@tangle-network/agent-eval 0.200.0 → 0.201.0

package/dist/{paired-promotion-decision-D5B1Cxc2.js.map → paired-promotion-decision-DhrmjdV8.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-promotion-decision-D5B1Cxc2.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.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 * Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n */\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n  BOOTSTRAP_GATE_MIN_N,\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 test also imposes its own\n   *  minimum: bootstrap eligibility for a continuous mean, or the exact-test\n   *  minimum for binary and explicit median targets. This is not a power guarantee. */\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  /** Declared binary support {0, binaryScale}, including when all observations\n   *  are zero. Must be finite and positive; every observation must use that\n   *  support. Uses the risk-difference mean, so cannot accompany 'median'.\n   *  Omitted: infer binary support from observed positive values. */\n  binaryScale?: number\n  /**\n   * Declare the outcome continuous: never infer a two-point support from the\n   * observed values, so the estimator is fixed before the data is seen. A\n   * sealed call on a continuous measure must route to the bootstrap whatever\n   * the sample happens to look like; with inference on, a sample whose values\n   * all land on {0, s} is silently decided on the score interval instead.\n   * Measured under discovery's E1 alternative (deltas in {+1, -0.5, 0} on a\n   * zero baseline) at 20 pairs: 5.4 % of 2,000 simulated samples (seed 11)\n   * re-route without this flag, 0 with it; the exact rate is 5.54 %\n   * (`paired-promotion-power.test.ts`). Cannot accompany `binaryScale`.\n   */\n  continuous?: boolean\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  /** Declared or inferred positive level ({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 for the target, confidence, and caller floor. */\n  minimumPairs: number\n  /** n >= minimumPairs. */\n  sufficient: boolean\n  /** The deciding interval is zero-width at numeric precision 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  /** Declared or inferred positive binary level; null off the binary path. */\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  declaredBinaryScale?: number,\n  continuous = false,\n): PairedDecisionShape {\n  if (continuous && declaredBinaryScale !== undefined) {\n    throw new Error('pairedDecisionShape: continuous cannot accompany binaryScale')\n  }\n  if (declaredBinaryScale !== undefined) {\n    if (!Number.isFinite(declaredBinaryScale) || declaredBinaryScale <= 0) {\n      throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n    }\n    if (statistic === 'median') {\n      throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n    }\n    for (const [name, arm] of [\n      ['before', before],\n      ['after', after],\n    ] as const) {\n      for (let i = 0; i < arm.length; i++) {\n        const value = arm[i]!\n        if (value !== 0 && value !== declaredBinaryScale) {\n          throw new Error(\n            `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  if (continuous) {\n    return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${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 { binaryScale, tieFraction } = pairedDecisionShape(\n    before,\n    after,\n    options.statistic,\n    options.binaryScale,\n    options.continuous ?? false,\n  )\n  const estimatorMinimum =\n    binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n  const n = before.length\n  const sufficient = n >= minimumPairs\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: minimumPairs,\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          ? bootstrapStatistic === 'mean'\n            ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n              ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n            : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n              ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n          : '',\n    }\n  }\n\n  // Equivalent deltas can differ by rounding after subtraction and averaging.\n  // Use the interval's own scale so tiny score units retain the same decision\n  // as a rescaled copy of the evidence.\n  const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n  const intervalTolerance = intervalScale * Number.EPSILON * 8\n  const indeterminate =\n    !Number.isFinite(core.low) ||\n    !Number.isFinite(core.high) ||\n    core.high - core.low <= intervalTolerance\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"],"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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACuDA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACA,aAAa,OACQ;CACrB,IAAI,cAAc,wBAAwB,KAAA,GACxC,MAAM,IAAI,MAAM,8DAA8D;CAEhF,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,IAAI,YACF,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;CAEvE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,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,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,aACR,QAAQ,cAAc,KACxB;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,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;EACZ,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,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,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"}
+{"version":3,"file":"paired-promotion-decision-DhrmjdV8.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.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 * Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n */\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n  BOOTSTRAP_GATE_MIN_N,\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 test also imposes its own\n   *  minimum: bootstrap eligibility for a continuous mean, or the exact-test\n   *  minimum for binary and explicit median targets. This is not a power guarantee. */\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  /** Declared binary support {0, binaryScale}, including when all observations\n   *  are zero. Must be finite and positive; every observation must use that\n   *  support. Uses the risk-difference mean, so cannot accompany 'median'.\n   *  Omitted: infer binary support from observed positive values. */\n  binaryScale?: number\n  /**\n   * Declare the outcome continuous: never infer a two-point support from the\n   * observed values, so the estimator is fixed before the data is seen. A\n   * sealed call on a continuous measure must route to the bootstrap whatever\n   * the sample happens to look like; with inference on, a sample whose values\n   * all land on {0, s} is silently decided on the score interval instead.\n   * Measured under discovery's E1 alternative (deltas in {+1, -0.5, 0} on a\n   * zero baseline) at 20 pairs: 5.4 % of 2,000 simulated samples (seed 11)\n   * re-route without this flag, 0 with it; the exact rate is 5.54 %\n   * (`paired-promotion-power.test.ts`). Cannot accompany `binaryScale`.\n   */\n  continuous?: boolean\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  /** Declared or inferred positive level ({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 for the target, confidence, and caller floor. */\n  minimumPairs: number\n  /** n >= minimumPairs. */\n  sufficient: boolean\n  /** The deciding interval is zero-width at numeric precision 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  /** Declared or inferred positive binary level; null off the binary path. */\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  declaredBinaryScale?: number,\n  continuous = false,\n): PairedDecisionShape {\n  if (continuous && declaredBinaryScale !== undefined) {\n    throw new Error('pairedDecisionShape: continuous cannot accompany binaryScale')\n  }\n  if (declaredBinaryScale !== undefined) {\n    if (!Number.isFinite(declaredBinaryScale) || declaredBinaryScale <= 0) {\n      throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n    }\n    if (statistic === 'median') {\n      throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n    }\n    for (const [name, arm] of [\n      ['before', before],\n      ['after', after],\n    ] as const) {\n      for (let i = 0; i < arm.length; i++) {\n        const value = arm[i]!\n        if (value !== 0 && value !== declaredBinaryScale) {\n          throw new Error(\n            `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  if (continuous) {\n    return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${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 { binaryScale, tieFraction } = pairedDecisionShape(\n    before,\n    after,\n    options.statistic,\n    options.binaryScale,\n    options.continuous ?? false,\n  )\n  const estimatorMinimum =\n    binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n  const n = before.length\n  const sufficient = n >= minimumPairs\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: minimumPairs,\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          ? bootstrapStatistic === 'mean'\n            ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n              ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n            : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n              ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n          : '',\n    }\n  }\n\n  // Equivalent deltas can differ by rounding after subtraction and averaging.\n  // Use the interval's own scale so tiny score units retain the same decision\n  // as a rescaled copy of the evidence.\n  const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n  const intervalTolerance = intervalScale * Number.EPSILON * 8\n  const indeterminate =\n    !Number.isFinite(core.low) ||\n    !Number.isFinite(core.high) ||\n    core.high - core.low <= intervalTolerance\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"],"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;CAGF,MAAM,cAAc,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAAS;CAClF,MAAM,QAAQ,eAAe,aAAa,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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACuDA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACA,aAAa,OACQ;CACrB,IAAI,cAAc,wBAAwB,KAAA,GACxC,MAAM,IAAI,MAAM,8DAA8D;CAEhF,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,IAAI,YACF,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;CAEvE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,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,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,aACR,QAAQ,cAAc,KACxB;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,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;EACZ,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,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,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"}

package/dist/{paired-tests-C8iCsioC.js → paired-tests-Ccr16gcW.js}

@@ -1,6 +1,6 @@
 import { s as ValidationError } from "./errors-Dngq5h35.js";
-import { a as medianInPlace, i as makeRng, n as binomialHalfUpperTail, t as assertFiniteSample } from "./internal-BMFSR8Ns.js";
-import { t as studentTCdf } from "./student-t-BA-Uy51p.js";
+import { a as medianInPlace, i as makeRng, n as binomialHalfUpperTail, t as assertFiniteSample } from "./internal-BOe3zmjQ.js";
+import { t as studentTCdf } from "./student-t-BBZYka0E.js";
 //#region src/statistics/paired-tests.ts
 /**
 * Paired significance tests on continuous scores: the paired t-test, the
@@ -210,4 +210,4 @@ const DECISION_PAIRED_DELTA_STATISTIC = "mean";
 //#endregion
 export { pairedSignTest as a, pairedDeltaTieFraction as i, DECISION_PAIRED_DELTA_STATISTIC as n, pairedTTest as o, pairedBootstrap as r, BOOTSTRAP_GATE_MIN_N as t };
 
-//# sourceMappingURL=paired-tests-C8iCsioC.js.map
+//# sourceMappingURL=paired-tests-Ccr16gcW.js.map

package/dist/{paired-tests-C8iCsioC.js.map → paired-tests-Ccr16gcW.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-tests-C8iCsioC.js","names":[],"sources":["../src/statistics/paired-tests.ts"],"sourcesContent":["/**\n * Paired significance tests on continuous scores: the paired t-test, the\n * promotion-gate paired bootstrap, the exact sign test, and the package-wide\n * paired-delta decision statistic.\n */\n\nimport { ValidationError } from '../errors'\nimport { studentTCdf } from '../math/student-t'\nimport { assertFiniteSample, binomialHalfUpperTail, makeRng, medianInPlace } from './internal'\n\nexport interface PairedTTestResult {\n  /** Null when the statistic is undefined — see {@link pairedTTest}. */\n  t: number | null\n  df: number\n  /** Null exactly when `t` is null. */\n  p: number | null\n}\n\n/**\n * Paired t-test — before/after measurements on the SAME items.\n * Pairing removes inter-item variance, giving tighter significance than\n * an unpaired test when comparing prompt v1 vs prompt v2 on identical\n * scenarios.\n *\n * Returns `t = p = null` where the statistic is undefined: fewer than two\n * pairs, or a non-zero constant delta whose observed variance is zero. A\n * constant shift carries no information about the variance it would have to\n * be compared against, so the honest answer is \"undefined\", not `p = 0` —\n * three observations cannot buy absolute certainty. This is the same contract\n * {@link pairedCohensDz} states for the same condition. An all-zero delta is\n * different: it is a measured null, and returns `t = 0, p = 1`.\n */\nexport function pairedTTest(before: number[], after: number[]): PairedTTestResult {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedTTest: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  assertFiniteSample('pairedTTest', 'before', before)\n  assertFiniteSample('pairedTTest', 'after', after)\n  const n = before.length\n  if (n < 2) return { t: null, df: 0, p: null }\n\n  const diffs = before.map((b, i) => after[i]! - b)\n  const mean = diffs.reduce((a, b) => a + b, 0) / n\n  const variance = diffs.reduce((acc, d) => acc + (d - mean) ** 2, 0) / (n - 1)\n  const se = Math.sqrt(variance / n)\n  if (se === 0) {\n    return mean === 0 ? { t: 0, df: n - 1, p: 1 } : { t: null, df: n - 1, p: null }\n  }\n\n  const t = mean / se\n  const df = n - 1\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\n// ── Paired bootstrap (promotion-gate effect size) ────────────────────\n\nexport interface PairedBootstrapResult {\n  /** Number of paired observations. */\n  n: number\n  /** Median of paired deltas (after − before). */\n  median: number\n  /** Mean of paired deltas. */\n  mean: number\n  /** Lower bound of the bootstrap CI on the chosen statistic. */\n  low: number\n  /** Upper bound of the bootstrap CI on the chosen statistic. */\n  high: number\n  /** Confidence level used (e.g. 0.95). */\n  confidence: number\n  /** Number of bootstrap resamples used. */\n  resamples: number\n  /** False below {@link BOOTSTRAP_GATE_MIN_N}. See {@link pairedBootstrap}. */\n  gateEligible: boolean\n}\n\n/**\n * Pairs below which a percentile bootstrap interval is descriptive spread only.\n *\n * `P(low > 0)` under a true null, against a nominal 2.5 %, measured over 4000\n * seeded trials: 13.53 % at n = 3, 3.52 % at n = 10, 3.10 % at n = 20 on the\n * median; 13.85 %, 4.90 %, 3.80 % on the mean. This is intrinsic to resampling\n * three points, not an implementation error — scipy's BCa gives 16.0 % on the\n * same n = 3 data — so no change to the estimator moves it. Below this floor\n * the decision belongs to the exact sign test or exact signed-rank test.\n */\nexport const BOOTSTRAP_GATE_MIN_N = 20\n\nexport interface PairedBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Statistic to bootstrap. Default 'median'. */\n  statistic?: 'median' | 'mean'\n  /** Deterministic seed. If omitted, derived from the deltas so the interval\n   *  is reproducible regardless. */\n  seed?: number\n}\n\n/**\n * Paired bootstrap on (after − before) deltas. Returns a CI on the chosen\n * statistic (median by default); pairs are resampled with replacement. Throws\n * on unequal sample sizes.\n *\n * `low > threshold` carries the stated confidence ONLY at `n ≥\n * {@link BOOTSTRAP_GATE_MIN_N}`, which `gateEligible` reports. Below it the\n * check fires under a true null several times more often than nominal, so the\n * interval is descriptive spread and a promotion must not turn on it.\n */\nexport function pairedBootstrap(\n  before: number[],\n  after: number[],\n  opts: PairedBootstrapOptions = {},\n): PairedBootstrapResult {\n  if (before.length !== after.length) {\n    throw new Error(`pairedBootstrap: unequal sample sizes (${before.length} vs ${after.length})`)\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const statistic = opts.statistic ?? 'median'\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedBootstrap: confidence must be in (0,1), got ${confidence}`)\n  }\n\n  const n = before.length\n  const deltas = before.map((b, i) => after[i]! - b)\n  const gateEligible = n >= BOOTSTRAP_GATE_MIN_N\n  if (n === 0) {\n    return { n: 0, median: 0, mean: 0, low: 0, high: 0, confidence, resamples, gateEligible }\n  }\n  if (n === 1) {\n    const d = deltas[0]!\n    return { n: 1, median: d, mean: d, low: d, high: d, confidence, resamples, gateEligible }\n  }\n\n  const rng = makeRng(opts.seed, deltas)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    if (statistic === 'mean') {\n      let sum = 0\n      for (let k = 0; k < n; k++) {\n        sum += deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = sum / n\n    } else {\n      const acc = new Array<number>(n)\n      for (let k = 0; k < n; k++) {\n        acc[k] = deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = medianInPlace(acc)\n    }\n  }\n  samples.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n\n  return {\n    n,\n    median: medianInPlace([...deltas]),\n    mean: deltas.reduce((s, x) => s + x, 0) / n,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n    gateEligible,\n  }\n}\n\n/** Pre-registered direction for a one-sided paired sign test. */\nexport type SignTestAlternative = 'greater' | 'less'\n\n/** Exact one-sided sign-test result for paired numeric differences. */\nexport interface PairedSignTestResult {\n  /** Total supplied differences, including zero ties. */\n  n: number\n  /** Strictly positive differences. */\n  positive: number\n  /** Strictly negative differences. */\n  negative: number\n  /** Zero differences excluded from the binomial test. */\n  ties: number\n  /** Non-zero differences used by the binomial test. */\n  nNonTies: number\n  /** Direction of the pre-registered alternative hypothesis. */\n  alternative: SignTestAlternative\n  /** Exact one-sided p-value under P(positive) = P(negative) = 0.5. */\n  pValue: number\n}\n\n/**\n * Exact one-sided sign test over paired differences.\n *\n * Pass `after[i] - before[i]` for each matched item. `alternative = 'greater'`\n * tests whether positive signs are more likely than negative signs and returns\n * `P(Binomial(nNonTies, 0.5) >= positive)`. `alternative = 'less'` treats\n * negative signs as successes instead. With a continuous difference\n * distribution this is the usual directional median test. Exact zero\n * differences are ties and do not enter the binomial denominator. All-tie and\n * empty inputs return p = 1. Every input difference must be finite, and the\n * direction must be chosen explicitly so a caller cannot select it after\n * seeing the signs.\n */\nexport function pairedSignTest(\n  differences: readonly number[],\n  alternative: SignTestAlternative,\n): PairedSignTestResult {\n  if (alternative !== 'greater' && alternative !== 'less') {\n    throw new ValidationError(\n      `pairedSignTest: alternative must be 'greater' or 'less', got ${alternative}`,\n    )\n  }\n\n  let positive = 0\n  let negative = 0\n  let ties = 0\n  for (let i = 0; i < differences.length; i++) {\n    const difference = differences[i]!\n    if (!Number.isFinite(difference)) {\n      throw new ValidationError(\n        `pairedSignTest: difference at index ${i} must be finite, got ${difference}`,\n      )\n    }\n    if (difference > 0) positive++\n    else if (difference < 0) negative++\n    else ties++\n  }\n\n  const nNonTies = positive + negative\n  const successes = alternative === 'greater' ? positive : negative\n  return {\n    n: differences.length,\n    positive,\n    negative,\n    ties,\n    nNonTies,\n    alternative,\n    pValue: binomialHalfUpperTail(successes, nNonTies),\n  }\n}\n\n/** Fraction of paired observations whose delta is an exact tie (|after − before|\n *  < 1e-9). Throws on unequal sample sizes; 0 pairs ⇒ 0. */\nexport function pairedDeltaTieFraction(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number {\n  if (before.length !== after.length) {\n    throw new Error(\n      `pairedDeltaTieFraction: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const n = before.length\n  if (n === 0) return 0\n  let ties = 0\n  for (let i = 0; i < n; i++) {\n    if (Math.abs(after[i]! - before[i]!) < 1e-9) ties++\n  }\n  return ties / n\n}\n\n/**\n * The paired-delta statistic a DECISION is computed on, package-wide.\n *\n * The mean paired delta is the estimator that answers the question a promotion\n * gate asks — \"by how much did the candidate move the score\" — in the caller's\n * own units, and it equals the aggregate lift everyone quotes. The MEDIAN\n * answers a different question and loses the answer to this one in every regime\n * eval data actually lands in:\n *   - TWO-POINT (pass/fail) outcomes on any encoding: the delta vector lives in\n *     {−s, 0, +s} dominated by zeros, so the median and its whole bootstrap CI\n *     are pinned at exactly 0 however large the shift. (Decide these on\n *     {@link pairedRiskDifferenceExact} instead — same estimand, exact interval.)\n *   - TIE-DOMINATED outcomes: at half the pairs tied the sample median is 0 by\n *     construction, and `ci.low > threshold` then answers \"no\" forever at a\n *     non-negative threshold and \"yes\" forever at a negative one.\n *   - LOW-CARDINALITY outcomes, even well below half ties: judge dimensions on\n *     integer 0-100, and block scores like {⅔, 1} from averaging pass/fail\n *     leaves, put the median on a coarse lattice whose bootstrap percentiles\n *     land on atoms. Measured: 26 blocks of 3 pass/fail leaves carrying a real\n *     +12.8pp lift, only 23% of pairs tied, gives a median CI of [0, 0.333] —\n *     lower bound exactly 0, so a gate at threshold 0 refuses a real lift.\n * That last case is why there is no tie-fraction threshold here: any cutoff on\n * ties leaves the lattice case open on the other side of it.\n *\n * `heldoutSignificance` has defaulted to the mean since #316 for the same\n * reason. The median remains available per call site for callers who\n * specifically want outlier robustness and accept the blindness.\n */\nexport const DECISION_PAIRED_DELTA_STATISTIC: 'mean' = 'mean'\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AAgCA,SAAgB,YAAY,QAAkB,OAAoC;CAChF,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,sCAAsC,OAAO,OAAO,MAAM,MAAM,OAAO,EACzE;CAEF,mBAAmB,eAAe,UAAU,MAAM;CAClD,mBAAmB,eAAe,SAAS,KAAK;CAChD,MAAM,IAAI,OAAO;CACjB,IAAI,IAAI,GAAG,OAAO;EAAE,GAAG;EAAM,IAAI;EAAG,GAAG;CAAK;CAE5C,MAAM,QAAQ,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CAChD,MAAM,OAAO,MAAM,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAChD,MAAM,WAAW,MAAM,QAAQ,KAAK,MAAM,OAAO,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI;CAC3E,MAAM,KAAK,KAAK,KAAK,WAAW,CAAC;CACjC,IAAI,OAAO,GACT,OAAO,SAAS,IAAI;EAAE,GAAG;EAAG,IAAI,IAAI;EAAG,GAAG;CAAE,IAAI;EAAE,GAAG;EAAM,IAAI,IAAI;EAAG,GAAG;CAAK;CAGhF,MAAM,IAAI,OAAO;CACjB,MAAM,KAAK,IAAI;CAEf,OAAO;EAAE;EAAG;EAAI,GADN,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;CAC5B;AACpB;;;;;;;;;;;AAiCA,MAAa,uBAAuB;;;;;;;;;;;AAwBpC,SAAgB,gBACd,QACA,OACA,OAA+B,CAAC,GACT;CACvB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MAAM,0CAA0C,OAAO,OAAO,MAAM,MAAM,OAAO,EAAE;CAE/F,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,qDAAqD,YAAY;CAGnF,MAAM,IAAI,OAAO;CACjB,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,eAAe,KAAA;CACrB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,QAAQ;EAAG,MAAM;EAAG,KAAK;EAAG,MAAM;EAAG;EAAY;EAAW;CAAa;CAE1F,IAAI,MAAM,GAAG;EACX,MAAM,IAAI,OAAO;EACjB,OAAO;GAAE,GAAG;GAAG,QAAQ;GAAG,MAAM;GAAG,KAAK;GAAG,MAAM;GAAG;GAAY;GAAW;EAAa;CAC1F;CAEA,MAAM,MAAM,QAAQ,KAAK,MAAM,MAAM;CACrC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAC7B,IAAI,cAAc,QAAQ;EACxB,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,OAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEpC,QAAQ,KAAK,MAAM;CACrB,OAAO;EACL,MAAM,MAAM,IAAI,MAAc,CAAC;EAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEtC,QAAQ,KAAK,cAAc,GAAG;CAChC;CAEF,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAE5B,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;CACjD,MAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;CAElF,OAAO;EACL;EACA,QAAQ,cAAc,CAAC,GAAG,MAAM,CAAC;EACjC,MAAM,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;EAC1C,KAAK,QAAQ;EACb,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM;EACtC;EACA;EACA;CACF;AACF;;;;;;;;;;;;;;AAoCA,SAAgB,eACd,aACA,aACsB;CACtB,IAAI,gBAAgB,aAAa,gBAAgB,QAC/C,MAAM,IAAI,gBACR,gEAAgE,aAClE;CAGF,IAAI,WAAW;CACf,IAAI,WAAW;CACf,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;EAC3C,MAAM,aAAa,YAAY;EAC/B,IAAI,CAAC,OAAO,SAAS,UAAU,GAC7B,MAAM,IAAI,gBACR,uCAAuC,EAAE,uBAAuB,YAClE;EAEF,IAAI,aAAa,GAAG;OACf,IAAI,aAAa,GAAG;OACpB;CACP;CAEA,MAAM,WAAW,WAAW;CAC5B,MAAM,YAAY,gBAAgB,YAAY,WAAW;CACzD,OAAO;EACL,GAAG,YAAY;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ,sBAAsB,WAAW,QAAQ;CACnD;AACF;;;AAIA,SAAgB,uBACd,QACA,OACQ;CACR,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,iDAAiD,OAAO,OAAO,MAAM,MAAM,OAAO,EACpF;CAEF,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO;CACpB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,IAAI,MAAM,KAAM,OAAO,EAAG,IAAI,MAAM;CAE/C,OAAO,OAAO;AAChB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,MAAa,kCAA0C"}
+{"version":3,"file":"paired-tests-Ccr16gcW.js","names":[],"sources":["../src/statistics/paired-tests.ts"],"sourcesContent":["/**\n * Paired significance tests on continuous scores: the paired t-test, the\n * promotion-gate paired bootstrap, the exact sign test, and the package-wide\n * paired-delta decision statistic.\n */\n\nimport { ValidationError } from '../errors'\nimport { studentTCdf } from '../math/student-t'\nimport { assertFiniteSample, binomialHalfUpperTail, makeRng, medianInPlace } from './internal'\n\nexport interface PairedTTestResult {\n  /** Null when the statistic is undefined — see {@link pairedTTest}. */\n  t: number | null\n  df: number\n  /** Null exactly when `t` is null. */\n  p: number | null\n}\n\n/**\n * Paired t-test — before/after measurements on the SAME items.\n * Pairing removes inter-item variance, giving tighter significance than\n * an unpaired test when comparing prompt v1 vs prompt v2 on identical\n * scenarios.\n *\n * Returns `t = p = null` where the statistic is undefined: fewer than two\n * pairs, or a non-zero constant delta whose observed variance is zero. A\n * constant shift carries no information about the variance it would have to\n * be compared against, so the honest answer is \"undefined\", not `p = 0` —\n * three observations cannot buy absolute certainty. This is the same contract\n * {@link pairedCohensDz} states for the same condition. An all-zero delta is\n * different: it is a measured null, and returns `t = 0, p = 1`.\n */\nexport function pairedTTest(before: number[], after: number[]): PairedTTestResult {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `pairedTTest: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  assertFiniteSample('pairedTTest', 'before', before)\n  assertFiniteSample('pairedTTest', 'after', after)\n  const n = before.length\n  if (n < 2) return { t: null, df: 0, p: null }\n\n  const diffs = before.map((b, i) => after[i]! - b)\n  const mean = diffs.reduce((a, b) => a + b, 0) / n\n  const variance = diffs.reduce((acc, d) => acc + (d - mean) ** 2, 0) / (n - 1)\n  const se = Math.sqrt(variance / n)\n  if (se === 0) {\n    return mean === 0 ? { t: 0, df: n - 1, p: 1 } : { t: null, df: n - 1, p: null }\n  }\n\n  const t = mean / se\n  const df = n - 1\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\n// ── Paired bootstrap (promotion-gate effect size) ────────────────────\n\nexport interface PairedBootstrapResult {\n  /** Number of paired observations. */\n  n: number\n  /** Median of paired deltas (after − before). */\n  median: number\n  /** Mean of paired deltas. */\n  mean: number\n  /** Lower bound of the bootstrap CI on the chosen statistic. */\n  low: number\n  /** Upper bound of the bootstrap CI on the chosen statistic. */\n  high: number\n  /** Confidence level used (e.g. 0.95). */\n  confidence: number\n  /** Number of bootstrap resamples used. */\n  resamples: number\n  /** False below {@link BOOTSTRAP_GATE_MIN_N}. See {@link pairedBootstrap}. */\n  gateEligible: boolean\n}\n\n/**\n * Pairs below which a percentile bootstrap interval is descriptive spread only.\n *\n * `P(low > 0)` under a true null, against a nominal 2.5 %, measured over 4000\n * seeded trials: 13.53 % at n = 3, 3.52 % at n = 10, 3.10 % at n = 20 on the\n * median; 13.85 %, 4.90 %, 3.80 % on the mean. This is intrinsic to resampling\n * three points, not an implementation error — scipy's BCa gives 16.0 % on the\n * same n = 3 data — so no change to the estimator moves it. Below this floor\n * the decision belongs to the exact sign test or exact signed-rank test.\n */\nexport const BOOTSTRAP_GATE_MIN_N = 20\n\nexport interface PairedBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Statistic to bootstrap. Default 'median'. */\n  statistic?: 'median' | 'mean'\n  /** Deterministic seed. If omitted, derived from the deltas so the interval\n   *  is reproducible regardless. */\n  seed?: number\n}\n\n/**\n * Paired bootstrap on (after − before) deltas. Returns a CI on the chosen\n * statistic (median by default); pairs are resampled with replacement. Throws\n * on unequal sample sizes.\n *\n * `low > threshold` carries the stated confidence ONLY at `n ≥\n * {@link BOOTSTRAP_GATE_MIN_N}`, which `gateEligible` reports. Below it the\n * check fires under a true null several times more often than nominal, so the\n * interval is descriptive spread and a promotion must not turn on it.\n */\nexport function pairedBootstrap(\n  before: number[],\n  after: number[],\n  opts: PairedBootstrapOptions = {},\n): PairedBootstrapResult {\n  if (before.length !== after.length) {\n    throw new Error(`pairedBootstrap: unequal sample sizes (${before.length} vs ${after.length})`)\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const statistic = opts.statistic ?? 'median'\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedBootstrap: confidence must be in (0,1), got ${confidence}`)\n  }\n\n  const n = before.length\n  const deltas = before.map((b, i) => after[i]! - b)\n  const gateEligible = n >= BOOTSTRAP_GATE_MIN_N\n  if (n === 0) {\n    return { n: 0, median: 0, mean: 0, low: 0, high: 0, confidence, resamples, gateEligible }\n  }\n  if (n === 1) {\n    const d = deltas[0]!\n    return { n: 1, median: d, mean: d, low: d, high: d, confidence, resamples, gateEligible }\n  }\n\n  const rng = makeRng(opts.seed, deltas)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    if (statistic === 'mean') {\n      let sum = 0\n      for (let k = 0; k < n; k++) {\n        sum += deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = sum / n\n    } else {\n      const acc = new Array<number>(n)\n      for (let k = 0; k < n; k++) {\n        acc[k] = deltas[Math.floor(rng() * n)]!\n      }\n      samples[b] = medianInPlace(acc)\n    }\n  }\n  samples.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n\n  return {\n    n,\n    median: medianInPlace([...deltas]),\n    mean: deltas.reduce((s, x) => s + x, 0) / n,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n    gateEligible,\n  }\n}\n\n/** Pre-registered direction for a one-sided paired sign test. */\nexport type SignTestAlternative = 'greater' | 'less'\n\n/** Exact one-sided sign-test result for paired numeric differences. */\nexport interface PairedSignTestResult {\n  /** Total supplied differences, including zero ties. */\n  n: number\n  /** Strictly positive differences. */\n  positive: number\n  /** Strictly negative differences. */\n  negative: number\n  /** Zero differences excluded from the binomial test. */\n  ties: number\n  /** Non-zero differences used by the binomial test. */\n  nNonTies: number\n  /** Direction of the pre-registered alternative hypothesis. */\n  alternative: SignTestAlternative\n  /** Exact one-sided p-value under P(positive) = P(negative) = 0.5. */\n  pValue: number\n}\n\n/**\n * Exact one-sided sign test over paired differences.\n *\n * Pass `after[i] - before[i]` for each matched item. `alternative = 'greater'`\n * tests whether positive signs are more likely than negative signs and returns\n * `P(Binomial(nNonTies, 0.5) >= positive)`. `alternative = 'less'` treats\n * negative signs as successes instead. With a continuous difference\n * distribution this is the usual directional median test. Exact zero\n * differences are ties and do not enter the binomial denominator. All-tie and\n * empty inputs return p = 1. Every input difference must be finite, and the\n * direction must be chosen explicitly so a caller cannot select it after\n * seeing the signs.\n */\nexport function pairedSignTest(\n  differences: readonly number[],\n  alternative: SignTestAlternative,\n): PairedSignTestResult {\n  if (alternative !== 'greater' && alternative !== 'less') {\n    throw new ValidationError(\n      `pairedSignTest: alternative must be 'greater' or 'less', got ${alternative}`,\n    )\n  }\n\n  let positive = 0\n  let negative = 0\n  let ties = 0\n  for (let i = 0; i < differences.length; i++) {\n    const difference = differences[i]!\n    if (!Number.isFinite(difference)) {\n      throw new ValidationError(\n        `pairedSignTest: difference at index ${i} must be finite, got ${difference}`,\n      )\n    }\n    if (difference > 0) positive++\n    else if (difference < 0) negative++\n    else ties++\n  }\n\n  const nNonTies = positive + negative\n  const successes = alternative === 'greater' ? positive : negative\n  return {\n    n: differences.length,\n    positive,\n    negative,\n    ties,\n    nNonTies,\n    alternative,\n    pValue: binomialHalfUpperTail(successes, nNonTies),\n  }\n}\n\n/** Fraction of paired observations whose delta is an exact tie (|after − before|\n *  < 1e-9). Throws on unequal sample sizes; 0 pairs ⇒ 0. */\nexport function pairedDeltaTieFraction(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number {\n  if (before.length !== after.length) {\n    throw new Error(\n      `pairedDeltaTieFraction: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const n = before.length\n  if (n === 0) return 0\n  let ties = 0\n  for (let i = 0; i < n; i++) {\n    if (Math.abs(after[i]! - before[i]!) < 1e-9) ties++\n  }\n  return ties / n\n}\n\n/**\n * The paired-delta statistic a DECISION is computed on, package-wide.\n *\n * The mean paired delta is the estimator that answers the question a promotion\n * gate asks — \"by how much did the candidate move the score\" — in the caller's\n * own units, and it equals the aggregate lift everyone quotes. The MEDIAN\n * answers a different question and loses the answer to this one in every regime\n * eval data actually lands in:\n *   - TWO-POINT (pass/fail) outcomes on any encoding: the delta vector lives in\n *     {−s, 0, +s} dominated by zeros, so the median and its whole bootstrap CI\n *     are pinned at exactly 0 however large the shift. (Decide these on\n *     {@link pairedRiskDifferenceExact} instead — same estimand, exact interval.)\n *   - TIE-DOMINATED outcomes: at half the pairs tied the sample median is 0 by\n *     construction, and `ci.low > threshold` then answers \"no\" forever at a\n *     non-negative threshold and \"yes\" forever at a negative one.\n *   - LOW-CARDINALITY outcomes, even well below half ties: judge dimensions on\n *     integer 0-100, and block scores like {⅔, 1} from averaging pass/fail\n *     leaves, put the median on a coarse lattice whose bootstrap percentiles\n *     land on atoms. Measured: 26 blocks of 3 pass/fail leaves carrying a real\n *     +12.8pp lift, only 23% of pairs tied, gives a median CI of [0, 0.333] —\n *     lower bound exactly 0, so a gate at threshold 0 refuses a real lift.\n * That last case is why there is no tie-fraction threshold here: any cutoff on\n * ties leaves the lattice case open on the other side of it.\n *\n * `heldoutSignificance` has defaulted to the mean since #316 for the same\n * reason. The median remains available per call site for callers who\n * specifically want outlier robustness and accept the blindness.\n */\nexport const DECISION_PAIRED_DELTA_STATISTIC: 'mean' = 'mean'\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AAgCA,SAAgB,YAAY,QAAkB,OAAoC;CAChF,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,sCAAsC,OAAO,OAAO,MAAM,MAAM,OAAO,EACzE;CAEF,mBAAmB,eAAe,UAAU,MAAM;CAClD,mBAAmB,eAAe,SAAS,KAAK;CAChD,MAAM,IAAI,OAAO;CACjB,IAAI,IAAI,GAAG,OAAO;EAAE,GAAG;EAAM,IAAI;EAAG,GAAG;CAAK;CAE5C,MAAM,QAAQ,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CAChD,MAAM,OAAO,MAAM,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAChD,MAAM,WAAW,MAAM,QAAQ,KAAK,MAAM,OAAO,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI;CAC3E,MAAM,KAAK,KAAK,KAAK,WAAW,CAAC;CACjC,IAAI,OAAO,GACT,OAAO,SAAS,IAAI;EAAE,GAAG;EAAG,IAAI,IAAI;EAAG,GAAG;CAAE,IAAI;EAAE,GAAG;EAAM,IAAI,IAAI;EAAG,GAAG;CAAK;CAGhF,MAAM,IAAI,OAAO;CACjB,MAAM,KAAK,IAAI;CAEf,OAAO;EAAE;EAAG;EAAI,GADN,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;CAC5B;AACpB;;;;;;;;;;;AAiCA,MAAa,uBAAuB;;;;;;;;;;;AAwBpC,SAAgB,gBACd,QACA,OACA,OAA+B,CAAC,GACT;CACvB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MAAM,0CAA0C,OAAO,OAAO,MAAM,MAAM,OAAO,EAAE;CAE/F,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,qDAAqD,YAAY;CAGnF,MAAM,IAAI,OAAO;CACjB,MAAM,SAAS,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC;CACjD,MAAM,eAAe,KAAA;CACrB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,QAAQ;EAAG,MAAM;EAAG,KAAK;EAAG,MAAM;EAAG;EAAY;EAAW;CAAa;CAE1F,IAAI,MAAM,GAAG;EACX,MAAM,IAAI,OAAO;EACjB,OAAO;GAAE,GAAG;GAAG,QAAQ;GAAG,MAAM;GAAG,KAAK;GAAG,MAAM;GAAG;GAAY;GAAW;EAAa;CAC1F;CAEA,MAAM,MAAM,QAAQ,KAAK,MAAM,MAAM;CACrC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAC7B,IAAI,cAAc,QAAQ;EACxB,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,OAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEpC,QAAQ,KAAK,MAAM;CACrB,OAAO;EACL,MAAM,MAAM,IAAI,MAAc,CAAC;EAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEtC,QAAQ,KAAK,cAAc,GAAG;CAChC;CAEF,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAE5B,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;CACjD,MAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;CAElF,OAAO;EACL;EACA,QAAQ,cAAc,CAAC,GAAG,MAAM,CAAC;EACjC,MAAM,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;EAC1C,KAAK,QAAQ;EACb,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM;EACtC;EACA;EACA;CACF;AACF;;;;;;;;;;;;;;AAoCA,SAAgB,eACd,aACA,aACsB;CACtB,IAAI,gBAAgB,aAAa,gBAAgB,QAC/C,MAAM,IAAI,gBACR,gEAAgE,aAClE;CAGF,IAAI,WAAW;CACf,IAAI,WAAW;CACf,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;EAC3C,MAAM,aAAa,YAAY;EAC/B,IAAI,CAAC,OAAO,SAAS,UAAU,GAC7B,MAAM,IAAI,gBACR,uCAAuC,EAAE,uBAAuB,YAClE;EAEF,IAAI,aAAa,GAAG;OACf,IAAI,aAAa,GAAG;OACpB;CACP;CAEA,MAAM,WAAW,WAAW;CAC5B,MAAM,YAAY,gBAAgB,YAAY,WAAW;CACzD,OAAO;EACL,GAAG,YAAY;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ,sBAAsB,WAAW,QAAQ;CACnD;AACF;;;AAIA,SAAgB,uBACd,QACA,OACQ;CACR,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,iDAAiD,OAAO,OAAO,MAAM,MAAM,OAAO,EACpF;CAEF,MAAM,IAAI,OAAO;CACjB,IAAI,MAAM,GAAG,OAAO;CACpB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,IAAI,KAAK,IAAI,MAAM,KAAM,OAAO,EAAG,IAAI,MAAM;CAE/C,OAAO,OAAO;AAChB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,MAAa,kCAA0C"}

package/dist/pipelines/index.d.ts

@@ -6,7 +6,7 @@ import { n as TrajectoryStep } from "../trajectory-Y1fmdlw7.js";
 import { c as computeToolUseMetrics, d as judgeAgreementView, f as BudgetBreachFinding, g as BaselineReport, h as BaselineOptions, i as toolWasteView, l as JudgeAgreementReport, m as budgetBreachView, n as ToolWasteOptions, p as BudgetBreachReport, r as ToolWasteReport, t as ToolWasteFinding, u as JudgePair } from "../tool-waste-CP5DO383.js";
 //#region src/pipelines/first-divergence.d.ts
 /** One step of a run, as the diff reads it. */
-interface DiffStep {
+export interface DiffStep {
   /** The step's id within its run. */
   id: string;
   name: string;
@@ -14,13 +14,13 @@ interface DiffStep {
   /** Fields compared once two steps pair, such as status, tool or model. */
   fields: Readonly<Record<string, string | number | boolean | null>>;
 }
-type StepPairing = 'id' | 'position' | 'name';
-interface FieldDifference {
+export type StepPairing = 'id' | 'position' | 'name';
+export interface FieldDifference {
   field: string;
   a: string | number | boolean | null;
   b: string | number | boolean | null;
 }
-interface StepPair {
+export interface StepPair {
   /** Index in run A. */
   a: number;
   /** Index in run B. */
@@ -29,8 +29,8 @@ interface StepPair {
   /** Empty when the paired steps agree on name, kind and every field. */
   differences: FieldDifference[];
 }
-type StepDivergenceKind = 'changed' | 'replaced' | 'only-in-a' | 'only-in-b' | 'reordered';
-interface StepDivergence {
+export type StepDivergenceKind = 'changed' | 'replaced' | 'only-in-a' | 'only-in-b' | 'reordered';
+export interface StepDivergence {
   /** The first index at which the runs stop agreeing; equals `commonPrefixLen`. */
   index: number;
   kind: StepDivergenceKind;
@@ -41,7 +41,7 @@ interface StepDivergence {
   differences: FieldDifference[];
   reason: string;
 }
-interface StepDiff {
+export interface StepDiff {
   /** Paired steps in run A order. */
   pairs: StepPair[];
   /** Indexes of run A steps with no partner in run B. */
@@ -53,8 +53,8 @@ interface StepDiff {
   /** Null when the runs agree step for step. */
   firstDivergence: StepDivergence | null;
 }
-declare function diffSteps(a: readonly DiffStep[], b: readonly DiffStep[]): StepDiff;
-interface DivergenceReport {
+export declare function diffSteps(a: readonly DiffStep[], b: readonly DiffStep[]): StepDiff;
+export interface DivergenceReport {
   runA: string;
   runB: string;
   /** Null when the runs agree step for step. */
@@ -66,31 +66,31 @@ interface DivergenceReport {
   /** The step diff behind the first divergence, over trajectory step indexes. */
   diff: StepDiff;
 }
-declare function firstDivergenceView(store: TraceStore, runA: string, runB: string): Promise<DivergenceReport>;
+export declare function firstDivergenceView(store: TraceStore, runA: string, runB: string): Promise<DivergenceReport>;
 /**
  * One run's spans as diff steps, in containment order: depth first, siblings
  * by start time, the same order `buildTrajectory` gives a trace store run.
  * Spans whose parent is absent, or whose parents form a cycle, start their own
  * subtree, so every span appears exactly once.
  */
-declare function diffStepsFromSpans(spans: readonly DiagnosisSpan[]): DiffStep[];
+export declare function diffStepsFromSpans(spans: readonly DiagnosisSpan[]): DiffStep[];
 //#endregion
 //#region src/pipelines/regression.d.ts
-interface RegressionSpec {
+export interface RegressionSpec {
   metric: string;
   higherIsBetter: boolean;
   /** Extract a scalar from a run. Omit it and `metric` must name one of
    *  `RUN_METRICS`; any other name is refused. */
   extract?: (run: Run, store: TraceStore) => Promise<number | null>;
 }
-interface RegressionOptions extends BaselineOptions {
+export interface RegressionOptions extends BaselineOptions {
   baseline: RunFilter;
   candidate: RunFilter;
 }
-declare function regressionView(store: TraceStore, metrics: RegressionSpec[], options: RegressionOptions): Promise<BaselineReport>;
+export declare function regressionView(store: TraceStore, metrics: RegressionSpec[], options: RegressionOptions): Promise<BaselineReport>;
 //#endregion
 //#region src/pipelines/stuck-loop.d.ts
-interface StuckLoopFinding {
+export interface StuckLoopFinding {
   runId: string;
   toolName: string;
   argHash: string;
@@ -102,12 +102,12 @@ interface StuckLoopFinding {
   /** Milliseconds between first and last call in the loop. */
   windowMs: number;
 }
-interface StuckLoopReport {
+export interface StuckLoopReport {
   findings: StuckLoopFinding[];
   affectedRunRatio: number;
   totalRuns: number;
 }
-interface StuckLoopOptions {
+export interface StuckLoopOptions {
   /** Minimum call count to flag a loop (default 3). */
   minOccurrences?: number;
   /** Maximum time between the first and last repeated call (default 60 seconds). */
@@ -120,7 +120,7 @@ interface StuckLoopOptions {
   /** Filter to a specific runId; omit to scan the entire corpus. */
   runId?: string;
 }
-declare function stuckLoopView(store: TraceStore, options?: StuckLoopOptions): Promise<StuckLoopReport>;
+export declare function stuckLoopView(store: TraceStore, options?: StuckLoopOptions): Promise<StuckLoopReport>;
 //#endregion
-export { BudgetBreachFinding, BudgetBreachReport, DiffStep, DivergenceReport, FailureCluster, FailureClusterReport, FieldDifference, JudgeAgreementReport, JudgePair, RegressionOptions, RegressionSpec, StepDiff, StepDivergence, StepDivergenceKind, StepPair, StepPairing, StuckLoopFinding, StuckLoopOptions, StuckLoopReport, ToolWasteFinding, ToolWasteOptions, ToolWasteReport, budgetBreachView, computeToolUseMetrics, diffSteps, diffStepsFromSpans, failureClusterView, firstDivergenceView, judgeAgreementView, regressionView, stuckLoopView, toolWasteView };
+export { BudgetBreachFinding, BudgetBreachReport, FailureCluster, FailureClusterReport, JudgeAgreementReport, JudgePair, ToolWasteFinding, ToolWasteOptions, ToolWasteReport, budgetBreachView, computeToolUseMetrics, failureClusterView, judgeAgreementView, toolWasteView };
 //# sourceMappingURL=index.d.ts.map

package/dist/pipelines/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/pipelines/first-divergence.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts"],"mappings":";;;;;;;;UAyBiB;;EAEf;EACA;EACA;;EAEA,QAAQ,SAAS;;KAGP;UAEK;EACf;EACA;EACA;;UAGe;;EAEf;;EAEA;EACA,UAAU;;EAEV,aAAa;;KAGH;UAEK;;EAEf;EACA,MAAM;;EAEN;EACA;;EAEA,aAAa;EACb;;UAGe;;EAEf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,iBAAiB;;iBAGH,UAAU,YAAY,YAAY,YAAY,aAAa;UAwI1D;EACf;EACA;;EAEA;EACA,QAAQ;EACR,QAAQ;EACR;EACA;;EAEA,MAAM;;iBAGc,oBACpB,OAAO,YACP,cACA,eACC,QAAQ;;;;;;;iBAmCK,mBAAmB,gBAAgB,kBAAkB;;;UC7PpD;EACf;EACA;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B,0BAA0B;EACzC,UAAU;EACV,WAAW;;iBAGS,eACpB,OAAO,YACP,SAAS,kBACT,SAAS,oBACR,QAAQ;;;UCFM;EACf;EACA;EACA;;EAEA;EACA;;EAEA;;EAEA;;UAGe;EACf,UAAU;EACV;EACA;;UAGe;;EAEf;;EAEA;;;;;EAKA;;EAEA;;iBAGoB,cACpB,OAAO,YACP,UAAS,mBACR,QAAQ"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/pipelines/first-divergence.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts"],"mappings":";;;;;;;;iBAyBiB;;EAEf;EACA;EACA;;EAEA,QAAQ,SAAS;;YAGP;iBAEK;EACf;EACA;EACA;;iBAGe;;EAEf;;EAEA;EACA,UAAU;;EAEV,aAAa;;YAGH;iBAEK;;EAEf;EACA,MAAM;;EAEN;EACA;;EAEA,aAAa;EACb;;iBAGe;;EAEf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,iBAAiB;;wBAGH,UAAU,YAAY,YAAY,YAAY,aAAa;iBAwI1D;EACf;EACA;;EAEA;EACA,QAAQ;EACR,QAAQ;EACR;EACA;;EAEA,MAAM;;wBAGc,oBACpB,OAAO,YACP,cACA,eACC,QAAQ;;;;;;;wBAmCK,mBAAmB,gBAAgB,kBAAkB;;;iBC7PpD;EACf;EACA;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;iBAG5B,0BAA0B;EACzC,UAAU;EACV,WAAW;;wBAGS,eACpB,OAAO,YACP,SAAS,kBACT,SAAS,oBACR,QAAQ;;;iBCFM;EACf;EACA;EACA;;EAEA;EACA;;EAEA;;EAEA;;iBAGe;EACf,UAAU;EACV;EACA;;iBAGe;;EAEf;;EAEA;;;;;EAKA;;EAEA;;wBAGoB,cACpB,OAAO,YACP,UAAS,mBACR,QAAQ"}

package/dist/pipelines/index.js

@@ -1,5 +1,5 @@
-import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "../tool-waste-CJDfxH45.js";
-import { t as compareToBaseline } from "../baseline-BC-eBZ7U.js";
+import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "../tool-waste-B-u3rzCz.js";
+import { t as compareToBaseline } from "../baseline-CVKsfWTg.js";
 import { a as isToolSpan } from "../schema-CSf6qWgZ.js";
 import { a as hasCapturedToolArgs, r as argHash, u as runMetricExtractor } from "../query-DlV0pdx4.js";
 import { t as buildTrajectory } from "../trajectory-D_7rLrvE.js";
@@ -203,7 +203,7 @@ function diffStepsFromSpans(spans) {
 async function regressionView(store, metrics, options) {
 	const baselineRuns = await store.listRuns(options.baseline);
 	const candidateRuns = await store.listRuns(options.candidate);
-	return compareToBaseline(await Promise.all(metrics.map(async (m) => {
+	const samples = await Promise.all(metrics.map(async (m) => {
 		const extract = m.extract ?? runMetricExtractor(m.metric);
 		const baseline = await extractAll(baselineRuns, extract, store);
 		const candidate = await extractAll(candidateRuns, extract, store);
@@ -213,7 +213,8 @@ async function regressionView(store, metrics, options) {
 			baseline,
 			candidate
 		};
-	})), options);
+	}));
+	return compareToBaseline(samples, options);
 }
 async function extractAll(runs, extract, store) {
 	const out = [];

package/dist/pipelines/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/first-divergence.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts"],"sourcesContent":["/**\n * Step diff between two runs, with the first divergence marked.\n *\n * Steps pair in three passes, each over the steps still unpaired:\n *   1. `id`: the same step id in both runs, e.g. a forked run that kept its\n *      parent's span ids.\n *   2. `position`: the same index with the same name and kind.\n *   3. `name`: the same name and kind anywhere, first unpaired step in order.\n * A step left unpaired exists in only one run. The first divergence is the\n * first index at which the two runs stop agreeing step for step, and it says\n * how: a paired step changed, one run replaced, added or removed a step, or the\n * runs ran the same steps in a different order.\n *\n * `diffSteps` reads plain `DiffStep` lists, so the same diff serves agent-eval\n * trace stores (`firstDivergenceView`) and flat OTLP spans\n * (`diffStepsFromSpans`). Use it to attribute \"why is variant B better?\" to a\n * step, and to compare sibling or parent and child runs of a search tree.\n */\n\nimport type { DiagnosisSpan } from '../diagnosis/spans'\nimport type { Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type TrajectoryStep } from '../trajectory'\n\n/** One step of a run, as the diff reads it. */\nexport interface DiffStep {\n  /** The step's id within its run. */\n  id: string\n  name: string\n  kind: string\n  /** Fields compared once two steps pair, such as status, tool or model. */\n  fields: Readonly<Record<string, string | number | boolean | null>>\n}\n\nexport type StepPairing = 'id' | 'position' | 'name'\n\nexport interface FieldDifference {\n  field: string\n  a: string | number | boolean | null\n  b: string | number | boolean | null\n}\n\nexport interface StepPair {\n  /** Index in run A. */\n  a: number\n  /** Index in run B. */\n  b: number\n  pairedBy: StepPairing\n  /** Empty when the paired steps agree on name, kind and every field. */\n  differences: FieldDifference[]\n}\n\nexport type StepDivergenceKind = 'changed' | 'replaced' | 'only-in-a' | 'only-in-b' | 'reordered'\n\nexport interface StepDivergence {\n  /** The first index at which the runs stop agreeing; equals `commonPrefixLen`. */\n  index: number\n  kind: StepDivergenceKind\n  /** Index of the step at `index` in each run, or null when that run has ended. */\n  a: number | null\n  b: number | null\n  /** Set when `kind` is `changed`. */\n  differences: FieldDifference[]\n  reason: string\n}\n\nexport interface StepDiff {\n  /** Paired steps in run A order. */\n  pairs: StepPair[]\n  /** Indexes of run A steps with no partner in run B. */\n  onlyInA: number[]\n  /** Indexes of run B steps with no partner in run A. */\n  onlyInB: number[]\n  /** Leading steps paired index for index with no difference. */\n  commonPrefixLen: number\n  /** Null when the runs agree step for step. */\n  firstDivergence: StepDivergence | null\n}\n\nexport function diffSteps(a: readonly DiffStep[], b: readonly DiffStep[]): StepDiff {\n  const partnerOfA = new Map<number, { b: number; pairedBy: StepPairing }>()\n  const pairedB = new Set<number>()\n  const pair = (i: number, j: number, pairedBy: StepPairing) => {\n    partnerOfA.set(i, { b: j, pairedBy })\n    pairedB.add(j)\n  }\n\n  const bById = new Map<string, number[]>()\n  b.forEach((step, j) => {\n    const list = bById.get(step.id)\n    if (list) list.push(j)\n    else bById.set(step.id, [j])\n  })\n  a.forEach((step, i) => {\n    const j = bById.get(step.id)?.find((candidate) => !pairedB.has(candidate))\n    if (j !== undefined) pair(i, j, 'id')\n  })\n  a.forEach((step, i) => {\n    if (partnerOfA.has(i) || i >= b.length || pairedB.has(i)) return\n    if (sameShape(step, b[i]!)) pair(i, i, 'position')\n  })\n  a.forEach((step, i) => {\n    if (partnerOfA.has(i)) return\n    const j = b.findIndex((candidate, index) => !pairedB.has(index) && sameShape(step, candidate))\n    if (j >= 0) pair(i, j, 'name')\n  })\n\n  const pairs: StepPair[] = [...partnerOfA.entries()]\n    .sort(([x], [y]) => x - y)\n    .map(([i, partner]) => ({\n      a: i,\n      b: partner.b,\n      pairedBy: partner.pairedBy,\n      differences: differencesBetween(a[i]!, b[partner.b]!),\n    }))\n  const pairByA = new Map(pairs.map((entry) => [entry.a, entry]))\n  const partnerOfB = new Map(pairs.map((entry) => [entry.b, entry.a]))\n\n  let prefix = 0\n  while (prefix < a.length && prefix < b.length) {\n    const entry = pairByA.get(prefix)\n    if (!entry || entry.b !== prefix || entry.differences.length > 0) break\n    prefix += 1\n  }\n\n  return {\n    pairs,\n    onlyInA: a.map((_, i) => i).filter((i) => !partnerOfA.has(i)),\n    onlyInB: b.map((_, j) => j).filter((j) => !pairedB.has(j)),\n    commonPrefixLen: prefix,\n    firstDivergence:\n      prefix === a.length && prefix === b.length\n        ? null\n        : divergenceAt(prefix, a, b, pairByA.get(prefix), partnerOfB),\n  }\n}\n\nfunction divergenceAt(\n  index: number,\n  a: readonly DiffStep[],\n  b: readonly DiffStep[],\n  pairOfA: StepPair | undefined,\n  partnerOfB: ReadonlyMap<number, number>,\n): StepDivergence {\n  const stepA = a[index]\n  const stepB = b[index]\n  const base = {\n    index,\n    a: stepA ? index : null,\n    b: stepB ? index : null,\n    differences: [] as FieldDifference[],\n  }\n  if (stepA && stepB && pairOfA?.b === index) {\n    return {\n      ...base,\n      kind: 'changed',\n      differences: pairOfA.differences,\n      reason: `at index ${index}, ${describe(stepA)} changed: ${pairOfA.differences\n        .map((d) => `${d.field} ${String(d.a)} vs ${String(d.b)}`)\n        .join('; ')}`,\n    }\n  }\n  const aUnpaired = stepA !== undefined && pairOfA === undefined\n  const bUnpaired = stepB !== undefined && !partnerOfB.has(index)\n  if (stepA && stepB && aUnpaired && bUnpaired) {\n    return {\n      ...base,\n      kind: 'replaced',\n      reason: `at index ${index}, A ran ${describe(stepA)} and B ran ${describe(stepB)}`,\n    }\n  }\n  if (stepA && aUnpaired) {\n    return {\n      ...base,\n      kind: 'only-in-a',\n      reason: `at index ${index}, only A has ${describe(stepA)}`,\n    }\n  }\n  if (stepB && bUnpaired) {\n    return {\n      ...base,\n      kind: 'only-in-b',\n      reason: `at index ${index}, only B has ${describe(stepB)}`,\n    }\n  }\n  return {\n    ...base,\n    kind: 'reordered',\n    reason: `at index ${index}, A ran ${describe(stepA!)} and B ran ${describe(stepB!)}; both steps occur in the other run at a different position`,\n  }\n}\n\nfunction sameShape(x: DiffStep, y: DiffStep): boolean {\n  return x.name === y.name && x.kind === y.kind\n}\n\nfunction differencesBetween(x: DiffStep, y: DiffStep): FieldDifference[] {\n  const differences: FieldDifference[] = []\n  if (x.kind !== y.kind) differences.push({ field: 'kind', a: x.kind, b: y.kind })\n  if (x.name !== y.name) differences.push({ field: 'name', a: x.name, b: y.name })\n  const fields = [...new Set([...Object.keys(x.fields), ...Object.keys(y.fields)])].sort()\n  for (const field of fields) {\n    const left = x.fields[field] ?? null\n    const right = y.fields[field] ?? null\n    if (left !== right) differences.push({ field, a: left, b: right })\n  }\n  return differences\n}\n\nfunction describe(step: DiffStep): string {\n  return `${step.kind} \"${step.name}\"`\n}\n\n// ── agent-eval trace store ────────────────────────────────────────────\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  /** Null when the runs agree step for step. */\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  commonPrefixLen: number\n  /** The step diff behind the first divergence, over trajectory step indexes. */\n  diff: StepDiff\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const diff = diffSteps(a.steps.map(trajectoryDiffStep), b.steps.map(trajectoryDiffStep))\n  const divergence = diff.firstDivergence\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: divergence?.index ?? null,\n    ...(divergence?.a != null ? { aStep: a.steps[divergence.a] } : {}),\n    ...(divergence?.b != null ? { bStep: b.steps[divergence.b] } : {}),\n    ...(divergence ? { reason: divergence.reason } : {}),\n    commonPrefixLen: diff.commonPrefixLen,\n    diff,\n  }\n}\n\nfunction trajectoryDiffStep(step: TrajectoryStep): DiffStep {\n  const span: Span = step.span\n  const fields: Record<string, string | number | boolean | null> = {\n    status: span.status ?? null,\n  }\n  if (span.kind === 'tool') fields.tool = span.toolName\n  if (span.kind === 'llm') fields.model = span.model\n  if (span.kind === 'judge') fields.dimension = span.dimension\n  return { id: span.spanId, name: span.name, kind: span.kind, fields }\n}\n\n// ── flat OTLP spans ───────────────────────────────────────────────────\n\n/**\n * One run's spans as diff steps, in containment order: depth first, siblings\n * by start time, the same order `buildTrajectory` gives a trace store run.\n * Spans whose parent is absent, or whose parents form a cycle, start their own\n * subtree, so every span appears exactly once.\n */\nexport function diffStepsFromSpans(spans: readonly DiagnosisSpan[]): DiffStep[] {\n  const ids = new Set(spans.map((span) => span.spanId))\n  const children = new Map<string | null, DiagnosisSpan[]>()\n  for (const span of spans) {\n    const parent =\n      span.parentSpanId !== null && ids.has(span.parentSpanId) ? span.parentSpanId : null\n    const list = children.get(parent)\n    if (list) list.push(span)\n    else children.set(parent, [span])\n  }\n  // Span ids are random per run, so two runs of the same logical trace can carry\n  // different ids for \"the same\" span. Siblings that share a start time (a common\n  // case: parallel steps a driver starts together) must still sort the same way\n  // in both runs, or the diff below reports a false `reordered` divergence purely\n  // from spanId noise. End time, kind and name are content, not per-run identity,\n  // so they break the tie before spanId, which is the last resort.\n  const byStart = (x: DiagnosisSpan, y: DiagnosisSpan) =>\n    (x.startMs ?? 0) - (y.startMs ?? 0) ||\n    (x.endMs ?? 0) - (y.endMs ?? 0) ||\n    x.kind.localeCompare(y.kind) ||\n    x.name.localeCompare(y.name) ||\n    x.spanId.localeCompare(y.spanId)\n  for (const list of children.values()) list.sort(byStart)\n  const ordered: DiffStep[] = []\n  const visited = new Set<string>()\n  const visit = (span: DiagnosisSpan) => {\n    if (visited.has(span.spanId)) return\n    visited.add(span.spanId)\n    ordered.push({\n      id: span.spanId,\n      name: span.name,\n      kind: span.kind,\n      fields: { status: span.status, tool: span.toolName, model: span.model },\n    })\n    for (const child of children.get(span.spanId) ?? []) visit(child)\n  }\n  for (const root of children.get(null) ?? []) visit(root)\n  for (const span of [...spans].sort(byStart)) visit(span)\n  return ordered\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { runMetricExtractor } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Omit it and `metric` must name one of\n   *  `RUN_METRICS`; any other name is refused. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? runMetricExtractor(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n"],"mappings":";;;;;;;AA+EA,SAAgB,UAAU,GAAwB,GAAkC;CAClF,MAAM,6BAAa,IAAI,IAAkD;CACzE,MAAM,0BAAU,IAAI,IAAY;CAChC,MAAM,QAAQ,GAAW,GAAW,aAA0B;EAC5D,WAAW,IAAI,GAAG;GAAE,GAAG;GAAG;EAAS,CAAC;EACpC,QAAQ,IAAI,CAAC;CACf;CAEA,MAAM,wBAAQ,IAAI,IAAsB;CACxC,EAAE,SAAS,MAAM,MAAM;EACrB,MAAM,OAAO,MAAM,IAAI,KAAK,EAAE;EAC9B,IAAI,MAAM,KAAK,KAAK,CAAC;OAChB,MAAM,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC;CAC7B,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,MAAM,IAAI,MAAM,IAAI,KAAK,EAAE,CAAC,EAAE,MAAM,cAAc,CAAC,QAAQ,IAAI,SAAS,CAAC;EACzE,IAAI,MAAM,KAAA,GAAW,KAAK,GAAG,GAAG,IAAI;CACtC,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,IAAI,WAAW,IAAI,CAAC,KAAK,KAAK,EAAE,UAAU,QAAQ,IAAI,CAAC,GAAG;EAC1D,IAAI,UAAU,MAAM,EAAE,EAAG,GAAG,KAAK,GAAG,GAAG,UAAU;CACnD,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,IAAI,WAAW,IAAI,CAAC,GAAG;EACvB,MAAM,IAAI,EAAE,WAAW,WAAW,UAAU,CAAC,QAAQ,IAAI,KAAK,KAAK,UAAU,MAAM,SAAS,CAAC;EAC7F,IAAI,KAAK,GAAG,KAAK,GAAG,GAAG,MAAM;CAC/B,CAAC;CAED,MAAM,QAAoB,CAAC,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChD,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC,CACzB,KAAK,CAAC,GAAG,cAAc;EACtB,GAAG;EACH,GAAG,QAAQ;EACX,UAAU,QAAQ;EAClB,aAAa,mBAAmB,EAAE,IAAK,EAAE,QAAQ,EAAG;CACtD,EAAE;CACJ,MAAM,UAAU,IAAI,IAAI,MAAM,KAAK,UAAU,CAAC,MAAM,GAAG,KAAK,CAAC,CAAC;CAC9D,MAAM,aAAa,IAAI,IAAI,MAAM,KAAK,UAAU,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC;CAEnE,IAAI,SAAS;CACb,OAAO,SAAS,EAAE,UAAU,SAAS,EAAE,QAAQ;EAC7C,MAAM,QAAQ,QAAQ,IAAI,MAAM;EAChC,IAAI,CAAC,SAAS,MAAM,MAAM,UAAU,MAAM,YAAY,SAAS,GAAG;EAClE,UAAU;CACZ;CAEA,OAAO;EACL;EACA,SAAS,EAAE,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,MAAM,CAAC,WAAW,IAAI,CAAC,CAAC;EAC5D,SAAS,EAAE,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,MAAM,CAAC,QAAQ,IAAI,CAAC,CAAC;EACzD,iBAAiB;EACjB,iBACE,WAAW,EAAE,UAAU,WAAW,EAAE,SAChC,OACA,aAAa,QAAQ,GAAG,GAAG,QAAQ,IAAI,MAAM,GAAG,UAAU;CAClE;AACF;AAEA,SAAS,aACP,OACA,GACA,GACA,SACA,YACgB;CAChB,MAAM,QAAQ,EAAE;CAChB,MAAM,QAAQ,EAAE;CAChB,MAAM,OAAO;EACX;EACA,GAAG,QAAQ,QAAQ;EACnB,GAAG,QAAQ,QAAQ;EACnB,aAAa,CAAC;CAChB;CACA,IAAI,SAAS,SAAS,SAAS,MAAM,OACnC,OAAO;EACL,GAAG;EACH,MAAM;EACN,aAAa,QAAQ;EACrB,QAAQ,YAAY,MAAM,IAAI,SAAS,KAAK,EAAE,YAAY,QAAQ,YAC/D,KAAK,MAAM,GAAG,EAAE,MAAM,GAAG,OAAO,EAAE,CAAC,EAAE,MAAM,OAAO,EAAE,CAAC,GAAG,CAAC,CACzD,KAAK,IAAI;CACd;CAEF,MAAM,YAAY,UAAU,KAAA,KAAa,YAAY,KAAA;CACrD,MAAM,YAAY,UAAU,KAAA,KAAa,CAAC,WAAW,IAAI,KAAK;CAC9D,IAAI,SAAS,SAAS,aAAa,WACjC,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,UAAU,SAAS,KAAK,EAAE,aAAa,SAAS,KAAK;CACjF;CAEF,IAAI,SAAS,WACX,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,eAAe,SAAS,KAAK;CACzD;CAEF,IAAI,SAAS,WACX,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,eAAe,SAAS,KAAK;CACzD;CAEF,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,UAAU,SAAS,KAAM,EAAE,aAAa,SAAS,KAAM,EAAE;CACrF;AACF;AAEA,SAAS,UAAU,GAAa,GAAsB;CACpD,OAAO,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE;AAC3C;AAEA,SAAS,mBAAmB,GAAa,GAAgC;CACvE,MAAM,cAAiC,CAAC;CACxC,IAAI,EAAE,SAAS,EAAE,MAAM,YAAY,KAAK;EAAE,OAAO;EAAQ,GAAG,EAAE;EAAM,GAAG,EAAE;CAAK,CAAC;CAC/E,IAAI,EAAE,SAAS,EAAE,MAAM,YAAY,KAAK;EAAE,OAAO;EAAQ,GAAG,EAAE;EAAM,GAAG,EAAE;CAAK,CAAC;CAC/E,MAAM,SAAS,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,EAAE,MAAM,GAAG,GAAG,OAAO,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,KAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,OAAO,EAAE,OAAO,UAAU;EAChC,MAAM,QAAQ,EAAE,OAAO,UAAU;EACjC,IAAI,SAAS,OAAO,YAAY,KAAK;GAAE;GAAO,GAAG;GAAM,GAAG;EAAM,CAAC;CACnE;CACA,OAAO;AACT;AAEA,SAAS,SAAS,MAAwB;CACxC,OAAO,GAAG,KAAK,KAAK,IAAI,KAAK,KAAK;AACpC;AAiBA,eAAsB,oBACpB,OACA,MACA,MAC2B;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,OAAO,UAAU,EAAE,MAAM,IAAI,kBAAkB,GAAG,EAAE,MAAM,IAAI,kBAAkB,CAAC;CACvF,MAAM,aAAa,KAAK;CACxB,OAAO;EACL;EACA;EACA,sBAAsB,YAAY,SAAS;EAC3C,GAAI,YAAY,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,GAAG,IAAI,CAAC;EAChE,GAAI,YAAY,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,GAAG,IAAI,CAAC;EAChE,GAAI,aAAa,EAAE,QAAQ,WAAW,OAAO,IAAI,CAAC;EAClD,iBAAiB,KAAK;EACtB;CACF;AACF;AAEA,SAAS,mBAAmB,MAAgC;CAC1D,MAAM,OAAa,KAAK;CACxB,MAAM,SAA2D,EAC/D,QAAQ,KAAK,UAAU,KACzB;CACA,IAAI,KAAK,SAAS,QAAQ,OAAO,OAAO,KAAK;CAC7C,IAAI,KAAK,SAAS,OAAO,OAAO,QAAQ,KAAK;CAC7C,IAAI,KAAK,SAAS,SAAS,OAAO,YAAY,KAAK;CACnD,OAAO;EAAE,IAAI,KAAK;EAAQ,MAAM,KAAK;EAAM,MAAM,KAAK;EAAM;CAAO;AACrE;;;;;;;AAUA,SAAgB,mBAAmB,OAA6C;CAC9E,MAAM,MAAM,IAAI,IAAI,MAAM,KAAK,SAAS,KAAK,MAAM,CAAC;CACpD,MAAM,2BAAW,IAAI,IAAoC;CACzD,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SACJ,KAAK,iBAAiB,QAAQ,IAAI,IAAI,KAAK,YAAY,IAAI,KAAK,eAAe;EACjF,MAAM,OAAO,SAAS,IAAI,MAAM;EAChC,IAAI,MAAM,KAAK,KAAK,IAAI;OACnB,SAAS,IAAI,QAAQ,CAAC,IAAI,CAAC;CAClC;CAOA,MAAM,WAAW,GAAkB,OAChC,EAAE,WAAW,MAAM,EAAE,WAAW,OAChC,EAAE,SAAS,MAAM,EAAE,SAAS,MAC7B,EAAE,KAAK,cAAc,EAAE,IAAI,KAC3B,EAAE,KAAK,cAAc,EAAE,IAAI,KAC3B,EAAE,OAAO,cAAc,EAAE,MAAM;CACjC,KAAK,MAAM,QAAQ,SAAS,OAAO,GAAG,KAAK,KAAK,OAAO;CACvD,MAAM,UAAsB,CAAC;CAC7B,MAAM,0BAAU,IAAI,IAAY;CAChC,MAAM,SAAS,SAAwB;EACrC,IAAI,QAAQ,IAAI,KAAK,MAAM,GAAG;EAC9B,QAAQ,IAAI,KAAK,MAAM;EACvB,QAAQ,KAAK;GACX,IAAI,KAAK;GACT,MAAM,KAAK;GACX,MAAM,KAAK;GACX,QAAQ;IAAE,QAAQ,KAAK;IAAQ,MAAM,KAAK;IAAU,OAAO,KAAK;GAAM;EACxE,CAAC;EACD,KAAK,MAAM,SAAS,SAAS,IAAI,KAAK,MAAM,KAAK,CAAC,GAAG,MAAM,KAAK;CAClE;CACA,KAAK,MAAM,QAAQ,SAAS,IAAI,IAAI,KAAK,CAAC,GAAG,MAAM,IAAI;CACvD,KAAK,MAAM,QAAQ,CAAC,GAAG,KAAK,CAAC,CAAC,KAAK,OAAO,GAAG,MAAM,IAAI;CACvD,OAAO;AACT;;;;;;;;;;;ACvRA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAS5D,OAAO,kBAAkB,MARH,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,mBAAmB,EAAE,MAAM;EACxD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH,GACkC,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;;;;;;;;;;;;ACzCA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/first-divergence.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts"],"sourcesContent":["/**\n * Step diff between two runs, with the first divergence marked.\n *\n * Steps pair in three passes, each over the steps still unpaired:\n *   1. `id`: the same step id in both runs, e.g. a forked run that kept its\n *      parent's span ids.\n *   2. `position`: the same index with the same name and kind.\n *   3. `name`: the same name and kind anywhere, first unpaired step in order.\n * A step left unpaired exists in only one run. The first divergence is the\n * first index at which the two runs stop agreeing step for step, and it says\n * how: a paired step changed, one run replaced, added or removed a step, or the\n * runs ran the same steps in a different order.\n *\n * `diffSteps` reads plain `DiffStep` lists, so the same diff serves agent-eval\n * trace stores (`firstDivergenceView`) and flat OTLP spans\n * (`diffStepsFromSpans`). Use it to attribute \"why is variant B better?\" to a\n * step, and to compare sibling or parent and child runs of a search tree.\n */\n\nimport type { DiagnosisSpan } from '../diagnosis/spans'\nimport type { Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type TrajectoryStep } from '../trajectory'\n\n/** One step of a run, as the diff reads it. */\nexport interface DiffStep {\n  /** The step's id within its run. */\n  id: string\n  name: string\n  kind: string\n  /** Fields compared once two steps pair, such as status, tool or model. */\n  fields: Readonly<Record<string, string | number | boolean | null>>\n}\n\nexport type StepPairing = 'id' | 'position' | 'name'\n\nexport interface FieldDifference {\n  field: string\n  a: string | number | boolean | null\n  b: string | number | boolean | null\n}\n\nexport interface StepPair {\n  /** Index in run A. */\n  a: number\n  /** Index in run B. */\n  b: number\n  pairedBy: StepPairing\n  /** Empty when the paired steps agree on name, kind and every field. */\n  differences: FieldDifference[]\n}\n\nexport type StepDivergenceKind = 'changed' | 'replaced' | 'only-in-a' | 'only-in-b' | 'reordered'\n\nexport interface StepDivergence {\n  /** The first index at which the runs stop agreeing; equals `commonPrefixLen`. */\n  index: number\n  kind: StepDivergenceKind\n  /** Index of the step at `index` in each run, or null when that run has ended. */\n  a: number | null\n  b: number | null\n  /** Set when `kind` is `changed`. */\n  differences: FieldDifference[]\n  reason: string\n}\n\nexport interface StepDiff {\n  /** Paired steps in run A order. */\n  pairs: StepPair[]\n  /** Indexes of run A steps with no partner in run B. */\n  onlyInA: number[]\n  /** Indexes of run B steps with no partner in run A. */\n  onlyInB: number[]\n  /** Leading steps paired index for index with no difference. */\n  commonPrefixLen: number\n  /** Null when the runs agree step for step. */\n  firstDivergence: StepDivergence | null\n}\n\nexport function diffSteps(a: readonly DiffStep[], b: readonly DiffStep[]): StepDiff {\n  const partnerOfA = new Map<number, { b: number; pairedBy: StepPairing }>()\n  const pairedB = new Set<number>()\n  const pair = (i: number, j: number, pairedBy: StepPairing) => {\n    partnerOfA.set(i, { b: j, pairedBy })\n    pairedB.add(j)\n  }\n\n  const bById = new Map<string, number[]>()\n  b.forEach((step, j) => {\n    const list = bById.get(step.id)\n    if (list) list.push(j)\n    else bById.set(step.id, [j])\n  })\n  a.forEach((step, i) => {\n    const j = bById.get(step.id)?.find((candidate) => !pairedB.has(candidate))\n    if (j !== undefined) pair(i, j, 'id')\n  })\n  a.forEach((step, i) => {\n    if (partnerOfA.has(i) || i >= b.length || pairedB.has(i)) return\n    if (sameShape(step, b[i]!)) pair(i, i, 'position')\n  })\n  a.forEach((step, i) => {\n    if (partnerOfA.has(i)) return\n    const j = b.findIndex((candidate, index) => !pairedB.has(index) && sameShape(step, candidate))\n    if (j >= 0) pair(i, j, 'name')\n  })\n\n  const pairs: StepPair[] = [...partnerOfA.entries()]\n    .sort(([x], [y]) => x - y)\n    .map(([i, partner]) => ({\n      a: i,\n      b: partner.b,\n      pairedBy: partner.pairedBy,\n      differences: differencesBetween(a[i]!, b[partner.b]!),\n    }))\n  const pairByA = new Map(pairs.map((entry) => [entry.a, entry]))\n  const partnerOfB = new Map(pairs.map((entry) => [entry.b, entry.a]))\n\n  let prefix = 0\n  while (prefix < a.length && prefix < b.length) {\n    const entry = pairByA.get(prefix)\n    if (!entry || entry.b !== prefix || entry.differences.length > 0) break\n    prefix += 1\n  }\n\n  return {\n    pairs,\n    onlyInA: a.map((_, i) => i).filter((i) => !partnerOfA.has(i)),\n    onlyInB: b.map((_, j) => j).filter((j) => !pairedB.has(j)),\n    commonPrefixLen: prefix,\n    firstDivergence:\n      prefix === a.length && prefix === b.length\n        ? null\n        : divergenceAt(prefix, a, b, pairByA.get(prefix), partnerOfB),\n  }\n}\n\nfunction divergenceAt(\n  index: number,\n  a: readonly DiffStep[],\n  b: readonly DiffStep[],\n  pairOfA: StepPair | undefined,\n  partnerOfB: ReadonlyMap<number, number>,\n): StepDivergence {\n  const stepA = a[index]\n  const stepB = b[index]\n  const base = {\n    index,\n    a: stepA ? index : null,\n    b: stepB ? index : null,\n    differences: [] as FieldDifference[],\n  }\n  if (stepA && stepB && pairOfA?.b === index) {\n    return {\n      ...base,\n      kind: 'changed',\n      differences: pairOfA.differences,\n      reason: `at index ${index}, ${describe(stepA)} changed: ${pairOfA.differences\n        .map((d) => `${d.field} ${String(d.a)} vs ${String(d.b)}`)\n        .join('; ')}`,\n    }\n  }\n  const aUnpaired = stepA !== undefined && pairOfA === undefined\n  const bUnpaired = stepB !== undefined && !partnerOfB.has(index)\n  if (stepA && stepB && aUnpaired && bUnpaired) {\n    return {\n      ...base,\n      kind: 'replaced',\n      reason: `at index ${index}, A ran ${describe(stepA)} and B ran ${describe(stepB)}`,\n    }\n  }\n  if (stepA && aUnpaired) {\n    return {\n      ...base,\n      kind: 'only-in-a',\n      reason: `at index ${index}, only A has ${describe(stepA)}`,\n    }\n  }\n  if (stepB && bUnpaired) {\n    return {\n      ...base,\n      kind: 'only-in-b',\n      reason: `at index ${index}, only B has ${describe(stepB)}`,\n    }\n  }\n  return {\n    ...base,\n    kind: 'reordered',\n    reason: `at index ${index}, A ran ${describe(stepA!)} and B ran ${describe(stepB!)}; both steps occur in the other run at a different position`,\n  }\n}\n\nfunction sameShape(x: DiffStep, y: DiffStep): boolean {\n  return x.name === y.name && x.kind === y.kind\n}\n\nfunction differencesBetween(x: DiffStep, y: DiffStep): FieldDifference[] {\n  const differences: FieldDifference[] = []\n  if (x.kind !== y.kind) differences.push({ field: 'kind', a: x.kind, b: y.kind })\n  if (x.name !== y.name) differences.push({ field: 'name', a: x.name, b: y.name })\n  const fields = [...new Set([...Object.keys(x.fields), ...Object.keys(y.fields)])].sort()\n  for (const field of fields) {\n    const left = x.fields[field] ?? null\n    const right = y.fields[field] ?? null\n    if (left !== right) differences.push({ field, a: left, b: right })\n  }\n  return differences\n}\n\nfunction describe(step: DiffStep): string {\n  return `${step.kind} \"${step.name}\"`\n}\n\n// ── agent-eval trace store ────────────────────────────────────────────\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  /** Null when the runs agree step for step. */\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  commonPrefixLen: number\n  /** The step diff behind the first divergence, over trajectory step indexes. */\n  diff: StepDiff\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const diff = diffSteps(a.steps.map(trajectoryDiffStep), b.steps.map(trajectoryDiffStep))\n  const divergence = diff.firstDivergence\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: divergence?.index ?? null,\n    ...(divergence?.a != null ? { aStep: a.steps[divergence.a] } : {}),\n    ...(divergence?.b != null ? { bStep: b.steps[divergence.b] } : {}),\n    ...(divergence ? { reason: divergence.reason } : {}),\n    commonPrefixLen: diff.commonPrefixLen,\n    diff,\n  }\n}\n\nfunction trajectoryDiffStep(step: TrajectoryStep): DiffStep {\n  const span: Span = step.span\n  const fields: Record<string, string | number | boolean | null> = {\n    status: span.status ?? null,\n  }\n  if (span.kind === 'tool') fields.tool = span.toolName\n  if (span.kind === 'llm') fields.model = span.model\n  if (span.kind === 'judge') fields.dimension = span.dimension\n  return { id: span.spanId, name: span.name, kind: span.kind, fields }\n}\n\n// ── flat OTLP spans ───────────────────────────────────────────────────\n\n/**\n * One run's spans as diff steps, in containment order: depth first, siblings\n * by start time, the same order `buildTrajectory` gives a trace store run.\n * Spans whose parent is absent, or whose parents form a cycle, start their own\n * subtree, so every span appears exactly once.\n */\nexport function diffStepsFromSpans(spans: readonly DiagnosisSpan[]): DiffStep[] {\n  const ids = new Set(spans.map((span) => span.spanId))\n  const children = new Map<string | null, DiagnosisSpan[]>()\n  for (const span of spans) {\n    const parent =\n      span.parentSpanId !== null && ids.has(span.parentSpanId) ? span.parentSpanId : null\n    const list = children.get(parent)\n    if (list) list.push(span)\n    else children.set(parent, [span])\n  }\n  // Span ids are random per run, so two runs of the same logical trace can carry\n  // different ids for \"the same\" span. Siblings that share a start time (a common\n  // case: parallel steps a driver starts together) must still sort the same way\n  // in both runs, or the diff below reports a false `reordered` divergence purely\n  // from spanId noise. End time, kind and name are content, not per-run identity,\n  // so they break the tie before spanId, which is the last resort.\n  const byStart = (x: DiagnosisSpan, y: DiagnosisSpan) =>\n    (x.startMs ?? 0) - (y.startMs ?? 0) ||\n    (x.endMs ?? 0) - (y.endMs ?? 0) ||\n    x.kind.localeCompare(y.kind) ||\n    x.name.localeCompare(y.name) ||\n    x.spanId.localeCompare(y.spanId)\n  for (const list of children.values()) list.sort(byStart)\n  const ordered: DiffStep[] = []\n  const visited = new Set<string>()\n  const visit = (span: DiagnosisSpan) => {\n    if (visited.has(span.spanId)) return\n    visited.add(span.spanId)\n    ordered.push({\n      id: span.spanId,\n      name: span.name,\n      kind: span.kind,\n      fields: { status: span.status, tool: span.toolName, model: span.model },\n    })\n    for (const child of children.get(span.spanId) ?? []) visit(child)\n  }\n  for (const root of children.get(null) ?? []) visit(root)\n  for (const span of [...spans].sort(byStart)) visit(span)\n  return ordered\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { runMetricExtractor } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Omit it and `metric` must name one of\n   *  `RUN_METRICS`; any other name is refused. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? runMetricExtractor(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n"],"mappings":";;;;;;;AA+EA,SAAgB,UAAU,GAAwB,GAAkC;CAClF,MAAM,6BAAa,IAAI,IAAkD;CACzE,MAAM,0BAAU,IAAI,IAAY;CAChC,MAAM,QAAQ,GAAW,GAAW,aAA0B;EAC5D,WAAW,IAAI,GAAG;GAAE,GAAG;GAAG;EAAS,CAAC;EACpC,QAAQ,IAAI,CAAC;CACf;CAEA,MAAM,wBAAQ,IAAI,IAAsB;CACxC,EAAE,SAAS,MAAM,MAAM;EACrB,MAAM,OAAO,MAAM,IAAI,KAAK,EAAE;EAC9B,IAAI,MAAM,KAAK,KAAK,CAAC;OAChB,MAAM,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC;CAC7B,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,MAAM,IAAI,MAAM,IAAI,KAAK,EAAE,CAAC,EAAE,MAAM,cAAc,CAAC,QAAQ,IAAI,SAAS,CAAC;EACzE,IAAI,MAAM,KAAA,GAAW,KAAK,GAAG,GAAG,IAAI;CACtC,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,IAAI,WAAW,IAAI,CAAC,KAAK,KAAK,EAAE,UAAU,QAAQ,IAAI,CAAC,GAAG;EAC1D,IAAI,UAAU,MAAM,EAAE,EAAG,GAAG,KAAK,GAAG,GAAG,UAAU;CACnD,CAAC;CACD,EAAE,SAAS,MAAM,MAAM;EACrB,IAAI,WAAW,IAAI,CAAC,GAAG;EACvB,MAAM,IAAI,EAAE,WAAW,WAAW,UAAU,CAAC,QAAQ,IAAI,KAAK,KAAK,UAAU,MAAM,SAAS,CAAC;EAC7F,IAAI,KAAK,GAAG,KAAK,GAAG,GAAG,MAAM;CAC/B,CAAC;CAED,MAAM,QAAoB,CAAC,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChD,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC,CACzB,KAAK,CAAC,GAAG,cAAc;EACtB,GAAG;EACH,GAAG,QAAQ;EACX,UAAU,QAAQ;EAClB,aAAa,mBAAmB,EAAE,IAAK,EAAE,QAAQ,EAAG;CACtD,EAAE;CACJ,MAAM,UAAU,IAAI,IAAI,MAAM,KAAK,UAAU,CAAC,MAAM,GAAG,KAAK,CAAC,CAAC;CAC9D,MAAM,aAAa,IAAI,IAAI,MAAM,KAAK,UAAU,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC;CAEnE,IAAI,SAAS;CACb,OAAO,SAAS,EAAE,UAAU,SAAS,EAAE,QAAQ;EAC7C,MAAM,QAAQ,QAAQ,IAAI,MAAM;EAChC,IAAI,CAAC,SAAS,MAAM,MAAM,UAAU,MAAM,YAAY,SAAS,GAAG;EAClE,UAAU;CACZ;CAEA,OAAO;EACL;EACA,SAAS,EAAE,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,MAAM,CAAC,WAAW,IAAI,CAAC,CAAC;EAC5D,SAAS,EAAE,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,MAAM,CAAC,QAAQ,IAAI,CAAC,CAAC;EACzD,iBAAiB;EACjB,iBACE,WAAW,EAAE,UAAU,WAAW,EAAE,SAChC,OACA,aAAa,QAAQ,GAAG,GAAG,QAAQ,IAAI,MAAM,GAAG,UAAU;CAClE;AACF;AAEA,SAAS,aACP,OACA,GACA,GACA,SACA,YACgB;CAChB,MAAM,QAAQ,EAAE;CAChB,MAAM,QAAQ,EAAE;CAChB,MAAM,OAAO;EACX;EACA,GAAG,QAAQ,QAAQ;EACnB,GAAG,QAAQ,QAAQ;EACnB,aAAa,CAAC;CAChB;CACA,IAAI,SAAS,SAAS,SAAS,MAAM,OACnC,OAAO;EACL,GAAG;EACH,MAAM;EACN,aAAa,QAAQ;EACrB,QAAQ,YAAY,MAAM,IAAI,SAAS,KAAK,EAAE,YAAY,QAAQ,YAC/D,KAAK,MAAM,GAAG,EAAE,MAAM,GAAG,OAAO,EAAE,CAAC,EAAE,MAAM,OAAO,EAAE,CAAC,GAAG,CAAC,CACzD,KAAK,IAAI;CACd;CAEF,MAAM,YAAY,UAAU,KAAA,KAAa,YAAY,KAAA;CACrD,MAAM,YAAY,UAAU,KAAA,KAAa,CAAC,WAAW,IAAI,KAAK;CAC9D,IAAI,SAAS,SAAS,aAAa,WACjC,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,UAAU,SAAS,KAAK,EAAE,aAAa,SAAS,KAAK;CACjF;CAEF,IAAI,SAAS,WACX,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,eAAe,SAAS,KAAK;CACzD;CAEF,IAAI,SAAS,WACX,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,eAAe,SAAS,KAAK;CACzD;CAEF,OAAO;EACL,GAAG;EACH,MAAM;EACN,QAAQ,YAAY,MAAM,UAAU,SAAS,KAAM,EAAE,aAAa,SAAS,KAAM,EAAE;CACrF;AACF;AAEA,SAAS,UAAU,GAAa,GAAsB;CACpD,OAAO,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE;AAC3C;AAEA,SAAS,mBAAmB,GAAa,GAAgC;CACvE,MAAM,cAAiC,CAAC;CACxC,IAAI,EAAE,SAAS,EAAE,MAAM,YAAY,KAAK;EAAE,OAAO;EAAQ,GAAG,EAAE;EAAM,GAAG,EAAE;CAAK,CAAC;CAC/E,IAAI,EAAE,SAAS,EAAE,MAAM,YAAY,KAAK;EAAE,OAAO;EAAQ,GAAG,EAAE;EAAM,GAAG,EAAE;CAAK,CAAC;CAC/E,MAAM,SAAS,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,EAAE,MAAM,GAAG,GAAG,OAAO,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,KAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,OAAO,EAAE,OAAO,UAAU;EAChC,MAAM,QAAQ,EAAE,OAAO,UAAU;EACjC,IAAI,SAAS,OAAO,YAAY,KAAK;GAAE;GAAO,GAAG;GAAM,GAAG;EAAM,CAAC;CACnE;CACA,OAAO;AACT;AAEA,SAAS,SAAS,MAAwB;CACxC,OAAO,GAAG,KAAK,KAAK,IAAI,KAAK,KAAK;AACpC;AAiBA,eAAsB,oBACpB,OACA,MACA,MAC2B;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,OAAO,UAAU,EAAE,MAAM,IAAI,kBAAkB,GAAG,EAAE,MAAM,IAAI,kBAAkB,CAAC;CACvF,MAAM,aAAa,KAAK;CACxB,OAAO;EACL;EACA;EACA,sBAAsB,YAAY,SAAS;EAC3C,GAAI,YAAY,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,GAAG,IAAI,CAAC;EAChE,GAAI,YAAY,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,GAAG,IAAI,CAAC;EAChE,GAAI,aAAa,EAAE,QAAQ,WAAW,OAAO,IAAI,CAAC;EAClD,iBAAiB,KAAK;EACtB;CACF;AACF;AAEA,SAAS,mBAAmB,MAAgC;CAC1D,MAAM,OAAa,KAAK;CACxB,MAAM,SAA2D,EAC/D,QAAQ,KAAK,UAAU,KACzB;CACA,IAAI,KAAK,SAAS,QAAQ,OAAO,OAAO,KAAK;CAC7C,IAAI,KAAK,SAAS,OAAO,OAAO,QAAQ,KAAK;CAC7C,IAAI,KAAK,SAAS,SAAS,OAAO,YAAY,KAAK;CACnD,OAAO;EAAE,IAAI,KAAK;EAAQ,MAAM,KAAK;EAAM,MAAM,KAAK;EAAM;CAAO;AACrE;;;;;;;AAUA,SAAgB,mBAAmB,OAA6C;CAC9E,MAAM,MAAM,IAAI,IAAI,MAAM,KAAK,SAAS,KAAK,MAAM,CAAC;CACpD,MAAM,2BAAW,IAAI,IAAoC;CACzD,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SACJ,KAAK,iBAAiB,QAAQ,IAAI,IAAI,KAAK,YAAY,IAAI,KAAK,eAAe;EACjF,MAAM,OAAO,SAAS,IAAI,MAAM;EAChC,IAAI,MAAM,KAAK,KAAK,IAAI;OACnB,SAAS,IAAI,QAAQ,CAAC,IAAI,CAAC;CAClC;CAOA,MAAM,WAAW,GAAkB,OAChC,EAAE,WAAW,MAAM,EAAE,WAAW,OAChC,EAAE,SAAS,MAAM,EAAE,SAAS,MAC7B,EAAE,KAAK,cAAc,EAAE,IAAI,KAC3B,EAAE,KAAK,cAAc,EAAE,IAAI,KAC3B,EAAE,OAAO,cAAc,EAAE,MAAM;CACjC,KAAK,MAAM,QAAQ,SAAS,OAAO,GAAG,KAAK,KAAK,OAAO;CACvD,MAAM,UAAsB,CAAC;CAC7B,MAAM,0BAAU,IAAI,IAAY;CAChC,MAAM,SAAS,SAAwB;EACrC,IAAI,QAAQ,IAAI,KAAK,MAAM,GAAG;EAC9B,QAAQ,IAAI,KAAK,MAAM;EACvB,QAAQ,KAAK;GACX,IAAI,KAAK;GACT,MAAM,KAAK;GACX,MAAM,KAAK;GACX,QAAQ;IAAE,QAAQ,KAAK;IAAQ,MAAM,KAAK;IAAU,OAAO,KAAK;GAAM;EACxE,CAAC;EACD,KAAK,MAAM,SAAS,SAAS,IAAI,KAAK,MAAM,KAAK,CAAC,GAAG,MAAM,KAAK;CAClE;CACA,KAAK,MAAM,QAAQ,SAAS,IAAI,IAAI,KAAK,CAAC,GAAG,MAAM,IAAI;CACvD,KAAK,MAAM,QAAQ,CAAC,GAAG,KAAK,CAAC,CAAC,KAAK,OAAO,GAAG,MAAM,IAAI;CACvD,OAAO;AACT;;;;;;;;;;;ACvRA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAC5D,MAAM,UAAU,MAAM,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,mBAAmB,EAAE,MAAM;EACxD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH;CACA,OAAO,kBAAkB,SAAS,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;;;;;;;;;;;;ACzCA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF"}

package/dist/{power-and-mde-K_CmO3dc.js → power-and-mde-bie6i965.js}

@@ -1,5 +1,5 @@
-import { s as zQuantile } from "./internal-BMFSR8Ns.js";
-import { u as normalCdf } from "./paired-arms-DKZFVPt8.js";
+import { s as zQuantile } from "./internal-BOe3zmjQ.js";
+import { u as normalCdf } from "./paired-arms-DpS_h8ON.js";
 //#region src/statistics/power-and-mde.ts
 /**
 * Required N per arm for a two-sample comparison at target effect size,
@@ -12,7 +12,8 @@ function requiredSampleSize(opts) {
 	if (!Number.isFinite(effect) || effect <= 0) return Infinity;
 	const alpha = opts.alpha ?? .05;
 	const power = opts.power ?? .8;
-	const n = 2 * ((zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / effect) ** 2;
+	const twoSided = opts.twoSided ?? true;
+	const n = 2 * ((zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / effect) ** 2;
 	return Math.ceil(n);
 }
 /**
@@ -29,7 +30,8 @@ function requiredPairedSampleSize(opts) {
 	if (!Number.isFinite(effect) || effect <= 0) return Infinity;
 	const alpha = opts.alpha ?? .05;
 	const power = opts.power ?? .8;
-	const zAlpha = zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha);
+	const twoSided = opts.twoSided ?? true;
+	const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha);
 	const zBeta = zQuantile(power);
 	return Math.ceil(((zAlpha + zBeta) / effect) ** 2);
 }
@@ -43,7 +45,8 @@ function pairedMde(opts) {
 	if (!Number.isFinite(opts.nPaired) || opts.nPaired <= 0) return Infinity;
 	const alpha = opts.alpha ?? .05;
 	const power = opts.power ?? .8;
-	return (zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / Math.sqrt(opts.nPaired);
+	const twoSided = opts.twoSided ?? true;
+	return (zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / Math.sqrt(opts.nPaired);
 }
 /**
 * Number of paired observations needed for a McNemar test to reach a target
@@ -94,4 +97,4 @@ function mcnemarPower(opts) {
 //#endregion
 export { requiredSampleSize as a, requiredPairedSampleSize as i, mcnemarRequiredN as n, pairedMde as r, mcnemarPower as t };
 
-//# sourceMappingURL=power-and-mde-K_CmO3dc.js.map
+//# sourceMappingURL=power-and-mde-bie6i965.js.map