@tangle-network/agent-eval 0.120.2 → 0.120.4

package/dist/chunk-DPZAEKA6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/summary-report.ts"],"sourcesContent":["/**\n * Reporting helpers — production summaries and paper-quality figures — sit alongside `reporter.ts` rather\n * than replacing it.\n *\n * Three artefacts:\n *\n *   - `summaryTable`           Markdown table of per-candidate means,\n *                            95% bootstrap CIs, BH-adjusted Wilcoxon\n *                            p-values, and Cohen's d versus a\n *                            comparator candidate.\n *   - `paretoChart`         Abstract spec for a cost vs quality\n *                            scatter, with gate decisions overlaid.\n *                            Returns numbers + labels — caller\n *                            chooses the plotting library.\n *   - `gainHistogram`\n *                            Per-item paired holdout deltas as a\n *                            histogram spec (bins + counts + median +\n *                            CI). Same \"data, not images\" contract.\n *\n * The figure types are PlotSpecs — JSON-friendly, library-agnostic.\n * They aren't React components and they aren't PNGs; they are\n * what you'd hand to vega-lite, plotly, matplotlib, or your own\n * Canvas renderer to draw the actual figure.\n */\n\nimport type { GateDecision } from './held-out-gate'\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport { canonicalize, hashJson } from './pre-registration'\nimport type { RunRecord } from './run-record'\nimport {\n  benjaminiHochberg,\n  cohensD,\n  confidenceInterval,\n  pairedBootstrap,\n  pairedMde,\n  wilcoxonSignedRank,\n} from './statistics'\n\n// ── summaryTable ───────────────────────────────────────────────────────\n\nexport interface SummaryTableOptions {\n  /** Comparator candidate id. Wilcoxon + Cohen's d are computed\n   *  versus this candidate. Required for paired stats columns. */\n  comparator?: string\n  /** Which split to read scores from. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the bootstrap CI on the mean. Default 0.95. */\n  confidence?: number\n  /** FDR for BH adjustment of the comparison p-values. Default 0.05. */\n  fdr?: number\n}\n\nexport interface SummaryTableRow {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  /** BH-adjusted q-value vs comparator. NaN if no comparator. */\n  qValue: number\n  /** Cohen's d vs comparator. NaN if no comparator. */\n  cohensD: number\n}\n\nexport interface SummaryTable {\n  rows: SummaryTableRow[]\n  comparator: string | null\n  split: 'search' | 'holdout'\n  /** Pre-rendered markdown — drop into a paper or PR. */\n  markdown: string\n}\n\n/**\n * Table 1 helper. Buckets runs by `candidateId`, computes mean +\n * bootstrap CI on the chosen split, and (when a comparator is given)\n * BH-adjusted Wilcoxon p + Cohen's d versus that comparator.\n */\nexport function summaryTable(runs: RunRecord[], opts: SummaryTableOptions = {}): SummaryTable {\n  const split = opts.split ?? 'holdout'\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const comparator = opts.comparator ?? null\n  const scoreField = split === 'holdout' ? 'holdoutScore' : 'searchScore'\n\n  const byCandidate = new Map<string, { runs: RunRecord[]; scores: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    const v = r.outcome[scoreField]\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    const bucket = byCandidate.get(r.candidateId) ?? { runs: [], scores: [] }\n    bucket.runs.push(r)\n    bucket.scores.push(v)\n    byCandidate.set(r.candidateId, bucket)\n  }\n\n  const candidateIds = [...byCandidate.keys()].sort()\n  const compRuns = comparator ? byCandidate.get(comparator) : undefined\n\n  // First pass: per-candidate means + CIs + raw p-values.\n  const tentative: Array<SummaryTableRow & { rawP: number }> = []\n  for (const id of candidateIds) {\n    const bucket = byCandidate.get(id)!\n    const ci = confidenceInterval(bucket.scores, confidence)\n    let rawP = Number.NaN\n    let d = Number.NaN\n    if (comparator && compRuns && id !== comparator) {\n      const paired = pairScoresByKey(bucket.runs, compRuns.runs, scoreField)\n      if (paired.before.length >= 6) {\n        rawP = wilcoxonSignedRank(paired.before, paired.after).p\n      }\n      d = cohensD(compRuns.scores, bucket.scores)\n    }\n    tentative.push({\n      candidateId: id,\n      n: bucket.scores.length,\n      mean: ci.mean,\n      ciLow: ci.lower,\n      ciHigh: ci.upper,\n      qValue: rawP,\n      cohensD: d,\n      rawP,\n    })\n  }\n\n  // BH-adjust across the comparison set (skip NaN rows / the\n  // comparator itself). Adjustment is a no-op when there are 0 or 1\n  // comparators.\n  if (comparator) {\n    const idxs: number[] = []\n    const ps: number[] = []\n    for (let i = 0; i < tentative.length; i++) {\n      const r = tentative[i]!\n      if (r.candidateId === comparator) continue\n      if (!Number.isFinite(r.rawP)) continue\n      idxs.push(i)\n      ps.push(r.rawP)\n    }\n    if (ps.length > 0) {\n      const { qValues } = benjaminiHochberg(ps, fdr)\n      for (let k = 0; k < idxs.length; k++) {\n        tentative[idxs[k]!]!.qValue = qValues[k]!\n      }\n    }\n  }\n\n  const rows = tentative.map(({ rawP: _rawP, ...rest }) => rest)\n  const markdown = renderSummaryTableMarkdown(rows, comparator, split)\n  return { rows, comparator, split, markdown }\n}\n\nfunction pairScoresByKey(\n  candidate: RunRecord[],\n  baseline: RunRecord[],\n  scoreField: 'searchScore' | 'holdoutScore',\n): { before: number[]; after: number[] } {\n  const baseIdx = new Map<string, number>()\n  for (const r of baseline) {\n    const v = r.outcome[scoreField]\n    if (typeof v === 'number' && Number.isFinite(v)) {\n      baseIdx.set(`${r.experimentId}::${r.seed}`, v)\n    }\n  }\n  const before: number[] = []\n  const after: number[] = []\n  for (const r of candidate) {\n    const v = r.outcome[scoreField]\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    const key = `${r.experimentId}::${r.seed}`\n    const b = baseIdx.get(key)\n    if (b === undefined) continue\n    before.push(b)\n    after.push(v)\n  }\n  return { before, after }\n}\n\nfunction renderSummaryTableMarkdown(\n  rows: SummaryTableRow[],\n  comparator: string | null,\n  split: 'search' | 'holdout',\n): string {\n  const lines: string[] = []\n  const cmpLabel = comparator ? ` (vs ${comparator})` : ''\n  lines.push(`Summary Table — ${split} split${cmpLabel}`)\n  lines.push('')\n  lines.push(\"| Candidate | N | Mean | 95% CI | q (BH) | Cohen's d |\")\n  lines.push('|---|---:|---:|---|---:|---:|')\n  for (const r of rows) {\n    const ci = `[${fmt(r.ciLow)}, ${fmt(r.ciHigh)}]`\n    const q = Number.isFinite(r.qValue) ? r.qValue.toFixed(4) : '—'\n    const d = Number.isFinite(r.cohensD) ? r.cohensD.toFixed(3) : '—'\n    lines.push(`| ${r.candidateId} | ${r.n} | ${fmt(r.mean)} | ${ci} | ${q} | ${d} |`)\n  }\n  return lines.join('\\n')\n}\n\n// ── paretoChart ─────────────────────────────────────────────────────\n\nexport interface ParetoPoint {\n  candidateId: string\n  /** Mean USD cost per run on the chosen split. */\n  cost: number\n  /** Mean score on the chosen split. */\n  quality: number\n  /** Number of runs that informed this point. */\n  n: number\n  /** Whether this candidate is on the Pareto frontier — high\n   *  quality, low cost, no dominator. */\n  onFrontier: boolean\n  /** Optional gate verdict for this candidate, if a `GateDecision`\n   *  for it was passed in. */\n  gate?: 'promote' | 'reject_few_runs' | 'reject_negative_delta' | 'reject_overfit_gap' | null\n}\n\nexport interface ParetoFigureSpec {\n  kind: 'pareto-cost-quality'\n  split: 'search' | 'holdout'\n  points: ParetoPoint[]\n  axes: { x: 'costUsd'; y: 'score' }\n}\n\n/**\n * Cost vs quality scatter spec. `gateDecisions` is keyed by\n * candidate id; if present, every point picks up the gate verdict\n * for overlay.\n */\nexport function paretoChart(\n  runs: RunRecord[],\n  opts: {\n    split?: 'search' | 'holdout'\n    gateDecisions?: Record<string, GateDecision>\n  } = {},\n): ParetoFigureSpec {\n  const split = opts.split ?? 'holdout'\n  const scoreField = split === 'holdout' ? 'holdoutScore' : 'searchScore'\n\n  const buckets = new Map<string, { cost: number[]; quality: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    const v = r.outcome[scoreField]\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    const bucket = buckets.get(r.candidateId) ?? { cost: [], quality: [] }\n    bucket.cost.push(r.costUsd)\n    bucket.quality.push(v)\n    buckets.set(r.candidateId, bucket)\n  }\n\n  const points: ParetoPoint[] = []\n  for (const [candidateId, bucket] of buckets.entries()) {\n    points.push({\n      candidateId,\n      cost: avg(bucket.cost),\n      quality: avg(bucket.quality),\n      n: bucket.cost.length,\n      onFrontier: false,\n      gate: opts.gateDecisions?.[candidateId]\n        ? gateLabel(opts.gateDecisions[candidateId]!)\n        : undefined,\n    })\n  }\n\n  // Pareto: minimize cost, maximize quality. A point is dominated if\n  // some other point has lower-or-equal cost AND higher-or-equal\n  // quality with strict inequality somewhere.\n  for (const p of points) {\n    p.onFrontier = !points.some((q) => q !== p && dominates(q, p))\n  }\n\n  return {\n    kind: 'pareto-cost-quality',\n    split,\n    axes: { x: 'costUsd', y: 'score' },\n    points,\n  }\n}\n\nfunction dominates(a: ParetoPoint, b: ParetoPoint): boolean {\n  return a.cost <= b.cost && a.quality >= b.quality && (a.cost < b.cost || a.quality > b.quality)\n}\n\nfunction gateLabel(d: GateDecision): ParetoPoint['gate'] {\n  if (d.promote) return 'promote'\n  if (d.rejectionCode === 'few_runs') return 'reject_few_runs'\n  if (d.rejectionCode === 'negative_delta') return 'reject_negative_delta'\n  if (d.rejectionCode === 'overfit_gap') return 'reject_overfit_gap'\n  return null\n}\n\n// ── gainHistogram ───────────────────────────────────────────\n\nexport interface GainDistributionBin {\n  /** Inclusive lower edge. */\n  lo: number\n  /** Exclusive upper edge (or inclusive if it's the last bin). */\n  hi: number\n  /** Number of pairs whose delta lands in this bin. */\n  count: number\n}\n\nexport interface GainDistributionFigureSpec {\n  kind: 'gain-distribution'\n  candidateId: string\n  comparator: string\n  split: 'search' | 'holdout'\n  /** Number of pairs used. */\n  n: number\n  bins: GainDistributionBin[]\n  median: number\n  ci: { low: number; high: number }\n}\n\nexport interface GainDistributionOptions {\n  /** Number of histogram bins. Default 11 (so the centre is exact at 0). */\n  bins?: number\n  /** Which split to use. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the CI. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Deterministic seed. */\n  seed?: number\n}\n\n/**\n * Held-out improvement distribution: per-pair delta (candidate −\n * comparator), histogrammed. Includes the bootstrap CI on the median\n * delta — same primitive the promotion gate uses.\n */\nexport function gainHistogram(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: GainDistributionOptions = {},\n): GainDistributionFigureSpec {\n  const split = opts.split ?? 'holdout'\n  const scoreField = split === 'holdout' ? 'holdoutScore' : 'searchScore'\n  const binCount = opts.bins ?? 11\n  if (binCount < 1) throw new Error('gainHistogram: bins must be ≥ 1')\n\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === split)\n  // pairScoresByKey returns before=baseline-score, after=candidate-score\n  // for each (experimentId, seed) pair where both sides recorded a\n  // valid score on this split. delta = after - before = candidate - baseline.\n  const { before, after } = pairScoresByKey(candidate, baseline, scoreField)\n  const n = before.length\n\n  if (n === 0) {\n    return {\n      kind: 'gain-distribution',\n      candidateId,\n      comparator,\n      split,\n      n: 0,\n      bins: [],\n      median: 0,\n      ci: { low: 0, high: 0 },\n    }\n  }\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const median = medianOfSorted(sortedDeltas)\n  const min = sortedDeltas[0]!\n  const max = sortedDeltas[sortedDeltas.length - 1]!\n\n  // Symmetric bins around the wider of (|min|, |max|) so the chart\n  // visually centres on zero without dropping outliers.\n  const bound = Math.max(Math.abs(min), Math.abs(max), 1e-6)\n  const lo = -bound\n  const hi = bound\n  const width = (hi - lo) / binCount\n  const bins: GainDistributionBin[] = []\n  for (let i = 0; i < binCount; i++) {\n    bins.push({ lo: lo + i * width, hi: lo + (i + 1) * width, count: 0 })\n  }\n  for (const d of deltas) {\n    let idx = Math.floor((d - lo) / width)\n    if (idx < 0) idx = 0\n    if (idx >= binCount) idx = binCount - 1\n    bins[idx]!.count += 1\n  }\n\n  const ci = pairedBootstrap(before, after, {\n    confidence: opts.confidence ?? 0.95,\n    resamples: opts.resamples ?? 2000,\n    statistic: 'median',\n    seed: opts.seed,\n  })\n\n  return {\n    kind: 'gain-distribution',\n    candidateId,\n    comparator,\n    split,\n    n,\n    bins,\n    median,\n    ci: { low: ci.low, high: ci.high },\n  }\n}\n\n// ── researchReport ───────────────────────────────────────────────────\n\nexport type ResearchReportDecision =\n  | 'promote'\n  | 'hold'\n  | 'reject'\n  | 'equivalent'\n  | 'needs_more_data'\n\n/**\n * Hard floor below which a paired comparison is treated as uninformative\n * regardless of `minPairs`. Mirrors the lower limit on Wilcoxon signed-rank\n * exact tables; below this the test has no power to separate effect sizes.\n */\nexport const RESEARCH_REPORT_HARD_PAIR_FLOOR = 6\n\nexport interface ResearchReportOptions {\n  /** Human-readable report title. */\n  title?: string\n  /** Comparator candidate id. Required for statistical decision guidance. */\n  comparator?: string\n  /** Which split to use for the primary decision. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level used by lower-level report helpers. Default 0.95. */\n  confidence?: number\n  /** FDR threshold for q-values. Default 0.05. */\n  fdr?: number\n  /**\n   * Soft floor on paired observations before issuing a directional\n   * promote / reject. Below this we report `needs_more_data` and surface the\n   * minimum detectable effect at the current N. Default 20 — chosen so the\n   * Wilcoxon signed-rank approximation is reasonable and so the paired\n   * bootstrap CI has non-degenerate coverage. Hard floor is enforced at\n   * `RESEARCH_REPORT_HARD_PAIR_FLOOR` (6) regardless of this value.\n   */\n  minPairs?: number\n  /**\n   * Region of Practical Equivalence on the paired delta. When a candidate's\n   * paired-delta CI is fully contained in `[low, high]`, the decision is\n   * `equivalent` rather than `hold`. Sourced from the domain owner — there is\n   * no statistically-defensible default.\n   */\n  rope?: { low: number; high: number }\n  /**\n   * Power for the minimum detectable effect (MDE) reported on each candidate.\n   * Default 0.8.\n   */\n  mdePower?: number\n  /**\n   * Two-sided alpha for the MDE. Default matches `fdr` so the reported MDE\n   * lines up with the test the report actually runs.\n   */\n  mdeAlpha?: number\n  /** Optional held-out gate decisions keyed by candidate id. */\n  gateDecisions?: Record<string, GateDecision>\n  /** Optional failure clusters from failureClusterView. */\n  failureClusters?: FailureClusterReport\n  /** Build gain histograms for these candidates. Defaults to all non-comparator candidates. */\n  candidateIds?: string[]\n  /** Deterministic bootstrap seed passed to gainHistogram and the posterior helper. */\n  seed?: number\n  /** Report timestamp. Defaults to current time. */\n  generatedAt?: string\n  /**\n   * Hash of a preregistered protocol (e.g. `signManifest({...}).contentHash`).\n   * Embedded verbatim in the report so the analysis can be cited as the\n   * preregistered one rather than a post-hoc fishing expedition.\n   */\n  preregistrationHash?: string\n}\n\nexport interface ResearchReportRecommendation {\n  decision: ResearchReportDecision\n  candidateId: string | null\n  rationale: string[]\n  risks: string[]\n  nextActions: string[]\n}\n\nexport interface ResearchReportCandidate {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  qValue: number\n  cohensD: number\n  meanDeltaVsComparator: number | null\n  pairedN: number\n  medianGain: number | null\n  meanGain: number | null\n  gainCi: { low: number; high: number } | null\n  /**\n   * Bayesian-bootstrap-style posterior summaries on the paired delta. Computed\n   * from the same resamples that produce the gain CI; interpretable as\n   * \"fraction of resamples in which the candidate beats the comparator on\n   * matched pairs.\"\n   */\n  prGreaterThanZero: number | null\n  prInRope: number | null\n  /**\n   * Minimum detectable effect (in score units) at the candidate's paired N,\n   * the configured power, and the configured alpha. Standardised by the\n   * observed paired-delta SD and inverted via `requiredSampleSize`. Reported\n   * for every candidate so a `needs_more_data` verdict is actionable.\n   */\n  mde: number | null\n  onParetoFrontier: boolean\n  gate?: ParetoPoint['gate']\n  decision: ResearchReportDecision\n  decisionReason: string\n}\n\nexport interface ResearchReportMethodology {\n  /**\n   * Plain-language assumptions the report depends on. Read these first when\n   * deciding whether the verdict is load-bearing for a launch decision.\n   */\n  assumptions: string[]\n  /** Tests and estimators the verdict was computed from. */\n  methods: string[]\n  /** Alternatives the author considered and why this report didn't take them. */\n  alternatives: string[]\n  /** Failure modes — when this report should NOT drive a decision. */\n  whenNotToApply: string[]\n  /** Citations for the methodological choices above. */\n  citations: string[]\n}\n\nexport interface ResearchReport {\n  kind: 'agent-eval-research-report'\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  /**\n   * SHA-256 over the canonicalised set of `(runId, candidateId, split)` triples\n   * the report was computed from, plus the comparator and split. Stable across\n   * key insertion order; recomputable by the reader to verify provenance.\n   */\n  runFingerprint: string\n  preregistrationHash: string | null\n  rope: { low: number; high: number } | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  charts: {\n    pareto: ParetoFigureSpec\n    gains: GainDistributionFigureSpec[]\n  }\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n  markdown: string\n  html: string\n}\n\n/**\n * Internal: paired posterior summary on (candidate − comparator) deltas.\n *\n * Returns the bootstrap CI on the median (matching `gainHistogram`) plus\n * Bayesian-flavoured posterior summaries Pr(Δ>0) and Pr(Δ∈ROPE) computed\n * from a Bayesian-bootstrap-flavoured resample distribution on the mean\n * (Rubin 1981 — non-informative bootstrap-prior duality), and the\n * minimum detectable paired effect at the configured power and α.\n *\n * `null` is returned when no paired observations exist; callers must\n * gate on `n` before consuming the bootstrap statistics.\n */\nfunction pairedPosterior(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: {\n    split: 'search' | 'holdout'\n    confidence: number\n    seed?: number\n    rope: { low: number; high: number } | null\n    mdePower: number\n    mdeAlpha: number\n  },\n): {\n  n: number\n  meanDelta: number\n  medianDelta: number\n  sdDelta: number\n  ci: { low: number; high: number }\n  prGreaterThanZero: number\n  prInRope: number | null\n  mde: number\n} | null {\n  const scoreField = opts.split === 'holdout' ? 'holdoutScore' : 'searchScore'\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === opts.split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === opts.split)\n  const { before, after } = pairScoresByKey(candidate, baseline, scoreField)\n  const n = before.length\n  if (n === 0) return null\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const meanDelta = deltas.reduce((s, x) => s + x, 0) / n\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const medianDelta = medianOfSorted(sortedDeltas)\n  const sdDelta = stdev(deltas, meanDelta)\n\n  const ci = pairedBootstrap(before, after, {\n    confidence: opts.confidence,\n    resamples: 2000,\n    statistic: 'median',\n    seed: opts.seed,\n  })\n\n  // Enumerate bootstrap-mean samples to derive posterior summaries on the\n  // mean delta. Same RNG family as `pairedBootstrap` but kept local so we can\n  // examine the full sample distribution rather than just quantiles.\n  const meanSamples = bootstrapMeanSamples(deltas, 2000, opts.seed)\n  const prGreaterThanZero =\n    meanSamples.length === 0 ? 0 : meanSamples.filter((s) => s > 0).length / meanSamples.length\n  const prInRope =\n    opts.rope === null || meanSamples.length === 0\n      ? null\n      : meanSamples.filter((s) => s >= opts.rope!.low && s <= opts.rope!.high).length /\n        meanSamples.length\n\n  const dStandardised = pairedMde({ nPaired: n, alpha: opts.mdeAlpha, power: opts.mdePower })\n  const mde = sdDelta === 0 ? 0 : dStandardised * sdDelta\n\n  return {\n    n,\n    meanDelta,\n    medianDelta,\n    sdDelta,\n    ci: { low: ci.low, high: ci.high },\n    prGreaterThanZero,\n    prInRope,\n    mde,\n  }\n}\n\nfunction bootstrapMeanSamples(deltas: number[], resamples: number, seed?: number): number[] {\n  const n = deltas.length\n  if (n === 0) return []\n  if (n === 1) return new Array<number>(resamples).fill(deltas[0]!)\n  const rng = seedRng(seed)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let sum = 0\n    for (let k = 0; k < n; k++) sum += deltas[Math.floor(rng() * n)]!\n    samples[b] = sum / n\n  }\n  return samples\n}\n\nfunction seedRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\nfunction stdev(xs: number[], mean: number): number {\n  if (xs.length < 2) return 0\n  let sse = 0\n  for (const x of xs) sse += (x - mean) ** 2\n  return Math.sqrt(sse / (xs.length - 1))\n}\n\n/**\n * Executive research report for CPO / AI-lead / launch-review consumption.\n *\n * Composes:\n *   - `summaryTable`         marginal stats with BH-FDR-adjusted q-values\n *   - `paretoChart`           cost-vs-quality frontier with gate overlay\n *   - `gainHistogram`         per-candidate paired-delta distribution\n *   - paired posterior (this file): bootstrap CI on median, Pr(Δ>0),\n *                              Pr(Δ∈ROPE), MDE at the configured power\n *\n * Decisions are made on paired evidence — never on marginal means alone —\n * and respect any held-out gate decision the caller passes through. The\n * report embeds a SHA-256 fingerprint of the input run set and, optionally,\n * the hash of a preregistered protocol so a downstream reader can verify\n * provenance and that the analysis was the preregistered one.\n *\n * Async because the fingerprint uses Web Crypto via `hashJson`; deterministic\n * for any fixed `runs`, `seed`, and ROPE.\n */\nexport async function researchReport(\n  runs: RunRecord[],\n  opts: ResearchReportOptions = {},\n): Promise<ResearchReport> {\n  const split = opts.split ?? 'holdout'\n  const comparator = opts.comparator ?? null\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const minPairs = Math.max(opts.minPairs ?? 20, RESEARCH_REPORT_HARD_PAIR_FLOOR)\n  const rope = opts.rope ?? null\n  const mdePower = opts.mdePower ?? 0.8\n  const mdeAlpha = opts.mdeAlpha ?? fdr\n  const title = opts.title ?? 'Agent Evaluation Research Report'\n  const generatedAt = opts.generatedAt ?? new Date().toISOString()\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  if (rope && !(Number.isFinite(rope.low) && Number.isFinite(rope.high) && rope.low <= rope.high)) {\n    throw new Error(\n      `researchReport: rope must satisfy low ≤ high with finite bounds, got ${JSON.stringify(rope)}`,\n    )\n  }\n\n  const summary = summaryTable(runs, {\n    comparator: comparator ?? undefined,\n    split,\n    confidence,\n    fdr,\n  })\n  const pareto = paretoChart(runs, { split, gateDecisions: opts.gateDecisions })\n  const candidateIds =\n    opts.candidateIds ?? summary.rows.map((r) => r.candidateId).filter((id) => id !== comparator)\n  const gains = comparator\n    ? candidateIds.map((id) =>\n        gainHistogram(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n        }),\n      )\n    : []\n\n  const gainByCandidate = new Map(gains.map((g) => [g.candidateId, g]))\n  const paretoByCandidate = new Map(pareto.points.map((p) => [p.candidateId, p]))\n  const posteriorByCandidate = new Map<string, ReturnType<typeof pairedPosterior>>()\n  if (comparator) {\n    for (const id of candidateIds) {\n      posteriorByCandidate.set(\n        id,\n        pairedPosterior(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n          rope,\n          mdePower,\n          mdeAlpha,\n        }),\n      )\n    }\n  }\n\n  const candidates = summary.rows\n    .map((row) => {\n      const gain = gainByCandidate.get(row.candidateId)\n      const point = paretoByCandidate.get(row.candidateId)\n      const posterior = posteriorByCandidate.get(row.candidateId) ?? null\n      const classified = classifyCandidate(row, {\n        comparator,\n        posterior,\n        point,\n        fdr,\n        minPairs,\n        rope,\n      })\n      return {\n        candidateId: row.candidateId,\n        n: row.n,\n        mean: row.mean,\n        ciLow: row.ciLow,\n        ciHigh: row.ciHigh,\n        qValue: row.qValue,\n        cohensD: row.cohensD,\n        meanDeltaVsComparator: posterior ? posterior.meanDelta : null,\n        pairedN: posterior?.n ?? gain?.n ?? 0,\n        medianGain: posterior ? posterior.medianDelta : gain ? gain.median : null,\n        meanGain: posterior ? posterior.meanDelta : null,\n        gainCi: posterior ? posterior.ci : gain ? gain.ci : null,\n        prGreaterThanZero: posterior ? posterior.prGreaterThanZero : null,\n        prInRope: posterior ? posterior.prInRope : null,\n        mde: posterior ? posterior.mde : null,\n        onParetoFrontier: point?.onFrontier ?? false,\n        gate: point?.gate,\n        decision: classified.decision,\n        decisionReason: classified.reason,\n      } satisfies ResearchReportCandidate\n    })\n    .sort((a, b) => {\n      const decisionRank = decisionWeight(b.decision) - decisionWeight(a.decision)\n      if (decisionRank !== 0) return decisionRank\n      return b.mean - a.mean\n    })\n\n  const recommendation = buildRecommendation(candidates, {\n    comparator,\n    failureClusters: opts.failureClusters,\n    rope,\n    minPairs,\n    preregistrationHash,\n  })\n  const executiveSummary = buildExecutiveSummary(candidates, recommendation, {\n    comparator,\n    split,\n    failureClusters: opts.failureClusters,\n    preregistrationHash,\n  })\n  const methodology = buildMethodology({\n    split,\n    comparator,\n    fdr,\n    minPairs,\n    rope,\n    confidence,\n    mdePower,\n    mdeAlpha,\n  })\n\n  const runFingerprint = await hashJson(\n    canonicalize({\n      triples: runs\n        .filter((r) => r.splitTag === split)\n        .map((r) => ({ runId: r.runId, candidateId: r.candidateId, splitTag: r.splitTag }))\n        .sort((a, b) => a.runId.localeCompare(b.runId)),\n      comparator,\n      split,\n    }),\n  )\n\n  const markdown = renderResearchMarkdown({\n    title,\n    generatedAt,\n    split,\n    comparator,\n    rope,\n    runFingerprint,\n    preregistrationHash,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    pareto,\n    gains,\n    methodology,\n    failureClusters: opts.failureClusters,\n  })\n  const html = renderResearchHtml(markdown, title)\n\n  return {\n    kind: 'agent-eval-research-report',\n    title,\n    generatedAt,\n    split,\n    comparator,\n    runFingerprint,\n    preregistrationHash,\n    rope,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    charts: { pareto, gains },\n    methodology,\n    failureClusters: opts.failureClusters,\n    markdown,\n    html,\n  }\n}\n\nfunction buildMethodology(ctx: {\n  split: 'search' | 'holdout'\n  comparator: string | null\n  fdr: number\n  minPairs: number\n  rope: { low: number; high: number } | null\n  confidence: number\n  mdePower: number\n  mdeAlpha: number\n}): ResearchReportMethodology {\n  const assumptions: string[] = [\n    'Pairs are matched by (experimentId, seed); the candidate and comparator see the same scenarios in the same order.',\n    'Paired deltas are exchangeable conditional on the matched scenario — no mid-run distribution shift.',\n    `Decisions are pre-specified at fdr=${ctx.fdr}, minPairs=${ctx.minPairs}, confidence=${ctx.confidence}; deviating from these post-hoc invalidates the false-discovery control.`,\n  ]\n  if (ctx.rope) {\n    assumptions.push(\n      `The Region of Practical Equivalence ${formatRope(ctx.rope)} is supplied by the domain owner; equivalent verdicts are only meaningful if that range is treated as the standing definition of \"no material difference.\"`,\n    )\n  }\n  if (ctx.comparator === null) {\n    assumptions.push('No comparator was configured; this run is descriptive, not causal.')\n  }\n  const methods: string[] = [\n    \"Marginal scores summarised with BH-FDR-adjusted Wilcoxon signed-rank q-values and Cohen's d via summaryTable.\",\n    'Paired evidence summarised with bootstrap CI on the median delta and Bayesian-bootstrap-style Pr(Δ>0) and Pr(Δ∈ROPE) on the mean delta.',\n    `Minimum detectable effect reported per candidate at α=${ctx.mdeAlpha} (two-sided), power=${ctx.mdePower}, standardised by the observed paired-delta SD.`,\n    'Pareto frontier flagged as a separate axis (cost vs quality); a candidate can be on-frontier without winning the paired test.',\n    'Held-out gate decisions, when supplied, override the statistical verdict in the reject direction.',\n  ]\n  const alternatives: string[] = [\n    'Paired t-test rejected: not robust to the heavy-tailed score distributions common in agent benchmarks.',\n    'Unpaired Mann–Whitney rejected: matched scenarios make pairing free; unpaired throws away that variance reduction.',\n    'Sequential / always-valid inference (e-values, mSPRT) is the right tool for iterative sweeps and is out of scope for this single-look report — preregister and run once, or wrap this report in an alpha-spending schedule.',\n    'Hierarchical Bayesian shrinkage across many candidates is future work; the current ranking uses raw paired statistics.',\n  ]\n  const whenNotToApply: string[] = [\n    `Paired N below ${RESEARCH_REPORT_HARD_PAIR_FLOOR} on any candidate — the bootstrap CI is degenerate.`,\n    'Comparator chosen post-hoc by inspecting the same data; q-values are no longer false-discovery-controlled.',\n    'Scenarios not drawn under a stable preregistered protocol; the report can describe the data but cannot anchor a launch decision.',\n    'Score distributions with mid-run shift (judge model swap, rubric change, infra outage) — pair exchangeability is violated.',\n  ]\n  const citations: string[] = [\n    'Benjamini, Y. & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. JRSS B, 57(1), 289–300.',\n    'Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1(6), 80–83.',\n    'Efron, B. (1979). Bootstrap methods: another look at the jackknife. Annals of Statistics, 7(1), 1–26.',\n    'Rubin, D. B. (1981). The Bayesian bootstrap. Annals of Statistics, 9(1), 130–134.',\n    'Kruschke, J. K. (2018). Rejecting or accepting parameter values in Bayesian estimation. Advances in Methods and Practices in Psychological Science, 1(2), 270–280. (ROPE.)',\n  ]\n  return { assumptions, methods, alternatives, whenNotToApply, citations }\n}\n\nfunction formatRope(rope: { low: number; high: number }): string {\n  return `[${fmt(rope.low)}, ${fmt(rope.high)}]`\n}\n\nfunction classifyCandidate(\n  row: SummaryTableRow,\n  ctx: {\n    comparator: string | null\n    posterior: ReturnType<typeof pairedPosterior> | null\n    point?: ParetoPoint\n    fdr: number\n    minPairs: number\n    rope: { low: number; high: number } | null\n  },\n): { decision: ResearchReportDecision; reason: string } {\n  if (ctx.comparator && row.candidateId === ctx.comparator) {\n    return { decision: 'hold', reason: 'Comparator baseline.' }\n  }\n  if (!ctx.comparator) {\n    return {\n      decision: ctx.point?.onFrontier ? 'hold' : 'needs_more_data',\n      reason:\n        'No comparator configured; report ranks candidates but cannot anchor a promotion call.',\n    }\n  }\n  // Held-out gate is authoritative against — promote requires statistical\n  // evidence even if the gate said `promote` (gate is necessary, not sufficient).\n  if (ctx.point?.gate && ctx.point.gate !== 'promote') {\n    return { decision: 'reject', reason: `Held-out gate returned ${ctx.point.gate}.` }\n  }\n  if (!ctx.posterior || ctx.posterior.n < RESEARCH_REPORT_HARD_PAIR_FLOOR) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior?.n ?? 0} paired observations; below hard floor of ${RESEARCH_REPORT_HARD_PAIR_FLOOR} for any paired inference.`,\n    }\n  }\n  const ci = ctx.posterior.ci\n  if (ctx.rope && ci.low >= ctx.rope.low && ci.high <= ctx.rope.high) {\n    return {\n      decision: 'equivalent',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] is fully inside ROPE ${formatRope(ctx.rope)}; candidate is practically equivalent to comparator.`,\n    }\n  }\n  const significant = Number.isFinite(row.qValue) && row.qValue <= ctx.fdr\n  const gainPositive = ci.low > 0\n  const gainNegative = ci.high < 0\n  if (gainNegative) {\n    return {\n      decision: 'reject',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] lies entirely below zero.`,\n    }\n  }\n  if (ctx.posterior.n < ctx.minPairs) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior.n} paired observations; minimum detectable effect at this N is ${fmt(ctx.posterior.mde)} score units (need ≥ ${ctx.minPairs} pairs to issue a directional verdict).`,\n    }\n  }\n  if (significant && gainPositive) {\n    return {\n      decision: 'promote',\n      reason: `BH-adjusted q=${fmt(row.qValue)} ≤ ${ctx.fdr} and paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] excludes zero; Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)}.`,\n    }\n  }\n  return {\n    decision: 'hold',\n    reason: `Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)} but CI [${fmt(ci.low)}, ${fmt(ci.high)}] crosses zero; effect not decisive at fdr=${ctx.fdr}.`,\n  }\n}\n\nfunction buildRecommendation(\n  candidates: ResearchReportCandidate[],\n  ctx: {\n    comparator: string | null\n    failureClusters?: FailureClusterReport\n    rope: { low: number; high: number } | null\n    minPairs: number\n    preregistrationHash: string | null\n  },\n): ResearchReportRecommendation {\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  const bestPromote = nonComparator.find((c) => c.decision === 'promote')\n  const bestEquivalent = nonComparator.find((c) => c.decision === 'equivalent')\n  const chosen = bestPromote ?? bestEquivalent ?? nonComparator[0] ?? null\n  const decision: ResearchReportDecision = bestPromote\n    ? 'promote'\n    : nonComparator.some((c) => c.decision === 'needs_more_data')\n      ? 'needs_more_data'\n      : bestEquivalent\n        ? 'equivalent'\n        : nonComparator.some((c) => c.decision === 'hold')\n          ? 'hold'\n          : 'reject'\n\n  const rationale: string[] = []\n  const risks: string[] = []\n  const nextActions: string[] = []\n\n  if (chosen) {\n    rationale.push(`${chosen.candidateId}: ${chosen.decisionReason}`)\n    if (chosen.gainCi) {\n      const probSummary =\n        chosen.prGreaterThanZero !== null ? `, Pr(Δ>0)=${fmt(chosen.prGreaterThanZero)}` : ''\n      rationale.push(\n        `Median paired gain CI: [${fmt(chosen.gainCi.low)}, ${fmt(chosen.gainCi.high)}]${probSummary}.`,\n      )\n    }\n    if (chosen.mde !== null && Number.isFinite(chosen.mde)) {\n      rationale.push(`MDE at current paired N=${chosen.pairedN}: ${fmt(chosen.mde)} score units.`)\n    }\n  }\n  if (!ctx.comparator) {\n    risks.push('No comparator was configured; verdict is descriptive, not causal.')\n    nextActions.push('Re-run with a stable comparator candidate for paired inference.')\n  }\n  if (!ctx.preregistrationHash) {\n    risks.push(\n      'No preregistration hash supplied; readers cannot verify the analysis was specified before data inspection.',\n    )\n    nextActions.push(\n      'Sign a HypothesisManifest before the next sweep and pass `preregistrationHash` so the report cites it.',\n    )\n  }\n  if (ctx.rope === null && nonComparator.length > 0) {\n    risks.push(\n      'No ROPE configured; the report cannot distinguish \"equivalent\" from \"inconclusive\".',\n    )\n    nextActions.push(\n      'Define a domain-specific Region of Practical Equivalence and pass it to lock in the equivalence threshold.',\n    )\n  }\n  const inconclusive = nonComparator.filter((c) => c.decision === 'needs_more_data')\n  if (inconclusive.length > 0) {\n    const worst = inconclusive.reduce((a, b) => (b.pairedN < a.pairedN ? b : a))\n    risks.push(\n      `${inconclusive.length} candidate(s) below soft floor (${ctx.minPairs} pairs); thinnest is ${worst.candidateId} with ${worst.pairedN}.`,\n    )\n    nextActions.push(\n      `Collect at least ${ctx.minPairs - worst.pairedN} more matched holdout runs for ${worst.candidateId}.`,\n    )\n  }\n  const rejected = nonComparator.filter((c) => c.decision === 'reject')\n  if (rejected.length > 0) {\n    risks.push(\n      `${rejected.length} candidate(s) failed the paired test or held-out gate; do not ship those variants.`,\n    )\n  }\n  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {\n    const top = ctx.failureClusters.clusters[0]!\n    risks.push(`Top failure cluster: ${top.failureClass} across ${top.runCount} run(s).`)\n    nextActions.push('Prioritize the largest failure cluster before broad rollout.')\n  }\n  if (decision === 'promote') {\n    nextActions.push('Ship behind the existing promotion gate and monitor canaries.')\n  } else if (decision === 'hold') {\n    nextActions.push('Keep current production candidate while expanding holdout evidence.')\n  } else if (decision === 'equivalent') {\n    nextActions.push(\n      'Either keep the comparator (no quality regression) or promote on cost/latency grounds — equivalence does not justify either; the choice is a product decision, not a stats one.',\n    )\n  } else if (decision === 'reject') {\n    nextActions.push(\n      'Do not promote this sweep; inspect failures and generate a revised candidate.',\n    )\n  }\n\n  return {\n    decision,\n    candidateId: chosen?.candidateId ?? null,\n    rationale,\n    risks,\n    nextActions,\n  }\n}\n\nfunction buildExecutiveSummary(\n  candidates: ResearchReportCandidate[],\n  recommendation: ResearchReportRecommendation,\n  ctx: {\n    comparator: string | null\n    split: 'search' | 'holdout'\n    failureClusters?: FailureClusterReport\n    preregistrationHash: string | null\n  },\n): string[] {\n  const lines: string[] = []\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  lines.push(\n    `Evaluated ${nonComparator.length} candidate(s) on the ${ctx.split} split${ctx.comparator ? ` against ${ctx.comparator}` : ''}.`,\n  )\n  lines.push(\n    `Recommendation: ${recommendation.decision}${recommendation.candidateId ? ` ${recommendation.candidateId}` : ''}.`,\n  )\n  const promoted = nonComparator.filter((c) => c.decision === 'promote').length\n  const held = nonComparator.filter((c) => c.decision === 'hold').length\n  const equivalent = nonComparator.filter((c) => c.decision === 'equivalent').length\n  const rejected = nonComparator.filter((c) => c.decision === 'reject').length\n  const more = nonComparator.filter((c) => c.decision === 'needs_more_data').length\n  lines.push(\n    `Decision mix: ${promoted} promote, ${equivalent} equivalent, ${held} hold, ${rejected} reject, ${more} need more data.`,\n  )\n  const frontier = nonComparator.filter((c) => c.onParetoFrontier).map((c) => c.candidateId)\n  if (frontier.length > 0) lines.push(`Pareto-frontier candidates: ${frontier.join(', ')}.`)\n  if (ctx.failureClusters) {\n    lines.push(\n      `Failure clustering found ${ctx.failureClusters.totalFailures}/${ctx.failureClusters.totalRuns} failed runs across ${ctx.failureClusters.clusters.length} reportable cluster(s).`,\n    )\n  }\n  lines.push(\n    ctx.preregistrationHash\n      ? `Preregistered analysis: ${ctx.preregistrationHash.slice(0, 12)}…`\n      : 'Analysis is post-hoc — no preregistration hash supplied.',\n  )\n  return lines\n}\n\nfunction renderResearchMarkdown(report: {\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  pareto: ParetoFigureSpec\n  gains: GainDistributionFigureSpec[]\n  rope: { low: number; high: number } | null\n  runFingerprint: string\n  preregistrationHash: string | null\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n}): string {\n  const lines: string[] = []\n  lines.push(`# ${report.title}`)\n  lines.push('')\n  lines.push(`**Generated:** ${report.generatedAt}`)\n  lines.push(`**Primary split:** ${report.split}`)\n  lines.push(`**Comparator:** ${report.comparator ?? 'not configured'}`)\n  lines.push(`**ROPE:** ${report.rope ? formatRope(report.rope) : 'not configured'}`)\n  lines.push(`**Run fingerprint:** \\`${report.runFingerprint}\\``)\n  lines.push(\n    `**Preregistration:** ${report.preregistrationHash ? `\\`${report.preregistrationHash}\\`` : 'none'}`,\n  )\n  lines.push('')\n  lines.push('## Executive Summary')\n  lines.push('')\n  for (const item of report.executiveSummary) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Recommendation')\n  lines.push('')\n  lines.push(`**Decision:** ${report.recommendation.decision}`)\n  lines.push(`**Candidate:** ${report.recommendation.candidateId ?? 'N/A'}`)\n  lines.push('')\n  lines.push('### Rationale')\n  lines.push('')\n  for (const item of report.recommendation.rationale) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Risks')\n  lines.push('')\n  for (const item of report.recommendation.risks.length\n    ? report.recommendation.risks\n    : ['No material report-level risks detected.']) {\n    lines.push(`- ${item}`)\n  }\n  lines.push('')\n  lines.push('### Next Actions')\n  lines.push('')\n  for (const item of report.recommendation.nextActions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Candidate Decision Table')\n  lines.push('')\n  lines.push(\n    '| Candidate | Decision | Mean | Δ̄ | Pr(Δ>0) | q | d | Paired N | Median Gain CI | MDE | Pareto | Gate |',\n  )\n  lines.push('|---|---|---:|---:|---:|---:|---:|---:|---|---:|---|---|')\n  for (const c of report.candidates) {\n    const delta = c.meanDeltaVsComparator === null ? '-' : signed(c.meanDeltaVsComparator)\n    const prGt = c.prGreaterThanZero === null ? '-' : c.prGreaterThanZero.toFixed(3)\n    const q = Number.isFinite(c.qValue) ? c.qValue.toFixed(4) : '-'\n    const d = Number.isFinite(c.cohensD) ? c.cohensD.toFixed(3) : '-'\n    const gain = c.gainCi ? `[${fmt(c.gainCi.low)}, ${fmt(c.gainCi.high)}]` : '-'\n    const mde = c.mde === null || !Number.isFinite(c.mde) ? '-' : fmt(c.mde)\n    lines.push(\n      `| ${c.candidateId} | ${c.decision} | ${fmt(c.mean)} | ${delta} | ${prGt} | ${q} | ${d} | ${c.pairedN} | ${gain} | ${mde} | ${c.onParetoFrontier ? 'yes' : 'no'} | ${c.gate ?? '-'} |`,\n    )\n  }\n  lines.push('')\n  lines.push('## Statistical Summary')\n  lines.push('')\n  lines.push(report.summary.markdown)\n  lines.push('')\n  lines.push('## Methodology')\n  lines.push('')\n  lines.push('### Assumptions')\n  lines.push('')\n  for (const item of report.methodology.assumptions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Methods')\n  lines.push('')\n  for (const item of report.methodology.methods) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Alternatives Considered')\n  lines.push('')\n  for (const item of report.methodology.alternatives) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### When NOT To Apply')\n  lines.push('')\n  for (const item of report.methodology.whenNotToApply) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Citations')\n  lines.push('')\n  for (const item of report.methodology.citations) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Chart Specs')\n  lines.push('')\n  lines.push(\n    'The report carries JSON chart specs for Pareto cost/quality and paired gain histograms.',\n  )\n  lines.push('')\n  lines.push('```json')\n  lines.push(JSON.stringify({ pareto: report.pareto, gains: report.gains }, null, 2))\n  lines.push('```')\n  if (report.failureClusters) {\n    lines.push('')\n    lines.push('## Failure Clusters')\n    lines.push('')\n    lines.push('| Failure Class | Runs | Scenarios | Tool | Example |')\n    lines.push('|---|---:|---:|---|---|')\n    for (const c of report.failureClusters.clusters.slice(0, 10)) {\n      lines.push(\n        `| ${c.failureClass} | ${c.runCount} | ${c.scenarioIds.length} | ${c.toolName ?? '-'} | ${escapePipes(c.exampleError ?? c.exampleRunId)} |`,\n      )\n    }\n  }\n  return lines.join('\\n')\n}\n\nfunction renderResearchHtml(markdown: string, title: string): string {\n  const body = markdownToHtml(markdown)\n  return [\n    '<!doctype html>',\n    '<html lang=\"en\">',\n    '<head>',\n    '<meta charset=\"utf-8\">',\n    '<meta name=\"viewport\" content=\"width=device-width, initial-scale=1\">',\n    `<title>${escapeHtml(title)}</title>`,\n    '<style>',\n    'body{font-family:Inter,ui-sans-serif,system-ui,-apple-system,BlinkMacSystemFont,\"Segoe UI\",sans-serif;margin:0;color:#172026;background:#f7f8f8;}',\n    'main{max-width:1080px;margin:0 auto;padding:40px 24px 64px;background:#fff;min-height:100vh;}',\n    'h1{font-size:34px;line-height:1.15;margin:0 0 20px;}h2{margin-top:34px;border-top:1px solid #d9dfdf;padding-top:22px;}h3{margin-top:22px;}',\n    'p,li{line-height:1.55;}table{border-collapse:collapse;width:100%;margin:16px 0;font-size:14px;}th,td{border:1px solid #d9dfdf;padding:8px;text-align:left;}th{background:#eef2f2;}',\n    'code,pre{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;}pre{overflow:auto;background:#111827;color:#f9fafb;padding:16px;border-radius:6px;}',\n    '</style>',\n    '</head>',\n    '<body><main>',\n    body,\n    '</main></body></html>',\n  ].join('\\n')\n}\n\nfunction markdownToHtml(markdown: string): string {\n  const lines = markdown.split('\\n')\n  const html: string[] = []\n  let inList = false\n  let inCode = false\n  let code: string[] = []\n  let table: string[] = []\n\n  const flushList = () => {\n    if (inList) {\n      html.push('</ul>')\n      inList = false\n    }\n  }\n  const flushTable = () => {\n    if (table.length === 0) return\n    html.push(renderMarkdownTable(table))\n    table = []\n  }\n\n  for (const line of lines) {\n    if (line.startsWith('```')) {\n      if (inCode) {\n        html.push(`<pre><code>${escapeHtml(code.join('\\n'))}</code></pre>`)\n        code = []\n        inCode = false\n      } else {\n        flushList()\n        flushTable()\n        inCode = true\n      }\n      continue\n    }\n    if (inCode) {\n      code.push(line)\n      continue\n    }\n    if (line.startsWith('|')) {\n      flushList()\n      table.push(line)\n      continue\n    }\n    flushTable()\n    if (line.startsWith('- ')) {\n      if (!inList) {\n        html.push('<ul>')\n        inList = true\n      }\n      html.push(`<li>${inlineMarkdown(line.slice(2))}</li>`)\n      continue\n    }\n    flushList()\n    if (line.startsWith('# ')) html.push(`<h1>${inlineMarkdown(line.slice(2))}</h1>`)\n    else if (line.startsWith('## ')) html.push(`<h2>${inlineMarkdown(line.slice(3))}</h2>`)\n    else if (line.startsWith('### ')) html.push(`<h3>${inlineMarkdown(line.slice(4))}</h3>`)\n    else if (line.trim() === '') html.push('')\n    else html.push(`<p>${inlineMarkdown(line)}</p>`)\n  }\n  flushList()\n  flushTable()\n  return html.join('\\n')\n}\n\nfunction renderMarkdownTable(lines: string[]): string {\n  const rows = lines\n    .filter((line) => !/^\\|[-:\\s|]+\\|$/.test(line))\n    .map((line) =>\n      line\n        .slice(1, -1)\n        .split('|')\n        .map((cell) => inlineMarkdown(cell.trim())),\n    )\n  if (rows.length === 0) return ''\n  const [head, ...body] = rows\n  const th = head!.map((cell) => `<th>${cell}</th>`).join('')\n  const trs = body\n    .map((row) => `<tr>${row.map((cell) => `<td>${cell}</td>`).join('')}</tr>`)\n    .join('\\n')\n  return `<table><thead><tr>${th}</tr></thead><tbody>${trs}</tbody></table>`\n}\n\nfunction inlineMarkdown(s: string): string {\n  return escapeHtml(s).replace(/\\*\\*([^*]+)\\*\\*/g, '<strong>$1</strong>')\n}\n\nfunction escapeHtml(s: string): string {\n  return s\n    .replace(/&/g, '&amp;')\n    .replace(/</g, '&lt;')\n    .replace(/>/g, '&gt;')\n    .replace(/\"/g, '&quot;')\n}\n\nfunction escapePipes(s: string): string {\n  return s.replace(/\\|/g, '\\\\|')\n}\n\nfunction decisionWeight(decision: ResearchReportDecision): number {\n  if (decision === 'promote') return 5\n  if (decision === 'equivalent') return 4\n  if (decision === 'hold') return 3\n  if (decision === 'needs_more_data') return 2\n  return 1\n}\n\nfunction signed(x: number): string {\n  return `${x >= 0 ? '+' : ''}${fmt(x)}`\n}\n\n// ── tiny helpers ─────────────────────────────────────────────────────\n\nfunction avg(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 medianOfSorted(sorted: number[]): number {\n  if (sorted.length === 0) return 0\n  const mid = Math.floor(sorted.length / 2)\n  return sorted.length % 2 === 0 ? (sorted[mid - 1]! + sorted[mid]!) / 2 : sorted[mid]!\n}\n\nfunction fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n"],"mappings":";;;;;;;;;;;;;;AA6EO,SAAS,aAAa,MAAmB,OAA4B,CAAC,GAAiB;AAC5F,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,aAAa,UAAU,YAAY,iBAAiB;AAE1D,QAAM,cAAc,oBAAI,IAAqD;AAC7E,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,aAAa,MAAO;AAC1B,UAAM,IAAI,EAAE,QAAQ,UAAU;AAC9B,QAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,EAAG;AAClD,UAAM,SAAS,YAAY,IAAI,EAAE,WAAW,KAAK,EAAE,MAAM,CAAC,GAAG,QAAQ,CAAC,EAAE;AACxE,WAAO,KAAK,KAAK,CAAC;AAClB,WAAO,OAAO,KAAK,CAAC;AACpB,gBAAY,IAAI,EAAE,aAAa,MAAM;AAAA,EACvC;AAEA,QAAM,eAAe,CAAC,GAAG,YAAY,KAAK,CAAC,EAAE,KAAK;AAClD,QAAM,WAAW,aAAa,YAAY,IAAI,UAAU,IAAI;AAG5D,QAAM,YAAuD,CAAC;AAC9D,aAAW,MAAM,cAAc;AAC7B,UAAM,SAAS,YAAY,IAAI,EAAE;AACjC,UAAM,KAAK,mBAAmB,OAAO,QAAQ,UAAU;AACvD,QAAI,OAAO,OAAO;AAClB,QAAI,IAAI,OAAO;AACf,QAAI,cAAc,YAAY,OAAO,YAAY;AAC/C,YAAM,SAAS,gBAAgB,OAAO,MAAM,SAAS,MAAM,UAAU;AACrE,UAAI,OAAO,OAAO,UAAU,GAAG;AAC7B,eAAO,mBAAmB,OAAO,QAAQ,OAAO,KAAK,EAAE;AAAA,MACzD;AACA,UAAI,QAAQ,SAAS,QAAQ,OAAO,MAAM;AAAA,IAC5C;AACA,cAAU,KAAK;AAAA,MACb,aAAa;AAAA,MACb,GAAG,OAAO,OAAO;AAAA,MACjB,MAAM,GAAG;AAAA,MACT,OAAO,GAAG;AAAA,MACV,QAAQ,GAAG;AAAA,MACX,QAAQ;AAAA,MACR,SAAS;AAAA,MACT;AAAA,IACF,CAAC;AAAA,EACH;AAKA,MAAI,YAAY;AACd,UAAM,OAAiB,CAAC;AACxB,UAAM,KAAe,CAAC;AACtB,aAAS,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AACzC,YAAM,IAAI,UAAU,CAAC;AACrB,UAAI,EAAE,gBAAgB,WAAY;AAClC,UAAI,CAAC,OAAO,SAAS,EAAE,IAAI,EAAG;AAC9B,WAAK,KAAK,CAAC;AACX,SAAG,KAAK,EAAE,IAAI;AAAA,IAChB;AACA,QAAI,GAAG,SAAS,GAAG;AACjB,YAAM,EAAE,QAAQ,IAAI,kBAAkB,IAAI,GAAG;AAC7C,eAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,kBAAU,KAAK,CAAC,CAAE,EAAG,SAAS,QAAQ,CAAC;AAAA,MACzC;AAAA,IACF;AAAA,EACF;AAEA,QAAM,OAAO,UAAU,IAAI,CAAC,EAAE,MAAM,OAAO,GAAG,KAAK,MAAM,IAAI;AAC7D,QAAM,WAAW,2BAA2B,MAAM,YAAY,KAAK;AACnE,SAAO,EAAE,MAAM,YAAY,OAAO,SAAS;AAC7C;AAEA,SAAS,gBACP,WACA,UACA,YACuC;AACvC,QAAM,UAAU,oBAAI,IAAoB;AACxC,aAAW,KAAK,UAAU;AACxB,UAAM,IAAI,EAAE,QAAQ,UAAU;AAC9B,QAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,GAAG;AAC/C,cAAQ,IAAI,GAAG,EAAE,YAAY,KAAK,EAAE,IAAI,IAAI,CAAC;AAAA,IAC/C;AAAA,EACF;AACA,QAAM,SAAmB,CAAC;AAC1B,QAAM,QAAkB,CAAC;AACzB,aAAW,KAAK,WAAW;AACzB,UAAM,IAAI,EAAE,QAAQ,UAAU;AAC9B,QAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,EAAG;AAClD,UAAM,MAAM,GAAG,EAAE,YAAY,KAAK,EAAE,IAAI;AACxC,UAAM,IAAI,QAAQ,IAAI,GAAG;AACzB,QAAI,MAAM,OAAW;AACrB,WAAO,KAAK,CAAC;AACb,UAAM,KAAK,CAAC;AAAA,EACd;AACA,SAAO,EAAE,QAAQ,MAAM;AACzB;AAEA,SAAS,2BACP,MACA,YACA,OACQ;AACR,QAAM,QAAkB,CAAC;AACzB,QAAM,WAAW,aAAa,QAAQ,UAAU,MAAM;AACtD,QAAM,KAAK,wBAAmB,KAAK,SAAS,QAAQ,EAAE;AACtD,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,wDAAwD;AACnE,QAAM,KAAK,+BAA+B;AAC1C,aAAW,KAAK,MAAM;AACpB,UAAM,KAAK,IAAI,IAAI,EAAE,KAAK,CAAC,KAAK,IAAI,EAAE,MAAM,CAAC;AAC7C,UAAM,IAAI,OAAO,SAAS,EAAE,MAAM,IAAI,EAAE,OAAO,QAAQ,CAAC,IAAI;AAC5D,UAAM,IAAI,OAAO,SAAS,EAAE,OAAO,IAAI,EAAE,QAAQ,QAAQ,CAAC,IAAI;AAC9D,UAAM,KAAK,KAAK,EAAE,WAAW,MAAM,EAAE,CAAC,MAAM,IAAI,EAAE,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,CAAC,IAAI;AAAA,EACnF;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAgCO,SAAS,YACd,MACA,OAGI,CAAC,GACa;AAClB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,aAAa,UAAU,YAAY,iBAAiB;AAE1D,QAAM,UAAU,oBAAI,IAAmD;AACvE,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,aAAa,MAAO;AAC1B,UAAM,IAAI,EAAE,QAAQ,UAAU;AAC9B,QAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,EAAG;AAClD,UAAM,SAAS,QAAQ,IAAI,EAAE,WAAW,KAAK,EAAE,MAAM,CAAC,GAAG,SAAS,CAAC,EAAE;AACrE,WAAO,KAAK,KAAK,EAAE,OAAO;AAC1B,WAAO,QAAQ,KAAK,CAAC;AACrB,YAAQ,IAAI,EAAE,aAAa,MAAM;AAAA,EACnC;AAEA,QAAM,SAAwB,CAAC;AAC/B,aAAW,CAAC,aAAa,MAAM,KAAK,QAAQ,QAAQ,GAAG;AACrD,WAAO,KAAK;AAAA,MACV;AAAA,MACA,MAAM,IAAI,OAAO,IAAI;AAAA,MACrB,SAAS,IAAI,OAAO,OAAO;AAAA,MAC3B,GAAG,OAAO,KAAK;AAAA,MACf,YAAY;AAAA,MACZ,MAAM,KAAK,gBAAgB,WAAW,IAClC,UAAU,KAAK,cAAc,WAAW,CAAE,IAC1C;AAAA,IACN,CAAC;AAAA,EACH;AAKA,aAAW,KAAK,QAAQ;AACtB,MAAE,aAAa,CAAC,OAAO,KAAK,CAAC,MAAM,MAAM,KAAK,UAAU,GAAG,CAAC,CAAC;AAAA,EAC/D;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IACN;AAAA,IACA,MAAM,EAAE,GAAG,WAAW,GAAG,QAAQ;AAAA,IACjC;AAAA,EACF;AACF;AAEA,SAAS,UAAU,GAAgB,GAAyB;AAC1D,SAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,WAAW,EAAE,YAAY,EAAE,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE;AACzF;AAEA,SAAS,UAAU,GAAsC;AACvD,MAAI,EAAE,QAAS,QAAO;AACtB,MAAI,EAAE,kBAAkB,WAAY,QAAO;AAC3C,MAAI,EAAE,kBAAkB,iBAAkB,QAAO;AACjD,MAAI,EAAE,kBAAkB,cAAe,QAAO;AAC9C,SAAO;AACT;AA2CO,SAAS,cACd,MACA,aACA,YACA,OAAgC,CAAC,GACL;AAC5B,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,aAAa,UAAU,YAAY,iBAAiB;AAC1D,QAAM,WAAW,KAAK,QAAQ;AAC9B,MAAI,WAAW,EAAG,OAAM,IAAI,MAAM,sCAAiC;AAEnE,QAAM,YAAY,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAAK;AAC1F,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAAK;AAIxF,QAAM,EAAE,QAAQ,MAAM,IAAI,gBAAgB,WAAW,UAAU,UAAU;AACzE,QAAM,IAAI,OAAO;AAEjB,MAAI,MAAM,GAAG;AACX,WAAO;AAAA,MACL,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAG;AAAA,MACH,MAAM,CAAC;AAAA,MACP,QAAQ;AAAA,MACR,IAAI,EAAE,KAAK,GAAG,MAAM,EAAE;AAAA,IACxB;AAAA,EACF;AAEA,QAAM,SAAS,OAAO,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,IAAK,CAAC;AACjD,QAAM,eAAe,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACrD,QAAM,SAAS,eAAe,YAAY;AAC1C,QAAM,MAAM,aAAa,CAAC;AAC1B,QAAM,MAAM,aAAa,aAAa,SAAS,CAAC;AAIhD,QAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,GAAG,GAAG,KAAK,IAAI,GAAG,GAAG,IAAI;AACzD,QAAM,KAAK,CAAC;AACZ,QAAM,KAAK;AACX,QAAM,SAAS,KAAK,MAAM;AAC1B,QAAM,OAA8B,CAAC;AACrC,WAAS,IAAI,GAAG,IAAI,UAAU,KAAK;AACjC,SAAK,KAAK,EAAE,IAAI,KAAK,IAAI,OAAO,IAAI,MAAM,IAAI,KAAK,OAAO,OAAO,EAAE,CAAC;AAAA,EACtE;AACA,aAAW,KAAK,QAAQ;AACtB,QAAI,MAAM,KAAK,OAAO,IAAI,MAAM,KAAK;AACrC,QAAI,MAAM,EAAG,OAAM;AACnB,QAAI,OAAO,SAAU,OAAM,WAAW;AACtC,SAAK,GAAG,EAAG,SAAS;AAAA,EACtB;AAEA,QAAM,KAAK,gBAAgB,QAAQ,OAAO;AAAA,IACxC,YAAY,KAAK,cAAc;AAAA,IAC/B,WAAW,KAAK,aAAa;AAAA,IAC7B,WAAW;AAAA,IACX,MAAM,KAAK;AAAA,EACb,CAAC;AAED,SAAO;AAAA,IACL,MAAM;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,IAAI,EAAE,KAAK,GAAG,KAAK,MAAM,GAAG,KAAK;AAAA,EACnC;AACF;AAgBO,IAAM,kCAAkC;AA2J/C,SAAS,gBACP,MACA,aACA,YACA,MAiBO;AACP,QAAM,aAAa,KAAK,UAAU,YAAY,iBAAiB;AAC/D,QAAM,YAAY,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAAK,KAAK;AAC/F,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAAK,KAAK;AAC7F,QAAM,EAAE,QAAQ,MAAM,IAAI,gBAAgB,WAAW,UAAU,UAAU;AACzE,QAAM,IAAI,OAAO;AACjB,MAAI,MAAM,EAAG,QAAO;AAEpB,QAAM,SAAS,OAAO,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,IAAK,CAAC;AACjD,QAAM,YAAY,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACtD,QAAM,eAAe,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACrD,QAAM,cAAc,eAAe,YAAY;AAC/C,QAAM,UAAU,MAAM,QAAQ,SAAS;AAEvC,QAAM,KAAK,gBAAgB,QAAQ,OAAO;AAAA,IACxC,YAAY,KAAK;AAAA,IACjB,WAAW;AAAA,IACX,WAAW;AAAA,IACX,MAAM,KAAK;AAAA,EACb,CAAC;AAKD,QAAM,cAAc,qBAAqB,QAAQ,KAAM,KAAK,IAAI;AAChE,QAAM,oBACJ,YAAY,WAAW,IAAI,IAAI,YAAY,OAAO,CAAC,MAAM,IAAI,CAAC,EAAE,SAAS,YAAY;AACvF,QAAM,WACJ,KAAK,SAAS,QAAQ,YAAY,WAAW,IACzC,OACA,YAAY,OAAO,CAAC,MAAM,KAAK,KAAK,KAAM,OAAO,KAAK,KAAK,KAAM,IAAI,EAAE,SACvE,YAAY;AAElB,QAAM,gBAAgB,UAAU,EAAE,SAAS,GAAG,OAAO,KAAK,UAAU,OAAO,KAAK,SAAS,CAAC;AAC1F,QAAM,MAAM,YAAY,IAAI,IAAI,gBAAgB;AAEhD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,IAAI,EAAE,KAAK,GAAG,KAAK,MAAM,GAAG,KAAK;AAAA,IACjC;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,qBAAqB,QAAkB,WAAmB,MAAyB;AAC1F,QAAM,IAAI,OAAO;AACjB,MAAI,MAAM,EAAG,QAAO,CAAC;AACrB,MAAI,MAAM,EAAG,QAAO,IAAI,MAAc,SAAS,EAAE,KAAK,OAAO,CAAC,CAAE;AAChE,QAAM,MAAM,QAAQ,IAAI;AACxB,QAAM,UAAU,IAAI,MAAc,SAAS;AAC3C,WAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,QAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,CAAC;AAC/D,YAAQ,CAAC,IAAI,MAAM;AAAA,EACrB;AACA,SAAO;AACT;AAEA,SAAS,QAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,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,MAAM,IAAc,MAAsB;AACjD,MAAI,GAAG,SAAS,EAAG,QAAO;AAC1B,MAAI,MAAM;AACV,aAAW,KAAK,GAAI,SAAQ,IAAI,SAAS;AACzC,SAAO,KAAK,KAAK,OAAO,GAAG,SAAS,EAAE;AACxC;AAqBA,eAAsB,eACpB,MACA,OAA8B,CAAC,GACN;AACzB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,WAAW,KAAK,IAAI,KAAK,YAAY,IAAI,+BAA+B;AAC9E,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,cAAc,KAAK,gBAAe,oBAAI,KAAK,GAAE,YAAY;AAC/D,QAAM,sBAAsB,KAAK,uBAAuB;AAExD,MAAI,QAAQ,EAAE,OAAO,SAAS,KAAK,GAAG,KAAK,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,OAAO,KAAK,OAAO;AAC/F,UAAM,IAAI;AAAA,MACR,6EAAwE,KAAK,UAAU,IAAI,CAAC;AAAA,IAC9F;AAAA,EACF;AAEA,QAAM,UAAU,aAAa,MAAM;AAAA,IACjC,YAAY,cAAc;AAAA,IAC1B;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AACD,QAAM,SAAS,YAAY,MAAM,EAAE,OAAO,eAAe,KAAK,cAAc,CAAC;AAC7E,QAAM,eACJ,KAAK,gBAAgB,QAAQ,KAAK,IAAI,CAAC,MAAM,EAAE,WAAW,EAAE,OAAO,CAAC,OAAO,OAAO,UAAU;AAC9F,QAAM,QAAQ,aACV,aAAa;AAAA,IAAI,CAAC,OAChB,cAAc,MAAM,IAAI,YAAY;AAAA,MAClC;AAAA,MACA;AAAA,MACA,MAAM,KAAK;AAAA,IACb,CAAC;AAAA,EACH,IACA,CAAC;AAEL,QAAM,kBAAkB,IAAI,IAAI,MAAM,IAAI,CAAC,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;AACpE,QAAM,oBAAoB,IAAI,IAAI,OAAO,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;AAC9E,QAAM,uBAAuB,oBAAI,IAAgD;AACjF,MAAI,YAAY;AACd,eAAW,MAAM,cAAc;AAC7B,2BAAqB;AAAA,QACnB;AAAA,QACA,gBAAgB,MAAM,IAAI,YAAY;AAAA,UACpC;AAAA,UACA;AAAA,UACA,MAAM,KAAK;AAAA,UACX;AAAA,UACA;AAAA,UACA;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAa,QAAQ,KACxB,IAAI,CAAC,QAAQ;AACZ,UAAM,OAAO,gBAAgB,IAAI,IAAI,WAAW;AAChD,UAAM,QAAQ,kBAAkB,IAAI,IAAI,WAAW;AACnD,UAAM,YAAY,qBAAqB,IAAI,IAAI,WAAW,KAAK;AAC/D,UAAM,aAAa,kBAAkB,KAAK;AAAA,MACxC;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AACD,WAAO;AAAA,MACL,aAAa,IAAI;AAAA,MACjB,GAAG,IAAI;AAAA,MACP,MAAM,IAAI;AAAA,MACV,OAAO,IAAI;AAAA,MACX,QAAQ,IAAI;AAAA,MACZ,QAAQ,IAAI;AAAA,MACZ,SAAS,IAAI;AAAA,MACb,uBAAuB,YAAY,UAAU,YAAY;AAAA,MACzD,SAAS,WAAW,KAAK,MAAM,KAAK;AAAA,MACpC,YAAY,YAAY,UAAU,cAAc,OAAO,KAAK,SAAS;AAAA,MACrE,UAAU,YAAY,UAAU,YAAY;AAAA,MAC5C,QAAQ,YAAY,UAAU,KAAK,OAAO,KAAK,KAAK;AAAA,MACpD,mBAAmB,YAAY,UAAU,oBAAoB;AAAA,MAC7D,UAAU,YAAY,UAAU,WAAW;AAAA,MAC3C,KAAK,YAAY,UAAU,MAAM;AAAA,MACjC,kBAAkB,OAAO,cAAc;AAAA,MACvC,MAAM,OAAO;AAAA,MACb,UAAU,WAAW;AAAA,MACrB,gBAAgB,WAAW;AAAA,IAC7B;AAAA,EACF,CAAC,EACA,KAAK,CAAC,GAAG,MAAM;AACd,UAAM,eAAe,eAAe,EAAE,QAAQ,IAAI,eAAe,EAAE,QAAQ;AAC3E,QAAI,iBAAiB,EAAG,QAAO;AAC/B,WAAO,EAAE,OAAO,EAAE;AAAA,EACpB,CAAC;AAEH,QAAM,iBAAiB,oBAAoB,YAAY;AAAA,IACrD;AAAA,IACA,iBAAiB,KAAK;AAAA,IACtB;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AACD,QAAM,mBAAmB,sBAAsB,YAAY,gBAAgB;AAAA,IACzE;AAAA,IACA;AAAA,IACA,iBAAiB,KAAK;AAAA,IACtB;AAAA,EACF,CAAC;AACD,QAAM,cAAc,iBAAiB;AAAA,IACnC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAED,QAAM,iBAAiB,MAAM;AAAA,IAC3B,aAAa;AAAA,MACX,SAAS,KACN,OAAO,CAAC,MAAM,EAAE,aAAa,KAAK,EAClC,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,aAAa,EAAE,aAAa,UAAU,EAAE,SAAS,EAAE,EACjF,KAAK,CAAC,GAAG,MAAM,EAAE,MAAM,cAAc,EAAE,KAAK,CAAC;AAAA,MAChD;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAEA,QAAM,WAAW,uBAAuB;AAAA,IACtC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,iBAAiB,KAAK;AAAA,EACxB,CAAC;AACD,QAAM,OAAO,mBAAmB,UAAU,KAAK;AAE/C,SAAO;AAAA,IACL,MAAM;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,EAAE,QAAQ,MAAM;AAAA,IACxB;AAAA,IACA,iBAAiB,KAAK;AAAA,IACtB;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,KASI;AAC5B,QAAM,cAAwB;AAAA,IAC5B;AAAA,IACA;AAAA,IACA,sCAAsC,IAAI,GAAG,cAAc,IAAI,QAAQ,gBAAgB,IAAI,UAAU;AAAA,EACvG;AACA,MAAI,IAAI,MAAM;AACZ,gBAAY;AAAA,MACV,uCAAuC,WAAW,IAAI,IAAI,CAAC;AAAA,IAC7D;AAAA,EACF;AACA,MAAI,IAAI,eAAe,MAAM;AAC3B,gBAAY,KAAK,oEAAoE;AAAA,EACvF;AACA,QAAM,UAAoB;AAAA,IACxB;AAAA,IACA;AAAA,IACA,8DAAyD,IAAI,QAAQ,uBAAuB,IAAI,QAAQ;AAAA,IACxG;AAAA,IACA;AAAA,EACF;AACA,QAAM,eAAyB;AAAA,IAC7B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,QAAM,iBAA2B;AAAA,IAC/B,kBAAkB,+BAA+B;AAAA,IACjD;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,QAAM,YAAsB;AAAA,IAC1B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO,EAAE,aAAa,SAAS,cAAc,gBAAgB,UAAU;AACzE;AAEA,SAAS,WAAW,MAA6C;AAC/D,SAAO,IAAI,IAAI,KAAK,GAAG,CAAC,KAAK,IAAI,KAAK,IAAI,CAAC;AAC7C;AAEA,SAAS,kBACP,KACA,KAQsD;AACtD,MAAI,IAAI,cAAc,IAAI,gBAAgB,IAAI,YAAY;AACxD,WAAO,EAAE,UAAU,QAAQ,QAAQ,uBAAuB;AAAA,EAC5D;AACA,MAAI,CAAC,IAAI,YAAY;AACnB,WAAO;AAAA,MACL,UAAU,IAAI,OAAO,aAAa,SAAS;AAAA,MAC3C,QACE;AAAA,IACJ;AAAA,EACF;AAGA,MAAI,IAAI,OAAO,QAAQ,IAAI,MAAM,SAAS,WAAW;AACnD,WAAO,EAAE,UAAU,UAAU,QAAQ,0BAA0B,IAAI,MAAM,IAAI,IAAI;AAAA,EACnF;AACA,MAAI,CAAC,IAAI,aAAa,IAAI,UAAU,IAAI,iCAAiC;AACvE,WAAO;AAAA,MACL,UAAU;AAAA,MACV,QAAQ,QAAQ,IAAI,WAAW,KAAK,CAAC,6CAA6C,+BAA+B;AAAA,IACnH;AAAA,EACF;AACA,QAAM,KAAK,IAAI,UAAU;AACzB,MAAI,IAAI,QAAQ,GAAG,OAAO,IAAI,KAAK,OAAO,GAAG,QAAQ,IAAI,KAAK,MAAM;AAClE,WAAO;AAAA,MACL,UAAU;AAAA,MACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC,0BAA0B,WAAW,IAAI,IAAI,CAAC;AAAA,IACxG;AAAA,EACF;AACA,QAAM,cAAc,OAAO,SAAS,IAAI,MAAM,KAAK,IAAI,UAAU,IAAI;AACrE,QAAM,eAAe,GAAG,MAAM;AAC9B,QAAM,eAAe,GAAG,OAAO;AAC/B,MAAI,cAAc;AAChB,WAAO;AAAA,MACL,UAAU;AAAA,MACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,MAAI,IAAI,UAAU,IAAI,IAAI,UAAU;AAClC,WAAO;AAAA,MACL,UAAU;AAAA,MACV,QAAQ,QAAQ,IAAI,UAAU,CAAC,gEAAgE,IAAI,IAAI,UAAU,GAAG,CAAC,6BAAwB,IAAI,QAAQ;AAAA,IAC3J;AAAA,EACF;AACA,MAAI,eAAe,cAAc;AAC/B,WAAO;AAAA,MACL,UAAU;AAAA,MACV,QAAQ,iBAAiB,IAAI,IAAI,MAAM,CAAC,WAAM,IAAI,GAAG,yBAAyB,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC,iCAA4B,IAAI,IAAI,UAAU,iBAAiB,CAAC;AAAA,IAC5K;AAAA,EACF;AACA,SAAO;AAAA,IACL,UAAU;AAAA,IACV,QAAQ,gBAAW,IAAI,IAAI,UAAU,iBAAiB,CAAC,YAAY,IAAI,GAAG,GAAG,CAAC,KAAK,IAAI,GAAG,IAAI,CAAC,8CAA8C,IAAI,GAAG;AAAA,EACtJ;AACF;AAEA,SAAS,oBACP,YACA,KAO8B;AAC9B,QAAM,gBAAgB,WAAW,OAAO,CAAC,MAAM,EAAE,gBAAgB,IAAI,UAAU;AAC/E,QAAM,cAAc,cAAc,KAAK,CAAC,MAAM,EAAE,aAAa,SAAS;AACtE,QAAM,iBAAiB,cAAc,KAAK,CAAC,MAAM,EAAE,aAAa,YAAY;AAC5E,QAAM,SAAS,eAAe,kBAAkB,cAAc,CAAC,KAAK;AACpE,QAAM,WAAmC,cACrC,YACA,cAAc,KAAK,CAAC,MAAM,EAAE,aAAa,iBAAiB,IACxD,oBACA,iBACE,eACA,cAAc,KAAK,CAAC,MAAM,EAAE,aAAa,MAAM,IAC7C,SACA;AAEV,QAAM,YAAsB,CAAC;AAC7B,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAwB,CAAC;AAE/B,MAAI,QAAQ;AACV,cAAU,KAAK,GAAG,OAAO,WAAW,KAAK,OAAO,cAAc,EAAE;AAChE,QAAI,OAAO,QAAQ;AACjB,YAAM,cACJ,OAAO,sBAAsB,OAAO,kBAAa,IAAI,OAAO,iBAAiB,CAAC,KAAK;AACrF,gBAAU;AAAA,QACR,2BAA2B,IAAI,OAAO,OAAO,GAAG,CAAC,KAAK,IAAI,OAAO,OAAO,IAAI,CAAC,IAAI,WAAW;AAAA,MAC9F;AAAA,IACF;AACA,QAAI,OAAO,QAAQ,QAAQ,OAAO,SAAS,OAAO,GAAG,GAAG;AACtD,gBAAU,KAAK,2BAA2B,OAAO,OAAO,KAAK,IAAI,OAAO,GAAG,CAAC,eAAe;AAAA,IAC7F;AAAA,EACF;AACA,MAAI,CAAC,IAAI,YAAY;AACnB,UAAM,KAAK,mEAAmE;AAC9E,gBAAY,KAAK,iEAAiE;AAAA,EACpF;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM;AAAA,MACJ;AAAA,IACF;AACA,gBAAY;AAAA,MACV;AAAA,IACF;AAAA,EACF;AACA,MAAI,IAAI,SAAS,QAAQ,cAAc,SAAS,GAAG;AACjD,UAAM;AAAA,MACJ;AAAA,IACF;AACA,gBAAY;AAAA,MACV;AAAA,IACF;AAAA,EACF;AACA,QAAM,eAAe,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,iBAAiB;AACjF,MAAI,aAAa,SAAS,GAAG;AAC3B,UAAM,QAAQ,aAAa,OAAO,CAAC,GAAG,MAAO,EAAE,UAAU,EAAE,UAAU,IAAI,CAAE;AAC3E,UAAM;AAAA,MACJ,GAAG,aAAa,MAAM,mCAAmC,IAAI,QAAQ,wBAAwB,MAAM,WAAW,SAAS,MAAM,OAAO;AAAA,IACtI;AACA,gBAAY;AAAA,MACV,oBAAoB,IAAI,WAAW,MAAM,OAAO,kCAAkC,MAAM,WAAW;AAAA,IACrG;AAAA,EACF;AACA,QAAM,WAAW,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ;AACpE,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM;AAAA,MACJ,GAAG,SAAS,MAAM;AAAA,IACpB;AAAA,EACF;AACA,MAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,SAAS,GAAG;AAClE,UAAM,MAAM,IAAI,gBAAgB,SAAS,CAAC;AAC1C,UAAM,KAAK,wBAAwB,IAAI,YAAY,WAAW,IAAI,QAAQ,UAAU;AACpF,gBAAY,KAAK,8DAA8D;AAAA,EACjF;AACA,MAAI,aAAa,WAAW;AAC1B,gBAAY,KAAK,+DAA+D;AAAA,EAClF,WAAW,aAAa,QAAQ;AAC9B,gBAAY,KAAK,qEAAqE;AAAA,EACxF,WAAW,aAAa,cAAc;AACpC,gBAAY;AAAA,MACV;AAAA,IACF;AAAA,EACF,WAAW,aAAa,UAAU;AAChC,gBAAY;AAAA,MACV;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,aAAa,QAAQ,eAAe;AAAA,IACpC;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,sBACP,YACA,gBACA,KAMU;AACV,QAAM,QAAkB,CAAC;AACzB,QAAM,gBAAgB,WAAW,OAAO,CAAC,MAAM,EAAE,gBAAgB,IAAI,UAAU;AAC/E,QAAM;AAAA,IACJ,aAAa,cAAc,MAAM,wBAAwB,IAAI,KAAK,SAAS,IAAI,aAAa,YAAY,IAAI,UAAU,KAAK,EAAE;AAAA,EAC/H;AACA,QAAM;AAAA,IACJ,mBAAmB,eAAe,QAAQ,GAAG,eAAe,cAAc,IAAI,eAAe,WAAW,KAAK,EAAE;AAAA,EACjH;AACA,QAAM,WAAW,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,SAAS,EAAE;AACvE,QAAM,OAAO,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,MAAM,EAAE;AAChE,QAAM,aAAa,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,YAAY,EAAE;AAC5E,QAAM,WAAW,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ,EAAE;AACtE,QAAM,OAAO,cAAc,OAAO,CAAC,MAAM,EAAE,aAAa,iBAAiB,EAAE;AAC3E,QAAM;AAAA,IACJ,iBAAiB,QAAQ,aAAa,UAAU,gBAAgB,IAAI,UAAU,QAAQ,YAAY,IAAI;AAAA,EACxG;AACA,QAAM,WAAW,cAAc,OAAO,CAAC,MAAM,EAAE,gBAAgB,EAAE,IAAI,CAAC,MAAM,EAAE,WAAW;AACzF,MAAI,SAAS,SAAS,EAAG,OAAM,KAAK,+BAA+B,SAAS,KAAK,IAAI,CAAC,GAAG;AACzF,MAAI,IAAI,iBAAiB;AACvB,UAAM;AAAA,MACJ,4BAA4B,IAAI,gBAAgB,aAAa,IAAI,IAAI,gBAAgB,SAAS,uBAAuB,IAAI,gBAAgB,SAAS,MAAM;AAAA,IAC1J;AAAA,EACF;AACA,QAAM;AAAA,IACJ,IAAI,sBACA,2BAA2B,IAAI,oBAAoB,MAAM,GAAG,EAAE,CAAC,WAC/D;AAAA,EACN;AACA,SAAO;AACT;AAEA,SAAS,uBAAuB,QAgBrB;AACT,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,KAAK,OAAO,KAAK,EAAE;AAC9B,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,kBAAkB,OAAO,WAAW,EAAE;AACjD,QAAM,KAAK,sBAAsB,OAAO,KAAK,EAAE;AAC/C,QAAM,KAAK,mBAAmB,OAAO,cAAc,gBAAgB,EAAE;AACrE,QAAM,KAAK,aAAa,OAAO,OAAO,WAAW,OAAO,IAAI,IAAI,gBAAgB,EAAE;AAClF,QAAM,KAAK,0BAA0B,OAAO,cAAc,IAAI;AAC9D,QAAM;AAAA,IACJ,wBAAwB,OAAO,sBAAsB,KAAK,OAAO,mBAAmB,OAAO,MAAM;AAAA,EACnG;AACA,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,sBAAsB;AACjC,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,iBAAkB,OAAM,KAAK,KAAK,IAAI,EAAE;AAClE,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,mBAAmB;AAC9B,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB,OAAO,eAAe,QAAQ,EAAE;AAC5D,QAAM,KAAK,kBAAkB,OAAO,eAAe,eAAe,KAAK,EAAE;AACzE,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,eAAe;AAC1B,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,eAAe,UAAW,OAAM,KAAK,KAAK,IAAI,EAAE;AAC1E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,WAAW;AACtB,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,eAAe,MAAM,SAC3C,OAAO,eAAe,QACtB,CAAC,0CAA0C,GAAG;AAChD,UAAM,KAAK,KAAK,IAAI,EAAE;AAAA,EACxB;AACA,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,kBAAkB;AAC7B,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,eAAe,YAAa,OAAM,KAAK,KAAK,IAAI,EAAE;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,6BAA6B;AACxC,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ;AAAA,EACF;AACA,QAAM,KAAK,0DAA0D;AACrE,aAAW,KAAK,OAAO,YAAY;AACjC,UAAM,QAAQ,EAAE,0BAA0B,OAAO,MAAM,OAAO,EAAE,qBAAqB;AACrF,UAAM,OAAO,EAAE,sBAAsB,OAAO,MAAM,EAAE,kBAAkB,QAAQ,CAAC;AAC/E,UAAM,IAAI,OAAO,SAAS,EAAE,MAAM,IAAI,EAAE,OAAO,QAAQ,CAAC,IAAI;AAC5D,UAAM,IAAI,OAAO,SAAS,EAAE,OAAO,IAAI,EAAE,QAAQ,QAAQ,CAAC,IAAI;AAC9D,UAAM,OAAO,EAAE,SAAS,IAAI,IAAI,EAAE,OAAO,GAAG,CAAC,KAAK,IAAI,EAAE,OAAO,IAAI,CAAC,MAAM;AAC1E,UAAM,MAAM,EAAE,QAAQ,QAAQ,CAAC,OAAO,SAAS,EAAE,GAAG,IAAI,MAAM,IAAI,EAAE,GAAG;AACvE,UAAM;AAAA,MACJ,KAAK,EAAE,WAAW,MAAM,EAAE,QAAQ,MAAM,IAAI,EAAE,IAAI,CAAC,MAAM,KAAK,MAAM,IAAI,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,MAAM,IAAI,MAAM,GAAG,MAAM,EAAE,mBAAmB,QAAQ,IAAI,MAAM,EAAE,QAAQ,GAAG;AAAA,IACpL;AAAA,EACF;AACA,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,wBAAwB;AACnC,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,OAAO,QAAQ,QAAQ;AAClC,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,gBAAgB;AAC3B,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB;AAC5B,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,YAAY,YAAa,OAAM,KAAK,KAAK,IAAI,EAAE;AACzE,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,aAAa;AACxB,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,YAAY,QAAS,OAAM,KAAK,KAAK,IAAI,EAAE;AACrE,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,6BAA6B;AACxC,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,YAAY,aAAc,OAAM,KAAK,KAAK,IAAI,EAAE;AAC1E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,uBAAuB;AAClC,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,YAAY,eAAgB,OAAM,KAAK,KAAK,IAAI,EAAE;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,eAAe;AAC1B,QAAM,KAAK,EAAE;AACb,aAAW,QAAQ,OAAO,YAAY,UAAW,OAAM,KAAK,KAAK,IAAI,EAAE;AACvE,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,gBAAgB;AAC3B,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ;AAAA,EACF;AACA,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,SAAS;AACpB,QAAM,KAAK,KAAK,UAAU,EAAE,QAAQ,OAAO,QAAQ,OAAO,OAAO,MAAM,GAAG,MAAM,CAAC,CAAC;AAClF,QAAM,KAAK,KAAK;AAChB,MAAI,OAAO,iBAAiB;AAC1B,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,qBAAqB;AAChC,UAAM,KAAK,EAAE;AACb,UAAM,KAAK,uDAAuD;AAClE,UAAM,KAAK,yBAAyB;AACpC,eAAW,KAAK,OAAO,gBAAgB,SAAS,MAAM,GAAG,EAAE,GAAG;AAC5D,YAAM;AAAA,QACJ,KAAK,EAAE,YAAY,MAAM,EAAE,QAAQ,MAAM,EAAE,YAAY,MAAM,MAAM,EAAE,YAAY,GAAG,MAAM,YAAY,EAAE,gBAAgB,EAAE,YAAY,CAAC;AAAA,MACzI;AAAA,IACF;AAAA,EACF;AACA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,mBAAmB,UAAkB,OAAuB;AACnE,QAAM,OAAO,eAAe,QAAQ;AACpC,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,UAAU,WAAW,KAAK,CAAC;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;AAEA,SAAS,eAAe,UAA0B;AAChD,QAAM,QAAQ,SAAS,MAAM,IAAI;AACjC,QAAM,OAAiB,CAAC;AACxB,MAAI,SAAS;AACb,MAAI,SAAS;AACb,MAAI,OAAiB,CAAC;AACtB,MAAI,QAAkB,CAAC;AAEvB,QAAM,YAAY,MAAM;AACtB,QAAI,QAAQ;AACV,WAAK,KAAK,OAAO;AACjB,eAAS;AAAA,IACX;AAAA,EACF;AACA,QAAM,aAAa,MAAM;AACvB,QAAI,MAAM,WAAW,EAAG;AACxB,SAAK,KAAK,oBAAoB,KAAK,CAAC;AACpC,YAAQ,CAAC;AAAA,EACX;AAEA,aAAW,QAAQ,OAAO;AACxB,QAAI,KAAK,WAAW,KAAK,GAAG;AAC1B,UAAI,QAAQ;AACV,aAAK,KAAK,cAAc,WAAW,KAAK,KAAK,IAAI,CAAC,CAAC,eAAe;AAClE,eAAO,CAAC;AACR,iBAAS;AAAA,MACX,OAAO;AACL,kBAAU;AACV,mBAAW;AACX,iBAAS;AAAA,MACX;AACA;AAAA,IACF;AACA,QAAI,QAAQ;AACV,WAAK,KAAK,IAAI;AACd;AAAA,IACF;AACA,QAAI,KAAK,WAAW,GAAG,GAAG;AACxB,gBAAU;AACV,YAAM,KAAK,IAAI;AACf;AAAA,IACF;AACA,eAAW;AACX,QAAI,KAAK,WAAW,IAAI,GAAG;AACzB,UAAI,CAAC,QAAQ;AACX,aAAK,KAAK,MAAM;AAChB,iBAAS;AAAA,MACX;AACA,WAAK,KAAK,OAAO,eAAe,KAAK,MAAM,CAAC,CAAC,CAAC,OAAO;AACrD;AAAA,IACF;AACA,cAAU;AACV,QAAI,KAAK,WAAW,IAAI,EAAG,MAAK,KAAK,OAAO,eAAe,KAAK,MAAM,CAAC,CAAC,CAAC,OAAO;AAAA,aACvE,KAAK,WAAW,KAAK,EAAG,MAAK,KAAK,OAAO,eAAe,KAAK,MAAM,CAAC,CAAC,CAAC,OAAO;AAAA,aAC7E,KAAK,WAAW,MAAM,EAAG,MAAK,KAAK,OAAO,eAAe,KAAK,MAAM,CAAC,CAAC,CAAC,OAAO;AAAA,aAC9E,KAAK,KAAK,MAAM,GAAI,MAAK,KAAK,EAAE;AAAA,QACpC,MAAK,KAAK,MAAM,eAAe,IAAI,CAAC,MAAM;AAAA,EACjD;AACA,YAAU;AACV,aAAW;AACX,SAAO,KAAK,KAAK,IAAI;AACvB;AAEA,SAAS,oBAAoB,OAAyB;AACpD,QAAM,OAAO,MACV,OAAO,CAAC,SAAS,CAAC,iBAAiB,KAAK,IAAI,CAAC,EAC7C;AAAA,IAAI,CAAC,SACJ,KACG,MAAM,GAAG,EAAE,EACX,MAAM,GAAG,EACT,IAAI,CAAC,SAAS,eAAe,KAAK,KAAK,CAAC,CAAC;AAAA,EAC9C;AACF,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,QAAM,CAAC,MAAM,GAAG,IAAI,IAAI;AACxB,QAAM,KAAK,KAAM,IAAI,CAAC,SAAS,OAAO,IAAI,OAAO,EAAE,KAAK,EAAE;AAC1D,QAAM,MAAM,KACT,IAAI,CAAC,QAAQ,OAAO,IAAI,IAAI,CAAC,SAAS,OAAO,IAAI,OAAO,EAAE,KAAK,EAAE,CAAC,OAAO,EACzE,KAAK,IAAI;AACZ,SAAO,qBAAqB,EAAE,uBAAuB,GAAG;AAC1D;AAEA,SAAS,eAAe,GAAmB;AACzC,SAAO,WAAW,CAAC,EAAE,QAAQ,oBAAoB,qBAAqB;AACxE;AAEA,SAAS,WAAW,GAAmB;AACrC,SAAO,EACJ,QAAQ,MAAM,OAAO,EACrB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,QAAQ;AAC3B;AAEA,SAAS,YAAY,GAAmB;AACtC,SAAO,EAAE,QAAQ,OAAO,KAAK;AAC/B;AAEA,SAAS,eAAe,UAA0C;AAChE,MAAI,aAAa,UAAW,QAAO;AACnC,MAAI,aAAa,aAAc,QAAO;AACtC,MAAI,aAAa,OAAQ,QAAO;AAChC,MAAI,aAAa,kBAAmB,QAAO;AAC3C,SAAO;AACT;AAEA,SAAS,OAAO,GAAmB;AACjC,SAAO,GAAG,KAAK,IAAI,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC;AACtC;AAIA,SAAS,IAAI,IAAsB;AACjC,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,eAAe,QAA0B;AAChD,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,QAAM,MAAM,KAAK,MAAM,OAAO,SAAS,CAAC;AACxC,SAAO,OAAO,SAAS,MAAM,KAAK,OAAO,MAAM,CAAC,IAAK,OAAO,GAAG,KAAM,IAAI,OAAO,GAAG;AACrF;AAEA,SAAS,IAAI,GAAmB;AAC9B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO,CAAC;AACxC,SAAO,EAAE,QAAQ,CAAC;AACpB;","names":[]}

package/dist/chunk-DTJ6QUQB.js

@@ -0,0 +1,131 @@
+import {
+  ValidationError
+} from "./chunk-ONWEPEDO.js";
+
+// src/rl/off-policy.ts
+function inverseProbabilityWeighting(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) {
+    return zeroEstimate();
+  }
+  const weights = [];
+  const weightedRewards = [];
+  let maxW = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(
+        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`
+      );
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    const r = clamp(t.reward, clip.low, clip.high);
+    weights.push(w);
+    weightedRewards.push(w * r);
+    if (w > maxW) maxW = w;
+  }
+  const n = weights.length;
+  const value = weightedRewards.reduce((s, x) => s + x, 0) / n;
+  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+  const sumW = weights.reduce((s, w) => s + w, 0);
+  const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  return {
+    value,
+    standardError: Math.sqrt(variance / n),
+    effectiveSampleSize: effN,
+    n,
+    maxImportanceWeight: maxW
+  };
+}
+function selfNormalizedImportanceWeighting(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) return zeroEstimate();
+  const weights = [];
+  const rewards = [];
+  let maxW = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(
+        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`
+      );
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    weights.push(w);
+    rewards.push(clamp(t.reward, clip.low, clip.high));
+    if (w > maxW) maxW = w;
+  }
+  const sumW = weights.reduce((s, w) => s + w, 0);
+  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i], 0);
+  const value = sumW === 0 ? 0 : sumWR / sumW;
+  const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  const phi = weights.map((w, i) => w * (rewards[i] - value));
+  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW);
+  return {
+    value,
+    standardError: Math.sqrt(variance),
+    effectiveSampleSize: effN,
+    n: trajectories.length,
+    maxImportanceWeight: maxW
+  };
+}
+function doublyRobust(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) return zeroEstimate();
+  const contributions = [];
+  let maxW = 0;
+  let sumW = 0;
+  let sumW2 = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`);
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    const r = clamp(t.reward, clip.low, clip.high);
+    const q = typeof t.qHat === "number" && Number.isFinite(t.qHat) ? clamp(t.qHat, clip.low, clip.high) : null;
+    if (q === null) {
+      contributions.push(w * r);
+    } else {
+      contributions.push(q + w * (r - q));
+    }
+    if (w > maxW) maxW = w;
+    sumW += w;
+    sumW2 += w * w;
+  }
+  const n = contributions.length;
+  const value = contributions.reduce((s, x) => s + x, 0) / n;
+  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  return {
+    value,
+    standardError: Math.sqrt(variance / n),
+    effectiveSampleSize: effN,
+    n,
+    maxImportanceWeight: maxW
+  };
+}
+function offPolicyEstimateAll(trajectories, opts = {}) {
+  return {
+    ips: inverseProbabilityWeighting(trajectories, opts),
+    snips: selfNormalizedImportanceWeighting(trajectories, opts),
+    dr: doublyRobust(trajectories, opts)
+  };
+}
+function zeroEstimate() {
+  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 };
+}
+function clamp(x, lo, hi) {
+  if (!Number.isFinite(x)) return lo;
+  return Math.max(lo, Math.min(hi, x));
+}
+
+export {
+  inverseProbabilityWeighting,
+  selfNormalizedImportanceWeighting,
+  doublyRobust,
+  offPolicyEstimateAll
+};
+//# sourceMappingURL=chunk-DTJ6QUQB.js.map

package/dist/chunk-DTJ6QUQB.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rl/off-policy.ts"],"sourcesContent":["/**\n * Off-policy evaluation primitives.\n *\n * Standard inverse-probability-weighted (IPS), self-normalized\n * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the\n * value of a *target* policy given trajectories collected under a\n * *behavior* policy. This is the canonical RL eval task: \"we have last\n * week's runs, we changed the policy — how would the new one do without\n * re-running?\"\n *\n * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan\n * & Joachims 2015 for SNIPS) but the *application* to LLM-agent\n * evaluation needs care:\n *\n *   - The \"policy\" is the (prompt, tool config, model snapshot) triple.\n *     Two policies have the same probability over an action *iff* their\n *     LLM call would emit the same token with the same probability —\n *     which is generally unknowable without the model log-probs.\n *   - For LLM agents, propensity scores must be supplied by the caller\n *     (logged in the trace, recovered from token log-probs, or estimated\n *     via a learned propensity model). We do NOT estimate propensity here.\n *   - Doubly-robust requires a Q-function (model-based reward predictor).\n *     We accept any callable; consumers pass either a tabular average,\n *     a regression fit, or a learned reward model.\n *\n * Bias / variance tradeoffs:\n *   - IPS: unbiased; high variance for small overlap, infinite variance\n *     when target has support outside behavior.\n *   - SNIPS: lower variance, slight bias; usually preferred in practice.\n *   - DR: doubly-robust — unbiased if either propensity OR Q-function is\n *     correct. Lowest practical variance when Q is decent. Use this.\n *\n * Caveat the panel will land: on the LLM-agent setting, propensity scores\n * recovered from token log-probs are noisy, the action space is enormous,\n * and overlap is often poor. These estimators are useful but not magic;\n * complement with `replayCampaign` (exact replay where the request hashes\n * match) for high-confidence answers and OPE for the gap.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface OffPolicyTrajectory {\n  /** Stable id, for traceability through the dataset. */\n  runId: string\n  /** Reward observed under the behavior policy (the realized outcome). */\n  reward: number\n  /**\n   * Behavior-policy probability of the action that was taken. For LLM\n   * agents this is typically `exp(sum(token_log_probs))` over the chosen\n   * trajectory. Must be in (0, 1].\n   */\n  behaviorProb: number\n  /**\n   * Target-policy probability of the same action. For replay-style\n   * counterfactual evaluation this is what the *new* policy would have\n   * assigned to the *old* trajectory. Must be in [0, 1].\n   */\n  targetProb: number\n  /**\n   * Optional model-based reward prediction at the same context. Used by\n   * `doublyRobust`. Set to `null` for IPS-only evaluation.\n   */\n  qHat?: number | null\n}\n\nexport interface OffPolicyEstimate {\n  /** Estimated value of the target policy. */\n  value: number\n  /** Standard error of the estimate. */\n  standardError: number\n  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */\n  effectiveSampleSize: number\n  /** Number of trajectories used. */\n  n: number\n  /**\n   * Diagnostic: maximum importance weight observed. Large values (>>10x\n   * mean) are a red flag — variance is dominated by a few outliers.\n   */\n  maxImportanceWeight: number\n}\n\nexport interface OffPolicyOptions {\n  /**\n   * Cap importance weights at this value (Ionides 2008 truncated IS) to\n   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).\n   * Set e.g. `10` for stable estimates when the policies are close.\n   */\n  weightCap?: number\n  /** Reward clipping range. Default `[0, 1]`. */\n  rewardClip?: { low: number; high: number }\n}\n\n/**\n * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator\n * of E[reward under target policy]. Variance scales with the spread of\n * target/behavior ratios.\n */\nexport function inverseProbabilityWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n\n  if (trajectories.length === 0) {\n    return zeroEstimate()\n  }\n\n  const weights: number[] = []\n  const weightedRewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    weights.push(w)\n    weightedRewards.push(w * r)\n    if (w > maxW) maxW = w\n  }\n  const n = weights.length\n  const value = weightedRewards.reduce((s, x) => s + x, 0) / n\n  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at\n * the cost of small bias (vanishing as N grows). The right default for\n * LLM-agent evaluation where overlap is often poor.\n */\nexport function selfNormalizedImportanceWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const weights: number[] = []\n  const rewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    weights.push(w)\n    rewards.push(clamp(t.reward, clip.low, clip.high))\n    if (w > maxW) maxW = w\n  }\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i]!, 0)\n  const value = sumW === 0 ? 0 : sumWR / sumW\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  // Influence-function-based SE for SNIPS (Owen 2013, Ch. 9).\n  const phi = weights.map((w, i) => w * (rewards[i]! - value))\n  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW)\n  return {\n    value,\n    standardError: Math.sqrt(variance),\n    effectiveSampleSize: effN,\n    n: trajectories.length,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).\n *\n *     V_DR = (1/N) * sum_i [ q_hat_i + (target_prob_i / behavior_prob_i) * (r_i - q_hat_i) ]\n *\n * Unbiased if EITHER:\n *   - the importance ratios are correct (IPS-style validity), OR\n *   - the Q-hat function is correct (model-based validity).\n *\n * In practice both are imperfect, but the residual bias is the *product*\n * of both errors — much smaller than either alone. This is why DR is the\n * default in production OPE pipelines.\n *\n * Requires `qHat` on every trajectory. If any are `null`, the estimator\n * falls back to SNIPS for those entries (loud-fallback behavior; the\n * report's `n` reflects the full set but `effectiveSampleSize` accounts\n * for the lost variance reduction).\n */\nexport function doublyRobust(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const contributions: number[] = []\n  let maxW = 0\n  let sumW = 0\n  let sumW2 = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`)\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    const q =\n      typeof t.qHat === 'number' && Number.isFinite(t.qHat)\n        ? clamp(t.qHat, clip.low, clip.high)\n        : null\n    if (q === null) {\n      contributions.push(w * r) // fallback: IPS for this entry\n    } else {\n      contributions.push(q + w * (r - q))\n    }\n    if (w > maxW) maxW = w\n    sumW += w\n    sumW2 += w * w\n  }\n  const n = contributions.length\n  const value = contributions.reduce((s, x) => s + x, 0) / n\n  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Convenience: run all three estimators and return them side-by-side.\n * The recommended diagnostic — agreement across estimators is a much\n * stronger signal than any single one.\n */\nexport function offPolicyEstimateAll(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): { ips: OffPolicyEstimate; snips: OffPolicyEstimate; dr: OffPolicyEstimate } {\n  return {\n    ips: inverseProbabilityWeighting(trajectories, opts),\n    snips: selfNormalizedImportanceWeighting(trajectories, opts),\n    dr: doublyRobust(trajectories, opts),\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction zeroEstimate(): OffPolicyEstimate {\n  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 }\n}\n\nfunction clamp(x: number, lo: number, hi: number): number {\n  if (!Number.isFinite(x)) return lo\n  return Math.max(lo, Math.min(hi, x))\n}\n"],"mappings":";;;;;AAiGO,SAAS,4BACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAElD,MAAI,aAAa,WAAW,GAAG;AAC7B,WAAO,aAAa;AAAA,EACtB;AAEA,QAAM,UAAoB,CAAC;AAC3B,QAAM,kBAA4B,CAAC;AACnC,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,gEAAgE,EAAE,KAAK;AAAA,MACzE;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,YAAQ,KAAK,CAAC;AACd,oBAAgB,KAAK,IAAI,CAAC;AAC1B,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,IAAI,QAAQ;AAClB,QAAM,QAAQ,gBAAgB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3D,QAAM,WAAW,gBAAgB,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC9F,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,kCACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,UAAoB,CAAC;AAC3B,QAAM,UAAoB,CAAC;AAC3B,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,sEAAsE,EAAE,KAAK;AAAA,MAC/E;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,YAAQ,KAAK,CAAC;AACd,YAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,CAAC;AACjD,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,GAAG,MAAM,IAAI,IAAI,QAAQ,CAAC,GAAI,CAAC;AAChE,QAAM,QAAQ,SAAS,IAAI,IAAI,QAAQ;AACvC,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,QAAM,MAAM,QAAQ,IAAI,CAAC,GAAG,MAAM,KAAK,QAAQ,CAAC,IAAK,MAAM;AAC3D,QAAM,WAAW,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO,IAAI;AAC7E,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,QAAQ;AAAA,IACjC,qBAAqB;AAAA,IACrB,GAAG,aAAa;AAAA,IAChB,qBAAqB;AAAA,EACvB;AACF;AAoBO,SAAS,aACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,gBAA0B,CAAC;AACjC,MAAI,OAAO;AACX,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI,gBAAgB,iDAAiD,EAAE,KAAK,GAAG;AAAA,IACvF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,UAAM,IACJ,OAAO,EAAE,SAAS,YAAY,OAAO,SAAS,EAAE,IAAI,IAChD,MAAM,EAAE,MAAM,KAAK,KAAK,KAAK,IAAI,IACjC;AACN,QAAI,MAAM,MAAM;AACd,oBAAc,KAAK,IAAI,CAAC;AAAA,IAC1B,OAAO;AACL,oBAAc,KAAK,IAAI,KAAK,IAAI,EAAE;AAAA,IACpC;AACA,QAAI,IAAI,KAAM,QAAO;AACrB,YAAQ;AACR,aAAS,IAAI;AAAA,EACf;AACA,QAAM,IAAI,cAAc;AACxB,QAAM,QAAQ,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACzD,QAAM,WAAW,cAAc,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC5F,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAC9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,qBACd,cACA,OAAyB,CAAC,GACmD;AAC7E,SAAO;AAAA,IACL,KAAK,4BAA4B,cAAc,IAAI;AAAA,IACnD,OAAO,kCAAkC,cAAc,IAAI;AAAA,IAC3D,IAAI,aAAa,cAAc,IAAI;AAAA,EACrC;AACF;AAIA,SAAS,eAAkC;AACzC,SAAO,EAAE,OAAO,GAAG,eAAe,GAAG,qBAAqB,GAAG,GAAG,GAAG,qBAAqB,EAAE;AAC5F;AAEA,SAAS,MAAM,GAAW,IAAY,IAAoB;AACxD,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,CAAC;AACrC;","names":[]}

package/dist/chunk-GGE4NNQT.js

@@ -0,0 +1,65 @@
+// src/trace/redact.ts
+var DEFAULT_REDACTION_RULES = [
+  { id: "email", pattern: /\b[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}\b/gi },
+  { id: "ssn", pattern: /\b\d{3}-\d{2}-\d{4}\b/g },
+  { id: "credit-card", pattern: /\b(?:\d[ -]*?){13,16}\b/g },
+  { id: "phone-us", pattern: /\b(?:\+?1[-.\s]?)?\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b/g },
+  { id: "ipv4", pattern: /\b(?:\d{1,3}\.){3}\d{1,3}\b/g },
+  { id: "aws-access-key", pattern: /\bAKIA[0-9A-Z]{16}\b/g },
+  { id: "bearer", pattern: /\bBearer\s+[A-Za-z0-9._~+/=-]{10,}/gi },
+  { id: "sk-key", pattern: /\bsk-[A-Za-z0-9_-]{10,}\b/g },
+  {
+    id: "private-key-block",
+    pattern: /-----BEGIN (?:RSA |EC |OPENSSH |DSA )?PRIVATE KEY-----[\s\S]*?-----END[^-]*-----/g
+  }
+];
+var REDACTION_VERSION = "1.0.0";
+function redactString(input, rules = DEFAULT_REDACTION_RULES) {
+  const byRule = {};
+  let redactionCount = 0;
+  let output = input;
+  for (const rule of rules) {
+    let hits = 0;
+    output = output.replace(rule.pattern, () => {
+      hits++;
+      return rule.replacement ?? `[redacted:${rule.id}]`;
+    });
+    if (hits > 0) {
+      byRule[rule.id] = hits;
+      redactionCount += hits;
+    }
+  }
+  return { output, report: { redactionCount, byRule } };
+}
+function redactValue(value, rules = DEFAULT_REDACTION_RULES, report = { redactionCount: 0, byRule: {} }) {
+  if (typeof value === "string") {
+    const { output, report: r } = redactString(value, rules);
+    report.redactionCount += r.redactionCount;
+    for (const [k, v] of Object.entries(r.byRule)) {
+      report.byRule[k] = (report.byRule[k] ?? 0) + v;
+    }
+    return { value: output, report };
+  }
+  if (Array.isArray(value)) {
+    return {
+      value: value.map((v) => redactValue(v, rules, report).value),
+      report
+    };
+  }
+  if (value !== null && typeof value === "object") {
+    const next = {};
+    for (const [k, v] of Object.entries(value)) {
+      next[k] = redactValue(v, rules, report).value;
+    }
+    return { value: next, report };
+  }
+  return { value, report };
+}
+
+export {
+  DEFAULT_REDACTION_RULES,
+  REDACTION_VERSION,
+  redactString,
+  redactValue
+};
+//# sourceMappingURL=chunk-GGE4NNQT.js.map

package/dist/chunk-GGE4NNQT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/trace/redact.ts"],"sourcesContent":["/**\n * Redaction — remove PII / secrets from trace payloads before persist.\n *\n * Pre-persistence rules mean raw traces in storage are already scrubbed.\n * Unredacted variants (for debugging / post-mortems) live in a separate\n * storage layer with stricter access controls; this module only covers\n * the default scrub-then-persist path.\n *\n * Rules compose: pass an array of `RedactionRule`, each is applied in\n * order. Strings that match get replaced with a tagged sentinel so the\n * eval framework can count how many redactions happened per run\n * (surfaced via `redaction_applied` events).\n */\n\nexport interface RedactionRule {\n  id: string\n  pattern: RegExp\n  /** Replacement — e.g. '[PII:email]'. Defaults to `[redacted:{id}]`. */\n  replacement?: string\n}\n\nexport interface RedactionReport {\n  redactionCount: number\n  byRule: Record<string, number>\n}\n\n/** OWASP / common-sense defaults — extend per-domain. */\nexport const DEFAULT_REDACTION_RULES: RedactionRule[] = [\n  { id: 'email', pattern: /\\b[A-Z0-9._%+-]+@[A-Z0-9.-]+\\.[A-Z]{2,}\\b/gi },\n  { id: 'ssn', pattern: /\\b\\d{3}-\\d{2}-\\d{4}\\b/g },\n  { id: 'credit-card', pattern: /\\b(?:\\d[ -]*?){13,16}\\b/g },\n  { id: 'phone-us', pattern: /\\b(?:\\+?1[-.\\s]?)?\\(?\\d{3}\\)?[-.\\s]?\\d{3}[-.\\s]?\\d{4}\\b/g },\n  { id: 'ipv4', pattern: /\\b(?:\\d{1,3}\\.){3}\\d{1,3}\\b/g },\n  { id: 'aws-access-key', pattern: /\\bAKIA[0-9A-Z]{16}\\b/g },\n  { id: 'bearer', pattern: /\\bBearer\\s+[A-Za-z0-9._~+/=-]{10,}/gi },\n  { id: 'sk-key', pattern: /\\bsk-[A-Za-z0-9_-]{10,}\\b/g },\n  {\n    id: 'private-key-block',\n    pattern: /-----BEGIN (?:RSA |EC |OPENSSH |DSA )?PRIVATE KEY-----[\\s\\S]*?-----END[^-]*-----/g,\n  },\n]\n\nexport const REDACTION_VERSION = '1.0.0'\n\n/**\n * Redact a single string. Returns the new string and a per-rule count of\n * how many substitutions fired.\n */\nexport function redactString(\n  input: string,\n  rules: RedactionRule[] = DEFAULT_REDACTION_RULES,\n): { output: string; report: RedactionReport } {\n  const byRule: Record<string, number> = {}\n  let redactionCount = 0\n  let output = input\n  for (const rule of rules) {\n    let hits = 0\n    output = output.replace(rule.pattern, () => {\n      hits++\n      return rule.replacement ?? `[redacted:${rule.id}]`\n    })\n    if (hits > 0) {\n      byRule[rule.id] = hits\n      redactionCount += hits\n    }\n  }\n  return { output, report: { redactionCount, byRule } }\n}\n\n/**\n * Walk a JSON-ish value applying `redactString` to every string leaf.\n * Arrays and plain objects are recursed; other types pass through\n * untouched. Circular references throw — traces should be tree-shaped.\n */\nexport function redactValue(\n  value: unknown,\n  rules: RedactionRule[] = DEFAULT_REDACTION_RULES,\n  report: RedactionReport = { redactionCount: 0, byRule: {} },\n): { value: unknown; report: RedactionReport } {\n  if (typeof value === 'string') {\n    const { output, report: r } = redactString(value, rules)\n    report.redactionCount += r.redactionCount\n    for (const [k, v] of Object.entries(r.byRule)) {\n      report.byRule[k] = (report.byRule[k] ?? 0) + v\n    }\n    return { value: output, report }\n  }\n  if (Array.isArray(value)) {\n    return {\n      value: value.map((v) => redactValue(v, rules, report).value),\n      report,\n    }\n  }\n  if (value !== null && typeof value === 'object') {\n    const next: Record<string, unknown> = {}\n    for (const [k, v] of Object.entries(value)) {\n      next[k] = redactValue(v, rules, report).value\n    }\n    return { value: next, report }\n  }\n  return { value, report }\n}\n"],"mappings":";AA2BO,IAAM,0BAA2C;AAAA,EACtD,EAAE,IAAI,SAAS,SAAS,8CAA8C;AAAA,EACtE,EAAE,IAAI,OAAO,SAAS,yBAAyB;AAAA,EAC/C,EAAE,IAAI,eAAe,SAAS,2BAA2B;AAAA,EACzD,EAAE,IAAI,YAAY,SAAS,2DAA2D;AAAA,EACtF,EAAE,IAAI,QAAQ,SAAS,+BAA+B;AAAA,EACtD,EAAE,IAAI,kBAAkB,SAAS,wBAAwB;AAAA,EACzD,EAAE,IAAI,UAAU,SAAS,uCAAuC;AAAA,EAChE,EAAE,IAAI,UAAU,SAAS,6BAA6B;AAAA,EACtD;AAAA,IACE,IAAI;AAAA,IACJ,SAAS;AAAA,EACX;AACF;AAEO,IAAM,oBAAoB;AAM1B,SAAS,aACd,OACA,QAAyB,yBACoB;AAC7C,QAAM,SAAiC,CAAC;AACxC,MAAI,iBAAiB;AACrB,MAAI,SAAS;AACb,aAAW,QAAQ,OAAO;AACxB,QAAI,OAAO;AACX,aAAS,OAAO,QAAQ,KAAK,SAAS,MAAM;AAC1C;AACA,aAAO,KAAK,eAAe,aAAa,KAAK,EAAE;AAAA,IACjD,CAAC;AACD,QAAI,OAAO,GAAG;AACZ,aAAO,KAAK,EAAE,IAAI;AAClB,wBAAkB;AAAA,IACpB;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,QAAQ,EAAE,gBAAgB,OAAO,EAAE;AACtD;AAOO,SAAS,YACd,OACA,QAAyB,yBACzB,SAA0B,EAAE,gBAAgB,GAAG,QAAQ,CAAC,EAAE,GACb;AAC7C,MAAI,OAAO,UAAU,UAAU;AAC7B,UAAM,EAAE,QAAQ,QAAQ,EAAE,IAAI,aAAa,OAAO,KAAK;AACvD,WAAO,kBAAkB,EAAE;AAC3B,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,EAAE,MAAM,GAAG;AAC7C,aAAO,OAAO,CAAC,KAAK,OAAO,OAAO,CAAC,KAAK,KAAK;AAAA,IAC/C;AACA,WAAO,EAAE,OAAO,QAAQ,OAAO;AAAA,EACjC;AACA,MAAI,MAAM,QAAQ,KAAK,GAAG;AACxB,WAAO;AAAA,MACL,OAAO,MAAM,IAAI,CAAC,MAAM,YAAY,GAAG,OAAO,MAAM,EAAE,KAAK;AAAA,MAC3D;AAAA,IACF;AAAA,EACF;AACA,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,OAAgC,CAAC;AACvC,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,KAAK,GAAG;AAC1C,WAAK,CAAC,IAAI,YAAY,GAAG,OAAO,MAAM,EAAE;AAAA,IAC1C;AACA,WAAO,EAAE,OAAO,MAAM,OAAO;AAAA,EAC/B;AACA,SAAO,EAAE,OAAO,OAAO;AACzB;","names":[]}