@tangle-network/agent-eval 0.25.0 → 0.27.2

package/dist/chunk-R5UQJNKC.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): { 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 = 1000\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(Math.random() * 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"],"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,MACmC;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;AACV,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,KAAK,OAAO,IAAI,CAAC,CAAC;AAAA,IAC7C;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;","names":["i","i"]}

package/dist/{chunk-SESZDQPX.js → chunk-RUI6SIHY.js}

@@ -3,7 +3,7 @@ import {
 } from "./chunk-4S4BM3QQ.js";
 import {
   researchReport
-} from "./chunk-2A5XJB43.js";
+} from "./chunk-5AKPEK5L.js";
 import {
   RunIntegrityError,
   assertRunCaptured
@@ -17,7 +17,7 @@ import {
 import {
   canonicalize,
   hashJson
-} from "./chunk-4F5DQN55.js";
+} from "./chunk-VSMTAMNK.js";
 
 // src/eval-campaign.ts
 var DEFAULT_INTEGRITY = {
@@ -284,4 +284,4 @@ function defaultRunId(params) {
 export {
   runEvalCampaign
 };
-//# sourceMappingURL=chunk-SESZDQPX.js.map
+//# sourceMappingURL=chunk-RUI6SIHY.js.map

package/dist/chunk-RUI6SIHY.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/eval-campaign.ts"],"sourcesContent":["/**\n * EvalCampaign — opinionated matrix runner that wires the four\n * capture-integrity directives by construction.\n *\n * The canonical benchmark shape — matrix runner → for each\n * (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the\n * run → analyze — has a bug class at the integration boundary: raw\n * events not captured, route silently wrong, integrity not asserted,\n * analyst never run. The directives in `SKILL.md § Capture integrity`\n * are the mitigations.\n *\n * `EvalCampaign` is the structural fix — consumers don't wire the\n * integrity surface themselves; the campaign owns it. Specifically:\n *\n *   - calls `assertLlmRoute` once at preflight before any work runs\n *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories\n *   - constructs the `TraceEmitter` with `onRunComplete: [analyst hook]`\n *   - hands the runner an `LlmClientOptions` pre-wired with the sink and\n *     trace context — the runner can't accidentally call an LLM without\n *     capturing the raw HTTP envelope\n *   - calls `assertRunCaptured` after every `endRun` and routes failures\n *     through a configurable policy (`throw` / `mark_failed` / `log`)\n *   - assembles per-run `RunRecord`s and runs `researchReport` at the end\n *     so the campaign artifact is launch-decision-grade by default\n *   - embeds the campaign fingerprint (a SHA-256 over the canonicalised\n *     run set) and optional `preregistrationHash` in the report\n *\n * The runner contract is intentionally narrow: produce a `CampaignRunOutcome`\n * given a fully-wired `CampaignRunContext`. Everything orchestration-shaped\n * lives in the campaign. This is the inversion-of-control point — consumers\n * stop writing matrix runners and start writing scenario-runners.\n *\n * Out of scope for v1 (tracked in `docs/research-report-methodology.md`):\n *\n *   - Distributed/cluster execution (concurrency is local async)\n *   - Adaptive sampling / sequential interim looks\n *   - Resume from partial state across crashes\n *   - LLM-call retry beyond what `LlmClient` already does\n */\n\nimport { assertLlmRoute, type LlmClientOptions, type LlmRouteRequirements } from './llm-client'\nimport { canonicalize, hashJson } from './pre-registration'\nimport type {\n  RunJudgeMetadata,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTokenUsage,\n} from './run-record'\nimport { type ResearchReport, type ResearchReportOptions, researchReport } from './summary-report'\nimport type { RunCompleteHook } from './trace/emitter'\nimport { TraceEmitter } from './trace/emitter'\nimport {\n  assertRunCaptured,\n  RunIntegrityError,\n  type RunIntegrityExpectations,\n  type RunIntegrityReport,\n} from './trace/integrity'\nimport { FileSystemRawProviderSink, type RawProviderSink } from './trace/raw-provider-sink'\nimport type { TraceStore } from './trace/store'\n\n// ── Public types ─────────────────────────────────────────────────────────\n\nexport interface CampaignVariant<V> {\n  id: string\n  payload: V\n}\n\nexport interface CampaignScenario {\n  scenarioId: string\n  /** Free-form metadata propagated to runs and reports. */\n  tags?: Record<string, string>\n}\n\nexport interface CampaignRunContext<V> {\n  /** Stable run id. The campaign generates this; the runner does not. */\n  runId: string\n  /** Logical experiment id (campaignId by default; overridable per-run via opts). */\n  experimentId: string\n  variant: V\n  variantId: string\n  scenarioId: string\n  scenarioTags: Record<string, string>\n  seed: number\n  splitTag: RunSplitTag\n  /**\n   * The TraceEmitter for this run, with `onRunComplete` hooks pre-wired\n   * (analyst auto-execution if configured, plus integrity check). The\n   * runner MUST call `emitter.startRun` before doing any work and either\n   * `emitter.endRun` or `emitter.abortRun` before returning.\n   */\n  emitter: TraceEmitter\n  store: TraceStore\n  rawSink: RawProviderSink\n  /**\n   * Pre-wired LLM client options — `rawSink` and `traceContext` are populated\n   * so any `callLlm(req, ctx.llmOpts)` automatically captures raw HTTP. The\n   * runner can spread additional fields if needed.\n   */\n  llmOpts: LlmClientOptions\n}\n\nexport interface CampaignRunOutcome {\n  /** Did the run pass? Mirrors `RunOutcome.pass` semantics. */\n  pass: boolean\n  /** Score for the run on its split. Maps to `searchScore` or `holdoutScore`. */\n  score: number\n  /** Mandatory cost in USD. Use 0 + raw.cost_unknown=1 only if truly unknown. */\n  costUsd: number\n  tokenUsage: RunTokenUsage\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** sha256 of the effective prompt sent to the model. */\n  promptHash: string\n  /** sha256 of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Optional extra numeric metrics to land in `outcome.raw`. */\n  raw?: Record<string, number>\n  /** Optional failure-taxonomy tag if the run failed. */\n  failureMode?: string\n  /** Optional judge metadata when a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n}\n\nexport type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>\n\nexport type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log'\n\nexport interface EvalCampaignOptions<V> {\n  /**\n   * Stable id for the campaign. Used as the default `experimentId` on\n   * every run, and folded into the campaign fingerprint.\n   */\n  campaignId: string\n  variants: CampaignVariant<V>[]\n  scenarios: CampaignScenario[]\n  /** Default `[0, 1, 2]`. */\n  seeds?: number[]\n  /** Default `'holdout'` — the split that anchors a launch decision. */\n  splitTag?: RunSplitTag\n  /** Git SHA the campaign is run against. Mandatory; `RunRecord` rejects unset. */\n  commitSha: string\n  /**\n   * LLM client config. Augmented per-run with `rawSink` and `traceContext`\n   * before being passed to the runner. The campaign asserts this config\n   * matches `routeRequirements` once at preflight.\n   */\n  llmOpts: LlmClientOptions\n  /**\n   * Default `{ requireExplicitBaseUrl: true, requireAuth: true }` — fail\n   * loud if the campaign would silently fall back to the public router or\n   * run unauthenticated. Override with an empty object to disable.\n   */\n  routeRequirements?: LlmRouteRequirements\n  /**\n   * Per-run TraceStore factory. Common shape: a fresh store per run keyed\n   * on `runId`. Implementations that share a store across the campaign\n   * are valid — the campaign only writes through `emitter`.\n   */\n  storeFactory: (params: CampaignFactoryParams) => TraceStore\n  /**\n   * Per-run RawProviderSink factory. Defaults to `FileSystemRawProviderSink`\n   * rooted at `${workDir}/raw-events/${runId}` if `workDir` is supplied;\n   * otherwise required. Forensic capture is non-negotiable in a campaign\n   * run — pass `NoopRawProviderSink` explicitly if you want to opt out.\n   */\n  rawSinkFactory?: (params: CampaignFactoryParams) => RawProviderSink\n  /**\n   * Filesystem root for default `rawSinkFactory`. Ignored if\n   * `rawSinkFactory` is supplied.\n   */\n  workDir?: string\n  /**\n   * Extra `onRunComplete` hooks the campaign appends (after its own\n   * integrity-check hook). Pass `traceAnalystOnRunComplete(...)` here.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /**\n   * Per-run integrity expectations. Defaults to:\n   *   `{ llmSpansMin: 1, requireRawCoverageOfLlmSpans: true, requireOutcome: true }`.\n   * Override (e.g. `{ llmSpansMin: 0 }`) for runs that don't call LLMs.\n   */\n  integrity?: RunIntegrityExpectations\n  /** Behaviour when integrity fails. Default `'mark_failed'`. */\n  onIntegrityFailure?: CampaignIntegrityPolicy\n  /**\n   * Per-run runner. Receives a fully-wired context; produces an outcome\n   * the campaign converts into a `RunRecord`.\n   */\n  runner: CampaignRunner<V>\n  /**\n   * If set, the campaign computes `researchReport` at the end. `comparator`\n   * is a `variantId`. Other fields are forwarded verbatim.\n   */\n  report?: { comparator?: string } & Omit<\n    ResearchReportOptions,\n    'comparator' | 'preregistrationHash' | 'generatedAt'\n  >\n  /**\n   * Hash of a signed `HypothesisManifest` (see `pre-registration.ts`).\n   * Embedded in the campaign fingerprint and the research report.\n   */\n  preregistrationHash?: string\n  /** Local concurrency. Default `1` (sequential). */\n  concurrency?: number\n  /**\n   * Override the time source. Tests pass a mock to make wallMs deterministic.\n   */\n  now?: () => number\n  /** Override the runId generator. Tests pin this. */\n  runId?: (params: CampaignFactoryParams) => string\n}\n\nexport interface CampaignFactoryParams {\n  campaignId: string\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n}\n\nexport interface FailedRun {\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n  reason: string\n  error?: string\n}\n\nexport interface EvalCampaignResult {\n  campaignId: string\n  /** SHA-256 over canonicalised `(variantIds, scenarioIds, seeds, comparator, splitTag, baseUrl, provider, preregistrationHash)`. */\n  campaignFingerprint: string\n  preregistrationHash: string | null\n  /** Successful runs only. Failed runs land in `failedRuns`. */\n  runs: RunRecord[]\n  /** Integrity reports for every successful run. */\n  integrityReports: RunIntegrityReport[]\n  failedRuns: FailedRun[]\n  /** Computed when `report` is set on options. */\n  report?: ResearchReport\n  startedAt: string\n  endedAt: string\n}\n\n// ── Implementation ───────────────────────────────────────────────────────\n\nconst DEFAULT_INTEGRITY: RunIntegrityExpectations = {\n  llmSpansMin: 1,\n  requireRawCoverageOfLlmSpans: true,\n  requireOutcome: true,\n}\n\nconst DEFAULT_ROUTE: LlmRouteRequirements = {\n  requireExplicitBaseUrl: true,\n  requireAuth: true,\n}\n\nexport async function runEvalCampaign<V>(\n  opts: EvalCampaignOptions<V>,\n): Promise<EvalCampaignResult> {\n  // ── Preflight ──────────────────────────────────────────────────────\n  assertLlmRoute(opts.llmOpts, opts.routeRequirements ?? DEFAULT_ROUTE)\n\n  if (opts.variants.length === 0) {\n    throw new Error('runEvalCampaign: variants must be non-empty.')\n  }\n  if (opts.scenarios.length === 0) {\n    throw new Error('runEvalCampaign: scenarios must be non-empty.')\n  }\n  const variantIds = new Set<string>()\n  for (const v of opts.variants) {\n    if (variantIds.has(v.id)) {\n      throw new Error(`runEvalCampaign: duplicate variant id \"${v.id}\".`)\n    }\n    variantIds.add(v.id)\n  }\n  const scenarioIds = new Set<string>()\n  for (const s of opts.scenarios) {\n    if (scenarioIds.has(s.scenarioId)) {\n      throw new Error(`runEvalCampaign: duplicate scenarioId \"${s.scenarioId}\".`)\n    }\n    scenarioIds.add(s.scenarioId)\n  }\n  if (opts.report?.comparator && !variantIds.has(opts.report.comparator)) {\n    throw new Error(\n      `runEvalCampaign: report.comparator \"${opts.report.comparator}\" is not a configured variantId.`,\n    )\n  }\n  if (!opts.commitSha) {\n    throw new Error('runEvalCampaign: commitSha is required (every RunRecord needs it).')\n  }\n\n  const seeds = opts.seeds ?? [0, 1, 2]\n  const splitTag: RunSplitTag = opts.splitTag ?? 'holdout'\n  const concurrency = Math.max(1, opts.concurrency ?? 1)\n  const integrity = { ...DEFAULT_INTEGRITY, ...(opts.integrity ?? {}) }\n  const onIntegrityFailure: CampaignIntegrityPolicy = opts.onIntegrityFailure ?? 'mark_failed'\n  const now = opts.now ?? (() => Date.now())\n  const baseUrl = (opts.llmOpts.baseUrl ?? '').replace(/\\/+$/, '')\n  const provider = opts.llmOpts.provider ?? null\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir)\n\n  // ── Fingerprint ────────────────────────────────────────────────────\n  const campaignFingerprint = await hashJson(\n    canonicalize({\n      campaignId: opts.campaignId,\n      variants: opts.variants.map((v) => v.id).sort(),\n      scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),\n      seeds: [...seeds].sort((a, b) => a - b),\n      splitTag,\n      comparator: opts.report?.comparator ?? null,\n      baseUrl,\n      provider,\n      preregistrationHash,\n    }),\n  )\n\n  // ── Plan the matrix ────────────────────────────────────────────────\n  type Cell = { variant: CampaignVariant<V>; scenario: CampaignScenario; seed: number }\n  const cells: Cell[] = []\n  for (const variant of opts.variants) {\n    for (const scenario of opts.scenarios) {\n      for (const seed of seeds) {\n        cells.push({ variant, scenario, seed })\n      }\n    }\n  }\n\n  const startedAt = new Date(now()).toISOString()\n  const runs: RunRecord[] = []\n  const integrityReports: RunIntegrityReport[] = []\n  const failedRuns: FailedRun[] = []\n\n  // ── Execute (bounded-concurrency worker pool) ──────────────────────\n  let cursor = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = cursor++\n      if (i >= cells.length) return\n      const cell = cells[i]!\n      try {\n        const result = await runOneCell(cell)\n        runs.push(result.record)\n        integrityReports.push(result.integrity)\n      } catch (err) {\n        if (err instanceof CellExecutionError) {\n          failedRuns.push(err.failed)\n          if (err.integrity) integrityReports.push(err.integrity)\n        } else {\n          // Genuine bug — not a runner failure, not an integrity failure.\n          // Surface it; don't silently mask.\n          throw err\n        }\n      }\n    }\n  }\n\n  async function runOneCell(\n    cell: Cell,\n  ): Promise<{ record: RunRecord; integrity: RunIntegrityReport }> {\n    const runId = (opts.runId ?? defaultRunId)({\n      campaignId: opts.campaignId,\n      runId: '', // unused by default generator\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    })\n    const factoryParams: CampaignFactoryParams = {\n      campaignId: opts.campaignId,\n      runId,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    }\n    const store = opts.storeFactory(factoryParams)\n    const rawSink = rawSinkFactory(factoryParams)\n\n    const emitter = new TraceEmitter(store, {\n      runId,\n      now: opts.now,\n      onRunComplete: opts.onRunComplete,\n    })\n\n    const llmOpts: LlmClientOptions = {\n      ...opts.llmOpts,\n      rawSink,\n      traceContext: { runId },\n    }\n\n    const ctx: CampaignRunContext<V> = {\n      runId,\n      experimentId: opts.campaignId,\n      variant: cell.variant.payload,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      scenarioTags: cell.scenario.tags ?? {},\n      seed: cell.seed,\n      splitTag,\n      emitter,\n      store,\n      rawSink,\n      llmOpts,\n    }\n\n    const wallStart = now()\n    let outcome: CampaignRunOutcome\n    try {\n      outcome = await opts.runner(ctx)\n    } catch (err) {\n      const message = err instanceof Error ? err.message : String(err)\n      // The runner threw mid-execution; give it a chance to have aborted.\n      try {\n        await emitter.abortRun(message)\n      } catch {\n        // Already aborted/ended; ignore.\n      }\n      throw new CellExecutionError({\n        runId,\n        variantId: cell.variant.id,\n        scenarioId: cell.scenario.scenarioId,\n        seed: cell.seed,\n        reason: 'runner_threw',\n        error: message,\n      })\n    }\n    const wallMs = now() - wallStart\n\n    const integrityReport = await assertRunCaptured(store, runId, { ...integrity, rawSink })\n    if (!integrityReport.ok) {\n      switch (onIntegrityFailure) {\n        case 'throw':\n          throw new RunIntegrityError(integrityReport)\n        case 'mark_failed':\n          throw new CellExecutionError(\n            {\n              runId,\n              variantId: cell.variant.id,\n              scenarioId: cell.scenario.scenarioId,\n              seed: cell.seed,\n              reason: 'integrity_failed',\n              error: integrityReport.issues.map((i) => i.code).join(', '),\n            },\n            integrityReport,\n          )\n        case 'log':\n          // Caller wants the run admitted with a flagged report; fall through.\n          break\n      }\n    }\n\n    const recordOutcome: RunOutcome = {\n      raw: outcome.raw ?? {},\n    }\n    if (splitTag === 'holdout') recordOutcome.holdoutScore = outcome.score\n    else recordOutcome.searchScore = outcome.score\n\n    const record: RunRecord = {\n      runId,\n      experimentId: opts.campaignId,\n      candidateId: cell.variant.id,\n      seed: cell.seed,\n      model: outcome.model,\n      promptHash: outcome.promptHash,\n      configHash: outcome.configHash,\n      commitSha: opts.commitSha,\n      wallMs,\n      costUsd: outcome.costUsd,\n      tokenUsage: outcome.tokenUsage,\n      judgeMetadata: outcome.judgeMetadata,\n      outcome: recordOutcome,\n      failureMode: outcome.failureMode,\n      splitTag,\n      scenarioId: cell.scenario.scenarioId,\n    }\n    return { record, integrity: integrityReport }\n  }\n\n  const workers = Array.from({ length: Math.min(concurrency, cells.length) }, () => worker())\n  await Promise.all(workers)\n\n  // ── Optional research report ───────────────────────────────────────\n  let report: ResearchReport | undefined\n  if (opts.report) {\n    const reportOpts: ResearchReportOptions = {\n      ...opts.report,\n      comparator: opts.report.comparator,\n      split: splitTag === 'dev' ? 'search' : splitTag,\n      generatedAt: new Date(now()).toISOString(),\n      preregistrationHash: preregistrationHash ?? undefined,\n    }\n    report = await researchReport(runs, reportOpts)\n  }\n\n  const endedAt = new Date(now()).toISOString()\n\n  return {\n    campaignId: opts.campaignId,\n    campaignFingerprint,\n    preregistrationHash,\n    runs,\n    integrityReports,\n    failedRuns,\n    report,\n    startedAt,\n    endedAt,\n  }\n}\n\n// ── Internal ─────────────────────────────────────────────────────────────\n\nclass CellExecutionError extends Error {\n  readonly failed: FailedRun\n  readonly integrity?: RunIntegrityReport\n  constructor(failed: FailedRun, integrity?: RunIntegrityReport) {\n    super(`cell ${failed.variantId}/${failed.scenarioId}@${failed.seed} failed: ${failed.reason}`)\n    this.failed = failed\n    this.integrity = integrity\n  }\n}\n\nfunction defaultRawSinkFactory(workDir: string | undefined) {\n  return (params: CampaignFactoryParams): RawProviderSink => {\n    if (!workDir) {\n      throw new Error(\n        'runEvalCampaign: rawSinkFactory not supplied and workDir not set. Pass either to enable raw provider capture, or pass `new NoopRawProviderSink()` via rawSinkFactory to opt out explicitly.',\n      )\n    }\n    return new FileSystemRawProviderSink({\n      dir: `${workDir}/raw-events/${params.runId}`,\n    })\n  }\n}\n\nfunction defaultRunId(params: CampaignFactoryParams): string {\n  // Stable across re-runs: fingerprint of (campaignId, variantId, scenarioId, seed).\n  // Caller can override via opts.runId for non-deterministic IDs.\n  const base = `${params.campaignId}::${params.variantId}::${params.scenarioId}::${params.seed}`\n  // Lightweight hex: we don't need crypto-grade here, just stability + uniqueness.\n  let h1 = 0x811c9dc5\n  let h2 = 0x12345678\n  for (let i = 0; i < base.length; i++) {\n    const c = base.charCodeAt(i)\n    h1 = Math.imul(h1 ^ c, 0x01000193) >>> 0\n    h2 = Math.imul(h2 ^ c, 0x9e3779b1) >>> 0\n  }\n  return `run-${h1.toString(16).padStart(8, '0')}${h2.toString(16).padStart(8, '0')}`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;AAwPA,IAAM,oBAA8C;AAAA,EAClD,aAAa;AAAA,EACb,8BAA8B;AAAA,EAC9B,gBAAgB;AAClB;AAEA,IAAM,gBAAsC;AAAA,EAC1C,wBAAwB;AAAA,EACxB,aAAa;AACf;AAEA,eAAsB,gBACpB,MAC6B;AAE7B,iBAAe,KAAK,SAAS,KAAK,qBAAqB,aAAa;AAEpE,MAAI,KAAK,SAAS,WAAW,GAAG;AAC9B,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AACA,MAAI,KAAK,UAAU,WAAW,GAAG;AAC/B,UAAM,IAAI,MAAM,+CAA+C;AAAA,EACjE;AACA,QAAM,aAAa,oBAAI,IAAY;AACnC,aAAW,KAAK,KAAK,UAAU;AAC7B,QAAI,WAAW,IAAI,EAAE,EAAE,GAAG;AACxB,YAAM,IAAI,MAAM,0CAA0C,EAAE,EAAE,IAAI;AAAA,IACpE;AACA,eAAW,IAAI,EAAE,EAAE;AAAA,EACrB;AACA,QAAM,cAAc,oBAAI,IAAY;AACpC,aAAW,KAAK,KAAK,WAAW;AAC9B,QAAI,YAAY,IAAI,EAAE,UAAU,GAAG;AACjC,YAAM,IAAI,MAAM,0CAA0C,EAAE,UAAU,IAAI;AAAA,IAC5E;AACA,gBAAY,IAAI,EAAE,UAAU;AAAA,EAC9B;AACA,MAAI,KAAK,QAAQ,cAAc,CAAC,WAAW,IAAI,KAAK,OAAO,UAAU,GAAG;AACtE,UAAM,IAAI;AAAA,MACR,uCAAuC,KAAK,OAAO,UAAU;AAAA,IAC/D;AAAA,EACF;AACA,MAAI,CAAC,KAAK,WAAW;AACnB,UAAM,IAAI,MAAM,oEAAoE;AAAA,EACtF;AAEA,QAAM,QAAQ,KAAK,SAAS,CAAC,GAAG,GAAG,CAAC;AACpC,QAAM,WAAwB,KAAK,YAAY;AAC/C,QAAM,cAAc,KAAK,IAAI,GAAG,KAAK,eAAe,CAAC;AACrD,QAAM,YAAY,EAAE,GAAG,mBAAmB,GAAI,KAAK,aAAa,CAAC,EAAG;AACpE,QAAM,qBAA8C,KAAK,sBAAsB;AAC/E,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,WAAW,KAAK,QAAQ,WAAW,IAAI,QAAQ,QAAQ,EAAE;AAC/D,QAAM,WAAW,KAAK,QAAQ,YAAY;AAC1C,QAAM,sBAAsB,KAAK,uBAAuB;AAExD,QAAM,iBAAiB,KAAK,kBAAkB,sBAAsB,KAAK,OAAO;AAGhF,QAAM,sBAAsB,MAAM;AAAA,IAChC,aAAa;AAAA,MACX,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK;AAAA,MAC9C,WAAW,KAAK,UAAU,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,KAAK;AAAA,MACxD,OAAO,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAAA,MACtC;AAAA,MACA,YAAY,KAAK,QAAQ,cAAc;AAAA,MACvC;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAIA,QAAM,QAAgB,CAAC;AACvB,aAAW,WAAW,KAAK,UAAU;AACnC,eAAW,YAAY,KAAK,WAAW;AACrC,iBAAW,QAAQ,OAAO;AACxB,cAAM,KAAK,EAAE,SAAS,UAAU,KAAK,CAAC;AAAA,MACxC;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAC9C,QAAM,OAAoB,CAAC;AAC3B,QAAM,mBAAyC,CAAC;AAChD,QAAM,aAA0B,CAAC;AAGjC,MAAI,SAAS;AACb,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,MAAM,OAAQ;AACvB,YAAM,OAAO,MAAM,CAAC;AACpB,UAAI;AACF,cAAM,SAAS,MAAM,WAAW,IAAI;AACpC,aAAK,KAAK,OAAO,MAAM;AACvB,yBAAiB,KAAK,OAAO,SAAS;AAAA,MACxC,SAAS,KAAK;AACZ,YAAI,eAAe,oBAAoB;AACrC,qBAAW,KAAK,IAAI,MAAM;AAC1B,cAAI,IAAI,UAAW,kBAAiB,KAAK,IAAI,SAAS;AAAA,QACxD,OAAO;AAGL,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,iBAAe,WACb,MAC+D;AAC/D,UAAM,SAAS,KAAK,SAAS,cAAc;AAAA,MACzC,YAAY,KAAK;AAAA,MACjB,OAAO;AAAA;AAAA,MACP,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,MAAM,KAAK;AAAA,IACb,CAAC;AACD,UAAM,gBAAuC;AAAA,MAC3C,YAAY,KAAK;AAAA,MACjB;AAAA,MACA,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,MAAM,KAAK;AAAA,IACb;AACA,UAAM,QAAQ,KAAK,aAAa,aAAa;AAC7C,UAAM,UAAU,eAAe,aAAa;AAE5C,UAAM,UAAU,IAAI,aAAa,OAAO;AAAA,MACtC;AAAA,MACA,KAAK,KAAK;AAAA,MACV,eAAe,KAAK;AAAA,IACtB,CAAC;AAED,UAAM,UAA4B;AAAA,MAChC,GAAG,KAAK;AAAA,MACR;AAAA,MACA,cAAc,EAAE,MAAM;AAAA,IACxB;AAEA,UAAM,MAA6B;AAAA,MACjC;AAAA,MACA,cAAc,KAAK;AAAA,MACnB,SAAS,KAAK,QAAQ;AAAA,MACtB,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,cAAc,KAAK,SAAS,QAAQ,CAAC;AAAA,MACrC,MAAM,KAAK;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAEA,UAAM,YAAY,IAAI;AACtB,QAAI;AACJ,QAAI;AACF,gBAAU,MAAM,KAAK,OAAO,GAAG;AAAA,IACjC,SAAS,KAAK;AACZ,YAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAE/D,UAAI;AACF,cAAM,QAAQ,SAAS,OAAO;AAAA,MAChC,QAAQ;AAAA,MAER;AACA,YAAM,IAAI,mBAAmB;AAAA,QAC3B;AAAA,QACA,WAAW,KAAK,QAAQ;AAAA,QACxB,YAAY,KAAK,SAAS;AAAA,QAC1B,MAAM,KAAK;AAAA,QACX,QAAQ;AAAA,QACR,OAAO;AAAA,MACT,CAAC;AAAA,IACH;AACA,UAAM,SAAS,IAAI,IAAI;AAEvB,UAAM,kBAAkB,MAAM,kBAAkB,OAAO,OAAO,EAAE,GAAG,WAAW,QAAQ,CAAC;AACvF,QAAI,CAAC,gBAAgB,IAAI;AACvB,cAAQ,oBAAoB;AAAA,QAC1B,KAAK;AACH,gBAAM,IAAI,kBAAkB,eAAe;AAAA,QAC7C,KAAK;AACH,gBAAM,IAAI;AAAA,YACR;AAAA,cACE;AAAA,cACA,WAAW,KAAK,QAAQ;AAAA,cACxB,YAAY,KAAK,SAAS;AAAA,cAC1B,MAAM,KAAK;AAAA,cACX,QAAQ;AAAA,cACR,OAAO,gBAAgB,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,IAAI;AAAA,YAC5D;AAAA,YACA;AAAA,UACF;AAAA,QACF,KAAK;AAEH;AAAA,MACJ;AAAA,IACF;AAEA,UAAM,gBAA4B;AAAA,MAChC,KAAK,QAAQ,OAAO,CAAC;AAAA,IACvB;AACA,QAAI,aAAa,UAAW,eAAc,eAAe,QAAQ;AAAA,QAC5D,eAAc,cAAc,QAAQ;AAEzC,UAAM,SAAoB;AAAA,MACxB;AAAA,MACA,cAAc,KAAK;AAAA,MACnB,aAAa,KAAK,QAAQ;AAAA,MAC1B,MAAM,KAAK;AAAA,MACX,OAAO,QAAQ;AAAA,MACf,YAAY,QAAQ;AAAA,MACpB,YAAY,QAAQ;AAAA,MACpB,WAAW,KAAK;AAAA,MAChB;AAAA,MACA,SAAS,QAAQ;AAAA,MACjB,YAAY,QAAQ;AAAA,MACpB,eAAe,QAAQ;AAAA,MACvB,SAAS;AAAA,MACT,aAAa,QAAQ;AAAA,MACrB;AAAA,MACA,YAAY,KAAK,SAAS;AAAA,IAC5B;AACA,WAAO,EAAE,QAAQ,WAAW,gBAAgB;AAAA,EAC9C;AAEA,QAAM,UAAU,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,aAAa,MAAM,MAAM,EAAE,GAAG,MAAM,OAAO,CAAC;AAC1F,QAAM,QAAQ,IAAI,OAAO;AAGzB,MAAI;AACJ,MAAI,KAAK,QAAQ;AACf,UAAM,aAAoC;AAAA,MACxC,GAAG,KAAK;AAAA,MACR,YAAY,KAAK,OAAO;AAAA,MACxB,OAAO,aAAa,QAAQ,WAAW;AAAA,MACvC,aAAa,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAAA,MACzC,qBAAqB,uBAAuB;AAAA,IAC9C;AACA,aAAS,MAAM,eAAe,MAAM,UAAU;AAAA,EAChD;AAEA,QAAM,UAAU,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAE5C,SAAO;AAAA,IACL,YAAY,KAAK;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAIA,IAAM,qBAAN,cAAiC,MAAM;AAAA,EAC5B;AAAA,EACA;AAAA,EACT,YAAY,QAAmB,WAAgC;AAC7D,UAAM,QAAQ,OAAO,SAAS,IAAI,OAAO,UAAU,IAAI,OAAO,IAAI,YAAY,OAAO,MAAM,EAAE;AAC7F,SAAK,SAAS;AACd,SAAK,YAAY;AAAA,EACnB;AACF;AAEA,SAAS,sBAAsB,SAA6B;AAC1D,SAAO,CAAC,WAAmD;AACzD,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AACA,WAAO,IAAI,0BAA0B;AAAA,MACnC,KAAK,GAAG,OAAO,eAAe,OAAO,KAAK;AAAA,IAC5C,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aAAa,QAAuC;AAG3D,QAAM,OAAO,GAAG,OAAO,UAAU,KAAK,OAAO,SAAS,KAAK,OAAO,UAAU,KAAK,OAAO,IAAI;AAE5F,MAAI,KAAK;AACT,MAAI,KAAK;AACT,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,IAAI,KAAK,WAAW,CAAC;AAC3B,SAAK,KAAK,KAAK,KAAK,GAAG,QAAU,MAAM;AACvC,SAAK,KAAK,KAAK,KAAK,GAAG,UAAU,MAAM;AAAA,EACzC;AACA,SAAO,OAAO,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC,GAAG,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC;AACnF;","names":[]}

package/dist/{chunk-EDUKQ5AM.js → chunk-SZSBQUIJ.js}

@@ -4,7 +4,7 @@ import {
 import {
   pairedBootstrap,
   pairedWilcoxon
-} from "./chunk-2A5XJB43.js";
+} from "./chunk-5AKPEK5L.js";
 
 // src/feedback-trajectory.ts
 var DEFAULT_SPLIT_POLICY = {
@@ -1409,4 +1409,4 @@ export {
   CallbackResearcher,
   NoopResearcher
 };
-//# sourceMappingURL=chunk-EDUKQ5AM.js.map
+//# sourceMappingURL=chunk-SZSBQUIJ.js.map

package/dist/chunk-SZSBQUIJ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/feedback-trajectory.ts","../src/held-out-gate.ts","../src/pareto.ts","../src/prompt-evolution.ts","../src/multi-shot-optimization.ts","../src/reflective-mutation.ts","../src/researcher.ts"],"sourcesContent":["import type { ControlEvalResult, ControlRunResult, ControlStep } from './control-runtime'\nimport type { DatasetScenario, DatasetSplit } from './dataset'\n\nexport type FeedbackArtifactType =\n  | 'text'\n  | 'code'\n  | 'plan'\n  | 'research'\n  | 'action'\n  | 'ui'\n  | 'decision'\n  | 'data'\n  | 'other'\n\nexport type FeedbackLabelSource = 'user' | 'judge' | 'environment' | 'metric' | 'policy' | 'system'\n\nexport type FeedbackLabelKind =\n  | 'approve'\n  | 'reject'\n  | 'select'\n  | 'edit'\n  | 'rank'\n  | 'rate'\n  | 'comment'\n  | 'metric_outcome'\n  | 'policy_block'\n  | 'revision_request'\n\nexport type FeedbackSeverity = 'info' | 'warning' | 'error' | 'critical'\n\nexport interface FeedbackTask {\n  intent: string\n  context?: unknown\n}\n\nexport interface ProposedSideEffect {\n  type: string\n  risk?: 'low' | 'medium' | 'high'\n  costUsd?: number\n  externalSideEffect?: boolean\n  requiresApproval?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackLabel {\n  id?: string\n  source: FeedbackLabelSource\n  kind: FeedbackLabelKind\n  value: unknown\n  reason?: string\n  severity?: FeedbackSeverity\n  createdAt: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackAttempt {\n  id: string\n  stepIndex: number\n  artifactType: FeedbackArtifactType\n  artifact: unknown\n  options?: unknown[]\n  proposedAction?: ProposedSideEffect\n  evals?: ControlEvalResult[]\n  feedback?: FeedbackLabel[]\n  createdAt: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackOutcome {\n  success?: boolean\n  score?: number\n  metrics?: Record<string, number>\n  costUsd?: number\n  detail?: string\n  observedAt?: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackTrajectory {\n  id: string\n  projectId?: string\n  scenarioId?: string\n  task: FeedbackTask\n  attempts: FeedbackAttempt[]\n  labels: FeedbackLabel[]\n  outcome?: FeedbackOutcome\n  split?: DatasetSplit\n  tags?: Record<string, string>\n  createdAt: string\n  updatedAt?: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackTrajectoryStore {\n  save(trajectory: FeedbackTrajectory): Promise<void>\n  get(id: string): Promise<FeedbackTrajectory | null>\n  list(filter?: FeedbackTrajectoryFilter): Promise<FeedbackTrajectory[]>\n  appendAttempt(id: string, attempt: FeedbackAttempt): Promise<FeedbackTrajectory>\n  appendLabel(id: string, label: FeedbackLabel, attemptId?: string): Promise<FeedbackTrajectory>\n}\n\nexport interface FeedbackTrajectoryFilter {\n  projectId?: string\n  scenarioId?: string\n  split?: DatasetSplit\n  tag?: [string, string]\n}\n\nexport interface FeedbackSplitPolicy {\n  trainPct?: number\n  devPct?: number\n  testPct?: number\n  holdoutPct?: number\n}\n\nexport interface PreferenceMemoryEntry {\n  instruction: string\n  rationale: string\n  weight: number\n  sourceTrajectoryId: string\n  sourceLabelId?: string\n  category?: string\n}\n\nexport interface FeedbackOptimizerRow {\n  scenarioId: string\n  trajectoryId: string\n  labelKinds: FeedbackLabelKind[]\n  score?: number\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackReplayResult {\n  trajectoryId: string\n  pass: boolean\n  score?: number\n  labels: FeedbackLabel[]\n  outcome?: FeedbackOutcome\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackReplayAdapter {\n  replay(\n    trajectory: FeedbackTrajectory,\n  ):\n    | Promise<Omit<FeedbackReplayResult, 'trajectoryId'>>\n    | Omit<FeedbackReplayResult, 'trajectoryId'>\n}\n\nconst DEFAULT_SPLIT_POLICY: Required<FeedbackSplitPolicy> = {\n  trainPct: 70,\n  devPct: 15,\n  testPct: 10,\n  holdoutPct: 5,\n}\n\nexport class InMemoryFeedbackTrajectoryStore implements FeedbackTrajectoryStore {\n  private readonly trajectories = new Map<string, FeedbackTrajectory>()\n\n  async save(trajectory: FeedbackTrajectory): Promise<void> {\n    this.trajectories.set(trajectory.id, cloneTrajectory(trajectory))\n  }\n\n  async get(id: string): Promise<FeedbackTrajectory | null> {\n    const trajectory = this.trajectories.get(id)\n    return trajectory ? cloneTrajectory(trajectory) : null\n  }\n\n  async list(filter: FeedbackTrajectoryFilter = {}): Promise<FeedbackTrajectory[]> {\n    return [...this.trajectories.values()]\n      .filter((trajectory) => matchesFilter(trajectory, filter))\n      .map(cloneTrajectory)\n  }\n\n  async appendAttempt(id: string, attempt: FeedbackAttempt): Promise<FeedbackTrajectory> {\n    const trajectory = this.trajectories.get(id)\n    if (!trajectory)\n      throw new Error(`FeedbackTrajectoryStore.appendAttempt: unknown trajectory \"${id}\"`)\n    const next = cloneTrajectory({\n      ...trajectory,\n      attempts: [...trajectory.attempts, attempt],\n      updatedAt: attempt.createdAt,\n    })\n    this.trajectories.set(id, next)\n    return cloneTrajectory(next)\n  }\n\n  async appendLabel(\n    id: string,\n    label: FeedbackLabel,\n    attemptId?: string,\n  ): Promise<FeedbackTrajectory> {\n    const trajectory = this.trajectories.get(id)\n    if (!trajectory)\n      throw new Error(`FeedbackTrajectoryStore.appendLabel: unknown trajectory \"${id}\"`)\n    const attempts = attemptId\n      ? trajectory.attempts.map((attempt) =>\n          attempt.id === attemptId\n            ? { ...attempt, feedback: [...(attempt.feedback ?? []), label] }\n            : attempt,\n        )\n      : trajectory.attempts\n    const next = cloneTrajectory({\n      ...trajectory,\n      attempts,\n      labels: attemptId ? trajectory.labels : [...trajectory.labels, label],\n      updatedAt: label.createdAt,\n    })\n    this.trajectories.set(id, next)\n    return cloneTrajectory(next)\n  }\n}\n\nexport class FileSystemFeedbackTrajectoryStore implements FeedbackTrajectoryStore {\n  private readonly dir: string\n  private readonly memory = new InMemoryFeedbackTrajectoryStore()\n  private loaded = false\n\n  constructor(options: { dir: string }) {\n    this.dir = options.dir\n  }\n\n  async save(trajectory: FeedbackTrajectory): Promise<void> {\n    await this.load()\n    await this.memory.save(trajectory)\n    await this.append({ op: 'save', trajectory })\n  }\n\n  async get(id: string): Promise<FeedbackTrajectory | null> {\n    await this.load()\n    return this.memory.get(id)\n  }\n\n  async list(filter: FeedbackTrajectoryFilter = {}): Promise<FeedbackTrajectory[]> {\n    await this.load()\n    return this.memory.list(filter)\n  }\n\n  async appendAttempt(id: string, attempt: FeedbackAttempt): Promise<FeedbackTrajectory> {\n    await this.load()\n    const next = await this.memory.appendAttempt(id, attempt)\n    await this.append({ op: 'appendAttempt', id, attempt })\n    return next\n  }\n\n  async appendLabel(\n    id: string,\n    label: FeedbackLabel,\n    attemptId?: string,\n  ): Promise<FeedbackTrajectory> {\n    await this.load()\n    const next = await this.memory.appendLabel(id, label, attemptId)\n    await this.append({ op: 'appendLabel', id, label, attemptId })\n    return next\n  }\n\n  private async append(record: unknown): Promise<void> {\n    const { appendFile, mkdir } = await import('node:fs/promises')\n    const { join } = await import('node:path')\n    await mkdir(this.dir, { recursive: true })\n    await appendFile(\n      join(this.dir, 'feedback-trajectories.ndjson'),\n      `${JSON.stringify(record)}\\n`,\n      'utf8',\n    )\n  }\n\n  private async load(): Promise<void> {\n    if (this.loaded) return\n    const { readFile } = await import('node:fs/promises')\n    const { join } = await import('node:path')\n    const file = join(this.dir, 'feedback-trajectories.ndjson')\n    try {\n      const raw = await readFile(file, 'utf8')\n      for (const line of raw.split('\\n')) {\n        if (!line.trim()) continue\n        try {\n          const record = JSON.parse(line) as\n            | { op: 'save'; trajectory: FeedbackTrajectory }\n            | { op: 'appendAttempt'; id: string; attempt: FeedbackAttempt }\n            | { op: 'appendLabel'; id: string; label: FeedbackLabel; attemptId?: string }\n          if (record.op === 'save') await this.memory.save(record.trajectory)\n          if (record.op === 'appendAttempt')\n            await this.memory.appendAttempt(record.id, record.attempt)\n          if (record.op === 'appendLabel')\n            await this.memory.appendLabel(record.id, record.label, record.attemptId)\n        } catch {\n          /* corrupt records are skipped so one bad line does not discard the corpus */\n        }\n      }\n    } catch {\n      /* first run */\n    }\n    this.loaded = true\n  }\n}\n\nexport function createFeedbackTrajectory(input: {\n  id?: string\n  projectId?: string\n  scenarioId?: string\n  task: FeedbackTask\n  attempts?: FeedbackAttempt[]\n  labels?: FeedbackLabel[]\n  outcome?: FeedbackOutcome\n  split?: DatasetSplit\n  tags?: Record<string, string>\n  createdAt?: string\n  metadata?: Record<string, unknown>\n}): FeedbackTrajectory {\n  const createdAt = input.createdAt ?? new Date().toISOString()\n  const id =\n    input.id ??\n    `ft_${stableHash(`${input.projectId ?? ''}|${input.scenarioId ?? ''}|${input.task.intent}|${createdAt}`).toString(16)}`\n  return {\n    id,\n    projectId: input.projectId,\n    scenarioId: input.scenarioId,\n    task: input.task,\n    attempts: input.attempts ?? [],\n    labels: input.labels ?? [],\n    outcome: input.outcome,\n    split: input.split,\n    tags: input.tags,\n    createdAt,\n    metadata: input.metadata,\n  }\n}\n\nexport function assignFeedbackSplit(\n  trajectory: Pick<FeedbackTrajectory, 'id' | 'projectId' | 'scenarioId' | 'task'>,\n  policy: FeedbackSplitPolicy = {},\n): DatasetSplit {\n  const split = { ...DEFAULT_SPLIT_POLICY, ...policy }\n  const total = split.trainPct + split.devPct + split.testPct + split.holdoutPct\n  if (total <= 0) throw new Error('assignFeedbackSplit: split percentages must sum above zero')\n  const bucket =\n    stableHash(\n      `${trajectory.projectId ?? ''}|${trajectory.scenarioId ?? ''}|${trajectory.id}|${trajectory.task.intent}`,\n    ) % total\n  if (bucket < split.trainPct) return 'train'\n  if (bucket < split.trainPct + split.devPct) return 'dev'\n  if (bucket < split.trainPct + split.devPct + split.testPct) return 'test'\n  return 'holdout'\n}\n\nexport function withAssignedFeedbackSplit(\n  trajectory: FeedbackTrajectory,\n  policy?: FeedbackSplitPolicy,\n): FeedbackTrajectory {\n  return {\n    ...trajectory,\n    split: trajectory.split ?? assignFeedbackSplit(trajectory, policy),\n  }\n}\n\nexport function feedbackTrajectoryToDatasetScenario(\n  trajectory: FeedbackTrajectory,\n): DatasetScenario {\n  const withSplit = withAssignedFeedbackSplit(trajectory)\n  return {\n    id: withSplit.scenarioId ?? withSplit.id,\n    split: withSplit.split,\n    payload: withSplit,\n    tags: {\n      ...(withSplit.projectId ? { projectId: withSplit.projectId } : {}),\n      ...(withSplit.tags ?? {}),\n      source: 'feedback-trajectory',\n    },\n  }\n}\n\nexport function feedbackTrajectoriesToDatasetScenarios(\n  trajectories: FeedbackTrajectory[],\n): DatasetScenario[] {\n  return trajectories.map(feedbackTrajectoryToDatasetScenario)\n}\n\nexport function feedbackTrajectoryToOptimizerRow(\n  trajectory: FeedbackTrajectory,\n): FeedbackOptimizerRow {\n  const labels = allLabels(trajectory)\n  return {\n    scenarioId: trajectory.scenarioId ?? trajectory.id,\n    trajectoryId: trajectory.id,\n    labelKinds: [...new Set(labels.map((label) => label.kind))],\n    score: trajectory.outcome?.score ?? scoreFromLabels(labels),\n    metadata: {\n      projectId: trajectory.projectId,\n      split: trajectory.split,\n      intent: trajectory.task.intent,\n      attempts: trajectory.attempts.length,\n      outcome: trajectory.outcome,\n      labels,\n    },\n  }\n}\n\nexport function feedbackTrajectoriesToOptimizerRows(\n  trajectories: FeedbackTrajectory[],\n): FeedbackOptimizerRow[] {\n  return trajectories.map(feedbackTrajectoryToOptimizerRow)\n}\n\nexport async function replayFeedbackTrajectory(\n  trajectory: FeedbackTrajectory,\n  adapter: FeedbackReplayAdapter,\n): Promise<FeedbackReplayResult> {\n  try {\n    const result = await adapter.replay(trajectory)\n    return {\n      trajectoryId: trajectory.id,\n      ...result,\n    }\n  } catch (err) {\n    const createdAt = new Date().toISOString()\n    const message = err instanceof Error ? err.message : String(err)\n    return {\n      trajectoryId: trajectory.id,\n      pass: false,\n      labels: [\n        {\n          source: 'system',\n          kind: 'reject',\n          value: false,\n          reason: message,\n          severity: 'error',\n          createdAt,\n        },\n      ],\n      outcome: {\n        success: false,\n        score: 0,\n        detail: message,\n        observedAt: createdAt,\n      },\n      metadata: { replayError: true },\n    }\n  }\n}\n\nexport async function replayFeedbackTrajectories(\n  trajectories: FeedbackTrajectory[],\n  adapter: FeedbackReplayAdapter,\n): Promise<FeedbackReplayResult[]> {\n  const results: FeedbackReplayResult[] = []\n  for (const trajectory of trajectories) {\n    results.push(await replayFeedbackTrajectory(trajectory, adapter))\n  }\n  return results\n}\n\nexport function summarizePreferenceMemory(\n  trajectories: FeedbackTrajectory[],\n  options: { maxEntries?: number } = {},\n): PreferenceMemoryEntry[] {\n  const maxEntries = options.maxEntries ?? 20\n  const entries: PreferenceMemoryEntry[] = []\n  for (const trajectory of trajectories) {\n    for (const label of allLabels(trajectory)) {\n      const instruction = instructionFromLabel(trajectory, label)\n      if (!instruction) continue\n      entries.push({\n        instruction,\n        rationale: label.reason ?? `${label.kind} label from ${label.source}`,\n        weight: weightForLabel(label),\n        sourceTrajectoryId: trajectory.id,\n        sourceLabelId: label.id,\n        category: label.kind,\n      })\n    }\n  }\n\n  const byInstruction = new Map<string, PreferenceMemoryEntry>()\n  for (const entry of entries) {\n    const key = entry.instruction.toLowerCase().replace(/\\s+/g, ' ').trim()\n    const existing = byInstruction.get(key)\n    if (!existing || entry.weight > existing.weight) byInstruction.set(key, entry)\n  }\n  return [...byInstruction.values()].sort((a, b) => b.weight - a.weight).slice(0, maxEntries)\n}\n\nexport function renderPreferenceMemoryMarkdown(entries: PreferenceMemoryEntry[]): string {\n  const lines = ['# Preference Memory', '']\n  for (const entry of entries) {\n    lines.push(`- ${entry.instruction}`)\n    lines.push(`  Rationale: ${entry.rationale}`)\n    lines.push(`  Source: ${entry.sourceTrajectoryId}`)\n    lines.push('')\n  }\n  return `${lines.join('\\n').trim()}\\n`\n}\n\nexport function serializeFeedbackTrajectoriesJsonl(trajectories: FeedbackTrajectory[]): string {\n  return `${trajectories\n    .slice()\n    .sort((a, b) => a.id.localeCompare(b.id))\n    .map((trajectory) => JSON.stringify(canonicalize(trajectory)))\n    .join('\\n')}\\n`\n}\n\nexport function parseFeedbackTrajectoriesJsonl(jsonl: string): FeedbackTrajectory[] {\n  const trajectories: FeedbackTrajectory[] = []\n  for (const line of jsonl.split('\\n')) {\n    if (!line.trim()) continue\n    trajectories.push(JSON.parse(line) as FeedbackTrajectory)\n  }\n  return trajectories\n}\n\nexport function controlRunToFeedbackTrajectory<TState, TAction, TActionResult>(\n  run: ControlRunResult<TState, TAction, TActionResult>,\n  options: {\n    projectId?: string\n    scenarioId?: string\n    artifactType?: FeedbackArtifactType\n    artifactFromStep?: (step: ControlStep<TState, TAction, TActionResult>) => unknown\n    proposedActionFromStep?: (\n      step: ControlStep<TState, TAction, TActionResult>,\n    ) => ProposedSideEffect | undefined\n    createdAt?: string\n  } = {},\n): FeedbackTrajectory {\n  const createdAt = options.createdAt ?? new Date().toISOString()\n  const trajectoryId =\n    run.runId ?? `ft_control_${stableHash(`${run.intent}|${createdAt}`).toString(16)}`\n  return createFeedbackTrajectory({\n    id: trajectoryId,\n    projectId: options.projectId,\n    scenarioId: options.scenarioId,\n    task: { intent: run.intent },\n    createdAt,\n    attempts: run.steps.map((step) => ({\n      id: `${trajectoryId}_step_${step.index}`,\n      stepIndex: step.index,\n      artifactType: options.artifactType ?? 'action',\n      artifact: options.artifactFromStep?.(step) ?? step.actionOutcome?.result ?? step.decision,\n      proposedAction: options.proposedActionFromStep?.(step),\n      evals: step.evalsAfter,\n      createdAt: step.startedAt,\n      metadata: {\n        decision: step.decision,\n        actionOutcome: step.actionOutcome,\n      },\n    })),\n    labels: [\n      {\n        source: 'system',\n        kind: run.pass ? 'approve' : 'reject',\n        value: run.pass,\n        reason: run.reason,\n        severity: run.pass ? 'info' : 'error',\n        createdAt,\n      },\n    ],\n    outcome: {\n      success: run.pass,\n      score: run.score,\n      costUsd: run.spentCostUsd,\n      detail: run.reason,\n      observedAt: createdAt,\n      metadata: {\n        stoppedBy: run.stoppedBy,\n        failureClass: run.failureClass,\n      },\n    },\n  })\n}\n\nfunction allLabels(trajectory: FeedbackTrajectory): FeedbackLabel[] {\n  const labels = [\n    ...trajectory.labels,\n    ...trajectory.attempts.flatMap((attempt) => attempt.feedback ?? []),\n  ]\n  const seen = new Set<string>()\n  return labels.filter((label) => {\n    const key =\n      label.id ?? `${label.source}|${label.kind}|${label.createdAt}|${JSON.stringify(label.value)}`\n    if (seen.has(key)) return false\n    seen.add(key)\n    return true\n  })\n}\n\nfunction scoreFromLabels(labels: FeedbackLabel[]): number | undefined {\n  if (!labels.length) return undefined\n  const scored = labels\n    .map((label) => {\n      if (label.kind === 'approve' || label.kind === 'select') return 1\n      if (label.kind === 'reject' || label.kind === 'policy_block') return 0\n      if (label.kind === 'rate' && typeof label.value === 'number')\n        return Math.max(0, Math.min(1, label.value))\n      return undefined\n    })\n    .filter((value): value is number => typeof value === 'number')\n  if (!scored.length) return undefined\n  return Math.round((scored.reduce((sum, value) => sum + value, 0) / scored.length) * 1000) / 1000\n}\n\nfunction instructionFromLabel(\n  trajectory: FeedbackTrajectory,\n  label: FeedbackLabel,\n): string | undefined {\n  if (label.kind === 'reject' && label.reason)\n    return `Avoid outputs like \"${compact(trajectory.task.intent, 80)}\" when: ${label.reason}`\n  if (label.kind === 'revision_request' && label.reason)\n    return `Revise similar work by applying: ${label.reason}`\n  if (label.kind === 'select' && label.reason)\n    return `Prefer selected options for \"${compact(trajectory.task.intent, 80)}\" because: ${label.reason}`\n  if (label.kind === 'approve' && label.reason)\n    return `Repeat the pattern approved for \"${compact(trajectory.task.intent, 80)}\": ${label.reason}`\n  if (label.kind === 'comment' && label.reason) return label.reason\n  return undefined\n}\n\nfunction weightForLabel(label: FeedbackLabel): number {\n  const severity =\n    label.severity === 'critical'\n      ? 4\n      : label.severity === 'error'\n        ? 3\n        : label.severity === 'warning'\n          ? 2\n          : 1\n  const source =\n    label.source === 'user'\n      ? 3\n      : label.source === 'metric' || label.source === 'environment'\n        ? 2\n        : 1\n  return severity * source\n}\n\nfunction matchesFilter(trajectory: FeedbackTrajectory, filter: FeedbackTrajectoryFilter): boolean {\n  if (filter.projectId && trajectory.projectId !== filter.projectId) return false\n  if (filter.scenarioId && trajectory.scenarioId !== filter.scenarioId) return false\n  if (filter.split && trajectory.split !== filter.split) return false\n  if (filter.tag) {\n    const [key, value] = filter.tag\n    if (trajectory.tags?.[key] !== value) return false\n  }\n  return true\n}\n\nfunction cloneTrajectory(trajectory: FeedbackTrajectory): FeedbackTrajectory {\n  return JSON.parse(JSON.stringify(trajectory)) as FeedbackTrajectory\n}\n\nfunction compact(value: string, max: number): string {\n  const normalized = value.replace(/\\s+/g, ' ').trim()\n  return normalized.length > max ? `${normalized.slice(0, max).trim()}...` : normalized\n}\n\nfunction stableHash(input: string): number {\n  let hash = 2166136261\n  for (let i = 0; i < input.length; i += 1) {\n    hash ^= input.charCodeAt(i)\n    hash = Math.imul(hash, 16777619)\n  }\n  return hash >>> 0\n}\n\nfunction canonicalize(value: unknown): unknown {\n  if (value === null || typeof value !== 'object') return value\n  if (Array.isArray(value)) return value.map(canonicalize)\n  const out: Record<string, unknown> = {}\n  for (const key of Object.keys(value as Record<string, unknown>).sort()) {\n    out[key] = canonicalize((value as Record<string, unknown>)[key])\n  }\n  return out\n}\n","/**\n * HeldOutGate — first-class held-out paired-delta promotion gate.\n *\n * Encodes the \"honesty override\" pattern that lived inline in\n * `~/webb/redteam/scripts/agent-eval-autoresearch.ts:138–171`.\n * The optimizer's best-guess is one thing; what we should actually\n * ship is another. The gate is the line between them.\n *\n * A candidate is promoted iff ALL three pass:\n *\n *   1. **Productive runs**: the candidate has at least\n *      `minProductiveRuns` paired observations on items where BOTH\n *      candidate and baseline produced a real (non-silent) score.\n *   2. **Paired delta**: the lower bound of the bootstrap CI on the\n *      median per-item delta (candidate − baseline) on the HOLDOUT\n *      split is strictly greater than `pairedDeltaThreshold`.\n *   3. **Overfit gap**: the candidate's gap between search-split\n *      score and holdout-split score is no worse (more positive)\n *      than the baseline's gap by more than `overfitGapThreshold`.\n *      \"Better on search, worse on holdout\" is the canonical\n *      overfit pattern; this catches it.\n *\n * The decision carries a machine-readable `rejectionCode` plus an\n * `evidence` block with every number the gate looked at, so the\n * downstream researcher / paper / dashboard can re-derive the\n * verdict without re-running.\n *\n * See also:\n *   - `src/paired-stats.ts` for `pairedBootstrap` + `pairedWilcoxon`\n *   - `src/run-record.ts` for the input row schema\n *   - `src/reference-replay.ts` for the older, reference-replay-\n *     specific promotion path (still useful for replay-style evals).\n */\n\nimport { pairedBootstrap, pairedWilcoxon } from './paired-stats'\nimport type { RunRecord } from './run-record'\n\nexport type HeldOutGateRejectionCode = 'few_runs' | 'negative_delta' | 'overfit_gap'\n\nexport interface HeldOutGateConfig {\n  /** Minimum number of paired (candidate, baseline) holdout observations\n   *  required before the gate will even consider promoting. Default 3. */\n  minProductiveRuns?: number\n  /** The bootstrap-CI lower bound on the median paired holdout delta\n   *  must exceed this to promote. Default 0. */\n  pairedDeltaThreshold?: number\n  /** Maximum allowed worsening of (search − holdout) gap relative to\n   *  baseline. Default 0.15 (i.e. candidate may overfit by up to 15\n   *  absolute score points more than baseline before rejection). */\n  overfitGapThreshold?: number\n  /** Stable label of the baseline candidate. Required — paper-grade\n   *  evaluation never compares two unlabelled candidates. */\n  baselineKey: string\n  /** Confidence level for the bootstrap CI. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  bootstrapResamples?: number\n  /** Optional deterministic seed for the bootstrap. Default undefined\n   *  (Math.random). */\n  seed?: number\n}\n\nexport interface GateEvidence {\n  /** Number of paired (candidate, baseline) holdout observations used. */\n  productiveRuns: number\n  /** Median of (candidate − baseline) paired holdout deltas. */\n  medianPairedDelta: number\n  /** Bootstrap CI on the median paired holdout delta. */\n  pairedCI: { low: number; high: number }\n  /** Wilcoxon signed-rank p-value on the paired holdout deltas. */\n  pairedPValue: number\n  /** Mean candidate score on the search split (NaN if none). */\n  searchScore: number\n  /** Mean candidate score on the holdout split (NaN if none). */\n  holdoutScore: number\n  /** Candidate (search − holdout) gap. */\n  overfitGap: number\n  /** Baseline (search − holdout) gap. */\n  baselineOverfitGap: number\n}\n\nexport interface GateDecision {\n  /** Final promote/no-promote verdict. */\n  promote: boolean\n  /** The candidate that was evaluated. */\n  candidateId: string\n  /** The baseline it was compared against. */\n  baselineId: string\n  /** Every number the gate looked at, for audit + paper export. */\n  evidence: GateEvidence\n  /** Human-readable reason. */\n  reason: string\n  /** Machine-readable rejection code, or null on promote. */\n  rejectionCode: HeldOutGateRejectionCode | null\n}\n\n/**\n * Held-out paired-delta promotion gate. Construct once with config,\n * call `evaluate(candidateRuns, baselineRuns)` per (candidate,\n * baseline) pair. Stateless across calls.\n */\nexport class HeldOutGate {\n  private readonly minProductiveRuns: number\n  private readonly pairedDeltaThreshold: number\n  private readonly overfitGapThreshold: number\n  private readonly baselineKey: string\n  private readonly confidence: number\n  private readonly resamples: number\n  private readonly seed?: number\n\n  constructor(config: HeldOutGateConfig) {\n    if (!config.baselineKey) {\n      throw new Error('HeldOutGate: baselineKey is required')\n    }\n    this.minProductiveRuns = config.minProductiveRuns ?? 3\n    this.pairedDeltaThreshold = config.pairedDeltaThreshold ?? 0\n    this.overfitGapThreshold = config.overfitGapThreshold ?? 0.15\n    this.baselineKey = config.baselineKey\n    this.confidence = config.confidence ?? 0.95\n    this.resamples = config.bootstrapResamples ?? 2000\n    this.seed = config.seed\n  }\n\n  /** Decide whether `candidate` should replace `baseline`. Pairing\n   *  is by (experimentId, seed) — identical experiment + seed pairs\n   *  the candidate run with the matching baseline run. Pairs without\n   *  a holdout score on both sides are dropped. */\n  evaluate(candidate: RunRecord[], baseline: RunRecord[]): GateDecision {\n    const candidateId = inferCandidateId(candidate, this.baselineKey)\n    const baselineId = this.baselineKey\n\n    // Pair holdout runs by (experimentId, seed).\n    const baselineHoldoutByKey = indexHoldoutByKey(baseline)\n    const beforeHoldout: number[] = []\n    const afterHoldout: number[] = []\n    for (const run of candidate) {\n      if (run.splitTag !== 'holdout') continue\n      if (run.outcome.holdoutScore === undefined) continue\n      const key = pairKey(run)\n      const counterpart = baselineHoldoutByKey.get(key)\n      if (counterpart === undefined) continue\n      beforeHoldout.push(counterpart)\n      afterHoldout.push(run.outcome.holdoutScore)\n    }\n\n    const productiveRuns = beforeHoldout.length\n\n    // Always compute the gap numbers — useful even when we reject on\n    // few_runs (you want to see why).\n    const candidateSearchMean = mean(scores(candidate, 'searchScore', 'search'))\n    const candidateHoldoutMean = mean(scores(candidate, 'holdoutScore', 'holdout'))\n    const baselineSearchMean = mean(scores(baseline, 'searchScore', 'search'))\n    const baselineHoldoutMean = mean(scores(baseline, 'holdoutScore', 'holdout'))\n\n    const overfitGap = safeDiff(candidateSearchMean, candidateHoldoutMean)\n    const baselineOverfitGap = safeDiff(baselineSearchMean, baselineHoldoutMean)\n\n    // Few-runs gate.\n    if (productiveRuns < this.minProductiveRuns) {\n      return {\n        promote: false,\n        candidateId,\n        baselineId,\n        evidence: {\n          productiveRuns,\n          medianPairedDelta: productiveRuns > 0 ? medianDelta(beforeHoldout, afterHoldout) : 0,\n          pairedCI: { low: 0, high: 0 },\n          pairedPValue: 1,\n          searchScore: candidateSearchMean,\n          holdoutScore: candidateHoldoutMean,\n          overfitGap,\n          baselineOverfitGap,\n        },\n        reason: `few_runs: ${productiveRuns} paired holdout observation(s) < min ${this.minProductiveRuns}`,\n        rejectionCode: 'few_runs',\n      }\n    }\n\n    // Paired bootstrap on holdout deltas.\n    const ci = pairedBootstrap(beforeHoldout, afterHoldout, {\n      confidence: this.confidence,\n      resamples: this.resamples,\n      statistic: 'median',\n      seed: this.seed,\n    })\n    const wilcoxon = pairedWilcoxon(beforeHoldout, afterHoldout)\n\n    const evidence: GateEvidence = {\n      productiveRuns,\n      medianPairedDelta: ci.median,\n      pairedCI: { low: ci.low, high: ci.high },\n      pairedPValue: wilcoxon.p,\n      searchScore: candidateSearchMean,\n      holdoutScore: candidateHoldoutMean,\n      overfitGap,\n      baselineOverfitGap,\n    }\n\n    // Negative-delta gate (CI lower bound must clear the threshold).\n    if (!(ci.low > this.pairedDeltaThreshold)) {\n      return {\n        promote: false,\n        candidateId,\n        baselineId,\n        evidence,\n        reason:\n          `negative_delta: paired holdout median Δ=${fmt(ci.median)} ` +\n          `CI=[${fmt(ci.low)}, ${fmt(ci.high)}] does not clear threshold ${fmt(this.pairedDeltaThreshold)}`,\n        rejectionCode: 'negative_delta',\n      }\n    }\n\n    // Overfit-gap gate. We allow some absolute slack —\n    // candidate.gap ≤ baseline.gap + overfitGapThreshold.\n    if (\n      Number.isFinite(overfitGap) &&\n      Number.isFinite(baselineOverfitGap) &&\n      overfitGap > baselineOverfitGap + this.overfitGapThreshold\n    ) {\n      return {\n        promote: false,\n        candidateId,\n        baselineId,\n        evidence,\n        reason:\n          `overfit_gap: candidate gap=${fmt(overfitGap)} exceeds baseline gap=${fmt(baselineOverfitGap)} ` +\n          `by more than ${fmt(this.overfitGapThreshold)}`,\n        rejectionCode: 'overfit_gap',\n      }\n    }\n\n    return {\n      promote: true,\n      candidateId,\n      baselineId,\n      evidence,\n      reason:\n        `promote: paired holdout median Δ=${fmt(ci.median)} ` +\n        `CI=[${fmt(ci.low)}, ${fmt(ci.high)}] over ${productiveRuns} pairs; ` +\n        `overfit gap candidate=${fmt(overfitGap)} vs baseline=${fmt(baselineOverfitGap)}`,\n      rejectionCode: null,\n    }\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction inferCandidateId(candidate: RunRecord[], baselineKey: string): string {\n  for (const run of candidate) {\n    if (run.candidateId && run.candidateId !== baselineKey) return run.candidateId\n  }\n  // All candidate rows match the baseline key — caller mistake, but\n  // surface the symptom rather than throwing inside the gate.\n  return candidate[0]?.candidateId ?? '(unknown candidate)'\n}\n\nfunction indexHoldoutByKey(runs: RunRecord[]): Map<string, number> {\n  const out = new Map<string, number>()\n  for (const r of runs) {\n    if (r.splitTag !== 'holdout') continue\n    if (r.outcome.holdoutScore === undefined) continue\n    out.set(pairKey(r), r.outcome.holdoutScore)\n  }\n  return out\n}\n\nfunction pairKey(r: RunRecord): string {\n  return `${r.experimentId}::${r.seed}`\n}\n\nfunction scores(\n  runs: RunRecord[],\n  field: 'searchScore' | 'holdoutScore',\n  splitFilter: 'search' | 'holdout',\n): number[] {\n  const out: number[] = []\n  for (const r of runs) {\n    if (r.splitTag !== splitFilter) continue\n    const v = r.outcome[field]\n    if (typeof v === 'number' && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return Number.NaN\n  return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction safeDiff(a: number, b: number): number {\n  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN\n  return a - b\n}\n\nfunction medianDelta(before: number[], after: number[]): number {\n  const ds = before.map((b, i) => after[i]! - b).sort((x, y) => x - y)\n  if (ds.length === 0) return 0\n  const mid = Math.floor(ds.length / 2)\n  return ds.length % 2 === 0 ? (ds[mid - 1]! + ds[mid]!) / 2 : ds[mid]!\n}\n\nfunction fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/**\n * Pareto frontier — multi-objective optimization over candidate runs.\n *\n * Lifted from ADC pareto.ts and blueprint-agent frontier.ts. When you're\n * trading off (cost, latency, quality) or (passRate, tokenBudget,\n * ttfb), you rarely have a single \"winner\" — you have a set of\n * non-dominated candidates. This module exposes:\n *\n *   - `paretoFrontier`: filter a set of candidates to the non-dominated ones\n *   - `dominates`: does A dominate B across all objectives?\n *\n * Each objective is declared with a direction: 'maximize' (higher=better)\n * or 'minimize' (lower=better). Candidates are any object; pass an\n * `objective(candidate)` accessor.\n */\n\nexport type Direction = 'maximize' | 'minimize'\n\nexport interface Objective<T> {\n  /** Stable label used in reports. */\n  name: string\n  direction: Direction\n  value: (candidate: T) => number\n}\n\nexport interface ParetoResult<T> {\n  frontier: T[]\n  dominated: T[]\n  /** Index map: frontier[i] dominates each of dominatedBy[i]. */\n  dominanceMap: Array<{ dominator: T; dominated: T[] }>\n}\n\n/** Does candidate A weakly dominate B — strictly better on at least one objective and no worse on any? */\nexport function dominates<T>(a: T, b: T, objectives: Objective<T>[]): boolean {\n  let strictlyBetter = false\n  for (const obj of objectives) {\n    const av = obj.value(a)\n    const bv = obj.value(b)\n    if (!Number.isFinite(av) || !Number.isFinite(bv)) return false\n    const aIsBetter = obj.direction === 'maximize' ? av > bv : av < bv\n    const aIsWorse = obj.direction === 'maximize' ? av < bv : av > bv\n    if (aIsWorse) return false\n    if (aIsBetter) strictlyBetter = true\n  }\n  return strictlyBetter\n}\n\n/**\n * Compute the non-dominated frontier. Candidates with NaN/Infinity on any\n * objective are excluded (can't rank them). A candidate enters the frontier\n * iff no other candidate dominates it.\n */\nexport function paretoFrontier<T>(candidates: T[], objectives: Objective<T>[]): ParetoResult<T> {\n  if (objectives.length === 0) {\n    throw new Error('paretoFrontier: at least 1 objective required')\n  }\n  const valid = candidates.filter((c) => objectives.every((o) => Number.isFinite(o.value(c))))\n  const frontier: T[] = []\n  const dominated: T[] = []\n  for (const c of valid) {\n    const isDominated = valid.some((other) => other !== c && dominates(other, c, objectives))\n    if (isDominated) dominated.push(c)\n    else frontier.push(c)\n  }\n  const dominanceMap = frontier.map((d) => ({\n    dominator: d,\n    dominated: dominated.filter((x) => dominates(d, x, objectives)),\n  }))\n  return { frontier, dominated, dominanceMap }\n}\n\n/**\n * Weighted-sum scalarisation. Use as a tie-break / single-winner selector\n * when callers don't want to consume a frontier. Each objective contributes\n * its normalised value (0..1 via min-max across the candidate pool) times\n * its weight; missing weights default to 1/N.\n *\n * Direction is honoured automatically — `minimize` axes have their values\n * inverted before scaling so \"higher scalar = better\" always holds.\n */\nexport function scalarScore<T>(\n  candidates: T[],\n  objectives: Objective<T>[],\n  options: { weights?: Partial<Record<string, number>> } = {},\n): Array<{ candidate: T; score: number }> {\n  if (candidates.length === 0) return []\n  const weights = options.weights ?? {}\n  const totalWeight = objectives.reduce((s, o) => s + (weights[o.name] ?? 1), 0)\n\n  // Pre-compute min/max per objective for normalisation.\n  const ranges = objectives.map((obj) => {\n    const values = candidates.map((c) => obj.value(c)).filter((v) => Number.isFinite(v))\n    if (values.length === 0) return { min: 0, max: 1 }\n    const min = Math.min(...values)\n    const max = Math.max(...values)\n    return { min, max: max === min ? min + 1 : max }\n  })\n\n  return candidates.map((c) => {\n    let score = 0\n    objectives.forEach((obj, i) => {\n      const v = obj.value(c)\n      if (!Number.isFinite(v)) return\n      const { min, max } = ranges[i]!\n      const normalised = (v - min) / (max - min)\n      const directional = obj.direction === 'maximize' ? normalised : 1 - normalised\n      const weight = (weights[obj.name] ?? 1) / totalWeight\n      score += directional * weight\n    })\n    return { candidate: c, score }\n  })\n}\n\n/**\n * NSGA-II crowding distance — secondary sort for ties on the frontier.\n *\n * When the Pareto front collapses to a single point (or many candidates tie\n * on dominance), naive selection picks arbitrarily and the population\n * degenerates over generations. NSGA-II preserves diversity by preferring\n * candidates with more empty space around them on the frontier.\n *\n * Returns an array of `{ candidate, distance }` in the SAME order as the\n * input. Higher distance = more isolated = should be preferred when\n * preserving diversity.\n */\nexport function crowdingDistance<T>(\n  candidates: T[],\n  objectives: Objective<T>[],\n): Array<{ candidate: T; distance: number }> {\n  const distances = new Map<T, number>(candidates.map((c) => [c, 0]))\n\n  for (const obj of objectives) {\n    const sorted = [...candidates].sort((a, b) => obj.value(a) - obj.value(b))\n    const min = obj.value(sorted[0]!)\n    const max = obj.value(sorted[sorted.length - 1]!)\n    const range = max - min || 1\n\n    // Boundary points get infinity (always preferred for diversity).\n    distances.set(sorted[0]!, Infinity)\n    distances.set(sorted[sorted.length - 1]!, Infinity)\n    for (let i = 1; i < sorted.length - 1; i++) {\n      const prev = obj.value(sorted[i - 1]!)\n      const next = obj.value(sorted[i + 1]!)\n      const current = distances.get(sorted[i]!)!\n      if (current === Infinity) continue\n      distances.set(sorted[i]!, current + (next - prev) / range)\n    }\n  }\n\n  return candidates.map((c) => ({ candidate: c, distance: distances.get(c) ?? 0 }))\n}\n\n/**\n * Pareto frontier with tie-break by crowding distance — the canonical\n * NSGA-II selection step. Returns the frontier sorted by descending crowding\n * distance so callers can `.slice(0, k)` to pick K diverse winners.\n */\nexport function paretoFrontierWithCrowding<T>(\n  candidates: T[],\n  objectives: Objective<T>[],\n): Array<{ candidate: T; distance: number }> {\n  const { frontier } = paretoFrontier(candidates, objectives)\n  if (frontier.length === 0) return []\n  const distances = crowdingDistance(frontier, objectives)\n  return distances.sort((a, b) => b.distance - a.distance)\n}\n","/**\n * PromptEvolutionLoop — population-based reflective-mutation evolution.\n *\n * Above the existing `AxGepaSteeringOptimizer` (which RANKS variants),\n * this loop GENERATES variants. Each generation:\n *   1. Score the population across (variant × scenario × rep).\n *   2. Pick survivors from the Pareto frontier (with crowding-distance tie-break).\n *   3. Ask the mutator for replacements until population size is restored.\n *   4. Repeat for N generations OR until convergence.\n *\n * Domain-agnostic. Consumers supply:\n *   - A seed population of `EvolvableVariant`s.\n *   - A `ScoreAdapter` that runs (variant, scenario, rep) → `TrialResult`.\n *   - A `MutateAdapter` that produces children given trace evidence.\n *   - Pareto `Objective<TrialAggregate>[]` defining the multi-objective vector.\n *\n * The loop owns: population management, parallel scheduling (concurrency-\n * limited), Pareto selection with crowding distance, generation reporting.\n *\n * It does NOT own: rendering trials to a model, executing prompts, choosing\n * mutation primitives, persisting to disk. Those are the consumer's call.\n */\n\nimport { type Objective, paretoFrontierWithCrowding, scalarScore } from './pareto'\n\nexport interface EvolvableVariant<P = unknown> {\n  /** Stable id for the variant — surfaces in reports and trial results. */\n  id: string\n  /** Variant payload — interpretation is the consumer's responsibility. */\n  payload: P\n  /** Generation index (0 = seed, then 1, 2, ...). */\n  generation: number\n  /** Parent variant id when produced via mutation; absent for seeds. */\n  parentId?: string\n  /** Human label for reports. */\n  label: string\n  /** What the mutator was trying to fix. */\n  rationale?: string\n}\n\nexport interface TrialResult {\n  variantId: string\n  scenarioId: string\n  rep: number\n  ok: boolean\n  /** Primary scalar score the consumer cares about (e.g., recall, accuracy). */\n  score: number\n  /** Token cost (or any cost-like dimension). */\n  cost?: number\n  /** Wall time in ms. */\n  durationMs?: number\n  /** Free-form metric bag for objective accessors. */\n  metrics?: Record<string, number>\n  error?: string\n  /**\n   * Whether the judge LLM call(s) that produced this trial's score actually\n   * completed. `undefined` means \"consumer didn't report\"; `false` means\n   * the judge aborted/failed and the score is synthetic (typically 0 or\n   * partial). `aggregateTrials({mode: 'exclude-failed'})` skips these\n   * trials so a silent-zero judge can't pollute the composite.\n   */\n  judgeSucceeded?: boolean\n  /** Number of judge attempts (informational, populated by `withJudgeRetry`). */\n  judgeAttempts?: number\n  /** Last judge error message when `judgeSucceeded === false`. */\n  judgeError?: string\n}\n\n/** Aggregated trial summary for one (variant, scenario) pair across reps. */\nexport interface ScenarioAggregate {\n  variantId: string\n  scenarioId: string\n  meanScore: number\n  meanCost: number\n  meanDurationMs: number\n  okRate: number\n  trials: number\n  /** Mean of every numeric metric across reps. */\n  metrics: Record<string, number>\n}\n\n/** Aggregated trial summary for one variant across all scenarios. */\nexport interface VariantAggregate {\n  variantId: string\n  meanScore: number\n  meanCost: number\n  meanDurationMs: number\n  okRate: number\n  scenarios: ScenarioAggregate[]\n  /** Mean of every numeric metric, averaged across scenarios. */\n  metrics: Record<string, number>\n}\n\nexport interface ScoreAdapter<P = unknown> {\n  score(args: {\n    variant: EvolvableVariant<P>\n    scenarioId: string\n    rep: number\n  }): Promise<TrialResult>\n}\n\nexport interface MutateAdapter<P = unknown> {\n  mutate(args: {\n    parent: EvolvableVariant<P>\n    parentAggregate: VariantAggregate\n    topTrials: TrialResult[]\n    bottomTrials: TrialResult[]\n    childCount: number\n    generation: number\n  }): Promise<EvolvableVariant<P>[]>\n}\n\nexport interface PromptEvolutionConfig<P = unknown> {\n  runId: string\n  /** What component is being mutated — surfaces in reports + reflection prompts. */\n  target: string\n  seedVariants: EvolvableVariant<P>[]\n  scenarioIds: string[]\n  reps: number\n  generations: number\n  populationSize: number\n  /** Maximum concurrent score() calls. */\n  scoreConcurrency: number\n  scoreAdapter: ScoreAdapter<P>\n  mutateAdapter: MutateAdapter<P>\n  /** Pareto objectives over `VariantAggregate`. Ordered by importance. */\n  objectives: Objective<VariantAggregate>[]\n  /** Optional weights for the scalar tie-break selector (by objective name). */\n  scalarWeights?: Record<string, number>\n  /** Stop early if a generation produces no Pareto improvement. Default true. */\n  earlyStopOnNoImprovement?: boolean\n  onProgress?: (event: PromptEvolutionEvent) => void\n  /**\n   * Optional cache key for memoising scored (variantId, scenarioId, rep)\n   * tuples. When provided AND a cache instance is passed, repeated trials\n   * skip re-scoring. Cache keys are stable across runs.\n   */\n  cache?: TrialCache\n}\n\nexport interface TrialCache {\n  get(key: string): TrialResult | undefined\n  set(key: string, value: TrialResult): void\n}\n\nexport class InMemoryTrialCache implements TrialCache {\n  private store = new Map<string, TrialResult>()\n  get(key: string): TrialResult | undefined {\n    return this.store.get(key)\n  }\n  set(key: string, value: TrialResult): void {\n    this.store.set(key, value)\n  }\n  size(): number {\n    return this.store.size\n  }\n  clear(): void {\n    this.store.clear()\n  }\n}\n\nexport type PromptEvolutionEvent =\n  | { type: 'generation-start'; generation: number; populationSize: number }\n  | {\n      type: 'trial-complete'\n      generation: number\n      variantId: string\n      scenarioId: string\n      rep: number\n      ok: boolean\n      score: number\n      cached: boolean\n    }\n  | { type: 'generation-complete'; report: GenerationReport<unknown> }\n  | { type: 'converged'; generation: number; reason: string }\n\nexport interface GenerationReport<P = unknown> {\n  runId: string\n  target: string\n  generation: number\n  variants: EvolvableVariant<P>[]\n  aggregates: VariantAggregate[]\n  /** Frontier candidates, sorted by descending crowding distance. */\n  paretoFrontIds: string[]\n  /** Scalar-best variant id — used for the single \"winner\" if callers want one. */\n  winnerId: string\n  /** Trials that fed this generation (kept for downstream reporting). */\n  trials: TrialResult[]\n}\n\nexport interface PromptEvolutionResult<P = unknown> {\n  runId: string\n  target: string\n  generations: GenerationReport<P>[]\n  /** Best variant by scalar score in the final generation. */\n  bestVariant: EvolvableVariant<P>\n  /** Best aggregate (matches bestVariant). */\n  bestAggregate: VariantAggregate\n}\n\nexport async function runPromptEvolution<P>(\n  config: PromptEvolutionConfig<P>,\n): Promise<PromptEvolutionResult<P>> {\n  const generations: GenerationReport<P>[] = []\n  let population = [...config.seedVariants]\n  let bestVariant: EvolvableVariant<P> = population[0]!\n  let bestAggregate: VariantAggregate | null = null\n\n  for (let generation = 0; generation < config.generations; generation++) {\n    config.onProgress?.({ type: 'generation-start', generation, populationSize: population.length })\n\n    const trials = await scorePopulation(population, config, generation)\n    const aggregates = aggregateTrials(population, config.scenarioIds, trials)\n\n    const front = paretoFrontierWithCrowding(aggregates, config.objectives)\n    const frontIds = new Set(front.map((c) => c.candidate.variantId))\n\n    const scored = scalarScore(aggregates, config.objectives, { weights: config.scalarWeights })\n    scored.sort((a, b) => b.score - a.score)\n    const winnerId = scored[0]?.candidate.variantId ?? aggregates[0]?.variantId ?? population[0]!.id\n\n    const report: GenerationReport<P> = {\n      runId: config.runId,\n      target: config.target,\n      generation,\n      variants: population,\n      aggregates,\n      paretoFrontIds: front.map((c) => c.candidate.variantId),\n      winnerId,\n      trials,\n    }\n    generations.push(report)\n    config.onProgress?.({ type: 'generation-complete', report })\n\n    const winnerAgg = aggregates.find((a) => a.variantId === winnerId)\n    if (winnerAgg) {\n      const winner = population.find((v) => v.id === winnerId)\n      if (winner) bestVariant = winner\n      bestAggregate = winnerAgg\n    }\n\n    // Convergence: no Pareto-or-scalar improvement vs previous generation.\n    if (config.earlyStopOnNoImprovement !== false && generations.length >= 2) {\n      const prev = generations[generations.length - 2]!\n      const noChange =\n        prev.winnerId === winnerId && samePopulation(prev.paretoFrontIds, [...frontIds])\n      if (noChange) {\n        config.onProgress?.({\n          type: 'converged',\n          generation,\n          reason: 'no improvement vs previous generation',\n        })\n        break\n      }\n    }\n\n    if (generation === config.generations - 1) break\n\n    population = await nextPopulation(population, aggregates, trials, front, config, generation + 1)\n  }\n\n  return {\n    runId: config.runId,\n    target: config.target,\n    generations,\n    bestVariant,\n    bestAggregate:\n      bestAggregate ??\n      aggregateTrials(population, config.scenarioIds, []).find(\n        (a) => a.variantId === bestVariant.id,\n      )!,\n  }\n}\n\nasync function scorePopulation<P>(\n  population: EvolvableVariant<P>[],\n  config: PromptEvolutionConfig<P>,\n  generation: number,\n): Promise<TrialResult[]> {\n  const jobs: Array<() => Promise<TrialResult>> = []\n  for (const variant of population) {\n    for (const scenarioId of config.scenarioIds) {\n      for (let rep = 0; rep < config.reps; rep++) {\n        jobs.push(async () => {\n          const cacheKey = `${variant.id}|${scenarioId}|${rep}`\n          const cached = config.cache?.get(cacheKey)\n          if (cached) {\n            config.onProgress?.({\n              type: 'trial-complete',\n              generation,\n              variantId: variant.id,\n              scenarioId,\n              rep,\n              ok: cached.ok,\n              score: cached.score,\n              cached: true,\n            })\n            return cached\n          }\n          const result = await config.scoreAdapter.score({ variant, scenarioId, rep })\n          config.cache?.set(cacheKey, result)\n          config.onProgress?.({\n            type: 'trial-complete',\n            generation,\n            variantId: variant.id,\n            scenarioId,\n            rep,\n            ok: result.ok,\n            score: result.score,\n            cached: false,\n          })\n          return result\n        })\n      }\n    }\n  }\n  return runWithConcurrency(jobs, config.scoreConcurrency)\n}\n\nasync function runWithConcurrency<T>(\n  jobs: Array<() => Promise<T>>,\n  concurrency: number,\n): Promise<T[]> {\n  const results: T[] = new Array(jobs.length)\n  const limit = Math.max(1, concurrency)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= jobs.length) return\n      results[i] = await jobs[i]!()\n    }\n  }\n  await Promise.all(Array.from({ length: limit }, () => worker()))\n  return results\n}\n\nfunction aggregateTrials<P>(\n  population: EvolvableVariant<P>[],\n  scenarioIds: string[],\n  trials: TrialResult[],\n): VariantAggregate[] {\n  return population.map((variant) => {\n    const variantTrials = trials.filter((t) => t.variantId === variant.id)\n    const scenarios = scenarioIds.map<ScenarioAggregate>((sid) => {\n      const scenarioTrials = variantTrials.filter((t) => t.scenarioId === sid)\n      const okTrials = scenarioTrials.filter((t) => t.ok)\n      // Mean score must include every successfully-graded trial — a trial\n      // with score=0.6 and ok=false (below quality_bar) is real signal, not\n      // noise. Only `error` trials (agent crash, judge crash) carry a\n      // synthetic score and are excluded. okRate continues to reflect the\n      // pass/fail rate against the configured quality_bar.\n      const gradedTrials = scenarioTrials.filter((t) => !t.error)\n      const metrics = aggregateMetrics(gradedTrials.map((t) => t.metrics ?? {}))\n      return {\n        variantId: variant.id,\n        scenarioId: sid,\n        meanScore: mean(gradedTrials.map((t) => t.score)),\n        meanCost: mean(gradedTrials.map((t) => t.cost ?? 0)),\n        meanDurationMs: mean(gradedTrials.map((t) => t.durationMs ?? 0)),\n        okRate: scenarioTrials.length === 0 ? 0 : okTrials.length / scenarioTrials.length,\n        trials: scenarioTrials.length,\n        metrics,\n      }\n    })\n    return {\n      variantId: variant.id,\n      meanScore: mean(scenarios.map((s) => s.meanScore)),\n      meanCost: mean(scenarios.map((s) => s.meanCost)),\n      meanDurationMs: mean(scenarios.map((s) => s.meanDurationMs)),\n      okRate: mean(scenarios.map((s) => s.okRate)),\n      scenarios,\n      metrics: aggregateMetrics(scenarios.map((s) => s.metrics)),\n    }\n  })\n}\n\nfunction aggregateMetrics(rows: Array<Record<string, number>>): Record<string, number> {\n  const buckets = new Map<string, number[]>()\n  for (const row of rows) {\n    for (const [k, v] of Object.entries(row)) {\n      if (!Number.isFinite(v)) continue\n      const list = buckets.get(k) ?? []\n      list.push(v)\n      buckets.set(k, list)\n    }\n  }\n  const out: Record<string, number> = {}\n  for (const [k, list] of buckets) out[k] = mean(list)\n  return out\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  return xs.reduce((a, b) => a + b, 0) / xs.length\n}\n\nasync function nextPopulation<P>(\n  current: EvolvableVariant<P>[],\n  aggregates: VariantAggregate[],\n  trials: TrialResult[],\n  front: Array<{ candidate: VariantAggregate; distance: number }>,\n  config: PromptEvolutionConfig<P>,\n  nextGeneration: number,\n): Promise<EvolvableVariant<P>[]> {\n  const survivorIds = new Set(front.map((c) => c.candidate.variantId))\n  const survivors = current.filter((v) => survivorIds.has(v.id))\n\n  // Pick the best survivor (by scalar) as the mutation parent.\n  const ranked = scalarScore(aggregates, config.objectives, { weights: config.scalarWeights }).sort(\n    (a, b) => b.score - a.score,\n  )\n  const parentId = ranked[0]?.candidate.variantId ?? current[0]!.id\n  const parent = current.find((v) => v.id === parentId) ?? current[0]!\n  const parentAggregate = aggregates.find((a) => a.variantId === parent.id) ?? aggregates[0]!\n\n  const topTrials = topKTrialsByScore(trials, parent.id, 3)\n  const bottomTrials = bottomKTrialsByScore(trials, parent.id, 3)\n  const childCount = Math.max(0, config.populationSize - survivors.length)\n  let children: EvolvableVariant<P>[] = []\n  if (childCount > 0) {\n    children = await config.mutateAdapter.mutate({\n      parent,\n      parentAggregate,\n      topTrials,\n      bottomTrials,\n      childCount,\n      generation: nextGeneration,\n    })\n    children = children\n      .slice(0, childCount)\n      .map((c) => ({ ...c, generation: nextGeneration, parentId: parent.id }))\n  }\n  return [...survivors, ...children]\n}\n\nfunction topKTrialsByScore(trials: TrialResult[], variantId: string, k: number): TrialResult[] {\n  return trials\n    .filter((t) => t.variantId === variantId && t.ok)\n    .sort((a, b) => b.score - a.score)\n    .slice(0, k)\n}\n\nfunction bottomKTrialsByScore(trials: TrialResult[], variantId: string, k: number): TrialResult[] {\n  return trials\n    .filter((t) => t.variantId === variantId && t.ok)\n    .sort((a, b) => a.score - b.score)\n    .slice(0, k)\n}\n\nfunction samePopulation(a: string[], b: string[]): boolean {\n  if (a.length !== b.length) return false\n  const setA = new Set(a)\n  return b.every((id) => setA.has(id))\n}\n","/**\n * Multi-shot optimization adapter.\n *\n * This is the canonical bridge between variable-length agent trajectories\n * and `runPromptEvolution`. Apps provide four things:\n *\n *   - variants: prompt/config/tool-policy candidates\n *   - runner: executes one full task trajectory for a variant\n *   - scorer: turns that trajectory into score + actionable side information\n *   - mutator: proposes new variants from top/bottom scored trials\n *\n * The adapter owns the boring but easy-to-get-wrong glue: stable seeds,\n * score/cost objectives, error-to-trial conversion, ASI metric projection,\n * and optional paired holdout gating via `HeldOutGate`.\n */\n\nimport { type GateDecision, HeldOutGate, type HeldOutGateConfig } from './held-out-gate'\nimport type { Objective } from './pareto'\nimport {\n  type EvolvableVariant,\n  type PromptEvolutionEvent,\n  type PromptEvolutionResult,\n  runPromptEvolution,\n  type ScoreAdapter,\n  type TrialCache,\n  type TrialResult,\n  type VariantAggregate,\n} from './prompt-evolution'\nimport type { TrialTrace } from './reflective-mutation'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\n\nexport type MultiShotSplit = 'search' | 'dev' | 'holdout'\n\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\nexport type MultiShotVariant<P = unknown> = EvolvableVariant<P>\n\nexport interface ActionableSideInfo {\n  /** Stable expectation/check id when available. */\n  expectationId?: string\n  /** Human-readable diagnosis of what happened. */\n  message: string\n  severity?: AsiSeverity\n  /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n  evidence?: string\n  /** Prompt/tool/context surface likely responsible. */\n  responsibleSurface?: string\n  /** Suggested fix in natural language. */\n  suggestion?: string\n  /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n  matched?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport interface MultiShotTrace {\n  scenarioId: string\n  /** Full turn/tool trace. Shape is intentionally app-owned. */\n  turns?: unknown[]\n  toolCalls?: unknown[]\n  artifacts?: unknown[]\n  /** Compact final output or summary used by reflection prompts. */\n  transcript?: string\n  output?: unknown\n  metadata?: Record<string, unknown>\n}\n\nexport interface MultiShotRun {\n  trace: MultiShotTrace\n  costUsd?: number\n  durationMs?: number\n  tokenUsage?: { input?: number; output?: number; cached?: number }\n  metadata?: Record<string, unknown>\n}\n\nexport interface MultiShotRunInput<P = unknown> {\n  variant: EvolvableVariant<P>\n  scenarioId: string\n  rep: number\n  split: MultiShotSplit\n  /** Stable paired seed for baseline/candidate comparisons. */\n  seed: number\n}\n\nexport interface MultiShotRunner<P = unknown> {\n  run(input: MultiShotRunInput<P>): Promise<MultiShotRun> | MultiShotRun\n}\n\nexport interface MultiShotScore {\n  /** Primary score in [0,1]. The adapter clamps for safety. */\n  score: number\n  /** Pass/fail for top/bottom trial selection. Defaults to true. */\n  ok?: boolean\n  costUsd?: number\n  durationMs?: number\n  metrics?: Record<string, number>\n  asi?: ActionableSideInfo[]\n  /** Optional rich output shown to reflection mutators. */\n  emitted?: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface MultiShotScorer<P = unknown> {\n  score(\n    input: MultiShotRunInput<P> & { run: MultiShotRun },\n  ): Promise<MultiShotScore> | MultiShotScore\n}\n\nexport interface MultiShotTrialResult extends TrialResult {\n  split: MultiShotSplit\n  seed: number\n  trace?: MultiShotTrace\n  asi?: ActionableSideInfo[]\n  emitted?: string\n  metadata?: Record<string, unknown>\n}\n\nexport interface MultiShotMutateAdapter<P = unknown> {\n  mutate(args: {\n    parent: EvolvableVariant<P>\n    parentAggregate: VariantAggregate\n    topTrials: MultiShotTrialResult[]\n    bottomTrials: MultiShotTrialResult[]\n    childCount: number\n    generation: number\n  }): Promise<EvolvableVariant<P>[]>\n}\n\nexport interface MultiShotGateConfig<P = unknown> {\n  /** Search rows are optional, but enable HeldOutGate's overfit-gap check. */\n  searchScenarioIds?: string[]\n  holdoutScenarioIds: string[]\n  reps?: number\n  gate: HeldOutGateConfig\n  /** Convert scored trajectory runs into paper-grade RunRecords. */\n  toRunRecord(input: {\n    variant: EvolvableVariant<P>\n    scenarioId: string\n    rep: number\n    split: RunSplitTag\n    seed: number\n    trial: MultiShotTrialResult\n  }): RunRecord\n}\n\nexport interface MultiShotOptimizationConfig<P = unknown> {\n  runId: string\n  target: string\n  seedVariants: EvolvableVariant<P>[]\n  searchScenarioIds: string[]\n  reps: number\n  generations: number\n  populationSize: number\n  scoreConcurrency?: number\n  runner: MultiShotRunner<P>\n  scorer: MultiShotScorer<P>\n  mutateAdapter: MultiShotMutateAdapter<P>\n  objectives?: Objective<VariantAggregate>[]\n  scalarWeights?: Record<string, number>\n  cache?: TrialCache\n  earlyStopOnNoImprovement?: boolean\n  seedBase?: number\n  onProgress?: (event: PromptEvolutionEvent) => void\n  gate?: MultiShotGateConfig<P>\n}\n\nexport interface MultiShotGateResult {\n  decision: GateDecision\n  candidateRuns: RunRecord[]\n  baselineRuns: RunRecord[]\n}\n\nexport interface MultiShotOptimizationResult<P = unknown> {\n  evolution: PromptEvolutionResult<P>\n  /** Best candidate on the optimizer-visible search split. */\n  searchBestVariant: EvolvableVariant<P>\n  searchBestAggregate: VariantAggregate\n  /** Variant callers should actually ship after optional holdout gating. */\n  promotedVariant: EvolvableVariant<P>\n  promotedAggregate: VariantAggregate\n  /** Null when no gate was configured or the search-best candidate was the baseline. */\n  gate: MultiShotGateResult | null\n}\n\nexport async function runMultiShotOptimization<P>(\n  config: MultiShotOptimizationConfig<P>,\n): Promise<MultiShotOptimizationResult<P>> {\n  validateConfig(config)\n\n  const scoreAdapter: ScoreAdapter<P> = {\n    score: (args) => scoreOne(config, args.variant, args.scenarioId, args.rep, 'search'),\n  }\n\n  const evolution = await runPromptEvolution<P>({\n    runId: config.runId,\n    target: config.target,\n    seedVariants: config.seedVariants,\n    scenarioIds: config.searchScenarioIds,\n    reps: config.reps,\n    generations: config.generations,\n    populationSize: config.populationSize,\n    scoreConcurrency: config.scoreConcurrency ?? 1,\n    scoreAdapter,\n    mutateAdapter: {\n      mutate: (args) =>\n        config.mutateAdapter.mutate({\n          ...args,\n          topTrials: args.topTrials as MultiShotTrialResult[],\n          bottomTrials: args.bottomTrials as MultiShotTrialResult[],\n        }),\n    },\n    objectives: config.objectives ?? defaultMultiShotObjectives(),\n    scalarWeights: config.scalarWeights,\n    earlyStopOnNoImprovement: config.earlyStopOnNoImprovement,\n    cache: config.cache,\n    onProgress: config.onProgress,\n  })\n\n  let gate: MultiShotGateResult | null = null\n  const baseline = config.seedVariants[0]!\n  let promotedVariant = evolution.bestVariant\n  let promotedAggregate = evolution.bestAggregate\n  if (config.gate && evolution.bestVariant.id !== baseline.id) {\n    gate = await evaluateMultiShotGate(config, baseline, evolution.bestVariant)\n    if (!gate.decision.promote) {\n      promotedVariant = baseline\n      promotedAggregate = aggregateFor(evolution, baseline.id)\n    }\n  }\n\n  return {\n    evolution,\n    searchBestVariant: evolution.bestVariant,\n    searchBestAggregate: evolution.bestAggregate,\n    promotedVariant,\n    promotedAggregate,\n    gate,\n  }\n}\n\nexport function defaultMultiShotObjectives(): Objective<VariantAggregate>[] {\n  return [\n    { name: 'score', direction: 'maximize', value: (a) => a.meanScore },\n    { name: 'cost', direction: 'minimize', value: (a) => a.meanCost },\n  ]\n}\n\nexport function trialTraceFromMultiShotTrial(trial: MultiShotTrialResult): TrialTrace {\n  return {\n    id: `${trial.variantId}/${trial.scenarioId}/r${trial.rep}`,\n    score: trial.score,\n    inputName: trial.scenarioId,\n    expectations: (trial.asi ?? []).map((item, i) => ({\n      id: item.expectationId ?? `asi-${i}`,\n      phrase: item.message,\n      matched: item.matched ?? false,\n    })),\n    emitted: trial.emitted ?? traceExcerpt(trial.trace),\n    metrics: trial.metrics,\n  }\n}\n\nasync function evaluateMultiShotGate<P>(\n  config: MultiShotOptimizationConfig<P>,\n  baseline: EvolvableVariant<P>,\n  candidate: EvolvableVariant<P>,\n): Promise<MultiShotGateResult> {\n  const gateConfig = config.gate!\n  const reps = gateConfig.reps ?? config.reps\n  const candidateRuns: RunRecord[] = []\n  const baselineRuns: RunRecord[] = []\n\n  const searchIds = gateConfig.searchScenarioIds ?? config.searchScenarioIds\n  for (const scenarioId of searchIds) {\n    for (let rep = 0; rep < reps; rep++) {\n      const seed = seedFor(config, scenarioId, rep)\n      const baseTrial = await scoreOne(config, baseline, scenarioId, rep, 'search')\n      const candTrial = await scoreOne(config, candidate, scenarioId, rep, 'search')\n      baselineRuns.push(\n        toValidatedRecord(config, baseline, scenarioId, rep, 'search', seed, baseTrial),\n      )\n      candidateRuns.push(\n        toValidatedRecord(config, candidate, scenarioId, rep, 'search', seed, candTrial),\n      )\n    }\n  }\n\n  for (const scenarioId of gateConfig.holdoutScenarioIds) {\n    for (let rep = 0; rep < reps; rep++) {\n      const seed = seedFor(config, scenarioId, rep)\n      const baseTrial = await scoreOne(config, baseline, scenarioId, rep, 'holdout')\n      const candTrial = await scoreOne(config, candidate, scenarioId, rep, 'holdout')\n      baselineRuns.push(\n        toValidatedRecord(config, baseline, scenarioId, rep, 'holdout', seed, baseTrial),\n      )\n      candidateRuns.push(\n        toValidatedRecord(config, candidate, scenarioId, rep, 'holdout', seed, candTrial),\n      )\n    }\n  }\n\n  const decision = new HeldOutGate(gateConfig.gate).evaluate(candidateRuns, baselineRuns)\n  return { decision, candidateRuns, baselineRuns }\n}\n\nasync function scoreOne<P>(\n  config: MultiShotOptimizationConfig<P>,\n  variant: EvolvableVariant<P>,\n  scenarioId: string,\n  rep: number,\n  split: MultiShotSplit,\n): Promise<MultiShotTrialResult> {\n  const seed = seedFor(config, scenarioId, rep)\n  const input: MultiShotRunInput<P> = { variant, scenarioId, rep, split, seed }\n  try {\n    const run = await config.runner.run(input)\n    const scored = await config.scorer.score({ ...input, run })\n    const asi = scored.asi ?? []\n    return {\n      variantId: variant.id,\n      scenarioId,\n      rep,\n      ok: scored.ok ?? true,\n      score: clamp01(scored.score),\n      cost: scored.costUsd ?? run.costUsd ?? 0,\n      durationMs: scored.durationMs ?? run.durationMs ?? 0,\n      metrics: {\n        ...numericMetrics(scored.metrics),\n        ...asiMetrics(asi),\n      },\n      split,\n      seed,\n      trace: run.trace,\n      asi,\n      emitted: scored.emitted ?? traceExcerpt(run.trace),\n      metadata: scored.metadata,\n    }\n  } catch (err) {\n    return {\n      variantId: variant.id,\n      scenarioId,\n      rep,\n      ok: false,\n      score: 0,\n      cost: 0,\n      durationMs: 0,\n      metrics: { error: 1 },\n      error: err instanceof Error ? err.message : String(err),\n      split,\n      seed,\n      asi: [\n        {\n          severity: 'critical',\n          message: err instanceof Error ? err.message : String(err),\n          responsibleSurface: config.target,\n        },\n      ],\n      emitted: '',\n    }\n  }\n}\n\nfunction toValidatedRecord<P>(\n  config: MultiShotOptimizationConfig<P>,\n  variant: EvolvableVariant<P>,\n  scenarioId: string,\n  rep: number,\n  split: RunSplitTag,\n  seed: number,\n  trial: MultiShotTrialResult,\n): RunRecord {\n  const record = config.gate!.toRunRecord({ variant, scenarioId, rep, split, seed, trial })\n  return validateRunRecord(record)\n}\n\nfunction validateConfig<P>(config: MultiShotOptimizationConfig<P>): void {\n  if (!config.runId.trim()) throw new Error('runMultiShotOptimization: runId must not be empty')\n  if (!config.target.trim()) throw new Error('runMultiShotOptimization: target must not be empty')\n  if (config.seedVariants.length === 0) {\n    throw new Error('runMultiShotOptimization: seedVariants must not be empty')\n  }\n  if (config.searchScenarioIds.length === 0) {\n    throw new Error('runMultiShotOptimization: searchScenarioIds must not be empty')\n  }\n  requirePositiveInteger(config.reps, 'reps')\n  requirePositiveInteger(config.generations, 'generations')\n  requirePositiveInteger(config.populationSize, 'populationSize')\n  if (config.scoreConcurrency !== undefined)\n    requirePositiveInteger(config.scoreConcurrency, 'scoreConcurrency')\n  if (config.populationSize < config.seedVariants.length) {\n    throw new Error('runMultiShotOptimization: populationSize must be >= seedVariants.length')\n  }\n  assertUnique(\n    config.seedVariants.map((v) => v.id),\n    'seedVariants.id',\n  )\n  assertUnique(config.searchScenarioIds, 'searchScenarioIds')\n\n  if (config.gate) {\n    if (config.gate.holdoutScenarioIds.length === 0) {\n      throw new Error('runMultiShotOptimization: gate.holdoutScenarioIds must not be empty')\n    }\n    if (config.gate.reps !== undefined) requirePositiveInteger(config.gate.reps, 'gate.reps')\n    assertUnique(config.gate.holdoutScenarioIds, 'gate.holdoutScenarioIds')\n    if (config.gate.searchScenarioIds)\n      assertUnique(config.gate.searchScenarioIds, 'gate.searchScenarioIds')\n    const searchIds = new Set(config.searchScenarioIds)\n    for (const id of config.gate.holdoutScenarioIds) {\n      if (searchIds.has(id)) {\n        throw new Error(\n          `runMultiShotOptimization: holdout scenario \"${id}\" also appears in searchScenarioIds`,\n        )\n      }\n    }\n    const baselineId = config.seedVariants[0]!.id\n    if (config.gate.gate.baselineKey !== baselineId) {\n      throw new Error(\n        `runMultiShotOptimization: gate.gate.baselineKey must match first seed variant id \"${baselineId}\"`,\n      )\n    }\n  }\n}\n\nfunction requirePositiveInteger(value: number, name: string): void {\n  if (!Number.isInteger(value) || value <= 0) {\n    throw new Error(`runMultiShotOptimization: ${name} must be a positive integer`)\n  }\n}\n\nfunction assertUnique(values: string[], name: string): void {\n  const seen = new Set<string>()\n  for (const value of values) {\n    if (!value.trim())\n      throw new Error(`runMultiShotOptimization: ${name} must not contain empty values`)\n    if (seen.has(value)) throw new Error(`runMultiShotOptimization: duplicate ${name} \"${value}\"`)\n    seen.add(value)\n  }\n}\n\nfunction aggregateFor<P>(evolution: PromptEvolutionResult<P>, variantId: string): VariantAggregate {\n  const final = evolution.generations[evolution.generations.length - 1]\n  const aggregate = final?.aggregates.find((a) => a.variantId === variantId)\n  if (!aggregate) {\n    throw new Error(`runMultiShotOptimization: missing aggregate for variant \"${variantId}\"`)\n  }\n  return aggregate\n}\n\nfunction seedFor<P>(\n  config: MultiShotOptimizationConfig<P>,\n  scenarioId: string,\n  rep: number,\n): number {\n  const base = config.seedBase ?? 0\n  return (base + stableHash(`${scenarioId}\\x1f${rep}`)) % Number.MAX_SAFE_INTEGER\n}\n\nfunction stableHash(input: string): number {\n  let h = 2166136261\n  for (let i = 0; i < input.length; i++) {\n    h ^= input.charCodeAt(i)\n    h = Math.imul(h, 16777619)\n  }\n  return h >>> 0\n}\n\nfunction clamp01(n: number): number {\n  if (!Number.isFinite(n)) return 0\n  return Math.max(0, Math.min(1, n))\n}\n\nfunction numericMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const [k, v] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(v)) out[k] = v\n  }\n  return out\n}\n\nfunction asiMetrics(asi: ActionableSideInfo[]): Record<string, number> {\n  const out: Record<string, number> = { asi: asi.length }\n  for (const item of asi.slice(0, 1000)) {\n    const sev = normalizeSeverity(item.severity)\n    out[`asi.${sev}`] = (out[`asi.${sev}`] ?? 0) + 1\n    if (item.responsibleSurface) {\n      const key = `surface.${metricKeySegment(item.responsibleSurface)}`\n      out[key] = (out[key] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction normalizeSeverity(severity: AsiSeverity | undefined): AsiSeverity {\n  if (\n    severity === 'info' ||\n    severity === 'warning' ||\n    severity === 'error' ||\n    severity === 'critical'\n  ) {\n    return severity\n  }\n  return 'error'\n}\n\nfunction metricKeySegment(raw: string): string {\n  return (\n    raw\n      .trim()\n      .replace(/[^a-zA-Z0-9._-]+/g, '_')\n      .slice(0, 80) || 'unknown'\n  )\n}\n\nfunction traceExcerpt(trace: MultiShotTrace | undefined): string | undefined {\n  if (!trace) return undefined\n  if (typeof trace.output === 'string') return trace.output\n  if (trace.transcript) return trace.transcript\n  if (trace.turns) {\n    try {\n      const clipped = trace.turns.slice(0, 20)\n      const suffix =\n        trace.turns.length > clipped.length\n          ? ` ... ${trace.turns.length - clipped.length} more turn(s)`\n          : ''\n      return `${JSON.stringify(clipped).slice(0, 2000)}${suffix}`\n    } catch {\n      return '[unserializable trace turns]'\n    }\n  }\n  return undefined\n}\n","/**\n * Reflective mutation — primitives for trace-conditioned prompt rewriting.\n *\n * Used by `prompt-evolution.ts` (and any consumer running iterative\n * improvement). Given a parent prompt + concrete trace evidence (top trials,\n * bottom trials, missed expectations), produce an LLM-ready prompt that\n * proposes targeted mutations — not blind rephrasings.\n *\n * Why this lives outside `prompt-evolution.ts`: any consumer that wants to\n * run reflective rewriting WITHOUT the population/Pareto machinery can\n * import these primitives directly.\n *\n * Quality bar (vs. naive \"mutate this prompt\"):\n *   - Show parent ↔ children diff, not just one variant\n *   - Quote specific missed goldens with their match phrases\n *   - Surface the model's actual emitted output side-by-side with what was expected\n *   - Quote concrete mutation primitives so the model has a vocabulary\n */\n\nexport interface TrialTrace {\n  /** Stable id for the trial — surfaces in the prompt for grounding. */\n  id: string\n  /** Score the trial received on its primary metric. */\n  score: number\n  /** Candidate inputs the agent was given (e.g., the fixture or scenario). */\n  inputName?: string\n  /**\n   * Goldens / expectations this trial was tested against, with whether each\n   * was matched. The reflection prompt quotes the missed ones specifically.\n   */\n  expectations?: Array<{ id: string; phrase: string; matched: boolean }>\n  /** Free-form text — what the agent actually emitted (e.g., findings, plan). */\n  emitted?: string\n  /** Optional structured metrics (recall, precision, cost, latency). */\n  metrics?: Record<string, number>\n}\n\nexport interface ReflectionContext {\n  /** What is being mutated — appears in the system prompt for orientation. */\n  target: string\n  /** Current variant's payload — JSON-serialised for the prompt. */\n  parentPayload: unknown\n  /** Best-performing trials this generation. */\n  topTrials: TrialTrace[]\n  /** Worst-performing trials this generation — the missed-golden source. */\n  bottomTrials: TrialTrace[]\n  /** How many children the mutator should propose. */\n  childCount: number\n  /** Optional: domain-specific mutation primitives the model can pick from. */\n  mutationPrimitives?: string[]\n}\n\nexport const DEFAULT_MUTATION_PRIMITIVES: string[] = [\n  'Strengthen an imperative (\"should\" → \"must\")',\n  'Add a concrete example pulled from a missed-golden phrase',\n  'Remove a redundant rule that did not improve recall',\n  'Add a counterfactual (\"if X is missing, the score is capped at Y\")',\n  'Reorder sections so the highest-impact rule is first',\n  'Replace abstract language with a domain-specific noun the trial misses',\n]\n\n/**\n * Build the LLM-ready reflection prompt. Output is plain text — pass it as\n * the user message. The system message should be small and stable (e.g.\n * \"Output ONLY a JSON object matching the schema below.\").\n */\nexport function buildReflectionPrompt(ctx: ReflectionContext): string {\n  const primitives = ctx.mutationPrimitives ?? DEFAULT_MUTATION_PRIMITIVES\n  const sections: string[] = []\n\n  sections.push(`# Mutation target: ${ctx.target}`)\n  sections.push('')\n  sections.push(\n    `You are tuning the prompt component named \\`${ctx.target}\\`. The current variant is shown below; you have ${ctx.topTrials.length} top trials and ${ctx.bottomTrials.length} bottom trials as evidence. Propose ${ctx.childCount} mutation${ctx.childCount === 1 ? '' : 's'} that fix specific weaknesses visible in the bottom trials. Avoid blank rephrasings.`,\n  )\n  sections.push('')\n\n  sections.push('## Current variant')\n  sections.push('```json')\n  sections.push(JSON.stringify(ctx.parentPayload, null, 2))\n  sections.push('```')\n  sections.push('')\n\n  if (ctx.bottomTrials.length > 0) {\n    sections.push('## Failures (bottom trials) — what went wrong')\n    sections.push('')\n    for (const trial of ctx.bottomTrials) {\n      sections.push(\n        `### Trial \\`${trial.id}\\` — score ${trial.score.toFixed(2)}${trial.inputName ? ` (${trial.inputName})` : ''}`,\n      )\n      const missed = (trial.expectations ?? []).filter((e) => !e.matched)\n      if (missed.length > 0) {\n        sections.push('')\n        sections.push('**Missed expectations:**')\n        for (const m of missed) {\n          sections.push(`- \\`${m.id}\\`: should match phrase \\`${quote(m.phrase)}\\``)\n        }\n      }\n      if (trial.emitted) {\n        sections.push('')\n        sections.push('**What the agent emitted:**')\n        sections.push('```')\n        sections.push(truncate(trial.emitted, 600))\n        sections.push('```')\n      }\n      sections.push('')\n    }\n  }\n\n  if (ctx.topTrials.length > 0) {\n    sections.push('## Successes (top trials) — what to preserve')\n    sections.push('')\n    for (const trial of ctx.topTrials) {\n      sections.push(\n        `- \\`${trial.id}\\`: score ${trial.score.toFixed(2)}${trial.inputName ? ` (${trial.inputName})` : ''}`,\n      )\n    }\n    sections.push('')\n  }\n\n  sections.push('## Allowed mutation primitives')\n  sections.push('')\n  for (const p of primitives) sections.push(`- ${p}`)\n  sections.push('')\n\n  sections.push('## Output schema')\n  sections.push('')\n  sections.push('Respond with a JSON object — no prose, no markdown fences:')\n  sections.push('```json')\n  sections.push(\n    JSON.stringify(\n      {\n        proposals: [\n          {\n            label: '<short label, ≤ 40 chars>',\n            rationale: '<which failure this targets and which primitive you used>',\n            payload: '<full payload of the new variant — same shape as the current variant>',\n          },\n        ],\n      },\n      null,\n      2,\n    ),\n  )\n  sections.push('```')\n\n  return sections.join('\\n')\n}\n\nfunction truncate(s: string, max: number): string {\n  if (s.length <= max) return s\n  return `${s.slice(0, max)}… [truncated]`\n}\n\nfunction quote(s: string): string {\n  return s.replace(/`/g, '\\\\`')\n}\n\nexport interface ReflectionProposal {\n  label: string\n  rationale: string\n  payload: unknown\n}\n\n/**\n * Parse the model's JSON response back into proposals. Tolerates markdown\n * fences and surrounding prose. Returns at most `maxProposals`.\n */\n/**\n * Walk the input as JSON-aware (string vs not, escape-aware) and close\n * unclosed `{` / `[` in LIFO order at the tail. If the input was already\n * balanced returns it unchanged. If a string was open at end-of-input we\n * also close it with `\"` first, since a truncated string-mid-value is the\n * most common LLM cap-hit failure mode and JSON.parse cannot proceed\n * without one.\n *\n * Returns null when the structure is unrecoverable (e.g. depth would go\n * negative — that's an *over*-closed prefix, not a truncation).\n */\nfunction autoCloseTruncatedJson(raw: string): string | null {\n  const stack: Array<'{' | '['> = []\n  let inString = false\n  let escaped = false\n  for (const c of raw) {\n    if (escaped) {\n      escaped = false\n      continue\n    }\n    if (inString) {\n      if (c === '\\\\') {\n        escaped = true\n        continue\n      }\n      if (c === '\"') {\n        inString = false\n        continue\n      }\n      continue\n    }\n    if (c === '\"') {\n      inString = true\n      continue\n    }\n    if (c === '{' || c === '[') stack.push(c)\n    else if (c === '}') {\n      if (stack.pop() !== '{') return null\n    } else if (c === ']') {\n      if (stack.pop() !== '[') return null\n    }\n  }\n  if (stack.length === 0 && !inString) return raw\n  let suffix = ''\n  if (inString) suffix += '\"'\n  while (stack.length > 0) {\n    const opener = stack.pop()!\n    suffix += opener === '{' ? '}' : ']'\n  }\n  return raw + suffix\n}\n\nexport function parseReflectionResponse(raw: string, maxProposals?: number): ReflectionProposal[] {\n  let text = raw.trim()\n  if (text.startsWith('```')) text = text.replace(/^```(?:json)?\\n?/, '').replace(/\\n?```$/, '')\n\n  // Try to parse as either a JSON object `{proposals: [...]}` or a bare\n  // array `[...]`. LLMs frequently emit one or the other depending on how\n  // they read the schema example; accept both.\n  let parsed: unknown = null\n  const objectStart = text.indexOf('{')\n  const objectEnd = text.lastIndexOf('}')\n  const arrayStart = text.indexOf('[')\n  const arrayEnd = text.lastIndexOf(']')\n  // Prefer whichever delimiter comes first (the model committed to that shape).\n  const tryObjectFirst = objectStart >= 0 && (arrayStart < 0 || objectStart < arrayStart)\n  const candidates: string[] = []\n  if (tryObjectFirst) {\n    if (objectStart >= 0 && objectEnd > objectStart)\n      candidates.push(text.slice(objectStart, objectEnd + 1))\n    if (arrayStart >= 0 && arrayEnd > arrayStart)\n      candidates.push(text.slice(arrayStart, arrayEnd + 1))\n  } else {\n    if (arrayStart >= 0 && arrayEnd > arrayStart)\n      candidates.push(text.slice(arrayStart, arrayEnd + 1))\n    if (objectStart >= 0 && objectEnd > objectStart)\n      candidates.push(text.slice(objectStart, objectEnd + 1))\n  }\n  for (const slice of candidates) {\n    try {\n      parsed = JSON.parse(slice)\n      break\n    } catch {\n      // try next\n    }\n  }\n\n  // Truncation-tolerant fallback: LLMs frequently hit a max_tokens cap\n  // mid-emission, leaving N unclosed `}` / `]` at the tail. Close them in\n  // order from the deepest unclosed structure outward, by walking the\n  // candidate slice and tracking depth, then retrying JSON.parse. This\n  // recovers any complete proposals before the cutoff and drops the rest.\n  if (parsed == null) {\n    for (const slice of candidates) {\n      const closed = autoCloseTruncatedJson(slice)\n      if (closed != null && closed !== slice) {\n        try {\n          parsed = JSON.parse(closed)\n          break\n        } catch {\n          // give up on this candidate\n        }\n      }\n    }\n  }\n\n  if (parsed == null) return []\n\n  // Normalize: accept `{proposals: [...]}` or a bare array.\n  let proposalsRaw: unknown\n  if (Array.isArray(parsed)) {\n    proposalsRaw = parsed\n  } else if (parsed && typeof parsed === 'object') {\n    proposalsRaw = (parsed as { proposals?: unknown }).proposals\n  }\n  if (!Array.isArray(proposalsRaw)) return []\n\n  const out: ReflectionProposal[] = []\n  for (const p of proposalsRaw) {\n    if (!p || typeof p !== 'object') continue\n    const obj = p as { label?: unknown; rationale?: unknown; payload?: unknown }\n    if (!('payload' in obj)) continue\n    out.push({\n      label: typeof obj.label === 'string' ? obj.label : 'mutation',\n      rationale: typeof obj.rationale === 'string' ? obj.rationale : '',\n      payload: obj.payload,\n    })\n    if (maxProposals !== undefined && out.length >= maxProposals) break\n  }\n  return out\n}\n","/**\n * Researcher interface — stable hook for an external autonomous-research\n * agent to drive the meta-loop.\n *\n * Implementations live downstream (typically in a private repo that\n * runs the actual LLM). This package ships only the contract + a\n * `NoopResearcher` so consumers can wire the surface without being\n * forced to implement every method up front.\n *\n * The four methods mirror the four stages of the paper \"Two Loops,\n * Three Roles\":\n *\n *   inspectFailures   — given the observed runs, what failure modes\n *                       are present? (data → diagnosis)\n *   proposeChange     — given diagnosed failure modes, what\n *                       structural changes should we try?\n *                       (diagnosis → plan delta)\n *   applyChange       — fold the proposed deltas into a concrete\n *                       experiment plan against an existing baseline.\n *                       (plan delta → executable plan)\n *   evaluateChange    — run the plan, return runs + the gate verdict.\n *                       (executable plan → verdict)\n *\n * Composition is the discipline: a Researcher implementation MUST\n * keep these four steps separate and inspectable. Conflating\n * \"diagnose + propose + run\" into a single LLM call defeats the\n * point of the framework — you can't audit which step lied.\n *\n * THIS INTERFACE IS STABLE. Breaking changes require a new module\n * (e.g. `Researcher2`) so existing implementations keep working.\n */\n\nimport type { GateDecision } from './held-out-gate'\nimport type { RunRecord, RunSplitTag } from './run-record'\n\n/** A diagnosed failure mode with the run-IDs that exhibit it. */\nexport interface FailureMode {\n  /** Short machine-readable code. Must be stable across runs of the\n   *  same researcher to enable longitudinal tracking. */\n  code: string\n  /** Human-readable description for the paper / dashboard. */\n  description: string\n  evidence: {\n    /** Run IDs (from `RunRecord.runId`) where this failure mode was\n     *  observed. */\n    runIds: string[]\n    /** Number of run samples that informed the diagnosis. */\n    samples: number\n  }\n}\n\n/** A single steering change the researcher wants to try. */\nexport interface SteeringChange {\n  kind: 'reviewer_prompt' | 'skill_add' | 'skill_remove' | 'threshold' | 'budget'\n  /** Implementation-specific payload. Researcher implementations\n   *  define the schema — keep this `unknown` here to avoid coupling\n   *  the public interface to any one researcher's internal model. */\n  payload: unknown\n  /** Why the researcher proposed this change. Goes into the audit\n   *  trail next to the failure-mode evidence. */\n  rationale: string\n  /** Optional self-reported expected delta on the headline metric. */\n  expectedDelta?: number\n}\n\n/** A single experiment plan, mapped onto the search/holdout splits. */\nexport interface ExperimentPlan {\n  baselineCandidateId: string\n  proposedCandidateId: string\n  changes: SteeringChange[]\n  /** USD ceiling for the entire experiment. The runner must stop\n   *  before exceeding this and report a partial result. */\n  evaluationBudgetUsd: number\n  /** Item IDs (your dataset keys) for the search vs holdout splits. */\n  splits: { search: string[]; holdout: string[] }\n}\n\n/** Result of running a plan: every run, plus the gate verdict. */\nexport interface ExperimentResult {\n  plan: ExperimentPlan\n  runs: RunRecord[]\n  gateDecision: GateDecision\n}\n\n/**\n * The researcher loop. Stable, four-step, inspectable.\n *\n *   ┌──────────┐  inspectFailures  ┌──────────┐  proposeChange ┌──────────┐\n *   │   runs   │ ─────────────────▶│ failures │ ──────────────▶│ changes  │\n *   └──────────┘                   └──────────┘                └────┬─────┘\n *                                                                   │\n *                                                                   ▼\n *                              ┌────────────────┐  applyChange ┌────────┐\n *                              │ ExperimentPlan │ ◀────────────│  base  │\n *                              └────────┬───────┘              └────────┘\n *                                       │\n *                       evaluateChange  ▼\n *                              ┌────────────────┐\n *                              │ ExperimentResult│\n *                              └────────────────┘\n */\nexport interface Researcher {\n  inspectFailures(runs: RunRecord[]): Promise<FailureMode[]>\n  proposeChange(failures: FailureMode[]): Promise<SteeringChange[]>\n  applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan>\n  evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult>\n}\n\nexport interface CallbackResearcherOptions {\n  inspectFailures: Researcher['inspectFailures']\n  proposeChange: Researcher['proposeChange']\n  applyChange: Researcher['applyChange']\n  evaluateChange: Researcher['evaluateChange']\n}\n\n/**\n * Minimal concrete researcher for tests, scripts, and small integrations.\n * Larger autonomous researchers can still implement `Researcher` directly.\n */\nexport class CallbackResearcher implements Researcher {\n  constructor(private readonly callbacks: CallbackResearcherOptions) {}\n\n  inspectFailures(runs: RunRecord[]): Promise<FailureMode[]> {\n    return this.callbacks.inspectFailures(runs)\n  }\n\n  proposeChange(failures: FailureMode[]): Promise<SteeringChange[]> {\n    return this.callbacks.proposeChange(failures)\n  }\n\n  applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan> {\n    return this.callbacks.applyChange(changes, baseline)\n  }\n\n  evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult> {\n    return this.callbacks.evaluateChange(plan)\n  }\n}\n\n/**\n * No-op researcher — fails loud on every method. Use as a placeholder\n * in code paths that wire the interface but don't have an implementation\n * yet. Importantly, this does NOT silently succeed: a no-op researcher\n * that returned empty arrays would muffle the loop's signal that\n * nobody implemented the brain.\n */\nexport class NoopResearcher implements Researcher {\n  private readonly hint: string\n\n  constructor(hint = 'NoopResearcher: no implementation wired') {\n    this.hint = hint\n  }\n\n  async inspectFailures(_runs: RunRecord[]): Promise<FailureMode[]> {\n    throw new Error(`${this.hint} (inspectFailures not implemented)`)\n  }\n\n  async proposeChange(_failures: FailureMode[]): Promise<SteeringChange[]> {\n    throw new Error(`${this.hint} (proposeChange not implemented)`)\n  }\n\n  async applyChange(\n    _changes: SteeringChange[],\n    _baseline: ExperimentPlan,\n  ): Promise<ExperimentPlan> {\n    throw new Error(`${this.hint} (applyChange not implemented)`)\n  }\n\n  async evaluateChange(_plan: ExperimentPlan): Promise<ExperimentResult> {\n    throw new Error(`${this.hint} (evaluateChange not implemented)`)\n  }\n}\n\n/** Re-export the split alias so callers don't have to import twice. */\nexport type { RunSplitTag }\n"],"mappings":";;;;;;;;;AAqJA,IAAM,uBAAsD;AAAA,EAC1D,UAAU;AAAA,EACV,QAAQ;AAAA,EACR,SAAS;AAAA,EACT,YAAY;AACd;AAEO,IAAM,kCAAN,MAAyE;AAAA,EAC7D,eAAe,oBAAI,IAAgC;AAAA,EAEpE,MAAM,KAAK,YAA+C;AACxD,SAAK,aAAa,IAAI,WAAW,IAAI,gBAAgB,UAAU,CAAC;AAAA,EAClE;AAAA,EAEA,MAAM,IAAI,IAAgD;AACxD,UAAM,aAAa,KAAK,aAAa,IAAI,EAAE;AAC3C,WAAO,aAAa,gBAAgB,UAAU,IAAI;AAAA,EACpD;AAAA,EAEA,MAAM,KAAK,SAAmC,CAAC,GAAkC;AAC/E,WAAO,CAAC,GAAG,KAAK,aAAa,OAAO,CAAC,EAClC,OAAO,CAAC,eAAe,cAAc,YAAY,MAAM,CAAC,EACxD,IAAI,eAAe;AAAA,EACxB;AAAA,EAEA,MAAM,cAAc,IAAY,SAAuD;AACrF,UAAM,aAAa,KAAK,aAAa,IAAI,EAAE;AAC3C,QAAI,CAAC;AACH,YAAM,IAAI,MAAM,8DAA8D,EAAE,GAAG;AACrF,UAAM,OAAO,gBAAgB;AAAA,MAC3B,GAAG;AAAA,MACH,UAAU,CAAC,GAAG,WAAW,UAAU,OAAO;AAAA,MAC1C,WAAW,QAAQ;AAAA,IACrB,CAAC;AACD,SAAK,aAAa,IAAI,IAAI,IAAI;AAC9B,WAAO,gBAAgB,IAAI;AAAA,EAC7B;AAAA,EAEA,MAAM,YACJ,IACA,OACA,WAC6B;AAC7B,UAAM,aAAa,KAAK,aAAa,IAAI,EAAE;AAC3C,QAAI,CAAC;AACH,YAAM,IAAI,MAAM,4DAA4D,EAAE,GAAG;AACnF,UAAM,WAAW,YACb,WAAW,SAAS;AAAA,MAAI,CAAC,YACvB,QAAQ,OAAO,YACX,EAAE,GAAG,SAAS,UAAU,CAAC,GAAI,QAAQ,YAAY,CAAC,GAAI,KAAK,EAAE,IAC7D;AAAA,IACN,IACA,WAAW;AACf,UAAM,OAAO,gBAAgB;AAAA,MAC3B,GAAG;AAAA,MACH;AAAA,MACA,QAAQ,YAAY,WAAW,SAAS,CAAC,GAAG,WAAW,QAAQ,KAAK;AAAA,MACpE,WAAW,MAAM;AAAA,IACnB,CAAC;AACD,SAAK,aAAa,IAAI,IAAI,IAAI;AAC9B,WAAO,gBAAgB,IAAI;AAAA,EAC7B;AACF;AAEO,IAAM,oCAAN,MAA2E;AAAA,EAC/D;AAAA,EACA,SAAS,IAAI,gCAAgC;AAAA,EACtD,SAAS;AAAA,EAEjB,YAAY,SAA0B;AACpC,SAAK,MAAM,QAAQ;AAAA,EACrB;AAAA,EAEA,MAAM,KAAK,YAA+C;AACxD,UAAM,KAAK,KAAK;AAChB,UAAM,KAAK,OAAO,KAAK,UAAU;AACjC,UAAM,KAAK,OAAO,EAAE,IAAI,QAAQ,WAAW,CAAC;AAAA,EAC9C;AAAA,EAEA,MAAM,IAAI,IAAgD;AACxD,UAAM,KAAK,KAAK;AAChB,WAAO,KAAK,OAAO,IAAI,EAAE;AAAA,EAC3B;AAAA,EAEA,MAAM,KAAK,SAAmC,CAAC,GAAkC;AAC/E,UAAM,KAAK,KAAK;AAChB,WAAO,KAAK,OAAO,KAAK,MAAM;AAAA,EAChC;AAAA,EAEA,MAAM,cAAc,IAAY,SAAuD;AACrF,UAAM,KAAK,KAAK;AAChB,UAAM,OAAO,MAAM,KAAK,OAAO,cAAc,IAAI,OAAO;AACxD,UAAM,KAAK,OAAO,EAAE,IAAI,iBAAiB,IAAI,QAAQ,CAAC;AACtD,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,YACJ,IACA,OACA,WAC6B;AAC7B,UAAM,KAAK,KAAK;AAChB,UAAM,OAAO,MAAM,KAAK,OAAO,YAAY,IAAI,OAAO,SAAS;AAC/D,UAAM,KAAK,OAAO,EAAE,IAAI,eAAe,IAAI,OAAO,UAAU,CAAC;AAC7D,WAAO;AAAA,EACT;AAAA,EAEA,MAAc,OAAO,QAAgC;AACnD,UAAM,EAAE,YAAY,MAAM,IAAI,MAAM,OAAO,aAAkB;AAC7D,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,MAAW;AACzC,UAAM,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;AACzC,UAAM;AAAA,MACJ,KAAK,KAAK,KAAK,8BAA8B;AAAA,MAC7C,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAc,OAAsB;AAClC,QAAI,KAAK,OAAQ;AACjB,UAAM,EAAE,SAAS,IAAI,MAAM,OAAO,aAAkB;AACpD,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,MAAW;AACzC,UAAM,OAAO,KAAK,KAAK,KAAK,8BAA8B;AAC1D,QAAI;AACF,YAAM,MAAM,MAAM,SAAS,MAAM,MAAM;AACvC,iBAAW,QAAQ,IAAI,MAAM,IAAI,GAAG;AAClC,YAAI,CAAC,KAAK,KAAK,EAAG;AAClB,YAAI;AACF,gBAAM,SAAS,KAAK,MAAM,IAAI;AAI9B,cAAI,OAAO,OAAO,OAAQ,OAAM,KAAK,OAAO,KAAK,OAAO,UAAU;AAClE,cAAI,OAAO,OAAO;AAChB,kBAAM,KAAK,OAAO,cAAc,OAAO,IAAI,OAAO,OAAO;AAC3D,cAAI,OAAO,OAAO;AAChB,kBAAM,KAAK,OAAO,YAAY,OAAO,IAAI,OAAO,OAAO,OAAO,SAAS;AAAA,QAC3E,QAAQ;AAAA,QAER;AAAA,MACF;AAAA,IACF,QAAQ;AAAA,IAER;AACA,SAAK,SAAS;AAAA,EAChB;AACF;AAEO,SAAS,yBAAyB,OAYlB;AACrB,QAAM,YAAY,MAAM,cAAa,oBAAI,KAAK,GAAE,YAAY;AAC5D,QAAM,KACJ,MAAM,MACN,MAAM,WAAW,GAAG,MAAM,aAAa,EAAE,IAAI,MAAM,cAAc,EAAE,IAAI,MAAM,KAAK,MAAM,IAAI,SAAS,EAAE,EAAE,SAAS,EAAE,CAAC;AACvH,SAAO;AAAA,IACL;AAAA,IACA,WAAW,MAAM;AAAA,IACjB,YAAY,MAAM;AAAA,IAClB,MAAM,MAAM;AAAA,IACZ,UAAU,MAAM,YAAY,CAAC;AAAA,IAC7B,QAAQ,MAAM,UAAU,CAAC;AAAA,IACzB,SAAS,MAAM;AAAA,IACf,OAAO,MAAM;AAAA,IACb,MAAM,MAAM;AAAA,IACZ;AAAA,IACA,UAAU,MAAM;AAAA,EAClB;AACF;AAEO,SAAS,oBACd,YACA,SAA8B,CAAC,GACjB;AACd,QAAM,QAAQ,EAAE,GAAG,sBAAsB,GAAG,OAAO;AACnD,QAAM,QAAQ,MAAM,WAAW,MAAM,SAAS,MAAM,UAAU,MAAM;AACpE,MAAI,SAAS,EAAG,OAAM,IAAI,MAAM,4DAA4D;AAC5F,QAAM,SACJ;AAAA,IACE,GAAG,WAAW,aAAa,EAAE,IAAI,WAAW,cAAc,EAAE,IAAI,WAAW,EAAE,IAAI,WAAW,KAAK,MAAM;AAAA,EACzG,IAAI;AACN,MAAI,SAAS,MAAM,SAAU,QAAO;AACpC,MAAI,SAAS,MAAM,WAAW,MAAM,OAAQ,QAAO;AACnD,MAAI,SAAS,MAAM,WAAW,MAAM,SAAS,MAAM,QAAS,QAAO;AACnE,SAAO;AACT;AAEO,SAAS,0BACd,YACA,QACoB;AACpB,SAAO;AAAA,IACL,GAAG;AAAA,IACH,OAAO,WAAW,SAAS,oBAAoB,YAAY,MAAM;AAAA,EACnE;AACF;AAEO,SAAS,oCACd,YACiB;AACjB,QAAM,YAAY,0BAA0B,UAAU;AACtD,SAAO;AAAA,IACL,IAAI,UAAU,cAAc,UAAU;AAAA,IACtC,OAAO,UAAU;AAAA,IACjB,SAAS;AAAA,IACT,MAAM;AAAA,MACJ,GAAI,UAAU,YAAY,EAAE,WAAW,UAAU,UAAU,IAAI,CAAC;AAAA,MAChE,GAAI,UAAU,QAAQ,CAAC;AAAA,MACvB,QAAQ;AAAA,IACV;AAAA,EACF;AACF;AAEO,SAAS,uCACd,cACmB;AACnB,SAAO,aAAa,IAAI,mCAAmC;AAC7D;AAEO,SAAS,iCACd,YACsB;AACtB,QAAM,SAAS,UAAU,UAAU;AACnC,SAAO;AAAA,IACL,YAAY,WAAW,cAAc,WAAW;AAAA,IAChD,cAAc,WAAW;AAAA,IACzB,YAAY,CAAC,GAAG,IAAI,IAAI,OAAO,IAAI,CAAC,UAAU,MAAM,IAAI,CAAC,CAAC;AAAA,IAC1D,OAAO,WAAW,SAAS,SAAS,gBAAgB,MAAM;AAAA,IAC1D,UAAU;AAAA,MACR,WAAW,WAAW;AAAA,MACtB,OAAO,WAAW;AAAA,MAClB,QAAQ,WAAW,KAAK;AAAA,MACxB,UAAU,WAAW,SAAS;AAAA,MAC9B,SAAS,WAAW;AAAA,MACpB;AAAA,IACF;AAAA,EACF;AACF;AAEO,SAAS,oCACd,cACwB;AACxB,SAAO,aAAa,IAAI,gCAAgC;AAC1D;AAEA,eAAsB,yBACpB,YACA,SAC+B;AAC/B,MAAI;AACF,UAAM,SAAS,MAAM,QAAQ,OAAO,UAAU;AAC9C,WAAO;AAAA,MACL,cAAc,WAAW;AAAA,MACzB,GAAG;AAAA,IACL;AAAA,EACF,SAAS,KAAK;AACZ,UAAM,aAAY,oBAAI,KAAK,GAAE,YAAY;AACzC,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,WAAO;AAAA,MACL,cAAc,WAAW;AAAA,MACzB,MAAM;AAAA,MACN,QAAQ;AAAA,QACN;AAAA,UACE,QAAQ;AAAA,UACR,MAAM;AAAA,UACN,OAAO;AAAA,UACP,QAAQ;AAAA,UACR,UAAU;AAAA,UACV;AAAA,QACF;AAAA,MACF;AAAA,MACA,SAAS;AAAA,QACP,SAAS;AAAA,QACT,OAAO;AAAA,QACP,QAAQ;AAAA,QACR,YAAY;AAAA,MACd;AAAA,MACA,UAAU,EAAE,aAAa,KAAK;AAAA,IAChC;AAAA,EACF;AACF;AAEA,eAAsB,2BACpB,cACA,SACiC;AACjC,QAAM,UAAkC,CAAC;AACzC,aAAW,cAAc,cAAc;AACrC,YAAQ,KAAK,MAAM,yBAAyB,YAAY,OAAO,CAAC;AAAA,EAClE;AACA,SAAO;AACT;AAEO,SAAS,0BACd,cACA,UAAmC,CAAC,GACX;AACzB,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,UAAmC,CAAC;AAC1C,aAAW,cAAc,cAAc;AACrC,eAAW,SAAS,UAAU,UAAU,GAAG;AACzC,YAAM,cAAc,qBAAqB,YAAY,KAAK;AAC1D,UAAI,CAAC,YAAa;AAClB,cAAQ,KAAK;AAAA,QACX;AAAA,QACA,WAAW,MAAM,UAAU,GAAG,MAAM,IAAI,eAAe,MAAM,MAAM;AAAA,QACnE,QAAQ,eAAe,KAAK;AAAA,QAC5B,oBAAoB,WAAW;AAAA,QAC/B,eAAe,MAAM;AAAA,QACrB,UAAU,MAAM;AAAA,MAClB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,gBAAgB,oBAAI,IAAmC;AAC7D,aAAW,SAAS,SAAS;AAC3B,UAAM,MAAM,MAAM,YAAY,YAAY,EAAE,QAAQ,QAAQ,GAAG,EAAE,KAAK;AACtE,UAAM,WAAW,cAAc,IAAI,GAAG;AACtC,QAAI,CAAC,YAAY,MAAM,SAAS,SAAS,OAAQ,eAAc,IAAI,KAAK,KAAK;AAAA,EAC/E;AACA,SAAO,CAAC,GAAG,cAAc,OAAO,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,MAAM,GAAG,UAAU;AAC5F;AAEO,SAAS,+BAA+B,SAA0C;AACvF,QAAM,QAAQ,CAAC,uBAAuB,EAAE;AACxC,aAAW,SAAS,SAAS;AAC3B,UAAM,KAAK,KAAK,MAAM,WAAW,EAAE;AACnC,UAAM,KAAK,gBAAgB,MAAM,SAAS,EAAE;AAC5C,UAAM,KAAK,aAAa,MAAM,kBAAkB,EAAE;AAClD,UAAM,KAAK,EAAE;AAAA,EACf;AACA,SAAO,GAAG,MAAM,KAAK,IAAI,EAAE,KAAK,CAAC;AAAA;AACnC;AAEO,SAAS,mCAAmC,cAA4C;AAC7F,SAAO,GAAG,aACP,MAAM,EACN,KAAK,CAAC,GAAG,MAAM,EAAE,GAAG,cAAc,EAAE,EAAE,CAAC,EACvC,IAAI,CAAC,eAAe,KAAK,UAAU,aAAa,UAAU,CAAC,CAAC,EAC5D,KAAK,IAAI,CAAC;AAAA;AACf;AAEO,SAAS,+BAA+B,OAAqC;AAClF,QAAM,eAAqC,CAAC;AAC5C,aAAW,QAAQ,MAAM,MAAM,IAAI,GAAG;AACpC,QAAI,CAAC,KAAK,KAAK,EAAG;AAClB,iBAAa,KAAK,KAAK,MAAM,IAAI,CAAuB;AAAA,EAC1D;AACA,SAAO;AACT;AAEO,SAAS,+BACd,KACA,UASI,CAAC,GACe;AACpB,QAAM,YAAY,QAAQ,cAAa,oBAAI,KAAK,GAAE,YAAY;AAC9D,QAAM,eACJ,IAAI,SAAS,cAAc,WAAW,GAAG,IAAI,MAAM,IAAI,SAAS,EAAE,EAAE,SAAS,EAAE,CAAC;AAClF,SAAO,yBAAyB;AAAA,IAC9B,IAAI;AAAA,IACJ,WAAW,QAAQ;AAAA,IACnB,YAAY,QAAQ;AAAA,IACpB,MAAM,EAAE,QAAQ,IAAI,OAAO;AAAA,IAC3B;AAAA,IACA,UAAU,IAAI,MAAM,IAAI,CAAC,UAAU;AAAA,MACjC,IAAI,GAAG,YAAY,SAAS,KAAK,KAAK;AAAA,MACtC,WAAW,KAAK;AAAA,MAChB,cAAc,QAAQ,gBAAgB;AAAA,MACtC,UAAU,QAAQ,mBAAmB,IAAI,KAAK,KAAK,eAAe,UAAU,KAAK;AAAA,MACjF,gBAAgB,QAAQ,yBAAyB,IAAI;AAAA,MACrD,OAAO,KAAK;AAAA,MACZ,WAAW,KAAK;AAAA,MAChB,UAAU;AAAA,QACR,UAAU,KAAK;AAAA,QACf,eAAe,KAAK;AAAA,MACtB;AAAA,IACF,EAAE;AAAA,IACF,QAAQ;AAAA,MACN;AAAA,QACE,QAAQ;AAAA,QACR,MAAM,IAAI,OAAO,YAAY;AAAA,QAC7B,OAAO,IAAI;AAAA,QACX,QAAQ,IAAI;AAAA,QACZ,UAAU,IAAI,OAAO,SAAS;AAAA,QAC9B;AAAA,MACF;AAAA,IACF;AAAA,IACA,SAAS;AAAA,MACP,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,SAAS,IAAI;AAAA,MACb,QAAQ,IAAI;AAAA,MACZ,YAAY;AAAA,MACZ,UAAU;AAAA,QACR,WAAW,IAAI;AAAA,QACf,cAAc,IAAI;AAAA,MACpB;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAEA,SAAS,UAAU,YAAiD;AAClE,QAAM,SAAS;AAAA,IACb,GAAG,WAAW;AAAA,IACd,GAAG,WAAW,SAAS,QAAQ,CAAC,YAAY,QAAQ,YAAY,CAAC,CAAC;AAAA,EACpE;AACA,QAAM,OAAO,oBAAI,IAAY;AAC7B,SAAO,OAAO,OAAO,CAAC,UAAU;AAC9B,UAAM,MACJ,MAAM,MAAM,GAAG,MAAM,MAAM,IAAI,MAAM,IAAI,IAAI,MAAM,SAAS,IAAI,KAAK,UAAU,MAAM,KAAK,CAAC;AAC7F,QAAI,KAAK,IAAI,GAAG,EAAG,QAAO;AAC1B,SAAK,IAAI,GAAG;AACZ,WAAO;AAAA,EACT,CAAC;AACH;AAEA,SAAS,gBAAgB,QAA6C;AACpE,MAAI,CAAC,OAAO,OAAQ,QAAO;AAC3B,QAAM,SAAS,OACZ,IAAI,CAAC,UAAU;AACd,QAAI,MAAM,SAAS,aAAa,MAAM,SAAS,SAAU,QAAO;AAChE,QAAI,MAAM,SAAS,YAAY,MAAM,SAAS,eAAgB,QAAO;AACrE,QAAI,MAAM,SAAS,UAAU,OAAO,MAAM,UAAU;AAClD,aAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,MAAM,KAAK,CAAC;AAC7C,WAAO;AAAA,EACT,CAAC,EACA,OAAO,CAAC,UAA2B,OAAO,UAAU,QAAQ;AAC/D,MAAI,CAAC,OAAO,OAAQ,QAAO;AAC3B,SAAO,KAAK,MAAO,OAAO,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO,SAAU,GAAI,IAAI;AAC9F;AAEA,SAAS,qBACP,YACA,OACoB;AACpB,MAAI,MAAM,SAAS,YAAY,MAAM;AACnC,WAAO,uBAAuB,QAAQ,WAAW,KAAK,QAAQ,EAAE,CAAC,WAAW,MAAM,MAAM;AAC1F,MAAI,MAAM,SAAS,sBAAsB,MAAM;AAC7C,WAAO,oCAAoC,MAAM,MAAM;AACzD,MAAI,MAAM,SAAS,YAAY,MAAM;AACnC,WAAO,gCAAgC,QAAQ,WAAW,KAAK,QAAQ,EAAE,CAAC,cAAc,MAAM,MAAM;AACtG,MAAI,MAAM,SAAS,aAAa,MAAM;AACpC,WAAO,oCAAoC,QAAQ,WAAW,KAAK,QAAQ,EAAE,CAAC,MAAM,MAAM,MAAM;AAClG,MAAI,MAAM,SAAS,aAAa,MAAM,OAAQ,QAAO,MAAM;AAC3D,SAAO;AACT;AAEA,SAAS,eAAe,OAA8B;AACpD,QAAM,WACJ,MAAM,aAAa,aACf,IACA,MAAM,aAAa,UACjB,IACA,MAAM,aAAa,YACjB,IACA;AACV,QAAM,SACJ,MAAM,WAAW,SACb,IACA,MAAM,WAAW,YAAY,MAAM,WAAW,gBAC5C,IACA;AACR,SAAO,WAAW;AACpB;AAEA,SAAS,cAAc,YAAgC,QAA2C;AAChG,MAAI,OAAO,aAAa,WAAW,cAAc,OAAO,UAAW,QAAO;AAC1E,MAAI,OAAO,cAAc,WAAW,eAAe,OAAO,WAAY,QAAO;AAC7E,MAAI,OAAO,SAAS,WAAW,UAAU,OAAO,MAAO,QAAO;AAC9D,MAAI,OAAO,KAAK;AACd,UAAM,CAAC,KAAK,KAAK,IAAI,OAAO;AAC5B,QAAI,WAAW,OAAO,GAAG,MAAM,MAAO,QAAO;AAAA,EAC/C;AACA,SAAO;AACT;AAEA,SAAS,gBAAgB,YAAoD;AAC3E,SAAO,KAAK,MAAM,KAAK,UAAU,UAAU,CAAC;AAC9C;AAEA,SAAS,QAAQ,OAAe,KAAqB;AACnD,QAAM,aAAa,MAAM,QAAQ,QAAQ,GAAG,EAAE,KAAK;AACnD,SAAO,WAAW,SAAS,MAAM,GAAG,WAAW,MAAM,GAAG,GAAG,EAAE,KAAK,CAAC,QAAQ;AAC7E;AAEA,SAAS,WAAW,OAAuB;AACzC,MAAI,OAAO;AACX,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,GAAG;AACxC,YAAQ,MAAM,WAAW,CAAC;AAC1B,WAAO,KAAK,KAAK,MAAM,QAAQ;AAAA,EACjC;AACA,SAAO,SAAS;AAClB;AAEA,SAAS,aAAa,OAAyB;AAC7C,MAAI,UAAU,QAAQ,OAAO,UAAU,SAAU,QAAO;AACxD,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,YAAY;AACvD,QAAM,MAA+B,CAAC;AACtC,aAAW,OAAO,OAAO,KAAK,KAAgC,EAAE,KAAK,GAAG;AACtE,QAAI,GAAG,IAAI,aAAc,MAAkC,GAAG,CAAC;AAAA,EACjE;AACA,SAAO;AACT;;;ACzjBO,IAAM,cAAN,MAAkB;AAAA,EACN;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEjB,YAAY,QAA2B;AACrC,QAAI,CAAC,OAAO,aAAa;AACvB,YAAM,IAAI,MAAM,sCAAsC;AAAA,IACxD;AACA,SAAK,oBAAoB,OAAO,qBAAqB;AACrD,SAAK,uBAAuB,OAAO,wBAAwB;AAC3D,SAAK,sBAAsB,OAAO,uBAAuB;AACzD,SAAK,cAAc,OAAO;AAC1B,SAAK,aAAa,OAAO,cAAc;AACvC,SAAK,YAAY,OAAO,sBAAsB;AAC9C,SAAK,OAAO,OAAO;AAAA,EACrB;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,SAAS,WAAwB,UAAqC;AACpE,UAAM,cAAc,iBAAiB,WAAW,KAAK,WAAW;AAChE,UAAM,aAAa,KAAK;AAGxB,UAAM,uBAAuB,kBAAkB,QAAQ;AACvD,UAAM,gBAA0B,CAAC;AACjC,UAAM,eAAyB,CAAC;AAChC,eAAW,OAAO,WAAW;AAC3B,UAAI,IAAI,aAAa,UAAW;AAChC,UAAI,IAAI,QAAQ,iBAAiB,OAAW;AAC5C,YAAM,MAAM,QAAQ,GAAG;AACvB,YAAM,cAAc,qBAAqB,IAAI,GAAG;AAChD,UAAI,gBAAgB,OAAW;AAC/B,oBAAc,KAAK,WAAW;AAC9B,mBAAa,KAAK,IAAI,QAAQ,YAAY;AAAA,IAC5C;AAEA,UAAM,iBAAiB,cAAc;AAIrC,UAAM,sBAAsB,KAAK,OAAO,WAAW,eAAe,QAAQ,CAAC;AAC3E,UAAM,uBAAuB,KAAK,OAAO,WAAW,gBAAgB,SAAS,CAAC;AAC9E,UAAM,qBAAqB,KAAK,OAAO,UAAU,eAAe,QAAQ,CAAC;AACzE,UAAM,sBAAsB,KAAK,OAAO,UAAU,gBAAgB,SAAS,CAAC;AAE5E,UAAM,aAAa,SAAS,qBAAqB,oBAAoB;AACrE,UAAM,qBAAqB,SAAS,oBAAoB,mBAAmB;AAG3E,QAAI,iBAAiB,KAAK,mBAAmB;AAC3C,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA,UAAU;AAAA,UACR;AAAA,UACA,mBAAmB,iBAAiB,IAAI,YAAY,eAAe,YAAY,IAAI;AAAA,UACnF,UAAU,EAAE,KAAK,GAAG,MAAM,EAAE;AAAA,UAC5B,cAAc;AAAA,UACd,aAAa;AAAA,UACb,cAAc;AAAA,UACd;AAAA,UACA;AAAA,QACF;AAAA,QACA,QAAQ,aAAa,cAAc,wCAAwC,KAAK,iBAAiB;AAAA,QACjG,eAAe;AAAA,MACjB;AAAA,IACF;AAGA,UAAM,KAAK,gBAAgB,eAAe,cAAc;AAAA,MACtD,YAAY,KAAK;AAAA,MACjB,WAAW,KAAK;AAAA,MAChB,WAAW;AAAA,MACX,MAAM,KAAK;AAAA,IACb,CAAC;AACD,UAAM,WAAW,eAAe,eAAe,YAAY;AAE3D,UAAM,WAAyB;AAAA,MAC7B;AAAA,MACA,mBAAmB,GAAG;AAAA,MACtB,UAAU,EAAE,KAAK,GAAG,KAAK,MAAM,GAAG,KAAK;AAAA,MACvC,cAAc,SAAS;AAAA,MACvB,aAAa;AAAA,MACb,cAAc;AAAA,MACd;AAAA,MACA;AAAA,IACF;AAGA,QAAI,EAAE,GAAG,MAAM,KAAK,uBAAuB;AACzC,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA;AAAA,QACA,QACE,gDAA2C,IAAI,GAAG,MAAM,CAAC,QAClD,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC,8BAA8B,IAAI,KAAK,oBAAoB,CAAC;AAAA,QACjG,eAAe;AAAA,MACjB;AAAA,IACF;AAIA,QACE,OAAO,SAAS,UAAU,KAC1B,OAAO,SAAS,kBAAkB,KAClC,aAAa,qBAAqB,KAAK,qBACvC;AACA,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA;AAAA,QACA,QACE,8BAA8B,IAAI,UAAU,CAAC,yBAAyB,IAAI,kBAAkB,CAAC,iBAC7E,IAAI,KAAK,mBAAmB,CAAC;AAAA,QAC/C,eAAe;AAAA,MACjB;AAAA,IACF;AAEA,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA;AAAA,MACA;AAAA,MACA,QACE,yCAAoC,IAAI,GAAG,MAAM,CAAC,QAC3C,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC,UAAU,cAAc,iCAClC,IAAI,UAAU,CAAC,gBAAgB,IAAI,kBAAkB,CAAC;AAAA,MACjF,eAAe;AAAA,IACjB;AAAA,EACF;AACF;AAIA,SAAS,iBAAiB,WAAwB,aAA6B;AAC7E,aAAW,OAAO,WAAW;AAC3B,QAAI,IAAI,eAAe,IAAI,gBAAgB,YAAa,QAAO,IAAI;AAAA,EACrE;AAGA,SAAO,UAAU,CAAC,GAAG,eAAe;AACtC;AAEA,SAAS,kBAAkB,MAAwC;AACjE,QAAM,MAAM,oBAAI,IAAoB;AACpC,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,aAAa,UAAW;AAC9B,QAAI,EAAE,QAAQ,iBAAiB,OAAW;AAC1C,QAAI,IAAI,QAAQ,CAAC,GAAG,EAAE,QAAQ,YAAY;AAAA,EAC5C;AACA,SAAO;AACT;AAEA,SAAS,QAAQ,GAAsB;AACrC,SAAO,GAAG,EAAE,YAAY,KAAK,EAAE,IAAI;AACrC;AAEA,SAAS,OACP,MACA,OACA,aACU;AACV,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,aAAa,YAAa;AAChC,UAAM,IAAI,EAAE,QAAQ,KAAK;AACzB,QAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,KAAK,CAAC;AAAA,EAC7D;AACA,SAAO;AACT;AAEA,SAAS,KAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,SAAS,GAAW,GAAmB;AAC9C,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO;AAC9D,SAAO,IAAI;AACb;AAEA,SAAS,YAAY,QAAkB,OAAyB;AAC9D,QAAM,KAAK,OAAO,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,IAAK,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACnE,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,QAAM,MAAM,KAAK,MAAM,GAAG,SAAS,CAAC;AACpC,SAAO,GAAG,SAAS,MAAM,KAAK,GAAG,MAAM,CAAC,IAAK,GAAG,GAAG,KAAM,IAAI,GAAG,GAAG;AACrE;AAEA,SAAS,IAAI,GAAmB;AAC9B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO,CAAC;AACxC,SAAO,EAAE,QAAQ,CAAC;AACpB;;;AC/QO,SAAS,UAAa,GAAM,GAAM,YAAqC;AAC5E,MAAI,iBAAiB;AACrB,aAAW,OAAO,YAAY;AAC5B,UAAM,KAAK,IAAI,MAAM,CAAC;AACtB,UAAM,KAAK,IAAI,MAAM,CAAC;AACtB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,CAAC,OAAO,SAAS,EAAE,EAAG,QAAO;AACzD,UAAM,YAAY,IAAI,cAAc,aAAa,KAAK,KAAK,KAAK;AAChE,UAAM,WAAW,IAAI,cAAc,aAAa,KAAK,KAAK,KAAK;AAC/D,QAAI,SAAU,QAAO;AACrB,QAAI,UAAW,kBAAiB;AAAA,EAClC;AACA,SAAO;AACT;AAOO,SAAS,eAAkB,YAAiB,YAA6C;AAC9F,MAAI,WAAW,WAAW,GAAG;AAC3B,UAAM,IAAI,MAAM,+CAA+C;AAAA,EACjE;AACA,QAAM,QAAQ,WAAW,OAAO,CAAC,MAAM,WAAW,MAAM,CAAC,MAAM,OAAO,SAAS,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC;AAC3F,QAAM,WAAgB,CAAC;AACvB,QAAM,YAAiB,CAAC;AACxB,aAAW,KAAK,OAAO;AACrB,UAAM,cAAc,MAAM,KAAK,CAAC,UAAU,UAAU,KAAK,UAAU,OAAO,GAAG,UAAU,CAAC;AACxF,QAAI,YAAa,WAAU,KAAK,CAAC;AAAA,QAC5B,UAAS,KAAK,CAAC;AAAA,EACtB;AACA,QAAM,eAAe,SAAS,IAAI,CAAC,OAAO;AAAA,IACxC,WAAW;AAAA,IACX,WAAW,UAAU,OAAO,CAAC,MAAM,UAAU,GAAG,GAAG,UAAU,CAAC;AAAA,EAChE,EAAE;AACF,SAAO,EAAE,UAAU,WAAW,aAAa;AAC7C;AAWO,SAAS,YACd,YACA,YACA,UAAyD,CAAC,GAClB;AACxC,MAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AACrC,QAAM,UAAU,QAAQ,WAAW,CAAC;AACpC,QAAM,cAAc,WAAW,OAAO,CAAC,GAAG,MAAM,KAAK,QAAQ,EAAE,IAAI,KAAK,IAAI,CAAC;AAG7E,QAAM,SAAS,WAAW,IAAI,CAAC,QAAQ;AACrC,UAAM,SAAS,WAAW,IAAI,CAAC,MAAM,IAAI,MAAM,CAAC,CAAC,EAAE,OAAO,CAAC,MAAM,OAAO,SAAS,CAAC,CAAC;AACnF,QAAI,OAAO,WAAW,EAAG,QAAO,EAAE,KAAK,GAAG,KAAK,EAAE;AACjD,UAAM,MAAM,KAAK,IAAI,GAAG,MAAM;AAC9B,UAAM,MAAM,KAAK,IAAI,GAAG,MAAM;AAC9B,WAAO,EAAE,KAAK,KAAK,QAAQ,MAAM,MAAM,IAAI,IAAI;AAAA,EACjD,CAAC;AAED,SAAO,WAAW,IAAI,CAAC,MAAM;AAC3B,QAAI,QAAQ;AACZ,eAAW,QAAQ,CAAC,KAAK,MAAM;AAC7B,YAAM,IAAI,IAAI,MAAM,CAAC;AACrB,UAAI,CAAC,OAAO,SAAS,CAAC,EAAG;AACzB,YAAM,EAAE,KAAK,IAAI,IAAI,OAAO,CAAC;AAC7B,YAAM,cAAc,IAAI,QAAQ,MAAM;AACtC,YAAM,cAAc,IAAI,cAAc,aAAa,aAAa,IAAI;AACpE,YAAM,UAAU,QAAQ,IAAI,IAAI,KAAK,KAAK;AAC1C,eAAS,cAAc;AAAA,IACzB,CAAC;AACD,WAAO,EAAE,WAAW,GAAG,MAAM;AAAA,EAC/B,CAAC;AACH;AAcO,SAAS,iBACd,YACA,YAC2C;AAC3C,QAAM,YAAY,IAAI,IAAe,WAAW,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAElE,aAAW,OAAO,YAAY;AAC5B,UAAM,SAAS,CAAC,GAAG,UAAU,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,MAAM,CAAC,IAAI,IAAI,MAAM,CAAC,CAAC;AACzE,UAAM,MAAM,IAAI,MAAM,OAAO,CAAC,CAAE;AAChC,UAAM,MAAM,IAAI,MAAM,OAAO,OAAO,SAAS,CAAC,CAAE;AAChD,UAAM,QAAQ,MAAM,OAAO;AAG3B,cAAU,IAAI,OAAO,CAAC,GAAI,QAAQ;AAClC,cAAU,IAAI,OAAO,OAAO,SAAS,CAAC,GAAI,QAAQ;AAClD,aAAS,IAAI,GAAG,IAAI,OAAO,SAAS,GAAG,KAAK;AAC1C,YAAM,OAAO,IAAI,MAAM,OAAO,IAAI,CAAC,CAAE;AACrC,YAAM,OAAO,IAAI,MAAM,OAAO,IAAI,CAAC,CAAE;AACrC,YAAM,UAAU,UAAU,IAAI,OAAO,CAAC,CAAE;AACxC,UAAI,YAAY,SAAU;AAC1B,gBAAU,IAAI,OAAO,CAAC,GAAI,WAAW,OAAO,QAAQ,KAAK;AAAA,IAC3D;AAAA,EACF;AAEA,SAAO,WAAW,IAAI,CAAC,OAAO,EAAE,WAAW,GAAG,UAAU,UAAU,IAAI,CAAC,KAAK,EAAE,EAAE;AAClF;AAOO,SAAS,2BACd,YACA,YAC2C;AAC3C,QAAM,EAAE,SAAS,IAAI,eAAe,YAAY,UAAU;AAC1D,MAAI,SAAS,WAAW,EAAG,QAAO,CAAC;AACnC,QAAM,YAAY,iBAAiB,UAAU,UAAU;AACvD,SAAO,UAAU,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AACzD;;;ACpBO,IAAM,qBAAN,MAA+C;AAAA,EAC5C,QAAQ,oBAAI,IAAyB;AAAA,EAC7C,IAAI,KAAsC;AACxC,WAAO,KAAK,MAAM,IAAI,GAAG;AAAA,EAC3B;AAAA,EACA,IAAI,KAAa,OAA0B;AACzC,SAAK,MAAM,IAAI,KAAK,KAAK;AAAA,EAC3B;AAAA,EACA,OAAe;AACb,WAAO,KAAK,MAAM;AAAA,EACpB;AAAA,EACA,QAAc;AACZ,SAAK,MAAM,MAAM;AAAA,EACnB;AACF;AAyCA,eAAsB,mBACpB,QACmC;AACnC,QAAM,cAAqC,CAAC;AAC5C,MAAI,aAAa,CAAC,GAAG,OAAO,YAAY;AACxC,MAAI,cAAmC,WAAW,CAAC;AACnD,MAAI,gBAAyC;AAE7C,WAAS,aAAa,GAAG,aAAa,OAAO,aAAa,cAAc;AACtE,WAAO,aAAa,EAAE,MAAM,oBAAoB,YAAY,gBAAgB,WAAW,OAAO,CAAC;AAE/F,UAAM,SAAS,MAAM,gBAAgB,YAAY,QAAQ,UAAU;AACnE,UAAM,aAAa,gBAAgB,YAAY,OAAO,aAAa,MAAM;AAEzE,UAAM,QAAQ,2BAA2B,YAAY,OAAO,UAAU;AACtE,UAAM,WAAW,IAAI,IAAI,MAAM,IAAI,CAAC,MAAM,EAAE,UAAU,SAAS,CAAC;AAEhE,UAAM,SAAS,YAAY,YAAY,OAAO,YAAY,EAAE,SAAS,OAAO,cAAc,CAAC;AAC3F,WAAO,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AACvC,UAAM,WAAW,OAAO,CAAC,GAAG,UAAU,aAAa,WAAW,CAAC,GAAG,aAAa,WAAW,CAAC,EAAG;AAE9F,UAAM,SAA8B;AAAA,MAClC,OAAO,OAAO;AAAA,MACd,QAAQ,OAAO;AAAA,MACf;AAAA,MACA,UAAU;AAAA,MACV;AAAA,MACA,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,UAAU,SAAS;AAAA,MACtD;AAAA,MACA;AAAA,IACF;AACA,gBAAY,KAAK,MAAM;AACvB,WAAO,aAAa,EAAE,MAAM,uBAAuB,OAAO,CAAC;AAE3D,UAAM,YAAY,WAAW,KAAK,CAAC,MAAM,EAAE,cAAc,QAAQ;AACjE,QAAI,WAAW;AACb,YAAM,SAAS,WAAW,KAAK,CAAC,MAAM,EAAE,OAAO,QAAQ;AACvD,UAAI,OAAQ,eAAc;AAC1B,sBAAgB;AAAA,IAClB;AAGA,QAAI,OAAO,6BAA6B,SAAS,YAAY,UAAU,GAAG;AACxE,YAAM,OAAO,YAAY,YAAY,SAAS,CAAC;AAC/C,YAAM,WACJ,KAAK,aAAa,YAAY,eAAe,KAAK,gBAAgB,CAAC,GAAG,QAAQ,CAAC;AACjF,UAAI,UAAU;AACZ,eAAO,aAAa;AAAA,UAClB,MAAM;AAAA,UACN;AAAA,UACA,QAAQ;AAAA,QACV,CAAC;AACD;AAAA,MACF;AAAA,IACF;AAEA,QAAI,eAAe,OAAO,cAAc,EAAG;AAE3C,iBAAa,MAAM,eAAe,YAAY,YAAY,QAAQ,OAAO,QAAQ,aAAa,CAAC;AAAA,EACjG;AAEA,SAAO;AAAA,IACL,OAAO,OAAO;AAAA,IACd,QAAQ,OAAO;AAAA,IACf;AAAA,IACA;AAAA,IACA,eACE,iBACA,gBAAgB,YAAY,OAAO,aAAa,CAAC,CAAC,EAAE;AAAA,MAClD,CAAC,MAAM,EAAE,cAAc,YAAY;AAAA,IACrC;AAAA,EACJ;AACF;AAEA,eAAe,gBACb,YACA,QACA,YACwB;AACxB,QAAM,OAA0C,CAAC;AACjD,aAAW,WAAW,YAAY;AAChC,eAAW,cAAc,OAAO,aAAa;AAC3C,eAAS,MAAM,GAAG,MAAM,OAAO,MAAM,OAAO;AAC1C,aAAK,KAAK,YAAY;AACpB,gBAAM,WAAW,GAAG,QAAQ,EAAE,IAAI,UAAU,IAAI,GAAG;AACnD,gBAAM,SAAS,OAAO,OAAO,IAAI,QAAQ;AACzC,cAAI,QAAQ;AACV,mBAAO,aAAa;AAAA,cAClB,MAAM;AAAA,cACN;AAAA,cACA,WAAW,QAAQ;AAAA,cACnB;AAAA,cACA;AAAA,cACA,IAAI,OAAO;AAAA,cACX,OAAO,OAAO;AAAA,cACd,QAAQ;AAAA,YACV,CAAC;AACD,mBAAO;AAAA,UACT;AACA,gBAAM,SAAS,MAAM,OAAO,aAAa,MAAM,EAAE,SAAS,YAAY,IAAI,CAAC;AAC3E,iBAAO,OAAO,IAAI,UAAU,MAAM;AAClC,iBAAO,aAAa;AAAA,YAClB,MAAM;AAAA,YACN;AAAA,YACA,WAAW,QAAQ;AAAA,YACnB;AAAA,YACA;AAAA,YACA,IAAI,OAAO;AAAA,YACX,OAAO,OAAO;AAAA,YACd,QAAQ;AAAA,UACV,CAAC;AACD,iBAAO;AAAA,QACT,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AACA,SAAO,mBAAmB,MAAM,OAAO,gBAAgB;AACzD;AAEA,eAAe,mBACb,MACA,aACc;AACd,QAAM,UAAe,IAAI,MAAM,KAAK,MAAM;AAC1C,QAAM,QAAQ,KAAK,IAAI,GAAG,WAAW;AACrC,MAAI,OAAO;AACX,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,KAAK,OAAQ;AACtB,cAAQ,CAAC,IAAI,MAAM,KAAK,CAAC,EAAG;AAAA,IAC9B;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,MAAM,GAAG,MAAM,OAAO,CAAC,CAAC;AAC/D,SAAO;AACT;AAEA,SAAS,gBACP,YACA,aACA,QACoB;AACpB,SAAO,WAAW,IAAI,CAAC,YAAY;AACjC,UAAM,gBAAgB,OAAO,OAAO,CAAC,MAAM,EAAE,cAAc,QAAQ,EAAE;AACrE,UAAM,YAAY,YAAY,IAAuB,CAAC,QAAQ;AAC5D,YAAM,iBAAiB,cAAc,OAAO,CAAC,MAAM,EAAE,eAAe,GAAG;AACvE,YAAM,WAAW,eAAe,OAAO,CAAC,MAAM,EAAE,EAAE;AAMlD,YAAM,eAAe,eAAe,OAAO,CAAC,MAAM,CAAC,EAAE,KAAK;AAC1D,YAAM,UAAU,iBAAiB,aAAa,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;AACzE,aAAO;AAAA,QACL,WAAW,QAAQ;AAAA,QACnB,YAAY;AAAA,QACZ,WAAWA,MAAK,aAAa,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAAA,QAChD,UAAUA,MAAK,aAAa,IAAI,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;AAAA,QACnD,gBAAgBA,MAAK,aAAa,IAAI,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC;AAAA,QAC/D,QAAQ,eAAe,WAAW,IAAI,IAAI,SAAS,SAAS,eAAe;AAAA,QAC3E,QAAQ,eAAe;AAAA,QACvB;AAAA,MACF;AAAA,IACF,CAAC;AACD,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB,WAAWA,MAAK,UAAU,IAAI,CAAC,MAAM,EAAE,SAAS,CAAC;AAAA,MACjD,UAAUA,MAAK,UAAU,IAAI,CAAC,MAAM,EAAE,QAAQ,CAAC;AAAA,MAC/C,gBAAgBA,MAAK,UAAU,IAAI,CAAC,MAAM,EAAE,cAAc,CAAC;AAAA,MAC3D,QAAQA,MAAK,UAAU,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC;AAAA,MAC3C;AAAA,MACA,SAAS,iBAAiB,UAAU,IAAI,CAAC,MAAM,EAAE,OAAO,CAAC;AAAA,IAC3D;AAAA,EACF,CAAC;AACH;AAEA,SAAS,iBAAiB,MAA6D;AACrF,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,OAAO,MAAM;AACtB,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,GAAG,GAAG;AACxC,UAAI,CAAC,OAAO,SAAS,CAAC,EAAG;AACzB,YAAM,OAAO,QAAQ,IAAI,CAAC,KAAK,CAAC;AAChC,WAAK,KAAK,CAAC;AACX,cAAQ,IAAI,GAAG,IAAI;AAAA,IACrB;AAAA,EACF;AACA,QAAM,MAA8B,CAAC;AACrC,aAAW,CAAC,GAAG,IAAI,KAAK,QAAS,KAAI,CAAC,IAAIA,MAAK,IAAI;AACnD,SAAO;AACT;AAEA,SAASA,MAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,eAAe,eACb,SACA,YACA,QACA,OACA,QACA,gBACgC;AAChC,QAAM,cAAc,IAAI,IAAI,MAAM,IAAI,CAAC,MAAM,EAAE,UAAU,SAAS,CAAC;AACnE,QAAM,YAAY,QAAQ,OAAO,CAAC,MAAM,YAAY,IAAI,EAAE,EAAE,CAAC;AAG7D,QAAM,SAAS,YAAY,YAAY,OAAO,YAAY,EAAE,SAAS,OAAO,cAAc,CAAC,EAAE;AAAA,IAC3F,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE;AAAA,EACxB;AACA,QAAM,WAAW,OAAO,CAAC,GAAG,UAAU,aAAa,QAAQ,CAAC,EAAG;AAC/D,QAAM,SAAS,QAAQ,KAAK,CAAC,MAAM,EAAE,OAAO,QAAQ,KAAK,QAAQ,CAAC;AAClE,QAAM,kBAAkB,WAAW,KAAK,CAAC,MAAM,EAAE,cAAc,OAAO,EAAE,KAAK,WAAW,CAAC;AAEzF,QAAM,YAAY,kBAAkB,QAAQ,OAAO,IAAI,CAAC;AACxD,QAAM,eAAe,qBAAqB,QAAQ,OAAO,IAAI,CAAC;AAC9D,QAAM,aAAa,KAAK,IAAI,GAAG,OAAO,iBAAiB,UAAU,MAAM;AACvE,MAAI,WAAkC,CAAC;AACvC,MAAI,aAAa,GAAG;AAClB,eAAW,MAAM,OAAO,cAAc,OAAO;AAAA,MAC3C;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,YAAY;AAAA,IACd,CAAC;AACD,eAAW,SACR,MAAM,GAAG,UAAU,EACnB,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,YAAY,gBAAgB,UAAU,OAAO,GAAG,EAAE;AAAA,EAC3E;AACA,SAAO,CAAC,GAAG,WAAW,GAAG,QAAQ;AACnC;AAEA,SAAS,kBAAkB,QAAuB,WAAmB,GAA0B;AAC7F,SAAO,OACJ,OAAO,CAAC,MAAM,EAAE,cAAc,aAAa,EAAE,EAAE,EAC/C,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,CAAC;AACf;AAEA,SAAS,qBAAqB,QAAuB,WAAmB,GAA0B;AAChG,SAAO,OACJ,OAAO,CAAC,MAAM,EAAE,cAAc,aAAa,EAAE,EAAE,EAC/C,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,CAAC;AACf;AAEA,SAAS,eAAe,GAAa,GAAsB;AACzD,MAAI,EAAE,WAAW,EAAE,OAAQ,QAAO;AAClC,QAAM,OAAO,IAAI,IAAI,CAAC;AACtB,SAAO,EAAE,MAAM,CAAC,OAAO,KAAK,IAAI,EAAE,CAAC;AACrC;;;AC/QA,eAAsB,yBACpB,QACyC;AACzC,iBAAe,MAAM;AAErB,QAAM,eAAgC;AAAA,IACpC,OAAO,CAAC,SAAS,SAAS,QAAQ,KAAK,SAAS,KAAK,YAAY,KAAK,KAAK,QAAQ;AAAA,EACrF;AAEA,QAAM,YAAY,MAAM,mBAAsB;AAAA,IAC5C,OAAO,OAAO;AAAA,IACd,QAAQ,OAAO;AAAA,IACf,cAAc,OAAO;AAAA,IACrB,aAAa,OAAO;AAAA,IACpB,MAAM,OAAO;AAAA,IACb,aAAa,OAAO;AAAA,IACpB,gBAAgB,OAAO;AAAA,IACvB,kBAAkB,OAAO,oBAAoB;AAAA,IAC7C;AAAA,IACA,eAAe;AAAA,MACb,QAAQ,CAAC,SACP,OAAO,cAAc,OAAO;AAAA,QAC1B,GAAG;AAAA,QACH,WAAW,KAAK;AAAA,QAChB,cAAc,KAAK;AAAA,MACrB,CAAC;AAAA,IACL;AAAA,IACA,YAAY,OAAO,cAAc,2BAA2B;AAAA,IAC5D,eAAe,OAAO;AAAA,IACtB,0BAA0B,OAAO;AAAA,IACjC,OAAO,OAAO;AAAA,IACd,YAAY,OAAO;AAAA,EACrB,CAAC;AAED,MAAI,OAAmC;AACvC,QAAM,WAAW,OAAO,aAAa,CAAC;AACtC,MAAI,kBAAkB,UAAU;AAChC,MAAI,oBAAoB,UAAU;AAClC,MAAI,OAAO,QAAQ,UAAU,YAAY,OAAO,SAAS,IAAI;AAC3D,WAAO,MAAM,sBAAsB,QAAQ,UAAU,UAAU,WAAW;AAC1E,QAAI,CAAC,KAAK,SAAS,SAAS;AAC1B,wBAAkB;AAClB,0BAAoB,aAAa,WAAW,SAAS,EAAE;AAAA,IACzD;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,mBAAmB,UAAU;AAAA,IAC7B,qBAAqB,UAAU;AAAA,IAC/B;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEO,SAAS,6BAA4D;AAC1E,SAAO;AAAA,IACL,EAAE,MAAM,SAAS,WAAW,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU;AAAA,IAClE,EAAE,MAAM,QAAQ,WAAW,YAAY,OAAO,CAAC,MAAM,EAAE,SAAS;AAAA,EAClE;AACF;AAEO,SAAS,6BAA6B,OAAyC;AACpF,SAAO;AAAA,IACL,IAAI,GAAG,MAAM,SAAS,IAAI,MAAM,UAAU,KAAK,MAAM,GAAG;AAAA,IACxD,OAAO,MAAM;AAAA,IACb,WAAW,MAAM;AAAA,IACjB,eAAe,MAAM,OAAO,CAAC,GAAG,IAAI,CAAC,MAAM,OAAO;AAAA,MAChD,IAAI,KAAK,iBAAiB,OAAO,CAAC;AAAA,MAClC,QAAQ,KAAK;AAAA,MACb,SAAS,KAAK,WAAW;AAAA,IAC3B,EAAE;AAAA,IACF,SAAS,MAAM,WAAW,aAAa,MAAM,KAAK;AAAA,IAClD,SAAS,MAAM;AAAA,EACjB;AACF;AAEA,eAAe,sBACb,QACA,UACA,WAC8B;AAC9B,QAAM,aAAa,OAAO;AAC1B,QAAM,OAAO,WAAW,QAAQ,OAAO;AACvC,QAAM,gBAA6B,CAAC;AACpC,QAAM,eAA4B,CAAC;AAEnC,QAAM,YAAY,WAAW,qBAAqB,OAAO;AACzD,aAAW,cAAc,WAAW;AAClC,aAAS,MAAM,GAAG,MAAM,MAAM,OAAO;AACnC,YAAM,OAAO,QAAQ,QAAQ,YAAY,GAAG;AAC5C,YAAM,YAAY,MAAM,SAAS,QAAQ,UAAU,YAAY,KAAK,QAAQ;AAC5E,YAAM,YAAY,MAAM,SAAS,QAAQ,WAAW,YAAY,KAAK,QAAQ;AAC7E,mBAAa;AAAA,QACX,kBAAkB,QAAQ,UAAU,YAAY,KAAK,UAAU,MAAM,SAAS;AAAA,MAChF;AACA,oBAAc;AAAA,QACZ,kBAAkB,QAAQ,WAAW,YAAY,KAAK,UAAU,MAAM,SAAS;AAAA,MACjF;AAAA,IACF;AAAA,EACF;AAEA,aAAW,cAAc,WAAW,oBAAoB;AACtD,aAAS,MAAM,GAAG,MAAM,MAAM,OAAO;AACnC,YAAM,OAAO,QAAQ,QAAQ,YAAY,GAAG;AAC5C,YAAM,YAAY,MAAM,SAAS,QAAQ,UAAU,YAAY,KAAK,SAAS;AAC7E,YAAM,YAAY,MAAM,SAAS,QAAQ,WAAW,YAAY,KAAK,SAAS;AAC9E,mBAAa;AAAA,QACX,kBAAkB,QAAQ,UAAU,YAAY,KAAK,WAAW,MAAM,SAAS;AAAA,MACjF;AACA,oBAAc;AAAA,QACZ,kBAAkB,QAAQ,WAAW,YAAY,KAAK,WAAW,MAAM,SAAS;AAAA,MAClF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,WAAW,IAAI,YAAY,WAAW,IAAI,EAAE,SAAS,eAAe,YAAY;AACtF,SAAO,EAAE,UAAU,eAAe,aAAa;AACjD;AAEA,eAAe,SACb,QACA,SACA,YACA,KACA,OAC+B;AAC/B,QAAM,OAAO,QAAQ,QAAQ,YAAY,GAAG;AAC5C,QAAM,QAA8B,EAAE,SAAS,YAAY,KAAK,OAAO,KAAK;AAC5E,MAAI;AACF,UAAM,MAAM,MAAM,OAAO,OAAO,IAAI,KAAK;AACzC,UAAM,SAAS,MAAM,OAAO,OAAO,MAAM,EAAE,GAAG,OAAO,IAAI,CAAC;AAC1D,UAAM,MAAM,OAAO,OAAO,CAAC;AAC3B,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB;AAAA,MACA;AAAA,MACA,IAAI,OAAO,MAAM;AAAA,MACjB,OAAO,QAAQ,OAAO,KAAK;AAAA,MAC3B,MAAM,OAAO,WAAW,IAAI,WAAW;AAAA,MACvC,YAAY,OAAO,cAAc,IAAI,cAAc;AAAA,MACnD,SAAS;AAAA,QACP,GAAG,eAAe,OAAO,OAAO;AAAA,QAChC,GAAG,WAAW,GAAG;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,MACA,OAAO,IAAI;AAAA,MACX;AAAA,MACA,SAAS,OAAO,WAAW,aAAa,IAAI,KAAK;AAAA,MACjD,UAAU,OAAO;AAAA,IACnB;AAAA,EACF,SAAS,KAAK;AACZ,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB;AAAA,MACA;AAAA,MACA,IAAI;AAAA,MACJ,OAAO;AAAA,MACP,MAAM;AAAA,MACN,YAAY;AAAA,MACZ,SAAS,EAAE,OAAO,EAAE;AAAA,MACpB,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACtD;AAAA,MACA;AAAA,MACA,KAAK;AAAA,QACH;AAAA,UACE,UAAU;AAAA,UACV,SAAS,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,UACxD,oBAAoB,OAAO;AAAA,QAC7B;AAAA,MACF;AAAA,MACA,SAAS;AAAA,IACX;AAAA,EACF;AACF;AAEA,SAAS,kBACP,QACA,SACA,YACA,KACA,OACA,MACA,OACW;AACX,QAAM,SAAS,OAAO,KAAM,YAAY,EAAE,SAAS,YAAY,KAAK,OAAO,MAAM,MAAM,CAAC;AACxF,SAAO,kBAAkB,MAAM;AACjC;AAEA,SAAS,eAAkB,QAA8C;AACvE,MAAI,CAAC,OAAO,MAAM,KAAK,EAAG,OAAM,IAAI,MAAM,mDAAmD;AAC7F,MAAI,CAAC,OAAO,OAAO,KAAK,EAAG,OAAM,IAAI,MAAM,oDAAoD;AAC/F,MAAI,OAAO,aAAa,WAAW,GAAG;AACpC,UAAM,IAAI,MAAM,0DAA0D;AAAA,EAC5E;AACA,MAAI,OAAO,kBAAkB,WAAW,GAAG;AACzC,UAAM,IAAI,MAAM,+DAA+D;AAAA,EACjF;AACA,yBAAuB,OAAO,MAAM,MAAM;AAC1C,yBAAuB,OAAO,aAAa,aAAa;AACxD,yBAAuB,OAAO,gBAAgB,gBAAgB;AAC9D,MAAI,OAAO,qBAAqB;AAC9B,2BAAuB,OAAO,kBAAkB,kBAAkB;AACpE,MAAI,OAAO,iBAAiB,OAAO,aAAa,QAAQ;AACtD,UAAM,IAAI,MAAM,yEAAyE;AAAA,EAC3F;AACA;AAAA,IACE,OAAO,aAAa,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,IACnC;AAAA,EACF;AACA,eAAa,OAAO,mBAAmB,mBAAmB;AAE1D,MAAI,OAAO,MAAM;AACf,QAAI,OAAO,KAAK,mBAAmB,WAAW,GAAG;AAC/C,YAAM,IAAI,MAAM,qEAAqE;AAAA,IACvF;AACA,QAAI,OAAO,KAAK,SAAS,OAAW,wBAAuB,OAAO,KAAK,MAAM,WAAW;AACxF,iBAAa,OAAO,KAAK,oBAAoB,yBAAyB;AACtE,QAAI,OAAO,KAAK;AACd,mBAAa,OAAO,KAAK,mBAAmB,wBAAwB;AACtE,UAAM,YAAY,IAAI,IAAI,OAAO,iBAAiB;AAClD,eAAW,MAAM,OAAO,KAAK,oBAAoB;AAC/C,UAAI,UAAU,IAAI,EAAE,GAAG;AACrB,cAAM,IAAI;AAAA,UACR,+CAA+C,EAAE;AAAA,QACnD;AAAA,MACF;AAAA,IACF;AACA,UAAM,aAAa,OAAO,aAAa,CAAC,EAAG;AAC3C,QAAI,OAAO,KAAK,KAAK,gBAAgB,YAAY;AAC/C,YAAM,IAAI;AAAA,QACR,qFAAqF,UAAU;AAAA,MACjG;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,uBAAuB,OAAe,MAAoB;AACjE,MAAI,CAAC,OAAO,UAAU,KAAK,KAAK,SAAS,GAAG;AAC1C,UAAM,IAAI,MAAM,6BAA6B,IAAI,6BAA6B;AAAA,EAChF;AACF;AAEA,SAAS,aAAa,QAAkB,MAAoB;AAC1D,QAAM,OAAO,oBAAI,IAAY;AAC7B,aAAW,SAAS,QAAQ;AAC1B,QAAI,CAAC,MAAM,KAAK;AACd,YAAM,IAAI,MAAM,6BAA6B,IAAI,gCAAgC;AACnF,QAAI,KAAK,IAAI,KAAK,EAAG,OAAM,IAAI,MAAM,uCAAuC,IAAI,KAAK,KAAK,GAAG;AAC7F,SAAK,IAAI,KAAK;AAAA,EAChB;AACF;AAEA,SAAS,aAAgB,WAAqC,WAAqC;AACjG,QAAM,QAAQ,UAAU,YAAY,UAAU,YAAY,SAAS,CAAC;AACpE,QAAM,YAAY,OAAO,WAAW,KAAK,CAAC,MAAM,EAAE,cAAc,SAAS;AACzE,MAAI,CAAC,WAAW;AACd,UAAM,IAAI,MAAM,4DAA4D,SAAS,GAAG;AAAA,EAC1F;AACA,SAAO;AACT;AAEA,SAAS,QACP,QACA,YACA,KACQ;AACR,QAAM,OAAO,OAAO,YAAY;AAChC,UAAQ,OAAOC,YAAW,GAAG,UAAU,IAAO,GAAG,EAAE,KAAK,OAAO;AACjE;AAEA,SAASA,YAAW,OAAuB;AACzC,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,SAAK,MAAM,WAAW,CAAC;AACvB,QAAI,KAAK,KAAK,GAAG,QAAQ;AAAA,EAC3B;AACA,SAAO,MAAM;AACf;AAEA,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,SAAqE;AAC3F,QAAM,MAA8B,CAAC;AACrC,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,WAAW,CAAC,CAAC,GAAG;AAClD,QAAI,OAAO,SAAS,CAAC,EAAG,KAAI,CAAC,IAAI;AAAA,EACnC;AACA,SAAO;AACT;AAEA,SAAS,WAAW,KAAmD;AACrE,QAAM,MAA8B,EAAE,KAAK,IAAI,OAAO;AACtD,aAAW,QAAQ,IAAI,MAAM,GAAG,GAAI,GAAG;AACrC,UAAM,MAAM,kBAAkB,KAAK,QAAQ;AAC3C,QAAI,OAAO,GAAG,EAAE,KAAK,IAAI,OAAO,GAAG,EAAE,KAAK,KAAK;AAC/C,QAAI,KAAK,oBAAoB;AAC3B,YAAM,MAAM,WAAW,iBAAiB,KAAK,kBAAkB,CAAC;AAChE,UAAI,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK;AAAA,IAC/B;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,kBAAkB,UAAgD;AACzE,MACE,aAAa,UACb,aAAa,aACb,aAAa,WACb,aAAa,YACb;AACA,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEA,SAAS,iBAAiB,KAAqB;AAC7C,SACE,IACG,KAAK,EACL,QAAQ,qBAAqB,GAAG,EAChC,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEA,SAAS,aAAa,OAAuD;AAC3E,MAAI,CAAC,MAAO,QAAO;AACnB,MAAI,OAAO,MAAM,WAAW,SAAU,QAAO,MAAM;AACnD,MAAI,MAAM,WAAY,QAAO,MAAM;AACnC,MAAI,MAAM,OAAO;AACf,QAAI;AACF,YAAM,UAAU,MAAM,MAAM,MAAM,GAAG,EAAE;AACvC,YAAM,SACJ,MAAM,MAAM,SAAS,QAAQ,SACzB,QAAQ,MAAM,MAAM,SAAS,QAAQ,MAAM,kBAC3C;AACN,aAAO,GAAG,KAAK,UAAU,OAAO,EAAE,MAAM,GAAG,GAAI,CAAC,GAAG,MAAM;AAAA,IAC3D,QAAQ;AACN,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;;;AC7dO,IAAM,8BAAwC;AAAA,EACnD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAOO,SAAS,sBAAsB,KAAgC;AACpE,QAAM,aAAa,IAAI,sBAAsB;AAC7C,QAAM,WAAqB,CAAC;AAE5B,WAAS,KAAK,sBAAsB,IAAI,MAAM,EAAE;AAChD,WAAS,KAAK,EAAE;AAChB,WAAS;AAAA,IACP,+CAA+C,IAAI,MAAM,oDAAoD,IAAI,UAAU,MAAM,mBAAmB,IAAI,aAAa,MAAM,uCAAuC,IAAI,UAAU,YAAY,IAAI,eAAe,IAAI,KAAK,GAAG;AAAA,EAC7Q;AACA,WAAS,KAAK,EAAE;AAEhB,WAAS,KAAK,oBAAoB;AAClC,WAAS,KAAK,SAAS;AACvB,WAAS,KAAK,KAAK,UAAU,IAAI,eAAe,MAAM,CAAC,CAAC;AACxD,WAAS,KAAK,KAAK;AACnB,WAAS,KAAK,EAAE;AAEhB,MAAI,IAAI,aAAa,SAAS,GAAG;AAC/B,aAAS,KAAK,oDAA+C;AAC7D,aAAS,KAAK,EAAE;AAChB,eAAW,SAAS,IAAI,cAAc;AACpC,eAAS;AAAA,QACP,eAAe,MAAM,EAAE,mBAAc,MAAM,MAAM,QAAQ,CAAC,CAAC,GAAG,MAAM,YAAY,KAAK,MAAM,SAAS,MAAM,EAAE;AAAA,MAC9G;AACA,YAAM,UAAU,MAAM,gBAAgB,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,EAAE,OAAO;AAClE,UAAI,OAAO,SAAS,GAAG;AACrB,iBAAS,KAAK,EAAE;AAChB,iBAAS,KAAK,0BAA0B;AACxC,mBAAW,KAAK,QAAQ;AACtB,mBAAS,KAAK,OAAO,EAAE,EAAE,6BAA6B,MAAM,EAAE,MAAM,CAAC,IAAI;AAAA,QAC3E;AAAA,MACF;AACA,UAAI,MAAM,SAAS;AACjB,iBAAS,KAAK,EAAE;AAChB,iBAAS,KAAK,6BAA6B;AAC3C,iBAAS,KAAK,KAAK;AACnB,iBAAS,KAAK,SAAS,MAAM,SAAS,GAAG,CAAC;AAC1C,iBAAS,KAAK,KAAK;AAAA,MACrB;AACA,eAAS,KAAK,EAAE;AAAA,IAClB;AAAA,EACF;AAEA,MAAI,IAAI,UAAU,SAAS,GAAG;AAC5B,aAAS,KAAK,mDAA8C;AAC5D,aAAS,KAAK,EAAE;AAChB,eAAW,SAAS,IAAI,WAAW;AACjC,eAAS;AAAA,QACP,OAAO,MAAM,EAAE,aAAa,MAAM,MAAM,QAAQ,CAAC,CAAC,GAAG,MAAM,YAAY,KAAK,MAAM,SAAS,MAAM,EAAE;AAAA,MACrG;AAAA,IACF;AACA,aAAS,KAAK,EAAE;AAAA,EAClB;AAEA,WAAS,KAAK,gCAAgC;AAC9C,WAAS,KAAK,EAAE;AAChB,aAAW,KAAK,WAAY,UAAS,KAAK,KAAK,CAAC,EAAE;AAClD,WAAS,KAAK,EAAE;AAEhB,WAAS,KAAK,kBAAkB;AAChC,WAAS,KAAK,EAAE;AAChB,WAAS,KAAK,iEAA4D;AAC1E,WAAS,KAAK,SAAS;AACvB,WAAS;AAAA,IACP,KAAK;AAAA,MACH;AAAA,QACE,WAAW;AAAA,UACT;AAAA,YACE,OAAO;AAAA,YACP,WAAW;AAAA,YACX,SAAS;AAAA,UACX;AAAA,QACF;AAAA,MACF;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,WAAS,KAAK,KAAK;AAEnB,SAAO,SAAS,KAAK,IAAI;AAC3B;AAEA,SAAS,SAAS,GAAW,KAAqB;AAChD,MAAI,EAAE,UAAU,IAAK,QAAO;AAC5B,SAAO,GAAG,EAAE,MAAM,GAAG,GAAG,CAAC;AAC3B;AAEA,SAAS,MAAM,GAAmB;AAChC,SAAO,EAAE,QAAQ,MAAM,KAAK;AAC9B;AAuBA,SAAS,uBAAuB,KAA4B;AAC1D,QAAM,QAA0B,CAAC;AACjC,MAAI,WAAW;AACf,MAAI,UAAU;AACd,aAAW,KAAK,KAAK;AACnB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,UAAU;AACZ,UAAI,MAAM,MAAM;AACd,kBAAU;AACV;AAAA,MACF;AACA,UAAI,MAAM,KAAK;AACb,mBAAW;AACX;AAAA,MACF;AACA;AAAA,IACF;AACA,QAAI,MAAM,KAAK;AACb,iBAAW;AACX;AAAA,IACF;AACA,QAAI,MAAM,OAAO,MAAM,IAAK,OAAM,KAAK,CAAC;AAAA,aAC/B,MAAM,KAAK;AAClB,UAAI,MAAM,IAAI,MAAM,IAAK,QAAO;AAAA,IAClC,WAAW,MAAM,KAAK;AACpB,UAAI,MAAM,IAAI,MAAM,IAAK,QAAO;AAAA,IAClC;AAAA,EACF;AACA,MAAI,MAAM,WAAW,KAAK,CAAC,SAAU,QAAO;AAC5C,MAAI,SAAS;AACb,MAAI,SAAU,WAAU;AACxB,SAAO,MAAM,SAAS,GAAG;AACvB,UAAM,SAAS,MAAM,IAAI;AACzB,cAAU,WAAW,MAAM,MAAM;AAAA,EACnC;AACA,SAAO,MAAM;AACf;AAEO,SAAS,wBAAwB,KAAa,cAA6C;AAChG,MAAI,OAAO,IAAI,KAAK;AACpB,MAAI,KAAK,WAAW,KAAK,EAAG,QAAO,KAAK,QAAQ,oBAAoB,EAAE,EAAE,QAAQ,WAAW,EAAE;AAK7F,MAAI,SAAkB;AACtB,QAAM,cAAc,KAAK,QAAQ,GAAG;AACpC,QAAM,YAAY,KAAK,YAAY,GAAG;AACtC,QAAM,aAAa,KAAK,QAAQ,GAAG;AACnC,QAAM,WAAW,KAAK,YAAY,GAAG;AAErC,QAAM,iBAAiB,eAAe,MAAM,aAAa,KAAK,cAAc;AAC5E,QAAM,aAAuB,CAAC;AAC9B,MAAI,gBAAgB;AAClB,QAAI,eAAe,KAAK,YAAY;AAClC,iBAAW,KAAK,KAAK,MAAM,aAAa,YAAY,CAAC,CAAC;AACxD,QAAI,cAAc,KAAK,WAAW;AAChC,iBAAW,KAAK,KAAK,MAAM,YAAY,WAAW,CAAC,CAAC;AAAA,EACxD,OAAO;AACL,QAAI,cAAc,KAAK,WAAW;AAChC,iBAAW,KAAK,KAAK,MAAM,YAAY,WAAW,CAAC,CAAC;AACtD,QAAI,eAAe,KAAK,YAAY;AAClC,iBAAW,KAAK,KAAK,MAAM,aAAa,YAAY,CAAC,CAAC;AAAA,EAC1D;AACA,aAAW,SAAS,YAAY;AAC9B,QAAI;AACF,eAAS,KAAK,MAAM,KAAK;AACzB;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAOA,MAAI,UAAU,MAAM;AAClB,eAAW,SAAS,YAAY;AAC9B,YAAM,SAAS,uBAAuB,KAAK;AAC3C,UAAI,UAAU,QAAQ,WAAW,OAAO;AACtC,YAAI;AACF,mBAAS,KAAK,MAAM,MAAM;AAC1B;AAAA,QACF,QAAQ;AAAA,QAER;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,MAAI,UAAU,KAAM,QAAO,CAAC;AAG5B,MAAI;AACJ,MAAI,MAAM,QAAQ,MAAM,GAAG;AACzB,mBAAe;AAAA,EACjB,WAAW,UAAU,OAAO,WAAW,UAAU;AAC/C,mBAAgB,OAAmC;AAAA,EACrD;AACA,MAAI,CAAC,MAAM,QAAQ,YAAY,EAAG,QAAO,CAAC;AAE1C,QAAM,MAA4B,CAAC;AACnC,aAAW,KAAK,cAAc;AAC5B,QAAI,CAAC,KAAK,OAAO,MAAM,SAAU;AACjC,UAAM,MAAM;AACZ,QAAI,EAAE,aAAa,KAAM;AACzB,QAAI,KAAK;AAAA,MACP,OAAO,OAAO,IAAI,UAAU,WAAW,IAAI,QAAQ;AAAA,MACnD,WAAW,OAAO,IAAI,cAAc,WAAW,IAAI,YAAY;AAAA,MAC/D,SAAS,IAAI;AAAA,IACf,CAAC;AACD,QAAI,iBAAiB,UAAa,IAAI,UAAU,aAAc;AAAA,EAChE;AACA,SAAO;AACT;;;ACnLO,IAAM,qBAAN,MAA+C;AAAA,EACpD,YAA6B,WAAsC;AAAtC;AAAA,EAAuC;AAAA,EAAvC;AAAA,EAE7B,gBAAgB,MAA2C;AACzD,WAAO,KAAK,UAAU,gBAAgB,IAAI;AAAA,EAC5C;AAAA,EAEA,cAAc,UAAoD;AAChE,WAAO,KAAK,UAAU,cAAc,QAAQ;AAAA,EAC9C;AAAA,EAEA,YAAY,SAA2B,UAAmD;AACxF,WAAO,KAAK,UAAU,YAAY,SAAS,QAAQ;AAAA,EACrD;AAAA,EAEA,eAAe,MAAiD;AAC9D,WAAO,KAAK,UAAU,eAAe,IAAI;AAAA,EAC3C;AACF;AASO,IAAM,iBAAN,MAA2C;AAAA,EAC/B;AAAA,EAEjB,YAAY,OAAO,2CAA2C;AAC5D,SAAK,OAAO;AAAA,EACd;AAAA,EAEA,MAAM,gBAAgB,OAA4C;AAChE,UAAM,IAAI,MAAM,GAAG,KAAK,IAAI,oCAAoC;AAAA,EAClE;AAAA,EAEA,MAAM,cAAc,WAAqD;AACvE,UAAM,IAAI,MAAM,GAAG,KAAK,IAAI,kCAAkC;AAAA,EAChE;AAAA,EAEA,MAAM,YACJ,UACA,WACyB;AACzB,UAAM,IAAI,MAAM,GAAG,KAAK,IAAI,gCAAgC;AAAA,EAC9D;AAAA,EAEA,MAAM,eAAe,OAAkD;AACrE,UAAM,IAAI,MAAM,GAAG,KAAK,IAAI,mCAAmC;AAAA,EACjE;AACF;","names":["mean","stableHash"]}

package/dist/{chunk-4F5DQN55.js → chunk-VSMTAMNK.js}

@@ -50,4 +50,4 @@ export {
   verifyManifest,
   evaluateHypothesis
 };
-//# sourceMappingURL=chunk-4F5DQN55.js.map
+//# sourceMappingURL=chunk-VSMTAMNK.js.map

package/dist/chunk-VSMTAMNK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/pre-registration.ts"],"sourcesContent":["/**\n * Pre-registered hypotheses — declare what you're testing BEFORE the\n * run, check it AFTER. Prevents p-hacking, optional stopping, and the\n * \"we ran until it looked good\" failure mode.\n *\n * Manifest is a plain JSON-friendly object. Sign it with a content hash\n * + timestamp; the registered record becomes immutable. Post-run,\n * evaluate the manifest against observed results — the library refuses\n * to let you re-interpret a different metric as the declared one.\n */\n\nexport interface HypothesisManifest {\n  id: string\n  /** Human prose — goes into the audit trail. */\n  hypothesis: string\n  /** Metric the hypothesis claims to move. */\n  metric: string\n  /** 'increase' = candidate should score higher than baseline; 'decrease' = lower. */\n  direction: 'increase' | 'decrease'\n  /** Minimum effect size to count (same units as the metric). */\n  minEffect: number\n  /** Alpha threshold. */\n  alpha: number\n  /** Target statistical power at which sample size was pre-computed. */\n  power: number\n  /** Declared N per arm before running. */\n  preRegisteredN: number\n  /** ISO8601 timestamp the manifest was registered. */\n  registeredAt: string\n  /** Optional identifiers to tie into the trace corpus. */\n  baselineLabel?: string\n  candidateLabel?: string\n}\n\n/**\n * Identifier for the hashing scheme used to produce `contentHash`.\n *\n * `'sha256-content'` — sha256 hex over the canonicalized manifest with\n * the `contentHash` and `algo` fields stripped. Held as a string union\n * so future schemes can be added without breaking parsers; SignedManifest\n * values without `algo` deserialize cleanly because the field is optional.\n */\nexport type SignedManifestAlgo = 'sha256-content'\n\nexport interface SignedManifest extends HypothesisManifest {\n  /** sha256 hex of canonicalized manifest (everything except contentHash and algo). */\n  contentHash: string\n  /**\n   * Algorithm string describing how `contentHash` was produced.\n   *\n   * Optional on the type so serialized manifests without it still parse,\n   * but ALWAYS populated by {@link signManifest}. Consumers that want to\n   * enforce a known algorithm should reject manifests where this field\n   * is missing or unrecognized.\n   */\n  algo?: SignedManifestAlgo\n}\n\nexport interface HypothesisResult {\n  manifest: SignedManifest\n  observedN: number\n  observedEffect: number\n  observedPValue: number\n  /** True iff the observed effect hits the pre-declared direction with\n   *  magnitude ≥ minEffect AND p < alpha. */\n  confirmed: boolean\n  /** Enumerated reasons the hypothesis was rejected (each a machine-tag). */\n  rejectionReasons: Array<\n    'wrong_direction' | 'effect_too_small' | 'not_significant' | 'undersampled'\n  >\n  notes?: string\n}\n\n/**\n * Deterministic JSON canonicalization — sort object keys recursively.\n *\n * Two semantically-equal objects produce byte-identical canonicalized output;\n * this is what makes a content-hash stable across encoders, key insertion\n * orders, and runtime versions. Exported for any consumer that needs the same\n * canonicalization guarantee outside the manifest-signing path (e.g., signing\n * an artifact bundle, hashing a dataset version, etc.).\n */\nexport function canonicalize(v: unknown): unknown {\n  if (v === null || typeof v !== 'object') return v\n  if (Array.isArray(v)) return v.map(canonicalize)\n  const keys = Object.keys(v as Record<string, unknown>).sort()\n  const out: Record<string, unknown> = {}\n  for (const k of keys) out[k] = canonicalize((v as Record<string, unknown>)[k])\n  return out\n}\n\n/**\n * SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.\n *\n * The same primitive `signManifest` and `verifyManifest` are built on, exposed\n * directly so consumers signing arbitrary structured content (artifact bundles,\n * production packets, dataset manifests, etc.) don't have to re-derive\n * canonicalize+sha256 from scratch.\n *\n * Stable across:\n *   - object key insertion order (canonicalization sorts keys recursively)\n *   - encoder choice (UTF-8 via TextEncoder, fixed)\n *   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)\n *\n * Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,\n * which takes a string input and returns a truncated 12-char prompt id.\n * Use `hashJson` when you mean \"canonicalize then hash.\"\n *\n * @example\n *   const hash = await hashJson({ id: '1', kind: 'spec' })\n *   // 'a3f1...' (64 hex chars)\n */\nexport async function hashJson<T>(obj: T): Promise<string> {\n  const canonical = canonicalize(obj)\n  const bytes = new TextEncoder().encode(JSON.stringify(canonical))\n  const digest = await globalThis.crypto.subtle.digest('SHA-256', bytes)\n  return Array.from(new Uint8Array(digest))\n    .map((b) => b.toString(16).padStart(2, '0'))\n    .join('')\n}\n\n/**\n * Sign a manifest with a SHA-256 content hash.\n *\n * The hash covers the canonicalized manifest with the `contentHash`\n * and `algo` fields stripped; this lets verifiers re-sign the rest and\n * compare. Returned manifest always carries `algo: 'sha256-content'`\n * so downstream consumers can identify the scheme; manifests without\n * `algo` still verify because it is stripped before hashing on both sides.\n */\nexport async function signManifest(m: HypothesisManifest): Promise<SignedManifest> {\n  const hash = await hashJson(m)\n  return { ...m, contentHash: hash, algo: 'sha256-content' }\n}\n\n/**\n * Verify that a signed manifest has not been tampered with.\n *\n * Strips `contentHash` and `algo` before re-signing so manifests without\n * `algo` verify identically to ones that carry it.\n */\nexport async function verifyManifest(m: SignedManifest): Promise<boolean> {\n  const { contentHash, algo: _algo, ...rest } = m\n  void _algo\n  const resigned = await signManifest(rest)\n  return resigned.contentHash === contentHash\n}\n\n/**\n * Evaluate a pre-registered hypothesis against observed results.\n * Mechanical — no re-interpretation permitted.\n */\nexport async function evaluateHypothesis(\n  manifest: SignedManifest,\n  observed: { n: number; effect: number; pValue: number },\n): Promise<HypothesisResult> {\n  if (!(await verifyManifest(manifest))) {\n    throw new Error('evaluateHypothesis: manifest content hash mismatch (tampered)')\n  }\n  const reasons: HypothesisResult['rejectionReasons'] = []\n  const directionOk = manifest.direction === 'increase' ? observed.effect > 0 : observed.effect < 0\n  if (!directionOk) reasons.push('wrong_direction')\n  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push('effect_too_small')\n  if (observed.pValue >= manifest.alpha) reasons.push('not_significant')\n  if (observed.n < manifest.preRegisteredN) reasons.push('undersampled')\n  return {\n    manifest,\n    observedN: observed.n,\n    observedEffect: observed.effect,\n    observedPValue: observed.pValue,\n    confirmed: reasons.length === 0,\n    rejectionReasons: reasons,\n  }\n}\n"],"mappings":";AAkFO,SAAS,aAAa,GAAqB;AAChD,MAAI,MAAM,QAAQ,OAAO,MAAM,SAAU,QAAO;AAChD,MAAI,MAAM,QAAQ,CAAC,EAAG,QAAO,EAAE,IAAI,YAAY;AAC/C,QAAM,OAAO,OAAO,KAAK,CAA4B,EAAE,KAAK;AAC5D,QAAM,MAA+B,CAAC;AACtC,aAAW,KAAK,KAAM,KAAI,CAAC,IAAI,aAAc,EAA8B,CAAC,CAAC;AAC7E,SAAO;AACT;AAuBA,eAAsB,SAAY,KAAyB;AACzD,QAAM,YAAY,aAAa,GAAG;AAClC,QAAM,QAAQ,IAAI,YAAY,EAAE,OAAO,KAAK,UAAU,SAAS,CAAC;AAChE,QAAM,SAAS,MAAM,WAAW,OAAO,OAAO,OAAO,WAAW,KAAK;AACrE,SAAO,MAAM,KAAK,IAAI,WAAW,MAAM,CAAC,EACrC,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC,EAC1C,KAAK,EAAE;AACZ;AAWA,eAAsB,aAAa,GAAgD;AACjF,QAAM,OAAO,MAAM,SAAS,CAAC;AAC7B,SAAO,EAAE,GAAG,GAAG,aAAa,MAAM,MAAM,iBAAiB;AAC3D;AAQA,eAAsB,eAAe,GAAqC;AACxE,QAAM,EAAE,aAAa,MAAM,OAAO,GAAG,KAAK,IAAI;AAC9C,OAAK;AACL,QAAM,WAAW,MAAM,aAAa,IAAI;AACxC,SAAO,SAAS,gBAAgB;AAClC;AAMA,eAAsB,mBACpB,UACA,UAC2B;AAC3B,MAAI,CAAE,MAAM,eAAe,QAAQ,GAAI;AACrC,UAAM,IAAI,MAAM,+DAA+D;AAAA,EACjF;AACA,QAAM,UAAgD,CAAC;AACvD,QAAM,cAAc,SAAS,cAAc,aAAa,SAAS,SAAS,IAAI,SAAS,SAAS;AAChG,MAAI,CAAC,YAAa,SAAQ,KAAK,iBAAiB;AAChD,MAAI,KAAK,IAAI,SAAS,MAAM,IAAI,SAAS,UAAW,SAAQ,KAAK,kBAAkB;AACnF,MAAI,SAAS,UAAU,SAAS,MAAO,SAAQ,KAAK,iBAAiB;AACrE,MAAI,SAAS,IAAI,SAAS,eAAgB,SAAQ,KAAK,cAAc;AACrE,SAAO;AAAA,IACL;AAAA,IACA,WAAW,SAAS;AAAA,IACpB,gBAAgB,SAAS;AAAA,IACzB,gBAAgB,SAAS;AAAA,IACzB,WAAW,QAAQ,WAAW;AAAA,IAC9B,kBAAkB;AAAA,EACpB;AACF;","names":[]}

package/dist/{chunk-5LBB5B3Z.js → chunk-XFZCM5Z3.js}

@@ -974,4 +974,4 @@ export {
   createApp,
   startServer
 };
-//# sourceMappingURL=chunk-5LBB5B3Z.js.map
+//# sourceMappingURL=chunk-XFZCM5Z3.js.map