@tangle-network/agent-eval 0.216.1 → 0.218.1

package/dist/summary-report-BsXJoYX5.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"summary-report-BsXJoYX5.js","names":["avg","hi"],"sources":["../src/research-report-render.ts","../src/research-report-analysis.ts","../src/statistics/effect-sizes.ts","../src/summary-report-evidence.ts","../src/summary-report-figures.ts","../src/research-report.ts"],"sourcesContent":["/** Markdown research reports and their shared numeric text formatting. */\n\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport type {\n  ResearchReportCandidate,\n  ResearchReportMethodology,\n  ResearchReportRecommendation,\n} from './research-report'\nimport type {\n  GainDistributionFigureSpec,\n  ParetoFigureSpec,\n  SummaryTable,\n} from './summary-report-figures'\n\nexport function renderResearchMarkdown(report: {\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  pareto: ParetoFigureSpec\n  gains: GainDistributionFigureSpec[]\n  rope: { low: number; high: number } | null\n  runFingerprint: string\n  preregistrationHash: string | null\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n}): string {\n  const lines: string[] = []\n  lines.push(`# ${report.title}`)\n  lines.push('')\n  lines.push(`**Generated:** ${report.generatedAt}`)\n  lines.push(`**Primary split:** ${report.split}`)\n  lines.push(`**Comparator:** ${report.comparator ?? 'not configured'}`)\n  lines.push(`**ROPE:** ${report.rope ? formatRope(report.rope) : 'not configured'}`)\n  lines.push(`**Run fingerprint:** \\`${report.runFingerprint}\\``)\n  lines.push(\n    `**Preregistration:** ${report.preregistrationHash ? `\\`${report.preregistrationHash}\\`` : 'none'}`,\n  )\n  lines.push('')\n  lines.push('## Executive Summary')\n  lines.push('')\n  for (const item of report.executiveSummary) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Recommendation')\n  lines.push('')\n  lines.push(`**Decision:** ${report.recommendation.decision}`)\n  lines.push(`**Candidate:** ${report.recommendation.candidateId ?? 'N/A'}`)\n  lines.push('')\n  lines.push('### Rationale')\n  lines.push('')\n  for (const item of report.recommendation.rationale) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Risks')\n  lines.push('')\n  for (const item of report.recommendation.risks.length\n    ? report.recommendation.risks\n    : ['No material report-level risks detected.']) {\n    lines.push(`- ${item}`)\n  }\n  lines.push('')\n  lines.push('### Next Actions')\n  lines.push('')\n  for (const item of report.recommendation.nextActions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Candidate Decision Table')\n  lines.push('')\n  lines.push(\n    '| Candidate | Decision | Mean | Δ̄ | Pr(Δ>0) | q | dz | Paired N | Median Gain CI | MDE | Pareto | Gate |',\n  )\n  lines.push('|---|---|---:|---:|---:|---:|---:|---:|---|---:|---|---|')\n  for (const c of report.candidates) {\n    const delta = c.meanDeltaVsComparator === null ? '-' : signed(c.meanDeltaVsComparator)\n    const prGt = c.prGreaterThanZero === null ? '-' : c.prGreaterThanZero.toFixed(3)\n    const q = c.qValue !== null ? c.qValue.toFixed(4) : '-'\n    const d = c.cohensD !== null ? c.cohensD.toFixed(3) : '-'\n    const gain = c.gainCi ? `[${fmt(c.gainCi.low)}, ${fmt(c.gainCi.high)}]` : '-'\n    const mde = c.mde === null || !Number.isFinite(c.mde) ? '-' : fmt(c.mde)\n    lines.push(\n      `| ${c.candidateId} | ${c.decision} | ${fmt(c.mean)} | ${delta} | ${prGt} | ${q} | ${d} | ${c.pairedN} | ${gain} | ${mde} | ${c.onParetoFrontier ? 'yes' : 'no'} | ${c.gate ?? '-'} |`,\n    )\n  }\n  lines.push('')\n  lines.push('## Statistical Summary')\n  lines.push('')\n  lines.push(report.summary.markdown)\n  lines.push('')\n  lines.push('## Methodology')\n  lines.push('')\n  lines.push('### Assumptions')\n  lines.push('')\n  for (const item of report.methodology.assumptions) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Methods')\n  lines.push('')\n  for (const item of report.methodology.methods) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Alternatives Considered')\n  lines.push('')\n  for (const item of report.methodology.alternatives) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### When NOT To Apply')\n  lines.push('')\n  for (const item of report.methodology.whenNotToApply) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('### Citations')\n  lines.push('')\n  for (const item of report.methodology.citations) lines.push(`- ${item}`)\n  lines.push('')\n  lines.push('## Chart Specs')\n  lines.push('')\n  lines.push(\n    'The report carries JSON chart specs for Pareto cost/quality and paired gain histograms.',\n  )\n  lines.push('')\n  lines.push('```json')\n  lines.push(JSON.stringify({ pareto: report.pareto, gains: report.gains }, null, 2))\n  lines.push('```')\n  if (report.failureClusters) {\n    lines.push('')\n    lines.push('## Failure Clusters')\n    lines.push('')\n    lines.push('| Failure Class | Runs | Scenarios | Tool | Example |')\n    lines.push('|---|---:|---:|---|---|')\n    for (const c of report.failureClusters.clusters.slice(0, 10)) {\n      lines.push(\n        `| ${c.failureClass} | ${c.runCount} | ${c.scenarioIds.length} | ${c.toolName ?? '-'} | ${escapePipes(c.exampleError ?? c.exampleRunId)} |`,\n      )\n    }\n  }\n  return lines.join('\\n')\n}\n\nfunction escapePipes(s: string): string {\n  return s.replace(/\\|/g, '\\\\|')\n}\n\nfunction signed(x: number): string {\n  return `${x >= 0 ? '+' : ''}${fmt(x)}`\n}\n\nexport function formatRope(rope: { low: number; high: number }): string {\n  return `[${fmt(rope.low)}, ${fmt(rope.high)}]`\n}\n\nexport function fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/** Evidence-based report verdicts, recommendations, and methodology. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport type {\n  ResearchReportCandidate,\n  ResearchReportDecision,\n  ResearchReportMethodology,\n  ResearchReportRecommendation,\n} from './research-report'\nimport { fmt, formatRope } from './research-report-render'\nimport type { pairedPosterior } from './summary-report-evidence'\nimport type { ParetoPoint, SummaryTableRow } from './summary-report-figures'\n\n/**\n * Hard floor below which a paired comparison is treated as uninformative\n * regardless of `minPairs`. Mirrors the lower limit on Wilcoxon signed-rank\n * exact tables; below this the test has no power to separate effect sizes.\n */\nexport const RESEARCH_REPORT_HARD_PAIR_FLOOR = 6\n\nexport function buildMethodology(ctx: {\n  split: 'search' | 'holdout'\n  comparator: string | null\n  fdr: number\n  minPairs: number\n  rope: { low: number; high: number } | null\n  confidence: number\n  mdePower: number\n  mdeAlpha: number\n}): ResearchReportMethodology {\n  const assumptions: string[] = [\n    'Pairs are matched by (experimentId, scenarioId, seed); input order does not affect the comparison.',\n    'Paired deltas are exchangeable conditional on the matched scenario — no mid-run distribution shift.',\n    `Decisions are pre-specified at fdr=${ctx.fdr}, minPairs=${ctx.minPairs}, confidence=${ctx.confidence}; deviating from these post-hoc invalidates the false-discovery control.`,\n  ]\n  if (ctx.rope) {\n    assumptions.push(\n      `The Region of Practical Equivalence ${formatRope(ctx.rope)} is supplied by the domain owner; equivalent verdicts are only meaningful if that range is treated as the standing definition of \"no material difference.\"`,\n    )\n  }\n  if (ctx.comparator === null) {\n    assumptions.push('No comparator was configured; this run is descriptive, not causal.')\n  }\n  const methods: string[] = [\n    \"Marginal scores summarised with BH-FDR-adjusted Wilcoxon signed-rank q-values and paired Cohen's dz via summaryTable.\",\n    'Paired evidence summarised with a frequentist bootstrap CI on the median delta and a Dirichlet-weight Bayesian bootstrap for Pr(Δ>0) and Pr(Δ∈ROPE) on the mean delta.',\n    `Minimum detectable effect reported per candidate at α=${ctx.mdeAlpha} (two-sided), power=${ctx.mdePower}, standardised by the observed paired-delta SD.`,\n    'Pareto frontier flagged as a separate axis (cost vs quality); a candidate can be on-frontier without winning the paired test.',\n    'Held-out gate decisions, when supplied, override the statistical verdict in the reject direction.',\n  ]\n  const alternatives: string[] = [\n    'Paired t-test rejected: not robust to the heavy-tailed score distributions common in agent benchmarks.',\n    'Unpaired Mann–Whitney rejected: matched scenarios make pairing free; unpaired throws away that variance reduction.',\n    'Sequential / always-valid inference (e-values, mSPRT) is the right tool for iterative sweeps and is out of scope for this single-look report — preregister and run once, or wrap this report in an alpha-spending schedule.',\n    'Hierarchical Bayesian shrinkage across many candidates is future work; the current ranking uses raw paired statistics.',\n  ]\n  const whenNotToApply: string[] = [\n    `Paired N below ${RESEARCH_REPORT_HARD_PAIR_FLOOR} on any candidate — the bootstrap CI is degenerate.`,\n    'Comparator chosen post-hoc by inspecting the same data; q-values are no longer false-discovery-controlled.',\n    'Scenarios not drawn under a stable preregistered protocol; the report can describe the data but cannot anchor a launch decision.',\n    'Score distributions with mid-run shift (judge model swap, rubric change, infra outage) — pair exchangeability is violated.',\n  ]\n  const citations: string[] = [\n    'Benjamini, Y. & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. JRSS B, 57(1), 289–300.',\n    'Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1(6), 80–83.',\n    'Efron, B. (1979). Bootstrap methods: another look at the jackknife. Annals of Statistics, 7(1), 1–26.',\n    'Rubin, D. B. (1981). The Bayesian bootstrap. Annals of Statistics, 9(1), 130–134.',\n    'Kruschke, J. K. (2018). Rejecting or accepting parameter values in Bayesian estimation. Advances in Methods and Practices in Psychological Science, 1(2), 270–280. (ROPE.)',\n  ]\n  return { assumptions, methods, alternatives, whenNotToApply, citations }\n}\n\nexport function classifyCandidate(\n  row: SummaryTableRow,\n  ctx: {\n    comparator: string | null\n    posterior: ReturnType<typeof pairedPosterior> | null\n    point?: ParetoPoint\n    gateDecision?: HeldOutGateDecision\n    split: 'search' | 'holdout'\n    fdr: number\n    minPairs: number\n    rope: { low: number; high: number } | null\n  },\n): { decision: ResearchReportDecision; reason: string } {\n  if (ctx.comparator && row.candidateId === ctx.comparator) {\n    return { decision: 'hold', reason: 'Comparator baseline.' }\n  }\n  if (ctx.gateDecision && !ctx.gateDecision.promote) {\n    const code = ctx.gateDecision.rejectionCode ?? 'rejected'\n    return {\n      decision: 'reject',\n      reason: `Held-out gate returned ${code}: ${ctx.gateDecision.reason}`,\n    }\n  }\n  if (!ctx.comparator) {\n    return {\n      decision: ctx.point?.onFrontier ? 'hold' : 'needs_more_data',\n      reason:\n        'No comparator configured; report ranks candidates but cannot anchor a promotion call.',\n    }\n  }\n  if (!ctx.posterior || ctx.posterior.n < RESEARCH_REPORT_HARD_PAIR_FLOOR) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior?.n ?? 0} paired observations; below hard floor of ${RESEARCH_REPORT_HARD_PAIR_FLOOR} for any paired inference.`,\n    }\n  }\n  const ci = ctx.posterior.ci\n  if (ctx.split === 'search') {\n    return {\n      decision: 'hold',\n      reason:\n        'Search-split evidence is descriptive; release decisions require matched holdout runs.',\n    }\n  }\n  // A ZERO-WIDTH interval cannot support any directional or equivalence claim:\n  // every resample was identical, so it says nothing about how far the estimate\n  // could be wrong. It is the shape that sits fully inside any ROPE containing\n  // it (a fake \"practically equivalent\") and clears any threshold below it (a\n  // fake \"promote\"). Both are an absence of evidence, so it is disqualified from\n  // each claim rather than short-circuiting the whole classifier — the\n  // sample-size and MDE diagnostics below are still the more useful answer when\n  // they apply.\n  const degenerate = !Number.isFinite(ci.low) || !Number.isFinite(ci.high) || ci.low === ci.high\n  if (!degenerate && ctx.rope && ci.low >= ctx.rope.low && ci.high <= ctx.rope.high) {\n    return {\n      decision: 'equivalent',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] is fully inside ROPE ${formatRope(ctx.rope)}; candidate is practically equivalent to comparator.`,\n    }\n  }\n  const gainPositive = !degenerate && ci.low > 0\n  const gainNegative = !degenerate && ci.high < 0\n  if (gainNegative) {\n    return {\n      decision: 'reject',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] lies entirely below zero.`,\n    }\n  }\n  if (ctx.posterior.n < ctx.minPairs) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Only ${ctx.posterior.n} paired observations; minimum detectable effect at this N is ${fmt(ctx.posterior.mde)} score units (need ≥ ${ctx.minPairs} pairs to issue a directional verdict).`,\n    }\n  }\n  if (degenerate) {\n    return {\n      decision: 'needs_more_data',\n      reason: `Paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] has zero width — every resample was identical, so the interval carries no information about how far the estimate could be wrong and cannot support a promote or equivalence call.`,\n    }\n  }\n  if (row.qValue !== null && row.qValue <= ctx.fdr && gainPositive) {\n    return {\n      decision: 'promote',\n      reason: `BH-adjusted q=${fmt(row.qValue)} ≤ ${ctx.fdr} and paired-delta CI [${fmt(ci.low)}, ${fmt(ci.high)}] excludes zero; Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)}.`,\n    }\n  }\n  return {\n    decision: 'hold',\n    reason: `Pr(Δ>0)=${fmt(ctx.posterior.prGreaterThanZero)} but CI [${fmt(ci.low)}, ${fmt(ci.high)}] crosses zero; effect not decisive at fdr=${ctx.fdr}.`,\n  }\n}\n\nexport function buildRecommendation(\n  candidates: ResearchReportCandidate[],\n  ctx: {\n    comparator: string | null\n    failureClusters?: FailureClusterReport\n    rope: { low: number; high: number } | null\n    minPairs: number\n    preregistrationHash: string | null\n  },\n): ResearchReportRecommendation {\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  const bestPromote = nonComparator.find((c) => c.decision === 'promote')\n  const bestEquivalent = nonComparator.find((c) => c.decision === 'equivalent')\n  const chosen = bestPromote ?? bestEquivalent ?? nonComparator[0] ?? null\n  const decision: ResearchReportDecision = bestPromote\n    ? 'promote'\n    : nonComparator.some((c) => c.decision === 'needs_more_data')\n      ? 'needs_more_data'\n      : bestEquivalent\n        ? 'equivalent'\n        : nonComparator.some((c) => c.decision === 'hold')\n          ? 'hold'\n          : 'reject'\n\n  const rationale: string[] = []\n  const risks: string[] = []\n  const nextActions: string[] = []\n\n  if (chosen) {\n    rationale.push(`${chosen.candidateId}: ${chosen.decisionReason}`)\n    if (chosen.gainCi) {\n      const probSummary =\n        chosen.prGreaterThanZero !== null ? `, Pr(Δ>0)=${fmt(chosen.prGreaterThanZero)}` : ''\n      rationale.push(\n        `Median paired gain CI: [${fmt(chosen.gainCi.low)}, ${fmt(chosen.gainCi.high)}]${probSummary}.`,\n      )\n    }\n    if (chosen.mde !== null && Number.isFinite(chosen.mde)) {\n      rationale.push(`MDE at current paired N=${chosen.pairedN}: ${fmt(chosen.mde)} score units.`)\n    }\n  }\n  if (!ctx.comparator) {\n    risks.push('No comparator was configured; verdict is descriptive, not causal.')\n    nextActions.push('Re-run with a stable comparator candidate for paired inference.')\n  }\n  if (!ctx.preregistrationHash) {\n    risks.push(\n      'No preregistration hash supplied; readers cannot verify the analysis was specified before data inspection.',\n    )\n    nextActions.push(\n      'Sign a HypothesisManifest before the next sweep and pass `preregistrationHash` so the report cites it.',\n    )\n  }\n  if (ctx.rope === null && nonComparator.length > 0) {\n    risks.push(\n      'No ROPE configured; the report cannot distinguish \"equivalent\" from \"inconclusive\".',\n    )\n    nextActions.push(\n      'Define a domain-specific Region of Practical Equivalence and pass it to lock in the equivalence threshold.',\n    )\n  }\n  const inconclusive = nonComparator.filter((c) => c.decision === 'needs_more_data')\n  if (inconclusive.length > 0) {\n    const worst = inconclusive.reduce((a, b) => (b.pairedN < a.pairedN ? b : a))\n    risks.push(\n      `${inconclusive.length} candidate(s) below soft floor (${ctx.minPairs} pairs); thinnest is ${worst.candidateId} with ${worst.pairedN}.`,\n    )\n    nextActions.push(\n      `Collect at least ${ctx.minPairs - worst.pairedN} more matched holdout runs for ${worst.candidateId}.`,\n    )\n  }\n  const rejected = nonComparator.filter((c) => c.decision === 'reject')\n  if (rejected.length > 0) {\n    risks.push(\n      `${rejected.length} candidate(s) failed the paired test or held-out gate; do not ship those variants.`,\n    )\n  }\n  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {\n    const top = ctx.failureClusters.clusters[0]!\n    risks.push(`Top failure cluster: ${top.failureClass} across ${top.runCount} run(s).`)\n    nextActions.push('Prioritize the largest failure cluster before broad rollout.')\n  }\n  if (decision === 'promote') {\n    nextActions.push('Ship behind the existing promotion gate and monitor canaries.')\n  } else if (decision === 'hold') {\n    nextActions.push('Keep current production candidate while expanding holdout evidence.')\n  } else if (decision === 'equivalent') {\n    nextActions.push(\n      'Either keep the comparator (no quality regression) or promote on cost/latency grounds — equivalence does not justify either; the choice is a product decision, not a stats one.',\n    )\n  } else if (decision === 'reject') {\n    nextActions.push(\n      'Do not promote this sweep; inspect failures and generate a revised candidate.',\n    )\n  }\n\n  return {\n    decision,\n    candidateId: chosen?.candidateId ?? null,\n    rationale,\n    risks,\n    nextActions,\n  }\n}\n\nexport function buildExecutiveSummary(\n  candidates: ResearchReportCandidate[],\n  recommendation: ResearchReportRecommendation,\n  ctx: {\n    comparator: string | null\n    split: 'search' | 'holdout'\n    failureClusters?: FailureClusterReport\n    preregistrationHash: string | null\n  },\n): string[] {\n  const lines: string[] = []\n  const nonComparator = candidates.filter((c) => c.candidateId !== ctx.comparator)\n  lines.push(\n    `Evaluated ${nonComparator.length} candidate(s) on the ${ctx.split} split${ctx.comparator ? ` against ${ctx.comparator}` : ''}.`,\n  )\n  lines.push(\n    `Recommendation: ${recommendation.decision}${recommendation.candidateId ? ` ${recommendation.candidateId}` : ''}.`,\n  )\n  const promoted = nonComparator.filter((c) => c.decision === 'promote').length\n  const held = nonComparator.filter((c) => c.decision === 'hold').length\n  const equivalent = nonComparator.filter((c) => c.decision === 'equivalent').length\n  const rejected = nonComparator.filter((c) => c.decision === 'reject').length\n  const more = nonComparator.filter((c) => c.decision === 'needs_more_data').length\n  lines.push(\n    `Decision mix: ${promoted} promote, ${equivalent} equivalent, ${held} hold, ${rejected} reject, ${more} need more data.`,\n  )\n  const frontier = nonComparator.filter((c) => c.onParetoFrontier).map((c) => c.candidateId)\n  if (frontier.length > 0) lines.push(`Pareto-frontier candidates: ${frontier.join(', ')}.`)\n  if (ctx.failureClusters) {\n    lines.push(\n      `Failure clustering found ${ctx.failureClusters.totalFailures}/${ctx.failureClusters.totalRuns} failed runs across ${ctx.failureClusters.clusters.length} reportable cluster(s).`,\n    )\n  }\n  lines.push(\n    ctx.preregistrationHash\n      ? `Preregistered analysis: ${ctx.preregistrationHash.slice(0, 12)}…`\n      : 'Analysis is post-hoc — no preregistration hash supplied.',\n  )\n  return lines\n}\n\nexport function decisionWeight(decision: ResearchReportDecision): number {\n  if (decision === 'promote') return 5\n  if (decision === 'equivalent') return 4\n  if (decision === 'hold') return 3\n  if (decision === 'needs_more_data') return 2\n  return 1\n}\n","/**\n * Standardized effect sizes: Cohen's d family and Cliff's delta, with the\n * shared degenerate-input contract (null where the effect is undefined).\n */\n\nimport { ValidationError } from '../errors'\n\n/**\n * Cohen's d — standardized effect size for two independent groups.\n * Positive d means group b has higher mean than group a.\n * Rule of thumb: |d| < 0.2 negligible, 0.2–0.5 small, 0.5–0.8 medium, > 0.8 large.\n *\n * Returns null where the standardized effect is undefined: fewer than two\n * observations in either group, or a zero pooled standard deviation with\n * unequal means. Null is NOT \"no effect\" — zero within-group spread across a\n * real mean gap is an unbounded effect, the opposite of negligible. Equal\n * means with zero spread is a genuine 0. Same contract as\n * {@link pairedCohensDz}.\n */\nexport function cohensD(a: number[], b: number[]): number | null {\n  if (a.length < 2 || b.length < 2) return null\n  const meanA = a.reduce((x, y) => x + y, 0) / a.length\n  const meanB = b.reduce((x, y) => x + y, 0) / b.length\n  const varA = a.reduce((acc, x) => acc + (x - meanA) ** 2, 0) / (a.length - 1)\n  const varB = b.reduce((acc, x) => acc + (x - meanB) ** 2, 0) / (b.length - 1)\n  const pooled = Math.sqrt(\n    ((a.length - 1) * varA + (b.length - 1) * varB) / (a.length + b.length - 2),\n  )\n  if (pooled === 0) return meanB === meanA ? 0 : null\n  return (meanB - meanA) / pooled\n}\n\n/**\n * Cohen's dz for paired observations: mean(after - before) divided by the\n * sample standard deviation of those within-pair deltas.\n *\n * Returns null when fewer than two pairs exist or a non-zero constant delta\n * has zero observed variance. In that case the standardized effect is\n * undefined, not an arbitrarily large finite number.\n */\nexport function pairedCohensDz(before: number[], after: number[]): number | null {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedCohensDz: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  if (before.length < 2) return null\n  const deltas = before.map((value, index) => after[index]! - value)\n  if (deltas.some((value) => !Number.isFinite(value))) {\n    throw new ValidationError('pairedCohensDz: all paired values must be finite')\n  }\n  const meanDelta = deltas.reduce((sum, value) => sum + value, 0) / deltas.length\n  const variance =\n    deltas.reduce((sum, value) => sum + (value - meanDelta) ** 2, 0) / (deltas.length - 1)\n  const standardDeviation = Math.sqrt(variance)\n  const scale = Math.max(1, Math.abs(meanDelta), ...deltas.map(Math.abs))\n  if (standardDeviation <= Number.EPSILON * scale) return meanDelta === 0 ? 0 : null\n  return meanDelta / standardDeviation\n}\n\nexport type CliffsMagnitude = 'negligible' | 'small' | 'medium' | 'large'\n\n/**\n * Cliff's delta — a non-parametric effect size for two independent samples.\n * `δ = (#(after > before) − #(after < before)) / (n_before · n_after)`,\n * ranging [-1, 1]. Positive ⇒ `after` tends to exceed `before` (improvement).\n *\n * Distribution-free counterpart to Cohen's d: no normality assumption, robust\n * to the bounded/skewed score distributions judges produce. Pairs with\n * `pairedBootstrap` / `wilcoxonSignedRank` for the non-parametric reporting\n * path. Returns 0 when either sample is empty.\n */\nexport function cliffsDelta(before: number[], after: number[]): number {\n  const n = before.length * after.length\n  if (n === 0) return 0\n  let dominance = 0\n  for (const a of after) {\n    for (const b of before) {\n      if (a > b) dominance += 1\n      else if (a < b) dominance -= 1\n    }\n  }\n  return dominance / n\n}\n\n/**\n * Map a Cliff's delta to a qualitative magnitude using the standard\n * Romano et al. thresholds (|δ|): <0.147 negligible, <0.33 small,\n * <0.474 medium, else large.\n */\nexport function interpretCliffs(delta: number): CliffsMagnitude {\n  const d = Math.abs(delta)\n  if (d < 0.147) return 'negligible'\n  if (d < 0.33) return 'small'\n  if (d < 0.474) return 'medium'\n  return 'large'\n}\n","/** Matched split scores and paired statistical evidence for report consumers. */\n\nimport { pairRunRecords } from './paired-arms'\nimport { observedSplitScore, type ScorePreference } from './rollout/reward'\nimport type { RunRecord } from './run-record'\nimport { pairedBootstrap, pairedMde } from './statistics'\nimport { makeRng } from './statistics/internal'\n\nexport function pairScoresByKey(\n  candidate: RunRecord[],\n  baseline: RunRecord[],\n  split: ScorePreference,\n): {\n  before: number[]\n  after: number[]\n  unpairedCandidateN: number\n  unpairedComparatorN: number\n} {\n  // RAW, split-exact reads through `observedSplitScore` — the single allowed\n  // spelling of the split-score derivation (see rollout/score-derivation-guard).\n  const scoreOf = (run: RunRecord): number | undefined => {\n    const v = observedSplitScore(run, split)\n    return typeof v === 'number' && Number.isFinite(v) ? v : undefined\n  }\n  const scoredCandidate = candidate.filter((run) => scoreOf(run) !== undefined)\n  const scoredBaseline = baseline.filter((run) => scoreOf(run) !== undefined)\n  const pairing = pairRunRecords(scoredBaseline, scoredCandidate)\n  return {\n    before: pairing.pairs.map((pair) => scoreOf(pair.baseline) as number),\n    after: pairing.pairs.map((pair) => scoreOf(pair.treatment) as number),\n    unpairedCandidateN: pairing.unpairedTreatment.length,\n    unpairedComparatorN: pairing.unpairedBaseline.length,\n  }\n}\n\n/**\n * Internal: paired posterior summary on (candidate − comparator) deltas.\n *\n * Returns the bootstrap CI on the median (matching `gainHistogram`) plus\n * Bayesian-bootstrap posterior summaries Pr(Δ>0) and Pr(Δ∈ROPE) computed\n * from Dirichlet(1, ..., 1) weights on the observed paired deltas\n * (Rubin 1981), and the\n * minimum detectable paired effect at the configured power and α.\n *\n * `null` is returned when no paired observations exist; callers must\n * gate on `n` before consuming the bootstrap statistics.\n */\nexport function pairedPosterior(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: {\n    split: 'search' | 'holdout'\n    confidence: number\n    seed?: number\n    rope: { low: number; high: number } | null\n    mdePower: number\n    mdeAlpha: number\n  },\n): {\n  n: number\n  meanDelta: number\n  medianDelta: number\n  sdDelta: number\n  ci: { low: number; high: number }\n  prGreaterThanZero: number\n  prInRope: number | null\n  mde: number\n} | null {\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === opts.split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === opts.split)\n  const { before, after } = pairScoresByKey(candidate, baseline, opts.split)\n  const n = before.length\n  if (n === 0) return null\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const meanDelta = deltas.reduce((s, x) => s + x, 0) / n\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const medianDelta = 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  const meanSamples = bayesianBootstrapMeanSamples(deltas, 2000, opts.seed)\n  const prGreaterThanZero =\n    meanSamples.length === 0 ? 0 : meanSamples.filter((s) => s > 0).length / meanSamples.length\n  const prInRope =\n    opts.rope === null || meanSamples.length === 0\n      ? null\n      : meanSamples.filter((s) => s >= opts.rope!.low && s <= opts.rope!.high).length /\n        meanSamples.length\n\n  const dStandardised = pairedMde({ nPaired: n, alpha: opts.mdeAlpha, power: opts.mdePower })\n  const mde = sdDelta === 0 ? 0 : dStandardised * sdDelta\n\n  return {\n    n,\n    meanDelta,\n    medianDelta,\n    sdDelta,\n    ci: { low: ci.low, high: ci.high },\n    prGreaterThanZero,\n    prInRope,\n    mde,\n  }\n}\n\nfunction bayesianBootstrapMeanSamples(\n  deltas: number[],\n  resamples: number,\n  seed?: number,\n): number[] {\n  const n = deltas.length\n  if (n === 0) return []\n  if (n === 1) return new Array<number>(resamples).fill(deltas[0]!)\n  const rng = makeRng(seed, deltas)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let weightedSum = 0\n    let weightSum = 0\n    for (let k = 0; k < n; k++) {\n      const weight = -Math.log(Math.max(Number.MIN_VALUE, rng()))\n      weightedSum += deltas[k]! * weight\n      weightSum += weight\n    }\n    samples[b] = weightedSum / weightSum\n  }\n  return samples\n}\n\nfunction stdev(xs: number[], mean: number): number {\n  if (xs.length < 2) return 0\n  let sse = 0\n  for (const x of xs) sse += (x - mean) ** 2\n  return Math.sqrt(sse / (xs.length - 1))\n}\n\nexport function 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","/** Tables and library-independent chart specifications from scored runs. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport { fmt } from './research-report-render'\nimport { observedSplitScore } from './rollout/reward'\nimport type { RunRecord } from './run-record'\nimport {\n  benjaminiHochberg,\n  confidenceInterval,\n  pairedBootstrap,\n  pairedCohensDz,\n  wilcoxonSignedRank,\n} from './statistics'\nimport { mean as avg } from './statistics/mean'\nimport { medianOfSorted, pairScoresByKey } from './summary-report-evidence'\n\n// ── summaryTable ───────────────────────────────────────────────────────\n\nexport interface SummaryTableOptions {\n  /** Comparator candidate id. Wilcoxon + paired Cohen's dz are computed\n   *  versus this candidate. Required for paired stats columns. */\n  comparator?: string\n  /** Which split to read scores from. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the bootstrap CI on the mean. Default 0.95. */\n  confidence?: number\n  /** FDR for BH adjustment of the comparison p-values. Default 0.05. */\n  fdr?: number\n}\n\nexport interface SummaryTableRow {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  /** BH-adjusted q-value vs comparator, or null when unavailable. */\n  qValue: number | null\n  /** Paired Cohen's dz vs comparator, or null when the paired variance is zero. */\n  cohensD: number | null\n  /** Matched observations used for paired comparison, or null on the comparator row. */\n  pairedN: number | null\n  /** Candidate observations without a comparator match. */\n  unpairedCandidateN: number | null\n  /** Comparator observations without a candidate match. */\n  unpairedComparatorN: number | null\n}\n\nexport interface SummaryTable {\n  rows: SummaryTableRow[]\n  comparator: string | null\n  split: 'search' | 'holdout'\n  /** Pre-rendered markdown — drop into a paper or PR. */\n  markdown: string\n}\n\n/**\n * Table 1 helper. Buckets runs by `candidateId`, computes mean +\n * bootstrap CI on the chosen split, and (when a comparator is given)\n * BH-adjusted Wilcoxon p + paired Cohen's dz versus that comparator.\n */\nexport function summaryTable(runs: RunRecord[], opts: SummaryTableOptions = {}): SummaryTable {\n  const split = opts.split ?? 'holdout'\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const comparator = opts.comparator ?? null\n\n  const byCandidate = new Map<string, { runs: RunRecord[]; scores: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    // RAW (`observedSplitScore`): every figure in this module reports what the\n    // runs scored. Zeroing a gamed run here would erase it from the very table\n    // a reviewer reads to notice it, and the split is already filtered above,\n    // so the exact-split reader is the right one — no cross-split fallback.\n    const v = observedSplitScore(r, split)\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    const bucket = byCandidate.get(r.candidateId) ?? { runs: [], scores: [] }\n    bucket.runs.push(r)\n    bucket.scores.push(v)\n    byCandidate.set(r.candidateId, bucket)\n  }\n\n  const candidateIds = [...byCandidate.keys()].sort()\n  const compRuns = comparator ? byCandidate.get(comparator) : undefined\n\n  // First pass: per-candidate means + CIs + raw p-values.\n  const tentative: Array<SummaryTableRow & { rawP: number | null }> = []\n  for (const id of candidateIds) {\n    const bucket = byCandidate.get(id)!\n    const ci = confidenceInterval(bucket.scores, confidence)\n    let rawP: number | null = null\n    let d: number | null = null\n    let pairedN: number | null = null\n    let unpairedCandidateN: number | null = null\n    let unpairedComparatorN: number | null = null\n    if (comparator && compRuns && id !== comparator) {\n      const paired = pairScoresByKey(bucket.runs, compRuns.runs, split)\n      pairedN = paired.before.length\n      unpairedCandidateN = paired.unpairedCandidateN\n      unpairedComparatorN = paired.unpairedComparatorN\n      if (paired.before.length >= 6) {\n        rawP = wilcoxonSignedRank(paired.before, paired.after).p\n      }\n      d = pairedCohensDz(paired.before, paired.after)\n    }\n    tentative.push({\n      candidateId: id,\n      n: bucket.scores.length,\n      mean: ci.mean,\n      ciLow: ci.lower,\n      ciHigh: ci.upper,\n      qValue: rawP,\n      cohensD: d,\n      pairedN,\n      unpairedCandidateN,\n      unpairedComparatorN,\n      rawP,\n    })\n  }\n\n  // BH-adjust across the comparison set (skip NaN rows / the\n  // comparator itself). Adjustment is a no-op when there are 0 or 1\n  // comparators.\n  if (comparator) {\n    const idxs: number[] = []\n    const ps: number[] = []\n    for (let i = 0; i < tentative.length; i++) {\n      const r = tentative[i]!\n      if (r.candidateId === comparator) continue\n      if (r.rawP === null || !Number.isFinite(r.rawP)) continue\n      idxs.push(i)\n      ps.push(r.rawP)\n    }\n    if (ps.length > 0) {\n      const { qValues } = benjaminiHochberg(ps, fdr)\n      for (let k = 0; k < idxs.length; k++) {\n        tentative[idxs[k]!]!.qValue = qValues[k]!\n      }\n    }\n  }\n\n  const rows = tentative.map(({ rawP: _rawP, ...rest }) => rest)\n  const markdown = renderSummaryTableMarkdown(rows, comparator, split)\n  return { rows, comparator, split, markdown }\n}\n\nfunction renderSummaryTableMarkdown(\n  rows: SummaryTableRow[],\n  comparator: string | null,\n  split: 'search' | 'holdout',\n): string {\n  const lines: string[] = []\n  const cmpLabel = comparator ? ` (vs ${comparator})` : ''\n  lines.push(`Summary Table — ${split} split${cmpLabel}`)\n  lines.push('')\n  lines.push(\"| Candidate | N | Pairs | Unpaired | Mean | 95% CI | q (BH) | Cohen's dz |\")\n  lines.push('|---|---:|---:|---:|---:|---|---:|---:|')\n  for (const r of rows) {\n    const ci = `[${fmt(r.ciLow)}, ${fmt(r.ciHigh)}]`\n    const q = r.qValue !== null ? r.qValue.toFixed(4) : '—'\n    const d = r.cohensD !== null ? r.cohensD.toFixed(3) : '—'\n    const paired = r.pairedN ?? '—'\n    const unpaired =\n      r.unpairedCandidateN === null || r.unpairedComparatorN === null\n        ? '—'\n        : `${r.unpairedCandidateN}/${r.unpairedComparatorN}`\n    lines.push(\n      `| ${r.candidateId} | ${r.n} | ${paired} | ${unpaired} | ${fmt(r.mean)} | ${ci} | ${q} | ${d} |`,\n    )\n  }\n  return lines.join('\\n')\n}\n\n// ── paretoChart ─────────────────────────────────────────────────────\n\nexport interface ParetoPoint {\n  candidateId: string\n  /** Mean USD cost per run on the chosen split. */\n  cost: number\n  /** Mean score on the chosen split. */\n  quality: number\n  /** Number of runs that informed this point. */\n  n: number\n  /** Whether this candidate is on the Pareto frontier — high\n   *  quality, low cost, no dominator. */\n  onFrontier: boolean\n  /** Optional gate verdict for this candidate, if a `HeldOutGateDecision`\n   *  for it was passed in. */\n  gate?: 'promote' | 'reject'\n}\n\nexport interface ParetoFigureSpec {\n  kind: 'pareto-cost-quality'\n  split: 'search' | 'holdout'\n  points: ParetoPoint[]\n  axes: { x: 'costUsd'; y: 'score' }\n}\n\n/**\n * Cost vs quality scatter spec. `gateDecisions` is keyed by\n * candidate id; if present, every point picks up the gate verdict\n * for overlay.\n */\nexport function paretoChart(\n  runs: RunRecord[],\n  opts: {\n    split?: 'search' | 'holdout'\n    gateDecisions?: Record<string, HeldOutGateDecision>\n  } = {},\n): ParetoFigureSpec {\n  const split = opts.split ?? 'holdout'\n\n  const buckets = new Map<string, { cost: number[]; quality: number[] }>()\n  for (const r of runs) {\n    if (r.splitTag !== split) continue\n    const v = observedSplitScore(r, split)\n    if (typeof v !== 'number' || !Number.isFinite(v)) continue\n    if (r.costUsd === null) continue\n    const bucket = buckets.get(r.candidateId) ?? { cost: [], quality: [] }\n    bucket.cost.push(r.costUsd)\n    bucket.quality.push(v)\n    buckets.set(r.candidateId, bucket)\n  }\n\n  const points: ParetoPoint[] = []\n  for (const [candidateId, bucket] of buckets.entries()) {\n    points.push({\n      candidateId,\n      cost: avg(bucket.cost),\n      quality: avg(bucket.quality),\n      n: bucket.cost.length,\n      onFrontier: false,\n      gate: opts.gateDecisions?.[candidateId]\n        ? gateLabel(opts.gateDecisions[candidateId]!)\n        : undefined,\n    })\n  }\n\n  // Pareto: minimize cost, maximize quality. A point is dominated if\n  // some other point has lower-or-equal cost AND higher-or-equal\n  // quality with strict inequality somewhere.\n  for (const p of points) {\n    p.onFrontier = !points.some((q) => q !== p && dominates(q, p))\n  }\n\n  return {\n    kind: 'pareto-cost-quality',\n    split,\n    axes: { x: 'costUsd', y: 'score' },\n    points,\n  }\n}\n\nfunction dominates(a: ParetoPoint, b: ParetoPoint): boolean {\n  return a.cost <= b.cost && a.quality >= b.quality && (a.cost < b.cost || a.quality > b.quality)\n}\n\nexport function gateLabel(d: HeldOutGateDecision): ParetoPoint['gate'] {\n  return d.promote ? 'promote' : 'reject'\n}\n\n// ── gainHistogram ───────────────────────────────────────────\n\nexport interface GainDistributionBin {\n  /** Inclusive lower edge. */\n  lo: number\n  /** Exclusive upper edge (or inclusive if it's the last bin). */\n  hi: number\n  /** Number of pairs whose delta lands in this bin. */\n  count: number\n}\n\nexport interface GainDistributionFigureSpec {\n  kind: 'gain-distribution'\n  candidateId: string\n  comparator: string\n  split: 'search' | 'holdout'\n  /** Number of pairs used. */\n  n: number\n  /** Candidate rows without a comparator match. */\n  unpairedCandidateN: number\n  /** Comparator rows without a candidate match. */\n  unpairedComparatorN: number\n  bins: GainDistributionBin[]\n  median: number | null\n  ci: { low: number; high: number } | null\n}\n\nexport interface GainDistributionOptions {\n  /** Number of histogram bins. Default 11 (so the centre is exact at 0). */\n  bins?: number\n  /** Which split to use. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level for the CI. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Deterministic seed. */\n  seed?: number\n}\n\n/**\n * Held-out improvement distribution: per-pair delta (candidate −\n * comparator), histogrammed. Includes the bootstrap CI on the median\n * delta — same primitive the promotion gate uses.\n */\nexport function gainHistogram(\n  runs: RunRecord[],\n  candidateId: string,\n  comparator: string,\n  opts: GainDistributionOptions = {},\n): GainDistributionFigureSpec {\n  const split = opts.split ?? 'holdout'\n  const binCount = opts.bins ?? 11\n  if (binCount < 1) throw new Error('gainHistogram: bins must be ≥ 1')\n\n  const candidate = runs.filter((r) => r.candidateId === candidateId && r.splitTag === split)\n  const baseline = runs.filter((r) => r.candidateId === comparator && r.splitTag === split)\n  // pairScoresByKey returns before=baseline-score, after=candidate-score\n  // for each canonical work-identity pair where both sides recorded a\n  // valid score on this split. delta = after - before = candidate - baseline.\n  const { before, after, unpairedCandidateN, unpairedComparatorN } = pairScoresByKey(\n    candidate,\n    baseline,\n    split,\n  )\n  const n = before.length\n\n  if (n === 0) {\n    return {\n      kind: 'gain-distribution',\n      candidateId,\n      comparator,\n      split,\n      n: 0,\n      unpairedCandidateN,\n      unpairedComparatorN,\n      bins: [],\n      median: null,\n      ci: null,\n    }\n  }\n\n  const deltas = before.map((b, i) => after[i]! - b)\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const median = 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    unpairedCandidateN,\n    unpairedComparatorN,\n    bins,\n    median,\n    ci: { low: ci.low, high: ci.high },\n  }\n}\n","/** Assemble research report evidence, guidance, figures, and rendered documents. */\n\nimport type { HeldOutGateDecision } from './held-out-gate'\nimport type { FailureClusterReport } from './pipelines/failure-cluster'\nimport { hashJson } from './pre-registration'\nimport { renderResearchHtml } from './reporting-html'\nimport {\n  buildExecutiveSummary,\n  buildMethodology,\n  buildRecommendation,\n  classifyCandidate,\n  decisionWeight,\n  RESEARCH_REPORT_HARD_PAIR_FLOOR,\n} from './research-report-analysis'\nimport { renderResearchMarkdown } from './research-report-render'\nimport type { RunRecord } from './run-record'\nimport { pairedPosterior } from './summary-report-evidence'\nimport {\n  type GainDistributionFigureSpec,\n  gainHistogram,\n  gateLabel,\n  type ParetoFigureSpec,\n  type ParetoPoint,\n  paretoChart,\n  type SummaryTable,\n  summaryTable,\n} from './summary-report-figures'\n\nexport { RESEARCH_REPORT_HARD_PAIR_FLOOR } from './research-report-analysis'\n\nexport type ResearchReportDecision =\n  | 'promote'\n  | 'hold'\n  | 'reject'\n  | 'equivalent'\n  | 'needs_more_data'\n\nexport interface ResearchReportOptions {\n  /** Human-readable report title. */\n  title?: string\n  /** Comparator candidate id. Required for statistical decision guidance. */\n  comparator?: string\n  /** Which split to use for the primary decision. Default 'holdout'. */\n  split?: 'search' | 'holdout'\n  /** Confidence level used by lower-level report helpers. Default 0.95. */\n  confidence?: number\n  /** FDR threshold for q-values. Default 0.05. */\n  fdr?: number\n  /**\n   * Soft floor on paired observations before issuing a directional\n   * promote / reject. Below this we report `needs_more_data` and surface the\n   * minimum detectable effect at the current N. Default 20 — chosen so the\n   * Wilcoxon signed-rank approximation is reasonable and so the paired\n   * bootstrap CI has non-degenerate coverage. Hard floor is enforced at\n   * `RESEARCH_REPORT_HARD_PAIR_FLOOR` (6) regardless of this value.\n   */\n  minPairs?: number\n  /**\n   * Region of Practical Equivalence on the paired delta. When a candidate's\n   * paired-delta CI is fully contained in `[low, high]`, the decision is\n   * `equivalent` rather than `hold`. Sourced from the domain owner — there is\n   * no statistically-defensible default.\n   */\n  rope?: { low: number; high: number }\n  /**\n   * Power for the minimum detectable effect (MDE) reported on each candidate.\n   * Default 0.8.\n   */\n  mdePower?: number\n  /**\n   * Two-sided alpha for the MDE. Default matches `fdr` so the reported MDE\n   * lines up with the test the report actually runs.\n   */\n  mdeAlpha?: number\n  /** Optional held-out gate decisions keyed by candidate id. */\n  gateDecisions?: Record<string, HeldOutGateDecision>\n  /** Optional failure clusters from failureClusterView. */\n  failureClusters?: FailureClusterReport\n  /** Build gain histograms for these candidates. Defaults to all non-comparator candidates. */\n  candidateIds?: string[]\n  /** Deterministic bootstrap seed passed to gainHistogram and the posterior helper. */\n  seed?: number\n  /** Report timestamp. Defaults to current time. */\n  generatedAt?: string\n  /**\n   * Hash of a preregistered protocol (e.g. `signManifest({...}).contentHash`).\n   * Embedded verbatim in the report so the analysis can be cited as the\n   * preregistered one rather than a post-hoc fishing expedition.\n   */\n  preregistrationHash?: string\n}\n\nexport interface ResearchReportRecommendation {\n  decision: ResearchReportDecision\n  candidateId: string | null\n  rationale: string[]\n  risks: string[]\n  nextActions: string[]\n}\n\nexport interface ResearchReportCandidate {\n  candidateId: string\n  n: number\n  mean: number\n  ciLow: number\n  ciHigh: number\n  qValue: number | null\n  cohensD: number | null\n  meanDeltaVsComparator: number | null\n  pairedN: number\n  medianGain: number | null\n  meanGain: number | null\n  gainCi: { low: number; high: number } | null\n  /**\n   * Bayesian-bootstrap posterior summaries on the paired mean delta.\n   * Dirichlet(1, ..., 1) weights represent uncertainty over the empirical\n   * distribution of matched deltas.\n   */\n  prGreaterThanZero: number | null\n  prInRope: number | null\n  /**\n   * Minimum detectable effect (in score units) at the candidate's paired N,\n   * the configured power, and the configured alpha. Standardised by the\n   * observed paired-delta SD and inverted via `requiredSampleSize`. Reported\n   * for every candidate so a `needs_more_data` verdict is actionable.\n   */\n  mde: number | null\n  onParetoFrontier: boolean\n  gate?: ParetoPoint['gate']\n  decision: ResearchReportDecision\n  decisionReason: string\n}\n\nexport interface ResearchReportMethodology {\n  /**\n   * Plain-language assumptions the report depends on. Read these first when\n   * deciding whether the verdict is load-bearing for a launch decision.\n   */\n  assumptions: string[]\n  /** Tests and estimators the verdict was computed from. */\n  methods: string[]\n  /** Alternatives the author considered and why this report didn't take them. */\n  alternatives: string[]\n  /** Failure modes — when this report should NOT drive a decision. */\n  whenNotToApply: string[]\n  /** Citations for the methodological choices above. */\n  citations: string[]\n}\n\nexport interface ResearchReport {\n  kind: 'agent-eval-research-report'\n  title: string\n  generatedAt: string\n  split: 'search' | 'holdout'\n  comparator: string | null\n  /**\n   * SHA-256 over the canonicalised set of `(runId, candidateId, split)` triples\n   * the report was computed from, plus the comparator and split. Stable across\n   * key insertion order; recomputable by the reader to verify provenance.\n   */\n  runFingerprint: string\n  preregistrationHash: string | null\n  rope: { low: number; high: number } | null\n  executiveSummary: string[]\n  recommendation: ResearchReportRecommendation\n  candidates: ResearchReportCandidate[]\n  summary: SummaryTable\n  charts: {\n    pareto: ParetoFigureSpec\n    gains: GainDistributionFigureSpec[]\n  }\n  methodology: ResearchReportMethodology\n  failureClusters?: FailureClusterReport\n  markdown: string\n  html: string\n}\n\n/**\n * Executive research report for CPO / AI-lead / launch-review consumption.\n *\n * Composes:\n *   - `summaryTable`         marginal stats with BH-FDR-adjusted q-values\n *   - `paretoChart`           cost-vs-quality frontier with gate overlay\n *   - `gainHistogram`         per-candidate paired-delta distribution\n *   - paired posterior (this file): bootstrap CI on median, Bayesian-bootstrap Pr(Δ>0),\n *                              Pr(Δ∈ROPE), MDE at the configured power\n *\n * Decisions are made on paired evidence — never on marginal means alone —\n * and respect any held-out gate decision the caller passes through. The\n * report embeds a SHA-256 fingerprint of the input run set and, optionally,\n * the hash of a preregistered protocol so a downstream reader can verify\n * provenance and that the analysis was the preregistered one.\n *\n * Async because the fingerprint uses Web Crypto via `hashJson`; deterministic\n * for any fixed `runs`, `seed`, and ROPE.\n */\nexport async function researchReport(\n  runs: RunRecord[],\n  opts: ResearchReportOptions = {},\n): Promise<ResearchReport> {\n  const split = opts.split ?? 'holdout'\n  const comparator = opts.comparator ?? null\n  const confidence = opts.confidence ?? 0.95\n  const fdr = opts.fdr ?? 0.05\n  const minPairs = Math.max(opts.minPairs ?? 20, RESEARCH_REPORT_HARD_PAIR_FLOOR)\n  const rope = opts.rope ?? null\n  const mdePower = opts.mdePower ?? 0.8\n  const mdeAlpha = opts.mdeAlpha ?? fdr\n  const title = opts.title ?? 'Agent Evaluation Research Report'\n  const generatedAt = opts.generatedAt ?? new Date().toISOString()\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  if (rope && !(Number.isFinite(rope.low) && Number.isFinite(rope.high) && rope.low <= rope.high)) {\n    throw new Error(\n      `researchReport: rope must satisfy low ≤ high with finite bounds, got ${JSON.stringify(rope)}`,\n    )\n  }\n\n  const summary = summaryTable(runs, {\n    comparator: comparator ?? undefined,\n    split,\n    confidence,\n    fdr,\n  })\n  const pareto = paretoChart(runs, { split, gateDecisions: opts.gateDecisions })\n  const candidateIds =\n    opts.candidateIds ?? summary.rows.map((r) => r.candidateId).filter((id) => id !== comparator)\n  const gains = comparator\n    ? candidateIds.map((id) =>\n        gainHistogram(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n        }),\n      )\n    : []\n\n  const gainByCandidate = new Map(gains.map((g) => [g.candidateId, g]))\n  const paretoByCandidate = new Map(pareto.points.map((p) => [p.candidateId, p]))\n  const posteriorByCandidate = new Map<string, ReturnType<typeof pairedPosterior>>()\n  if (comparator) {\n    for (const id of candidateIds) {\n      posteriorByCandidate.set(\n        id,\n        pairedPosterior(runs, id, comparator, {\n          split,\n          confidence,\n          seed: opts.seed,\n          rope,\n          mdePower,\n          mdeAlpha,\n        }),\n      )\n    }\n  }\n\n  const candidates = summary.rows\n    .map((row) => {\n      const gain = gainByCandidate.get(row.candidateId)\n      const point = paretoByCandidate.get(row.candidateId)\n      const posterior = posteriorByCandidate.get(row.candidateId) ?? null\n      const gateDecision = opts.gateDecisions?.[row.candidateId]\n      const classified = classifyCandidate(row, {\n        comparator,\n        posterior,\n        point,\n        gateDecision,\n        split,\n        fdr,\n        minPairs,\n        rope,\n      })\n      return {\n        candidateId: row.candidateId,\n        n: row.n,\n        mean: row.mean,\n        ciLow: row.ciLow,\n        ciHigh: row.ciHigh,\n        qValue: row.qValue,\n        cohensD: row.cohensD,\n        meanDeltaVsComparator: posterior ? posterior.meanDelta : null,\n        pairedN: posterior?.n ?? gain?.n ?? 0,\n        medianGain: posterior ? posterior.medianDelta : gain ? gain.median : null,\n        meanGain: posterior ? posterior.meanDelta : null,\n        gainCi: posterior ? posterior.ci : gain ? gain.ci : null,\n        prGreaterThanZero: posterior ? posterior.prGreaterThanZero : null,\n        prInRope: posterior ? posterior.prInRope : null,\n        mde: posterior ? posterior.mde : null,\n        onParetoFrontier: point?.onFrontier ?? false,\n        gate: gateDecision ? gateLabel(gateDecision) : undefined,\n        decision: classified.decision,\n        decisionReason: classified.reason,\n      } satisfies ResearchReportCandidate\n    })\n    .sort((a, b) => {\n      const decisionRank = decisionWeight(b.decision) - decisionWeight(a.decision)\n      if (decisionRank !== 0) return decisionRank\n      return b.mean - a.mean\n    })\n\n  const recommendation = buildRecommendation(candidates, {\n    comparator,\n    failureClusters: opts.failureClusters,\n    rope,\n    minPairs,\n    preregistrationHash,\n  })\n  const executiveSummary = buildExecutiveSummary(candidates, recommendation, {\n    comparator,\n    split,\n    failureClusters: opts.failureClusters,\n    preregistrationHash,\n  })\n  const methodology = buildMethodology({\n    split,\n    comparator,\n    fdr,\n    minPairs,\n    rope,\n    confidence,\n    mdePower,\n    mdeAlpha,\n  })\n\n  const runFingerprint = await hashJson({\n    triples: runs\n      .filter((r) => r.splitTag === split)\n      .map((r) => ({ runId: r.runId, candidateId: r.candidateId, splitTag: r.splitTag }))\n      .sort((a, b) => a.runId.localeCompare(b.runId)),\n    comparator,\n    split,\n  })\n\n  const markdown = renderResearchMarkdown({\n    title,\n    generatedAt,\n    split,\n    comparator,\n    rope,\n    runFingerprint,\n    preregistrationHash,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    pareto,\n    gains,\n    methodology,\n    failureClusters: opts.failureClusters,\n  })\n  const html = renderResearchHtml(markdown, title)\n\n  return {\n    kind: 'agent-eval-research-report',\n    title,\n    generatedAt,\n    split,\n    comparator,\n    runFingerprint,\n    preregistrationHash,\n    rope,\n    executiveSummary,\n    recommendation,\n    candidates,\n    summary,\n    charts: { pareto, gains },\n    methodology,\n    failureClusters: opts.failureClusters,\n    markdown,\n    html,\n  }\n}\n"],"mappings":";;;;;;;;;;;AAcA,SAAgB,uBAAuB,QAgB5B;CACT,MAAM,QAAkB,CAAC;CACzB,MAAM,KAAK,KAAK,OAAO,OAAO;CAC9B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,kBAAkB,OAAO,aAAa;CACjD,MAAM,KAAK,sBAAsB,OAAO,OAAO;CAC/C,MAAM,KAAK,mBAAmB,OAAO,cAAc,kBAAkB;CACrE,MAAM,KAAK,aAAa,OAAO,OAAO,WAAW,OAAO,IAAI,IAAI,kBAAkB;CAClF,MAAM,KAAK,0BAA0B,OAAO,eAAe,GAAG;CAC9D,MAAM,KACJ,wBAAwB,OAAO,sBAAsB,KAAK,OAAO,oBAAoB,MAAM,QAC7F;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,sBAAsB;CACjC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,kBAAkB,MAAM,KAAK,KAAK,MAAM;CAClE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,mBAAmB;CAC9B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,iBAAiB,OAAO,eAAe,UAAU;CAC5D,MAAM,KAAK,kBAAkB,OAAO,eAAe,eAAe,OAAO;CACzE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,eAAe;CAC1B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,WAAW,MAAM,KAAK,KAAK,MAAM;CAC1E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,WAAW;CACtB,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,MAAM,SAC3C,OAAO,eAAe,QACtB,CAAC,0CAA0C,GAC7C,MAAM,KAAK,KAAK,MAAM;CAExB,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,kBAAkB;CAC7B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,eAAe,aAAa,MAAM,KAAK,KAAK,MAAM;CAC5E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,6BAA6B;CACxC,MAAM,KAAK,EAAE;CACb,MAAM,KACJ,2GACF;CACA,MAAM,KAAK,0DAA0D;CACrE,KAAK,MAAM,KAAK,OAAO,YAAY;EACjC,MAAM,QAAQ,EAAE,0BAA0B,OAAO,MAAM,OAAO,EAAE,qBAAqB;EACrF,MAAM,OAAO,EAAE,sBAAsB,OAAO,MAAM,EAAE,kBAAkB,QAAQ,CAAC;EAC/E,MAAM,IAAI,EAAE,WAAW,OAAO,EAAE,OAAO,QAAQ,CAAC,IAAI;EACpD,MAAM,IAAI,EAAE,YAAY,OAAO,EAAE,QAAQ,QAAQ,CAAC,IAAI;EACtD,MAAM,OAAO,EAAE,SAAS,IAAI,IAAI,EAAE,OAAO,GAAG,EAAE,IAAI,IAAI,EAAE,OAAO,IAAI,EAAE,KAAK;EAC1E,MAAM,MAAM,EAAE,QAAQ,QAAQ,CAAC,OAAO,SAAS,EAAE,GAAG,IAAI,MAAM,IAAI,EAAE,GAAG;EACvE,MAAM,KACJ,KAAK,EAAE,YAAY,KAAK,EAAE,SAAS,KAAK,IAAI,EAAE,IAAI,EAAE,KAAK,MAAM,KAAK,KAAK,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,EAAE,mBAAmB,QAAQ,KAAK,KAAK,EAAE,QAAQ,IAAI,GACrL;CACF;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,wBAAwB;CACnC,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,OAAO,QAAQ,QAAQ;CAClC,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,gBAAgB;CAC3B,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,iBAAiB;CAC5B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,aAAa,MAAM,KAAK,KAAK,MAAM;CACzE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,aAAa;CACxB,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,SAAS,MAAM,KAAK,KAAK,MAAM;CACrE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,6BAA6B;CACxC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,cAAc,MAAM,KAAK,KAAK,MAAM;CAC1E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,uBAAuB;CAClC,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,gBAAgB,MAAM,KAAK,KAAK,MAAM;CAC5E,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,eAAe;CAC1B,MAAM,KAAK,EAAE;CACb,KAAK,MAAM,QAAQ,OAAO,YAAY,WAAW,MAAM,KAAK,KAAK,MAAM;CACvE,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,gBAAgB;CAC3B,MAAM,KAAK,EAAE;CACb,MAAM,KACJ,yFACF;CACA,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,SAAS;CACpB,MAAM,KAAK,KAAK,UAAU;EAAE,QAAQ,OAAO;EAAQ,OAAO,OAAO;CAAM,GAAG,MAAM,CAAC,CAAC;CAClF,MAAM,KAAK,KAAK;CAChB,IAAI,OAAO,iBAAiB;EAC1B,MAAM,KAAK,EAAE;EACb,MAAM,KAAK,qBAAqB;EAChC,MAAM,KAAK,EAAE;EACb,MAAM,KAAK,uDAAuD;EAClE,MAAM,KAAK,yBAAyB;EACpC,KAAK,MAAM,KAAK,OAAO,gBAAgB,SAAS,MAAM,GAAG,EAAE,GACzD,MAAM,KACJ,KAAK,EAAE,aAAa,KAAK,EAAE,SAAS,KAAK,EAAE,YAAY,OAAO,KAAK,EAAE,YAAY,IAAI,KAAK,YAAY,EAAE,gBAAgB,EAAE,YAAY,EAAE,GAC1I;CAEJ;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,YAAY,GAAmB;CACtC,OAAO,EAAE,QAAQ,OAAO,KAAK;AAC/B;AAEA,SAAS,OAAO,GAAmB;CACjC,OAAO,GAAG,KAAK,IAAI,MAAM,KAAK,IAAI,CAAC;AACrC;AAEA,SAAgB,WAAW,MAA6C;CACtE,OAAO,IAAI,IAAI,KAAK,GAAG,EAAE,IAAI,IAAI,KAAK,IAAI,EAAE;AAC9C;AAEA,SAAgB,IAAI,GAAmB;CACrC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO,OAAO,CAAC;CACxC,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;ACpIA,MAAa,kCAAkC;AAE/C,SAAgB,iBAAiB,KASH;CAC5B,MAAM,cAAwB;EAC5B;EACA;EACA,sCAAsC,IAAI,IAAI,aAAa,IAAI,SAAS,eAAe,IAAI,WAAW;CACxG;CACA,IAAI,IAAI,MACN,YAAY,KACV,uCAAuC,WAAW,IAAI,IAAI,EAAE,2JAC9D;CAEF,IAAI,IAAI,eAAe,MACrB,YAAY,KAAK,oEAAoE;CA4BvF,OAAO;EAAE;EAAa,SAAA;GAzBpB;GACA;GACA,yDAAyD,IAAI,SAAS,sBAAsB,IAAI,SAAS;GACzG;GACA;EAqB0B;EAAG,cAAA;GAlB7B;GACA;GACA;GACA;EAewC;EAAG,gBAAA;GAZ3C;GACA;GACA;GACA;EASwD;EAAG,WAAA;GAN3D;GACA;GACA;GACA;GACA;EAEmE;CAAE;AACzE;AAEA,SAAgB,kBACd,KACA,KAUsD;CACtD,IAAI,IAAI,cAAc,IAAI,gBAAgB,IAAI,YAC5C,OAAO;EAAE,UAAU;EAAQ,QAAQ;CAAuB;CAE5D,IAAI,IAAI,gBAAgB,CAAC,IAAI,aAAa,SAExC,OAAO;EACL,UAAU;EACV,QAAQ,0BAHG,IAAI,aAAa,iBAAiB,WAGN,IAAI,IAAI,aAAa;CAC9D;CAEF,IAAI,CAAC,IAAI,YACP,OAAO;EACL,UAAU,IAAI,OAAO,aAAa,SAAS;EAC3C,QACE;CACJ;CAEF,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,IAAA,GAClC,OAAO;EACL,UAAU;EACV,QAAQ,QAAQ,IAAI,WAAW,KAAK,EAAE;CACxC;CAEF,MAAM,KAAK,IAAI,UAAU;CACzB,IAAI,IAAI,UAAU,UAChB,OAAO;EACL,UAAU;EACV,QACE;CACJ;CAUF,MAAM,aAAa,CAAC,OAAO,SAAS,GAAG,GAAG,KAAK,CAAC,OAAO,SAAS,GAAG,IAAI,KAAK,GAAG,QAAQ,GAAG;CAC1F,IAAI,CAAC,cAAc,IAAI,QAAQ,GAAG,OAAO,IAAI,KAAK,OAAO,GAAG,QAAQ,IAAI,KAAK,MAC3E,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,yBAAyB,WAAW,IAAI,IAAI,EAAE;CACzG;CAEF,MAAM,eAAe,CAAC,cAAc,GAAG,MAAM;CAE7C,IADqB,CAAC,cAAc,GAAG,OAAO,GAE5C,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE;CAC3D;CAEF,IAAI,IAAI,UAAU,IAAI,IAAI,UACxB,OAAO;EACL,UAAU;EACV,QAAQ,QAAQ,IAAI,UAAU,EAAE,+DAA+D,IAAI,IAAI,UAAU,GAAG,EAAE,uBAAuB,IAAI,SAAS;CAC5J;CAEF,IAAI,YACF,OAAO;EACL,UAAU;EACV,QAAQ,oBAAoB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE;CAC3D;CAEF,IAAI,IAAI,WAAW,QAAQ,IAAI,UAAU,IAAI,OAAO,cAClD,OAAO;EACL,UAAU;EACV,QAAQ,iBAAiB,IAAI,IAAI,MAAM,EAAE,KAAK,IAAI,IAAI,wBAAwB,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,2BAA2B,IAAI,IAAI,UAAU,iBAAiB,EAAE;CAC7K;CAEF,OAAO;EACL,UAAU;EACV,QAAQ,WAAW,IAAI,IAAI,UAAU,iBAAiB,EAAE,WAAW,IAAI,GAAG,GAAG,EAAE,IAAI,IAAI,GAAG,IAAI,EAAE,6CAA6C,IAAI,IAAI;CACvJ;AACF;AAEA,SAAgB,oBACd,YACA,KAO8B;CAC9B,MAAM,gBAAgB,WAAW,QAAQ,MAAM,EAAE,gBAAgB,IAAI,UAAU;CAC/E,MAAM,cAAc,cAAc,MAAM,MAAM,EAAE,aAAa,SAAS;CACtE,MAAM,iBAAiB,cAAc,MAAM,MAAM,EAAE,aAAa,YAAY;CAC5E,MAAM,SAAS,eAAe,kBAAkB,cAAc,MAAM;CACpE,MAAM,WAAmC,cACrC,YACA,cAAc,MAAM,MAAM,EAAE,aAAa,iBAAiB,IACxD,oBACA,iBACE,eACA,cAAc,MAAM,MAAM,EAAE,aAAa,MAAM,IAC7C,SACA;CAEV,MAAM,YAAsB,CAAC;CAC7B,MAAM,QAAkB,CAAC;CACzB,MAAM,cAAwB,CAAC;CAE/B,IAAI,QAAQ;EACV,UAAU,KAAK,GAAG,OAAO,YAAY,IAAI,OAAO,gBAAgB;EAChE,IAAI,OAAO,QAAQ;GACjB,MAAM,cACJ,OAAO,sBAAsB,OAAO,aAAa,IAAI,OAAO,iBAAiB,MAAM;GACrF,UAAU,KACR,2BAA2B,IAAI,OAAO,OAAO,GAAG,EAAE,IAAI,IAAI,OAAO,OAAO,IAAI,EAAE,GAAG,YAAY,EAC/F;EACF;EACA,IAAI,OAAO,QAAQ,QAAQ,OAAO,SAAS,OAAO,GAAG,GACnD,UAAU,KAAK,2BAA2B,OAAO,QAAQ,IAAI,IAAI,OAAO,GAAG,EAAE,cAAc;CAE/F;CACA,IAAI,CAAC,IAAI,YAAY;EACnB,MAAM,KAAK,mEAAmE;EAC9E,YAAY,KAAK,iEAAiE;CACpF;CACA,IAAI,CAAC,IAAI,qBAAqB;EAC5B,MAAM,KACJ,4GACF;EACA,YAAY,KACV,wGACF;CACF;CACA,IAAI,IAAI,SAAS,QAAQ,cAAc,SAAS,GAAG;EACjD,MAAM,KACJ,yFACF;EACA,YAAY,KACV,4GACF;CACF;CACA,MAAM,eAAe,cAAc,QAAQ,MAAM,EAAE,aAAa,iBAAiB;CACjF,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,QAAQ,aAAa,QAAQ,GAAG,MAAO,EAAE,UAAU,EAAE,UAAU,IAAI,CAAE;EAC3E,MAAM,KACJ,GAAG,aAAa,OAAO,kCAAkC,IAAI,SAAS,uBAAuB,MAAM,YAAY,QAAQ,MAAM,QAAQ,EACvI;EACA,YAAY,KACV,oBAAoB,IAAI,WAAW,MAAM,QAAQ,iCAAiC,MAAM,YAAY,EACtG;CACF;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,QAAQ;CACpE,IAAI,SAAS,SAAS,GACpB,MAAM,KACJ,GAAG,SAAS,OAAO,mFACrB;CAEF,IAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,SAAS,GAAG;EAClE,MAAM,MAAM,IAAI,gBAAgB,SAAS;EACzC,MAAM,KAAK,wBAAwB,IAAI,aAAa,UAAU,IAAI,SAAS,SAAS;EACpF,YAAY,KAAK,8DAA8D;CACjF;CACA,IAAI,aAAa,WACf,YAAY,KAAK,+DAA+D;MAC3E,IAAI,aAAa,QACtB,YAAY,KAAK,qEAAqE;MACjF,IAAI,aAAa,cACtB,YAAY,KACV,iLACF;MACK,IAAI,aAAa,UACtB,YAAY,KACV,+EACF;CAGF,OAAO;EACL;EACA,aAAa,QAAQ,eAAe;EACpC;EACA;EACA;CACF;AACF;AAEA,SAAgB,sBACd,YACA,gBACA,KAMU;CACV,MAAM,QAAkB,CAAC;CACzB,MAAM,gBAAgB,WAAW,QAAQ,MAAM,EAAE,gBAAgB,IAAI,UAAU;CAC/E,MAAM,KACJ,aAAa,cAAc,OAAO,uBAAuB,IAAI,MAAM,QAAQ,IAAI,aAAa,YAAY,IAAI,eAAe,GAAG,EAChI;CACA,MAAM,KACJ,mBAAmB,eAAe,WAAW,eAAe,cAAc,IAAI,eAAe,gBAAgB,GAAG,EAClH;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACvE,MAAM,OAAO,cAAc,QAAQ,MAAM,EAAE,aAAa,MAAM,CAAC,CAAC;CAChE,MAAM,aAAa,cAAc,QAAQ,MAAM,EAAE,aAAa,YAAY,CAAC,CAAC;CAC5E,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,OAAO,cAAc,QAAQ,MAAM,EAAE,aAAa,iBAAiB,CAAC,CAAC;CAC3E,MAAM,KACJ,iBAAiB,SAAS,YAAY,WAAW,eAAe,KAAK,SAAS,SAAS,WAAW,KAAK,iBACzG;CACA,MAAM,WAAW,cAAc,QAAQ,MAAM,EAAE,gBAAgB,CAAC,CAAC,KAAK,MAAM,EAAE,WAAW;CACzF,IAAI,SAAS,SAAS,GAAG,MAAM,KAAK,+BAA+B,SAAS,KAAK,IAAI,EAAE,EAAE;CACzF,IAAI,IAAI,iBACN,MAAM,KACJ,4BAA4B,IAAI,gBAAgB,cAAc,GAAG,IAAI,gBAAgB,UAAU,sBAAsB,IAAI,gBAAgB,SAAS,OAAO,wBAC3J;CAEF,MAAM,KACJ,IAAI,sBACA,2BAA2B,IAAI,oBAAoB,MAAM,GAAG,EAAE,EAAE,KAChE,0DACN;CACA,OAAO;AACT;AAEA,SAAgB,eAAe,UAA0C;CACvE,IAAI,aAAa,WAAW,OAAO;CACnC,IAAI,aAAa,cAAc,OAAO;CACtC,IAAI,aAAa,QAAQ,OAAO;CAChC,IAAI,aAAa,mBAAmB,OAAO;CAC3C,OAAO;AACT;;;;;;;;;;;;;;;;;;;ACzSA,SAAgB,QAAQ,GAAa,GAA4B;CAC/D,IAAI,EAAE,SAAS,KAAK,EAAE,SAAS,GAAG,OAAO;CACzC,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;CAC/C,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE;CAC/C,MAAM,OAAO,EAAE,QAAQ,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;CAC3E,MAAM,OAAO,EAAE,QAAQ,KAAK,MAAM,OAAO,IAAI,UAAU,GAAG,CAAC,KAAK,EAAE,SAAS;CAC3E,MAAM,SAAS,KAAK,OAChB,EAAE,SAAS,KAAK,QAAQ,EAAE,SAAS,KAAK,SAAS,EAAE,SAAS,EAAE,SAAS,EAC3E;CACA,IAAI,WAAW,GAAG,OAAO,UAAU,QAAQ,IAAI;CAC/C,QAAQ,QAAQ,SAAS;AAC3B;;;;;;;;;AAUA,SAAgB,eAAe,QAAkB,OAAgC;CAC/E,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,yCAAyC,OAAO,OAAO,MAAM,MAAM,OAAO,EAC5E;CAEF,IAAI,OAAO,SAAS,GAAG,OAAO;CAC9B,MAAM,SAAS,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,KAAK;CACjE,IAAI,OAAO,MAAM,UAAU,CAAC,OAAO,SAAS,KAAK,CAAC,GAChD,MAAM,IAAI,gBAAgB,kDAAkD;CAE9E,MAAM,YAAY,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;CACzE,MAAM,WACJ,OAAO,QAAQ,KAAK,UAAU,OAAO,QAAQ,cAAc,GAAG,CAAC,KAAK,OAAO,SAAS;CACtF,MAAM,oBAAoB,KAAK,KAAK,QAAQ;CAC5C,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,SAAS,GAAG,GAAG,OAAO,IAAI,KAAK,GAAG,CAAC;CACtE,IAAI,qBAAqB,OAAO,UAAU,OAAO,OAAO,cAAc,IAAI,IAAI;CAC9E,OAAO,YAAY;AACrB;;;;;;;;;;;AAcA,SAAgB,YAAY,QAAkB,OAAyB;CACrE,MAAM,IAAI,OAAO,SAAS,MAAM;CAChC,IAAI,MAAM,GAAG,OAAO;CACpB,IAAI,YAAY;CAChB,KAAK,MAAM,KAAK,OACd,KAAK,MAAM,KAAK,QACd,IAAI,IAAI,GAAG,aAAa;MACnB,IAAI,IAAI,GAAG,aAAa;CAGjC,OAAO,YAAY;AACrB;;;;;;AAOA,SAAgB,gBAAgB,OAAgC;CAC9D,MAAM,IAAI,KAAK,IAAI,KAAK;CACxB,IAAI,IAAI,MAAO,OAAO;CACtB,IAAI,IAAI,KAAM,OAAO;CACrB,IAAI,IAAI,MAAO,OAAO;CACtB,OAAO;AACT;;;;ACxFA,SAAgB,gBACd,WACA,UACA,OAMA;CAGA,MAAM,WAAW,QAAuC;EACtD,MAAM,IAAI,mBAAmB,KAAK,KAAK;EACvC,OAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI,KAAA;CAC3D;CACA,MAAM,kBAAkB,UAAU,QAAQ,QAAQ,QAAQ,GAAG,MAAM,KAAA,CAAS;CAC5E,MAAM,iBAAiB,SAAS,QAAQ,QAAQ,QAAQ,GAAG,MAAM,KAAA,CAAS;CAC1E,MAAM,UAAU,eAAe,gBAAgB,eAAe;CAC9D,OAAO;EACL,QAAQ,QAAQ,MAAM,KAAK,SAAS,QAAQ,KAAK,QAAQ,CAAW;EACpE,OAAO,QAAQ,MAAM,KAAK,SAAS,QAAQ,KAAK,SAAS,CAAW;EACpE,oBAAoB,QAAQ,kBAAkB;EAC9C,qBAAqB,QAAQ,iBAAiB;CAChD;AACF;;;;;;;;;;;;;AAcA,SAAgB,gBACd,MACA,aACA,YACA,MAiBO;CAGP,MAAM,EAAE,QAAQ,UAAU,gBAFR,KAAK,QAAQ,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAAK,KAExC,GADjC,KAAK,QAAQ,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAAK,KAC5B,GAAG,KAAK,KAAK;CACzE,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO;CAEpB,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,YAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAEtD,MAAM,cAAc,eADC,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CACN,CAAC;CAC/C,MAAM,UAAU,MAAM,QAAQ,SAAS;CAEvC,MAAM,KAAK,gBAAgB,QAAQ,OAAO;EACxC,YAAY,KAAK;EACjB,WAAW;EACX,WAAW;EACX,MAAM,KAAK;CACb,CAAC;CAED,MAAM,cAAc,6BAA6B,QAAQ,KAAM,KAAK,IAAI;CACxE,MAAM,oBACJ,YAAY,WAAW,IAAI,IAAI,YAAY,QAAQ,MAAM,IAAI,CAAC,CAAC,CAAC,SAAS,YAAY;CACvF,MAAM,WACJ,KAAK,SAAS,QAAQ,YAAY,WAAW,IACzC,OACA,YAAY,QAAQ,MAAM,KAAK,KAAK,KAAM,OAAO,KAAK,KAAK,KAAM,IAAI,CAAC,CAAC,SACvE,YAAY;CAElB,MAAM,gBAAgB,UAAU;EAAE,SAAS;EAAG,OAAO,KAAK;EAAU,OAAO,KAAK;CAAS,CAAC;CAC1F,MAAM,MAAM,YAAY,IAAI,IAAI,gBAAgB;CAEhD,OAAO;EACL;EACA;EACA;EACA;EACA,IAAI;GAAE,KAAK,GAAG;GAAK,MAAM,GAAG;EAAK;EACjC;EACA;EACA;CACF;AACF;AAEA,SAAS,6BACP,QACA,WACA,MACU;CACV,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO,CAAC;CACrB,IAAI,MAAM,GAAG,OAAO,IAAI,MAAc,SAAS,CAAC,CAAC,KAAK,OAAO,EAAG;CAChE,MAAM,MAAM,QAAQ,MAAM,MAAM;CAChC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,IAAI,cAAc;EAClB,IAAI,YAAY;EAChB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,OAAO,WAAW,IAAI,CAAC,CAAC;GAC1D,eAAe,OAAO,KAAM;GAC5B,aAAa;EACf;EACA,QAAQ,KAAK,cAAc;CAC7B;CACA,OAAO;AACT;AAEA,SAAS,MAAM,IAAc,MAAsB;CACjD,IAAI,GAAG,SAAS,GAAG,OAAO;CAC1B,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,IAAI,QAAQ,IAAI,SAAS;CACzC,OAAO,KAAK,KAAK,OAAO,GAAG,SAAS,EAAE;AACxC;AAEA,SAAgB,eAAe,QAA0B;CACvD,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,MAAM,KAAK,MAAM,OAAO,SAAS,CAAC;CACxC,OAAO,OAAO,SAAS,MAAM,KAAK,OAAO,MAAM,KAAM,OAAO,QAAS,IAAI,OAAO;AAClF;;;;;;;;ACrFA,SAAgB,aAAa,MAAmB,OAA4B,CAAC,GAAiB;CAC5F,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,aAAa,KAAK,cAAc;CAEtC,MAAM,8BAAc,IAAI,IAAqD;CAC7E,KAAK,MAAM,KAAK,MAAM;EACpB,IAAI,EAAE,aAAa,OAAO;EAK1B,MAAM,IAAI,mBAAmB,GAAG,KAAK;EACrC,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,MAAM,SAAS,YAAY,IAAI,EAAE,WAAW,KAAK;GAAE,MAAM,CAAC;GAAG,QAAQ,CAAC;EAAE;EACxE,OAAO,KAAK,KAAK,CAAC;EAClB,OAAO,OAAO,KAAK,CAAC;EACpB,YAAY,IAAI,EAAE,aAAa,MAAM;CACvC;CAEA,MAAM,eAAe,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;CAClD,MAAM,WAAW,aAAa,YAAY,IAAI,UAAU,IAAI,KAAA;CAG5D,MAAM,YAA8D,CAAC;CACrE,KAAK,MAAM,MAAM,cAAc;EAC7B,MAAM,SAAS,YAAY,IAAI,EAAE;EACjC,MAAM,KAAK,mBAAmB,OAAO,QAAQ,UAAU;EACvD,IAAI,OAAsB;EAC1B,IAAI,IAAmB;EACvB,IAAI,UAAyB;EAC7B,IAAI,qBAAoC;EACxC,IAAI,sBAAqC;EACzC,IAAI,cAAc,YAAY,OAAO,YAAY;GAC/C,MAAM,SAAS,gBAAgB,OAAO,MAAM,SAAS,MAAM,KAAK;GAChE,UAAU,OAAO,OAAO;GACxB,qBAAqB,OAAO;GAC5B,sBAAsB,OAAO;GAC7B,IAAI,OAAO,OAAO,UAAU,GAC1B,OAAO,mBAAmB,OAAO,QAAQ,OAAO,KAAK,CAAC,CAAC;GAEzD,IAAI,eAAe,OAAO,QAAQ,OAAO,KAAK;EAChD;EACA,UAAU,KAAK;GACb,aAAa;GACb,GAAG,OAAO,OAAO;GACjB,MAAM,GAAG;GACT,OAAO,GAAG;GACV,QAAQ,GAAG;GACX,QAAQ;GACR,SAAS;GACT;GACA;GACA;GACA;EACF,CAAC;CACH;CAKA,IAAI,YAAY;EACd,MAAM,OAAiB,CAAC;EACxB,MAAM,KAAe,CAAC;EACtB,KAAK,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;GACzC,MAAM,IAAI,UAAU;GACpB,IAAI,EAAE,gBAAgB,YAAY;GAClC,IAAI,EAAE,SAAS,QAAQ,CAAC,OAAO,SAAS,EAAE,IAAI,GAAG;GACjD,KAAK,KAAK,CAAC;GACX,GAAG,KAAK,EAAE,IAAI;EAChB;EACA,IAAI,GAAG,SAAS,GAAG;GACjB,MAAM,EAAE,YAAY,kBAAkB,IAAI,GAAG;GAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAC/B,UAAU,KAAK,GAAI,CAAE,SAAS,QAAQ;EAE1C;CACF;CAEA,MAAM,OAAO,UAAU,KAAK,EAAE,MAAM,OAAO,GAAG,WAAW,IAAI;CAE7D,OAAO;EAAE;EAAM;EAAY;EAAO,UADjB,2BAA2B,MAAM,YAAY,KACrB;CAAE;AAC7C;AAEA,SAAS,2BACP,MACA,YACA,OACQ;CACR,MAAM,QAAkB,CAAC;CACzB,MAAM,WAAW,aAAa,QAAQ,WAAW,KAAK;CACtD,MAAM,KAAK,mBAAmB,MAAM,QAAQ,UAAU;CACtD,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,4EAA4E;CACvF,MAAM,KAAK,yCAAyC;CACpD,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,KAAK,IAAI,IAAI,EAAE,KAAK,EAAE,IAAI,IAAI,EAAE,MAAM,EAAE;EAC9C,MAAM,IAAI,EAAE,WAAW,OAAO,EAAE,OAAO,QAAQ,CAAC,IAAI;EACpD,MAAM,IAAI,EAAE,YAAY,OAAO,EAAE,QAAQ,QAAQ,CAAC,IAAI;EACtD,MAAM,SAAS,EAAE,WAAW;EAC5B,MAAM,WACJ,EAAE,uBAAuB,QAAQ,EAAE,wBAAwB,OACvD,MACA,GAAG,EAAE,mBAAmB,GAAG,EAAE;EACnC,MAAM,KACJ,KAAK,EAAE,YAAY,KAAK,EAAE,EAAE,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI,EAAE,IAAI,EAAE,KAAK,GAAG,KAAK,EAAE,KAAK,EAAE,GAC/F;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;;;;;;AAgCA,SAAgB,YACd,MACA,OAGI,CAAC,GACa;CAClB,MAAM,QAAQ,KAAK,SAAS;CAE5B,MAAM,0BAAU,IAAI,IAAmD;CACvE,KAAK,MAAM,KAAK,MAAM;EACpB,IAAI,EAAE,aAAa,OAAO;EAC1B,MAAM,IAAI,mBAAmB,GAAG,KAAK;EACrC,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,IAAI,EAAE,YAAY,MAAM;EACxB,MAAM,SAAS,QAAQ,IAAI,EAAE,WAAW,KAAK;GAAE,MAAM,CAAC;GAAG,SAAS,CAAC;EAAE;EACrE,OAAO,KAAK,KAAK,EAAE,OAAO;EAC1B,OAAO,QAAQ,KAAK,CAAC;EACrB,QAAQ,IAAI,EAAE,aAAa,MAAM;CACnC;CAEA,MAAM,SAAwB,CAAC;CAC/B,KAAK,MAAM,CAAC,aAAa,WAAW,QAAQ,QAAQ,GAClD,OAAO,KAAK;EACV;EACA,MAAMA,KAAI,OAAO,IAAI;EACrB,SAASA,KAAI,OAAO,OAAO;EAC3B,GAAG,OAAO,KAAK;EACf,YAAY;EACZ,MAAM,KAAK,gBAAgB,eACvB,UAAU,KAAK,cAAc,YAAa,IAC1C,KAAA;CACN,CAAC;CAMH,KAAK,MAAM,KAAK,QACd,EAAE,aAAa,CAAC,OAAO,MAAM,MAAM,MAAM,KAAK,UAAU,GAAG,CAAC,CAAC;CAG/D,OAAO;EACL,MAAM;EACN;EACA,MAAM;GAAE,GAAG;GAAW,GAAG;EAAQ;EACjC;CACF;AACF;AAEA,SAAS,UAAU,GAAgB,GAAyB;CAC1D,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,WAAW,EAAE,YAAY,EAAE,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE;AACzF;AAEA,SAAgB,UAAU,GAA6C;CACrE,OAAO,EAAE,UAAU,YAAY;AACjC;;;;;;AA+CA,SAAgB,cACd,MACA,aACA,YACA,OAAgC,CAAC,GACL;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,QAAQ;CAC9B,IAAI,WAAW,GAAG,MAAM,IAAI,MAAM,iCAAiC;CAOnE,MAAM,EAAE,QAAQ,OAAO,oBAAoB,wBAAwB,gBALjD,KAAK,QAAQ,MAAM,EAAE,gBAAgB,eAAe,EAAE,aAAa,KAMnF,GALe,KAAK,QAAQ,MAAM,EAAE,gBAAgB,cAAc,EAAE,aAAa,KAMjF,GACA,KACF;CACA,MAAM,IAAI,OAAO;CAEjB,IAAI,MAAM,GACR,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA,GAAG;EACH;EACA;EACA,MAAM,CAAC;EACP,QAAQ;EACR,IAAI;CACN;CAGF,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,eAAe,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACrD,MAAM,SAAS,eAAe,YAAY;CAC1C,MAAM,MAAM,aAAa;CACzB,MAAM,MAAM,aAAa,aAAa,SAAS;CAI/C,MAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,GAAG,GAAG,KAAK,IAAI,GAAG,GAAG,IAAI;CACzD,MAAM,KAAK,CAAC;CAEZ,MAAM,SAASC,QAAK,MAAM;CAC1B,MAAM,OAA8B,CAAC;CACrC,KAAK,IAAI,IAAI,GAAG,IAAI,UAAU,KAC5B,KAAK,KAAK;EAAE,IAAI,KAAK,IAAI;EAAO,IAAI,MAAM,IAAI,KAAK;EAAO,OAAO;CAAE,CAAC;CAEtE,KAAK,MAAM,KAAK,QAAQ;EACtB,IAAI,MAAM,KAAK,OAAO,IAAI,MAAM,KAAK;EACrC,IAAI,MAAM,GAAG,MAAM;EACnB,IAAI,OAAO,UAAU,MAAM,WAAW;EACtC,KAAK,IAAI,CAAE,SAAS;CACtB;CAEA,MAAM,KAAK,gBAAgB,QAAQ,OAAO;EACxC,YAAY,KAAK,cAAc;EAC/B,WAAW,KAAK,aAAa;EAC7B,WAAW;EACX,MAAM,KAAK;CACb,CAAC;CAED,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,IAAI;GAAE,KAAK,GAAG;GAAK,MAAM,GAAG;EAAK;CACnC;AACF;;;;;;;;;;;;;;;;;;;;;;AC7LA,eAAsB,eACpB,MACA,OAA8B,CAAC,GACN;CACzB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,WAAW,KAAK,IAAI,KAAK,YAAY,IAAA,CAAmC;CAC9E,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,cAAc,KAAK,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;CAC/D,MAAM,sBAAsB,KAAK,uBAAuB;CAExD,IAAI,QAAQ,EAAE,OAAO,SAAS,KAAK,GAAG,KAAK,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,OAAO,KAAK,OACxF,MAAM,IAAI,MACR,wEAAwE,KAAK,UAAU,IAAI,GAC7F;CAGF,MAAM,UAAU,aAAa,MAAM;EACjC,YAAY,cAAc,KAAA;EAC1B;EACA;EACA;CACF,CAAC;CACD,MAAM,SAAS,YAAY,MAAM;EAAE;EAAO,eAAe,KAAK;CAAc,CAAC;CAC7E,MAAM,eACJ,KAAK,gBAAgB,QAAQ,KAAK,KAAK,MAAM,EAAE,WAAW,CAAC,CAAC,QAAQ,OAAO,OAAO,UAAU;CAC9F,MAAM,QAAQ,aACV,aAAa,KAAK,OAChB,cAAc,MAAM,IAAI,YAAY;EAClC;EACA;EACA,MAAM,KAAK;CACb,CAAC,CACH,IACA,CAAC;CAEL,MAAM,kBAAkB,IAAI,IAAI,MAAM,KAAK,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;CACpE,MAAM,oBAAoB,IAAI,IAAI,OAAO,OAAO,KAAK,MAAM,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;CAC9E,MAAM,uCAAuB,IAAI,IAAgD;CACjF,IAAI,YACF,KAAK,MAAM,MAAM,cACf,qBAAqB,IACnB,IACA,gBAAgB,MAAM,IAAI,YAAY;EACpC;EACA;EACA,MAAM,KAAK;EACX;EACA;EACA;CACF,CAAC,CACH;CAIJ,MAAM,aAAa,QAAQ,KACxB,KAAK,QAAQ;EACZ,MAAM,OAAO,gBAAgB,IAAI,IAAI,WAAW;EAChD,MAAM,QAAQ,kBAAkB,IAAI,IAAI,WAAW;EACnD,MAAM,YAAY,qBAAqB,IAAI,IAAI,WAAW,KAAK;EAC/D,MAAM,eAAe,KAAK,gBAAgB,IAAI;EAC9C,MAAM,aAAa,kBAAkB,KAAK;GACxC;GACA;GACA;GACA;GACA;GACA;GACA;GACA;EACF,CAAC;EACD,OAAO;GACL,aAAa,IAAI;GACjB,GAAG,IAAI;GACP,MAAM,IAAI;GACV,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb,uBAAuB,YAAY,UAAU,YAAY;GACzD,SAAS,WAAW,KAAK,MAAM,KAAK;GACpC,YAAY,YAAY,UAAU,cAAc,OAAO,KAAK,SAAS;GACrE,UAAU,YAAY,UAAU,YAAY;GAC5C,QAAQ,YAAY,UAAU,KAAK,OAAO,KAAK,KAAK;GACpD,mBAAmB,YAAY,UAAU,oBAAoB;GAC7D,UAAU,YAAY,UAAU,WAAW;GAC3C,KAAK,YAAY,UAAU,MAAM;GACjC,kBAAkB,OAAO,cAAc;GACvC,MAAM,eAAe,UAAU,YAAY,IAAI,KAAA;GAC/C,UAAU,WAAW;GACrB,gBAAgB,WAAW;EAC7B;CACF,CAAC,CAAC,CACD,MAAM,GAAG,MAAM;EACd,MAAM,eAAe,eAAe,EAAE,QAAQ,IAAI,eAAe,EAAE,QAAQ;EAC3E,IAAI,iBAAiB,GAAG,OAAO;EAC/B,OAAO,EAAE,OAAO,EAAE;CACpB,CAAC;CAEH,MAAM,iBAAiB,oBAAoB,YAAY;EACrD;EACA,iBAAiB,KAAK;EACtB;EACA;EACA;CACF,CAAC;CACD,MAAM,mBAAmB,sBAAsB,YAAY,gBAAgB;EACzE;EACA;EACA,iBAAiB,KAAK;EACtB;CACF,CAAC;CACD,MAAM,cAAc,iBAAiB;EACnC;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF,CAAC;CAED,MAAM,iBAAiB,MAAM,SAAS;EACpC,SAAS,KACN,QAAQ,MAAM,EAAE,aAAa,KAAK,CAAC,CACnC,KAAK,OAAO;GAAE,OAAO,EAAE;GAAO,aAAa,EAAE;GAAa,UAAU,EAAE;EAAS,EAAE,CAAC,CAClF,MAAM,GAAG,MAAM,EAAE,MAAM,cAAc,EAAE,KAAK,CAAC;EAChD;EACA;CACF,CAAC;CAED,MAAM,WAAW,uBAAuB;EACtC;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,iBAAiB,KAAK;CACxB,CAAC;CACD,MAAM,OAAO,mBAAmB,UAAU,KAAK;CAE/C,OAAO;EACL,MAAM;EACN;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;GAAE;GAAQ;EAAM;EACxB;EACA,iBAAiB,KAAK;EACtB;EACA;CACF;AACF"}

package/dist/{summary-report-bdFDXq0b.d.ts → summary-report-CkbRcSuu.d.ts}

@@ -1,7 +1,7 @@
-import { a as RunRecord } from "./run-record-C1E24G2A.js";
-import { r as PairedDecisionStatistic } from "./paired-promotion-decision-DPKz8ZnF.js";
-import "./reward-CpCmAM8L.js";
-import { n as FailureClusterReport } from "./failure-cluster-CbgEDcFJ.js";
+import { a as RunRecord } from "./run-record-CaxbLq6Y.js";
+import { r as PairedDecisionStatistic } from "./paired-promotion-decision-CQMJ5xGN.js";
+import "./reward-DrZ71vnf.js";
+import { n as FailureClusterReport } from "./failure-cluster-DN0t63Z0.js";
 //#region src/held-out-gate.d.ts
 type HeldOutGateRejectionCode = 'few_runs' | 'incomplete_coverage' | 'missing_split_scores' | 'missing_cost' | 'negative_delta' | 'indeterminate_delta' | 'overfit_gap' | 'cost_ceiling';
 /**
@@ -192,7 +192,7 @@ interface GateEvidence {
    */
   realnessGatedRuns: number;
 }
-interface GateDecision {
+interface HeldOutGateDecision {
   /** Final promote/no-promote verdict. */
   promote: boolean;
   /** The candidate that was evaluated. */
@@ -226,7 +226,22 @@ declare class HeldOutGate {
   /** Decide whether `candidate` should replace `baseline`.
    *  Pairing is by `(experimentId, scenarioId, seed)`.
    *  Missing or duplicate identities throw instead of comparing by position. */
-  evaluate(candidate: RunRecord[], baseline: RunRecord[]): GateDecision;
+  evaluate(candidate: RunRecord[], baseline: RunRecord[]): HeldOutGateDecision;
+  /** Refusals decided before any interval is computed: absent split scores,
+   *  incomplete coverage, too few paired runs. */
+  private refuseOnMissingEvidence;
+  /** The paired holdout delta, decided by the one shared rule (see
+   *  `paired-promotion-decision.ts`): the estimator comes from the outcome's
+   *  shape, McNemar vetoes at a non-negative threshold, and a zero-width
+   *  interval is refused. `medianPairedDelta` stays the literal median on
+   *  EVERY path so the number that made the old gate blind remains visible. */
+  private decidePairedDelta;
+  /** Refusals read off the paired delta: a zero-width interval, too few pairs
+   *  for the chosen estimator, or an interval that does not clear the threshold. */
+  private refuseOnDelta;
+  /** The overfit-gap and cost-ceiling refusals, run after the quality gates so
+   *  a cost-driven rejection never masks a quality failure. */
+  private refuseOnOverfitOrCost;
   /**
    * Which paired statistic decides this data, and the shape facts behind it —
    * for the early exits, where no interval is computed. Delegates to
@@ -289,7 +304,7 @@ interface ParetoPoint {
   /** Whether this candidate is on the Pareto frontier — high
    *  quality, low cost, no dominator. */
   onFrontier: boolean;
-  /** Optional gate verdict for this candidate, if a `GateDecision`
+  /** Optional gate verdict for this candidate, if a `HeldOutGateDecision`
    *  for it was passed in. */
   gate?: 'promote' | 'reject';
 }
@@ -302,15 +317,6 @@ interface ParetoFigureSpec {
     y: 'score';
   };
 }
-/**
- * Cost vs quality scatter spec. `gateDecisions` is keyed by
- * candidate id; if present, every point picks up the gate verdict
- * for overlay.
- */
-declare function paretoChart(runs: RunRecord[], opts?: {
-  split?: 'search' | 'holdout';
-  gateDecisions?: Record<string, GateDecision>;
-}): ParetoFigureSpec;
 interface GainDistributionBin {
   /** Inclusive lower edge. */
   lo: number;
@@ -407,7 +413,7 @@ interface ResearchReportOptions {
    */
   mdeAlpha?: number;
   /** Optional held-out gate decisions keyed by candidate id. */
-  gateDecisions?: Record<string, GateDecision>;
+  gateDecisions?: Record<string, HeldOutGateDecision>;
   /** Optional failure clusters from failureClusterView. */
   failureClusters?: FailureClusterReport;
   /** Build gain histograms for these candidates. Defaults to all non-comparator candidates. */
@@ -531,5 +537,5 @@ interface ResearchReport {
  */
 declare function researchReport(runs: RunRecord[], opts?: ResearchReportOptions): Promise<ResearchReport>;
 //#endregion
-export { HeldOutGateConfig as C, HeldOutGate as S, SplitCoverage as T, gainHistogram as _, ResearchReportOptions as a, GateDecision as b, RESEARCH_REPORT_HARD_PAIR_FLOOR as c, GainDistributionOptions as d, ParetoFigureSpec as f, SummaryTableRow as g, SummaryTableOptions as h, ResearchReportMethodology as i, GainDistributionBin as l, SummaryTable as m, ResearchReportCandidate as n, ResearchReportRecommendation as o, ParetoPoint as p, ResearchReportDecision as r, researchReport as s, ResearchReport as t, GainDistributionFigureSpec as u, paretoChart as v, HeldOutGateRejectionCode as w, GateEvidence as x, summaryTable as y };
-//# sourceMappingURL=summary-report-bdFDXq0b.d.ts.map
+export { HeldOutGateRejectionCode as C, HeldOutGateDecision as S, gainHistogram as _, ResearchReportOptions as a, HeldOutGate as b, RESEARCH_REPORT_HARD_PAIR_FLOOR as c, GainDistributionOptions as d, ParetoFigureSpec as f, SummaryTableRow as g, SummaryTableOptions as h, ResearchReportMethodology as i, GainDistributionBin as l, SummaryTable as m, ResearchReportCandidate as n, ResearchReportRecommendation as o, ParetoPoint as p, ResearchReportDecision as r, researchReport as s, ResearchReport as t, GainDistributionFigureSpec as u, summaryTable as v, SplitCoverage as w, HeldOutGateConfig as x, GateEvidence as y };
+//# sourceMappingURL=summary-report-CkbRcSuu.d.ts.map

package/dist/summary-report-CkbRcSuu.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"summary-report-CkbRcSuu.d.ts","names":[],"sources":["../src/held-out-gate.ts","../src/summary-report-figures.ts","../src/research-report-analysis.ts","../src/research-report.ts"],"mappings":";;;;;KAuIY;;;;;;;;UAiBK;;EAEf;;;EAGA;;;;EAIA;;EAEA;;EAEA;;;EAGA;;UAGe;;;;EAIf;;;;;EAKA;;;;;;;;;;;EAWA;;;;EAIA;;;EAGA;;EAEA;;EAEA;;;;EAIA;;;;;;;;;;;;;;;;;EAiBA;;;;;;;;;;;;;;EAcA;;;;KAKG,iBAAiB;UAEL;;EAEf;;EAEA;;EAEA;;;;;;;;EAQA,iBAAiB;;;;;EAKjB,gBAAgB;;;;;EAKhB;;;;;;;;;EASA,gBAAgB;;;EAGhB;;;EAGA;IAAY;IAAa;;;;EAGzB;;;;;;;EAOA;IAAW;IAAW;IAAW;IAAqB;;;;;;EAKtD;;;;;EAKA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA;;;EAGA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA,UAAU;;EAEV;;EAEA,eAAe;;;;;;;cAQJ;mBACM;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;EAEjB,YAAY,QAAQ;;;;EAoCpB,SAAS,WAAW,aAAa,UAAU,cAAc;;;UAiCjD;;;;;;UAoEA;;;UAiCA;;;UAkEA;;;;;;;UAqDA;;;;UCxnBO;;;EAGf;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,MAAM;EACN;EACA;;EAEA;;;;;;;iBAQc,aAAa,MAAM,aAAa,OAAM,sBAA2B;UAkHhE;EACf;;EAEA;;EAEA;;EAEA;;;EAGA;;;EAGA;;UAGe;EACf;EACA;EACA,QAAQ;EACR;IAAQ;IAAc;;;UAoEP;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,MAAM;EACN;EACA;IAAM;IAAa;;;UAGJ;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;iBAQc,cACd,MAAM,aACN,qBACA,oBACA,OAAM,0BACL;;;;;;;;cCpSU;;;KCWD;UAOK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;EASA;;;;;;;EAOA;IAAS;IAAa;;;;;;EAKtB;;;;;EAKA;;EAEA,gBAAgB,eAAe;;EAE/B,kBAAkB;;EAElB;;EAEA;;EAEA;;;;;;EAMA;;UAGe;EACf,UAAU;EACV;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IAAU;IAAa;;;;;;;EAMvB;EACA;;;;;;;EAOA;EACA;EACA,OAAO;EACP,UAAU;EACV;;UAGe;;;;;EAKf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;;;;;EAMA;EACA;EACA;IAAQ;IAAa;;EACrB;EACA,gBAAgB;EAChB,YAAY;EACZ,SAAS;EACT;IACE,QAAQ;IACR,OAAO;;EAET,aAAa;EACb,kBAAkB;EAClB;EACA;;;;;;;;;;;;;;;;;;;;;iBAsBoB,eACpB,MAAM,aACN,OAAM,wBACL,QAAQ"}

package/dist/supervisor-run/index.d.ts

@@ -1,5 +1,5 @@
 import { b as RolloutSplit } from "../schema-C-ziyJD7.js";
-import { A as WallDistribution, C as SupervisorRunSources, D as SupervisorStatusSource, E as SupervisorRunTreeGapCode, F as showMeasured, I as unavailable, M as WorkerNativeSession, N as WorkerTurnCoverage, O as TerminalFailure, P as isUnavailable, S as SupervisorRunRollup, T as SupervisorRunTreeGap, _ as SpendMeasurements, a as NamedResourceReceipt, b as SupervisorRunReader, c as PatchStats, d as RoleSpend, f as RollupCellRow, g as SpendMeasurement, h as SourceLimits, i as NO_SOURCE_LIMITS, j as WorkerLogSource, k as Unavailable, l as PerWorkerRow, m as SUPERVISOR_RUN_SCHEMA, n as EconomicsMetrics, o as OrchestrationMetrics, p as SUPERVISOR_RUN_ROLLUP_SCHEMA, r as Measured, s as OutcomeMetrics, t as DecisionMetrics, u as ResourceSpendRecord, v as SteerBreakdown, w as SupervisorRunTree, x as SupervisorRunReport, y as SupervisorRunNodeRole } from "../types-BqkVmHfv.js";
+import { A as WorkerTurnCoverage, C as SupervisorRunTreeGapCode, D as WallDistribution, E as Unavailable, M as showMeasured, N as unavailable, O as WorkerLogSource, S as SupervisorRunTreeGap, T as TerminalFailure, _ as SupervisorRunReader, a as OrchestrationMetrics, b as SupervisorRunSources, c as PerWorkerRow, d as RollupCellRow, f as SourceLimits, g as SupervisorRunNodeRole, h as SteerBreakdown, i as NamedResourceReceipt, j as isUnavailable, k as WorkerNativeSession, l as ResourceSpendRecord, m as SpendMeasurements, n as EconomicsMetrics, o as OutcomeMetrics, p as SpendMeasurement, r as Measured, s as PatchStats, t as DecisionMetrics, u as RoleSpend, v as SupervisorRunReport, w as SupervisorStatusSource, x as SupervisorRunTree, y as SupervisorRunRollup } from "../types-CmXY0PQf.js";
 //#region src/supervisor-run/analyze.d.ts
 /**
  * Analyze already-read supervisor-run bytes. Pure and synchronous: same bytes
@@ -8,6 +8,64 @@ import { A as WallDistribution, C as SupervisorRunSources, D as SupervisorStatus
 export declare function analyzeSupervisorRunSources(src: SupervisorRunSources, now?: () => number): SupervisorRunReport;
 //#endregion
 //#region src/supervisor-run/claude-code-reader.d.ts
+/**
+ * Supervision-tree reader over a THIRD-PARTY harness: Claude Code.
+ *
+ * The Runtime reader reads a supervisor journal designed for this analysis.
+ * for this analysis. This reader reads a harness we do not control, whose
+ * transcript was designed for replaying a chat — and recovers the same tree
+ * from it. If both produce a `SupervisorRunSources`, the tree model is a
+ * property of multi-agent runs, not of our journal format.
+ *
+ * ## Where the tree hides in a Claude Code transcript
+ *
+ * | Tree fact | Claude Code evidence |
+ * |---|---|
+ * | spawn      | assistant `tool_use` (`Agent` / `Task`), answered by a `tool_result` whose `toolUseResult.agentId` names the child |
+ * | settle     | a `<task-notification>` block in a later user line: `<task-id>` = agentId, `<status>` |
+ * | steer      | assistant `tool_use` (`SendMessage`) with `input.to` = agentId — mid-task, to a LIVE child |
+ * | delivered  | that steer's `tool_result` carrying `success` / `resumedAgentId` |
+ * | cancel     | assistant `tool_use` (`TaskStop`) targeting an agentId |
+ * | brain spend| `message.usage` on the main thread's assistant lines |
+ * | worker spend| `message.usage` inside `<session>/subagents/agent-<id>.jsonl` |
+ * | depth      | a child transcript that itself contains `Agent` tool_use lines |
+ *
+ * Every one of those is read from the normalized session of @tangle-network/harness-sessions — the SAME
+ * parser `src/rollout/readers/claude-jsonl.ts` uses for solo rollouts. There
+ * is no second transcript parser.
+ *
+ * ## What Claude Code cannot say
+ *
+ * It records tokens but never a price, runs no per-worker verify, and keeps no
+ * per-worker patch. Those are declared once in `limits`, so the analyzer
+ * reports `unavailable — <reason>` instead of the $0 / 0-accepted that summing
+ * an empty field would produce. See `SourceLimits`.
+ *
+ * ## Metric coverage vs the Runtime journal
+ *
+ * Measured on a real 52-agent session (fixture:
+ * `tests/fixtures/supervisor-run/claude-code-session-*`).
+ *
+ * | Metric | Runtime | Claude Code | Why |
+ * |---|---|---|---|
+ * | workersSpawned / Settled / Cancelled | full | full | spawn tool_use + task-notification + TaskStop |
+ * | steers / steersDelivered / steersByWorker | full | full | `SendMessage`; delivery from its tool_result |
+ * | waves / waveSizes / maxConcurrency | full | full | derived from spawn/settle instants |
+ * | respawns / repeatedLabels | full | full | same derivation |
+ * | delegationDepth | full | full | a child transcript's own spawn calls |
+ * | timeToFirstSpawn / supervisorWall | full | full | transcript instants |
+ * | idleMs / idlePct / workerUtilization | full | PARTIAL | an agent that never notifies is counted live to the end of the transcript |
+ * | observeThenRespawn / respawnWithoutEvidence | full | full | ordering of spawn vs settle instants |
+ * | workerEvidenceBytes | full | PARTIAL | the child's closing message; 0 for pruned transcripts |
+ * | brain tokens in/out + cache | full | full | main-thread `message.usage` |
+ * | worker tokens in/out + cache | via harness join | full or unavailable | totals are refused if any spawned transcript was pruned; retained per-worker rows remain available |
+ * | perWorker wall | full | full | spawn → settle instants |
+ * | accepted / rejected / emptyPass / settledVerdicts | full | NONE | no per-worker verify step exists |
+ * | brain/worker/total usd, costPerAcceptedPatch | full | NONE | transcripts carry no price |
+ * | patch stats, delivered, verifyPass/Rc | full | NONE | no diff is handed back |
+ * | judgeResolved / Score / Passed / Total | full | NONE | no judge in the loop |
+ * | driverSteerCalls, brainTruncations | full | NONE | no outer driver log, no per-call finish_reason tap |
+ */
 interface ClaudeCodeReaderOptions {
   /** The main session transcript: `~/.claude/projects/<slug>/<sessionId>.jsonl`. */
   readonly transcriptPath: string;
@@ -23,13 +81,6 @@ interface ClaudeCodeReaderOptions {
   readonly steerTools?: readonly string[];
   readonly cancelTools?: readonly string[];
 }
-/**
- * Read a Claude Code session (plus its subagent transcripts) as supervision-tree
- * source bytes. Never throws on a missing artifact.
- */
-export declare function readClaudeCodeSupervisorRun(opts: ClaudeCodeReaderOptions): Promise<SupervisorRunSources>;
-/** A `SupervisorRunReader` over a Claude Code session — the same contract Runtime implements. */
-export declare function claudeCodeSupervisorRunReader(opts: ClaudeCodeReaderOptions): SupervisorRunReader;
 //#endregion
 //#region src/supervisor-run/integrity-types.d.ts
 export declare const SUPERVISOR_RUN_INTEGRITY_SCHEMA = "tangle.supervisor-run-integrity@1";
@@ -93,18 +144,6 @@ interface WriteSupervisorRunOptions {
    */
   readonly reportDir?: string;
 }
-/**
- * Read a completed run, write `run-report.json` + `run-report.md` beside its
- * artifacts, and append the headline block to the run log. Never throws on a
- * missing artifact — a run that produced nothing still yields a report whose
- * every metric says why.
- */
-export declare function writeSupervisorRunReport(runDir: string, opts?: WriteSupervisorRunOptions): Promise<SupervisorRunReport>;
-/**
- * Best-effort wrapper for a hot path: a reporting failure must never kill a run
- * that already produced real work. Returns null and logs the reason instead.
- */
-export declare function writeSupervisorRunReportSafe(runDir: string, opts?: WriteSupervisorRunOptions): Promise<SupervisorRunReport | null>;
 /**
  * Every Runtime supervisor run below `root`.
  *
@@ -139,11 +178,6 @@ interface SupervisorRunCommandIo {
   readonly stdout: (text: string) => void;
   readonly stderr: (text: string) => void;
 }
-/**
- * CLI driver for `agent-eval supervisor-run`. `argv` is everything after the
- * command name, so `['report', runDir, ...flags]`. Returns the process exit code.
- */
-export declare function runSupervisorRunCommand(argv: readonly string[], io?: SupervisorRunCommandIo): Promise<number>;
 //#endregion
 //#region src/supervisor-run/source-facts.d.ts
 interface Tokens {
@@ -425,8 +459,6 @@ interface RuntimeReaderOptions {
  * verdict.
  */
 export declare function readRuntimeSupervisorRun(runDir: string, opts?: RuntimeReaderOptions): Promise<SupervisorRunSources>;
-/** The agent-runtime file-backed layout as a `SupervisorRunReader`. */
-export declare function runtimeSupervisorRunReader(runDir: string, opts?: RuntimeReaderOptions): SupervisorRunReader;
 /**
  * True when a directory holds any file only Runtime's durable layer writes:
  * the spawn journal, the observer journal, or the failure record. A run that
@@ -468,5 +500,5 @@ interface TerminalRecord {
 }
 export declare function readTerminalRecord(input: TerminalRecordInput): TerminalRecord;
 //#endregion
-export { type ClaudeCodeReaderOptions, type CloseRow, type DecisionMetrics, type EconomicsMetrics, type Measured, NO_SOURCE_LIMITS, type NamedResourceReceipt, type OrchestrationMetrics, type OutcomeMetrics, type PatchStats, type PerWorkerRow, type ResourceSpendRecord, type RoleSpend, type RollupCellRow, type RuntimeReaderOptions, SUPERVISOR_RUN_ROLLUP_SCHEMA, SUPERVISOR_RUN_SCHEMA, type SourceLimits, type SpawnRow, type SpendMeasurement, type SpendMeasurements, type SteerAcknowledgementFact, type SteerBreakdown, type SteerRequestFact, type SupervisorJournalDialect, type SupervisorRolloutOptions, type SupervisorRunCommandIo, type SupervisorRunIntegrityEvidence, type SupervisorRunIntegrityIssue, type SupervisorRunIntegrityIssueCode, type SupervisorRunIntegrityOptions, type SupervisorRunIntegrityReport, type SupervisorRunIntegritySeverity, type SupervisorRunNodeRole, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunReportFormat, type SupervisorRunRollup, type SupervisorRunSourceOnlyCheckCode, type SupervisorRunSources, type SupervisorRunTree, type SupervisorRunTreeGap, type SupervisorRunTreeGapCode, type SupervisorStatusSource, type SupervisorTreeFacts, type TerminalFailure, type TerminalRecord, type TerminalRecordInput, type Unavailable, type WallDistribution, type WorkerLogFacts, type WorkerLogSource, type WorkerNativeSession, type WorkerTurnCoverage, type WriteSupervisorRunOptions, isUnavailable, showMeasured, unavailable };
+export { type ClaudeCodeReaderOptions, type CloseRow, type DecisionMetrics, type EconomicsMetrics, type Measured, type NamedResourceReceipt, type OrchestrationMetrics, type OutcomeMetrics, type PatchStats, type PerWorkerRow, type ResourceSpendRecord, type RoleSpend, type RollupCellRow, type RuntimeReaderOptions, type SourceLimits, type SpawnRow, type SpendMeasurement, type SpendMeasurements, type SteerAcknowledgementFact, type SteerBreakdown, type SteerRequestFact, type SupervisorJournalDialect, type SupervisorRolloutOptions, type SupervisorRunCommandIo, type SupervisorRunIntegrityEvidence, type SupervisorRunIntegrityIssue, type SupervisorRunIntegrityIssueCode, type SupervisorRunIntegrityOptions, type SupervisorRunIntegrityReport, type SupervisorRunIntegritySeverity, type SupervisorRunNodeRole, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunReportFormat, type SupervisorRunRollup, type SupervisorRunSourceOnlyCheckCode, type SupervisorRunSources, type SupervisorRunTree, type SupervisorRunTreeGap, type SupervisorRunTreeGapCode, type SupervisorStatusSource, type SupervisorTreeFacts, type TerminalFailure, type TerminalRecord, type TerminalRecordInput, type Unavailable, type WallDistribution, type WorkerLogFacts, type WorkerLogSource, type WorkerNativeSession, type WorkerTurnCoverage, type WriteSupervisorRunOptions, isUnavailable, showMeasured, unavailable };
 //# sourceMappingURL=index.d.ts.map

package/dist/supervisor-run/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/patch-stats.ts","../../src/supervisor-run/reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/report-command.ts","../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/rollup.ts","../../src/supervisor-run/runtime-reader.ts","../../src/supervisor-run/terminal-record.ts"],"mappings":";;;;;;;wBAmDgB,4BACd,KAAK,sBACL,qBACC;;;UC6Bc;;WAEN;;;;;WAKA;WACA;WACA;WACA;WACA;WACA;WACA;;;;;;wBAoMW,4BACpB,MAAM,0BACL,QAAQ;;wBAuVK,8BAA8B,MAAM,0BAA0B;;;qBC3nBjE;KAED;KAEA;KAKA;UAqCK;WACN;WACA;;UAGM;WACN,MAAM;WACN;WACA,UAAU;WACV;WACA;WACA;WACA,mBAAmB;WACnB;WACA,WAAW,SAAS;;UAGd;WACN,eAAe;WACf;WACA;WACA,MAAM;WACN,iBAAiB;;UAGX;;WAEN;;;;;;;;wBC9BK,8BACd,OAAO,uBAAuB,mBAC9B,UAAS,gCACR;;;;wBC5Ca,WAAW,eAAe;;;;;;;;wBCgBpB,qBACpB,gBAAgB,sBAAsB,uBACrC,QAAQ;UAYM;;WAEN;;WAEA;;;;;;WAMA;;;;;;;;wBASW,yBACpB,gBACA,OAAM,4BACL,QAAQ;;;;;wBAwCW,6BACpB,gBACA,OAAM,4BACL,QAAQ;;;;;;;wBAiBW,sBAAsB,eAAe;;;;;;;wBC/F3C,4BAA4B,GAAG;wBA8E/B,4BAA4B,GAAG;wBAsL/B,+BACd,QAAQ,qBACR;;;;;;;;;;;;;KC1QU;;UAQK;WACN,SAAS;WACT,SAAS;;;;;;wBAwDE,wBACpB,yBACA,KAAI,yBACH;;;UCvDO;EACR;EACA;EACA;EACA;;EAEA;;;;;;EAMA;;UAGe;WACN,YAAY,kBAAkB;EACvC,QAAQ;;;;;;EAMR;EACA;;;;;;EAMA;;UA6Ge;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN,2BAA2B;WAC3B;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;;;;;;;IAOP;IACA;IACA;IACA;;IAEA;;IAEA;;IAEA;IACA;;IAEA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;WAMA;WACA;;;;;;;;;;;;;;;;;;KAmDC;wBAiFI,oBAAoB,KAAK,uBAAuB;;;UCva/C;;WAEN;;WAEA,QAAQ;;WAER;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA;WACA;WACA;WACA;;WAEA;;;;;;;wBAwBK,0BACd,KAAK,sBACL,OAAM,2BACL;;;;;;;;wBCnFa,qBAAqB,kBAAkB,wBAAwB;;;UCyrB9D;;;;;;WAMN;;;;;;;;;;;;;;;;;;;wBA0EW,yBACpB,gBACA,OAAM,uBACL,QAAQ;;wBAoFK,2BACd,gBACA,OAAM,uBACL;;;;;;wBASmB,0BAA0B,iBAAiB;;;;qBCj2BpD;;qBAIA;UAEI;;WAEN,OAAO;;WAEP,QAAQ;;;;;WAKR,SAAS;;UAGH;WACN,WAAW;WACX,iBAAiB,SAAS;;;;;WAK1B,WAAW;;WAEX,SAAS,SAAS;;;;;WAKlB;;wBAmBK,mBAAmB,OAAO,sBAAsB"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/patch-stats.ts","../../src/supervisor-run/reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/report-command.ts","../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/rollup.ts","../../src/supervisor-run/runtime-reader.ts","../../src/supervisor-run/terminal-record.ts"],"mappings":";;;;;;;wBAuCgB,4BACd,KAAK,sBACL,qBACC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UCgCc;;WAEN;;;;;WAKA;WACA;WACA;WACA;WACA;WACA;WACA;;;;qBCrFE;KAED;KAEA;KAKA;UAqCK;WACN;WACA;;UAGM;WACN,MAAM;WACN;WACA,UAAU;WACV;WACA;WACA;WACA,mBAAmB;WACnB;WACA,WAAW,SAAS;;UAGd;WACN,eAAe;WACf;WACA;WACA,MAAM;WACN,iBAAiB;;UAGX;;WAEN;;;;;;;;wBC9BK,8BACd,OAAO,uBAAuB,mBAC9B,UAAS,gCACR;;;;wBC5Ca,WAAW,eAAe;;;;;;;;wBCgBpB,qBACpB,gBAAgB,sBAAsB,uBACrC,QAAQ;UAYM;;WAEN;;WAEA;;;;;;WAMA;;;;;;;;wBAeW,sBAAsB,eAAe;;;;;;;wBCtC3C,4BAA4B,GAAG;wBA8E/B,4BAA4B,GAAG;wBAsL/B,+BACd,QAAQ,qBACR;;;;;;;;;;;;;KC1QU;;UAQK;WACN,SAAS;WACT,SAAS;;;;UCIV;EACR;EACA;EACA;EACA;;EAEA;;;;;;EAMA;;UAGQ;WACC,YAAY,kBAAkB;EACvC,QAAQ;;;;;;EAMR;EACA;;;;;;EAMA;;UA6Ge;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN,2BAA2B;WAC3B;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;;;;;;;IAOP;IACA;IACA;IACA;;IAEA;;IAEA;;IAEA;IACA;;IAEA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;WAMA;WACA;;;;;;;;;;;;;;;;;;KAmDC;wBAkcI,oBAAoB,KAAK,uBAAuB;;;UClxB/C;;WAEN;;WAEA,QAAQ;;WAER;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA;WACA;WACA;WACA;;WAEA;;;;;;;wBAwBK,0BACd,KAAK,sBACL,OAAM,2BACL;;;;;;;;wBCzFa,qBAAqB,kBAAkB,wBAAwB;;;UCwrB9D;;;;;;WAMN;;;;;;;;;;;;;;;;;;;wBA0EW,yBACpB,gBACA,OAAM,uBACL,QAAQ;;;;;;wBA0FW,0BAA0B,iBAAiB;;;;qBC11BpD;;qBAIA;UAEI;;WAEN,OAAO;;WAEP,QAAQ;;;;;WAKR,SAAS;;UAGH;WACN,WAAW;WACX,iBAAiB,SAAS;;;;;WAK1B,WAAW;;WAEX,SAAS,SAAS;;;;;WAKlB;;wBAmBK,mBAAmB,OAAO,sBAAsB"}