@tangle-network/agent-eval 0.69.0 → 0.71.0

package/dist/rl.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/off-policy.ts","../src/rl/preferences.ts","../src/rl/run-record-adapters.ts","../src/rl/tournament.ts","../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/adversarial.ts","../src/rl/exporters.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/process-reward.ts","../src/rl/rl-campaign.ts"],"sourcesContent":["/**\n * Test-time compute scaling curves.\n *\n * The test-time-compute frontier paper (Snell et al. 2024) and the\n * subsequent o1-style scaling work both show that LLM-agent capability\n * is a function of the compute budget at inference, not just of the\n * training run. The right way to characterize a candidate is therefore\n * a *curve* — score at compute budgets {1×, 4×, 16×, …} — not a single\n * point.\n *\n * This module ships:\n *\n *   1. The compute-curve harness — `runComputeCurve(runner, budgets)` —\n *      that evaluates one candidate at a sequence of compute budgets\n *      and returns the (compute, score) curve.\n *   2. A best-of-N evaluator — `bestOfN(runner, n, scoreFn)` — the\n *      simplest test-time-compute scaling primitive: sample N\n *      independent rollouts, return the best.\n *   3. A self-consistency evaluator — `selfConsistency(runner, n)` —\n *      the majority-vote variant of best-of-N for tasks with a small\n *      categorical answer space.\n *   4. Pareto-frontier extraction over multiple candidates — given\n *      (candidate, compute, score) tuples, return the set of\n *      candidate-compute combinations that aren't dominated.\n *\n * Caveat: \"compute\" here is the caller's notion of a compute unit. For\n * agent eval that's typically wall-time × parallelism, or token budget,\n * or LLM-call count. We accept whatever the caller provides; the curve\n * is on whatever axis they pick.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface ComputeCurveBudget {\n  /** Identifier — for the report. Common: '1x', '4x', '16x'. */\n  id: string\n  /** Numeric value on the chosen axis (tokens, calls, USD, ms — caller picks). */\n  cost: number\n  /** Free-form metadata (the caller can carry per-budget config). */\n  meta?: Record<string, unknown>\n}\n\nexport interface ComputeCurvePoint {\n  budgetId: string\n  cost: number\n  score: number\n  /** Number of underlying samples used at this budget. */\n  samples: number\n  /** Optional spread / variance information. */\n  std?: number\n  /** Any extra metrics the runner returned. */\n  metrics?: Record<string, number>\n}\n\nexport interface ComputeCurve {\n  candidateId: string\n  points: ComputeCurvePoint[]\n  /** Rough exponent fit: score ≈ a + b * log(cost). Useful for \"how steep is the curve?\" */\n  logSlope: number | null\n  /** Best (highest-score) point on the curve. */\n  best: ComputeCurvePoint\n}\n\nexport interface RunComputeCurveOptions {\n  candidateId: string\n  budgets: ComputeCurveBudget[]\n  /**\n   * Run the candidate at one budget. Returns the realized score plus\n   * optional spread + extra metrics.\n   */\n  runAtBudget: (budget: ComputeCurveBudget) => Promise<{\n    score: number\n    samples: number\n    std?: number\n    metrics?: Record<string, number>\n  }>\n}\n\nexport async function runComputeCurve(opts: RunComputeCurveOptions): Promise<ComputeCurve> {\n  const points: ComputeCurvePoint[] = []\n  for (const budget of opts.budgets) {\n    const r = await opts.runAtBudget(budget)\n    points.push({\n      budgetId: budget.id,\n      cost: budget.cost,\n      score: r.score,\n      samples: r.samples,\n      std: r.std,\n      metrics: r.metrics,\n    })\n  }\n  const sorted = [...points].sort((a, b) => a.cost - b.cost)\n  const logSlope = sorted.length >= 2 ? fitLogSlope(sorted) : null\n  const best = points.reduce((a, b) => (b.score > a.score ? b : a))\n  return { candidateId: opts.candidateId, points: sorted, logSlope, best }\n}\n\nexport interface ComputeBestOfNOptions<O> {\n  /** Number of independent samples to draw. */\n  n: number\n  /** Sampler — produces one rollout. */\n  sample: (sampleIdx: number) => Promise<O>\n  /** Score one rollout. */\n  scoreFn: (rollout: O) => Promise<number> | number\n}\n\nexport interface ComputeBestOfNResult<O> {\n  best: O\n  bestScore: number\n  scores: number[]\n  meanScore: number\n  /** Index of the best rollout, for diagnostics. */\n  bestIndex: number\n}\n\n/** The simplest test-time scaling primitive. */\nexport async function bestOfN<O>(opts: ComputeBestOfNOptions<O>): Promise<ComputeBestOfNResult<O>> {\n  if (opts.n <= 0) throw new ValidationError('bestOfN: n must be > 0')\n  const rollouts: O[] = []\n  const scores: number[] = []\n  for (let i = 0; i < opts.n; i++) {\n    const r = await opts.sample(i)\n    rollouts.push(r)\n    scores.push(await opts.scoreFn(r))\n  }\n  let bestIndex = 0\n  for (let i = 1; i < scores.length; i++) if (scores[i]! > scores[bestIndex]!) bestIndex = i\n  const meanScore = scores.reduce((s, x) => s + x, 0) / scores.length\n  return {\n    best: rollouts[bestIndex]!,\n    bestScore: scores[bestIndex]!,\n    scores,\n    meanScore,\n    bestIndex,\n  }\n}\n\nexport interface SelfConsistencyOptions<O> {\n  n: number\n  sample: (sampleIdx: number) => Promise<O>\n  /** Extract the canonical answer key (string) from a rollout. */\n  answerKey: (rollout: O) => string\n}\n\nexport interface SelfConsistencyResult<O> {\n  /** Modal answer (the majority vote). */\n  answer: string\n  /** Fraction of samples voting for the modal answer in [0, 1]. */\n  agreement: number\n  /** Histogram of all answers. */\n  histogram: Record<string, number>\n  /** A representative rollout that voted for the modal answer. */\n  representative: O\n  /** All rollouts. */\n  rollouts: O[]\n}\n\n/**\n * Self-consistency / majority-vote test-time scaling. For tasks with a\n * small categorical answer space (math problems, multiple choice).\n */\nexport async function selfConsistency<O>(\n  opts: SelfConsistencyOptions<O>,\n): Promise<SelfConsistencyResult<O>> {\n  if (opts.n <= 0) throw new ValidationError('selfConsistency: n must be > 0')\n  const rollouts: O[] = []\n  const histogram: Record<string, number> = {}\n  for (let i = 0; i < opts.n; i++) {\n    const r = await opts.sample(i)\n    rollouts.push(r)\n    const key = opts.answerKey(r)\n    histogram[key] = (histogram[key] ?? 0) + 1\n  }\n  let answer = ''\n  let max = -1\n  for (const [k, v] of Object.entries(histogram)) {\n    if (v > max) {\n      max = v\n      answer = k\n    }\n  }\n  const representative = rollouts.find((r) => opts.answerKey(r) === answer) ?? rollouts[0]!\n  return {\n    answer,\n    agreement: max / opts.n,\n    histogram,\n    representative,\n    rollouts,\n  }\n}\n\n/**\n * Pareto frontier over (candidate, compute, score) tuples. A point is on\n * the frontier iff no other point dominates it in both score (higher\n * better) and cost (lower better). Returns the frontier sorted ascending\n * by cost.\n */\nexport interface ParetoPointInput {\n  candidateId: string\n  budgetId: string\n  cost: number\n  score: number\n}\n\nexport function paretoFrontier(points: ParetoPointInput[]): ParetoPointInput[] {\n  const onFrontier: ParetoPointInput[] = []\n  for (const p of points) {\n    const dominated = points.some(\n      (q) =>\n        q !== p && q.cost <= p.cost && q.score >= p.score && (q.cost < p.cost || q.score > p.score),\n    )\n    if (!dominated) onFrontier.push(p)\n  }\n  return onFrontier.sort((a, b) => a.cost - b.cost)\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction fitLogSlope(points: ComputeCurvePoint[]): number {\n  // OLS slope of score on log(cost). Used as a single-number summary of\n  // how much marginal compute helps. Positive = score improves with\n  // compute; near-zero = capability ceiling reached.\n  const xs = points.map((p) => Math.log(Math.max(1e-12, p.cost)))\n  const ys = points.map((p) => p.score)\n  const n = xs.length\n  const mx = xs.reduce((s, x) => s + x, 0) / n\n  const my = ys.reduce((s, y) => s + y, 0) / n\n  let num = 0\n  let den = 0\n  for (let i = 0; i < n; i++) {\n    num += (xs[i]! - mx) * (ys[i]! - my)\n    den += (xs[i]! - mx) ** 2\n  }\n  return den === 0 ? 0 : num / den\n}\n","/**\n * Contamination probe — held-out perturbation tests.\n *\n * The bug class: once a benchmark scenario set is published, models train\n * on it, and your scores become invalid. SWE-Bench-Verified, GPQA, and\n * MMLU-Pro all exist because their predecessors got contaminated within\n * months. The right defense is to keep a held-out *perturbed* version of\n * every scenario — same task, slightly different surface — and check\n * whether scores diverge significantly. Genuine capability transfers; rote\n * memorization doesn't.\n *\n * This module ships the probe contract:\n *\n *   1. A `ScenarioPerturbation` strategy type — function that produces a\n *      perturbed scenario from an original.\n *   2. `runContaminationProbe({ originals, perturbed, scoreFn })` — runs\n *      both halves and reports per-scenario score divergence + a global\n *      contamination verdict via paired Wilcoxon.\n *   3. Several stock perturbations: `renameVariables`, `shuffleOrder`,\n *      `paraphrasePrompt`, `injectIrrelevantClause`. Each preserves the\n *      task's structural difficulty while breaking surface memorization.\n *\n * The verdict is conservative: if the perturbed-vs-original score\n * difference is statistically significant (BH-adjusted p < 0.05) AND\n * the median drop is > 5 percentage points, we flag *contamination\n * suspected*. False positives are possible (the perturbation might\n * actually be harder); the default is to flag for review, not to\n * autoreject.\n */\n\nimport { ValidationError } from '../errors'\nimport { benjaminiHochberg, wilcoxonSignedRank } from '../statistics'\n\nexport type ScenarioPerturbationKind =\n  | 'rename_variables'\n  | 'shuffle_order'\n  | 'paraphrase'\n  | 'inject_irrelevant_clause'\n  | 'custom'\n\nexport interface ScenarioPerturbation<S> {\n  kind: ScenarioPerturbationKind\n  /** Apply to one scenario, return its perturbed sibling. */\n  apply: (scenario: S) => Promise<S> | S\n  /** Optional id — for the report. */\n  id?: string\n}\n\nexport interface ContaminationProbeInput<S> {\n  /** Identity of every scenario. The probe's `runFingerprint` keys on these. */\n  scenarioId: (s: S) => string\n  /** Original scenarios. */\n  originals: S[]\n  /**\n   * Either pre-computed perturbations (one per original, same order) OR a\n   * `perturbation` strategy that synthesizes them on the fly.\n   */\n  perturbed?: S[]\n  perturbation?: ScenarioPerturbation<S>\n  /**\n   * Run the policy/agent against one scenario and return a scalar score\n   * in [0, 1]. The probe doesn't care what the policy is — that's the\n   * caller's contract.\n   */\n  scoreFn: (s: S) => Promise<number>\n}\n\nexport interface ContaminationProbeOptions {\n  /** Drop scores below this from the probe; treats partial failures separately. Default 0. */\n  scoreFloor?: number\n  /**\n   * BH-FDR threshold for declaring contamination on each per-scenario\n   * delta. Default 0.05.\n   */\n  fdr?: number\n  /**\n   * Minimum median per-scenario drop to flag global contamination. Default\n   * 0.05 (5 percentage points). Smaller drops may be noise.\n   */\n  minMedianDrop?: number\n}\n\nexport interface ContaminationProbeReport {\n  perScenario: Array<{\n    scenarioId: string\n    originalScore: number\n    perturbedScore: number\n    delta: number // perturbed - original (negative = drop)\n    /** Per-scenario q-value (single-test BH for a single scenario). Mainly for display. */\n    qValue: number\n  }>\n  /** Wilcoxon paired-test on the deltas. */\n  pairedTest: { w: number; p: number }\n  medianDelta: number\n  meanDelta: number\n  contaminationSuspected: boolean\n  reason: string\n  /** Number of scenarios processed. */\n  n: number\n}\n\nexport async function runContaminationProbe<S>(\n  input: ContaminationProbeInput<S>,\n  opts: ContaminationProbeOptions = {},\n): Promise<ContaminationProbeReport> {\n  const fdr = opts.fdr ?? 0.05\n  const minMedianDrop = opts.minMedianDrop ?? 0.05\n  const floor = opts.scoreFloor ?? 0\n\n  if (!input.perturbed && !input.perturbation) {\n    throw new ValidationError(\n      'runContaminationProbe: must supply either `perturbed` or `perturbation`.',\n    )\n  }\n  const perturbed: S[] =\n    input.perturbed ?? (await Promise.all(input.originals.map((s) => input.perturbation!.apply(s))))\n  if (perturbed.length !== input.originals.length) {\n    throw new ValidationError(\n      `runContaminationProbe: perturbed length ${perturbed.length} ≠ originals ${input.originals.length}`,\n    )\n  }\n\n  // Score both halves.\n  const origScores = await Promise.all(input.originals.map((s) => input.scoreFn(s)))\n  const pertScores = await Promise.all(perturbed.map((s) => input.scoreFn(s)))\n\n  const perScenario = input.originals.map((s, i) => ({\n    scenarioId: input.scenarioId(s),\n    originalScore: origScores[i]!,\n    perturbedScore: pertScores[i]!,\n    delta: pertScores[i]! - origScores[i]!,\n    qValue: NaN,\n  }))\n\n  // Drop scenarios below the floor (partial failures we don't trust).\n  const valid = perScenario.filter((p) => p.originalScore >= floor && p.perturbedScore >= floor)\n  if (valid.length < 4) {\n    return {\n      perScenario,\n      pairedTest: { w: 0, p: 1 },\n      medianDelta: 0,\n      meanDelta: 0,\n      contaminationSuspected: false,\n      reason: `insufficient valid scenarios (n=${valid.length}, need ≥ 4)`,\n      n: valid.length,\n    }\n  }\n\n  const origValid = valid.map((p) => p.originalScore)\n  const pertValid = valid.map((p) => p.perturbedScore)\n  const pairedTest = wilcoxonSignedRank(origValid, pertValid)\n  const deltas = valid.map((p) => p.delta)\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const median = sortedDeltas[Math.floor(sortedDeltas.length / 2)]!\n  const mean = deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n  // Per-scenario q-values via BH on a synthetic per-scenario p-value\n  // (one-sample bootstrap; we use the absolute delta normalized by median\n  // as a coarse signal — this is a display aid, the load-bearing test\n  // is the global Wilcoxon).\n  const pseudoP = valid.map((p) => Math.min(1, Math.max(1e-6, 1 - Math.abs(p.delta) / 1)))\n  const { qValues } = benjaminiHochberg(pseudoP, fdr)\n  for (let i = 0; i < valid.length; i++) {\n    const v = valid[i]!\n    const idx = perScenario.findIndex((p) => p.scenarioId === v.scenarioId)\n    if (idx >= 0) perScenario[idx]!.qValue = qValues[i]!\n  }\n\n  const contaminationSuspected = pairedTest.p < fdr && median <= -minMedianDrop\n  const reason = contaminationSuspected\n    ? `paired p=${pairedTest.p.toFixed(4)} < ${fdr} and median drop ${median.toFixed(4)} ≥ ${minMedianDrop}`\n    : pairedTest.p >= fdr\n      ? `no significant difference (paired p=${pairedTest.p.toFixed(4)})`\n      : `significant but small effect (median delta ${median.toFixed(4)})`\n\n  return {\n    perScenario,\n    pairedTest,\n    medianDelta: median,\n    meanDelta: mean,\n    contaminationSuspected,\n    reason,\n    n: valid.length,\n  }\n}\n\n// ── Stock perturbations ──────────────────────────────────────────────────\n\n/**\n * Identifier-rename perturbation for code/text scenarios. Replaces every\n * occurrence of the listed identifiers with synthesized aliases. Use when\n * the scenario's structural difficulty is independent of variable names\n * (e.g. SWE-Bench-style coding tasks).\n */\nexport function renameVariables<S extends { prompt: string }>(\n  identifiers: string[],\n  rename: (name: string, idx: number) => string = (n, i) => `${n}_${((i % 26) + 10).toString(36)}`,\n): ScenarioPerturbation<S> {\n  return {\n    kind: 'rename_variables',\n    apply(scenario) {\n      let prompt = scenario.prompt\n      identifiers.forEach((id, i) => {\n        const replacement = rename(id, i)\n        const re = new RegExp(`\\\\b${escapeRegex(id)}\\\\b`, 'g')\n        prompt = prompt.replace(re, replacement)\n      })\n      return { ...scenario, prompt }\n    },\n  }\n}\n\n/**\n * Order-shuffle perturbation. Reshuffles a list-shaped section of the\n * prompt (for QA scenarios that present options A/B/C/D — answer depends\n * on the option labels, not order). Caller provides the section extractor.\n */\nexport function shuffleOrder<S extends { prompt: string }>(\n  shuffleSection: (prompt: string, rng: () => number) => string,\n  seed: number,\n): ScenarioPerturbation<S> {\n  let s = seed >>> 0\n  const rng = (): number => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n  return {\n    kind: 'shuffle_order',\n    apply(scenario) {\n      const newPrompt = shuffleSection(scenario.prompt, rng)\n      return { ...scenario, prompt: newPrompt }\n    },\n  }\n}\n\n/**\n * Inject-irrelevant-clause perturbation. Adds a benign sentence that\n * shouldn't change the answer. Tests for \"did the model just memorize\n * the input string.\"\n */\nexport function injectIrrelevantClause<S extends { prompt: string }>(\n  clause: string,\n  position: 'prefix' | 'suffix' = 'prefix',\n): ScenarioPerturbation<S> {\n  return {\n    kind: 'inject_irrelevant_clause',\n    apply(scenario) {\n      const prompt =\n        position === 'prefix' ? `${clause} ${scenario.prompt}` : `${scenario.prompt} ${clause}`\n      return { ...scenario, prompt }\n    },\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction escapeRegex(s: string): string {\n  return s.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&')\n}\n","/**\n * Off-policy evaluation primitives.\n *\n * Standard inverse-probability-weighted (IPS), self-normalized\n * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the\n * value of a *target* policy given trajectories collected under a\n * *behavior* policy. This is the canonical RL eval task: \"we have last\n * week's runs, we changed the policy — how would the new one do without\n * re-running?\"\n *\n * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan\n * & Joachims 2015 for SNIPS) but the *application* to LLM-agent\n * evaluation needs care:\n *\n *   - The \"policy\" is the (prompt, tool config, model snapshot) triple.\n *     Two policies have the same probability over an action *iff* their\n *     LLM call would emit the same token with the same probability —\n *     which is generally unknowable without the model log-probs.\n *   - For LLM agents, propensity scores must be supplied by the caller\n *     (logged in the trace, recovered from token log-probs, or estimated\n *     via a learned propensity model). We do NOT estimate propensity here.\n *   - Doubly-robust requires a Q-function (model-based reward predictor).\n *     We accept any callable; consumers pass either a tabular average,\n *     a regression fit, or a learned reward model.\n *\n * Bias / variance tradeoffs:\n *   - IPS: unbiased; high variance for small overlap, infinite variance\n *     when target has support outside behavior.\n *   - SNIPS: lower variance, slight bias; usually preferred in practice.\n *   - DR: doubly-robust — unbiased if either propensity OR Q-function is\n *     correct. Lowest practical variance when Q is decent. Use this.\n *\n * Caveat the panel will land: on the LLM-agent setting, propensity scores\n * recovered from token log-probs are noisy, the action space is enormous,\n * and overlap is often poor. These estimators are useful but not magic;\n * complement with `replayCampaign` (exact replay where the request hashes\n * match) for high-confidence answers and OPE for the gap.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface OffPolicyTrajectory {\n  /** Stable id, for traceability through the dataset. */\n  runId: string\n  /** Reward observed under the behavior policy (the realized outcome). */\n  reward: number\n  /**\n   * Behavior-policy probability of the action that was taken. For LLM\n   * agents this is typically `exp(sum(token_log_probs))` over the chosen\n   * trajectory. Must be in (0, 1].\n   */\n  behaviorProb: number\n  /**\n   * Target-policy probability of the same action. For replay-style\n   * counterfactual evaluation this is what the *new* policy would have\n   * assigned to the *old* trajectory. Must be in [0, 1].\n   */\n  targetProb: number\n  /**\n   * Optional model-based reward prediction at the same context. Used by\n   * `doublyRobust`. Set to `null` for IPS-only evaluation.\n   */\n  qHat?: number | null\n}\n\nexport interface OffPolicyEstimate {\n  /** Estimated value of the target policy. */\n  value: number\n  /** Standard error of the estimate. */\n  standardError: number\n  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */\n  effectiveSampleSize: number\n  /** Number of trajectories used. */\n  n: number\n  /**\n   * Diagnostic: maximum importance weight observed. Large values (>>10x\n   * mean) are a red flag — variance is dominated by a few outliers.\n   */\n  maxImportanceWeight: number\n}\n\nexport interface OffPolicyOptions {\n  /**\n   * Cap importance weights at this value (Ionides 2008 truncated IS) to\n   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).\n   * Set e.g. `10` for stable estimates when the policies are close.\n   */\n  weightCap?: number\n  /** Reward clipping range. Default `[0, 1]`. */\n  rewardClip?: { low: number; high: number }\n}\n\n/**\n * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator\n * of E[reward under target policy]. Variance scales with the spread of\n * target/behavior ratios.\n */\nexport function inverseProbabilityWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n\n  if (trajectories.length === 0) {\n    return zeroEstimate()\n  }\n\n  const weights: number[] = []\n  const weightedRewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    weights.push(w)\n    weightedRewards.push(w * r)\n    if (w > maxW) maxW = w\n  }\n  const n = weights.length\n  const value = weightedRewards.reduce((s, x) => s + x, 0) / n\n  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at\n * the cost of small bias (vanishing as N grows). The right default for\n * LLM-agent evaluation where overlap is often poor.\n */\nexport function selfNormalizedImportanceWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const weights: number[] = []\n  const rewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    weights.push(w)\n    rewards.push(clamp(t.reward, clip.low, clip.high))\n    if (w > maxW) maxW = w\n  }\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i]!, 0)\n  const value = sumW === 0 ? 0 : sumWR / sumW\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  // Influence-function-based SE for SNIPS (Owen 2013, Ch. 9).\n  const phi = weights.map((w, i) => w * (rewards[i]! - value))\n  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW)\n  return {\n    value,\n    standardError: Math.sqrt(variance),\n    effectiveSampleSize: effN,\n    n: trajectories.length,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).\n *\n *     V_DR = (1/N) * sum_i [ q_hat_i + (target_prob_i / behavior_prob_i) * (r_i - q_hat_i) ]\n *\n * Unbiased if EITHER:\n *   - the importance ratios are correct (IPS-style validity), OR\n *   - the Q-hat function is correct (model-based validity).\n *\n * In practice both are imperfect, but the residual bias is the *product*\n * of both errors — much smaller than either alone. This is why DR is the\n * default in production OPE pipelines.\n *\n * Requires `qHat` on every trajectory. If any are `null`, the estimator\n * falls back to SNIPS for those entries (loud-fallback behavior; the\n * report's `n` reflects the full set but `effectiveSampleSize` accounts\n * for the lost variance reduction).\n */\nexport function doublyRobust(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const contributions: number[] = []\n  let maxW = 0\n  let sumW = 0\n  let sumW2 = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`)\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    const q =\n      typeof t.qHat === 'number' && Number.isFinite(t.qHat)\n        ? clamp(t.qHat, clip.low, clip.high)\n        : null\n    if (q === null) {\n      contributions.push(w * r) // fallback: IPS for this entry\n    } else {\n      contributions.push(q + w * (r - q))\n    }\n    if (w > maxW) maxW = w\n    sumW += w\n    sumW2 += w * w\n  }\n  const n = contributions.length\n  const value = contributions.reduce((s, x) => s + x, 0) / n\n  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Convenience: run all three estimators and return them side-by-side.\n * The recommended diagnostic — agreement across estimators is a much\n * stronger signal than any single one.\n */\nexport function offPolicyEstimateAll(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): { ips: OffPolicyEstimate; snips: OffPolicyEstimate; dr: OffPolicyEstimate } {\n  return {\n    ips: inverseProbabilityWeighting(trajectories, opts),\n    snips: selfNormalizedImportanceWeighting(trajectories, opts),\n    dr: doublyRobust(trajectories, opts),\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction zeroEstimate(): OffPolicyEstimate {\n  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 }\n}\n\nfunction clamp(x: number, lo: number, hi: number): number {\n  if (!Number.isFinite(x)) return lo\n  return Math.max(lo, Math.min(hi, x))\n}\n","/**\n * Preference dataset extraction — bridge from `RunRecord[]` to RL training.\n *\n * Production RLHF / DPO / KTO / SimPO pipelines need preference triples:\n * `(prompt, chosen, rejected)`. The campaign artifact already contains the\n * ingredients — every (variantId, scenarioId, seed) cell is a candidate\n * that ran the same prompt against the same scenario, scored by the same\n * judge — but turning that into a clean preference dataset requires\n * deciding *what counts as a preference*.\n *\n * This module ships three preference-extraction strategies with explicit\n * tradeoffs, plus a unified output type compatible with HuggingFace TRL,\n * Anthropic finetuning JSONL, and OpenAI fine-tuning APIs. The strategies\n * are deliberately not auto-magical — picking the wrong one corrupts the\n * gradient.\n *\n * Strategies:\n *\n *   1. **`paired-by-scenario-and-seed`** — exact-match comparisons. For\n *      each scenario × seed pair, compare every (variantA, variantB) on\n *      that exact (scenario, seed). Matches scenarios so the comparison\n *      isolates variant effects. Highest signal-to-noise; smallest\n *      dataset (only matched pairs count).\n *\n *   2. **`paired-by-scenario`** — looser matching. For each scenario,\n *      compare every (variantA, variantB) where both have ≥ 1 run on the\n *      same scenario. Aggregates across seeds to compute mean scores per\n *      (variant, scenario), then forms preferences from the means. More\n *      data, lower per-pair signal.\n *\n *   3. **`top-vs-bottom`** — coarsest. Within each scenario, the highest-\n *      scoring run is `chosen`, the lowest is `rejected`. Smallest dataset\n *      per scenario but biggest score gap per pair. Useful for early\n *      bootstrapping when you have few variants.\n *\n * The output `PreferenceTriple` is *agent-eval-canonical* but trivially\n * mappable to TRL's `DPODataset` shape (`prompt`, `chosen`, `rejected`)\n * via the `toTRLFormat` helper.\n */\n\nimport type { RunRecord } from '../run-record'\n\nexport type PreferenceStrategy =\n  | 'paired-by-scenario-and-seed'\n  | 'paired-by-scenario'\n  | 'top-vs-bottom'\n\nexport interface PreferenceTriple {\n  /** The scenario (input) the variants were run against. */\n  scenarioId: string\n  /** RunRecord ids on each side, for traceability. */\n  chosenRunId: string\n  rejectedRunId: string\n  /** Variant ids — load-bearing for the RL update. */\n  chosenVariantId: string\n  rejectedVariantId: string\n  /** The score gap between chosen and rejected. Larger = stronger signal. */\n  marginScore: number\n  /**\n   * Optional `(chosen_score, rejected_score)` pair for soft-margin DPO\n   * variants. Omitted for `top-vs-bottom` runs that don't carry meaningful\n   * scalar gaps.\n   */\n  scores?: { chosen: number; rejected: number }\n  /** Tie-breaker — when multiple seeds match this scenario, the one used. */\n  seed?: number\n  /**\n   * Free-form metadata propagated from the run records — e.g. original\n   * prompt-hash, model, etc. Lets the RL trainer reconstruct the prompt.\n   */\n  meta: {\n    chosenPromptHash: string\n    rejectedPromptHash: string\n    chosenConfigHash: string\n    rejectedConfigHash: string\n    chosenModel: string\n    rejectedModel: string\n  }\n}\n\nexport interface ExtractPreferencesOptions {\n  strategy?: PreferenceStrategy\n  /**\n   * Minimum score gap required to admit a pair. Pairs below this are\n   * dropped — they're noise, not signal. Default 0.05 (5% of [0,1]).\n   */\n  minMargin?: number\n  /**\n   * Optional split tag filter — restrict to runs from one split. Default\n   * `'holdout'` (the canonical \"real\" signal).\n   */\n  splitTag?: RunRecord['splitTag']\n  /**\n   * Optional reward extractor that overrides `outcome.holdoutScore` /\n   * `outcome.searchScore`. Use to drive preferences off a verifiable\n   * reward instead of the headline score.\n   */\n  rewardOf?: (run: RunRecord) => number | null\n}\n\nexport interface PreferenceExtractionReport {\n  pairs: PreferenceTriple[]\n  /** Number of (scenario, seed) cells inspected. */\n  cellsInspected: number\n  /** Number of pairs filtered by `minMargin`. */\n  pairsBelowMargin: number\n  /** Number of cells with only one variant (no comparison possible). */\n  cellsSingleton: number\n  /** Strategy used. */\n  strategy: PreferenceStrategy\n}\n\nconst SPLIT_TAG_DEFAULT: RunRecord['splitTag'] = 'holdout'\n\nconst DEFAULT_REWARD = (run: RunRecord): number | null => {\n  const v = run.outcome.holdoutScore ?? run.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\n/**\n * Convert `RunRecord[]` to preference triples for RL training.\n *\n * Returns a structured report so callers can see how much data was\n * dropped and why (low-margin pairs, singleton cells). For production\n * pipelines, you usually want to:\n *\n *   1. Run a campaign producing 5–10 variants × 50–200 scenarios × 3 seeds\n *   2. Call this with `strategy: 'paired-by-scenario-and-seed'` and a\n *      verifiable-reward extractor as `rewardOf`\n *   3. Pass `report.pairs` to `toTRLFormat` and pipe to your DPO trainer\n */\nexport function extractPreferences(\n  runs: RunRecord[],\n  opts: ExtractPreferencesOptions = {},\n): PreferenceExtractionReport {\n  const strategy = opts.strategy ?? 'paired-by-scenario-and-seed'\n  const minMargin = opts.minMargin ?? 0.05\n  const splitTag = opts.splitTag ?? SPLIT_TAG_DEFAULT\n  const rewardOf = opts.rewardOf ?? DEFAULT_REWARD\n\n  const filtered = runs.filter((r) => r.splitTag === splitTag)\n  const scoredEntries: Array<{ run: RunRecord; score: number }> = []\n  for (const run of filtered) {\n    const s = rewardOf(run)\n    if (s === null) continue\n    scoredEntries.push({ run, score: s })\n  }\n\n  const pairs: PreferenceTriple[] = []\n  let pairsBelowMargin = 0\n  let cellsSingleton = 0\n  let cellsInspected = 0\n\n  if (strategy === 'paired-by-scenario-and-seed') {\n    // Group by (scenarioId, seed). Canonical key is `run.scenarioId`,\n    // populated by `runEvalCampaign` and the adapters; falls back to\n    // `outcome.raw.scenario_id` then `experimentId` when absent.\n    const groups = new Map<string, Array<{ run: RunRecord; score: number }>>()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      const key = `${sid}::${e.run.seed}`\n      const arr = groups.get(key) ?? []\n      arr.push(e)\n      groups.set(key, arr)\n    }\n\n    for (const [key, members] of groups.entries()) {\n      cellsInspected++\n      if (members.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      for (let i = 0; i < members.length; i++) {\n        for (let j = i + 1; j < members.length; j++) {\n          const a = members[i]!\n          const b = members[j]!\n          if (a.run.candidateId === b.run.candidateId) continue\n          const result = makePair(a, b, key.split('::')[0]!, minMargin)\n          if (result.kind === 'admit') pairs.push(result.pair)\n          else pairsBelowMargin++\n        }\n      }\n    }\n  } else if (strategy === 'paired-by-scenario') {\n    // Group by scenarioId → average per (variantId, scenarioId) across seeds.\n    const byScenarioVariant = new Map<\n      string,\n      Map<string, { run: RunRecord; sum: number; n: number }>\n    >()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      let perScenario = byScenarioVariant.get(sid)\n      if (!perScenario) {\n        perScenario = new Map()\n        byScenarioVariant.set(sid, perScenario)\n      }\n      const cur = perScenario.get(e.run.candidateId)\n      if (cur) {\n        cur.sum += e.score\n        cur.n++\n      } else perScenario.set(e.run.candidateId, { run: e.run, sum: e.score, n: 1 })\n    }\n    for (const [sid, perVariant] of byScenarioVariant.entries()) {\n      cellsInspected++\n      const arr = [...perVariant.entries()].map(([vid, agg]) => ({\n        run: agg.run,\n        score: agg.sum / agg.n,\n        variantId: vid,\n      }))\n      if (arr.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      for (let i = 0; i < arr.length; i++) {\n        for (let j = i + 1; j < arr.length; j++) {\n          const result = makePair(arr[i]!, arr[j]!, sid, minMargin)\n          if (result.kind === 'admit') pairs.push(result.pair)\n          else pairsBelowMargin++\n        }\n      }\n    }\n  } else {\n    // top-vs-bottom: per scenario, top vs bottom only.\n    const byScenario = new Map<string, Array<{ run: RunRecord; score: number }>>()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      const arr = byScenario.get(sid) ?? []\n      arr.push(e)\n      byScenario.set(sid, arr)\n    }\n    for (const [sid, arr] of byScenario.entries()) {\n      cellsInspected++\n      if (arr.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      const sorted = [...arr].sort((a, b) => a.score - b.score)\n      const top = sorted[sorted.length - 1]!\n      const bot = sorted[0]!\n      if (top.run.candidateId === bot.run.candidateId) {\n        cellsSingleton++\n        continue\n      }\n      const result = makePair(bot, top, sid, minMargin)\n      if (result.kind === 'admit') pairs.push(result.pair)\n      else pairsBelowMargin++\n    }\n  }\n\n  return { pairs, cellsInspected, pairsBelowMargin, cellsSingleton, strategy }\n}\n\n/**\n * TRL-compatible export. TRL's `DPODataset` is `{ prompt, chosen, rejected }`\n * but the prompt isn't stored on the RunRecord — only its hash. The caller\n * passes a `promptOf(promptHash)` lookup that the TRL trainer can use.\n */\nexport function toTRLFormat(\n  triples: PreferenceTriple[],\n  promptOf: (hash: string) => string,\n): Array<{ prompt: string; chosen: string; rejected: string }> {\n  return triples.map((t) => ({\n    prompt: promptOf(t.meta.chosenPromptHash),\n    chosen: t.meta.chosenPromptHash, // caller substitutes the model output via the runId map\n    rejected: t.meta.rejectedPromptHash,\n  }))\n}\n\n/**\n * Anthropic finetuning JSONL export — `{ system, user, assistant_chosen, assistant_rejected }`\n * shape. Same caveat as TRL: prompt + outputs are content the caller has\n * to map back from the run record / raw event log.\n */\nexport function toAnthropicFormat(\n  triples: PreferenceTriple[],\n): Array<{ scenarioId: string; chosenRunId: string; rejectedRunId: string; margin: number }> {\n  return triples.map((t) => ({\n    scenarioId: t.scenarioId,\n    chosenRunId: t.chosenRunId,\n    rejectedRunId: t.rejectedRunId,\n    margin: t.marginScore,\n  }))\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makePair(\n  a: { run: RunRecord; score: number },\n  b: { run: RunRecord; score: number },\n  scenarioId: string,\n  minMargin: number,\n): { kind: 'admit'; pair: PreferenceTriple } | { kind: 'reject' } {\n  const margin = Math.abs(a.score - b.score)\n  if (margin < minMargin) return { kind: 'reject' }\n  const [chosen, rejected] = a.score > b.score ? [a, b] : [b, a]\n  return {\n    kind: 'admit',\n    pair: {\n      scenarioId,\n      chosenRunId: chosen.run.runId,\n      rejectedRunId: rejected.run.runId,\n      chosenVariantId: chosen.run.candidateId,\n      rejectedVariantId: rejected.run.candidateId,\n      marginScore: chosen.score - rejected.score,\n      scores: { chosen: chosen.score, rejected: rejected.score },\n      seed: chosen.run.seed === rejected.run.seed ? chosen.run.seed : undefined,\n      meta: {\n        chosenPromptHash: chosen.run.promptHash,\n        rejectedPromptHash: rejected.run.promptHash,\n        chosenConfigHash: chosen.run.configHash,\n        rejectedConfigHash: rejected.run.configHash,\n        chosenModel: chosen.run.model,\n        rejectedModel: rejected.run.model,\n      },\n    },\n  }\n}\n\n/**\n * Canonical scenario key for a RunRecord. Three-tier fallback:\n *   1. `run.scenarioId` — populated by `runEvalCampaign` and every adapter\n *   2. `run.outcome.raw.scenario_id` — string or numeric, when present\n *   3. `run.experimentId` — worst-case bucket\n */\nfunction scenarioOf(run: RunRecord): string {\n  if (typeof run.scenarioId === 'string' && run.scenarioId.length > 0) return run.scenarioId\n  const fromRaw = run.outcome.raw.scenario_id\n  if (typeof fromRaw === 'number' && Number.isFinite(fromRaw)) return String(fromRaw)\n  if (typeof fromRaw === 'string') return fromRaw\n  return run.experimentId\n}\n","/**\n * Adapters: convert measurement outputs into the canonical `RunRecord[]`\n * artifact that `replayCache`, `pairedEvalueSequence`, and\n * `rubricPredictiveValidity` consume. Two sources:\n *   - `campaignToRunRecords` — the campaign substrate's per-cell results\n *     (the modern path: `runCampaign` / `runImprovementLoop` → records).\n *   - `verificationReportToRunRecord` — a `MultiLayerVerifier` report.\n *\n * Adapters are thin and explicit — every mandatory `RunRecord` field comes\n * from a caller-supplied context (`commitSha`, `model`, `promptHash`,\n * `configHash`) plus the cell's runtime data. The validator still rejects\n * bare-alias model strings — the caller snapshot-pins.\n */\n\nimport type { CampaignResult } from '../campaign'\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport type { RunRecord, RunSplitTag } from '../run-record'\n\nexport interface AdapterContext {\n  /** Logical experiment id — typically the campaign or sweep identifier. */\n  experimentId: string\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** Hash of the effective prompt sent to the model. */\n  promptHash: string\n  /** Hash of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Default split tag. Default `'search'`. */\n  splitTag?: RunSplitTag\n  /** Default cost in USD when the source doesn't record one. Default `0`. */\n  defaultCostUsd?: number\n}\n\n/**\n * Convert a `CampaignResult` into canonical `RunRecord[]` — one record per\n * scored cell. The cell's mean judge composite becomes the split score; every\n * judge dimension is carried through to `outcome.raw`. A cell that errored\n * becomes a record with `failureMode: 'cell_error'` (kept, not dropped — an\n * unscored cell is signal). `candidateId` identifies the measured surface\n * (defaults to the campaign manifest hash).\n */\nexport function campaignToRunRecords(\n  campaign: CampaignResult,\n  ctx: AdapterContext & { candidateId?: string },\n): RunRecord[] {\n  const splitTag = ctx.splitTag ?? 'search'\n  const candidateId = ctx.candidateId ?? campaign.manifestHash\n  return campaign.cells.map((cell) => {\n    const composites = Object.values(cell.judgeScores).map((s) => s.composite)\n    const score =\n      composites.length > 0 ? composites.reduce((a, b) => a + b, 0) / composites.length : 0\n    const raw: Record<string, number> = { rep: cell.rep, duration_ms: cell.durationMs }\n    for (const judge of Object.values(cell.judgeScores)) {\n      for (const [dim, value] of Object.entries(judge.dimensions)) {\n        if (Number.isFinite(value)) raw[`dim.${dim}`] = value\n      }\n    }\n    if (typeof cell.generation === 'number') raw.generation = cell.generation\n    const outcome: RunRecord['outcome'] = { raw }\n    if (splitTag === 'holdout') outcome.holdoutScore = score\n    else outcome.searchScore = score\n    return {\n      runId: cell.cellId,\n      experimentId: ctx.experimentId,\n      candidateId,\n      seed: cell.seed,\n      model: ctx.model,\n      promptHash: ctx.promptHash,\n      configHash: ctx.configHash,\n      commitSha: ctx.commitSha,\n      wallMs: cell.durationMs,\n      costUsd: Number.isFinite(cell.costUsd) ? cell.costUsd : (ctx.defaultCostUsd ?? 0),\n      tokenUsage: { input: 0, output: 0 },\n      outcome,\n      failureMode: cell.error ? 'cell_error' : undefined,\n      splitTag,\n      scenarioId: cell.scenarioId,\n    }\n  })\n}\n\n/**\n * Convert a `MultiLayerVerifier` `VerificationReport` into a `RunRecord`.\n * `outcome.searchScore` (or `holdoutScore`) is `report.blendedScore`;\n * `outcome.raw` carries every layer's score + a pass indicator; `failureMode`\n * is the first failing layer's reason.\n */\nexport function verificationReportToRunRecord(\n  report: VerificationReport,\n  ctx: AdapterContext & { candidateId: string; scenarioId?: string },\n  opts: { runId?: string } = {},\n): RunRecord {\n  const splitTag = ctx.splitTag ?? 'search'\n  const runId = opts.runId ?? `run-${ctx.candidateId}-${ctx.experimentId}-${report.startedAt}`\n\n  const raw: Record<string, number> = {\n    pass_count: report.passCount,\n    fail_count: report.failCount,\n    error_count: report.errorCount,\n    skipped_count: report.skippedCount,\n    duration_ms: report.durationMs,\n    blended_score: report.blendedScore,\n  }\n  for (const layer of report.layers) {\n    if (typeof layer.score === 'number') raw[`layer.${layer.layer}`] = layer.score\n    raw[`layer_${layer.layer}_pass`] = layer.status === 'pass' ? 1 : 0\n    if (layer.diagnostics) {\n      for (const [k, v] of Object.entries(layer.diagnostics)) {\n        if (typeof v === 'number' && Number.isFinite(v)) raw[`layer.${layer.layer}.${k}`] = v\n      }\n    }\n  }\n\n  const firstFail = report.layers.find((l) => l.status === 'fail' || l.status === 'error')\n  const outcome: RunRecord['outcome'] = { raw }\n  if (splitTag === 'holdout') outcome.holdoutScore = report.blendedScore\n  else outcome.searchScore = report.blendedScore\n\n  return {\n    runId,\n    experimentId: ctx.experimentId,\n    candidateId: ctx.candidateId,\n    seed: 0,\n    model: ctx.model,\n    promptHash: ctx.promptHash,\n    configHash: ctx.configHash,\n    commitSha: ctx.commitSha,\n    wallMs: report.durationMs,\n    costUsd: ctx.defaultCostUsd ?? 0,\n    tokenUsage: { input: 0, output: 0 },\n    outcome,\n    failureMode: firstFail ? failureModeFromLayer(firstFail) : undefined,\n    splitTag,\n    scenarioId: ctx.scenarioId,\n  }\n}\n\nfunction failureModeFromLayer(layer: LayerResult): string {\n  if (layer.status === 'error') return `layer_${layer.layer}_error`\n  if (layer.status === 'fail') return `layer_${layer.layer}_fail`\n  if (layer.status === 'timeout') return `layer_${layer.layer}_timeout`\n  return `layer_${layer.layer}_${layer.status}`\n}\n","/**\n * Bradley-Terry / Elo tournament evaluation.\n *\n * For multi-candidate sweeps, comparing every candidate's score against\n * a fixed comparator wastes information — the comparator becomes a high-\n * variance reference and rank flips between near-tied middle-rank\n * candidates are dominated by noise. Pairwise tournaments fix this:\n * every (i, j) pair contributes a comparison to a Bradley-Terry MLE that\n * estimates each candidate's strength on a unified scale.\n *\n * For online updating (rolling campaigns where new candidates arrive\n * over time), we also ship classical Elo with configurable K-factor.\n *\n * References:\n *   - Bradley, R. A., Terry, M. E. (1952). Rank analysis of incomplete\n *     block designs. Biometrika, 39(3/4), 324–345.\n *   - Hunter, D. R. (2004). MM algorithms for generalized Bradley-Terry\n *     models. Annals of Statistics, 32(1), 384–406. (The MLE algorithm\n *     used here.)\n *   - Elo, A. E. (1978). The Rating of Chess Players, Past and Present.\n *\n * This is a useful primitive because most LLM-eval communities (Chatbot\n * Arena, AlpacaEval, ELO-style ablation) have converged on pairwise\n * tournament eval as the most sample-efficient and most rank-stable\n * method when you have many candidates.\n */\n\nexport interface PairwiseOutcome {\n  /** Winner candidate id. */\n  winner: string\n  /** Loser candidate id. */\n  loser: string\n  /**\n   * Optional draw flag. When true, both candidates get half-credit\n   * (Bradley-Terry handles draws as half-wins for each side).\n   */\n  draw?: boolean\n  /**\n   * Optional weight — useful if some pairwise comparisons are stronger\n   * signals than others (e.g. a paired test with a wider score gap is\n   * a more confident comparison). Default 1.\n   */\n  weight?: number\n}\n\nexport interface BradleyTerryRating {\n  candidateId: string\n  /** Latent strength θ ≥ 0 from the BT MLE. */\n  strength: number\n  /** Log-strength = log(θ) — interpretable on a linear scale. */\n  logStrength: number\n  /** Number of pairwise comparisons this candidate appears in. */\n  n: number\n  /** Win count (+ 0.5 per draw). */\n  wins: number\n}\n\nexport interface BradleyTerryFit {\n  ratings: BradleyTerryRating[]\n  /** Iterations of the MM algorithm before convergence. */\n  iterations: number\n  /** Final maximum |θ_new - θ_old| / θ_old. */\n  finalDelta: number\n  converged: boolean\n}\n\n/**\n * Bradley-Terry MLE via Hunter's MM algorithm.\n *\n * Iteration: θ_i^new = W_i / Σ_{j ≠ i} N_ij / (θ_i + θ_j)\n *   where W_i = wins by i (+ 0.5 per draw), N_ij = total comparisons.\n *\n * Returns log-strengths normalized so the smallest is 0 (any constant\n * offset is unobservable in BT — only differences are identified).\n */\nexport function fitBradleyTerry(\n  outcomes: PairwiseOutcome[],\n  opts: { tolerance?: number; maxIterations?: number; smoothing?: number } = {},\n): BradleyTerryFit {\n  const tol = opts.tolerance ?? 1e-6\n  const maxIter = opts.maxIterations ?? 256\n  // Small positive default — Hunter's MM degenerates when a candidate has\n  // zero wins (θ → 0 → log → -∞). 0.1 is negligible against real win counts\n  // (~1 win / 10 comparisons) and keeps the iteration well-conditioned.\n  // Override to 0 if the comparison set is guaranteed strongly connected.\n  const smoothing = opts.smoothing ?? 0.1\n\n  const candidates = new Set<string>()\n  for (const o of outcomes) {\n    candidates.add(o.winner)\n    candidates.add(o.loser)\n  }\n  const ids = [...candidates].sort()\n  const idx = new Map(ids.map((id, i) => [id, i]))\n  const n = ids.length\n  if (n === 0) return { ratings: [], iterations: 0, finalDelta: 0, converged: true }\n  if (n === 1) {\n    return {\n      ratings: [{ candidateId: ids[0]!, strength: 1, logStrength: 0, n: 0, wins: 0 }],\n      iterations: 0,\n      finalDelta: 0,\n      converged: true,\n    }\n  }\n\n  // Build win matrix W[i][j] = (weighted) wins of i over j, plus half for draws.\n  // Build comparison matrix N[i][j] = total weighted comparisons between i and j.\n  const W: number[][] = Array.from({ length: n }, () => new Array<number>(n).fill(0))\n  const N: number[][] = Array.from({ length: n }, () => new Array<number>(n).fill(0))\n  for (const o of outcomes) {\n    const i = idx.get(o.winner)!\n    const j = idx.get(o.loser)!\n    const w = o.weight ?? 1\n    if (o.draw) {\n      W[i]![j]! += 0.5 * w\n      W[j]![i]! += 0.5 * w\n    } else {\n      W[i]![j]! += w\n    }\n    N[i]![j]! += w\n    N[j]![i]! += w\n  }\n\n  // Per-candidate total wins.\n  const winsTotal = new Array<number>(n).fill(0)\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < n; j++) winsTotal[i]! += W[i]![j]!\n    winsTotal[i]! += smoothing // tiny smoothing to keep θ positive\n  }\n  const compsTotal = new Array<number>(n).fill(0)\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < n; j++) compsTotal[i]! += N[i]![j]!\n  }\n\n  // MM iterations.\n  let theta = new Array<number>(n).fill(1)\n  let iter = 0\n  let delta = Infinity\n  for (; iter < maxIter; iter++) {\n    const newTheta = new Array<number>(n)\n    for (let i = 0; i < n; i++) {\n      let denom = 0\n      for (let j = 0; j < n; j++) {\n        if (j === i) continue\n        if (N[i]![j]! === 0) continue\n        denom += N[i]![j]! / (theta[i]! + theta[j]!)\n      }\n      newTheta[i] = denom === 0 ? theta[i]! : winsTotal[i]! / denom\n    }\n    // Normalize so geometric mean = 1 (numerical stability).\n    let logSum = 0\n    for (let i = 0; i < n; i++) logSum += Math.log(Math.max(1e-300, newTheta[i]!))\n    const norm = Math.exp(logSum / n)\n    for (let i = 0; i < n; i++) newTheta[i] = newTheta[i]! / norm\n\n    delta = 0\n    for (let i = 0; i < n; i++) {\n      const d = Math.abs(newTheta[i]! - theta[i]!) / Math.max(1e-12, theta[i]!)\n      if (d > delta) delta = d\n    }\n    theta = newTheta\n    if (delta < tol) break\n  }\n\n  const minLog = Math.min(...theta.map((t) => Math.log(Math.max(1e-300, t))))\n  const ratings: BradleyTerryRating[] = ids.map((id, i) => ({\n    candidateId: id,\n    strength: theta[i]!,\n    logStrength: Math.log(Math.max(1e-300, theta[i]!)) - minLog,\n    n: compsTotal[i]!,\n    wins: winsTotal[i]! - smoothing,\n  }))\n\n  return {\n    ratings: ratings.sort((a, b) => b.strength - a.strength),\n    iterations: iter,\n    finalDelta: delta,\n    converged: delta < tol,\n  }\n}\n\n/**\n * Online Elo updates. Use when comparisons arrive over time and you want\n * a running rating without re-fitting the full BT MLE on every update.\n *\n * Initialize ratings to `defaultRating` (1500 by default). Each call to\n * `applyEloUpdate` mutates the map in place and returns the deltas so\n * the caller can log per-comparison rating changes.\n */\nexport interface EloOptions {\n  /** Default rating for unseen candidates. Default 1500. */\n  defaultRating?: number\n  /** K-factor controls the step size. Default 32 (FIDE-ish). */\n  kFactor?: number\n}\n\nexport function applyEloUpdate(\n  ratings: Map<string, number>,\n  outcome: PairwiseOutcome,\n  opts: EloOptions = {},\n): { winnerDelta: number; loserDelta: number } {\n  const defaultRating = opts.defaultRating ?? 1500\n  const k = opts.kFactor ?? 32\n\n  const rW = ratings.get(outcome.winner) ?? defaultRating\n  const rL = ratings.get(outcome.loser) ?? defaultRating\n\n  const expectedW = 1 / (1 + 10 ** ((rL - rW) / 400))\n  const scoreW = outcome.draw ? 0.5 : 1\n  const scoreL = outcome.draw ? 0.5 : 0\n  const w = outcome.weight ?? 1\n\n  const winnerDelta = k * w * (scoreW - expectedW)\n  const loserDelta = k * w * (scoreL - (1 - expectedW))\n\n  ratings.set(outcome.winner, rW + winnerDelta)\n  ratings.set(outcome.loser, rL + loserDelta)\n\n  return { winnerDelta, loserDelta }\n}\n\n/**\n * Build pairwise outcomes from the campaign artifact: for every scenario\n * shared by two candidates, the higher-scoring run wins. Useful when you\n * want a tournament view of an existing campaign without an additional\n * pairwise judge call.\n */\nexport interface BuildPairwiseFromCampaignInput {\n  runs: Array<{\n    candidateId: string\n    /** Stable identifier for the matching unit (typically scenarioId). */\n    matchKey: string\n    score: number\n  }>\n  /**\n   * Tied-score margin. Below this, the comparison is a draw. Default 0\n   * (no ties).\n   */\n  drawMargin?: number\n}\n\nexport function buildPairwiseFromCampaign(\n  input: BuildPairwiseFromCampaignInput,\n): PairwiseOutcome[] {\n  const drawMargin = input.drawMargin ?? 0\n  const byKey = new Map<string, Array<{ candidateId: string; score: number }>>()\n  for (const r of input.runs) {\n    const arr = byKey.get(r.matchKey) ?? []\n    arr.push({ candidateId: r.candidateId, score: r.score })\n    byKey.set(r.matchKey, arr)\n  }\n  const outcomes: PairwiseOutcome[] = []\n  for (const arr of byKey.values()) {\n    for (let i = 0; i < arr.length; i++) {\n      for (let j = i + 1; j < arr.length; j++) {\n        const a = arr[i]!\n        const b = arr[j]!\n        if (a.candidateId === b.candidateId) continue\n        const margin = Math.abs(a.score - b.score)\n        if (margin <= drawMargin) {\n          outcomes.push({ winner: a.candidateId, loser: b.candidateId, draw: true, weight: 1 })\n        } else {\n          const [winner, loser] = a.score > b.score ? [a, b] : [b, a]\n          outcomes.push({ winner: winner.candidateId, loser: loser.candidateId, weight: margin })\n        }\n      }\n    }\n  }\n  return outcomes\n}\n","/**\n * Adaptive curriculum / active scenario selection.\n *\n * Fixed scenario sets waste sample budget on cells the policy already\n * passes (no information left) and cells the policy never passes (no\n * gradient available either). Active learning over scenarios fixes this\n * by allocating the next sample budget to cells where the policy's\n * outcome is *uncertain* — those carry the most decision-relevant signal.\n *\n * This module ships two complementary strategies:\n *\n *   1. **Variance-based** — score each (variant, scenario) cell by the\n *      empirical variance of past observations. Allocate next-round budget\n *      proportional to variance. Standard active-learning-by-uncertainty\n *      heuristic; works well when the policy is non-deterministic and\n *      cells differ in observation noise.\n *\n *   2. **Bandit-based (Thompson sampling)** — model each (variant,\n *      scenario) cell as a Beta-Bernoulli arm; sample a posterior; pick\n *      cells whose posterior mean is closest to the per-scenario decision\n *      threshold. The right primitive when scenarios are\n *      \"pass/fail\" rather than continuous, and when promotion gates fire\n *      at a known threshold (e.g., 0.5).\n *\n * The output is a *next-round budget allocation* — a list of (variant,\n * scenario, count) triples. The consumer's matrix runner consumes the\n * allocation, runs those cells, feeds the new observations back. Loop.\n *\n * Out of scope (deliberate): scenario *generation* — that's the\n * adversarial primitive's job. This module allocates over an existing\n * scenario pool.\n */\n\nimport type { RunRecord } from '../run-record'\n\nexport interface CellObservation {\n  variantId: string\n  scenarioId: string\n  /** Observed score in [0, 1]. */\n  score: number\n  /** For Bernoulli arms — derive from the score with a threshold if needed. */\n  pass?: boolean\n}\n\nexport interface CurriculumAllocation {\n  variantId: string\n  scenarioId: string\n  /** How many additional reps to run on this cell. */\n  count: number\n  /** Strategy-specific reason for the allocation. */\n  reason: string\n}\n\nexport interface VarianceCurriculumOptions {\n  /** Total reps to allocate across all cells. */\n  budget: number\n  /**\n   * Smoothing prior on variance — keeps the allocator from concentrating\n   * on a cell with one observation just because its 1-sample variance is\n   * 0. Default 0.05.\n   */\n  variancePrior?: number\n  /**\n   * Minimum reps per cell — even when the variance estimate is low, give\n   * every cell at least this many. Default 1.\n   */\n  floorPerCell?: number\n}\n\n/**\n * Variance-proportional allocation. For each cell, estimate variance from\n * past observations + a prior, then allocate the budget proportional to\n * (sqrt(variance) + 1/sqrt(n)) — a classical optimal-allocation rule\n * (Neyman 1934) that balances \"explore noisy cells\" with \"explore\n * under-sampled cells.\"\n */\nexport function varianceBasedCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: VarianceCurriculumOptions,\n): CurriculumAllocation[] {\n  const variancePrior = opts.variancePrior ?? 0.05\n  const floor = opts.floorPerCell ?? 1\n  const budget = opts.budget\n\n  const grouped = new Map<string, number[]>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const arr = grouped.get(k) ?? []\n    arr.push(o.score)\n    grouped.set(k, arr)\n  }\n\n  const cellStats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const samples = grouped.get(k) ?? []\n    const n = samples.length\n    const mean = n === 0 ? 0.5 : samples.reduce((s, v) => s + v, 0) / n\n    const variance =\n      n < 2\n        ? variancePrior\n        : samples.reduce((s, v) => s + (v - mean) ** 2, 0) / (n - 1) + variancePrior\n    // Neyman optimal allocation: weight ∝ √variance; add √(1/n) to break\n    // ties toward under-sampled cells.\n    const weight = Math.sqrt(variance) + 1 / Math.sqrt(Math.max(1, n))\n    return { variantId: c.variantId, scenarioId: c.scenarioId, n, mean, variance, weight }\n  })\n\n  // Reserve floor*N for the floor; allocate the rest proportional to weight.\n  const floorTotal = floor * cellStats.length\n  if (floorTotal >= budget) {\n    const each = Math.max(1, Math.floor(budget / Math.max(1, cellStats.length)))\n    return cellStats.map((c) => ({\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: each,\n      reason: `floor allocation (budget tight; n=${c.n})`,\n    }))\n  }\n  const remaining = budget - floorTotal\n  const totalWeight = cellStats.reduce((s, c) => s + c.weight, 0)\n  return cellStats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * remaining)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: floor + proportional,\n      reason: `variance ${c.variance.toFixed(3)} (n=${c.n}, mean=${c.mean.toFixed(3)})`,\n    }\n  })\n}\n\nexport interface ThompsonCurriculumOptions {\n  budget: number\n  /**\n   * The per-scenario decision threshold. Cells whose posterior mean is\n   * closest to this get the most budget — that's where the next observation\n   * has the highest information value for the gate decision. Default 0.5.\n   */\n  decisionThreshold?: number\n  /** Beta prior parameters. Default α=β=1 (uniform). */\n  priorAlpha?: number\n  priorBeta?: number\n  /** Seed the Thompson sampler. Default unset (Math.random). */\n  seed?: number\n}\n\n/**\n * Thompson-sampling-style allocation for pass/fail cells. For each cell:\n *\n *   - Maintain Beta(α + passes, β + failures) posterior on pass-rate\n *   - Allocation weight ∝ exp(-((sampledMean - threshold) / σ)^2):\n *     cells whose sampled posterior straddles the decision boundary get\n *     the most weight; cells already clearly above or below get less.\n *\n * This is the right primitive when promotion gates fire at a known\n * threshold and you want to sharpen the posterior near the boundary.\n */\nexport function thompsonCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: ThompsonCurriculumOptions,\n): CurriculumAllocation[] {\n  const threshold = opts.decisionThreshold ?? 0.5\n  const alpha0 = opts.priorAlpha ?? 1\n  const beta0 = opts.priorBeta ?? 1\n  const rng = makeRng(opts.seed)\n\n  const grouped = new Map<string, { passes: number; failures: number }>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const pass = o.pass ?? o.score >= threshold\n    if (pass) cur.passes += 1\n    else cur.failures += 1\n    grouped.set(k, cur)\n  }\n\n  const stats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const a = alpha0 + cur.passes\n    const b = beta0 + cur.failures\n    // Sample a single Beta draw — the Thompson signal.\n    const sampled = sampleBeta(a, b, rng)\n    const distance = Math.abs(sampled - threshold)\n    // Information-near-threshold weight: closer = higher.\n    // Use Gaussian-shaped kernel with σ tuned to posterior std.\n    const variance = (a * b) / ((a + b) ** 2 * (a + b + 1))\n    const sigma = Math.max(0.05, Math.sqrt(variance))\n    const weight = Math.exp(-((distance / sigma) ** 2))\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      n: cur.passes + cur.failures,\n      sampled,\n      sigma,\n      weight,\n      a,\n      b,\n    }\n  })\n\n  const totalWeight = stats.reduce((s, c) => s + c.weight, 0)\n  return stats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * opts.budget)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: Math.max(0, proportional),\n      reason: `Beta(${c.a.toFixed(1)},${c.b.toFixed(1)}) sample=${c.sampled.toFixed(3)} (target ${threshold})`,\n    }\n  })\n}\n\n/** Convenience: extract `CellObservation[]` directly from `RunRecord[]`. */\nexport function observationsFromRunRecords(\n  runs: RunRecord[],\n  opts: { passThreshold?: number; useHoldout?: boolean } = {},\n): CellObservation[] {\n  const threshold = opts.passThreshold ?? 0.5\n  const useHoldout = opts.useHoldout ?? true\n  const out: CellObservation[] = []\n  for (const r of runs) {\n    if (!r.scenarioId) continue\n    const score = useHoldout\n      ? (r.outcome.holdoutScore ?? r.outcome.searchScore)\n      : (r.outcome.searchScore ?? r.outcome.holdoutScore)\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    out.push({\n      variantId: r.candidateId,\n      scenarioId: r.scenarioId,\n      score,\n      pass: score >= threshold,\n    })\n  }\n  return out\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/**\n * Sample from Beta(α, β) via the Marsaglia–Tsang method using two Gamma\n * variates. Accuracy is good for α, β > 1; we floor the parameters at 1\n * to avoid degenerate cases.\n */\nfunction sampleBeta(alpha: number, beta: number, rng: () => number): number {\n  const a = Math.max(1, alpha)\n  const b = Math.max(1, beta)\n  const x = sampleGamma(a, rng)\n  const y = sampleGamma(b, rng)\n  return x / (x + y)\n}\n\nfunction sampleGamma(shape: number, rng: () => number): number {\n  // Marsaglia–Tsang for shape ≥ 1.\n  const d = shape - 1 / 3\n  const c = 1 / Math.sqrt(9 * d)\n  while (true) {\n    let x: number\n    let v: number\n    do {\n      const u1 = rng() || 1e-12\n      const u2 = rng() || 1e-12\n      // Box-Muller for a normal sample.\n      x = Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2)\n      v = 1 + c * x\n    } while (v <= 0)\n    v = v * v * v\n    const u = rng()\n    if (u < 1 - 0.0331 * x ** 4) return d * v\n    if (Math.log(u) < 0.5 * x * x + d * (1 - v + Math.log(v))) return d * v\n  }\n}\n","/**\n * Sample-efficient adaptation evaluation.\n *\n * For foundation-model-based agents, the load-bearing capability isn't\n * raw end-state performance — it's *how fast the agent reaches that\n * performance from cold start*. The same model with a worse prompt that\n * adapts in 5 demonstrations beats the same model with a better prompt\n * that needs 50. Standard meta-learning eval (Finn et al., MAML, RL² lit)\n * reports an *adaptation curve*: score after k=0, 1, 2, 4, 8, 16, …\n * in-context examples or fine-tune steps.\n *\n * This module ships:\n *\n *   1. `runAdaptationCurve` — given a runner that takes k demonstrations\n *      and returns a score, produce the (k, score) curve.\n *   2. `compareAdaptationCurves` — paired comparison across two policies.\n *      Returns per-k delta with bootstrap CIs and an \"area-under-curve\"\n *      summary statistic.\n *   3. `firstPassK` — for pass/fail evaluation, the minimum k at which\n *      the policy reliably passes (≥ pass-rate threshold over reps).\n *\n * Use cases:\n *   - Compare two prompt designs that have similar end-state performance\n *     but different in-context efficiency.\n *   - Decide between fine-tuning and prompting based on adaptation cost.\n *   - Detect when a policy \"memorizes\" k=0 inputs vs. genuinely adapts.\n */\n\nexport interface AdaptationRunner<S> {\n  /**\n   * Runs the policy on `scenario` with `k` demonstrations. Returns a\n   * scalar score in [0, 1]. The runner is responsible for any caching;\n   * the harness calls it once per (scenario, k, rep) cell.\n   */\n  run(args: { scenario: S; k: number; rep: number }): Promise<number>\n}\n\nexport interface RunAdaptationCurveOptions<S> {\n  scenarios: S[]\n  /** Number-of-shots to evaluate at. Default `[0, 1, 2, 4, 8, 16]`. */\n  ks?: number[]\n  /** Reps per (scenario, k) cell. Default 3. */\n  reps?: number\n  runner: AdaptationRunner<S>\n  /** Pass-rate threshold for `firstPassK` reporting. Default 0.5. */\n  passThreshold?: number\n}\n\nexport interface AdaptationPoint {\n  k: number\n  meanScore: number\n  passRate: number\n  std: number\n  n: number\n  /** Per-scenario means at this k. */\n  perScenario: Array<{ scenarioId: string; meanScore: number; passes: number; total: number }>\n}\n\nexport interface AdaptationCurve {\n  points: AdaptationPoint[]\n  /**\n   * Smallest `k` at which `passRate ≥ passThreshold`. `null` if no `k`\n   * tested reaches it.\n   */\n  firstPassK: number | null\n  /**\n   * Area under the (k, meanScore) curve, normalized by max-k. A\n   * single-number summary of \"how well does this policy adapt from\n   * cold-start to fully-conditioned.\" Higher = better adapter.\n   */\n  adaptationArea: number\n}\n\nexport async function runAdaptationCurve<S extends { scenarioId?: string }>(\n  opts: RunAdaptationCurveOptions<S>,\n): Promise<AdaptationCurve> {\n  const ks = opts.ks ?? [0, 1, 2, 4, 8, 16]\n  const reps = opts.reps ?? 3\n  const passThreshold = opts.passThreshold ?? 0.5\n  const sortedKs = [...ks].sort((a, b) => a - b)\n\n  const points: AdaptationPoint[] = []\n  for (const k of sortedKs) {\n    const perScenario: AdaptationPoint['perScenario'] = []\n    const allScores: number[] = []\n    let totalPasses = 0\n    let totalAttempts = 0\n    for (const scenario of opts.scenarios) {\n      const sid = scenario.scenarioId ?? `scenario-${opts.scenarios.indexOf(scenario)}`\n      const scores: number[] = []\n      let passes = 0\n      for (let r = 0; r < reps; r++) {\n        const score = await opts.runner.run({ scenario, k, rep: r })\n        scores.push(score)\n        if (score >= passThreshold) passes++\n        allScores.push(score)\n        if (score >= passThreshold) totalPasses++\n        totalAttempts++\n      }\n      const meanS = scores.reduce((s, v) => s + v, 0) / scores.length\n      perScenario.push({ scenarioId: sid, meanScore: meanS, passes, total: scores.length })\n    }\n    const meanScore = allScores.reduce((s, v) => s + v, 0) / Math.max(1, allScores.length)\n    const variance =\n      allScores.length < 2\n        ? 0\n        : allScores.reduce((s, v) => s + (v - meanScore) ** 2, 0) / (allScores.length - 1)\n    points.push({\n      k,\n      meanScore,\n      passRate: totalPasses / Math.max(1, totalAttempts),\n      std: Math.sqrt(variance),\n      n: allScores.length,\n      perScenario,\n    })\n  }\n\n  const firstPassK = points.find((p) => p.passRate >= passThreshold)?.k ?? null\n  const maxK = sortedKs[sortedKs.length - 1] ?? 1\n  // Trapezoidal area under the (k, meanScore) curve, normalized by k-range.\n  let area = 0\n  for (let i = 1; i < points.length; i++) {\n    const x1 = points[i - 1]!.k\n    const x2 = points[i]!.k\n    const y1 = points[i - 1]!.meanScore\n    const y2 = points[i]!.meanScore\n    area += ((y1 + y2) / 2) * (x2 - x1)\n  }\n  const adaptationArea = maxK === 0 ? 0 : area / maxK\n\n  return { points, firstPassK, adaptationArea }\n}\n\nexport interface CompareCurvesResult {\n  perK: Array<{\n    k: number\n    deltaMean: number\n    aLow: number\n    aHigh: number\n    bLow: number\n    bHigh: number\n  }>\n  areaDelta: number\n  firstPassKDelta: number | null\n  /** Verdict: 'a_better' | 'b_better' | 'similar'. */\n  verdict: 'a_better' | 'b_better' | 'similar'\n  /** Rationale, ready to render. */\n  rationale: string\n}\n\n/**\n * Paired comparison of two adaptation curves. Per-k deltas with 95%\n * bootstrap CIs (constructed from each curve's `perScenario` per-k means\n * — the bootstrap unit is the scenario, not the rep).\n */\nexport function compareAdaptationCurves(\n  a: AdaptationCurve,\n  b: AdaptationCurve,\n  opts: { confidence?: number; bootstrapResamples?: number; seed?: number } = {},\n): CompareCurvesResult {\n  const conf = opts.confidence ?? 0.95\n  const resamples = opts.bootstrapResamples ?? 500\n  const rng = makeRng(opts.seed)\n\n  const perK: CompareCurvesResult['perK'] = []\n  for (const ap of a.points) {\n    const bp = b.points.find((p) => p.k === ap.k)\n    if (!bp) continue\n    const aMeans = ap.perScenario.map((s) => s.meanScore)\n    const bMeans = bp.perScenario.map((s) => s.meanScore)\n    const aCi = bootstrapMeanCi(aMeans, resamples, conf, rng)\n    const bCi = bootstrapMeanCi(bMeans, resamples, conf, rng)\n    perK.push({\n      k: ap.k,\n      deltaMean: ap.meanScore - bp.meanScore,\n      aLow: aCi.low,\n      aHigh: aCi.high,\n      bLow: bCi.low,\n      bHigh: bCi.high,\n    })\n  }\n\n  const areaDelta = a.adaptationArea - b.adaptationArea\n  const firstPassKDelta =\n    a.firstPassK !== null && b.firstPassK !== null\n      ? b.firstPassK - a.firstPassK // smaller k for a means a adapts faster (positive delta)\n      : null\n\n  // Composite verdict: positive area delta + most per-k deltas in same\n  // direction → that side wins. Within ε of zero on both → similar.\n  const meanDelta = perK.reduce((s, p) => s + p.deltaMean, 0) / Math.max(1, perK.length)\n  let verdict: CompareCurvesResult['verdict']\n  if (Math.abs(meanDelta) < 0.02 && Math.abs(areaDelta) < 0.02) verdict = 'similar'\n  else if (meanDelta > 0 && areaDelta > 0) verdict = 'a_better'\n  else if (meanDelta < 0 && areaDelta < 0) verdict = 'b_better'\n  else verdict = 'similar'\n\n  const rationale =\n    `mean per-k delta=${meanDelta.toFixed(3)}, area delta=${areaDelta.toFixed(3)}` +\n    (firstPassKDelta !== null ? `, first-pass-k delta=${firstPassKDelta}` : '')\n\n  return { perK, areaDelta, firstPassKDelta, verdict, rationale }\n}\n\n/** First k at which the curve's per-scenario pass rate reliably hits the threshold. */\nexport function firstPassK(curve: AdaptationCurve, threshold = 0.5): number | null {\n  return curve.points.find((p) => p.passRate >= threshold)?.k ?? null\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\nfunction bootstrapMeanCi(\n  xs: number[],\n  resamples: number,\n  confidence: number,\n  rng: () => number,\n): { low: number; high: number } {\n  if (xs.length < 2) return { low: xs[0] ?? 0, high: xs[0] ?? 0 }\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let sum = 0\n    for (let i = 0; i < xs.length; i++) sum += xs[Math.floor(rng() * xs.length)]!\n    samples[b] = sum / xs.length\n  }\n  samples.sort((a, b) => a - b)\n  const alpha = 1 - confidence\n  return {\n    low: samples[Math.floor((alpha / 2) * resamples)]!,\n    high: samples[Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)]!,\n  }\n}\n","/**\n * Adversarial scenario search.\n *\n * Capability evaluation on a fixed scenario set measures performance on\n * the distribution someone curated. Production failure modes live in the\n * tail — inputs the curator didn't think of, or actively avoided. The\n * adversarial-search primitive actively looks for them: starting from a\n * pool of scenarios where the policy already passes, it mutates them\n * (paraphrase, edge-case substitution, compositional combination) and\n * keeps the mutations that *break* the policy.\n *\n * This is not magic. It's the simplest version of the loop that AdA\n * (Open-Ended Adaptation, DeepMind 2023), POET, and Anthropic's\n * auto-jailbreak rigs all run: hill-climb against a failure indicator,\n * keep the survivors, repeat. We ship the harness; consumers supply the\n * mutation strategies and the failure detector.\n *\n * Why ship this in agent-eval and not as a separate red-team tool: every\n * piece of the standard adversarial loop is already in this package\n * (`runEvalCampaign` for the matrix run, `RawProviderSink` for capture,\n * `assertRunCaptured` for integrity, `pairedEvalueSequence` for stop\n * criteria). The adversarial primitive is just the *scenario-mutation\n * meta-loop* on top of that machinery.\n */\n\nexport interface AdversarialScenario<S> {\n  /** Stable id — used for deduplication and lineage tracking. */\n  id: string\n  /** Generation index — 0 for seeds, 1 for first round of mutations, etc. */\n  generation: number\n  /** Lineage — id of the parent scenario this was mutated from, if any. */\n  parentId: string | null\n  scenario: S\n  /** Score on the policy under test. Lower = adversarial signal. */\n  score: number | null\n  /** Strategy that produced this mutation, for diagnostics. */\n  mutationStrategy: string | null\n}\n\nexport interface AdversarialMutation<S> {\n  id: string\n  /**\n   * Mutate one scenario. Return null to skip; return one or more new\n   * scenarios. The harness deduplicates by `mutateScenarioId(scenario)`.\n   */\n  mutate(parent: S, rng: () => number): Promise<S[]> | S[]\n}\n\nexport interface AdversarialSearchOptions<S> {\n  /** Initial scenarios — typically those the policy currently passes. */\n  seeds: S[]\n  /** Stable identifier extraction. */\n  mutateScenarioId: (s: S) => string\n  /** Mutation strategies. */\n  mutations: AdversarialMutation<S>[]\n  /**\n   * Run the policy under test against one scenario, return a scalar score\n   * in [0, 1]. Lower = adversarial signal.\n   */\n  scoreFn: (s: S) => Promise<number>\n  /**\n   * Threshold below which a scenario counts as a \"failure\" worth keeping.\n   * Default 0.5.\n   */\n  failureThreshold?: number\n  /** Number of mutation rounds. Default 3. */\n  rounds?: number\n  /** Children per parent per round. Default 4. */\n  childrenPerParent?: number\n  /** Maximum total scenarios examined. Default Infinity. */\n  budget?: number\n  /** Seed for the deterministic RNG. Default 1. */\n  seed?: number\n}\n\nexport interface AdversarialSearchReport<S> {\n  scenarios: AdversarialScenario<S>[]\n  /** Discovered failures sorted by score ascending. */\n  failures: AdversarialScenario<S>[]\n  /** Round-by-round counts. */\n  byGeneration: Array<{ generation: number; total: number; failures: number; meanScore: number }>\n  /** Total scoreFn invocations consumed. */\n  scoreCalls: number\n}\n\nexport async function adversarialScenarioSearch<S>(\n  opts: AdversarialSearchOptions<S>,\n): Promise<AdversarialSearchReport<S>> {\n  const failureThreshold = opts.failureThreshold ?? 0.5\n  const rounds = opts.rounds ?? 3\n  const children = opts.childrenPerParent ?? 4\n  const budget = opts.budget ?? Number.POSITIVE_INFINITY\n  const seed = opts.seed ?? 1\n  const rng = mulberry32(seed)\n\n  const scenarios: AdversarialScenario<S>[] = []\n  const seen = new Set<string>()\n  let scoreCalls = 0\n\n  // Seed generation.\n  for (const s of opts.seeds) {\n    const id = opts.mutateScenarioId(s)\n    if (seen.has(id)) continue\n    seen.add(id)\n    if (scoreCalls >= budget) break\n    const score = await opts.scoreFn(s)\n    scoreCalls++\n    scenarios.push({\n      id,\n      generation: 0,\n      parentId: null,\n      scenario: s,\n      score,\n      mutationStrategy: null,\n    })\n  }\n\n  // Mutation rounds.\n  for (let g = 1; g <= rounds; g++) {\n    if (scoreCalls >= budget) break\n    const parents = scenarios.filter((s) => s.generation === g - 1)\n    for (const parent of parents) {\n      for (const mutation of opts.mutations) {\n        if (scoreCalls >= budget) break\n        const produced = await mutation.mutate(parent.scenario, rng)\n        const childArr = Array.isArray(produced) ? produced : [produced]\n        for (let k = 0; k < Math.min(children, childArr.length); k++) {\n          if (scoreCalls >= budget) break\n          const child = childArr[k]!\n          const cid = opts.mutateScenarioId(child)\n          if (seen.has(cid)) continue\n          seen.add(cid)\n          const cscore = await opts.scoreFn(child)\n          scoreCalls++\n          scenarios.push({\n            id: cid,\n            generation: g,\n            parentId: parent.id,\n            scenario: child,\n            score: cscore,\n            mutationStrategy: mutation.id,\n          })\n        }\n      }\n    }\n  }\n\n  const failures = scenarios\n    .filter((s) => s.score !== null && s.score < failureThreshold)\n    .sort((a, b) => (a.score ?? 0) - (b.score ?? 0))\n\n  const byGeneration: AdversarialSearchReport<S>['byGeneration'] = []\n  const maxGen = scenarios.reduce((m, s) => Math.max(m, s.generation), 0)\n  for (let g = 0; g <= maxGen; g++) {\n    const gens = scenarios.filter((s) => s.generation === g)\n    if (gens.length === 0) continue\n    const fails = gens.filter((s) => s.score !== null && s.score < failureThreshold).length\n    const meanScore = gens.reduce((sum, s) => sum + (s.score ?? 0), 0) / gens.length\n    byGeneration.push({ generation: g, total: gens.length, failures: fails, meanScore })\n  }\n\n  return { scenarios, failures, byGeneration, scoreCalls }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mulberry32(seed: number): () => number {\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n","/**\n * Trainer-format exporters.\n *\n * agent-eval produces canonical artifacts (`RunRecord[]`, `PreferenceTriple[]`,\n * `StepReward[]`, `PrmTrainingTriple[]`). RL training pipelines consume\n * different shapes — Hugging Face TRL, Prime Intellect's prime-rl, OpenAI\n * fine-tuning, Anthropic finetuning, OpenRLHF, verl. Each has its own\n * JSONL conventions. Rather than ship N adapters, this module ships the\n * canonical formats most production pipelines accept and ergonomic helpers\n * for the rest.\n *\n * Shapes:\n *   - **DPO / IPO / KTO** — `{prompt, chosen, rejected}` JSONL. Consumed\n *     by HuggingFace TRL, prime-rl's offline DPO, OpenRLHF.\n *   - **GRPO offline** — `{prompt, completions[], rewards[]}` JSONL.\n *     Consumed by prime-rl GRPO, verl, OpenRLHF.\n *   - **SFT** — `{messages[]}` JSONL with chosen completion as the final\n *     assistant turn. Consumed by HF SFT trainers, OpenAI fine-tuning,\n *     Anthropic finetuning.\n *   - **PRM** — `{prompt, prefix_steps[], chosen_step, rejected_step}` JSONL.\n *     Consumed by Lightman-style PRM trainers and prime-rl's PRM mode.\n *\n * Why ship this in agent-eval rather than a separate adapter package: the\n * canonical artifacts (`RunRecord[]`, `PreferenceTriple[]`, etc.) are\n * agent-eval's contract; without first-party exporters consumers reverse-\n * engineer the mapping every release. The exporters codify it.\n *\n * The exporters take callbacks for any field that isn't on the canonical\n * artifact (specifically: prompt + completion text, since the package\n * stores only their hashes by design — full text is the consumer's\n * trace store / raw event log).\n */\n\nimport type { RunRecord } from '../run-record'\nimport type { PreferenceTriple } from './preferences'\nimport type { PrmTrainingTriple, StepReward } from './process-reward'\n\n// ── DPO / IPO / KTO ──────────────────────────────────────────────────────\n\nexport interface DpoLookups {\n  /** Resolve the prompt text for a run (typically from a trace store / raw event sink). */\n  promptOf: (runId: string) => string | Promise<string>\n  /** Resolve the assistant completion text for a run. */\n  completionOf: (runId: string) => string | Promise<string>\n}\n\nexport interface DpoExportRow {\n  prompt: string\n  chosen: string\n  rejected: string\n  /** Carried-through margin. Some KTO / IPO variants use this. */\n  margin?: number\n  /** Free-form metadata for downstream filtering / sharding. */\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert preference triples to TRL-compatible DPO rows. The shape\n * `{prompt, chosen, rejected}` is the canonical HuggingFace DPODataset\n * entry; every major DPO trainer accepts it.\n */\nexport async function toDpoRows(\n  triples: PreferenceTriple[],\n  lookups: DpoLookups,\n): Promise<DpoExportRow[]> {\n  const out: DpoExportRow[] = []\n  for (const t of triples) {\n    const [prompt, chosen, rejected] = await Promise.all([\n      Promise.resolve(lookups.promptOf(t.chosenRunId)),\n      Promise.resolve(lookups.completionOf(t.chosenRunId)),\n      Promise.resolve(lookups.completionOf(t.rejectedRunId)),\n    ])\n    out.push({\n      prompt,\n      chosen,\n      rejected,\n      margin: t.marginScore,\n      meta: {\n        scenarioId: t.scenarioId,\n        chosenVariantId: t.chosenVariantId,\n        rejectedVariantId: t.rejectedVariantId,\n        chosenRunId: t.chosenRunId,\n        rejectedRunId: t.rejectedRunId,\n        chosenModel: t.meta.chosenModel,\n        rejectedModel: t.meta.rejectedModel,\n      },\n    })\n  }\n  return out\n}\n\n/** Serialize DPO rows as JSONL. One line per row. */\nexport function toDpoJsonl(rows: DpoExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── GRPO offline ─────────────────────────────────────────────────────────\n\nexport interface GrpoLookups {\n  promptOf: (runId: string) => string | Promise<string>\n  completionOf: (runId: string) => string | Promise<string>\n  /** Optional: derive a custom reward from the run. Defaults to score. */\n  rewardOf?: (run: RunRecord) => number | null\n}\n\nexport interface GrpoExportRow {\n  prompt: string\n  completions: string[]\n  rewards: number[]\n  /** runIds in the same order as `completions[]` for traceability. */\n  runIds: string[]\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert RunRecord[] grouped by `(scenarioId)` into GRPO offline rows —\n * one row per scenario, with one completion per run on that scenario.\n *\n * GRPO (Shao et al. 2024 / DeepSeek-R1) trains on relative advantages\n * within a group of completions for the same prompt; this is the\n * canonical input format.\n */\nexport async function toGrpoRows(\n  runs: RunRecord[],\n  lookups: GrpoLookups,\n): Promise<GrpoExportRow[]> {\n  const rewardOf = lookups.rewardOf ?? defaultReward\n  const grouped = new Map<string, RunRecord[]>()\n  for (const r of runs) {\n    const sid = r.scenarioId ?? r.experimentId\n    const arr = grouped.get(sid) ?? []\n    arr.push(r)\n    grouped.set(sid, arr)\n  }\n\n  const rows: GrpoExportRow[] = []\n  for (const [scenarioId, group] of grouped.entries()) {\n    if (group.length === 0) continue\n    // Resolve prompt once per group (assumes all runs in a group share the prompt).\n    const prompt = await Promise.resolve(lookups.promptOf(group[0]!.runId))\n    const completions: string[] = []\n    const rewards: number[] = []\n    const runIds: string[] = []\n    for (const r of group) {\n      const reward = rewardOf(r)\n      if (reward === null) continue\n      const completion = await Promise.resolve(lookups.completionOf(r.runId))\n      completions.push(completion)\n      rewards.push(reward)\n      runIds.push(r.runId)\n    }\n    if (completions.length === 0) continue\n    rows.push({\n      prompt,\n      completions,\n      rewards,\n      runIds,\n      meta: {\n        scenarioId,\n        n: completions.length,\n        meanReward: rewards.reduce((s, x) => s + x, 0) / rewards.length,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toGrpoJsonl(rows: GrpoExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── SFT ──────────────────────────────────────────────────────────────────\n\nexport interface SftLookups {\n  promptOf: (runId: string) => string | Promise<string>\n  completionOf: (runId: string) => string | Promise<string>\n  /** Optional system message. Default omits. */\n  systemOf?: (run: RunRecord) => string | null | undefined\n  /** Filter — return false to skip the run (e.g., low score, failed cases). */\n  include?: (run: RunRecord) => boolean\n}\n\nexport interface SftExportRow {\n  messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }>\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert RunRecord[] into Hugging Face / OpenAI / Anthropic-style\n * conversational SFT rows. By default every record becomes one row;\n * pass `include` to filter (e.g., keep only `score >= 0.8` for\n * rejection-sampling SFT).\n */\nexport async function toSftRows(runs: RunRecord[], lookups: SftLookups): Promise<SftExportRow[]> {\n  const include = lookups.include ?? (() => true)\n  const rows: SftExportRow[] = []\n  for (const r of runs) {\n    if (!include(r)) continue\n    const system = lookups.systemOf?.(r)\n    const [prompt, completion] = await Promise.all([\n      Promise.resolve(lookups.promptOf(r.runId)),\n      Promise.resolve(lookups.completionOf(r.runId)),\n    ])\n    const messages: SftExportRow['messages'] = []\n    if (system) messages.push({ role: 'system', content: system })\n    messages.push({ role: 'user', content: prompt })\n    messages.push({ role: 'assistant', content: completion })\n    rows.push({\n      messages,\n      meta: {\n        runId: r.runId,\n        candidateId: r.candidateId,\n        scenarioId: r.scenarioId,\n        score: r.outcome.holdoutScore ?? r.outcome.searchScore,\n        model: r.model,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toSftJsonl(rows: SftExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── PRM ──────────────────────────────────────────────────────────────────\n\nexport interface PrmLookups {\n  /** Resolve the prompt text for a run. */\n  promptOf: (runId: string) => string | Promise<string>\n  /** Resolve the trajectory step text for a (runId, spanId) pair. */\n  stepTextOf: (runId: string, spanId: string) => string | Promise<string>\n  /** Optional: sequence of prefix span ids leading up to the divergence. */\n  prefixOf?: (runId: string, prefixStepIndex: number) => string[] | Promise<string[]>\n}\n\nexport interface PrmExportRow {\n  prompt: string\n  /** Span ids for the steps before divergence — caller resolves text via `stepTextOf`. */\n  prefixSpanIds: string[]\n  prefixStepText: string[]\n  chosenStep: string\n  rejectedStep: string\n  chosenReward: number\n  rejectedReward: number\n  marginScore: number\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert PRM training triples to JSONL rows. Caller's `stepTextOf`\n * callback resolves span text from the consumer's trace store.\n */\nexport async function toPrmRows(\n  triples: PrmTrainingTriple[],\n  lookups: PrmLookups,\n): Promise<PrmExportRow[]> {\n  const rows: PrmExportRow[] = []\n  for (const t of triples) {\n    const prompt = await Promise.resolve(lookups.promptOf(t.prefixRunId))\n    const prefixSpanIds = lookups.prefixOf\n      ? await Promise.resolve(lookups.prefixOf(t.prefixRunId, t.prefixStepIndex))\n      : []\n    const prefixStepText: string[] = []\n    for (const spanId of prefixSpanIds) {\n      prefixStepText.push(await Promise.resolve(lookups.stepTextOf(t.prefixRunId, spanId)))\n    }\n    const chosenStep = await Promise.resolve(lookups.stepTextOf(t.prefixRunId, t.chosenSpanId))\n    const rejectedStep = await Promise.resolve(\n      lookups.stepTextOf(t.rejectedRunId, t.rejectedSpanId),\n    )\n    rows.push({\n      prompt,\n      prefixSpanIds,\n      prefixStepText,\n      chosenStep,\n      rejectedStep,\n      chosenReward: t.chosenReward,\n      rejectedReward: t.rejectedReward,\n      marginScore: t.marginScore,\n      meta: {\n        prefixRunId: t.prefixRunId,\n        rejectedRunId: t.rejectedRunId,\n        prefixStepIndex: t.prefixStepIndex,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toPrmJsonl(rows: PrmExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── Step rewards (for value-function regression) ─────────────────────────\n\nexport interface StepRewardJsonlRow {\n  runId: string\n  spanId: string\n  stepIndex: number\n  reward: number\n  determinism: 'deterministic' | 'probabilistic'\n  weight: number\n}\n\nexport function stepRewardsToJsonl(stepRewards: StepReward[]): string {\n  const rows: StepRewardJsonlRow[] = stepRewards.map((s) => ({\n    runId: s.runId,\n    spanId: s.spanId,\n    stepIndex: s.stepIndex,\n    reward: s.reward,\n    determinism: s.determinism,\n    weight: s.weight ?? 1,\n  }))\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction defaultReward(run: RunRecord): number | null {\n  const v = run.outcome.holdoutScore ?? run.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n","/**\n * `PredictiveValidityResearcher` — concrete `Researcher` implementation\n * that drives selection from outcome-anchored predictive validity.\n *\n * Each method:\n *\n *   - `inspectFailures(runs)` — synthesizes failure modes from the\n *     bottom-quartile of `RunRecord`s on the configured proxy reward.\n *   - `proposeChange(failures)` — proposes steering changes that target\n *     the rubrics with the lowest predictive validity (decorative ones).\n *     Either reduce their weight in the composite, or recalibrate them.\n *   - `applyChange(changes, baseline)` — merges the proposed steering\n *     into the experiment plan.\n *   - `evaluateChange(plan)` — re-runs the predictive-validity check on\n *     the post-change runs and reports the delta.\n *\n * The result is a closed loop: the rubric weights drift toward the ones\n * that actually predict deployment outcomes, automatically. Pair with\n * `runRLCampaign` for the full auto-research story.\n */\n\nimport type { GateDecision } from '../held-out-gate'\nimport type { OutcomeStore } from '../meta-eval/outcome-store'\nimport {\n  type RubricPredictiveValidityReport,\n  rubricPredictiveValidity,\n} from '../meta-eval/rubric-predictive-validity'\nimport type {\n  ExperimentPlan,\n  ExperimentResult,\n  FailureMode,\n  Researcher,\n  SteeringChange,\n} from '../researcher'\nimport type { RunRecord } from '../run-record'\n\nexport interface PredictiveValidityResearcherOptions {\n  outcomes: OutcomeStore\n  outcomeMetrics: string[]\n  /** Score threshold below which a run counts as a \"failure.\" Default 0.5. */\n  failureThreshold?: number\n  /** Spearman bucket below which a rubric is \"decorative.\" Default 0.4. */\n  decorativeThreshold?: number\n  /** Optional steering-namespace prefix for proposed changes. Default `'rubric_weight'`. */\n  steeringNamespace?: string\n  /** Override the rubric set the researcher inspects. Default: every numeric `outcome.raw` key seen. */\n  rubrics?: string[]\n  /**\n   * Snapshot stash hook — called with the most recent predictive-validity\n   * report. Useful when a downstream system wants to log rubric drift over\n   * time. Default no-op.\n   */\n  onReport?: (report: RubricPredictiveValidityReport) => void | Promise<void>\n}\n\n/**\n * Concrete `Researcher` driven by `rubricPredictiveValidity`. The brain:\n * rubrics that don't predict deployment outcomes don't earn weight.\n */\nexport class PredictiveValidityResearcher implements Researcher {\n  private opts: PredictiveValidityResearcherOptions\n  private lastReport: RubricPredictiveValidityReport | null = null\n\n  constructor(opts: PredictiveValidityResearcherOptions) {\n    this.opts = opts\n  }\n\n  async inspectFailures(runs: RunRecord[]): Promise<FailureMode[]> {\n    const threshold = this.opts.failureThreshold ?? 0.5\n    const failures: FailureMode[] = []\n    const failingRuns = runs.filter((r) => {\n      const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n      return typeof score === 'number' && score < threshold\n    })\n    if (failingRuns.length === 0) return failures\n\n    // Group failures by candidateId — the researcher's primary handle is\n    // \"this candidate is producing low-scoring outputs in this scenario.\"\n    const grouped = new Map<string, RunRecord[]>()\n    for (const r of failingRuns) {\n      const arr = grouped.get(r.candidateId) ?? []\n      arr.push(r)\n      grouped.set(r.candidateId, arr)\n    }\n\n    for (const [candidateId, group] of grouped.entries()) {\n      const meanScore =\n        group.reduce((s, r) => {\n          const x = r.outcome.holdoutScore ?? r.outcome.searchScore ?? 0\n          return s + x\n        }, 0) / group.length\n      failures.push({\n        code: `low-score-${candidateId}`,\n        description: `${candidateId} scored < ${threshold} on ${group.length} run(s) (mean ${meanScore.toFixed(3)})`,\n        evidence: {\n          runIds: group.slice(0, 8).map((r) => r.runId),\n          samples: group.length,\n        },\n      })\n    }\n    return failures\n  }\n\n  async proposeChange(failures: FailureMode[]): Promise<SteeringChange[]> {\n    if (failures.length === 0) return []\n\n    // Without a prior report, return a single \"collect more outcome data\"\n    // change — the researcher refuses to reweight rubrics from zero evidence.\n    if (this.lastReport === null) {\n      return [\n        {\n          kind: 'threshold',\n          payload: { directive: 'researcher.collect-more-outcomes' },\n          rationale:\n            'predictive-validity researcher has no prior report; cannot recommend rubric reweighting until at least one report exists',\n        },\n      ]\n    }\n\n    const decorativeThreshold = this.opts.decorativeThreshold ?? 0.4\n    const changes: SteeringChange[] = []\n\n    for (const ranking of this.lastReport.ranked) {\n      if (ranking.verdict === 'load_bearing') continue\n      if (Math.abs(ranking.spearman) >= decorativeThreshold) continue\n      changes.push({\n        kind: 'reviewer_prompt',\n        payload: {\n          rubric: ranking.rubric,\n          action: 'down-weight',\n          spearman: ranking.spearman,\n          bestOutcome: ranking.bestOutcome,\n        },\n        rationale: `predictive-validity Spearman=${ranking.spearman.toFixed(3)} vs ${ranking.bestOutcome} (decorative); recommend down-weighting`,\n        expectedDelta: -Math.max(0, 0.05 - Math.abs(ranking.spearman)),\n      })\n    }\n    for (const ranking of this.lastReport.ranked.slice(0, 1)) {\n      if (ranking.verdict !== 'load_bearing') continue\n      changes.push({\n        kind: 'reviewer_prompt',\n        payload: {\n          rubric: ranking.rubric,\n          action: 'up-weight',\n          spearman: ranking.spearman,\n          bestOutcome: ranking.bestOutcome,\n        },\n        rationale: `predictive-validity Spearman=${ranking.spearman.toFixed(3)} vs ${ranking.bestOutcome} (load-bearing); recommend up-weighting`,\n        expectedDelta: Math.max(0, Math.abs(ranking.spearman) - 0.5) * 0.1,\n      })\n    }\n    return changes\n  }\n\n  async applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan> {\n    // Merge proposed changes into the plan's `changes` array, preserving\n    // any changes the baseline already had.\n    return {\n      ...baseline,\n      changes: [...baseline.changes, ...changes],\n    }\n  }\n\n  async evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult> {\n    // The researcher contract takes a *plan* and returns a *result* —\n    // implementations that only understand re-scoring runs can produce a\n    // \"no-op\" gate decision and let the caller drive the actual sweep.\n    // Real evaluators (CallbackResearcher) execute the plan; we report.\n    const emptyGate: GateDecision = {\n      promote: false,\n      candidateId: plan.proposedCandidateId,\n      baselineId: plan.baselineCandidateId,\n      evidence: {\n        productiveRuns: 0,\n        medianPairedDelta: 0,\n        pairedCI: { low: 0, high: 0 },\n        pairedPValue: 1,\n        searchScore: 0,\n        holdoutScore: 0,\n        overfitGap: 0,\n        baselineOverfitGap: 0,\n        medianCandidateCost: Number.NaN,\n        medianBaselineCost: Number.NaN,\n      },\n      reason:\n        'predictive-validity researcher does not execute plans; the caller is expected to run the sweep and call rubricPredictiveValidity directly with the resulting RunRecord[].',\n      rejectionCode: 'few_runs',\n    }\n    return {\n      plan,\n      runs: [],\n      gateDecision: emptyGate,\n    }\n  }\n\n  /**\n   * Run the predictive-validity check explicitly against a fresh RunRecord\n   * set. Updates the researcher's cached report so subsequent\n   * `proposeChange` calls have evidence to draw from.\n   */\n  async runValidityCheck(runs: RunRecord[]): Promise<RubricPredictiveValidityReport> {\n    const report = await rubricPredictiveValidity({\n      runs,\n      outcomes: this.opts.outcomes,\n      outcomeMetrics: this.opts.outcomeMetrics,\n      rubrics: this.opts.rubrics,\n    })\n    if (this.opts.onReport) await this.opts.onReport(report)\n    this.lastReport = report\n    return report\n  }\n\n  /**\n   * Force-feed a predictive-validity report into the researcher state —\n   * useful when the consumer ran the report out-of-band and wants the\n   * researcher's later proposals informed by it.\n   */\n  setReport(report: RubricPredictiveValidityReport): void {\n    this.lastReport = report\n  }\n\n  getLastReport(): RubricPredictiveValidityReport | null {\n    return this.lastReport\n  }\n}\n","/**\n * Process reward extraction — step-level credit assignment from trace spans.\n *\n * RL on long-horizon agents needs *step-level* rewards, not run-level\n * ones. The classic credit-assignment problem (Sutton & Barto) requires\n * knowing which sub-decisions in a trajectory contributed to the\n * outcome. Modern systems (DeepSeek-R1, OpenAI o-series, Lightman et al.\n * \"Let's Verify Step by Step\" 2023) train *process reward models* (PRMs)\n * that score every step, then do RL with the PRM as the reward signal.\n *\n * This module extracts `StepReward[]` from trace spans — one per\n * meaningful step — and ships:\n *\n *   1. `extractStepRewards(store, runId, opts)` — span → step-reward\n *      conversion using configurable per-span scorers (LLM judge over the\n *      span output, deterministic checkers, or a learned PRM).\n *   2. `runwiseStepRewardSummary(stepRewards)` — aggregate the per-step\n *      signal into a credit-assignment-aware run-level score.\n *   3. `prmTrainingPairs(stepRewards, options)` — produce the\n *      `(prefix, suffix_chosen, suffix_rejected)` triples that PRM\n *      training pipelines consume.\n *\n * What we ship: the *extraction* and *aggregation* infrastructure plus\n * the data shape PRM training expects. We do NOT ship the actual PRM\n * training (gradient descent over a transformer is out of scope for a\n * TS package). The interface is the contract; downstream consumers wire\n * their preferred trainer.\n *\n * Caveat the panel will land: this is descriptive credit assignment\n * (which steps correlate with outcome), not causal credit assignment\n * (which steps caused outcome). For causal claims you need\n * counterfactual rollouts or a learned dynamics model. Future work; the\n * descriptive version is what production PRM training actually uses.\n */\n\nimport type { Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface StepReward {\n  /** Trace span this reward attaches to. */\n  spanId: string\n  runId: string\n  /** Index in the trajectory (0-based, in started-at order). */\n  stepIndex: number\n  /** Span kind (typically 'tool', 'llm', 'judge'). */\n  kind: Span['kind']\n  /** Span name — for the consumer's downstream filtering. */\n  name: string\n  /** Step-level reward in [0, 1]. */\n  reward: number\n  /**\n   * Determinism class. Mirrors the verifiable-reward distinction:\n   * deterministic = test/compile/schema check; probabilistic = LLM judge.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /** Optional rationale / evidence — the trainer typically discards. */\n  rationale?: string\n  /** Optional weight — how much this step contributes to credit assignment. */\n  weight?: number\n}\n\nexport interface StepScorer {\n  /** Span kinds this scorer applies to. */\n  appliesTo: Span['kind'][]\n  /** Returns null to skip the span; returns a `StepReward` shape (without index/runId/spanId, which are filled in). */\n  score(span: Span): Promise<Omit<StepReward, 'spanId' | 'runId' | 'stepIndex'>> | null | undefined\n}\n\nexport interface ExtractStepRewardsOptions {\n  /**\n   * Ordered list of scorers. Each span runs through scorers in order;\n   * the first non-null result wins. If no scorer applies, the span is\n   * skipped (not all spans are training-worthy).\n   */\n  scorers: StepScorer[]\n  /** Optional filter — return null to drop the span entirely before scoring. */\n  preFilter?: (span: Span) => boolean\n}\n\nexport async function extractStepRewards(\n  store: TraceStore,\n  runId: string,\n  opts: ExtractStepRewardsOptions,\n): Promise<StepReward[]> {\n  const spans = await store.spans({ runId })\n  const ordered = [...spans].sort((a, b) => a.startedAt - b.startedAt)\n  const out: StepReward[] = []\n  let idx = 0\n  for (const span of ordered) {\n    if (opts.preFilter && !opts.preFilter(span)) continue\n    let scored: Awaited<ReturnType<StepScorer['score']>> = null\n    for (const s of opts.scorers) {\n      if (!s.appliesTo.includes(span.kind)) continue\n      const r = await s.score(span)\n      if (r) {\n        scored = r\n        break\n      }\n    }\n    if (!scored) continue\n    out.push({\n      spanId: span.spanId,\n      runId,\n      stepIndex: idx++,\n      kind: span.kind,\n      name: span.name,\n      reward: scored.reward,\n      determinism: scored.determinism,\n      rationale: scored.rationale,\n      weight: scored.weight,\n    })\n  }\n  return out\n}\n\nexport interface RunwiseStepSummary {\n  runId: string\n  totalSteps: number\n  meanReward: number\n  /** Sum-of-rewards (weighted by `weight ?? 1`). Use as the run-level proxy. */\n  sumWeightedReward: number\n  /** Fraction of steps where reward < 0.5 — proxy for \"where the policy was wrong.\" */\n  failureFraction: number\n  /** Maximum drop in reward between consecutive steps — diagnoses a step where things went sideways. */\n  worstStepDelta: number\n  worstStepIndex: number | null\n}\n\nexport function runwiseStepRewardSummary(stepRewards: StepReward[]): RunwiseStepSummary {\n  if (stepRewards.length === 0) {\n    return {\n      runId: '',\n      totalSteps: 0,\n      meanReward: 0,\n      sumWeightedReward: 0,\n      failureFraction: 0,\n      worstStepDelta: 0,\n      worstStepIndex: null,\n    }\n  }\n  const runId = stepRewards[0]!.runId\n  let sumW = 0\n  let sumWR = 0\n  let failures = 0\n  let worstDelta = 0\n  let worstIdx: number | null = null\n  let prev = stepRewards[0]!.reward\n  for (let i = 0; i < stepRewards.length; i++) {\n    const s = stepRewards[i]!\n    const w = s.weight ?? 1\n    sumW += w\n    sumWR += w * s.reward\n    if (s.reward < 0.5) failures++\n    if (i > 0) {\n      const delta = s.reward - prev\n      if (delta < worstDelta) {\n        worstDelta = delta\n        worstIdx = i\n      }\n      prev = s.reward\n    } else {\n      prev = s.reward\n    }\n  }\n  return {\n    runId,\n    totalSteps: stepRewards.length,\n    meanReward: sumW === 0 ? 0 : sumWR / sumW,\n    sumWeightedReward: sumWR,\n    failureFraction: failures / stepRewards.length,\n    worstStepDelta: worstDelta,\n    worstStepIndex: worstIdx,\n  }\n}\n\nexport interface PrmTrainingTriple {\n  /** Prefix run-id (or composite key) — the trajectory up to step k-1. */\n  prefixRunId: string\n  prefixStepIndex: number\n  /** The step that came next on a high-reward trajectory. */\n  chosenSpanId: string\n  chosenReward: number\n  /** A step from a divergent low-reward trajectory at the same prefix length. */\n  rejectedSpanId: string\n  rejectedReward: number\n  /** The prefix run came from this run; the rejected step came from `rejectedRunId`. */\n  rejectedRunId: string\n  marginScore: number\n}\n\n/**\n * Build PRM training triples. The shape: pair runs that share an early\n * prefix (same scenario, same first N steps) and diverge later — at the\n * point of divergence, the high-reward run's next step is `chosen`, the\n * low-reward run's next step is `rejected`. This is the canonical PRM\n * training data shape from Lightman et al. and DeepSeek-R1 process\n * supervision.\n *\n * Implementation note: we don't have a way to detect \"same prefix\" in\n * the general agent setting (token-level prefixes require hashing model\n * outputs). The current heuristic groups by `(scenarioId, prefixSpanName\n * sequence)` — runs are paired when their first K span names match. For\n * production use this should be replaced with a proper trajectory-prefix\n * hash; the heuristic is good enough for early-stage scaffolding.\n */\nexport function prmTrainingPairs(\n  stepRewardsByRun: Map<string, StepReward[]>,\n  opts: { minMargin?: number; minPrefixLength?: number } = {},\n): PrmTrainingTriple[] {\n  const minMargin = opts.minMargin ?? 0.2\n  const minPrefix = opts.minPrefixLength ?? 1\n  const runs = [...stepRewardsByRun.entries()].map(([runId, steps]) => ({ runId, steps }))\n  const triples: PrmTrainingTriple[] = []\n\n  for (let i = 0; i < runs.length; i++) {\n    for (let j = i + 1; j < runs.length; j++) {\n      const a = runs[i]!\n      const b = runs[j]!\n      const minLen = Math.min(a.steps.length, b.steps.length)\n      if (minLen < minPrefix + 1) continue\n\n      // Find the first index where the trajectories diverge: either by\n      // step structure (kind/name mismatch) OR by reward gap ≥ minMargin.\n      // Names that match but rewards that differ ARE divergence — that's\n      // the canonical PRM training case (same step structure, different\n      // outcomes via state/context).\n      let divergenceIdx = -1\n      for (let k = 0; k < minLen; k++) {\n        const sa = a.steps[k]!\n        const sb = b.steps[k]!\n        const structuralDivergence = sa.kind !== sb.kind || sa.name !== sb.name\n        const rewardGap = Math.abs(sa.reward - sb.reward)\n        if (structuralDivergence || rewardGap >= minMargin) {\n          divergenceIdx = k\n          break\n        }\n      }\n      if (divergenceIdx < 0) continue\n      if (divergenceIdx < minPrefix) continue\n\n      const aNext = a.steps[divergenceIdx]!\n      const bNext = b.steps[divergenceIdx]!\n      const margin = Math.abs(aNext.reward - bNext.reward)\n      if (margin < minMargin) continue\n\n      const chosen = aNext.reward > bNext.reward ? aNext : bNext\n      const rejected = aNext.reward > bNext.reward ? bNext : aNext\n      const chosenRun = aNext.reward > bNext.reward ? a.runId : b.runId\n      const rejectedRun = aNext.reward > bNext.reward ? b.runId : a.runId\n      triples.push({\n        prefixRunId: chosenRun,\n        prefixStepIndex: divergenceIdx - 1,\n        chosenSpanId: chosen.spanId,\n        chosenReward: chosen.reward,\n        rejectedSpanId: rejected.spanId,\n        rejectedReward: rejected.reward,\n        rejectedRunId: rejectedRun,\n        marginScore: chosen.reward - rejected.reward,\n      })\n    }\n  }\n  return triples\n}\n","/**\n * `runRLCampaign` — top-level orchestrator that runs the matrix and\n * produces every RL-ready artifact in one call.\n *\n * Wires:\n *   1. `runEvalCampaign` for the matrix run (capture, integrity, hooks)\n *   2. `extractVerifiableReward` over each run, separating deterministic\n *      from probabilistic reward sources for the trainer\n *   3. `extractPreferences` to produce DPO/PPO/KTO triples\n *   4. `evaluateInterimReleaseConfidence` over paired deltas (anytime-valid)\n *   5. `rubricPredictiveValidity` against an outcome store, when provided\n *   6. `detectRewardHacking` as a standing hygiene check\n *   7. Trainer-format export rows ready for prime-rl / TRL / verl\n *\n * The output `RLCampaignResult` is a single, audit-ready artifact: every\n * stage's output is in there. The consumer's downstream fits in a single\n * line: pass `result.preferences` to their DPO trainer, `result.grpoRows`\n * to GRPO, `result.runs` plus `result.rewardSignals` to a custom RL loop.\n */\n\nimport {\n  type EvalCampaignOptions,\n  type EvalCampaignResult,\n  runEvalCampaign,\n} from '../eval-campaign'\nimport type { OutcomeStore } from '../meta-eval/outcome-store'\nimport {\n  type RubricPredictiveValidityReport,\n  rubricPredictiveValidity,\n} from '../meta-eval/rubric-predictive-validity'\nimport type { RunRecord } from '../run-record'\nimport { evaluateInterimReleaseConfidence, type InterimReleaseConfidence } from '../sequential'\nimport {\n  type DpoExportRow,\n  type DpoLookups,\n  type GrpoExportRow,\n  type GrpoLookups,\n  type SftExportRow,\n  type SftLookups,\n  toDpoRows,\n  toGrpoRows,\n  toSftRows,\n} from './exporters'\nimport {\n  type ExtractPreferencesOptions,\n  extractPreferences,\n  type PreferenceExtractionReport,\n} from './preferences'\nimport { detectRewardHacking, type RewardHackingReport } from './reward-hacking'\nimport {\n  extractVerifiableRewardsFromRecords,\n  type VerifiableReward,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport interface RunRLCampaignOptions<V> extends EvalCampaignOptions<V> {\n  /** Preference-extraction options. Default uses paired-by-scenario-and-seed with min-margin 0.05. */\n  preferences?: ExtractPreferencesOptions\n  /** Verifiable-reward extraction options. */\n  verifiableReward?: VerifiableRewardExtractionOptions\n  /** Outcome store + metric names — when supplied, runs `rubricPredictiveValidity` post-campaign. */\n  outcomeStore?: OutcomeStore\n  outcomeMetrics?: string[]\n  /** Anytime-valid sequential evaluation options. */\n  sequential?: { alpha?: number; bound?: number; rope?: { low: number; high: number } }\n  /** Trainer-format export lookups. When provided, the orchestrator builds the corresponding rows. */\n  trainerExport?: {\n    dpo?: DpoLookups\n    grpo?: GrpoLookups\n    sft?: SftLookups\n  }\n}\n\nexport interface RLCampaignResult<V> {\n  campaign: EvalCampaignResult\n  /** Per-run verifiable reward (deterministic when available, probabilistic fallback otherwise). */\n  rewardSignals: Array<{ runId: string; reward: VerifiableReward | null }>\n  /** Preference extraction report. */\n  preferences: PreferenceExtractionReport\n  /** Anytime-valid interim verdict over the paired deltas (vs comparator). */\n  interimConfidence: InterimReleaseConfidence | null\n  /** Standing reward-hacking hygiene check. */\n  rewardHacking: RewardHackingReport\n  /** Predictive validity, when an outcome store was supplied. */\n  predictiveValidity: RubricPredictiveValidityReport | null\n  /** Trainer-export rows, populated only for the formats the caller requested via `trainerExport`. */\n  trainerRows: {\n    dpo?: DpoExportRow[]\n    grpo?: GrpoExportRow[]\n    sft?: SftExportRow[]\n  }\n  /**\n   * One-line top-level summary the consumer can log.\n   */\n  summary: string\n  /**\n   * Convenience type-tag — consumers can branch on `result.kind`.\n   */\n  kind: 'agent-eval-rl-campaign'\n  unusedVariant?: V\n}\n\nexport async function runRLCampaign<V>(\n  opts: RunRLCampaignOptions<V>,\n): Promise<RLCampaignResult<V>> {\n  // ── 1. Run the matrix ──────────────────────────────────────────────\n  const campaign = await runEvalCampaign(opts)\n\n  // ── 2. Extract reward signals (deterministic-first) ────────────────\n  const rewardSignals = extractVerifiableRewardsFromRecords(\n    campaign.runs,\n    opts.verifiableReward ?? {},\n  )\n\n  // ── 3. Extract preference triples ──────────────────────────────────\n  const preferences = extractPreferences(campaign.runs, {\n    strategy: opts.preferences?.strategy ?? 'paired-by-scenario-and-seed',\n    minMargin: opts.preferences?.minMargin ?? 0.05,\n    splitTag: opts.preferences?.splitTag ?? opts.splitTag ?? 'holdout',\n    rewardOf: opts.preferences?.rewardOf,\n  })\n\n  // ── 4. Sequential / anytime-valid interim verdict ──────────────────\n  let interimConfidence: InterimReleaseConfidence | null = null\n  if (opts.report?.comparator) {\n    const comparator = opts.report.comparator\n    const deltaSeries = collectPairedDeltaSeries(campaign.runs, comparator)\n    if (deltaSeries.some((s) => s.deltas.length > 0)) {\n      interimConfidence = evaluateInterimReleaseConfidence({\n        deltaSeries,\n        alpha: opts.sequential?.alpha,\n        bound: opts.sequential?.bound,\n        rope: opts.sequential?.rope ?? opts.report?.rope,\n      })\n    }\n  }\n\n  // ── 5. Standing reward-hacking hygiene ─────────────────────────────\n  const rewardHacking = detectRewardHacking({\n    runs: campaign.runs,\n    verifiableRewardOptions: opts.verifiableReward,\n  })\n\n  // ── 6. Predictive validity (when outcomes are supplied) ────────────\n  let predictiveValidity: RubricPredictiveValidityReport | null = null\n  if (opts.outcomeStore && opts.outcomeMetrics && opts.outcomeMetrics.length > 0) {\n    predictiveValidity = await rubricPredictiveValidity({\n      runs: campaign.runs,\n      outcomes: opts.outcomeStore,\n      outcomeMetrics: opts.outcomeMetrics,\n    })\n  }\n\n  // ── 7. Trainer-format export ───────────────────────────────────────\n  const trainerRows: RLCampaignResult<V>['trainerRows'] = {}\n  if (opts.trainerExport?.dpo) {\n    trainerRows.dpo = await toDpoRows(preferences.pairs, opts.trainerExport.dpo)\n  }\n  if (opts.trainerExport?.grpo) {\n    trainerRows.grpo = await toGrpoRows(campaign.runs, opts.trainerExport.grpo)\n  }\n  if (opts.trainerExport?.sft) {\n    trainerRows.sft = await toSftRows(campaign.runs, opts.trainerExport.sft)\n  }\n\n  const summary = buildSummary({\n    campaign,\n    preferences,\n    interimConfidence,\n    rewardHacking,\n    predictiveValidity,\n  })\n\n  return {\n    campaign,\n    rewardSignals,\n    preferences,\n    interimConfidence,\n    rewardHacking,\n    predictiveValidity,\n    trainerRows,\n    summary,\n    kind: 'agent-eval-rl-campaign',\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction collectPairedDeltaSeries(\n  runs: RunRecord[],\n  comparator: string,\n): Array<{ candidateId: string; deltas: number[] }> {\n  // Pair on (scenarioId, seed). For each candidate that isn't the comparator,\n  // compute candidate.score - comparator.score on matching cells.\n  const baseline = new Map<string, number>()\n  for (const r of runs) {\n    if (r.candidateId !== comparator) continue\n    const sid = r.scenarioId ?? r.experimentId\n    const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    baseline.set(`${sid}::${r.seed}`, score)\n  }\n  const byCandidate = new Map<string, number[]>()\n  for (const r of runs) {\n    if (r.candidateId === comparator) continue\n    const sid = r.scenarioId ?? r.experimentId\n    const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    const baseScore = baseline.get(`${sid}::${r.seed}`)\n    if (typeof baseScore !== 'number') continue\n    const arr = byCandidate.get(r.candidateId) ?? []\n    arr.push(score - baseScore)\n    byCandidate.set(r.candidateId, arr)\n  }\n  return [...byCandidate.entries()].map(([candidateId, deltas]) => ({ candidateId, deltas }))\n}\n\nfunction buildSummary(args: {\n  campaign: EvalCampaignResult\n  preferences: PreferenceExtractionReport\n  interimConfidence: InterimReleaseConfidence | null\n  rewardHacking: RewardHackingReport\n  predictiveValidity: RubricPredictiveValidityReport | null\n}): string {\n  const c = args.campaign\n  const lines = [\n    `${c.campaignId}: ${c.runs.length} successful runs / ${c.failedRuns.length} failed (fingerprint ${c.campaignFingerprint.slice(0, 12)}…)`,\n    `preferences: ${args.preferences.pairs.length} (${args.preferences.strategy}, ${args.preferences.pairsBelowMargin} below margin)`,\n  ]\n  if (args.interimConfidence) {\n    lines.push(\n      `sequential verdict: ${args.interimConfidence.recommendation.decision}` +\n        (args.interimConfidence.recommendation.candidateId\n          ? ` ${args.interimConfidence.recommendation.candidateId}`\n          : ''),\n    )\n  }\n  lines.push(\n    `reward-hacking: ${args.rewardHacking.verdict} (${args.rewardHacking.findings.length} signals checked)`,\n  )\n  if (args.predictiveValidity) {\n    const top = args.predictiveValidity.ranked[0]\n    lines.push(\n      `top-rubric: ${top?.rubric ?? 'none'} ρ=${(top?.spearman ?? 0).toFixed(2)} (${top?.verdict ?? 'no data'})`,\n    )\n  }\n  return lines.join(' | ')\n}\n\n// Re-export `runEvalCampaign` so consumers can pick the lower-level\n// primitive without flipping import paths.\nexport { runEvalCampaign } from '../eval-campaign'\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8EA,eAAsB,gBAAgB,MAAqD;AACzF,QAAM,SAA8B,CAAC;AACrC,aAAW,UAAU,KAAK,SAAS;AACjC,UAAM,IAAI,MAAM,KAAK,YAAY,MAAM;AACvC,WAAO,KAAK;AAAA,MACV,UAAU,OAAO;AAAA,MACjB,MAAM,OAAO;AAAA,MACb,OAAO,EAAE;AAAA,MACT,SAAS,EAAE;AAAA,MACX,KAAK,EAAE;AAAA,MACP,SAAS,EAAE;AAAA,IACb,CAAC;AAAA,EACH;AACA,QAAM,SAAS,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AACzD,QAAM,WAAW,OAAO,UAAU,IAAI,YAAY,MAAM,IAAI;AAC5D,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAO,EAAE,QAAQ,EAAE,QAAQ,IAAI,CAAE;AAChE,SAAO,EAAE,aAAa,KAAK,aAAa,QAAQ,QAAQ,UAAU,KAAK;AACzE;AAqBA,eAAsB,QAAW,MAAkE;AACjG,MAAI,KAAK,KAAK,EAAG,OAAM,IAAI,gBAAgB,wBAAwB;AACnE,QAAM,WAAgB,CAAC;AACvB,QAAM,SAAmB,CAAC;AAC1B,WAAS,IAAI,GAAG,IAAI,KAAK,GAAG,KAAK;AAC/B,UAAM,IAAI,MAAM,KAAK,OAAO,CAAC;AAC7B,aAAS,KAAK,CAAC;AACf,WAAO,KAAK,MAAM,KAAK,QAAQ,CAAC,CAAC;AAAA,EACnC;AACA,MAAI,YAAY;AAChB,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,IAAK,KAAI,OAAO,CAAC,IAAK,OAAO,SAAS,EAAI,aAAY;AACzF,QAAM,YAAY,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAC7D,SAAO;AAAA,IACL,MAAM,SAAS,SAAS;AAAA,IACxB,WAAW,OAAO,SAAS;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AA0BA,eAAsB,gBACpB,MACmC;AACnC,MAAI,KAAK,KAAK,EAAG,OAAM,IAAI,gBAAgB,gCAAgC;AAC3E,QAAM,WAAgB,CAAC;AACvB,QAAM,YAAoC,CAAC;AAC3C,WAAS,IAAI,GAAG,IAAI,KAAK,GAAG,KAAK;AAC/B,UAAM,IAAI,MAAM,KAAK,OAAO,CAAC;AAC7B,aAAS,KAAK,CAAC;AACf,UAAM,MAAM,KAAK,UAAU,CAAC;AAC5B,cAAU,GAAG,KAAK,UAAU,GAAG,KAAK,KAAK;AAAA,EAC3C;AACA,MAAI,SAAS;AACb,MAAI,MAAM;AACV,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,SAAS,GAAG;AAC9C,QAAI,IAAI,KAAK;AACX,YAAM;AACN,eAAS;AAAA,IACX;AAAA,EACF;AACA,QAAM,iBAAiB,SAAS,KAAK,CAAC,MAAM,KAAK,UAAU,CAAC,MAAM,MAAM,KAAK,SAAS,CAAC;AACvF,SAAO;AAAA,IACL;AAAA,IACA,WAAW,MAAM,KAAK;AAAA,IACtB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAeO,SAAS,eAAe,QAAgD;AAC7E,QAAM,aAAiC,CAAC;AACxC,aAAW,KAAK,QAAQ;AACtB,UAAM,YAAY,OAAO;AAAA,MACvB,CAAC,MACC,MAAM,KAAK,EAAE,QAAQ,EAAE,QAAQ,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE;AAAA,IACzF;AACA,QAAI,CAAC,UAAW,YAAW,KAAK,CAAC;AAAA,EACnC;AACA,SAAO,WAAW,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClD;AAIA,SAAS,YAAY,QAAqC;AAIxD,QAAM,KAAK,OAAO,IAAI,CAAC,MAAM,KAAK,IAAI,KAAK,IAAI,OAAO,EAAE,IAAI,CAAC,CAAC;AAC9D,QAAM,KAAK,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK;AACpC,QAAM,IAAI,GAAG;AACb,QAAM,KAAK,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3C,QAAM,KAAK,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3C,MAAI,MAAM;AACV,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAQ,GAAG,CAAC,IAAK,OAAO,GAAG,CAAC,IAAK;AACjC,YAAQ,GAAG,CAAC,IAAK,OAAO;AAAA,EAC1B;AACA,SAAO,QAAQ,IAAI,IAAI,MAAM;AAC/B;;;ACrIA,eAAsB,sBACpB,OACA,OAAkC,CAAC,GACA;AACnC,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,QAAQ,KAAK,cAAc;AAEjC,MAAI,CAAC,MAAM,aAAa,CAAC,MAAM,cAAc;AAC3C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,YACJ,MAAM,aAAc,MAAM,QAAQ,IAAI,MAAM,UAAU,IAAI,CAAC,MAAM,MAAM,aAAc,MAAM,CAAC,CAAC,CAAC;AAChG,MAAI,UAAU,WAAW,MAAM,UAAU,QAAQ;AAC/C,UAAM,IAAI;AAAA,MACR,2CAA2C,UAAU,MAAM,qBAAgB,MAAM,UAAU,MAAM;AAAA,IACnG;AAAA,EACF;AAGA,QAAM,aAAa,MAAM,QAAQ,IAAI,MAAM,UAAU,IAAI,CAAC,MAAM,MAAM,QAAQ,CAAC,CAAC,CAAC;AACjF,QAAM,aAAa,MAAM,QAAQ,IAAI,UAAU,IAAI,CAAC,MAAM,MAAM,QAAQ,CAAC,CAAC,CAAC;AAE3E,QAAM,cAAc,MAAM,UAAU,IAAI,CAAC,GAAG,OAAO;AAAA,IACjD,YAAY,MAAM,WAAW,CAAC;AAAA,IAC9B,eAAe,WAAW,CAAC;AAAA,IAC3B,gBAAgB,WAAW,CAAC;AAAA,IAC5B,OAAO,WAAW,CAAC,IAAK,WAAW,CAAC;AAAA,IACpC,QAAQ;AAAA,EACV,EAAE;AAGF,QAAM,QAAQ,YAAY,OAAO,CAAC,MAAM,EAAE,iBAAiB,SAAS,EAAE,kBAAkB,KAAK;AAC7F,MAAI,MAAM,SAAS,GAAG;AACpB,WAAO;AAAA,MACL;AAAA,MACA,YAAY,EAAE,GAAG,GAAG,GAAG,EAAE;AAAA,MACzB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,wBAAwB;AAAA,MACxB,QAAQ,mCAAmC,MAAM,MAAM;AAAA,MACvD,GAAG,MAAM;AAAA,IACX;AAAA,EACF;AAEA,QAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,aAAa;AAClD,QAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,cAAc;AACnD,QAAM,aAAa,mBAAmB,WAAW,SAAS;AAC1D,QAAM,SAAS,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AACvC,QAAM,eAAe,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACrD,QAAM,SAAS,aAAa,KAAK,MAAM,aAAa,SAAS,CAAC,CAAC;AAC/D,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAMxD,QAAM,UAAU,MAAM,IAAI,CAAC,MAAM,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,IAAI,KAAK,IAAI,EAAE,KAAK,IAAI,CAAC,CAAC,CAAC;AACvF,QAAM,EAAE,QAAQ,IAAI,kBAAkB,SAAS,GAAG;AAClD,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAM,IAAI,MAAM,CAAC;AACjB,UAAM,MAAM,YAAY,UAAU,CAAC,MAAM,EAAE,eAAe,EAAE,UAAU;AACtE,QAAI,OAAO,EAAG,aAAY,GAAG,EAAG,SAAS,QAAQ,CAAC;AAAA,EACpD;AAEA,QAAM,yBAAyB,WAAW,IAAI,OAAO,UAAU,CAAC;AAChE,QAAM,SAAS,yBACX,YAAY,WAAW,EAAE,QAAQ,CAAC,CAAC,MAAM,GAAG,oBAAoB,OAAO,QAAQ,CAAC,CAAC,WAAM,aAAa,KACpG,WAAW,KAAK,MACd,uCAAuC,WAAW,EAAE,QAAQ,CAAC,CAAC,MAC9D,8CAA8C,OAAO,QAAQ,CAAC,CAAC;AAErE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,aAAa;AAAA,IACb,WAAW;AAAA,IACX;AAAA,IACA;AAAA,IACA,GAAG,MAAM;AAAA,EACX;AACF;AAUO,SAAS,gBACd,aACA,SAAgD,CAAC,GAAG,MAAM,GAAG,CAAC,KAAM,IAAI,KAAM,IAAI,SAAS,EAAE,CAAC,IACrE;AACzB,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,UAAI,SAAS,SAAS;AACtB,kBAAY,QAAQ,CAAC,IAAI,MAAM;AAC7B,cAAM,cAAc,OAAO,IAAI,CAAC;AAChC,cAAM,KAAK,IAAI,OAAO,MAAM,YAAY,EAAE,CAAC,OAAO,GAAG;AACrD,iBAAS,OAAO,QAAQ,IAAI,WAAW;AAAA,MACzC,CAAC;AACD,aAAO,EAAE,GAAG,UAAU,OAAO;AAAA,IAC/B;AAAA,EACF;AACF;AAOO,SAAS,aACd,gBACA,MACyB;AACzB,MAAI,IAAI,SAAS;AACjB,QAAM,MAAM,MAAc;AACxB,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,YAAM,YAAY,eAAe,SAAS,QAAQ,GAAG;AACrD,aAAO,EAAE,GAAG,UAAU,QAAQ,UAAU;AAAA,IAC1C;AAAA,EACF;AACF;AAOO,SAAS,uBACd,QACA,WAAgC,UACP;AACzB,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,YAAM,SACJ,aAAa,WAAW,GAAG,MAAM,IAAI,SAAS,MAAM,KAAK,GAAG,SAAS,MAAM,IAAI,MAAM;AACvF,aAAO,EAAE,GAAG,UAAU,OAAO;AAAA,IAC/B;AAAA,EACF;AACF;AAIA,SAAS,YAAY,GAAmB;AACtC,SAAO,EAAE,QAAQ,uBAAuB,MAAM;AAChD;;;ACpKO,SAAS,4BACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAElD,MAAI,aAAa,WAAW,GAAG;AAC7B,WAAO,aAAa;AAAA,EACtB;AAEA,QAAM,UAAoB,CAAC;AAC3B,QAAM,kBAA4B,CAAC;AACnC,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,gEAAgE,EAAE,KAAK;AAAA,MACzE;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,YAAQ,KAAK,CAAC;AACd,oBAAgB,KAAK,IAAI,CAAC;AAC1B,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,IAAI,QAAQ;AAClB,QAAM,QAAQ,gBAAgB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3D,QAAM,WAAW,gBAAgB,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC9F,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,kCACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,UAAoB,CAAC;AAC3B,QAAM,UAAoB,CAAC;AAC3B,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,sEAAsE,EAAE,KAAK;AAAA,MAC/E;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,YAAQ,KAAK,CAAC;AACd,YAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,CAAC;AACjD,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,GAAG,MAAM,IAAI,IAAI,QAAQ,CAAC,GAAI,CAAC;AAChE,QAAM,QAAQ,SAAS,IAAI,IAAI,QAAQ;AACvC,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,QAAM,MAAM,QAAQ,IAAI,CAAC,GAAG,MAAM,KAAK,QAAQ,CAAC,IAAK,MAAM;AAC3D,QAAM,WAAW,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO,IAAI;AAC7E,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,QAAQ;AAAA,IACjC,qBAAqB;AAAA,IACrB,GAAG,aAAa;AAAA,IAChB,qBAAqB;AAAA,EACvB;AACF;AAoBO,SAAS,aACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,gBAA0B,CAAC;AACjC,MAAI,OAAO;AACX,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI,gBAAgB,iDAAiD,EAAE,KAAK,GAAG;AAAA,IACvF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,UAAM,IACJ,OAAO,EAAE,SAAS,YAAY,OAAO,SAAS,EAAE,IAAI,IAChD,MAAM,EAAE,MAAM,KAAK,KAAK,KAAK,IAAI,IACjC;AACN,QAAI,MAAM,MAAM;AACd,oBAAc,KAAK,IAAI,CAAC;AAAA,IAC1B,OAAO;AACL,oBAAc,KAAK,IAAI,KAAK,IAAI,EAAE;AAAA,IACpC;AACA,QAAI,IAAI,KAAM,QAAO;AACrB,YAAQ;AACR,aAAS,IAAI;AAAA,EACf;AACA,QAAM,IAAI,cAAc;AACxB,QAAM,QAAQ,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACzD,QAAM,WAAW,cAAc,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC5F,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAC9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,qBACd,cACA,OAAyB,CAAC,GACmD;AAC7E,SAAO;AAAA,IACL,KAAK,4BAA4B,cAAc,IAAI;AAAA,IACnD,OAAO,kCAAkC,cAAc,IAAI;AAAA,IAC3D,IAAI,aAAa,cAAc,IAAI;AAAA,EACrC;AACF;AAIA,SAAS,eAAkC;AACzC,SAAO,EAAE,OAAO,GAAG,eAAe,GAAG,qBAAqB,GAAG,GAAG,GAAG,qBAAqB,EAAE;AAC5F;AAEA,SAAS,MAAM,GAAW,IAAY,IAAoB;AACxD,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,CAAC;AACrC;;;AC9JA,IAAM,oBAA2C;AAEjD,IAAM,iBAAiB,CAAC,QAAkC;AACxD,QAAM,IAAI,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AAClD,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAcO,SAAS,mBACd,MACA,OAAkC,CAAC,GACP;AAC5B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,WAAW,KAAK,YAAY;AAElC,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ;AAC3D,QAAM,gBAA0D,CAAC;AACjE,aAAW,OAAO,UAAU;AAC1B,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,MAAM,KAAM;AAChB,kBAAc,KAAK,EAAE,KAAK,OAAO,EAAE,CAAC;AAAA,EACtC;AAEA,QAAM,QAA4B,CAAC;AACnC,MAAI,mBAAmB;AACvB,MAAI,iBAAiB;AACrB,MAAI,iBAAiB;AAErB,MAAI,aAAa,+BAA+B;AAI9C,UAAM,SAAS,oBAAI,IAAsD;AACzE,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,YAAM,MAAM,GAAG,GAAG,KAAK,EAAE,IAAI,IAAI;AACjC,YAAM,MAAM,OAAO,IAAI,GAAG,KAAK,CAAC;AAChC,UAAI,KAAK,CAAC;AACV,aAAO,IAAI,KAAK,GAAG;AAAA,IACrB;AAEA,eAAW,CAAC,KAAK,OAAO,KAAK,OAAO,QAAQ,GAAG;AAC7C;AACA,UAAI,QAAQ,SAAS,GAAG;AACtB;AACA;AAAA,MACF;AACA,eAAS,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,iBAAS,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AAC3C,gBAAM,IAAI,QAAQ,CAAC;AACnB,gBAAM,IAAI,QAAQ,CAAC;AACnB,cAAI,EAAE,IAAI,gBAAgB,EAAE,IAAI,YAAa;AAC7C,gBAAM,SAAS,SAAS,GAAG,GAAG,IAAI,MAAM,IAAI,EAAE,CAAC,GAAI,SAAS;AAC5D,cAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,cAC9C;AAAA,QACP;AAAA,MACF;AAAA,IACF;AAAA,EACF,WAAW,aAAa,sBAAsB;AAE5C,UAAM,oBAAoB,oBAAI,IAG5B;AACF,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,UAAI,cAAc,kBAAkB,IAAI,GAAG;AAC3C,UAAI,CAAC,aAAa;AAChB,sBAAc,oBAAI,IAAI;AACtB,0BAAkB,IAAI,KAAK,WAAW;AAAA,MACxC;AACA,YAAM,MAAM,YAAY,IAAI,EAAE,IAAI,WAAW;AAC7C,UAAI,KAAK;AACP,YAAI,OAAO,EAAE;AACb,YAAI;AAAA,MACN,MAAO,aAAY,IAAI,EAAE,IAAI,aAAa,EAAE,KAAK,EAAE,KAAK,KAAK,EAAE,OAAO,GAAG,EAAE,CAAC;AAAA,IAC9E;AACA,eAAW,CAAC,KAAK,UAAU,KAAK,kBAAkB,QAAQ,GAAG;AAC3D;AACA,YAAM,MAAM,CAAC,GAAG,WAAW,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG,OAAO;AAAA,QACzD,KAAK,IAAI;AAAA,QACT,OAAO,IAAI,MAAM,IAAI;AAAA,QACrB,WAAW;AAAA,MACb,EAAE;AACF,UAAI,IAAI,SAAS,GAAG;AAClB;AACA;AAAA,MACF;AACA,eAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,iBAAS,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACvC,gBAAM,SAAS,SAAS,IAAI,CAAC,GAAI,IAAI,CAAC,GAAI,KAAK,SAAS;AACxD,cAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,cAC9C;AAAA,QACP;AAAA,MACF;AAAA,IACF;AAAA,EACF,OAAO;AAEL,UAAM,aAAa,oBAAI,IAAsD;AAC7E,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,YAAM,MAAM,WAAW,IAAI,GAAG,KAAK,CAAC;AACpC,UAAI,KAAK,CAAC;AACV,iBAAW,IAAI,KAAK,GAAG;AAAA,IACzB;AACA,eAAW,CAAC,KAAK,GAAG,KAAK,WAAW,QAAQ,GAAG;AAC7C;AACA,UAAI,IAAI,SAAS,GAAG;AAClB;AACA;AAAA,MACF;AACA,YAAM,SAAS,CAAC,GAAG,GAAG,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AACxD,YAAM,MAAM,OAAO,OAAO,SAAS,CAAC;AACpC,YAAM,MAAM,OAAO,CAAC;AACpB,UAAI,IAAI,IAAI,gBAAgB,IAAI,IAAI,aAAa;AAC/C;AACA;AAAA,MACF;AACA,YAAM,SAAS,SAAS,KAAK,KAAK,KAAK,SAAS;AAChD,UAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,UAC9C;AAAA,IACP;AAAA,EACF;AAEA,SAAO,EAAE,OAAO,gBAAgB,kBAAkB,gBAAgB,SAAS;AAC7E;AAOO,SAAS,YACd,SACA,UAC6D;AAC7D,SAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,IACzB,QAAQ,SAAS,EAAE,KAAK,gBAAgB;AAAA,IACxC,QAAQ,EAAE,KAAK;AAAA;AAAA,IACf,UAAU,EAAE,KAAK;AAAA,EACnB,EAAE;AACJ;AAOO,SAAS,kBACd,SAC2F;AAC3F,SAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,IACzB,YAAY,EAAE;AAAA,IACd,aAAa,EAAE;AAAA,IACf,eAAe,EAAE;AAAA,IACjB,QAAQ,EAAE;AAAA,EACZ,EAAE;AACJ;AAIA,SAAS,SACP,GACA,GACA,YACA,WACgE;AAChE,QAAM,SAAS,KAAK,IAAI,EAAE,QAAQ,EAAE,KAAK;AACzC,MAAI,SAAS,UAAW,QAAO,EAAE,MAAM,SAAS;AAChD,QAAM,CAAC,QAAQ,QAAQ,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC;AAC7D,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM;AAAA,MACJ;AAAA,MACA,aAAa,OAAO,IAAI;AAAA,MACxB,eAAe,SAAS,IAAI;AAAA,MAC5B,iBAAiB,OAAO,IAAI;AAAA,MAC5B,mBAAmB,SAAS,IAAI;AAAA,MAChC,aAAa,OAAO,QAAQ,SAAS;AAAA,MACrC,QAAQ,EAAE,QAAQ,OAAO,OAAO,UAAU,SAAS,MAAM;AAAA,MACzD,MAAM,OAAO,IAAI,SAAS,SAAS,IAAI,OAAO,OAAO,IAAI,OAAO;AAAA,MAChE,MAAM;AAAA,QACJ,kBAAkB,OAAO,IAAI;AAAA,QAC7B,oBAAoB,SAAS,IAAI;AAAA,QACjC,kBAAkB,OAAO,IAAI;AAAA,QAC7B,oBAAoB,SAAS,IAAI;AAAA,QACjC,aAAa,OAAO,IAAI;AAAA,QACxB,eAAe,SAAS,IAAI;AAAA,MAC9B;AAAA,IACF;AAAA,EACF;AACF;AAQA,SAAS,WAAW,KAAwB;AAC1C,MAAI,OAAO,IAAI,eAAe,YAAY,IAAI,WAAW,SAAS,EAAG,QAAO,IAAI;AAChF,QAAM,UAAU,IAAI,QAAQ,IAAI;AAChC,MAAI,OAAO,YAAY,YAAY,OAAO,SAAS,OAAO,EAAG,QAAO,OAAO,OAAO;AAClF,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,SAAO,IAAI;AACb;;;AC/RO,SAAS,qBACd,UACA,KACa;AACb,QAAM,WAAW,IAAI,YAAY;AACjC,QAAM,cAAc,IAAI,eAAe,SAAS;AAChD,SAAO,SAAS,MAAM,IAAI,CAAC,SAAS;AAClC,UAAM,aAAa,OAAO,OAAO,KAAK,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AACzE,UAAM,QACJ,WAAW,SAAS,IAAI,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW,SAAS;AACtF,UAAM,MAA8B,EAAE,KAAK,KAAK,KAAK,aAAa,KAAK,WAAW;AAClF,eAAW,SAAS,OAAO,OAAO,KAAK,WAAW,GAAG;AACnD,iBAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,MAAM,UAAU,GAAG;AAC3D,YAAI,OAAO,SAAS,KAAK,EAAG,KAAI,OAAO,GAAG,EAAE,IAAI;AAAA,MAClD;AAAA,IACF;AACA,QAAI,OAAO,KAAK,eAAe,SAAU,KAAI,aAAa,KAAK;AAC/D,UAAM,UAAgC,EAAE,IAAI;AAC5C,QAAI,aAAa,UAAW,SAAQ,eAAe;AAAA,QAC9C,SAAQ,cAAc;AAC3B,WAAO;AAAA,MACL,OAAO,KAAK;AAAA,MACZ,cAAc,IAAI;AAAA,MAClB;AAAA,MACA,MAAM,KAAK;AAAA,MACX,OAAO,IAAI;AAAA,MACX,YAAY,IAAI;AAAA,MAChB,YAAY,IAAI;AAAA,MAChB,WAAW,IAAI;AAAA,MACf,QAAQ,KAAK;AAAA,MACb,SAAS,OAAO,SAAS,KAAK,OAAO,IAAI,KAAK,UAAW,IAAI,kBAAkB;AAAA,MAC/E,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC;AAAA,MACA,aAAa,KAAK,QAAQ,eAAe;AAAA,MACzC;AAAA,MACA,YAAY,KAAK;AAAA,IACnB;AAAA,EACF,CAAC;AACH;AAQO,SAAS,8BACd,QACA,KACA,OAA2B,CAAC,GACjB;AACX,QAAM,WAAW,IAAI,YAAY;AACjC,QAAM,QAAQ,KAAK,SAAS,OAAO,IAAI,WAAW,IAAI,IAAI,YAAY,IAAI,OAAO,SAAS;AAE1F,QAAM,MAA8B;AAAA,IAClC,YAAY,OAAO;AAAA,IACnB,YAAY,OAAO;AAAA,IACnB,aAAa,OAAO;AAAA,IACpB,eAAe,OAAO;AAAA,IACtB,aAAa,OAAO;AAAA,IACpB,eAAe,OAAO;AAAA,EACxB;AACA,aAAW,SAAS,OAAO,QAAQ;AACjC,QAAI,OAAO,MAAM,UAAU,SAAU,KAAI,SAAS,MAAM,KAAK,EAAE,IAAI,MAAM;AACzE,QAAI,SAAS,MAAM,KAAK,OAAO,IAAI,MAAM,WAAW,SAAS,IAAI;AACjE,QAAI,MAAM,aAAa;AACrB,iBAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,MAAM,WAAW,GAAG;AACtD,YAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,SAAS,MAAM,KAAK,IAAI,CAAC,EAAE,IAAI;AAAA,MACtF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,OAAO,OAAO,KAAK,CAAC,MAAM,EAAE,WAAW,UAAU,EAAE,WAAW,OAAO;AACvF,QAAM,UAAgC,EAAE,IAAI;AAC5C,MAAI,aAAa,UAAW,SAAQ,eAAe,OAAO;AAAA,MACrD,SAAQ,cAAc,OAAO;AAElC,SAAO;AAAA,IACL;AAAA,IACA,cAAc,IAAI;AAAA,IAClB,aAAa,IAAI;AAAA,IACjB,MAAM;AAAA,IACN,OAAO,IAAI;AAAA,IACX,YAAY,IAAI;AAAA,IAChB,YAAY,IAAI;AAAA,IAChB,WAAW,IAAI;AAAA,IACf,QAAQ,OAAO;AAAA,IACf,SAAS,IAAI,kBAAkB;AAAA,IAC/B,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,IAClC;AAAA,IACA,aAAa,YAAY,qBAAqB,SAAS,IAAI;AAAA,IAC3D;AAAA,IACA,YAAY,IAAI;AAAA,EAClB;AACF;AAEA,SAAS,qBAAqB,OAA4B;AACxD,MAAI,MAAM,WAAW,QAAS,QAAO,SAAS,MAAM,KAAK;AACzD,MAAI,MAAM,WAAW,OAAQ,QAAO,SAAS,MAAM,KAAK;AACxD,MAAI,MAAM,WAAW,UAAW,QAAO,SAAS,MAAM,KAAK;AAC3D,SAAO,SAAS,MAAM,KAAK,IAAI,MAAM,MAAM;AAC7C;;;ACrEO,SAAS,gBACd,UACA,OAA2E,CAAC,GAC3D;AACjB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,UAAU,KAAK,iBAAiB;AAKtC,QAAM,YAAY,KAAK,aAAa;AAEpC,QAAM,aAAa,oBAAI,IAAY;AACnC,aAAW,KAAK,UAAU;AACxB,eAAW,IAAI,EAAE,MAAM;AACvB,eAAW,IAAI,EAAE,KAAK;AAAA,EACxB;AACA,QAAM,MAAM,CAAC,GAAG,UAAU,EAAE,KAAK;AACjC,QAAM,MAAM,IAAI,IAAI,IAAI,IAAI,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC;AAC/C,QAAM,IAAI,IAAI;AACd,MAAI,MAAM,EAAG,QAAO,EAAE,SAAS,CAAC,GAAG,YAAY,GAAG,YAAY,GAAG,WAAW,KAAK;AACjF,MAAI,MAAM,GAAG;AACX,WAAO;AAAA,MACL,SAAS,CAAC,EAAE,aAAa,IAAI,CAAC,GAAI,UAAU,GAAG,aAAa,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC;AAAA,MAC9E,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,IACb;AAAA,EACF;AAIA,QAAM,IAAgB,MAAM,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC,CAAC;AAClF,QAAM,IAAgB,MAAM,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC,CAAC;AAClF,aAAW,KAAK,UAAU;AACxB,UAAM,IAAI,IAAI,IAAI,EAAE,MAAM;AAC1B,UAAM,IAAI,IAAI,IAAI,EAAE,KAAK;AACzB,UAAM,IAAI,EAAE,UAAU;AACtB,QAAI,EAAE,MAAM;AACV,QAAE,CAAC,EAAG,CAAC,KAAM,MAAM;AACnB,QAAE,CAAC,EAAG,CAAC,KAAM,MAAM;AAAA,IACrB,OAAO;AACL,QAAE,CAAC,EAAG,CAAC,KAAM;AAAA,IACf;AACA,MAAE,CAAC,EAAG,CAAC,KAAM;AACb,MAAE,CAAC,EAAG,CAAC,KAAM;AAAA,EACf;AAGA,QAAM,YAAY,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AAC7C,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,WAAU,CAAC,KAAM,EAAE,CAAC,EAAG,CAAC;AACpD,cAAU,CAAC,KAAM;AAAA,EACnB;AACA,QAAM,aAAa,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AAC9C,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,YAAW,CAAC,KAAM,EAAE,CAAC,EAAG,CAAC;AAAA,EACvD;AAGA,MAAI,QAAQ,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AACvC,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,SAAO,OAAO,SAAS,QAAQ;AAC7B,UAAM,WAAW,IAAI,MAAc,CAAC;AACpC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAI,QAAQ;AACZ,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAI,MAAM,EAAG;AACb,YAAI,EAAE,CAAC,EAAG,CAAC,MAAO,EAAG;AACrB,iBAAS,EAAE,CAAC,EAAG,CAAC,KAAM,MAAM,CAAC,IAAK,MAAM,CAAC;AAAA,MAC3C;AACA,eAAS,CAAC,IAAI,UAAU,IAAI,MAAM,CAAC,IAAK,UAAU,CAAC,IAAK;AAAA,IAC1D;AAEA,QAAI,SAAS;AACb,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,WAAU,KAAK,IAAI,KAAK,IAAI,QAAQ,SAAS,CAAC,CAAE,CAAC;AAC7E,UAAM,OAAO,KAAK,IAAI,SAAS,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,UAAS,CAAC,IAAI,SAAS,CAAC,IAAK;AAEzD,YAAQ;AACR,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,IAAI,KAAK,IAAI,SAAS,CAAC,IAAK,MAAM,CAAC,CAAE,IAAI,KAAK,IAAI,OAAO,MAAM,CAAC,CAAE;AACxE,UAAI,IAAI,MAAO,SAAQ;AAAA,IACzB;AACA,YAAQ;AACR,QAAI,QAAQ,IAAK;AAAA,EACnB;AAEA,QAAM,SAAS,KAAK,IAAI,GAAG,MAAM,IAAI,CAAC,MAAM,KAAK,IAAI,KAAK,IAAI,QAAQ,CAAC,CAAC,CAAC,CAAC;AAC1E,QAAM,UAAgC,IAAI,IAAI,CAAC,IAAI,OAAO;AAAA,IACxD,aAAa;AAAA,IACb,UAAU,MAAM,CAAC;AAAA,IACjB,aAAa,KAAK,IAAI,KAAK,IAAI,QAAQ,MAAM,CAAC,CAAE,CAAC,IAAI;AAAA,IACrD,GAAG,WAAW,CAAC;AAAA,IACf,MAAM,UAAU,CAAC,IAAK;AAAA,EACxB,EAAE;AAEF,SAAO;AAAA,IACL,SAAS,QAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AAAA,IACvD,YAAY;AAAA,IACZ,YAAY;AAAA,IACZ,WAAW,QAAQ;AAAA,EACrB;AACF;AAiBO,SAAS,eACd,SACA,SACA,OAAmB,CAAC,GACyB;AAC7C,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,IAAI,KAAK,WAAW;AAE1B,QAAM,KAAK,QAAQ,IAAI,QAAQ,MAAM,KAAK;AAC1C,QAAM,KAAK,QAAQ,IAAI,QAAQ,KAAK,KAAK;AAEzC,QAAM,YAAY,KAAK,IAAI,QAAQ,KAAK,MAAM;AAC9C,QAAM,SAAS,QAAQ,OAAO,MAAM;AACpC,QAAM,SAAS,QAAQ,OAAO,MAAM;AACpC,QAAM,IAAI,QAAQ,UAAU;AAE5B,QAAM,cAAc,IAAI,KAAK,SAAS;AACtC,QAAM,aAAa,IAAI,KAAK,UAAU,IAAI;AAE1C,UAAQ,IAAI,QAAQ,QAAQ,KAAK,WAAW;AAC5C,UAAQ,IAAI,QAAQ,OAAO,KAAK,UAAU;AAE1C,SAAO,EAAE,aAAa,WAAW;AACnC;AAsBO,SAAS,0BACd,OACmB;AACnB,QAAM,aAAa,MAAM,cAAc;AACvC,QAAM,QAAQ,oBAAI,IAA2D;AAC7E,aAAW,KAAK,MAAM,MAAM;AAC1B,UAAM,MAAM,MAAM,IAAI,EAAE,QAAQ,KAAK,CAAC;AACtC,QAAI,KAAK,EAAE,aAAa,EAAE,aAAa,OAAO,EAAE,MAAM,CAAC;AACvD,UAAM,IAAI,EAAE,UAAU,GAAG;AAAA,EAC3B;AACA,QAAM,WAA8B,CAAC;AACrC,aAAW,OAAO,MAAM,OAAO,GAAG;AAChC,aAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,eAAS,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACvC,cAAM,IAAI,IAAI,CAAC;AACf,cAAM,IAAI,IAAI,CAAC;AACf,YAAI,EAAE,gBAAgB,EAAE,YAAa;AACrC,cAAM,SAAS,KAAK,IAAI,EAAE,QAAQ,EAAE,KAAK;AACzC,YAAI,UAAU,YAAY;AACxB,mBAAS,KAAK,EAAE,QAAQ,EAAE,aAAa,OAAO,EAAE,aAAa,MAAM,MAAM,QAAQ,EAAE,CAAC;AAAA,QACtF,OAAO;AACL,gBAAM,CAAC,QAAQ,KAAK,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC;AAC1D,mBAAS,KAAK,EAAE,QAAQ,OAAO,aAAa,OAAO,MAAM,aAAa,QAAQ,OAAO,CAAC;AAAA,QACxF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;;;ACjMO,SAAS,wBACd,cACA,gBACA,MACwB;AACxB,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,QAAQ,KAAK,gBAAgB;AACnC,QAAM,SAAS,KAAK;AAEpB,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,CAAC;AAC/B,QAAI,KAAK,EAAE,KAAK;AAChB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,YAAY,eAAe,IAAI,CAAC,MAAM;AAC1C,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,UAAU,QAAQ,IAAI,CAAC,KAAK,CAAC;AACnC,UAAM,IAAI,QAAQ;AAClB,UAAM,OAAO,MAAM,IAAI,MAAM,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClE,UAAM,WACJ,IAAI,IACA,gBACA,QAAQ,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI,KAAK;AAGnE,UAAM,SAAS,KAAK,KAAK,QAAQ,IAAI,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,CAAC,CAAC;AACjE,WAAO,EAAE,WAAW,EAAE,WAAW,YAAY,EAAE,YAAY,GAAG,MAAM,UAAU,OAAO;AAAA,EACvF,CAAC;AAGD,QAAM,aAAa,QAAQ,UAAU;AACrC,MAAI,cAAc,QAAQ;AACxB,UAAM,OAAO,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC,CAAC;AAC3E,WAAO,UAAU,IAAI,CAAC,OAAO;AAAA,MAC3B,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO;AAAA,MACP,QAAQ,qCAAqC,EAAE,CAAC;AAAA,IAClD,EAAE;AAAA,EACJ;AACA,QAAM,YAAY,SAAS;AAC3B,QAAM,cAAc,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC9D,SAAO,UAAU,IAAI,CAAC,MAAM;AAC1B,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,SAAS;AAC5F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,QAAQ,YAAY,EAAE,SAAS,QAAQ,CAAC,CAAC,OAAO,EAAE,CAAC,UAAU,EAAE,KAAK,QAAQ,CAAC,CAAC;AAAA,IAChF;AAAA,EACF,CAAC;AACH;AA4BO,SAAS,mBACd,cACA,gBACA,MACwB;AACxB,QAAM,YAAY,KAAK,qBAAqB;AAC5C,QAAM,SAAS,KAAK,cAAc;AAClC,QAAM,QAAQ,KAAK,aAAa;AAChC,QAAM,MAAM,QAAQ,KAAK,IAAI;AAE7B,QAAM,UAAU,oBAAI,IAAkD;AACtE,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,OAAO,EAAE,QAAQ,EAAE,SAAS;AAClC,QAAI,KAAM,KAAI,UAAU;AAAA,QACnB,KAAI,YAAY;AACrB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,QAAQ,eAAe,IAAI,CAAC,MAAM;AACtC,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,IAAI,SAAS,IAAI;AACvB,UAAM,IAAI,QAAQ,IAAI;AAEtB,UAAM,UAAU,WAAW,GAAG,GAAG,GAAG;AACpC,UAAM,WAAW,KAAK,IAAI,UAAU,SAAS;AAG7C,UAAM,WAAY,IAAI,MAAO,IAAI,MAAM,KAAK,IAAI,IAAI;AACpD,UAAM,QAAQ,KAAK,IAAI,MAAM,KAAK,KAAK,QAAQ,CAAC;AAChD,UAAM,SAAS,KAAK,IAAI,GAAG,WAAW,UAAU,EAAE;AAClD,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,GAAG,IAAI,SAAS,IAAI;AAAA,MACpB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAED,QAAM,cAAc,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC1D,SAAO,MAAM,IAAI,CAAC,MAAM;AACtB,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,KAAK,MAAM;AAC9F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,KAAK,IAAI,GAAG,YAAY;AAAA,MAC/B,QAAQ,QAAQ,EAAE,EAAE,QAAQ,CAAC,CAAC,IAAI,EAAE,EAAE,QAAQ,CAAC,CAAC,YAAY,EAAE,QAAQ,QAAQ,CAAC,CAAC,YAAY,SAAS;AAAA,IACvG;AAAA,EACF,CAAC;AACH;AAGO,SAAS,2BACd,MACA,OAAyD,CAAC,GACvC;AACnB,QAAM,YAAY,KAAK,iBAAiB;AACxC,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,MAAyB,CAAC;AAChC,aAAW,KAAK,MAAM;AACpB,QAAI,CAAC,EAAE,WAAY;AACnB,UAAM,QAAQ,aACT,EAAE,QAAQ,gBAAgB,EAAE,QAAQ,cACpC,EAAE,QAAQ,eAAe,EAAE,QAAQ;AACxC,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,QAAI,KAAK;AAAA,MACP,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd;AAAA,MACA,MAAM,SAAS;AAAA,IACjB,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAOA,SAAS,WAAW,OAAe,MAAc,KAA2B;AAC1E,QAAM,IAAI,KAAK,IAAI,GAAG,KAAK;AAC3B,QAAM,IAAI,KAAK,IAAI,GAAG,IAAI;AAC1B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,SAAO,KAAK,IAAI;AAClB;AAEA,SAAS,YAAY,OAAe,KAA2B;AAE7D,QAAM,IAAI,QAAQ,IAAI;AACtB,QAAM,IAAI,IAAI,KAAK,KAAK,IAAI,CAAC;AAC7B,SAAO,MAAM;AACX,QAAI;AACJ,QAAI;AACJ,OAAG;AACD,YAAM,KAAK,IAAI,KAAK;AACpB,YAAM,KAAK,IAAI,KAAK;AAEpB,UAAI,KAAK,KAAK,KAAK,KAAK,IAAI,EAAE,CAAC,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK,EAAE;AAC5D,UAAI,IAAI,IAAI;AAAA,IACd,SAAS,KAAK;AACd,QAAI,IAAI,IAAI;AACZ,UAAM,IAAI,IAAI;AACd,QAAI,IAAI,IAAI,SAAS,KAAK,EAAG,QAAO,IAAI;AACxC,QAAI,KAAK,IAAI,CAAC,IAAI,MAAM,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,IAAI,CAAC,GAAI,QAAO,IAAI;AAAA,EACxE;AACF;;;ACpNA,eAAsB,mBACpB,MAC0B;AAC1B,QAAM,KAAK,KAAK,MAAM,CAAC,GAAG,GAAG,GAAG,GAAG,GAAG,EAAE;AACxC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,WAAW,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAE7C,QAAM,SAA4B,CAAC;AACnC,aAAW,KAAK,UAAU;AACxB,UAAM,cAA8C,CAAC;AACrD,UAAM,YAAsB,CAAC;AAC7B,QAAI,cAAc;AAClB,QAAI,gBAAgB;AACpB,eAAW,YAAY,KAAK,WAAW;AACrC,YAAM,MAAM,SAAS,cAAc,YAAY,KAAK,UAAU,QAAQ,QAAQ,CAAC;AAC/E,YAAM,SAAmB,CAAC;AAC1B,UAAI,SAAS;AACb,eAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,cAAM,QAAQ,MAAM,KAAK,OAAO,IAAI,EAAE,UAAU,GAAG,KAAK,EAAE,CAAC;AAC3D,eAAO,KAAK,KAAK;AACjB,YAAI,SAAS,cAAe;AAC5B,kBAAU,KAAK,KAAK;AACpB,YAAI,SAAS,cAAe;AAC5B;AAAA,MACF;AACA,YAAM,QAAQ,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACzD,kBAAY,KAAK,EAAE,YAAY,KAAK,WAAW,OAAO,QAAQ,OAAO,OAAO,OAAO,CAAC;AAAA,IACtF;AACA,UAAM,YAAY,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,UAAU,MAAM;AACrF,UAAM,WACJ,UAAU,SAAS,IACf,IACA,UAAU,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,cAAc,GAAG,CAAC,KAAK,UAAU,SAAS;AACpF,WAAO,KAAK;AAAA,MACV;AAAA,MACA;AAAA,MACA,UAAU,cAAc,KAAK,IAAI,GAAG,aAAa;AAAA,MACjD,KAAK,KAAK,KAAK,QAAQ;AAAA,MACvB,GAAG,UAAU;AAAA,MACb;AAAA,IACF,CAAC;AAAA,EACH;AAEA,QAAMA,cAAa,OAAO,KAAK,CAAC,MAAM,EAAE,YAAY,aAAa,GAAG,KAAK;AACzE,QAAM,OAAO,SAAS,SAAS,SAAS,CAAC,KAAK;AAE9C,MAAI,OAAO;AACX,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,UAAM,KAAK,OAAO,IAAI,CAAC,EAAG;AAC1B,UAAM,KAAK,OAAO,CAAC,EAAG;AACtB,UAAM,KAAK,OAAO,IAAI,CAAC,EAAG;AAC1B,UAAM,KAAK,OAAO,CAAC,EAAG;AACtB,aAAU,KAAK,MAAM,KAAM,KAAK;AAAA,EAClC;AACA,QAAM,iBAAiB,SAAS,IAAI,IAAI,OAAO;AAE/C,SAAO,EAAE,QAAQ,YAAAA,aAAY,eAAe;AAC9C;AAwBO,SAAS,wBACd,GACA,GACA,OAA4E,CAAC,GACxD;AACrB,QAAM,OAAO,KAAK,cAAc;AAChC,QAAM,YAAY,KAAK,sBAAsB;AAC7C,QAAM,MAAMC,SAAQ,KAAK,IAAI;AAE7B,QAAM,OAAoC,CAAC;AAC3C,aAAW,MAAM,EAAE,QAAQ;AACzB,UAAM,KAAK,EAAE,OAAO,KAAK,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAC5C,QAAI,CAAC,GAAI;AACT,UAAM,SAAS,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,SAAS;AACpD,UAAM,SAAS,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,SAAS;AACpD,UAAM,MAAM,gBAAgB,QAAQ,WAAW,MAAM,GAAG;AACxD,UAAM,MAAM,gBAAgB,QAAQ,WAAW,MAAM,GAAG;AACxD,SAAK,KAAK;AAAA,MACR,GAAG,GAAG;AAAA,MACN,WAAW,GAAG,YAAY,GAAG;AAAA,MAC7B,MAAM,IAAI;AAAA,MACV,OAAO,IAAI;AAAA,MACX,MAAM,IAAI;AAAA,MACV,OAAO,IAAI;AAAA,IACb,CAAC;AAAA,EACH;AAEA,QAAM,YAAY,EAAE,iBAAiB,EAAE;AACvC,QAAM,kBACJ,EAAE,eAAe,QAAQ,EAAE,eAAe,OACtC,EAAE,aAAa,EAAE,aACjB;AAIN,QAAM,YAAY,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,WAAW,CAAC,IAAI,KAAK,IAAI,GAAG,KAAK,MAAM;AACrF,MAAI;AACJ,MAAI,KAAK,IAAI,SAAS,IAAI,QAAQ,KAAK,IAAI,SAAS,IAAI,KAAM,WAAU;AAAA,WAC/D,YAAY,KAAK,YAAY,EAAG,WAAU;AAAA,WAC1C,YAAY,KAAK,YAAY,EAAG,WAAU;AAAA,MAC9C,WAAU;AAEf,QAAM,YACJ,oBAAoB,UAAU,QAAQ,CAAC,CAAC,gBAAgB,UAAU,QAAQ,CAAC,CAAC,MAC3E,oBAAoB,OAAO,wBAAwB,eAAe,KAAK;AAE1E,SAAO,EAAE,MAAM,WAAW,iBAAiB,SAAS,UAAU;AAChE;AAGO,SAAS,WAAW,OAAwB,YAAY,KAAoB;AACjF,SAAO,MAAM,OAAO,KAAK,CAAC,MAAM,EAAE,YAAY,SAAS,GAAG,KAAK;AACjE;AAIA,SAASA,SAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAEA,SAAS,gBACP,IACA,WACA,YACA,KAC+B;AAC/B,MAAI,GAAG,SAAS,EAAG,QAAO,EAAE,KAAK,GAAG,CAAC,KAAK,GAAG,MAAM,GAAG,CAAC,KAAK,EAAE;AAC9D,QAAM,UAAU,IAAI,MAAc,SAAS;AAC3C,WAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,QAAO,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC3E,YAAQ,CAAC,IAAI,MAAM,GAAG;AAAA,EACxB;AACA,UAAQ,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC5B,QAAM,QAAQ,IAAI;AAClB,SAAO;AAAA,IACL,KAAK,QAAQ,KAAK,MAAO,QAAQ,IAAK,SAAS,CAAC;AAAA,IAChD,MAAM,QAAQ,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC,CAAC;AAAA,EACnF;AACF;;;AC7JA,eAAsB,0BACpB,MACqC;AACrC,QAAM,mBAAmB,KAAK,oBAAoB;AAClD,QAAM,SAAS,KAAK,UAAU;AAC9B,QAAM,WAAW,KAAK,qBAAqB;AAC3C,QAAM,SAAS,KAAK,UAAU,OAAO;AACrC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,MAAM,WAAW,IAAI;AAE3B,QAAM,YAAsC,CAAC;AAC7C,QAAM,OAAO,oBAAI,IAAY;AAC7B,MAAI,aAAa;AAGjB,aAAW,KAAK,KAAK,OAAO;AAC1B,UAAM,KAAK,KAAK,iBAAiB,CAAC;AAClC,QAAI,KAAK,IAAI,EAAE,EAAG;AAClB,SAAK,IAAI,EAAE;AACX,QAAI,cAAc,OAAQ;AAC1B,UAAM,QAAQ,MAAM,KAAK,QAAQ,CAAC;AAClC;AACA,cAAU,KAAK;AAAA,MACb;AAAA,MACA,YAAY;AAAA,MACZ,UAAU;AAAA,MACV,UAAU;AAAA,MACV;AAAA,MACA,kBAAkB;AAAA,IACpB,CAAC;AAAA,EACH;AAGA,WAAS,IAAI,GAAG,KAAK,QAAQ,KAAK;AAChC,QAAI,cAAc,OAAQ;AAC1B,UAAM,UAAU,UAAU,OAAO,CAAC,MAAM,EAAE,eAAe,IAAI,CAAC;AAC9D,eAAW,UAAU,SAAS;AAC5B,iBAAW,YAAY,KAAK,WAAW;AACrC,YAAI,cAAc,OAAQ;AAC1B,cAAM,WAAW,MAAM,SAAS,OAAO,OAAO,UAAU,GAAG;AAC3D,cAAM,WAAW,MAAM,QAAQ,QAAQ,IAAI,WAAW,CAAC,QAAQ;AAC/D,iBAAS,IAAI,GAAG,IAAI,KAAK,IAAI,UAAU,SAAS,MAAM,GAAG,KAAK;AAC5D,cAAI,cAAc,OAAQ;AAC1B,gBAAM,QAAQ,SAAS,CAAC;AACxB,gBAAM,MAAM,KAAK,iBAAiB,KAAK;AACvC,cAAI,KAAK,IAAI,GAAG,EAAG;AACnB,eAAK,IAAI,GAAG;AACZ,gBAAM,SAAS,MAAM,KAAK,QAAQ,KAAK;AACvC;AACA,oBAAU,KAAK;AAAA,YACb,IAAI;AAAA,YACJ,YAAY;AAAA,YACZ,UAAU,OAAO;AAAA,YACjB,UAAU;AAAA,YACV,OAAO;AAAA,YACP,kBAAkB,SAAS;AAAA,UAC7B,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,EAAE,QAAQ,gBAAgB,EAC5D,KAAK,CAAC,GAAG,OAAO,EAAE,SAAS,MAAM,EAAE,SAAS,EAAE;AAEjD,QAAM,eAA2D,CAAC;AAClE,QAAM,SAAS,UAAU,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,UAAU,GAAG,CAAC;AACtE,WAAS,IAAI,GAAG,KAAK,QAAQ,KAAK;AAChC,UAAM,OAAO,UAAU,OAAO,CAAC,MAAM,EAAE,eAAe,CAAC;AACvD,QAAI,KAAK,WAAW,EAAG;AACvB,UAAM,QAAQ,KAAK,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,EAAE,QAAQ,gBAAgB,EAAE;AACjF,UAAM,YAAY,KAAK,OAAO,CAAC,KAAK,MAAM,OAAO,EAAE,SAAS,IAAI,CAAC,IAAI,KAAK;AAC1E,iBAAa,KAAK,EAAE,YAAY,GAAG,OAAO,KAAK,QAAQ,UAAU,OAAO,UAAU,CAAC;AAAA,EACrF;AAEA,SAAO,EAAE,WAAW,UAAU,cAAc,WAAW;AACzD;AAIA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;;;AClHA,eAAsB,UACpB,SACA,SACyB;AACzB,QAAM,MAAsB,CAAC;AAC7B,aAAW,KAAK,SAAS;AACvB,UAAM,CAAC,QAAQ,QAAQ,QAAQ,IAAI,MAAM,QAAQ,IAAI;AAAA,MACnD,QAAQ,QAAQ,QAAQ,SAAS,EAAE,WAAW,CAAC;AAAA,MAC/C,QAAQ,QAAQ,QAAQ,aAAa,EAAE,WAAW,CAAC;AAAA,MACnD,QAAQ,QAAQ,QAAQ,aAAa,EAAE,aAAa,CAAC;AAAA,IACvD,CAAC;AACD,QAAI,KAAK;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ,EAAE;AAAA,MACV,MAAM;AAAA,QACJ,YAAY,EAAE;AAAA,QACd,iBAAiB,EAAE;AAAA,QACnB,mBAAmB,EAAE;AAAA,QACrB,aAAa,EAAE;AAAA,QACf,eAAe,EAAE;AAAA,QACjB,aAAa,EAAE,KAAK;AAAA,QACpB,eAAe,EAAE,KAAK;AAAA,MACxB;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAGO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AA4BA,eAAsB,WACpB,MACA,SAC0B;AAC1B,QAAM,WAAW,QAAQ,YAAY;AACrC,QAAM,UAAU,oBAAI,IAAyB;AAC7C,aAAW,KAAK,MAAM;AACpB,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,MAAM,QAAQ,IAAI,GAAG,KAAK,CAAC;AACjC,QAAI,KAAK,CAAC;AACV,YAAQ,IAAI,KAAK,GAAG;AAAA,EACtB;AAEA,QAAM,OAAwB,CAAC;AAC/B,aAAW,CAAC,YAAY,KAAK,KAAK,QAAQ,QAAQ,GAAG;AACnD,QAAI,MAAM,WAAW,EAAG;AAExB,UAAM,SAAS,MAAM,QAAQ,QAAQ,QAAQ,SAAS,MAAM,CAAC,EAAG,KAAK,CAAC;AACtE,UAAM,cAAwB,CAAC;AAC/B,UAAM,UAAoB,CAAC;AAC3B,UAAM,SAAmB,CAAC;AAC1B,eAAW,KAAK,OAAO;AACrB,YAAM,SAAS,SAAS,CAAC;AACzB,UAAI,WAAW,KAAM;AACrB,YAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,aAAa,EAAE,KAAK,CAAC;AACtE,kBAAY,KAAK,UAAU;AAC3B,cAAQ,KAAK,MAAM;AACnB,aAAO,KAAK,EAAE,KAAK;AAAA,IACrB;AACA,QAAI,YAAY,WAAW,EAAG;AAC9B,SAAK,KAAK;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,MAAM;AAAA,QACJ;AAAA,QACA,GAAG,YAAY;AAAA,QACf,YAAY,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,QAAQ;AAAA,MAC3D;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,YAAY,MAA+B;AACzD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAwBA,eAAsB,UAAU,MAAmB,SAA8C;AAC/F,QAAM,UAAU,QAAQ,YAAY,MAAM;AAC1C,QAAM,OAAuB,CAAC;AAC9B,aAAW,KAAK,MAAM;AACpB,QAAI,CAAC,QAAQ,CAAC,EAAG;AACjB,UAAM,SAAS,QAAQ,WAAW,CAAC;AACnC,UAAM,CAAC,QAAQ,UAAU,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC7C,QAAQ,QAAQ,QAAQ,SAAS,EAAE,KAAK,CAAC;AAAA,MACzC,QAAQ,QAAQ,QAAQ,aAAa,EAAE,KAAK,CAAC;AAAA,IAC/C,CAAC;AACD,UAAM,WAAqC,CAAC;AAC5C,QAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,OAAO,CAAC;AAC7D,aAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAC/C,aAAS,KAAK,EAAE,MAAM,aAAa,SAAS,WAAW,CAAC;AACxD,SAAK,KAAK;AAAA,MACR;AAAA,MACA,MAAM;AAAA,QACJ,OAAO,EAAE;AAAA,QACT,aAAa,EAAE;AAAA,QACf,YAAY,EAAE;AAAA,QACd,OAAO,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAAA,QAC3C,OAAO,EAAE;AAAA,MACX;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AA8BA,eAAsB,UACpB,SACA,SACyB;AACzB,QAAM,OAAuB,CAAC;AAC9B,aAAW,KAAK,SAAS;AACvB,UAAM,SAAS,MAAM,QAAQ,QAAQ,QAAQ,SAAS,EAAE,WAAW,CAAC;AACpE,UAAM,gBAAgB,QAAQ,WAC1B,MAAM,QAAQ,QAAQ,QAAQ,SAAS,EAAE,aAAa,EAAE,eAAe,CAAC,IACxE,CAAC;AACL,UAAM,iBAA2B,CAAC;AAClC,eAAW,UAAU,eAAe;AAClC,qBAAe,KAAK,MAAM,QAAQ,QAAQ,QAAQ,WAAW,EAAE,aAAa,MAAM,CAAC,CAAC;AAAA,IACtF;AACA,UAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,WAAW,EAAE,aAAa,EAAE,YAAY,CAAC;AAC1F,UAAM,eAAe,MAAM,QAAQ;AAAA,MACjC,QAAQ,WAAW,EAAE,eAAe,EAAE,cAAc;AAAA,IACtD;AACA,SAAK,KAAK;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,cAAc,EAAE;AAAA,MAChB,gBAAgB,EAAE;AAAA,MAClB,aAAa,EAAE;AAAA,MACf,MAAM;AAAA,QACJ,aAAa,EAAE;AAAA,QACf,eAAe,EAAE;AAAA,QACjB,iBAAiB,EAAE;AAAA,MACrB;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAaO,SAAS,mBAAmB,aAAmC;AACpE,QAAM,OAA6B,YAAY,IAAI,CAAC,OAAO;AAAA,IACzD,OAAO,EAAE;AAAA,IACT,QAAQ,EAAE;AAAA,IACV,WAAW,EAAE;AAAA,IACb,QAAQ,EAAE;AAAA,IACV,aAAa,EAAE;AAAA,IACf,QAAQ,EAAE,UAAU;AAAA,EACtB,EAAE;AACF,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAIA,SAAS,cAAc,KAA+B;AACpD,QAAM,IAAI,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AAClD,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;;;ACvQO,IAAM,+BAAN,MAAyD;AAAA,EACtD;AAAA,EACA,aAAoD;AAAA,EAE5D,YAAY,MAA2C;AACrD,SAAK,OAAO;AAAA,EACd;AAAA,EAEA,MAAM,gBAAgB,MAA2C;AAC/D,UAAM,YAAY,KAAK,KAAK,oBAAoB;AAChD,UAAM,WAA0B,CAAC;AACjC,UAAM,cAAc,KAAK,OAAO,CAAC,MAAM;AACrC,YAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,aAAO,OAAO,UAAU,YAAY,QAAQ;AAAA,IAC9C,CAAC;AACD,QAAI,YAAY,WAAW,EAAG,QAAO;AAIrC,UAAM,UAAU,oBAAI,IAAyB;AAC7C,eAAW,KAAK,aAAa;AAC3B,YAAM,MAAM,QAAQ,IAAI,EAAE,WAAW,KAAK,CAAC;AAC3C,UAAI,KAAK,CAAC;AACV,cAAQ,IAAI,EAAE,aAAa,GAAG;AAAA,IAChC;AAEA,eAAW,CAAC,aAAa,KAAK,KAAK,QAAQ,QAAQ,GAAG;AACpD,YAAM,YACJ,MAAM,OAAO,CAAC,GAAG,MAAM;AACrB,cAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ,eAAe;AAC7D,eAAO,IAAI;AAAA,MACb,GAAG,CAAC,IAAI,MAAM;AAChB,eAAS,KAAK;AAAA,QACZ,MAAM,aAAa,WAAW;AAAA,QAC9B,aAAa,GAAG,WAAW,aAAa,SAAS,OAAO,MAAM,MAAM,iBAAiB,UAAU,QAAQ,CAAC,CAAC;AAAA,QACzG,UAAU;AAAA,UACR,QAAQ,MAAM,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,UAC5C,SAAS,MAAM;AAAA,QACjB;AAAA,MACF,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,cAAc,UAAoD;AACtE,QAAI,SAAS,WAAW,EAAG,QAAO,CAAC;AAInC,QAAI,KAAK,eAAe,MAAM;AAC5B,aAAO;AAAA,QACL;AAAA,UACE,MAAM;AAAA,UACN,SAAS,EAAE,WAAW,mCAAmC;AAAA,UACzD,WACE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAEA,UAAM,sBAAsB,KAAK,KAAK,uBAAuB;AAC7D,UAAM,UAA4B,CAAC;AAEnC,eAAW,WAAW,KAAK,WAAW,QAAQ;AAC5C,UAAI,QAAQ,YAAY,eAAgB;AACxC,UAAI,KAAK,IAAI,QAAQ,QAAQ,KAAK,oBAAqB;AACvD,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,SAAS;AAAA,UACP,QAAQ,QAAQ;AAAA,UAChB,QAAQ;AAAA,UACR,UAAU,QAAQ;AAAA,UAClB,aAAa,QAAQ;AAAA,QACvB;AAAA,QACA,WAAW,gCAAgC,QAAQ,SAAS,QAAQ,CAAC,CAAC,OAAO,QAAQ,WAAW;AAAA,QAChG,eAAe,CAAC,KAAK,IAAI,GAAG,OAAO,KAAK,IAAI,QAAQ,QAAQ,CAAC;AAAA,MAC/D,CAAC;AAAA,IACH;AACA,eAAW,WAAW,KAAK,WAAW,OAAO,MAAM,GAAG,CAAC,GAAG;AACxD,UAAI,QAAQ,YAAY,eAAgB;AACxC,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,SAAS;AAAA,UACP,QAAQ,QAAQ;AAAA,UAChB,QAAQ;AAAA,UACR,UAAU,QAAQ;AAAA,UAClB,aAAa,QAAQ;AAAA,QACvB;AAAA,QACA,WAAW,gCAAgC,QAAQ,SAAS,QAAQ,CAAC,CAAC,OAAO,QAAQ,WAAW;AAAA,QAChG,eAAe,KAAK,IAAI,GAAG,KAAK,IAAI,QAAQ,QAAQ,IAAI,GAAG,IAAI;AAAA,MACjE,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,YAAY,SAA2B,UAAmD;AAG9F,WAAO;AAAA,MACL,GAAG;AAAA,MACH,SAAS,CAAC,GAAG,SAAS,SAAS,GAAG,OAAO;AAAA,IAC3C;AAAA,EACF;AAAA,EAEA,MAAM,eAAe,MAAiD;AAKpE,UAAM,YAA0B;AAAA,MAC9B,SAAS;AAAA,MACT,aAAa,KAAK;AAAA,MAClB,YAAY,KAAK;AAAA,MACjB,UAAU;AAAA,QACR,gBAAgB;AAAA,QAChB,mBAAmB;AAAA,QACnB,UAAU,EAAE,KAAK,GAAG,MAAM,EAAE;AAAA,QAC5B,cAAc;AAAA,QACd,aAAa;AAAA,QACb,cAAc;AAAA,QACd,YAAY;AAAA,QACZ,oBAAoB;AAAA,QACpB,qBAAqB,OAAO;AAAA,QAC5B,oBAAoB,OAAO;AAAA,MAC7B;AAAA,MACA,QACE;AAAA,MACF,eAAe;AAAA,IACjB;AACA,WAAO;AAAA,MACL;AAAA,MACA,MAAM,CAAC;AAAA,MACP,cAAc;AAAA,IAChB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBAAiB,MAA4D;AACjF,UAAM,SAAS,MAAM,yBAAyB;AAAA,MAC5C;AAAA,MACA,UAAU,KAAK,KAAK;AAAA,MACpB,gBAAgB,KAAK,KAAK;AAAA,MAC1B,SAAS,KAAK,KAAK;AAAA,IACrB,CAAC;AACD,QAAI,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,SAAS,MAAM;AACvD,SAAK,aAAa;AAClB,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,UAAU,QAA8C;AACtD,SAAK,aAAa;AAAA,EACpB;AAAA,EAEA,gBAAuD;AACrD,WAAO,KAAK;AAAA,EACd;AACF;;;ACjJA,eAAsB,mBACpB,OACA,OACA,MACuB;AACvB,QAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;AACzC,QAAM,UAAU,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACnE,QAAM,MAAoB,CAAC;AAC3B,MAAI,MAAM;AACV,aAAW,QAAQ,SAAS;AAC1B,QAAI,KAAK,aAAa,CAAC,KAAK,UAAU,IAAI,EAAG;AAC7C,QAAI,SAAmD;AACvD,eAAW,KAAK,KAAK,SAAS;AAC5B,UAAI,CAAC,EAAE,UAAU,SAAS,KAAK,IAAI,EAAG;AACtC,YAAM,IAAI,MAAM,EAAE,MAAM,IAAI;AAC5B,UAAI,GAAG;AACL,iBAAS;AACT;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,OAAQ;AACb,QAAI,KAAK;AAAA,MACP,QAAQ,KAAK;AAAA,MACb;AAAA,MACA,WAAW;AAAA,MACX,MAAM,KAAK;AAAA,MACX,MAAM,KAAK;AAAA,MACX,QAAQ,OAAO;AAAA,MACf,aAAa,OAAO;AAAA,MACpB,WAAW,OAAO;AAAA,MAClB,QAAQ,OAAO;AAAA,IACjB,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAeO,SAAS,yBAAyB,aAA+C;AACtF,MAAI,YAAY,WAAW,GAAG;AAC5B,WAAO;AAAA,MACL,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,mBAAmB;AAAA,MACnB,iBAAiB;AAAA,MACjB,gBAAgB;AAAA,MAChB,gBAAgB;AAAA,IAClB;AAAA,EACF;AACA,QAAM,QAAQ,YAAY,CAAC,EAAG;AAC9B,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,aAAa;AACjB,MAAI,WAA0B;AAC9B,MAAI,OAAO,YAAY,CAAC,EAAG;AAC3B,WAAS,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;AAC3C,UAAM,IAAI,YAAY,CAAC;AACvB,UAAM,IAAI,EAAE,UAAU;AACtB,YAAQ;AACR,aAAS,IAAI,EAAE;AACf,QAAI,EAAE,SAAS,IAAK;AACpB,QAAI,IAAI,GAAG;AACT,YAAM,QAAQ,EAAE,SAAS;AACzB,UAAI,QAAQ,YAAY;AACtB,qBAAa;AACb,mBAAW;AAAA,MACb;AACA,aAAO,EAAE;AAAA,IACX,OAAO;AACL,aAAO,EAAE;AAAA,IACX;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,YAAY;AAAA,IACxB,YAAY,SAAS,IAAI,IAAI,QAAQ;AAAA,IACrC,mBAAmB;AAAA,IACnB,iBAAiB,WAAW,YAAY;AAAA,IACxC,gBAAgB;AAAA,IAChB,gBAAgB;AAAA,EAClB;AACF;AAgCO,SAAS,iBACd,kBACA,OAAyD,CAAC,GACrC;AACrB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,YAAY,KAAK,mBAAmB;AAC1C,QAAM,OAAO,CAAC,GAAG,iBAAiB,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,EAAE,OAAO,MAAM,EAAE;AACvF,QAAM,UAA+B,CAAC;AAEtC,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,aAAS,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACxC,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;AACtD,UAAI,SAAS,YAAY,EAAG;AAO5B,UAAI,gBAAgB;AACpB,eAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,cAAM,KAAK,EAAE,MAAM,CAAC;AACpB,cAAM,KAAK,EAAE,MAAM,CAAC;AACpB,cAAM,uBAAuB,GAAG,SAAS,GAAG,QAAQ,GAAG,SAAS,GAAG;AACnE,cAAM,YAAY,KAAK,IAAI,GAAG,SAAS,GAAG,MAAM;AAChD,YAAI,wBAAwB,aAAa,WAAW;AAClD,0BAAgB;AAChB;AAAA,QACF;AAAA,MACF;AACA,UAAI,gBAAgB,EAAG;AACvB,UAAI,gBAAgB,UAAW;AAE/B,YAAM,QAAQ,EAAE,MAAM,aAAa;AACnC,YAAM,QAAQ,EAAE,MAAM,aAAa;AACnC,YAAM,SAAS,KAAK,IAAI,MAAM,SAAS,MAAM,MAAM;AACnD,UAAI,SAAS,UAAW;AAExB,YAAM,SAAS,MAAM,SAAS,MAAM,SAAS,QAAQ;AACrD,YAAM,WAAW,MAAM,SAAS,MAAM,SAAS,QAAQ;AACvD,YAAM,YAAY,MAAM,SAAS,MAAM,SAAS,EAAE,QAAQ,EAAE;AAC5D,YAAM,cAAc,MAAM,SAAS,MAAM,SAAS,EAAE,QAAQ,EAAE;AAC9D,cAAQ,KAAK;AAAA,QACX,aAAa;AAAA,QACb,iBAAiB,gBAAgB;AAAA,QACjC,cAAc,OAAO;AAAA,QACrB,cAAc,OAAO;AAAA,QACrB,gBAAgB,SAAS;AAAA,QACzB,gBAAgB,SAAS;AAAA,QACzB,eAAe;AAAA,QACf,aAAa,OAAO,SAAS,SAAS;AAAA,MACxC,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;;;AChKA,eAAsB,cACpB,MAC8B;AAE9B,QAAM,WAAW,MAAM,gBAAgB,IAAI;AAG3C,QAAM,gBAAgB;AAAA,IACpB,SAAS;AAAA,IACT,KAAK,oBAAoB,CAAC;AAAA,EAC5B;AAGA,QAAM,cAAc,mBAAmB,SAAS,MAAM;AAAA,IACpD,UAAU,KAAK,aAAa,YAAY;AAAA,IACxC,WAAW,KAAK,aAAa,aAAa;AAAA,IAC1C,UAAU,KAAK,aAAa,YAAY,KAAK,YAAY;AAAA,IACzD,UAAU,KAAK,aAAa;AAAA,EAC9B,CAAC;AAGD,MAAI,oBAAqD;AACzD,MAAI,KAAK,QAAQ,YAAY;AAC3B,UAAM,aAAa,KAAK,OAAO;AAC/B,UAAM,cAAc,yBAAyB,SAAS,MAAM,UAAU;AACtE,QAAI,YAAY,KAAK,CAAC,MAAM,EAAE,OAAO,SAAS,CAAC,GAAG;AAChD,0BAAoB,iCAAiC;AAAA,QACnD;AAAA,QACA,OAAO,KAAK,YAAY;AAAA,QACxB,OAAO,KAAK,YAAY;AAAA,QACxB,MAAM,KAAK,YAAY,QAAQ,KAAK,QAAQ;AAAA,MAC9C,CAAC;AAAA,IACH;AAAA,EACF;AAGA,QAAM,gBAAgB,oBAAoB;AAAA,IACxC,MAAM,SAAS;AAAA,IACf,yBAAyB,KAAK;AAAA,EAChC,CAAC;AAGD,MAAI,qBAA4D;AAChE,MAAI,KAAK,gBAAgB,KAAK,kBAAkB,KAAK,eAAe,SAAS,GAAG;AAC9E,yBAAqB,MAAM,yBAAyB;AAAA,MAClD,MAAM,SAAS;AAAA,MACf,UAAU,KAAK;AAAA,MACf,gBAAgB,KAAK;AAAA,IACvB,CAAC;AAAA,EACH;AAGA,QAAM,cAAkD,CAAC;AACzD,MAAI,KAAK,eAAe,KAAK;AAC3B,gBAAY,MAAM,MAAM,UAAU,YAAY,OAAO,KAAK,cAAc,GAAG;AAAA,EAC7E;AACA,MAAI,KAAK,eAAe,MAAM;AAC5B,gBAAY,OAAO,MAAM,WAAW,SAAS,MAAM,KAAK,cAAc,IAAI;AAAA,EAC5E;AACA,MAAI,KAAK,eAAe,KAAK;AAC3B,gBAAY,MAAM,MAAM,UAAU,SAAS,MAAM,KAAK,cAAc,GAAG;AAAA,EACzE;AAEA,QAAM,UAAU,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM;AAAA,EACR;AACF;AAIA,SAAS,yBACP,MACA,YACkD;AAGlD,QAAM,WAAW,oBAAI,IAAoB;AACzC,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,gBAAgB,WAAY;AAClC,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,aAAS,IAAI,GAAG,GAAG,KAAK,EAAE,IAAI,IAAI,KAAK;AAAA,EACzC;AACA,QAAM,cAAc,oBAAI,IAAsB;AAC9C,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,gBAAgB,WAAY;AAClC,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,UAAM,YAAY,SAAS,IAAI,GAAG,GAAG,KAAK,EAAE,IAAI,EAAE;AAClD,QAAI,OAAO,cAAc,SAAU;AACnC,UAAM,MAAM,YAAY,IAAI,EAAE,WAAW,KAAK,CAAC;AAC/C,QAAI,KAAK,QAAQ,SAAS;AAC1B,gBAAY,IAAI,EAAE,aAAa,GAAG;AAAA,EACpC;AACA,SAAO,CAAC,GAAG,YAAY,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,aAAa,MAAM,OAAO,EAAE,aAAa,OAAO,EAAE;AAC5F;AAEA,SAAS,aAAa,MAMX;AACT,QAAM,IAAI,KAAK;AACf,QAAM,QAAQ;AAAA,IACZ,GAAG,EAAE,UAAU,KAAK,EAAE,KAAK,MAAM,sBAAsB,EAAE,WAAW,MAAM,wBAAwB,EAAE,oBAAoB,MAAM,GAAG,EAAE,CAAC;AAAA,IACpI,gBAAgB,KAAK,YAAY,MAAM,MAAM,KAAK,KAAK,YAAY,QAAQ,KAAK,KAAK,YAAY,gBAAgB;AAAA,EACnH;AACA,MAAI,KAAK,mBAAmB;AAC1B,UAAM;AAAA,MACJ,uBAAuB,KAAK,kBAAkB,eAAe,QAAQ,MAClE,KAAK,kBAAkB,eAAe,cACnC,IAAI,KAAK,kBAAkB,eAAe,WAAW,KACrD;AAAA,IACR;AAAA,EACF;AACA,QAAM;AAAA,IACJ,mBAAmB,KAAK,cAAc,OAAO,KAAK,KAAK,cAAc,SAAS,MAAM;AAAA,EACtF;AACA,MAAI,KAAK,oBAAoB;AAC3B,UAAM,MAAM,KAAK,mBAAmB,OAAO,CAAC;AAC5C,UAAM;AAAA,MACJ,eAAe,KAAK,UAAU,MAAM,YAAO,KAAK,YAAY,GAAG,QAAQ,CAAC,CAAC,KAAK,KAAK,WAAW,SAAS;AAAA,IACzG;AAAA,EACF;AACA,SAAO,MAAM,KAAK,KAAK;AACzB;","names":["firstPassK","makeRng"]}
+{"version":3,"sources":["../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/off-policy.ts","../src/rl/preferences.ts","../src/rl/run-record-adapters.ts","../src/rl/tournament.ts","../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/adversarial.ts","../src/rl/corpus.ts","../src/rl/exporters.ts","../src/rl/dataset.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/process-reward.ts","../src/rl/rl-campaign.ts"],"sourcesContent":["/**\n * Test-time compute scaling curves.\n *\n * The test-time-compute frontier paper (Snell et al. 2024) and the\n * subsequent o1-style scaling work both show that LLM-agent capability\n * is a function of the compute budget at inference, not just of the\n * training run. The right way to characterize a candidate is therefore\n * a *curve* — score at compute budgets {1×, 4×, 16×, …} — not a single\n * point.\n *\n * This module ships:\n *\n *   1. The compute-curve harness — `runComputeCurve(runner, budgets)` —\n *      that evaluates one candidate at a sequence of compute budgets\n *      and returns the (compute, score) curve.\n *   2. A best-of-N evaluator — `bestOfN(runner, n, scoreFn)` — the\n *      simplest test-time-compute scaling primitive: sample N\n *      independent rollouts, return the best.\n *   3. A self-consistency evaluator — `selfConsistency(runner, n)` —\n *      the majority-vote variant of best-of-N for tasks with a small\n *      categorical answer space.\n *   4. Pareto-frontier extraction over multiple candidates — given\n *      (candidate, compute, score) tuples, return the set of\n *      candidate-compute combinations that aren't dominated.\n *\n * Caveat: \"compute\" here is the caller's notion of a compute unit. For\n * agent eval that's typically wall-time × parallelism, or token budget,\n * or LLM-call count. We accept whatever the caller provides; the curve\n * is on whatever axis they pick.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface ComputeCurveBudget {\n  /** Identifier — for the report. Common: '1x', '4x', '16x'. */\n  id: string\n  /** Numeric value on the chosen axis (tokens, calls, USD, ms — caller picks). */\n  cost: number\n  /** Free-form metadata (the caller can carry per-budget config). */\n  meta?: Record<string, unknown>\n}\n\nexport interface ComputeCurvePoint {\n  budgetId: string\n  cost: number\n  score: number\n  /** Number of underlying samples used at this budget. */\n  samples: number\n  /** Optional spread / variance information. */\n  std?: number\n  /** Any extra metrics the runner returned. */\n  metrics?: Record<string, number>\n}\n\nexport interface ComputeCurve {\n  candidateId: string\n  points: ComputeCurvePoint[]\n  /** Rough exponent fit: score ≈ a + b * log(cost). Useful for \"how steep is the curve?\" */\n  logSlope: number | null\n  /** Best (highest-score) point on the curve. */\n  best: ComputeCurvePoint\n}\n\nexport interface RunComputeCurveOptions {\n  candidateId: string\n  budgets: ComputeCurveBudget[]\n  /**\n   * Run the candidate at one budget. Returns the realized score plus\n   * optional spread + extra metrics.\n   */\n  runAtBudget: (budget: ComputeCurveBudget) => Promise<{\n    score: number\n    samples: number\n    std?: number\n    metrics?: Record<string, number>\n  }>\n}\n\nexport async function runComputeCurve(opts: RunComputeCurveOptions): Promise<ComputeCurve> {\n  const points: ComputeCurvePoint[] = []\n  for (const budget of opts.budgets) {\n    const r = await opts.runAtBudget(budget)\n    points.push({\n      budgetId: budget.id,\n      cost: budget.cost,\n      score: r.score,\n      samples: r.samples,\n      std: r.std,\n      metrics: r.metrics,\n    })\n  }\n  const sorted = [...points].sort((a, b) => a.cost - b.cost)\n  const logSlope = sorted.length >= 2 ? fitLogSlope(sorted) : null\n  const best = points.reduce((a, b) => (b.score > a.score ? b : a))\n  return { candidateId: opts.candidateId, points: sorted, logSlope, best }\n}\n\nexport interface ComputeBestOfNOptions<O> {\n  /** Number of independent samples to draw. */\n  n: number\n  /** Sampler — produces one rollout. */\n  sample: (sampleIdx: number) => Promise<O>\n  /** Score one rollout. */\n  scoreFn: (rollout: O) => Promise<number> | number\n}\n\nexport interface ComputeBestOfNResult<O> {\n  best: O\n  bestScore: number\n  scores: number[]\n  meanScore: number\n  /** Index of the best rollout, for diagnostics. */\n  bestIndex: number\n}\n\n/** The simplest test-time scaling primitive. */\nexport async function bestOfN<O>(opts: ComputeBestOfNOptions<O>): Promise<ComputeBestOfNResult<O>> {\n  if (opts.n <= 0) throw new ValidationError('bestOfN: n must be > 0')\n  const rollouts: O[] = []\n  const scores: number[] = []\n  for (let i = 0; i < opts.n; i++) {\n    const r = await opts.sample(i)\n    rollouts.push(r)\n    scores.push(await opts.scoreFn(r))\n  }\n  let bestIndex = 0\n  for (let i = 1; i < scores.length; i++) if (scores[i]! > scores[bestIndex]!) bestIndex = i\n  const meanScore = scores.reduce((s, x) => s + x, 0) / scores.length\n  return {\n    best: rollouts[bestIndex]!,\n    bestScore: scores[bestIndex]!,\n    scores,\n    meanScore,\n    bestIndex,\n  }\n}\n\nexport interface SelfConsistencyOptions<O> {\n  n: number\n  sample: (sampleIdx: number) => Promise<O>\n  /** Extract the canonical answer key (string) from a rollout. */\n  answerKey: (rollout: O) => string\n}\n\nexport interface SelfConsistencyResult<O> {\n  /** Modal answer (the majority vote). */\n  answer: string\n  /** Fraction of samples voting for the modal answer in [0, 1]. */\n  agreement: number\n  /** Histogram of all answers. */\n  histogram: Record<string, number>\n  /** A representative rollout that voted for the modal answer. */\n  representative: O\n  /** All rollouts. */\n  rollouts: O[]\n}\n\n/**\n * Self-consistency / majority-vote test-time scaling. For tasks with a\n * small categorical answer space (math problems, multiple choice).\n */\nexport async function selfConsistency<O>(\n  opts: SelfConsistencyOptions<O>,\n): Promise<SelfConsistencyResult<O>> {\n  if (opts.n <= 0) throw new ValidationError('selfConsistency: n must be > 0')\n  const rollouts: O[] = []\n  const histogram: Record<string, number> = {}\n  for (let i = 0; i < opts.n; i++) {\n    const r = await opts.sample(i)\n    rollouts.push(r)\n    const key = opts.answerKey(r)\n    histogram[key] = (histogram[key] ?? 0) + 1\n  }\n  let answer = ''\n  let max = -1\n  for (const [k, v] of Object.entries(histogram)) {\n    if (v > max) {\n      max = v\n      answer = k\n    }\n  }\n  const representative = rollouts.find((r) => opts.answerKey(r) === answer) ?? rollouts[0]!\n  return {\n    answer,\n    agreement: max / opts.n,\n    histogram,\n    representative,\n    rollouts,\n  }\n}\n\n/**\n * Pareto frontier over (candidate, compute, score) tuples. A point is on\n * the frontier iff no other point dominates it in both score (higher\n * better) and cost (lower better). Returns the frontier sorted ascending\n * by cost.\n */\nexport interface ParetoPointInput {\n  candidateId: string\n  budgetId: string\n  cost: number\n  score: number\n}\n\nexport function paretoFrontier(points: ParetoPointInput[]): ParetoPointInput[] {\n  const onFrontier: ParetoPointInput[] = []\n  for (const p of points) {\n    const dominated = points.some(\n      (q) =>\n        q !== p && q.cost <= p.cost && q.score >= p.score && (q.cost < p.cost || q.score > p.score),\n    )\n    if (!dominated) onFrontier.push(p)\n  }\n  return onFrontier.sort((a, b) => a.cost - b.cost)\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction fitLogSlope(points: ComputeCurvePoint[]): number {\n  // OLS slope of score on log(cost). Used as a single-number summary of\n  // how much marginal compute helps. Positive = score improves with\n  // compute; near-zero = capability ceiling reached.\n  const xs = points.map((p) => Math.log(Math.max(1e-12, p.cost)))\n  const ys = points.map((p) => p.score)\n  const n = xs.length\n  const mx = xs.reduce((s, x) => s + x, 0) / n\n  const my = ys.reduce((s, y) => s + y, 0) / n\n  let num = 0\n  let den = 0\n  for (let i = 0; i < n; i++) {\n    num += (xs[i]! - mx) * (ys[i]! - my)\n    den += (xs[i]! - mx) ** 2\n  }\n  return den === 0 ? 0 : num / den\n}\n","/**\n * Contamination probe — held-out perturbation tests.\n *\n * The bug class: once a benchmark scenario set is published, models train\n * on it, and your scores become invalid. SWE-Bench-Verified, GPQA, and\n * MMLU-Pro all exist because their predecessors got contaminated within\n * months. The right defense is to keep a held-out *perturbed* version of\n * every scenario — same task, slightly different surface — and check\n * whether scores diverge significantly. Genuine capability transfers; rote\n * memorization doesn't.\n *\n * This module ships the probe contract:\n *\n *   1. A `ScenarioPerturbation` strategy type — function that produces a\n *      perturbed scenario from an original.\n *   2. `runContaminationProbe({ originals, perturbed, scoreFn })` — runs\n *      both halves and reports per-scenario score divergence + a global\n *      contamination verdict via paired Wilcoxon.\n *   3. Several stock perturbations: `renameVariables`, `shuffleOrder`,\n *      `paraphrasePrompt`, `injectIrrelevantClause`. Each preserves the\n *      task's structural difficulty while breaking surface memorization.\n *\n * The verdict is conservative: if the perturbed-vs-original score\n * difference is statistically significant (BH-adjusted p < 0.05) AND\n * the median drop is > 5 percentage points, we flag *contamination\n * suspected*. False positives are possible (the perturbation might\n * actually be harder); the default is to flag for review, not to\n * autoreject.\n */\n\nimport { ValidationError } from '../errors'\nimport { benjaminiHochberg, wilcoxonSignedRank } from '../statistics'\n\nexport type ScenarioPerturbationKind =\n  | 'rename_variables'\n  | 'shuffle_order'\n  | 'paraphrase'\n  | 'inject_irrelevant_clause'\n  | 'custom'\n\nexport interface ScenarioPerturbation<S> {\n  kind: ScenarioPerturbationKind\n  /** Apply to one scenario, return its perturbed sibling. */\n  apply: (scenario: S) => Promise<S> | S\n  /** Optional id — for the report. */\n  id?: string\n}\n\nexport interface ContaminationProbeInput<S> {\n  /** Identity of every scenario. The probe's `runFingerprint` keys on these. */\n  scenarioId: (s: S) => string\n  /** Original scenarios. */\n  originals: S[]\n  /**\n   * Either pre-computed perturbations (one per original, same order) OR a\n   * `perturbation` strategy that synthesizes them on the fly.\n   */\n  perturbed?: S[]\n  perturbation?: ScenarioPerturbation<S>\n  /**\n   * Run the policy/agent against one scenario and return a scalar score\n   * in [0, 1]. The probe doesn't care what the policy is — that's the\n   * caller's contract.\n   */\n  scoreFn: (s: S) => Promise<number>\n}\n\nexport interface ContaminationProbeOptions {\n  /** Drop scores below this from the probe; treats partial failures separately. Default 0. */\n  scoreFloor?: number\n  /**\n   * BH-FDR threshold for declaring contamination on each per-scenario\n   * delta. Default 0.05.\n   */\n  fdr?: number\n  /**\n   * Minimum median per-scenario drop to flag global contamination. Default\n   * 0.05 (5 percentage points). Smaller drops may be noise.\n   */\n  minMedianDrop?: number\n}\n\nexport interface ContaminationProbeReport {\n  perScenario: Array<{\n    scenarioId: string\n    originalScore: number\n    perturbedScore: number\n    delta: number // perturbed - original (negative = drop)\n    /** Per-scenario q-value (single-test BH for a single scenario). Mainly for display. */\n    qValue: number\n  }>\n  /** Wilcoxon paired-test on the deltas. */\n  pairedTest: { w: number; p: number }\n  medianDelta: number\n  meanDelta: number\n  contaminationSuspected: boolean\n  reason: string\n  /** Number of scenarios processed. */\n  n: number\n}\n\nexport async function runContaminationProbe<S>(\n  input: ContaminationProbeInput<S>,\n  opts: ContaminationProbeOptions = {},\n): Promise<ContaminationProbeReport> {\n  const fdr = opts.fdr ?? 0.05\n  const minMedianDrop = opts.minMedianDrop ?? 0.05\n  const floor = opts.scoreFloor ?? 0\n\n  if (!input.perturbed && !input.perturbation) {\n    throw new ValidationError(\n      'runContaminationProbe: must supply either `perturbed` or `perturbation`.',\n    )\n  }\n  const perturbed: S[] =\n    input.perturbed ?? (await Promise.all(input.originals.map((s) => input.perturbation!.apply(s))))\n  if (perturbed.length !== input.originals.length) {\n    throw new ValidationError(\n      `runContaminationProbe: perturbed length ${perturbed.length} ≠ originals ${input.originals.length}`,\n    )\n  }\n\n  // Score both halves.\n  const origScores = await Promise.all(input.originals.map((s) => input.scoreFn(s)))\n  const pertScores = await Promise.all(perturbed.map((s) => input.scoreFn(s)))\n\n  const perScenario = input.originals.map((s, i) => ({\n    scenarioId: input.scenarioId(s),\n    originalScore: origScores[i]!,\n    perturbedScore: pertScores[i]!,\n    delta: pertScores[i]! - origScores[i]!,\n    qValue: NaN,\n  }))\n\n  // Drop scenarios below the floor (partial failures we don't trust).\n  const valid = perScenario.filter((p) => p.originalScore >= floor && p.perturbedScore >= floor)\n  if (valid.length < 4) {\n    return {\n      perScenario,\n      pairedTest: { w: 0, p: 1 },\n      medianDelta: 0,\n      meanDelta: 0,\n      contaminationSuspected: false,\n      reason: `insufficient valid scenarios (n=${valid.length}, need ≥ 4)`,\n      n: valid.length,\n    }\n  }\n\n  const origValid = valid.map((p) => p.originalScore)\n  const pertValid = valid.map((p) => p.perturbedScore)\n  const pairedTest = wilcoxonSignedRank(origValid, pertValid)\n  const deltas = valid.map((p) => p.delta)\n  const sortedDeltas = [...deltas].sort((a, b) => a - b)\n  const median = sortedDeltas[Math.floor(sortedDeltas.length / 2)]!\n  const mean = deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n  // Per-scenario q-values via BH on a synthetic per-scenario p-value\n  // (one-sample bootstrap; we use the absolute delta normalized by median\n  // as a coarse signal — this is a display aid, the load-bearing test\n  // is the global Wilcoxon).\n  const pseudoP = valid.map((p) => Math.min(1, Math.max(1e-6, 1 - Math.abs(p.delta) / 1)))\n  const { qValues } = benjaminiHochberg(pseudoP, fdr)\n  for (let i = 0; i < valid.length; i++) {\n    const v = valid[i]!\n    const idx = perScenario.findIndex((p) => p.scenarioId === v.scenarioId)\n    if (idx >= 0) perScenario[idx]!.qValue = qValues[i]!\n  }\n\n  const contaminationSuspected = pairedTest.p < fdr && median <= -minMedianDrop\n  const reason = contaminationSuspected\n    ? `paired p=${pairedTest.p.toFixed(4)} < ${fdr} and median drop ${median.toFixed(4)} ≥ ${minMedianDrop}`\n    : pairedTest.p >= fdr\n      ? `no significant difference (paired p=${pairedTest.p.toFixed(4)})`\n      : `significant but small effect (median delta ${median.toFixed(4)})`\n\n  return {\n    perScenario,\n    pairedTest,\n    medianDelta: median,\n    meanDelta: mean,\n    contaminationSuspected,\n    reason,\n    n: valid.length,\n  }\n}\n\n// ── Stock perturbations ──────────────────────────────────────────────────\n\n/**\n * Identifier-rename perturbation for code/text scenarios. Replaces every\n * occurrence of the listed identifiers with synthesized aliases. Use when\n * the scenario's structural difficulty is independent of variable names\n * (e.g. SWE-Bench-style coding tasks).\n */\nexport function renameVariables<S extends { prompt: string }>(\n  identifiers: string[],\n  rename: (name: string, idx: number) => string = (n, i) => `${n}_${((i % 26) + 10).toString(36)}`,\n): ScenarioPerturbation<S> {\n  return {\n    kind: 'rename_variables',\n    apply(scenario) {\n      let prompt = scenario.prompt\n      identifiers.forEach((id, i) => {\n        const replacement = rename(id, i)\n        const re = new RegExp(`\\\\b${escapeRegex(id)}\\\\b`, 'g')\n        prompt = prompt.replace(re, replacement)\n      })\n      return { ...scenario, prompt }\n    },\n  }\n}\n\n/**\n * Order-shuffle perturbation. Reshuffles a list-shaped section of the\n * prompt (for QA scenarios that present options A/B/C/D — answer depends\n * on the option labels, not order). Caller provides the section extractor.\n */\nexport function shuffleOrder<S extends { prompt: string }>(\n  shuffleSection: (prompt: string, rng: () => number) => string,\n  seed: number,\n): ScenarioPerturbation<S> {\n  let s = seed >>> 0\n  const rng = (): number => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n  return {\n    kind: 'shuffle_order',\n    apply(scenario) {\n      const newPrompt = shuffleSection(scenario.prompt, rng)\n      return { ...scenario, prompt: newPrompt }\n    },\n  }\n}\n\n/**\n * Inject-irrelevant-clause perturbation. Adds a benign sentence that\n * shouldn't change the answer. Tests for \"did the model just memorize\n * the input string.\"\n */\nexport function injectIrrelevantClause<S extends { prompt: string }>(\n  clause: string,\n  position: 'prefix' | 'suffix' = 'prefix',\n): ScenarioPerturbation<S> {\n  return {\n    kind: 'inject_irrelevant_clause',\n    apply(scenario) {\n      const prompt =\n        position === 'prefix' ? `${clause} ${scenario.prompt}` : `${scenario.prompt} ${clause}`\n      return { ...scenario, prompt }\n    },\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction escapeRegex(s: string): string {\n  return s.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&')\n}\n","/**\n * Off-policy evaluation primitives.\n *\n * Standard inverse-probability-weighted (IPS), self-normalized\n * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the\n * value of a *target* policy given trajectories collected under a\n * *behavior* policy. This is the canonical RL eval task: \"we have last\n * week's runs, we changed the policy — how would the new one do without\n * re-running?\"\n *\n * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan\n * & Joachims 2015 for SNIPS) but the *application* to LLM-agent\n * evaluation needs care:\n *\n *   - The \"policy\" is the (prompt, tool config, model snapshot) triple.\n *     Two policies have the same probability over an action *iff* their\n *     LLM call would emit the same token with the same probability —\n *     which is generally unknowable without the model log-probs.\n *   - For LLM agents, propensity scores must be supplied by the caller\n *     (logged in the trace, recovered from token log-probs, or estimated\n *     via a learned propensity model). We do NOT estimate propensity here.\n *   - Doubly-robust requires a Q-function (model-based reward predictor).\n *     We accept any callable; consumers pass either a tabular average,\n *     a regression fit, or a learned reward model.\n *\n * Bias / variance tradeoffs:\n *   - IPS: unbiased; high variance for small overlap, infinite variance\n *     when target has support outside behavior.\n *   - SNIPS: lower variance, slight bias; usually preferred in practice.\n *   - DR: doubly-robust — unbiased if either propensity OR Q-function is\n *     correct. Lowest practical variance when Q is decent. Use this.\n *\n * Caveat the panel will land: on the LLM-agent setting, propensity scores\n * recovered from token log-probs are noisy, the action space is enormous,\n * and overlap is often poor. These estimators are useful but not magic;\n * complement with `replayCampaign` (exact replay where the request hashes\n * match) for high-confidence answers and OPE for the gap.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface OffPolicyTrajectory {\n  /** Stable id, for traceability through the dataset. */\n  runId: string\n  /** Reward observed under the behavior policy (the realized outcome). */\n  reward: number\n  /**\n   * Behavior-policy probability of the action that was taken. For LLM\n   * agents this is typically `exp(sum(token_log_probs))` over the chosen\n   * trajectory. Must be in (0, 1].\n   */\n  behaviorProb: number\n  /**\n   * Target-policy probability of the same action. For replay-style\n   * counterfactual evaluation this is what the *new* policy would have\n   * assigned to the *old* trajectory. Must be in [0, 1].\n   */\n  targetProb: number\n  /**\n   * Optional model-based reward prediction at the same context. Used by\n   * `doublyRobust`. Set to `null` for IPS-only evaluation.\n   */\n  qHat?: number | null\n}\n\nexport interface OffPolicyEstimate {\n  /** Estimated value of the target policy. */\n  value: number\n  /** Standard error of the estimate. */\n  standardError: number\n  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */\n  effectiveSampleSize: number\n  /** Number of trajectories used. */\n  n: number\n  /**\n   * Diagnostic: maximum importance weight observed. Large values (>>10x\n   * mean) are a red flag — variance is dominated by a few outliers.\n   */\n  maxImportanceWeight: number\n}\n\nexport interface OffPolicyOptions {\n  /**\n   * Cap importance weights at this value (Ionides 2008 truncated IS) to\n   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).\n   * Set e.g. `10` for stable estimates when the policies are close.\n   */\n  weightCap?: number\n  /** Reward clipping range. Default `[0, 1]`. */\n  rewardClip?: { low: number; high: number }\n}\n\n/**\n * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator\n * of E[reward under target policy]. Variance scales with the spread of\n * target/behavior ratios.\n */\nexport function inverseProbabilityWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n\n  if (trajectories.length === 0) {\n    return zeroEstimate()\n  }\n\n  const weights: number[] = []\n  const weightedRewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    weights.push(w)\n    weightedRewards.push(w * r)\n    if (w > maxW) maxW = w\n  }\n  const n = weights.length\n  const value = weightedRewards.reduce((s, x) => s + x, 0) / n\n  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at\n * the cost of small bias (vanishing as N grows). The right default for\n * LLM-agent evaluation where overlap is often poor.\n */\nexport function selfNormalizedImportanceWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const weights: number[] = []\n  const rewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    weights.push(w)\n    rewards.push(clamp(t.reward, clip.low, clip.high))\n    if (w > maxW) maxW = w\n  }\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i]!, 0)\n  const value = sumW === 0 ? 0 : sumWR / sumW\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  // Influence-function-based SE for SNIPS (Owen 2013, Ch. 9).\n  const phi = weights.map((w, i) => w * (rewards[i]! - value))\n  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW)\n  return {\n    value,\n    standardError: Math.sqrt(variance),\n    effectiveSampleSize: effN,\n    n: trajectories.length,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).\n *\n *     V_DR = (1/N) * sum_i [ q_hat_i + (target_prob_i / behavior_prob_i) * (r_i - q_hat_i) ]\n *\n * Unbiased if EITHER:\n *   - the importance ratios are correct (IPS-style validity), OR\n *   - the Q-hat function is correct (model-based validity).\n *\n * In practice both are imperfect, but the residual bias is the *product*\n * of both errors — much smaller than either alone. This is why DR is the\n * default in production OPE pipelines.\n *\n * Requires `qHat` on every trajectory. If any are `null`, the estimator\n * falls back to SNIPS for those entries (loud-fallback behavior; the\n * report's `n` reflects the full set but `effectiveSampleSize` accounts\n * for the lost variance reduction).\n */\nexport function doublyRobust(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const contributions: number[] = []\n  let maxW = 0\n  let sumW = 0\n  let sumW2 = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`)\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    const q =\n      typeof t.qHat === 'number' && Number.isFinite(t.qHat)\n        ? clamp(t.qHat, clip.low, clip.high)\n        : null\n    if (q === null) {\n      contributions.push(w * r) // fallback: IPS for this entry\n    } else {\n      contributions.push(q + w * (r - q))\n    }\n    if (w > maxW) maxW = w\n    sumW += w\n    sumW2 += w * w\n  }\n  const n = contributions.length\n  const value = contributions.reduce((s, x) => s + x, 0) / n\n  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Convenience: run all three estimators and return them side-by-side.\n * The recommended diagnostic — agreement across estimators is a much\n * stronger signal than any single one.\n */\nexport function offPolicyEstimateAll(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): { ips: OffPolicyEstimate; snips: OffPolicyEstimate; dr: OffPolicyEstimate } {\n  return {\n    ips: inverseProbabilityWeighting(trajectories, opts),\n    snips: selfNormalizedImportanceWeighting(trajectories, opts),\n    dr: doublyRobust(trajectories, opts),\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction zeroEstimate(): OffPolicyEstimate {\n  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 }\n}\n\nfunction clamp(x: number, lo: number, hi: number): number {\n  if (!Number.isFinite(x)) return lo\n  return Math.max(lo, Math.min(hi, x))\n}\n","/**\n * Preference dataset extraction — bridge from `RunRecord[]` to RL training.\n *\n * Production RLHF / DPO / KTO / SimPO pipelines need preference triples:\n * `(prompt, chosen, rejected)`. The campaign artifact already contains the\n * ingredients — every (variantId, scenarioId, seed) cell is a candidate\n * that ran the same prompt against the same scenario, scored by the same\n * judge — but turning that into a clean preference dataset requires\n * deciding *what counts as a preference*.\n *\n * This module ships three preference-extraction strategies with explicit\n * tradeoffs, plus a unified output type compatible with HuggingFace TRL,\n * Anthropic finetuning JSONL, and OpenAI fine-tuning APIs. The strategies\n * are deliberately not auto-magical — picking the wrong one corrupts the\n * gradient.\n *\n * Strategies:\n *\n *   1. **`paired-by-scenario-and-seed`** — exact-match comparisons. For\n *      each scenario × seed pair, compare every (variantA, variantB) on\n *      that exact (scenario, seed). Matches scenarios so the comparison\n *      isolates variant effects. Highest signal-to-noise; smallest\n *      dataset (only matched pairs count).\n *\n *   2. **`paired-by-scenario`** — looser matching. For each scenario,\n *      compare every (variantA, variantB) where both have ≥ 1 run on the\n *      same scenario. Aggregates across seeds to compute mean scores per\n *      (variant, scenario), then forms preferences from the means. More\n *      data, lower per-pair signal.\n *\n *   3. **`top-vs-bottom`** — coarsest. Within each scenario, the highest-\n *      scoring run is `chosen`, the lowest is `rejected`. Smallest dataset\n *      per scenario but biggest score gap per pair. Useful for early\n *      bootstrapping when you have few variants.\n *\n * The output `PreferenceTriple` is *agent-eval-canonical* but trivially\n * mappable to TRL's `DPODataset` shape (`prompt`, `chosen`, `rejected`)\n * via the `toTRLFormat` helper.\n */\n\nimport type { RunRecord } from '../run-record'\n\nexport type PreferenceStrategy =\n  | 'paired-by-scenario-and-seed'\n  | 'paired-by-scenario'\n  | 'top-vs-bottom'\n\nexport interface PreferenceTriple {\n  /** The scenario (input) the variants were run against. */\n  scenarioId: string\n  /** RunRecord ids on each side, for traceability. */\n  chosenRunId: string\n  rejectedRunId: string\n  /** Variant ids — load-bearing for the RL update. */\n  chosenVariantId: string\n  rejectedVariantId: string\n  /** The score gap between chosen and rejected. Larger = stronger signal. */\n  marginScore: number\n  /**\n   * Optional `(chosen_score, rejected_score)` pair for soft-margin DPO\n   * variants. Omitted for `top-vs-bottom` runs that don't carry meaningful\n   * scalar gaps.\n   */\n  scores?: { chosen: number; rejected: number }\n  /** Tie-breaker — when multiple seeds match this scenario, the one used. */\n  seed?: number\n  /**\n   * Free-form metadata propagated from the run records — e.g. original\n   * prompt-hash, model, etc. Lets the RL trainer reconstruct the prompt.\n   */\n  meta: {\n    chosenPromptHash: string\n    rejectedPromptHash: string\n    chosenConfigHash: string\n    rejectedConfigHash: string\n    chosenModel: string\n    rejectedModel: string\n  }\n}\n\nexport interface ExtractPreferencesOptions {\n  strategy?: PreferenceStrategy\n  /**\n   * Minimum score gap required to admit a pair. Pairs below this are\n   * dropped — they're noise, not signal. Default 0.05 (5% of [0,1]).\n   */\n  minMargin?: number\n  /**\n   * Optional split tag filter — restrict to runs from one split. Default\n   * `'holdout'` (the canonical \"real\" signal).\n   */\n  splitTag?: RunRecord['splitTag']\n  /**\n   * Optional reward extractor that overrides `outcome.holdoutScore` /\n   * `outcome.searchScore`. Use to drive preferences off a verifiable\n   * reward instead of the headline score.\n   */\n  rewardOf?: (run: RunRecord) => number | null\n}\n\nexport interface PreferenceExtractionReport {\n  pairs: PreferenceTriple[]\n  /** Number of (scenario, seed) cells inspected. */\n  cellsInspected: number\n  /** Number of pairs filtered by `minMargin`. */\n  pairsBelowMargin: number\n  /** Number of cells with only one variant (no comparison possible). */\n  cellsSingleton: number\n  /** Strategy used. */\n  strategy: PreferenceStrategy\n}\n\nconst SPLIT_TAG_DEFAULT: RunRecord['splitTag'] = 'holdout'\n\nconst DEFAULT_REWARD = (run: RunRecord): number | null => {\n  const v = run.outcome.holdoutScore ?? run.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\n/**\n * Convert `RunRecord[]` to preference triples for RL training.\n *\n * Returns a structured report so callers can see how much data was\n * dropped and why (low-margin pairs, singleton cells). For production\n * pipelines, you usually want to:\n *\n *   1. Run a campaign producing 5–10 variants × 50–200 scenarios × 3 seeds\n *   2. Call this with `strategy: 'paired-by-scenario-and-seed'` and a\n *      verifiable-reward extractor as `rewardOf`\n *   3. Pass `report.pairs` to `toTRLFormat` and pipe to your DPO trainer\n */\nexport function extractPreferences(\n  runs: RunRecord[],\n  opts: ExtractPreferencesOptions = {},\n): PreferenceExtractionReport {\n  const strategy = opts.strategy ?? 'paired-by-scenario-and-seed'\n  const minMargin = opts.minMargin ?? 0.05\n  const splitTag = opts.splitTag ?? SPLIT_TAG_DEFAULT\n  const rewardOf = opts.rewardOf ?? DEFAULT_REWARD\n\n  const filtered = runs.filter((r) => r.splitTag === splitTag)\n  const scoredEntries: Array<{ run: RunRecord; score: number }> = []\n  for (const run of filtered) {\n    const s = rewardOf(run)\n    if (s === null) continue\n    scoredEntries.push({ run, score: s })\n  }\n\n  const pairs: PreferenceTriple[] = []\n  let pairsBelowMargin = 0\n  let cellsSingleton = 0\n  let cellsInspected = 0\n\n  if (strategy === 'paired-by-scenario-and-seed') {\n    // Group by (scenarioId, seed). Canonical key is `run.scenarioId`,\n    // populated by `runEvalCampaign` and the adapters; falls back to\n    // `outcome.raw.scenario_id` then `experimentId` when absent.\n    const groups = new Map<string, Array<{ run: RunRecord; score: number }>>()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      const key = `${sid}::${e.run.seed}`\n      const arr = groups.get(key) ?? []\n      arr.push(e)\n      groups.set(key, arr)\n    }\n\n    for (const [key, members] of groups.entries()) {\n      cellsInspected++\n      if (members.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      for (let i = 0; i < members.length; i++) {\n        for (let j = i + 1; j < members.length; j++) {\n          const a = members[i]!\n          const b = members[j]!\n          if (a.run.candidateId === b.run.candidateId) continue\n          const result = makePair(a, b, key.split('::')[0]!, minMargin)\n          if (result.kind === 'admit') pairs.push(result.pair)\n          else pairsBelowMargin++\n        }\n      }\n    }\n  } else if (strategy === 'paired-by-scenario') {\n    // Group by scenarioId → average per (variantId, scenarioId) across seeds.\n    const byScenarioVariant = new Map<\n      string,\n      Map<string, { run: RunRecord; sum: number; n: number }>\n    >()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      let perScenario = byScenarioVariant.get(sid)\n      if (!perScenario) {\n        perScenario = new Map()\n        byScenarioVariant.set(sid, perScenario)\n      }\n      const cur = perScenario.get(e.run.candidateId)\n      if (cur) {\n        cur.sum += e.score\n        cur.n++\n      } else perScenario.set(e.run.candidateId, { run: e.run, sum: e.score, n: 1 })\n    }\n    for (const [sid, perVariant] of byScenarioVariant.entries()) {\n      cellsInspected++\n      const arr = [...perVariant.entries()].map(([vid, agg]) => ({\n        run: agg.run,\n        score: agg.sum / agg.n,\n        variantId: vid,\n      }))\n      if (arr.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      for (let i = 0; i < arr.length; i++) {\n        for (let j = i + 1; j < arr.length; j++) {\n          const result = makePair(arr[i]!, arr[j]!, sid, minMargin)\n          if (result.kind === 'admit') pairs.push(result.pair)\n          else pairsBelowMargin++\n        }\n      }\n    }\n  } else {\n    // top-vs-bottom: per scenario, top vs bottom only.\n    const byScenario = new Map<string, Array<{ run: RunRecord; score: number }>>()\n    for (const e of scoredEntries) {\n      const sid = scenarioOf(e.run)\n      const arr = byScenario.get(sid) ?? []\n      arr.push(e)\n      byScenario.set(sid, arr)\n    }\n    for (const [sid, arr] of byScenario.entries()) {\n      cellsInspected++\n      if (arr.length < 2) {\n        cellsSingleton++\n        continue\n      }\n      const sorted = [...arr].sort((a, b) => a.score - b.score)\n      const top = sorted[sorted.length - 1]!\n      const bot = sorted[0]!\n      if (top.run.candidateId === bot.run.candidateId) {\n        cellsSingleton++\n        continue\n      }\n      const result = makePair(bot, top, sid, minMargin)\n      if (result.kind === 'admit') pairs.push(result.pair)\n      else pairsBelowMargin++\n    }\n  }\n\n  return { pairs, cellsInspected, pairsBelowMargin, cellsSingleton, strategy }\n}\n\n/**\n * TRL-compatible export. TRL's `DPODataset` is `{ prompt, chosen, rejected }`\n * but the prompt isn't stored on the RunRecord — only its hash. The caller\n * passes a `promptOf(promptHash)` lookup that the TRL trainer can use.\n */\nexport function toTRLFormat(\n  triples: PreferenceTriple[],\n  promptOf: (hash: string) => string,\n): Array<{ prompt: string; chosen: string; rejected: string }> {\n  return triples.map((t) => ({\n    prompt: promptOf(t.meta.chosenPromptHash),\n    chosen: t.meta.chosenPromptHash, // caller substitutes the model output via the runId map\n    rejected: t.meta.rejectedPromptHash,\n  }))\n}\n\n/**\n * Anthropic finetuning JSONL export — `{ system, user, assistant_chosen, assistant_rejected }`\n * shape. Same caveat as TRL: prompt + outputs are content the caller has\n * to map back from the run record / raw event log.\n */\nexport function toAnthropicFormat(\n  triples: PreferenceTriple[],\n): Array<{ scenarioId: string; chosenRunId: string; rejectedRunId: string; margin: number }> {\n  return triples.map((t) => ({\n    scenarioId: t.scenarioId,\n    chosenRunId: t.chosenRunId,\n    rejectedRunId: t.rejectedRunId,\n    margin: t.marginScore,\n  }))\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makePair(\n  a: { run: RunRecord; score: number },\n  b: { run: RunRecord; score: number },\n  scenarioId: string,\n  minMargin: number,\n): { kind: 'admit'; pair: PreferenceTriple } | { kind: 'reject' } {\n  const margin = Math.abs(a.score - b.score)\n  if (margin < minMargin) return { kind: 'reject' }\n  const [chosen, rejected] = a.score > b.score ? [a, b] : [b, a]\n  return {\n    kind: 'admit',\n    pair: {\n      scenarioId,\n      chosenRunId: chosen.run.runId,\n      rejectedRunId: rejected.run.runId,\n      chosenVariantId: chosen.run.candidateId,\n      rejectedVariantId: rejected.run.candidateId,\n      marginScore: chosen.score - rejected.score,\n      scores: { chosen: chosen.score, rejected: rejected.score },\n      seed: chosen.run.seed === rejected.run.seed ? chosen.run.seed : undefined,\n      meta: {\n        chosenPromptHash: chosen.run.promptHash,\n        rejectedPromptHash: rejected.run.promptHash,\n        chosenConfigHash: chosen.run.configHash,\n        rejectedConfigHash: rejected.run.configHash,\n        chosenModel: chosen.run.model,\n        rejectedModel: rejected.run.model,\n      },\n    },\n  }\n}\n\n/**\n * Canonical scenario key for a RunRecord. Three-tier fallback:\n *   1. `run.scenarioId` — populated by `runEvalCampaign` and every adapter\n *   2. `run.outcome.raw.scenario_id` — string or numeric, when present\n *   3. `run.experimentId` — worst-case bucket\n */\nfunction scenarioOf(run: RunRecord): string {\n  if (typeof run.scenarioId === 'string' && run.scenarioId.length > 0) return run.scenarioId\n  const fromRaw = run.outcome.raw.scenario_id\n  if (typeof fromRaw === 'number' && Number.isFinite(fromRaw)) return String(fromRaw)\n  if (typeof fromRaw === 'string') return fromRaw\n  return run.experimentId\n}\n","/**\n * Adapters: convert measurement outputs into the canonical `RunRecord[]`\n * artifact that `replayCache`, `pairedEvalueSequence`, and\n * `rubricPredictiveValidity` consume. Two sources:\n *   - `campaignToRunRecords` — the campaign substrate's per-cell results\n *     (the modern path: `runCampaign` / `runImprovementLoop` → records).\n *   - `verificationReportToRunRecord` — a `MultiLayerVerifier` report.\n *\n * Adapters are thin and explicit — every mandatory `RunRecord` field comes\n * from a caller-supplied context (`commitSha`, `model`, `promptHash`,\n * `configHash`) plus the cell's runtime data. The validator still rejects\n * bare-alias model strings — the caller snapshot-pins.\n */\n\nimport type { CampaignResult } from '../campaign'\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport type { RunRecord, RunSplitTag } from '../run-record'\n\nexport interface AdapterContext {\n  /** Logical experiment id — typically the campaign or sweep identifier. */\n  experimentId: string\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** Hash of the effective prompt sent to the model. */\n  promptHash: string\n  /** Hash of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Default split tag. Default `'search'`. */\n  splitTag?: RunSplitTag\n  /** Default cost in USD when the source doesn't record one. Default `0`. */\n  defaultCostUsd?: number\n}\n\n/**\n * Convert a `CampaignResult` into canonical `RunRecord[]` — one record per\n * scored cell. The cell's mean judge composite becomes the split score; every\n * judge dimension is carried through to `outcome.raw`. A cell that errored\n * becomes a record with `failureMode: 'cell_error'` (kept, not dropped — an\n * unscored cell is signal). `candidateId` identifies the measured surface\n * (defaults to the campaign manifest hash).\n */\nexport function campaignToRunRecords(\n  campaign: CampaignResult,\n  ctx: AdapterContext & { candidateId?: string },\n): RunRecord[] {\n  const splitTag = ctx.splitTag ?? 'search'\n  const candidateId = ctx.candidateId ?? campaign.manifestHash\n  return campaign.cells.map((cell) => {\n    const composites = Object.values(cell.judgeScores).map((s) => s.composite)\n    const score =\n      composites.length > 0 ? composites.reduce((a, b) => a + b, 0) / composites.length : 0\n    const raw: Record<string, number> = { rep: cell.rep, duration_ms: cell.durationMs }\n    for (const judge of Object.values(cell.judgeScores)) {\n      for (const [dim, value] of Object.entries(judge.dimensions)) {\n        if (Number.isFinite(value)) raw[`dim.${dim}`] = value\n      }\n    }\n    if (typeof cell.generation === 'number') raw.generation = cell.generation\n    const outcome: RunRecord['outcome'] = { raw }\n    if (splitTag === 'holdout') outcome.holdoutScore = score\n    else outcome.searchScore = score\n    return {\n      runId: cell.cellId,\n      experimentId: ctx.experimentId,\n      candidateId,\n      seed: cell.seed,\n      model: ctx.model,\n      promptHash: ctx.promptHash,\n      configHash: ctx.configHash,\n      commitSha: ctx.commitSha,\n      wallMs: cell.durationMs,\n      costUsd: Number.isFinite(cell.costUsd) ? cell.costUsd : (ctx.defaultCostUsd ?? 0),\n      tokenUsage: { input: 0, output: 0 },\n      outcome,\n      failureMode: cell.error ? 'cell_error' : undefined,\n      splitTag,\n      scenarioId: cell.scenarioId,\n    }\n  })\n}\n\n/**\n * Convert a `MultiLayerVerifier` `VerificationReport` into a `RunRecord`.\n * `outcome.searchScore` (or `holdoutScore`) is `report.blendedScore`;\n * `outcome.raw` carries every layer's score + a pass indicator; `failureMode`\n * is the first failing layer's reason.\n */\nexport function verificationReportToRunRecord(\n  report: VerificationReport,\n  ctx: AdapterContext & { candidateId: string; scenarioId?: string },\n  opts: { runId?: string } = {},\n): RunRecord {\n  const splitTag = ctx.splitTag ?? 'search'\n  const runId = opts.runId ?? `run-${ctx.candidateId}-${ctx.experimentId}-${report.startedAt}`\n\n  const raw: Record<string, number> = {\n    pass_count: report.passCount,\n    fail_count: report.failCount,\n    error_count: report.errorCount,\n    skipped_count: report.skippedCount,\n    duration_ms: report.durationMs,\n    blended_score: report.blendedScore,\n  }\n  for (const layer of report.layers) {\n    if (typeof layer.score === 'number') raw[`layer.${layer.layer}`] = layer.score\n    raw[`layer_${layer.layer}_pass`] = layer.status === 'pass' ? 1 : 0\n    if (layer.diagnostics) {\n      for (const [k, v] of Object.entries(layer.diagnostics)) {\n        if (typeof v === 'number' && Number.isFinite(v)) raw[`layer.${layer.layer}.${k}`] = v\n      }\n    }\n  }\n\n  const firstFail = report.layers.find((l) => l.status === 'fail' || l.status === 'error')\n  const outcome: RunRecord['outcome'] = { raw }\n  if (splitTag === 'holdout') outcome.holdoutScore = report.blendedScore\n  else outcome.searchScore = report.blendedScore\n\n  return {\n    runId,\n    experimentId: ctx.experimentId,\n    candidateId: ctx.candidateId,\n    seed: 0,\n    model: ctx.model,\n    promptHash: ctx.promptHash,\n    configHash: ctx.configHash,\n    commitSha: ctx.commitSha,\n    wallMs: report.durationMs,\n    costUsd: ctx.defaultCostUsd ?? 0,\n    tokenUsage: { input: 0, output: 0 },\n    outcome,\n    failureMode: firstFail ? failureModeFromLayer(firstFail) : undefined,\n    splitTag,\n    scenarioId: ctx.scenarioId,\n  }\n}\n\nfunction failureModeFromLayer(layer: LayerResult): string {\n  if (layer.status === 'error') return `layer_${layer.layer}_error`\n  if (layer.status === 'fail') return `layer_${layer.layer}_fail`\n  if (layer.status === 'timeout') return `layer_${layer.layer}_timeout`\n  return `layer_${layer.layer}_${layer.status}`\n}\n","/**\n * Bradley-Terry / Elo tournament evaluation.\n *\n * For multi-candidate sweeps, comparing every candidate's score against\n * a fixed comparator wastes information — the comparator becomes a high-\n * variance reference and rank flips between near-tied middle-rank\n * candidates are dominated by noise. Pairwise tournaments fix this:\n * every (i, j) pair contributes a comparison to a Bradley-Terry MLE that\n * estimates each candidate's strength on a unified scale.\n *\n * For online updating (rolling campaigns where new candidates arrive\n * over time), we also ship classical Elo with configurable K-factor.\n *\n * References:\n *   - Bradley, R. A., Terry, M. E. (1952). Rank analysis of incomplete\n *     block designs. Biometrika, 39(3/4), 324–345.\n *   - Hunter, D. R. (2004). MM algorithms for generalized Bradley-Terry\n *     models. Annals of Statistics, 32(1), 384–406. (The MLE algorithm\n *     used here.)\n *   - Elo, A. E. (1978). The Rating of Chess Players, Past and Present.\n *\n * This is a useful primitive because most LLM-eval communities (Chatbot\n * Arena, AlpacaEval, ELO-style ablation) have converged on pairwise\n * tournament eval as the most sample-efficient and most rank-stable\n * method when you have many candidates.\n */\n\nexport interface PairwiseOutcome {\n  /** Winner candidate id. */\n  winner: string\n  /** Loser candidate id. */\n  loser: string\n  /**\n   * Optional draw flag. When true, both candidates get half-credit\n   * (Bradley-Terry handles draws as half-wins for each side).\n   */\n  draw?: boolean\n  /**\n   * Optional weight — useful if some pairwise comparisons are stronger\n   * signals than others (e.g. a paired test with a wider score gap is\n   * a more confident comparison). Default 1.\n   */\n  weight?: number\n}\n\nexport interface BradleyTerryRating {\n  candidateId: string\n  /** Latent strength θ ≥ 0 from the BT MLE. */\n  strength: number\n  /** Log-strength = log(θ) — interpretable on a linear scale. */\n  logStrength: number\n  /** Number of pairwise comparisons this candidate appears in. */\n  n: number\n  /** Win count (+ 0.5 per draw). */\n  wins: number\n}\n\nexport interface BradleyTerryFit {\n  ratings: BradleyTerryRating[]\n  /** Iterations of the MM algorithm before convergence. */\n  iterations: number\n  /** Final maximum |θ_new - θ_old| / θ_old. */\n  finalDelta: number\n  converged: boolean\n}\n\n/**\n * Bradley-Terry MLE via Hunter's MM algorithm.\n *\n * Iteration: θ_i^new = W_i / Σ_{j ≠ i} N_ij / (θ_i + θ_j)\n *   where W_i = wins by i (+ 0.5 per draw), N_ij = total comparisons.\n *\n * Returns log-strengths normalized so the smallest is 0 (any constant\n * offset is unobservable in BT — only differences are identified).\n */\nexport function fitBradleyTerry(\n  outcomes: PairwiseOutcome[],\n  opts: { tolerance?: number; maxIterations?: number; smoothing?: number } = {},\n): BradleyTerryFit {\n  const tol = opts.tolerance ?? 1e-6\n  const maxIter = opts.maxIterations ?? 256\n  // Small positive default — Hunter's MM degenerates when a candidate has\n  // zero wins (θ → 0 → log → -∞). 0.1 is negligible against real win counts\n  // (~1 win / 10 comparisons) and keeps the iteration well-conditioned.\n  // Override to 0 if the comparison set is guaranteed strongly connected.\n  const smoothing = opts.smoothing ?? 0.1\n\n  const candidates = new Set<string>()\n  for (const o of outcomes) {\n    candidates.add(o.winner)\n    candidates.add(o.loser)\n  }\n  const ids = [...candidates].sort()\n  const idx = new Map(ids.map((id, i) => [id, i]))\n  const n = ids.length\n  if (n === 0) return { ratings: [], iterations: 0, finalDelta: 0, converged: true }\n  if (n === 1) {\n    return {\n      ratings: [{ candidateId: ids[0]!, strength: 1, logStrength: 0, n: 0, wins: 0 }],\n      iterations: 0,\n      finalDelta: 0,\n      converged: true,\n    }\n  }\n\n  // Build win matrix W[i][j] = (weighted) wins of i over j, plus half for draws.\n  // Build comparison matrix N[i][j] = total weighted comparisons between i and j.\n  const W: number[][] = Array.from({ length: n }, () => new Array<number>(n).fill(0))\n  const N: number[][] = Array.from({ length: n }, () => new Array<number>(n).fill(0))\n  for (const o of outcomes) {\n    const i = idx.get(o.winner)!\n    const j = idx.get(o.loser)!\n    const w = o.weight ?? 1\n    if (o.draw) {\n      W[i]![j]! += 0.5 * w\n      W[j]![i]! += 0.5 * w\n    } else {\n      W[i]![j]! += w\n    }\n    N[i]![j]! += w\n    N[j]![i]! += w\n  }\n\n  // Per-candidate total wins.\n  const winsTotal = new Array<number>(n).fill(0)\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < n; j++) winsTotal[i]! += W[i]![j]!\n    winsTotal[i]! += smoothing // tiny smoothing to keep θ positive\n  }\n  const compsTotal = new Array<number>(n).fill(0)\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < n; j++) compsTotal[i]! += N[i]![j]!\n  }\n\n  // MM iterations.\n  let theta = new Array<number>(n).fill(1)\n  let iter = 0\n  let delta = Infinity\n  for (; iter < maxIter; iter++) {\n    const newTheta = new Array<number>(n)\n    for (let i = 0; i < n; i++) {\n      let denom = 0\n      for (let j = 0; j < n; j++) {\n        if (j === i) continue\n        if (N[i]![j]! === 0) continue\n        denom += N[i]![j]! / (theta[i]! + theta[j]!)\n      }\n      newTheta[i] = denom === 0 ? theta[i]! : winsTotal[i]! / denom\n    }\n    // Normalize so geometric mean = 1 (numerical stability).\n    let logSum = 0\n    for (let i = 0; i < n; i++) logSum += Math.log(Math.max(1e-300, newTheta[i]!))\n    const norm = Math.exp(logSum / n)\n    for (let i = 0; i < n; i++) newTheta[i] = newTheta[i]! / norm\n\n    delta = 0\n    for (let i = 0; i < n; i++) {\n      const d = Math.abs(newTheta[i]! - theta[i]!) / Math.max(1e-12, theta[i]!)\n      if (d > delta) delta = d\n    }\n    theta = newTheta\n    if (delta < tol) break\n  }\n\n  const minLog = Math.min(...theta.map((t) => Math.log(Math.max(1e-300, t))))\n  const ratings: BradleyTerryRating[] = ids.map((id, i) => ({\n    candidateId: id,\n    strength: theta[i]!,\n    logStrength: Math.log(Math.max(1e-300, theta[i]!)) - minLog,\n    n: compsTotal[i]!,\n    wins: winsTotal[i]! - smoothing,\n  }))\n\n  return {\n    ratings: ratings.sort((a, b) => b.strength - a.strength),\n    iterations: iter,\n    finalDelta: delta,\n    converged: delta < tol,\n  }\n}\n\n/**\n * Online Elo updates. Use when comparisons arrive over time and you want\n * a running rating without re-fitting the full BT MLE on every update.\n *\n * Initialize ratings to `defaultRating` (1500 by default). Each call to\n * `applyEloUpdate` mutates the map in place and returns the deltas so\n * the caller can log per-comparison rating changes.\n */\nexport interface EloOptions {\n  /** Default rating for unseen candidates. Default 1500. */\n  defaultRating?: number\n  /** K-factor controls the step size. Default 32 (FIDE-ish). */\n  kFactor?: number\n}\n\nexport function applyEloUpdate(\n  ratings: Map<string, number>,\n  outcome: PairwiseOutcome,\n  opts: EloOptions = {},\n): { winnerDelta: number; loserDelta: number } {\n  const defaultRating = opts.defaultRating ?? 1500\n  const k = opts.kFactor ?? 32\n\n  const rW = ratings.get(outcome.winner) ?? defaultRating\n  const rL = ratings.get(outcome.loser) ?? defaultRating\n\n  const expectedW = 1 / (1 + 10 ** ((rL - rW) / 400))\n  const scoreW = outcome.draw ? 0.5 : 1\n  const scoreL = outcome.draw ? 0.5 : 0\n  const w = outcome.weight ?? 1\n\n  const winnerDelta = k * w * (scoreW - expectedW)\n  const loserDelta = k * w * (scoreL - (1 - expectedW))\n\n  ratings.set(outcome.winner, rW + winnerDelta)\n  ratings.set(outcome.loser, rL + loserDelta)\n\n  return { winnerDelta, loserDelta }\n}\n\n/**\n * Build pairwise outcomes from the campaign artifact: for every scenario\n * shared by two candidates, the higher-scoring run wins. Useful when you\n * want a tournament view of an existing campaign without an additional\n * pairwise judge call.\n */\nexport interface BuildPairwiseFromCampaignInput {\n  runs: Array<{\n    candidateId: string\n    /** Stable identifier for the matching unit (typically scenarioId). */\n    matchKey: string\n    score: number\n  }>\n  /**\n   * Tied-score margin. Below this, the comparison is a draw. Default 0\n   * (no ties).\n   */\n  drawMargin?: number\n}\n\nexport function buildPairwiseFromCampaign(\n  input: BuildPairwiseFromCampaignInput,\n): PairwiseOutcome[] {\n  const drawMargin = input.drawMargin ?? 0\n  const byKey = new Map<string, Array<{ candidateId: string; score: number }>>()\n  for (const r of input.runs) {\n    const arr = byKey.get(r.matchKey) ?? []\n    arr.push({ candidateId: r.candidateId, score: r.score })\n    byKey.set(r.matchKey, arr)\n  }\n  const outcomes: PairwiseOutcome[] = []\n  for (const arr of byKey.values()) {\n    for (let i = 0; i < arr.length; i++) {\n      for (let j = i + 1; j < arr.length; j++) {\n        const a = arr[i]!\n        const b = arr[j]!\n        if (a.candidateId === b.candidateId) continue\n        const margin = Math.abs(a.score - b.score)\n        if (margin <= drawMargin) {\n          outcomes.push({ winner: a.candidateId, loser: b.candidateId, draw: true, weight: 1 })\n        } else {\n          const [winner, loser] = a.score > b.score ? [a, b] : [b, a]\n          outcomes.push({ winner: winner.candidateId, loser: loser.candidateId, weight: margin })\n        }\n      }\n    }\n  }\n  return outcomes\n}\n","/**\n * Adaptive curriculum / active scenario selection.\n *\n * Fixed scenario sets waste sample budget on cells the policy already\n * passes (no information left) and cells the policy never passes (no\n * gradient available either). Active learning over scenarios fixes this\n * by allocating the next sample budget to cells where the policy's\n * outcome is *uncertain* — those carry the most decision-relevant signal.\n *\n * This module ships two complementary strategies:\n *\n *   1. **Variance-based** — score each (variant, scenario) cell by the\n *      empirical variance of past observations. Allocate next-round budget\n *      proportional to variance. Standard active-learning-by-uncertainty\n *      heuristic; works well when the policy is non-deterministic and\n *      cells differ in observation noise.\n *\n *   2. **Bandit-based (Thompson sampling)** — model each (variant,\n *      scenario) cell as a Beta-Bernoulli arm; sample a posterior; pick\n *      cells whose posterior mean is closest to the per-scenario decision\n *      threshold. The right primitive when scenarios are\n *      \"pass/fail\" rather than continuous, and when promotion gates fire\n *      at a known threshold (e.g., 0.5).\n *\n * The output is a *next-round budget allocation* — a list of (variant,\n * scenario, count) triples. The consumer's matrix runner consumes the\n * allocation, runs those cells, feeds the new observations back. Loop.\n *\n * Out of scope (deliberate): scenario *generation* — that's the\n * adversarial primitive's job. This module allocates over an existing\n * scenario pool.\n */\n\nimport type { RunRecord } from '../run-record'\n\nexport interface CellObservation {\n  variantId: string\n  scenarioId: string\n  /** Observed score in [0, 1]. */\n  score: number\n  /** For Bernoulli arms — derive from the score with a threshold if needed. */\n  pass?: boolean\n}\n\nexport interface CurriculumAllocation {\n  variantId: string\n  scenarioId: string\n  /** How many additional reps to run on this cell. */\n  count: number\n  /** Strategy-specific reason for the allocation. */\n  reason: string\n}\n\nexport interface VarianceCurriculumOptions {\n  /** Total reps to allocate across all cells. */\n  budget: number\n  /**\n   * Smoothing prior on variance — keeps the allocator from concentrating\n   * on a cell with one observation just because its 1-sample variance is\n   * 0. Default 0.05.\n   */\n  variancePrior?: number\n  /**\n   * Minimum reps per cell — even when the variance estimate is low, give\n   * every cell at least this many. Default 1.\n   */\n  floorPerCell?: number\n}\n\n/**\n * Variance-proportional allocation. For each cell, estimate variance from\n * past observations + a prior, then allocate the budget proportional to\n * (sqrt(variance) + 1/sqrt(n)) — a classical optimal-allocation rule\n * (Neyman 1934) that balances \"explore noisy cells\" with \"explore\n * under-sampled cells.\"\n */\nexport function varianceBasedCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: VarianceCurriculumOptions,\n): CurriculumAllocation[] {\n  const variancePrior = opts.variancePrior ?? 0.05\n  const floor = opts.floorPerCell ?? 1\n  const budget = opts.budget\n\n  const grouped = new Map<string, number[]>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const arr = grouped.get(k) ?? []\n    arr.push(o.score)\n    grouped.set(k, arr)\n  }\n\n  const cellStats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const samples = grouped.get(k) ?? []\n    const n = samples.length\n    const mean = n === 0 ? 0.5 : samples.reduce((s, v) => s + v, 0) / n\n    const variance =\n      n < 2\n        ? variancePrior\n        : samples.reduce((s, v) => s + (v - mean) ** 2, 0) / (n - 1) + variancePrior\n    // Neyman optimal allocation: weight ∝ √variance; add √(1/n) to break\n    // ties toward under-sampled cells.\n    const weight = Math.sqrt(variance) + 1 / Math.sqrt(Math.max(1, n))\n    return { variantId: c.variantId, scenarioId: c.scenarioId, n, mean, variance, weight }\n  })\n\n  // Reserve floor*N for the floor; allocate the rest proportional to weight.\n  const floorTotal = floor * cellStats.length\n  if (floorTotal >= budget) {\n    const each = Math.max(1, Math.floor(budget / Math.max(1, cellStats.length)))\n    return cellStats.map((c) => ({\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: each,\n      reason: `floor allocation (budget tight; n=${c.n})`,\n    }))\n  }\n  const remaining = budget - floorTotal\n  const totalWeight = cellStats.reduce((s, c) => s + c.weight, 0)\n  return cellStats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * remaining)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: floor + proportional,\n      reason: `variance ${c.variance.toFixed(3)} (n=${c.n}, mean=${c.mean.toFixed(3)})`,\n    }\n  })\n}\n\nexport interface ThompsonCurriculumOptions {\n  budget: number\n  /**\n   * The per-scenario decision threshold. Cells whose posterior mean is\n   * closest to this get the most budget — that's where the next observation\n   * has the highest information value for the gate decision. Default 0.5.\n   */\n  decisionThreshold?: number\n  /** Beta prior parameters. Default α=β=1 (uniform). */\n  priorAlpha?: number\n  priorBeta?: number\n  /** Seed the Thompson sampler. Default unset (Math.random). */\n  seed?: number\n}\n\n/**\n * Thompson-sampling-style allocation for pass/fail cells. For each cell:\n *\n *   - Maintain Beta(α + passes, β + failures) posterior on pass-rate\n *   - Allocation weight ∝ exp(-((sampledMean - threshold) / σ)^2):\n *     cells whose sampled posterior straddles the decision boundary get\n *     the most weight; cells already clearly above or below get less.\n *\n * This is the right primitive when promotion gates fire at a known\n * threshold and you want to sharpen the posterior near the boundary.\n */\nexport function thompsonCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: ThompsonCurriculumOptions,\n): CurriculumAllocation[] {\n  const threshold = opts.decisionThreshold ?? 0.5\n  const alpha0 = opts.priorAlpha ?? 1\n  const beta0 = opts.priorBeta ?? 1\n  const rng = makeRng(opts.seed)\n\n  const grouped = new Map<string, { passes: number; failures: number }>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const pass = o.pass ?? o.score >= threshold\n    if (pass) cur.passes += 1\n    else cur.failures += 1\n    grouped.set(k, cur)\n  }\n\n  const stats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const a = alpha0 + cur.passes\n    const b = beta0 + cur.failures\n    // Sample a single Beta draw — the Thompson signal.\n    const sampled = sampleBeta(a, b, rng)\n    const distance = Math.abs(sampled - threshold)\n    // Information-near-threshold weight: closer = higher.\n    // Use Gaussian-shaped kernel with σ tuned to posterior std.\n    const variance = (a * b) / ((a + b) ** 2 * (a + b + 1))\n    const sigma = Math.max(0.05, Math.sqrt(variance))\n    const weight = Math.exp(-((distance / sigma) ** 2))\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      n: cur.passes + cur.failures,\n      sampled,\n      sigma,\n      weight,\n      a,\n      b,\n    }\n  })\n\n  const totalWeight = stats.reduce((s, c) => s + c.weight, 0)\n  return stats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * opts.budget)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: Math.max(0, proportional),\n      reason: `Beta(${c.a.toFixed(1)},${c.b.toFixed(1)}) sample=${c.sampled.toFixed(3)} (target ${threshold})`,\n    }\n  })\n}\n\n/** Convenience: extract `CellObservation[]` directly from `RunRecord[]`. */\nexport function observationsFromRunRecords(\n  runs: RunRecord[],\n  opts: { passThreshold?: number; useHoldout?: boolean } = {},\n): CellObservation[] {\n  const threshold = opts.passThreshold ?? 0.5\n  const useHoldout = opts.useHoldout ?? true\n  const out: CellObservation[] = []\n  for (const r of runs) {\n    if (!r.scenarioId) continue\n    const score = useHoldout\n      ? (r.outcome.holdoutScore ?? r.outcome.searchScore)\n      : (r.outcome.searchScore ?? r.outcome.holdoutScore)\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    out.push({\n      variantId: r.candidateId,\n      scenarioId: r.scenarioId,\n      score,\n      pass: score >= threshold,\n    })\n  }\n  return out\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/**\n * Sample from Beta(α, β) via the Marsaglia–Tsang method using two Gamma\n * variates. Accuracy is good for α, β > 1; we floor the parameters at 1\n * to avoid degenerate cases.\n */\nfunction sampleBeta(alpha: number, beta: number, rng: () => number): number {\n  const a = Math.max(1, alpha)\n  const b = Math.max(1, beta)\n  const x = sampleGamma(a, rng)\n  const y = sampleGamma(b, rng)\n  return x / (x + y)\n}\n\nfunction sampleGamma(shape: number, rng: () => number): number {\n  // Marsaglia–Tsang for shape ≥ 1.\n  const d = shape - 1 / 3\n  const c = 1 / Math.sqrt(9 * d)\n  while (true) {\n    let x: number\n    let v: number\n    do {\n      const u1 = rng() || 1e-12\n      const u2 = rng() || 1e-12\n      // Box-Muller for a normal sample.\n      x = Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2)\n      v = 1 + c * x\n    } while (v <= 0)\n    v = v * v * v\n    const u = rng()\n    if (u < 1 - 0.0331 * x ** 4) return d * v\n    if (Math.log(u) < 0.5 * x * x + d * (1 - v + Math.log(v))) return d * v\n  }\n}\n","/**\n * Sample-efficient adaptation evaluation.\n *\n * For foundation-model-based agents, the load-bearing capability isn't\n * raw end-state performance — it's *how fast the agent reaches that\n * performance from cold start*. The same model with a worse prompt that\n * adapts in 5 demonstrations beats the same model with a better prompt\n * that needs 50. Standard meta-learning eval (Finn et al., MAML, RL² lit)\n * reports an *adaptation curve*: score after k=0, 1, 2, 4, 8, 16, …\n * in-context examples or fine-tune steps.\n *\n * This module ships:\n *\n *   1. `runAdaptationCurve` — given a runner that takes k demonstrations\n *      and returns a score, produce the (k, score) curve.\n *   2. `compareAdaptationCurves` — paired comparison across two policies.\n *      Returns per-k delta with bootstrap CIs and an \"area-under-curve\"\n *      summary statistic.\n *   3. `firstPassK` — for pass/fail evaluation, the minimum k at which\n *      the policy reliably passes (≥ pass-rate threshold over reps).\n *\n * Use cases:\n *   - Compare two prompt designs that have similar end-state performance\n *     but different in-context efficiency.\n *   - Decide between fine-tuning and prompting based on adaptation cost.\n *   - Detect when a policy \"memorizes\" k=0 inputs vs. genuinely adapts.\n */\n\nexport interface AdaptationRunner<S> {\n  /**\n   * Runs the policy on `scenario` with `k` demonstrations. Returns a\n   * scalar score in [0, 1]. The runner is responsible for any caching;\n   * the harness calls it once per (scenario, k, rep) cell.\n   */\n  run(args: { scenario: S; k: number; rep: number }): Promise<number>\n}\n\nexport interface RunAdaptationCurveOptions<S> {\n  scenarios: S[]\n  /** Number-of-shots to evaluate at. Default `[0, 1, 2, 4, 8, 16]`. */\n  ks?: number[]\n  /** Reps per (scenario, k) cell. Default 3. */\n  reps?: number\n  runner: AdaptationRunner<S>\n  /** Pass-rate threshold for `firstPassK` reporting. Default 0.5. */\n  passThreshold?: number\n}\n\nexport interface AdaptationPoint {\n  k: number\n  meanScore: number\n  passRate: number\n  std: number\n  n: number\n  /** Per-scenario means at this k. */\n  perScenario: Array<{ scenarioId: string; meanScore: number; passes: number; total: number }>\n}\n\nexport interface AdaptationCurve {\n  points: AdaptationPoint[]\n  /**\n   * Smallest `k` at which `passRate ≥ passThreshold`. `null` if no `k`\n   * tested reaches it.\n   */\n  firstPassK: number | null\n  /**\n   * Area under the (k, meanScore) curve, normalized by max-k. A\n   * single-number summary of \"how well does this policy adapt from\n   * cold-start to fully-conditioned.\" Higher = better adapter.\n   */\n  adaptationArea: number\n}\n\nexport async function runAdaptationCurve<S extends { scenarioId?: string }>(\n  opts: RunAdaptationCurveOptions<S>,\n): Promise<AdaptationCurve> {\n  const ks = opts.ks ?? [0, 1, 2, 4, 8, 16]\n  const reps = opts.reps ?? 3\n  const passThreshold = opts.passThreshold ?? 0.5\n  const sortedKs = [...ks].sort((a, b) => a - b)\n\n  const points: AdaptationPoint[] = []\n  for (const k of sortedKs) {\n    const perScenario: AdaptationPoint['perScenario'] = []\n    const allScores: number[] = []\n    let totalPasses = 0\n    let totalAttempts = 0\n    for (const scenario of opts.scenarios) {\n      const sid = scenario.scenarioId ?? `scenario-${opts.scenarios.indexOf(scenario)}`\n      const scores: number[] = []\n      let passes = 0\n      for (let r = 0; r < reps; r++) {\n        const score = await opts.runner.run({ scenario, k, rep: r })\n        scores.push(score)\n        if (score >= passThreshold) passes++\n        allScores.push(score)\n        if (score >= passThreshold) totalPasses++\n        totalAttempts++\n      }\n      const meanS = scores.reduce((s, v) => s + v, 0) / scores.length\n      perScenario.push({ scenarioId: sid, meanScore: meanS, passes, total: scores.length })\n    }\n    const meanScore = allScores.reduce((s, v) => s + v, 0) / Math.max(1, allScores.length)\n    const variance =\n      allScores.length < 2\n        ? 0\n        : allScores.reduce((s, v) => s + (v - meanScore) ** 2, 0) / (allScores.length - 1)\n    points.push({\n      k,\n      meanScore,\n      passRate: totalPasses / Math.max(1, totalAttempts),\n      std: Math.sqrt(variance),\n      n: allScores.length,\n      perScenario,\n    })\n  }\n\n  const firstPassK = points.find((p) => p.passRate >= passThreshold)?.k ?? null\n  const maxK = sortedKs[sortedKs.length - 1] ?? 1\n  // Trapezoidal area under the (k, meanScore) curve, normalized by k-range.\n  let area = 0\n  for (let i = 1; i < points.length; i++) {\n    const x1 = points[i - 1]!.k\n    const x2 = points[i]!.k\n    const y1 = points[i - 1]!.meanScore\n    const y2 = points[i]!.meanScore\n    area += ((y1 + y2) / 2) * (x2 - x1)\n  }\n  const adaptationArea = maxK === 0 ? 0 : area / maxK\n\n  return { points, firstPassK, adaptationArea }\n}\n\nexport interface CompareCurvesResult {\n  perK: Array<{\n    k: number\n    deltaMean: number\n    aLow: number\n    aHigh: number\n    bLow: number\n    bHigh: number\n  }>\n  areaDelta: number\n  firstPassKDelta: number | null\n  /** Verdict: 'a_better' | 'b_better' | 'similar'. */\n  verdict: 'a_better' | 'b_better' | 'similar'\n  /** Rationale, ready to render. */\n  rationale: string\n}\n\n/**\n * Paired comparison of two adaptation curves. Per-k deltas with 95%\n * bootstrap CIs (constructed from each curve's `perScenario` per-k means\n * — the bootstrap unit is the scenario, not the rep).\n */\nexport function compareAdaptationCurves(\n  a: AdaptationCurve,\n  b: AdaptationCurve,\n  opts: { confidence?: number; bootstrapResamples?: number; seed?: number } = {},\n): CompareCurvesResult {\n  const conf = opts.confidence ?? 0.95\n  const resamples = opts.bootstrapResamples ?? 500\n  const rng = makeRng(opts.seed)\n\n  const perK: CompareCurvesResult['perK'] = []\n  for (const ap of a.points) {\n    const bp = b.points.find((p) => p.k === ap.k)\n    if (!bp) continue\n    const aMeans = ap.perScenario.map((s) => s.meanScore)\n    const bMeans = bp.perScenario.map((s) => s.meanScore)\n    const aCi = bootstrapMeanCi(aMeans, resamples, conf, rng)\n    const bCi = bootstrapMeanCi(bMeans, resamples, conf, rng)\n    perK.push({\n      k: ap.k,\n      deltaMean: ap.meanScore - bp.meanScore,\n      aLow: aCi.low,\n      aHigh: aCi.high,\n      bLow: bCi.low,\n      bHigh: bCi.high,\n    })\n  }\n\n  const areaDelta = a.adaptationArea - b.adaptationArea\n  const firstPassKDelta =\n    a.firstPassK !== null && b.firstPassK !== null\n      ? b.firstPassK - a.firstPassK // smaller k for a means a adapts faster (positive delta)\n      : null\n\n  // Composite verdict: positive area delta + most per-k deltas in same\n  // direction → that side wins. Within ε of zero on both → similar.\n  const meanDelta = perK.reduce((s, p) => s + p.deltaMean, 0) / Math.max(1, perK.length)\n  let verdict: CompareCurvesResult['verdict']\n  if (Math.abs(meanDelta) < 0.02 && Math.abs(areaDelta) < 0.02) verdict = 'similar'\n  else if (meanDelta > 0 && areaDelta > 0) verdict = 'a_better'\n  else if (meanDelta < 0 && areaDelta < 0) verdict = 'b_better'\n  else verdict = 'similar'\n\n  const rationale =\n    `mean per-k delta=${meanDelta.toFixed(3)}, area delta=${areaDelta.toFixed(3)}` +\n    (firstPassKDelta !== null ? `, first-pass-k delta=${firstPassKDelta}` : '')\n\n  return { perK, areaDelta, firstPassKDelta, verdict, rationale }\n}\n\n/** First k at which the curve's per-scenario pass rate reliably hits the threshold. */\nexport function firstPassK(curve: AdaptationCurve, threshold = 0.5): number | null {\n  return curve.points.find((p) => p.passRate >= threshold)?.k ?? null\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\nfunction bootstrapMeanCi(\n  xs: number[],\n  resamples: number,\n  confidence: number,\n  rng: () => number,\n): { low: number; high: number } {\n  if (xs.length < 2) return { low: xs[0] ?? 0, high: xs[0] ?? 0 }\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let sum = 0\n    for (let i = 0; i < xs.length; i++) sum += xs[Math.floor(rng() * xs.length)]!\n    samples[b] = sum / xs.length\n  }\n  samples.sort((a, b) => a - b)\n  const alpha = 1 - confidence\n  return {\n    low: samples[Math.floor((alpha / 2) * resamples)]!,\n    high: samples[Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)]!,\n  }\n}\n","/**\n * Adversarial scenario search.\n *\n * Capability evaluation on a fixed scenario set measures performance on\n * the distribution someone curated. Production failure modes live in the\n * tail — inputs the curator didn't think of, or actively avoided. The\n * adversarial-search primitive actively looks for them: starting from a\n * pool of scenarios where the policy already passes, it mutates them\n * (paraphrase, edge-case substitution, compositional combination) and\n * keeps the mutations that *break* the policy.\n *\n * This is not magic. It's the simplest version of the loop that AdA\n * (Open-Ended Adaptation, DeepMind 2023), POET, and Anthropic's\n * auto-jailbreak rigs all run: hill-climb against a failure indicator,\n * keep the survivors, repeat. We ship the harness; consumers supply the\n * mutation strategies and the failure detector.\n *\n * Why ship this in agent-eval and not as a separate red-team tool: every\n * piece of the standard adversarial loop is already in this package\n * (`runEvalCampaign` for the matrix run, `RawProviderSink` for capture,\n * `assertRunCaptured` for integrity, `pairedEvalueSequence` for stop\n * criteria). The adversarial primitive is just the *scenario-mutation\n * meta-loop* on top of that machinery.\n */\n\nexport interface AdversarialScenario<S> {\n  /** Stable id — used for deduplication and lineage tracking. */\n  id: string\n  /** Generation index — 0 for seeds, 1 for first round of mutations, etc. */\n  generation: number\n  /** Lineage — id of the parent scenario this was mutated from, if any. */\n  parentId: string | null\n  scenario: S\n  /** Score on the policy under test. Lower = adversarial signal. */\n  score: number | null\n  /** Strategy that produced this mutation, for diagnostics. */\n  mutationStrategy: string | null\n}\n\nexport interface AdversarialMutation<S> {\n  id: string\n  /**\n   * Mutate one scenario. Return null to skip; return one or more new\n   * scenarios. The harness deduplicates by `mutateScenarioId(scenario)`.\n   */\n  mutate(parent: S, rng: () => number): Promise<S[]> | S[]\n}\n\nexport interface AdversarialSearchOptions<S> {\n  /** Initial scenarios — typically those the policy currently passes. */\n  seeds: S[]\n  /** Stable identifier extraction. */\n  mutateScenarioId: (s: S) => string\n  /** Mutation strategies. */\n  mutations: AdversarialMutation<S>[]\n  /**\n   * Run the policy under test against one scenario, return a scalar score\n   * in [0, 1]. Lower = adversarial signal.\n   */\n  scoreFn: (s: S) => Promise<number>\n  /**\n   * Threshold below which a scenario counts as a \"failure\" worth keeping.\n   * Default 0.5.\n   */\n  failureThreshold?: number\n  /** Number of mutation rounds. Default 3. */\n  rounds?: number\n  /** Children per parent per round. Default 4. */\n  childrenPerParent?: number\n  /** Maximum total scenarios examined. Default Infinity. */\n  budget?: number\n  /** Seed for the deterministic RNG. Default 1. */\n  seed?: number\n}\n\nexport interface AdversarialSearchReport<S> {\n  scenarios: AdversarialScenario<S>[]\n  /** Discovered failures sorted by score ascending. */\n  failures: AdversarialScenario<S>[]\n  /** Round-by-round counts. */\n  byGeneration: Array<{ generation: number; total: number; failures: number; meanScore: number }>\n  /** Total scoreFn invocations consumed. */\n  scoreCalls: number\n}\n\nexport async function adversarialScenarioSearch<S>(\n  opts: AdversarialSearchOptions<S>,\n): Promise<AdversarialSearchReport<S>> {\n  const failureThreshold = opts.failureThreshold ?? 0.5\n  const rounds = opts.rounds ?? 3\n  const children = opts.childrenPerParent ?? 4\n  const budget = opts.budget ?? Number.POSITIVE_INFINITY\n  const seed = opts.seed ?? 1\n  const rng = mulberry32(seed)\n\n  const scenarios: AdversarialScenario<S>[] = []\n  const seen = new Set<string>()\n  let scoreCalls = 0\n\n  // Seed generation.\n  for (const s of opts.seeds) {\n    const id = opts.mutateScenarioId(s)\n    if (seen.has(id)) continue\n    seen.add(id)\n    if (scoreCalls >= budget) break\n    const score = await opts.scoreFn(s)\n    scoreCalls++\n    scenarios.push({\n      id,\n      generation: 0,\n      parentId: null,\n      scenario: s,\n      score,\n      mutationStrategy: null,\n    })\n  }\n\n  // Mutation rounds.\n  for (let g = 1; g <= rounds; g++) {\n    if (scoreCalls >= budget) break\n    const parents = scenarios.filter((s) => s.generation === g - 1)\n    for (const parent of parents) {\n      for (const mutation of opts.mutations) {\n        if (scoreCalls >= budget) break\n        const produced = await mutation.mutate(parent.scenario, rng)\n        const childArr = Array.isArray(produced) ? produced : [produced]\n        for (let k = 0; k < Math.min(children, childArr.length); k++) {\n          if (scoreCalls >= budget) break\n          const child = childArr[k]!\n          const cid = opts.mutateScenarioId(child)\n          if (seen.has(cid)) continue\n          seen.add(cid)\n          const cscore = await opts.scoreFn(child)\n          scoreCalls++\n          scenarios.push({\n            id: cid,\n            generation: g,\n            parentId: parent.id,\n            scenario: child,\n            score: cscore,\n            mutationStrategy: mutation.id,\n          })\n        }\n      }\n    }\n  }\n\n  const failures = scenarios\n    .filter((s) => s.score !== null && s.score < failureThreshold)\n    .sort((a, b) => (a.score ?? 0) - (b.score ?? 0))\n\n  const byGeneration: AdversarialSearchReport<S>['byGeneration'] = []\n  const maxGen = scenarios.reduce((m, s) => Math.max(m, s.generation), 0)\n  for (let g = 0; g <= maxGen; g++) {\n    const gens = scenarios.filter((s) => s.generation === g)\n    if (gens.length === 0) continue\n    const fails = gens.filter((s) => s.score !== null && s.score < failureThreshold).length\n    const meanScore = gens.reduce((sum, s) => sum + (s.score ?? 0), 0) / gens.length\n    byGeneration.push({ generation: g, total: gens.length, failures: fails, meanScore })\n  }\n\n  return { scenarios, failures, byGeneration, scoreCalls }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mulberry32(seed: number): () => number {\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n","/**\n * RL corpus — the durable, append-only accumulation of graded RunRecords that\n * every eval run deposits BY DEFAULT.\n *\n * The dataset is the free exhaust of the normal eval process: we run evals\n * constantly to get an agent production-ready, and those runs already produce\n * graded trajectories. Instead of writing them to an ephemeral run dir and\n * throwing them away, `appendToCorpus` accumulates them into a durable corpus;\n * `buildDatasetFromCorpus` later harvests the whole corpus into a publishable\n * bundle. No separate data-collection campaign — the data accrues from work we\n * do anyway. This is the \"best things for free by our process\" layer.\n *\n * Trajectory text rides on the record as top-level `prompt` / `completion`\n * (what the eval harnesses capture; the RunRecord validator ignores the extra\n * keys). The harvest reads them directly — no trace store round-trip needed.\n */\n\nimport { appendFileSync, existsSync, mkdirSync, readFileSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport type { RunRecord } from '../run-record'\nimport { buildRlDataset, type RlDatasetBundle, type RlDatasetConfig } from './dataset'\n\n/** A corpus record is a RunRecord carrying the trajectory text the harness\n *  captured. `prompt`/`completion` are top-level (the validator ignores extras). */\nexport type CorpusRecord = RunRecord & { prompt?: string; completion?: string }\n\nexport interface CorpusAppendResult {\n  appended: number\n  /** Skipped because a record with the same runId was already in the corpus\n   *  (idempotent appends — NOT re-run collapsing; re-runs get fresh runIds). */\n  skipped: number\n  total: number\n}\n\n/**\n * Append graded records to the corpus (append-only JSONL). Deduplicates by\n * `runId` against what's already on disk so re-running the same harness is\n * idempotent. Creates the file and parent dir. This is the call every eval\n * harness makes by default after producing its records.\n */\nexport function appendToCorpus(records: CorpusRecord[], corpusPath: string): CorpusAppendResult {\n  mkdirSync(dirname(corpusPath), { recursive: true })\n  const existing = existsSync(corpusPath) ? readCorpus(corpusPath) : []\n  const seen = new Set(existing.map((r) => r.runId))\n  const lines: string[] = []\n  let appended = 0\n  let skipped = 0\n  for (const r of records) {\n    if (seen.has(r.runId)) {\n      skipped++\n      continue\n    }\n    seen.add(r.runId)\n    lines.push(JSON.stringify(r))\n    appended++\n  }\n  if (lines.length > 0) appendFileSync(corpusPath, `${lines.join('\\n')}\\n`)\n  return { appended, skipped, total: existing.length + appended }\n}\n\n/** Read the full corpus. Returns [] if the corpus does not exist yet. */\nexport function readCorpus(corpusPath: string): CorpusRecord[] {\n  if (!existsSync(corpusPath)) return []\n  const out: CorpusRecord[] = []\n  for (const line of readFileSync(corpusPath, 'utf8').split('\\n')) {\n    if (line.trim()) out.push(JSON.parse(line) as CorpusRecord)\n  }\n  return out\n}\n\nfunction rewardOf(r: CorpusRecord): number {\n  const v = r.outcome.holdoutScore ?? r.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : 0\n}\n\nexport interface HarvestOptions {\n  /** Keep only records scoring >= this (rejection-sampling for SFT). */\n  minScore?: number\n  /** Keep only these splits (e.g. ['holdout'] for an eval-only dataset). */\n  splits?: RunRecord['splitTag'][]\n}\n\n/**\n * Harvest the accumulated corpus into a publishable RL dataset bundle. Reads\n * trajectory text from each record's top-level `prompt`/`completion`; records\n * missing either are excluded (a graded score with no trajectory can't train).\n * Optionally filters by score / split. Throws (via buildRlDataset) if nothing\n * survives — an empty dataset must never be published.\n */\nexport async function buildDatasetFromCorpus(\n  corpusPath: string,\n  config: RlDatasetConfig,\n  opts: HarvestOptions = {},\n): Promise<RlDatasetBundle> {\n  let records = readCorpus(corpusPath).filter(\n    (r) => typeof r.prompt === 'string' && typeof r.completion === 'string',\n  )\n  if (opts.splits) records = records.filter((r) => opts.splits!.includes(r.splitTag))\n  if (opts.minScore != null) records = records.filter((r) => rewardOf(r) >= opts.minScore!)\n\n  const text = new Map(\n    records.map((r) => [r.runId, { prompt: r.prompt!, completion: r.completion! }]),\n  )\n  const lookups = {\n    promptOf: (id: string) => text.get(id)?.prompt ?? '',\n    completionOf: (id: string) => text.get(id)?.completion ?? '',\n  }\n  return buildRlDataset(records, lookups, config)\n}\n","/**\n * Trainer-format exporters.\n *\n * agent-eval produces canonical artifacts (`RunRecord[]`, `PreferenceTriple[]`,\n * `StepReward[]`, `PrmTrainingTriple[]`). RL training pipelines consume\n * different shapes — Hugging Face TRL, Prime Intellect's prime-rl, OpenAI\n * fine-tuning, Anthropic finetuning, OpenRLHF, verl. Each has its own\n * JSONL conventions. Rather than ship N adapters, this module ships the\n * canonical formats most production pipelines accept and ergonomic helpers\n * for the rest.\n *\n * Shapes:\n *   - **DPO / IPO / KTO** — `{prompt, chosen, rejected}` JSONL. Consumed\n *     by HuggingFace TRL, prime-rl's offline DPO, OpenRLHF.\n *   - **GRPO offline** — `{prompt, completions[], rewards[]}` JSONL.\n *     Consumed by prime-rl GRPO, verl, OpenRLHF.\n *   - **SFT** — `{messages[]}` JSONL with chosen completion as the final\n *     assistant turn. Consumed by HF SFT trainers, OpenAI fine-tuning,\n *     Anthropic finetuning.\n *   - **PRM** — `{prompt, prefix_steps[], chosen_step, rejected_step}` JSONL.\n *     Consumed by Lightman-style PRM trainers and prime-rl's PRM mode.\n *\n * Why ship this in agent-eval rather than a separate adapter package: the\n * canonical artifacts (`RunRecord[]`, `PreferenceTriple[]`, etc.) are\n * agent-eval's contract; without first-party exporters consumers reverse-\n * engineer the mapping every release. The exporters codify it.\n *\n * The exporters take callbacks for any field that isn't on the canonical\n * artifact (specifically: prompt + completion text, since the package\n * stores only their hashes by design — full text is the consumer's\n * trace store / raw event log).\n */\n\nimport type { RunRecord } from '../run-record'\nimport type { PreferenceTriple } from './preferences'\nimport type { PrmTrainingTriple, StepReward } from './process-reward'\n\n// ── DPO / IPO / KTO ──────────────────────────────────────────────────────\n\nexport interface DpoLookups {\n  /** Resolve the prompt text for a run (typically from a trace store / raw event sink). */\n  promptOf: (runId: string) => string | Promise<string>\n  /** Resolve the assistant completion text for a run. */\n  completionOf: (runId: string) => string | Promise<string>\n}\n\nexport interface DpoExportRow {\n  prompt: string\n  chosen: string\n  rejected: string\n  /** Carried-through margin. Some KTO / IPO variants use this. */\n  margin?: number\n  /** Free-form metadata for downstream filtering / sharding. */\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert preference triples to TRL-compatible DPO rows. The shape\n * `{prompt, chosen, rejected}` is the canonical HuggingFace DPODataset\n * entry; every major DPO trainer accepts it.\n */\nexport async function toDpoRows(\n  triples: PreferenceTriple[],\n  lookups: DpoLookups,\n): Promise<DpoExportRow[]> {\n  const out: DpoExportRow[] = []\n  for (const t of triples) {\n    const [prompt, chosen, rejected] = await Promise.all([\n      Promise.resolve(lookups.promptOf(t.chosenRunId)),\n      Promise.resolve(lookups.completionOf(t.chosenRunId)),\n      Promise.resolve(lookups.completionOf(t.rejectedRunId)),\n    ])\n    out.push({\n      prompt,\n      chosen,\n      rejected,\n      margin: t.marginScore,\n      meta: {\n        scenarioId: t.scenarioId,\n        chosenVariantId: t.chosenVariantId,\n        rejectedVariantId: t.rejectedVariantId,\n        chosenRunId: t.chosenRunId,\n        rejectedRunId: t.rejectedRunId,\n        chosenModel: t.meta.chosenModel,\n        rejectedModel: t.meta.rejectedModel,\n      },\n    })\n  }\n  return out\n}\n\n/** Serialize DPO rows as JSONL. One line per row. */\nexport function toDpoJsonl(rows: DpoExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── GRPO offline ─────────────────────────────────────────────────────────\n\nexport interface GrpoLookups {\n  promptOf: (runId: string) => string | Promise<string>\n  completionOf: (runId: string) => string | Promise<string>\n  /** Optional: derive a custom reward from the run. Defaults to score. */\n  rewardOf?: (run: RunRecord) => number | null\n}\n\nexport interface GrpoExportRow {\n  prompt: string\n  completions: string[]\n  rewards: number[]\n  /** runIds in the same order as `completions[]` for traceability. */\n  runIds: string[]\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert RunRecord[] grouped by `(scenarioId)` into GRPO offline rows —\n * one row per scenario, with one completion per run on that scenario.\n *\n * GRPO (Shao et al. 2024 / DeepSeek-R1) trains on relative advantages\n * within a group of completions for the same prompt; this is the\n * canonical input format.\n */\nexport async function toGrpoRows(\n  runs: RunRecord[],\n  lookups: GrpoLookups,\n): Promise<GrpoExportRow[]> {\n  const rewardOf = lookups.rewardOf ?? defaultReward\n  const grouped = new Map<string, RunRecord[]>()\n  for (const r of runs) {\n    const sid = r.scenarioId ?? r.experimentId\n    const arr = grouped.get(sid) ?? []\n    arr.push(r)\n    grouped.set(sid, arr)\n  }\n\n  const rows: GrpoExportRow[] = []\n  for (const [scenarioId, group] of grouped.entries()) {\n    if (group.length === 0) continue\n    // Resolve prompt once per group (assumes all runs in a group share the prompt).\n    const prompt = await Promise.resolve(lookups.promptOf(group[0]!.runId))\n    const completions: string[] = []\n    const rewards: number[] = []\n    const runIds: string[] = []\n    for (const r of group) {\n      const reward = rewardOf(r)\n      if (reward === null) continue\n      const completion = await Promise.resolve(lookups.completionOf(r.runId))\n      completions.push(completion)\n      rewards.push(reward)\n      runIds.push(r.runId)\n    }\n    if (completions.length === 0) continue\n    rows.push({\n      prompt,\n      completions,\n      rewards,\n      runIds,\n      meta: {\n        scenarioId,\n        n: completions.length,\n        meanReward: rewards.reduce((s, x) => s + x, 0) / rewards.length,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toGrpoJsonl(rows: GrpoExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── SFT ──────────────────────────────────────────────────────────────────\n\nexport interface SftLookups {\n  promptOf: (runId: string) => string | Promise<string>\n  completionOf: (runId: string) => string | Promise<string>\n  /** Optional system message. Default omits. */\n  systemOf?: (run: RunRecord) => string | null | undefined\n  /** Filter — return false to skip the run (e.g., low score, failed cases). */\n  include?: (run: RunRecord) => boolean\n}\n\nexport interface SftExportRow {\n  messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }>\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert RunRecord[] into Hugging Face / OpenAI / Anthropic-style\n * conversational SFT rows. By default every record becomes one row;\n * pass `include` to filter (e.g., keep only `score >= 0.8` for\n * rejection-sampling SFT).\n */\nexport async function toSftRows(runs: RunRecord[], lookups: SftLookups): Promise<SftExportRow[]> {\n  const include = lookups.include ?? (() => true)\n  const rows: SftExportRow[] = []\n  for (const r of runs) {\n    if (!include(r)) continue\n    const system = lookups.systemOf?.(r)\n    const [prompt, completion] = await Promise.all([\n      Promise.resolve(lookups.promptOf(r.runId)),\n      Promise.resolve(lookups.completionOf(r.runId)),\n    ])\n    const messages: SftExportRow['messages'] = []\n    if (system) messages.push({ role: 'system', content: system })\n    messages.push({ role: 'user', content: prompt })\n    messages.push({ role: 'assistant', content: completion })\n    rows.push({\n      messages,\n      meta: {\n        runId: r.runId,\n        candidateId: r.candidateId,\n        scenarioId: r.scenarioId,\n        score: r.outcome.holdoutScore ?? r.outcome.searchScore,\n        model: r.model,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toSftJsonl(rows: SftExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── PRM ──────────────────────────────────────────────────────────────────\n\nexport interface PrmLookups {\n  /** Resolve the prompt text for a run. */\n  promptOf: (runId: string) => string | Promise<string>\n  /** Resolve the trajectory step text for a (runId, spanId) pair. */\n  stepTextOf: (runId: string, spanId: string) => string | Promise<string>\n  /** Optional: sequence of prefix span ids leading up to the divergence. */\n  prefixOf?: (runId: string, prefixStepIndex: number) => string[] | Promise<string[]>\n}\n\nexport interface PrmExportRow {\n  prompt: string\n  /** Span ids for the steps before divergence — caller resolves text via `stepTextOf`. */\n  prefixSpanIds: string[]\n  prefixStepText: string[]\n  chosenStep: string\n  rejectedStep: string\n  chosenReward: number\n  rejectedReward: number\n  marginScore: number\n  meta?: Record<string, unknown>\n}\n\n/**\n * Convert PRM training triples to JSONL rows. Caller's `stepTextOf`\n * callback resolves span text from the consumer's trace store.\n */\nexport async function toPrmRows(\n  triples: PrmTrainingTriple[],\n  lookups: PrmLookups,\n): Promise<PrmExportRow[]> {\n  const rows: PrmExportRow[] = []\n  for (const t of triples) {\n    const prompt = await Promise.resolve(lookups.promptOf(t.prefixRunId))\n    const prefixSpanIds = lookups.prefixOf\n      ? await Promise.resolve(lookups.prefixOf(t.prefixRunId, t.prefixStepIndex))\n      : []\n    const prefixStepText: string[] = []\n    for (const spanId of prefixSpanIds) {\n      prefixStepText.push(await Promise.resolve(lookups.stepTextOf(t.prefixRunId, spanId)))\n    }\n    const chosenStep = await Promise.resolve(lookups.stepTextOf(t.prefixRunId, t.chosenSpanId))\n    const rejectedStep = await Promise.resolve(\n      lookups.stepTextOf(t.rejectedRunId, t.rejectedSpanId),\n    )\n    rows.push({\n      prompt,\n      prefixSpanIds,\n      prefixStepText,\n      chosenStep,\n      rejectedStep,\n      chosenReward: t.chosenReward,\n      rejectedReward: t.rejectedReward,\n      marginScore: t.marginScore,\n      meta: {\n        prefixRunId: t.prefixRunId,\n        rejectedRunId: t.rejectedRunId,\n        prefixStepIndex: t.prefixStepIndex,\n      },\n    })\n  }\n  return rows\n}\n\nexport function toPrmJsonl(rows: PrmExportRow[]): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── Step rewards (for value-function regression) ─────────────────────────\n\nexport interface StepRewardJsonlRow {\n  runId: string\n  spanId: string\n  stepIndex: number\n  reward: number\n  determinism: 'deterministic' | 'probabilistic'\n  weight: number\n}\n\nexport function stepRewardsToJsonl(stepRewards: StepReward[]): string {\n  const rows: StepRewardJsonlRow[] = stepRewards.map((s) => ({\n    runId: s.runId,\n    spanId: s.spanId,\n    stepIndex: s.stepIndex,\n    reward: s.reward,\n    determinism: s.determinism,\n    weight: s.weight ?? 1,\n  }))\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length > 0 ? '\\n' : '')\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction defaultReward(run: RunRecord): number | null {\n  const v = run.outcome.holdoutScore ?? run.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n","/**\n * RL dataset packaging + datasheet — the publishable, sellable bundle.\n *\n * The format exporters (`toGrpoRows` / `toSftRows` / `toDpoRows`) already\n * produce trainer-ready shapes (prime-rl GRPO, TRL DPO, conversational SFT).\n * What turns that into a dataset someone can PUBLISH or BUY is the provenance\n * + a datasheet: which models produced it, which prompt/agent versions, how the\n * reward was derived (deterministic verifiable vs probabilistic judge — the\n * credibility axis a buyer checks first), the split discipline, the reward\n * distribution, the quality gates, the license, and the intended/out-of-scope\n * uses. This module computes those facts from the `RunRecord[]` and renders a\n * \"Datasheet for Datasets\" (Gebru et al. 2018) card alongside the format files.\n *\n * It composes the existing `rl/exporters` — it does not reimplement any trainer\n * format. The renderers token-identity step (DeepSeek/Kimi/Qwen tokenization\n * with per-token loss masks) is a downstream Python stage that consumes the\n * `messages`/`completions` this bundle emits.\n */\n\nimport type { RunRecord, RunSplitTag } from '../run-record'\nimport {\n  type DpoLookups,\n  type GrpoLookups,\n  type SftLookups,\n  toDpoJsonl,\n  toDpoRows,\n  toGrpoJsonl,\n  toGrpoRows,\n  toSftJsonl,\n  toSftRows,\n} from './exporters'\nimport type { PreferenceTriple } from './preferences'\n\nexport type RewardKind = 'deterministic' | 'probabilistic' | 'mixed'\nexport type DatasetFormat = 'grpo' | 'sft' | 'dpo'\n\n/** Caller-declared context — the qualitative half of the datasheet that can't\n *  be computed from records. */\nexport interface RlDatasetConfig {\n  name: string\n  version: string\n  /** Product/task domain, e.g. 'legal-m&a', 'tax-1040'. */\n  domain: string\n  /** SPDX id or a named commercial license. Required — an unlicensed dataset\n   *  cannot be published or sold. */\n  license: string\n  /** How the reward was produced. `kind: 'deterministic'` (a test/schema/XPath\n   *  decided it) is the credibility signal; 'probabilistic' = LLM-judge. */\n  reward: { kind: RewardKind; source: string; description: string }\n  intendedUse: string\n  outOfScope: string\n  limitations: string\n  /** ISO timestamp — passed in (the substrate forbids Date.now()). */\n  createdAtIso: string\n  /** Default: ['grpo', 'sft']. */\n  formats?: DatasetFormat[]\n  /** Quality gates already run, recorded on the card for the buyer. */\n  qualityGates?: {\n    contaminationProbe?: 'passed' | 'failed' | 'not-run'\n    dedup?: boolean\n    verifiableRewardFilter?: boolean\n  }\n}\n\nexport interface RewardStats {\n  n: number\n  mean: number\n  median: number\n  min: number\n  max: number\n  std: number\n}\n\nexport interface RlDatasetStats {\n  records: number\n  /** Record count per split — a publishable dataset must declare its holdout. */\n  splits: Record<RunSplitTag, number>\n  reward: RewardStats\n  /** Distinct snapshot-pinned models that produced the trajectories. */\n  models: string[]\n  /** Distinct effective-prompt hashes (the agent profile/prompt versions). */\n  promptHashes: string[]\n  commitShas: string[]\n  totalTokens: { input: number; output: number }\n  totalCostUsd: number\n}\n\nexport interface RlDatasetManifest extends RlDatasetConfig {\n  formats: DatasetFormat[]\n  rowCounts: Partial<Record<DatasetFormat, number>>\n  stats: RlDatasetStats\n}\n\nexport interface RlDatasetBundle {\n  manifest: RlDatasetManifest\n  /** Relative filename -> contents. Write these to a directory to publish. */\n  files: Record<string, string>\n}\n\nfunction reward(r: RunRecord): number | null {\n  const v = r.outcome.holdoutScore ?? r.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nfunction distinct(xs: string[]): string[] {\n  return [...new Set(xs)].sort()\n}\n\nfunction computeRewardStats(values: number[]): RewardStats {\n  if (values.length === 0) return { n: 0, mean: 0, median: 0, min: 0, max: 0, std: 0 }\n  const sorted = [...values].sort((a, b) => a - b)\n  const n = sorted.length\n  const mean = sorted.reduce((s, x) => s + x, 0) / n\n  const mid = Math.floor(n / 2)\n  const median = n % 2 === 0 ? (sorted[mid - 1]! + sorted[mid]!) / 2 : sorted[mid]!\n  const variance = sorted.reduce((s, x) => s + (x - mean) ** 2, 0) / n\n  return { n, mean, median, min: sorted[0]!, max: sorted[n - 1]!, std: Math.sqrt(variance) }\n}\n\nfunction computeStats(records: RunRecord[]): RlDatasetStats {\n  const splits: Record<RunSplitTag, number> = { search: 0, dev: 0, holdout: 0 }\n  let inTok = 0\n  let outTok = 0\n  let cost = 0\n  const rewards: number[] = []\n  for (const r of records) {\n    splits[r.splitTag] = (splits[r.splitTag] ?? 0) + 1\n    inTok += r.tokenUsage.input\n    outTok += r.tokenUsage.output\n    cost += r.costUsd\n    const rw = reward(r)\n    if (rw !== null) rewards.push(rw)\n  }\n  return {\n    records: records.length,\n    splits,\n    reward: computeRewardStats(rewards),\n    models: distinct(records.map((r) => r.model)),\n    promptHashes: distinct(records.map((r) => r.promptHash)),\n    commitShas: distinct(records.map((r) => r.commitSha)),\n    totalTokens: { input: inTok, output: outTok },\n    totalCostUsd: cost,\n  }\n}\n\n/**\n * Package graded `RunRecord[]` into a publishable RL dataset bundle: the\n * trainer-format JSONL files + a manifest + a datasheet. DPO requires\n * pre-extracted preference triples (pass `preferences`); GRPO/SFT derive from\n * the records directly via the supplied lookups. Throws on an empty corpus —\n * an empty dataset must never be published.\n */\nexport async function buildRlDataset(\n  records: RunRecord[],\n  lookups: GrpoLookups & SftLookups,\n  config: RlDatasetConfig,\n  preferences?: { triples: PreferenceTriple[]; lookups: DpoLookups },\n): Promise<RlDatasetBundle> {\n  if (records.length === 0) {\n    throw new Error('buildRlDataset: no records — refusing to package an empty dataset')\n  }\n  const formats = config.formats ?? ['grpo', 'sft']\n  const files: Record<string, string> = {}\n  const rowCounts: Partial<Record<DatasetFormat, number>> = {}\n\n  if (formats.includes('grpo')) {\n    const rows = await toGrpoRows(records, lookups)\n    files['train.grpo.jsonl'] = toGrpoJsonl(rows)\n    rowCounts.grpo = rows.length\n  }\n  if (formats.includes('sft')) {\n    const rows = await toSftRows(records, lookups)\n    files['train.sft.jsonl'] = toSftJsonl(rows)\n    rowCounts.sft = rows.length\n  }\n  if (formats.includes('dpo')) {\n    if (!preferences) {\n      throw new Error(\"buildRlDataset: format 'dpo' requires `preferences` (triples + lookups)\")\n    }\n    const rows = await toDpoRows(preferences.triples, preferences.lookups)\n    files['train.dpo.jsonl'] = toDpoJsonl(rows)\n    rowCounts.dpo = rows.length\n  }\n\n  const manifest: RlDatasetManifest = {\n    ...config,\n    formats,\n    rowCounts,\n    stats: computeStats(records),\n  }\n  files['manifest.json'] = `${JSON.stringify(manifest, null, 2)}\\n`\n  files['DATASHEET.md'] = datasheetToMarkdown(manifest)\n  return { manifest, files }\n}\n\nfunction pct(x: number): string {\n  return `${(x * 100).toFixed(1)}%`\n}\n\n/** Render the \"Datasheet for Datasets\" card — the artifact a buyer reads. */\nexport function datasheetToMarkdown(m: RlDatasetManifest): string {\n  const s = m.stats\n  const total = s.records || 1\n  const splitLines = (['search', 'dev', 'holdout'] as RunSplitTag[])\n    .map((k) => `  - \\`${k}\\`: ${s.splits[k]} (${pct(s.splits[k] / total)})`)\n    .join('\\n')\n  const deterministic = m.reward.kind === 'deterministic'\n  return [\n    `# Dataset: ${m.name} \\`v${m.version}\\``,\n    '',\n    `**Domain:** ${m.domain} · **Created:** ${m.createdAtIso} · **License:** ${m.license}`,\n    '',\n    '## Reward provenance',\n    `- **Kind:** ${m.reward.kind}${deterministic ? ' ✅ (decidable — not judge-noise)' : ''}`,\n    `- **Source:** ${m.reward.source}`,\n    `- **Description:** ${m.reward.description}`,\n    '',\n    '## Composition',\n    `- **Records (trajectories):** ${s.records}`,\n    `- **Formats:** ${m.formats.map((f) => `${f} (${m.rowCounts[f] ?? 0} rows)`).join(', ')}`,\n    '- **Splits:**',\n    splitLines,\n    '',\n    '## Reward distribution',\n    `- n=${s.reward.n} · mean=${s.reward.mean.toFixed(3)} · median=${s.reward.median.toFixed(3)} · min=${s.reward.min.toFixed(3)} · max=${s.reward.max.toFixed(3)} · std=${s.reward.std.toFixed(3)}`,\n    '',\n    '## Provenance',\n    `- **Models:** ${s.models.join(', ')}`,\n    `- **Prompt/agent versions (sha256):** ${s.promptHashes.length} distinct`,\n    `- **Commits:** ${s.commitShas.join(', ')}`,\n    `- **Tokens:** ${s.totalTokens.input} in / ${s.totalTokens.output} out · **Cost:** $${s.totalCostUsd.toFixed(2)}`,\n    '',\n    '## Quality gates',\n    `- Contamination probe: ${m.qualityGates?.contaminationProbe ?? 'not-run'}`,\n    `- Dedup: ${m.qualityGates?.dedup ? 'yes' : 'no'} · Verifiable-reward filter: ${m.qualityGates?.verifiableRewardFilter ? 'yes' : 'no'}`,\n    '',\n    '## Recommended uses',\n    m.intendedUse,\n    '',\n    '## Out of scope',\n    m.outOfScope,\n    '',\n    '## Limitations',\n    m.limitations,\n    '',\n    '## Token rendering',\n    'For RL/SFT training, tokenize with the per-model renderer (DeepSeek-V3 / Kimi-K2 / Qwen3) to preserve token identity and per-token loss masks across tool-call turns — see `renderers` (PrimeIntellect). The `messages` / `completions` here are the renderer input.',\n    '',\n  ].join('\\n')\n}\n","/**\n * `PredictiveValidityResearcher` — concrete `Researcher` implementation\n * that drives selection from outcome-anchored predictive validity.\n *\n * Each method:\n *\n *   - `inspectFailures(runs)` — synthesizes failure modes from the\n *     bottom-quartile of `RunRecord`s on the configured proxy reward.\n *   - `proposeChange(failures)` — proposes steering changes that target\n *     the rubrics with the lowest predictive validity (decorative ones).\n *     Either reduce their weight in the composite, or recalibrate them.\n *   - `applyChange(changes, baseline)` — merges the proposed steering\n *     into the experiment plan.\n *   - `evaluateChange(plan)` — re-runs the predictive-validity check on\n *     the post-change runs and reports the delta.\n *\n * The result is a closed loop: the rubric weights drift toward the ones\n * that actually predict deployment outcomes, automatically. Pair with\n * `runRLCampaign` for the full auto-research story.\n */\n\nimport type { GateDecision } from '../held-out-gate'\nimport type { OutcomeStore } from '../meta-eval/outcome-store'\nimport {\n  type RubricPredictiveValidityReport,\n  rubricPredictiveValidity,\n} from '../meta-eval/rubric-predictive-validity'\nimport type {\n  ExperimentPlan,\n  ExperimentResult,\n  FailureMode,\n  Researcher,\n  SteeringChange,\n} from '../researcher'\nimport type { RunRecord } from '../run-record'\n\nexport interface PredictiveValidityResearcherOptions {\n  outcomes: OutcomeStore\n  outcomeMetrics: string[]\n  /** Score threshold below which a run counts as a \"failure.\" Default 0.5. */\n  failureThreshold?: number\n  /** Spearman bucket below which a rubric is \"decorative.\" Default 0.4. */\n  decorativeThreshold?: number\n  /** Optional steering-namespace prefix for proposed changes. Default `'rubric_weight'`. */\n  steeringNamespace?: string\n  /** Override the rubric set the researcher inspects. Default: every numeric `outcome.raw` key seen. */\n  rubrics?: string[]\n  /**\n   * Snapshot stash hook — called with the most recent predictive-validity\n   * report. Useful when a downstream system wants to log rubric drift over\n   * time. Default no-op.\n   */\n  onReport?: (report: RubricPredictiveValidityReport) => void | Promise<void>\n}\n\n/**\n * Concrete `Researcher` driven by `rubricPredictiveValidity`. The brain:\n * rubrics that don't predict deployment outcomes don't earn weight.\n */\nexport class PredictiveValidityResearcher implements Researcher {\n  private opts: PredictiveValidityResearcherOptions\n  private lastReport: RubricPredictiveValidityReport | null = null\n\n  constructor(opts: PredictiveValidityResearcherOptions) {\n    this.opts = opts\n  }\n\n  async inspectFailures(runs: RunRecord[]): Promise<FailureMode[]> {\n    const threshold = this.opts.failureThreshold ?? 0.5\n    const failures: FailureMode[] = []\n    const failingRuns = runs.filter((r) => {\n      const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n      return typeof score === 'number' && score < threshold\n    })\n    if (failingRuns.length === 0) return failures\n\n    // Group failures by candidateId — the researcher's primary handle is\n    // \"this candidate is producing low-scoring outputs in this scenario.\"\n    const grouped = new Map<string, RunRecord[]>()\n    for (const r of failingRuns) {\n      const arr = grouped.get(r.candidateId) ?? []\n      arr.push(r)\n      grouped.set(r.candidateId, arr)\n    }\n\n    for (const [candidateId, group] of grouped.entries()) {\n      const meanScore =\n        group.reduce((s, r) => {\n          const x = r.outcome.holdoutScore ?? r.outcome.searchScore ?? 0\n          return s + x\n        }, 0) / group.length\n      failures.push({\n        code: `low-score-${candidateId}`,\n        description: `${candidateId} scored < ${threshold} on ${group.length} run(s) (mean ${meanScore.toFixed(3)})`,\n        evidence: {\n          runIds: group.slice(0, 8).map((r) => r.runId),\n          samples: group.length,\n        },\n      })\n    }\n    return failures\n  }\n\n  async proposeChange(failures: FailureMode[]): Promise<SteeringChange[]> {\n    if (failures.length === 0) return []\n\n    // Without a prior report, return a single \"collect more outcome data\"\n    // change — the researcher refuses to reweight rubrics from zero evidence.\n    if (this.lastReport === null) {\n      return [\n        {\n          kind: 'threshold',\n          payload: { directive: 'researcher.collect-more-outcomes' },\n          rationale:\n            'predictive-validity researcher has no prior report; cannot recommend rubric reweighting until at least one report exists',\n        },\n      ]\n    }\n\n    const decorativeThreshold = this.opts.decorativeThreshold ?? 0.4\n    const changes: SteeringChange[] = []\n\n    for (const ranking of this.lastReport.ranked) {\n      if (ranking.verdict === 'load_bearing') continue\n      if (Math.abs(ranking.spearman) >= decorativeThreshold) continue\n      changes.push({\n        kind: 'reviewer_prompt',\n        payload: {\n          rubric: ranking.rubric,\n          action: 'down-weight',\n          spearman: ranking.spearman,\n          bestOutcome: ranking.bestOutcome,\n        },\n        rationale: `predictive-validity Spearman=${ranking.spearman.toFixed(3)} vs ${ranking.bestOutcome} (decorative); recommend down-weighting`,\n        expectedDelta: -Math.max(0, 0.05 - Math.abs(ranking.spearman)),\n      })\n    }\n    for (const ranking of this.lastReport.ranked.slice(0, 1)) {\n      if (ranking.verdict !== 'load_bearing') continue\n      changes.push({\n        kind: 'reviewer_prompt',\n        payload: {\n          rubric: ranking.rubric,\n          action: 'up-weight',\n          spearman: ranking.spearman,\n          bestOutcome: ranking.bestOutcome,\n        },\n        rationale: `predictive-validity Spearman=${ranking.spearman.toFixed(3)} vs ${ranking.bestOutcome} (load-bearing); recommend up-weighting`,\n        expectedDelta: Math.max(0, Math.abs(ranking.spearman) - 0.5) * 0.1,\n      })\n    }\n    return changes\n  }\n\n  async applyChange(changes: SteeringChange[], baseline: ExperimentPlan): Promise<ExperimentPlan> {\n    // Merge proposed changes into the plan's `changes` array, preserving\n    // any changes the baseline already had.\n    return {\n      ...baseline,\n      changes: [...baseline.changes, ...changes],\n    }\n  }\n\n  async evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult> {\n    // The researcher contract takes a *plan* and returns a *result* —\n    // implementations that only understand re-scoring runs can produce a\n    // \"no-op\" gate decision and let the caller drive the actual sweep.\n    // Real evaluators (CallbackResearcher) execute the plan; we report.\n    const emptyGate: GateDecision = {\n      promote: false,\n      candidateId: plan.proposedCandidateId,\n      baselineId: plan.baselineCandidateId,\n      evidence: {\n        productiveRuns: 0,\n        medianPairedDelta: 0,\n        pairedCI: { low: 0, high: 0 },\n        pairedPValue: 1,\n        searchScore: 0,\n        holdoutScore: 0,\n        overfitGap: 0,\n        baselineOverfitGap: 0,\n        medianCandidateCost: Number.NaN,\n        medianBaselineCost: Number.NaN,\n      },\n      reason:\n        'predictive-validity researcher does not execute plans; the caller is expected to run the sweep and call rubricPredictiveValidity directly with the resulting RunRecord[].',\n      rejectionCode: 'few_runs',\n    }\n    return {\n      plan,\n      runs: [],\n      gateDecision: emptyGate,\n    }\n  }\n\n  /**\n   * Run the predictive-validity check explicitly against a fresh RunRecord\n   * set. Updates the researcher's cached report so subsequent\n   * `proposeChange` calls have evidence to draw from.\n   */\n  async runValidityCheck(runs: RunRecord[]): Promise<RubricPredictiveValidityReport> {\n    const report = await rubricPredictiveValidity({\n      runs,\n      outcomes: this.opts.outcomes,\n      outcomeMetrics: this.opts.outcomeMetrics,\n      rubrics: this.opts.rubrics,\n    })\n    if (this.opts.onReport) await this.opts.onReport(report)\n    this.lastReport = report\n    return report\n  }\n\n  /**\n   * Force-feed a predictive-validity report into the researcher state —\n   * useful when the consumer ran the report out-of-band and wants the\n   * researcher's later proposals informed by it.\n   */\n  setReport(report: RubricPredictiveValidityReport): void {\n    this.lastReport = report\n  }\n\n  getLastReport(): RubricPredictiveValidityReport | null {\n    return this.lastReport\n  }\n}\n","/**\n * Process reward extraction — step-level credit assignment from trace spans.\n *\n * RL on long-horizon agents needs *step-level* rewards, not run-level\n * ones. The classic credit-assignment problem (Sutton & Barto) requires\n * knowing which sub-decisions in a trajectory contributed to the\n * outcome. Modern systems (DeepSeek-R1, OpenAI o-series, Lightman et al.\n * \"Let's Verify Step by Step\" 2023) train *process reward models* (PRMs)\n * that score every step, then do RL with the PRM as the reward signal.\n *\n * This module extracts `StepReward[]` from trace spans — one per\n * meaningful step — and ships:\n *\n *   1. `extractStepRewards(store, runId, opts)` — span → step-reward\n *      conversion using configurable per-span scorers (LLM judge over the\n *      span output, deterministic checkers, or a learned PRM).\n *   2. `runwiseStepRewardSummary(stepRewards)` — aggregate the per-step\n *      signal into a credit-assignment-aware run-level score.\n *   3. `prmTrainingPairs(stepRewards, options)` — produce the\n *      `(prefix, suffix_chosen, suffix_rejected)` triples that PRM\n *      training pipelines consume.\n *\n * What we ship: the *extraction* and *aggregation* infrastructure plus\n * the data shape PRM training expects. We do NOT ship the actual PRM\n * training (gradient descent over a transformer is out of scope for a\n * TS package). The interface is the contract; downstream consumers wire\n * their preferred trainer.\n *\n * Caveat the panel will land: this is descriptive credit assignment\n * (which steps correlate with outcome), not causal credit assignment\n * (which steps caused outcome). For causal claims you need\n * counterfactual rollouts or a learned dynamics model. Future work; the\n * descriptive version is what production PRM training actually uses.\n */\n\nimport type { Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface StepReward {\n  /** Trace span this reward attaches to. */\n  spanId: string\n  runId: string\n  /** Index in the trajectory (0-based, in started-at order). */\n  stepIndex: number\n  /** Span kind (typically 'tool', 'llm', 'judge'). */\n  kind: Span['kind']\n  /** Span name — for the consumer's downstream filtering. */\n  name: string\n  /** Step-level reward in [0, 1]. */\n  reward: number\n  /**\n   * Determinism class. Mirrors the verifiable-reward distinction:\n   * deterministic = test/compile/schema check; probabilistic = LLM judge.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /** Optional rationale / evidence — the trainer typically discards. */\n  rationale?: string\n  /** Optional weight — how much this step contributes to credit assignment. */\n  weight?: number\n}\n\nexport interface StepScorer {\n  /** Span kinds this scorer applies to. */\n  appliesTo: Span['kind'][]\n  /** Returns null to skip the span; returns a `StepReward` shape (without index/runId/spanId, which are filled in). */\n  score(span: Span): Promise<Omit<StepReward, 'spanId' | 'runId' | 'stepIndex'>> | null | undefined\n}\n\nexport interface ExtractStepRewardsOptions {\n  /**\n   * Ordered list of scorers. Each span runs through scorers in order;\n   * the first non-null result wins. If no scorer applies, the span is\n   * skipped (not all spans are training-worthy).\n   */\n  scorers: StepScorer[]\n  /** Optional filter — return null to drop the span entirely before scoring. */\n  preFilter?: (span: Span) => boolean\n}\n\nexport async function extractStepRewards(\n  store: TraceStore,\n  runId: string,\n  opts: ExtractStepRewardsOptions,\n): Promise<StepReward[]> {\n  const spans = await store.spans({ runId })\n  const ordered = [...spans].sort((a, b) => a.startedAt - b.startedAt)\n  const out: StepReward[] = []\n  let idx = 0\n  for (const span of ordered) {\n    if (opts.preFilter && !opts.preFilter(span)) continue\n    let scored: Awaited<ReturnType<StepScorer['score']>> = null\n    for (const s of opts.scorers) {\n      if (!s.appliesTo.includes(span.kind)) continue\n      const r = await s.score(span)\n      if (r) {\n        scored = r\n        break\n      }\n    }\n    if (!scored) continue\n    out.push({\n      spanId: span.spanId,\n      runId,\n      stepIndex: idx++,\n      kind: span.kind,\n      name: span.name,\n      reward: scored.reward,\n      determinism: scored.determinism,\n      rationale: scored.rationale,\n      weight: scored.weight,\n    })\n  }\n  return out\n}\n\nexport interface RunwiseStepSummary {\n  runId: string\n  totalSteps: number\n  meanReward: number\n  /** Sum-of-rewards (weighted by `weight ?? 1`). Use as the run-level proxy. */\n  sumWeightedReward: number\n  /** Fraction of steps where reward < 0.5 — proxy for \"where the policy was wrong.\" */\n  failureFraction: number\n  /** Maximum drop in reward between consecutive steps — diagnoses a step where things went sideways. */\n  worstStepDelta: number\n  worstStepIndex: number | null\n}\n\nexport function runwiseStepRewardSummary(stepRewards: StepReward[]): RunwiseStepSummary {\n  if (stepRewards.length === 0) {\n    return {\n      runId: '',\n      totalSteps: 0,\n      meanReward: 0,\n      sumWeightedReward: 0,\n      failureFraction: 0,\n      worstStepDelta: 0,\n      worstStepIndex: null,\n    }\n  }\n  const runId = stepRewards[0]!.runId\n  let sumW = 0\n  let sumWR = 0\n  let failures = 0\n  let worstDelta = 0\n  let worstIdx: number | null = null\n  let prev = stepRewards[0]!.reward\n  for (let i = 0; i < stepRewards.length; i++) {\n    const s = stepRewards[i]!\n    const w = s.weight ?? 1\n    sumW += w\n    sumWR += w * s.reward\n    if (s.reward < 0.5) failures++\n    if (i > 0) {\n      const delta = s.reward - prev\n      if (delta < worstDelta) {\n        worstDelta = delta\n        worstIdx = i\n      }\n      prev = s.reward\n    } else {\n      prev = s.reward\n    }\n  }\n  return {\n    runId,\n    totalSteps: stepRewards.length,\n    meanReward: sumW === 0 ? 0 : sumWR / sumW,\n    sumWeightedReward: sumWR,\n    failureFraction: failures / stepRewards.length,\n    worstStepDelta: worstDelta,\n    worstStepIndex: worstIdx,\n  }\n}\n\nexport interface PrmTrainingTriple {\n  /** Prefix run-id (or composite key) — the trajectory up to step k-1. */\n  prefixRunId: string\n  prefixStepIndex: number\n  /** The step that came next on a high-reward trajectory. */\n  chosenSpanId: string\n  chosenReward: number\n  /** A step from a divergent low-reward trajectory at the same prefix length. */\n  rejectedSpanId: string\n  rejectedReward: number\n  /** The prefix run came from this run; the rejected step came from `rejectedRunId`. */\n  rejectedRunId: string\n  marginScore: number\n}\n\n/**\n * Build PRM training triples. The shape: pair runs that share an early\n * prefix (same scenario, same first N steps) and diverge later — at the\n * point of divergence, the high-reward run's next step is `chosen`, the\n * low-reward run's next step is `rejected`. This is the canonical PRM\n * training data shape from Lightman et al. and DeepSeek-R1 process\n * supervision.\n *\n * Implementation note: we don't have a way to detect \"same prefix\" in\n * the general agent setting (token-level prefixes require hashing model\n * outputs). The current heuristic groups by `(scenarioId, prefixSpanName\n * sequence)` — runs are paired when their first K span names match. For\n * production use this should be replaced with a proper trajectory-prefix\n * hash; the heuristic is good enough for early-stage scaffolding.\n */\nexport function prmTrainingPairs(\n  stepRewardsByRun: Map<string, StepReward[]>,\n  opts: { minMargin?: number; minPrefixLength?: number } = {},\n): PrmTrainingTriple[] {\n  const minMargin = opts.minMargin ?? 0.2\n  const minPrefix = opts.minPrefixLength ?? 1\n  const runs = [...stepRewardsByRun.entries()].map(([runId, steps]) => ({ runId, steps }))\n  const triples: PrmTrainingTriple[] = []\n\n  for (let i = 0; i < runs.length; i++) {\n    for (let j = i + 1; j < runs.length; j++) {\n      const a = runs[i]!\n      const b = runs[j]!\n      const minLen = Math.min(a.steps.length, b.steps.length)\n      if (minLen < minPrefix + 1) continue\n\n      // Find the first index where the trajectories diverge: either by\n      // step structure (kind/name mismatch) OR by reward gap ≥ minMargin.\n      // Names that match but rewards that differ ARE divergence — that's\n      // the canonical PRM training case (same step structure, different\n      // outcomes via state/context).\n      let divergenceIdx = -1\n      for (let k = 0; k < minLen; k++) {\n        const sa = a.steps[k]!\n        const sb = b.steps[k]!\n        const structuralDivergence = sa.kind !== sb.kind || sa.name !== sb.name\n        const rewardGap = Math.abs(sa.reward - sb.reward)\n        if (structuralDivergence || rewardGap >= minMargin) {\n          divergenceIdx = k\n          break\n        }\n      }\n      if (divergenceIdx < 0) continue\n      if (divergenceIdx < minPrefix) continue\n\n      const aNext = a.steps[divergenceIdx]!\n      const bNext = b.steps[divergenceIdx]!\n      const margin = Math.abs(aNext.reward - bNext.reward)\n      if (margin < minMargin) continue\n\n      const chosen = aNext.reward > bNext.reward ? aNext : bNext\n      const rejected = aNext.reward > bNext.reward ? bNext : aNext\n      const chosenRun = aNext.reward > bNext.reward ? a.runId : b.runId\n      const rejectedRun = aNext.reward > bNext.reward ? b.runId : a.runId\n      triples.push({\n        prefixRunId: chosenRun,\n        prefixStepIndex: divergenceIdx - 1,\n        chosenSpanId: chosen.spanId,\n        chosenReward: chosen.reward,\n        rejectedSpanId: rejected.spanId,\n        rejectedReward: rejected.reward,\n        rejectedRunId: rejectedRun,\n        marginScore: chosen.reward - rejected.reward,\n      })\n    }\n  }\n  return triples\n}\n","/**\n * `runRLCampaign` — top-level orchestrator that runs the matrix and\n * produces every RL-ready artifact in one call.\n *\n * Wires:\n *   1. `runEvalCampaign` for the matrix run (capture, integrity, hooks)\n *   2. `extractVerifiableReward` over each run, separating deterministic\n *      from probabilistic reward sources for the trainer\n *   3. `extractPreferences` to produce DPO/PPO/KTO triples\n *   4. `evaluateInterimReleaseConfidence` over paired deltas (anytime-valid)\n *   5. `rubricPredictiveValidity` against an outcome store, when provided\n *   6. `detectRewardHacking` as a standing hygiene check\n *   7. Trainer-format export rows ready for prime-rl / TRL / verl\n *\n * The output `RLCampaignResult` is a single, audit-ready artifact: every\n * stage's output is in there. The consumer's downstream fits in a single\n * line: pass `result.preferences` to their DPO trainer, `result.grpoRows`\n * to GRPO, `result.runs` plus `result.rewardSignals` to a custom RL loop.\n */\n\nimport {\n  type EvalCampaignOptions,\n  type EvalCampaignResult,\n  runEvalCampaign,\n} from '../eval-campaign'\nimport type { OutcomeStore } from '../meta-eval/outcome-store'\nimport {\n  type RubricPredictiveValidityReport,\n  rubricPredictiveValidity,\n} from '../meta-eval/rubric-predictive-validity'\nimport type { RunRecord } from '../run-record'\nimport { evaluateInterimReleaseConfidence, type InterimReleaseConfidence } from '../sequential'\nimport {\n  type DpoExportRow,\n  type DpoLookups,\n  type GrpoExportRow,\n  type GrpoLookups,\n  type SftExportRow,\n  type SftLookups,\n  toDpoRows,\n  toGrpoRows,\n  toSftRows,\n} from './exporters'\nimport {\n  type ExtractPreferencesOptions,\n  extractPreferences,\n  type PreferenceExtractionReport,\n} from './preferences'\nimport { detectRewardHacking, type RewardHackingReport } from './reward-hacking'\nimport {\n  extractVerifiableRewardsFromRecords,\n  type VerifiableReward,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport interface RunRLCampaignOptions<V> extends EvalCampaignOptions<V> {\n  /** Preference-extraction options. Default uses paired-by-scenario-and-seed with min-margin 0.05. */\n  preferences?: ExtractPreferencesOptions\n  /** Verifiable-reward extraction options. */\n  verifiableReward?: VerifiableRewardExtractionOptions\n  /** Outcome store + metric names — when supplied, runs `rubricPredictiveValidity` post-campaign. */\n  outcomeStore?: OutcomeStore\n  outcomeMetrics?: string[]\n  /** Anytime-valid sequential evaluation options. */\n  sequential?: { alpha?: number; bound?: number; rope?: { low: number; high: number } }\n  /** Trainer-format export lookups. When provided, the orchestrator builds the corresponding rows. */\n  trainerExport?: {\n    dpo?: DpoLookups\n    grpo?: GrpoLookups\n    sft?: SftLookups\n  }\n}\n\nexport interface RLCampaignResult<V> {\n  campaign: EvalCampaignResult\n  /** Per-run verifiable reward (deterministic when available, probabilistic fallback otherwise). */\n  rewardSignals: Array<{ runId: string; reward: VerifiableReward | null }>\n  /** Preference extraction report. */\n  preferences: PreferenceExtractionReport\n  /** Anytime-valid interim verdict over the paired deltas (vs comparator). */\n  interimConfidence: InterimReleaseConfidence | null\n  /** Standing reward-hacking hygiene check. */\n  rewardHacking: RewardHackingReport\n  /** Predictive validity, when an outcome store was supplied. */\n  predictiveValidity: RubricPredictiveValidityReport | null\n  /** Trainer-export rows, populated only for the formats the caller requested via `trainerExport`. */\n  trainerRows: {\n    dpo?: DpoExportRow[]\n    grpo?: GrpoExportRow[]\n    sft?: SftExportRow[]\n  }\n  /**\n   * One-line top-level summary the consumer can log.\n   */\n  summary: string\n  /**\n   * Convenience type-tag — consumers can branch on `result.kind`.\n   */\n  kind: 'agent-eval-rl-campaign'\n  unusedVariant?: V\n}\n\nexport async function runRLCampaign<V>(\n  opts: RunRLCampaignOptions<V>,\n): Promise<RLCampaignResult<V>> {\n  // ── 1. Run the matrix ──────────────────────────────────────────────\n  const campaign = await runEvalCampaign(opts)\n\n  // ── 2. Extract reward signals (deterministic-first) ────────────────\n  const rewardSignals = extractVerifiableRewardsFromRecords(\n    campaign.runs,\n    opts.verifiableReward ?? {},\n  )\n\n  // ── 3. Extract preference triples ──────────────────────────────────\n  const preferences = extractPreferences(campaign.runs, {\n    strategy: opts.preferences?.strategy ?? 'paired-by-scenario-and-seed',\n    minMargin: opts.preferences?.minMargin ?? 0.05,\n    splitTag: opts.preferences?.splitTag ?? opts.splitTag ?? 'holdout',\n    rewardOf: opts.preferences?.rewardOf,\n  })\n\n  // ── 4. Sequential / anytime-valid interim verdict ──────────────────\n  let interimConfidence: InterimReleaseConfidence | null = null\n  if (opts.report?.comparator) {\n    const comparator = opts.report.comparator\n    const deltaSeries = collectPairedDeltaSeries(campaign.runs, comparator)\n    if (deltaSeries.some((s) => s.deltas.length > 0)) {\n      interimConfidence = evaluateInterimReleaseConfidence({\n        deltaSeries,\n        alpha: opts.sequential?.alpha,\n        bound: opts.sequential?.bound,\n        rope: opts.sequential?.rope ?? opts.report?.rope,\n      })\n    }\n  }\n\n  // ── 5. Standing reward-hacking hygiene ─────────────────────────────\n  const rewardHacking = detectRewardHacking({\n    runs: campaign.runs,\n    verifiableRewardOptions: opts.verifiableReward,\n  })\n\n  // ── 6. Predictive validity (when outcomes are supplied) ────────────\n  let predictiveValidity: RubricPredictiveValidityReport | null = null\n  if (opts.outcomeStore && opts.outcomeMetrics && opts.outcomeMetrics.length > 0) {\n    predictiveValidity = await rubricPredictiveValidity({\n      runs: campaign.runs,\n      outcomes: opts.outcomeStore,\n      outcomeMetrics: opts.outcomeMetrics,\n    })\n  }\n\n  // ── 7. Trainer-format export ───────────────────────────────────────\n  const trainerRows: RLCampaignResult<V>['trainerRows'] = {}\n  if (opts.trainerExport?.dpo) {\n    trainerRows.dpo = await toDpoRows(preferences.pairs, opts.trainerExport.dpo)\n  }\n  if (opts.trainerExport?.grpo) {\n    trainerRows.grpo = await toGrpoRows(campaign.runs, opts.trainerExport.grpo)\n  }\n  if (opts.trainerExport?.sft) {\n    trainerRows.sft = await toSftRows(campaign.runs, opts.trainerExport.sft)\n  }\n\n  const summary = buildSummary({\n    campaign,\n    preferences,\n    interimConfidence,\n    rewardHacking,\n    predictiveValidity,\n  })\n\n  return {\n    campaign,\n    rewardSignals,\n    preferences,\n    interimConfidence,\n    rewardHacking,\n    predictiveValidity,\n    trainerRows,\n    summary,\n    kind: 'agent-eval-rl-campaign',\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction collectPairedDeltaSeries(\n  runs: RunRecord[],\n  comparator: string,\n): Array<{ candidateId: string; deltas: number[] }> {\n  // Pair on (scenarioId, seed). For each candidate that isn't the comparator,\n  // compute candidate.score - comparator.score on matching cells.\n  const baseline = new Map<string, number>()\n  for (const r of runs) {\n    if (r.candidateId !== comparator) continue\n    const sid = r.scenarioId ?? r.experimentId\n    const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    baseline.set(`${sid}::${r.seed}`, score)\n  }\n  const byCandidate = new Map<string, number[]>()\n  for (const r of runs) {\n    if (r.candidateId === comparator) continue\n    const sid = r.scenarioId ?? r.experimentId\n    const score = r.outcome.holdoutScore ?? r.outcome.searchScore\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    const baseScore = baseline.get(`${sid}::${r.seed}`)\n    if (typeof baseScore !== 'number') continue\n    const arr = byCandidate.get(r.candidateId) ?? []\n    arr.push(score - baseScore)\n    byCandidate.set(r.candidateId, arr)\n  }\n  return [...byCandidate.entries()].map(([candidateId, deltas]) => ({ candidateId, deltas }))\n}\n\nfunction buildSummary(args: {\n  campaign: EvalCampaignResult\n  preferences: PreferenceExtractionReport\n  interimConfidence: InterimReleaseConfidence | null\n  rewardHacking: RewardHackingReport\n  predictiveValidity: RubricPredictiveValidityReport | null\n}): string {\n  const c = args.campaign\n  const lines = [\n    `${c.campaignId}: ${c.runs.length} successful runs / ${c.failedRuns.length} failed (fingerprint ${c.campaignFingerprint.slice(0, 12)}…)`,\n    `preferences: ${args.preferences.pairs.length} (${args.preferences.strategy}, ${args.preferences.pairsBelowMargin} below margin)`,\n  ]\n  if (args.interimConfidence) {\n    lines.push(\n      `sequential verdict: ${args.interimConfidence.recommendation.decision}` +\n        (args.interimConfidence.recommendation.candidateId\n          ? ` ${args.interimConfidence.recommendation.candidateId}`\n          : ''),\n    )\n  }\n  lines.push(\n    `reward-hacking: ${args.rewardHacking.verdict} (${args.rewardHacking.findings.length} signals checked)`,\n  )\n  if (args.predictiveValidity) {\n    const top = args.predictiveValidity.ranked[0]\n    lines.push(\n      `top-rubric: ${top?.rubric ?? 'none'} ρ=${(top?.spearman ?? 0).toFixed(2)} (${top?.verdict ?? 'no data'})`,\n    )\n  }\n  return lines.join(' | ')\n}\n\n// Re-export `runEvalCampaign` so consumers can pick the lower-level\n// primitive without flipping import paths.\nexport { runEvalCampaign } from '../eval-campaign'\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8EA,eAAsB,gBAAgB,MAAqD;AACzF,QAAM,SAA8B,CAAC;AACrC,aAAW,UAAU,KAAK,SAAS;AACjC,UAAM,IAAI,MAAM,KAAK,YAAY,MAAM;AACvC,WAAO,KAAK;AAAA,MACV,UAAU,OAAO;AAAA,MACjB,MAAM,OAAO;AAAA,MACb,OAAO,EAAE;AAAA,MACT,SAAS,EAAE;AAAA,MACX,KAAK,EAAE;AAAA,MACP,SAAS,EAAE;AAAA,IACb,CAAC;AAAA,EACH;AACA,QAAM,SAAS,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AACzD,QAAM,WAAW,OAAO,UAAU,IAAI,YAAY,MAAM,IAAI;AAC5D,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAO,EAAE,QAAQ,EAAE,QAAQ,IAAI,CAAE;AAChE,SAAO,EAAE,aAAa,KAAK,aAAa,QAAQ,QAAQ,UAAU,KAAK;AACzE;AAqBA,eAAsB,QAAW,MAAkE;AACjG,MAAI,KAAK,KAAK,EAAG,OAAM,IAAI,gBAAgB,wBAAwB;AACnE,QAAM,WAAgB,CAAC;AACvB,QAAM,SAAmB,CAAC;AAC1B,WAAS,IAAI,GAAG,IAAI,KAAK,GAAG,KAAK;AAC/B,UAAM,IAAI,MAAM,KAAK,OAAO,CAAC;AAC7B,aAAS,KAAK,CAAC;AACf,WAAO,KAAK,MAAM,KAAK,QAAQ,CAAC,CAAC;AAAA,EACnC;AACA,MAAI,YAAY;AAChB,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,IAAK,KAAI,OAAO,CAAC,IAAK,OAAO,SAAS,EAAI,aAAY;AACzF,QAAM,YAAY,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAC7D,SAAO;AAAA,IACL,MAAM,SAAS,SAAS;AAAA,IACxB,WAAW,OAAO,SAAS;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AA0BA,eAAsB,gBACpB,MACmC;AACnC,MAAI,KAAK,KAAK,EAAG,OAAM,IAAI,gBAAgB,gCAAgC;AAC3E,QAAM,WAAgB,CAAC;AACvB,QAAM,YAAoC,CAAC;AAC3C,WAAS,IAAI,GAAG,IAAI,KAAK,GAAG,KAAK;AAC/B,UAAM,IAAI,MAAM,KAAK,OAAO,CAAC;AAC7B,aAAS,KAAK,CAAC;AACf,UAAM,MAAM,KAAK,UAAU,CAAC;AAC5B,cAAU,GAAG,KAAK,UAAU,GAAG,KAAK,KAAK;AAAA,EAC3C;AACA,MAAI,SAAS;AACb,MAAI,MAAM;AACV,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,SAAS,GAAG;AAC9C,QAAI,IAAI,KAAK;AACX,YAAM;AACN,eAAS;AAAA,IACX;AAAA,EACF;AACA,QAAM,iBAAiB,SAAS,KAAK,CAAC,MAAM,KAAK,UAAU,CAAC,MAAM,MAAM,KAAK,SAAS,CAAC;AACvF,SAAO;AAAA,IACL;AAAA,IACA,WAAW,MAAM,KAAK;AAAA,IACtB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAeO,SAAS,eAAe,QAAgD;AAC7E,QAAM,aAAiC,CAAC;AACxC,aAAW,KAAK,QAAQ;AACtB,UAAM,YAAY,OAAO;AAAA,MACvB,CAAC,MACC,MAAM,KAAK,EAAE,QAAQ,EAAE,QAAQ,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE;AAAA,IACzF;AACA,QAAI,CAAC,UAAW,YAAW,KAAK,CAAC;AAAA,EACnC;AACA,SAAO,WAAW,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClD;AAIA,SAAS,YAAY,QAAqC;AAIxD,QAAM,KAAK,OAAO,IAAI,CAAC,MAAM,KAAK,IAAI,KAAK,IAAI,OAAO,EAAE,IAAI,CAAC,CAAC;AAC9D,QAAM,KAAK,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK;AACpC,QAAM,IAAI,GAAG;AACb,QAAM,KAAK,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3C,QAAM,KAAK,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3C,MAAI,MAAM;AACV,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAQ,GAAG,CAAC,IAAK,OAAO,GAAG,CAAC,IAAK;AACjC,YAAQ,GAAG,CAAC,IAAK,OAAO;AAAA,EAC1B;AACA,SAAO,QAAQ,IAAI,IAAI,MAAM;AAC/B;;;ACrIA,eAAsB,sBACpB,OACA,OAAkC,CAAC,GACA;AACnC,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,QAAQ,KAAK,cAAc;AAEjC,MAAI,CAAC,MAAM,aAAa,CAAC,MAAM,cAAc;AAC3C,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,YACJ,MAAM,aAAc,MAAM,QAAQ,IAAI,MAAM,UAAU,IAAI,CAAC,MAAM,MAAM,aAAc,MAAM,CAAC,CAAC,CAAC;AAChG,MAAI,UAAU,WAAW,MAAM,UAAU,QAAQ;AAC/C,UAAM,IAAI;AAAA,MACR,2CAA2C,UAAU,MAAM,qBAAgB,MAAM,UAAU,MAAM;AAAA,IACnG;AAAA,EACF;AAGA,QAAM,aAAa,MAAM,QAAQ,IAAI,MAAM,UAAU,IAAI,CAAC,MAAM,MAAM,QAAQ,CAAC,CAAC,CAAC;AACjF,QAAM,aAAa,MAAM,QAAQ,IAAI,UAAU,IAAI,CAAC,MAAM,MAAM,QAAQ,CAAC,CAAC,CAAC;AAE3E,QAAM,cAAc,MAAM,UAAU,IAAI,CAAC,GAAG,OAAO;AAAA,IACjD,YAAY,MAAM,WAAW,CAAC;AAAA,IAC9B,eAAe,WAAW,CAAC;AAAA,IAC3B,gBAAgB,WAAW,CAAC;AAAA,IAC5B,OAAO,WAAW,CAAC,IAAK,WAAW,CAAC;AAAA,IACpC,QAAQ;AAAA,EACV,EAAE;AAGF,QAAM,QAAQ,YAAY,OAAO,CAAC,MAAM,EAAE,iBAAiB,SAAS,EAAE,kBAAkB,KAAK;AAC7F,MAAI,MAAM,SAAS,GAAG;AACpB,WAAO;AAAA,MACL;AAAA,MACA,YAAY,EAAE,GAAG,GAAG,GAAG,EAAE;AAAA,MACzB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,wBAAwB;AAAA,MACxB,QAAQ,mCAAmC,MAAM,MAAM;AAAA,MACvD,GAAG,MAAM;AAAA,IACX;AAAA,EACF;AAEA,QAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,aAAa;AAClD,QAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,cAAc;AACnD,QAAM,aAAa,mBAAmB,WAAW,SAAS;AAC1D,QAAM,SAAS,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AACvC,QAAM,eAAe,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACrD,QAAM,SAAS,aAAa,KAAK,MAAM,aAAa,SAAS,CAAC,CAAC;AAC/D,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAMxD,QAAM,UAAU,MAAM,IAAI,CAAC,MAAM,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,IAAI,KAAK,IAAI,EAAE,KAAK,IAAI,CAAC,CAAC,CAAC;AACvF,QAAM,EAAE,QAAQ,IAAI,kBAAkB,SAAS,GAAG;AAClD,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAM,IAAI,MAAM,CAAC;AACjB,UAAM,MAAM,YAAY,UAAU,CAAC,MAAM,EAAE,eAAe,EAAE,UAAU;AACtE,QAAI,OAAO,EAAG,aAAY,GAAG,EAAG,SAAS,QAAQ,CAAC;AAAA,EACpD;AAEA,QAAM,yBAAyB,WAAW,IAAI,OAAO,UAAU,CAAC;AAChE,QAAM,SAAS,yBACX,YAAY,WAAW,EAAE,QAAQ,CAAC,CAAC,MAAM,GAAG,oBAAoB,OAAO,QAAQ,CAAC,CAAC,WAAM,aAAa,KACpG,WAAW,KAAK,MACd,uCAAuC,WAAW,EAAE,QAAQ,CAAC,CAAC,MAC9D,8CAA8C,OAAO,QAAQ,CAAC,CAAC;AAErE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,aAAa;AAAA,IACb,WAAW;AAAA,IACX;AAAA,IACA;AAAA,IACA,GAAG,MAAM;AAAA,EACX;AACF;AAUO,SAAS,gBACd,aACA,SAAgD,CAAC,GAAG,MAAM,GAAG,CAAC,KAAM,IAAI,KAAM,IAAI,SAAS,EAAE,CAAC,IACrE;AACzB,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,UAAI,SAAS,SAAS;AACtB,kBAAY,QAAQ,CAAC,IAAI,MAAM;AAC7B,cAAM,cAAc,OAAO,IAAI,CAAC;AAChC,cAAM,KAAK,IAAI,OAAO,MAAM,YAAY,EAAE,CAAC,OAAO,GAAG;AACrD,iBAAS,OAAO,QAAQ,IAAI,WAAW;AAAA,MACzC,CAAC;AACD,aAAO,EAAE,GAAG,UAAU,OAAO;AAAA,IAC/B;AAAA,EACF;AACF;AAOO,SAAS,aACd,gBACA,MACyB;AACzB,MAAI,IAAI,SAAS;AACjB,QAAM,MAAM,MAAc;AACxB,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,YAAM,YAAY,eAAe,SAAS,QAAQ,GAAG;AACrD,aAAO,EAAE,GAAG,UAAU,QAAQ,UAAU;AAAA,IAC1C;AAAA,EACF;AACF;AAOO,SAAS,uBACd,QACA,WAAgC,UACP;AACzB,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,UAAU;AACd,YAAM,SACJ,aAAa,WAAW,GAAG,MAAM,IAAI,SAAS,MAAM,KAAK,GAAG,SAAS,MAAM,IAAI,MAAM;AACvF,aAAO,EAAE,GAAG,UAAU,OAAO;AAAA,IAC/B;AAAA,EACF;AACF;AAIA,SAAS,YAAY,GAAmB;AACtC,SAAO,EAAE,QAAQ,uBAAuB,MAAM;AAChD;;;ACpKO,SAAS,4BACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAElD,MAAI,aAAa,WAAW,GAAG;AAC7B,WAAO,aAAa;AAAA,EACtB;AAEA,QAAM,UAAoB,CAAC;AAC3B,QAAM,kBAA4B,CAAC;AACnC,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,gEAAgE,EAAE,KAAK;AAAA,MACzE;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,YAAQ,KAAK,CAAC;AACd,oBAAgB,KAAK,IAAI,CAAC;AAC1B,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,IAAI,QAAQ;AAClB,QAAM,QAAQ,gBAAgB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3D,QAAM,WAAW,gBAAgB,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC9F,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,kCACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,UAAoB,CAAC;AAC3B,QAAM,UAAoB,CAAC;AAC3B,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,sEAAsE,EAAE,KAAK;AAAA,MAC/E;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,YAAQ,KAAK,CAAC;AACd,YAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,CAAC;AACjD,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,GAAG,MAAM,IAAI,IAAI,QAAQ,CAAC,GAAI,CAAC;AAChE,QAAM,QAAQ,SAAS,IAAI,IAAI,QAAQ;AACvC,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,QAAM,MAAM,QAAQ,IAAI,CAAC,GAAG,MAAM,KAAK,QAAQ,CAAC,IAAK,MAAM;AAC3D,QAAM,WAAW,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO,IAAI;AAC7E,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,QAAQ;AAAA,IACjC,qBAAqB;AAAA,IACrB,GAAG,aAAa;AAAA,IAChB,qBAAqB;AAAA,EACvB;AACF;AAoBO,SAAS,aACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,gBAA0B,CAAC;AACjC,MAAI,OAAO;AACX,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI,gBAAgB,iDAAiD,EAAE,KAAK,GAAG;AAAA,IACvF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,UAAM,IACJ,OAAO,EAAE,SAAS,YAAY,OAAO,SAAS,EAAE,IAAI,IAChD,MAAM,EAAE,MAAM,KAAK,KAAK,KAAK,IAAI,IACjC;AACN,QAAI,MAAM,MAAM;AACd,oBAAc,KAAK,IAAI,CAAC;AAAA,IAC1B,OAAO;AACL,oBAAc,KAAK,IAAI,KAAK,IAAI,EAAE;AAAA,IACpC;AACA,QAAI,IAAI,KAAM,QAAO;AACrB,YAAQ;AACR,aAAS,IAAI;AAAA,EACf;AACA,QAAM,IAAI,cAAc;AACxB,QAAM,QAAQ,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACzD,QAAM,WAAW,cAAc,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC5F,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAC9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,qBACd,cACA,OAAyB,CAAC,GACmD;AAC7E,SAAO;AAAA,IACL,KAAK,4BAA4B,cAAc,IAAI;AAAA,IACnD,OAAO,kCAAkC,cAAc,IAAI;AAAA,IAC3D,IAAI,aAAa,cAAc,IAAI;AAAA,EACrC;AACF;AAIA,SAAS,eAAkC;AACzC,SAAO,EAAE,OAAO,GAAG,eAAe,GAAG,qBAAqB,GAAG,GAAG,GAAG,qBAAqB,EAAE;AAC5F;AAEA,SAAS,MAAM,GAAW,IAAY,IAAoB;AACxD,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,CAAC;AACrC;;;AC9JA,IAAM,oBAA2C;AAEjD,IAAM,iBAAiB,CAAC,QAAkC;AACxD,QAAM,IAAI,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AAClD,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAcO,SAAS,mBACd,MACA,OAAkC,CAAC,GACP;AAC5B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,WAAW,KAAK,YAAY;AAClC,QAAMA,YAAW,KAAK,YAAY;AAElC,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ;AAC3D,QAAM,gBAA0D,CAAC;AACjE,aAAW,OAAO,UAAU;AAC1B,UAAM,IAAIA,UAAS,GAAG;AACtB,QAAI,MAAM,KAAM;AAChB,kBAAc,KAAK,EAAE,KAAK,OAAO,EAAE,CAAC;AAAA,EACtC;AAEA,QAAM,QAA4B,CAAC;AACnC,MAAI,mBAAmB;AACvB,MAAI,iBAAiB;AACrB,MAAI,iBAAiB;AAErB,MAAI,aAAa,+BAA+B;AAI9C,UAAM,SAAS,oBAAI,IAAsD;AACzE,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,YAAM,MAAM,GAAG,GAAG,KAAK,EAAE,IAAI,IAAI;AACjC,YAAM,MAAM,OAAO,IAAI,GAAG,KAAK,CAAC;AAChC,UAAI,KAAK,CAAC;AACV,aAAO,IAAI,KAAK,GAAG;AAAA,IACrB;AAEA,eAAW,CAAC,KAAK,OAAO,KAAK,OAAO,QAAQ,GAAG;AAC7C;AACA,UAAI,QAAQ,SAAS,GAAG;AACtB;AACA;AAAA,MACF;AACA,eAAS,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AACvC,iBAAS,IAAI,IAAI,GAAG,IAAI,QAAQ,QAAQ,KAAK;AAC3C,gBAAM,IAAI,QAAQ,CAAC;AACnB,gBAAM,IAAI,QAAQ,CAAC;AACnB,cAAI,EAAE,IAAI,gBAAgB,EAAE,IAAI,YAAa;AAC7C,gBAAM,SAAS,SAAS,GAAG,GAAG,IAAI,MAAM,IAAI,EAAE,CAAC,GAAI,SAAS;AAC5D,cAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,cAC9C;AAAA,QACP;AAAA,MACF;AAAA,IACF;AAAA,EACF,WAAW,aAAa,sBAAsB;AAE5C,UAAM,oBAAoB,oBAAI,IAG5B;AACF,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,UAAI,cAAc,kBAAkB,IAAI,GAAG;AAC3C,UAAI,CAAC,aAAa;AAChB,sBAAc,oBAAI,IAAI;AACtB,0BAAkB,IAAI,KAAK,WAAW;AAAA,MACxC;AACA,YAAM,MAAM,YAAY,IAAI,EAAE,IAAI,WAAW;AAC7C,UAAI,KAAK;AACP,YAAI,OAAO,EAAE;AACb,YAAI;AAAA,MACN,MAAO,aAAY,IAAI,EAAE,IAAI,aAAa,EAAE,KAAK,EAAE,KAAK,KAAK,EAAE,OAAO,GAAG,EAAE,CAAC;AAAA,IAC9E;AACA,eAAW,CAAC,KAAK,UAAU,KAAK,kBAAkB,QAAQ,GAAG;AAC3D;AACA,YAAM,MAAM,CAAC,GAAG,WAAW,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG,OAAO;AAAA,QACzD,KAAK,IAAI;AAAA,QACT,OAAO,IAAI,MAAM,IAAI;AAAA,QACrB,WAAW;AAAA,MACb,EAAE;AACF,UAAI,IAAI,SAAS,GAAG;AAClB;AACA;AAAA,MACF;AACA,eAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,iBAAS,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACvC,gBAAM,SAAS,SAAS,IAAI,CAAC,GAAI,IAAI,CAAC,GAAI,KAAK,SAAS;AACxD,cAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,cAC9C;AAAA,QACP;AAAA,MACF;AAAA,IACF;AAAA,EACF,OAAO;AAEL,UAAM,aAAa,oBAAI,IAAsD;AAC7E,eAAW,KAAK,eAAe;AAC7B,YAAM,MAAM,WAAW,EAAE,GAAG;AAC5B,YAAM,MAAM,WAAW,IAAI,GAAG,KAAK,CAAC;AACpC,UAAI,KAAK,CAAC;AACV,iBAAW,IAAI,KAAK,GAAG;AAAA,IACzB;AACA,eAAW,CAAC,KAAK,GAAG,KAAK,WAAW,QAAQ,GAAG;AAC7C;AACA,UAAI,IAAI,SAAS,GAAG;AAClB;AACA;AAAA,MACF;AACA,YAAM,SAAS,CAAC,GAAG,GAAG,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AACxD,YAAM,MAAM,OAAO,OAAO,SAAS,CAAC;AACpC,YAAM,MAAM,OAAO,CAAC;AACpB,UAAI,IAAI,IAAI,gBAAgB,IAAI,IAAI,aAAa;AAC/C;AACA;AAAA,MACF;AACA,YAAM,SAAS,SAAS,KAAK,KAAK,KAAK,SAAS;AAChD,UAAI,OAAO,SAAS,QAAS,OAAM,KAAK,OAAO,IAAI;AAAA,UAC9C;AAAA,IACP;AAAA,EACF;AAEA,SAAO,EAAE,OAAO,gBAAgB,kBAAkB,gBAAgB,SAAS;AAC7E;AAOO,SAAS,YACd,SACA,UAC6D;AAC7D,SAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,IACzB,QAAQ,SAAS,EAAE,KAAK,gBAAgB;AAAA,IACxC,QAAQ,EAAE,KAAK;AAAA;AAAA,IACf,UAAU,EAAE,KAAK;AAAA,EACnB,EAAE;AACJ;AAOO,SAAS,kBACd,SAC2F;AAC3F,SAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,IACzB,YAAY,EAAE;AAAA,IACd,aAAa,EAAE;AAAA,IACf,eAAe,EAAE;AAAA,IACjB,QAAQ,EAAE;AAAA,EACZ,EAAE;AACJ;AAIA,SAAS,SACP,GACA,GACA,YACA,WACgE;AAChE,QAAM,SAAS,KAAK,IAAI,EAAE,QAAQ,EAAE,KAAK;AACzC,MAAI,SAAS,UAAW,QAAO,EAAE,MAAM,SAAS;AAChD,QAAM,CAAC,QAAQ,QAAQ,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC;AAC7D,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM;AAAA,MACJ;AAAA,MACA,aAAa,OAAO,IAAI;AAAA,MACxB,eAAe,SAAS,IAAI;AAAA,MAC5B,iBAAiB,OAAO,IAAI;AAAA,MAC5B,mBAAmB,SAAS,IAAI;AAAA,MAChC,aAAa,OAAO,QAAQ,SAAS;AAAA,MACrC,QAAQ,EAAE,QAAQ,OAAO,OAAO,UAAU,SAAS,MAAM;AAAA,MACzD,MAAM,OAAO,IAAI,SAAS,SAAS,IAAI,OAAO,OAAO,IAAI,OAAO;AAAA,MAChE,MAAM;AAAA,QACJ,kBAAkB,OAAO,IAAI;AAAA,QAC7B,oBAAoB,SAAS,IAAI;AAAA,QACjC,kBAAkB,OAAO,IAAI;AAAA,QAC7B,oBAAoB,SAAS,IAAI;AAAA,QACjC,aAAa,OAAO,IAAI;AAAA,QACxB,eAAe,SAAS,IAAI;AAAA,MAC9B;AAAA,IACF;AAAA,EACF;AACF;AAQA,SAAS,WAAW,KAAwB;AAC1C,MAAI,OAAO,IAAI,eAAe,YAAY,IAAI,WAAW,SAAS,EAAG,QAAO,IAAI;AAChF,QAAM,UAAU,IAAI,QAAQ,IAAI;AAChC,MAAI,OAAO,YAAY,YAAY,OAAO,SAAS,OAAO,EAAG,QAAO,OAAO,OAAO;AAClF,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,SAAO,IAAI;AACb;;;AC/RO,SAAS,qBACd,UACA,KACa;AACb,QAAM,WAAW,IAAI,YAAY;AACjC,QAAM,cAAc,IAAI,eAAe,SAAS;AAChD,SAAO,SAAS,MAAM,IAAI,CAAC,SAAS;AAClC,UAAM,aAAa,OAAO,OAAO,KAAK,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AACzE,UAAM,QACJ,WAAW,SAAS,IAAI,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW,SAAS;AACtF,UAAM,MAA8B,EAAE,KAAK,KAAK,KAAK,aAAa,KAAK,WAAW;AAClF,eAAW,SAAS,OAAO,OAAO,KAAK,WAAW,GAAG;AACnD,iBAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,MAAM,UAAU,GAAG;AAC3D,YAAI,OAAO,SAAS,KAAK,EAAG,KAAI,OAAO,GAAG,EAAE,IAAI;AAAA,MAClD;AAAA,IACF;AACA,QAAI,OAAO,KAAK,eAAe,SAAU,KAAI,aAAa,KAAK;AAC/D,UAAM,UAAgC,EAAE,IAAI;AAC5C,QAAI,aAAa,UAAW,SAAQ,eAAe;AAAA,QAC9C,SAAQ,cAAc;AAC3B,WAAO;AAAA,MACL,OAAO,KAAK;AAAA,MACZ,cAAc,IAAI;AAAA,MAClB;AAAA,MACA,MAAM,KAAK;AAAA,MACX,OAAO,IAAI;AAAA,MACX,YAAY,IAAI;AAAA,MAChB,YAAY,IAAI;AAAA,MAChB,WAAW,IAAI;AAAA,MACf,QAAQ,KAAK;AAAA,MACb,SAAS,OAAO,SAAS,KAAK,OAAO,IAAI,KAAK,UAAW,IAAI,kBAAkB;AAAA,MAC/E,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC;AAAA,MACA,aAAa,KAAK,QAAQ,eAAe;AAAA,MACzC;AAAA,MACA,YAAY,KAAK;AAAA,IACnB;AAAA,EACF,CAAC;AACH;AAQO,SAAS,8BACd,QACA,KACA,OAA2B,CAAC,GACjB;AACX,QAAM,WAAW,IAAI,YAAY;AACjC,QAAM,QAAQ,KAAK,SAAS,OAAO,IAAI,WAAW,IAAI,IAAI,YAAY,IAAI,OAAO,SAAS;AAE1F,QAAM,MAA8B;AAAA,IAClC,YAAY,OAAO;AAAA,IACnB,YAAY,OAAO;AAAA,IACnB,aAAa,OAAO;AAAA,IACpB,eAAe,OAAO;AAAA,IACtB,aAAa,OAAO;AAAA,IACpB,eAAe,OAAO;AAAA,EACxB;AACA,aAAW,SAAS,OAAO,QAAQ;AACjC,QAAI,OAAO,MAAM,UAAU,SAAU,KAAI,SAAS,MAAM,KAAK,EAAE,IAAI,MAAM;AACzE,QAAI,SAAS,MAAM,KAAK,OAAO,IAAI,MAAM,WAAW,SAAS,IAAI;AACjE,QAAI,MAAM,aAAa;AACrB,iBAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,MAAM,WAAW,GAAG;AACtD,YAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,SAAS,MAAM,KAAK,IAAI,CAAC,EAAE,IAAI;AAAA,MACtF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,OAAO,OAAO,KAAK,CAAC,MAAM,EAAE,WAAW,UAAU,EAAE,WAAW,OAAO;AACvF,QAAM,UAAgC,EAAE,IAAI;AAC5C,MAAI,aAAa,UAAW,SAAQ,eAAe,OAAO;AAAA,MACrD,SAAQ,cAAc,OAAO;AAElC,SAAO;AAAA,IACL;AAAA,IACA,cAAc,IAAI;AAAA,IAClB,aAAa,IAAI;AAAA,IACjB,MAAM;AAAA,IACN,OAAO,IAAI;AAAA,IACX,YAAY,IAAI;AAAA,IAChB,YAAY,IAAI;AAAA,IAChB,WAAW,IAAI;AAAA,IACf,QAAQ,OAAO;AAAA,IACf,SAAS,IAAI,kBAAkB;AAAA,IAC/B,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,IAClC;AAAA,IACA,aAAa,YAAY,qBAAqB,SAAS,IAAI;AAAA,IAC3D;AAAA,IACA,YAAY,IAAI;AAAA,EAClB;AACF;AAEA,SAAS,qBAAqB,OAA4B;AACxD,MAAI,MAAM,WAAW,QAAS,QAAO,SAAS,MAAM,KAAK;AACzD,MAAI,MAAM,WAAW,OAAQ,QAAO,SAAS,MAAM,KAAK;AACxD,MAAI,MAAM,WAAW,UAAW,QAAO,SAAS,MAAM,KAAK;AAC3D,SAAO,SAAS,MAAM,KAAK,IAAI,MAAM,MAAM;AAC7C;;;ACrEO,SAAS,gBACd,UACA,OAA2E,CAAC,GAC3D;AACjB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,UAAU,KAAK,iBAAiB;AAKtC,QAAM,YAAY,KAAK,aAAa;AAEpC,QAAM,aAAa,oBAAI,IAAY;AACnC,aAAW,KAAK,UAAU;AACxB,eAAW,IAAI,EAAE,MAAM;AACvB,eAAW,IAAI,EAAE,KAAK;AAAA,EACxB;AACA,QAAM,MAAM,CAAC,GAAG,UAAU,EAAE,KAAK;AACjC,QAAM,MAAM,IAAI,IAAI,IAAI,IAAI,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC;AAC/C,QAAM,IAAI,IAAI;AACd,MAAI,MAAM,EAAG,QAAO,EAAE,SAAS,CAAC,GAAG,YAAY,GAAG,YAAY,GAAG,WAAW,KAAK;AACjF,MAAI,MAAM,GAAG;AACX,WAAO;AAAA,MACL,SAAS,CAAC,EAAE,aAAa,IAAI,CAAC,GAAI,UAAU,GAAG,aAAa,GAAG,GAAG,GAAG,MAAM,EAAE,CAAC;AAAA,MAC9E,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,IACb;AAAA,EACF;AAIA,QAAM,IAAgB,MAAM,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC,CAAC;AAClF,QAAM,IAAgB,MAAM,KAAK,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC,CAAC;AAClF,aAAW,KAAK,UAAU;AACxB,UAAM,IAAI,IAAI,IAAI,EAAE,MAAM;AAC1B,UAAM,IAAI,IAAI,IAAI,EAAE,KAAK;AACzB,UAAM,IAAI,EAAE,UAAU;AACtB,QAAI,EAAE,MAAM;AACV,QAAE,CAAC,EAAG,CAAC,KAAM,MAAM;AACnB,QAAE,CAAC,EAAG,CAAC,KAAM,MAAM;AAAA,IACrB,OAAO;AACL,QAAE,CAAC,EAAG,CAAC,KAAM;AAAA,IACf;AACA,MAAE,CAAC,EAAG,CAAC,KAAM;AACb,MAAE,CAAC,EAAG,CAAC,KAAM;AAAA,EACf;AAGA,QAAM,YAAY,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AAC7C,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,WAAU,CAAC,KAAM,EAAE,CAAC,EAAG,CAAC;AACpD,cAAU,CAAC,KAAM;AAAA,EACnB;AACA,QAAM,aAAa,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AAC9C,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,YAAW,CAAC,KAAM,EAAE,CAAC,EAAG,CAAC;AAAA,EACvD;AAGA,MAAI,QAAQ,IAAI,MAAc,CAAC,EAAE,KAAK,CAAC;AACvC,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,SAAO,OAAO,SAAS,QAAQ;AAC7B,UAAM,WAAW,IAAI,MAAc,CAAC;AACpC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAI,QAAQ;AACZ,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAI,MAAM,EAAG;AACb,YAAI,EAAE,CAAC,EAAG,CAAC,MAAO,EAAG;AACrB,iBAAS,EAAE,CAAC,EAAG,CAAC,KAAM,MAAM,CAAC,IAAK,MAAM,CAAC;AAAA,MAC3C;AACA,eAAS,CAAC,IAAI,UAAU,IAAI,MAAM,CAAC,IAAK,UAAU,CAAC,IAAK;AAAA,IAC1D;AAEA,QAAI,SAAS;AACb,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,WAAU,KAAK,IAAI,KAAK,IAAI,QAAQ,SAAS,CAAC,CAAE,CAAC;AAC7E,UAAM,OAAO,KAAK,IAAI,SAAS,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,UAAS,CAAC,IAAI,SAAS,CAAC,IAAK;AAEzD,YAAQ;AACR,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,IAAI,KAAK,IAAI,SAAS,CAAC,IAAK,MAAM,CAAC,CAAE,IAAI,KAAK,IAAI,OAAO,MAAM,CAAC,CAAE;AACxE,UAAI,IAAI,MAAO,SAAQ;AAAA,IACzB;AACA,YAAQ;AACR,QAAI,QAAQ,IAAK;AAAA,EACnB;AAEA,QAAM,SAAS,KAAK,IAAI,GAAG,MAAM,IAAI,CAAC,MAAM,KAAK,IAAI,KAAK,IAAI,QAAQ,CAAC,CAAC,CAAC,CAAC;AAC1E,QAAM,UAAgC,IAAI,IAAI,CAAC,IAAI,OAAO;AAAA,IACxD,aAAa;AAAA,IACb,UAAU,MAAM,CAAC;AAAA,IACjB,aAAa,KAAK,IAAI,KAAK,IAAI,QAAQ,MAAM,CAAC,CAAE,CAAC,IAAI;AAAA,IACrD,GAAG,WAAW,CAAC;AAAA,IACf,MAAM,UAAU,CAAC,IAAK;AAAA,EACxB,EAAE;AAEF,SAAO;AAAA,IACL,SAAS,QAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AAAA,IACvD,YAAY;AAAA,IACZ,YAAY;AAAA,IACZ,WAAW,QAAQ;AAAA,EACrB;AACF;AAiBO,SAAS,eACd,SACA,SACA,OAAmB,CAAC,GACyB;AAC7C,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,IAAI,KAAK,WAAW;AAE1B,QAAM,KAAK,QAAQ,IAAI,QAAQ,MAAM,KAAK;AAC1C,QAAM,KAAK,QAAQ,IAAI,QAAQ,KAAK,KAAK;AAEzC,QAAM,YAAY,KAAK,IAAI,QAAQ,KAAK,MAAM;AAC9C,QAAM,SAAS,QAAQ,OAAO,MAAM;AACpC,QAAM,SAAS,QAAQ,OAAO,MAAM;AACpC,QAAM,IAAI,QAAQ,UAAU;AAE5B,QAAM,cAAc,IAAI,KAAK,SAAS;AACtC,QAAM,aAAa,IAAI,KAAK,UAAU,IAAI;AAE1C,UAAQ,IAAI,QAAQ,QAAQ,KAAK,WAAW;AAC5C,UAAQ,IAAI,QAAQ,OAAO,KAAK,UAAU;AAE1C,SAAO,EAAE,aAAa,WAAW;AACnC;AAsBO,SAAS,0BACd,OACmB;AACnB,QAAM,aAAa,MAAM,cAAc;AACvC,QAAM,QAAQ,oBAAI,IAA2D;AAC7E,aAAW,KAAK,MAAM,MAAM;AAC1B,UAAM,MAAM,MAAM,IAAI,EAAE,QAAQ,KAAK,CAAC;AACtC,QAAI,KAAK,EAAE,aAAa,EAAE,aAAa,OAAO,EAAE,MAAM,CAAC;AACvD,UAAM,IAAI,EAAE,UAAU,GAAG;AAAA,EAC3B;AACA,QAAM,WAA8B,CAAC;AACrC,aAAW,OAAO,MAAM,OAAO,GAAG;AAChC,aAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,eAAS,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACvC,cAAM,IAAI,IAAI,CAAC;AACf,cAAM,IAAI,IAAI,CAAC;AACf,YAAI,EAAE,gBAAgB,EAAE,YAAa;AACrC,cAAM,SAAS,KAAK,IAAI,EAAE,QAAQ,EAAE,KAAK;AACzC,YAAI,UAAU,YAAY;AACxB,mBAAS,KAAK,EAAE,QAAQ,EAAE,aAAa,OAAO,EAAE,aAAa,MAAM,MAAM,QAAQ,EAAE,CAAC;AAAA,QACtF,OAAO;AACL,gBAAM,CAAC,QAAQ,KAAK,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC;AAC1D,mBAAS,KAAK,EAAE,QAAQ,OAAO,aAAa,OAAO,MAAM,aAAa,QAAQ,OAAO,CAAC;AAAA,QACxF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;;;ACjMO,SAAS,wBACd,cACA,gBACA,MACwB;AACxB,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,QAAQ,KAAK,gBAAgB;AACnC,QAAM,SAAS,KAAK;AAEpB,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,CAAC;AAC/B,QAAI,KAAK,EAAE,KAAK;AAChB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,YAAY,eAAe,IAAI,CAAC,MAAM;AAC1C,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,UAAU,QAAQ,IAAI,CAAC,KAAK,CAAC;AACnC,UAAM,IAAI,QAAQ;AAClB,UAAM,OAAO,MAAM,IAAI,MAAM,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClE,UAAM,WACJ,IAAI,IACA,gBACA,QAAQ,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI,KAAK;AAGnE,UAAM,SAAS,KAAK,KAAK,QAAQ,IAAI,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,CAAC,CAAC;AACjE,WAAO,EAAE,WAAW,EAAE,WAAW,YAAY,EAAE,YAAY,GAAG,MAAM,UAAU,OAAO;AAAA,EACvF,CAAC;AAGD,QAAM,aAAa,QAAQ,UAAU;AACrC,MAAI,cAAc,QAAQ;AACxB,UAAM,OAAO,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC,CAAC;AAC3E,WAAO,UAAU,IAAI,CAAC,OAAO;AAAA,MAC3B,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO;AAAA,MACP,QAAQ,qCAAqC,EAAE,CAAC;AAAA,IAClD,EAAE;AAAA,EACJ;AACA,QAAM,YAAY,SAAS;AAC3B,QAAM,cAAc,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC9D,SAAO,UAAU,IAAI,CAAC,MAAM;AAC1B,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,SAAS;AAC5F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,QAAQ,YAAY,EAAE,SAAS,QAAQ,CAAC,CAAC,OAAO,EAAE,CAAC,UAAU,EAAE,KAAK,QAAQ,CAAC,CAAC;AAAA,IAChF;AAAA,EACF,CAAC;AACH;AA4BO,SAAS,mBACd,cACA,gBACA,MACwB;AACxB,QAAM,YAAY,KAAK,qBAAqB;AAC5C,QAAM,SAAS,KAAK,cAAc;AAClC,QAAM,QAAQ,KAAK,aAAa;AAChC,QAAM,MAAM,QAAQ,KAAK,IAAI;AAE7B,QAAM,UAAU,oBAAI,IAAkD;AACtE,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,OAAO,EAAE,QAAQ,EAAE,SAAS;AAClC,QAAI,KAAM,KAAI,UAAU;AAAA,QACnB,KAAI,YAAY;AACrB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,QAAQ,eAAe,IAAI,CAAC,MAAM;AACtC,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,IAAI,SAAS,IAAI;AACvB,UAAM,IAAI,QAAQ,IAAI;AAEtB,UAAM,UAAU,WAAW,GAAG,GAAG,GAAG;AACpC,UAAM,WAAW,KAAK,IAAI,UAAU,SAAS;AAG7C,UAAM,WAAY,IAAI,MAAO,IAAI,MAAM,KAAK,IAAI,IAAI;AACpD,UAAM,QAAQ,KAAK,IAAI,MAAM,KAAK,KAAK,QAAQ,CAAC;AAChD,UAAM,SAAS,KAAK,IAAI,GAAG,WAAW,UAAU,EAAE;AAClD,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,GAAG,IAAI,SAAS,IAAI;AAAA,MACpB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAED,QAAM,cAAc,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC1D,SAAO,MAAM,IAAI,CAAC,MAAM;AACtB,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,KAAK,MAAM;AAC9F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,KAAK,IAAI,GAAG,YAAY;AAAA,MAC/B,QAAQ,QAAQ,EAAE,EAAE,QAAQ,CAAC,CAAC,IAAI,EAAE,EAAE,QAAQ,CAAC,CAAC,YAAY,EAAE,QAAQ,QAAQ,CAAC,CAAC,YAAY,SAAS;AAAA,IACvG;AAAA,EACF,CAAC;AACH;AAGO,SAAS,2BACd,MACA,OAAyD,CAAC,GACvC;AACnB,QAAM,YAAY,KAAK,iBAAiB;AACxC,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,MAAyB,CAAC;AAChC,aAAW,KAAK,MAAM;AACpB,QAAI,CAAC,EAAE,WAAY;AACnB,UAAM,QAAQ,aACT,EAAE,QAAQ,gBAAgB,EAAE,QAAQ,cACpC,EAAE,QAAQ,eAAe,EAAE,QAAQ;AACxC,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,QAAI,KAAK;AAAA,MACP,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd;AAAA,MACA,MAAM,SAAS;AAAA,IACjB,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAOA,SAAS,WAAW,OAAe,MAAc,KAA2B;AAC1E,QAAM,IAAI,KAAK,IAAI,GAAG,KAAK;AAC3B,QAAM,IAAI,KAAK,IAAI,GAAG,IAAI;AAC1B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,SAAO,KAAK,IAAI;AAClB;AAEA,SAAS,YAAY,OAAe,KAA2B;AAE7D,QAAM,IAAI,QAAQ,IAAI;AACtB,QAAM,IAAI,IAAI,KAAK,KAAK,IAAI,CAAC;AAC7B,SAAO,MAAM;AACX,QAAI;AACJ,QAAI;AACJ,OAAG;AACD,YAAM,KAAK,IAAI,KAAK;AACpB,YAAM,KAAK,IAAI,KAAK;AAEpB,UAAI,KAAK,KAAK,KAAK,KAAK,IAAI,EAAE,CAAC,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK,EAAE;AAC5D,UAAI,IAAI,IAAI;AAAA,IACd,SAAS,KAAK;AACd,QAAI,IAAI,IAAI;AACZ,UAAM,IAAI,IAAI;AACd,QAAI,IAAI,IAAI,SAAS,KAAK,EAAG,QAAO,IAAI;AACxC,QAAI,KAAK,IAAI,CAAC,IAAI,MAAM,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,IAAI,CAAC,GAAI,QAAO,IAAI;AAAA,EACxE;AACF;;;ACpNA,eAAsB,mBACpB,MAC0B;AAC1B,QAAM,KAAK,KAAK,MAAM,CAAC,GAAG,GAAG,GAAG,GAAG,GAAG,EAAE;AACxC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,WAAW,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAE7C,QAAM,SAA4B,CAAC;AACnC,aAAW,KAAK,UAAU;AACxB,UAAM,cAA8C,CAAC;AACrD,UAAM,YAAsB,CAAC;AAC7B,QAAI,cAAc;AAClB,QAAI,gBAAgB;AACpB,eAAW,YAAY,KAAK,WAAW;AACrC,YAAM,MAAM,SAAS,cAAc,YAAY,KAAK,UAAU,QAAQ,QAAQ,CAAC;AAC/E,YAAM,SAAmB,CAAC;AAC1B,UAAI,SAAS;AACb,eAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,cAAM,QAAQ,MAAM,KAAK,OAAO,IAAI,EAAE,UAAU,GAAG,KAAK,EAAE,CAAC;AAC3D,eAAO,KAAK,KAAK;AACjB,YAAI,SAAS,cAAe;AAC5B,kBAAU,KAAK,KAAK;AACpB,YAAI,SAAS,cAAe;AAC5B;AAAA,MACF;AACA,YAAM,QAAQ,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACzD,kBAAY,KAAK,EAAE,YAAY,KAAK,WAAW,OAAO,QAAQ,OAAO,OAAO,OAAO,CAAC;AAAA,IACtF;AACA,UAAM,YAAY,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,UAAU,MAAM;AACrF,UAAM,WACJ,UAAU,SAAS,IACf,IACA,UAAU,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,cAAc,GAAG,CAAC,KAAK,UAAU,SAAS;AACpF,WAAO,KAAK;AAAA,MACV;AAAA,MACA;AAAA,MACA,UAAU,cAAc,KAAK,IAAI,GAAG,aAAa;AAAA,MACjD,KAAK,KAAK,KAAK,QAAQ;AAAA,MACvB,GAAG,UAAU;AAAA,MACb;AAAA,IACF,CAAC;AAAA,EACH;AAEA,QAAMC,cAAa,OAAO,KAAK,CAAC,MAAM,EAAE,YAAY,aAAa,GAAG,KAAK;AACzE,QAAM,OAAO,SAAS,SAAS,SAAS,CAAC,KAAK;AAE9C,MAAI,OAAO;AACX,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,UAAM,KAAK,OAAO,IAAI,CAAC,EAAG;AAC1B,UAAM,KAAK,OAAO,CAAC,EAAG;AACtB,UAAM,KAAK,OAAO,IAAI,CAAC,EAAG;AAC1B,UAAM,KAAK,OAAO,CAAC,EAAG;AACtB,aAAU,KAAK,MAAM,KAAM,KAAK;AAAA,EAClC;AACA,QAAM,iBAAiB,SAAS,IAAI,IAAI,OAAO;AAE/C,SAAO,EAAE,QAAQ,YAAAA,aAAY,eAAe;AAC9C;AAwBO,SAAS,wBACd,GACA,GACA,OAA4E,CAAC,GACxD;AACrB,QAAM,OAAO,KAAK,cAAc;AAChC,QAAM,YAAY,KAAK,sBAAsB;AAC7C,QAAM,MAAMC,SAAQ,KAAK,IAAI;AAE7B,QAAM,OAAoC,CAAC;AAC3C,aAAW,MAAM,EAAE,QAAQ;AACzB,UAAM,KAAK,EAAE,OAAO,KAAK,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAC5C,QAAI,CAAC,GAAI;AACT,UAAM,SAAS,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,SAAS;AACpD,UAAM,SAAS,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,SAAS;AACpD,UAAM,MAAM,gBAAgB,QAAQ,WAAW,MAAM,GAAG;AACxD,UAAM,MAAM,gBAAgB,QAAQ,WAAW,MAAM,GAAG;AACxD,SAAK,KAAK;AAAA,MACR,GAAG,GAAG;AAAA,MACN,WAAW,GAAG,YAAY,GAAG;AAAA,MAC7B,MAAM,IAAI;AAAA,MACV,OAAO,IAAI;AAAA,MACX,MAAM,IAAI;AAAA,MACV,OAAO,IAAI;AAAA,IACb,CAAC;AAAA,EACH;AAEA,QAAM,YAAY,EAAE,iBAAiB,EAAE;AACvC,QAAM,kBACJ,EAAE,eAAe,QAAQ,EAAE,eAAe,OACtC,EAAE,aAAa,EAAE,aACjB;AAIN,QAAM,YAAY,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,WAAW,CAAC,IAAI,KAAK,IAAI,GAAG,KAAK,MAAM;AACrF,MAAI;AACJ,MAAI,KAAK,IAAI,SAAS,IAAI,QAAQ,KAAK,IAAI,SAAS,IAAI,KAAM,WAAU;AAAA,WAC/D,YAAY,KAAK,YAAY,EAAG,WAAU;AAAA,WAC1C,YAAY,KAAK,YAAY,EAAG,WAAU;AAAA,MAC9C,WAAU;AAEf,QAAM,YACJ,oBAAoB,UAAU,QAAQ,CAAC,CAAC,gBAAgB,UAAU,QAAQ,CAAC,CAAC,MAC3E,oBAAoB,OAAO,wBAAwB,eAAe,KAAK;AAE1E,SAAO,EAAE,MAAM,WAAW,iBAAiB,SAAS,UAAU;AAChE;AAGO,SAAS,WAAW,OAAwB,YAAY,KAAoB;AACjF,SAAO,MAAM,OAAO,KAAK,CAAC,MAAM,EAAE,YAAY,SAAS,GAAG,KAAK;AACjE;AAIA,SAASA,SAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAEA,SAAS,gBACP,IACA,WACA,YACA,KAC+B;AAC/B,MAAI,GAAG,SAAS,EAAG,QAAO,EAAE,KAAK,GAAG,CAAC,KAAK,GAAG,MAAM,GAAG,CAAC,KAAK,EAAE;AAC9D,QAAM,UAAU,IAAI,MAAc,SAAS;AAC3C,WAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,QAAO,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC3E,YAAQ,CAAC,IAAI,MAAM,GAAG;AAAA,EACxB;AACA,UAAQ,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC5B,QAAM,QAAQ,IAAI;AAClB,SAAO;AAAA,IACL,KAAK,QAAQ,KAAK,MAAO,QAAQ,IAAK,SAAS,CAAC;AAAA,IAChD,MAAM,QAAQ,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC,CAAC;AAAA,EACnF;AACF;;;AC7JA,eAAsB,0BACpB,MACqC;AACrC,QAAM,mBAAmB,KAAK,oBAAoB;AAClD,QAAM,SAAS,KAAK,UAAU;AAC9B,QAAM,WAAW,KAAK,qBAAqB;AAC3C,QAAM,SAAS,KAAK,UAAU,OAAO;AACrC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,MAAM,WAAW,IAAI;AAE3B,QAAM,YAAsC,CAAC;AAC7C,QAAM,OAAO,oBAAI,IAAY;AAC7B,MAAI,aAAa;AAGjB,aAAW,KAAK,KAAK,OAAO;AAC1B,UAAM,KAAK,KAAK,iBAAiB,CAAC;AAClC,QAAI,KAAK,IAAI,EAAE,EAAG;AAClB,SAAK,IAAI,EAAE;AACX,QAAI,cAAc,OAAQ;AAC1B,UAAM,QAAQ,MAAM,KAAK,QAAQ,CAAC;AAClC;AACA,cAAU,KAAK;AAAA,MACb;AAAA,MACA,YAAY;AAAA,MACZ,UAAU;AAAA,MACV,UAAU;AAAA,MACV;AAAA,MACA,kBAAkB;AAAA,IACpB,CAAC;AAAA,EACH;AAGA,WAAS,IAAI,GAAG,KAAK,QAAQ,KAAK;AAChC,QAAI,cAAc,OAAQ;AAC1B,UAAM,UAAU,UAAU,OAAO,CAAC,MAAM,EAAE,eAAe,IAAI,CAAC;AAC9D,eAAW,UAAU,SAAS;AAC5B,iBAAW,YAAY,KAAK,WAAW;AACrC,YAAI,cAAc,OAAQ;AAC1B,cAAM,WAAW,MAAM,SAAS,OAAO,OAAO,UAAU,GAAG;AAC3D,cAAM,WAAW,MAAM,QAAQ,QAAQ,IAAI,WAAW,CAAC,QAAQ;AAC/D,iBAAS,IAAI,GAAG,IAAI,KAAK,IAAI,UAAU,SAAS,MAAM,GAAG,KAAK;AAC5D,cAAI,cAAc,OAAQ;AAC1B,gBAAM,QAAQ,SAAS,CAAC;AACxB,gBAAM,MAAM,KAAK,iBAAiB,KAAK;AACvC,cAAI,KAAK,IAAI,GAAG,EAAG;AACnB,eAAK,IAAI,GAAG;AACZ,gBAAM,SAAS,MAAM,KAAK,QAAQ,KAAK;AACvC;AACA,oBAAU,KAAK;AAAA,YACb,IAAI;AAAA,YACJ,YAAY;AAAA,YACZ,UAAU,OAAO;AAAA,YACjB,UAAU;AAAA,YACV,OAAO;AAAA,YACP,kBAAkB,SAAS;AAAA,UAC7B,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,WAAW,UACd,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,EAAE,QAAQ,gBAAgB,EAC5D,KAAK,CAAC,GAAG,OAAO,EAAE,SAAS,MAAM,EAAE,SAAS,EAAE;AAEjD,QAAM,eAA2D,CAAC;AAClE,QAAM,SAAS,UAAU,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,UAAU,GAAG,CAAC;AACtE,WAAS,IAAI,GAAG,KAAK,QAAQ,KAAK;AAChC,UAAM,OAAO,UAAU,OAAO,CAAC,MAAM,EAAE,eAAe,CAAC;AACvD,QAAI,KAAK,WAAW,EAAG;AACvB,UAAM,QAAQ,KAAK,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,EAAE,QAAQ,gBAAgB,EAAE;AACjF,UAAM,YAAY,KAAK,OAAO,CAAC,KAAK,MAAM,OAAO,EAAE,SAAS,IAAI,CAAC,IAAI,KAAK;AAC1E,iBAAa,KAAK,EAAE,YAAY,GAAG,OAAO,KAAK,QAAQ,UAAU,OAAO,UAAU,CAAC;AAAA,EACrF;AAEA,SAAO,EAAE,WAAW,UAAU,cAAc,WAAW;AACzD;AAIA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;;;AC9JA,SAAS,gBAAgB,YAAY,WAAW,oBAAoB;AACpE,SAAS,eAAe;;;AC2CxB,eAAsB,UACpB,SACA,SACyB;AACzB,QAAM,MAAsB,CAAC;AAC7B,aAAW,KAAK,SAAS;AACvB,UAAM,CAAC,QAAQ,QAAQ,QAAQ,IAAI,MAAM,QAAQ,IAAI;AAAA,MACnD,QAAQ,QAAQ,QAAQ,SAAS,EAAE,WAAW,CAAC;AAAA,MAC/C,QAAQ,QAAQ,QAAQ,aAAa,EAAE,WAAW,CAAC;AAAA,MACnD,QAAQ,QAAQ,QAAQ,aAAa,EAAE,aAAa,CAAC;AAAA,IACvD,CAAC;AACD,QAAI,KAAK;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ,EAAE;AAAA,MACV,MAAM;AAAA,QACJ,YAAY,EAAE;AAAA,QACd,iBAAiB,EAAE;AAAA,QACnB,mBAAmB,EAAE;AAAA,QACrB,aAAa,EAAE;AAAA,QACf,eAAe,EAAE;AAAA,QACjB,aAAa,EAAE,KAAK;AAAA,QACpB,eAAe,EAAE,KAAK;AAAA,MACxB;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAGO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AA4BA,eAAsB,WACpB,MACA,SAC0B;AAC1B,QAAMC,YAAW,QAAQ,YAAY;AACrC,QAAM,UAAU,oBAAI,IAAyB;AAC7C,aAAW,KAAK,MAAM;AACpB,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,MAAM,QAAQ,IAAI,GAAG,KAAK,CAAC;AACjC,QAAI,KAAK,CAAC;AACV,YAAQ,IAAI,KAAK,GAAG;AAAA,EACtB;AAEA,QAAM,OAAwB,CAAC;AAC/B,aAAW,CAAC,YAAY,KAAK,KAAK,QAAQ,QAAQ,GAAG;AACnD,QAAI,MAAM,WAAW,EAAG;AAExB,UAAM,SAAS,MAAM,QAAQ,QAAQ,QAAQ,SAAS,MAAM,CAAC,EAAG,KAAK,CAAC;AACtE,UAAM,cAAwB,CAAC;AAC/B,UAAM,UAAoB,CAAC;AAC3B,UAAM,SAAmB,CAAC;AAC1B,eAAW,KAAK,OAAO;AACrB,YAAMC,UAASD,UAAS,CAAC;AACzB,UAAIC,YAAW,KAAM;AACrB,YAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,aAAa,EAAE,KAAK,CAAC;AACtE,kBAAY,KAAK,UAAU;AAC3B,cAAQ,KAAKA,OAAM;AACnB,aAAO,KAAK,EAAE,KAAK;AAAA,IACrB;AACA,QAAI,YAAY,WAAW,EAAG;AAC9B,SAAK,KAAK;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,MAAM;AAAA,QACJ;AAAA,QACA,GAAG,YAAY;AAAA,QACf,YAAY,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,QAAQ;AAAA,MAC3D;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,YAAY,MAA+B;AACzD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAwBA,eAAsB,UAAU,MAAmB,SAA8C;AAC/F,QAAM,UAAU,QAAQ,YAAY,MAAM;AAC1C,QAAM,OAAuB,CAAC;AAC9B,aAAW,KAAK,MAAM;AACpB,QAAI,CAAC,QAAQ,CAAC,EAAG;AACjB,UAAM,SAAS,QAAQ,WAAW,CAAC;AACnC,UAAM,CAAC,QAAQ,UAAU,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC7C,QAAQ,QAAQ,QAAQ,SAAS,EAAE,KAAK,CAAC;AAAA,MACzC,QAAQ,QAAQ,QAAQ,aAAa,EAAE,KAAK,CAAC;AAAA,IAC/C,CAAC;AACD,UAAM,WAAqC,CAAC;AAC5C,QAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,OAAO,CAAC;AAC7D,aAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAC/C,aAAS,KAAK,EAAE,MAAM,aAAa,SAAS,WAAW,CAAC;AACxD,SAAK,KAAK;AAAA,MACR;AAAA,MACA,MAAM;AAAA,QACJ,OAAO,EAAE;AAAA,QACT,aAAa,EAAE;AAAA,QACf,YAAY,EAAE;AAAA,QACd,OAAO,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAAA,QAC3C,OAAO,EAAE;AAAA,MACX;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AA8BA,eAAsB,UACpB,SACA,SACyB;AACzB,QAAM,OAAuB,CAAC;AAC9B,aAAW,KAAK,SAAS;AACvB,UAAM,SAAS,MAAM,QAAQ,QAAQ,QAAQ,SAAS,EAAE,WAAW,CAAC;AACpE,UAAM,gBAAgB,QAAQ,WAC1B,MAAM,QAAQ,QAAQ,QAAQ,SAAS,EAAE,aAAa,EAAE,eAAe,CAAC,IACxE,CAAC;AACL,UAAM,iBAA2B,CAAC;AAClC,eAAW,UAAU,eAAe;AAClC,qBAAe,KAAK,MAAM,QAAQ,QAAQ,QAAQ,WAAW,EAAE,aAAa,MAAM,CAAC,CAAC;AAAA,IACtF;AACA,UAAM,aAAa,MAAM,QAAQ,QAAQ,QAAQ,WAAW,EAAE,aAAa,EAAE,YAAY,CAAC;AAC1F,UAAM,eAAe,MAAM,QAAQ;AAAA,MACjC,QAAQ,WAAW,EAAE,eAAe,EAAE,cAAc;AAAA,IACtD;AACA,SAAK,KAAK;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,cAAc,EAAE;AAAA,MAChB,gBAAgB,EAAE;AAAA,MAClB,aAAa,EAAE;AAAA,MACf,MAAM;AAAA,QACJ,aAAa,EAAE;AAAA,QACf,eAAe,EAAE;AAAA,QACjB,iBAAiB,EAAE;AAAA,MACrB;AAAA,IACF,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,WAAW,MAA8B;AACvD,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAaO,SAAS,mBAAmB,aAAmC;AACpE,QAAM,OAA6B,YAAY,IAAI,CAAC,OAAO;AAAA,IACzD,OAAO,EAAE;AAAA,IACT,QAAQ,EAAE;AAAA,IACV,WAAW,EAAE;AAAA,IACb,QAAQ,EAAE;AAAA,IACV,aAAa,EAAE;AAAA,IACf,QAAQ,EAAE,UAAU;AAAA,EACtB,EAAE;AACF,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO;AACnF;AAIA,SAAS,cAAc,KAA+B;AACpD,QAAM,IAAI,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AAClD,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;;;AC/NA,SAAS,OAAO,GAA6B;AAC3C,QAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAC9C,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEA,SAAS,SAAS,IAAwB;AACxC,SAAO,CAAC,GAAG,IAAI,IAAI,EAAE,CAAC,EAAE,KAAK;AAC/B;AAEA,SAAS,mBAAmB,QAA+B;AACzD,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,GAAG,GAAG,MAAM,GAAG,QAAQ,GAAG,KAAK,GAAG,KAAK,GAAG,KAAK,EAAE;AACnF,QAAM,SAAS,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC/C,QAAM,IAAI,OAAO;AACjB,QAAM,OAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACjD,QAAM,MAAM,KAAK,MAAM,IAAI,CAAC;AAC5B,QAAM,SAAS,IAAI,MAAM,KAAK,OAAO,MAAM,CAAC,IAAK,OAAO,GAAG,KAAM,IAAI,OAAO,GAAG;AAC/E,QAAM,WAAW,OAAO,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,IAAI;AACnE,SAAO,EAAE,GAAG,MAAM,QAAQ,KAAK,OAAO,CAAC,GAAI,KAAK,OAAO,IAAI,CAAC,GAAI,KAAK,KAAK,KAAK,QAAQ,EAAE;AAC3F;AAEA,SAAS,aAAa,SAAsC;AAC1D,QAAM,SAAsC,EAAE,QAAQ,GAAG,KAAK,GAAG,SAAS,EAAE;AAC5E,MAAI,QAAQ;AACZ,MAAI,SAAS;AACb,MAAI,OAAO;AACX,QAAM,UAAoB,CAAC;AAC3B,aAAW,KAAK,SAAS;AACvB,WAAO,EAAE,QAAQ,KAAK,OAAO,EAAE,QAAQ,KAAK,KAAK;AACjD,aAAS,EAAE,WAAW;AACtB,cAAU,EAAE,WAAW;AACvB,YAAQ,EAAE;AACV,UAAM,KAAK,OAAO,CAAC;AACnB,QAAI,OAAO,KAAM,SAAQ,KAAK,EAAE;AAAA,EAClC;AACA,SAAO;AAAA,IACL,SAAS,QAAQ;AAAA,IACjB;AAAA,IACA,QAAQ,mBAAmB,OAAO;AAAA,IAClC,QAAQ,SAAS,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAAA,IAC5C,cAAc,SAAS,QAAQ,IAAI,CAAC,MAAM,EAAE,UAAU,CAAC;AAAA,IACvD,YAAY,SAAS,QAAQ,IAAI,CAAC,MAAM,EAAE,SAAS,CAAC;AAAA,IACpD,aAAa,EAAE,OAAO,OAAO,QAAQ,OAAO;AAAA,IAC5C,cAAc;AAAA,EAChB;AACF;AASA,eAAsB,eACpB,SACA,SACA,QACA,aAC0B;AAC1B,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI,MAAM,wEAAmE;AAAA,EACrF;AACA,QAAM,UAAU,OAAO,WAAW,CAAC,QAAQ,KAAK;AAChD,QAAM,QAAgC,CAAC;AACvC,QAAM,YAAoD,CAAC;AAE3D,MAAI,QAAQ,SAAS,MAAM,GAAG;AAC5B,UAAM,OAAO,MAAM,WAAW,SAAS,OAAO;AAC9C,UAAM,kBAAkB,IAAI,YAAY,IAAI;AAC5C,cAAU,OAAO,KAAK;AAAA,EACxB;AACA,MAAI,QAAQ,SAAS,KAAK,GAAG;AAC3B,UAAM,OAAO,MAAM,UAAU,SAAS,OAAO;AAC7C,UAAM,iBAAiB,IAAI,WAAW,IAAI;AAC1C,cAAU,MAAM,KAAK;AAAA,EACvB;AACA,MAAI,QAAQ,SAAS,KAAK,GAAG;AAC3B,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,yEAAyE;AAAA,IAC3F;AACA,UAAM,OAAO,MAAM,UAAU,YAAY,SAAS,YAAY,OAAO;AACrE,UAAM,iBAAiB,IAAI,WAAW,IAAI;AAC1C,cAAU,MAAM,KAAK;AAAA,EACvB;AAEA,QAAM,WAA8B;AAAA,IAClC,GAAG;AAAA,IACH;AAAA,IACA;AAAA,IACA,OAAO,aAAa,OAAO;AAAA,EAC7B;AACA,QAAM,eAAe,IAAI,GAAG,KAAK,UAAU,UAAU,MAAM,CAAC,CAAC;AAAA;AAC7D,QAAM,cAAc,IAAI,oBAAoB,QAAQ;AACpD,SAAO,EAAE,UAAU,MAAM;AAC3B;AAEA,SAAS,IAAI,GAAmB;AAC9B,SAAO,IAAI,IAAI,KAAK,QAAQ,CAAC,CAAC;AAChC;AAGO,SAAS,oBAAoB,GAA8B;AAChE,QAAM,IAAI,EAAE;AACZ,QAAM,QAAQ,EAAE,WAAW;AAC3B,QAAM,aAAc,CAAC,UAAU,OAAO,SAAS,EAC5C,IAAI,CAAC,MAAM,SAAS,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,KAAK,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,CAAC,GAAG,EACvE,KAAK,IAAI;AACZ,QAAM,gBAAgB,EAAE,OAAO,SAAS;AACxC,SAAO;AAAA,IACL,cAAc,EAAE,IAAI,OAAO,EAAE,OAAO;AAAA,IACpC;AAAA,IACA,eAAe,EAAE,MAAM,sBAAmB,EAAE,YAAY,sBAAmB,EAAE,OAAO;AAAA,IACpF;AAAA,IACA;AAAA,IACA,eAAe,EAAE,OAAO,IAAI,GAAG,gBAAgB,+CAAqC,EAAE;AAAA,IACtF,iBAAiB,EAAE,OAAO,MAAM;AAAA,IAChC,sBAAsB,EAAE,OAAO,WAAW;AAAA,IAC1C;AAAA,IACA;AAAA,IACA,iCAAiC,EAAE,OAAO;AAAA,IAC1C,kBAAkB,EAAE,QAAQ,IAAI,CAAC,MAAM,GAAG,CAAC,KAAK,EAAE,UAAU,CAAC,KAAK,CAAC,QAAQ,EAAE,KAAK,IAAI,CAAC;AAAA,IACvF;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,OAAO,EAAE,OAAO,CAAC,cAAW,EAAE,OAAO,KAAK,QAAQ,CAAC,CAAC,gBAAa,EAAE,OAAO,OAAO,QAAQ,CAAC,CAAC,aAAU,EAAE,OAAO,IAAI,QAAQ,CAAC,CAAC,aAAU,EAAE,OAAO,IAAI,QAAQ,CAAC,CAAC,aAAU,EAAE,OAAO,IAAI,QAAQ,CAAC,CAAC;AAAA,IAC9L;AAAA,IACA;AAAA,IACA,iBAAiB,EAAE,OAAO,KAAK,IAAI,CAAC;AAAA,IACpC,yCAAyC,EAAE,aAAa,MAAM;AAAA,IAC9D,kBAAkB,EAAE,WAAW,KAAK,IAAI,CAAC;AAAA,IACzC,iBAAiB,EAAE,YAAY,KAAK,SAAS,EAAE,YAAY,MAAM,wBAAqB,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IAC/G;AAAA,IACA;AAAA,IACA,0BAA0B,EAAE,cAAc,sBAAsB,SAAS;AAAA,IACzE,YAAY,EAAE,cAAc,QAAQ,QAAQ,IAAI,mCAAgC,EAAE,cAAc,yBAAyB,QAAQ,IAAI;AAAA,IACrI;AAAA,IACA;AAAA,IACA,EAAE;AAAA,IACF;AAAA,IACA;AAAA,IACA,EAAE;AAAA,IACF;AAAA,IACA;AAAA,IACA,EAAE;AAAA,IACF;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;;;AFjNO,SAAS,eAAe,SAAyB,YAAwC;AAC9F,YAAU,QAAQ,UAAU,GAAG,EAAE,WAAW,KAAK,CAAC;AAClD,QAAM,WAAW,WAAW,UAAU,IAAI,WAAW,UAAU,IAAI,CAAC;AACpE,QAAM,OAAO,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACjD,QAAM,QAAkB,CAAC;AACzB,MAAI,WAAW;AACf,MAAI,UAAU;AACd,aAAW,KAAK,SAAS;AACvB,QAAI,KAAK,IAAI,EAAE,KAAK,GAAG;AACrB;AACA;AAAA,IACF;AACA,SAAK,IAAI,EAAE,KAAK;AAChB,UAAM,KAAK,KAAK,UAAU,CAAC,CAAC;AAC5B;AAAA,EACF;AACA,MAAI,MAAM,SAAS,EAAG,gBAAe,YAAY,GAAG,MAAM,KAAK,IAAI,CAAC;AAAA,CAAI;AACxE,SAAO,EAAE,UAAU,SAAS,OAAO,SAAS,SAAS,SAAS;AAChE;AAGO,SAAS,WAAW,YAAoC;AAC7D,MAAI,CAAC,WAAW,UAAU,EAAG,QAAO,CAAC;AACrC,QAAM,MAAsB,CAAC;AAC7B,aAAW,QAAQ,aAAa,YAAY,MAAM,EAAE,MAAM,IAAI,GAAG;AAC/D,QAAI,KAAK,KAAK,EAAG,KAAI,KAAK,KAAK,MAAM,IAAI,CAAiB;AAAA,EAC5D;AACA,SAAO;AACT;AAEA,SAAS,SAAS,GAAyB;AACzC,QAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAC9C,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAgBA,eAAsB,uBACpB,YACA,QACA,OAAuB,CAAC,GACE;AAC1B,MAAI,UAAU,WAAW,UAAU,EAAE;AAAA,IACnC,CAAC,MAAM,OAAO,EAAE,WAAW,YAAY,OAAO,EAAE,eAAe;AAAA,EACjE;AACA,MAAI,KAAK,OAAQ,WAAU,QAAQ,OAAO,CAAC,MAAM,KAAK,OAAQ,SAAS,EAAE,QAAQ,CAAC;AAClF,MAAI,KAAK,YAAY,KAAM,WAAU,QAAQ,OAAO,CAAC,MAAM,SAAS,CAAC,KAAK,KAAK,QAAS;AAExF,QAAM,OAAO,IAAI;AAAA,IACf,QAAQ,IAAI,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,QAAS,YAAY,EAAE,WAAY,CAAC,CAAC;AAAA,EAChF;AACA,QAAM,UAAU;AAAA,IACd,UAAU,CAAC,OAAe,KAAK,IAAI,EAAE,GAAG,UAAU;AAAA,IAClD,cAAc,CAAC,OAAe,KAAK,IAAI,EAAE,GAAG,cAAc;AAAA,EAC5D;AACA,SAAO,eAAe,SAAS,SAAS,MAAM;AAChD;;;AGjDO,IAAM,+BAAN,MAAyD;AAAA,EACtD;AAAA,EACA,aAAoD;AAAA,EAE5D,YAAY,MAA2C;AACrD,SAAK,OAAO;AAAA,EACd;AAAA,EAEA,MAAM,gBAAgB,MAA2C;AAC/D,UAAM,YAAY,KAAK,KAAK,oBAAoB;AAChD,UAAM,WAA0B,CAAC;AACjC,UAAM,cAAc,KAAK,OAAO,CAAC,MAAM;AACrC,YAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,aAAO,OAAO,UAAU,YAAY,QAAQ;AAAA,IAC9C,CAAC;AACD,QAAI,YAAY,WAAW,EAAG,QAAO;AAIrC,UAAM,UAAU,oBAAI,IAAyB;AAC7C,eAAW,KAAK,aAAa;AAC3B,YAAM,MAAM,QAAQ,IAAI,EAAE,WAAW,KAAK,CAAC;AAC3C,UAAI,KAAK,CAAC;AACV,cAAQ,IAAI,EAAE,aAAa,GAAG;AAAA,IAChC;AAEA,eAAW,CAAC,aAAa,KAAK,KAAK,QAAQ,QAAQ,GAAG;AACpD,YAAM,YACJ,MAAM,OAAO,CAAC,GAAG,MAAM;AACrB,cAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ,eAAe;AAC7D,eAAO,IAAI;AAAA,MACb,GAAG,CAAC,IAAI,MAAM;AAChB,eAAS,KAAK;AAAA,QACZ,MAAM,aAAa,WAAW;AAAA,QAC9B,aAAa,GAAG,WAAW,aAAa,SAAS,OAAO,MAAM,MAAM,iBAAiB,UAAU,QAAQ,CAAC,CAAC;AAAA,QACzG,UAAU;AAAA,UACR,QAAQ,MAAM,MAAM,GAAG,CAAC,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,UAC5C,SAAS,MAAM;AAAA,QACjB;AAAA,MACF,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,cAAc,UAAoD;AACtE,QAAI,SAAS,WAAW,EAAG,QAAO,CAAC;AAInC,QAAI,KAAK,eAAe,MAAM;AAC5B,aAAO;AAAA,QACL;AAAA,UACE,MAAM;AAAA,UACN,SAAS,EAAE,WAAW,mCAAmC;AAAA,UACzD,WACE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAEA,UAAM,sBAAsB,KAAK,KAAK,uBAAuB;AAC7D,UAAM,UAA4B,CAAC;AAEnC,eAAW,WAAW,KAAK,WAAW,QAAQ;AAC5C,UAAI,QAAQ,YAAY,eAAgB;AACxC,UAAI,KAAK,IAAI,QAAQ,QAAQ,KAAK,oBAAqB;AACvD,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,SAAS;AAAA,UACP,QAAQ,QAAQ;AAAA,UAChB,QAAQ;AAAA,UACR,UAAU,QAAQ;AAAA,UAClB,aAAa,QAAQ;AAAA,QACvB;AAAA,QACA,WAAW,gCAAgC,QAAQ,SAAS,QAAQ,CAAC,CAAC,OAAO,QAAQ,WAAW;AAAA,QAChG,eAAe,CAAC,KAAK,IAAI,GAAG,OAAO,KAAK,IAAI,QAAQ,QAAQ,CAAC;AAAA,MAC/D,CAAC;AAAA,IACH;AACA,eAAW,WAAW,KAAK,WAAW,OAAO,MAAM,GAAG,CAAC,GAAG;AACxD,UAAI,QAAQ,YAAY,eAAgB;AACxC,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,SAAS;AAAA,UACP,QAAQ,QAAQ;AAAA,UAChB,QAAQ;AAAA,UACR,UAAU,QAAQ;AAAA,UAClB,aAAa,QAAQ;AAAA,QACvB;AAAA,QACA,WAAW,gCAAgC,QAAQ,SAAS,QAAQ,CAAC,CAAC,OAAO,QAAQ,WAAW;AAAA,QAChG,eAAe,KAAK,IAAI,GAAG,KAAK,IAAI,QAAQ,QAAQ,IAAI,GAAG,IAAI;AAAA,MACjE,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,YAAY,SAA2B,UAAmD;AAG9F,WAAO;AAAA,MACL,GAAG;AAAA,MACH,SAAS,CAAC,GAAG,SAAS,SAAS,GAAG,OAAO;AAAA,IAC3C;AAAA,EACF;AAAA,EAEA,MAAM,eAAe,MAAiD;AAKpE,UAAM,YAA0B;AAAA,MAC9B,SAAS;AAAA,MACT,aAAa,KAAK;AAAA,MAClB,YAAY,KAAK;AAAA,MACjB,UAAU;AAAA,QACR,gBAAgB;AAAA,QAChB,mBAAmB;AAAA,QACnB,UAAU,EAAE,KAAK,GAAG,MAAM,EAAE;AAAA,QAC5B,cAAc;AAAA,QACd,aAAa;AAAA,QACb,cAAc;AAAA,QACd,YAAY;AAAA,QACZ,oBAAoB;AAAA,QACpB,qBAAqB,OAAO;AAAA,QAC5B,oBAAoB,OAAO;AAAA,MAC7B;AAAA,MACA,QACE;AAAA,MACF,eAAe;AAAA,IACjB;AACA,WAAO;AAAA,MACL;AAAA,MACA,MAAM,CAAC;AAAA,MACP,cAAc;AAAA,IAChB;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBAAiB,MAA4D;AACjF,UAAM,SAAS,MAAM,yBAAyB;AAAA,MAC5C;AAAA,MACA,UAAU,KAAK,KAAK;AAAA,MACpB,gBAAgB,KAAK,KAAK;AAAA,MAC1B,SAAS,KAAK,KAAK;AAAA,IACrB,CAAC;AACD,QAAI,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,SAAS,MAAM;AACvD,SAAK,aAAa;AAClB,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,UAAU,QAA8C;AACtD,SAAK,aAAa;AAAA,EACpB;AAAA,EAEA,gBAAuD;AACrD,WAAO,KAAK;AAAA,EACd;AACF;;;ACjJA,eAAsB,mBACpB,OACA,OACA,MACuB;AACvB,QAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;AACzC,QAAM,UAAU,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACnE,QAAM,MAAoB,CAAC;AAC3B,MAAI,MAAM;AACV,aAAW,QAAQ,SAAS;AAC1B,QAAI,KAAK,aAAa,CAAC,KAAK,UAAU,IAAI,EAAG;AAC7C,QAAI,SAAmD;AACvD,eAAW,KAAK,KAAK,SAAS;AAC5B,UAAI,CAAC,EAAE,UAAU,SAAS,KAAK,IAAI,EAAG;AACtC,YAAM,IAAI,MAAM,EAAE,MAAM,IAAI;AAC5B,UAAI,GAAG;AACL,iBAAS;AACT;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,OAAQ;AACb,QAAI,KAAK;AAAA,MACP,QAAQ,KAAK;AAAA,MACb;AAAA,MACA,WAAW;AAAA,MACX,MAAM,KAAK;AAAA,MACX,MAAM,KAAK;AAAA,MACX,QAAQ,OAAO;AAAA,MACf,aAAa,OAAO;AAAA,MACpB,WAAW,OAAO;AAAA,MAClB,QAAQ,OAAO;AAAA,IACjB,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAeO,SAAS,yBAAyB,aAA+C;AACtF,MAAI,YAAY,WAAW,GAAG;AAC5B,WAAO;AAAA,MACL,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,mBAAmB;AAAA,MACnB,iBAAiB;AAAA,MACjB,gBAAgB;AAAA,MAChB,gBAAgB;AAAA,IAClB;AAAA,EACF;AACA,QAAM,QAAQ,YAAY,CAAC,EAAG;AAC9B,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,aAAa;AACjB,MAAI,WAA0B;AAC9B,MAAI,OAAO,YAAY,CAAC,EAAG;AAC3B,WAAS,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;AAC3C,UAAM,IAAI,YAAY,CAAC;AACvB,UAAM,IAAI,EAAE,UAAU;AACtB,YAAQ;AACR,aAAS,IAAI,EAAE;AACf,QAAI,EAAE,SAAS,IAAK;AACpB,QAAI,IAAI,GAAG;AACT,YAAM,QAAQ,EAAE,SAAS;AACzB,UAAI,QAAQ,YAAY;AACtB,qBAAa;AACb,mBAAW;AAAA,MACb;AACA,aAAO,EAAE;AAAA,IACX,OAAO;AACL,aAAO,EAAE;AAAA,IACX;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,YAAY;AAAA,IACxB,YAAY,SAAS,IAAI,IAAI,QAAQ;AAAA,IACrC,mBAAmB;AAAA,IACnB,iBAAiB,WAAW,YAAY;AAAA,IACxC,gBAAgB;AAAA,IAChB,gBAAgB;AAAA,EAClB;AACF;AAgCO,SAAS,iBACd,kBACA,OAAyD,CAAC,GACrC;AACrB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,YAAY,KAAK,mBAAmB;AAC1C,QAAM,OAAO,CAAC,GAAG,iBAAiB,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,OAAO,KAAK,OAAO,EAAE,OAAO,MAAM,EAAE;AACvF,QAAM,UAA+B,CAAC;AAEtC,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,aAAS,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACxC,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,IAAI,KAAK,CAAC;AAChB,YAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;AACtD,UAAI,SAAS,YAAY,EAAG;AAO5B,UAAI,gBAAgB;AACpB,eAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,cAAM,KAAK,EAAE,MAAM,CAAC;AACpB,cAAM,KAAK,EAAE,MAAM,CAAC;AACpB,cAAM,uBAAuB,GAAG,SAAS,GAAG,QAAQ,GAAG,SAAS,GAAG;AACnE,cAAM,YAAY,KAAK,IAAI,GAAG,SAAS,GAAG,MAAM;AAChD,YAAI,wBAAwB,aAAa,WAAW;AAClD,0BAAgB;AAChB;AAAA,QACF;AAAA,MACF;AACA,UAAI,gBAAgB,EAAG;AACvB,UAAI,gBAAgB,UAAW;AAE/B,YAAM,QAAQ,EAAE,MAAM,aAAa;AACnC,YAAM,QAAQ,EAAE,MAAM,aAAa;AACnC,YAAM,SAAS,KAAK,IAAI,MAAM,SAAS,MAAM,MAAM;AACnD,UAAI,SAAS,UAAW;AAExB,YAAM,SAAS,MAAM,SAAS,MAAM,SAAS,QAAQ;AACrD,YAAM,WAAW,MAAM,SAAS,MAAM,SAAS,QAAQ;AACvD,YAAM,YAAY,MAAM,SAAS,MAAM,SAAS,EAAE,QAAQ,EAAE;AAC5D,YAAM,cAAc,MAAM,SAAS,MAAM,SAAS,EAAE,QAAQ,EAAE;AAC9D,cAAQ,KAAK;AAAA,QACX,aAAa;AAAA,QACb,iBAAiB,gBAAgB;AAAA,QACjC,cAAc,OAAO;AAAA,QACrB,cAAc,OAAO;AAAA,QACrB,gBAAgB,SAAS;AAAA,QACzB,gBAAgB,SAAS;AAAA,QACzB,eAAe;AAAA,QACf,aAAa,OAAO,SAAS,SAAS;AAAA,MACxC,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;;;AChKA,eAAsB,cACpB,MAC8B;AAE9B,QAAM,WAAW,MAAM,gBAAgB,IAAI;AAG3C,QAAM,gBAAgB;AAAA,IACpB,SAAS;AAAA,IACT,KAAK,oBAAoB,CAAC;AAAA,EAC5B;AAGA,QAAM,cAAc,mBAAmB,SAAS,MAAM;AAAA,IACpD,UAAU,KAAK,aAAa,YAAY;AAAA,IACxC,WAAW,KAAK,aAAa,aAAa;AAAA,IAC1C,UAAU,KAAK,aAAa,YAAY,KAAK,YAAY;AAAA,IACzD,UAAU,KAAK,aAAa;AAAA,EAC9B,CAAC;AAGD,MAAI,oBAAqD;AACzD,MAAI,KAAK,QAAQ,YAAY;AAC3B,UAAM,aAAa,KAAK,OAAO;AAC/B,UAAM,cAAc,yBAAyB,SAAS,MAAM,UAAU;AACtE,QAAI,YAAY,KAAK,CAAC,MAAM,EAAE,OAAO,SAAS,CAAC,GAAG;AAChD,0BAAoB,iCAAiC;AAAA,QACnD;AAAA,QACA,OAAO,KAAK,YAAY;AAAA,QACxB,OAAO,KAAK,YAAY;AAAA,QACxB,MAAM,KAAK,YAAY,QAAQ,KAAK,QAAQ;AAAA,MAC9C,CAAC;AAAA,IACH;AAAA,EACF;AAGA,QAAM,gBAAgB,oBAAoB;AAAA,IACxC,MAAM,SAAS;AAAA,IACf,yBAAyB,KAAK;AAAA,EAChC,CAAC;AAGD,MAAI,qBAA4D;AAChE,MAAI,KAAK,gBAAgB,KAAK,kBAAkB,KAAK,eAAe,SAAS,GAAG;AAC9E,yBAAqB,MAAM,yBAAyB;AAAA,MAClD,MAAM,SAAS;AAAA,MACf,UAAU,KAAK;AAAA,MACf,gBAAgB,KAAK;AAAA,IACvB,CAAC;AAAA,EACH;AAGA,QAAM,cAAkD,CAAC;AACzD,MAAI,KAAK,eAAe,KAAK;AAC3B,gBAAY,MAAM,MAAM,UAAU,YAAY,OAAO,KAAK,cAAc,GAAG;AAAA,EAC7E;AACA,MAAI,KAAK,eAAe,MAAM;AAC5B,gBAAY,OAAO,MAAM,WAAW,SAAS,MAAM,KAAK,cAAc,IAAI;AAAA,EAC5E;AACA,MAAI,KAAK,eAAe,KAAK;AAC3B,gBAAY,MAAM,MAAM,UAAU,SAAS,MAAM,KAAK,cAAc,GAAG;AAAA,EACzE;AAEA,QAAM,UAAU,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM;AAAA,EACR;AACF;AAIA,SAAS,yBACP,MACA,YACkD;AAGlD,QAAM,WAAW,oBAAI,IAAoB;AACzC,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,gBAAgB,WAAY;AAClC,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,aAAS,IAAI,GAAG,GAAG,KAAK,EAAE,IAAI,IAAI,KAAK;AAAA,EACzC;AACA,QAAM,cAAc,oBAAI,IAAsB;AAC9C,aAAW,KAAK,MAAM;AACpB,QAAI,EAAE,gBAAgB,WAAY;AAClC,UAAM,MAAM,EAAE,cAAc,EAAE;AAC9B,UAAM,QAAQ,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAClD,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,UAAM,YAAY,SAAS,IAAI,GAAG,GAAG,KAAK,EAAE,IAAI,EAAE;AAClD,QAAI,OAAO,cAAc,SAAU;AACnC,UAAM,MAAM,YAAY,IAAI,EAAE,WAAW,KAAK,CAAC;AAC/C,QAAI,KAAK,QAAQ,SAAS;AAC1B,gBAAY,IAAI,EAAE,aAAa,GAAG;AAAA,EACpC;AACA,SAAO,CAAC,GAAG,YAAY,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,aAAa,MAAM,OAAO,EAAE,aAAa,OAAO,EAAE;AAC5F;AAEA,SAAS,aAAa,MAMX;AACT,QAAM,IAAI,KAAK;AACf,QAAM,QAAQ;AAAA,IACZ,GAAG,EAAE,UAAU,KAAK,EAAE,KAAK,MAAM,sBAAsB,EAAE,WAAW,MAAM,wBAAwB,EAAE,oBAAoB,MAAM,GAAG,EAAE,CAAC;AAAA,IACpI,gBAAgB,KAAK,YAAY,MAAM,MAAM,KAAK,KAAK,YAAY,QAAQ,KAAK,KAAK,YAAY,gBAAgB;AAAA,EACnH;AACA,MAAI,KAAK,mBAAmB;AAC1B,UAAM;AAAA,MACJ,uBAAuB,KAAK,kBAAkB,eAAe,QAAQ,MAClE,KAAK,kBAAkB,eAAe,cACnC,IAAI,KAAK,kBAAkB,eAAe,WAAW,KACrD;AAAA,IACR;AAAA,EACF;AACA,QAAM;AAAA,IACJ,mBAAmB,KAAK,cAAc,OAAO,KAAK,KAAK,cAAc,SAAS,MAAM;AAAA,EACtF;AACA,MAAI,KAAK,oBAAoB;AAC3B,UAAM,MAAM,KAAK,mBAAmB,OAAO,CAAC;AAC5C,UAAM;AAAA,MACJ,eAAe,KAAK,UAAU,MAAM,YAAO,KAAK,YAAY,GAAG,QAAQ,CAAC,CAAC,KAAK,KAAK,WAAW,SAAS;AAAA,IACzG;AAAA,EACF;AACA,SAAO,MAAM,KAAK,KAAK;AACzB;","names":["rewardOf","firstPassK","makeRng","rewardOf","reward"]}