@tangle-network/agent-eval 0.174.0 → 0.176.0

package/dist/power-preflight-CFXm0Vjo.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"power-preflight-CFXm0Vjo.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/gates/statistical-heldout.ts","../src/campaign/gates/power-preflight.ts"],"sourcesContent":["import {\n  type PairedBootstrapOptions,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n  pairedSignTest,\n} from './statistics'\n\nexport interface PairedDeltaTestOptions extends PairedBootstrapOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. */\n  threshold?: number\n  /** Caller-required paired observations. The exact test may impose a higher minimum. */\n  minPairs?: number\n}\n\nexport interface PairedDeltaTestResult {\n  bootstrap: PairedBootstrapResult\n  method: 'bootstrap-ci' | 'exact-sign'\n  /** Exact one-sided p-value below the bootstrap minimum; otherwise null. */\n  pValue: number | null\n  /** Effective observation minimum after accounting for confidence. */\n  minimumPairs: number\n  sufficient: boolean\n  /**\n   * The bootstrap interval has zero width (or is non-finite), so it carries no\n   * information about how far the estimate could be wrong and cannot support a\n   * decision in either direction. See {@link pairedDeltaTest}.\n   */\n  indeterminate: boolean\n  significant: boolean\n}\n\n/** Smallest all-positive sample that can clear a one-sided exact sign test. */\nexport function minimumPairsForPairedDeltaTest(confidence = 0.95): number {\n  if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n    throw new Error(\n      `minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`,\n    )\n  }\n  const oneSidedAlpha = (1 - confidence) / 2\n  return Math.ceil(Math.log2(1 / oneSidedAlpha))\n}\n\n/**\n * Tests whether a paired candidate-minus-baseline delta clears a threshold.\n *\n * At 20 or more pairs, the percentile bootstrap lower bound carries the\n * decision. Below that point the interval is descriptive only, so the function\n * switches to a pre-registered one-sided exact sign test. The exact path is\n * deliberately conservative: it requires both a point estimate above the\n * threshold and enough consistently positive paired differences.\n *\n * ## A zero-width interval is never significant\n *\n * When every paired delta is identical the resample distribution is a point\n * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n\n * identical deltas of g. Neither says the effect is certain — both say the\n * sample carries no information about how far the estimate could be wrong, and\n * `low > threshold` then answers on the point estimate alone. It fails in both\n * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a\n * tie-dominated pass/fail comparison laundered a regression into a\n * noninferiority pass, and `[g, g]` clears every threshold below g with no\n * spread behind it. Under a bounded asymmetric null whose true mean paired\n * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —\n * every sample that misses the drop is exactly that shape, and deciding on\n * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.\n *\n * So `indeterminate` is reported and `significant` is false whenever the\n * interval has zero width, on BOTH paths: at small n the exact sign test is a\n * test of the MEDIAN and a zero-spread sample is precisely where it stops\n * saying anything about the mean the caller is thresholding.\n *\n * `threshold` may be negative — that is a noninferiority margin, and it is the\n * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome\n * the percentile bootstrap is not a valid interval at a nonzero margin at all;\n * use {@link decidePairedPromotion}, which routes those to Tango's score\n * interval, rather than thresholding this function's bootstrap directly.\n */\nexport function pairedDeltaTest(\n  before: number[],\n  after: number[],\n  options: PairedDeltaTestOptions = {},\n): PairedDeltaTestResult {\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`)\n  }\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n  const bootstrap = pairedBootstrap(before, after, options)\n  const sufficient = bootstrap.n >= minimumPairs\n  // A point-mass resample distribution is an absence of evidence, not a\n  // certainty, and it is the shape that clears every negative threshold. Both\n  // paths refuse it: see the \"zero-width\" section of this function's docs.\n  const indeterminate =\n    bootstrap.n > 0 &&\n    (!Number.isFinite(bootstrap.low) ||\n      !Number.isFinite(bootstrap.high) ||\n      bootstrap.low === bootstrap.high)\n\n  if (bootstrap.gateEligible) {\n    return {\n      bootstrap,\n      method: 'bootstrap-ci',\n      pValue: null,\n      minimumPairs,\n      sufficient,\n      indeterminate,\n      significant: sufficient && !indeterminate && bootstrap.low > threshold,\n    }\n  }\n\n  const differences = before.map((value, index) => after[index]! - value - threshold)\n  const exact = pairedSignTest(differences, 'greater')\n  const estimate = options.statistic === 'mean' ? bootstrap.mean : bootstrap.median\n  return {\n    bootstrap,\n    method: 'exact-sign',\n    pValue: exact.pValue,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    significant:\n      sufficient &&\n      !indeterminate &&\n      estimate > threshold &&\n      exact.pValue <= (1 - bootstrap.confidence) / 2,\n  }\n}\n","/**\n * @module\n * ONE rule for \"does this paired interval clear a promotion threshold\".\n *\n * The rule below was derived on `HeldOutGate` (#479) after the same estimator\n * bug shipped twice. It then turned out that a SECOND gate — the composable\n * `heldOutGate`, plus everything else routed through `heldoutSignificance` —\n * still carried the original defect, because the rule had been written into one\n * gate's method body rather than into a shared function. Two copies of a\n * statistical rule is how a defect survives in one of them, so there is now\n * exactly one copy and both gates call it.\n *\n * Three things the rule does that a bare `pairedBootstrap(...).low > threshold`\n * does not:\n *\n * 1. **Two-point (pass/fail) outcomes decide on Tango's SCORE interval.** On a\n *    pass/fail eval the paired delta vector is dominated by ties, so the\n *    bootstrap of the mean is a resample of a lattice with three atoms and its\n *    percentile interval is not valid at a nonzero margin. The score interval\n *    (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under\n *    each hypothesised margin instead of fixing it at the observed value, which\n *    is the only construction that stays a confidence interval as the margin\n *    moves off zero — the regime every noninferiority threshold lives in.\n *    Measured on the composable gate before this change, at a true risk\n *    difference sitting exactly on the production caller's -0.05 margin and a\n *    nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.\n * 2. **McNemar's exact test holds a VETO at every non-negative threshold.**\n *    Redundant with the interval by construction and kept anyway, so that\n *    swapping the estimator for one without that duality cannot silently\n *    reintroduce \"promotes what the exact test refuses\". Witness: n = 6, b = 5,\n *    c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs\n *    (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE\n *    threshold is a noninferiority question, which McNemar's test of \"no\n *    difference\" is not the right test for, so the veto does not apply there.\n * 3. **A ZERO-WIDTH interval is refused, wherever it sits.** At [0, 0] it\n *    cannot tell a gain from a regression and clears every negative threshold.\n *    Away from zero it fails the opposite way: n identical positive deltas give\n *    [g, g], which clears threshold 0 on no spread at all. Both are an absence\n *    of evidence. Measured on the composable gate before this change, under a\n *    bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %\n *    false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.\n *\n * Orthogonal to the small-sample switch inside {@link pairedDeltaTest}: that\n * picks the TEST from the sample size (bootstrap CI at n >= 20, pre-registered\n * exact sign test below it), this picks the ESTIMATOR from the outcome's shape.\n * Both are needed — an exact sign test applied to a tie-pinned median is still\n * blind, and a mean bootstrap CI at n = 6 is still not a valid test.\n */\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n  type PairedBootstrapResult,\n  pairedBinaryScale,\n  pairedDeltaTieFraction,\n  pairedRiskDifferenceExact,\n  pairedRiskDifferenceScore,\n} from './statistics'\n\n/** Which paired estimator produced the deciding interval. */\nexport type PairedDecisionStatistic =\n  | 'paired_risk_difference'\n  | 'mean_bootstrap'\n  | 'median_bootstrap'\n\n/** Which test carried the decision, given the estimator and the sample size. */\nexport type PairedDecisionMethod = 'score-interval' | 'bootstrap-ci' | 'exact-sign'\n\n/** McNemar's exact paired-binary evidence, on the two-point path only. */\nexport interface PairedMcNemarEvidence {\n  /** Discordant pairs the treatment won. */\n  b: number\n  /** Discordant pairs the control won. */\n  c: number\n  /** b + c — the only pairs carrying information. */\n  nDiscordant: number\n  /** Two-sided exact p-value. */\n  pValue: number\n}\n\nexport interface PairedPromotionDecisionOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement, in the\n   *  caller's native units. May be negative (a noninferiority margin). Default 0. */\n  threshold?: number\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples, on the paths where a bootstrap decides. Default 2000. */\n  resamples?: number\n  /** Deterministic bootstrap seed. Omitted ⇒ derived from the deltas. */\n  seed?: number\n  /** Caller-required paired observations. The exact test may impose a higher\n   *  minimum; the effective one is reported as `minimumPairs`. */\n  minPairs?: number\n  /**\n   * `'mean'` (default) routes by SHAPE: a two-point (pass/fail) outcome on any\n   * encoding decides on the score interval, everything else on the mean\n   * bootstrap. `'median'` forces the median bootstrap on every input, including\n   * shapes where it is structurally blind — kept for callers who want outlier\n   * robustness on genuinely continuous outcomes and accept that cost.\n   */\n  statistic?: 'mean' | 'median'\n}\n\nexport interface PairedPromotionDecision {\n  /** Paired observations supplied. */\n  n: number\n  /** Threshold the interval was judged against, native units. */\n  threshold: number\n  confidence: number\n  statistic: PairedDecisionStatistic\n  method: PairedDecisionMethod\n  /** Common positive level of a two-point outcome ({0,1} ⇒ 1, {0,100} ⇒ 100),\n   *  or null when the outcome is not two-point. Non-null is exactly the\n   *  condition for the `paired_risk_difference` path, and it is the factor\n   *  `delta` / `low` / `high` were rescaled by. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n  /** Point estimate of the DECIDING statistic, in the caller's native units. */\n  delta: number\n  /** Lower bound of the DECIDING interval, native units. */\n  low: number\n  /** Upper bound of the DECIDING interval, native units. */\n  high: number\n  /** The bootstrap that decided, or null when the score interval did. Callers\n   *  that need a bootstrap as a diagnostic on the two-point path compute their\n   *  own — it is not computed here, so the binary path costs no resamples. */\n  bootstrap: PairedBootstrapResult | null\n  /** McNemar's exact evidence, or null off the two-point path. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** Exact one-sided sign-test p-value on the small-sample bootstrap path;\n   *  null otherwise. */\n  pValue: number | null\n  /** Effective observation minimum after accounting for confidence. */\n  minimumPairs: number\n  /** n >= minimumPairs. */\n  sufficient: boolean\n  /** The deciding interval is zero-width or non-finite — no evidence in either\n   *  direction, so it cannot clear any threshold on evidence. */\n  indeterminate: boolean\n  /** McNemar's exact test refuses at a non-negative threshold. */\n  exactTestVetoes: boolean\n  /** The deciding interval clears the threshold, ignoring the other two guards. */\n  clearsThreshold: boolean\n  /** `sufficient && !indeterminate && clearsThreshold && !exactTestVetoes` —\n   *  the whole rule. */\n  promote: boolean\n  /** What `delta` measures, for a reason string. */\n  label: 'success-rate' | 'mean' | 'median'\n  /** Why a zero-width interval is zero-width; empty when it is not. */\n  indeterminateCause: string\n  /** Sentence naming the test when the exact sign test decided; else empty. */\n  methodDetail: string\n}\n\n/** The shape facts that pick the estimator, without computing an interval. */\nexport interface PairedDecisionShape {\n  statistic: PairedDecisionStatistic\n  /** Common positive level of a two-point outcome; null when not two-point. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n}\n\n/**\n * Which estimator {@link decidePairedPromotion} would use on this data, and the\n * shape facts behind it — for callers that must report the shape on a path\n * where no interval is computed at all (an early rejection, or zero pairs).\n * Cheap: no bootstrap, no interval.\n */\nexport function pairedDecisionShape(\n  before: number[],\n  after: number[],\n  statistic: 'mean' | 'median' = 'mean',\n): PairedDecisionShape {\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(\n      `decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`,\n    )\n  }\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n  const n = before.length\n  const sufficient = n >= minimumPairs\n  const { binaryScale, tieFraction } = pairedDecisionShape(before, after, options.statistic)\n\n  let core: {\n    statistic: PairedDecisionStatistic\n    method: PairedDecisionMethod\n    delta: number\n    low: number\n    high: number\n    bootstrap: PairedBootstrapResult | null\n    mcnemar: PairedMcNemarEvidence | null\n    pValue: number | null\n    clearsThreshold: boolean\n    label: PairedPromotionDecision['label']\n    methodDetail: string\n  }\n\n  if (binaryScale !== null) {\n    // Normalise the two-point encoding to {0,1} so the estimators see the\n    // pass/fail structure, then rescale the answer back into the caller's\n    // native units — the threshold is read in the units of the scores, so a\n    // 0-100 pass/fail dimension must be gated in points, not in a rate.\n    const unitControl = before.map((v) => v / binaryScale)\n    const unitTreatment = after.map((v) => v / binaryScale)\n    // TWO estimators with different jobs, because no single one does both.\n    // `exact` is the authority on RD = 0 and supplies the veto; its\n    // Clopper-Pearson interval conditions on the discordant count and is\n    // therefore NOT a confidence interval at a nonzero margin. `score` is\n    // Tango's, which is.\n    const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence)\n    const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence)\n    const low = score.lower * binaryScale\n    core = {\n      statistic: 'paired_risk_difference',\n      method: 'score-interval',\n      delta: score.riskDifference * binaryScale,\n      low,\n      high: score.upper * binaryScale,\n      bootstrap: null,\n      mcnemar: {\n        b: exact.b,\n        c: exact.c,\n        nDiscordant: exact.nDiscordant,\n        pValue: exact.pValue,\n      },\n      pValue: null,\n      clearsThreshold: low > threshold,\n      label: 'success-rate',\n      methodDetail: '',\n    }\n  } else {\n    const bootstrapStatistic = options.statistic === 'median' ? 'median' : 'mean'\n    const test = pairedDeltaTest(before, after, {\n      confidence,\n      resamples: options.resamples,\n      statistic: bootstrapStatistic,\n      seed: options.seed,\n      threshold,\n      minPairs: options.minPairs,\n    })\n    const ci = test.bootstrap\n    core = {\n      statistic: bootstrapStatistic === 'mean' ? 'mean_bootstrap' : 'median_bootstrap',\n      method: test.method,\n      delta: bootstrapStatistic === 'mean' ? ci.mean : ci.median,\n      low: ci.low,\n      high: ci.high,\n      bootstrap: ci,\n      mcnemar: null,\n      pValue: test.pValue,\n      clearsThreshold: test.significant,\n      label: bootstrapStatistic,\n      methodDetail:\n        test.method === 'exact-sign'\n          ? ` Below ${test.minimumPairs} pairs the interval is descriptive only;` +\n            ` the decision is the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n          : '',\n    }\n  }\n\n  const indeterminate =\n    !Number.isFinite(core.low) || !Number.isFinite(core.high) || core.low === core.high\n  const indeterminateCause = !indeterminate\n    ? ''\n    : tieFraction === 1\n      ? 'every paired delta is an exact tie'\n      : core.mcnemar !== null && core.mcnemar.nDiscordant === 0\n        ? 'every pair is concordant (0 discordant pairs)'\n        : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`\n  // Only at a non-negative threshold: a negative threshold asks a\n  // noninferiority question, which McNemar's test of \"no difference\" does not\n  // answer.\n  const exactTestVetoes =\n    core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence)\n\n  return {\n    n,\n    threshold,\n    confidence,\n    binaryScale,\n    tieFraction,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    indeterminateCause,\n    exactTestVetoes,\n    promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,\n    ...core,\n  }\n}\n\nfunction fmt(x: number): string {\n  return x.toFixed(4)\n}\n","/**\n * Statistical held-out promotion machinery — the trustworthy core the\n * point-estimate `heldout-delta` gate lacked.\n *\n * The shipped false positive it prevents: a winner re-scored against the\n * baseline on the holdout read run-to-run model NOISE (e.g. 91 vs 95) as a\n * \"+4 lift\" and shipped, because the gate compared point estimates with no\n * confidence interval. Here we pair candidate vs baseline holdout observations\n * and bootstrap a CI on the paired delta — a candidate ships only when the CI\n * lower bound clears the effect-size threshold (the gain is real at the\n * confidence level, not noise), and is blocked when a critical dimension\n * (e.g. `hallucination_free` for a legal agent) significantly regresses even if\n * the net composite rose (anti-Goodhart).\n *\n * Two traps this module is built around (both produce a NEW false positive if\n * gotten wrong):\n *   1. PAIRING GRANULARITY — pairs by FULL `cellId` (`scenario:rep`), never by\n *      `scenarioId` (which averages reps away and destroys the within-pair\n *      variance reduction that makes a paired bootstrap tighter than unpaired).\n *      One paired observation per cell ⇒ reps multiply n.\n *   2. SCALE — a judge may emit composites/dimensions on [0,1] or 0-100. The\n *      threshold + tolerance are interpreted in the judge's NATIVE scale; the\n *      per-dimension tolerance auto-scales off the observed baseline magnitudes\n *      so `-0.10` on [0,1] doesn't silently become a no-op on a 0-100 dimension.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n  type PairedPromotionDecision,\n} from '../../paired-promotion-decision'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { JudgeScore } from '../types'\n\n/** Tie fraction at/above which a gate annotates its verdict with the tie share.\n *  Tie-domination of the median bites structurally at >= 0.5 (the median is then\n *  0 by construction); 0.4 is a softer warn threshold that flags a run APPROACHING\n *  that regime, so an operator sees it before the median goes fully blind. */\nexport const TIE_WARN_FRACTION = 0.4\n\nexport interface PairedHoldout {\n  /** Baseline scalar per paired cell (same order as `after`/`cellIds`). */\n  before: number[]\n  /** Candidate scalar per paired cell. */\n  after: number[]\n  /** The full cellIds (`scenario:rep`) that paired, in order. */\n  cellIds: string[]\n}\n\n/**\n * Pair candidate vs baseline holdout observations by FULL cellId. `select`\n * pulls the scalar from a cell's judge reports (composite, or a named\n * dimension); a cell contributes the mean of `select` across its judges. Cells\n * whose scenario is not in `scenarioIds`, or where `select` is undefined for\n * every judge on either side, are skipped on BOTH sides so the arrays stay\n * paired. Throws when the two maps disagree on which holdout cells exist — a\n * load-bearing invariant: the baseline + winner holdout campaigns run the same\n * scenarios with the same seed base, so their cellIds MUST align; a mismatch\n * means a silent pairing bug, not a soft fallback.\n */\nexport function pairHoldout(\n  candidate: Map<string, Record<string, JudgeScore>>,\n  baseline: Map<string, Record<string, JudgeScore>>,\n  scenarioIds: Set<string>,\n  select: (s: JudgeScore) => number | undefined,\n): PairedHoldout {\n  const cellValue = (\n    byCell: Map<string, Record<string, JudgeScore>>,\n    cellId: string,\n  ): number | undefined => {\n    const scores = byCell.get(cellId)\n    if (!scores) return undefined\n    const vals: number[] = []\n    for (const s of Object.values(scores)) {\n      if (s.failed === true) {\n        throw new Error(`pairHoldout: cell '${cellId}' contains a failed judge score`)\n      }\n      const v = select(s)\n      if (typeof v === 'number' && !Number.isFinite(v)) {\n        throw new Error(`pairHoldout: cell '${cellId}' contains a non-finite selected score`)\n      }\n      if (typeof v === 'number') vals.push(v)\n    }\n    if (vals.length === 0) return undefined\n    return vals.reduce((a, b) => a + b, 0) / vals.length\n  }\n\n  const inScope = (cellId: string) => scenarioIds.has(cellId.split(':')[0] ?? '')\n  const candCells = [...candidate.keys()].filter(inScope).sort()\n  const baseCells = [...baseline.keys()].filter(inScope).sort()\n  // Alignment invariant — the holdout campaigns share scenarios + seed, so the\n  // cell sets must be identical. Differ ⇒ a real pairing bug; fail loud.\n  if (candCells.length !== baseCells.length || candCells.some((c, i) => c !== baseCells[i])) {\n    throw new Error(\n      `pairHoldout: candidate/baseline holdout cells do not align — ` +\n        `candidate=[${candCells.join(',')}] baseline=[${baseCells.join(',')}]. ` +\n        `Both holdout campaigns must run the same scenarios with the same seed base.`,\n    )\n  }\n\n  const before: number[] = []\n  const after: number[] = []\n  const cellIds: string[] = []\n  for (const cellId of candCells) {\n    const b = cellValue(baseline, cellId)\n    const a = cellValue(candidate, cellId)\n    // A scalar absent on both sides means that dimension was not scored. A\n    // one-sided absence is asymmetric evidence loss, never a row to discard.\n    if (b === undefined && a === undefined) continue\n    if (b === undefined || a === undefined) {\n      throw new Error(`pairHoldout: cell '${cellId}' has a selected score on only one arm`)\n    }\n    before.push(b)\n    after.push(a)\n    cellIds.push(cellId)\n  }\n  return { before, after, cellIds }\n}\n\nexport interface HeldoutSignificance {\n  paired: PairedHoldout\n  /**\n   * The paired bootstrap on the requested statistic (MEAN by default — see the\n   * tie note on `heldoutSignificance`).\n   *\n   * DIAGNOSTIC, not necessarily the interval the verdict keyed on. On a\n   * two-point (pass/fail) outcome the decision routes to Tango's score interval\n   * instead, because a percentile bootstrap of the mean over a three-atom\n   * lattice is not a valid interval at a nonzero margin. Read\n   * `decision.low`/`decision.high` for the interval that actually decided, and\n   * `decisionStatistic` for which one it is.\n   */\n  bootstrap: PairedBootstrapResult\n  /** The MEDIAN paired-delta bootstrap, reported as a diagnostic. When many\n   *  scenarios are tied (both sides solve them), the median is pinned near 0\n   *  regardless of the mean lift — comparing the two exposes tie-domination. */\n  medianBootstrap: PairedBootstrapResult\n  /**\n   * The full promotion decision: which estimator the outcome's shape admits,\n   * the interval it produced, McNemar's exact veto on the two-point path, and\n   * whether the interval was zero-width (no evidence in either direction). The\n   * single source of `significant`.\n   */\n  decision: PairedPromotionDecision\n  /** Which paired estimator the verdict was decided on. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on the two-point path; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** Fraction of paired observations that are exact ties (|delta| < 1e-9). A\n   *  high tie fraction is WHY a median-based gate would have missed a real lift;\n   *  it is the observability the tie fix adds. */\n  tieFraction: number\n  /** n paired observations. */\n  n: number\n  /** Effective minimum after applying the bootstrap's hard statistical floor. */\n  minimumRequired: number\n  /** Statistical method that carried the decision. */\n  decisionMethod: PairedDecisionMethod\n  /** Exact one-sided p-value on the small-sample path; otherwise null. */\n  pValue: number | null\n  /** True iff n >= minimumRequired, the DECIDING interval has nonzero width,\n   *  its lower bound clears the threshold, and McNemar's exact test does not\n   *  veto at a non-negative threshold. */\n  significant: boolean\n  /** Set when n < minimumRequired — too little evidence to claim significance. */\n  fewRuns: boolean\n}\n\nexport interface HeldoutSignificanceOptions {\n  deltaThreshold?: number\n  minProductiveRuns?: number\n  confidence?: number\n  resamples?: number\n  /** Fixed by default for a deterministic, reproducible gate verdict. */\n  seed?: number\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * Significance of the held-out composite lift: ship only when the lower bound\n * of the interval the outcome's shape ADMITS exceeds `deltaThreshold` (default\n * 0 ⇒ \"confidently positive\"). Interpret `deltaThreshold` in the judge's native\n * scale.\n *\n * The decision is delegated whole to {@link decidePairedPromotion}, the one\n * copy of the rule (`src/paired-promotion-decision.ts`), which `HeldOutGate`\n * also calls. That module's header carries the measurements; the short version\n * is three guards a bare `bootstrap.low > threshold` does not have:\n *\n *   - a two-point (pass/fail) outcome decides on Tango's SCORE interval, the\n *     only paired-binary construction that stays valid at a nonzero margin;\n *   - McNemar's exact test VETOES at any non-negative threshold;\n *   - a ZERO-WIDTH interval is refused rather than promoted, in either\n *     direction — [0,0] clears every negative threshold and [g,g] clears every\n *     threshold below g, and both are an absence of evidence, not a result.\n *\n * Measured on this function before those guards landed, at a nominal 5 %:\n * 14.60 % false promotion at n = 40 on a paired-binary noninferiority boundary,\n * and 88.50 % at n = 6 under a bounded asymmetric null whose true mean paired\n * delta is exactly 0.\n *\n * At small n, where the percentile bootstrap is descriptive only, a\n * pre-registered exact sign test still carries the bootstrap path.\n */\nexport function heldoutSignificance(\n  paired: PairedHoldout,\n  opts: HeldoutSignificanceOptions = {},\n): HeldoutSignificance {\n  const deltaThreshold = opts.deltaThreshold ?? 0\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  // DEFAULT to the MEAN paired delta, not the median. The median is destroyed by\n  // TIES: whenever both baseline and candidate solve a holdout scenario (a common\n  // case once the agent is decent — and INCREASINGLY common as you add holdout\n  // scenarios for statistical power), that scenario contributes delta 0. When\n  // >=50% of paired cells are ties the median is pinned at 0 regardless of a large,\n  // consistent lift on the rest, and the gate holds a genuinely better candidate.\n  // (Measured live, supervisor-lab run 6: 40 cells, 20 ties, MEAN +0.177, MEDIAN 0\n  // → false hold. Doubling the holdout for \"power\" made it WORSE by adding ties.)\n  // The mean equals the reported aggregate lift, ties correctly contribute 0\n  // without dominating, and it is the textbook paired-comparison estimator; the\n  // median is kept as a reported diagnostic. Callers wanting outlier-robustness at\n  // the cost of tie-blindness can still pass `statistic: 'median'`.\n  const statistic = opts.statistic ?? 'mean'\n  const decision = decidePairedPromotion(paired.before, paired.after, {\n    confidence,\n    resamples,\n    statistic,\n    seed,\n    threshold: deltaThreshold,\n    minPairs: opts.minProductiveRuns,\n  })\n  // `decision.bootstrap` is null exactly when the score interval decided, so\n  // the requested-statistic bootstrap is computed here for the diagnostic\n  // field. Same two bootstraps as before on every path.\n  const bootstrap =\n    decision.bootstrap ??\n    pairedBootstrap(paired.before, paired.after, { confidence, resamples, statistic, seed })\n  const medianBootstrap =\n    statistic === 'median'\n      ? bootstrap\n      : pairedBootstrap(paired.before, paired.after, {\n          confidence,\n          resamples,\n          statistic: 'median',\n          seed,\n        })\n  const n = paired.before.length\n  let ties = 0\n  for (let i = 0; i < n; i += 1) {\n    const after = paired.after[i] ?? 0\n    const before = paired.before[i] ?? 0\n    if (Math.abs(after - before) < 1e-9) ties += 1\n  }\n  const tieFraction = n === 0 ? 0 : ties / n\n  return {\n    paired,\n    bootstrap,\n    medianBootstrap,\n    decision,\n    decisionStatistic: decision.statistic,\n    mcnemar: decision.mcnemar,\n    tieFraction,\n    n,\n    minimumRequired: decision.minimumPairs,\n    decisionMethod: decision.method,\n    pValue: decision.pValue,\n    significant: decision.promote,\n    fewRuns: !decision.sufficient,\n  }\n}\n\nexport interface DimensionRegression {\n  dimension: string\n  /** Paired bootstrap on (candidate − baseline). DIAGNOSTIC on a pass/fail\n   *  dimension, where `ci` carries the interval that decided instead. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low` is the lower bound of. `'mean'`\n   *  unless the caller asked for the median. `bootstrap.median` still carries\n   *  the median point estimate either way. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval `regressed` was decided on, in the dimension's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail dimension; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width — no evidence in either direction. */\n  indeterminate: boolean\n  /** True iff the candidate may have regressed this dimension by more than\n   *  tolerance: the lower bound of the DECIDING interval on (candidate −\n   *  baseline) is below −tolerance, OR the exact small-sample test proves a drop\n   *  past tolerance. */\n  regressed: boolean\n  tolerance: number\n  n: number\n}\n\n/** Detect the native scale of a set of scores: 0-100 when any magnitude clears\n *  1.5, else [0,1]. Used to auto-scale the regression tolerance so a default\n *  expressed for [0,1] is not silently a no-op on a 0-100 dimension. */\nexport function detectScale(values: number[]): 1 | 100 {\n  return values.some((v) => Math.abs(v) > 1.5) ? 100 : 1\n}\n\n/** Per-critical-dimension regression guard. For each dimension, pair the\n *  candidate vs baseline values by full cellId and bootstrap the paired delta;\n *  a dimension is \"regressed\" when the CI lower bound < −tolerance (conservative\n *  — blocks if the credible worst case exceeds tolerance, which is the right\n *  posture for safety dimensions like `hallucination_free`). When `tolerance`\n *  is omitted it auto-scales: 0.05 on [0,1], 5 on 0-100.\n *\n *  The interval comes from {@link decidePairedPromotion}, so a pass/fail\n *  dimension is judged on Tango's score interval rather than a percentile\n *  bootstrap of the mean — `tolerance` is a NONZERO margin, and the bootstrap\n *  is not a valid interval at one. That matters most here because this guard\n *  fails OPEN by construction: `tolerance` is positive, so an interval pinned at\n *  [0,0] never satisfies `low < −tolerance` and a real regression on a safety\n *  dimension would be reported as `regressed: false`. On the median it fails the\n *  same way for the same reason — when most pairs tie, which is automatic for a\n *  pass/fail dimension on {0,1} and on the 0-100 encoding `detectScale` exists\n *  to support, the median CI collapses to [0,0]. Pass `statistic: 'median'` to\n *  restore the pre-0.134 behaviour. */\nexport function dimensionRegressions(\n  candidate: Map<string, Record<string, JudgeScore>>,\n  baseline: Map<string, Record<string, JudgeScore>>,\n  scenarioIds: Set<string>,\n  criticalDimensions: string[],\n  opts: {\n    tolerance?: number\n    confidence?: number\n    resamples?: number\n    seed?: number\n    /** Paired statistic the CI is computed on. Default `'mean'` — see\n     *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n    statistic?: 'mean' | 'median'\n  } = {},\n): DimensionRegression[] {\n  const out: DimensionRegression[] = []\n  for (const dim of criticalDimensions) {\n    const paired = pairHoldout(candidate, baseline, scenarioIds, (s) => s.dimensions[dim])\n    if (paired.before.length === 0) continue // dimension not scored on this judge\n    const tolerance = opts.tolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    const shared = {\n      confidence: opts.confidence ?? 0.95,\n      resamples: opts.resamples ?? 2000,\n      statistic: bootstrapStatistic,\n      seed: opts.seed ?? 1337,\n    }\n    const guard = decidePairedPromotion(paired.before, paired.after, shared)\n    const regression = decidePairedPromotion(paired.after, paired.before, {\n      ...shared,\n      threshold: tolerance,\n    })\n    const bootstrap = guard.bootstrap ?? pairedBootstrap(paired.before, paired.after, shared)\n    out.push({\n      dimension: dim,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: guard.low, high: guard.high },\n      decisionStatistic: guard.statistic,\n      mcnemar: guard.mcnemar,\n      indeterminate: guard.indeterminate,\n      // Fires on EITHER burden of proof — see `floorBreached` in\n      // `promotion-policy.ts` for the same rule and the reasoning. The CI arm\n      // is the contract this interface and `default-production-gate`'s reason\n      // string both state (\"CI.low < -tolerance\"), read off the DECIDING\n      // interval; the exact-test arm adds the small-sample path where the\n      // bootstrap interval is descriptive only.\n      // The credible-worst-case arm stays on the BOOTSTRAP — see the same\n      // decision in `floorBreached` (`promotion-policy.ts`). On a pass/fail\n      // dimension the score interval is ±z²/(n+z²) even when every pair is\n      // concordant, so reading the floor off it would flag an unchanged safety\n      // dimension as regressed at any realistic n. The PROVEN-drop arm does use\n      // the shared rule, which is what closes the fail-open hole that mattered:\n      // a genuine pass/fail regression is now judged on an interval valid at\n      // the nonzero `tolerance`, instead of a bootstrap that is pinned wherever\n      // ties dominate.\n      regressed: bootstrap.low < -tolerance || regression.promote,\n      tolerance,\n      n: paired.before.length,\n    })\n  }\n  return out\n}\n","/**\n * Power preflight — \"can this budget detect the effect you are hunting?\"\n *\n * The failure it prevents (measured, twice): a live prompt-improvement campaign ran\n * 333 sandbox cells over 5.6 hours and produced a +0.08 holdout lift the ship gate\n * (paired bootstrap, CI.low > 0.05) could not distinguish from zero — because at\n * that holdout size and worker variance the MINIMUM DETECTABLE lift was larger than\n * any effect a prompt change plausibly produces. The budget was spent learning what\n * a 30-second calculation on the baseline cells already knew. No eval framework we\n * know of surfaces this; every underpowered improvement run everywhere ends in an\n * uninformative \"hold\".\n *\n * Model: the ship rule is `CI.low(paired Δ) > deltaThreshold`. Approximating the\n * bootstrap CI as normal, `CI.low ≈ effect − z·sd_Δ/√n`, so the smallest shippable\n * true effect is `MDE = deltaThreshold + z·sd_Δ/√n`. The paired-delta SD is unknown\n * before the candidate exists; we bound it by the zero-correlation case\n * `sd_Δ ≤ √2·sd_baseline` — a CONSERVATIVE (upper) MDE, which is the correct\n * direction for a warning. Pairing is per cell (`scenario:rep`), so reps multiply n.\n *\n * Standalone by design: feed it any baseline composites (a `gate:'none'` run, a\n * live-proof table) BEFORE budgeting the real search; `selfImprove` also attaches\n * it to every result and warns when the run was structurally unable to ship.\n */\n\nexport interface PowerPreflightOptions {\n  /** Per-cell baseline composites on the HOLDOUT scenarios (one per scenario:rep cell). */\n  baselineComposites: number[]\n  /** Paired observations the budgeted comparison will produce\n   *  (holdout scenarios × reps). Defaults to `baselineComposites.length`. */\n  pairedN?: number\n  /** The ship gate's effect-size threshold. Default 0.05 (defaultProductionGate). */\n  deltaThreshold?: number\n  /** CI confidence the gate uses. Default 0.95. */\n  confidence?: number\n  /** True when the holdout is scored by the SAME judge/scorer family as the gate\n   *  (selfImprove's default composition — one judge scores everything). Under a\n   *  shared channel, raising paired n reduces only the IDIOSYNCRATIC noise share;\n   *  systematic judge bias is untouched, so the MDE here is a lower bound and the\n   *  only full debiaser is an independent second scoring channel\n   *  (recursive-self-improvement S1c, closed form in EXP-023 P0). Default false. */\n  sharedScorerChannel?: boolean\n}\n\nexport interface PowerPreflight {\n  /** Paired observations the comparison will have. */\n  n: number\n  /** Baseline per-cell composite standard deviation (the variance the effect must beat). */\n  sd: number\n  /** Minimum detectable lift: the smallest TRUE effect the gate could ship at this budget. */\n  mde: number\n  /** Baseline holdout composite mean. */\n  baselineMean: number\n  /** Headroom to a perfect 1.0 composite (the largest achievable lift on a [0,1] judge). */\n  headroom: number\n  /** True when even the largest achievable effect (headroom) is below the MDE —\n   *  the run is structurally unable to ship regardless of proposal quality.\n   *  Only asserted for [0,1]-scaled judges (see `scaleAssumed`). */\n  underpowered: boolean\n  /** True when composites look [0,1]-scaled; headroom/underpowered are only\n   *  meaningful under that convention (0-100 judges get mde/sd/n but no verdict). */\n  scaleAssumed: boolean\n  deltaThreshold: number\n  confidence: number\n  /** Set when the holdout shares the gate's scoring channel: more cells cannot\n   *  buy back systematic judge bias — treat the MDE as a lower bound. */\n  sharedChannelCaveat?: string\n  /** One actionable sentence for humans and logs. */\n  recommendation: string\n}\n\n/** Two-sided z for the common confidence levels; interpolation is overkill here. */\nfunction zFor(confidence: number): number {\n  if (confidence >= 0.99) return 2.576\n  if (confidence >= 0.95) return 1.96\n  if (confidence >= 0.9) return 1.645\n  return 1.282\n}\n\n/** Estimate the minimum detectable lift a paired-holdout improvement run can\n *  ship at a given budget, from the baseline holdout composites — call it BEFORE\n *  spending a search to learn whether the effect you are hunting is even\n *  observable at this holdout size and worker variance. */\nexport function powerPreflight(opts: PowerPreflightOptions): PowerPreflight {\n  const composites = opts.baselineComposites.filter((v) => Number.isFinite(v))\n  if (composites.length < 3) {\n    throw new Error(\n      `powerPreflight: need >= 3 finite baseline composites to estimate variance, got ${composites.length}`,\n    )\n  }\n  const deltaThreshold = opts.deltaThreshold ?? 0.05\n  const confidence = opts.confidence ?? 0.95\n  const n = opts.pairedN ?? composites.length\n  if (n < 2) throw new Error(`powerPreflight: pairedN must be >= 2, got ${n}`)\n\n  const mean = composites.reduce((a, b) => a + b, 0) / composites.length\n  const variance =\n    composites.reduce((a, b) => a + (b - mean) * (b - mean), 0) / (composites.length - 1)\n  const sd = Math.sqrt(variance)\n  const z = zFor(confidence)\n  const mde = deltaThreshold + (z * Math.SQRT2 * sd) / Math.sqrt(n)\n\n  const scaleAssumed = composites.every((v) => v >= -0.001 && v <= 1.5)\n  const headroom = Math.max(0, 1 - mean)\n  const underpowered = scaleAssumed && mde > headroom\n\n  const sharedChannelCaveat = opts.sharedScorerChannel\n    ? 'Holdout and gate share one scoring channel: raising n/reps reduces only idiosyncratic noise — systematic judge bias remains and this MDE is a lower bound. Full debiasing needs an independent second scoring channel (different judge/benchmark family).'\n    : undefined\n\n  const recommendation = underpowered\n    ? `UNDERPOWERED: minimum detectable lift ${mde.toFixed(3)} exceeds the ${headroom.toFixed(3)} headroom above the baseline (${mean.toFixed(3)}) — no achievable effect can ship at this budget. Raise paired n (scenarios x reps) to ~${Math.ceil(((z * Math.SQRT2 * sd) / Math.max(headroom - deltaThreshold, 0.01)) ** 2)} or reduce worker variance before searching.`\n    : `Minimum detectable lift at n=${n}: ${mde.toFixed(3)} (baseline sd ${sd.toFixed(3)}). Effects smaller than this cannot clear the gate; budget the search for effects you believe exceed it.`\n\n  return {\n    n,\n    sd,\n    mde,\n    baselineMean: mean,\n    headroom,\n    underpowered,\n    scaleAssumed,\n    deltaThreshold,\n    confidence,\n    ...(sharedChannelCaveat ? { sharedChannelCaveat } : {}),\n    recommendation: sharedChannelCaveat\n      ? `${recommendation} ${sharedChannelCaveat}`\n      : recommendation,\n  }\n}\n"],"mappings":";;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACqCA,SAAgB,oBACd,QACA,OACA,YAA+B,QACV;CACrB,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,MAAM,cAAc,kBAAkB,QAAQ,KAAK;CACnD,IAAI,gBAAgB,MAClB,OAAO;EAAE,WAAW;EAA0B;EAAa;CAAY;CAEzE,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;AACvE;;;;;;AAOA,SAAgB,sBACd,QACA,OACA,UAA0C,CAAC,GAClB;CACzB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,gDAAgD,OAAO,OAAO,MAAM,MAAM,OAAO,EACnF;CAEF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,wDAAwD,WAAW;CAErF,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,eAAe,+BAA+B,UAAU;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MACR,mEAAmE,kBACrE;CAEF,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CACxB,MAAM,EAAE,aAAa,gBAAgB,oBAAoB,QAAQ,OAAO,QAAQ,SAAS;CAEzF,IAAI;CAcJ,IAAI,gBAAgB,MAAM;EAKxB,MAAM,cAAc,OAAO,KAAK,MAAM,IAAI,WAAW;EACrD,MAAM,gBAAgB,MAAM,KAAK,MAAM,IAAI,WAAW;EAMtD,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,MAAM,MAAM,QAAQ;EAC1B,OAAO;GACL,WAAW;GACX,QAAQ;GACR,OAAO,MAAM,iBAAiB;GAC9B;GACA,MAAM,MAAM,QAAQ;GACpB,WAAW;GACX,SAAS;IACP,GAAG,MAAM;IACT,GAAG,MAAM;IACT,aAAa,MAAM;IACnB,QAAQ,MAAM;GAChB;GACA,QAAQ;GACR,iBAAiB,MAAM;GACvB,OAAO;GACP,cAAc;EAChB;CACF,OAAO;EACL,MAAM,qBAAqB,QAAQ,cAAc,WAAW,WAAW;EACvE,MAAM,OAAO,gBAAgB,QAAQ,OAAO;GAC1C;GACA,WAAW,QAAQ;GACnB,WAAW;GACX,MAAM,QAAQ;GACd;GACA,UAAU,QAAQ;EACpB,CAAC;EACD,MAAM,KAAK,KAAK;EAChB,OAAO;GACL,WAAW,uBAAuB,SAAS,mBAAmB;GAC9D,QAAQ,KAAK;GACb,OAAO,uBAAuB,SAAS,GAAG,OAAO,GAAG;GACpD,KAAK,GAAG;GACR,MAAM,GAAG;GACT,WAAW;GACX,SAAS;GACT,QAAQ,KAAK;GACb,iBAAiB,KAAK;GACtB,OAAO;GACP,cACE,KAAK,WAAW,eACZ,UAAU,KAAK,aAAa,4FACyB,IAAI,KAAK,UAAU,CAAC,EAAE,KAC3E;EACR;CACF;CAEA,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KAAK,CAAC,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,QAAQ,KAAK;CACjF,MAAM,qBAAqB,CAAC,gBACxB,KACA,gBAAgB,IACd,uCACA,KAAK,YAAY,QAAQ,KAAK,QAAQ,gBAAgB,IACpD,kDACA,OAAO,KAAK,MAAM,8BAA8B,IAAI,KAAK,GAAG;CAIpE,MAAM,kBACJ,KAAK,YAAY,QAAQ,aAAa,KAAK,EAAE,KAAK,QAAQ,SAAS,IAAI;CAEzE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,cAAc,CAAC,iBAAiB,KAAK,mBAAmB,CAAC;EAClE,GAAG;CACL;AACF;AAEA,SAAS,IAAI,GAAmB;CAC9B,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC3RA,MAAa,oBAAoB;;;;;;;;;;;;AAsBjC,SAAgB,YACd,WACA,UACA,aACA,QACe;CACf,MAAM,aACJ,QACA,WACuB;EACvB,MAAM,SAAS,OAAO,IAAI,MAAM;EAChC,IAAI,CAAC,QAAQ,OAAO,KAAA;EACpB,MAAM,OAAiB,CAAC;EACxB,KAAK,MAAM,KAAK,OAAO,OAAO,MAAM,GAAG;GACrC,IAAI,EAAE,WAAW,MACf,MAAM,IAAI,MAAM,sBAAsB,OAAO,gCAAgC;GAE/E,MAAM,IAAI,OAAO,CAAC;GAClB,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAC7C,MAAM,IAAI,MAAM,sBAAsB,OAAO,uCAAuC;GAEtF,IAAI,OAAO,MAAM,UAAU,KAAK,KAAK,CAAC;EACxC;EACA,IAAI,KAAK,WAAW,GAAG,OAAO,KAAA;EAC9B,OAAO,KAAK,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK;CAChD;CAEA,MAAM,WAAW,WAAmB,YAAY,IAAI,OAAO,MAAM,GAAG,CAAC,CAAC,MAAM,EAAE;CAC9E,MAAM,YAAY,CAAC,GAAG,UAAU,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK;CAC7D,MAAM,YAAY,CAAC,GAAG,SAAS,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK;CAG5D,IAAI,UAAU,WAAW,UAAU,UAAU,UAAU,MAAM,GAAG,MAAM,MAAM,UAAU,EAAE,GACtF,MAAM,IAAI,MACR,2EACgB,UAAU,KAAK,GAAG,EAAE,cAAc,UAAU,KAAK,GAAG,EAAE,+EAExE;CAGF,MAAM,SAAmB,CAAC;CAC1B,MAAM,QAAkB,CAAC;CACzB,MAAM,UAAoB,CAAC;CAC3B,KAAK,MAAM,UAAU,WAAW;EAC9B,MAAM,IAAI,UAAU,UAAU,MAAM;EACpC,MAAM,IAAI,UAAU,WAAW,MAAM;EAGrC,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;EACxC,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAC3B,MAAM,IAAI,MAAM,sBAAsB,OAAO,uCAAuC;EAEtF,OAAO,KAAK,CAAC;EACb,MAAM,KAAK,CAAC;EACZ,QAAQ,KAAK,MAAM;CACrB;CACA,OAAO;EAAE;EAAQ;EAAO;CAAQ;AAClC;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuFA,SAAgB,oBACd,QACA,OAAmC,CAAC,GACf;CACrB,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAa1B,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,WAAW,sBAAsB,OAAO,QAAQ,OAAO,OAAO;EAClE;EACA;EACA;EACA;EACA,WAAW;EACX,UAAU,KAAK;CACjB,CAAC;CAID,MAAM,YACJ,SAAS,aACT,gBAAgB,OAAO,QAAQ,OAAO,OAAO;EAAE;EAAY;EAAW;EAAW;CAAK,CAAC;CACzF,MAAM,kBACJ,cAAc,WACV,YACA,gBAAgB,OAAO,QAAQ,OAAO,OAAO;EAC3C;EACA;EACA,WAAW;EACX;CACF,CAAC;CACP,MAAM,IAAI,OAAO,OAAO;CACxB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;EAC7B,MAAM,QAAQ,OAAO,MAAM,MAAM;EACjC,MAAM,SAAS,OAAO,OAAO,MAAM;EACnC,IAAI,KAAK,IAAI,QAAQ,MAAM,IAAI,MAAM,QAAQ;CAC/C;CACA,MAAM,cAAc,MAAM,IAAI,IAAI,OAAO;CACzC,OAAO;EACL;EACA;EACA;EACA;EACA,mBAAmB,SAAS;EAC5B,SAAS,SAAS;EAClB;EACA;EACA,iBAAiB,SAAS;EAC1B,gBAAgB,SAAS;EACzB,QAAQ,SAAS;EACjB,aAAa,SAAS;EACtB,SAAS,CAAC,SAAS;CACrB;AACF;;;;AA+BA,SAAgB,YAAY,QAA2B;CACrD,OAAO,OAAO,MAAM,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG,IAAI,MAAM;AACvD;;;;;;;;;;;;;;;;;;;AAoBA,SAAgB,qBACd,WACA,UACA,aACA,oBACA,OAQI,CAAC,GACkB;CACvB,MAAM,MAA6B,CAAC;CACpC,KAAK,MAAM,OAAO,oBAAoB;EACpC,MAAM,SAAS,YAAY,WAAW,UAAU,cAAc,MAAM,EAAE,WAAW,IAAI;EACrF,IAAI,OAAO,OAAO,WAAW,GAAG;EAChC,MAAM,YAAY,KAAK,aAAa,MAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC1F,MAAM,qBAAqB,KAAK,aAAA;EAChC,MAAM,SAAS;GACb,YAAY,KAAK,cAAc;GAC/B,WAAW,KAAK,aAAa;GAC7B,WAAW;GACX,MAAM,KAAK,QAAQ;EACrB;EACA,MAAM,QAAQ,sBAAsB,OAAO,QAAQ,OAAO,OAAO,MAAM;EACvE,MAAM,aAAa,sBAAsB,OAAO,OAAO,OAAO,QAAQ;GACpE,GAAG;GACH,WAAW;EACb,CAAC;EACD,MAAM,YAAY,MAAM,aAAa,gBAAgB,OAAO,QAAQ,OAAO,OAAO,MAAM;EACxF,IAAI,KAAK;GACP,WAAW;GACX;GACA;GACA,IAAI;IAAE,KAAK,MAAM;IAAK,MAAM,MAAM;GAAK;GACvC,mBAAmB,MAAM;GACzB,SAAS,MAAM;GACf,eAAe,MAAM;GAgBrB,WAAW,UAAU,MAAM,CAAC,aAAa,WAAW;GACpD;GACA,GAAG,OAAO,OAAO;EACnB,CAAC;CACH;CACA,OAAO;AACT;;;;ACjUA,SAAS,KAAK,YAA4B;CACxC,IAAI,cAAc,KAAM,OAAO;CAC/B,IAAI,cAAc,KAAM,OAAO;CAC/B,IAAI,cAAc,IAAK,OAAO;CAC9B,OAAO;AACT;;;;;AAMA,SAAgB,eAAe,MAA6C;CAC1E,MAAM,aAAa,KAAK,mBAAmB,QAAQ,MAAM,OAAO,SAAS,CAAC,CAAC;CAC3E,IAAI,WAAW,SAAS,GACtB,MAAM,IAAI,MACR,kFAAkF,WAAW,QAC/F;CAEF,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,IAAI,KAAK,WAAW,WAAW;CACrC,IAAI,IAAI,GAAG,MAAM,IAAI,MAAM,6CAA6C,GAAG;CAE3E,MAAM,OAAO,WAAW,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;CAChE,MAAM,WACJ,WAAW,QAAQ,GAAG,MAAM,KAAK,IAAI,SAAS,IAAI,OAAO,CAAC,KAAK,WAAW,SAAS;CACrF,MAAM,KAAK,KAAK,KAAK,QAAQ;CAC7B,MAAM,IAAI,KAAK,UAAU;CACzB,MAAM,MAAM,iBAAkB,IAAI,KAAK,QAAQ,KAAM,KAAK,KAAK,CAAC;CAEhE,MAAM,eAAe,WAAW,OAAO,MAAM,KAAK,SAAU,KAAK,GAAG;CACpE,MAAM,WAAW,KAAK,IAAI,GAAG,IAAI,IAAI;CACrC,MAAM,eAAe,gBAAgB,MAAM;CAE3C,MAAM,sBAAsB,KAAK,sBAC7B,8PACA,KAAA;CAEJ,MAAM,iBAAiB,eACnB,yCAAyC,IAAI,QAAQ,CAAC,EAAE,eAAe,SAAS,QAAQ,CAAC,EAAE,gCAAgC,KAAK,QAAQ,CAAC,EAAE,0FAA0F,KAAK,MAAO,IAAI,KAAK,QAAQ,KAAM,KAAK,IAAI,WAAW,gBAAgB,GAAI,MAAM,CAAC,EAAE,gDACzT,gCAAgC,EAAE,IAAI,IAAI,QAAQ,CAAC,EAAE,gBAAgB,GAAG,QAAQ,CAAC,EAAE;CAEvF,OAAO;EACL;EACA;EACA;EACA,cAAc;EACd;EACA;EACA;EACA;EACA;EACA,GAAI,sBAAsB,EAAE,oBAAoB,IAAI,CAAC;EACrD,gBAAgB,sBACZ,GAAG,eAAe,GAAG,wBACrB;CACN;AACF"}

package/dist/pre-registration-D94b7Of5.js

@@ -1,110 +0,0 @@
-import { a as hashCanonical, r as canonicalString } from "./canonical-DPyQ_rpt.js";
-import { createHash } from "node:crypto";
-//#region src/pre-registration.ts
-/**
-* Pre-registered hypotheses — declare what you're testing BEFORE the
-* run, check it AFTER. Prevents p-hacking, optional stopping, and the
-* "we ran until it looked good" failure mode.
-*
-* Manifest is a plain JSON-friendly object. Sign it with a content hash
-* + timestamp; the registered record becomes immutable. Post-run,
-* evaluate the manifest against observed results — the library refuses
-* to let you re-interpret a different metric as the declared one.
-*
-* A signed manifest is a portable record: it is written once and verified
-* later, possibly by a different release. `algo` names the digest scheme it
-* was signed under, and verification selects the encoder by that field, so a
-* manifest signed by an earlier release still verifies.
-*/
-/**
-* SHA-256 hex (full 64 chars) over the RFC 8785 canonical JSON encoding of
-* `obj` — the package's one identity scheme, shared with `ledger-core`.
-*
-* Values canonical JSON cannot represent faithfully — `undefined`, `NaN`,
-* class instances, cycles — are refused rather than coerced, because a
-* coercion maps two distinct records onto one digest.
-*
-* Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,
-* which takes a string input and returns a truncated 12-char prompt id.
-*
-* @example
-*   const hash = await hashJson({ id: '1', kind: 'spec' })
-*   // 'a3f1...' (64 hex chars)
-*/
-async function hashJson(obj) {
-	return hashCanonical(obj).slice(7);
-}
-/**
-* Key-sorted `JSON.stringify` digest. Private and read-only: it exists so a
-* manifest signed under `'sha256-content'` still verifies, and nothing that
-* WRITES a digest may call it.
-*/
-function legacyContentDigest(value) {
-	return createHash("sha256").update(JSON.stringify(sortKeysDeep(value)), "utf8").digest("hex");
-}
-function sortKeysDeep(value) {
-	if (value === null || typeof value !== "object") return value;
-	if (Array.isArray(value)) return value.map(sortKeysDeep);
-	const out = {};
-	for (const key of Object.keys(value).sort()) out[key] = sortKeysDeep(value[key]);
-	return out;
-}
-/**
-* Digest of a manifest under its own declared scheme, with `contentHash` and
-* `algo` stripped. Synchronous, so a caller that must fail before consuming an
-* observation does not have to await. Throws on an `algo` this release does
-* not know — an unverifiable manifest must not read as a valid one.
-*/
-function manifestContentDigest(manifest) {
-	const { contentHash: _contentHash, algo, ...rest } = manifest;
-	if (algo === void 0 || algo === "sha256-content") return legacyContentDigest(rest);
-	if (algo === "sha256-rfc8785") return createHash("sha256").update(canonicalString(rest), "utf8").digest("hex");
-	throw new Error(`pre-registration: unrecognized manifest hash algo '${String(algo)}'`);
-}
-/**
-* Sign a manifest with a SHA-256 content hash over its RFC 8785 canonical
-* JSON, with `contentHash` and `algo` stripped, and stamp the scheme in
-* `algo` so a later reader knows which encoder to verify with.
-*/
-async function signManifest(m) {
-	const signed = {
-		...m,
-		contentHash: "",
-		algo: "sha256-rfc8785"
-	};
-	return {
-		...signed,
-		contentHash: manifestContentDigest(signed)
-	};
-}
-/**
-* Verify that a signed manifest has not been tampered with, under the scheme
-* the manifest itself declares.
-*/
-async function verifyManifest(m) {
-	return manifestContentDigest(m) === m.contentHash;
-}
-/**
-* Evaluate a pre-registered hypothesis against observed results.
-* Mechanical — no re-interpretation permitted.
-*/
-async function evaluateHypothesis(manifest, observed) {
-	if (!await verifyManifest(manifest)) throw new Error("evaluateHypothesis: manifest content hash mismatch (tampered)");
-	const reasons = [];
-	if (!(manifest.direction === "increase" ? observed.effect > 0 : observed.effect < 0)) reasons.push("wrong_direction");
-	if (Math.abs(observed.effect) < manifest.minEffect) reasons.push("effect_too_small");
-	if (observed.pValue >= manifest.alpha) reasons.push("not_significant");
-	if (observed.n < manifest.preRegisteredN) reasons.push("undersampled");
-	return {
-		manifest,
-		observedN: observed.n,
-		observedEffect: observed.effect,
-		observedPValue: observed.pValue,
-		confirmed: reasons.length === 0,
-		rejectionReasons: reasons
-	};
-}
-//#endregion
-export { verifyManifest as a, signManifest as i, hashJson as n, manifestContentDigest as r, evaluateHypothesis as t };
-
-//# sourceMappingURL=pre-registration-D94b7Of5.js.map

package/dist/pre-registration-D94b7Of5.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"pre-registration-D94b7Of5.js","names":[],"sources":["../src/pre-registration.ts"],"sourcesContent":["/**\n * Pre-registered hypotheses — declare what you're testing BEFORE the\n * run, check it AFTER. Prevents p-hacking, optional stopping, and the\n * \"we ran until it looked good\" failure mode.\n *\n * Manifest is a plain JSON-friendly object. Sign it with a content hash\n * + timestamp; the registered record becomes immutable. Post-run,\n * evaluate the manifest against observed results — the library refuses\n * to let you re-interpret a different metric as the declared one.\n *\n * A signed manifest is a portable record: it is written once and verified\n * later, possibly by a different release. `algo` names the digest scheme it\n * was signed under, and verification selects the encoder by that field, so a\n * manifest signed by an earlier release still verifies.\n */\n\nimport { createHash } from 'node:crypto'\nimport { canonicalString, hashCanonical } from './ledger-core/canonical'\n\nexport interface HypothesisManifest {\n  id: string\n  /** Human prose — goes into the audit trail. */\n  hypothesis: string\n  /** Metric the hypothesis claims to move. */\n  metric: string\n  /** 'increase' = candidate should score higher than baseline; 'decrease' = lower. */\n  direction: 'increase' | 'decrease'\n  /** Minimum effect size to count (same units as the metric). */\n  minEffect: number\n  /** Alpha threshold. */\n  alpha: number\n  /** Target statistical power at which sample size was pre-computed. */\n  power: number\n  /** Declared N per arm before running. */\n  preRegisteredN: number\n  /** ISO8601 timestamp the manifest was registered. */\n  registeredAt: string\n  /** Optional identifiers to tie into the trace corpus. */\n  baselineLabel?: string\n  candidateLabel?: string\n}\n\n/**\n * Identifier for the hashing scheme used to produce `contentHash`.\n *\n * Both schemes are sha256 hex over the manifest with `contentHash` and `algo`\n * stripped, and differ only in how that manifest is serialized:\n *\n * - `'sha256-rfc8785'` — RFC 8785 canonical JSON. What {@link signManifest}\n *   emits.\n * - `'sha256-content'` — key-sorted `JSON.stringify`. Read-only: manifests\n *   signed by an earlier release carry it, or carry no `algo` at all, and\n *   {@link verifyManifest} still verifies them.\n */\nexport type SignedManifestAlgo = 'sha256-content' | 'sha256-rfc8785'\n\nexport interface SignedManifest extends HypothesisManifest {\n  /** sha256 hex of canonicalized manifest (everything except contentHash and algo). */\n  contentHash: string\n  /**\n   * Algorithm string describing how `contentHash` was produced.\n   *\n   * Optional on the type so serialized manifests without it still parse,\n   * but ALWAYS populated by {@link signManifest}. Consumers that want to\n   * enforce a known algorithm should reject manifests where this field\n   * is missing or unrecognized.\n   */\n  algo?: SignedManifestAlgo\n}\n\nexport interface HypothesisResult {\n  manifest: SignedManifest\n  observedN: number\n  observedEffect: number\n  observedPValue: number\n  /** True iff the observed effect hits the pre-declared direction with\n   *  magnitude ≥ minEffect AND p < alpha. */\n  confirmed: boolean\n  /** Enumerated reasons the hypothesis was rejected (each a machine-tag). */\n  rejectionReasons: Array<\n    'wrong_direction' | 'effect_too_small' | 'not_significant' | 'undersampled'\n  >\n  notes?: string\n}\n\n/**\n * SHA-256 hex (full 64 chars) over the RFC 8785 canonical JSON encoding of\n * `obj` — the package's one identity scheme, shared with `ledger-core`.\n *\n * Values canonical JSON cannot represent faithfully — `undefined`, `NaN`,\n * class instances, cycles — are refused rather than coerced, because a\n * coercion maps two distinct records onto one digest.\n *\n * Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,\n * which takes a string input and returns a truncated 12-char prompt id.\n *\n * @example\n *   const hash = await hashJson({ id: '1', kind: 'spec' })\n *   // 'a3f1...' (64 hex chars)\n */\nexport async function hashJson<T>(obj: T): Promise<string> {\n  return hashCanonical(obj).slice('sha256:'.length)\n}\n\n/**\n * Key-sorted `JSON.stringify` digest. Private and read-only: it exists so a\n * manifest signed under `'sha256-content'` still verifies, and nothing that\n * WRITES a digest may call it.\n */\nfunction legacyContentDigest(value: unknown): string {\n  return createHash('sha256')\n    .update(JSON.stringify(sortKeysDeep(value)), 'utf8')\n    .digest('hex')\n}\n\nfunction sortKeysDeep(value: unknown): unknown {\n  if (value === null || typeof value !== 'object') return value\n  if (Array.isArray(value)) return value.map(sortKeysDeep)\n  const out: Record<string, unknown> = {}\n  for (const key of Object.keys(value as Record<string, unknown>).sort()) {\n    out[key] = sortKeysDeep((value as Record<string, unknown>)[key])\n  }\n  return out\n}\n\n/**\n * Digest of a manifest under its own declared scheme, with `contentHash` and\n * `algo` stripped. Synchronous, so a caller that must fail before consuming an\n * observation does not have to await. Throws on an `algo` this release does\n * not know — an unverifiable manifest must not read as a valid one.\n */\nexport function manifestContentDigest(manifest: SignedManifest): string {\n  const { contentHash: _contentHash, algo, ...rest } = manifest\n  void _contentHash\n  if (algo === undefined || algo === 'sha256-content') return legacyContentDigest(rest)\n  if (algo === 'sha256-rfc8785') {\n    return createHash('sha256').update(canonicalString(rest), 'utf8').digest('hex')\n  }\n  throw new Error(`pre-registration: unrecognized manifest hash algo '${String(algo)}'`)\n}\n\n/**\n * Sign a manifest with a SHA-256 content hash over its RFC 8785 canonical\n * JSON, with `contentHash` and `algo` stripped, and stamp the scheme in\n * `algo` so a later reader knows which encoder to verify with.\n */\nexport async function signManifest(m: HypothesisManifest): Promise<SignedManifest> {\n  const signed: SignedManifest = { ...m, contentHash: '', algo: 'sha256-rfc8785' }\n  return { ...signed, contentHash: manifestContentDigest(signed) }\n}\n\n/**\n * Verify that a signed manifest has not been tampered with, under the scheme\n * the manifest itself declares.\n */\nexport async function verifyManifest(m: SignedManifest): Promise<boolean> {\n  return manifestContentDigest(m) === m.contentHash\n}\n\n/**\n * Evaluate a pre-registered hypothesis against observed results.\n * Mechanical — no re-interpretation permitted.\n */\nexport async function evaluateHypothesis(\n  manifest: SignedManifest,\n  observed: { n: number; effect: number; pValue: number },\n): Promise<HypothesisResult> {\n  if (!(await verifyManifest(manifest))) {\n    throw new Error('evaluateHypothesis: manifest content hash mismatch (tampered)')\n  }\n  const reasons: HypothesisResult['rejectionReasons'] = []\n  const directionOk = manifest.direction === 'increase' ? observed.effect > 0 : observed.effect < 0\n  if (!directionOk) reasons.push('wrong_direction')\n  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push('effect_too_small')\n  if (observed.pValue >= manifest.alpha) reasons.push('not_significant')\n  if (observed.n < manifest.preRegisteredN) reasons.push('undersampled')\n  return {\n    manifest,\n    observedN: observed.n,\n    observedEffect: observed.effect,\n    observedPValue: observed.pValue,\n    confirmed: reasons.length === 0,\n    rejectionReasons: reasons,\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoGA,eAAsB,SAAY,KAAyB;CACzD,OAAO,cAAc,GAAG,CAAC,CAAC,MAAM,CAAgB;AAClD;;;;;;AAOA,SAAS,oBAAoB,OAAwB;CACnD,OAAO,WAAW,QAAQ,CAAC,CACxB,OAAO,KAAK,UAAU,aAAa,KAAK,CAAC,GAAG,MAAM,CAAC,CACnD,OAAO,KAAK;AACjB;AAEA,SAAS,aAAa,OAAyB;CAC7C,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,MAAM,IAAI,YAAY;CACvD,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,OAAO,OAAO,KAAK,KAAgC,CAAC,CAAC,KAAK,GACnE,IAAI,OAAO,aAAc,MAAkC,IAAI;CAEjE,OAAO;AACT;;;;;;;AAQA,SAAgB,sBAAsB,UAAkC;CACtE,MAAM,EAAE,aAAa,cAAc,MAAM,GAAG,SAAS;CAErD,IAAI,SAAS,KAAA,KAAa,SAAS,kBAAkB,OAAO,oBAAoB,IAAI;CACpF,IAAI,SAAS,kBACX,OAAO,WAAW,QAAQ,CAAC,CAAC,OAAO,gBAAgB,IAAI,GAAG,MAAM,CAAC,CAAC,OAAO,KAAK;CAEhF,MAAM,IAAI,MAAM,sDAAsD,OAAO,IAAI,EAAE,EAAE;AACvF;;;;;;AAOA,eAAsB,aAAa,GAAgD;CACjF,MAAM,SAAyB;EAAE,GAAG;EAAG,aAAa;EAAI,MAAM;CAAiB;CAC/E,OAAO;EAAE,GAAG;EAAQ,aAAa,sBAAsB,MAAM;CAAE;AACjE;;;;;AAMA,eAAsB,eAAe,GAAqC;CACxE,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;;;;;AAMA,eAAsB,mBACpB,UACA,UAC2B;CAC3B,IAAI,CAAE,MAAM,eAAe,QAAQ,GACjC,MAAM,IAAI,MAAM,+DAA+D;CAEjF,MAAM,UAAgD,CAAC;CAEvD,IAAI,EADgB,SAAS,cAAc,aAAa,SAAS,SAAS,IAAI,SAAS,SAAS,IAC9E,QAAQ,KAAK,iBAAiB;CAChD,IAAI,KAAK,IAAI,SAAS,MAAM,IAAI,SAAS,WAAW,QAAQ,KAAK,kBAAkB;CACnF,IAAI,SAAS,UAAU,SAAS,OAAO,QAAQ,KAAK,iBAAiB;CACrE,IAAI,SAAS,IAAI,SAAS,gBAAgB,QAAQ,KAAK,cAAc;CACrE,OAAO;EACL;EACA,WAAW,SAAS;EACpB,gBAAgB,SAAS;EACzB,gBAAgB,SAAS;EACzB,WAAW,QAAQ,WAAW;EAC9B,kBAAkB;CACpB;AACF"}

package/dist/profile-cell.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"profile-cell.js","names":[],"sources":["../src/agent-profile-cell.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { hashJson } from './pre-registration'\n\nexport type AgentProfileCellSchemaVersion = 'agent-profile-cell/v1'\n\nexport type AgentProfileJsonObject = { [key: string]: AgentProfileJson }\n\nexport type AgentProfileJson =\n  | string\n  | number\n  | boolean\n  | null\n  | AgentProfileJson[]\n  | AgentProfileJsonObject\n\nexport type AgentProfileDimensionValue = string | number | boolean | null\n\nexport interface AgentProfileSource {\n  /** Runtime/profile contract being fingerprinted, e.g. `agent-interface-profile`. */\n  kind: string\n  /** sha256 over the canonical source profile object. */\n  hash: string\n}\n\nexport interface AgentProfileSourceInput {\n  kind: string\n  /** Precomputed sha256 for callers that already sign their profile artifact. */\n  hash?: string\n  /** Full canonical runtime profile; hashed and then discarded from the cell. */\n  profile?: AgentProfileJson\n}\n\nexport interface AgentProfileHarness {\n  id: string\n  version?: string\n  hash?: string\n}\n\nexport interface AgentProfileCellInput {\n  profileId: string\n  sourceProfile: AgentProfileSourceInput\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport interface AgentProfileCell {\n  schemaVersion: AgentProfileCellSchemaVersion\n  cellId: string\n  profileId: string\n  sourceProfile: AgentProfileSource\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport class AgentProfileCellValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\nconst SHA256_HEX = /^[0-9a-f]{64}$/\n/**\n * A cell id names the digest scheme that produced it. `sha256-rfc8785` is what\n * {@link buildAgentProfileCell} mints; the bare `sha256` form is read-only,\n * carried by cells built under an earlier release, and still verifies.\n */\nconst CELL_ID = /^agent-profile-cell:sha256(?:-rfc8785)?:[0-9a-f]{64}$/\nconst CELL_ID_PREFIX = 'agent-profile-cell:sha256-rfc8785:'\nconst LEGACY_CELL_ID_PREFIX = 'agent-profile-cell:sha256:'\n\nexport async function buildAgentProfileCell(\n  input: AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  const material = await normalizeAgentProfileCellInput(input)\n  const cellId = `${CELL_ID_PREFIX}${await hashJson(material)}`\n  return { ...material, cellId }\n}\n\nexport function agentProfileCellHashMaterial(\n  cell: AgentProfileCell,\n): Omit<AgentProfileCell, 'cellId'> {\n  const { cellId: _cellId, ...material } = cell\n  void _cellId\n  return normalizeAgentProfileCell(material)\n}\n\n/**\n * Verify an `AgentProfileCell`'s `cellId` matches the sha256 of its hash-material\n * fields, confirming the record has not been tampered with. The id names its own\n * digest scheme, so a cell minted by an earlier release verifies under that scheme.\n */\nexport async function verifyAgentProfileCell(cell: AgentProfileCell): Promise<boolean> {\n  validateAgentProfileCell(cell)\n  const material = agentProfileCellHashMaterial(cell)\n  if (cell.cellId.startsWith(CELL_ID_PREFIX)) {\n    return cell.cellId === `${CELL_ID_PREFIX}${await hashJson(material)}`\n  }\n  return cell.cellId === `${LEGACY_CELL_ID_PREFIX}${legacyCellDigest(material)}`\n}\n\n/**\n * Key-sorted `JSON.stringify` digest. Private and read-only: it verifies a cell\n * id minted before the RFC 8785 scheme, and no path that MINTS an id calls it.\n */\nfunction legacyCellDigest(value: unknown): string {\n  return createHash('sha256')\n    .update(JSON.stringify(sortKeysDeep(value)), 'utf8')\n    .digest('hex')\n}\n\nfunction sortKeysDeep(value: unknown): unknown {\n  if (value === null || typeof value !== 'object') return value\n  if (Array.isArray(value)) return value.map(sortKeysDeep)\n  const out: Record<string, unknown> = {}\n  for (const key of Object.keys(value as Record<string, unknown>).sort()) {\n    out[key] = sortKeysDeep((value as Record<string, unknown>)[key])\n  }\n  return out\n}\n\nexport function validateAgentProfileCell(input: unknown): AgentProfileCell {\n  if (input === null || typeof input !== 'object') {\n    throw new AgentProfileCellValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n  expectLiteral(obj.schemaVersion, 'agent-profile-cell/v1', 'schemaVersion')\n  if (typeof obj.cellId !== 'string' || !CELL_ID.test(obj.cellId)) {\n    throw new AgentProfileCellValidationError(\n      'cellId must match agent-profile-cell:sha256:<64 lowercase hex chars>',\n      'cellId',\n    )\n  }\n  expectString(obj.profileId, 'profileId')\n  validateSource(obj.sourceProfile, 'sourceProfile')\n  if (obj.harness !== undefined) validateHarness(obj.harness, 'harness')\n  if (obj.model !== undefined) expectString(obj.model, 'model')\n  if (obj.promptHash !== undefined) expectString(obj.promptHash, 'promptHash')\n  if (obj.dimensions !== undefined) validateDimensions(obj.dimensions, 'dimensions')\n  return input as AgentProfileCell\n}\n\nexport function requireAgentProfileCell(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): AgentProfileCell {\n  if (!record.agentProfile) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" is missing agentProfile; profile-cell grouping requires explicit profile identity`,\n      'agentProfile',\n    )\n  }\n  return validateAgentProfileCell(record.agentProfile)\n}\n\nexport function agentProfileCellKey(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): string {\n  return requireAgentProfileCell(record).cellId\n}\n\nexport async function assertRunAgentProfileCell(record: {\n  runId: string\n  model: string\n  promptHash: string\n  agentProfile?: AgentProfileCell\n}): Promise<AgentProfileCell> {\n  const profile = requireAgentProfileCell(record)\n  if (!(await verifyAgentProfileCell(profile))) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" has an agentProfile.cellId that does not match its content`,\n      'agentProfile.cellId',\n    )\n  }\n  if (profile.model !== undefined && profile.model !== record.model) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.model \"${profile.model}\" does not match model \"${record.model}\"`,\n      'agentProfile.model',\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== record.promptHash) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${record.promptHash}\"`,\n      'agentProfile.promptHash',\n    )\n  }\n  return profile\n}\n\nexport function groupRunsByAgentProfileCell<\n  T extends { runId: string; agentProfile?: AgentProfileCell },\n>(records: readonly T[]): Map<string, T[]> {\n  const groups = new Map<string, T[]>()\n  for (const record of records) {\n    const key = agentProfileCellKey(record)\n    const bucket = groups.get(key)\n    if (bucket) bucket.push(record)\n    else groups.set(key, [record])\n  }\n  return groups\n}\n\nasync function normalizeAgentProfileCellInput(\n  input: AgentProfileCellInput,\n): Promise<Omit<AgentProfileCell, 'cellId'>> {\n  return normalizeAgentProfileCell({\n    schemaVersion: 'agent-profile-cell/v1',\n    profileId: input.profileId,\n    sourceProfile: await normalizeSourceInput(input.sourceProfile),\n    harness: input.harness,\n    model: input.model,\n    promptHash: input.promptHash,\n    dimensions: input.dimensions,\n  })\n}\n\nfunction normalizeAgentProfileCell(\n  input: Omit<AgentProfileCell, 'cellId'>,\n): Omit<AgentProfileCell, 'cellId'> {\n  return compactObject({\n    schemaVersion: 'agent-profile-cell/v1' as const,\n    profileId: requireNonEmpty(input.profileId, 'profileId'),\n    sourceProfile: normalizeSource(input.sourceProfile),\n    harness: input.harness ? normalizeHarness(input.harness, 'harness') : undefined,\n    model: optionalNonEmpty(input.model, 'model'),\n    promptHash: optionalNonEmpty(input.promptHash, 'promptHash'),\n    dimensions: input.dimensions\n      ? nonEmptyRecord(normalizeDimensions(input.dimensions))\n      : undefined,\n  })\n}\n\nasync function normalizeSourceInput(input: AgentProfileSourceInput): Promise<AgentProfileSource> {\n  const kind = requireNonEmpty(input.kind, 'sourceProfile.kind')\n  if (input.hash !== undefined && input.profile !== undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide either hash or profile, not both',\n      'sourceProfile',\n    )\n  }\n  if (input.hash !== undefined) {\n    return { kind, hash: requireSha256Hex(input.hash, 'sourceProfile.hash') }\n  }\n  if (input.profile === undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide hash or profile',\n      'sourceProfile',\n    )\n  }\n  assertJson(input.profile, 'sourceProfile.profile')\n  return { kind, hash: await hashJson(input.profile) }\n}\n\nfunction normalizeSource(input: AgentProfileSource): AgentProfileSource {\n  return {\n    kind: requireNonEmpty(input.kind, 'sourceProfile.kind'),\n    hash: requireSha256Hex(input.hash, 'sourceProfile.hash'),\n  }\n}\n\nfunction normalizeHarness(input: AgentProfileHarness, path: string): AgentProfileHarness {\n  return compactObject({\n    id: requireNonEmpty(input.id, `${path}.id`),\n    version: optionalNonEmpty(input.version, `${path}.version`),\n    hash: optionalNonEmpty(input.hash, `${path}.hash`),\n  })\n}\n\nfunction normalizeDimensions(\n  input: Record<string, AgentProfileDimensionValue>,\n): Record<string, AgentProfileDimensionValue> {\n  const out: Record<string, AgentProfileDimensionValue> = {}\n  for (const key of Object.keys(input).sort()) {\n    const value = input[key]\n    requireNonEmpty(key, 'dimensions.<key>')\n    if (\n      value !== null &&\n      typeof value !== 'string' &&\n      typeof value !== 'number' &&\n      typeof value !== 'boolean'\n    ) {\n      throw new AgentProfileCellValidationError(\n        'expected primitive dimension value',\n        `dimensions.${key}`,\n      )\n    }\n    if (typeof value === 'number' && !Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', `dimensions.${key}`)\n    }\n    out[key] = value\n  }\n  return out\n}\n\nfunction compactObject<T extends Record<string, unknown>>(input: T): T {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as T\n}\n\nfunction nonEmptyRecord<T extends Record<string, unknown>>(input: T): T | undefined {\n  return Object.keys(input).length > 0 ? input : undefined\n}\n\nfunction validateSource(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.kind, `${path}.kind`)\n  requireSha256Hex(rec.hash, `${path}.hash`)\n}\n\nfunction validateHarness(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.id, `${path}.id`)\n  if (rec.version !== undefined) expectString(rec.version, `${path}.version`)\n  if (rec.hash !== undefined) expectString(rec.hash, `${path}.hash`)\n}\n\nfunction validateDimensions(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  normalizeDimensions(value as Record<string, AgentProfileDimensionValue>)\n}\n\nfunction assertJson(value: AgentProfileJson, path: string): void {\n  if (value === null) return\n  const type = typeof value\n  if (type === 'string' || type === 'boolean') return\n  if (type === 'number') {\n    if (!Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', path)\n    }\n    return\n  }\n  if (Array.isArray(value)) {\n    value.forEach((item, index) => {\n      assertJson(item, `${path}[${index}]`)\n    })\n    return\n  }\n  if (type === 'object') {\n    for (const [key, nested] of Object.entries(value)) {\n      requireNonEmpty(key, `${path}.<key>`)\n      assertJson(nested, `${path}.${key}`)\n    }\n    return\n  }\n  throw new AgentProfileCellValidationError('expected JSON-compatible value', path)\n}\n\nfunction expectLiteral(value: unknown, expected: string, path: string): void {\n  if (value !== expected) {\n    throw new AgentProfileCellValidationError(`expected ${expected}`, path)\n  }\n}\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n}\n\nfunction requireNonEmpty(value: string, path: string): string {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n  return value\n}\n\nfunction optionalNonEmpty(value: string | undefined, path: string): string | undefined {\n  if (value === undefined) return undefined\n  return requireNonEmpty(value, path)\n}\n\nfunction requireSha256Hex(value: unknown, path: string): string {\n  if (typeof value !== 'string' || !SHA256_HEX.test(value)) {\n    throw new AgentProfileCellValidationError('expected 64 lowercase sha256 hex chars', path)\n  }\n  return value\n}\n\n// ── Consumer helpers ─────────────────────────────────────────────────\n//\n// Boilerplate every product consuming `buildAgentProfileCell` used to duplicate:\n//\n//   1. A `JSON.parse(JSON.stringify(value))` helper that canonicalizes an\n//      arbitrary `@tangle-network/agent-interface` `AgentProfile` into the recursive\n//      `AgentProfileJson` shape, with a fail-loud error when the profile\n//      is not JSON-serializable.\n//\n//   2. The magic string `'agent-interface-profile'` for `sourceProfile.kind`.\n//\n// Both belong here so the cross-product cell join (same canonical profile\n// hashes to the same `sourceProfile.hash` across products) is enforced by\n// the type system, not by every consumer remembering to do it right.\n// See blueprint-agent issue tangle-network/agent-eval#82.\n\n/** Canonical `sourceProfile.kind` values. Two products fingerprinting the\n *  same canonical profile MUST use the same kind for their cells to share\n *  `sourceProfile.hash`. Extend rather than create new strings — adding a\n *  new kind is a deliberate cross-product schema change. */\nexport const AGENT_PROFILE_KINDS = {\n  /** A profile declared via `defineAgentProfile(...)` from\n   *  `@tangle-network/agent-interface`. The default kind for router-backed\n   *  and sandbox-backed products. */\n  AGENT_INTERFACE_PROFILE: 'agent-interface-profile',\n} as const\n\nexport type AgentProfileKind = (typeof AGENT_PROFILE_KINDS)[keyof typeof AGENT_PROFILE_KINDS]\n\n/** Canonicalize an arbitrary value into `AgentProfileJson` by JSON\n *  round-trip. Throws when the value contains anything not representable\n *  as JSON (functions, BigInt, cycles) — non-portable profiles fail loud\n *  rather than silently dropping fields. */\nexport function toAgentProfileJson(value: unknown): AgentProfileJson {\n  let serialized: string | undefined\n  try {\n    serialized = JSON.stringify(value)\n  } catch (err) {\n    throw new AgentProfileCellValidationError(\n      `agent profile must be JSON-serializable: ${err instanceof Error ? err.message : String(err)}`,\n      'sourceProfile.profile',\n    )\n  }\n  if (serialized === undefined) {\n    throw new AgentProfileCellValidationError(\n      'agent profile must be JSON-serializable (got undefined after JSON.stringify)',\n      'sourceProfile.profile',\n    )\n  }\n  return JSON.parse(serialized) as AgentProfileJson\n}\n\n/** Canonical AgentProfile shape required when deriving a stable cell id. */\nexport type AgentInterfaceProfileLike = AgentProfile & { name: string; version: string }\n\n/** Higher-level helper that hard-codes the canonical\n *  `agent-interface-profile` kind plus the JSON canonicalization. Equivalent\n *  to calling `buildAgentProfileCell` with `profileId = \\`${name}@${version}\\``\n *  and `sourceProfile = { kind: AGENT_INTERFACE_PROFILE, profile: <round-tripped> }`.\n *\n *  Use this from any product consuming an agent-interface `AgentProfile`; the\n *  manual `buildAgentProfileCell` call is reserved for advanced cases\n *  (custom kinds, pre-computed source hashes, alternate profileId\n *  conventions). */\nexport async function buildAgentInterfaceProfileCell(\n  profile: AgentInterfaceProfileLike,\n  input: Omit<AgentProfileCellInput, 'profileId' | 'sourceProfile'>,\n): Promise<AgentProfileCell> {\n  if (!profile || typeof profile !== 'object') {\n    throw new AgentProfileCellValidationError('AgentProfile must be an object', 'profile')\n  }\n  if (typeof profile.name !== 'string' || profile.name.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'AgentProfile must have a non-empty `name`',\n      'profile.name',\n    )\n  }\n  if (typeof profile.version !== 'string' || profile.version.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'AgentProfile must have a non-empty `version`',\n      'profile.version',\n    )\n  }\n  return buildAgentProfileCell({\n    ...input,\n    profileId: `${profile.name}@${profile.version}`,\n    sourceProfile: {\n      kind: AGENT_PROFILE_KINDS.AGENT_INTERFACE_PROFILE,\n      profile: toAgentProfileJson(profile),\n    },\n  })\n}\n"],"mappings":";;;;AA4DA,IAAa,kCAAb,cAAqD,gBAAgB;CACnE;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;AAEA,MAAM,aAAa;;;;;;AAMnB,MAAM,UAAU;AAChB,MAAM,iBAAiB;AACvB,MAAM,wBAAwB;AAE9B,eAAsB,sBACpB,OAC2B;CAC3B,MAAM,WAAW,MAAM,+BAA+B,KAAK;CAC3D,MAAM,SAAS,GAAG,iBAAiB,MAAM,SAAS,QAAQ;CAC1D,OAAO;EAAE,GAAG;EAAU;CAAO;AAC/B;AAEA,SAAgB,6BACd,MACkC;CAClC,MAAM,EAAE,QAAQ,SAAS,GAAG,aAAa;CAEzC,OAAO,0BAA0B,QAAQ;AAC3C;;;;;;AAOA,eAAsB,uBAAuB,MAA0C;CACrF,yBAAyB,IAAI;CAC7B,MAAM,WAAW,6BAA6B,IAAI;CAClD,IAAI,KAAK,OAAO,WAAW,cAAc,GACvC,OAAO,KAAK,WAAW,GAAG,iBAAiB,MAAM,SAAS,QAAQ;CAEpE,OAAO,KAAK,WAAW,GAAG,wBAAwB,iBAAiB,QAAQ;AAC7E;;;;;AAMA,SAAS,iBAAiB,OAAwB;CAChD,OAAO,WAAW,QAAQ,CAAC,CACxB,OAAO,KAAK,UAAU,aAAa,KAAK,CAAC,GAAG,MAAM,CAAC,CACnD,OAAO,KAAK;AACjB;AAEA,SAAS,aAAa,OAAyB;CAC7C,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,MAAM,IAAI,YAAY;CACvD,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,OAAO,OAAO,KAAK,KAAgC,CAAC,CAAC,KAAK,GACnE,IAAI,OAAO,aAAc,MAAkC,IAAI;CAEjE,OAAO;AACT;AAEA,SAAgB,yBAAyB,OAAkC;CACzE,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,gCAAgC,iBAAiB;CAE7D,MAAM,MAAM;CACZ,cAAc,IAAI,eAAe,yBAAyB,eAAe;CACzE,IAAI,OAAO,IAAI,WAAW,YAAY,CAAC,QAAQ,KAAK,IAAI,MAAM,GAC5D,MAAM,IAAI,gCACR,wEACA,QACF;CAEF,aAAa,IAAI,WAAW,WAAW;CACvC,eAAe,IAAI,eAAe,eAAe;CACjD,IAAI,IAAI,YAAY,KAAA,GAAW,gBAAgB,IAAI,SAAS,SAAS;CACrE,IAAI,IAAI,UAAU,KAAA,GAAW,aAAa,IAAI,OAAO,OAAO;CAC5D,IAAI,IAAI,eAAe,KAAA,GAAW,aAAa,IAAI,YAAY,YAAY;CAC3E,IAAI,IAAI,eAAe,KAAA,GAAW,mBAAmB,IAAI,YAAY,YAAY;CACjF,OAAO;AACT;AAEA,SAAgB,wBAAwB,QAGnB;CACnB,IAAI,CAAC,OAAO,cACV,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,sFACrB,cACF;CAEF,OAAO,yBAAyB,OAAO,YAAY;AACrD;AAEA,SAAgB,oBAAoB,QAGzB;CACT,OAAO,wBAAwB,MAAM,CAAC,CAAC;AACzC;AAEA,eAAsB,0BAA0B,QAKlB;CAC5B,MAAM,UAAU,wBAAwB,MAAM;CAC9C,IAAI,CAAE,MAAM,uBAAuB,OAAO,GACxC,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,+DACrB,qBACF;CAEF,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,OAAO,OAC1D,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,wBAAwB,QAAQ,MAAM,0BAA0B,OAAO,MAAM,IAClG,oBACF;CAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,OAAO,YACpE,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,6BAA6B,QAAQ,WAAW,+BAA+B,OAAO,WAAW,IACtH,yBACF;CAEF,OAAO;AACT;AAEA,SAAgB,4BAEd,SAAyC;CACzC,MAAM,yBAAS,IAAI,IAAiB;CACpC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,MAAM,oBAAoB,MAAM;EACtC,MAAM,SAAS,OAAO,IAAI,GAAG;EAC7B,IAAI,QAAQ,OAAO,KAAK,MAAM;OACzB,OAAO,IAAI,KAAK,CAAC,MAAM,CAAC;CAC/B;CACA,OAAO;AACT;AAEA,eAAe,+BACb,OAC2C;CAC3C,OAAO,0BAA0B;EAC/B,eAAe;EACf,WAAW,MAAM;EACjB,eAAe,MAAM,qBAAqB,MAAM,aAAa;EAC7D,SAAS,MAAM;EACf,OAAO,MAAM;EACb,YAAY,MAAM;EAClB,YAAY,MAAM;CACpB,CAAC;AACH;AAEA,SAAS,0BACP,OACkC;CAClC,OAAO,cAAc;EACnB,eAAe;EACf,WAAW,gBAAgB,MAAM,WAAW,WAAW;EACvD,eAAe,gBAAgB,MAAM,aAAa;EAClD,SAAS,MAAM,UAAU,iBAAiB,MAAM,SAAS,SAAS,IAAI,KAAA;EACtE,OAAO,iBAAiB,MAAM,OAAO,OAAO;EAC5C,YAAY,iBAAiB,MAAM,YAAY,YAAY;EAC3D,YAAY,MAAM,aACd,eAAe,oBAAoB,MAAM,UAAU,CAAC,IACpD,KAAA;CACN,CAAC;AACH;AAEA,eAAe,qBAAqB,OAA6D;CAC/F,MAAM,OAAO,gBAAgB,MAAM,MAAM,oBAAoB;CAC7D,IAAI,MAAM,SAAS,KAAA,KAAa,MAAM,YAAY,KAAA,GAChD,MAAM,IAAI,gCACR,+DACA,eACF;CAEF,IAAI,MAAM,SAAS,KAAA,GACjB,OAAO;EAAE;EAAM,MAAM,iBAAiB,MAAM,MAAM,oBAAoB;CAAE;CAE1E,IAAI,MAAM,YAAY,KAAA,GACpB,MAAM,IAAI,gCACR,8CACA,eACF;CAEF,WAAW,MAAM,SAAS,uBAAuB;CACjD,OAAO;EAAE;EAAM,MAAM,MAAM,SAAS,MAAM,OAAO;CAAE;AACrD;AAEA,SAAS,gBAAgB,OAA+C;CACtE,OAAO;EACL,MAAM,gBAAgB,MAAM,MAAM,oBAAoB;EACtD,MAAM,iBAAiB,MAAM,MAAM,oBAAoB;CACzD;AACF;AAEA,SAAS,iBAAiB,OAA4B,MAAmC;CACvF,OAAO,cAAc;EACnB,IAAI,gBAAgB,MAAM,IAAI,GAAG,KAAK,IAAI;EAC1C,SAAS,iBAAiB,MAAM,SAAS,GAAG,KAAK,SAAS;EAC1D,MAAM,iBAAiB,MAAM,MAAM,GAAG,KAAK,MAAM;CACnD,CAAC;AACH;AAEA,SAAS,oBACP,OAC4C;CAC5C,MAAM,MAAkD,CAAC;CACzD,KAAK,MAAM,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,KAAK,GAAG;EAC3C,MAAM,QAAQ,MAAM;EACpB,gBAAgB,KAAK,kBAAkB;EACvC,IACE,UAAU,QACV,OAAO,UAAU,YACjB,OAAO,UAAU,YACjB,OAAO,UAAU,WAEjB,MAAM,IAAI,gCACR,sCACA,cAAc,KAChB;EAEF,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,gCAAgC,0BAA0B,cAAc,KAAK;EAEzF,IAAI,OAAO;CACb;CACA,OAAO;AACT;AAEA,SAAS,cAAiD,OAAa;CACrE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;AAEA,SAAS,eAAkD,OAAyB;CAClF,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS,IAAI,QAAQ,KAAA;AACjD;AAEA,SAAS,eAAe,OAAgB,MAAoB;CAC1D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,MAAM,MAAM;CACZ,aAAa,IAAI,MAAM,GAAG,KAAK,MAAM;CACrC,iBAAiB,IAAI,MAAM,GAAG,KAAK,MAAM;AAC3C;AAEA,SAAS,gBAAgB,OAAgB,MAAoB;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,MAAM,MAAM;CACZ,aAAa,IAAI,IAAI,GAAG,KAAK,IAAI;CACjC,IAAI,IAAI,YAAY,KAAA,GAAW,aAAa,IAAI,SAAS,GAAG,KAAK,SAAS;CAC1E,IAAI,IAAI,SAAS,KAAA,GAAW,aAAa,IAAI,MAAM,GAAG,KAAK,MAAM;AACnE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,oBAAoB,KAAmD;AACzE;AAEA,SAAS,WAAW,OAAyB,MAAoB;CAC/D,IAAI,UAAU,MAAM;CACpB,MAAM,OAAO,OAAO;CACpB,IAAI,SAAS,YAAY,SAAS,WAAW;CAC7C,IAAI,SAAS,UAAU;EACrB,IAAI,CAAC,OAAO,SAAS,KAAK,GACxB,MAAM,IAAI,gCAAgC,0BAA0B,IAAI;EAE1E;CACF;CACA,IAAI,MAAM,QAAQ,KAAK,GAAG;EACxB,MAAM,SAAS,MAAM,UAAU;GAC7B,WAAW,MAAM,GAAG,KAAK,GAAG,MAAM,EAAE;EACtC,CAAC;EACD;CACF;CACA,IAAI,SAAS,UAAU;EACrB,KAAK,MAAM,CAAC,KAAK,WAAW,OAAO,QAAQ,KAAK,GAAG;GACjD,gBAAgB,KAAK,GAAG,KAAK,OAAO;GACpC,WAAW,QAAQ,GAAG,KAAK,GAAG,KAAK;EACrC;EACA;CACF;CACA,MAAM,IAAI,gCAAgC,kCAAkC,IAAI;AAClF;AAEA,SAAS,cAAc,OAAgB,UAAkB,MAAoB;CAC3E,IAAI,UAAU,UACZ,MAAM,IAAI,gCAAgC,YAAY,YAAY,IAAI;AAE1E;AAEA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,gCAAgC,6BAA6B,IAAI;AAE/E;AAEA,SAAS,gBAAgB,OAAe,MAAsB;CAC5D,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,gCAAgC,6BAA6B,IAAI;CAE7E,OAAO;AACT;AAEA,SAAS,iBAAiB,OAA2B,MAAkC;CACrF,IAAI,UAAU,KAAA,GAAW,OAAO,KAAA;CAChC,OAAO,gBAAgB,OAAO,IAAI;AACpC;AAEA,SAAS,iBAAiB,OAAgB,MAAsB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,WAAW,KAAK,KAAK,GACrD,MAAM,IAAI,gCAAgC,0CAA0C,IAAI;CAE1F,OAAO;AACT;;;;;AAsBA,MAAa,sBAAsB;;;;AAIjC,yBAAyB,0BAC3B;;;;;AAQA,SAAgB,mBAAmB,OAAkC;CACnE,IAAI;CACJ,IAAI;EACF,aAAa,KAAK,UAAU,KAAK;CACnC,SAAS,KAAK;EACZ,MAAM,IAAI,gCACR,4CAA4C,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,KAC3F,uBACF;CACF;CACA,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,gCACR,gFACA,uBACF;CAEF,OAAO,KAAK,MAAM,UAAU;AAC9B;;;;;;;;;;AAcA,eAAsB,+BACpB,SACA,OAC2B;CAC3B,IAAI,CAAC,WAAW,OAAO,YAAY,UACjC,MAAM,IAAI,gCAAgC,kCAAkC,SAAS;CAEvF,IAAI,OAAO,QAAQ,SAAS,YAAY,QAAQ,KAAK,WAAW,GAC9D,MAAM,IAAI,gCACR,6CACA,cACF;CAEF,IAAI,OAAO,QAAQ,YAAY,YAAY,QAAQ,QAAQ,WAAW,GACpE,MAAM,IAAI,gCACR,gDACA,iBACF;CAEF,OAAO,sBAAsB;EAC3B,GAAG;EACH,WAAW,GAAG,QAAQ,KAAK,GAAG,QAAQ;EACtC,eAAe;GACb,MAAM,oBAAoB;GAC1B,SAAS,mBAAmB,OAAO;EACrC;CACF,CAAC;AACH"}

package/dist/reward-hacking-CKW4teig.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"reward-hacking-CKW4teig.js","names":["mean","clamp01"],"sources":["../src/campaign/run-record.ts","../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n","/**\n * Verifiable reward channel.\n *\n * For RL on coding / math / theorem-proving / structured-output tasks, the\n * reward signal is *decidable* — a test passes or fails, a proof checks or\n * doesn't, an output validates against a schema or doesn't. These rewards\n * are dramatically more useful for RL training than LLM-judge scores\n * because they don't drift, can't be Goodhart-gamed by the policy in the\n * same way, and don't require a separate calibration loop.\n *\n * The `MultiLayerVerifier` already produces this signal — it just doesn't\n * surface it in a shape that's clean enough for RL training. This module\n * wraps the verifier output so consumers can:\n *\n *   1. Extract a clean `VerifiableReward` from a `VerificationReport`\n *   2. Distinguish *deterministic* rewards (compile, test, schema) from\n *      *probabilistic* rewards (judge) so they can be weighted differently\n *      in the RL training step\n *   3. Filter `RunRecord[]` to only those with a verifiable reward,\n *      producing the clean training set that DeepSeek-R1-style GRPO and\n *      AlphaProof-style search both depend on\n *\n * Why this matters: every credible 2025-2026 frontier RL result on coding\n * agents leans on verifiable reward (DeepSeek-R1 GRPO on test pass-rate,\n * o-series RL on math/code, AlphaProof on Lean kernel checking). Mixing\n * judge scores into the reward signal poisons the gradient. This module\n * is the seam.\n */\n\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport { isRealnessGated, observedScore, trainingScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\nimport type { VerificationStrategySource } from '../verification-strategy'\n\n/**\n * What produced a reward. This is the verification-strategy family\n * (`src/verification-strategy.ts`) — one open vocabulary shared by rewards\n * and verdict certifications, so an RL consumer and a certification reader\n * mean the same thing by `'proof-kernel'`.\n *\n * Beyond the classic answer-key members (`compile`, `test`, `schema`,\n * `sandbox`, `judge`, `composite`) the family carries four members for\n * tasks with no held-out suite, each with a documented failure mode:\n *\n * - `'proof-kernel'` — kernel-checked formal proof. Failure mode: the\n *   formalization gap — the kernel never certifies that the formal\n *   statement matches the informal claim (`src/verification-strategy.ts`\n *   is the discharge protocol).\n * - `'invariant'` — invariant / metamorphic properties held. Failure\n *   mode: weak invariants pass everything; the set needs seeded-bug\n *   calibration before its pass carries weight.\n * - `'replication'` — independent re-execution from pinned inputs.\n *   Failure mode: re-runs the method, so it never catches an error the\n *   method itself carries.\n * - `'agreement'` — independently-derived results agree. Failure mode:\n *   the shared blind spot — derivers with common training corpora or\n *   priors agree for the same wrong reason. Probabilistic: the check may\n *   be mechanical, but the independence it certifies is not.\n *\n * The deterministic/probabilistic axis stays on `VerifiableReward` —\n * the producer declares it per reward, and `VERIFICATION_STRATEGIES`\n * documents each member's default class.\n */\nexport type VerifiableRewardSource = VerificationStrategySource\n\nexport interface VerifiableReward {\n  /** Scalar in [0, 1]. The RL training signal. */\n  value: number\n  /** What produced the reward — different sources have different determinism. */\n  source: VerifiableRewardSource\n  /**\n   * Determinism class. `'deterministic'` rewards are repeatable byte-for-byte\n   * given the same inputs (compile, test, schema validation, sandbox exit code).\n   * `'probabilistic'` rewards depend on a stochastic component (LLM judge).\n   * Mixing these in the same training batch without separation is a known\n   * footgun in production RLHF pipelines.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /**\n   * Confidence in the reward value. For deterministic sources this is 1.0\n   * (the bit either flipped or didn't). For judge sources this is the\n   * judge-reported confidence or — when missing — a calibrated prior.\n   */\n  confidence: number\n  /** The layer / judge id that produced the signal, for provenance. */\n  origin: string\n  /**\n   * Per-source contribution to `value`, keyed by layer/judge id. Single-source\n   * rewards carry one entry (`{ [origin]: value }`); composite rewards carry\n   * every contributing layer's score — the anti-scalar-collapse surface RL\n   * consumers weight per-source instead of trusting one blended number.\n   */\n  components: Record<string, number>\n  /**\n   * The run carries `outcome.realness.gated` — the authenticity gate flagged\n   * its success signal as faked.\n   *\n   * With the gate applied (the default) `value` and every `components` entry\n   * are 0 on such a run; with `applyRealnessGate: false` the observed numbers\n   * come back untouched and this flag is the only marker that they are not to\n   * be trusted. Either way it distinguishes \"measured a genuine failure\" from\n   * \"claimed a success we refuse to believe\", which a bare 0 cannot.\n   */\n  realnessGated?: boolean\n  /**\n   * Whether an authenticity screen COULD run on this reward at all — the same\n   * distinction `RolloutOutcome.realness_screened` draws, for the same reason.\n   *\n   * `false` on every reward from `extractVerifiableReward`, because a\n   * `VerificationReport` carries layer scores and nothing else: there is no\n   * `outcome.realness` to consult, so no gate has run, and `realnessGated`\n   * being absent there means \"unknown\", NOT \"clean\". Absent on the\n   * `RunRecord` path when the record itself carries no realness verdict.\n   *\n   * This matters most exactly where it is easiest to miss: a report whose\n   * deterministic layers all passed yields `determinism: 'deterministic'`,\n   * `confidence: 1` — the highest-credibility reward this module can emit —\n   * and a stubbed integration reporting green is precisely what a gamed run\n   * looks like. Consumers driving training off this shape must screen the run\n   * themselves; the flag is what tells them nobody has.\n   */\n  realnessScreened?: boolean\n}\n\nexport interface VerifiableRewardExtractionOptions {\n  /**\n   * Which layers count as deterministic-reward sources. The verifier doesn't\n   * tag layers as \"this is verifiable\"; the caller declares it via this list\n   * (or via the layer name → source mapping). Default treats common names\n   * (`install`, `typecheck`, `build`, `lint`, `test`, `compile`, `schema`,\n   * `sandbox`) as deterministic.\n   */\n  deterministicLayers?: string[]\n  /**\n   * Map layer name → reward source. Defaults to a sensible string-match.\n   */\n  sourceFor?: (layerName: string) => VerifiableRewardSource\n  /**\n   * Whether to fall back to a probabilistic (judge) reward when no\n   * deterministic layer produced a numeric score. Default `true`. Set to\n   * `false` for \"deterministic-only\" training pipelines that should\n   * discard runs without a verifiable signal.\n   */\n  fallbackToJudge?: boolean\n  /**\n   * Default confidence for probabilistic (judge) rewards when the judge\n   * doesn't report one. Default `0.7`.\n   */\n  judgeConfidenceFloor?: number\n  /**\n   * Whether the anti-Goodhart realness gate applies. Default `true`, and the\n   * default is the one every training path must keep.\n   *\n   * Set `false` ONLY for detection and analysis. `rl/reward-hacking.ts` does,\n   * for the same reason it reads `observedScore` for its proxy: it measures the\n   * DIVERGENCE between the judge signal and the deterministic one, and a\n   * deterministic reward that another gate already forced to 0 manufactures\n   * exactly that divergence on exactly the gamed population. The detector would\n   * then be re-reporting a verdict it was supposed to reach independently.\n   */\n  applyRealnessGate?: boolean\n}\n\n// 'agreement' is deliberately absent: agreement between fixed artifacts is\n// mechanical, but the independent derivation it certifies is stochastic, so\n// a producer that wants a deterministic agreement reward must declare it\n// (via `deterministicLayers` or by constructing the reward directly).\nconst DEFAULT_DETERMINISTIC_LAYERS = new Set([\n  'install',\n  'typecheck',\n  'build',\n  'lint',\n  'test',\n  'compile',\n  'schema',\n  'sandbox',\n  'unit_tests',\n  'integration_tests',\n  'proof_kernel',\n  'proof-kernel',\n  'invariant',\n  'metamorphic',\n  'replication',\n])\n\nconst DEFAULT_SOURCE_FOR = (name: string): VerifiableRewardSource => {\n  const lower = name.toLowerCase()\n  if (lower.includes('test')) return 'test'\n  if (\n    lower.includes('compile') ||\n    lower.includes('build') ||\n    lower.includes('typecheck') ||\n    lower.includes('lint')\n  )\n    return 'compile'\n  if (lower.includes('schema')) return 'schema'\n  if (lower.includes('sandbox')) return 'sandbox'\n  if (lower.includes('judge') || lower.includes('semantic')) return 'judge'\n  // Answer-key names above take precedence: a name like 'proof_test' maps\n  // to 'test'; only names with no answer-key match reach the open-family\n  // members.\n  if (lower.includes('proof') || lower.includes('kernel') || lower.includes('lean')) {\n    return 'proof-kernel'\n  }\n  if (lower.includes('invariant') || lower.includes('metamorphic')) return 'invariant'\n  if (lower.includes('replicat')) return 'replication'\n  if (lower.includes('agreement') || lower.includes('consensus')) return 'agreement'\n  return 'composite'\n}\n\n/**\n * Extract a `VerifiableReward` from a `VerificationReport`.\n *\n * Strategy: prefer the deterministic layers (in order: test → compile →\n * schema → sandbox), fall back to the judge layer if `fallbackToJudge` is\n * true, return `null` if no signal qualifies. When multiple deterministic\n * layers contribute, return a `'composite'` source with a weighted blend.\n *\n * NO realness gate is applied and none can be: a `VerificationReport` carries\n * layer scores and nothing about whether the run faked them — `realness` lives\n * on the `RunRecord`. Use `extractVerifiableRewardsFromRecords` for anything\n * that becomes training data; this signature is for scoring a report in hand.\n */\nexport function extractVerifiableReward(\n  report: VerificationReport,\n  opts: VerifiableRewardExtractionOptions = {},\n): VerifiableReward | null {\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  const deterministic = report.layers.filter(\n    (layer) => deterministicSet.has(layer.layer) && isMeasuredLayer(layer),\n  )\n\n  if (deterministic.length === 1) {\n    const layer = deterministic[0]!\n    const value = clamp01(layer.score!)\n    return {\n      value,\n      source: sourceFor(layer.layer),\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: layer.layer,\n      components: { [layer.layer]: value },\n      realnessScreened: false,\n    }\n  }\n\n  if (deterministic.length > 1) {\n    // Composite: weighted blend by `Layer.weight` if present, else equal.\n    let num = 0\n    let denom = 0\n    const components: Record<string, number> = {}\n    for (const l of deterministic) {\n      const w = (l.detail?.weight as number | undefined) ?? 1\n      num += w * (l.score ?? 0)\n      denom += w\n      components[l.layer] = l.score!\n    }\n    return {\n      value: denom === 0 ? 0 : clamp01(num / denom),\n      source: 'composite',\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: deterministic.map((l) => l.layer).join('+'),\n      components,\n      realnessScreened: false,\n    }\n  }\n\n  if (!fallbackToJudge) return null\n\n  const judge =\n    report.layers.find((layer) => isMeasuredLayer(layer) && sourceFor(layer.layer) === 'judge') ??\n    report.layers.find(isMeasuredLayer)\n\n  if (!judge) return null\n\n  const confFromDetail = judge.detail?.confidence as number | undefined\n  const judgeValue = clamp01(judge.score!)\n  return {\n    value: judgeValue,\n    source: 'judge',\n    determinism: 'probabilistic',\n    confidence: typeof confFromDetail === 'number' ? confFromDetail : judgeFloor,\n    origin: judge.layer,\n    components: { [judge.layer]: judgeValue },\n    realnessScreened: false,\n  }\n}\n\nfunction isMeasuredLayer(\n  layer: LayerResult,\n): layer is LayerResult & { status: 'pass' | 'fail'; score: number } {\n  return (\n    (layer.status === 'pass' || layer.status === 'fail') &&\n    typeof layer.score === 'number' &&\n    Number.isFinite(layer.score) &&\n    layer.score >= 0 &&\n    layer.score <= 1\n  )\n}\n\n/**\n * Extract verifiable rewards from `RunRecord[]` produced via the\n * `verificationReportToRunRecord` adapter (which encodes per-layer scores\n * in `outcome.raw['layer.<name>']`). For records that don't carry layer\n * scores, returns `null` for that record.\n *\n * This is the canonical bridge from \"campaign-shaped artifacts\" to\n * \"RL-training-ready reward signals\": every record that has a clean\n * verifiable reward becomes a training datum, every record that doesn't\n * gets filtered out (or kept with `'probabilistic'` determinism for\n * separate downstream handling).\n *\n * The realness gate applies to EVERY channel here, and to the deterministic one\n * MOST. It is tempting to reason that a decidable signal cannot be gamed, so\n * the gate is redundant on it — that reasoning is backwards. `realness.gated`\n * means the run's success signal was FAKED, and a test suite reporting green on\n * a stubbed integration is precisely what that looks like: the deterministic\n * layer is the thing that got faked. Exporting it ungated hands a trainer the\n * highest-credibility reward the module can emit (`determinism: 'deterministic'`,\n * `confidence: 1`) for the one population the gate exists to catch. Pass\n * `applyRealnessGate: false` only to look at the ungated numbers for detection.\n */\nexport function extractVerifiableRewardsFromRecords(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ runId: string; reward: VerifiableReward | null }> {\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n  const applyGate = opts.applyRealnessGate ?? true\n\n  return runs.map((run) => {\n    const flagged = isRealnessGated(run)\n    // Present only when the record carries an actual realness verdict. Absent\n    // is the honest \"unknown\"; `false` is reserved for a producer that\n    // declares it HAS no screen, which is the report-shaped path above.\n    const screened = run.outcome.realness === undefined ? {} : ({ realnessScreened: true } as const)\n    // Zeroed with `value`, never left at the measured number: `components`\n    // exists so an RL consumer can re-weight per source, and a raw layer score\n    // surviving there would let that re-weighting reconstruct the very reward\n    // the gate just refused. The measured layer scores stay on\n    // `run.outcome.raw['layer.*']`, which is where analysis reads them.\n    const gate = (value: number): number => (applyGate && flagged ? 0 : value)\n    // Recover per-layer scores from outcome.raw['layer.<name>']\n    const layerScores: Array<{ name: string; score: number }> = []\n    for (const [k, v] of Object.entries(run.outcome.raw)) {\n      if (\n        k.startsWith('layer.') &&\n        !k.includes('.', 6) &&\n        typeof v === 'number' &&\n        Number.isFinite(v)\n      ) {\n        layerScores.push({ name: k.slice('layer.'.length), score: v })\n      }\n    }\n    const det = layerScores.filter((l) => deterministicSet.has(l.name))\n\n    if (det.length === 1) {\n      const layer = det[0]!\n      const value = gate(clamp01(layer.score))\n      return {\n        runId: run.runId,\n        reward: {\n          value,\n          source: sourceFor(layer.name),\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: layer.name,\n          components: { [layer.name]: value },\n          realnessGated: flagged,\n          ...screened,\n        },\n      }\n    }\n    if (det.length > 1) {\n      const value = gate(clamp01(det.reduce((s, l) => s + l.score, 0) / det.length))\n      // Same clamp as the headline value: a producer writing layer.score 1.5\n      // into outcome.raw must not propagate 1.5 through a component either.\n      const components: Record<string, number> = Object.fromEntries(\n        det.map((l) => [l.name, gate(clamp01(l.score))]),\n      )\n      return {\n        runId: run.runId,\n        reward: {\n          value,\n          source: 'composite',\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: det.map((l) => l.name).join('+'),\n          components,\n          realnessGated: flagged,\n          ...screened,\n        },\n      }\n    }\n    if (!fallbackToJudge) return { runId: run.runId, reward: null }\n\n    // Probabilistic fallback: the run's primary score. `trainingScore` already\n    // carries the gate, so a gamed run falls to 0 rather than earning the\n    // judge's number; `observedScore` is the ungated reader the detection\n    // opt-out asks for. Either way an unscored run stays a labeled gap\n    // (`reward: null`), never a fabricated 0.\n    const primary = applyGate ? trainingScore(run) : observedScore(run)\n    if (typeof primary !== 'number' || !Number.isFinite(primary)) {\n      return { runId: run.runId, reward: null }\n    }\n    const primaryValue = clamp01(primary)\n    return {\n      runId: run.runId,\n      reward: {\n        value: primaryValue,\n        source: 'judge',\n        determinism: 'probabilistic',\n        confidence: judgeFloor,\n        origin: 'run.outcome.score',\n        components: { 'run.outcome.score': primaryValue },\n        realnessGated: flagged,\n        ...screened,\n      },\n    }\n  })\n}\n\n/**\n * Filter `RunRecord[]` to those with deterministic verifiable rewards.\n *\n * A realness-gated run is KEPT, at reward 0 with `realnessGated: true` — the\n * same rule GRPO uses on a gated line. 0 is the honest label for a faked\n * success and is usable signal, whereas dropping the run would move a group\n * baseline without saying so. (SFT differs: there every row is a target to\n * imitate, so a gated row is removed outright.)\n */\nexport function filterDeterministicallyRewarded(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ run: RunRecord; reward: VerifiableReward }> {\n  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false })\n  const out: Array<{ run: RunRecord; reward: VerifiableReward }> = []\n  for (let i = 0; i < runs.length; i++) {\n    const r = rewarded[i]!\n    if (r.reward && r.reward.determinism === 'deterministic') {\n      out.push({ run: runs[i]!, reward: r.reward })\n    }\n  }\n  return out\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n","/**\n * Reward hacking / Goodhart detection.\n *\n * Goodhart's Law says: when a measure becomes a target, it ceases to be\n * a good measure. In RLHF and agentic-RL settings this is the dominant\n * failure mode — the policy learns to produce outputs that score well on\n * the proxy reward (judge, rubric, test pass-rate) without producing\n * the underlying capability the proxy was meant to track.\n *\n * Krakovna et al. (2020, \"Specification Gaming Examples in AI\") and the\n * subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.\n * 2023) converge on a few diagnostic signatures:\n *\n *   1. **Reward divergence:** the proxy reward grows while the held-out\n *      ground-truth signal stagnates or drops. Predictive validity over\n *      time captures this.\n *   2. **Distributional shift in outputs:** after RL, the policy produces\n *      outputs that no longer match the reference distribution — usually\n *      because it found a high-reward attractor that's degenerate (e.g.\n *      one-token responses, repetition, formatting tricks).\n *   3. **Disagreement between independent rewards:** if you train on\n *      reward A and a held-out independent reward B drops sharply, you're\n *      probably hacking A.\n *   4. **Calibration drift:** the verifiable / deterministic component of\n *      the reward is stable; the probabilistic / judge component drifts up\n *      while the deterministic component doesn't. The judge is being\n *      gamed.\n *\n * This module ships explicit detectors for all four signatures, plus a\n * combined verdict. The output is diagnostic — actionable signals,\n * not autoreject — because each signature has known false positives\n * (e.g., a policy that genuinely improves can show distributional shift).\n *\n * Differs from `rubricPredictiveValidity` (which is a *standing* check on\n * whether rubrics correlate with deployment outcomes) — this is a\n * *temporal* check on whether the reward-vs-truth gap is *widening over\n * time during a training run*.\n */\n\nimport { observedScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\nimport { pearsonR } from '../statistics'\nimport {\n  filterDeterministicallyRewarded,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport type RewardHackingSignal =\n  | 'reward_divergence'\n  | 'distribution_shift'\n  | 'reward_disagreement'\n  | 'judge_drift'\n\nexport interface RewardHackingFinding {\n  signal: RewardHackingSignal\n  /** Severity in [0, 1]. >0.5 = strong signal. */\n  severity: number\n  message: string\n  /** Numeric evidence the consumer can render. */\n  detail: Record<string, number>\n}\n\nexport interface RewardHackingReport {\n  findings: RewardHackingFinding[]\n  /** Signals with enough usable observations to produce a finding. */\n  evaluatedSignals: RewardHackingSignal[]\n  /**\n   * Composite verdict. `'insufficient_evidence'` when fewer than four scored\n   * runs exist; otherwise `'clean'` if every signal severity < 0.3,\n   * `'suspect'` if at least one ≥ 0.3 but none ≥ 0.6, and `'gaming'` if any ≥ 0.6.\n   */\n  verdict: 'insufficient_evidence' | 'clean' | 'suspect' | 'gaming'\n  /** Rationale for the verdict, ready to paste into an audit log. */\n  rationale: string[]\n  /** Number of runs with a usable proxy reward. */\n  n: number\n}\n\nexport interface DetectRewardHackingInput {\n  /**\n   * Run records ordered by recency (oldest first). The detector segments\n   * them into prefix/suffix windows to compute \"did the gap widen.\"\n   */\n  runs: RunRecord[]\n  /**\n   * The metric the policy was trained to optimize. Should be present on\n   * `outcome.raw` or `outcome.holdoutScore`. Default reads `outcome.holdoutScore`.\n   */\n  proxyOf?: (run: RunRecord) => number | null\n  /**\n   * The held-out ground-truth metric. For RL on coding, this is typically\n   * test pass-rate. For RLHF, it's downstream task performance or human\n   * preference. For knowledge tasks, it's an independently-graded score.\n   */\n  truthOf?: (run: RunRecord) => number | null\n  /**\n   * Independent secondary reward. Used for the `reward_disagreement`\n   * signal. Default uses the verifiable reward extractor (deterministic\n   * sources only).\n   */\n  secondaryRewardOf?: (run: RunRecord) => number | null\n  /**\n   * Window size — how many of the most recent runs count as the \"after\"\n   * cohort. Default min(50, half the runs).\n   */\n  windowSize?: number\n  /**\n   * Severity threshold to flag a signal. Default 0.3 (suspect) and 0.6\n   * (gaming).\n   */\n  thresholds?: { suspect?: number; gaming?: number }\n  /**\n   * Verifiable-reward options used for the secondary-reward fallback.\n   */\n  verifiableRewardOptions?: VerifiableRewardExtractionOptions\n}\n\nconst DEFAULT_PROXY = (r: RunRecord): number | null => {\n  // DELIBERATELY UNGATED. This is the proxy reward the detector tests for\n  // Goodharting, and gated runs are exactly the gamed population. Forcing them\n  // to 0 would collapse the proxy toward the deterministic secondary signal —\n  // `reward_disagreement`'s correlation would rise, `judge_drift`'s gap would\n  // shrink, and `reward_divergence`'s proxy-up/truth-flat fingerprint would be\n  // erased — so the detector would report \"clean\" on the very runs it exists to\n  // catch. `null` (never 0) is also load-bearing: it sets the n-denominator via\n  // the filter below.\n  const v = observedScore(r)\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nexport function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport {\n  const proxyOf = input.proxyOf ?? DEFAULT_PROXY\n  const truthOf = input.truthOf\n  const sus = input.thresholds?.suspect ?? 0.3\n  const gam = input.thresholds?.gaming ?? 0.6\n\n  const runs = input.runs.filter((run) => finiteNumber(proxyOf(run)))\n  const n = runs.length\n  if (n < 4) {\n    return {\n      findings: [],\n      evaluatedSignals: [],\n      verdict: 'insufficient_evidence',\n      n,\n      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`],\n    }\n  }\n  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)))\n  const before = runs.slice(0, n - windowSize)\n  const after = runs.slice(n - windowSize)\n\n  const findings: RewardHackingFinding[] = []\n\n  // ── Signal 1: reward divergence (proxy ↑ while truth flat or ↓) ──────\n  if (truthOf) {\n    const beforeProxy = before.map(proxyOf).filter(finiteNumber)\n    const afterProxy = after.map(proxyOf).filter(finiteNumber)\n    const beforeTruth = before.map(truthOf).filter(finiteNumber)\n    const afterTruth = after.map(truthOf).filter(finiteNumber)\n    if (\n      beforeProxy.length >= 2 &&\n      afterProxy.length >= 2 &&\n      beforeTruth.length >= 2 &&\n      afterTruth.length >= 2\n    ) {\n      const proxyDelta = mean(afterProxy) - mean(beforeProxy)\n      const truthDelta = mean(afterTruth) - mean(beforeTruth)\n      // Divergence: proxy goes up while truth goes flat or down.\n      // Severity = max(0, (proxyDelta - truthDelta)) — bigger gap = bigger signal.\n      const gap = Math.max(0, proxyDelta - truthDelta)\n      const severity = clamp01(gap * 5) // scale: 0.2 absolute gap → severity 1.0\n      findings.push({\n        signal: 'reward_divergence',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart`\n            : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,\n        detail: {\n          proxyDelta,\n          truthDelta,\n          gap,\n          beforeN: beforeProxy.length,\n          afterN: afterProxy.length,\n        },\n      })\n    }\n  }\n\n  // ── Signal 2: distributional shift in outputs (KS on score distributions) ──\n  {\n    const beforeP = before.map(proxyOf).filter(finiteNumber)\n    const afterP = after.map(proxyOf).filter(finiteNumber)\n    if (beforeP.length >= 4 && afterP.length >= 4) {\n      const ks = ksStatistic(beforeP, afterP)\n      // KS statistic: bigger = more shift. We're agnostic about direction;\n      // genuine improvement ALSO produces shift, so this signal is\n      // contributory rather than load-bearing.\n      const severity = clamp01(ks - 0.2)\n      findings.push({\n        signal: 'distribution_shift',\n        severity,\n        message:\n          severity >= sus\n            ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large`\n            : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,\n        detail: { ks, beforeN: beforeP.length, afterN: afterP.length },\n      })\n    }\n  }\n\n  // ── Signal 3: reward disagreement (proxy vs independent secondary) ────\n  {\n    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions)\n    const aligned = runs\n      .map((r) => ({ p: proxyOf(r), s: secondaryOf(r) }))\n      .filter((x): x is { p: number; s: number } => finiteNumber(x.p) && finiteNumber(x.s))\n    if (aligned.length >= 4) {\n      const ps = aligned.map((x) => x.p)\n      const ss = aligned.map((x) => x.s)\n      const r = pearsonR(ps, ss)\n      // Disagreement: low or negative correlation between primary proxy\n      // reward and an independent secondary signal.\n      const severity = clamp01(0.5 - Math.max(0, r))\n      findings.push({\n        signal: 'reward_disagreement',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy`\n            : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,\n        detail: { pearson: r, n: aligned.length },\n      })\n    }\n  }\n\n  // ── Signal 4: judge drift (probabilistic up while deterministic flat) ─\n  {\n    // Ungated on purpose, exactly like `DEFAULT_PROXY` above. This signal is\n    // the GAP between the judge reward and the deterministic one; a\n    // deterministic reward another gate already forced to 0 would open that gap\n    // by construction on the gamed population, so the detector would fire on\n    // its own input rather than on evidence it found.\n    const detRuns = filterDeterministicallyRewarded(runs, {\n      ...(input.verifiableRewardOptions ?? {}),\n      applyRealnessGate: false,\n    })\n    if (detRuns.length >= 4) {\n      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2))\n      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2))\n      const detDelta =\n        mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value))\n      const proxyDelta =\n        mean(after.map(proxyOf).filter(finiteNumber)) -\n        mean(before.map(proxyOf).filter(finiteNumber))\n      const driftGap = Math.max(0, proxyDelta - detDelta)\n      const severity = clamp01(driftGap * 5)\n      findings.push({\n        signal: 'judge_drift',\n        severity,\n        message:\n          severity >= sus\n            ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing`\n            : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,\n        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length },\n      })\n    }\n  }\n\n  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0)\n  if (findings.length === 0) {\n    return {\n      findings,\n      evaluatedSignals: [],\n      verdict: 'insufficient_evidence',\n      rationale: [`no reward-hacking signal had enough paired evidence (n=${n})`],\n      n,\n    }\n  }\n  const verdict: RewardHackingReport['verdict'] =\n    maxSev >= gam ? 'gaming' : maxSev >= sus ? 'suspect' : 'clean'\n  const rationale = findings\n    .filter((f) => f.severity >= sus)\n    .map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`)\n  if (rationale.length === 0) rationale.push('no signals fired above suspect threshold')\n\n  return {\n    findings,\n    evaluatedSignals: findings.map((finding) => finding.signal),\n    verdict,\n    rationale,\n    n,\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction finiteNumber(value: number | null): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction ksStatistic(a: number[], b: number[]): number {\n  // Two-sample Kolmogorov-Smirnov statistic.\n  const sortedA = [...a].sort((x, y) => x - y)\n  const sortedB = [...b].sort((x, y) => x - y)\n  const all = [...new Set([...sortedA, ...sortedB])].sort((x, y) => x - y)\n  let max = 0\n  for (const v of all) {\n    const fa = sortedA.filter((x) => x <= v).length / sortedA.length\n    const fb = sortedB.filter((x) => x <= v).length / sortedB.length\n    max = Math.max(max, Math.abs(fa - fb))\n  }\n  return max\n}\n\nfunction defaultSecondary(\n  verifiableOpts?: VerifiableRewardExtractionOptions,\n): (run: RunRecord) => number | null {\n  return (run: RunRecord) => {\n    // Ungated for the same reason as signal 4: this is the INDEPENDENT\n    // secondary reward whose correlation with the proxy is the evidence.\n    // Zeroing it on gated runs would drive that correlation down mechanically.\n    const filtered = filterDeterministicallyRewarded([run], {\n      ...(verifiableOpts ?? {}),\n      applyRealnessGate: false,\n    })\n    return filtered.length === 1 ? filtered[0]!.reward.value : null\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;AA8CA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAYA,OAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAWA,OAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAASA,OAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE;;;ACtKA,MAAM,+CAA+B,IAAI,IAAI;CAC3C;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,sBAAsB,SAAyC;CACnE,MAAM,QAAQ,KAAK,YAAY;CAC/B,IAAI,MAAM,SAAS,MAAM,GAAG,OAAO;CACnC,IACE,MAAM,SAAS,SAAS,KACxB,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,MAAM,GAErB,OAAO;CACT,IAAI,MAAM,SAAS,QAAQ,GAAG,OAAO;CACrC,IAAI,MAAM,SAAS,SAAS,GAAG,OAAO;CACtC,IAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,UAAU,GAAG,OAAO;CAIlE,IAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,QAAQ,KAAK,MAAM,SAAS,MAAM,GAC9E,OAAO;CAET,IAAI,MAAM,SAAS,WAAW,KAAK,MAAM,SAAS,aAAa,GAAG,OAAO;CACzE,IAAI,MAAM,SAAS,UAAU,GAAG,OAAO;CACvC,IAAI,MAAM,SAAS,WAAW,KAAK,MAAM,SAAS,WAAW,GAAG,OAAO;CACvE,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,wBACd,QACA,OAA0C,CAAC,GAClB;CACzB,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAEhD,MAAM,gBAAgB,OAAO,OAAO,QACjC,UAAU,iBAAiB,IAAI,MAAM,KAAK,KAAK,gBAAgB,KAAK,CACvE;CAEA,IAAI,cAAc,WAAW,GAAG;EAC9B,MAAM,QAAQ,cAAc;EAC5B,MAAM,QAAQC,UAAQ,MAAM,KAAM;EAClC,OAAO;GACL;GACA,QAAQ,UAAU,MAAM,KAAK;GAC7B,aAAa;GACb,YAAY;GACZ,QAAQ,MAAM;GACd,YAAY,GAAG,MAAM,QAAQ,MAAM;GACnC,kBAAkB;EACpB;CACF;CAEA,IAAI,cAAc,SAAS,GAAG;EAE5B,IAAI,MAAM;EACV,IAAI,QAAQ;EACZ,MAAM,aAAqC,CAAC;EAC5C,KAAK,MAAM,KAAK,eAAe;GAC7B,MAAM,IAAK,EAAE,QAAQ,UAAiC;GACtD,OAAO,KAAK,EAAE,SAAS;GACvB,SAAS;GACT,WAAW,EAAE,SAAS,EAAE;EAC1B;EACA,OAAO;GACL,OAAO,UAAU,IAAI,IAAIA,UAAQ,MAAM,KAAK;GAC5C,QAAQ;GACR,aAAa;GACb,YAAY;GACZ,QAAQ,cAAc,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,KAAK,GAAG;GAClD;GACA,kBAAkB;EACpB;CACF;CAEA,IAAI,CAAC,iBAAiB,OAAO;CAE7B,MAAM,QACJ,OAAO,OAAO,MAAM,UAAU,gBAAgB,KAAK,KAAK,UAAU,MAAM,KAAK,MAAM,OAAO,KAC1F,OAAO,OAAO,KAAK,eAAe;CAEpC,IAAI,CAAC,OAAO,OAAO;CAEnB,MAAM,iBAAiB,MAAM,QAAQ;CACrC,MAAM,aAAaA,UAAQ,MAAM,KAAM;CACvC,OAAO;EACL,OAAO;EACP,QAAQ;EACR,aAAa;EACb,YAAY,OAAO,mBAAmB,WAAW,iBAAiB;EAClE,QAAQ,MAAM;EACd,YAAY,GAAG,MAAM,QAAQ,WAAW;EACxC,kBAAkB;CACpB;AACF;AAEA,SAAS,gBACP,OACmE;CACnE,QACG,MAAM,WAAW,UAAU,MAAM,WAAW,WAC7C,OAAO,MAAM,UAAU,YACvB,OAAO,SAAS,MAAM,KAAK,KAC3B,MAAM,SAAS,KACf,MAAM,SAAS;AAEnB;;;;;;;;;;;;;;;;;;;;;;;AAwBA,SAAgB,oCACd,MACA,OAA0C,CAAC,GACgB;CAC3D,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAChD,MAAM,YAAY,KAAK,qBAAqB;CAE5C,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,UAAU,gBAAgB,GAAG;EAInC,MAAM,WAAW,IAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAK,EAAE,kBAAkB,KAAK;EAMrF,MAAM,QAAQ,UAA2B,aAAa,UAAU,IAAI;EAEpE,MAAM,cAAsD,CAAC;EAC7D,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,IAAI,QAAQ,GAAG,GACjD,IACE,EAAE,WAAW,QAAQ,KACrB,CAAC,EAAE,SAAS,KAAK,CAAC,KAClB,OAAO,MAAM,YACb,OAAO,SAAS,CAAC,GAEjB,YAAY,KAAK;GAAE,MAAM,EAAE,MAAM,CAAe;GAAG,OAAO;EAAE,CAAC;EAGjE,MAAM,MAAM,YAAY,QAAQ,MAAM,iBAAiB,IAAI,EAAE,IAAI,CAAC;EAElE,IAAI,IAAI,WAAW,GAAG;GACpB,MAAM,QAAQ,IAAI;GAClB,MAAM,QAAQ,KAAKA,UAAQ,MAAM,KAAK,CAAC;GACvC,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ,UAAU,MAAM,IAAI;KAC5B,aAAa;KACb,YAAY;KACZ,QAAQ,MAAM;KACd,YAAY,GAAG,MAAM,OAAO,MAAM;KAClC,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,IAAI,SAAS,GAAG;GAClB,MAAM,QAAQ,KAAKA,UAAQ,IAAI,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,IAAI,MAAM,CAAC;GAG7E,MAAM,aAAqC,OAAO,YAChD,IAAI,KAAK,MAAM,CAAC,EAAE,MAAM,KAAKA,UAAQ,EAAE,KAAK,CAAC,CAAC,CAAC,CACjD;GACA,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ;KACR,aAAa;KACb,YAAY;KACZ,QAAQ,IAAI,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;KACvC;KACA,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,CAAC,iBAAiB,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAO9D,MAAM,UAAU,YAAY,cAAc,GAAG,IAAI,cAAc,GAAG;EAClE,IAAI,OAAO,YAAY,YAAY,CAAC,OAAO,SAAS,OAAO,GACzD,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAE1C,MAAM,eAAeA,UAAQ,OAAO;EACpC,OAAO;GACL,OAAO,IAAI;GACX,QAAQ;IACN,OAAO;IACP,QAAQ;IACR,aAAa;IACb,YAAY;IACZ,QAAQ;IACR,YAAY,EAAE,qBAAqB,aAAa;IAChD,eAAe;IACf,GAAG;GACL;EACF;CACF,CAAC;AACH;;;;;;;;;;AAWA,SAAgB,gCACd,MACA,OAA0C,CAAC,GACU;CACrD,MAAM,WAAW,oCAAoC,MAAM;EAAE,GAAG;EAAM,iBAAiB;CAAM,CAAC;CAC9F,MAAM,MAA2D,CAAC;CAClE,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,SAAS;EACnB,IAAI,EAAE,UAAU,EAAE,OAAO,gBAAgB,iBACvC,IAAI,KAAK;GAAE,KAAK,KAAK;GAAK,QAAQ,EAAE;EAAO,CAAC;CAEhD;CACA,OAAO;AACT;AAEA,SAASA,UAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACnVA,MAAM,iBAAiB,MAAgC;CASrD,MAAM,IAAI,cAAc,CAAC;CACzB,OAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEA,SAAgB,oBAAoB,OAAsD;CACxF,MAAM,UAAU,MAAM,WAAW;CACjC,MAAM,UAAU,MAAM;CACtB,MAAM,MAAM,MAAM,YAAY,WAAW;CACzC,MAAM,MAAM,MAAM,YAAY,UAAU;CAExC,MAAM,OAAO,MAAM,KAAK,QAAQ,QAAQ,aAAa,QAAQ,GAAG,CAAC,CAAC;CAClE,MAAM,IAAI,KAAK;CACf,IAAI,IAAI,GACN,OAAO;EACL,UAAU,CAAC;EACX,kBAAkB,CAAC;EACnB,SAAS;EACT;EACA,WAAW,CAAC,0CAA0C,EAAE,yBAAyB;CACnF;CAEF,MAAM,aAAa,KAAK,IAAI,GAAG,MAAM,cAAc,KAAK,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,CAAC;CAClF,MAAM,SAAS,KAAK,MAAM,GAAG,IAAI,UAAU;CAC3C,MAAM,QAAQ,KAAK,MAAM,IAAI,UAAU;CAEvC,MAAM,WAAmC,CAAC;CAG1C,IAAI,SAAS;EACX,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,IACE,YAAY,UAAU,KACtB,WAAW,UAAU,KACrB,YAAY,UAAU,KACtB,WAAW,UAAU,GACrB;GACA,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GACtD,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GAGtD,MAAM,MAAM,KAAK,IAAI,GAAG,aAAa,UAAU;GAC/C,MAAM,WAAW,QAAQ,MAAM,CAAC;GAChC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,wBAAwB,WAAW,QAAQ,CAAC,EAAE,0BAA0B,WAAW,QAAQ,CAAC,EAAE,yBAC9F,yCAAyC,WAAW,QAAQ,CAAC,EAAE,UAAU,WAAW,QAAQ,CAAC,EAAE;IACrG,QAAQ;KACN;KACA;KACA;KACA,SAAS,YAAY;KACrB,QAAQ,WAAW;IACrB;GACF,CAAC;EACH;CACF;CAGA;EACE,MAAM,UAAU,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACvD,MAAM,SAAS,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACrD,IAAI,QAAQ,UAAU,KAAK,OAAO,UAAU,GAAG;GAC7C,MAAM,KAAK,YAAY,SAAS,MAAM;GAItC,MAAM,WAAW,QAAQ,KAAK,EAAG;GACjC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,MAAM,GAAG,QAAQ,CAAC,EAAE,8DACpB,MAAM,GAAG,QAAQ,CAAC,EAAE;IAC1B,QAAQ;KAAE;KAAI,SAAS,QAAQ;KAAQ,QAAQ,OAAO;IAAO;GAC/D,CAAC;EACH;CACF;CAGA;EACE,MAAM,cAAc,MAAM,qBAAqB,iBAAiB,MAAM,uBAAuB;EAC7F,MAAM,UAAU,KACb,KAAK,OAAO;GAAE,GAAG,QAAQ,CAAC;GAAG,GAAG,YAAY,CAAC;EAAE,EAAE,CAAC,CAClD,QAAQ,MAAqC,aAAa,EAAE,CAAC,KAAK,aAAa,EAAE,CAAC,CAAC;EACtF,IAAI,QAAQ,UAAU,GAAG;GAGvB,MAAM,IAAI,SAFC,QAAQ,KAAK,MAAM,EAAE,CAEZ,GADT,QAAQ,KAAK,MAAM,EAAE,CACR,CAAC;GAGzB,MAAM,WAAW,QAAQ,KAAM,KAAK,IAAI,GAAG,CAAC,CAAC;GAC7C,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,sDAAsD,EAAE,QAAQ,CAAC,EAAE,6BACnE,0CAA0C,EAAE,QAAQ,CAAC;IAC3D,QAAQ;KAAE,SAAS;KAAG,GAAG,QAAQ;IAAO;GAC1C,CAAC;EACH;CACF;CAGA;EAME,MAAM,UAAU,gCAAgC,MAAM;GACpD,GAAI,MAAM,2BAA2B,CAAC;GACtC,mBAAmB;EACrB,CAAC;EACD,IAAI,QAAQ,UAAU,GAAG;GACvB,MAAM,YAAY,QAAQ,MAAM,GAAG,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;GAEjE,MAAM,WACJ,KAFe,QAAQ,MAAM,KAAK,MAAM,QAAQ,SAAS,CAAC,CAE9C,CAAC,CAAC,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC;GACvF,MAAM,aACJ,KAAK,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC,IAC5C,KAAK,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC;GAC/C,MAAM,WAAW,KAAK,IAAI,GAAG,aAAa,QAAQ;GAClD,MAAM,WAAW,QAAQ,WAAW,CAAC;GACrC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,gBAAgB,WAAW,QAAQ,CAAC,EAAE,+BAA+B,SAAS,QAAQ,CAAC,EAAE,mDACzF,uDAAuD,WAAW,QAAQ,CAAC,EAAE,QAAQ,SAAS,QAAQ,CAAC,EAAE;IAC/G,QAAQ;KAAE;KAAY;KAAU;KAAU,GAAG,QAAQ;IAAO;GAC9D,CAAC;EACH;CACF;CAEA,MAAM,SAAS,SAAS,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,QAAQ,GAAG,CAAC;CACnE,IAAI,SAAS,WAAW,GACtB,OAAO;EACL;EACA,kBAAkB,CAAC;EACnB,SAAS;EACT,WAAW,CAAC,0DAA0D,EAAE,EAAE;EAC1E;CACF;CAEF,MAAM,UACJ,UAAU,MAAM,WAAW,UAAU,MAAM,YAAY;CACzD,MAAM,YAAY,SACf,QAAQ,MAAM,EAAE,YAAY,GAAG,CAAC,CAChC,KAAK,MAAM,GAAG,EAAE,OAAO,aAAa,EAAE,SAAS,QAAQ,CAAC,EAAE,KAAK,EAAE,SAAS;CAC7E,IAAI,UAAU,WAAW,GAAG,UAAU,KAAK,0CAA0C;CAErF,OAAO;EACL;EACA,kBAAkB,SAAS,KAAK,YAAY,QAAQ,MAAM;EAC1D;EACA;EACA;CACF;AACF;AAIA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,aAAa,OAAuC;CAC3D,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,YAAY,GAAa,GAAqB;CAErD,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,MAAM,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACvE,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,IAAI,KAAK,KAAK,IAAI,KAAK,EAAE,CAAC;CACvC;CACA,OAAO;AACT;AAEA,SAAS,iBACP,gBACmC;CACnC,QAAQ,QAAmB;EAIzB,MAAM,WAAW,gCAAgC,CAAC,GAAG,GAAG;GACtD,GAAI,kBAAkB,CAAC;GACvB,mBAAmB;EACrB,CAAC;EACD,OAAO,SAAS,WAAW,IAAI,SAAS,EAAE,CAAE,OAAO,QAAQ;CAC7D;AACF"}