = 0\n  for (let j = 0; j < k; j++) ssC += (colMeans[j]! - grand) ** 2\n  ssC *= n\n  let ssT = 0\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < k; j++) ssT += (rows[i]![j]! - grand) ** 2\n  }\n  const ssE = ssT - ssR - ssC\n\n  const dfR = n - 1\n  const dfC = k - 1\n  const dfE = (n - 1) * (k - 1)\n  const msR = ssR / dfR\n  const msC = ssC / dfC\n  const msE = dfE > 0 ? ssE / dfE : 0\n\n  const denom = msR + (k - 1) * msE + (k * (msC - msE)) / n\n  if (denom === 0) {\n    // Degenerate (all values equal): convention = 1 if rows identical,\n    // else 0. Identical rows ⇒ msR = 0 too, so msR − msE near 0.\n    return msR === 0 && msE === 0 ? 1 : 0\n  }\n  return (msR - msE) / denom\n}\n\n/** Average pairwise statistic over all rater pairs. */\nfunction avgPairwise(rows: number[][], fn: (a: number[], b: number[]) => number): number {\n  const k = rows[0]?.length ?? 0\n  if (k < 2) return NaN\n  let sum = 0\n  let pairs = 0\n  for (let i = 0; i < k; i++) {\n    for (let j = i + 1; j < k; j++) {\n      const a = rows.map((row) => row[i]!)\n      const b = rows.map((row) => row[j]!)\n      const r = fn(a, b)\n      if (Number.isFinite(r)) {\n        sum += r\n        pairs++\n      }\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\n/** Spearman rank correlation. Ties get average ranks. */\nfunction spearmanR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return NaN\n  return pearsonR(rankWithTies(a), rankWithTies(b))\n}\n\nfunction rankWithTies(xs: number[]): number[] {\n  const n = xs.length\n  const indexed = xs.map((v, i) => ({ v, i }))\n  indexed.sort((x, y) => x.v - y.v)\n  const ranks = new Array(n).fill(0)\n  let i = 0\n  while (i < n) {\n    let j = i\n    while (j + 1 < n && indexed[j + 1]!.v === indexed[i]!.v) j++\n    // Average rank for ties (ranks are 1-indexed).\n    const avg = (i + j) / 2 + 1\n    for (let k = i; k <= j; k++) ranks[indexed[k]!.i] = avg\n    i = j + 1\n  }\n  return ranks\n}\n\n/** Seeded PRNG — Mulberry32. Deterministic across platforms. */\nfunction mulberry32(seed: number): () => number {\n  let a = seed >>> 0\n  return () => {\n    a = (a + 0x6d2b79f5) >>> 0\n    let t = a\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\nfunction percentileBounds(ciLevel: number): [number, number] {\n  const tail = (1 - ciLevel) / 2\n  return [tail, 1 - tail]\n}\n\n/** Linear-interpolated quantile of a pre-sorted ascending array. */\nfunction quantile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return NaN\n  if (sorted.length === 1) return sorted[0]!\n  const pos = q * (sorted.length - 1)\n  const lo = Math.floor(pos)\n  const hi = Math.ceil(pos)\n  if (lo === hi) return sorted[lo]!\n  const frac = pos - lo\n  return sorted[lo]! * (1 - frac) + sorted[hi]! * frac\n}\n","import { ValidationError } from './errors'\nimport {\n  type ContinuousAgreement,\n  type ContinuousAgreementOptions,\n  continuousAgreement,\n} from './judge-calibration'\nimport type { JudgeScore } from './types'\n\n/** Dimensions where lower raw score = better outcome (inverted semantics) */\nconst INVERTED_DIMENSIONS = new Set(['hallucination', 'false_confidence', 'worst_failure'])\n\n/**\n * Normalize scores so all dimensions follow \"higher = better\".\n * Inverted dimensions (hallucination, false_confidence, worst_failure)\n * already use inverted scoring in the prompt (10 = no hallucination),\n * but this function ensures consistency if raw scores leak through.\n */\nexport function normalizeScores(scores: JudgeScore[]): JudgeScore[] {\n  return scores.map((s) => {\n    if (INVERTED_DIMENSIONS.has(s.dimension)) {\n      return s\n    }\n    return s\n  })\n}\n\n/** Weighted mean — falls back to uniform weights when omitted */\nexport function weightedMean(scores: { score: number; weight?: number }[]): number {\n  if (scores.length === 0) return 0\n  let totalWeight = 0\n  let weightedSum = 0\n  for (const { score, weight } of scores) {\n    const w = weight ?? 1\n    weightedSum += score * w\n    totalWeight += w\n  }\n  return totalWeight > 0 ? weightedSum / totalWeight : 0\n}\n\n/** Bootstrap confidence interval */\nexport function confidenceInterval(\n  scores: number[],\n  confidence = 0.95,\n  opts: { seed?: number; resamples?: number } = {},\n): { mean: number; lower: number; upper: number } {\n  if (scores.length === 0) return { mean: 0, lower: 0, upper: 0 }\n  if (scores.length === 1) return { mean: scores[0]!, lower: scores[0]!, upper: scores[0]! }\n\n  const n = scores.length\n  const mean = scores.reduce((a, b) => a + b, 0) / n\n\n  const B = opts.resamples ?? 1000\n  const rng = makeRng(opts.seed)\n  const bootstrapMeans: number[] = []\n\n  for (let i = 0; i < B; i++) {\n    let sum = 0\n    for (let j = 0; j < n; j++) {\n      sum += scores[Math.floor(rng() * n)]!\n    }\n    bootstrapMeans.push(sum / n)\n  }\n\n  bootstrapMeans.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowerIdx = Math.floor((alpha / 2) * B)\n  const upperIdx = Math.floor((1 - alpha / 2) * B) - 1\n\n  return {\n    mean,\n    lower: bootstrapMeans[lowerIdx]!,\n    upper: bootstrapMeans[Math.min(upperIdx, B - 1)]!,\n  }\n}\n\n/**\n * Inter-rater reliability — simplified Krippendorff's alpha.\n *\n * Each inner array is one judge's scores for all items.\n * All arrays must have the same length (same items scored).\n */\nexport function interRaterReliability(judgeScores: JudgeScore[][]): number {\n  if (judgeScores.length < 2) return 1\n\n  // Group scores by dimension across judges\n  const dimensionMap = new Map<string, number[][]>()\n  for (const judgeSet of judgeScores) {\n    for (const s of judgeSet) {\n      if (!dimensionMap.has(s.dimension)) dimensionMap.set(s.dimension, [])\n      const arr = dimensionMap.get(s.dimension)!\n      if (arr.length === 0 || arr[arr.length - 1]!.length >= judgeScores.length) {\n        arr.push([s.score])\n      } else {\n        arr[arr.length - 1]!.push(s.score)\n      }\n    }\n  }\n\n  // Collect all paired ratings\n  const allValues: number[] = []\n  const pairDiffs: number[] = []\n\n  for (const items of dimensionMap.values()) {\n    for (const ratings of items) {\n      if (ratings.length < 2) continue\n      for (const v of ratings) allValues.push(v)\n      for (let i = 0; i < ratings.length; i++) {\n        for (let j = i + 1; j < ratings.length; j++) {\n          pairDiffs.push((ratings[i]! - ratings[j]!) ** 2)\n        }\n      }\n    }\n  }\n\n  if (pairDiffs.length === 0 || allValues.length < 2) return 1\n\n  const observedDisagreement = pairDiffs.reduce((a, b) => a + b, 0) / pairDiffs.length\n\n  // Expected disagreement from all possible pairings of values\n  let expectedDisagreement = 0\n  let expectedCount = 0\n  for (let i = 0; i < allValues.length; i++) {\n    for (let j = i + 1; j < allValues.length; j++) {\n      expectedDisagreement += (allValues[i]! - allValues[j]!) ** 2\n      expectedCount++\n    }\n  }\n  expectedDisagreement = expectedCount > 0 ? expectedDisagreement / expectedCount : 0\n\n  if (expectedDisagreement === 0) return 1\n  return 1 - observedDisagreement / expectedDisagreement\n}\n\n/**\n * Mann-Whitney U test for comparing two independent groups.\n * Returns U statistic and approximate p-value (normal approximation).\n */\nexport function mannWhitneyU(a: number[], b: number[]): { u: number; p: number } {\n  if (a.length === 0 || b.length === 0) return { u: 0, p: 1 }\n\n  const n1 = a.length\n  const n2 = b.length\n\n  // Rank all values together\n  const combined = [\n    ...a.map((v) => ({ v, group: 'a' as const })),\n    ...b.map((v) => ({ v, group: 'b' as const })),\n  ].sort((x, y) => x.v - y.v)\n\n  // Assign ranks with tie handling\n  const ranks: number[] = new Array(combined.length)\n  let i = 0\n  while (i < combined.length) {\n    let j = i\n    while (j < combined.length && combined[j]!.v === combined[i]!.v) j++\n    const avgRank = (i + 1 + j) / 2\n    for (let k = i; k < j; k++) ranks[k] = avgRank\n    i = j\n  }\n\n  // Sum ranks for group a\n  let r1 = 0\n  for (let k = 0; k < combined.length; k++) {\n    if (combined[k]!.group === 'a') r1 += ranks[k]!\n  }\n\n  const u1 = r1 - (n1 * (n1 + 1)) / 2\n  const u2 = n1 * n2 - u1\n  const u = Math.min(u1, u2)\n\n  // Normal approximation for p-value\n  const mu = (n1 * n2) / 2\n  const sigma = Math.sqrt((n1 * n2 * (n1 + n2 + 1)) / 12)\n\n  if (sigma === 0) return { u, p: 1 }\n\n  const z = Math.abs(u - mu) / sigma\n  // Two-tailed p-value from z-score (approximation)\n  const p = 2 * (1 - normalCdf(z))\n\n  return { u, p }\n}\n\n/** Partial credit: returns 0-1 ratio of current toward target */\nexport function partialCredit(current: number, target: number): number {\n  if (target <= 0) return 1\n  return Math.min(1, Math.max(0, current / target))\n}\n\n/**\n * Paired t-test — before/after measurements on the SAME items.\n * Pairing removes inter-item variance, giving tighter significance than\n * an unpaired test when comparing prompt v1 vs prompt v2 on identical\n * scenarios.\n */\nexport function pairedTTest(\n  before: number[],\n  after: number[],\n): { t: number; df: number; p: number } {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedTTest: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const n = before.length\n  if (n < 2) return { t: 0, df: 0, p: 1 }\n\n  const diffs = before.map((b, i) => after[i]! - b)\n  const mean = diffs.reduce((a, b) => a + b, 0) / n\n  const variance = diffs.reduce((acc, d) => acc + (d - mean) ** 2, 0) / (n - 1)\n  const se = Math.sqrt(variance / n)\n  if (se === 0) return { t: mean === 0 ? 0 : Infinity, df: n - 1, p: mean === 0 ? 1 : 0 }\n\n  const t = mean / se\n  const df = n - 1\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\n/**\n * Wilcoxon signed-rank test — paired non-parametric alternative.\n * Use when the differences aren't normally distributed.\n */\nexport function wilcoxonSignedRank(before: number[], after: number[]): { w: number; p: number } {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `wilcoxonSignedRank: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const diffs = before.map((b, i) => after[i]! - b).filter((d) => d !== 0)\n  const n = diffs.length\n  if (n < 6) return { w: 0, p: 1 }\n\n  const absRanks = diffs\n    .map((d, i) => ({ abs: Math.abs(d), sign: Math.sign(d), i }))\n    .sort((a, b) => a.abs - b.abs)\n  const ranks: number[] = new Array(n)\n  let i = 0\n  while (i < n) {\n    let j = i\n    while (j < n && absRanks[j]!.abs === absRanks[i]!.abs) j++\n    const avg = (i + 1 + j) / 2\n    for (let k = i; k < j; k++) ranks[absRanks[k]!.i] = avg\n    i = j\n  }\n  let wPlus = 0\n  for (let k = 0; k < n; k++) if (diffs[k]! > 0) wPlus += ranks[k]!\n\n  const mean = (n * (n + 1)) / 4\n  const variance = (n * (n + 1) * (2 * n + 1)) / 24\n  const z = (wPlus - mean) / Math.sqrt(variance)\n  const p = 2 * (1 - normalCdf(Math.abs(z)))\n  return { w: wPlus, p }\n}\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 */\nexport function cohensD(a: number[], b: number[]): number {\n  if (a.length < 2 || b.length < 2) return 0\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 0\n  return (meanB - meanA) / pooled\n}\n\n// ── Corpus-wide inter-rater agreement ──────────────────────────────\n//\n// The legacy `interRaterReliability(judgeScores)` computes a within-item\n// Krippendorff α — multiple judges score *the same item* and we ask\n// \"how much do their scores agree on that item?\" Useful for a single\n// scenario, but it cannot answer \"how reliable are these judges across\n// the whole evaluation corpus?\"\n//\n// `corpusInterRaterAgreement` does the corpus-wide question properly.\n// Inputs are flat per-(item, judge, dimension) score records. For each\n// dimension we pivot to a complete [n_items × n_judges] matrix and feed\n// it to the ICC(2,1) + κ_w machinery already validated in\n// `judge-calibration.ts`. An overall pooled metric averages the\n// per-dimension ICC/κ across dimensions.\n\nexport interface CorpusScoreRecord {\n  /** Stable identifier for the rated item (scenario, span, turn, …). */\n  itemId: string\n  /** Identifier for the judge that produced this score. */\n  judgeName: string\n  /** Dimension name (matches `JudgeScore.dimension`). */\n  dimension: string\n  /** Numeric score; must be finite. */\n  score: number\n}\n\nexport interface CorpusAgreementPerDimension extends ContinuousAgreement {\n  dimension: string\n  /** Item IDs that contributed to this dimension's matrix (every judge scored them). */\n  itemIds: string[]\n  /** Judge IDs that contributed to this dimension's matrix. */\n  judgeIds: string[]\n}\n\nexport interface CorpusAgreementReport {\n  /** Per-dimension ICC(2,1) + κ_w + Pearson + Spearman + bootstrap CIs. */\n  perDimension: CorpusAgreementPerDimension[]\n  /** Mean ICC across dimensions (NaN if no dimension yielded a finite ICC). */\n  overallIcc: number\n  /** Mean weighted κ across dimensions (NaN if none finite). */\n  overallWeightedKappa: number\n  /** Dimensions evaluated (sorted). */\n  dimensions: string[]\n  /** Judges seen across the corpus (sorted). */\n  judgeIds: string[]\n}\n\nexport interface CorpusAgreementOptions extends ContinuousAgreementOptions {\n  /**\n   * Restrict the audit to these dimensions. Default = every dimension\n   * that appears in the input. A dimension named here but absent from\n   * the input throws — silent omission would corrupt the overall metric.\n   */\n  dimensions?: string[]\n  /**\n   * Restrict the audit to these judges. Default = every judge that\n   * appears in the input. A judge named here but absent from a\n   * dimension throws (see \"fail loud\" below).\n   */\n  judges?: string[]\n}\n\n/**\n * Corpus-wide inter-rater agreement across N items × M judges × D dimensions.\n *\n * For each dimension, builds the [n_items][n_judges] matrix of scores\n * (keeping only items every judge rated on that dimension), then runs\n * `continuousAgreement` to get ICC(2,1), κ_w, Pearson, Spearman, and\n * bootstrap CIs. Reports a pooled mean across dimensions as a single\n * \"is this judge panel reliable on this corpus?\" number.\n *\n * Fail-loud contract:\n *   - Empty input throws.\n *   - Fewer than 2 judges or fewer than 2 items per dimension throws.\n *   - A judge present in some dimensions but with zero scored items on\n *     another dimension throws (would silently shrink the matrix).\n *   - Duplicate (itemId, judgeName, dimension) records throw.\n */\nexport function corpusInterRaterAgreement(\n  records: CorpusScoreRecord[],\n  opts: CorpusAgreementOptions = {},\n): CorpusAgreementReport {\n  if (records.length === 0) {\n    throw new ValidationError('corpusInterRaterAgreement: no score records supplied')\n  }\n\n  const judgesSeen = new Set<string>()\n  const dimsSeen = new Set<string>()\n  // dimension → judge → itemId → score\n  const grid = new Map<string, Map<string, Map<string, number>>>()\n\n  for (const r of records) {\n    if (!Number.isFinite(r.score)) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: non-finite score for (item=${r.itemId}, judge=${r.judgeName}, dim=${r.dimension})`,\n      )\n    }\n    judgesSeen.add(r.judgeName)\n    dimsSeen.add(r.dimension)\n    const byJudge = grid.get(r.dimension) ?? new Map<string, Map<string, number>>()\n    const byItem = byJudge.get(r.judgeName) ?? new Map<string, number>()\n    if (byItem.has(r.itemId)) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: duplicate record for (item=${r.itemId}, judge=${r.judgeName}, dim=${r.dimension})`,\n      )\n    }\n    byItem.set(r.itemId, r.score)\n    byJudge.set(r.judgeName, byItem)\n    grid.set(r.dimension, byJudge)\n  }\n\n  const targetDims = opts.dimensions ?? [...dimsSeen].sort()\n  for (const d of targetDims) {\n    if (!dimsSeen.has(d)) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: dimension '${d}' was requested but no records carry it`,\n      )\n    }\n  }\n  const targetJudges = opts.judges ? [...opts.judges] : [...judgesSeen].sort()\n  for (const j of targetJudges) {\n    if (!judgesSeen.has(j)) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: judge '${j}' was requested but produced no records`,\n      )\n    }\n  }\n  if (targetJudges.length < 2) {\n    throw new ValidationError(\n      `corpusInterRaterAgreement: need ≥2 judges, got ${targetJudges.length}`,\n    )\n  }\n\n  const perDimension: CorpusAgreementPerDimension[] = []\n  const iccs: number[] = []\n  const kappas: number[] = []\n\n  for (const dim of targetDims) {\n    const byJudge = grid.get(dim)!\n    // Fail loud: every requested judge must have scored ≥1 item on this dim.\n    const judgeItemCounts: Record<string, number> = {}\n    for (const j of targetJudges) {\n      const m = byJudge.get(j)\n      judgeItemCounts[j] = m?.size ?? 0\n    }\n    const emptyJudges = targetJudges.filter((j) => judgeItemCounts[j] === 0)\n    if (emptyJudges.length > 0) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: dimension '${dim}' has no scores from judge(s) ${emptyJudges.join(', ')} (counts: ${JSON.stringify(judgeItemCounts)})`,\n      )\n    }\n\n    // Items rated by *every* requested judge on this dim.\n    let commonItems: Set<string> | null = null\n    for (const j of targetJudges) {\n      const ids = new Set(byJudge.get(j)!.keys())\n      if (commonItems === null) {\n        commonItems = ids\n      } else {\n        const prev: Set<string> = commonItems\n        commonItems = new Set([...prev].filter((x) => ids.has(x)))\n      }\n    }\n    const sortedItems = [...(commonItems ?? new Set<string>())].sort()\n    if (sortedItems.length < 2) {\n      throw new ValidationError(\n        `corpusInterRaterAgreement: dimension '${dim}' has ${sortedItems.length} item(s) rated by all ${targetJudges.length} judges (need ≥2)`,\n      )\n    }\n\n    const matrix: number[][] = sortedItems.map((itemId) =>\n      targetJudges.map((j) => byJudge.get(j)!.get(itemId)!),\n    )\n    const agreement = continuousAgreement(matrix, opts)\n    perDimension.push({\n      ...agreement,\n      dimension: dim,\n      itemIds: sortedItems,\n      judgeIds: [...targetJudges],\n    })\n    if (Number.isFinite(agreement.icc)) iccs.push(agreement.icc)\n    if (Number.isFinite(agreement.weightedKappa)) kappas.push(agreement.weightedKappa)\n  }\n\n  const mean = (xs: number[]) =>\n    xs.length === 0 ? Number.NaN : xs.reduce((a, b) => a + b, 0) / xs.length\n  return {\n    perDimension,\n    overallIcc: mean(iccs),\n    overallWeightedKappa: mean(kappas),\n    dimensions: targetDims,\n    judgeIds: targetJudges,\n  }\n}\n\n/**\n * Convenience adapter for `JudgeScore[]` data keyed externally by item.\n *\n * Use when you have per-item arrays of `JudgeScore[]` (e.g. one\n * `ScenarioResult.judgeScores` per scenario) and want corpus-wide\n * agreement without manually flattening. `itemId` must be unique per\n * row of `itemsScores`.\n */\nexport function corpusInterRaterAgreementFromJudgeScores(\n  itemsScores: Array<{ itemId: string; scores: JudgeScore[] }>,\n  opts: CorpusAgreementOptions = {},\n): CorpusAgreementReport {\n  const records: CorpusScoreRecord[] = []\n  const seen = new Set<string>()\n  for (const { itemId, scores } of itemsScores) {\n    if (seen.has(itemId)) {\n      throw new ValidationError(\n        `corpusInterRaterAgreementFromJudgeScores: duplicate itemId '${itemId}'`,\n      )\n    }\n    seen.add(itemId)\n    for (const s of scores) {\n      records.push({\n        itemId,\n        judgeName: s.judgeName,\n        dimension: s.dimension,\n        score: s.score,\n      })\n    }\n  }\n  return corpusInterRaterAgreement(records, opts)\n}\n\n/** Student-t CDF approximation via Abramowitz-Stegun series. */\nfunction studentTCdf(t: number, df: number): number {\n  if (df <= 0) return 0.5\n  if (df > 100) return normalCdf(t)\n  const x = df / (df + t * t)\n  const a = df / 2\n  const b = 0.5\n  const ib = incompleteBeta(x, a, b)\n  return t >= 0 ? 1 - 0.5 * ib : 0.5 * ib\n}\n\n/** Regularized incomplete beta function via continued fraction (Lentz). */\nfunction incompleteBeta(x: number, a: number, b: number): number {\n  if (x <= 0) return 0\n  if (x >= 1) return 1\n  const lnBeta = lnGamma(a) + lnGamma(b) - lnGamma(a + b)\n  const front = Math.exp(Math.log(x) * a + Math.log(1 - x) * b - lnBeta) / a\n  const maxIter = 200\n  const eps = 3e-7\n  let c = 1\n  let d = 1 - ((a + b) * x) / (a + 1)\n  if (Math.abs(d) < 1e-30) d = 1e-30\n  d = 1 / d\n  let f = d\n  for (let m = 1; m <= maxIter; m++) {\n    const m2 = 2 * m\n    let num = (m * (b - m) * x) / ((a + m2 - 1) * (a + m2))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    f *= d * c\n    num = -((a + m) * (a + b + m) * x) / ((a + m2) * (a + m2 + 1))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    const delta = d * c\n    f *= delta\n    if (Math.abs(delta - 1) < eps) break\n  }\n  return front * f\n}\n\n/** Lanczos approximation to ln Γ(z). */\nfunction lnGamma(z: number): number {\n  const g = 7\n  const coefs = [\n    0.99999999999980993, 676.5203681218851, -1259.1392167224028, 771.32342877765313,\n    -176.61502916214059, 12.507343278686905, -0.13857109526572012, 9.9843695780195716e-6,\n    1.5056327351493116e-7,\n  ]\n  if (z < 0.5) {\n    return Math.log(Math.PI / Math.sin(Math.PI * z)) - lnGamma(1 - z)\n  }\n  z -= 1\n  let x = coefs[0]!\n  for (let i = 1; i < g + 2; i++) x += coefs[i]! / (z + i)\n  const t = z + g + 0.5\n  return 0.5 * Math.log(2 * Math.PI) + (z + 0.5) * Math.log(t) - t + Math.log(x)\n}\n\n// Standard normal CDF approximation (Abramowitz and Stegun)\nfunction normalCdf(x: number): number {\n  const a1 = 0.254829592\n  const a2 = -0.284496736\n  const a3 = 1.421413741\n  const a4 = -1.453152027\n  const a5 = 1.061405429\n  const p = 0.3275911\n\n  const sign = x < 0 ? -1 : 1\n  const absX = Math.abs(x)\n  const t = 1 / (1 + p * absX)\n  const y = 1 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp((-absX * absX) / 2)\n\n  return 0.5 * (1 + sign * y)\n}\n\n// ── Power analysis + multiple-comparison correction ──────────────────\n\n/**\n * Required N per arm for a two-sample comparison at target effect size,\n * alpha, and power. Normal-approximation formula:\n *   n = 2 * ( (z_{1-α/2} + z_{1-β}) / d )^2\n * where d is Cohen's d. Returns Infinity for effect ≤ 0.\n */\nexport function requiredSampleSize(opts: {\n  effect: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const effect = opts.effect\n  if (!Number.isFinite(effect) || effect <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  const n = 2 * ((zAlpha + zBeta) / effect) ** 2\n  return Math.ceil(n)\n}\n\n/**\n * Minimum detectable paired effect (standardised units) for a target paired\n * sample size: d_min = (z_{1-α/2} + z_β) / sqrt(n_paired). Multiply by\n * sd(deltas) for score units; treat as a lower bound — Wilcoxon and bootstrap\n * have asymptotic relative efficiency below 1 vs the t-test on heavy tails.\n */\nexport function pairedMde(opts: {\n  nPaired: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  if (!Number.isFinite(opts.nPaired) || opts.nPaired <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  return (zAlpha + zBeta) / Math.sqrt(opts.nPaired)\n}\n\n/** Bonferroni adjustment: multiply every p-value by the test count, clamp at 1. */\nexport function bonferroni(\n  pValues: number[],\n  alpha = 0.05,\n): { adjusted: number[]; significant: boolean[] } {\n  const k = pValues.length\n  const adjusted = pValues.map((p) => Math.min(1, p * k))\n  const significant = adjusted.map((p) => p < alpha)\n  return { adjusted, significant }\n}\n\n/**\n * Benjamini–Hochberg false discovery rate. Returns adjusted q-values and\n * significance at the target FDR; handles ties and preserves q monotonicity.\n */\nexport function benjaminiHochberg(\n  pValues: number[],\n  fdr = 0.05,\n): { qValues: number[]; significant: boolean[] } {\n  const n = pValues.length\n  if (n === 0) return { qValues: [], significant: [] }\n  const indexed = pValues.map((p, i) => ({ p, i })).sort((a, b) => a.p - b.p)\n  const q = new Array<number>(n)\n  let minRight = 1\n  for (let k = n - 1; k >= 0; k--) {\n    const rank = k + 1\n    const entry = indexed[k]!\n    const raw = (entry.p * n) / rank\n    const bounded = Math.min(minRight, raw)\n    minRight = bounded\n    q[entry.i] = Math.min(1, bounded)\n  }\n  const significant = q.map((v) => v < fdr)\n  return { qValues: q, significant }\n}\n\n// ── Paired bootstrap (promotion-gate effect size) ────────────────────\n\nexport interface PairedBootstrapResult {\n  /** Number of paired observations. */\n  n: number\n  /** Median of paired deltas (after − before). */\n  median: number\n  /** Mean of paired deltas. */\n  mean: number\n  /** Lower bound of the bootstrap CI on the chosen statistic. */\n  low: number\n  /** Upper bound of the bootstrap CI on the chosen statistic. */\n  high: number\n  /** Confidence level used (e.g. 0.95). */\n  confidence: number\n  /** Number of bootstrap resamples used. */\n  resamples: number\n}\n\nexport interface PairedBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Statistic to bootstrap. Default 'median'. */\n  statistic?: 'median' | 'mean'\n  /** Deterministic seed. If omitted, uses Math.random(). */\n  seed?: number\n}\n\n/**\n * Paired bootstrap on (after − before) deltas. Returns a CI on the chosen\n * statistic (median by default); pairs are resampled with replacement. The\n * lower bound is what the promotion gate checks — `low > threshold` means the\n * gain is real at the confidence level. Throws on unequal sample sizes.\n */\nexport function pairedBootstrap(\n  before: number[],\n  after: number[],\n  opts: PairedBootstrapOptions = {},\n): PairedBootstrapResult {\n  if (before.length !== after.length) {\n    throw new Error(`pairedBootstrap: unequal sample sizes (${before.length} vs ${after.length})`)\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const statistic = opts.statistic ?? 'median'\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedBootstrap: confidence must be in (0,1), got ${confidence}`)\n  }\n\n  const n = before.length\n  const deltas = before.map((b, i) => after[i]! - b)\n  if (n === 0) {\n    return { n: 0, median: 0, mean: 0, low: 0, high: 0, confidence, resamples }\n  }\n  if (n === 1) {\n    const d = deltas[0]!\n    return { n: 1, median: d, mean: d, low: d, high: d, confidence, resamples }\n  }\n\n  const rng = makeRng(opts.seed)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    if (statistic === 'mean') {\n      let sum = 0\n      for (let k = 0; k < n; k++) {\n        sum += deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = sum / n\n    } else {\n      const acc = new Array<number>(n)\n      for (let k = 0; k < n; k++) {\n        acc[k] = deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = medianInPlace(acc)\n    }\n  }\n  samples.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n\n  return {\n    n,\n    median: medianInPlace([...deltas]),\n    mean: deltas.reduce((s, x) => s + x, 0) / n,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n  }\n}\n\n// ── private stats helpers ────────────────────────────────────────────\n\n/** Standard-normal inverse CDF (Acklam approximation). */\nfunction zQuantile(p: number): number {\n  if (p <= 0 || p >= 1) {\n    if (p === 0) return -Infinity\n    if (p === 1) return Infinity\n    return NaN\n  }\n  const a = [\n    -3.969683028665376e1, 2.209460984245205e2, -2.759285104469687e2, 1.38357751867269e2,\n    -3.066479806614716e1, 2.506628277459239,\n  ]\n  const b = [\n    -5.447609879822406e1, 1.615858368580409e2, -1.556989798598866e2, 6.680131188771972e1,\n    -1.328068155288572e1,\n  ]\n  const c = [\n    -7.784894002430293e-3, -3.223964580411365e-1, -2.400758277161838, -2.549732539343734,\n    4.374664141464968, 2.938163982698783,\n  ]\n  const d = [7.784695709041462e-3, 3.224671290700398e-1, 2.445134137142996, 3.754408661907416]\n  const pLow = 0.02425\n  const pHigh = 1 - pLow\n  let q: number\n  let r: number\n  if (p < pLow) {\n    q = Math.sqrt(-2 * Math.log(p))\n    return (\n      (((((c[0]! * q + c[1]!) * q + c[2]!) * q + c[3]!) * q + c[4]!) * q + c[5]!) /\n      ((((d[0]! * q + d[1]!) * q + d[2]!) * q + d[3]!) * q + 1)\n    )\n  }\n  if (p <= pHigh) {\n    q = p - 0.5\n    r = q * q\n    return (\n      ((((((a[0]! * r + a[1]!) * r + a[2]!) * r + a[3]!) * r + a[4]!) * r + a[5]!) * q) /\n      (((((b[0]! * r + b[1]!) * r + b[2]!) * r + b[3]!) * r + b[4]!) * r + 1)\n    )\n  }\n  q = Math.sqrt(-2 * Math.log(1 - p))\n  return (\n    -(((((c[0]! * q + c[1]!) * q + c[2]!) * q + c[3]!) * q + c[4]!) * q + c[5]!) /\n    ((((d[0]! * q + d[1]!) * q + d[2]!) * q + d[3]!) * q + 1)\n  )\n}\n\nfunction medianInPlace(xs: number[]): number {\n  if (xs.length === 0) return 0\n  xs.sort((a, b) => a - b)\n  const mid = Math.floor(xs.length / 2)\n  return xs.length % 2 === 0 ? (xs[mid - 1]! + xs[mid]!) / 2 : xs[mid]!\n}\n\n/** Tiny seedable PRNG (mulberry32) — deterministic bootstrap resampling, not cryptographic. */\nfunction makeRng(seed: number | undefined): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed | 0 || 0x9e3779b9\n  return () => {\n    s = (s + 0x6d2b79f5) | 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n"],"mappings":";;;;;AA8CO,SAAS,eACd,QACA,WACmB;AACnB,QAAM,MAAM,oBAAI,IAAsC;AACtD,aAAW,KAAK,OAAQ,KAAI,IAAI,EAAE,QAAQ,EAAE,GAAG,EAAE,YAAY,GAAG,IAAI,CAAC;AACrE,aAAW,KAAK,WAAW;AACzB,UAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;AAC9B,QAAI,MAAO,OAAM,IAAI,EAAE;AAAA,EACzB;AACA,QAAM,SAAS,CAAC,GAAG,IAAI,OAAO,CAAC,EAAE,OAAO,CAAC,MAAM,OAAO,SAAS,EAAE,CAAC,CAAC;AACnE,QAAM,IAAI,OAAO;AACjB,MAAI,IAAI,GAAG;AACT,WAAO,EAAE,GAAG,SAAS,KAAK,OAAO,KAAK,KAAK,KAAK,YAAY,CAAC,EAAE;AAAA,EACjE;AACA,QAAM,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC;AACpC,QAAM,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,CAAC;AACpC,QAAM,UAAU,SAAS,QAAQ,MAAM;AACvC,QAAM,QAAQ,cAAc,OAAO,IAAI,KAAK,KAAK,GAAG,OAAO,IAAI,KAAK,KAAK,CAAC;AAC1E,QAAM,WAAW,OAAO,IAAI,CAAC,MAAM,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;AACtD,QAAM,MAAM,SAAS,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClD,QAAM,QAAQ,CAAC,GAAG,IAAI,QAAQ,CAAC,EAC5B,OAAO,CAAC,CAAC,EAAE,CAAC,MAAM,OAAO,SAAS,EAAE,CAAC,CAAC,EACtC,IAAI,CAAC,CAAC,QAAQ,CAAC,OAAO,EAAE,QAAQ,OAAO,EAAE,GAAG,OAAO,EAAE,GAAG,OAAO,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC,EAAE,EAAE,EACrF,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,CAAC;AACb,SAAO,EAAE,GAAG,SAAS,OAAO,KAAK,YAAY,MAAM;AACrD;AAeO,SAAS,eAAe,QAAgD;AAC7E,QAAM,QAAQ,oBAAI,IAAiD;AACnE,aAAW,KAAK,QAAQ;AACtB,UAAM,OAAO,MAAM,IAAI,EAAE,MAAM,KAAK,CAAC;AACrC,QAAI,EAAE,qBAAqB,QAAS,MAAK,QAAQ,EAAE;AAAA,aAC1C,EAAE,qBAAqB,SAAU,MAAK,SAAS,EAAE;AAC1D,UAAM,IAAI,EAAE,QAAQ,IAAI;AAAA,EAC1B;AACA,QAAM,SAAmB,CAAC;AAC1B,aAAW,EAAE,OAAO,OAAO,KAAK,MAAM,OAAO,GAAG;AAC9C,QAAI,UAAU,UAAa,WAAW,OAAW,QAAO,KAAK,QAAQ,MAAM;AAAA,EAC7E;AACA,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,UAAU,GAAG,GAAG,EAAE;AACpD,SAAO,EAAE,UAAU,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO,QAAQ,GAAG,OAAO,OAAO;AACzF;AAQO,SAAS,cACd,SACqB;AACrB,QAAM,IAAI,QAAQ;AAClB,MAAI,IAAI,EAAG,QAAO,EAAE,SAAS,KAAK,EAAE;AACpC,SAAO;AAAA,IACL,SAAS;AAAA,MACP,QAAQ,IAAI,CAAC,MAAM,EAAE,SAAS;AAAA,MAC9B,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,IAC5B;AAAA,IACA;AAAA,EACF;AACF;AAeO,SAAS,eACd,SACsB;AACtB,QAAM,MAAM,QAAQ,OAAO,CAAC,MAAM,EAAE,QAAQ,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK;AAChE,QAAM,OAAO,QAAQ,OAAO,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK;AAClE,MAAI,IAAI,WAAW,KAAK,KAAK,WAAW;AACtC,WAAO,EAAE,cAAc,GAAG,iBAAiB,GAAG,WAAW,GAAG,GAAG,EAAE;AACnE,QAAM,SAAS,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,IAAI;AACpD,QAAM,UAAU,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK;AACvD,SAAO;AAAA,IACL,cAAc;AAAA,IACd,iBAAiB;AAAA,IACjB,WAAW,SAAS;AAAA,IACpB,GAAG,QAAQ;AAAA,EACb;AACF;AAIA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,EAAG,QAAO;AAClD,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,QAAM,KAAK,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC5C,MAAI,MAAM,GACR,KAAK,GACL,KAAK;AACP,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,WAAO,KAAK;AACZ,UAAM,KAAK;AACX,UAAM,KAAK;AAAA,EACb;AACA,MAAI,OAAO,KAAK,OAAO,EAAG,QAAO,OAAO,KAAK,OAAO,IAAI,IAAI;AAC5D,SAAO,MAAM,KAAK,KAAK,KAAK,EAAE;AAChC;AAGA,SAAS,cAAc,GAAa,GAAqB;AACvD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAG,QAAO;AACpD,QAAM,MAAM,KAAK,IAAI,GAAG,GAAG,GAAG,CAAC;AAC/B,QAAM,MAAM,KAAK,IAAI,GAAG,GAAG,GAAG,CAAC;AAC/B,QAAM,IAAI,MAAM,MAAM;AACtB,MAAI,IAAI,EAAG,QAAO;AAClB,QAAM,WAAuB,MAAM,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,MAAM,CAAC,EAAE,KAAK,CAAC,CAAC;AACjF,QAAM,UAAU,IAAI,MAAM,CAAC,EAAE,KAAK,CAAC;AACnC,QAAM,UAAU,IAAI,MAAM,CAAC,EAAE,KAAK,CAAC;AACnC,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,MAAM,SAAS,EAAE;AACvB,QAAI,EAAE,KAAK,IAAI,EAAE,KAAK,KAAK;AAC3B,YAAQ,EAAE;AACV,YAAQ,EAAE;AAAA,EACZ;AACA,MAAI,MAAM;AACV,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,KAAK,IAAI,MAAM,KAAK,IAAI,MAAM;AACpC,YAAM,WAAY,QAAQ,CAAC,IAAI,QAAQ,CAAC,IAAK,EAAE;AAC/C,aAAO,IAAI,SAAS,CAAC,EAAG,CAAC;AACzB,aAAO,IAAI;AAAA,IACb;AAAA,EACF;AACA,MAAI,QAAQ,EAAG,QAAO;AACtB,SAAO,IAAI,MAAM;AACnB;AAqEO,SAAS,oBACd,QACA,OAAmC,CAAC,GACf;AACrB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,UAAU,KAAK,WAAW;AAEhC,QAAM,SAAS,OAAO,OAAO,CAAC,QAAQ,IAAI,UAAU,KAAK,IAAI,MAAM,CAAC,MAAM,OAAO,SAAS,CAAC,CAAC,CAAC;AAC7F,QAAM,SAAS,OAAO,CAAC,GAAG,UAAU;AAEpC,QAAM,QAAQ,OAAO,OAAO,CAAC,QAAQ,IAAI,WAAW,MAAM;AAC1D,QAAM,SAAS,MAAM;AAErB,MAAI,SAAS,KAAK,SAAS,GAAG;AAC5B,WAAO;AAAA,MACL,eAAe;AAAA,MACf,KAAK;AAAA,MACL,SAAS;AAAA,MACT,UAAU;AAAA,MACV,IAAI,EAAE,KAAK,CAAC,KAAK,GAAG,GAAG,eAAe,CAAC,KAAK,GAAG,EAAE;AAAA,MACjD,GAAG;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,QAAM,QAAQ,wBAAwB,OAAO,OAAO;AACpD,QAAM,MAAM,MAAM,KAAK;AACvB,QAAM,UAAU,YAAY,OAAO,QAAQ;AAC3C,QAAM,WAAW,YAAY,OAAO,SAAS;AAE7C,QAAM,QAA0B,CAAC,KAAK,GAAG;AACzC,QAAM,UAA4B,CAAC,KAAK,GAAG;AAC3C,MAAI,YAAY,GAAG;AACjB,UAAM,MAAM,WAAW,IAAI;AAC3B,UAAM,OAAiB,CAAC;AACxB,UAAM,SAAmB,CAAC;AAC1B,aAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,YAAM,SAAqB,IAAI,MAAM,MAAM;AAC3C,eAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,eAAO,CAAC,IAAI,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC;AAAA,MAC9C;AACA,YAAM,OAAO,MAAM,MAAM;AACzB,YAAM,KAAK,wBAAwB,QAAQ,OAAO;AAClD,UAAI,OAAO,SAAS,IAAI,EAAG,MAAK,KAAK,IAAI;AACzC,UAAI,OAAO,SAAS,EAAE,EAAG,QAAO,KAAK,EAAE;AAAA,IACzC;AACA,UAAM,CAAC,IAAI,EAAE,IAAI,iBAAiB,OAAO;AACzC,QAAI,KAAK,SAAS,GAAG;AACnB,WAAK,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACzB,YAAM,CAAC,IAAI,SAAS,MAAM,EAAE;AAC5B,YAAM,CAAC,IAAI,SAAS,MAAM,EAAE;AAAA,IAC9B;AACA,QAAI,OAAO,SAAS,GAAG;AACrB,aAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,cAAQ,CAAC,IAAI,SAAS,QAAQ,EAAE;AAChC,cAAQ,CAAC,IAAI,SAAS,QAAQ,EAAE;AAAA,IAClC;AAAA,EACF;AAEA,SAAO;AAAA,IACL,eAAe;AAAA,IACf;AAAA,IACA;AAAA,IACA;AAAA,IACA,IAAI,EAAE,KAAK,OAAO,eAAe,QAAQ;AAAA,IACzC,GAAG;AAAA,IACH;AAAA,EACF;AACF;AAmBO,SAAS,yBACd,QACA,WACA,OAAmC,CAAC,GACP;AAC7B,QAAM,OAAO,eAAe,QAAQ,SAAS;AAC7C,QAAM,MAAM,oBAAI,IAAsC;AACtD,aAAW,KAAK,OAAQ,KAAI,IAAI,EAAE,QAAQ,EAAE,GAAG,EAAE,YAAY,GAAG,IAAI,CAAC;AACrE,aAAW,KAAK,WAAW;AACzB,UAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;AAC9B,QAAI,MAAO,OAAM,IAAI,EAAE;AAAA,EACzB;AACA,QAAM,OAAmB,CAAC;AAC1B,aAAW,KAAK,IAAI,OAAO,GAAG;AAC5B,QAAI,OAAO,SAAS,EAAE,CAAC,EAAG,MAAK,KAAK,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;AAAA,EAChD;AACA,QAAM,YAAY,oBAAoB,MAAM,IAAI;AAChD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,yBAAyB,UAAU;AAAA,IACnC,KAAK,UAAU;AAAA,IACf,UAAU,UAAU;AAAA,IACpB,IAAI,UAAU;AAAA,EAChB;AACF;AAgBA,SAAS,wBAAwB,MAAkB,QAAwC;AACzF,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,QAAM,SAAS,KAAK,CAAC,EAAG;AACxB,MAAI,SAAS,EAAG,QAAO;AACvB,QAAM,MACJ,WAAW,WACP,CAAC,GAAW,MAAc,KAAK,IAAI,IAAI,CAAC,IACxC,CAAC,GAAW,OAAe,IAAI,MAAM;AAC3C,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,WAAS,KAAK,GAAG,KAAK,QAAQ,MAAM;AAClC,aAAS,KAAK,KAAK,GAAG,KAAK,QAAQ,MAAM;AACvC,YAAM,IAAI,KAAK,IAAI,CAAC,QAAQ,IAAI,EAAE,CAAE;AACpC,YAAM,IAAI,KAAK,IAAI,CAAC,QAAQ,IAAI,EAAE,CAAE;AACpC,YAAM,IAAI,EAAE;AACZ,UAAI,MAAM;AACV,eAAS,IAAI,GAAG,IAAI,GAAG,IAAK,QAAO,IAAI,EAAE,CAAC,GAAI,EAAE,CAAC,CAAE;AACnD,aAAO;AAEP,UAAI,MAAM;AACV,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,iBAAS,IAAI,GAAG,IAAI,GAAG,IAAK,QAAO,IAAI,EAAE,CAAC,GAAI,EAAE,CAAC,CAAE;AAAA,MACrD;AACA,aAAO,IAAI;AACX,UAAI,QAAQ,GAAG;AAGb,eAAO,QAAQ,IAAI,IAAI;AAAA,MACzB,OAAO;AACL,eAAO,IAAI,MAAM;AAAA,MACnB;AACA;AAAA,IACF;AAAA,EACF;AACA,SAAO,UAAU,IAAI,MAAM,MAAM;AACnC;AAUA,SAAS,MAAM,MAA0B;AACvC,QAAM,IAAI,KAAK;AACf,MAAI,IAAI,EAAG,QAAO;AAClB,QAAM,IAAI,KAAK,CAAC,EAAG;AACnB,MAAI,IAAI,EAAG,QAAO;AAGlB,QAAM,WAAW,KAAK,IAAI,CAAC,QAAQ,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC;AAErE,QAAM,WAAW,IAAI,MAAM,CAAC,EAAE,KAAK,CAAC;AACpC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,QAAI,IAAI;AACR,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,MAAK,KAAK,CAAC,EAAG,CAAC;AAC3C,aAAS,CAAC,IAAI,IAAI;AAAA,EACpB;AAEA,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,UAAS,SAAS,CAAC;AAC/C,WAAS;AAET,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,SAAS,CAAC,IAAK,UAAU;AAC7D,SAAO;AACP,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,SAAS,CAAC,IAAK,UAAU;AAC7D,SAAO;AACP,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,KAAK,CAAC,EAAG,CAAC,IAAK,UAAU;AAAA,EAC/D;AACA,QAAM,MAAM,MAAM,MAAM;AAExB,QAAM,MAAM,IAAI;AAChB,QAAM,MAAM,IAAI;AAChB,QAAM,OAAO,IAAI,MAAM,IAAI;AAC3B,QAAM,MAAM,MAAM;AAClB,QAAM,MAAM,MAAM;AAClB,QAAM,MAAM,MAAM,IAAI,MAAM,MAAM;AAElC,QAAM,QAAQ,OAAO,IAAI,KAAK,MAAO,KAAK,MAAM,OAAQ;AACxD,MAAI,UAAU,GAAG;AAGf,WAAO,QAAQ,KAAK,QAAQ,IAAI,IAAI;AAAA,EACtC;AACA,UAAQ,MAAM,OAAO;AACvB;AAGA,SAAS,YAAY,MAAkB,IAAkD;AACvF,QAAM,IAAI,KAAK,CAAC,GAAG,UAAU;AAC7B,MAAI,IAAI,EAAG,QAAO;AAClB,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;AAC9B,YAAM,IAAI,KAAK,IAAI,CAAC,QAAQ,IAAI,CAAC,CAAE;AACnC,YAAM,IAAI,KAAK,IAAI,CAAC,QAAQ,IAAI,CAAC,CAAE;AACnC,YAAM,IAAI,GAAG,GAAG,CAAC;AACjB,UAAI,OAAO,SAAS,CAAC,GAAG;AACtB,eAAO;AACP;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO,UAAU,IAAI,MAAM,MAAM;AACnC;AAGA,SAAS,UAAU,GAAa,GAAqB;AACnD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,EAAG,QAAO;AAClD,SAAO,SAAS,aAAa,CAAC,GAAG,aAAa,CAAC,CAAC;AAClD;AAEA,SAAS,aAAa,IAAwB;AAC5C,QAAM,IAAI,GAAG;AACb,QAAM,UAAU,GAAG,IAAI,CAAC,GAAGA,QAAO,EAAE,GAAG,GAAAA,GAAE,EAAE;AAC3C,UAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AAChC,QAAM,QAAQ,IAAI,MAAM,CAAC,EAAE,KAAK,CAAC;AACjC,MAAI,IAAI;AACR,SAAO,IAAI,GAAG;AACZ,QAAI,IAAI;AACR,WAAO,IAAI,IAAI,KAAK,QAAQ,IAAI,CAAC,EAAG,MAAM,QAAQ,CAAC,EAAG,EAAG;AAEzD,UAAM,OAAO,IAAI,KAAK,IAAI;AAC1B,aAAS,IAAI,GAAG,KAAK,GAAG,IAAK,OAAM,QAAQ,CAAC,EAAG,CAAC,IAAI;AACpD,QAAI,IAAI;AAAA,EACV;AACA,SAAO;AACT;AAGA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAEA,SAAS,iBAAiB,SAAmC;AAC3D,QAAM,QAAQ,IAAI,WAAW;AAC7B,SAAO,CAAC,MAAM,IAAI,IAAI;AACxB;AAGA,SAAS,SAAS,QAAkB,GAAmB;AACrD,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,OAAO,WAAW,EAAG,QAAO,OAAO,CAAC;AACxC,QAAM,MAAM,KAAK,OAAO,SAAS;AACjC,QAAM,KAAK,KAAK,MAAM,GAAG;AACzB,QAAM,KAAK,KAAK,KAAK,GAAG;AACxB,MAAI,OAAO,GAAI,QAAO,OAAO,EAAE;AAC/B,QAAM,OAAO,MAAM;AACnB,SAAO,OAAO,EAAE,KAAM,IAAI,QAAQ,OAAO,EAAE,IAAK;AAClD;;;AC5iBA,IAAM,sBAAsB,oBAAI,IAAI,CAAC,iBAAiB,oBAAoB,eAAe,CAAC;AAQnF,SAAS,gBAAgB,QAAoC;AAClE,SAAO,OAAO,IAAI,CAAC,MAAM;AACvB,QAAI,oBAAoB,IAAI,EAAE,SAAS,GAAG;AACxC,aAAO;AAAA,IACT;AACA,WAAO;AAAA,EACT,CAAC;AACH;AAGO,SAAS,aAAa,QAAsD;AACjF,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,cAAc;AAClB,MAAI,cAAc;AAClB,aAAW,EAAE,OAAO,OAAO,KAAK,QAAQ;AACtC,UAAM,IAAI,UAAU;AACpB,mBAAe,QAAQ;AACvB,mBAAe;AAAA,EACjB;AACA,SAAO,cAAc,IAAI,cAAc,cAAc;AACvD;AAGO,SAAS,mBACd,QACA,aAAa,MACb,OAA8C,CAAC,GACC;AAChD,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,MAAM,GAAG,OAAO,GAAG,OAAO,EAAE;AAC9D,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,MAAM,OAAO,CAAC,GAAI,OAAO,OAAO,CAAC,GAAI,OAAO,OAAO,CAAC,EAAG;AAEzF,QAAM,IAAI,OAAO;AACjB,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAEjD,QAAM,IAAI,KAAK,aAAa;AAC5B,QAAM,MAAM,QAAQ,KAAK,IAAI;AAC7B,QAAM,iBAA2B,CAAC;AAElC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,CAAC;AAAA,IACrC;AACA,mBAAe,KAAK,MAAM,CAAC;AAAA,EAC7B;AAEA,iBAAe,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAEnC,QAAM,QAAQ,IAAI;AAClB,QAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,CAAC;AAC3C,QAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,CAAC,IAAI;AAEnD,SAAO;AAAA,IACL;AAAA,IACA,OAAO,eAAe,QAAQ;AAAA,IAC9B,OAAO,eAAe,KAAK,IAAI,UAAU,IAAI,CAAC,CAAC;AAAA,EACjD;AACF;AAQO,SAAS,sBAAsB,aAAqC;AACzE,MAAI,YAAY,SAAS,EAAG,QAAO;AAGnC,QAAM,eAAe,oBAAI,IAAwB;AACjD,aAAW,YAAY,aAAa;AAClC,eAAW,KAAK,UAAU;AACxB,UAAI,CAAC,aAAa,IAAI,EAAE,SAAS,EAAG,cAAa,IAAI,EAAE,WAAW,CAAC,CAAC;AACpE,YAAM,MAAM,aAAa,IAAI,EAAE,SAAS;AACxC,UAAI,IAAI,WAAW,KAAK,IAAI,IAAI,SAAS,CAAC,EAAG,UAAU,YAAY,QAAQ;AACzE,YAAI,KAAK,CAAC,EAAE,KAAK,CAAC;AAAA,MACpB,OAAO;AACL,YAAI,IAAI,SAAS,CAAC,EAAG,KAAK,EAAE,KAAK;AAAA,MACnC;AAAA,IACF;AAAA,EACF;AAGA,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAsB,CAAC;AAE7B,aAAW,SAAS,aAAa,OAAO,GAAG;AACzC,eAAW,WAAW,OAAO;AAC3B,UAAI,QAAQ,SAAS,EAAG;AACxB,iBAAW,KAAK,QAAS,WAAU,KAAK,CAAC;AACzC,eAAS,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,iBAAS,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AAC3C,oBAAU,MAAM,QAAQ,CAAC,IAAK,QAAQ,CAAC,MAAO,CAAC;AAAA,QACjD;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,MAAI,UAAU,WAAW,KAAK,UAAU,SAAS,EAAG,QAAO;AAE3D,QAAM,uBAAuB,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,UAAU;AAG9E,MAAI,uBAAuB;AAC3B,MAAI,gBAAgB;AACpB,WAAS,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AACzC,aAAS,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AAC7C,+BAAyB,UAAU,CAAC,IAAK,UAAU,CAAC,MAAO;AAC3D;AAAA,IACF;AAAA,EACF;AACA,yBAAuB,gBAAgB,IAAI,uBAAuB,gBAAgB;AAElF,MAAI,yBAAyB,EAAG,QAAO;AACvC,SAAO,IAAI,uBAAuB;AACpC;AAMO,SAAS,aAAa,GAAa,GAAuC;AAC/E,MAAI,EAAE,WAAW,KAAK,EAAE,WAAW,EAAG,QAAO,EAAE,GAAG,GAAG,GAAG,EAAE;AAE1D,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,EAAE;AAGb,QAAM,WAAW;AAAA,IACf,GAAG,EAAE,IAAI,CAAC,OAAO,EAAE,GAAG,OAAO,IAAa,EAAE;AAAA,IAC5C,GAAG,EAAE,IAAI,CAAC,OAAO,EAAE,GAAG,OAAO,IAAa,EAAE;AAAA,EAC9C,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AAG1B,QAAM,QAAkB,IAAI,MAAM,SAAS,MAAM;AACjD,MAAI,IAAI;AACR,SAAO,IAAI,SAAS,QAAQ;AAC1B,QAAI,IAAI;AACR,WAAO,IAAI,SAAS,UAAU,SAAS,CAAC,EAAG,MAAM,SAAS,CAAC,EAAG,EAAG;AACjE,UAAM,WAAW,IAAI,IAAI,KAAK;AAC9B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,OAAM,CAAC,IAAI;AACvC,QAAI;AAAA,EACN;AAGA,MAAI,KAAK;AACT,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,QAAI,SAAS,CAAC,EAAG,UAAU,IAAK,OAAM,MAAM,CAAC;AAAA,EAC/C;AAEA,QAAM,KAAK,KAAM,MAAM,KAAK,KAAM;AAClC,QAAM,KAAK,KAAK,KAAK;AACrB,QAAM,IAAI,KAAK,IAAI,IAAI,EAAE;AAGzB,QAAM,KAAM,KAAK,KAAM;AACvB,QAAM,QAAQ,KAAK,KAAM,KAAK,MAAM,KAAK,KAAK,KAAM,EAAE;AAEtD,MAAI,UAAU,EAAG,QAAO,EAAE,GAAG,GAAG,EAAE;AAElC,QAAM,IAAI,KAAK,IAAI,IAAI,EAAE,IAAI;AAE7B,QAAM,IAAI,KAAK,IAAI,UAAU,CAAC;AAE9B,SAAO,EAAE,GAAG,EAAE;AAChB;AAGO,SAAS,cAAc,SAAiB,QAAwB;AACrE,MAAI,UAAU,EAAG,QAAO;AACxB,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC;AAClD;AAQO,SAAS,YACd,QACA,OACsC;AACtC,MAAI,OAAO,WAAW,MAAM,QAAQ;AAClC,UAAM,IAAI;AAAA,MACR,sCAAsC,OAAO,MAAM,OAAO,MAAM,MAAM;AAAA,IACxE;AAAA,EACF;AACA,QAAM,IAAI,OAAO;AACjB,MAAI,IAAI,EAAG,QAAO,EAAE,GAAG,GAAG,IAAI,GAAG,GAAG,EAAE;AAEtC,QAAM,QAAQ,OAAO,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,IAAK,CAAC;AAChD,QAAM,OAAO,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAChD,QAAM,WAAW,MAAM,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI;AAC3E,QAAM,KAAK,KAAK,KAAK,WAAW,CAAC;AACjC,MAAI,OAAO,EAAG,QAAO,EAAE,GAAG,SAAS,IAAI,IAAI,UAAU,IAAI,IAAI,GAAG,GAAG,SAAS,IAAI,IAAI,EAAE;AAEtF,QAAM,IAAI,OAAO;AACjB,QAAM,KAAK,IAAI;AACf,QAAM,IAAI,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;AAC9C,SAAO,EAAE,GAAG,IAAI,EAAE;AACpB;AAMO,SAAS,mBAAmB,QAAkB,OAA2C;AAC9F,MAAI,OAAO,WAAW,MAAM,QAAQ;AAClC,UAAM,IAAI;AAAA,MACR,6CAA6C,OAAO,MAAM,OAAO,MAAM,MAAM;AAAA,IAC/E;AAAA,EACF;AACA,QAAM,QAAQ,OAAO,IAAI,CAAC,GAAGC,OAAM,MAAMA,EAAC,IAAK,CAAC,EAAE,OAAO,CAAC,MAAM,MAAM,CAAC;AACvE,QAAM,IAAI,MAAM;AAChB,MAAI,IAAI,EAAG,QAAO,EAAE,GAAG,GAAG,GAAG,EAAE;AAE/B,QAAM,WAAW,MACd,IAAI,CAAC,GAAGA,QAAO,EAAE,KAAK,KAAK,IAAI,CAAC,GAAG,MAAM,KAAK,KAAK,CAAC,GAAG,GAAAA,GAAE,EAAE,EAC3D,KAAK,CAAC,GAAG,MAAM,EAAE,MAAM,EAAE,GAAG;AAC/B,QAAM,QAAkB,IAAI,MAAM,CAAC;AACnC,MAAI,IAAI;AACR,SAAO,IAAI,GAAG;AACZ,QAAI,IAAI;AACR,WAAO,IAAI,KAAK,SAAS,CAAC,EAAG,QAAQ,SAAS,CAAC,EAAG,IAAK;AACvD,UAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,OAAM,SAAS,CAAC,EAAG,CAAC,IAAI;AACpD,QAAI;AAAA,EACN;AACA,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,MAAM,CAAC,IAAK,EAAG,UAAS,MAAM,CAAC;AAE/D,QAAM,OAAQ,KAAK,IAAI,KAAM;AAC7B,QAAM,WAAY,KAAK,IAAI,MAAM,IAAI,IAAI,KAAM;AAC/C,QAAM,KAAK,QAAQ,QAAQ,KAAK,KAAK,QAAQ;AAC7C,QAAM,IAAI,KAAK,IAAI,UAAU,KAAK,IAAI,CAAC,CAAC;AACxC,SAAO,EAAE,GAAG,OAAO,EAAE;AACvB;AAOO,SAAS,QAAQ,GAAa,GAAqB;AACxD,MAAI,EAAE,SAAS,KAAK,EAAE,SAAS,EAAG,QAAO;AACzC,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC/C,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;AAC/C,QAAM,OAAO,EAAE,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;AAC3E,QAAM,OAAO,EAAE,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;AAC3E,QAAM,SAAS,KAAK;AAAA,MAChB,EAAE,SAAS,KAAK,QAAQ,EAAE,SAAS,KAAK,SAAS,EAAE,SAAS,EAAE,SAAS;AAAA,EAC3E;AACA,MAAI,WAAW,EAAG,QAAO;AACzB,UAAQ,QAAQ,SAAS;AAC3B;AAgFO,SAAS,0BACd,SACA,OAA+B,CAAC,GACT;AACvB,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI,gBAAgB,sDAAsD;AAAA,EAClF;AAEA,QAAM,aAAa,oBAAI,IAAY;AACnC,QAAM,WAAW,oBAAI,IAAY;AAEjC,QAAM,OAAO,oBAAI,IAA8C;AAE/D,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAAG;AAC7B,YAAM,IAAI;AAAA,QACR,yDAAyD,EAAE,MAAM,WAAW,EAAE,SAAS,SAAS,EAAE,SAAS;AAAA,MAC7G;AAAA,IACF;AACA,eAAW,IAAI,EAAE,SAAS;AAC1B,aAAS,IAAI,EAAE,SAAS;AACxB,UAAM,UAAU,KAAK,IAAI,EAAE,SAAS,KAAK,oBAAI,IAAiC;AAC9E,UAAM,SAAS,QAAQ,IAAI,EAAE,SAAS,KAAK,oBAAI,IAAoB;AACnE,QAAI,OAAO,IAAI,EAAE,MAAM,GAAG;AACxB,YAAM,IAAI;AAAA,QACR,yDAAyD,EAAE,MAAM,WAAW,EAAE,SAAS,SAAS,EAAE,SAAS;AAAA,MAC7G;AAAA,IACF;AACA,WAAO,IAAI,EAAE,QAAQ,EAAE,KAAK;AAC5B,YAAQ,IAAI,EAAE,WAAW,MAAM;AAC/B,SAAK,IAAI,EAAE,WAAW,OAAO;AAAA,EAC/B;AAEA,QAAM,aAAa,KAAK,cAAc,CAAC,GAAG,QAAQ,EAAE,KAAK;AACzD,aAAW,KAAK,YAAY;AAC1B,QAAI,CAAC,SAAS,IAAI,CAAC,GAAG;AACpB,YAAM,IAAI;AAAA,QACR,yCAAyC,CAAC;AAAA,MAC5C;AAAA,IACF;AAAA,EACF;AACA,QAAM,eAAe,KAAK,SAAS,CAAC,GAAG,KAAK,MAAM,IAAI,CAAC,GAAG,UAAU,EAAE,KAAK;AAC3E,aAAW,KAAK,cAAc;AAC5B,QAAI,CAAC,WAAW,IAAI,CAAC,GAAG;AACtB,YAAM,IAAI;AAAA,QACR,qCAAqC,CAAC;AAAA,MACxC;AAAA,IACF;AAAA,EACF;AACA,MAAI,aAAa,SAAS,GAAG;AAC3B,UAAM,IAAI;AAAA,MACR,uDAAkD,aAAa,MAAM;AAAA,IACvE;AAAA,EACF;AAEA,QAAM,eAA8C,CAAC;AACrD,QAAM,OAAiB,CAAC;AACxB,QAAM,SAAmB,CAAC;AAE1B,aAAW,OAAO,YAAY;AAC5B,UAAM,UAAU,KAAK,IAAI,GAAG;AAE5B,UAAM,kBAA0C,CAAC;AACjD,eAAW,KAAK,cAAc;AAC5B,YAAM,IAAI,QAAQ,IAAI,CAAC;AACvB,sBAAgB,CAAC,IAAI,GAAG,QAAQ;AAAA,IAClC;AACA,UAAM,cAAc,aAAa,OAAO,CAAC,MAAM,gBAAgB,CAAC,MAAM,CAAC;AACvE,QAAI,YAAY,SAAS,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,yCAAyC,GAAG,iCAAiC,YAAY,KAAK,IAAI,CAAC,aAAa,KAAK,UAAU,eAAe,CAAC;AAAA,MACjJ;AAAA,IACF;AAGA,QAAI,cAAkC;AACtC,eAAW,KAAK,cAAc;AAC5B,YAAM,MAAM,IAAI,IAAI,QAAQ,IAAI,CAAC,EAAG,KAAK,CAAC;AAC1C,UAAI,gBAAgB,MAAM;AACxB,sBAAc;AAAA,MAChB,OAAO;AACL,cAAM,OAAoB;AAC1B,sBAAc,IAAI,IAAI,CAAC,GAAG,IAAI,EAAE,OAAO,CAAC,MAAM,IAAI,IAAI,CAAC,CAAC,CAAC;AAAA,MAC3D;AAAA,IACF;AACA,UAAM,cAAc,CAAC,GAAI,eAAe,oBAAI,IAAY,CAAE,EAAE,KAAK;AACjE,QAAI,YAAY,SAAS,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,yCAAyC,GAAG,SAAS,YAAY,MAAM,yBAAyB,aAAa,MAAM;AAAA,MACrH;AAAA,IACF;AAEA,UAAM,SAAqB,YAAY;AAAA,MAAI,CAAC,WAC1C,aAAa,IAAI,CAAC,MAAM,QAAQ,IAAI,CAAC,EAAG,IAAI,MAAM,CAAE;AAAA,IACtD;AACA,UAAM,YAAY,oBAAoB,QAAQ,IAAI;AAClD,iBAAa,KAAK;AAAA,MAChB,GAAG;AAAA,MACH,WAAW;AAAA,MACX,SAAS;AAAA,MACT,UAAU,CAAC,GAAG,YAAY;AAAA,IAC5B,CAAC;AACD,QAAI,OAAO,SAAS,UAAU,GAAG,EAAG,MAAK,KAAK,UAAU,GAAG;AAC3D,QAAI,OAAO,SAAS,UAAU,aAAa,EAAG,QAAO,KAAK,UAAU,aAAa;AAAA,EACnF;AAEA,QAAM,OAAO,CAAC,OACZ,GAAG,WAAW,IAAI,OAAO,MAAM,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AACpE,SAAO;AAAA,IACL;AAAA,IACA,YAAY,KAAK,IAAI;AAAA,IACrB,sBAAsB,KAAK,MAAM;AAAA,IACjC,YAAY;AAAA,IACZ,UAAU;AAAA,EACZ;AACF;AAUO,SAAS,yCACd,aACA,OAA+B,CAAC,GACT;AACvB,QAAM,UAA+B,CAAC;AACtC,QAAM,OAAO,oBAAI,IAAY;AAC7B,aAAW,EAAE,QAAQ,OAAO,KAAK,aAAa;AAC5C,QAAI,KAAK,IAAI,MAAM,GAAG;AACpB,YAAM,IAAI;AAAA,QACR,+DAA+D,MAAM;AAAA,MACvE;AAAA,IACF;AACA,SAAK,IAAI,MAAM;AACf,eAAW,KAAK,QAAQ;AACtB,cAAQ,KAAK;AAAA,QACX;AAAA,QACA,WAAW,EAAE;AAAA,QACb,WAAW,EAAE;AAAA,QACb,OAAO,EAAE;AAAA,MACX,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO,0BAA0B,SAAS,IAAI;AAChD;AAGA,SAAS,YAAY,GAAW,IAAoB;AAClD,MAAI,MAAM,EAAG,QAAO;AACpB,MAAI,KAAK,IAAK,QAAO,UAAU,CAAC;AAChC,QAAM,IAAI,MAAM,KAAK,IAAI;AACzB,QAAM,IAAI,KAAK;AACf,QAAM,IAAI;AACV,QAAM,KAAK,eAAe,GAAG,GAAG,CAAC;AACjC,SAAO,KAAK,IAAI,IAAI,MAAM,KAAK,MAAM;AACvC;AAGA,SAAS,eAAe,GAAW,GAAW,GAAmB;AAC/D,MAAI,KAAK,EAAG,QAAO;AACnB,MAAI,KAAK,EAAG,QAAO;AACnB,QAAM,SAAS,QAAQ,CAAC,IAAI,QAAQ,CAAC,IAAI,QAAQ,IAAI,CAAC;AACtD,QAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,IAAI,IAAI,MAAM,IAAI;AACzE,QAAM,UAAU;AAChB,QAAM,MAAM;AACZ,MAAI,IAAI;AACR,MAAI,IAAI,KAAM,IAAI,KAAK,KAAM,IAAI;AACjC,MAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,MAAI,IAAI;AACR,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,KAAK,SAAS,KAAK;AACjC,UAAM,KAAK,IAAI;AACf,QAAI,MAAO,KAAK,IAAI,KAAK,MAAO,IAAI,KAAK,MAAM,IAAI;AACnD,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,SAAK,IAAI;AACT,UAAM,GAAG,IAAI,MAAM,IAAI,IAAI,KAAK,OAAO,IAAI,OAAO,IAAI,KAAK;AAC3D,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,UAAM,QAAQ,IAAI;AAClB,SAAK;AACL,QAAI,KAAK,IAAI,QAAQ,CAAC,IAAI,IAAK;AAAA,EACjC;AACA,SAAO,QAAQ;AACjB;AAGA,SAAS,QAAQ,GAAmB;AAClC,QAAM,IAAI;AACV,QAAM,QAAQ;AAAA,IACZ;AAAA,IAAqB;AAAA,IAAmB;AAAA,IAAqB;AAAA,IAC7D;AAAA,IAAqB;AAAA,IAAoB;AAAA,IAAsB;AAAA,IAC/D;AAAA,EACF;AACA,MAAI,IAAI,KAAK;AACX,WAAO,KAAK,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,QAAQ,IAAI,CAAC;AAAA,EAClE;AACA,OAAK;AACL,MAAI,IAAI,MAAM,CAAC;AACf,WAAS,IAAI,GAAG,IAAI,IAAI,GAAG,IAAK,MAAK,MAAM,CAAC,KAAM,IAAI;AACtD,QAAM,IAAI,IAAI,IAAI;AAClB,SAAO,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,KAAK,IAAI,OAAO,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,CAAC;AAC/E;AAGA,SAAS,UAAU,GAAmB;AACpC,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,IAAI;AAEV,QAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,QAAM,OAAO,KAAK,IAAI,CAAC;AACvB,QAAM,IAAI,KAAK,IAAI,IAAI;AACvB,QAAM,IAAI,QAAQ,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,KAAK,IAAK,CAAC,OAAO,OAAQ,CAAC;AAE9F,SAAO,OAAO,IAAI,OAAO;AAC3B;AAUO,SAAS,mBAAmB,MAKxB;AACT,QAAM,SAAS,KAAK;AACpB,MAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,EAAG,QAAO;AACpD,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;AAC7D,QAAM,QAAQ,UAAU,KAAK;AAC7B,QAAM,IAAI,MAAM,SAAS,SAAS,WAAW;AAC7C,SAAO,KAAK,KAAK,CAAC;AACpB;AAQO,SAAS,UAAU,MAKf;AACT,MAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,WAAW,EAAG,QAAO;AAChE,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;AAC7D,QAAM,QAAQ,UAAU,KAAK;AAC7B,UAAQ,SAAS,SAAS,KAAK,KAAK,KAAK,OAAO;AAClD;AAGO,SAAS,WACd,SACA,QAAQ,MACwC;AAChD,QAAM,IAAI,QAAQ;AAClB,QAAM,WAAW,QAAQ,IAAI,CAAC,MAAM,KAAK,IAAI,GAAG,IAAI,CAAC,CAAC;AACtD,QAAM,cAAc,SAAS,IAAI,CAAC,MAAM,IAAI,KAAK;AACjD,SAAO,EAAE,UAAU,YAAY;AACjC;AAMO,SAAS,kBACd,SACA,MAAM,MACyC;AAC/C,QAAM,IAAI,QAAQ;AAClB,MAAI,MAAM,EAAG,QAAO,EAAE,SAAS,CAAC,GAAG,aAAa,CAAC,EAAE;AACnD,QAAM,UAAU,QAAQ,IAAI,CAAC,GAAG,OAAO,EAAE,GAAG,EAAE,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AAC1E,QAAM,IAAI,IAAI,MAAc,CAAC;AAC7B,MAAI,WAAW;AACf,WAAS,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK;AAC/B,UAAM,OAAO,IAAI;AACjB,UAAM,QAAQ,QAAQ,CAAC;AACvB,UAAM,MAAO,MAAM,IAAI,IAAK;AAC5B,UAAM,UAAU,KAAK,IAAI,UAAU,GAAG;AACtC,eAAW;AACX,MAAE,MAAM,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO;AAAA,EAClC;AACA,QAAM,cAAc,EAAE,IAAI,CAAC,MAAM,IAAI,GAAG;AACxC,SAAO,EAAE,SAAS,GAAG,YAAY;AACnC;AAsCO,SAAS,gBACd,QACA,OACA,OAA+B,CAAC,GACT;AACvB,MAAI,OAAO,WAAW,MAAM,QAAQ;AAClC,UAAM,IAAI,MAAM,0CAA0C,OAAO,MAAM,OAAO,MAAM,MAAM,GAAG;AAAA,EAC/F;AACA,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,YAAY,KAAK,aAAa;AACpC,MAAI,cAAc,KAAK,cAAc,GAAG;AACtC,UAAM,IAAI,MAAM,qDAAqD,UAAU,EAAE;AAAA,EACnF;AAEA,QAAM,IAAI,OAAO;AACjB,QAAM,SAAS,OAAO,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,IAAK,CAAC;AACjD,MAAI,MAAM,GAAG;AACX,WAAO,EAAE,GAAG,GAAG,QAAQ,GAAG,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG,YAAY,UAAU;AAAA,EAC5E;AACA,MAAI,MAAM,GAAG;AACX,UAAM,IAAI,OAAO,CAAC;AAClB,WAAO,EAAE,GAAG,GAAG,QAAQ,GAAG,MAAM,GAAG,KAAK,GAAG,MAAM,GAAG,YAAY,UAAU;AAAA,EAC5E;AAEA,QAAM,MAAM,QAAQ,KAAK,IAAI;AAC7B,QAAM,UAAU,IAAI,MAAc,SAAS;AAC3C,WAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,QAAI,cAAc,QAAQ;AACxB,UAAI,MAAM;AACV,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,eAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,CAAC;AAAA,MACrC;AACA,cAAQ,CAAC,IAAI,MAAM;AAAA,IACrB,OAAO;AACL,YAAM,MAAM,IAAI,MAAc,CAAC;AAC/B,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAI,CAAC,IAAI,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,CAAC;AAAA,MACvC;AACA,cAAQ,CAAC,IAAI,cAAc,GAAG;AAAA,IAChC;AAAA,EACF;AACA,UAAQ,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAE5B,QAAM,QAAQ,IAAI;AAClB,QAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;AACjD,QAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;AAElF,SAAO;AAAA,IACL;AAAA,IACA,QAAQ,cAAc,CAAC,GAAG,MAAM,CAAC;AAAA,IACjC,MAAM,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAAA,IAC1C,KAAK,QAAQ,MAAM;AAAA,IACnB,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM,CAAC;AAAA,IACvC;AAAA,IACA;AAAA,EACF;AACF;AAKA,SAAS,UAAU,GAAmB;AACpC,MAAI,KAAK,KAAK,KAAK,GAAG;AACpB,QAAI,MAAM,EAAG,QAAO;AACpB,QAAI,MAAM,EAAG,QAAO;AACpB,WAAO;AAAA,EACT;AACA,QAAM,IAAI;AAAA,IACR;AAAA,IAAsB;AAAA,IAAqB;AAAA,IAAsB;AAAA,IACjE;AAAA,IAAsB;AAAA,EACxB;AACA,QAAM,IAAI;AAAA,IACR;AAAA,IAAsB;AAAA,IAAqB;AAAA,IAAsB;AAAA,IACjE;AAAA,EACF;AACA,QAAM,IAAI;AAAA,IACR;AAAA,IAAuB;AAAA,IAAuB;AAAA,IAAoB;AAAA,IAClE;AAAA,IAAmB;AAAA,EACrB;AACA,QAAM,IAAI,CAAC,sBAAsB,oBAAsB,mBAAmB,iBAAiB;AAC3F,QAAM,OAAO;AACb,QAAM,QAAQ,IAAI;AAClB,MAAI;AACJ,MAAI;AACJ,MAAI,IAAI,MAAM;AACZ,QAAI,KAAK,KAAK,KAAK,KAAK,IAAI,CAAC,CAAC;AAC9B,gBACO,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,SACpE,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI;AAAA,EAE3D;AACA,MAAI,KAAK,OAAO;AACd,QAAI,IAAI;AACR,QAAI,IAAI;AACR,gBACQ,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,SAC1E,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI;AAAA,EAEzE;AACA,MAAI,KAAK,KAAK,KAAK,KAAK,IAAI,IAAI,CAAC,CAAC;AAClC,SACE,MAAM,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,SACrE,EAAE,CAAC,IAAK,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI,EAAE,CAAC,KAAM,IAAI;AAE3D;AAEA,SAAS,cAAc,IAAsB;AAC3C,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,KAAG,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,QAAM,MAAM,KAAK,MAAM,GAAG,SAAS,CAAC;AACpC,SAAO,GAAG,SAAS,MAAM,KAAK,GAAG,MAAM,CAAC,IAAK,GAAG,GAAG,KAAM,IAAI,GAAG,GAAG;AACrE;AAGA,SAAS,QAAQ,MAAwC;AACvD,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,OAAO,KAAK;AACpB,SAAO,MAAM;AACX,QAAK,IAAI,aAAc;AACvB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;","names":["i","i"]}

package/dist/chunk-YV7J7X5N.js

@@ -0,0 +1,313 @@
+// src/rl/verifiable-reward.ts
+var DEFAULT_DETERMINISTIC_LAYERS = /* @__PURE__ */ new Set([
+  "install",
+  "typecheck",
+  "build",
+  "lint",
+  "test",
+  "compile",
+  "schema",
+  "sandbox",
+  "unit_tests",
+  "integration_tests"
+]);
+var DEFAULT_SOURCE_FOR = (name) => {
+  const lower = name.toLowerCase();
+  if (lower.includes("test")) return "test";
+  if (lower.includes("compile") || lower.includes("build") || lower.includes("typecheck") || lower.includes("lint"))
+    return "compile";
+  if (lower.includes("schema")) return "schema";
+  if (lower.includes("sandbox")) return "sandbox";
+  if (lower.includes("judge") || lower.includes("semantic")) return "judge";
+  return "composite";
+};
+function extractVerifiableReward(report, opts = {}) {
+  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
+  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
+  const fallbackToJudge = opts.fallbackToJudge ?? true;
+  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7;
+  const deterministic = report.layers.filter(
+    (l) => deterministicSet.has(l.layer) && typeof l.score === "number" && Number.isFinite(l.score)
+  );
+  if (deterministic.length === 1) {
+    const layer = deterministic[0];
+    return {
+      value: clamp01(layer.score),
+      source: sourceFor(layer.layer),
+      determinism: "deterministic",
+      confidence: 1,
+      origin: layer.layer,
+      breakdown: layerBreakdown(layer)
+    };
+  }
+  if (deterministic.length > 1) {
+    let num = 0;
+    let denom = 0;
+    const breakdown = {};
+    for (const l of deterministic) {
+      const w = l.detail?.weight ?? 1;
+      num += w * (l.score ?? 0);
+      denom += w;
+      breakdown[l.layer] = l.score;
+    }
+    return {
+      value: denom === 0 ? 0 : clamp01(num / denom),
+      source: "composite",
+      determinism: "deterministic",
+      confidence: 1,
+      origin: deterministic.map((l) => l.layer).join("+"),
+      breakdown
+    };
+  }
+  if (!fallbackToJudge) return null;
+  const judge = report.layers.find(
+    (l) => typeof l.score === "number" && Number.isFinite(l.score) && sourceFor(l.layer) === "judge"
+  ) ?? report.layers.find((l) => typeof l.score === "number" && Number.isFinite(l.score));
+  if (!judge) return null;
+  const confFromDetail = judge.detail?.confidence;
+  return {
+    value: clamp01(judge.score),
+    source: "judge",
+    determinism: "probabilistic",
+    confidence: typeof confFromDetail === "number" ? confFromDetail : judgeFloor,
+    origin: judge.layer,
+    breakdown: layerBreakdown(judge)
+  };
+}
+function extractVerifiableRewardsFromRecords(runs, opts = {}) {
+  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
+  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
+  const fallbackToJudge = opts.fallbackToJudge ?? true;
+  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7;
+  return runs.map((run) => {
+    const layerScores = [];
+    for (const [k, v] of Object.entries(run.outcome.raw)) {
+      if (k.startsWith("layer.") && !k.includes(".", 6) && typeof v === "number" && Number.isFinite(v)) {
+        layerScores.push({ name: k.slice("layer.".length), score: v });
+      }
+    }
+    const det = layerScores.filter((l) => deterministicSet.has(l.name));
+    if (det.length === 1) {
+      const layer = det[0];
+      return {
+        runId: run.runId,
+        reward: {
+          value: clamp01(layer.score),
+          source: sourceFor(layer.name),
+          determinism: "deterministic",
+          confidence: 1,
+          origin: layer.name
+        }
+      };
+    }
+    if (det.length > 1) {
+      const value = det.reduce((s, l) => s + l.score, 0) / det.length;
+      const breakdown = Object.fromEntries(
+        det.map((l) => [l.name, l.score])
+      );
+      return {
+        runId: run.runId,
+        reward: {
+          value: clamp01(value),
+          source: "composite",
+          determinism: "deterministic",
+          confidence: 1,
+          origin: det.map((l) => l.name).join("+"),
+          breakdown
+        }
+      };
+    }
+    if (!fallbackToJudge) return { runId: run.runId, reward: null };
+    const primary = run.outcome.holdoutScore ?? run.outcome.searchScore;
+    if (typeof primary !== "number" || !Number.isFinite(primary)) {
+      return { runId: run.runId, reward: null };
+    }
+    return {
+      runId: run.runId,
+      reward: {
+        value: clamp01(primary),
+        source: "judge",
+        determinism: "probabilistic",
+        confidence: judgeFloor,
+        origin: "run.outcome.score"
+      }
+    };
+  });
+}
+function filterDeterministicallyRewarded(runs, opts = {}) {
+  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false });
+  const out = [];
+  for (let i = 0; i < runs.length; i++) {
+    const r = rewarded[i];
+    if (r.reward && r.reward.determinism === "deterministic") {
+      out.push({ run: runs[i], reward: r.reward });
+    }
+  }
+  return out;
+}
+function clamp01(x) {
+  if (!Number.isFinite(x)) return 0;
+  return Math.max(0, Math.min(1, x));
+}
+function layerBreakdown(l) {
+  const out = {};
+  if (l.diagnostics) {
+    for (const [k, v] of Object.entries(l.diagnostics)) {
+      if (typeof v === "number" && Number.isFinite(v)) out[k] = v;
+    }
+  }
+  return out;
+}
+
+// src/rl/reward-hacking.ts
+var DEFAULT_PROXY = (r) => {
+  const v = r.outcome.holdoutScore ?? r.outcome.searchScore;
+  return typeof v === "number" && Number.isFinite(v) ? v : null;
+};
+function detectRewardHacking(input) {
+  const proxyOf = input.proxyOf ?? DEFAULT_PROXY;
+  const truthOf = input.truthOf;
+  const sus = input.thresholds?.suspect ?? 0.3;
+  const gam = input.thresholds?.gaming ?? 0.6;
+  const runs = input.runs.filter((r) => proxyOf(r) !== null);
+  const n = runs.length;
+  if (n < 4) {
+    return {
+      findings: [],
+      verdict: "clean",
+      n,
+      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`]
+    };
+  }
+  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)));
+  const before = runs.slice(0, n - windowSize);
+  const after = runs.slice(n - windowSize);
+  const findings = [];
+  if (truthOf) {
+    const beforeProxy = before.map(proxyOf).filter((v) => typeof v === "number");
+    const afterProxy = after.map(proxyOf).filter((v) => typeof v === "number");
+    const beforeTruth = before.map(truthOf).filter((v) => typeof v === "number");
+    const afterTruth = after.map(truthOf).filter((v) => typeof v === "number");
+    if (beforeProxy.length >= 2 && afterProxy.length >= 2 && beforeTruth.length >= 2 && afterTruth.length >= 2) {
+      const proxyDelta = mean(afterProxy) - mean(beforeProxy);
+      const truthDelta = mean(afterTruth) - mean(beforeTruth);
+      const gap = Math.max(0, proxyDelta - truthDelta);
+      const severity = clamp012(gap * 5);
+      findings.push({
+        signal: "reward_divergence",
+        severity,
+        message: severity >= sus ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} \u2014 potential Goodhart` : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,
+        detail: {
+          proxyDelta,
+          truthDelta,
+          gap,
+          beforeN: beforeProxy.length,
+          afterN: afterProxy.length
+        }
+      });
+    }
+  }
+  {
+    const beforeP = before.map(proxyOf).filter((v) => typeof v === "number");
+    const afterP = after.map(proxyOf).filter((v) => typeof v === "number");
+    if (beforeP.length >= 4 && afterP.length >= 4) {
+      const ks = ksStatistic(beforeP, afterP);
+      const severity = clamp012(ks - 0.2);
+      findings.push({
+        signal: "distribution_shift",
+        severity,
+        message: severity >= sus ? `KS=${ks.toFixed(3)} between before/after windows \u2014 distributional shift large` : `KS=${ks.toFixed(3)} between before/after windows \u2014 within-distribution drift`,
+        detail: { ks, beforeN: beforeP.length, afterN: afterP.length }
+      });
+    }
+  }
+  {
+    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions);
+    const aligned = runs.map((r) => ({ p: proxyOf(r), s: secondaryOf(r) })).filter(
+      (x) => typeof x.p === "number" && typeof x.s === "number"
+    );
+    if (aligned.length >= 4) {
+      const ps = aligned.map((x) => x.p);
+      const ss = aligned.map((x) => x.s);
+      const r = pearsonR(ps, ss);
+      const severity = clamp012(0.5 - Math.max(0, r));
+      findings.push({
+        signal: "reward_disagreement",
+        severity,
+        message: severity >= sus ? `proxy and independent secondary reward correlate \u03C1=${r.toFixed(3)} \u2014 possibly hacking proxy` : `proxy and secondary reward correlate \u03C1=${r.toFixed(3)}`,
+        detail: { pearson: r, n: aligned.length }
+      });
+    }
+  }
+  {
+    const detRuns = filterDeterministicallyRewarded(runs, input.verifiableRewardOptions ?? {});
+    if (detRuns.length >= 4) {
+      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2));
+      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2));
+      const detDelta = mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value));
+      const proxyDelta = mean(after.map(proxyOf).filter((v) => typeof v === "number")) - mean(before.map(proxyOf).filter((v) => typeof v === "number"));
+      const driftGap = Math.max(0, proxyDelta - detDelta);
+      const severity = clamp012(driftGap * 5);
+      findings.push({
+        signal: "judge_drift",
+        severity,
+        message: severity >= sus ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} \u2014 judge drifting up without verifiable backing` : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,
+        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length }
+      });
+    }
+  }
+  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0);
+  const verdict = maxSev >= gam ? "gaming" : maxSev >= sus ? "suspect" : "clean";
+  const rationale = findings.filter((f) => f.severity >= sus).map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} \u2014 ${f.message}`);
+  if (rationale.length === 0) rationale.push("no signals fired above suspect threshold");
+  return { findings, verdict, rationale, n };
+}
+function mean(xs) {
+  if (xs.length === 0) return 0;
+  return xs.reduce((s, x) => s + x, 0) / xs.length;
+}
+function clamp012(x) {
+  if (!Number.isFinite(x)) return 0;
+  return Math.max(0, Math.min(1, x));
+}
+function pearsonR(a, b) {
+  if (a.length !== b.length || a.length < 2) return 0;
+  const ma = mean(a);
+  const mb = mean(b);
+  let num = 0, da = 0, db = 0;
+  for (let i = 0; i < a.length; i++) {
+    const xa = a[i] - ma;
+    const xb = b[i] - mb;
+    num += xa * xb;
+    da += xa * xa;
+    db += xb * xb;
+  }
+  if (da === 0 || db === 0) return 0;
+  return num / Math.sqrt(da * db);
+}
+function ksStatistic(a, b) {
+  const sortedA = [...a].sort((x, y) => x - y);
+  const sortedB = [...b].sort((x, y) => x - y);
+  const all = [.../* @__PURE__ */ new Set([...sortedA, ...sortedB])].sort((x, y) => x - y);
+  let max = 0;
+  for (const v of all) {
+    const fa = sortedA.filter((x) => x <= v).length / sortedA.length;
+    const fb = sortedB.filter((x) => x <= v).length / sortedB.length;
+    max = Math.max(max, Math.abs(fa - fb));
+  }
+  return max;
+}
+function defaultSecondary(verifiableOpts) {
+  return (run) => {
+    const filtered = filterDeterministicallyRewarded([run], verifiableOpts ?? {});
+    return filtered.length === 1 ? filtered[0].reward.value : null;
+  };
+}
+
+export {
+  extractVerifiableReward,
+  extractVerifiableRewardsFromRecords,
+  filterDeterministicallyRewarded,
+  detectRewardHacking
+};
+//# sourceMappingURL=chunk-YV7J7X5N.js.map

package/dist/chunk-YV7J7X5N.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"sourcesContent":["/**\n * Verifiable reward channel.\n *\n * For RL on coding / math / theorem-proving / structured-output tasks, the\n * reward signal is *decidable* — a test passes or fails, a proof checks or\n * doesn't, an output validates against a schema or doesn't. These rewards\n * are dramatically more useful for RL training than LLM-judge scores\n * because they don't drift, can't be Goodhart-gamed by the policy in the\n * same way, and don't require a separate calibration loop.\n *\n * The `MultiLayerVerifier` already produces this signal — it just doesn't\n * surface it in a shape that's clean enough for RL training. This module\n * wraps the verifier output so consumers can:\n *\n *   1. Extract a clean `VerifiableReward` from a `VerificationReport`\n *   2. Distinguish *deterministic* rewards (compile, test, schema) from\n *      *probabilistic* rewards (judge) so they can be weighted differently\n *      in the RL training step\n *   3. Filter `RunRecord[]` to only those with a verifiable reward,\n *      producing the clean training set that DeepSeek-R1-style GRPO and\n *      AlphaProof-style search both depend on\n *\n * Why this matters: every credible 2025-2026 frontier RL result on coding\n * agents leans on verifiable reward (DeepSeek-R1 GRPO on test pass-rate,\n * o-series RL on math/code, AlphaProof on Lean kernel checking). Mixing\n * judge scores into the reward signal poisons the gradient. This module\n * is the seam.\n */\n\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport type { RunRecord } from '../run-record'\n\nexport type VerifiableRewardSource =\n  | 'compile' // typecheck / build / lint passed\n  | 'test' // unit / integration test pass-rate\n  | 'schema' // structured output validates\n  | 'sandbox' // sandbox exec exit code\n  | 'judge' // LLM judge — probabilistic, included for completeness\n  | 'composite' // weighted blend across multiple of the above\n\nexport interface VerifiableReward {\n  /** Scalar in [0, 1]. The RL training signal. */\n  value: number\n  /** What produced the reward — different sources have different determinism. */\n  source: VerifiableRewardSource\n  /**\n   * Determinism class. `'deterministic'` rewards are repeatable byte-for-byte\n   * given the same inputs (compile, test, schema validation, sandbox exit code).\n   * `'probabilistic'` rewards depend on a stochastic component (LLM judge).\n   * Mixing these in the same training batch without separation is a known\n   * footgun in production RLHF pipelines.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /**\n   * Confidence in the reward value. For deterministic sources this is 1.0\n   * (the bit either flipped or didn't). For judge sources this is the\n   * judge-reported confidence or — when missing — a calibrated prior.\n   */\n  confidence: number\n  /** The layer / judge id that produced the signal, for provenance. */\n  origin: string\n  /**\n   * Any per-source breakdown the consumer might want — e.g. `{ tests_passed: 7, tests_total: 10 }`.\n   */\n  breakdown?: Record<string, number>\n}\n\nexport interface VerifiableRewardExtractionOptions {\n  /**\n   * Which layers count as deterministic-reward sources. The verifier doesn't\n   * tag layers as \"this is verifiable\"; the caller declares it via this list\n   * (or via the layer name → source mapping). Default treats common names\n   * (`install`, `typecheck`, `build`, `lint`, `test`, `compile`, `schema`,\n   * `sandbox`) as deterministic.\n   */\n  deterministicLayers?: string[]\n  /**\n   * Map layer name → reward source. Defaults to a sensible string-match.\n   */\n  sourceFor?: (layerName: string) => VerifiableRewardSource\n  /**\n   * Whether to fall back to a probabilistic (judge) reward when no\n   * deterministic layer produced a numeric score. Default `true`. Set to\n   * `false` for \"deterministic-only\" training pipelines that should\n   * discard runs without a verifiable signal.\n   */\n  fallbackToJudge?: boolean\n  /**\n   * Default confidence for probabilistic (judge) rewards when the judge\n   * doesn't report one. Default `0.7`.\n   */\n  judgeConfidenceFloor?: number\n}\n\nconst DEFAULT_DETERMINISTIC_LAYERS = new Set([\n  'install',\n  'typecheck',\n  'build',\n  'lint',\n  'test',\n  'compile',\n  'schema',\n  'sandbox',\n  'unit_tests',\n  'integration_tests',\n])\n\nconst DEFAULT_SOURCE_FOR = (name: string): VerifiableRewardSource => {\n  const lower = name.toLowerCase()\n  if (lower.includes('test')) return 'test'\n  if (\n    lower.includes('compile') ||\n    lower.includes('build') ||\n    lower.includes('typecheck') ||\n    lower.includes('lint')\n  )\n    return 'compile'\n  if (lower.includes('schema')) return 'schema'\n  if (lower.includes('sandbox')) return 'sandbox'\n  if (lower.includes('judge') || lower.includes('semantic')) return 'judge'\n  return 'composite'\n}\n\n/**\n * Extract a `VerifiableReward` from a `VerificationReport`.\n *\n * Strategy: prefer the deterministic layers (in order: test → compile →\n * schema → sandbox), fall back to the judge layer if `fallbackToJudge` is\n * true, return `null` if no signal qualifies. When multiple deterministic\n * layers contribute, return a `'composite'` source with a weighted blend.\n */\nexport function extractVerifiableReward(\n  report: VerificationReport,\n  opts: VerifiableRewardExtractionOptions = {},\n): VerifiableReward | null {\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  const deterministic = report.layers.filter(\n    (l) => deterministicSet.has(l.layer) && typeof l.score === 'number' && Number.isFinite(l.score),\n  )\n\n  if (deterministic.length === 1) {\n    const layer = deterministic[0]!\n    return {\n      value: clamp01(layer.score!),\n      source: sourceFor(layer.layer),\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: layer.layer,\n      breakdown: layerBreakdown(layer),\n    }\n  }\n\n  if (deterministic.length > 1) {\n    // Composite: weighted blend by `Layer.weight` if present, else equal.\n    let num = 0\n    let denom = 0\n    const breakdown: Record<string, number> = {}\n    for (const l of deterministic) {\n      const w = (l.detail?.weight as number | undefined) ?? 1\n      num += w * (l.score ?? 0)\n      denom += w\n      breakdown[l.layer] = l.score!\n    }\n    return {\n      value: denom === 0 ? 0 : clamp01(num / denom),\n      source: 'composite',\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: deterministic.map((l) => l.layer).join('+'),\n      breakdown,\n    }\n  }\n\n  if (!fallbackToJudge) return null\n\n  const judge =\n    report.layers.find(\n      (l) =>\n        typeof l.score === 'number' && Number.isFinite(l.score) && sourceFor(l.layer) === 'judge',\n    ) ?? report.layers.find((l) => typeof l.score === 'number' && Number.isFinite(l.score))\n\n  if (!judge) return null\n\n  const confFromDetail = judge.detail?.confidence as number | undefined\n  return {\n    value: clamp01(judge.score!),\n    source: 'judge',\n    determinism: 'probabilistic',\n    confidence: typeof confFromDetail === 'number' ? confFromDetail : judgeFloor,\n    origin: judge.layer,\n    breakdown: layerBreakdown(judge),\n  }\n}\n\n/**\n * Extract verifiable rewards from `RunRecord[]` produced via the\n * `verificationReportToRunRecord` adapter (which encodes per-layer scores\n * in `outcome.raw['layer.<name>']`). For records that don't carry layer\n * scores, returns `null` for that record.\n *\n * This is the canonical bridge from \"campaign-shaped artifacts\" to\n * \"RL-training-ready reward signals\": every record that has a clean\n * verifiable reward becomes a training datum, every record that doesn't\n * gets filtered out (or kept with `'probabilistic'` determinism for\n * separate downstream handling).\n */\nexport function extractVerifiableRewardsFromRecords(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ runId: string; reward: VerifiableReward | null }> {\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  return runs.map((run) => {\n    // Recover per-layer scores from outcome.raw['layer.<name>']\n    const layerScores: Array<{ name: string; score: number }> = []\n    for (const [k, v] of Object.entries(run.outcome.raw)) {\n      if (\n        k.startsWith('layer.') &&\n        !k.includes('.', 6) &&\n        typeof v === 'number' &&\n        Number.isFinite(v)\n      ) {\n        layerScores.push({ name: k.slice('layer.'.length), score: v })\n      }\n    }\n    const det = layerScores.filter((l) => deterministicSet.has(l.name))\n\n    if (det.length === 1) {\n      const layer = det[0]!\n      return {\n        runId: run.runId,\n        reward: {\n          value: clamp01(layer.score),\n          source: sourceFor(layer.name),\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: layer.name,\n        },\n      }\n    }\n    if (det.length > 1) {\n      const value = det.reduce((s, l) => s + l.score, 0) / det.length\n      const breakdown: Record<string, number> = Object.fromEntries(\n        det.map((l) => [l.name, l.score]),\n      )\n      return {\n        runId: run.runId,\n        reward: {\n          value: clamp01(value),\n          source: 'composite',\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: det.map((l) => l.name).join('+'),\n          breakdown,\n        },\n      }\n    }\n    if (!fallbackToJudge) return { runId: run.runId, reward: null }\n\n    // Probabilistic fallback: use the run's primary score.\n    const primary = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (typeof primary !== 'number' || !Number.isFinite(primary)) {\n      return { runId: run.runId, reward: null }\n    }\n    return {\n      runId: run.runId,\n      reward: {\n        value: clamp01(primary),\n        source: 'judge',\n        determinism: 'probabilistic',\n        confidence: judgeFloor,\n        origin: 'run.outcome.score',\n      },\n    }\n  })\n}\n\n/** Filter `RunRecord[]` to those with deterministic verifiable rewards. */\nexport function filterDeterministicallyRewarded(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ run: RunRecord; reward: VerifiableReward }> {\n  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false })\n  const out: Array<{ run: RunRecord; reward: VerifiableReward }> = []\n  for (let i = 0; i < runs.length; i++) {\n    const r = rewarded[i]!\n    if (r.reward && r.reward.determinism === 'deterministic') {\n      out.push({ run: runs[i]!, reward: r.reward })\n    }\n  }\n  return out\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction layerBreakdown(l: LayerResult): Record<string, number> {\n  const out: Record<string, number> = {}\n  if (l.diagnostics) {\n    for (const [k, v] of Object.entries(l.diagnostics)) {\n      if (typeof v === 'number' && Number.isFinite(v)) out[k] = v\n    }\n  }\n  return out\n}\n","/**\n * Reward hacking / Goodhart detection.\n *\n * Goodhart's Law says: when a measure becomes a target, it ceases to be\n * a good measure. In RLHF and agentic-RL settings this is the dominant\n * failure mode — the policy learns to produce outputs that score well on\n * the proxy reward (judge, rubric, test pass-rate) without producing\n * the underlying capability the proxy was meant to track.\n *\n * Krakovna et al. (2020, \"Specification Gaming Examples in AI\") and the\n * subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.\n * 2023) converge on a few diagnostic signatures:\n *\n *   1. **Reward divergence:** the proxy reward grows while the held-out\n *      ground-truth signal stagnates or drops. Predictive validity over\n *      time captures this.\n *   2. **Distributional shift in outputs:** after RL, the policy produces\n *      outputs that no longer match the reference distribution — usually\n *      because it found a high-reward attractor that's degenerate (e.g.\n *      one-token responses, repetition, formatting tricks).\n *   3. **Disagreement between independent rewards:** if you train on\n *      reward A and a held-out independent reward B drops sharply, you're\n *      probably hacking A.\n *   4. **Calibration drift:** the verifiable / deterministic component of\n *      the reward is stable; the probabilistic / judge component drifts up\n *      while the deterministic component doesn't. The judge is being\n *      gamed.\n *\n * This module ships explicit detectors for all four signatures, plus a\n * combined verdict. The output is diagnostic — actionable signals,\n * not autoreject — because each signature has known false positives\n * (e.g., a policy that genuinely improves can show distributional shift).\n *\n * Differs from `rubricPredictiveValidity` (which is a *standing* check on\n * whether rubrics correlate with deployment outcomes) — this is a\n * *temporal* check on whether the reward-vs-truth gap is *widening over\n * time during a training run*.\n */\n\nimport type { RunRecord } from '../run-record'\nimport {\n  filterDeterministicallyRewarded,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport type RewardHackingSignal =\n  | 'reward_divergence'\n  | 'distribution_shift'\n  | 'reward_disagreement'\n  | 'judge_drift'\n\nexport interface RewardHackingFinding {\n  signal: RewardHackingSignal\n  /** Severity in [0, 1]. >0.5 = strong signal. */\n  severity: number\n  message: string\n  /** Numeric evidence the consumer can render. */\n  detail: Record<string, number>\n}\n\nexport interface RewardHackingReport {\n  findings: RewardHackingFinding[]\n  /**\n   * Composite verdict. `'clean'` if every signal severity < 0.3;\n   * `'suspect'` if at least one ≥ 0.3 but none ≥ 0.6; `'gaming'` if any ≥ 0.6.\n   */\n  verdict: 'clean' | 'suspect' | 'gaming'\n  /** Rationale for the verdict, ready to paste into an audit log. */\n  rationale: string[]\n  /** Number of paired (proxy, truth) data points the report saw. */\n  n: number\n}\n\nexport interface DetectRewardHackingInput {\n  /**\n   * Run records ordered by recency (oldest first). The detector segments\n   * them into prefix/suffix windows to compute \"did the gap widen.\"\n   */\n  runs: RunRecord[]\n  /**\n   * The metric the policy was trained to optimize. Should be present on\n   * `outcome.raw` or `outcome.holdoutScore`. Default reads `outcome.holdoutScore`.\n   */\n  proxyOf?: (run: RunRecord) => number | null\n  /**\n   * The held-out ground-truth metric. For RL on coding, this is typically\n   * test pass-rate. For RLHF, it's downstream task performance or human\n   * preference. For knowledge tasks, it's an independently-graded score.\n   */\n  truthOf?: (run: RunRecord) => number | null\n  /**\n   * Independent secondary reward. Used for the `reward_disagreement`\n   * signal. Default uses the verifiable reward extractor (deterministic\n   * sources only).\n   */\n  secondaryRewardOf?: (run: RunRecord) => number | null\n  /**\n   * Window size — how many of the most recent runs count as the \"after\"\n   * cohort. Default min(50, half the runs).\n   */\n  windowSize?: number\n  /**\n   * Severity threshold to flag a signal. Default 0.3 (suspect) and 0.6\n   * (gaming).\n   */\n  thresholds?: { suspect?: number; gaming?: number }\n  /**\n   * Verifiable-reward options used for the secondary-reward fallback.\n   */\n  verifiableRewardOptions?: VerifiableRewardExtractionOptions\n}\n\nconst DEFAULT_PROXY = (r: RunRecord): number | null => {\n  const v = r.outcome.holdoutScore ?? r.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nexport function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport {\n  const proxyOf = input.proxyOf ?? DEFAULT_PROXY\n  const truthOf = input.truthOf\n  const sus = input.thresholds?.suspect ?? 0.3\n  const gam = input.thresholds?.gaming ?? 0.6\n\n  const runs = input.runs.filter((r) => proxyOf(r) !== null)\n  const n = runs.length\n  if (n < 4) {\n    return {\n      findings: [],\n      verdict: 'clean',\n      n,\n      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`],\n    }\n  }\n  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)))\n  const before = runs.slice(0, n - windowSize)\n  const after = runs.slice(n - windowSize)\n\n  const findings: RewardHackingFinding[] = []\n\n  // ── Signal 1: reward divergence (proxy ↑ while truth flat or ↓) ──────\n  if (truthOf) {\n    const beforeProxy = before.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const afterProxy = after.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const beforeTruth = before.map(truthOf).filter((v): v is number => typeof v === 'number')\n    const afterTruth = after.map(truthOf).filter((v): v is number => typeof v === 'number')\n    if (\n      beforeProxy.length >= 2 &&\n      afterProxy.length >= 2 &&\n      beforeTruth.length >= 2 &&\n      afterTruth.length >= 2\n    ) {\n      const proxyDelta = mean(afterProxy) - mean(beforeProxy)\n      const truthDelta = mean(afterTruth) - mean(beforeTruth)\n      // Divergence: proxy goes up while truth goes flat or down.\n      // Severity = max(0, (proxyDelta - truthDelta)) — bigger gap = bigger signal.\n      const gap = Math.max(0, proxyDelta - truthDelta)\n      const severity = clamp01(gap * 5) // scale: 0.2 absolute gap → severity 1.0\n      findings.push({\n        signal: 'reward_divergence',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart`\n            : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,\n        detail: {\n          proxyDelta,\n          truthDelta,\n          gap,\n          beforeN: beforeProxy.length,\n          afterN: afterProxy.length,\n        },\n      })\n    }\n  }\n\n  // ── Signal 2: distributional shift in outputs (KS on score distributions) ──\n  {\n    const beforeP = before.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const afterP = after.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    if (beforeP.length >= 4 && afterP.length >= 4) {\n      const ks = ksStatistic(beforeP, afterP)\n      // KS statistic: bigger = more shift. We're agnostic about direction;\n      // genuine improvement ALSO produces shift, so this signal is\n      // contributory rather than load-bearing.\n      const severity = clamp01(ks - 0.2)\n      findings.push({\n        signal: 'distribution_shift',\n        severity,\n        message:\n          severity >= sus\n            ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large`\n            : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,\n        detail: { ks, beforeN: beforeP.length, afterN: afterP.length },\n      })\n    }\n  }\n\n  // ── Signal 3: reward disagreement (proxy vs independent secondary) ────\n  {\n    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions)\n    const aligned = runs\n      .map((r) => ({ p: proxyOf(r), s: secondaryOf(r) }))\n      .filter(\n        (x): x is { p: number; s: number } => typeof x.p === 'number' && typeof x.s === 'number',\n      )\n    if (aligned.length >= 4) {\n      const ps = aligned.map((x) => x.p)\n      const ss = aligned.map((x) => x.s)\n      const r = pearsonR(ps, ss)\n      // Disagreement: low or negative correlation between primary proxy\n      // reward and an independent secondary signal.\n      const severity = clamp01(0.5 - Math.max(0, r))\n      findings.push({\n        signal: 'reward_disagreement',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy`\n            : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,\n        detail: { pearson: r, n: aligned.length },\n      })\n    }\n  }\n\n  // ── Signal 4: judge drift (probabilistic up while deterministic flat) ─\n  {\n    const detRuns = filterDeterministicallyRewarded(runs, input.verifiableRewardOptions ?? {})\n    if (detRuns.length >= 4) {\n      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2))\n      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2))\n      const detDelta =\n        mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value))\n      const proxyDelta =\n        mean(after.map(proxyOf).filter((v): v is number => typeof v === 'number')) -\n        mean(before.map(proxyOf).filter((v): v is number => typeof v === 'number'))\n      const driftGap = Math.max(0, proxyDelta - detDelta)\n      const severity = clamp01(driftGap * 5)\n      findings.push({\n        signal: 'judge_drift',\n        severity,\n        message:\n          severity >= sus\n            ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing`\n            : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,\n        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length },\n      })\n    }\n  }\n\n  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0)\n  const verdict: RewardHackingReport['verdict'] =\n    maxSev >= gam ? 'gaming' : maxSev >= sus ? 'suspect' : 'clean'\n  const rationale = findings\n    .filter((f) => f.severity >= sus)\n    .map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`)\n  if (rationale.length === 0) rationale.push('no signals fired above suspect threshold')\n\n  return { findings, verdict, rationale, n }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction pearsonR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return 0\n  const ma = mean(a)\n  const mb = mean(b)\n  let num = 0,\n    da = 0,\n    db = 0\n  for (let i = 0; i < a.length; i++) {\n    const xa = a[i]! - ma\n    const xb = b[i]! - mb\n    num += xa * xb\n    da += xa * xa\n    db += xb * xb\n  }\n  if (da === 0 || db === 0) return 0\n  return num / Math.sqrt(da * db)\n}\n\nfunction ksStatistic(a: number[], b: number[]): number {\n  // Two-sample Kolmogorov-Smirnov statistic.\n  const sortedA = [...a].sort((x, y) => x - y)\n  const sortedB = [...b].sort((x, y) => x - y)\n  const all = [...new Set([...sortedA, ...sortedB])].sort((x, y) => x - y)\n  let max = 0\n  for (const v of all) {\n    const fa = sortedA.filter((x) => x <= v).length / sortedA.length\n    const fb = sortedB.filter((x) => x <= v).length / sortedB.length\n    max = Math.max(max, Math.abs(fa - fb))\n  }\n  return max\n}\n\nfunction defaultSecondary(\n  verifiableOpts?: VerifiableRewardExtractionOptions,\n): (run: RunRecord) => number | null {\n  return (run: RunRecord) => {\n    const filtered = filterDeterministicallyRewarded([run], verifiableOpts ?? {})\n    return filtered.length === 1 ? filtered[0]!.reward.value : null\n  }\n}\n"],"mappings":";AA8FA,IAAM,+BAA+B,oBAAI,IAAI;AAAA,EAC3C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,qBAAqB,CAAC,SAAyC;AACnE,QAAM,QAAQ,KAAK,YAAY;AAC/B,MAAI,MAAM,SAAS,MAAM,EAAG,QAAO;AACnC,MACE,MAAM,SAAS,SAAS,KACxB,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,MAAM;AAErB,WAAO;AACT,MAAI,MAAM,SAAS,QAAQ,EAAG,QAAO;AACrC,MAAI,MAAM,SAAS,SAAS,EAAG,QAAO;AACtC,MAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,UAAU,EAAG,QAAO;AAClE,SAAO;AACT;AAUO,SAAS,wBACd,QACA,OAA0C,CAAC,GAClB;AACzB,QAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;AAC9F,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,QAAM,aAAa,KAAK,wBAAwB;AAEhD,QAAM,gBAAgB,OAAO,OAAO;AAAA,IAClC,CAAC,MAAM,iBAAiB,IAAI,EAAE,KAAK,KAAK,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK;AAAA,EAChG;AAEA,MAAI,cAAc,WAAW,GAAG;AAC9B,UAAM,QAAQ,cAAc,CAAC;AAC7B,WAAO;AAAA,MACL,OAAO,QAAQ,MAAM,KAAM;AAAA,MAC3B,QAAQ,UAAU,MAAM,KAAK;AAAA,MAC7B,aAAa;AAAA,MACb,YAAY;AAAA,MACZ,QAAQ,MAAM;AAAA,MACd,WAAW,eAAe,KAAK;AAAA,IACjC;AAAA,EACF;AAEA,MAAI,cAAc,SAAS,GAAG;AAE5B,QAAI,MAAM;AACV,QAAI,QAAQ;AACZ,UAAM,YAAoC,CAAC;AAC3C,eAAW,KAAK,eAAe;AAC7B,YAAM,IAAK,EAAE,QAAQ,UAAiC;AACtD,aAAO,KAAK,EAAE,SAAS;AACvB,eAAS;AACT,gBAAU,EAAE,KAAK,IAAI,EAAE;AAAA,IACzB;AACA,WAAO;AAAA,MACL,OAAO,UAAU,IAAI,IAAI,QAAQ,MAAM,KAAK;AAAA,MAC5C,QAAQ;AAAA,MACR,aAAa;AAAA,MACb,YAAY;AAAA,MACZ,QAAQ,cAAc,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,KAAK,GAAG;AAAA,MAClD;AAAA,IACF;AAAA,EACF;AAEA,MAAI,CAAC,gBAAiB,QAAO;AAE7B,QAAM,QACJ,OAAO,OAAO;AAAA,IACZ,CAAC,MACC,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK,KAAK,UAAU,EAAE,KAAK,MAAM;AAAA,EACtF,KAAK,OAAO,OAAO,KAAK,CAAC,MAAM,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK,CAAC;AAExF,MAAI,CAAC,MAAO,QAAO;AAEnB,QAAM,iBAAiB,MAAM,QAAQ;AACrC,SAAO;AAAA,IACL,OAAO,QAAQ,MAAM,KAAM;AAAA,IAC3B,QAAQ;AAAA,IACR,aAAa;AAAA,IACb,YAAY,OAAO,mBAAmB,WAAW,iBAAiB;AAAA,IAClE,QAAQ,MAAM;AAAA,IACd,WAAW,eAAe,KAAK;AAAA,EACjC;AACF;AAcO,SAAS,oCACd,MACA,OAA0C,CAAC,GACgB;AAC3D,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;AAC9F,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,QAAM,aAAa,KAAK,wBAAwB;AAEhD,SAAO,KAAK,IAAI,CAAC,QAAQ;AAEvB,UAAM,cAAsD,CAAC;AAC7D,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,IAAI,QAAQ,GAAG,GAAG;AACpD,UACE,EAAE,WAAW,QAAQ,KACrB,CAAC,EAAE,SAAS,KAAK,CAAC,KAClB,OAAO,MAAM,YACb,OAAO,SAAS,CAAC,GACjB;AACA,oBAAY,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAM,GAAG,OAAO,EAAE,CAAC;AAAA,MAC/D;AAAA,IACF;AACA,UAAM,MAAM,YAAY,OAAO,CAAC,MAAM,iBAAiB,IAAI,EAAE,IAAI,CAAC;AAElE,QAAI,IAAI,WAAW,GAAG;AACpB,YAAM,QAAQ,IAAI,CAAC;AACnB,aAAO;AAAA,QACL,OAAO,IAAI;AAAA,QACX,QAAQ;AAAA,UACN,OAAO,QAAQ,MAAM,KAAK;AAAA,UAC1B,QAAQ,UAAU,MAAM,IAAI;AAAA,UAC5B,aAAa;AAAA,UACb,YAAY;AAAA,UACZ,QAAQ,MAAM;AAAA,QAChB;AAAA,MACF;AAAA,IACF;AACA,QAAI,IAAI,SAAS,GAAG;AAClB,YAAM,QAAQ,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,IAAI;AACzD,YAAM,YAAoC,OAAO;AAAA,QAC/C,IAAI,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC;AAAA,MAClC;AACA,aAAO;AAAA,QACL,OAAO,IAAI;AAAA,QACX,QAAQ;AAAA,UACN,OAAO,QAAQ,KAAK;AAAA,UACpB,QAAQ;AAAA,UACR,aAAa;AAAA,UACb,YAAY;AAAA,UACZ,QAAQ,IAAI,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG;AAAA,UACvC;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,gBAAiB,QAAO,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK;AAG9D,UAAM,UAAU,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACxD,QAAI,OAAO,YAAY,YAAY,CAAC,OAAO,SAAS,OAAO,GAAG;AAC5D,aAAO,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK;AAAA,IAC1C;AACA,WAAO;AAAA,MACL,OAAO,IAAI;AAAA,MACX,QAAQ;AAAA,QACN,OAAO,QAAQ,OAAO;AAAA,QACtB,QAAQ;AAAA,QACR,aAAa;AAAA,QACb,YAAY;AAAA,QACZ,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAGO,SAAS,gCACd,MACA,OAA0C,CAAC,GACU;AACrD,QAAM,WAAW,oCAAoC,MAAM,EAAE,GAAG,MAAM,iBAAiB,MAAM,CAAC;AAC9F,QAAM,MAA2D,CAAC;AAClE,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,IAAI,SAAS,CAAC;AACpB,QAAI,EAAE,UAAU,EAAE,OAAO,gBAAgB,iBAAiB;AACxD,UAAI,KAAK,EAAE,KAAK,KAAK,CAAC,GAAI,QAAQ,EAAE,OAAO,CAAC;AAAA,IAC9C;AAAA,EACF;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,eAAe,GAAwC;AAC9D,QAAM,MAA8B,CAAC;AACrC,MAAI,EAAE,aAAa;AACjB,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,EAAE,WAAW,GAAG;AAClD,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,CAAC,IAAI;AAAA,IAC5D;AAAA,EACF;AACA,SAAO;AACT;;;AC3MA,IAAM,gBAAgB,CAAC,MAAgC;AACrD,QAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAC9C,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEO,SAAS,oBAAoB,OAAsD;AACxF,QAAM,UAAU,MAAM,WAAW;AACjC,QAAM,UAAU,MAAM;AACtB,QAAM,MAAM,MAAM,YAAY,WAAW;AACzC,QAAM,MAAM,MAAM,YAAY,UAAU;AAExC,QAAM,OAAO,MAAM,KAAK,OAAO,CAAC,MAAM,QAAQ,CAAC,MAAM,IAAI;AACzD,QAAM,IAAI,KAAK;AACf,MAAI,IAAI,GAAG;AACT,WAAO;AAAA,MACL,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT;AAAA,MACA,WAAW,CAAC,0CAA0C,CAAC,0BAA0B;AAAA,IACnF;AAAA,EACF;AACA,QAAM,aAAa,KAAK,IAAI,GAAG,MAAM,cAAc,KAAK,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,CAAC;AAClF,QAAM,SAAS,KAAK,MAAM,GAAG,IAAI,UAAU;AAC3C,QAAM,QAAQ,KAAK,MAAM,IAAI,UAAU;AAEvC,QAAM,WAAmC,CAAC;AAG1C,MAAI,SAAS;AACX,UAAM,cAAc,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACxF,UAAM,aAAa,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACtF,UAAM,cAAc,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACxF,UAAM,aAAa,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACtF,QACE,YAAY,UAAU,KACtB,WAAW,UAAU,KACrB,YAAY,UAAU,KACtB,WAAW,UAAU,GACrB;AACA,YAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;AACtD,YAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;AAGtD,YAAM,MAAM,KAAK,IAAI,GAAG,aAAa,UAAU;AAC/C,YAAM,WAAWA,SAAQ,MAAM,CAAC;AAChC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,wBAAwB,WAAW,QAAQ,CAAC,CAAC,2BAA2B,WAAW,QAAQ,CAAC,CAAC,+BAC7F,yCAAyC,WAAW,QAAQ,CAAC,CAAC,WAAW,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpG,QAAQ;AAAA,UACN;AAAA,UACA;AAAA,UACA;AAAA,UACA,SAAS,YAAY;AAAA,UACrB,QAAQ,WAAW;AAAA,QACrB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,UAAU,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACpF,UAAM,SAAS,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AAClF,QAAI,QAAQ,UAAU,KAAK,OAAO,UAAU,GAAG;AAC7C,YAAM,KAAK,YAAY,SAAS,MAAM;AAItC,YAAM,WAAWA,SAAQ,KAAK,GAAG;AACjC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,MAAM,GAAG,QAAQ,CAAC,CAAC,oEACnB,MAAM,GAAG,QAAQ,CAAC,CAAC;AAAA,QACzB,QAAQ,EAAE,IAAI,SAAS,QAAQ,QAAQ,QAAQ,OAAO,OAAO;AAAA,MAC/D,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,cAAc,MAAM,qBAAqB,iBAAiB,MAAM,uBAAuB;AAC7F,UAAM,UAAU,KACb,IAAI,CAAC,OAAO,EAAE,GAAG,QAAQ,CAAC,GAAG,GAAG,YAAY,CAAC,EAAE,EAAE,EACjD;AAAA,MACC,CAAC,MAAqC,OAAO,EAAE,MAAM,YAAY,OAAO,EAAE,MAAM;AAAA,IAClF;AACF,QAAI,QAAQ,UAAU,GAAG;AACvB,YAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,CAAC;AACjC,YAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,CAAC;AACjC,YAAM,IAAI,SAAS,IAAI,EAAE;AAGzB,YAAM,WAAWA,SAAQ,MAAM,KAAK,IAAI,GAAG,CAAC,CAAC;AAC7C,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,2DAAsD,EAAE,QAAQ,CAAC,CAAC,mCAClE,+CAA0C,EAAE,QAAQ,CAAC,CAAC;AAAA,QAC5D,QAAQ,EAAE,SAAS,GAAG,GAAG,QAAQ,OAAO;AAAA,MAC1C,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,UAAU,gCAAgC,MAAM,MAAM,2BAA2B,CAAC,CAAC;AACzF,QAAI,QAAQ,UAAU,GAAG;AACvB,YAAM,YAAY,QAAQ,MAAM,GAAG,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;AACjE,YAAM,WAAW,QAAQ,MAAM,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;AAC7D,YAAM,WACJ,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,MAAM,EAAE,OAAO,KAAK,CAAC;AACvF,YAAM,aACJ,KAAK,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,CAAC,IACzE,KAAK,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,CAAC;AAC5E,YAAM,WAAW,KAAK,IAAI,GAAG,aAAa,QAAQ;AAClD,YAAM,WAAWA,SAAQ,WAAW,CAAC;AACrC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,gBAAgB,WAAW,QAAQ,CAAC,CAAC,gCAAgC,SAAS,QAAQ,CAAC,CAAC,yDACxF,uDAAuD,WAAW,QAAQ,CAAC,CAAC,SAAS,SAAS,QAAQ,CAAC,CAAC;AAAA,QAC9G,QAAQ,EAAE,YAAY,UAAU,UAAU,GAAG,QAAQ,OAAO;AAAA,MAC9D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,SAAS,SAAS,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,QAAQ,GAAG,CAAC;AACnE,QAAM,UACJ,UAAU,MAAM,WAAW,UAAU,MAAM,YAAY;AACzD,QAAM,YAAY,SACf,OAAO,CAAC,MAAM,EAAE,YAAY,GAAG,EAC/B,IAAI,CAAC,MAAM,GAAG,EAAE,MAAM,cAAc,EAAE,SAAS,QAAQ,CAAC,CAAC,WAAM,EAAE,OAAO,EAAE;AAC7E,MAAI,UAAU,WAAW,EAAG,WAAU,KAAK,0CAA0C;AAErF,SAAO,EAAE,UAAU,SAAS,WAAW,EAAE;AAC3C;AAIA,SAAS,KAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAASA,SAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,EAAG,QAAO;AAClD,QAAM,KAAK,KAAK,CAAC;AACjB,QAAM,KAAK,KAAK,CAAC;AACjB,MAAI,MAAM,GACR,KAAK,GACL,KAAK;AACP,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,WAAO,KAAK;AACZ,UAAM,KAAK;AACX,UAAM,KAAK;AAAA,EACb;AACA,MAAI,OAAO,KAAK,OAAO,EAAG,QAAO;AACjC,SAAO,MAAM,KAAK,KAAK,KAAK,EAAE;AAChC;AAEA,SAAS,YAAY,GAAa,GAAqB;AAErD,QAAM,UAAU,CAAC,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,UAAU,CAAC,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,MAAM,CAAC,GAAG,oBAAI,IAAI,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvE,MAAI,MAAM;AACV,aAAW,KAAK,KAAK;AACnB,UAAM,KAAK,QAAQ,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,SAAS,QAAQ;AAC1D,UAAM,KAAK,QAAQ,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,SAAS,QAAQ;AAC1D,UAAM,KAAK,IAAI,KAAK,KAAK,IAAI,KAAK,EAAE,CAAC;AAAA,EACvC;AACA,SAAO;AACT;AAEA,SAAS,iBACP,gBACmC;AACnC,SAAO,CAAC,QAAmB;AACzB,UAAM,WAAW,gCAAgC,CAAC,GAAG,GAAG,kBAAkB,CAAC,CAAC;AAC5E,WAAO,SAAS,WAAW,IAAI,SAAS,CAAC,EAAG,OAAO,QAAQ;AAAA,EAC7D;AACF;","names":["clamp01"]}

package/dist/{control-DVrmvM_k.d.ts → control-CmLJk3IG.d.ts}

@@ -1,4 +1,4 @@
-import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-iATEAHmc.js';
+import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-Dvy-bt7x.js';
 import { C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig } from './control-runtime-BZ_lVLYW.js';
 import { T as TraceEmitter } from './emitter-DP_cSSiw.js';
 import { T as TraceStore, F as FailureClass } from './store-Db2Bv8Cf.js';

package/dist/control.d.ts

@@ -1,7 +1,7 @@
-export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-DVrmvM_k.js';
+export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-CmLJk3IG.js';
 export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-BZ_lVLYW.js';
-import './feedback-trajectory-iATEAHmc.js';
-import './dataset-ueRVTUoY.js';
+import './feedback-trajectory-Dvy-bt7x.js';
+import './dataset-BlwAtYYf.js';
 import './errors-mje_cKOs.js';
 import './emitter-DP_cSSiw.js';
 import './store-Db2Bv8Cf.js';

package/dist/control.js

@@ -4,7 +4,7 @@ import {
   runProposeReview,
   runProposeReviewAsControlLoop,
   scoreFromEvals
-} from "./chunk-7PR3WPWE.js";
+} from "./chunk-L7XMNXLO.js";
 import {
   allCriticalPassed,
   objectiveEval,
@@ -13,7 +13,7 @@ import {
   stopOnRepeatedAction,
   subjectiveEval
 } from "./chunk-NCRFYPS3.js";
-import "./chunk-TSPOEDM3.js";
+import "./chunk-BWZEGTES.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";
 import "./chunk-QYJT52YW.js";