@tangle-network/agent-eval 0.133.1 → 0.133.3

package/dist/{paired-arms-6XItKzd1.js.map → paired-arms-CA_8pN01.js.map}

@@ -1 +1 @@
-{"version":3,"file":"paired-arms-6XItKzd1.js","names":[],"sources":["../src/paired-arms.ts"],"sourcesContent":["/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuGA,SAAgB,SAAS,MAA+B,MAAuC;CAC7F,MAAM,EAAE,aAAa,iBAAiB;CACtC,IAAI,gBAAgB,cAClB,MAAM,IAAI,gBACR,oDAAoD,YAAY,wCAClE;CAIF,MAAM,wBAAQ,IAAI,IAAyC;CAC3D,MAAM,2BAAW,IAAI,IAAY;CACjC,KAAK,MAAM,OAAO,MAAM;EACtB,SAAS,IAAI,IAAI,GAAG;EACpB,IAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,cAAc;EACzD,MAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,qBAAK,IAAI,IAA4B;EACpE,MAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,KAAK,GAAG;EACd,MAAM,IAAI,IAAI,SAAS,KAAK;EAC5B,MAAM,IAAI,IAAI,KAAK,KAAK;CAC1B;CAEA,KAAK,MAAM,OAAO,CAAC,aAAa,YAAY,GAC1C,IAAI,CAAC,MAAM,IAAI,GAAG,GAEhB,MAAM,IAAI,gBAAgB,8BAA8B,IAAI,mBAD/C,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,KAAK,SACoC,EAAE;CAI1F,MAAM,gBAAgB,MAAM,IAAI,WAAW;CAC3C,MAAM,iBAAiB,MAAM,IAAI,YAAY;CAE7C,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,MAAM,QAAuB,CAAC;CAC9B,MAAM,mBAAmC,CAAC;CAC1C,MAAM,oBAAoC,CAAC;CAC3C,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;EAE1C,IAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;GAClC,IAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;IACpC,MAAM,WAAW,EAAE;IACnB,MAAM,YAAY,EAAE;IACpB,IAAI,SAAS,WAAW,KAAA,KAAa,UAAU,WAAW,KAAA,GAEtD;SAAA,SAAS,WAAW,KAAA,KACpB,UAAU,WAAW,KAAA,KACrB,SAAS,WAAW,UAAU,QAC9B;MACA,iBAAiB,KAAK,QAAQ;MAC9B,kBAAkB,KAAK,SAAS;MAChC;KACF;;IAEF,MAAM,KAAK;KAAE;KAAS,UAAU;KAAG;KAAU;IAAU,CAAC;GAC1D,OAAO;IACL,iBAAiB,KAAK,GAAG,CAAC;IAC1B,kBAAkB,KAAK,GAAG,CAAC;GAC7B;GACA;EACF;EAEA,MAAM,SAAS,cAAc,GAAG,SAAS,WAAW;EACpD,MAAM,SAAS,cAAc,GAAG,SAAS,YAAY;EACrD,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;EACxE,IAAI,WAAW;EACf,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,WAAW,OAAO,IAAI,MAAM;GAClC,MAAM,YAAY,OAAO,IAAI,MAAM;GACnC,IAAI,aAAa,KAAA,KAAa,cAAc,KAAA,GAC1C,MAAM,KAAK;IAAE;IAAS,UAAU;IAAY;IAAU;GAAU,CAAC;QAC5D,IAAI,aAAa,KAAA,GACtB,iBAAiB,KAAK,QAAQ;QACzB,IAAI,cAAc,KAAA,GACvB,kBAAkB,KAAK,SAAS;EAEpC;CACF;CAEA,OAAO;EAAE;EAAO;EAAkB;CAAkB;AACtD;;;AAIA,SAAS,cACP,OACA,SACA,KAC2B;CAC3B,MAAM,wBAAQ,IAAI,IAA0B;CAC5C,KAAK,MAAM,OAAO,OAAO;EACvB,IAAI,IAAI,WAAW,KAAA,GACjB,MAAM,IAAI,gBACR,sBAAsB,QAAQ,mDAAmD,IAAI,8KAGvF;EAEF,IAAI,MAAM,IAAI,IAAI,MAAM,GACtB,MAAM,IAAI,gBACR,+BAA+B,IAAI,OAAO,iBAAiB,QAAQ,YAAY,IAAI,iEAErF;EAEF,MAAM,IAAI,IAAI,QAAQ,GAAG;CAC3B;CACA,OAAO;AACT;;;;;;;;;;;;;;AAiEA,SAAgB,kBACd,MACA,MACsB;CACtB,MAAM,EAAE,OAAO,kBAAkB,sBAAsB,SAAS,MAAM,IAAI;CAE1E,IAAI,cAAwC;CAC5C,MAAM,eAAyB,CAAC;CAChC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,SAAS,SAAS,KAAA,KAAa,KAAK,UAAU,SAAS,KAAA,GAAW;EAC3E,aAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;EAC5C,cAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;CAChD;CACA,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,KAAK,QAAQ,cAAc,aAAa;EAC9C,cAAc;GACZ,KAAK,GAAG;GACR,KAAK,GAAG;GACR,SAAS;GACT,gBAAgB,qBAAqB,cAAc,aAAa;EAClE;CACF;CAaA,MAAM,gBAVJ,KAAK,eACL,CACE,GAAG,IAAI,IACL,MAAM,SAAS,MAAM,CACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC,GACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC,CAC1C,CAAC,CACH,CACF,CAAC,CAAC,KAAK,EAAA,CAE6C,KAAK,SAAS;EAClE,MAAM,SAAmB,CAAC;EAC1B,MAAM,QAAkB,CAAC;EACzB,IAAI,WAAW;EACf,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,IAAI,YAAY,KAAK,UAAU,IAAI;GACzC,MAAM,IAAI,YAAY,KAAK,WAAW,IAAI;GAC1C,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;IACtC;IACA;GACF;GACA,OAAO,KAAK,CAAC;GACb,MAAM,KAAK,CAAC;EACd;EACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;EAC9F,OAAO;GACL;GACA,GAAG,OAAO;GACV;GACA,aAAa,aAAa,UAAU;GACpC,WAAW,aAAa,QAAQ;GAChC;GACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;EACzE;CACF,CAAC;CAED,OAAO;EACL,QAAQ,MAAM;EACd,mBAAmB,iBAAiB;EACpC,oBAAoB,kBAAkB;EACtC;EACA;CACF;AACF;;;;;;;;AA2BA,SAAgB,eACd,cACA,eACsB;CACtB,MAAM,eAAe,iBAAiB,cAAc,UAAU;CAC9D,MAAM,gBAAgB,iBAAiB,eAAe,WAAW;CACjE,yBAAyB,cAAc,UAAU;CACjD,yBAAyB,eAAe,WAAW;CACnD,IAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GACxD,OAAO;EACL,OAAO,CAAC;EACR,kBAAkB,aAAa,KAAK,QAAQ,IAAI,GAAG;EACnD,mBAAmB,cAAc,KAAK,QAAQ,IAAI,GAAG;CACvD;CAGF,MAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;EAC3D,aAAa;EACb,cAAc;CAChB,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,KAAK,SAAS;GAChC,MAAM,WAAW,KAAK;GACtB,MAAM,YAAY,KAAK;GACvB,OAAO;IACL,SAAS,KAAK;IACd,QAAQ,SAAS;IACjB,UAAU,SAAS;IACnB,WAAW,UAAU;GACvB;EACF,CAAC;EACD,kBAAkB,OAAO,iBAAiB,KAAK,QAAS,IAAwB,GAAG;EACnF,mBAAmB,OAAO,kBAAkB,KAAK,QAAS,IAAwB,GAAG;CACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;CACpF,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,aAAa,IAAI,WAAW,KAAK;EACvC,IAAI,CAAC,YACH,MAAM,IAAI,gBACR,wBAAwB,IAAI,MAAM,+EACpC;EAEF,OAAO;GACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;GACtD,QAAQ,OAAO,IAAI,IAAI;GACvB;GACA;EACF;CACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;CACrF,MAAM,4BAAY,IAAI,IAA+B;CACrD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;EAC7C,MAAM,KAAK,GAAG;EACd,UAAU,IAAI,IAAI,SAAS,KAAK;CAClC;CACA,KAAK,MAAM,CAAC,SAAS,UAAU,WAC7B,IAAI,MAAM,SAAS,GAAG,cAAc,OAAO,SAAS,GAAG;AAE3D;AAEA,SAAS,YAAY,KAAmB,MAAkC;CACxE,MAAM,IAAI,IAAI,UAAU;CACxB,IAAI,MAAM,KAAA,GAAW,OAAO,KAAA;CAC5B,IAAI,CAAC,OAAO,SAAS,CAAC,GACpB,MAAM,IAAI,gBACR,mDAAmD,KAAK,gBAAgB,IAAI,QAAQ,UAAU,IAAI,IAAI,MAAM,GAC9G;CAEF,OAAO;AACT"}
+{"version":3,"file":"paired-arms-CA_8pN01.js","names":[],"sources":["../src/paired-arms.ts"],"sourcesContent":["/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuGA,SAAgB,SAAS,MAA+B,MAAuC;CAC7F,MAAM,EAAE,aAAa,iBAAiB;CACtC,IAAI,gBAAgB,cAClB,MAAM,IAAI,gBACR,oDAAoD,YAAY,wCAClE;CAIF,MAAM,wBAAQ,IAAI,IAAyC;CAC3D,MAAM,2BAAW,IAAI,IAAY;CACjC,KAAK,MAAM,OAAO,MAAM;EACtB,SAAS,IAAI,IAAI,GAAG;EACpB,IAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,cAAc;EACzD,MAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,qBAAK,IAAI,IAA4B;EACpE,MAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,KAAK,GAAG;EACd,MAAM,IAAI,IAAI,SAAS,KAAK;EAC5B,MAAM,IAAI,IAAI,KAAK,KAAK;CAC1B;CAEA,KAAK,MAAM,OAAO,CAAC,aAAa,YAAY,GAC1C,IAAI,CAAC,MAAM,IAAI,GAAG,GAEhB,MAAM,IAAI,gBAAgB,8BAA8B,IAAI,mBAD/C,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,KAAK,SACoC,EAAE;CAI1F,MAAM,gBAAgB,MAAM,IAAI,WAAW;CAC3C,MAAM,iBAAiB,MAAM,IAAI,YAAY;CAE7C,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,MAAM,QAAuB,CAAC;CAC9B,MAAM,mBAAmC,CAAC;CAC1C,MAAM,oBAAoC,CAAC;CAC3C,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;EAE1C,IAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;GAClC,IAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;IACpC,MAAM,WAAW,EAAE;IACnB,MAAM,YAAY,EAAE;IACpB,IAAI,SAAS,WAAW,KAAA,KAAa,UAAU,WAAW,KAAA,GAEtD;SAAA,SAAS,WAAW,KAAA,KACpB,UAAU,WAAW,KAAA,KACrB,SAAS,WAAW,UAAU,QAC9B;MACA,iBAAiB,KAAK,QAAQ;MAC9B,kBAAkB,KAAK,SAAS;MAChC;KACF;;IAEF,MAAM,KAAK;KAAE;KAAS,UAAU;KAAG;KAAU;IAAU,CAAC;GAC1D,OAAO;IACL,iBAAiB,KAAK,GAAG,CAAC;IAC1B,kBAAkB,KAAK,GAAG,CAAC;GAC7B;GACA;EACF;EAEA,MAAM,SAAS,cAAc,GAAG,SAAS,WAAW;EACpD,MAAM,SAAS,cAAc,GAAG,SAAS,YAAY;EACrD,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;EACxE,IAAI,WAAW;EACf,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,WAAW,OAAO,IAAI,MAAM;GAClC,MAAM,YAAY,OAAO,IAAI,MAAM;GACnC,IAAI,aAAa,KAAA,KAAa,cAAc,KAAA,GAC1C,MAAM,KAAK;IAAE;IAAS,UAAU;IAAY;IAAU;GAAU,CAAC;QAC5D,IAAI,aAAa,KAAA,GACtB,iBAAiB,KAAK,QAAQ;QACzB,IAAI,cAAc,KAAA,GACvB,kBAAkB,KAAK,SAAS;EAEpC;CACF;CAEA,OAAO;EAAE;EAAO;EAAkB;CAAkB;AACtD;;;AAIA,SAAS,cACP,OACA,SACA,KAC2B;CAC3B,MAAM,wBAAQ,IAAI,IAA0B;CAC5C,KAAK,MAAM,OAAO,OAAO;EACvB,IAAI,IAAI,WAAW,KAAA,GACjB,MAAM,IAAI,gBACR,sBAAsB,QAAQ,mDAAmD,IAAI,8KAGvF;EAEF,IAAI,MAAM,IAAI,IAAI,MAAM,GACtB,MAAM,IAAI,gBACR,+BAA+B,IAAI,OAAO,iBAAiB,QAAQ,YAAY,IAAI,iEAErF;EAEF,MAAM,IAAI,IAAI,QAAQ,GAAG;CAC3B;CACA,OAAO;AACT;;;;;;;;;;;;;;AAiEA,SAAgB,kBACd,MACA,MACsB;CACtB,MAAM,EAAE,OAAO,kBAAkB,sBAAsB,SAAS,MAAM,IAAI;CAE1E,IAAI,cAAwC;CAC5C,MAAM,eAAyB,CAAC;CAChC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,SAAS,SAAS,KAAA,KAAa,KAAK,UAAU,SAAS,KAAA,GAAW;EAC3E,aAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;EAC5C,cAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;CAChD;CACA,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,KAAK,QAAQ,cAAc,aAAa;EAC9C,cAAc;GACZ,KAAK,GAAG;GACR,KAAK,GAAG;GACR,SAAS;GACT,gBAAgB,qBAAqB,cAAc,aAAa;EAClE;CACF;CAaA,MAAM,gBAVJ,KAAK,eACL,CACE,GAAG,IAAI,IACL,MAAM,SAAS,MAAM,CACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC,GACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC,CAC1C,CAAC,CACH,CACF,CAAC,CAAC,KAAK,EAAA,CAE6C,KAAK,SAAS;EAClE,MAAM,SAAmB,CAAC;EAC1B,MAAM,QAAkB,CAAC;EACzB,IAAI,WAAW;EACf,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,IAAI,YAAY,KAAK,UAAU,IAAI;GACzC,MAAM,IAAI,YAAY,KAAK,WAAW,IAAI;GAC1C,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;IACtC;IACA;GACF;GACA,OAAO,KAAK,CAAC;GACb,MAAM,KAAK,CAAC;EACd;EACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;EAC9F,OAAO;GACL;GACA,GAAG,OAAO;GACV;GACA,aAAa,aAAa,UAAU;GACpC,WAAW,aAAa,QAAQ;GAChC;GACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;EACzE;CACF,CAAC;CAED,OAAO;EACL,QAAQ,MAAM;EACd,mBAAmB,iBAAiB;EACpC,oBAAoB,kBAAkB;EACtC;EACA;CACF;AACF;;;;;;;;AA2BA,SAAgB,eACd,cACA,eACsB;CACtB,MAAM,eAAe,iBAAiB,cAAc,UAAU;CAC9D,MAAM,gBAAgB,iBAAiB,eAAe,WAAW;CACjE,yBAAyB,cAAc,UAAU;CACjD,yBAAyB,eAAe,WAAW;CACnD,IAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GACxD,OAAO;EACL,OAAO,CAAC;EACR,kBAAkB,aAAa,KAAK,QAAQ,IAAI,GAAG;EACnD,mBAAmB,cAAc,KAAK,QAAQ,IAAI,GAAG;CACvD;CAGF,MAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;EAC3D,aAAa;EACb,cAAc;CAChB,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,KAAK,SAAS;GAChC,MAAM,WAAW,KAAK;GACtB,MAAM,YAAY,KAAK;GACvB,OAAO;IACL,SAAS,KAAK;IACd,QAAQ,SAAS;IACjB,UAAU,SAAS;IACnB,WAAW,UAAU;GACvB;EACF,CAAC;EACD,kBAAkB,OAAO,iBAAiB,KAAK,QAAS,IAAwB,GAAG;EACnF,mBAAmB,OAAO,kBAAkB,KAAK,QAAS,IAAwB,GAAG;CACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;CACpF,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,aAAa,IAAI,WAAW,KAAK;EACvC,IAAI,CAAC,YACH,MAAM,IAAI,gBACR,wBAAwB,IAAI,MAAM,+EACpC;EAEF,OAAO;GACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;GACtD,QAAQ,OAAO,IAAI,IAAI;GACvB;GACA;EACF;CACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;CACrF,MAAM,4BAAY,IAAI,IAA+B;CACrD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;EAC7C,MAAM,KAAK,GAAG;EACd,UAAU,IAAI,IAAI,SAAS,KAAK;CAClC;CACA,KAAK,MAAM,CAAC,SAAS,UAAU,WAC7B,IAAI,MAAM,SAAS,GAAG,cAAc,OAAO,SAAS,GAAG;AAE3D;AAEA,SAAS,YAAY,KAAmB,MAAkC;CACxE,MAAM,IAAI,IAAI,UAAU;CACxB,IAAI,MAAM,KAAA,GAAW,OAAO,KAAA;CAC5B,IAAI,CAAC,OAAO,SAAS,CAAC,GACpB,MAAM,IAAI,gBACR,mDAAmD,KAAK,gBAAgB,IAAI,QAAQ,UAAU,IAAI,IAAI,MAAM,GAC9G;CAEF,OAAO;AACT"}

package/dist/pipelines/index.d.ts

@@ -2,7 +2,7 @@ import { S as ToolSpan, f as Run, r as BudgetSpec } from "../schema-BtVldJ3T.js"
 import { a as RunFilter, s as TraceStore } from "../store-CT9YIIve.js";
 import { o as llmSpans } from "../query-CJ_DX8vl.js";
 import { n as FailureClusterReport, r as failureClusterView, t as FailureCluster } from "../failure-cluster-CqcvCcdR.js";
-import { l as TrajectoryStep, m as computeToolUseMetrics, n as BaselineReport, t as BaselineOptions } from "../baseline-hG3K85h4.js";
+import { d as TrajectoryStep, g as computeToolUseMetrics, n as BaselineReport, t as BaselineOptions } from "../baseline-D_fT6277.js";
 //#region src/pipelines/budget-breach.d.ts
 interface BudgetBreachFinding {
   runId: string;

package/dist/pipelines/index.js

@@ -1,9 +1,10 @@
 import { t as executionTrackByLane } from "../execution-tracks-CpgFPpS5.js";
-import { E as pearsonR, u as interRaterReliability } from "../statistics-DWM_AyLe.js";
+import { h as interRaterReliability, j as pearsonR } from "../statistics-RwRNu2__.js";
 import { s as isToolSpan } from "../schema-CRhEY1SO.js";
 import { t as buildTrajectory } from "../trajectory-D_7rLrvE.js";
+import { t as compareToBaseline } from "../baseline-BaPxoROc.js";
 import { i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, s as runFailureClass, t as aggregateLlm } from "../query-Di7eEQ79.js";
-import { a as DEFAULT_RULES, i as computeToolUseMetrics, o as classifyFailure, t as compareToBaseline } from "../baseline-DcX5hQDv.js";
+import { n as DEFAULT_RULES, r as classifyFailure, t as computeToolUseMetrics } from "../tool-use-metrics-DEGMKycK.js";
 //#region src/pipelines/budget-breach.ts
 async function budgetBreachView(store, options = {}) {
 	const runs = await store.listRuns({

package/dist/pipelines/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/budget-breach.ts","../../src/pipelines/failure-cluster.ts","../../src/pipelines/first-divergence.ts","../../src/pipelines/judge-agreement.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts","../../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EAEpD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO,QAAA,MADrB,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACL,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;ACvFA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,KAAK,QAAQ,cAAc;CACjC,MAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;CACtD,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,MAAM,QAAQ,EAAE,MAAM;EACtB,MAAM,QAAQ,EAAE,MAAM;EACtB,IAAI,CAAC,GAAG,OAAO,KAAK,GAClB,OAAO;GACL;GACA;GACA,sBAAsB;GACtB;GACA;GACA,QAAQ,mBAAmB,OAAO,KAAK;GACvC,iBAAiB;EACnB;CAEJ;CACA,IAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAC7B,OAAO;EAAE;EAAM;EAAM,sBAAsB;EAAM,iBAAiB;CAAO;CAG3E,MAAM,SADqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI,EAAA,CAC5C,MAAM,SAAS;CAGpC,MAAM,SACJ,WAAW,IACP,0CAA0C,MAAM,gCAChD,sBAAsB,MAAM,4BAA4B,SAAS;CACvE,OAAO;EACL;EACA;EACA,sBAAsB;EACtB,OAAO,EAAE,MAAM;EACf,OAAO,EAAE,MAAM;EACf;EACA,iBAAiB;CACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO;CACxC,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,QAAQ,OAAO,EAAE,KAAK,aAAa,EAAE,KAAK;CACxF,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,OAAO,OAAO,EAAE,KAAK,UAAU,EAAE,KAAK;CACnF,IAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS,SAC7C,OAAO,EAAE,KAAK,cAAc,EAAE,KAAK;CACrC,OAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO,QAAQ,EAAE,KAAK,KAAK,MAAM,EAAE,KAAK;CACzE,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UACjF,OAAO,QAAQ,EAAE,KAAK,SAAS,MAAM,EAAE,KAAK;CAE9C,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAC5E,OAAO,SAAS,EAAE,KAAK,MAAM,MAAM,EAAE,KAAK;CAE5C,OAAO,SAAS,EAAE,KAAK,KAAK,QAAQ,EAAE,KAAK,KAAK;AAClD;;;;;;;;;;;;AC9DA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;;ACrEA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAS5D,OAAO,kBAAkB,MARH,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;EACpD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH,GACkC,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBACH,OAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;GAElD,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;ACvEA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF;;;;;;;;;;;;;;AC/LA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAIrC,OAAO,KAAK,GAHA,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IACU,EAAE,IAAI,OAAO,IAAI;CAEpC,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/budget-breach.ts","../../src/pipelines/failure-cluster.ts","../../src/pipelines/first-divergence.ts","../../src/pipelines/judge-agreement.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts","../../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EAEpD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO,QAAA,MADrB,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACL,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;ACvFA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,KAAK,QAAQ,cAAc;CACjC,MAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;CACtD,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,MAAM,QAAQ,EAAE,MAAM;EACtB,MAAM,QAAQ,EAAE,MAAM;EACtB,IAAI,CAAC,GAAG,OAAO,KAAK,GAClB,OAAO;GACL;GACA;GACA,sBAAsB;GACtB;GACA;GACA,QAAQ,mBAAmB,OAAO,KAAK;GACvC,iBAAiB;EACnB;CAEJ;CACA,IAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAC7B,OAAO;EAAE;EAAM;EAAM,sBAAsB;EAAM,iBAAiB;CAAO;CAG3E,MAAM,SADqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI,EAAA,CAC5C,MAAM,SAAS;CAGpC,MAAM,SACJ,WAAW,IACP,0CAA0C,MAAM,gCAChD,sBAAsB,MAAM,4BAA4B,SAAS;CACvE,OAAO;EACL;EACA;EACA,sBAAsB;EACtB,OAAO,EAAE,MAAM;EACf,OAAO,EAAE,MAAM;EACf;EACA,iBAAiB;CACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO;CACxC,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,QAAQ,OAAO,EAAE,KAAK,aAAa,EAAE,KAAK;CACxF,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,OAAO,OAAO,EAAE,KAAK,UAAU,EAAE,KAAK;CACnF,IAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS,SAC7C,OAAO,EAAE,KAAK,cAAc,EAAE,KAAK;CACrC,OAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO,QAAQ,EAAE,KAAK,KAAK,MAAM,EAAE,KAAK;CACzE,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UACjF,OAAO,QAAQ,EAAE,KAAK,SAAS,MAAM,EAAE,KAAK;CAE9C,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAC5E,OAAO,SAAS,EAAE,KAAK,MAAM,MAAM,EAAE,KAAK;CAE5C,OAAO,SAAS,EAAE,KAAK,KAAK,QAAQ,EAAE,KAAK,KAAK;AAClD;;;;;;;;;;;;AC9DA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;;ACrEA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAS5D,OAAO,kBAAkB,MARH,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;EACpD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH,GACkC,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBACH,OAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;GAElD,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;ACvEA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF;;;;;;;;;;;;;;AC/LA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAIrC,OAAO,KAAK,GAHA,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IACU,EAAE,IAAI,OAAO,IAAI;CAEpC,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}

package/dist/{release-report-wuilQkvK.js → release-report-BVZBmRZp.js}

@@ -2,7 +2,7 @@ import { c as VerificationError, s as ValidationError } from "./errors-8YnH8WlF.
 import { r as observedSplitScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
 import { t as FAILURE_CLASSES } from "./schema-CRhEY1SO.js";
 import { s as validateRunRecord } from "./run-record-BIwU2wdV.js";
-import { a as summaryTable } from "./summary-report-Ci17nIdU.js";
+import { a as summaryTable } from "./summary-report-BxtossFi.js";
 //#region src/release-confidence.ts
 const DEFAULT_THRESHOLDS = {
 	requireCorpus: true,
@@ -600,4 +600,4 @@ function duration(value) {
 //#endregion
 export { evaluateReleaseConfidence as a, assertReleaseConfidence as i, bootstrapCi as n, judgeReplayGate as r, renderReleaseReport as t };
 
-//# sourceMappingURL=release-report-wuilQkvK.js.map
+//# sourceMappingURL=release-report-BVZBmRZp.js.map