@tangle-network/agent-eval 0.40.5 → 0.42.0

package/dist/chunk-MAZ26DC7.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/sequential.ts"],"sourcesContent":["/**\n * Always-valid sequential evaluation.\n *\n * `researchReport` assumes a single pre-specified analysis. Real\n * consumers run campaigns weekly / nightly / per-PR; each new run silently\n * inflates the false-discovery rate, because the BH-FDR guarantee is for\n * the *first* look, not the 47th. Without time-uniform inference,\n * launch-decision teams either (a) don't peek, which forfeits the cost\n * advantage of stop-when-decisive, or (b) peek and pretend they didn't,\n * which forfeits scientific validity.\n *\n * This module ships **e-value-based confidence sequences** for paired\n * bounded outcomes. The methodology is the predictable plug-in betting\n * martingale of Waudby-Smith & Ramdas (2024) — provably valid at *any*\n * stopping time. Concretely:\n *\n *   For paired deltas D_1, D_2, … ∈ [-c, c] with the null H_0: E[D] ≤ 0,\n *   a betting fraction λ_i is chosen using only D_{1..i-1} (predictable\n *   plug-in), and the running e-value is\n *\n *     E_t = ∏_{i=1}^{t} (1 + λ_i · D_i)\n *\n *   E_t is a non-negative martingale under H_0 with E[E_t] ≤ 1, so by\n *   Ville's inequality, P(∃ t : E_t ≥ 1/α) ≤ α — we can reject the null\n *   at any time without inflating the type-I error.\n *\n * Combined with `runEvalCampaign`, every consumer running rolling\n * campaigns gains the ability to ship the moment evidence is decisive,\n * stop-early on dead-on-arrival variants, and accumulate evidence across\n * partial runs without spending the FDR budget. No new sweep is wasted.\n *\n * References:\n *   - Howard, S. R., Ramdas, A., McAuliffe, J., Sekhon, J. (2021).\n *     Time-uniform, nonparametric, nonasymptotic confidence sequences.\n *     Annals of Statistics, 49(2), 1055–1080.\n *   - Waudby-Smith, I., Ramdas, A. (2024). Estimating means of bounded\n *     random variables by betting. JRSS B, 86(1), 1–27.\n */\n\nexport type SequentialDecision = 'promote_now' | 'continue' | 'reject_now' | 'equivalent'\n\nexport interface PairedEvalueOptions {\n  /**\n   * Bound on |delta|. Default 1 (matching most score scales). Must satisfy\n   * c > 0; deltas outside [-c, c] are clipped with a warning attached to\n   * the return value.\n   */\n  bound?: number\n  /** Target Type-I error. Default 0.05. */\n  alpha?: number\n  /**\n   * Region of Practical Equivalence on the *mean* paired delta. When\n   * supplied, the verdict can return `'equivalent'` once the running\n   * confidence sequence on the mean is fully contained in [low, high].\n   */\n  rope?: { low: number; high: number }\n  /** Initial bet shrinkage (0 < scale ≤ 1). Default 0.5 — empirically robust. */\n  initialBetShrinkage?: number\n}\n\nexport interface PairedEvalueStep {\n  /** 1-indexed observation count. */\n  t: number\n  delta: number\n  /** Running e-value E_t = ∏ (1 + λ_i · D_i). */\n  evalue: number\n  /** Time-uniform p-value at stopping time t. */\n  pValue: number\n  /** Lower bound of the empirical Bernstein confidence sequence at level 1-α. */\n  csLow: number\n  csHigh: number\n  /** Verdict at this stopping time. */\n  decision: SequentialDecision\n}\n\nexport interface PairedEvalueSequence {\n  steps: PairedEvalueStep[]\n  /** The decision at the final step. */\n  finalDecision: SequentialDecision\n  /** Index (1-based) at which a non-`continue` decision first fired, or null. */\n  decisionFiredAt: number | null\n  /** True if any deltas were clipped to [-bound, bound]. */\n  clipped: boolean\n}\n\n/**\n * Run the paired e-value sequence over an in-order delta stream.\n *\n * Use for *streaming* / interim analyses: pass the deltas you have so\n * far, get the verdict at every prefix length. The decision is\n * monotone-stable in the sense that once `'reject_now'` or `'promote_now'`\n * fires, the verdict at later steps remains decisive (the e-value is a\n * non-negative martingale; once it crosses the threshold, it's crossed).\n */\nexport function pairedEvalueSequence(\n  deltas: number[],\n  opts: PairedEvalueOptions = {},\n): PairedEvalueSequence {\n  const c = opts.bound ?? 1\n  const alpha = opts.alpha ?? 0.05\n  const initialShrink = opts.initialBetShrinkage ?? 0.5\n  const rope = opts.rope ?? null\n  if (c <= 0) throw new Error('pairedEvalueSequence: bound must be > 0')\n  if (alpha <= 0 || alpha >= 1) throw new Error('pairedEvalueSequence: alpha must be in (0,1)')\n  if (rope && !(Number.isFinite(rope.low) && Number.isFinite(rope.high) && rope.low <= rope.high)) {\n    throw new Error('pairedEvalueSequence: rope must satisfy low ≤ high')\n  }\n\n  const steps: PairedEvalueStep[] = []\n  let clipped = false\n  let evalue = 1\n  let decisionFiredAt: number | null = null\n\n  // Running statistics (using only D_{1..i-1} for the bet → predictable plug-in).\n  let sum = 0\n  let sumSq = 0\n  let count = 0\n\n  for (let i = 0; i < deltas.length; i++) {\n    let d = deltas[i]!\n    if (d < -c || d > c) {\n      d = Math.max(-c, Math.min(c, d))\n      clipped = true\n    }\n\n    // Predictable plug-in bet (positive λ tests for E[D] > 0; we run a two-sided\n    // test by tracking the symmetric e-value via |bet|).\n    // λ_i ∝ mean / (variance + bound^2). Shrink early to avoid overbetting.\n    const muHat = count === 0 ? 0 : sum / count\n    const varHat = count === 0 ? c * c : Math.max(1e-12, sumSq / count - muHat * muHat)\n    const t = i + 1\n    const shrink = initialShrink * Math.min(1, count / 32) // anneal toward 1\n    let lambda = (muHat / (varHat + c * c)) * shrink\n    // Clip to ensure 1 + λ·D > 0 for all |D| ≤ c (so the e-value stays non-negative).\n    const lambdaMax = 0.99 / c\n    if (lambda > lambdaMax) lambda = lambdaMax\n    if (lambda < -lambdaMax) lambda = -lambdaMax\n\n    evalue = evalue * (1 + lambda * d)\n    if (!Number.isFinite(evalue) || evalue < 0) evalue = 0\n\n    sum += d\n    sumSq += d * d\n    count += 1\n\n    const pValue = Math.min(1, 1 / Math.max(evalue, 1e-300))\n\n    // Empirical Bernstein confidence sequence on the mean. Howard et al.\n    // (2021), Theorem 4.4 with σ̂² the running sample variance and a\n    // calibration constant tuned for two-sided coverage at level 1 - α.\n    const cs = empiricalBernsteinCs(sum, sumSq, count, c, alpha)\n\n    let decision: SequentialDecision = 'continue'\n    if (rope && cs.low >= rope.low && cs.high <= rope.high) decision = 'equivalent'\n    else if (evalue >= 2 / alpha && muHat > 0) decision = 'promote_now'\n    else if (evalue >= 2 / alpha && muHat < 0) decision = 'reject_now'\n    else if (rope && cs.high < rope.low) decision = 'reject_now'\n\n    if (decision !== 'continue' && decisionFiredAt === null) decisionFiredAt = t\n\n    steps.push({ t, delta: d, evalue, pValue, csLow: cs.low, csHigh: cs.high, decision })\n  }\n\n  const finalDecision = steps.length === 0 ? 'continue' : steps[steps.length - 1]!.decision\n  return { steps, finalDecision, decisionFiredAt, clipped }\n}\n\nexport interface InterimReleaseConfidenceInput {\n  /**\n   * One delta series per candidate (paired deltas vs comparator). Order\n   * within a series is the order the campaigns were run.\n   */\n  deltaSeries: Array<{ candidateId: string; deltas: number[] }>\n  alpha?: number\n  bound?: number\n  rope?: { low: number; high: number }\n}\n\nexport interface InterimReleaseConfidence {\n  candidates: Array<{\n    candidateId: string\n    decision: SequentialDecision\n    decisionFiredAt: number | null\n    finalEvalue: number\n    finalPValue: number\n    pairs: number\n    csLow: number\n    csHigh: number\n  }>\n  /**\n   * Campaign-level recommendation: pick the strongest 'promote_now', else\n   * 'continue' if any candidate is still live, else 'reject_now' if every\n   * candidate is dead, else 'equivalent'.\n   */\n  recommendation: { decision: SequentialDecision; candidateId: string | null }\n}\n\n/**\n * Run interim sequential analyses across many candidates at once,\n * preserving the time-uniform α guarantee for each candidate's series and\n * synthesising a campaign-level recommendation. Designed to be called on\n * every campaign tick — the recommendation is anytime-valid.\n */\nexport function evaluateInterimReleaseConfidence(\n  input: InterimReleaseConfidenceInput,\n): InterimReleaseConfidence {\n  const candidates = input.deltaSeries.map((s) => {\n    const seq = pairedEvalueSequence(s.deltas, {\n      alpha: input.alpha,\n      bound: input.bound,\n      rope: input.rope,\n    })\n    const last = seq.steps[seq.steps.length - 1]\n    return {\n      candidateId: s.candidateId,\n      decision: seq.finalDecision,\n      decisionFiredAt: seq.decisionFiredAt,\n      finalEvalue: last?.evalue ?? 1,\n      finalPValue: last?.pValue ?? 1,\n      pairs: seq.steps.length,\n      csLow: last?.csLow ?? Number.NEGATIVE_INFINITY,\n      csHigh: last?.csHigh ?? Number.POSITIVE_INFINITY,\n    }\n  })\n\n  const promote = candidates.find((c) => c.decision === 'promote_now')\n  if (promote)\n    return {\n      candidates,\n      recommendation: { decision: 'promote_now', candidateId: promote.candidateId },\n    }\n  const live = candidates.find((c) => c.decision === 'continue')\n  if (live) return { candidates, recommendation: { decision: 'continue', candidateId: null } }\n  const equiv = candidates.find((c) => c.decision === 'equivalent')\n  if (equiv)\n    return {\n      candidates,\n      recommendation: { decision: 'equivalent', candidateId: equiv.candidateId },\n    }\n  return { candidates, recommendation: { decision: 'reject_now', candidateId: null } }\n}\n\n// ── Internals ────────────────────────────────────────────────────────────\n\n/**\n * Empirical Bernstein confidence sequence on the mean of bounded variables.\n * Adapted from Howard et al. (2021) §4.4. Provides a time-uniform CI on\n * the running mean; valid at every stopping time.\n */\nfunction empiricalBernsteinCs(\n  sum: number,\n  sumSq: number,\n  n: number,\n  bound: number,\n  alpha: number,\n): { low: number; high: number } {\n  if (n === 0) return { low: -bound, high: bound }\n  const mean = sum / n\n  const variance = Math.max(0, sumSq / n - mean * mean)\n  // Iterated-log calibration constant. The 1.7 exponent matches the\n  // recommended choice in Howard et al. for two-sided coverage at level\n  // 1 - α with mild log-corrections; tightening further requires a\n  // tuned mixture and is out of scope.\n  const psi = Math.log(2 / alpha) + 1.7 * Math.log(Math.log(Math.max(Math.E, n)) + 1)\n  const radius = Math.sqrt((2 * variance * psi) / n) + (3 * bound * psi) / n\n  return { low: mean - radius, high: mean + radius }\n}\n"],"mappings":";AA8FO,SAAS,qBACd,QACA,OAA4B,CAAC,GACP;AACtB,QAAM,IAAI,KAAK,SAAS;AACxB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,gBAAgB,KAAK,uBAAuB;AAClD,QAAM,OAAO,KAAK,QAAQ;AAC1B,MAAI,KAAK,EAAG,OAAM,IAAI,MAAM,yCAAyC;AACrE,MAAI,SAAS,KAAK,SAAS,EAAG,OAAM,IAAI,MAAM,8CAA8C;AAC5F,MAAI,QAAQ,EAAE,OAAO,SAAS,KAAK,GAAG,KAAK,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,OAAO,KAAK,OAAO;AAC/F,UAAM,IAAI,MAAM,yDAAoD;AAAA,EACtE;AAEA,QAAM,QAA4B,CAAC;AACnC,MAAI,UAAU;AACd,MAAI,SAAS;AACb,MAAI,kBAAiC;AAGrC,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AAEZ,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,QAAI,IAAI,OAAO,CAAC;AAChB,QAAI,IAAI,CAAC,KAAK,IAAI,GAAG;AACnB,UAAI,KAAK,IAAI,CAAC,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AAC/B,gBAAU;AAAA,IACZ;AAKA,UAAM,QAAQ,UAAU,IAAI,IAAI,MAAM;AACtC,UAAM,SAAS,UAAU,IAAI,IAAI,IAAI,KAAK,IAAI,OAAO,QAAQ,QAAQ,QAAQ,KAAK;AAClF,UAAM,IAAI,IAAI;AACd,UAAM,SAAS,gBAAgB,KAAK,IAAI,GAAG,QAAQ,EAAE;AACrD,QAAI,SAAU,SAAS,SAAS,IAAI,KAAM;AAE1C,UAAM,YAAY,OAAO;AACzB,QAAI,SAAS,UAAW,UAAS;AACjC,QAAI,SAAS,CAAC,UAAW,UAAS,CAAC;AAEnC,aAAS,UAAU,IAAI,SAAS;AAChC,QAAI,CAAC,OAAO,SAAS,MAAM,KAAK,SAAS,EAAG,UAAS;AAErD,WAAO;AACP,aAAS,IAAI;AACb,aAAS;AAET,UAAM,SAAS,KAAK,IAAI,GAAG,IAAI,KAAK,IAAI,QAAQ,MAAM,CAAC;AAKvD,UAAM,KAAK,qBAAqB,KAAK,OAAO,OAAO,GAAG,KAAK;AAE3D,QAAI,WAA+B;AACnC,QAAI,QAAQ,GAAG,OAAO,KAAK,OAAO,GAAG,QAAQ,KAAK,KAAM,YAAW;AAAA,aAC1D,UAAU,IAAI,SAAS,QAAQ,EAAG,YAAW;AAAA,aAC7C,UAAU,IAAI,SAAS,QAAQ,EAAG,YAAW;AAAA,aAC7C,QAAQ,GAAG,OAAO,KAAK,IAAK,YAAW;AAEhD,QAAI,aAAa,cAAc,oBAAoB,KAAM,mBAAkB;AAE3E,UAAM,KAAK,EAAE,GAAG,OAAO,GAAG,QAAQ,QAAQ,OAAO,GAAG,KAAK,QAAQ,GAAG,MAAM,SAAS,CAAC;AAAA,EACtF;AAEA,QAAM,gBAAgB,MAAM,WAAW,IAAI,aAAa,MAAM,MAAM,SAAS,CAAC,EAAG;AACjF,SAAO,EAAE,OAAO,eAAe,iBAAiB,QAAQ;AAC1D;AAsCO,SAAS,iCACd,OAC0B;AAC1B,QAAM,aAAa,MAAM,YAAY,IAAI,CAAC,MAAM;AAC9C,UAAM,MAAM,qBAAqB,EAAE,QAAQ;AAAA,MACzC,OAAO,MAAM;AAAA,MACb,OAAO,MAAM;AAAA,MACb,MAAM,MAAM;AAAA,IACd,CAAC;AACD,UAAM,OAAO,IAAI,MAAM,IAAI,MAAM,SAAS,CAAC;AAC3C,WAAO;AAAA,MACL,aAAa,EAAE;AAAA,MACf,UAAU,IAAI;AAAA,MACd,iBAAiB,IAAI;AAAA,MACrB,aAAa,MAAM,UAAU;AAAA,MAC7B,aAAa,MAAM,UAAU;AAAA,MAC7B,OAAO,IAAI,MAAM;AAAA,MACjB,OAAO,MAAM,SAAS,OAAO;AAAA,MAC7B,QAAQ,MAAM,UAAU,OAAO;AAAA,IACjC;AAAA,EACF,CAAC;AAED,QAAM,UAAU,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,aAAa;AACnE,MAAI;AACF,WAAO;AAAA,MACL;AAAA,MACA,gBAAgB,EAAE,UAAU,eAAe,aAAa,QAAQ,YAAY;AAAA,IAC9E;AACF,QAAM,OAAO,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU;AAC7D,MAAI,KAAM,QAAO,EAAE,YAAY,gBAAgB,EAAE,UAAU,YAAY,aAAa,KAAK,EAAE;AAC3F,QAAM,QAAQ,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,YAAY;AAChE,MAAI;AACF,WAAO;AAAA,MACL;AAAA,MACA,gBAAgB,EAAE,UAAU,cAAc,aAAa,MAAM,YAAY;AAAA,IAC3E;AACF,SAAO,EAAE,YAAY,gBAAgB,EAAE,UAAU,cAAc,aAAa,KAAK,EAAE;AACrF;AASA,SAAS,qBACP,KACA,OACA,GACA,OACA,OAC+B;AAC/B,MAAI,MAAM,EAAG,QAAO,EAAE,KAAK,CAAC,OAAO,MAAM,MAAM;AAC/C,QAAM,OAAO,MAAM;AACnB,QAAM,WAAW,KAAK,IAAI,GAAG,QAAQ,IAAI,OAAO,IAAI;AAKpD,QAAM,MAAM,KAAK,IAAI,IAAI,KAAK,IAAI,MAAM,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,CAAC,CAAC,IAAI,CAAC;AAClF,QAAM,SAAS,KAAK,KAAM,IAAI,WAAW,MAAO,CAAC,IAAK,IAAI,QAAQ,MAAO;AACzE,SAAO,EAAE,KAAK,OAAO,QAAQ,MAAM,OAAO,OAAO;AACnD;","names":[]}

package/dist/chunk-NKLGKF2Q.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"sourcesContent":["/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n *   - Dataset manifests prove corpus/version coverage.\n *   - RunRecord rows prove reproducible search/holdout outcomes.\n *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n *   - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport type { ActionableSideInfo, MultiShotTrialResult } from './multi-shot-optimization'\nimport type { RunRecord, RunSplitTag } from './run-record'\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n  | 'corpus'\n  | 'quality'\n  | 'generalization'\n  | 'diagnostics'\n  | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n  scenarioId: string\n  candidateId?: string\n  split?: RunSplitTag\n  score?: number\n  ok?: boolean\n  turnCount?: number\n  costUsd?: number\n  durationMs?: number\n  failureMode?: string\n  asi?: ActionableSideInfo[]\n  metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n  /** Require a Dataset manifest or explicit scenarios. Default true. */\n  requireCorpus?: boolean\n  minScenarioCount?: number\n  minSearchRuns?: number\n  minHoldoutRuns?: number\n  /** Require at least one holdout scenario/run. Default true. */\n  requireHoldout?: boolean\n  minPassRate?: number\n  minMeanScore?: number\n  /** Search mean may exceed holdout mean by at most this much. */\n  maxOverfitGap?: number\n  maxMeanCostUsd?: number\n  maxP95WallMs?: number\n  /** Low-score/failed rows must carry ASI. Default true. */\n  requireAsiForFailures?: boolean\n  /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n  failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n  target: string\n  candidateId?: string\n  baselineId?: string\n  dataset?: DatasetManifest\n  scenarios?: readonly DatasetScenario[]\n  runs?: readonly RunRecord[]\n  traces?: readonly ReleaseTraceEvidence[]\n  gateDecision?: GateDecision | null\n  thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n  name: ReleaseConfidenceAxisName\n  status: ReleaseConfidenceStatus\n  score: number\n  detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n  axis: ReleaseConfidenceAxisName\n  severity: 'critical' | 'warning'\n  code: string\n  detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n  scenarioCount: number\n  searchRuns: number\n  holdoutRuns: number\n  passRate: number\n  meanScore: number\n  searchMeanScore: number\n  holdoutMeanScore: number\n  overfitGap: number\n  meanCostUsd: number\n  p95WallMs: number\n  failedRows: number\n  failuresWithAsi: number\n  singleShotTraces: number\n  multiShotTraces: number\n  splitCounts: Record<DatasetSplit, number>\n  domainCounts: Record<string, number>\n  failureModeCounts: Record<string, number>\n  responsibleSurfaceCounts: Record<string, number>\n}\n\nexport interface ReleaseConfidenceScorecard {\n  target: string\n  candidateId: string | null\n  baselineId: string | null\n  status: ReleaseConfidenceStatus\n  promote: boolean\n  axes: ReleaseConfidenceAxis[]\n  issues: ReleaseConfidenceIssue[]\n  metrics: ReleaseConfidenceMetrics\n  dataset: DatasetManifest | null\n  gateDecision: GateDecision | null\n  summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n  requireCorpus: true,\n  minScenarioCount: 1,\n  minSearchRuns: 1,\n  minHoldoutRuns: 1,\n  requireHoldout: true,\n  minPassRate: 0.8,\n  minMeanScore: 0.7,\n  maxOverfitGap: 0.15,\n  maxMeanCostUsd: Number.POSITIVE_INFINITY,\n  maxP95WallMs: Number.POSITIVE_INFINITY,\n  requireAsiForFailures: true,\n  failureScoreThreshold: 0.5,\n}\n\nexport function releaseTraceEvidenceFromMultiShotTrials(\n  trials: readonly MultiShotTrialResult[],\n): ReleaseTraceEvidence[] {\n  return trials.map((trial) => ({\n    scenarioId: trial.scenarioId,\n    candidateId: trial.variantId,\n    split: trial.split === 'holdout' ? 'holdout' : trial.split === 'dev' ? 'dev' : 'search',\n    score: trial.score,\n    ok: trial.ok,\n    turnCount: Array.isArray(trial.trace?.turns) ? trial.trace.turns.length : undefined,\n    costUsd: trial.cost,\n    durationMs: trial.durationMs,\n    failureMode: trial.error ? 'runtime_error' : undefined,\n    asi: trial.asi,\n    metadata: trial.metadata,\n  }))\n}\n\nexport function evaluateReleaseConfidence(\n  input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n  const candidateId = input.candidateId ?? null\n  const runs = filterCandidate(input.runs ?? [], candidateId, input.baselineId)\n  const traces = filterTraceCandidate(input.traces ?? [], candidateId, input.baselineId)\n  const scenarios = input.scenarios ?? []\n  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n  const searchScores = scoresFor(runs, 'search')\n  const holdoutScores = scoresFor(runs, 'holdout')\n  const allScores = [...searchScores, ...holdoutScores]\n  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n  const scoreUniverse = allScores.length > 0 ? allScores : traceScores\n  const searchRuns = runs.filter((r) => r.splitTag === 'search').length\n  const holdoutRuns = runs.filter((r) => r.splitTag === 'holdout').length\n  const searchMeanScore = mean(searchScores)\n  const holdoutMeanScore = mean(holdoutScores)\n  const metrics: ReleaseConfidenceMetrics = {\n    scenarioCount,\n    searchRuns,\n    holdoutRuns,\n    passRate: passRate(runs, traces, thresholds.failureScoreThreshold),\n    meanScore: mean(scoreUniverse),\n    searchMeanScore,\n    holdoutMeanScore,\n    overfitGap: safeDiff(searchMeanScore, holdoutMeanScore),\n    meanCostUsd: mean([\n      ...runs.map((r) => r.costUsd),\n      ...traces.map((t) => t.costUsd).filter(isFiniteNumber),\n    ]),\n    p95WallMs: percentile(\n      [...runs.map((r) => r.wallMs), ...traces.map((t) => t.durationMs).filter(isFiniteNumber)],\n      0.95,\n    ),\n    failedRows: failedRows(runs, traces, thresholds.failureScoreThreshold).length,\n    failuresWithAsi: failedRows(runs, traces, thresholds.failureScoreThreshold).filter(\n      (row) => row.hasAsi,\n    ).length,\n    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n    splitCounts,\n    domainCounts: countDomains(scenarios),\n    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),\n    responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n  }\n\n  const issues: ReleaseConfidenceIssue[] = []\n  checkCorpus(input, thresholds, metrics, issues)\n  checkQuality(thresholds, metrics, issues)\n  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n  checkDiagnostics(thresholds, metrics, issues)\n  checkEfficiency(thresholds, metrics, issues)\n\n  const axes = buildAxes(metrics, thresholds, input.gateDecision ?? null, issues)\n  const status = issues.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : issues.length > 0\n      ? 'warn'\n      : 'pass'\n\n  return {\n    target: input.target,\n    candidateId,\n    baselineId: input.baselineId ?? null,\n    status,\n    promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n    axes,\n    issues,\n    metrics,\n    dataset: input.dataset ?? null,\n    gateDecision: input.gateDecision ?? null,\n    summary: renderSummary(input.target, status, metrics, issues),\n  }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n  const scorecard = evaluateReleaseConfidence(input)\n  if (scorecard.status === 'fail') {\n    throw new VerificationError(scorecard.summary)\n  }\n  return scorecard\n}\n\nfunction filterCandidate(\n  runs: readonly RunRecord[],\n  candidateId: string | null,\n  baselineId?: string,\n): RunRecord[] {\n  if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n  return [...runs]\n}\n\nfunction filterTraceCandidate(\n  traces: readonly ReleaseTraceEvidence[],\n  candidateId: string | null,\n  baselineId?: string,\n): ReleaseTraceEvidence[] {\n  if (candidateId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n  if (baselineId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n  return [...traces]\n}\n\nfunction checkCorpus(\n  input: ReleaseConfidenceInput,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_corpus',\n      detail: 'No Dataset manifest or scenarios supplied.',\n    })\n  }\n  if (metrics.scenarioCount < thresholds.minScenarioCount) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'few_scenarios',\n      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n    })\n  }\n  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_holdout_split',\n      detail: 'Corpus has no holdout scenarios.',\n    })\n  }\n}\n\nfunction checkQuality(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.searchRuns < thresholds.minSearchRuns) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'few_search_runs',\n      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n    })\n  }\n  if (metrics.passRate < thresholds.minPassRate) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_pass_rate',\n      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n    })\n  }\n  if (metrics.meanScore < thresholds.minMeanScore) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_mean_score',\n      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n    })\n  }\n}\n\nfunction checkGeneralization(\n  gateDecision: GateDecision | null,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'few_holdout_runs',\n      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n    })\n  }\n  if (Number.isFinite(metrics.overfitGap) && metrics.overfitGap > thresholds.maxOverfitGap) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'overfit_gap',\n      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n    })\n  }\n  if (gateDecision && !gateDecision.promote) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n      detail: gateDecision.reason,\n    })\n  }\n}\n\nfunction checkDiagnostics(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (!thresholds.requireAsiForFailures) return\n  if (metrics.failedRows > metrics.failuresWithAsi) {\n    issues.push({\n      axis: 'diagnostics',\n      severity: 'critical',\n      code: 'missing_failure_asi',\n      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n    })\n  }\n}\n\nfunction checkEfficiency(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'cost_budget',\n      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n    })\n  }\n  if (metrics.p95WallMs > thresholds.maxP95WallMs) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'latency_budget',\n      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n    })\n  }\n}\n\nfunction buildAxes(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  gateDecision: GateDecision | null,\n  issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n  return [\n    axis(\n      'corpus',\n      issues,\n      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n    ),\n    axis(\n      'quality',\n      issues,\n      Math.min(metrics.passRate, metrics.meanScore),\n      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n    ),\n    axis(\n      'generalization',\n      issues,\n      gateDecision && !gateDecision.promote\n        ? 0\n        : gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n    ),\n    axis(\n      'diagnostics',\n      issues,\n      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n    ),\n    axis(\n      'efficiency',\n      issues,\n      efficiencyScore(metrics, thresholds),\n      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n    ),\n  ]\n}\n\nfunction axis(\n  name: ReleaseConfidenceAxisName,\n  issues: ReleaseConfidenceIssue[],\n  score: number,\n  detail: string,\n): ReleaseConfidenceAxis {\n  const own = issues.filter((i) => i.axis === name)\n  const status = own.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : own.length > 0\n      ? 'warn'\n      : 'pass'\n  return { name, status, score: bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n  const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n  for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n  return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const scenario of scenarios) {\n    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n    out[domain] = (out[domain] ?? 0) + 1\n  }\n  return out\n}\n\nfunction countFailureModes(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const mode = run.failureMode ?? 'low_score'\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      const mode = trace.failureMode ?? (trace.ok === false ? 'not_ok' : 'low_score')\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const trace of traces) {\n    for (const asi of trace.asi ?? []) {\n      const surface = asi.responsibleSurface ?? 'unknown'\n      out[surface] = (out[surface] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction failedRows(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Array<{ hasAsi: boolean }> {\n  const out: Array<{ hasAsi: boolean }> = []\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const asiMetric = run.outcome.raw.asi\n      out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n    }\n  }\n  return out\n}\n\nfunction passRate(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): number {\n  const outcomes = [\n    ...runs.map((run) => {\n      const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n      return !run.failureMode && score !== undefined && score >= threshold\n    }),\n    ...traces.map(\n      (trace) => trace.ok !== false && (trace.score === undefined || trace.score >= threshold),\n    ),\n  ]\n  if (outcomes.length === 0) return 0\n  return outcomes.filter(Boolean).length / outcomes.length\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n  return runs\n    .filter((run) => run.splitTag === split)\n    .map((run) => (split === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore))\n    .filter(isFiniteNumber)\n}\n\nfunction mean(xs: readonly number[]): number {\n  if (xs.length === 0) return Number.NaN\n  return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentile(xs: readonly number[], p: number): number {\n  if (xs.length === 0) return Number.NaN\n  const sorted = [...xs].sort((a, b) => a - b)\n  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction safeDiff(a: number, b: number): number {\n  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN\n  return a - b\n}\n\nfunction gapScore(gap: number, maxGap: number): number {\n  if (!Number.isFinite(gap)) return 0\n  if (maxGap <= 0) return gap <= 0 ? 1 : 0\n  return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n): number {\n  const cost =\n    Number.isFinite(thresholds.maxMeanCostUsd) && Number.isFinite(metrics.meanCostUsd)\n      ? bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n      : 1\n  const latency =\n    Number.isFinite(thresholds.maxP95WallMs) && Number.isFinite(metrics.p95WallMs)\n      ? bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n      : 1\n  return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n  target: string,\n  status: ReleaseConfidenceStatus,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): string {\n  const prefix = `release confidence ${status}: ${target}`\n  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n  if (issues.length === 0) return `${prefix}; ${metricText}`\n  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n *   - Bootstrap mean baseline vs candidate (1k resamples).\n *   - Compute the delta distribution; pass if the lower CI bound > 0.\n *   - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)\n *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)\n *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n  baselineMean: number\n  candidateMean: number\n  /** candidateMean - baselineMean, point estimate. */\n  delta: number\n  /** Lower bound of the (1 - alpha) CI on the delta. */\n  ciLower: number\n  /** Upper bound of the (1 - alpha) CI on the delta. */\n  ciUpper: number\n  /** Number of bootstrap resamples used. */\n  iterations: number\n  alpha: number\n  verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n  /** Confidence level alpha (default 0.05 → 95% CI). */\n  alpha?: number\n  /** Number of resamples (default 1000). */\n  iterations?: number\n  /**\n   * Minimum total samples (baseline + candidate) below which we always\n   * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n   * Default 6 (combined).\n   */\n  minTotalSamples?: number\n  /** RNG seed for reproducibility. Default: Math.random. */\n  seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n  baseline: number[],\n  candidate: number[],\n  options: BootstrapOptions = {},\n): BootstrapResult {\n  const alpha = options.alpha ?? 0.05\n  const iterations = options.iterations ?? 1000\n  const minTotal = options.minTotalSamples ?? 6\n  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n  const baselineMean = mean(baseline)\n  const candidateMean = mean(candidate)\n  const delta = candidateMean - baselineMean\n\n  if (\n    baseline.length + candidate.length < minTotal ||\n    baseline.length === 0 ||\n    candidate.length === 0\n  ) {\n    return {\n      baselineMean,\n      candidateMean,\n      delta,\n      ciLower: -Infinity,\n      ciUpper: Infinity,\n      iterations: 0,\n      alpha,\n      verdict: 'INCONCLUSIVE',\n    }\n  }\n\n  const deltas: number[] = new Array(iterations)\n  for (let i = 0; i < iterations; i++) {\n    const bResample = resample(baseline, rng)\n    const cResample = resample(candidate, rng)\n    deltas[i] = mean(cResample) - mean(bResample)\n  }\n  deltas.sort((a, b) => a - b)\n  const lowerIdx = Math.floor((alpha / 2) * iterations)\n  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n  const ciLower = deltas[Math.max(0, lowerIdx)]!\n  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n  let verdict: Verdict\n  if (ciLower > 0) verdict = 'ADVANCE'\n  else if (ciUpper < 0) verdict = 'REVERT'\n  else if (delta >= 0) verdict = 'KEEP'\n  else verdict = 'INCONCLUSIVE'\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ciLower,\n    ciUpper,\n    iterations,\n    alpha,\n    verdict,\n  }\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  let s = 0\n  for (const x of xs) s += x\n  return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n  const out = new Array(xs.length)\n  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n  return out\n}\n\n/** Mulberry32 — fast deterministic PRNG. Stable across runs given the same seed. */\nfunction mulberry32(seed: number): () => number {\n  let t = seed >>> 0\n  return () => {\n    t += 0x6d2b79f5\n    let r = t\n    r = Math.imul(r ^ (r >>> 15), r | 1)\n    r ^= r + Math.imul(r ^ (r >>> 7), r | 61)\n    return ((r ^ (r >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n  let h = 2166136261\n  for (const x of [...a, ...b]) {\n    const view = new Float64Array([x])\n    const bytes = new Uint8Array(view.buffer)\n    for (const byte of bytes) {\n      h ^= byte\n      h = Math.imul(h, 16777619)\n    }\n  }\n  return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n  baselineOutputs: TOutput[]\n  candidateOutputs: TOutput[]\n  /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n  judge: (output: TOutput) => Promise<number> | number\n  alpha?: number\n  iterations?: number\n  /** RNG seed for reproducibility. */\n  seed?: number\n  /** Maximum concurrent judge calls. Default 4. */\n  judgeConcurrency?: number\n}\n\nexport async function judgeReplayGate<TOutput>(\n  args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n  const concurrency = args.judgeConcurrency ?? 4\n  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n  const ci = bootstrapCi(baselineScores, candidateScores, {\n    ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n    ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n    ...(args.seed !== undefined ? { seed: args.seed } : {}),\n  })\n  return {\n    ...ci,\n    baselineSamples: baselineScores.length,\n    candidateSamples: candidateScores.length,\n  }\n}\n\nasync function scoreAll<TOutput>(\n  outputs: TOutput[],\n  judge: (output: TOutput) => Promise<number> | number,\n  concurrency: number,\n): Promise<number[]> {\n  const results: number[] = new Array(outputs.length)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= outputs.length) return\n      const v = await judge(outputs[i]!)\n      results[i] = Number.isFinite(v) ? v : 0\n    }\n  }\n  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n  return results\n}\n","import type { ReleaseConfidenceScorecard } from './release-confidence'\nimport type { RunRecord } from './run-record'\nimport { summaryTable } from './summary-report'\n\nexport interface RenderReleaseReportOptions {\n  title?: string\n  runs?: readonly RunRecord[]\n  comparator?: string\n  traceAnalystFindings?: readonly string[]\n  nextActions?: readonly string[]\n}\n\nexport function renderReleaseReport(\n  scorecard: ReleaseConfidenceScorecard,\n  options: RenderReleaseReportOptions = {},\n): string {\n  const title = options.title ?? `Release Report: ${scorecard.target}`\n  const lines: string[] = []\n  lines.push(`# ${title}`)\n  lines.push('')\n  lines.push(`Status: **${scorecard.status.toUpperCase()}**`)\n  lines.push(`Promote: **${scorecard.promote ? 'yes' : 'no'}**`)\n  if (scorecard.candidateId) lines.push(`Candidate: \\`${scorecard.candidateId}\\``)\n  if (scorecard.baselineId) lines.push(`Baseline: \\`${scorecard.baselineId}\\``)\n  lines.push('')\n  lines.push(scorecard.summary)\n  lines.push('')\n\n  lines.push('## Metrics')\n  lines.push('')\n  lines.push('| Metric | Value |')\n  lines.push('|---|---:|')\n  lines.push(`| Scenarios | ${scorecard.metrics.scenarioCount} |`)\n  lines.push(`| Search runs | ${scorecard.metrics.searchRuns} |`)\n  lines.push(`| Holdout runs | ${scorecard.metrics.holdoutRuns} |`)\n  lines.push(`| Pass rate | ${pct(scorecard.metrics.passRate)} |`)\n  lines.push(`| Mean score | ${num(scorecard.metrics.meanScore)} |`)\n  lines.push(`| Search mean | ${num(scorecard.metrics.searchMeanScore)} |`)\n  lines.push(`| Holdout mean | ${num(scorecard.metrics.holdoutMeanScore)} |`)\n  lines.push(`| Overfit gap | ${num(scorecard.metrics.overfitGap)} |`)\n  lines.push(`| Mean cost | $${num(scorecard.metrics.meanCostUsd)} |`)\n  lines.push(`| p95 wall time | ${Math.round(scorecard.metrics.p95WallMs)} ms |`)\n  lines.push('')\n\n  if (scorecard.issues.length > 0) {\n    lines.push('## Issues')\n    lines.push('')\n    for (const issue of scorecard.issues) {\n      lines.push(`- **${issue.severity}** \\`${issue.code}\\` (${issue.axis}): ${issue.detail}`)\n    }\n    lines.push('')\n  }\n\n  const surfaces = entries(scorecard.metrics.responsibleSurfaceCounts)\n  if (surfaces.length > 0) {\n    lines.push('## Responsible Surfaces')\n    lines.push('')\n    for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`)\n    lines.push('')\n  }\n\n  const failures = entries(scorecard.metrics.failureModeCounts)\n  if (failures.length > 0) {\n    lines.push('## Failure Modes')\n    lines.push('')\n    for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`)\n    lines.push('')\n  }\n\n  if (options.runs && options.runs.length > 0) {\n    lines.push('## Run Summary')\n    lines.push('')\n    lines.push(\n      summaryTable([...options.runs], {\n        comparator: options.comparator ?? scorecard.baselineId ?? undefined,\n        split: 'holdout',\n      }).markdown,\n    )\n    lines.push('')\n  }\n\n  if (options.traceAnalystFindings && options.traceAnalystFindings.length > 0) {\n    lines.push('## TraceAnalyst Findings')\n    lines.push('')\n    for (const finding of options.traceAnalystFindings) lines.push(`- ${finding}`)\n    lines.push('')\n  }\n\n  const nextActions = options.nextActions ?? defaultNextActions(scorecard)\n  if (nextActions.length > 0) {\n    lines.push('## Next Actions')\n    lines.push('')\n    for (const action of nextActions) lines.push(`- ${action}`)\n    lines.push('')\n  }\n\n  return `${lines.join('\\n').trimEnd()}\\n`\n}\n\nfunction defaultNextActions(scorecard: ReleaseConfidenceScorecard): string[] {\n  if (scorecard.promote) return ['Promote the candidate and keep canaries enabled.']\n  return scorecard.issues\n    .filter((issue) => issue.severity === 'critical')\n    .map((issue) => `Resolve ${issue.code}: ${issue.detail}`)\n}\n\nfunction entries(values: Record<string, number>): Array<[string, number]> {\n  return Object.entries(values)\n    .filter(([, count]) => count > 0)\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n}\n\nfunction pct(value: number): string {\n  return Number.isFinite(value) ? `${(value * 100).toFixed(1)}%` : 'n/a'\n}\n\nfunction num(value: number): string {\n  return Number.isFinite(value) ? value.toFixed(3) : 'n/a'\n}\n"],"mappings":";;;;;;;;AA2HA,IAAM,qBAA4D;AAAA,EAChE,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,aAAa;AAAA,EACb,cAAc;AAAA,EACd,eAAe;AAAA,EACf,gBAAgB,OAAO;AAAA,EACvB,cAAc,OAAO;AAAA,EACrB,uBAAuB;AAAA,EACvB,uBAAuB;AACzB;AAEO,SAAS,wCACd,QACwB;AACxB,SAAO,OAAO,IAAI,CAAC,WAAW;AAAA,IAC5B,YAAY,MAAM;AAAA,IAClB,aAAa,MAAM;AAAA,IACnB,OAAO,MAAM,UAAU,YAAY,YAAY,MAAM,UAAU,QAAQ,QAAQ;AAAA,IAC/E,OAAO,MAAM;AAAA,IACb,IAAI,MAAM;AAAA,IACV,WAAW,MAAM,QAAQ,MAAM,OAAO,KAAK,IAAI,MAAM,MAAM,MAAM,SAAS;AAAA,IAC1E,SAAS,MAAM;AAAA,IACf,YAAY,MAAM;AAAA,IAClB,aAAa,MAAM,QAAQ,kBAAkB;AAAA,IAC7C,KAAK,MAAM;AAAA,IACX,UAAU,MAAM;AAAA,EAClB,EAAE;AACJ;AAEO,SAAS,0BACd,OAC4B;AAC5B,QAAM,aAAa,EAAE,GAAG,oBAAoB,GAAG,MAAM,WAAW;AAChE,QAAM,cAAc,MAAM,eAAe;AACzC,QAAM,OAAO,gBAAgB,MAAM,QAAQ,CAAC,GAAG,aAAa,MAAM,UAAU;AAC5E,QAAM,SAAS,qBAAqB,MAAM,UAAU,CAAC,GAAG,aAAa,MAAM,UAAU;AACrF,QAAM,YAAY,MAAM,aAAa,CAAC;AACtC,QAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;AAChE,QAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;AAC/E,QAAM,eAAe,UAAU,MAAM,QAAQ;AAC7C,QAAM,gBAAgB,UAAU,MAAM,SAAS;AAC/C,QAAM,YAAY,CAAC,GAAG,cAAc,GAAG,aAAa;AACpD,QAAM,cAAc,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,OAAO,cAAc;AACpE,QAAM,gBAAgB,UAAU,SAAS,IAAI,YAAY;AACzD,QAAM,aAAa,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ,EAAE;AAC/D,QAAM,cAAc,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,SAAS,EAAE;AACjE,QAAM,kBAAkB,KAAK,YAAY;AACzC,QAAM,mBAAmB,KAAK,aAAa;AAC3C,QAAM,UAAoC;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA,UAAU,SAAS,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACjE,WAAW,KAAK,aAAa;AAAA,IAC7B;AAAA,IACA;AAAA,IACA,YAAY,SAAS,iBAAiB,gBAAgB;AAAA,IACtD,aAAa,KAAK;AAAA,MAChB,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,OAAO;AAAA,MAC5B,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,cAAc;AAAA,IACvD,CAAC;AAAA,IACD,WAAW;AAAA,MACT,CAAC,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,OAAO,cAAc,CAAC;AAAA,MACxF;AAAA,IACF;AAAA,IACA,YAAY,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,IACvE,iBAAiB,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,MAC1E,CAAC,QAAQ,IAAI;AAAA,IACf,EAAE;AAAA,IACF,kBAAkB,OAAO,OAAO,CAAC,MAAM,EAAE,cAAc,CAAC,EAAE;AAAA,IAC1D,iBAAiB,OAAO,OAAO,CAAC,OAAO,EAAE,aAAa,KAAK,CAAC,EAAE;AAAA,IAC9D;AAAA,IACA,cAAc,aAAa,SAAS;AAAA,IACpC,mBAAmB,kBAAkB,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACnF,0BAA0B,yBAAyB,MAAM;AAAA,EAC3D;AAEA,QAAM,SAAmC,CAAC;AAC1C,cAAY,OAAO,YAAY,SAAS,MAAM;AAC9C,eAAa,YAAY,SAAS,MAAM;AACxC,sBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;AAC3E,mBAAiB,YAAY,SAAS,MAAM;AAC5C,kBAAgB,YAAY,SAAS,MAAM;AAE3C,QAAM,OAAO,UAAU,SAAS,YAAY,MAAM,gBAAgB,MAAM,MAAM;AAC9E,QAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;AAEN,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd;AAAA,IACA,YAAY,MAAM,cAAc;AAAA,IAChC;AAAA,IACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;AAAA,IACjF;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,MAAM,WAAW;AAAA,IAC1B,cAAc,MAAM,gBAAgB;AAAA,IACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;AAAA,EAC9D;AACF;AAEO,SAAS,wBAAwB,OAA2D;AACjG,QAAM,YAAY,0BAA0B,KAAK;AACjD,MAAI,UAAU,WAAW,QAAQ;AAC/B,UAAM,IAAI,kBAAkB,UAAU,OAAO;AAAA,EAC/C;AACA,SAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;AACb,MAAI,YAAa,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,WAAW;AACxE,MAAI,WAAY,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAU;AACtE,SAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;AACxB,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,WAAW;AAC1F,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,UAAU;AACzF,SAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;AACN,MAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GAAG;AACtF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,gBAAgB,WAAW,kBAAkB;AACvD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,sBAAsB,WAAW,gBAAgB;AAAA,IACnF,CAAC;AAAA,EACH;AACA,MAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAAG;AAClE,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,aAAa,WAAW,eAAe;AACjD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,UAAU,wBAAwB,WAAW,aAAa;AAAA,IAC/E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,WAAW,WAAW,aAAa;AAC7C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,CAAC,MAAM,IAAI,WAAW,WAAW,CAAC;AAAA,IAC5E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;AACN,MAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAAgB;AAChF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,WAAW,yBAAyB,WAAW,cAAc;AAAA,IAClF,CAAC;AAAA,EACH;AACA,MAAI,OAAO,SAAS,QAAQ,UAAU,KAAK,QAAQ,aAAa,WAAW,eAAe;AACxF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,CAAC,MAAM,IAAI,WAAW,aAAa,CAAC;AAAA,IAC1F,CAAC;AAAA,EACH;AACA,MAAI,gBAAgB,CAAC,aAAa,SAAS;AACzC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM,QAAQ,aAAa,iBAAiB,QAAQ;AAAA,MACpD,QAAQ,aAAa;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,YACA,SACA,QACM;AACN,MAAI,CAAC,WAAW,sBAAuB;AACvC,MAAI,QAAQ,aAAa,QAAQ,iBAAiB;AAChD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,eAAe;AAAA,IACzD,CAAC;AAAA,EACH;AACF;AAEA,SAAS,gBACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,cAAc,WAAW,gBAAgB;AACnD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,CAAC,MAAM,IAAI,WAAW,cAAc,CAAC;AAAA,IACrF,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,UACP,SACA,YACA,cACA,QACyB;AACzB,SAAO;AAAA,IACL;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC;AAAA,MACxE,GAAG,QAAQ,aAAa,uBAAuB,QAAQ,YAAY,OAAO;AAAA,IAC5E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS;AAAA,MAC5C,YAAY,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IACvE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,CAAC,aAAa,UAC1B,IACA,SAAS,QAAQ,YAAY,WAAW,aAAa;AAAA,MACzD,eAAe,QAAQ,WAAW,eAAe,IAAI,QAAQ,UAAU,CAAC;AAAA,IAC1E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ;AAAA,MACjE,mBAAmB,QAAQ,eAAe,IAAI,QAAQ,UAAU;AAAA,IAClE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,SAAS,UAAU;AAAA,MACnC,eAAe,IAAI,QAAQ,WAAW,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IAC7E;AAAA,EACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;AACvB,QAAM,MAAM,OAAO,OAAO,CAAC,MAAM,EAAE,SAAS,IAAI;AAChD,QAAM,SAAS,IAAI,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;AACN,SAAO,EAAE,MAAM,QAAQ,OAAO,QAAQ,KAAK,GAAG,OAAO;AACvD;AAEA,SAAS,oBAAoB,WAAqE;AAChG,QAAM,SAAuC,EAAE,OAAO,GAAG,KAAK,GAAG,MAAM,GAAG,SAAS,EAAE;AACrF,aAAW,YAAY,UAAW,QAAO,SAAS,SAAS,OAAO;AAClE,SAAO;AACT;AAEA,SAAS,aAAa,WAA+D;AACnF,QAAM,MAA8B,CAAC;AACrC,aAAW,YAAY,WAAW;AAChC,UAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;AACnE,QAAI,MAAM,KAAK,IAAI,MAAM,KAAK,KAAK;AAAA,EACrC;AACA,SAAO;AACT;AAEA,SAAS,kBACP,MACA,QACA,WACwB;AACxB,QAAM,MAA8B,CAAC;AACrC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,OAAO,IAAI,eAAe;AAChC,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,YAAM,OAAO,MAAM,gBAAgB,MAAM,OAAO,QAAQ,WAAW;AACnE,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;AACjG,QAAM,MAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,eAAW,OAAO,MAAM,OAAO,CAAC,GAAG;AACjC,YAAM,UAAU,IAAI,sBAAsB;AAC1C,UAAI,OAAO,KAAK,IAAI,OAAO,KAAK,KAAK;AAAA,IACvC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;AAC5B,QAAM,MAAkC,CAAC;AACzC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,YAAY,IAAI,QAAQ,IAAI;AAClC,UAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,UAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;AAAA,IACnD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,SACP,MACA,QACA,WACQ;AACR,QAAM,WAAW;AAAA,IACf,GAAG,KAAK,IAAI,CAAC,QAAQ;AACnB,YAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,aAAO,CAAC,IAAI,eAAe,UAAU,UAAa,SAAS;AAAA,IAC7D,CAAC;AAAA,IACD,GAAG,OAAO;AAAA,MACR,CAAC,UAAU,MAAM,OAAO,UAAU,MAAM,UAAU,UAAa,MAAM,SAAS;AAAA,IAChF;AAAA,EACF;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,SAAO,SAAS,OAAO,OAAO,EAAE,SAAS,SAAS;AACpD;AAEA,SAAS,UAAU,MAA4B,OAA8B;AAC3E,SAAO,KACJ,OAAO,CAAC,QAAQ,IAAI,aAAa,KAAK,EACtC,IAAI,CAAC,QAAS,UAAU,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ,WAAY,EACvF,OAAO,cAAc;AAC1B;AAEA,SAAS,KAAK,IAA+B;AAC3C,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,WAAW,IAAuB,GAAmB;AAC5D,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,SAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC,CAAC;AAC1F;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,SAAS,GAAW,GAAmB;AAC9C,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO;AAC9D,SAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAa,QAAwB;AACrD,MAAI,CAAC,OAAO,SAAS,GAAG,EAAG,QAAO;AAClC,MAAI,UAAU,EAAG,QAAO,OAAO,IAAI,IAAI;AACvC,SAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACQ;AACR,QAAM,OACJ,OAAO,SAAS,WAAW,cAAc,KAAK,OAAO,SAAS,QAAQ,WAAW,IAC7E,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IACxE;AACN,QAAM,UACJ,OAAO,SAAS,WAAW,YAAY,KAAK,OAAO,SAAS,QAAQ,SAAS,IACzE,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACpE;AACN,SAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;AACR,QAAM,SAAS,sBAAsB,MAAM,KAAK,MAAM;AACtD,QAAM,aAAa,aAAa,QAAQ,aAAa,eAAe,QAAQ,UAAU,gBAAgB,QAAQ,WAAW,aAAa,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAC/L,MAAI,OAAO,WAAW,EAAG,QAAO,GAAG,MAAM,KAAK,UAAU;AACxD,SAAO,GAAG,MAAM,KAAK,UAAU,YAAY,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAChF;AAEA,SAAS,IAAI,GAAmB;AAC9B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO,CAAC;AACxC,SAAO,EAAE,QAAQ,CAAC;AACpB;;;ACziBO,SAAS,YACd,UACA,WACA,UAA4B,CAAC,GACZ;AACjB,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,WAAW,QAAQ,mBAAmB;AAC5C,QAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;AAEpE,QAAM,eAAeA,MAAK,QAAQ;AAClC,QAAM,gBAAgBA,MAAK,SAAS;AACpC,QAAM,QAAQ,gBAAgB;AAE9B,MACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GACrB;AACA,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,SAAS;AAAA,MACT,YAAY;AAAA,MACZ;AAAA,MACA,SAAS;AAAA,IACX;AAAA,EACF;AAEA,QAAM,SAAmB,IAAI,MAAM,UAAU;AAC7C,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,YAAY,SAAS,UAAU,GAAG;AACxC,UAAM,YAAY,SAAS,WAAW,GAAG;AACzC,WAAO,CAAC,IAAIA,MAAK,SAAS,IAAIA,MAAK,SAAS;AAAA,EAC9C;AACA,SAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,QAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;AACpD,QAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;AAC5D,QAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ,CAAC;AAC5C,QAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ,CAAC;AAEzD,MAAI;AACJ,MAAI,UAAU,EAAG,WAAU;AAAA,WAClB,UAAU,EAAG,WAAU;AAAA,WACvB,SAAS,EAAG,WAAU;AAAA,MAC1B,WAAU;AAEf,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAASA,MAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,MAAI,IAAI;AACR,aAAW,KAAK,GAAI,MAAK;AACzB,SAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;AAC3D,QAAM,MAAM,IAAI,MAAM,GAAG,MAAM;AAC/B,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,KAAI,CAAC,IAAI,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC7E,SAAO;AACT;AAGA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAGA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,IAAI;AACR,aAAW,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;AAC5B,UAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;AACjC,UAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;AACxC,eAAW,QAAQ,OAAO;AACxB,WAAK;AACL,UAAI,KAAK,KAAK,GAAG,QAAQ;AAAA,IAC3B;AAAA,EACF;AACA,SAAO,MAAM;AACf;AA8BA,eAAsB,gBACpB,MACkF;AAClF,QAAM,cAAc,KAAK,oBAAoB;AAC7C,QAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;AACnF,QAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;AACrF,QAAM,KAAK,YAAY,gBAAgB,iBAAiB;AAAA,IACtD,GAAI,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;AAAA,IACxD,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,IACvE,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,EACvD,CAAC;AACD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,iBAAiB,eAAe;AAAA,IAChC,kBAAkB,gBAAgB;AAAA,EACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;AACnB,QAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;AAClD,MAAI,OAAO;AACX,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,QAAQ,OAAQ;AACzB,YAAM,IAAI,MAAM,MAAM,QAAQ,CAAC,CAAE;AACjC,cAAQ,CAAC,IAAI,OAAO,SAAS,CAAC,IAAI,IAAI;AAAA,IACxC;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,GAAG,MAAM,OAAO,CAAC,CAAC;AAClF,SAAO;AACT;;;ACvNO,SAAS,oBACd,WACA,UAAsC,CAAC,GAC/B;AACR,QAAM,QAAQ,QAAQ,SAAS,mBAAmB,UAAU,MAAM;AAClE,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,KAAK,KAAK,EAAE;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,aAAa,UAAU,OAAO,YAAY,CAAC,IAAI;AAC1D,QAAM,KAAK,cAAc,UAAU,UAAU,QAAQ,IAAI,IAAI;AAC7D,MAAI,UAAU,YAAa,OAAM,KAAK,gBAAgB,UAAU,WAAW,IAAI;AAC/E,MAAI,UAAU,WAAY,OAAM,KAAK,eAAe,UAAU,UAAU,IAAI;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,UAAU,OAAO;AAC5B,QAAM,KAAK,EAAE;AAEb,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,oBAAoB;AAC/B,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,iBAAiB,UAAU,QAAQ,aAAa,IAAI;AAC/D,QAAM,KAAK,mBAAmB,UAAU,QAAQ,UAAU,IAAI;AAC9D,QAAM,KAAK,oBAAoB,UAAU,QAAQ,WAAW,IAAI;AAChE,QAAM,KAAK,iBAAiB,IAAI,UAAU,QAAQ,QAAQ,CAAC,IAAI;AAC/D,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,SAAS,CAAC,IAAI;AACjE,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,eAAe,CAAC,IAAI;AACxE,QAAM,KAAK,oBAAoB,IAAI,UAAU,QAAQ,gBAAgB,CAAC,IAAI;AAC1E,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,UAAU,CAAC,IAAI;AACnE,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,WAAW,CAAC,IAAI;AACnE,QAAM,KAAK,qBAAqB,KAAK,MAAM,UAAU,QAAQ,SAAS,CAAC,OAAO;AAC9E,QAAM,KAAK,EAAE;AAEb,MAAI,UAAU,OAAO,SAAS,GAAG;AAC/B,UAAM,KAAK,WAAW;AACtB,UAAM,KAAK,EAAE;AACb,eAAW,SAAS,UAAU,QAAQ;AACpC,YAAM,KAAK,OAAO,MAAM,QAAQ,QAAQ,MAAM,IAAI,OAAO,MAAM,IAAI,MAAM,MAAM,MAAM,EAAE;AAAA,IACzF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,wBAAwB;AACnE,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,yBAAyB;AACpC,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,SAAS,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,OAAO,KAAK,KAAK,EAAE;AAC5E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,iBAAiB;AAC5D,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,kBAAkB;AAC7B,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,MAAM,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,IAAI,KAAK,KAAK,EAAE;AACtE,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,QAAQ,QAAQ,KAAK,SAAS,GAAG;AAC3C,UAAM,KAAK,gBAAgB;AAC3B,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,aAAa,CAAC,GAAG,QAAQ,IAAI,GAAG;AAAA,QAC9B,YAAY,QAAQ,cAAc,UAAU,cAAc;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC,EAAE;AAAA,IACL;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,wBAAwB,QAAQ,qBAAqB,SAAS,GAAG;AAC3E,UAAM,KAAK,0BAA0B;AACrC,UAAM,KAAK,EAAE;AACb,eAAW,WAAW,QAAQ,qBAAsB,OAAM,KAAK,KAAK,OAAO,EAAE;AAC7E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,cAAc,QAAQ,eAAe,mBAAmB,SAAS;AACvE,MAAI,YAAY,SAAS,GAAG;AAC1B,UAAM,KAAK,iBAAiB;AAC5B,UAAM,KAAK,EAAE;AACb,eAAW,UAAU,YAAa,OAAM,KAAK,KAAK,MAAM,EAAE;AAC1D,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,CAAC;AAAA;AACtC;AAEA,SAAS,mBAAmB,WAAiD;AAC3E,MAAI,UAAU,QAAS,QAAO,CAAC,kDAAkD;AACjF,SAAO,UAAU,OACd,OAAO,CAAC,UAAU,MAAM,aAAa,UAAU,EAC/C,IAAI,CAAC,UAAU,WAAW,MAAM,IAAI,KAAK,MAAM,MAAM,EAAE;AAC5D;AAEA,SAAS,QAAQ,QAAyD;AACxE,SAAO,OAAO,QAAQ,MAAM,EACzB,OAAO,CAAC,CAAC,EAAE,KAAK,MAAM,QAAQ,CAAC,EAC/B,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC;AAC3D;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC,MAAM;AACnE;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,MAAM,QAAQ,CAAC,IAAI;AACrD;","names":["mean"]}

package/dist/multi-layer-verifier-BNi4-8lR.d.ts

@@ -1,141 +0,0 @@
-/**
- * Multi-layer verifier — ordered pipeline of verification layers.
- *
- * Different contract from {@link JudgeRunner} (which runs parallel
- * specs against a sandbox). MultiLayerVerifier is a DAG of layers
- * (install → typecheck → build → lint → serve → semantic → …) with
- * dependency-based skip, per-layer findings, soft-fail semantics, and
- * an aggregated `blendedScore` across all passed layers.
- *
- * Use when you want:
- *   - ordered stages where a failing upstream stage skips downstream ones
- *   - each stage produces rich `findings` (severity + message + evidence)
- *   - a single composite score across stages with per-stage weights
- *   - soft-fail stages whose failure doesn't abort the pipeline
- *
- * Use {@link JudgeRunner} when you want:
- *   - N independent judges running in parallel against the same artifact
- *   - no inter-judge dependencies
- *   - boolean `passed` per judge + overall
- *
- * Both primitives compose — JudgeRunner can be invoked as a single
- * layer inside a MultiLayerVerifier if that suits the caller.
- */
-type LayerStatus = 'pass' | 'fail' | 'skipped' | 'error' | 'timeout';
-type Severity = 'critical' | 'major' | 'minor' | 'info';
-interface Finding {
-    severity: Severity;
-    message: string;
-    evidence?: string;
-    /** Optional layer name the finding belongs to (set by the verifier if omitted). */
-    layer?: string;
-    /**
-     * Free-form structured payload — used by `multiToolchainLayer` to attach
-     * `{ adapter: 'pnpm' }`, by judges to attach evidence pointers, etc.
-     * Renderers MAY interrogate; agent-eval primitives never assume shape.
-     */
-    detail?: Record<string, unknown>;
-}
-interface LayerResult {
-    layer: string;
-    status: LayerStatus;
-    /** 0..1 score, optional — layers that don't produce a numeric score omit. */
-    score?: number;
-    durationMs: number;
-    findings: Finding[];
-    /** Short human-readable summary (one line). */
-    reason?: string;
-    /**
-     * Numeric layer-level diagnostics: error counts, warning counts,
-     * cyclomatic complexity, total adapter wall-time, etc. Keyed by
-     * diagnostic name; null = "diagnostic not applicable / not measured."
-     * Renderers that know the keys can display them; ones that don't,
-     * ignore. Free-form on purpose — consumers type the value shape in
-     * their own namespace.
-     */
-    diagnostics?: Record<string, number | null>;
-    /** Any rich per-layer detail — rendered as-is by consumers that know the layer. */
-    detail?: Record<string, unknown>;
-}
-interface VerifyContext<Env = unknown> {
-    /** Per-run opaque context the caller provides. Layers destructure what they need. */
-    env: Env;
-    /** Previously-computed results from layers that already ran. */
-    prior: Record<string, LayerResult>;
-    /** Signal — if aborted, layers MUST bail within reasonable wall. */
-    signal: AbortSignal;
-}
-interface Layer<Env = unknown> {
-    name: string;
-    /** Stages that must have `status: 'pass'` before this layer runs. */
-    dependsOn?: string[];
-    /**
-     * Weight in the composite `blendedScore`. Default 1.0. Layers with weight 0
-     * contribute findings but not score.
-     */
-    weight?: number;
-    /**
-     * If true, a `fail` status contributes to `blendedScore` (as 0) instead of
-     * being dropped — use for layers whose failure is a real signal. Default:
-     * fail drops from numerator + denominator, matching VB's existing semantics.
-     */
-    failContributesToScore?: boolean;
-    /** Optional per-layer wall-cap in ms. Honored by the verifier (AbortSignal). */
-    capMs?: number;
-    run: (ctx: VerifyContext<Env>) => Promise<LayerResult> | LayerResult;
-}
-interface VerifyOptions<Env = unknown> {
-    env: Env;
-    /**
-     * Overall wall cap. Default: sum of layer capMs, or Infinity if any layer
-     * omits a cap. The verifier short-circuits remaining layers on overall cap.
-     */
-    overallCapMs?: number;
-    /** Called with each layer result as it completes. */
-    onLayer?: (result: LayerResult) => void;
-}
-interface VerificationReport {
-    layers: LayerResult[];
-    passCount: number;
-    failCount: number;
-    skippedCount: number;
-    errorCount: number;
-    /** True iff at least one scored layer ran AND every scored layer passed. */
-    allPass: boolean;
-    /**
-     * Weighted mean of `score` across contributing layers. 0 when no layers
-     * contributed. See {@link Layer.failContributesToScore} for fail semantics.
-     */
-    blendedScore: number;
-    durationMs: number;
-    startedAt: string;
-    finishedAt: string;
-}
-/**
- * Grade a semantic-concept-style judge result into a single layer status.
- *
- * Pass when overall score >= threshold AND no critical-severity concept gap.
- * Fail otherwise. Use inside a `Layer.run` when wrapping a concept judge.
- *
- * Generalized from VerticalBench H3 fix: `failingConcepts.length === 0` was
- * too strict — a single concept at 6/10 failed the entire layer despite
- * overall score being >= 0.7. Now we trust the judge's own `severity` field:
- * `critical` findings veto; `major`/`minor` reduce the score but don't veto.
- */
-declare function gradeSemanticStatus(input: {
-    score: number;
-    findings: Array<{
-        severity: Severity;
-        present?: boolean;
-        score?: number;
-    }>;
-    available: boolean;
-    threshold?: number;
-}): LayerStatus;
-declare class MultiLayerVerifier<Env = unknown> {
-    private readonly layers;
-    constructor(layers: Layer<Env>[]);
-    run(opts: VerifyOptions<Env>): Promise<VerificationReport>;
-}
-
-export { type Finding as F, type Layer as L, MultiLayerVerifier as M, type Severity as S, type VerificationReport as V, type VerifyOptions as a, type LayerResult as b, type VerifyContext as c, type LayerStatus as d, gradeSemanticStatus as g };

package/dist/optimization.d.ts

@@ -1,11 +0,0 @@
-export { C as CallbackResearcher, a as CallbackResearcherOptions, b as CampaignFactoryParams, c as CampaignIntegrityPolicy, d as CampaignRunContext, e as CampaignRunOutcome, f as CampaignRunner, g as CampaignScenario, h as CampaignVariant, E as EvalCampaignOptions, i as EvalCampaignResult, j as ExperimentPlan, k as ExperimentResult, F as FailedRun, l as FailureMode, N as NoopResearcher, R as Researcher, S as SteeringChange, r as runEvalCampaign } from './researcher-DeZ_EArp.js';
-export { F as FeedbackArtifactType, a as FeedbackAttempt, b as FeedbackLabel, c as FeedbackLabelKind, d as FeedbackLabelSource, e as FeedbackOptimizerRow, f as FeedbackOutcome, g as FeedbackReplayAdapter, h as FeedbackReplayResult, i as FeedbackSeverity, j as FeedbackSplitPolicy, k as FeedbackTask, l as FeedbackTrajectory, m as FeedbackTrajectoryFilter, n as FeedbackTrajectoryStore, o as FileSystemFeedbackTrajectoryStore, I as InMemoryFeedbackTrajectoryStore, P as PreferenceMemoryEntry, p as ProposedSideEffect, q as assignFeedbackSplit, r as controlRunToFeedbackTrajectory, s as createFeedbackTrajectory, t as feedbackTrajectoriesToDatasetScenarios, u as feedbackTrajectoriesToOptimizerRows, v as feedbackTrajectoryToDatasetScenario, w as feedbackTrajectoryToOptimizerRow, x as parseFeedbackTrajectoriesJsonl, y as renderPreferenceMemoryMarkdown, z as replayFeedbackTrajectories, A as replayFeedbackTrajectory, B as serializeFeedbackTrajectoriesJsonl, C as summarizePreferenceMemory, D as withAssignedFeedbackSplit } from './feedback-trajectory-Dvy-bt7x.js';
-export { A as ActionableSideInfo, a as AsiSeverity, D as DEFAULT_MUTATION_PRIMITIVES, E as EvolvableVariant, G as GenerationReport, I as InMemoryTrialCache, M as MultiShotGateConfig, b as MultiShotGateResult, c as MultiShotMutateAdapter, d as MultiShotOptimizationConfig, e as MultiShotOptimizationResult, f as MultiShotRun, g as MultiShotRunInput, h as MultiShotRunner, i as MultiShotScore, j as MultiShotScorer, k as MultiShotSplit, l as MultiShotTrace, m as MultiShotTrialResult, n as MultiShotVariant, o as MutateAdapter, P as PromptEvolutionConfig, p as PromptEvolutionEvent, q as PromptEvolutionResult, R as ReflectionContext, r as ReflectionProposal, S as ScenarioAggregate, s as ScoreAdapter, T as TrialCache, t as TrialResult, u as TrialTrace, V as VariantAggregate, v as buildReflectionPrompt, w as defaultMultiShotObjectives, x as parseReflectionResponse, y as runMultiShotOptimization, z as runPromptEvolution, B as trialTraceFromMultiShotTrial } from './summary-report-DuZXOk7K.js';
-import './run-record-BGY6bHRh.js';
-import './errors-mje_cKOs.js';
-import './integrity-DYR5gWlb.js';
-import './store-Db2Bv8Cf.js';
-import './emitter-DP_cSSiw.js';
-import './control-runtime-BZ_lVLYW.js';
-import './dataset-BlwAtYYf.js';
-import './failure-cluster-Cw65_5FY.js';

package/dist/optimization.js

@@ -1,71 +0,0 @@
-import {
-  CallbackResearcher,
-  DEFAULT_MUTATION_PRIMITIVES,
-  FileSystemFeedbackTrajectoryStore,
-  InMemoryFeedbackTrajectoryStore,
-  InMemoryTrialCache,
-  NoopResearcher,
-  assignFeedbackSplit,
-  buildReflectionPrompt,
-  controlRunToFeedbackTrajectory,
-  createFeedbackTrajectory,
-  defaultMultiShotObjectives,
-  feedbackTrajectoriesToDatasetScenarios,
-  feedbackTrajectoriesToOptimizerRows,
-  feedbackTrajectoryToDatasetScenario,
-  feedbackTrajectoryToOptimizerRow,
-  parseFeedbackTrajectoriesJsonl,
-  parseReflectionResponse,
-  renderPreferenceMemoryMarkdown,
-  replayFeedbackTrajectories,
-  replayFeedbackTrajectory,
-  runMultiShotOptimization,
-  runPromptEvolution,
-  serializeFeedbackTrajectoriesJsonl,
-  summarizePreferenceMemory,
-  trialTraceFromMultiShotTrial,
-  withAssignedFeedbackSplit
-} from "./chunk-DMW5VENN.js";
-import {
-  runEvalCampaign
-} from "./chunk-LCIDRYGP.js";
-import "./chunk-VXNVVBZO.js";
-import "./chunk-BWZEGTES.js";
-import "./chunk-EGIPWXHL.js";
-import "./chunk-WP7SY7AI.js";
-import "./chunk-UBPIXOC4.js";
-import "./chunk-PC4UYEBM.js";
-import "./chunk-TVVP3ZZQ.js";
-import "./chunk-VSMTAMNK.js";
-import "./chunk-QYJT52YW.js";
-import "./chunk-PZ5AY32C.js";
-export {
-  CallbackResearcher,
-  DEFAULT_MUTATION_PRIMITIVES,
-  FileSystemFeedbackTrajectoryStore,
-  InMemoryFeedbackTrajectoryStore,
-  InMemoryTrialCache,
-  NoopResearcher,
-  assignFeedbackSplit,
-  buildReflectionPrompt,
-  controlRunToFeedbackTrajectory,
-  createFeedbackTrajectory,
-  defaultMultiShotObjectives,
-  feedbackTrajectoriesToDatasetScenarios,
-  feedbackTrajectoriesToOptimizerRows,
-  feedbackTrajectoryToDatasetScenario,
-  feedbackTrajectoryToOptimizerRow,
-  parseFeedbackTrajectoriesJsonl,
-  parseReflectionResponse,
-  renderPreferenceMemoryMarkdown,
-  replayFeedbackTrajectories,
-  replayFeedbackTrajectory,
-  runEvalCampaign,
-  runMultiShotOptimization,
-  runPromptEvolution,
-  serializeFeedbackTrajectoriesJsonl,
-  summarizePreferenceMemory,
-  trialTraceFromMultiShotTrial,
-  withAssignedFeedbackSplit
-};
-//# sourceMappingURL=optimization.js.map

package/dist/optimization.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/sequential-5iSVfzl2.d.ts

@@ -1,139 +0,0 @@
-/**
- * Always-valid sequential evaluation.
- *
- * `researchReport` assumes a single pre-specified analysis. Real
- * consumers run campaigns weekly / nightly / per-PR; each new run silently
- * inflates the false-discovery rate, because the BH-FDR guarantee is for
- * the *first* look, not the 47th. Without time-uniform inference,
- * launch-decision teams either (a) don't peek, which forfeits the cost
- * advantage of stop-when-decisive, or (b) peek and pretend they didn't,
- * which forfeits scientific validity.
- *
- * This module ships **e-value-based confidence sequences** for paired
- * bounded outcomes. The methodology is the predictable plug-in betting
- * martingale of Waudby-Smith & Ramdas (2024) — provably valid at *any*
- * stopping time. Concretely:
- *
- *   For paired deltas D_1, D_2, … ∈ [-c, c] with the null H_0: E[D] ≤ 0,
- *   a betting fraction λ_i is chosen using only D_{1..i-1} (predictable
- *   plug-in), and the running e-value is
- *
- *     E_t = ∏_{i=1}^{t} (1 + λ_i · D_i)
- *
- *   E_t is a non-negative martingale under H_0 with E[E_t] ≤ 1, so by
- *   Ville's inequality, P(∃ t : E_t ≥ 1/α) ≤ α — we can reject the null
- *   at any time without inflating the type-I error.
- *
- * Combined with `runEvalCampaign`, every consumer running rolling
- * campaigns gains the ability to ship the moment evidence is decisive,
- * stop-early on dead-on-arrival variants, and accumulate evidence across
- * partial runs without spending the FDR budget. No new sweep is wasted.
- *
- * References:
- *   - Howard, S. R., Ramdas, A., McAuliffe, J., Sekhon, J. (2021).
- *     Time-uniform, nonparametric, nonasymptotic confidence sequences.
- *     Annals of Statistics, 49(2), 1055–1080.
- *   - Waudby-Smith, I., Ramdas, A. (2024). Estimating means of bounded
- *     random variables by betting. JRSS B, 86(1), 1–27.
- */
-type SequentialDecision = 'promote_now' | 'continue' | 'reject_now' | 'equivalent';
-interface PairedEvalueOptions {
-    /**
-     * Bound on |delta|. Default 1 (matching most score scales). Must satisfy
-     * c > 0; deltas outside [-c, c] are clipped with a warning attached to
-     * the return value.
-     */
-    bound?: number;
-    /** Target Type-I error. Default 0.05. */
-    alpha?: number;
-    /**
-     * Region of Practical Equivalence on the *mean* paired delta. When
-     * supplied, the verdict can return `'equivalent'` once the running
-     * confidence sequence on the mean is fully contained in [low, high].
-     */
-    rope?: {
-        low: number;
-        high: number;
-    };
-    /** Initial bet shrinkage (0 < scale ≤ 1). Default 0.5 — empirically robust. */
-    initialBetShrinkage?: number;
-}
-interface PairedEvalueStep {
-    /** 1-indexed observation count. */
-    t: number;
-    delta: number;
-    /** Running e-value E_t = ∏ (1 + λ_i · D_i). */
-    evalue: number;
-    /** Time-uniform p-value at stopping time t. */
-    pValue: number;
-    /** Lower bound of the empirical Bernstein confidence sequence at level 1-α. */
-    csLow: number;
-    csHigh: number;
-    /** Verdict at this stopping time. */
-    decision: SequentialDecision;
-}
-interface PairedEvalueSequence {
-    steps: PairedEvalueStep[];
-    /** The decision at the final step. */
-    finalDecision: SequentialDecision;
-    /** Index (1-based) at which a non-`continue` decision first fired, or null. */
-    decisionFiredAt: number | null;
-    /** True if any deltas were clipped to [-bound, bound]. */
-    clipped: boolean;
-}
-/**
- * Run the paired e-value sequence over an in-order delta stream.
- *
- * Use for *streaming* / interim analyses: pass the deltas you have so
- * far, get the verdict at every prefix length. The decision is
- * monotone-stable in the sense that once `'reject_now'` or `'promote_now'`
- * fires, the verdict at later steps remains decisive (the e-value is a
- * non-negative martingale; once it crosses the threshold, it's crossed).
- */
-declare function pairedEvalueSequence(deltas: number[], opts?: PairedEvalueOptions): PairedEvalueSequence;
-interface InterimReleaseConfidenceInput {
-    /**
-     * One delta series per candidate (paired deltas vs comparator). Order
-     * within a series is the order the campaigns were run.
-     */
-    deltaSeries: Array<{
-        candidateId: string;
-        deltas: number[];
-    }>;
-    alpha?: number;
-    bound?: number;
-    rope?: {
-        low: number;
-        high: number;
-    };
-}
-interface InterimReleaseConfidence {
-    candidates: Array<{
-        candidateId: string;
-        decision: SequentialDecision;
-        decisionFiredAt: number | null;
-        finalEvalue: number;
-        finalPValue: number;
-        pairs: number;
-        csLow: number;
-        csHigh: number;
-    }>;
-    /**
-     * Campaign-level recommendation: pick the strongest 'promote_now', else
-     * 'continue' if any candidate is still live, else 'reject_now' if every
-     * candidate is dead, else 'equivalent'.
-     */
-    recommendation: {
-        decision: SequentialDecision;
-        candidateId: string | null;
-    };
-}
-/**
- * Run interim sequential analyses across many candidates at once,
- * preserving the time-uniform α guarantee for each candidate's series and
- * synthesising a campaign-level recommendation. Designed to be called on
- * every campaign tick — the recommendation is anytime-valid.
- */
-declare function evaluateInterimReleaseConfidence(input: InterimReleaseConfidenceInput): InterimReleaseConfidence;
-
-export { type InterimReleaseConfidence as I, type PairedEvalueOptions as P, type SequentialDecision as S, type InterimReleaseConfidenceInput as a, type PairedEvalueSequence as b, type PairedEvalueStep as c, evaluateInterimReleaseConfidence as e, pairedEvalueSequence as p };