@tangle-network/agent-eval 1.0.0 → 1.0.1

package/dist/{paired-binary-BR0nwHOA.js.map → paired-binary-B477lksS.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-binary-BR0nwHOA.js","names":[],"sources":["../src/statistics/paired-binary.ts"],"sourcesContent":["import { regularizedIncompleteBeta } from '../math/special-functions'\nimport { binomialSignTwoSided, zQuantile } from './internal'\n\n// ── Binomial proportion + paired-binary + coding-eval estimators ─────\n//\n// The paired family above (pairedBootstrap/pairedTTest/wilcoxonSignedRank)\n// operates on continuous scores. Pass/fail A/B comparisons — \"does treatment\n// X raise the success RATE vs control\" — are binary and paired, so they need\n// their own correct estimators: McNemar for significance (only the discordant\n// pairs carry signal), the paired risk difference for effect size, Wilson for\n// a single-arm proportion CI, and pass@k for the standard k-sample coding-eval\n// metric. The normal approximation is wrong for proportions near 0/1 and for\n// the small discordant counts typical of eval runs, so these are exact /\n// Wilson-based, not Wald.\n\n/** A binomial proportion estimate with a confidence interval. */\nexport interface ProportionInterval {\n  /** Point estimate successes / n; 0 is a compatibility placeholder when n = 0. */\n  estimate: number\n  /** Lower bound, clamped to [0, 1]. */\n  lower: number\n  /** Upper bound, clamped to [0, 1]. */\n  upper: number\n}\n\n/**\n * Wilson score interval for a binomial proportion. Correct at small n and near\n * 0/1, where the normal (Wald) approximation produces bounds outside [0, 1] and\n * understates coverage. Use this for any pass-rate / hit-rate / realness-rate\n * CI — the continuous `confidenceInterval` assumes the wrong distribution for a\n * proportion. Counts must be finite nonnegative integers, successes cannot exceed\n * n, and confidence must lie strictly between zero and one. With n = 0 the\n * interval is the full [0, 1] support: no rate is ruled out. The estimate stays\n * 0 for compatibility, not as a measured failure rate; retain n alongside it.\n */\nexport function wilson(successes: number, n: number, confidence = 0.95): ProportionInterval {\n  if (!Number.isSafeInteger(n) || n < 0) {\n    throw new Error('wilson: n must be a finite nonnegative integer')\n  }\n  if (!Number.isSafeInteger(successes) || successes < 0 || successes > n) {\n    throw new Error(`wilson: successes (${successes}) must be in [0, ${n}]`)\n  }\n  if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n    throw new Error('wilson: confidence must be between zero and one')\n  }\n  if (n === 0) return { estimate: 0, lower: 0, upper: 1 }\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const p = successes / n\n  const z2 = z * z\n  const denom = 1 + z2 / n\n  const center = (p + z2 / (2 * n)) / denom\n  const half = (z * Math.sqrt((p * (1 - p) + z2 / (4 * n)) / n)) / denom\n  return {\n    estimate: p,\n    lower: Math.max(0, center - half),\n    upper: Math.min(1, center + half),\n  }\n}\n\n/**\n * Are these per-item outcomes binary (every value exactly 0 or 1)?\n *\n * The discriminator a promotion gate needs before choosing a paired statistic.\n * On binary outcomes the paired delta vector lives in {-1, 0, +1} and is\n * normally dominated by zeros (both arms solve, or both arms miss, most items),\n * so its MEDIAN is pinned at exactly 0 no matter how large the real shift in\n * success rate is — and a bootstrap CI on that median collapses to [0, 0].\n * A gate keying on `ci.low > threshold` is then structurally unable to see\n * either a gain or a regression. Detect this shape and switch to the\n * paired-binary estimators ({@link mcnemar}, {@link pairedRiskDifference})\n * instead of silently answering \"no\" forever.\n *\n * Empty input is NOT binary: there is no evidence of the outcome's shape, and\n * defaulting an empty vector into the binary branch would pick a statistic on\n * no data at all.\n *\n * NOT the right discriminator for a gate. It recognises the literal {0, 1}\n * encoding and nothing else, so a pass/fail dimension emitted on 0-100 — which\n * judges in this codebase do routinely — reads as non-binary, and a single\n * partial-credit score in an otherwise pass/fail vector flips it to false while\n * leaving the median just as blind. Gates want {@link pairedBinaryScale} (any\n * two-point encoding). This predicate remains for callers that specifically\n * mean \"literally 0/1\".\n */\nfunction isBinaryOutcomeVector(values: ArrayLike<number>): boolean {\n  if (values.length === 0) return false\n  for (let i = 0; i < values.length; i++) {\n    const v = values[i]!\n    if (v !== 0 && v !== 1) return false\n  }\n  return true\n}\n\n/** Result of a McNemar paired-binary significance test. */\nexport interface McNemarResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs (b + c) — the only ones that carry signal. */\n  nDiscordant: number\n  /** Pairs where treatment succeeded and control failed (\"newly correct\"). */\n  b: number\n  /** Pairs where control succeeded and treatment failed (\"newly wrong\"). */\n  c: number\n  /** Continuity-corrected chi-square statistic (reference; exact p drives the call). */\n  statistic: number\n  /** Two-sided p-value. Exact (binomial sign test on discordant pairs). */\n  pValue: number\n}\n\n/**\n * McNemar's test for paired binary outcomes — the correct significance test for\n * \"does treatment change the success rate vs control on the SAME items\". Only\n * discordant pairs (one arm right, the other wrong) carry information; concordant\n * pairs are uninformative, so a paired t-test / two-proportion z-test on the raw\n * rates is wrong here. The p-value is exact: under H0 the b \"treatment-wins\" are\n * Binomial(b + c, 0.5), so the two-sided p is the doubled binomial tail — correct\n * at the small discordant counts typical of eval runs (no continuity-corrected\n * chi-square approximation needed, though it is returned as `statistic` for\n * reference). Inputs are paired 0/1 (or boolean) arrays, control first to match\n * the module's (before, after) convention. Throws on unequal lengths.\n */\nexport function mcnemar(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n): McNemarResult {\n  if (control.length !== treatment.length) {\n    throw new Error(`mcnemar: unequal sample sizes (${control.length} vs ${treatment.length})`)\n  }\n  const n = control.length\n  let b = 0 // treatment 1, control 0\n  let c = 0 // treatment 0, control 1\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const nDiscordant = b + c\n  const statistic = nDiscordant === 0 ? 0 : (Math.abs(b - c) - 1) ** 2 / nDiscordant\n  return { n, nDiscordant, b, c, statistic, pValue: binomialSignTwoSided(b, c) }\n}\n\n/** A paired binary effect size (treatment rate − control rate) with a CI. */\nexport interface RiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Lower bound of the CI, clamped to [-1, 1]. */\n  lower: number\n  /** Upper bound of the CI, clamped to [-1, 1]. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Paired risk difference (the effect-size companion to {@link mcnemar}): the\n * change in success rate p(treatment) − p(control) on matched items, which for\n * paired binary data equals (b − c) / n. The CI uses the paired variance from\n * the discordant counts, not the independent-samples formula (which overstates\n * the interval by ignoring the pairing). Inputs are paired 0/1 (or boolean)\n * arrays, control first. Throws on unequal lengths. With no pairs, returns the\n * full [-1, 1] support and a compatibility point estimate of 0, not equivalence.\n *\n * REPORTING ONLY — do NOT decide a promotion on this interval. The CI is a Wald\n * normal approximation, which badly UNDERCOVERS when only a handful of pairs are\n * discordant: at n = 3 with b = 2, c = 0 it returns [0.133, 1.000], excluding 0,\n * while McNemar's exact test on the same data gives p = 0.50. A gate keying on\n * `lower > 0` would promote noise. Use {@link pairedRiskDifferenceExact}, whose\n * interval is dual to the exact test by construction, for any decision.\n */\nexport function pairedRiskDifference(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): RiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifference: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  const n = control.length\n  if (n === 0) return { n: 0, b: 0, c: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const rd = (b - c) / n\n  const variance = (b + c - (b - c) ** 2 / n) / (n * n)\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const half = z * Math.sqrt(Math.max(0, variance))\n  return {\n    n,\n    b,\n    c,\n    riskDifference: rd,\n    lower: Math.max(-1, rd - half),\n    upper: Math.min(1, rd + half),\n    confidence,\n  }\n}\n\n/** A paired binary effect size with an EXACT interval and the exact test that\n *  bounds it — one object so a caller cannot read the estimate without the\n *  significance it is entitled to. */\nexport interface ExactRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c) — the only ones carrying information. */\n  nDiscordant: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Exact conditional CI lower bound: -1 with no pairs; 0 with only concordant pairs. */\n  lower: number\n  /** Exact conditional CI upper bound: 1 with no pairs; 0 with only concordant pairs. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n  /** McNemar's exact two-sided p-value on the same discordant counts. */\n  pValue: number\n}\n\n/**\n * Paired risk difference with the EXACT CONDITIONAL interval — the estimator a\n * promotion gate may decide on.\n *\n * Conditional on the number of discordant pairs m = b + c, the treatment-win\n * count b is Binomial(m, π) with π = P(treatment wins | discordant), and the\n * risk difference is an exact reparameterisation: RD = (2π − 1)·m/n. So a\n * Clopper-Pearson exact interval for π maps straight onto RD. This buys the\n * property the Wald interval in {@link pairedRiskDifference} does not have:\n *\n *   **`lower > 0` ⟺ McNemar's exact test rejects at α = 1 − confidence.**\n *\n * Clopper-Pearson excludes π = 0.5 exactly when the two-sided exact binomial\n * test of π = 0.5 rejects, and that test IS {@link mcnemar}'s p-value — so the\n * interval and the test can never disagree, and a gate keyed on `lower` cannot\n * promote what the exact test refuses. The exact p is returned in the same\n * object so the two are impossible to compute apart.\n *\n * The interval is conservative (exact intervals over-cover; conditioning on m\n * discards the concordant pairs' information about m itself). That is the\n * correct direction for a promotion gate: it refuses more often, never less.\n *\n * With n = 0, returns the full [-1, 1] support, p = 1, and a compatibility\n * point estimate of 0. With n > 0 and m = 0, π is not identified: the result\n * is the degenerate [0, 0] with p = 1. That is NOT evidence of equivalence —\n * callers must treat a zero-width interval as \"cannot decide\", not as \"no\n * difference\". Inputs are paired 0/1 (or boolean) arrays, control first.\n * Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceExact(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ExactRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceExact: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceExact: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return {\n      n: 0,\n      b: 0,\n      c: 0,\n      nDiscordant: 0,\n      riskDifference: 0,\n      lower: -1,\n      upper: 1,\n      confidence,\n      pValue: 1,\n    }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const m = b + c\n  const riskDifference = (b - c) / n\n  const pValue = binomialSignTwoSided(b, c)\n  if (m === 0) {\n    return { n, b, c, nDiscordant: 0, riskDifference: 0, lower: 0, upper: 0, confidence, pValue }\n  }\n  const alpha = 1 - confidence\n  const piLow = b === 0 ? 0 : betaQuantile(alpha / 2, b, m - b + 1)\n  const piHigh = b === m ? 1 : betaQuantile(1 - alpha / 2, b + 1, m - b)\n  const scale = m / n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: m,\n    riskDifference,\n    lower: Math.max(-1, (2 * piLow - 1) * scale),\n    upper: Math.min(1, (2 * piHigh - 1) * scale),\n    confidence,\n    pValue,\n  }\n}\n\n/** Inverse regularized incomplete beta by bisection on\n *  {@link regularizedIncompleteBeta}, which is monotone increasing in x. 80\n *  halvings of [0,1] resolve to ~8e-25, far past the continued fraction's own\n *  3e-15 tolerance, so the quantile is as exact as the CDF it inverts. */\nfunction betaQuantile(p: number, a: number, b: number): number {\n  if (p <= 0) return 0\n  if (p >= 1) return 1\n  let lo = 0\n  let hi = 1\n  for (let i = 0; i < 80; i++) {\n    const mid = (lo + hi) / 2\n    if (regularizedIncompleteBeta(mid, a, b) < p) lo = mid\n    else hi = mid\n  }\n  return (lo + hi) / 2\n}\n\n/** A paired binary effect size with an interval that is valid at a NONZERO\n *  margin — the estimator a noninferiority decision may be made on. */\nexport interface ScoreRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c). */\n  nDiscordant: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Score-interval lower bound on the population risk difference. */\n  lower: number\n  /** Score-interval upper bound on the population risk difference. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Constrained MLE of q = P(treatment loses) under the hypothesis RD = `delta`.\n *\n * Profiling the two concordant cells out of the multinomial leaves\n * `L(q) = b·log(q+delta) + c·log(q) + e·log(1 − 2q − delta)` with `e = n − b − c`,\n * whose stationary point is the positive root of\n * `2n·q² − [(b + c) − delta·(b + 3c + 2e)]·q − c·delta·(1 − delta) = 0`.\n * At `delta = 0` this returns `(b + c) / 2n`, the familiar null.\n */\nfunction constrainedLossRate(b: number, c: number, n: number, delta: number): number {\n  const e = n - b - c\n  const quadratic = 2 * n\n  const linear = -(b + c - delta * (b + 3 * c + 2 * e))\n  const constant = -c * delta * (1 - delta)\n  const discriminant = linear * linear - 4 * quadratic * constant\n  const root = discriminant > 0 ? Math.sqrt(discriminant) : 0\n  const q = (-linear + root) / (2 * quadratic)\n  // Clamp into the region where all four cell probabilities stay non-negative.\n  return Math.min(Math.max(q, Math.max(0, -delta)), Math.max(0, (1 - delta) / 2))\n}\n\n/** Tango's score statistic for H0: RD = `delta`. `Var(b − c) = n·(2q + delta −\n *  delta²)` under that hypothesis, evaluated at the constrained MLE of q. */\nfunction tangoScore(b: number, c: number, n: number, delta: number): number {\n  const numerator = b - c - n * delta\n  const q = constrainedLossRate(b, c, n, delta)\n  const variance = n * (2 * q + delta * (1 - delta))\n  if (!(variance > 0)) {\n    if (numerator === 0) return 0\n    return numerator > 0 ? Number.POSITIVE_INFINITY : Number.NEGATIVE_INFINITY\n  }\n  return numerator / Math.sqrt(variance)\n}\n\n/**\n * Paired risk difference with TANGO'S (1998) SCORE INTERVAL — the estimator a\n * promotion gate may decide on **at a nonzero margin**.\n *\n * {@link pairedRiskDifferenceExact} conditions on the observed discordant count\n * `m = b + c`, builds a Clopper-Pearson interval for the win share among those\n * `m` pairs, and multiplies by the observed `m/n`. That is exact for testing\n * RD = 0 — it is dual to McNemar — but it is NOT a confidence interval for the\n * population risk difference at a nonzero margin, because the sampling\n * variability of `m/n` itself is discarded. The gap is not academic: with the\n * production caller's `pairedDeltaThreshold: -0.05`, a process whose true risk\n * difference sits exactly on that margin clears a nominal-95 % `lower > margin`\n * check 24.75 % of the time at n = 40 and 43.95 % at n = 76 (2000 replicates\n * each) when the conditional interval decides.\n *\n * Tango's interval inverts the score test of RD = delta, which estimates the\n * nuisance loss rate under each hypothesised delta instead of fixing it at the\n * observed value, so `m` contributes its own uncertainty. It is the method\n * `ratesci::scorepairci` uses for paired risk-difference noninferiority, and it\n * is not conditional, so it stays valid as the margin moves away from zero.\n *\n * The bounds are found by bisecting `tangoScore(delta) = ±z` — the score is\n * monotone decreasing in delta, so each crossing is unique. Inputs are paired\n * 0/1 (or boolean) arrays, control first. Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceScore(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ScoreRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceScore: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceScore: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return { n: 0, b: 0, c: 0, nDiscordant: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const riskDifference = (b - c) / n\n  const z = zQuantile(1 - (1 - confidence) / 2)\n\n  // Lower bound: the smallest delta still inside the interval, i.e. the root of\n  // score(delta) = +z on [-1, riskDifference]. score(riskDifference) = 0 < z, so\n  // the right endpoint is always inside and the bisection is well posed.\n  let lo = -1\n  let hi = riskDifference\n  // 64 halvings resolve the [-1, 1] search interval below 6e-20, past the\n  // precision available to the score calculation. More iterations only\n  // repeat identical floating-point values while multiplying every gate\n  // decision's cost.\n  for (let i = 0; i < 64; i++) {\n    const mid = (lo + hi) / 2\n    if (tangoScore(b, c, n, mid) > z) lo = mid\n    else hi = mid\n  }\n  const lower = (lo + hi) / 2\n\n  // Upper bound: root of score(delta) = -z on [riskDifference, 1].\n  let ulo = riskDifference\n  let uhi = 1\n  for (let i = 0; i < 64; i++) {\n    const mid = (ulo + uhi) / 2\n    if (tangoScore(b, c, n, mid) > -z) ulo = mid\n    else uhi = mid\n  }\n  const upper = (ulo + uhi) / 2\n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: b + c,\n    riskDifference,\n    lower: Math.max(-1, lower),\n    upper: Math.min(1, upper),\n    confidence,\n  }\n}\n\n/**\n * The common positive level `s` such that EVERY value across both paired arms is\n * exactly 0 or `s` — i.e. the outcome is pass/fail, whatever encoding it arrived\n * in. Returns null when the outcomes are not two-point, when the two arms use\n * different levels, or when no positive value was observed at all (all-zero\n * arms: the level is not identified, and there is nothing to decide anyway).\n *\n * This is the scale-aware successor to {@link isBinaryOutcomeVector}, which only\n * recognises literal {0, 1}. Judges in this codebase emit dimensions on 0-100 as\n * well as [0,1] (see `detectScale` in `campaign/gates/statistical-heldout.ts`),\n * so a pass/fail dimension routinely arrives as {0, 100} and a {0,1}-only test\n * silently sends it down the median path that cannot see it. Any positive level\n * is accepted, not just 1 and 100: for a two-point {0, s} outcome the mean paired\n * delta is exactly s·(b − c)/n, so the binary estimators apply after dividing by\n * s and rescaling the result back into the caller's native units.\n *\n * Non-finite values ⇒ null: an unusable outcome must not be classified as a\n * clean pass/fail shape.\n */\nexport function pairedBinaryScale(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number | null {\n  let level: number | null = null\n  for (const arm of [before, after]) {\n    for (let i = 0; i < arm.length; i++) {\n      const v = arm[i]!\n      if (!Number.isFinite(v)) return null\n      if (v === 0) continue\n      if (v < 0) return null\n      if (level === null) level = v\n      else if (v !== level) return null\n    }\n  }\n  return level\n}\n\n/**\n * Unbiased pass@k for code generation (Chen et al. 2021, \"Evaluating Large\n * Language Models Trained on Code\"). Given `n` independent samples for one\n * problem of which `c` pass, the probability that at least one of a random k of\n * them passes is 1 − C(n−c, k) / C(n, k). Estimating pass@k as \"did any of the\n * first k pass\" is biased high at small n; this is the variance-reduced estimator\n * averaged implicitly over all k-subsets. Average the per-problem values across\n * the suite for the corpus pass@k. Computed in the numerically stable product\n * form. Requires 1 ≤ k ≤ n and 0 ≤ c ≤ n.\n */\nexport function passAtK(n: number, c: number, k: number): number {\n  if (!Number.isInteger(n) || !Number.isInteger(c) || !Number.isInteger(k)) {\n    throw new Error(`passAtK: n, c, k must be integers (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (k < 1 || k > n || c < 0 || c > n) {\n    throw new Error(`passAtK: require 1 ≤ k ≤ n and 0 ≤ c ≤ n (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (n - c < k) return 1\n  let prob = 1\n  for (let i = n - c + 1; i <= n; i++) prob *= 1 - k / i\n  return 1 - prob\n}\n"],"mappings":";;;;;;;;;;;;AAmCA,SAAgB,OAAO,WAAmB,GAAW,aAAa,KAA0B;CAC1F,IAAI,CAAC,OAAO,cAAc,CAAC,KAAK,IAAI,GAClC,MAAM,IAAI,MAAM,gDAAgD;CAElE,IAAI,CAAC,OAAO,cAAc,SAAS,KAAK,YAAY,KAAK,YAAY,GACnE,MAAM,IAAI,MAAM,sBAAsB,UAAU,mBAAmB,EAAE,EAAE;CAEzE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MAAM,iDAAiD;CAEnE,IAAI,MAAM,GAAG,OAAO;EAAE,UAAU;EAAG,OAAO;EAAG,OAAO;CAAE;CACtD,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAC5C,MAAM,IAAI,YAAY;CACtB,MAAM,KAAK,IAAI;CACf,MAAM,QAAQ,IAAI,KAAK;CACvB,MAAM,UAAU,IAAI,MAAM,IAAI,MAAM;CACpC,MAAM,OAAQ,IAAI,KAAK,MAAM,KAAK,IAAI,KAAK,MAAM,IAAI,MAAM,CAAC,IAAK;CACjE,OAAO;EACL,UAAU;EACV,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;EAChC,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;CAClC;AACF;;;;;;;;;;;;;AAgEA,SAAgB,QACd,SACA,WACe;CACf,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MAAM,kCAAkC,QAAQ,OAAO,MAAM,UAAU,OAAO,EAAE;CAE5F,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,cAAc,IAAI;CACxB,MAAM,YAAY,gBAAgB,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,IAAI,MAAM,IAAI;CACvE,OAAO;EAAE;EAAG;EAAa;EAAG;EAAG;EAAW,QAAQ,qBAAqB,GAAG,CAAC;CAAE;AAC/E;;;;;;;;;;;;;;;;;AAoCA,SAAgB,qBACd,SACA,WACA,aAAa,KACS;CACtB,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,+CAA+C,QAAQ,OAAO,MAAM,UAAU,OAAO,EACvF;CAEF,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GAAG,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAC3F,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,MAAM,IAAI,KAAK;CACrB,MAAM,YAAY,IAAI,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI;CAEnD,MAAM,OADI,UAAU,KAAK,IAAI,cAAc,CAC9B,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,CAAC;CAChD,OAAO;EACL;EACA;EACA;EACA,gBAAgB;EAChB,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI;EAC7B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI;EAC5B;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EACL,GAAG;EACH,GAAG;EACH,GAAG;EACH,aAAa;EACb,gBAAgB;EAChB,OAAO;EACP,OAAO;EACP;EACA,QAAQ;CACV;CAEF,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,IAAI,IAAI;CACd,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,SAAS,qBAAqB,GAAG,CAAC;CACxC,IAAI,MAAM,GACR,OAAO;EAAE;EAAG;EAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;EAAY;CAAO;CAE9F,MAAM,QAAQ,IAAI;CAClB,MAAM,QAAQ,MAAM,IAAI,IAAI,aAAa,QAAQ,GAAG,GAAG,IAAI,IAAI,CAAC;CAChE,MAAM,SAAS,MAAM,IAAI,IAAI,aAAa,IAAI,QAAQ,GAAG,IAAI,GAAG,IAAI,CAAC;CACrE,MAAM,QAAQ,IAAI;CAClB,OAAO;EACL;EACA;EACA;EACA,aAAa;EACb;EACA,OAAO,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,KAAK;EAC3C,OAAO,KAAK,IAAI,IAAI,IAAI,SAAS,KAAK,KAAK;EAC3C;EACA;CACF;AACF;;;;;AAMA,SAAS,aAAa,GAAW,GAAW,GAAmB;CAC7D,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,0BAA0B,KAAK,GAAG,CAAC,IAAI,GAAG,KAAK;OAC9C,KAAK;CACZ;CACA,QAAQ,KAAK,MAAM;AACrB;;;;;;;;;;AAgCA,SAAS,oBAAoB,GAAW,GAAW,GAAW,OAAuB;CACnF,MAAM,IAAI,IAAI,IAAI;CAClB,MAAM,YAAY,IAAI;CACtB,MAAM,SAAS,EAAE,IAAI,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI;CAClD,MAAM,WAAW,CAAC,IAAI,SAAS,IAAI;CACnC,MAAM,eAAe,SAAS,SAAS,IAAI,YAAY;CACvD,MAAM,OAAO,eAAe,IAAI,KAAK,KAAK,YAAY,IAAI;CAC1D,MAAM,KAAK,CAAC,SAAS,SAAS,IAAI;CAElC,OAAO,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,KAAK,IAAI,IAAI,IAAI,SAAS,CAAC,CAAC;AAChF;;;AAIA,SAAS,WAAW,GAAW,GAAW,GAAW,OAAuB;CAC1E,MAAM,YAAY,IAAI,IAAI,IAAI;CAE9B,MAAM,WAAW,KAAK,IADZ,oBAAoB,GAAG,GAAG,GAAG,KACb,IAAI,SAAS,IAAI;CAC3C,IAAI,EAAE,WAAW,IAAI;EACnB,IAAI,cAAc,GAAG,OAAO;EAC5B,OAAO,YAAY,IAAI,OAAO,oBAAoB,OAAO;CAC3D;CACA,OAAO,YAAY,KAAK,KAAK,QAAQ;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAEhG,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAK5C,IAAI,KAAK;CACT,IAAI,KAAK;CAKT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,GAAG,KAAK;OAClC,KAAK;CACZ;CACA,MAAM,SAAS,KAAK,MAAM;CAG1B,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,MAAM,OAAO;EAC1B,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,CAAC,GAAG,MAAM;OACpC,MAAM;CACb;CACA,MAAM,SAAS,MAAM,OAAO;CAE5B,OAAO;EACL;EACA;EACA;EACA,aAAa,IAAI;EACjB;EACA,OAAO,KAAK,IAAI,IAAI,KAAK;EACzB,OAAO,KAAK,IAAI,GAAG,KAAK;EACxB;CACF;AACF;;;;;;;;;;;;;;;;;;;;AAqBA,SAAgB,kBACd,QACA,OACe;CACf,IAAI,QAAuB;CAC3B,KAAK,MAAM,OAAO,CAAC,QAAQ,KAAK,GAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;EAChC,IAAI,MAAM,GAAG;EACb,IAAI,IAAI,GAAG,OAAO;EAClB,IAAI,UAAU,MAAM,QAAQ;OACvB,IAAI,MAAM,OAAO,OAAO;CAC/B;CAEF,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,QAAQ,GAAW,GAAW,GAAmB;CAC/D,IAAI,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,GACrE,MAAM,IAAI,MAAM,4CAA4C,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAElF,IAAI,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,GACjC,MAAM,IAAI,MAAM,mDAAmD,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAEzF,IAAI,IAAI,IAAI,GAAG,OAAO;CACtB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,QAAQ,IAAI,IAAI;CACrD,OAAO,IAAI;AACb"}
+{"version":3,"file":"paired-binary-B477lksS.js","names":[],"sources":["../src/statistics/paired-binary.ts"],"sourcesContent":["import { regularizedIncompleteBeta } from '../math/special-functions'\nimport { binomialSignTwoSided, zQuantile } from './internal'\n\n// ── Binomial proportion + paired-binary + coding-eval estimators ─────\n//\n// The paired family above (pairedBootstrap/pairedTTest/wilcoxonSignedRank)\n// operates on continuous scores. Pass/fail A/B comparisons — \"does treatment\n// X raise the success RATE vs control\" — are binary and paired, so they need\n// their own correct estimators: McNemar for significance (only the discordant\n// pairs carry signal), the paired risk difference for effect size, Wilson for\n// a single-arm proportion CI, and pass@k for the standard k-sample coding-eval\n// metric. The normal approximation is wrong for proportions near 0/1 and for\n// the small discordant counts typical of eval runs, so these are exact /\n// Wilson-based, not Wald.\n\n/** A binomial proportion estimate with a confidence interval. */\nexport interface ProportionInterval {\n  /** Point estimate successes / n; 0 is a compatibility placeholder when n = 0. */\n  estimate: number\n  /** Lower bound, clamped to [0, 1]. */\n  lower: number\n  /** Upper bound, clamped to [0, 1]. */\n  upper: number\n}\n\n/**\n * Wilson score interval for a binomial proportion. Correct at small n and near\n * 0/1, where the normal (Wald) approximation produces bounds outside [0, 1] and\n * understates coverage. Use this for any pass-rate / hit-rate / realness-rate\n * CI — the continuous `confidenceInterval` assumes the wrong distribution for a\n * proportion. Counts must be finite nonnegative integers, successes cannot exceed\n * n, and confidence must lie strictly between zero and one. With n = 0 the\n * interval is the full [0, 1] support: no rate is ruled out. The estimate stays\n * 0 for compatibility, not as a measured failure rate; retain n alongside it.\n */\nexport function wilson(successes: number, n: number, confidence = 0.95): ProportionInterval {\n  if (!Number.isSafeInteger(n) || n < 0) {\n    throw new Error('wilson: n must be a finite nonnegative integer')\n  }\n  if (!Number.isSafeInteger(successes) || successes < 0 || successes > n) {\n    throw new Error(`wilson: successes (${successes}) must be in [0, ${n}]`)\n  }\n  if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n    throw new Error('wilson: confidence must be between zero and one')\n  }\n  if (n === 0) return { estimate: 0, lower: 0, upper: 1 }\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const p = successes / n\n  const z2 = z * z\n  const denom = 1 + z2 / n\n  const center = (p + z2 / (2 * n)) / denom\n  const half = (z * Math.sqrt((p * (1 - p) + z2 / (4 * n)) / n)) / denom\n  return {\n    estimate: p,\n    lower: Math.max(0, center - half),\n    upper: Math.min(1, center + half),\n  }\n}\n\n/**\n * Are these per-item outcomes binary (every value exactly 0 or 1)?\n *\n * The discriminator a promotion gate needs before choosing a paired statistic.\n * On binary outcomes the paired delta vector lives in {-1, 0, +1} and is\n * normally dominated by zeros (both arms solve, or both arms miss, most items),\n * so its MEDIAN is pinned at exactly 0 no matter how large the real shift in\n * success rate is — and a bootstrap CI on that median collapses to [0, 0].\n * A gate keying on `ci.low > threshold` is then structurally unable to see\n * either a gain or a regression. Detect this shape and switch to the\n * paired-binary estimators ({@link mcnemar}, {@link pairedRiskDifference})\n * instead of silently answering \"no\" forever.\n *\n * Empty input is NOT binary: there is no evidence of the outcome's shape, and\n * defaulting an empty vector into the binary branch would pick a statistic on\n * no data at all.\n *\n * NOT the right discriminator for a gate. It recognises the literal {0, 1}\n * encoding and nothing else, so a pass/fail dimension emitted on 0-100 — which\n * judges in this codebase do routinely — reads as non-binary, and a single\n * partial-credit score in an otherwise pass/fail vector flips it to false while\n * leaving the median just as blind. Gates want {@link pairedBinaryScale} (any\n * two-point encoding). This predicate remains for callers that specifically\n * mean \"literally 0/1\".\n */\nfunction isBinaryOutcomeVector(values: ArrayLike<number>): boolean {\n  if (values.length === 0) return false\n  for (let i = 0; i < values.length; i++) {\n    const v = values[i]!\n    if (v !== 0 && v !== 1) return false\n  }\n  return true\n}\n\n/** Result of a McNemar paired-binary significance test. */\nexport interface McNemarResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs (b + c) — the only ones that carry signal. */\n  nDiscordant: number\n  /** Pairs where treatment succeeded and control failed (\"newly correct\"). */\n  b: number\n  /** Pairs where control succeeded and treatment failed (\"newly wrong\"). */\n  c: number\n  /** Continuity-corrected chi-square statistic (reference; exact p drives the call). */\n  statistic: number\n  /** Two-sided p-value. Exact (binomial sign test on discordant pairs). */\n  pValue: number\n}\n\n/**\n * McNemar's test for paired binary outcomes — the correct significance test for\n * \"does treatment change the success rate vs control on the SAME items\". Only\n * discordant pairs (one arm right, the other wrong) carry information; concordant\n * pairs are uninformative, so a paired t-test / two-proportion z-test on the raw\n * rates is wrong here. The p-value is exact: under H0 the b \"treatment-wins\" are\n * Binomial(b + c, 0.5), so the two-sided p is the doubled binomial tail — correct\n * at the small discordant counts typical of eval runs (no continuity-corrected\n * chi-square approximation needed, though it is returned as `statistic` for\n * reference). Inputs are paired 0/1 (or boolean) arrays, control first to match\n * the module's (before, after) convention. Throws on unequal lengths.\n */\nexport function mcnemar(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n): McNemarResult {\n  if (control.length !== treatment.length) {\n    throw new Error(`mcnemar: unequal sample sizes (${control.length} vs ${treatment.length})`)\n  }\n  const n = control.length\n  let b = 0 // treatment 1, control 0\n  let c = 0 // treatment 0, control 1\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const nDiscordant = b + c\n  const statistic = nDiscordant === 0 ? 0 : (Math.abs(b - c) - 1) ** 2 / nDiscordant\n  return { n, nDiscordant, b, c, statistic, pValue: binomialSignTwoSided(b, c) }\n}\n\n/** A paired binary effect size (treatment rate − control rate) with a CI. */\nexport interface RiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Lower bound of the CI, clamped to [-1, 1]. */\n  lower: number\n  /** Upper bound of the CI, clamped to [-1, 1]. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Paired risk difference (the effect-size companion to {@link mcnemar}): the\n * change in success rate p(treatment) − p(control) on matched items, which for\n * paired binary data equals (b − c) / n. The CI uses the paired variance from\n * the discordant counts, not the independent-samples formula (which overstates\n * the interval by ignoring the pairing). Inputs are paired 0/1 (or boolean)\n * arrays, control first. Throws on unequal lengths. With no pairs, returns the\n * full [-1, 1] support and a compatibility point estimate of 0, not equivalence.\n *\n * REPORTING ONLY — do NOT decide a promotion on this interval. The CI is a Wald\n * normal approximation, which badly UNDERCOVERS when only a handful of pairs are\n * discordant: at n = 3 with b = 2, c = 0 it returns [0.133, 1.000], excluding 0,\n * while McNemar's exact test on the same data gives p = 0.50. A gate keying on\n * `lower > 0` would promote noise. Use {@link pairedRiskDifferenceExact}, whose\n * interval is dual to the exact test by construction, for any decision.\n */\nexport function pairedRiskDifference(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): RiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifference: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  const n = control.length\n  if (n === 0) return { n: 0, b: 0, c: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const rd = (b - c) / n\n  const variance = (b + c - (b - c) ** 2 / n) / (n * n)\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const half = z * Math.sqrt(Math.max(0, variance))\n  return {\n    n,\n    b,\n    c,\n    riskDifference: rd,\n    lower: Math.max(-1, rd - half),\n    upper: Math.min(1, rd + half),\n    confidence,\n  }\n}\n\n/** A paired binary effect size with an EXACT interval and the exact test that\n *  bounds it — one object so a caller cannot read the estimate without the\n *  significance it is entitled to. */\nexport interface ExactRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c) — the only ones carrying information. */\n  nDiscordant: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Exact conditional CI lower bound: -1 with no pairs; 0 with only concordant pairs. */\n  lower: number\n  /** Exact conditional CI upper bound: 1 with no pairs; 0 with only concordant pairs. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n  /** McNemar's exact two-sided p-value on the same discordant counts. */\n  pValue: number\n}\n\n/**\n * Paired risk difference with the EXACT CONDITIONAL interval — the estimator a\n * promotion gate may decide on.\n *\n * Conditional on the number of discordant pairs m = b + c, the treatment-win\n * count b is Binomial(m, π) with π = P(treatment wins | discordant), and the\n * risk difference is an exact reparameterisation: RD = (2π − 1)·m/n. So a\n * Clopper-Pearson exact interval for π maps straight onto RD. This buys the\n * property the Wald interval in {@link pairedRiskDifference} does not have:\n *\n *   **`lower > 0` ⟺ McNemar's exact test rejects at α = 1 − confidence.**\n *\n * Clopper-Pearson excludes π = 0.5 exactly when the two-sided exact binomial\n * test of π = 0.5 rejects, and that test IS {@link mcnemar}'s p-value — so the\n * interval and the test can never disagree, and a gate keyed on `lower` cannot\n * promote what the exact test refuses. The exact p is returned in the same\n * object so the two are impossible to compute apart.\n *\n * The interval is conservative (exact intervals over-cover; conditioning on m\n * discards the concordant pairs' information about m itself). That is the\n * correct direction for a promotion gate: it refuses more often, never less.\n *\n * With n = 0, returns the full [-1, 1] support, p = 1, and a compatibility\n * point estimate of 0. With n > 0 and m = 0, π is not identified: the result\n * is the degenerate [0, 0] with p = 1. That is NOT evidence of equivalence —\n * callers must treat a zero-width interval as \"cannot decide\", not as \"no\n * difference\". Inputs are paired 0/1 (or boolean) arrays, control first.\n * Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceExact(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ExactRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceExact: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceExact: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return {\n      n: 0,\n      b: 0,\n      c: 0,\n      nDiscordant: 0,\n      riskDifference: 0,\n      lower: -1,\n      upper: 1,\n      confidence,\n      pValue: 1,\n    }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const m = b + c\n  const riskDifference = (b - c) / n\n  const pValue = binomialSignTwoSided(b, c)\n  if (m === 0) {\n    return { n, b, c, nDiscordant: 0, riskDifference: 0, lower: 0, upper: 0, confidence, pValue }\n  }\n  const alpha = 1 - confidence\n  const piLow = b === 0 ? 0 : betaQuantile(alpha / 2, b, m - b + 1)\n  const piHigh = b === m ? 1 : betaQuantile(1 - alpha / 2, b + 1, m - b)\n  const scale = m / n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: m,\n    riskDifference,\n    lower: Math.max(-1, (2 * piLow - 1) * scale),\n    upper: Math.min(1, (2 * piHigh - 1) * scale),\n    confidence,\n    pValue,\n  }\n}\n\n/** Inverse regularized incomplete beta by bisection on\n *  {@link regularizedIncompleteBeta}, which is monotone increasing in x. 80\n *  halvings of [0,1] resolve to ~8e-25, far past the continued fraction's own\n *  3e-15 tolerance, so the quantile is as exact as the CDF it inverts. */\nfunction betaQuantile(p: number, a: number, b: number): number {\n  if (p <= 0) return 0\n  if (p >= 1) return 1\n  let lo = 0\n  let hi = 1\n  for (let i = 0; i < 80; i++) {\n    const mid = (lo + hi) / 2\n    if (regularizedIncompleteBeta(mid, a, b) < p) lo = mid\n    else hi = mid\n  }\n  return (lo + hi) / 2\n}\n\n/** A paired binary effect size with an interval that is valid at a NONZERO\n *  margin — the estimator a noninferiority decision may be made on. */\nexport interface ScoreRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c). */\n  nDiscordant: number\n  /** Paired risk difference (b − c) / n; 0 is a placeholder when n = 0. */\n  riskDifference: number\n  /** Score-interval lower bound on the population risk difference. */\n  lower: number\n  /** Score-interval upper bound on the population risk difference. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Constrained MLE of q = P(treatment loses) under the hypothesis RD = `delta`.\n *\n * Profiling the two concordant cells out of the multinomial leaves\n * `L(q) = b·log(q+delta) + c·log(q) + e·log(1 − 2q − delta)` with `e = n − b − c`,\n * whose stationary point is the positive root of\n * `2n·q² − [(b + c) − delta·(b + 3c + 2e)]·q − c·delta·(1 − delta) = 0`.\n * At `delta = 0` this returns `(b + c) / 2n`, the familiar null.\n */\nfunction constrainedLossRate(b: number, c: number, n: number, delta: number): number {\n  const e = n - b - c\n  const quadratic = 2 * n\n  const linear = -(b + c - delta * (b + 3 * c + 2 * e))\n  const constant = -c * delta * (1 - delta)\n  const discriminant = linear * linear - 4 * quadratic * constant\n  const root = discriminant > 0 ? Math.sqrt(discriminant) : 0\n  const q = (-linear + root) / (2 * quadratic)\n  // Clamp into the region where all four cell probabilities stay non-negative.\n  return Math.min(Math.max(q, Math.max(0, -delta)), Math.max(0, (1 - delta) / 2))\n}\n\n/** Tango's score statistic for H0: RD = `delta`. `Var(b − c) = n·(2q + delta −\n *  delta²)` under that hypothesis, evaluated at the constrained MLE of q. */\nfunction tangoScore(b: number, c: number, n: number, delta: number): number {\n  const numerator = b - c - n * delta\n  const q = constrainedLossRate(b, c, n, delta)\n  const variance = n * (2 * q + delta * (1 - delta))\n  if (!(variance > 0)) {\n    if (numerator === 0) return 0\n    return numerator > 0 ? Number.POSITIVE_INFINITY : Number.NEGATIVE_INFINITY\n  }\n  return numerator / Math.sqrt(variance)\n}\n\n/**\n * Paired risk difference with TANGO'S (1998) SCORE INTERVAL — the estimator a\n * promotion gate may decide on **at a nonzero margin**.\n *\n * {@link pairedRiskDifferenceExact} conditions on the observed discordant count\n * `m = b + c`, builds a Clopper-Pearson interval for the win share among those\n * `m` pairs, and multiplies by the observed `m/n`. That is exact for testing\n * RD = 0 — it is dual to McNemar — but it is NOT a confidence interval for the\n * population risk difference at a nonzero margin, because the sampling\n * variability of `m/n` itself is discarded. The gap is not academic: with the\n * production caller's `pairedDeltaThreshold: -0.05`, a process whose true risk\n * difference sits exactly on that margin clears a nominal-95 % `lower > margin`\n * check 24.75 % of the time at n = 40 and 43.95 % at n = 76 (2000 replicates\n * each) when the conditional interval decides.\n *\n * Tango's interval inverts the score test of RD = delta, which estimates the\n * nuisance loss rate under each hypothesised delta instead of fixing it at the\n * observed value, so `m` contributes its own uncertainty. It is the method\n * `ratesci::scorepairci` uses for paired risk-difference noninferiority, and it\n * is not conditional, so it stays valid as the margin moves away from zero.\n *\n * The bounds are found by bisecting `tangoScore(delta) = ±z` — the score is\n * monotone decreasing in delta, so each crossing is unique. Inputs are paired\n * 0/1 (or boolean) arrays, control first. Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceScore(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ScoreRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceScore: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceScore: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return { n: 0, b: 0, c: 0, nDiscordant: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const riskDifference = (b - c) / n\n  const z = zQuantile(1 - (1 - confidence) / 2)\n\n  // Lower bound: the smallest delta still inside the interval, i.e. the root of\n  // score(delta) = +z on [-1, riskDifference]. score(riskDifference) = 0 < z, so\n  // the right endpoint is always inside and the bisection is well posed.\n  let lo = -1\n  let hi = riskDifference\n  // 64 halvings resolve the [-1, 1] search interval below 6e-20, past the\n  // precision available to the score calculation. More iterations only\n  // repeat identical floating-point values while multiplying every gate\n  // decision's cost.\n  for (let i = 0; i < 64; i++) {\n    const mid = (lo + hi) / 2\n    if (tangoScore(b, c, n, mid) > z) lo = mid\n    else hi = mid\n  }\n  const lower = (lo + hi) / 2\n\n  // Upper bound: root of score(delta) = -z on [riskDifference, 1].\n  let ulo = riskDifference\n  let uhi = 1\n  for (let i = 0; i < 64; i++) {\n    const mid = (ulo + uhi) / 2\n    if (tangoScore(b, c, n, mid) > -z) ulo = mid\n    else uhi = mid\n  }\n  const upper = (ulo + uhi) / 2\n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: b + c,\n    riskDifference,\n    lower: Math.max(-1, lower),\n    upper: Math.min(1, upper),\n    confidence,\n  }\n}\n\n/**\n * The common positive level `s` such that EVERY value across both paired arms is\n * exactly 0 or `s` — i.e. the outcome is pass/fail, whatever encoding it arrived\n * in. Returns null when the outcomes are not two-point, when the two arms use\n * different levels, or when no positive value was observed at all (all-zero\n * arms: the level is not identified, and there is nothing to decide anyway).\n *\n * This is the scale-aware successor to {@link isBinaryOutcomeVector}, which only\n * recognises literal {0, 1}. Judges in this codebase emit dimensions on 0-100 as\n * well as [0,1] (see `detectScale` in `campaign/gates/statistical-heldout.ts`),\n * so a pass/fail dimension routinely arrives as {0, 100} and a {0,1}-only test\n * silently sends it down the median path that cannot see it. Any positive level\n * is accepted, not just 1 and 100: for a two-point {0, s} outcome the mean paired\n * delta is exactly s·(b − c)/n, so the binary estimators apply after dividing by\n * s and rescaling the result back into the caller's native units.\n *\n * Non-finite values ⇒ null: an unusable outcome must not be classified as a\n * clean pass/fail shape.\n */\nexport function pairedBinaryScale(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number | null {\n  let level: number | null = null\n  for (const arm of [before, after]) {\n    for (let i = 0; i < arm.length; i++) {\n      const v = arm[i]!\n      if (!Number.isFinite(v)) return null\n      if (v === 0) continue\n      if (v < 0) return null\n      if (level === null) level = v\n      else if (v !== level) return null\n    }\n  }\n  return level\n}\n\n/**\n * Unbiased pass@k for code generation (Chen et al. 2021, \"Evaluating Large\n * Language Models Trained on Code\"). Given `n` independent samples for one\n * problem of which `c` pass, the probability that at least one of a random k of\n * them passes is 1 − C(n−c, k) / C(n, k). Estimating pass@k as \"did any of the\n * first k pass\" is biased high at small n; this is the variance-reduced estimator\n * averaged implicitly over all k-subsets. Average the per-problem values across\n * the suite for the corpus pass@k. Computed in the numerically stable product\n * form. Requires 1 ≤ k ≤ n and 0 ≤ c ≤ n.\n */\nexport function passAtK(n: number, c: number, k: number): number {\n  if (!Number.isInteger(n) || !Number.isInteger(c) || !Number.isInteger(k)) {\n    throw new Error(`passAtK: n, c, k must be integers (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (k < 1 || k > n || c < 0 || c > n) {\n    throw new Error(`passAtK: require 1 ≤ k ≤ n and 0 ≤ c ≤ n (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (n - c < k) return 1\n  let prob = 1\n  for (let i = n - c + 1; i <= n; i++) prob *= 1 - k / i\n  return 1 - prob\n}\n"],"mappings":";;;;;;;;;;;;AAmCA,SAAgB,OAAO,WAAmB,GAAW,aAAa,KAA0B;CAC1F,IAAI,CAAC,OAAO,cAAc,CAAC,KAAK,IAAI,GAClC,MAAM,IAAI,MAAM,gDAAgD;CAElE,IAAI,CAAC,OAAO,cAAc,SAAS,KAAK,YAAY,KAAK,YAAY,GACnE,MAAM,IAAI,MAAM,sBAAsB,UAAU,mBAAmB,EAAE,EAAE;CAEzE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MAAM,iDAAiD;CAEnE,IAAI,MAAM,GAAG,OAAO;EAAE,UAAU;EAAG,OAAO;EAAG,OAAO;CAAE;CACtD,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAC5C,MAAM,IAAI,YAAY;CACtB,MAAM,KAAK,IAAI;CACf,MAAM,QAAQ,IAAI,KAAK;CACvB,MAAM,UAAU,IAAI,MAAM,IAAI,MAAM;CACpC,MAAM,OAAQ,IAAI,KAAK,MAAM,KAAK,IAAI,KAAK,MAAM,IAAI,MAAM,CAAC,IAAK;CACjE,OAAO;EACL,UAAU;EACV,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;EAChC,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;CAClC;AACF;;;;;;;;;;;;;AAgEA,SAAgB,QACd,SACA,WACe;CACf,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MAAM,kCAAkC,QAAQ,OAAO,MAAM,UAAU,OAAO,EAAE;CAE5F,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,cAAc,IAAI;CACxB,MAAM,YAAY,gBAAgB,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,IAAI,MAAM,IAAI;CACvE,OAAO;EAAE;EAAG;EAAa;EAAG;EAAG;EAAW,QAAQ,qBAAqB,GAAG,CAAC;CAAE;AAC/E;;;;;;;;;;;;;;;;;AAoCA,SAAgB,qBACd,SACA,WACA,aAAa,KACS;CACtB,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,+CAA+C,QAAQ,OAAO,MAAM,UAAU,OAAO,EACvF;CAEF,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GAAG,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAC3F,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,MAAM,IAAI,KAAK;CACrB,MAAM,YAAY,IAAI,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI;CAEnD,MAAM,OADI,UAAU,KAAK,IAAI,cAAc,CAC9B,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,CAAC;CAChD,OAAO;EACL;EACA;EACA;EACA,gBAAgB;EAChB,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI;EAC7B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI;EAC5B;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EACL,GAAG;EACH,GAAG;EACH,GAAG;EACH,aAAa;EACb,gBAAgB;EAChB,OAAO;EACP,OAAO;EACP;EACA,QAAQ;CACV;CAEF,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,IAAI,IAAI;CACd,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,SAAS,qBAAqB,GAAG,CAAC;CACxC,IAAI,MAAM,GACR,OAAO;EAAE;EAAG;EAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;EAAY;CAAO;CAE9F,MAAM,QAAQ,IAAI;CAClB,MAAM,QAAQ,MAAM,IAAI,IAAI,aAAa,QAAQ,GAAG,GAAG,IAAI,IAAI,CAAC;CAChE,MAAM,SAAS,MAAM,IAAI,IAAI,aAAa,IAAI,QAAQ,GAAG,IAAI,GAAG,IAAI,CAAC;CACrE,MAAM,QAAQ,IAAI;CAClB,OAAO;EACL;EACA;EACA;EACA,aAAa;EACb;EACA,OAAO,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,KAAK;EAC3C,OAAO,KAAK,IAAI,IAAI,IAAI,SAAS,KAAK,KAAK;EAC3C;EACA;CACF;AACF;;;;;AAMA,SAAS,aAAa,GAAW,GAAW,GAAmB;CAC7D,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,0BAA0B,KAAK,GAAG,CAAC,IAAI,GAAG,KAAK;OAC9C,KAAK;CACZ;CACA,QAAQ,KAAK,MAAM;AACrB;;;;;;;;;;AAgCA,SAAS,oBAAoB,GAAW,GAAW,GAAW,OAAuB;CACnF,MAAM,IAAI,IAAI,IAAI;CAClB,MAAM,YAAY,IAAI;CACtB,MAAM,SAAS,EAAE,IAAI,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI;CAClD,MAAM,WAAW,CAAC,IAAI,SAAS,IAAI;CACnC,MAAM,eAAe,SAAS,SAAS,IAAI,YAAY;CACvD,MAAM,OAAO,eAAe,IAAI,KAAK,KAAK,YAAY,IAAI;CAC1D,MAAM,KAAK,CAAC,SAAS,SAAS,IAAI;CAElC,OAAO,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,KAAK,IAAI,IAAI,IAAI,SAAS,CAAC,CAAC;AAChF;;;AAIA,SAAS,WAAW,GAAW,GAAW,GAAW,OAAuB;CAC1E,MAAM,YAAY,IAAI,IAAI,IAAI;CAE9B,MAAM,WAAW,KAAK,IADZ,oBAAoB,GAAG,GAAG,GAAG,KACb,IAAI,SAAS,IAAI;CAC3C,IAAI,EAAE,WAAW,IAAI;EACnB,IAAI,cAAc,GAAG,OAAO;EAC5B,OAAO,YAAY,IAAI,OAAO,oBAAoB,OAAO;CAC3D;CACA,OAAO,YAAY,KAAK,KAAK,QAAQ;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAEhG,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAK5C,IAAI,KAAK;CACT,IAAI,KAAK;CAKT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,GAAG,KAAK;OAClC,KAAK;CACZ;CACA,MAAM,SAAS,KAAK,MAAM;CAG1B,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,MAAM,OAAO;EAC1B,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,CAAC,GAAG,MAAM;OACpC,MAAM;CACb;CACA,MAAM,SAAS,MAAM,OAAO;CAE5B,OAAO;EACL;EACA;EACA;EACA,aAAa,IAAI;EACjB;EACA,OAAO,KAAK,IAAI,IAAI,KAAK;EACzB,OAAO,KAAK,IAAI,GAAG,KAAK;EACxB;CACF;AACF;;;;;;;;;;;;;;;;;;;;AAqBA,SAAgB,kBACd,QACA,OACe;CACf,IAAI,QAAuB;CAC3B,KAAK,MAAM,OAAO,CAAC,QAAQ,KAAK,GAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;EAChC,IAAI,MAAM,GAAG;EACb,IAAI,IAAI,GAAG,OAAO;EAClB,IAAI,UAAU,MAAM,QAAQ;OACvB,IAAI,MAAM,OAAO,OAAO;CAC/B;CAEF,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,QAAQ,GAAW,GAAW,GAAmB;CAC/D,IAAI,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,GACrE,MAAM,IAAI,MAAM,4CAA4C,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAElF,IAAI,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,GACjC,MAAM,IAAI,MAAM,mDAAmD,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAEzF,IAAI,IAAI,IAAI,GAAG,OAAO;CACtB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,QAAQ,IAAI,IAAI;CACrD,OAAO,IAAI;AACb"}

package/dist/{paired-promotion-decision-BiVumgvM.js → paired-promotion-decision-5Pl_qJLX.js}

@@ -1,5 +1,5 @@
-import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, n as pairedBinaryScale } from "./paired-binary-BR0nwHOA.js";
-import { a as pairedSignTest, i as pairedDeltaTieFraction, r as pairedBootstrap } from "./paired-tests-C2zUqgUW.js";
+import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, n as pairedBinaryScale } from "./paired-binary-B477lksS.js";
+import { a as pairedSignTest, i as pairedDeltaTieFraction, r as pairedBootstrap } from "./paired-tests-kASEKpc-.js";
 // src/paired-delta-test.ts
 /** Smallest all-positive sample that can clear a one-sided exact sign test. */
 function minimumPairsForPairedDeltaTest(confidence = .95) {
@@ -253,4 +253,4 @@ function fmt(x) {
 }
 export { pairedDeltaTest as i, pairedDecisionShape as n, minimumPairsForPairedDeltaTest as r, decidePairedPromotion as t };
 
-//# sourceMappingURL=paired-promotion-decision-BiVumgvM.js.map
+//# sourceMappingURL=paired-promotion-decision-5Pl_qJLX.js.map

package/dist/{paired-promotion-decision-BiVumgvM.js.map → paired-promotion-decision-5Pl_qJLX.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-promotion-decision-BiVumgvM.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"}
+{"version":3,"file":"paired-promotion-decision-5Pl_qJLX.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-C2zUqgUW.js → paired-tests-kASEKpc-.js}

@@ -1,39 +1,5 @@
 import { s as ValidationError } from "./errors-DeEEV9Vx.js";
-import { a as medianInPlace, i as makeRng, l as regularizedIncompleteBeta, n as binomialHalfUpperTail, t as assertFiniteSample } from "./internal-v6XfOeJW.js";
-// src/math/student-t.ts
-/**
-* Student-t CDF via the regularized incomplete beta function.
-*/
-function studentTCdf(t, degreesOfFreedom) {
-	if (degreesOfFreedom <= 0) return .5;
-	const x = degreesOfFreedom / (degreesOfFreedom + t * t);
-	const beta = regularizedIncompleteBeta(x, degreesOfFreedom / 2, .5);
-	return t >= 0 ? 1 - .5 * beta : .5 * beta;
-}
-/**
-* Inverse Student-t CDF, solved against {@link studentTCdf}.
-*
-* The CDF is monotone, so bracket expansion followed by bisection is stable
-* across fractional degrees of freedom and does not need a separate
-* approximation with a different error profile.
-*/
-function studentTQuantile(probability, degreesOfFreedom) {
-	if (!Number.isFinite(probability) || probability < 0 || probability > 1) throw new RangeError(`studentTQuantile: probability must be in [0,1], got ${probability}`);
-	if (!Number.isFinite(degreesOfFreedom) || degreesOfFreedom <= 0) throw new RangeError(`studentTQuantile: degreesOfFreedom must be positive and finite, got ${degreesOfFreedom}`);
-	if (probability === 0) return Number.NEGATIVE_INFINITY;
-	if (probability === 1) return Number.POSITIVE_INFINITY;
-	if (probability === .5) return 0;
-	if (probability < .5) return -studentTQuantile(1 - probability, degreesOfFreedom);
-	let low = 0;
-	let high = 1;
-	while (studentTCdf(high, degreesOfFreedom) < probability) high *= 2;
-	for (let iteration = 0; iteration < 64; iteration++) {
-		const middle = (low + high) / 2;
-		if (studentTCdf(middle, degreesOfFreedom) < probability) low = middle;
-		else high = middle;
-	}
-	return (low + high) / 2;
-}
+import { a as binomialHalfUpperTail, c as medianInPlace, d as studentTCdf, i as assertFiniteSample, s as makeRng } from "./mean-DDFhslBt.js";
 // src/statistics/paired-tests.ts
 /**
 * Paired significance tests on continuous scores: the paired t-test, the
@@ -240,6 +206,6 @@ function pairedDeltaTieFraction(before, after) {
 * specifically want outlier robustness and accept the blindness.
 */
 const DECISION_PAIRED_DELTA_STATISTIC = "mean";
-export { pairedSignTest as a, studentTQuantile as c, pairedDeltaTieFraction as i, DECISION_PAIRED_DELTA_STATISTIC as n, pairedTTest as o, pairedBootstrap as r, studentTCdf as s, BOOTSTRAP_GATE_MIN_N as t };
+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-C2zUqgUW.js.map
+//# sourceMappingURL=paired-tests-kASEKpc-.js.map

package/dist/paired-tests-kASEKpc-.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"paired-tests-kASEKpc-.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/{pareto-7_LEmYGG.js → pareto-DY_x3HHg.js}

@@ -1,12 +1,12 @@
 import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-CicJlINA.js";
+import { t as mean } from "./mean-DDFhslBt.js";
 import { t as mulberry32 } from "./random-DrzcgC7F.js";
-import { t as confidenceInterval } from "./descriptive-cPwITVfy.js";
-import { s as wilson } from "./paired-binary-BR0nwHOA.js";
-import { n as DECISION_PAIRED_DELTA_STATISTIC } from "./paired-tests-C2zUqgUW.js";
-import { t as mean } from "./mean-BPuUNznQ.js";
-import { i as pairedDeltaTest, r as minimumPairsForPairedDeltaTest, t as decidePairedPromotion } from "./paired-promotion-decision-BiVumgvM.js";
+import { t as confidenceInterval } from "./descriptive-BTzLL-aj.js";
+import { s as wilson } from "./paired-binary-B477lksS.js";
+import { n as DECISION_PAIRED_DELTA_STATISTIC } from "./paired-tests-kASEKpc-.js";
+import { i as pairedDeltaTest, r as minimumPairsForPairedDeltaTest, t as decidePairedPromotion } from "./paired-promotion-decision-5Pl_qJLX.js";
 import { a as SearchLedgerError } from "./search-ledger-file-DTBD6X-y.js";
-import { s as redactText } from "./redact-D6APJS1d.js";
+import { c as redactText } from "./redact-zp_ILC4l.js";
 import { l as searchUnitScores, p as USD_TOLERANCE, t as FileSearchLedger, u as searchCellId } from "./search-ledger-BsVi-Ngv.js";
 import { t as acquireSingleRunLock } from "./single-run-lock-BGlW3OSk.js";
 // src/statistics/normalize-signed-zero.ts
@@ -2609,4 +2609,4 @@ function paretoFrontierWithCrowding(candidates, objectives) {
 }
 export { normalizeSignedZero as A, SEARCH_ESTIMATOR as C, searchEstimateMethod as D, searchCellSetDigest as E, searchPosterior as O, requiredPairsForPairedPromotion as S, estimateNodeFromCells as T, planSearchClaim as _, searchDivergence as a, verifySearchClaim as b, MIN_OUTCOMES_FOR_WEIGHT as c, rankingSplit as d, summarizeSamples as f, decideSearchClaim as g, SEARCH_CLAIM_RULE as h, runSearch as i, seedFromDigest as k, operatorYield as l, searchExpansionIndex as m, paretoFrontier as n, searchPolicyView as o, SEARCH_KERNEL_SOURCE as p, paretoFrontierWithCrowding as r, EXPANSION_OPERATORS as s, dominates as t, INSUFFICIENT_FROM as u, searchClaimDecision as v, estimateNode as w, pairedPromotionPower as x, searchClaimReserveUsd as y };
 
-//# sourceMappingURL=pareto-7_LEmYGG.js.map
+//# sourceMappingURL=pareto-DY_x3HHg.js.map