@tangle-network/agent-eval 0.181.0 → 0.183.0

package/dist/{run-record-Br-Yzt_k.js.map → run-record-DPa2PeHx.js.map}

@@ -1 +1 @@
-{"version":3,"file":"run-record-Br-Yzt_k.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.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\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): PairedDecisionShape {\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 ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${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  )\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","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,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,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,WACV;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;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}
+{"version":3,"file":"run-record-DPa2PeHx.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.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\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): PairedDecisionShape {\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 ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${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  )\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","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,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,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,WACV;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;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}

package/dist/{semantic-concept-judge-Bm5JDEKO.js → semantic-concept-judge-_3Fp6Q_3.js}

@@ -1,9 +1,9 @@
-import { i as CostLedger } from "./cost-ledger-B1qx30B4.js";
+import { i as CostLedger } from "./cost-ledger-_IbkejCF.js";
 import { u as Mutex } from "./journal-Cs9f7385.js";
-import { M as RawAnalystFindingSchema } from "./kind-factory-D5HOW3R0.js";
-import { c as extractJsonPayload } from "./llm-client-CGlSi8sb.js";
-import { n as paidJsonChat, t as paidChat } from "./chat-json-call-B_Xv2oJK.js";
-import { n as decodeRawFindingArray } from "./dspy-rlm-engine-S53V0HhE.js";
+import { M as RawAnalystFindingSchema } from "./kind-factory-DxTzg32u.js";
+import { c as extractJsonPayload } from "./llm-client-DDhZ29Vm.js";
+import { n as paidJsonChat, t as paidChat } from "./chat-json-call-C-DnDxEy.js";
+import { n as decodeRawFindingArray } from "./dspy-rlm-engine-8pgBSiy7.js";
 import { appendFileSync, existsSync, mkdirSync, readFileSync } from "node:fs";
 import { z } from "zod";
 import { dirname } from "node:path";
@@ -777,4 +777,4 @@ async function runSemanticConceptJudge(input, options) {
 //#endregion
 export { diffFindings as a, defaultIsMaterial as i, runSemanticConceptJudge as n, createChatTraceEngine as o, FindingsStore as r, SEMANTIC_CONCEPT_JUDGE_VERSION as t };
 
-//# sourceMappingURL=semantic-concept-judge-Bm5JDEKO.js.map
+//# sourceMappingURL=semantic-concept-judge-_3Fp6Q_3.js.map

package/dist/{semantic-concept-judge-Bm5JDEKO.js.map → semantic-concept-judge-_3Fp6Q_3.js.map}

@@ -1 +1 @@
-{"version":3,"file":"semantic-concept-judge-Bm5JDEKO.js","names":[],"sources":["../src/analyst/chat-trace-engine.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * A `TraceAnalysisEngine` that runs entirely inside Node against a\n * caller-owned `ChatClient`.\n *\n * The other engine in this package, `createDspyRlmTraceEngine`, reaches the\n * DSPy RLM through a Python subprocess. That made every model-backed analyst\n * unreachable for a consumer that already owns a model seam and no Python:\n * `createTraceAnalyst` requires an engine, the only exported constructor\n * required a Python runner, so `buildDefaultAnalystRegistry()` registered the\n * deterministic analyst alone and `analystsFromRegistry` refused the result.\n *\n * This engine closes that path. It drives the same investigation contract —\n * bounded trace tools, a prose answer, a strict findings array — with native\n * function calls over the transport the caller already bound with\n * `createChatClient`. A caller holding a bare\n * `(request: LlmCallRequest) => Promise<LlmCallResult>` adapts it in one line:\n *\n *   createChatClient({ transport: 'custom', chat: call, defaultModel, maximumAttempts })\n *\n * agent-eval still executes no paid model and holds no provider credential.\n * Every call goes through `paidChat`, so the shared cost ledger reserves the\n * priced maximum before the call and settles the receipt after it.\n *\n * Shape of one investigation:\n *   1. Investigation turns. The model reads the trace store through the tools\n *      its analyst kind was given. A turn that requests no tool ends the\n *      phase; an exhausted iteration or tool budget also ends it.\n *   2. One report turn. The model returns `{ answer, findings }` as JSON,\n *      decoded by the same `decodeRawFindingArray` the Python bridge uses, so\n *      a row this engine accepts is a row that engine could report.\n */\n\nimport { z } from 'zod'\nimport { paidChat } from '../chat-json-call'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport {\n  extractJsonPayload,\n  type LlmCallRequest,\n  type LlmMessage,\n  type LlmThinkingMode,\n  type LlmToolCall,\n  type LlmToolDefinition,\n} from '../llm-client'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\nimport type { ChatClient } from './chat-client'\nimport {\n  DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n  type TraceAnalysisEngine,\n  type TraceAnalysisEngineRequest,\n  type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { RawAnalystFindingSchema } from './finding-signature'\n\n/** Bumped whenever this engine's execution behavior changes. */\nconst CHAT_TRACE_ENGINE_VERSION = '1.0.0'\nconst ENGINE_ID = 'chat-trace'\n\n/** Marker appended to a tool result cut down to the retained-output budget. */\nconst TRUNCATION_MARKER = '\\n…[truncated to the analyst maxOutputChars budget]'\n\nexport interface ChatTraceEngineOptions {\n  /**\n   * Caller-owned transport. Build it with `createChatClient`; agent-eval never\n   * receives the provider credential.\n   */\n  chat: ChatClient\n  /** Model for every call. Defaults to the client's `defaultModel`. */\n  model?: string\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  /** Completion cap per turn. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */\n  maxOutputTokens?: number\n  /** Sampling temperature. Omitted from the request when absent. */\n  temperature?: number\n  /** Provider reasoning mode. Omitted when the provider default should apply. */\n  thinking?: LlmThinkingMode\n  /** Per-call deadline handed to the transport. */\n  requestTimeoutMs?: number\n  /**\n   * How the report turn asks for structured output.\n   *\n   * `json-object` sends JSON mode, which every OpenAI-compatible endpoint\n   * accepts, and relies on the finding grammar the analyst kind already puts\n   * in its instructions. `json-schema` sends the generated response schema as\n   * well, which stricter endpoints honour and older ones reject. Default:\n   * `json-object`.\n   */\n  reportFormat?: 'json-object' | 'json-schema'\n}\n\n/**\n * The report envelope. Its `findings` rows are generated from\n * `RawAnalystFindingSchema`, so the schema offered to a provider and the\n * decoder that accepts the answer cannot drift apart.\n */\nfunction buildReportJsonSchema(): { name: string; schema: Record<string, unknown> } {\n  const row = z.toJSONSchema(RawAnalystFindingSchema, { target: 'draft-7' }) as Record<\n    string,\n    unknown\n  >\n  delete row.$schema\n  return {\n    name: 'trace_analysis_report',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      required: ['answer', 'findings'],\n      properties: {\n        answer: { type: 'string', description: 'Direct prose answer to the question.' },\n        findings: { type: 'array', items: row },\n      },\n    },\n  }\n}\n\nconst REPORT_JSON_SCHEMA = buildReportJsonSchema()\n\n/**\n * Run bounded recursive trace analysis in-process over a caller-owned chat\n * transport. No Python, no subprocess, no loopback proxy.\n */\nexport function createChatTraceEngine(options: ChatTraceEngineOptions): TraceAnalysisEngine {\n  const model = resolveModel(options)\n  const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS\n  if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n    throw new TypeError('chat trace engine maxOutputTokens must be a positive safe integer')\n  }\n  if (\n    options.temperature !== undefined &&\n    (!Number.isFinite(options.temperature) || options.temperature < 0)\n  ) {\n    throw new TypeError('chat trace engine temperature must be a non-negative finite number')\n  }\n  if (\n    options.requestTimeoutMs !== undefined &&\n    (!Number.isSafeInteger(options.requestTimeoutMs) || options.requestTimeoutMs <= 0)\n  ) {\n    throw new TypeError('chat trace engine requestTimeoutMs must be a positive safe integer')\n  }\n  const reportFormat = options.reportFormat ?? 'json-object'\n  if (reportFormat !== 'json-object' && reportFormat !== 'json-schema') {\n    throw new TypeError(`chat trace engine reportFormat must be json-object or json-schema`)\n  }\n\n  return {\n    id: ENGINE_ID,\n    description:\n      'In-process recursive trace analysis over a caller-owned ChatClient with native tool calls.',\n    model,\n    version: CHAT_TRACE_ENGINE_VERSION,\n    executionConfig: {\n      kind: ENGINE_ID,\n      model,\n      transport: options.chat.transport,\n      maximum_attempts: options.chat.maximumAttempts ?? null,\n      pricing: options.pricing ? { ...options.pricing } : null,\n      max_output_tokens: maxOutputTokens,\n      temperature: options.temperature ?? null,\n      thinking: options.thinking ?? null,\n      request_timeout_ms: options.requestTimeoutMs ?? null,\n      report_format: reportFormat,\n      report_schema_name: REPORT_JSON_SCHEMA.name,\n    },\n    analyze: (request) => analyze(request, { ...options, model, maxOutputTokens, reportFormat }),\n  }\n}\n\ninterface ResolvedOptions extends ChatTraceEngineOptions {\n  model: string\n  maxOutputTokens: number\n  reportFormat: 'json-object' | 'json-schema'\n}\n\nasync function analyze(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n): Promise<TraceAnalysisEngineResult> {\n  // One model call is spent on the report, so a budget of one call would buy\n  // an answer with no investigation behind it. Refuse instead of pretending.\n  if (request.limits.maxLlmCalls < 2) {\n    throw new Error(\n      `chat trace engine reserves one model call for the report, so maxLlmCalls must be at least 2 (analyst '${request.analystId}' declared ${request.limits.maxLlmCalls})`,\n    )\n  }\n  const toolsByName = new Map(request.tools.map((tool) => [tool.name, tool]))\n  if (toolsByName.size !== request.tools.length) {\n    throw new Error(`chat trace engine received duplicate tool names for '${request.analystId}'`)\n  }\n  const toolDefinitions: LlmToolDefinition[] = request.tools.map((tool) => ({\n    type: 'function',\n    function: { name: tool.name, description: tool.description, parameters: tool.parameters },\n  }))\n  const investigationTurns = Math.max(\n    1,\n    Math.min(request.limits.maxIterations, request.limits.maxLlmCalls - 1),\n  )\n\n  const messages: LlmMessage[] = [\n    { role: 'system', content: systemPrompt(request, investigationTurns) },\n  ]\n  if (request.taskInputs) {\n    messages.push({ role: 'user', content: renderTaskInputs(request.taskInputs) })\n  }\n  messages.push({ role: 'user', content: request.question })\n\n  request.log?.('trace analyst engine started', {\n    engine: ENGINE_ID,\n    model: options.model,\n    transport: options.chat.transport,\n    tools: request.tools.map((tool) => tool.name),\n    limits: request.limits,\n  })\n\n  const trajectory: unknown[] = []\n  const servedModels = new Set<string>()\n  let modelCalls = 0\n  let toolCalls = 0\n  let truncatedToolResults = 0\n  let toolBudgetExhausted = false\n  let turnsUsed = 0\n  let stoppedOnAnswer = false\n\n  for (let turn = 0; turn < investigationTurns; turn++) {\n    turnsUsed = turn + 1\n    const response = await callModel(request, options, {\n      messages,\n      tools: toolDefinitions,\n      toolChoice: 'auto',\n      purpose: 'investigation',\n    })\n    modelCalls += 1\n    servedModels.add(response.servedModel ?? 'unreported')\n    const requested = response.toolCalls ?? []\n    messages.push({\n      role: 'assistant',\n      content: response.content,\n      ...(requested.length > 0 ? { toolCalls: requested } : {}),\n    })\n    trajectory.push({\n      turn: turnsUsed,\n      phase: 'investigation',\n      content: response.content,\n      tool_calls: requested.map((call) => ({ id: call.id, name: call.name })),\n      finish_reason: response.finishReason ?? null,\n    })\n    if (requested.length === 0) {\n      stoppedOnAnswer = true\n      break\n    }\n    for (const call of requested) {\n      if (toolCalls >= request.limits.maxToolCalls) {\n        toolBudgetExhausted = true\n        // Every requested id still needs an answer: a provider rejects the\n        // next turn when one tool call is left unanswered.\n        messages.push(toolMessage(call, exhaustedToolBudget(request.limits.maxToolCalls)))\n        continue\n      }\n      toolCalls += 1\n      const executed = await executeTool(call, toolsByName, request)\n      if (executed.truncated) truncatedToolResults += 1\n      messages.push(toolMessage(call, executed.payload))\n      trajectory.push({\n        turn: turnsUsed,\n        phase: 'tool',\n        name: call.name,\n        ok: executed.ok,\n        truncated: executed.truncated,\n        result_chars: executed.payload.length,\n      })\n    }\n    if (toolBudgetExhausted) break\n  }\n\n  messages.push({ role: 'user', content: reportPrompt(options.reportFormat) })\n  const report = await callModel(request, options, {\n    messages,\n    purpose: 'report',\n    json: options.reportFormat,\n  })\n  modelCalls += 1\n  servedModels.add(report.servedModel ?? 'unreported')\n  trajectory.push({\n    turn: turnsUsed + 1,\n    phase: 'report',\n    content: report.content,\n    finish_reason: report.finishReason ?? null,\n  })\n\n  const parsed = parseReport(report.content, request.analystId, (index, reason) => {\n    request.log?.('finding rejected: report row failed schema validation', {\n      engine: ENGINE_ID,\n      index,\n      reason,\n    })\n  })\n\n  const result: TraceAnalysisEngineResult = {\n    answer: parsed.answer,\n    findings: parsed.findings,\n    trajectory,\n    modelCalls,\n    toolCalls,\n    runtime: {\n      engine: ENGINE_ID,\n      model: options.model,\n      transport: options.chat.transport,\n      report_format: options.reportFormat,\n      investigation_turns: turnsUsed,\n      investigation_stopped_on_answer: stoppedOnAnswer,\n      tool_budget_exhausted: toolBudgetExhausted,\n      truncated_tool_results: truncatedToolResults,\n      rejected_findings: parsed.rejectedFindings,\n      // A gateway can answer `model: X` from another model on HTTP 200, so the\n      // ids the provider echoed travel with the result as evidence.\n      served_models: [...servedModels].sort(),\n      task_inputs: request.taskInputs ? 'prompt-delivered' : 'none',\n    },\n  }\n  request.log?.('trace analyst engine completed', {\n    engine: ENGINE_ID,\n    model_calls: result.modelCalls,\n    tool_calls: result.toolCalls,\n    findings: result.findings.length,\n  })\n  return result\n}\n\ninterface ModelTurn {\n  content: string\n  toolCalls?: LlmToolCall[]\n  servedModel?: string | null\n  finishReason?: string | null\n}\n\nasync function callModel(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n  turn: {\n    messages: LlmMessage[]\n    tools?: LlmToolDefinition[]\n    toolChoice?: LlmCallRequest['toolChoice']\n    purpose: 'investigation' | 'report'\n    json?: 'json-object' | 'json-schema'\n  },\n): Promise<ModelTurn> {\n  const chatRequest: LlmCallRequest = {\n    model: options.model,\n    // The transport may mutate what it is handed; the conversation is ours.\n    messages: turn.messages.map((message) => ({ ...message })),\n    maxTokens: options.maxOutputTokens,\n    ...(turn.tools && turn.tools.length > 0\n      ? { tools: turn.tools, ...(turn.toolChoice ? { toolChoice: turn.toolChoice } : {}) }\n      : {}),\n    ...(turn.json ? { jsonMode: true } : {}),\n    ...(turn.json === 'json-schema' ? { jsonSchema: REPORT_JSON_SCHEMA } : {}),\n    ...(options.temperature === undefined ? {} : { temperature: options.temperature }),\n    ...(options.thinking === undefined ? {} : { thinking: options.thinking }),\n    ...(options.requestTimeoutMs === undefined ? {} : { timeoutMs: options.requestTimeoutMs }),\n  }\n  const paid = await paidChat({\n    chat: options.chat,\n    request: chatRequest,\n    ledger: request.costLedger,\n    channel: 'analyst',\n    phase: request.costPhase,\n    actor: request.analystId,\n    ...(request.costTags ? { tags: request.costTags } : {}),\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    ...(request.signal ? { signal: request.signal } : {}),\n  })\n  if (!paid.succeeded) {\n    // The transport already spent its own attempts. A failure here ends the\n    // investigation: continuing would report an answer the model never gave.\n    throw new Error(\n      `chat trace engine ${turn.purpose} call failed for '${request.analystId}': ${paid.error.message}`,\n      { cause: paid.error },\n    )\n  }\n  return {\n    content: paid.response.content,\n    ...(paid.response.toolCalls ? { toolCalls: paid.response.toolCalls } : {}),\n    ...(paid.response.servedModel === undefined ? {} : { servedModel: paid.response.servedModel }),\n    ...(paid.response.finishReason === undefined\n      ? {}\n      : { finishReason: paid.response.finishReason }),\n  }\n}\n\ninterface ExecutedTool {\n  ok: boolean\n  truncated: boolean\n  payload: string\n}\n\nasync function executeTool(\n  call: LlmToolCall,\n  toolsByName: ReadonlyMap<string, TraceAnalysisToolDescriptor>,\n  request: TraceAnalysisEngineRequest,\n): Promise<ExecutedTool> {\n  const descriptor = toolsByName.get(call.name)\n  if (!descriptor) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `unknown tool '${call.name}'; available tools are ${[...toolsByName.keys()].sort().join(', ')}`,\n      ),\n    }\n  }\n  let args: unknown\n  try {\n    args = call.argumentsJson.trim() === '' ? {} : JSON.parse(call.argumentsJson)\n  } catch (error) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `arguments for '${call.name}' were not JSON: ${messageOf(error)}. Send valid JSON arguments.`,\n      ),\n    }\n  }\n  try {\n    const value = await descriptor.handler(\n      args,\n      request.signal ? { signal: request.signal } : undefined,\n    )\n    return truncateToolPayload(JSON.stringify(value ?? null), request.limits.maxOutputChars)\n  } catch (error) {\n    // A bounded store refuses an oversized or malformed read by design. That\n    // refusal is the model's next instruction, not the run's failure.\n    if (request.signal?.aborted) throw error\n    request.log?.('trace tool failed', {\n      engine: ENGINE_ID,\n      analyst_id: request.analystId,\n      tool: call.name,\n      reason: messageOf(error),\n    })\n    return { ok: false, truncated: false, payload: toolError(messageOf(error)) }\n  }\n}\n\nfunction truncateToolPayload(payload: string, maxOutputChars: number): ExecutedTool {\n  if (payload.length <= maxOutputChars) return { ok: true, truncated: false, payload }\n  const keep = Math.max(0, maxOutputChars - TRUNCATION_MARKER.length)\n  return { ok: true, truncated: true, payload: `${payload.slice(0, keep)}${TRUNCATION_MARKER}` }\n}\n\nfunction toolMessage(call: LlmToolCall, content: string): LlmMessage {\n  return { role: 'tool', toolCallId: call.id, content }\n}\n\nfunction toolError(reason: string): string {\n  return JSON.stringify({ error: reason })\n}\n\nfunction exhaustedToolBudget(maxToolCalls: number): string {\n  return toolError(\n    `the trace-tool budget of ${maxToolCalls} calls is spent; answer from the evidence already read`,\n  )\n}\n\nfunction systemPrompt(request: TraceAnalysisEngineRequest, investigationTurns: number): string {\n  return [\n    request.instructions.trim(),\n    [\n      'HOW THIS INVESTIGATION RUNS:',\n      `- You have ${investigationTurns} investigation turns and at most ${request.limits.maxToolCalls} trace-tool calls.`,\n      '- Call the trace tools to read the store. Never state a trace fact you did not read.',\n      `- A tool result longer than ${request.limits.maxOutputChars} characters is truncated; narrow the query instead of asking again.`,\n      '- A tool result carrying an \"error\" field is feedback: fix the call or take another route.',\n      '- Answer with no tool call once you have the evidence. You are then asked for the final report.',\n    ].join('\\n'),\n  ].join('\\n\\n')\n}\n\nfunction reportPrompt(reportFormat: 'json-object' | 'json-schema'): string {\n  return [\n    'Report now. Do not call any more tools.',\n    'Return one JSON object with exactly two fields:',\n    '  \"answer\": a direct prose answer to the question.',\n    '  \"findings\": an array of findings in the schema above. Emit [] when there is nothing to report.',\n    reportFormat === 'json-object'\n      ? 'Return the object alone, with no surrounding prose and no code fence.'\n      : 'Return the object in the response schema you were given.',\n  ].join('\\n')\n}\n\nfunction renderTaskInputs(taskInputs: Readonly<Record<string, unknown>>): string {\n  // The engine contract forbids dropping structured inputs. This engine has no\n  // code environment to bind them as variables, so they are delivered whole as\n  // conversation material and the runtime record says which way they arrived.\n  return [\n    'TASK INPUTS — structured material delivered with the question, not fetched through tools:',\n    JSON.stringify(taskInputs, null, 2),\n  ].join('\\n')\n}\n\ninterface ParsedReport {\n  answer: string\n  findings: TraceAnalysisEngineResult['findings']\n  rejectedFindings: number\n}\n\nfunction parseReport(\n  content: string,\n  analystId: string,\n  onRejectedFinding: (index: number, reason: string) => void,\n): ParsedReport {\n  const envelope = coerceReportEnvelope(content, analystId)\n  if (typeof envelope.answer !== 'string' || !envelope.answer.trim()) {\n    throw new Error(`chat trace engine report for '${analystId}' carried no answer`)\n  }\n  // Findings are model output: one malformed row is model noise, not an engine\n  // fault, and the rest of the paid investigation must survive it.\n  const decoded = decodeRawFindingArray(envelope.findings)\n  if (decoded.topLevelError !== undefined) {\n    throw new Error(\n      `chat trace engine report for '${analystId}' had a malformed findings array: ${decoded.topLevelError}`,\n    )\n  }\n  for (const rejection of decoded.rejected) {\n    onRejectedFinding(\n      rejection.index,\n      `${rejection.code}${rejection.path ? ` at ${rejection.path}` : ''}: ${rejection.message}`,\n    )\n  }\n  return {\n    answer: envelope.answer,\n    findings: decoded.accepted,\n    rejectedFindings: decoded.rejected.length,\n  }\n}\n\nfunction coerceReportEnvelope(content: string, analystId: string): Record<string, unknown> {\n  let value: unknown\n  try {\n    value = JSON.parse(extractJsonPayload(content))\n  } catch (error) {\n    throw new Error(`chat trace engine report for '${analystId}' was not JSON: ${messageOf(error)}`)\n  }\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new Error(`chat trace engine report for '${analystId}' was not a JSON object`)\n  }\n  return value as Record<string, unknown>\n}\n\nfunction resolveModel(options: ChatTraceEngineOptions): string {\n  const model = options.model ?? options.chat.defaultModel\n  if (typeof model !== 'string' || !model.trim() || model !== model.trim()) {\n    throw new TypeError(\n      'chat trace engine needs a model: pass ChatTraceEngineOptions.model or bind defaultModel on the ChatClient',\n    )\n  }\n  return model\n}\n\nfunction messageOf(error: unknown): string {\n  return error instanceof Error ? error.message : String(error)\n}\n","/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport type { ChatClient } from './analyst/chat-client'\nimport { paidJsonChat } from './chat-json-call'\nimport {\n  CostLedger,\n  type CostLedgerHandle,\n  type CostReceipt,\n  type CustomTokenPricing,\n} from './cost-ledger'\nimport type { LlmCallRequest } from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** Caller-owned transport. Required: agent-eval executes no paid model. */\n  chat: ChatClient\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: { maxPerFileChars: number; maxSourceChars: number; maxHtmlChars: number },\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions,\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts = {\n    chat: options.chat,\n    model: options.model ?? options.chat.defaultModel ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await paidJsonChat<{ summary: string; concepts: ConceptFinding[] }>({\n      chat: opts.chat,\n      request,\n      ledger: opts.costLedger,\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      tags: opts.costTags,\n      signal: opts.signal,\n      ...(opts.pricing ? { pricing: opts.pricing } : {}),\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions,\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,MAAM,4BAA4B;AAClC,MAAM,YAAY;;AAGlB,MAAM,oBAAoB;;;;;;AAqC1B,SAAS,wBAA2E;CAClF,MAAM,MAAM,EAAE,aAAa,yBAAyB,EAAE,QAAQ,UAAU,CAAC;CAIzE,OAAO,IAAI;CACX,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,sBAAsB;GACtB,UAAU,CAAC,UAAU,UAAU;GAC/B,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;IAAuC;IAC9E,UAAU;KAAE,MAAM;KAAS,OAAO;IAAI;GACxC;EACF;CACF;AACF;AAEA,MAAM,qBAAqB,sBAAsB;;;;;AAMjD,SAAgB,sBAAsB,SAAsD;CAC1F,MAAM,QAAQ,aAAa,OAAO;CAClC,MAAM,kBAAkB,QAAQ,mBAAA;CAChC,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,mEAAmE;CAEzF,IACE,QAAQ,gBAAgB,KAAA,MACvB,CAAC,OAAO,SAAS,QAAQ,WAAW,KAAK,QAAQ,cAAc,IAEhE,MAAM,IAAI,UAAU,oEAAoE;CAE1F,IACE,QAAQ,qBAAqB,KAAA,MAC5B,CAAC,OAAO,cAAc,QAAQ,gBAAgB,KAAK,QAAQ,oBAAoB,IAEhF,MAAM,IAAI,UAAU,oEAAoE;CAE1F,MAAM,eAAe,QAAQ,gBAAgB;CAC7C,IAAI,iBAAiB,iBAAiB,iBAAiB,eACrD,MAAM,IAAI,UAAU,mEAAmE;CAGzF,OAAO;EACL,IAAI;EACJ,aACE;EACF;EACA,SAAS;EACT,iBAAiB;GACf,MAAM;GACN;GACA,WAAW,QAAQ,KAAK;GACxB,kBAAkB,QAAQ,KAAK,mBAAmB;GAClD,SAAS,QAAQ,UAAU,EAAE,GAAG,QAAQ,QAAQ,IAAI;GACpD,mBAAmB;GACnB,aAAa,QAAQ,eAAe;GACpC,UAAU,QAAQ,YAAY;GAC9B,oBAAoB,QAAQ,oBAAoB;GAChD,eAAe;GACf,oBAAoB,mBAAmB;EACzC;EACA,UAAU,YAAY,QAAQ,SAAS;GAAE,GAAG;GAAS;GAAO;GAAiB;EAAa,CAAC;CAC7F;AACF;AAQA,eAAe,QACb,SACA,SACoC;CAGpC,IAAI,QAAQ,OAAO,cAAc,GAC/B,MAAM,IAAI,MACR,yGAAyG,QAAQ,UAAU,aAAa,QAAQ,OAAO,YAAY,EACrK;CAEF,MAAM,cAAc,IAAI,IAAI,QAAQ,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAC1E,IAAI,YAAY,SAAS,QAAQ,MAAM,QACrC,MAAM,IAAI,MAAM,wDAAwD,QAAQ,UAAU,EAAE;CAE9F,MAAM,kBAAuC,QAAQ,MAAM,KAAK,UAAU;EACxE,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAM,aAAa,KAAK;GAAa,YAAY,KAAK;EAAW;CAC1F,EAAE;CACF,MAAM,qBAAqB,KAAK,IAC9B,GACA,KAAK,IAAI,QAAQ,OAAO,eAAe,QAAQ,OAAO,cAAc,CAAC,CACvE;CAEA,MAAM,WAAyB,CAC7B;EAAE,MAAM;EAAU,SAAS,aAAa,SAAS,kBAAkB;CAAE,CACvE;CACA,IAAI,QAAQ,YACV,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,iBAAiB,QAAQ,UAAU;CAAE,CAAC;CAE/E,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,QAAQ;CAAS,CAAC;CAEzD,QAAQ,MAAM,gCAAgC;EAC5C,QAAQ;EACR,OAAO,QAAQ;EACf,WAAW,QAAQ,KAAK;EACxB,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;EAC5C,QAAQ,QAAQ;CAClB,CAAC;CAED,MAAM,aAAwB,CAAC;CAC/B,MAAM,+BAAe,IAAI,IAAY;CACrC,IAAI,aAAa;CACjB,IAAI,YAAY;CAChB,IAAI,uBAAuB;CAC3B,IAAI,sBAAsB;CAC1B,IAAI,YAAY;CAChB,IAAI,kBAAkB;CAEtB,KAAK,IAAI,OAAO,GAAG,OAAO,oBAAoB,QAAQ;EACpD,YAAY,OAAO;EACnB,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS;GACjD;GACA,OAAO;GACP,YAAY;GACZ,SAAS;EACX,CAAC;EACD,cAAc;EACd,aAAa,IAAI,SAAS,eAAe,YAAY;EACrD,MAAM,YAAY,SAAS,aAAa,CAAC;EACzC,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,SAAS;GAClB,GAAI,UAAU,SAAS,IAAI,EAAE,WAAW,UAAU,IAAI,CAAC;EACzD,CAAC;EACD,WAAW,KAAK;GACd,MAAM;GACN,OAAO;GACP,SAAS,SAAS;GAClB,YAAY,UAAU,KAAK,UAAU;IAAE,IAAI,KAAK;IAAI,MAAM,KAAK;GAAK,EAAE;GACtE,eAAe,SAAS,gBAAgB;EAC1C,CAAC;EACD,IAAI,UAAU,WAAW,GAAG;GAC1B,kBAAkB;GAClB;EACF;EACA,KAAK,MAAM,QAAQ,WAAW;GAC5B,IAAI,aAAa,QAAQ,OAAO,cAAc;IAC5C,sBAAsB;IAGtB,SAAS,KAAK,YAAY,MAAM,oBAAoB,QAAQ,OAAO,YAAY,CAAC,CAAC;IACjF;GACF;GACA,aAAa;GACb,MAAM,WAAW,MAAM,YAAY,MAAM,aAAa,OAAO;GAC7D,IAAI,SAAS,WAAW,wBAAwB;GAChD,SAAS,KAAK,YAAY,MAAM,SAAS,OAAO,CAAC;GACjD,WAAW,KAAK;IACd,MAAM;IACN,OAAO;IACP,MAAM,KAAK;IACX,IAAI,SAAS;IACb,WAAW,SAAS;IACpB,cAAc,SAAS,QAAQ;GACjC,CAAC;EACH;EACA,IAAI,qBAAqB;CAC3B;CAEA,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,aAAa,QAAQ,YAAY;CAAE,CAAC;CAC3E,MAAM,SAAS,MAAM,UAAU,SAAS,SAAS;EAC/C;EACA,SAAS;EACT,MAAM,QAAQ;CAChB,CAAC;CACD,cAAc;CACd,aAAa,IAAI,OAAO,eAAe,YAAY;CACnD,WAAW,KAAK;EACd,MAAM,YAAY;EAClB,OAAO;EACP,SAAS,OAAO;EAChB,eAAe,OAAO,gBAAgB;CACxC,CAAC;CAED,MAAM,SAAS,YAAY,OAAO,SAAS,QAAQ,YAAY,OAAO,WAAW;EAC/E,QAAQ,MAAM,yDAAyD;GACrE,QAAQ;GACR;GACA;EACF,CAAC;CACH,CAAC;CAED,MAAM,SAAoC;EACxC,QAAQ,OAAO;EACf,UAAU,OAAO;EACjB;EACA;EACA;EACA,SAAS;GACP,QAAQ;GACR,OAAO,QAAQ;GACf,WAAW,QAAQ,KAAK;GACxB,eAAe,QAAQ;GACvB,qBAAqB;GACrB,iCAAiC;GACjC,uBAAuB;GACvB,wBAAwB;GACxB,mBAAmB,OAAO;GAG1B,eAAe,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;GACtC,aAAa,QAAQ,aAAa,qBAAqB;EACzD;CACF;CACA,QAAQ,MAAM,kCAAkC;EAC9C,QAAQ;EACR,aAAa,OAAO;EACpB,YAAY,OAAO;EACnB,UAAU,OAAO,SAAS;CAC5B,CAAC;CACD,OAAO;AACT;AASA,eAAe,UACb,SACA,SACA,MAOoB;CACpB,MAAM,cAA8B;EAClC,OAAO,QAAQ;EAEf,UAAU,KAAK,SAAS,KAAK,aAAa,EAAE,GAAG,QAAQ,EAAE;EACzD,WAAW,QAAQ;EACnB,GAAI,KAAK,SAAS,KAAK,MAAM,SAAS,IAClC;GAAE,OAAO,KAAK;GAAO,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EAAG,IACjF,CAAC;EACL,GAAI,KAAK,OAAO,EAAE,UAAU,KAAK,IAAI,CAAC;EACtC,GAAI,KAAK,SAAS,gBAAgB,EAAE,YAAY,mBAAmB,IAAI,CAAC;EACxE,GAAI,QAAQ,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,YAAY;EAChF,GAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;EACvE,GAAI,QAAQ,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,WAAW,QAAQ,iBAAiB;CAC1F;CACA,MAAM,OAAO,MAAM,SAAS;EAC1B,MAAM,QAAQ;EACd,SAAS;EACT,QAAQ,QAAQ;EAChB,SAAS;EACT,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;EACrD,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;CACrD,CAAC;CACD,IAAI,CAAC,KAAK,WAGR,MAAM,IAAI,MACR,qBAAqB,KAAK,QAAQ,oBAAoB,QAAQ,UAAU,KAAK,KAAK,MAAM,WACxF,EAAE,OAAO,KAAK,MAAM,CACtB;CAEF,OAAO;EACL,SAAS,KAAK,SAAS;EACvB,GAAI,KAAK,SAAS,YAAY,EAAE,WAAW,KAAK,SAAS,UAAU,IAAI,CAAC;EACxE,GAAI,KAAK,SAAS,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,KAAK,SAAS,YAAY;EAC5F,GAAI,KAAK,SAAS,iBAAiB,KAAA,IAC/B,CAAC,IACD,EAAE,cAAc,KAAK,SAAS,aAAa;CACjD;AACF;AAQA,eAAe,YACb,MACA,aACA,SACuB;CACvB,MAAM,aAAa,YAAY,IAAI,KAAK,IAAI;CAC5C,IAAI,CAAC,YACH,OAAO;EACL,IAAI;EACJ,WAAW;EACX,SAAS,UACP,iBAAiB,KAAK,KAAK,yBAAyB,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC9F;CACF;CAEF,IAAI;CACJ,IAAI;EACF,OAAO,KAAK,cAAc,KAAK,MAAM,KAAK,CAAC,IAAI,KAAK,MAAM,KAAK,aAAa;CAC9E,SAAS,OAAO;EACd,OAAO;GACL,IAAI;GACJ,WAAW;GACX,SAAS,UACP,kBAAkB,KAAK,KAAK,mBAAmB,UAAU,KAAK,EAAE,6BAClE;EACF;CACF;CACA,IAAI;EACF,MAAM,QAAQ,MAAM,WAAW,QAC7B,MACA,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,KAAA,CAChD;EACA,OAAO,oBAAoB,KAAK,UAAU,SAAS,IAAI,GAAG,QAAQ,OAAO,cAAc;CACzF,SAAS,OAAO;EAGd,IAAI,QAAQ,QAAQ,SAAS,MAAM;EACnC,QAAQ,MAAM,qBAAqB;GACjC,QAAQ;GACR,YAAY,QAAQ;GACpB,MAAM,KAAK;GACX,QAAQ,UAAU,KAAK;EACzB,CAAC;EACD,OAAO;GAAE,IAAI;GAAO,WAAW;GAAO,SAAS,UAAU,UAAU,KAAK,CAAC;EAAE;CAC7E;AACF;AAEA,SAAS,oBAAoB,SAAiB,gBAAsC;CAClF,IAAI,QAAQ,UAAU,gBAAgB,OAAO;EAAE,IAAI;EAAM,WAAW;EAAO;CAAQ;CACnF,MAAM,OAAO,KAAK,IAAI,GAAG,iBAAiB,EAAwB;CAClE,OAAO;EAAE,IAAI;EAAM,WAAW;EAAM,SAAS,GAAG,QAAQ,MAAM,GAAG,IAAI,IAAI;CAAoB;AAC/F;AAEA,SAAS,YAAY,MAAmB,SAA6B;CACnE,OAAO;EAAE,MAAM;EAAQ,YAAY,KAAK;EAAI;CAAQ;AACtD;AAEA,SAAS,UAAU,QAAwB;CACzC,OAAO,KAAK,UAAU,EAAE,OAAO,OAAO,CAAC;AACzC;AAEA,SAAS,oBAAoB,cAA8B;CACzD,OAAO,UACL,4BAA4B,aAAa,uDAC3C;AACF;AAEA,SAAS,aAAa,SAAqC,oBAAoC;CAC7F,OAAO,CACL,QAAQ,aAAa,KAAK,GAC1B;EACE;EACA,cAAc,mBAAmB,mCAAmC,QAAQ,OAAO,aAAa;EAChG;EACA,+BAA+B,QAAQ,OAAO,eAAe;EAC7D;EACA;CACF,CAAC,CAAC,KAAK,IAAI,CACb,CAAC,CAAC,KAAK,MAAM;AACf;AAEA,SAAS,aAAa,cAAqD;CACzE,OAAO;EACL;EACA;EACA;EACA;EACA,iBAAiB,gBACb,0EACA;CACN,CAAC,CAAC,KAAK,IAAI;AACb;AAEA,SAAS,iBAAiB,YAAuD;CAI/E,OAAO,CACL,6FACA,KAAK,UAAU,YAAY,MAAM,CAAC,CACpC,CAAC,CAAC,KAAK,IAAI;AACb;AAQA,SAAS,YACP,SACA,WACA,mBACc;CACd,MAAM,WAAW,qBAAqB,SAAS,SAAS;CACxD,IAAI,OAAO,SAAS,WAAW,YAAY,CAAC,SAAS,OAAO,KAAK,GAC/D,MAAM,IAAI,MAAM,iCAAiC,UAAU,oBAAoB;CAIjF,MAAM,UAAU,sBAAsB,SAAS,QAAQ;CACvD,IAAI,QAAQ,kBAAkB,KAAA,GAC5B,MAAM,IAAI,MACR,iCAAiC,UAAU,oCAAoC,QAAQ,eACzF;CAEF,KAAK,MAAM,aAAa,QAAQ,UAC9B,kBACE,UAAU,OACV,GAAG,UAAU,OAAO,UAAU,OAAO,OAAO,UAAU,SAAS,GAAG,IAAI,UAAU,SAClF;CAEF,OAAO;EACL,QAAQ,SAAS;EACjB,UAAU,QAAQ;EAClB,kBAAkB,QAAQ,SAAS;CACrC;AACF;AAEA,SAAS,qBAAqB,SAAiB,WAA4C;CACzF,IAAI;CACJ,IAAI;EACF,QAAQ,KAAK,MAAM,mBAAmB,OAAO,CAAC;CAChD,SAAS,OAAO;EACd,MAAM,IAAI,MAAM,iCAAiC,UAAU,kBAAkB,UAAU,KAAK,GAAG;CACjG;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,MAAM,iCAAiC,UAAU,wBAAwB;CAErF,OAAO;AACT;AAEA,SAAS,aAAa,SAAyC;CAC7D,MAAM,QAAQ,QAAQ,SAAS,QAAQ,KAAK;CAC5C,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,UAAU,MAAM,KAAK,GACrE,MAAM,IAAI,UACR,2GACF;CAEF,OAAO;AACT;AAEA,SAAS,UAAU,OAAwB;CACzC,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;;;;;;;;;;;;;AChiBA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;AC9BA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAmCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,SACqC;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAAO;EACX,MAAM,QAAQ;EACd,OAAO,QAAQ,SAAS,QAAQ,KAAK,gBAAgB;EACrD,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,aAA8D;GAC/E,MAAM,KAAK;GACX;GACA,QAAQ,KAAK;GACb,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAClD,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU;EAElB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF"}
+{"version":3,"file":"semantic-concept-judge-_3Fp6Q_3.js","names":[],"sources":["../src/analyst/chat-trace-engine.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * A `TraceAnalysisEngine` that runs entirely inside Node against a\n * caller-owned `ChatClient`.\n *\n * The other engine in this package, `createDspyRlmTraceEngine`, reaches the\n * DSPy RLM through a Python subprocess. That made every model-backed analyst\n * unreachable for a consumer that already owns a model seam and no Python:\n * `createTraceAnalyst` requires an engine, the only exported constructor\n * required a Python runner, so `buildDefaultAnalystRegistry()` registered the\n * deterministic analyst alone and `analystsFromRegistry` refused the result.\n *\n * This engine closes that path. It drives the same investigation contract —\n * bounded trace tools, a prose answer, a strict findings array — with native\n * function calls over the transport the caller already bound with\n * `createChatClient`. A caller holding a bare\n * `(request: LlmCallRequest) => Promise<LlmCallResult>` adapts it in one line:\n *\n *   createChatClient({ transport: 'custom', chat: call, defaultModel, maximumAttempts })\n *\n * agent-eval still executes no paid model and holds no provider credential.\n * Every call goes through `paidChat`, so the shared cost ledger reserves the\n * priced maximum before the call and settles the receipt after it.\n *\n * Shape of one investigation:\n *   1. Investigation turns. The model reads the trace store through the tools\n *      its analyst kind was given. A turn that requests no tool ends the\n *      phase; an exhausted iteration or tool budget also ends it.\n *   2. One report turn. The model returns `{ answer, findings }` as JSON,\n *      decoded by the same `decodeRawFindingArray` the Python bridge uses, so\n *      a row this engine accepts is a row that engine could report.\n */\n\nimport { z } from 'zod'\nimport { paidChat } from '../chat-json-call'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport {\n  extractJsonPayload,\n  type LlmCallRequest,\n  type LlmMessage,\n  type LlmThinkingMode,\n  type LlmToolCall,\n  type LlmToolDefinition,\n} from '../llm-client'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\nimport type { ChatClient } from './chat-client'\nimport {\n  DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n  type TraceAnalysisEngine,\n  type TraceAnalysisEngineRequest,\n  type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { RawAnalystFindingSchema } from './finding-signature'\n\n/** Bumped whenever this engine's execution behavior changes. */\nconst CHAT_TRACE_ENGINE_VERSION = '1.0.0'\nconst ENGINE_ID = 'chat-trace'\n\n/** Marker appended to a tool result cut down to the retained-output budget. */\nconst TRUNCATION_MARKER = '\\n…[truncated to the analyst maxOutputChars budget]'\n\nexport interface ChatTraceEngineOptions {\n  /**\n   * Caller-owned transport. Build it with `createChatClient`; agent-eval never\n   * receives the provider credential.\n   */\n  chat: ChatClient\n  /** Model for every call. Defaults to the client's `defaultModel`. */\n  model?: string\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  /** Completion cap per turn. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */\n  maxOutputTokens?: number\n  /** Sampling temperature. Omitted from the request when absent. */\n  temperature?: number\n  /** Provider reasoning mode. Omitted when the provider default should apply. */\n  thinking?: LlmThinkingMode\n  /** Per-call deadline handed to the transport. */\n  requestTimeoutMs?: number\n  /**\n   * How the report turn asks for structured output.\n   *\n   * `json-object` sends JSON mode, which every OpenAI-compatible endpoint\n   * accepts, and relies on the finding grammar the analyst kind already puts\n   * in its instructions. `json-schema` sends the generated response schema as\n   * well, which stricter endpoints honour and older ones reject. Default:\n   * `json-object`.\n   */\n  reportFormat?: 'json-object' | 'json-schema'\n}\n\n/**\n * The report envelope. Its `findings` rows are generated from\n * `RawAnalystFindingSchema`, so the schema offered to a provider and the\n * decoder that accepts the answer cannot drift apart.\n */\nfunction buildReportJsonSchema(): { name: string; schema: Record<string, unknown> } {\n  const row = z.toJSONSchema(RawAnalystFindingSchema, { target: 'draft-7' }) as Record<\n    string,\n    unknown\n  >\n  delete row.$schema\n  return {\n    name: 'trace_analysis_report',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      required: ['answer', 'findings'],\n      properties: {\n        answer: { type: 'string', description: 'Direct prose answer to the question.' },\n        findings: { type: 'array', items: row },\n      },\n    },\n  }\n}\n\nconst REPORT_JSON_SCHEMA = buildReportJsonSchema()\n\n/**\n * Run bounded recursive trace analysis in-process over a caller-owned chat\n * transport. No Python, no subprocess, no loopback proxy.\n */\nexport function createChatTraceEngine(options: ChatTraceEngineOptions): TraceAnalysisEngine {\n  const model = resolveModel(options)\n  const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS\n  if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n    throw new TypeError('chat trace engine maxOutputTokens must be a positive safe integer')\n  }\n  if (\n    options.temperature !== undefined &&\n    (!Number.isFinite(options.temperature) || options.temperature < 0)\n  ) {\n    throw new TypeError('chat trace engine temperature must be a non-negative finite number')\n  }\n  if (\n    options.requestTimeoutMs !== undefined &&\n    (!Number.isSafeInteger(options.requestTimeoutMs) || options.requestTimeoutMs <= 0)\n  ) {\n    throw new TypeError('chat trace engine requestTimeoutMs must be a positive safe integer')\n  }\n  const reportFormat = options.reportFormat ?? 'json-object'\n  if (reportFormat !== 'json-object' && reportFormat !== 'json-schema') {\n    throw new TypeError(`chat trace engine reportFormat must be json-object or json-schema`)\n  }\n\n  return {\n    id: ENGINE_ID,\n    description:\n      'In-process recursive trace analysis over a caller-owned ChatClient with native tool calls.',\n    model,\n    version: CHAT_TRACE_ENGINE_VERSION,\n    executionConfig: {\n      kind: ENGINE_ID,\n      model,\n      transport: options.chat.transport,\n      maximum_attempts: options.chat.maximumAttempts ?? null,\n      pricing: options.pricing ? { ...options.pricing } : null,\n      max_output_tokens: maxOutputTokens,\n      temperature: options.temperature ?? null,\n      thinking: options.thinking ?? null,\n      request_timeout_ms: options.requestTimeoutMs ?? null,\n      report_format: reportFormat,\n      report_schema_name: REPORT_JSON_SCHEMA.name,\n    },\n    analyze: (request) => analyze(request, { ...options, model, maxOutputTokens, reportFormat }),\n  }\n}\n\ninterface ResolvedOptions extends ChatTraceEngineOptions {\n  model: string\n  maxOutputTokens: number\n  reportFormat: 'json-object' | 'json-schema'\n}\n\nasync function analyze(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n): Promise<TraceAnalysisEngineResult> {\n  // One model call is spent on the report, so a budget of one call would buy\n  // an answer with no investigation behind it. Refuse instead of pretending.\n  if (request.limits.maxLlmCalls < 2) {\n    throw new Error(\n      `chat trace engine reserves one model call for the report, so maxLlmCalls must be at least 2 (analyst '${request.analystId}' declared ${request.limits.maxLlmCalls})`,\n    )\n  }\n  const toolsByName = new Map(request.tools.map((tool) => [tool.name, tool]))\n  if (toolsByName.size !== request.tools.length) {\n    throw new Error(`chat trace engine received duplicate tool names for '${request.analystId}'`)\n  }\n  const toolDefinitions: LlmToolDefinition[] = request.tools.map((tool) => ({\n    type: 'function',\n    function: { name: tool.name, description: tool.description, parameters: tool.parameters },\n  }))\n  const investigationTurns = Math.max(\n    1,\n    Math.min(request.limits.maxIterations, request.limits.maxLlmCalls - 1),\n  )\n\n  const messages: LlmMessage[] = [\n    { role: 'system', content: systemPrompt(request, investigationTurns) },\n  ]\n  if (request.taskInputs) {\n    messages.push({ role: 'user', content: renderTaskInputs(request.taskInputs) })\n  }\n  messages.push({ role: 'user', content: request.question })\n\n  request.log?.('trace analyst engine started', {\n    engine: ENGINE_ID,\n    model: options.model,\n    transport: options.chat.transport,\n    tools: request.tools.map((tool) => tool.name),\n    limits: request.limits,\n  })\n\n  const trajectory: unknown[] = []\n  const servedModels = new Set<string>()\n  let modelCalls = 0\n  let toolCalls = 0\n  let truncatedToolResults = 0\n  let toolBudgetExhausted = false\n  let turnsUsed = 0\n  let stoppedOnAnswer = false\n\n  for (let turn = 0; turn < investigationTurns; turn++) {\n    turnsUsed = turn + 1\n    const response = await callModel(request, options, {\n      messages,\n      tools: toolDefinitions,\n      toolChoice: 'auto',\n      purpose: 'investigation',\n    })\n    modelCalls += 1\n    servedModels.add(response.servedModel ?? 'unreported')\n    const requested = response.toolCalls ?? []\n    messages.push({\n      role: 'assistant',\n      content: response.content,\n      ...(requested.length > 0 ? { toolCalls: requested } : {}),\n    })\n    trajectory.push({\n      turn: turnsUsed,\n      phase: 'investigation',\n      content: response.content,\n      tool_calls: requested.map((call) => ({ id: call.id, name: call.name })),\n      finish_reason: response.finishReason ?? null,\n    })\n    if (requested.length === 0) {\n      stoppedOnAnswer = true\n      break\n    }\n    for (const call of requested) {\n      if (toolCalls >= request.limits.maxToolCalls) {\n        toolBudgetExhausted = true\n        // Every requested id still needs an answer: a provider rejects the\n        // next turn when one tool call is left unanswered.\n        messages.push(toolMessage(call, exhaustedToolBudget(request.limits.maxToolCalls)))\n        continue\n      }\n      toolCalls += 1\n      const executed = await executeTool(call, toolsByName, request)\n      if (executed.truncated) truncatedToolResults += 1\n      messages.push(toolMessage(call, executed.payload))\n      trajectory.push({\n        turn: turnsUsed,\n        phase: 'tool',\n        name: call.name,\n        ok: executed.ok,\n        truncated: executed.truncated,\n        result_chars: executed.payload.length,\n      })\n    }\n    if (toolBudgetExhausted) break\n  }\n\n  messages.push({ role: 'user', content: reportPrompt(options.reportFormat) })\n  const report = await callModel(request, options, {\n    messages,\n    purpose: 'report',\n    json: options.reportFormat,\n  })\n  modelCalls += 1\n  servedModels.add(report.servedModel ?? 'unreported')\n  trajectory.push({\n    turn: turnsUsed + 1,\n    phase: 'report',\n    content: report.content,\n    finish_reason: report.finishReason ?? null,\n  })\n\n  const parsed = parseReport(report.content, request.analystId, (index, reason) => {\n    request.log?.('finding rejected: report row failed schema validation', {\n      engine: ENGINE_ID,\n      index,\n      reason,\n    })\n  })\n\n  const result: TraceAnalysisEngineResult = {\n    answer: parsed.answer,\n    findings: parsed.findings,\n    trajectory,\n    modelCalls,\n    toolCalls,\n    runtime: {\n      engine: ENGINE_ID,\n      model: options.model,\n      transport: options.chat.transport,\n      report_format: options.reportFormat,\n      investigation_turns: turnsUsed,\n      investigation_stopped_on_answer: stoppedOnAnswer,\n      tool_budget_exhausted: toolBudgetExhausted,\n      truncated_tool_results: truncatedToolResults,\n      rejected_findings: parsed.rejectedFindings,\n      // A gateway can answer `model: X` from another model on HTTP 200, so the\n      // ids the provider echoed travel with the result as evidence.\n      served_models: [...servedModels].sort(),\n      task_inputs: request.taskInputs ? 'prompt-delivered' : 'none',\n    },\n  }\n  request.log?.('trace analyst engine completed', {\n    engine: ENGINE_ID,\n    model_calls: result.modelCalls,\n    tool_calls: result.toolCalls,\n    findings: result.findings.length,\n  })\n  return result\n}\n\ninterface ModelTurn {\n  content: string\n  toolCalls?: LlmToolCall[]\n  servedModel?: string | null\n  finishReason?: string | null\n}\n\nasync function callModel(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n  turn: {\n    messages: LlmMessage[]\n    tools?: LlmToolDefinition[]\n    toolChoice?: LlmCallRequest['toolChoice']\n    purpose: 'investigation' | 'report'\n    json?: 'json-object' | 'json-schema'\n  },\n): Promise<ModelTurn> {\n  const chatRequest: LlmCallRequest = {\n    model: options.model,\n    // The transport may mutate what it is handed; the conversation is ours.\n    messages: turn.messages.map((message) => ({ ...message })),\n    maxTokens: options.maxOutputTokens,\n    ...(turn.tools && turn.tools.length > 0\n      ? { tools: turn.tools, ...(turn.toolChoice ? { toolChoice: turn.toolChoice } : {}) }\n      : {}),\n    ...(turn.json ? { jsonMode: true } : {}),\n    ...(turn.json === 'json-schema' ? { jsonSchema: REPORT_JSON_SCHEMA } : {}),\n    ...(options.temperature === undefined ? {} : { temperature: options.temperature }),\n    ...(options.thinking === undefined ? {} : { thinking: options.thinking }),\n    ...(options.requestTimeoutMs === undefined ? {} : { timeoutMs: options.requestTimeoutMs }),\n  }\n  const paid = await paidChat({\n    chat: options.chat,\n    request: chatRequest,\n    ledger: request.costLedger,\n    channel: 'analyst',\n    phase: request.costPhase,\n    actor: request.analystId,\n    ...(request.costTags ? { tags: request.costTags } : {}),\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    ...(request.signal ? { signal: request.signal } : {}),\n  })\n  if (!paid.succeeded) {\n    // The transport already spent its own attempts. A failure here ends the\n    // investigation: continuing would report an answer the model never gave.\n    throw new Error(\n      `chat trace engine ${turn.purpose} call failed for '${request.analystId}': ${paid.error.message}`,\n      { cause: paid.error },\n    )\n  }\n  return {\n    content: paid.response.content,\n    ...(paid.response.toolCalls ? { toolCalls: paid.response.toolCalls } : {}),\n    ...(paid.response.servedModel === undefined ? {} : { servedModel: paid.response.servedModel }),\n    ...(paid.response.finishReason === undefined\n      ? {}\n      : { finishReason: paid.response.finishReason }),\n  }\n}\n\ninterface ExecutedTool {\n  ok: boolean\n  truncated: boolean\n  payload: string\n}\n\nasync function executeTool(\n  call: LlmToolCall,\n  toolsByName: ReadonlyMap<string, TraceAnalysisToolDescriptor>,\n  request: TraceAnalysisEngineRequest,\n): Promise<ExecutedTool> {\n  const descriptor = toolsByName.get(call.name)\n  if (!descriptor) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `unknown tool '${call.name}'; available tools are ${[...toolsByName.keys()].sort().join(', ')}`,\n      ),\n    }\n  }\n  let args: unknown\n  try {\n    args = call.argumentsJson.trim() === '' ? {} : JSON.parse(call.argumentsJson)\n  } catch (error) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `arguments for '${call.name}' were not JSON: ${messageOf(error)}. Send valid JSON arguments.`,\n      ),\n    }\n  }\n  try {\n    const value = await descriptor.handler(\n      args,\n      request.signal ? { signal: request.signal } : undefined,\n    )\n    return truncateToolPayload(JSON.stringify(value ?? null), request.limits.maxOutputChars)\n  } catch (error) {\n    // A bounded store refuses an oversized or malformed read by design. That\n    // refusal is the model's next instruction, not the run's failure.\n    if (request.signal?.aborted) throw error\n    request.log?.('trace tool failed', {\n      engine: ENGINE_ID,\n      analyst_id: request.analystId,\n      tool: call.name,\n      reason: messageOf(error),\n    })\n    return { ok: false, truncated: false, payload: toolError(messageOf(error)) }\n  }\n}\n\nfunction truncateToolPayload(payload: string, maxOutputChars: number): ExecutedTool {\n  if (payload.length <= maxOutputChars) return { ok: true, truncated: false, payload }\n  const keep = Math.max(0, maxOutputChars - TRUNCATION_MARKER.length)\n  return { ok: true, truncated: true, payload: `${payload.slice(0, keep)}${TRUNCATION_MARKER}` }\n}\n\nfunction toolMessage(call: LlmToolCall, content: string): LlmMessage {\n  return { role: 'tool', toolCallId: call.id, content }\n}\n\nfunction toolError(reason: string): string {\n  return JSON.stringify({ error: reason })\n}\n\nfunction exhaustedToolBudget(maxToolCalls: number): string {\n  return toolError(\n    `the trace-tool budget of ${maxToolCalls} calls is spent; answer from the evidence already read`,\n  )\n}\n\nfunction systemPrompt(request: TraceAnalysisEngineRequest, investigationTurns: number): string {\n  return [\n    request.instructions.trim(),\n    [\n      'HOW THIS INVESTIGATION RUNS:',\n      `- You have ${investigationTurns} investigation turns and at most ${request.limits.maxToolCalls} trace-tool calls.`,\n      '- Call the trace tools to read the store. Never state a trace fact you did not read.',\n      `- A tool result longer than ${request.limits.maxOutputChars} characters is truncated; narrow the query instead of asking again.`,\n      '- A tool result carrying an \"error\" field is feedback: fix the call or take another route.',\n      '- Answer with no tool call once you have the evidence. You are then asked for the final report.',\n    ].join('\\n'),\n  ].join('\\n\\n')\n}\n\nfunction reportPrompt(reportFormat: 'json-object' | 'json-schema'): string {\n  return [\n    'Report now. Do not call any more tools.',\n    'Return one JSON object with exactly two fields:',\n    '  \"answer\": a direct prose answer to the question.',\n    '  \"findings\": an array of findings in the schema above. Emit [] when there is nothing to report.',\n    reportFormat === 'json-object'\n      ? 'Return the object alone, with no surrounding prose and no code fence.'\n      : 'Return the object in the response schema you were given.',\n  ].join('\\n')\n}\n\nfunction renderTaskInputs(taskInputs: Readonly<Record<string, unknown>>): string {\n  // The engine contract forbids dropping structured inputs. This engine has no\n  // code environment to bind them as variables, so they are delivered whole as\n  // conversation material and the runtime record says which way they arrived.\n  return [\n    'TASK INPUTS — structured material delivered with the question, not fetched through tools:',\n    JSON.stringify(taskInputs, null, 2),\n  ].join('\\n')\n}\n\ninterface ParsedReport {\n  answer: string\n  findings: TraceAnalysisEngineResult['findings']\n  rejectedFindings: number\n}\n\nfunction parseReport(\n  content: string,\n  analystId: string,\n  onRejectedFinding: (index: number, reason: string) => void,\n): ParsedReport {\n  const envelope = coerceReportEnvelope(content, analystId)\n  if (typeof envelope.answer !== 'string' || !envelope.answer.trim()) {\n    throw new Error(`chat trace engine report for '${analystId}' carried no answer`)\n  }\n  // Findings are model output: one malformed row is model noise, not an engine\n  // fault, and the rest of the paid investigation must survive it.\n  const decoded = decodeRawFindingArray(envelope.findings)\n  if (decoded.topLevelError !== undefined) {\n    throw new Error(\n      `chat trace engine report for '${analystId}' had a malformed findings array: ${decoded.topLevelError}`,\n    )\n  }\n  for (const rejection of decoded.rejected) {\n    onRejectedFinding(\n      rejection.index,\n      `${rejection.code}${rejection.path ? ` at ${rejection.path}` : ''}: ${rejection.message}`,\n    )\n  }\n  return {\n    answer: envelope.answer,\n    findings: decoded.accepted,\n    rejectedFindings: decoded.rejected.length,\n  }\n}\n\nfunction coerceReportEnvelope(content: string, analystId: string): Record<string, unknown> {\n  let value: unknown\n  try {\n    value = JSON.parse(extractJsonPayload(content))\n  } catch (error) {\n    throw new Error(`chat trace engine report for '${analystId}' was not JSON: ${messageOf(error)}`)\n  }\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new Error(`chat trace engine report for '${analystId}' was not a JSON object`)\n  }\n  return value as Record<string, unknown>\n}\n\nfunction resolveModel(options: ChatTraceEngineOptions): string {\n  const model = options.model ?? options.chat.defaultModel\n  if (typeof model !== 'string' || !model.trim() || model !== model.trim()) {\n    throw new TypeError(\n      'chat trace engine needs a model: pass ChatTraceEngineOptions.model or bind defaultModel on the ChatClient',\n    )\n  }\n  return model\n}\n\nfunction messageOf(error: unknown): string {\n  return error instanceof Error ? error.message : String(error)\n}\n","/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport type { ChatClient } from './analyst/chat-client'\nimport { paidJsonChat } from './chat-json-call'\nimport {\n  CostLedger,\n  type CostLedgerHandle,\n  type CostReceipt,\n  type CustomTokenPricing,\n} from './cost-ledger'\nimport type { LlmCallRequest } from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** Caller-owned transport. Required: agent-eval executes no paid model. */\n  chat: ChatClient\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: { maxPerFileChars: number; maxSourceChars: number; maxHtmlChars: number },\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions,\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts = {\n    chat: options.chat,\n    model: options.model ?? options.chat.defaultModel ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await paidJsonChat<{ summary: string; concepts: ConceptFinding[] }>({\n      chat: opts.chat,\n      request,\n      ledger: opts.costLedger,\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      tags: opts.costTags,\n      signal: opts.signal,\n      ...(opts.pricing ? { pricing: opts.pricing } : {}),\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions,\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,MAAM,4BAA4B;AAClC,MAAM,YAAY;;AAGlB,MAAM,oBAAoB;;;;;;AAqC1B,SAAS,wBAA2E;CAClF,MAAM,MAAM,EAAE,aAAa,yBAAyB,EAAE,QAAQ,UAAU,CAAC;CAIzE,OAAO,IAAI;CACX,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,sBAAsB;GACtB,UAAU,CAAC,UAAU,UAAU;GAC/B,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;IAAuC;IAC9E,UAAU;KAAE,MAAM;KAAS,OAAO;IAAI;GACxC;EACF;CACF;AACF;AAEA,MAAM,qBAAqB,sBAAsB;;;;;AAMjD,SAAgB,sBAAsB,SAAsD;CAC1F,MAAM,QAAQ,aAAa,OAAO;CAClC,MAAM,kBAAkB,QAAQ,mBAAA;CAChC,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,mEAAmE;CAEzF,IACE,QAAQ,gBAAgB,KAAA,MACvB,CAAC,OAAO,SAAS,QAAQ,WAAW,KAAK,QAAQ,cAAc,IAEhE,MAAM,IAAI,UAAU,oEAAoE;CAE1F,IACE,QAAQ,qBAAqB,KAAA,MAC5B,CAAC,OAAO,cAAc,QAAQ,gBAAgB,KAAK,QAAQ,oBAAoB,IAEhF,MAAM,IAAI,UAAU,oEAAoE;CAE1F,MAAM,eAAe,QAAQ,gBAAgB;CAC7C,IAAI,iBAAiB,iBAAiB,iBAAiB,eACrD,MAAM,IAAI,UAAU,mEAAmE;CAGzF,OAAO;EACL,IAAI;EACJ,aACE;EACF;EACA,SAAS;EACT,iBAAiB;GACf,MAAM;GACN;GACA,WAAW,QAAQ,KAAK;GACxB,kBAAkB,QAAQ,KAAK,mBAAmB;GAClD,SAAS,QAAQ,UAAU,EAAE,GAAG,QAAQ,QAAQ,IAAI;GACpD,mBAAmB;GACnB,aAAa,QAAQ,eAAe;GACpC,UAAU,QAAQ,YAAY;GAC9B,oBAAoB,QAAQ,oBAAoB;GAChD,eAAe;GACf,oBAAoB,mBAAmB;EACzC;EACA,UAAU,YAAY,QAAQ,SAAS;GAAE,GAAG;GAAS;GAAO;GAAiB;EAAa,CAAC;CAC7F;AACF;AAQA,eAAe,QACb,SACA,SACoC;CAGpC,IAAI,QAAQ,OAAO,cAAc,GAC/B,MAAM,IAAI,MACR,yGAAyG,QAAQ,UAAU,aAAa,QAAQ,OAAO,YAAY,EACrK;CAEF,MAAM,cAAc,IAAI,IAAI,QAAQ,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAC1E,IAAI,YAAY,SAAS,QAAQ,MAAM,QACrC,MAAM,IAAI,MAAM,wDAAwD,QAAQ,UAAU,EAAE;CAE9F,MAAM,kBAAuC,QAAQ,MAAM,KAAK,UAAU;EACxE,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAM,aAAa,KAAK;GAAa,YAAY,KAAK;EAAW;CAC1F,EAAE;CACF,MAAM,qBAAqB,KAAK,IAC9B,GACA,KAAK,IAAI,QAAQ,OAAO,eAAe,QAAQ,OAAO,cAAc,CAAC,CACvE;CAEA,MAAM,WAAyB,CAC7B;EAAE,MAAM;EAAU,SAAS,aAAa,SAAS,kBAAkB;CAAE,CACvE;CACA,IAAI,QAAQ,YACV,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,iBAAiB,QAAQ,UAAU;CAAE,CAAC;CAE/E,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,QAAQ;CAAS,CAAC;CAEzD,QAAQ,MAAM,gCAAgC;EAC5C,QAAQ;EACR,OAAO,QAAQ;EACf,WAAW,QAAQ,KAAK;EACxB,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;EAC5C,QAAQ,QAAQ;CAClB,CAAC;CAED,MAAM,aAAwB,CAAC;CAC/B,MAAM,+BAAe,IAAI,IAAY;CACrC,IAAI,aAAa;CACjB,IAAI,YAAY;CAChB,IAAI,uBAAuB;CAC3B,IAAI,sBAAsB;CAC1B,IAAI,YAAY;CAChB,IAAI,kBAAkB;CAEtB,KAAK,IAAI,OAAO,GAAG,OAAO,oBAAoB,QAAQ;EACpD,YAAY,OAAO;EACnB,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS;GACjD;GACA,OAAO;GACP,YAAY;GACZ,SAAS;EACX,CAAC;EACD,cAAc;EACd,aAAa,IAAI,SAAS,eAAe,YAAY;EACrD,MAAM,YAAY,SAAS,aAAa,CAAC;EACzC,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,SAAS;GAClB,GAAI,UAAU,SAAS,IAAI,EAAE,WAAW,UAAU,IAAI,CAAC;EACzD,CAAC;EACD,WAAW,KAAK;GACd,MAAM;GACN,OAAO;GACP,SAAS,SAAS;GAClB,YAAY,UAAU,KAAK,UAAU;IAAE,IAAI,KAAK;IAAI,MAAM,KAAK;GAAK,EAAE;GACtE,eAAe,SAAS,gBAAgB;EAC1C,CAAC;EACD,IAAI,UAAU,WAAW,GAAG;GAC1B,kBAAkB;GAClB;EACF;EACA,KAAK,MAAM,QAAQ,WAAW;GAC5B,IAAI,aAAa,QAAQ,OAAO,cAAc;IAC5C,sBAAsB;IAGtB,SAAS,KAAK,YAAY,MAAM,oBAAoB,QAAQ,OAAO,YAAY,CAAC,CAAC;IACjF;GACF;GACA,aAAa;GACb,MAAM,WAAW,MAAM,YAAY,MAAM,aAAa,OAAO;GAC7D,IAAI,SAAS,WAAW,wBAAwB;GAChD,SAAS,KAAK,YAAY,MAAM,SAAS,OAAO,CAAC;GACjD,WAAW,KAAK;IACd,MAAM;IACN,OAAO;IACP,MAAM,KAAK;IACX,IAAI,SAAS;IACb,WAAW,SAAS;IACpB,cAAc,SAAS,QAAQ;GACjC,CAAC;EACH;EACA,IAAI,qBAAqB;CAC3B;CAEA,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,aAAa,QAAQ,YAAY;CAAE,CAAC;CAC3E,MAAM,SAAS,MAAM,UAAU,SAAS,SAAS;EAC/C;EACA,SAAS;EACT,MAAM,QAAQ;CAChB,CAAC;CACD,cAAc;CACd,aAAa,IAAI,OAAO,eAAe,YAAY;CACnD,WAAW,KAAK;EACd,MAAM,YAAY;EAClB,OAAO;EACP,SAAS,OAAO;EAChB,eAAe,OAAO,gBAAgB;CACxC,CAAC;CAED,MAAM,SAAS,YAAY,OAAO,SAAS,QAAQ,YAAY,OAAO,WAAW;EAC/E,QAAQ,MAAM,yDAAyD;GACrE,QAAQ;GACR;GACA;EACF,CAAC;CACH,CAAC;CAED,MAAM,SAAoC;EACxC,QAAQ,OAAO;EACf,UAAU,OAAO;EACjB;EACA;EACA;EACA,SAAS;GACP,QAAQ;GACR,OAAO,QAAQ;GACf,WAAW,QAAQ,KAAK;GACxB,eAAe,QAAQ;GACvB,qBAAqB;GACrB,iCAAiC;GACjC,uBAAuB;GACvB,wBAAwB;GACxB,mBAAmB,OAAO;GAG1B,eAAe,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;GACtC,aAAa,QAAQ,aAAa,qBAAqB;EACzD;CACF;CACA,QAAQ,MAAM,kCAAkC;EAC9C,QAAQ;EACR,aAAa,OAAO;EACpB,YAAY,OAAO;EACnB,UAAU,OAAO,SAAS;CAC5B,CAAC;CACD,OAAO;AACT;AASA,eAAe,UACb,SACA,SACA,MAOoB;CACpB,MAAM,cAA8B;EAClC,OAAO,QAAQ;EAEf,UAAU,KAAK,SAAS,KAAK,aAAa,EAAE,GAAG,QAAQ,EAAE;EACzD,WAAW,QAAQ;EACnB,GAAI,KAAK,SAAS,KAAK,MAAM,SAAS,IAClC;GAAE,OAAO,KAAK;GAAO,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EAAG,IACjF,CAAC;EACL,GAAI,KAAK,OAAO,EAAE,UAAU,KAAK,IAAI,CAAC;EACtC,GAAI,KAAK,SAAS,gBAAgB,EAAE,YAAY,mBAAmB,IAAI,CAAC;EACxE,GAAI,QAAQ,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,YAAY;EAChF,GAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;EACvE,GAAI,QAAQ,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,WAAW,QAAQ,iBAAiB;CAC1F;CACA,MAAM,OAAO,MAAM,SAAS;EAC1B,MAAM,QAAQ;EACd,SAAS;EACT,QAAQ,QAAQ;EAChB,SAAS;EACT,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;EACrD,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;CACrD,CAAC;CACD,IAAI,CAAC,KAAK,WAGR,MAAM,IAAI,MACR,qBAAqB,KAAK,QAAQ,oBAAoB,QAAQ,UAAU,KAAK,KAAK,MAAM,WACxF,EAAE,OAAO,KAAK,MAAM,CACtB;CAEF,OAAO;EACL,SAAS,KAAK,SAAS;EACvB,GAAI,KAAK,SAAS,YAAY,EAAE,WAAW,KAAK,SAAS,UAAU,IAAI,CAAC;EACxE,GAAI,KAAK,SAAS,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,KAAK,SAAS,YAAY;EAC5F,GAAI,KAAK,SAAS,iBAAiB,KAAA,IAC/B,CAAC,IACD,EAAE,cAAc,KAAK,SAAS,aAAa;CACjD;AACF;AAQA,eAAe,YACb,MACA,aACA,SACuB;CACvB,MAAM,aAAa,YAAY,IAAI,KAAK,IAAI;CAC5C,IAAI,CAAC,YACH,OAAO;EACL,IAAI;EACJ,WAAW;EACX,SAAS,UACP,iBAAiB,KAAK,KAAK,yBAAyB,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC9F;CACF;CAEF,IAAI;CACJ,IAAI;EACF,OAAO,KAAK,cAAc,KAAK,MAAM,KAAK,CAAC,IAAI,KAAK,MAAM,KAAK,aAAa;CAC9E,SAAS,OAAO;EACd,OAAO;GACL,IAAI;GACJ,WAAW;GACX,SAAS,UACP,kBAAkB,KAAK,KAAK,mBAAmB,UAAU,KAAK,EAAE,6BAClE;EACF;CACF;CACA,IAAI;EACF,MAAM,QAAQ,MAAM,WAAW,QAC7B,MACA,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,KAAA,CAChD;EACA,OAAO,oBAAoB,KAAK,UAAU,SAAS,IAAI,GAAG,QAAQ,OAAO,cAAc;CACzF,SAAS,OAAO;EAGd,IAAI,QAAQ,QAAQ,SAAS,MAAM;EACnC,QAAQ,MAAM,qBAAqB;GACjC,QAAQ;GACR,YAAY,QAAQ;GACpB,MAAM,KAAK;GACX,QAAQ,UAAU,KAAK;EACzB,CAAC;EACD,OAAO;GAAE,IAAI;GAAO,WAAW;GAAO,SAAS,UAAU,UAAU,KAAK,CAAC;EAAE;CAC7E;AACF;AAEA,SAAS,oBAAoB,SAAiB,gBAAsC;CAClF,IAAI,QAAQ,UAAU,gBAAgB,OAAO;EAAE,IAAI;EAAM,WAAW;EAAO;CAAQ;CACnF,MAAM,OAAO,KAAK,IAAI,GAAG,iBAAiB,EAAwB;CAClE,OAAO;EAAE,IAAI;EAAM,WAAW;EAAM,SAAS,GAAG,QAAQ,MAAM,GAAG,IAAI,IAAI;CAAoB;AAC/F;AAEA,SAAS,YAAY,MAAmB,SAA6B;CACnE,OAAO;EAAE,MAAM;EAAQ,YAAY,KAAK;EAAI;CAAQ;AACtD;AAEA,SAAS,UAAU,QAAwB;CACzC,OAAO,KAAK,UAAU,EAAE,OAAO,OAAO,CAAC;AACzC;AAEA,SAAS,oBAAoB,cAA8B;CACzD,OAAO,UACL,4BAA4B,aAAa,uDAC3C;AACF;AAEA,SAAS,aAAa,SAAqC,oBAAoC;CAC7F,OAAO,CACL,QAAQ,aAAa,KAAK,GAC1B;EACE;EACA,cAAc,mBAAmB,mCAAmC,QAAQ,OAAO,aAAa;EAChG;EACA,+BAA+B,QAAQ,OAAO,eAAe;EAC7D;EACA;CACF,CAAC,CAAC,KAAK,IAAI,CACb,CAAC,CAAC,KAAK,MAAM;AACf;AAEA,SAAS,aAAa,cAAqD;CACzE,OAAO;EACL;EACA;EACA;EACA;EACA,iBAAiB,gBACb,0EACA;CACN,CAAC,CAAC,KAAK,IAAI;AACb;AAEA,SAAS,iBAAiB,YAAuD;CAI/E,OAAO,CACL,6FACA,KAAK,UAAU,YAAY,MAAM,CAAC,CACpC,CAAC,CAAC,KAAK,IAAI;AACb;AAQA,SAAS,YACP,SACA,WACA,mBACc;CACd,MAAM,WAAW,qBAAqB,SAAS,SAAS;CACxD,IAAI,OAAO,SAAS,WAAW,YAAY,CAAC,SAAS,OAAO,KAAK,GAC/D,MAAM,IAAI,MAAM,iCAAiC,UAAU,oBAAoB;CAIjF,MAAM,UAAU,sBAAsB,SAAS,QAAQ;CACvD,IAAI,QAAQ,kBAAkB,KAAA,GAC5B,MAAM,IAAI,MACR,iCAAiC,UAAU,oCAAoC,QAAQ,eACzF;CAEF,KAAK,MAAM,aAAa,QAAQ,UAC9B,kBACE,UAAU,OACV,GAAG,UAAU,OAAO,UAAU,OAAO,OAAO,UAAU,SAAS,GAAG,IAAI,UAAU,SAClF;CAEF,OAAO;EACL,QAAQ,SAAS;EACjB,UAAU,QAAQ;EAClB,kBAAkB,QAAQ,SAAS;CACrC;AACF;AAEA,SAAS,qBAAqB,SAAiB,WAA4C;CACzF,IAAI;CACJ,IAAI;EACF,QAAQ,KAAK,MAAM,mBAAmB,OAAO,CAAC;CAChD,SAAS,OAAO;EACd,MAAM,IAAI,MAAM,iCAAiC,UAAU,kBAAkB,UAAU,KAAK,GAAG;CACjG;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,MAAM,iCAAiC,UAAU,wBAAwB;CAErF,OAAO;AACT;AAEA,SAAS,aAAa,SAAyC;CAC7D,MAAM,QAAQ,QAAQ,SAAS,QAAQ,KAAK;CAC5C,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,UAAU,MAAM,KAAK,GACrE,MAAM,IAAI,UACR,2GACF;CAEF,OAAO;AACT;AAEA,SAAS,UAAU,OAAwB;CACzC,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;;;;;;;;;;;;;AChiBA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;AC9BA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAmCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,SACqC;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAAO;EACX,MAAM,QAAQ;EACd,OAAO,QAAQ,SAAS,QAAQ,KAAK,gBAAgB;EACrD,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,aAA8D;GAC/E,MAAM,KAAK;GACX;GACA,QAAQ,KAAK;GACb,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAClD,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU;EAElB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF"}

package/dist/{sequential-DAsyV2T9.js → sequential-D_xHrBZe.js}

@@ -1,7 +1,7 @@
 import { h as manifestContentDigest } from "./agent-profile-cell-Cv6UA-W_.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { t as eProcess } from "./sequential-eprocess-D1jKoihe.js";
-import { at as pairHoldout, tt as aggregatePairedHoldout } from "./campaign-evidence-B8oF9xQ6.js";
+import { at as pairHoldout, tt as aggregatePairedHoldout } from "./campaign-evidence-PYPUftRy.js";
 //#region src/campaign/gates/sequential.ts
 /**
 * Anytime-valid sequential promotion gate — an e-process (betting
@@ -339,4 +339,4 @@ function sequentialDecide(options = {}) {
 //#endregion
 export { sequentialPairedGate as n, sequentialDecide as t };
 
-//# sourceMappingURL=sequential-DAsyV2T9.js.map
+//# sourceMappingURL=sequential-D_xHrBZe.js.map