@tangle-network/agent-eval 1.0.0 → 1.0.2

package/dist/pipelines/index.d.ts

@@ -1,7 +1,7 @@
 import { f as Run } from "../schema-CkIvnmLU.js";
 import { a as RunFilter, s as TraceStore } from "../store-CMs5wGXP.js";
 import { n as FailureClusterReport, r as failureClusterView, t as FailureCluster } from "../failure-cluster-DOoa7CBg.js";
-import { t as DiagnosisSpan } from "../spans-vZoW__3q.js";
+import { t as DiagnosisSpan } from "../spans-CGjHouKM.js";
 import { _ as computeToolUseMetrics, a as JudgeAgreementReport, c as BudgetBreachFinding, f as TrajectoryStep, i as toolWasteView, l as BudgetBreachReport, n as ToolWasteOptions, o as JudgePair, r as ToolWasteReport, s as judgeAgreementView, t as ToolWasteFinding, u as budgetBreachView, v as BaselineOptions } from "../tool-waste-BFG1vwTV.js";
 /** One step of a run, as the diff reads it. */
 interface DiffStep {

package/dist/pipelines/index.js

@@ -1,7 +1,7 @@
 import { i as isToolSpan } from "../schema-D-xu3MU_.js";
 import { r as hasCapturedToolArgs, t as argHash } from "../query-CGXzO9d4.js";
 import { t as executionTrackByLane } from "../execution-tracks-DaDXYwAq.js";
-import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "../tool-waste-CLJ46AbT.js";
+import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "../tool-waste-CRJqOASv.js";
 // src/pipelines/first-divergence.ts
 function diffSteps(a, b) {
 	const partnerOfA = /* @__PURE__ */ new Map();

package/dist/{power-and-mde-CHvyDO_H.js → power-and-mde-CRxepy8y.js}

@@ -1,5 +1,5 @@
-import { o as normalCdf } from "./paired-arms-D5H_58e8.js";
-import { s as zQuantile } from "./internal-v6XfOeJW.js";
+import { o as normalCdf } from "./paired-arms-C0Vd9bGG.js";
+import { u as zQuantile } from "./mean-DDFhslBt.js";
 // src/statistics/power-and-mde.ts
 /**
 * Required N per arm for a two-sample comparison at target effect size,
@@ -96,4 +96,4 @@ function mcnemarPower(opts) {
 }
 export { requiredSampleSize as a, requiredPairedSampleSize as i, mcnemarRequiredN as n, pairedMde as r, mcnemarPower as t };
 
-//# sourceMappingURL=power-and-mde-CHvyDO_H.js.map
+//# sourceMappingURL=power-and-mde-CRxepy8y.js.map

package/dist/{power-and-mde-CHvyDO_H.js.map → power-and-mde-CRxepy8y.js.map}

@@ -1 +1 @@
-{"version":3,"file":"power-and-mde-CHvyDO_H.js","names":[],"sources":["../src/statistics/power-and-mde.ts"],"sourcesContent":["import { normalCdf } from '../math/normal'\nimport { zQuantile } from './internal'\n\n// ── Power analysis + minimum detectable effect ───────────────────────\n\n/**\n * Required N per arm for a two-sample comparison at target effect size,\n * alpha, and power. Normal-approximation formula:\n *   n = 2 * ( (z_{1-α/2} + z_{1-β}) / d )^2\n * where d is Cohen's d. Returns Infinity for effect ≤ 0.\n */\nexport function requiredSampleSize(opts: {\n  effect: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const effect = opts.effect\n  if (!Number.isFinite(effect) || effect <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  const n = 2 * ((zAlpha + zBeta) / effect) ** 2\n  return Math.ceil(n)\n}\n\n/**\n * Required number of paired observations for a target Cohen's dz.\n * Unlike the independent-groups formula, this has no two-arm factor of two.\n *\n * Normal quantiles with no t correction, so treat the result as a LOWER bound:\n * it returns 32 where the exact t-based answer is 34 at dz = 0.5, and 13 where\n * it is 15 at dz = 0.8 — a 6–13 % shortfall precisely in the range a caller\n * consults to decide whether 3–10 repetitions suffice.\n */\nexport function requiredPairedSampleSize(opts: {\n  effect: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const effect = opts.effect\n  if (!Number.isFinite(effect) || effect <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  return Math.ceil(((zAlpha + zBeta) / effect) ** 2)\n}\n\n/**\n * Minimum detectable paired effect (standardised units) for a target paired\n * sample size: d_min = (z_{1-α/2} + z_β) / sqrt(n_paired). Multiply by\n * sd(deltas) for score units; treat as a lower bound — Wilcoxon and bootstrap\n * have asymptotic relative efficiency below 1 vs the t-test on heavy tails.\n */\nexport function pairedMde(opts: {\n  nPaired: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  if (!Number.isFinite(opts.nPaired) || opts.nPaired <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  return (zAlpha + zBeta) / Math.sqrt(opts.nPaired)\n}\n\n/**\n * Number of paired observations needed for a McNemar test to reach a target\n * power — the pre-registration companion to {@link mcnemar}. Parametrised by the\n * expected discordant-cell probabilities `p10` (P[treatment wins on a pair]) and\n * `p01` (P[control wins]); concordant pairs carry no information, so the count\n * is driven entirely by the discordant rate. Lachin's (1992) asymptotic normal\n * approximation: with discordant rate `pDisc = p10 + p01` and marginal effect\n * `δ = p10 − p01`,\n *   n = ( z_{1-α/2}·√pDisc + z_{1-β}·√(pDisc − δ²) )² / δ².\n * Returns Infinity when there is no effect (p10 === p01). Asymptotic — at the\n * tiny discordant counts where the exact {@link mcnemar} differs from the normal\n * approximation, treat the result as a lower bound and prefer the discordant-pair\n * floor.\n */\nexport function mcnemarRequiredN(opts: {\n  p10: number\n  p01: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const { p10, p01 } = opts\n  if (p10 < 0 || p01 < 0 || p10 + p01 > 1) {\n    throw new Error(`mcnemarRequiredN: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`)\n  }\n  const delta = p10 - p01\n  if (delta === 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const pDisc = p10 + p01\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  const n =\n    (zAlpha * Math.sqrt(pDisc) + zBeta * Math.sqrt(Math.max(0, pDisc - delta * delta))) ** 2 /\n    (delta * delta)\n  return Math.ceil(n)\n}\n\n/**\n * Power of a McNemar test at a given number of paired observations, the inverse\n * of {@link mcnemarRequiredN} (same Lachin asymptotic model, same parameters).\n * Returns a value in [0, 1]; equals `alpha` when there is no effect.\n */\nexport function mcnemarPower(opts: {\n  p10: number\n  p01: number\n  nPairs: number\n  alpha?: number\n  twoSided?: boolean\n}): number {\n  const { p10, p01, nPairs } = opts\n  if (p10 < 0 || p01 < 0 || p10 + p01 > 1) {\n    throw new Error(`mcnemarPower: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`)\n  }\n  const alpha = opts.alpha ?? 0.05\n  const twoSided = opts.twoSided ?? true\n  const delta = p10 - p01\n  if (delta === 0 || nPairs <= 0) return alpha\n  const pDisc = p10 + p01\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const denom = Math.sqrt(Math.max(1e-12, pDisc - delta * delta))\n  const zBeta = (Math.sqrt(nPairs) * Math.abs(delta) - zAlpha * Math.sqrt(pDisc)) / denom\n  return Math.min(1, Math.max(0, normalCdf(zBeta)))\n}\n"],"mappings":";;;;;;;;;AAWA,SAAgB,mBAAmB,MAKxB;CACT,MAAM,SAAS,KAAK;CACpB,IAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG,OAAO;CACpD,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAGlC,MAAM,IAAI,MAFK,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAEnC,IADP,UAAU,KACK,KAAK,WAAW;CAC7C,OAAO,KAAK,KAAK,CAAC;AACpB;;;;;;;;;;AAWA,SAAgB,yBAAyB,MAK9B;CACT,MAAM,SAAS,KAAK;CACpB,IAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG,OAAO;CACpD,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,UAAU,KAAK;CAC7B,OAAO,KAAK,OAAO,SAAS,SAAS,WAAW,CAAC;AACnD;;;;;;;AAQA,SAAgB,UAAU,MAKf;CACT,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,WAAW,GAAG,OAAO;CAChE,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAGlC,QAFe,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAE3C,IADC,UAAU,KACH,KAAK,KAAK,KAAK,KAAK,OAAO;AAClD;;;;;;;;;;;;;;;AAgBA,SAAgB,iBAAiB,MAMtB;CACT,MAAM,EAAE,KAAK,QAAQ;CACrB,IAAI,MAAM,KAAK,MAAM,KAAK,MAAM,MAAM,GACpC,MAAM,IAAI,MAAM,8DAA8D,IAAI,IAAI,IAAI,EAAE;CAE9F,MAAM,QAAQ,MAAM;CACpB,IAAI,UAAU,GAAG,OAAO;CACxB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,MAAM;CACpB,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,UAAU,KAAK;CAC7B,MAAM,KACH,SAAS,KAAK,KAAK,KAAK,IAAI,QAAQ,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,QAAQ,KAAK,CAAC,MAAM,KACtF,QAAQ;CACX,OAAO,KAAK,KAAK,CAAC;AACpB;;;;;;AAOA,SAAgB,aAAa,MAMlB;CACT,MAAM,EAAE,KAAK,KAAK,WAAW;CAC7B,IAAI,MAAM,KAAK,MAAM,KAAK,MAAM,MAAM,GACpC,MAAM,IAAI,MAAM,0DAA0D,IAAI,IAAI,IAAI,EAAE;CAE1F,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,MAAM;CACpB,IAAI,UAAU,KAAK,UAAU,GAAG,OAAO;CACvC,MAAM,QAAQ,MAAM;CACpB,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,OAAO,QAAQ,QAAQ,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,KAAK,MAAM,IAAI,KAAK,IAAI,KAAK,IAAI,SAAS,KAAK,KAAK,KAAK,KAAK;CAClF,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,UAAU,KAAK,CAAC,CAAC;AAClD"}
+{"version":3,"file":"power-and-mde-CRxepy8y.js","names":[],"sources":["../src/statistics/power-and-mde.ts"],"sourcesContent":["import { normalCdf } from '../math/normal'\nimport { zQuantile } from './internal'\n\n// ── Power analysis + minimum detectable effect ───────────────────────\n\n/**\n * Required N per arm for a two-sample comparison at target effect size,\n * alpha, and power. Normal-approximation formula:\n *   n = 2 * ( (z_{1-α/2} + z_{1-β}) / d )^2\n * where d is Cohen's d. Returns Infinity for effect ≤ 0.\n */\nexport function requiredSampleSize(opts: {\n  effect: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const effect = opts.effect\n  if (!Number.isFinite(effect) || effect <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  const n = 2 * ((zAlpha + zBeta) / effect) ** 2\n  return Math.ceil(n)\n}\n\n/**\n * Required number of paired observations for a target Cohen's dz.\n * Unlike the independent-groups formula, this has no two-arm factor of two.\n *\n * Normal quantiles with no t correction, so treat the result as a LOWER bound:\n * it returns 32 where the exact t-based answer is 34 at dz = 0.5, and 13 where\n * it is 15 at dz = 0.8 — a 6–13 % shortfall precisely in the range a caller\n * consults to decide whether 3–10 repetitions suffice.\n */\nexport function requiredPairedSampleSize(opts: {\n  effect: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const effect = opts.effect\n  if (!Number.isFinite(effect) || effect <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  return Math.ceil(((zAlpha + zBeta) / effect) ** 2)\n}\n\n/**\n * Minimum detectable paired effect (standardised units) for a target paired\n * sample size: d_min = (z_{1-α/2} + z_β) / sqrt(n_paired). Multiply by\n * sd(deltas) for score units; treat as a lower bound — Wilcoxon and bootstrap\n * have asymptotic relative efficiency below 1 vs the t-test on heavy tails.\n */\nexport function pairedMde(opts: {\n  nPaired: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  if (!Number.isFinite(opts.nPaired) || opts.nPaired <= 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  return (zAlpha + zBeta) / Math.sqrt(opts.nPaired)\n}\n\n/**\n * Number of paired observations needed for a McNemar test to reach a target\n * power — the pre-registration companion to {@link mcnemar}. Parametrised by the\n * expected discordant-cell probabilities `p10` (P[treatment wins on a pair]) and\n * `p01` (P[control wins]); concordant pairs carry no information, so the count\n * is driven entirely by the discordant rate. Lachin's (1992) asymptotic normal\n * approximation: with discordant rate `pDisc = p10 + p01` and marginal effect\n * `δ = p10 − p01`,\n *   n = ( z_{1-α/2}·√pDisc + z_{1-β}·√(pDisc − δ²) )² / δ².\n * Returns Infinity when there is no effect (p10 === p01). Asymptotic — at the\n * tiny discordant counts where the exact {@link mcnemar} differs from the normal\n * approximation, treat the result as a lower bound and prefer the discordant-pair\n * floor.\n */\nexport function mcnemarRequiredN(opts: {\n  p10: number\n  p01: number\n  alpha?: number\n  power?: number\n  twoSided?: boolean\n}): number {\n  const { p10, p01 } = opts\n  if (p10 < 0 || p01 < 0 || p10 + p01 > 1) {\n    throw new Error(`mcnemarRequiredN: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`)\n  }\n  const delta = p10 - p01\n  if (delta === 0) return Infinity\n  const alpha = opts.alpha ?? 0.05\n  const power = opts.power ?? 0.8\n  const twoSided = opts.twoSided ?? true\n  const pDisc = p10 + p01\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const zBeta = zQuantile(power)\n  const n =\n    (zAlpha * Math.sqrt(pDisc) + zBeta * Math.sqrt(Math.max(0, pDisc - delta * delta))) ** 2 /\n    (delta * delta)\n  return Math.ceil(n)\n}\n\n/**\n * Power of a McNemar test at a given number of paired observations, the inverse\n * of {@link mcnemarRequiredN} (same Lachin asymptotic model, same parameters).\n * Returns a value in [0, 1]; equals `alpha` when there is no effect.\n */\nexport function mcnemarPower(opts: {\n  p10: number\n  p01: number\n  nPairs: number\n  alpha?: number\n  twoSided?: boolean\n}): number {\n  const { p10, p01, nPairs } = opts\n  if (p10 < 0 || p01 < 0 || p10 + p01 > 1) {\n    throw new Error(`mcnemarPower: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`)\n  }\n  const alpha = opts.alpha ?? 0.05\n  const twoSided = opts.twoSided ?? true\n  const delta = p10 - p01\n  if (delta === 0 || nPairs <= 0) return alpha\n  const pDisc = p10 + p01\n  const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha)\n  const denom = Math.sqrt(Math.max(1e-12, pDisc - delta * delta))\n  const zBeta = (Math.sqrt(nPairs) * Math.abs(delta) - zAlpha * Math.sqrt(pDisc)) / denom\n  return Math.min(1, Math.max(0, normalCdf(zBeta)))\n}\n"],"mappings":";;;;;;;;;AAWA,SAAgB,mBAAmB,MAKxB;CACT,MAAM,SAAS,KAAK;CACpB,IAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG,OAAO;CACpD,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAGlC,MAAM,IAAI,MAFK,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAEnC,IADP,UAAU,KACK,KAAK,WAAW;CAC7C,OAAO,KAAK,KAAK,CAAC;AACpB;;;;;;;;;;AAWA,SAAgB,yBAAyB,MAK9B;CACT,MAAM,SAAS,KAAK;CACpB,IAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG,OAAO;CACpD,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,UAAU,KAAK;CAC7B,OAAO,KAAK,OAAO,SAAS,SAAS,WAAW,CAAC;AACnD;;;;;;;AAQA,SAAgB,UAAU,MAKf;CACT,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,WAAW,GAAG,OAAO;CAChE,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAGlC,QAFe,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAE3C,IADC,UAAU,KACH,KAAK,KAAK,KAAK,KAAK,OAAO;AAClD;;;;;;;;;;;;;;;AAgBA,SAAgB,iBAAiB,MAMtB;CACT,MAAM,EAAE,KAAK,QAAQ;CACrB,IAAI,MAAM,KAAK,MAAM,KAAK,MAAM,MAAM,GACpC,MAAM,IAAI,MAAM,8DAA8D,IAAI,IAAI,IAAI,EAAE;CAE9F,MAAM,QAAQ,MAAM;CACpB,IAAI,UAAU,GAAG,OAAO;CACxB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,MAAM;CACpB,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,UAAU,KAAK;CAC7B,MAAM,KACH,SAAS,KAAK,KAAK,KAAK,IAAI,QAAQ,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,QAAQ,KAAK,CAAC,MAAM,KACtF,QAAQ;CACX,OAAO,KAAK,KAAK,CAAC;AACpB;;;;;;AAOA,SAAgB,aAAa,MAMlB;CACT,MAAM,EAAE,KAAK,KAAK,WAAW;CAC7B,IAAI,MAAM,KAAK,MAAM,KAAK,MAAM,MAAM,GACpC,MAAM,IAAI,MAAM,0DAA0D,IAAI,IAAI,IAAI,EAAE;CAE1F,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,QAAQ,MAAM;CACpB,IAAI,UAAU,KAAK,UAAU,GAAG,OAAO;CACvC,MAAM,QAAQ,MAAM;CACpB,MAAM,SAAS,UAAU,WAAW,IAAI,QAAQ,IAAI,IAAI,KAAK;CAC7D,MAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,OAAO,QAAQ,QAAQ,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,KAAK,MAAM,IAAI,KAAK,IAAI,KAAK,IAAI,SAAS,KAAK,KAAK,KAAK,KAAK;CAClF,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,UAAU,KAAK,CAAC,CAAC;AAClD"}

package/dist/{promotion-policy-D7B4_jEZ.js → promotion-policy-f79oklG_.js}

@@ -1,6 +1,6 @@
-import { r as pairedBootstrap } from "./paired-tests-C2zUqgUW.js";
-import { t as decidePairedPromotion } from "./paired-promotion-decision-BiVumgvM.js";
-import { o as pairHoldout, r as detectScale } from "./statistical-heldout-C9hz_AF9.js";
+import { r as pairedBootstrap } from "./paired-tests-kASEKpc-.js";
+import { t as decidePairedPromotion } from "./paired-promotion-decision-5Pl_qJLX.js";
+import { o as pairHoldout, r as detectScale } from "./statistical-heldout-DYF5f10K.js";
 // src/campaign/gates/promotion-policy.ts
 /**
 * Build paired evidence for each objective and apply a promotion policy.
@@ -170,4 +170,4 @@ function paretoSignificanceGate(options) {
 }
 export { paretoPolicy as n, paretoSignificanceGate as r, buildEvidenceVector as t };
 
-//# sourceMappingURL=promotion-policy-D7B4_jEZ.js.map
+//# sourceMappingURL=promotion-policy-f79oklG_.js.map

package/dist/{promotion-policy-D7B4_jEZ.js.map → promotion-policy-f79oklG_.js.map}

@@ -1 +1 @@
-{"version":3,"file":"promotion-policy-D7B4_jEZ.js","names":[],"sources":["../src/campaign/gates/promotion-policy.ts"],"sourcesContent":["/**\n * Build paired evidence for each objective and apply a promotion policy.\n * The default policy requires at least one gain and every regression floor to\n * clear. A floor can fail because a larger loss remains plausible; that does\n * not demonstrate an observed regression. Missing evidence stays unresolved.\n * Confidence intervals apply to each axis, without a multiplicity adjustment.\n *\n * Cost and latency remain aggregate constraints because GateContext does not\n * supply paired observations for them. Compose a budget gate when needed.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n} from '../../paired-promotion-decision'\nimport type { Direction } from '../../pareto'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** Declared binary support {0, binaryScale}, including zero-only observations.\n   *  Must be finite and positive; paired cell scores must be 0 or this scale.\n   *  Uses the risk-difference mean and rejects the 'median' statistic. */\n  binaryScale?: number\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance, or when the exact small-sample test proves\n   *  a drop past it. Defaults to 0.05 times the declared binary scale, or\n   *  auto-scales off observed magnitudes (0.05 on [0,1], 5 on 0-100). */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the shared paired decision rule.\n *  'regressed' includes uncertainty that prevents clearing the regression floor. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs' | 'indeterminate'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis.\n   *\n   *  DIAGNOSTIC on a pass/fail axis: there the verdict is decided on Tango's\n   *  score interval instead, because a percentile bootstrap over a three-atom\n   *  delta lattice is not a valid interval at the nonzero margin `floorTolerance`\n   *  and `gainThreshold` create. `ci` carries the interval that decided. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low`/`.high` bracket. `'mean'` unless the\n   *  caller asked for the median — on a pass/fail axis the median and its whole\n   *  CI are pinned at 0 by tie domination and can see neither a gain nor a\n   *  regression. `bootstrap.median` still carries the median point estimate. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval the axis verdict was actually decided on, good-direction and\n   *  in the axis's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail axis; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width or non-finite bounds. It cannot establish a gain or\n   *  clear a regression floor, regardless of the point estimate. */\n  indeterminate: boolean\n  /** Paired observations contributing to this axis. */\n  n: number\n  minimumRequired: number\n  decisionMethod: PairedDecisionMethod\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. The exact small-sample test may require more\n   *  observations at the selected confidence. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n  /** Paired statistic every axis CI is computed on. Default `'mean'` — see\n   *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const n = paired.before.length\n    const binaryScale = obj.binaryScale\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * (binaryScale ?? detectScale([...paired.before, ...paired.after]))\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Axes are decided on the MEAN paired delta — which for a pass/fail axis is\n    // exactly the change in success rate. The median is structurally blind on\n    // the shapes eval data lands in: with most pairs tied its bootstrap CI\n    // collapses to [0,0] and the axis reads 'flat', hiding real gains AND real\n    // regressions — pass/fail axes on ANY encoding ({0,1} and the 0-100 one\n    // `detectScale` exists for), and low-cardinality axes even below half ties.\n    // Orthogonal to `pairedDeltaTest`'s own small-sample switch: that picks the\n    // TEST (bootstrap CI at n ≥ 20, exact sign test below), this picks the\n    // ESTIMATOR the test is applied to. Both are needed — an exact sign test on\n    // a tie-pinned median is still blind.\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    // Both burdens of proof route through the ONE shared rule\n    // (`decidePairedPromotion`), so a pass/fail axis is judged on Tango's score\n    // interval — the only paired-binary construction valid at the nonzero\n    // margins `gainThreshold` / `floorTolerance` create — and a zero-width\n    // interval cannot buy a verdict in either direction.\n    const improvement = decidePairedPromotion(before, after, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: gainThreshold,\n      minPairs: opts.minProductiveRuns,\n      binaryScale,\n    })\n    const regression = decidePairedPromotion(after, before, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: floorTolerance,\n      minPairs: opts.minProductiveRuns,\n      binaryScale,\n    })\n    const bootstrap =\n      improvement.bootstrap ??\n      pairedBootstrap(before, after, {\n        confidence,\n        resamples,\n        statistic: bootstrapStatistic,\n        seed,\n      })\n    // A floor breach fires on EITHER burden of proof, because they cover\n    // different failures and the floor is the anti-Goodhart guard:\n    //   - `improvement.low < -floorTolerance` — the CREDIBLE WORST CASE exceeds\n    //     the tolerance. This is the contract `AxisEvidence.floorTolerance` and\n    //     `paretoPolicy`'s own reason string state, and it is the conservative\n    //     posture a safety axis needs: block unless the data can rule the\n    //     breach out, rather than waiting for the breach to be proven. Read off\n    //     the DECIDING interval, so a pass/fail axis is not screened by a\n    //     bootstrap that is pinned wherever ties dominate.\n    //   - `regression.promote` — a PROVEN drop past the tolerance. Adds the\n    //     small-sample path, where the decision is an exact sign test because\n    //     the bootstrap interval is descriptive only.\n    // A tied binary axis still has uncertainty about unseen discordant pairs.\n    // Its diagnostic bootstrap collapses to zero, so only the deciding score\n    // interval can establish that a regression stays within the declared floor.\n    const floorBreached = improvement.low < -floorTolerance || regression.promote\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict = !improvement.sufficient\n      ? 'few_runs'\n      : improvement.indeterminate\n        ? 'indeterminate'\n        : floorBreached\n          ? 'regressed'\n          : improvement.promote\n            ? 'improved'\n            : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: improvement.low, high: improvement.high },\n      decisionStatistic: improvement.statistic,\n      mcnemar: improvement.mcnemar,\n      indeterminate: improvement.indeterminate,\n      n,\n      minimumRequired: improvement.minimumPairs,\n      decisionMethod: improvement.method,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * Require a supported gain and every configured regression floor to clear.\n * A failed floor holds the candidate, including when uncertainty permits a loss.\n * Missing or indeterminate evidence requires more work; no gain holds release.\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    status:\n      ax.verdict === 'regressed'\n        ? ('fail' as const)\n        : ax.verdict === 'few_runs' || ax.verdict === 'indeterminate'\n          ? ('not_evaluated' as const)\n          : ('pass' as const),\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.ci.low,\n      ciHigh: ax.ci.high,\n      decisionStatistic: ax.decisionStatistic,\n      decisionMethod: ax.decisionMethod,\n      mcnemar: ax.mcnemar,\n      indeterminate: ax.indeterminate,\n      bootstrapCiLow: ax.bootstrap.low,\n      bootstrapCiHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const insufficient = ev.axes.filter(\n    (a) => a.verdict === 'few_runs' || a.verdict === 'indeterminate',\n  )\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // A gain on another axis cannot excuse an unresolved regression floor.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' did not clear its regression floor -${a.floorTolerance}: good-direction CI [${a.ci.low.toFixed(3)}, ${a.ci.high.toFixed(3)}] (n=${a.n})`,\n      )\n    }\n  } else if (insufficient.length > 0) {\n    // An unresolved axis cannot establish either a gain or a safe floor.\n    decision = 'need_more_work'\n    for (const a of insufficient) {\n      reasons.push(\n        a.verdict === 'few_runs'\n          ? `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`\n          : `objective '${a.name}' has an indeterminate deciding CI [${a.ci.low}, ${a.ci.high}] — insufficient evidence to clear its regression floor (n=${a.n})`,\n      )\n    }\n  } else if (improved.length > 0) {\n    decision = 'ship'\n    reasons.push(\n      `Supported objective gain: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.ci.low > 0 ? a.ci.low.toFixed(3) : a.bootstrap.mean.toFixed(3)} (CI.low ${a.ci.low.toFixed(3)})`,\n        )\n        .join(', ')}; every regression floor cleared`,\n    )\n  } else {\n    // Every floor cleared, but no objective demonstrated a significant gain.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: no objective shows a significant gain; every regression floor cleared',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAkIA,SAAgB,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;CAChB,IAAI,WAAW,WAAW,GACxB,MAAM,IAAI,MAAM,oDAAoD;CAEtE,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,IAAI,uBAAuB,IAAI;CAChD,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;CAE1D,MAAM,OAAuB,CAAC;CAC9B,KAAK,MAAM,OAAO,YAAY;EAC5B,IAAI;EACJ,IAAI,IAAI,OAAO,SAAS,aACtB,UAAU,MAAM,EAAE;OACb;GACL,MAAM,MAAM,IAAI,OAAO;GACvB,UAAU,MAAM,EAAE,WAAW;EAC/B;EACA,MAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;EAIzE,MAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;EACrE,MAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;EACnE,MAAM,IAAI,OAAO,OAAO;EACxB,MAAM,cAAc,IAAI;EACxB,MAAM,iBACJ,IAAI,kBAAkB,OAAQ,eAAe,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC9F,MAAM,gBAAgB,IAAI,iBAAiB;EAW3C,MAAM,qBAAqB,KAAK,aAAA;EAMhC,MAAM,cAAc,sBAAsB,QAAQ,OAAO;GACvD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;GACf;EACF,CAAC;EACD,MAAM,aAAa,sBAAsB,OAAO,QAAQ;GACtD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;GACf;EACF,CAAC;EACD,MAAM,YACJ,YAAY,aACZ,gBAAgB,QAAQ,OAAO;GAC7B;GACA;GACA,WAAW;GACX;EACF,CAAC;EAgBH,MAAM,gBAAgB,YAAY,MAAM,CAAC,kBAAkB,WAAW;EAMtE,MAAM,UAAuB,CAAC,YAAY,aACtC,aACA,YAAY,gBACV,kBACA,gBACE,cACA,YAAY,UACV,aACA;EACV,KAAK,KAAK;GACR,MAAM,IAAI;GACV,QAAQ,IAAI;GACZ,WAAW,IAAI;GACf;GACA;GACA,IAAI;IAAE,KAAK,YAAY;IAAK,MAAM,YAAY;GAAK;GACnD,mBAAmB,YAAY;GAC/B,SAAS,YAAY;GACrB,eAAe,YAAY;GAC3B;GACA,iBAAiB,YAAY;GAC7B,gBAAgB,YAAY;GAC5B;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC;CAEnD,OAAO;EAAE;EAAM,MADF,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;EAC1B,MAAM;GAAE,WAAW,IAAI,KAAK;GAAW,UAAU,IAAI,KAAK;EAAS;CAAE;AAC5F;;;;;;AAOA,MAAa,gBAAiC,OAAO;CACnD,MAAM,oBAAoB,GAAG,KAAK,KAAK,QAAQ;EAC7C,MAAM,aAAa,GAAG;EACtB,QACE,GAAG,YAAY,cACV,SACD,GAAG,YAAY,cAAc,GAAG,YAAY,kBACzC,kBACA;EACT,QAAQ;GACN,WAAW,GAAG;GACd,QAAQ,GAAG;GACX,SAAS,GAAG;GACZ,GAAG,GAAG;GACN,aAAa,GAAG,UAAU;GAC1B,OAAO,GAAG,GAAG;GACb,QAAQ,GAAG,GAAG;GACd,mBAAmB,GAAG;GACtB,gBAAgB,GAAG;GACnB,SAAS,GAAG;GACZ,eAAe,GAAG;GAClB,gBAAgB,GAAG,UAAU;GAC7B,iBAAiB,GAAG,UAAU;GAC9B,YAAY,GAAG,UAAU;GACzB,eAAe,GAAG;GAClB,gBAAgB,GAAG;EACrB;CACF,EAAE;CAEF,MAAM,YAAY,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,WAAW;CACjE,MAAM,eAAe,GAAG,KAAK,QAC1B,MAAM,EAAE,YAAY,cAAc,EAAE,YAAY,eACnD;CACA,MAAM,WAAW,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAE/D,IAAI;CACJ,MAAM,UAAoB,CAAC;CAC3B,IAAI,UAAU,SAAS,GAAG;EAExB,WAAW;EACX,KAAK,MAAM,KAAK,WACd,QAAQ,KACN,cAAc,EAAE,KAAK,wCAAwC,EAAE,eAAe,uBAAuB,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,IAAI,EAAE,GAAG,KAAK,QAAQ,CAAC,EAAE,OAAO,EAAE,EAAE,EAC/J;CAEJ,OAAO,IAAI,aAAa,SAAS,GAAG;EAElC,WAAW;EACX,KAAK,MAAM,KAAK,cACd,QAAQ,KACN,EAAE,YAAY,aACV,cAAc,EAAE,KAAK,eAAe,EAAE,EAAE,8DACxC,cAAc,EAAE,KAAK,sCAAsC,EAAE,GAAG,IAAI,IAAI,EAAE,GAAG,KAAK,6DAA6D,EAAE,EAAE,EACzJ;CAEJ,OAAO,IAAI,SAAS,SAAS,GAAG;EAC9B,WAAW;EACX,QAAQ,KACN,6BAA6B,SAC1B,KACE,MACC,IAAI,EAAE,KAAK,KAAK,EAAE,GAAG,MAAM,IAAI,EAAE,GAAG,IAAI,QAAQ,CAAC,IAAI,EAAE,UAAU,KAAK,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,EACpH,CAAC,CACA,KAAK,IAAI,EAAE,iCAChB;CACF,OAAO;EAEL,WAAW;EACX,QAAQ,KACN,8FACF;CACF;CAIA,MAAM,YAAY,GAAG,KAAK,MAAM,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK;CAChF,OAAO;EAAE;EAAU;EAAS;EAAmB,OAAO,WAAW,UAAU;CAAO;AACpF;;;;;;AAiBA,SAAgB,uBACd,SAC4B;CAC5B,IAAI,QAAQ,WAAW,WAAW,GAChC,MAAM,IAAI,MAAM,uDAAuD;CAEzE,MAAM,SAAS,QAAQ,UAAU;CACjC,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB,MAAM,OAAO,KAA6D;GACxE,MAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;GAC/D,OAAO,OAAO,EAAE;EAClB;CACF;AACF"}
+{"version":3,"file":"promotion-policy-f79oklG_.js","names":[],"sources":["../src/campaign/gates/promotion-policy.ts"],"sourcesContent":["/**\n * Build paired evidence for each objective and apply a promotion policy.\n * The default policy requires at least one gain and every regression floor to\n * clear. A floor can fail because a larger loss remains plausible; that does\n * not demonstrate an observed regression. Missing evidence stays unresolved.\n * Confidence intervals apply to each axis, without a multiplicity adjustment.\n *\n * Cost and latency remain aggregate constraints because GateContext does not\n * supply paired observations for them. Compose a budget gate when needed.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n} from '../../paired-promotion-decision'\nimport type { Direction } from '../../pareto'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** Declared binary support {0, binaryScale}, including zero-only observations.\n   *  Must be finite and positive; paired cell scores must be 0 or this scale.\n   *  Uses the risk-difference mean and rejects the 'median' statistic. */\n  binaryScale?: number\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance, or when the exact small-sample test proves\n   *  a drop past it. Defaults to 0.05 times the declared binary scale, or\n   *  auto-scales off observed magnitudes (0.05 on [0,1], 5 on 0-100). */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the shared paired decision rule.\n *  'regressed' includes uncertainty that prevents clearing the regression floor. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs' | 'indeterminate'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis.\n   *\n   *  DIAGNOSTIC on a pass/fail axis: there the verdict is decided on Tango's\n   *  score interval instead, because a percentile bootstrap over a three-atom\n   *  delta lattice is not a valid interval at the nonzero margin `floorTolerance`\n   *  and `gainThreshold` create. `ci` carries the interval that decided. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low`/`.high` bracket. `'mean'` unless the\n   *  caller asked for the median — on a pass/fail axis the median and its whole\n   *  CI are pinned at 0 by tie domination and can see neither a gain nor a\n   *  regression. `bootstrap.median` still carries the median point estimate. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval the axis verdict was actually decided on, good-direction and\n   *  in the axis's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail axis; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width or non-finite bounds. It cannot establish a gain or\n   *  clear a regression floor, regardless of the point estimate. */\n  indeterminate: boolean\n  /** Paired observations contributing to this axis. */\n  n: number\n  minimumRequired: number\n  decisionMethod: PairedDecisionMethod\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. The exact small-sample test may require more\n   *  observations at the selected confidence. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n  /** Paired statistic every axis CI is computed on. Default `'mean'` — see\n   *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const n = paired.before.length\n    const binaryScale = obj.binaryScale\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * (binaryScale ?? detectScale([...paired.before, ...paired.after]))\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Axes are decided on the MEAN paired delta — which for a pass/fail axis is\n    // exactly the change in success rate. The median is structurally blind on\n    // the shapes eval data lands in: with most pairs tied its bootstrap CI\n    // collapses to [0,0] and the axis reads 'flat', hiding real gains AND real\n    // regressions — pass/fail axes on ANY encoding ({0,1} and the 0-100 one\n    // `detectScale` exists for), and low-cardinality axes even below half ties.\n    // Orthogonal to `pairedDeltaTest`'s own small-sample switch: that picks the\n    // TEST (bootstrap CI at n ≥ 20, exact sign test below), this picks the\n    // ESTIMATOR the test is applied to. Both are needed — an exact sign test on\n    // a tie-pinned median is still blind.\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    // Both burdens of proof route through the ONE shared rule\n    // (`decidePairedPromotion`), so a pass/fail axis is judged on Tango's score\n    // interval — the only paired-binary construction valid at the nonzero\n    // margins `gainThreshold` / `floorTolerance` create — and a zero-width\n    // interval cannot buy a verdict in either direction.\n    const improvement = decidePairedPromotion(before, after, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: gainThreshold,\n      minPairs: opts.minProductiveRuns,\n      binaryScale,\n    })\n    const regression = decidePairedPromotion(after, before, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: floorTolerance,\n      minPairs: opts.minProductiveRuns,\n      binaryScale,\n    })\n    const bootstrap =\n      improvement.bootstrap ??\n      pairedBootstrap(before, after, {\n        confidence,\n        resamples,\n        statistic: bootstrapStatistic,\n        seed,\n      })\n    // A floor breach fires on EITHER burden of proof, because they cover\n    // different failures and the floor is the anti-Goodhart guard:\n    //   - `improvement.low < -floorTolerance` — the CREDIBLE WORST CASE exceeds\n    //     the tolerance. This is the contract `AxisEvidence.floorTolerance` and\n    //     `paretoPolicy`'s own reason string state, and it is the conservative\n    //     posture a safety axis needs: block unless the data can rule the\n    //     breach out, rather than waiting for the breach to be proven. Read off\n    //     the DECIDING interval, so a pass/fail axis is not screened by a\n    //     bootstrap that is pinned wherever ties dominate.\n    //   - `regression.promote` — a PROVEN drop past the tolerance. Adds the\n    //     small-sample path, where the decision is an exact sign test because\n    //     the bootstrap interval is descriptive only.\n    // A tied binary axis still has uncertainty about unseen discordant pairs.\n    // Its diagnostic bootstrap collapses to zero, so only the deciding score\n    // interval can establish that a regression stays within the declared floor.\n    const floorBreached = improvement.low < -floorTolerance || regression.promote\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict = !improvement.sufficient\n      ? 'few_runs'\n      : improvement.indeterminate\n        ? 'indeterminate'\n        : floorBreached\n          ? 'regressed'\n          : improvement.promote\n            ? 'improved'\n            : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: improvement.low, high: improvement.high },\n      decisionStatistic: improvement.statistic,\n      mcnemar: improvement.mcnemar,\n      indeterminate: improvement.indeterminate,\n      n,\n      minimumRequired: improvement.minimumPairs,\n      decisionMethod: improvement.method,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * Require a supported gain and every configured regression floor to clear.\n * A failed floor holds the candidate, including when uncertainty permits a loss.\n * Missing or indeterminate evidence requires more work; no gain holds release.\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    status:\n      ax.verdict === 'regressed'\n        ? ('fail' as const)\n        : ax.verdict === 'few_runs' || ax.verdict === 'indeterminate'\n          ? ('not_evaluated' as const)\n          : ('pass' as const),\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.ci.low,\n      ciHigh: ax.ci.high,\n      decisionStatistic: ax.decisionStatistic,\n      decisionMethod: ax.decisionMethod,\n      mcnemar: ax.mcnemar,\n      indeterminate: ax.indeterminate,\n      bootstrapCiLow: ax.bootstrap.low,\n      bootstrapCiHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const insufficient = ev.axes.filter(\n    (a) => a.verdict === 'few_runs' || a.verdict === 'indeterminate',\n  )\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // A gain on another axis cannot excuse an unresolved regression floor.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' did not clear its regression floor -${a.floorTolerance}: good-direction CI [${a.ci.low.toFixed(3)}, ${a.ci.high.toFixed(3)}] (n=${a.n})`,\n      )\n    }\n  } else if (insufficient.length > 0) {\n    // An unresolved axis cannot establish either a gain or a safe floor.\n    decision = 'need_more_work'\n    for (const a of insufficient) {\n      reasons.push(\n        a.verdict === 'few_runs'\n          ? `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`\n          : `objective '${a.name}' has an indeterminate deciding CI [${a.ci.low}, ${a.ci.high}] — insufficient evidence to clear its regression floor (n=${a.n})`,\n      )\n    }\n  } else if (improved.length > 0) {\n    decision = 'ship'\n    reasons.push(\n      `Supported objective gain: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.ci.low > 0 ? a.ci.low.toFixed(3) : a.bootstrap.mean.toFixed(3)} (CI.low ${a.ci.low.toFixed(3)})`,\n        )\n        .join(', ')}; every regression floor cleared`,\n    )\n  } else {\n    // Every floor cleared, but no objective demonstrated a significant gain.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: no objective shows a significant gain; every regression floor cleared',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAkIA,SAAgB,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;CAChB,IAAI,WAAW,WAAW,GACxB,MAAM,IAAI,MAAM,oDAAoD;CAEtE,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,IAAI,uBAAuB,IAAI;CAChD,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;CAE1D,MAAM,OAAuB,CAAC;CAC9B,KAAK,MAAM,OAAO,YAAY;EAC5B,IAAI;EACJ,IAAI,IAAI,OAAO,SAAS,aACtB,UAAU,MAAM,EAAE;OACb;GACL,MAAM,MAAM,IAAI,OAAO;GACvB,UAAU,MAAM,EAAE,WAAW;EAC/B;EACA,MAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;EAIzE,MAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;EACrE,MAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;EACnE,MAAM,IAAI,OAAO,OAAO;EACxB,MAAM,cAAc,IAAI;EACxB,MAAM,iBACJ,IAAI,kBAAkB,OAAQ,eAAe,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC9F,MAAM,gBAAgB,IAAI,iBAAiB;EAW3C,MAAM,qBAAqB,KAAK,aAAA;EAMhC,MAAM,cAAc,sBAAsB,QAAQ,OAAO;GACvD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;GACf;EACF,CAAC;EACD,MAAM,aAAa,sBAAsB,OAAO,QAAQ;GACtD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;GACf;EACF,CAAC;EACD,MAAM,YACJ,YAAY,aACZ,gBAAgB,QAAQ,OAAO;GAC7B;GACA;GACA,WAAW;GACX;EACF,CAAC;EAgBH,MAAM,gBAAgB,YAAY,MAAM,CAAC,kBAAkB,WAAW;EAMtE,MAAM,UAAuB,CAAC,YAAY,aACtC,aACA,YAAY,gBACV,kBACA,gBACE,cACA,YAAY,UACV,aACA;EACV,KAAK,KAAK;GACR,MAAM,IAAI;GACV,QAAQ,IAAI;GACZ,WAAW,IAAI;GACf;GACA;GACA,IAAI;IAAE,KAAK,YAAY;IAAK,MAAM,YAAY;GAAK;GACnD,mBAAmB,YAAY;GAC/B,SAAS,YAAY;GACrB,eAAe,YAAY;GAC3B;GACA,iBAAiB,YAAY;GAC7B,gBAAgB,YAAY;GAC5B;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC;CAEnD,OAAO;EAAE;EAAM,MADF,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;EAC1B,MAAM;GAAE,WAAW,IAAI,KAAK;GAAW,UAAU,IAAI,KAAK;EAAS;CAAE;AAC5F;;;;;;AAOA,MAAa,gBAAiC,OAAO;CACnD,MAAM,oBAAoB,GAAG,KAAK,KAAK,QAAQ;EAC7C,MAAM,aAAa,GAAG;EACtB,QACE,GAAG,YAAY,cACV,SACD,GAAG,YAAY,cAAc,GAAG,YAAY,kBACzC,kBACA;EACT,QAAQ;GACN,WAAW,GAAG;GACd,QAAQ,GAAG;GACX,SAAS,GAAG;GACZ,GAAG,GAAG;GACN,aAAa,GAAG,UAAU;GAC1B,OAAO,GAAG,GAAG;GACb,QAAQ,GAAG,GAAG;GACd,mBAAmB,GAAG;GACtB,gBAAgB,GAAG;GACnB,SAAS,GAAG;GACZ,eAAe,GAAG;GAClB,gBAAgB,GAAG,UAAU;GAC7B,iBAAiB,GAAG,UAAU;GAC9B,YAAY,GAAG,UAAU;GACzB,eAAe,GAAG;GAClB,gBAAgB,GAAG;EACrB;CACF,EAAE;CAEF,MAAM,YAAY,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,WAAW;CACjE,MAAM,eAAe,GAAG,KAAK,QAC1B,MAAM,EAAE,YAAY,cAAc,EAAE,YAAY,eACnD;CACA,MAAM,WAAW,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAE/D,IAAI;CACJ,MAAM,UAAoB,CAAC;CAC3B,IAAI,UAAU,SAAS,GAAG;EAExB,WAAW;EACX,KAAK,MAAM,KAAK,WACd,QAAQ,KACN,cAAc,EAAE,KAAK,wCAAwC,EAAE,eAAe,uBAAuB,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,IAAI,EAAE,GAAG,KAAK,QAAQ,CAAC,EAAE,OAAO,EAAE,EAAE,EAC/J;CAEJ,OAAO,IAAI,aAAa,SAAS,GAAG;EAElC,WAAW;EACX,KAAK,MAAM,KAAK,cACd,QAAQ,KACN,EAAE,YAAY,aACV,cAAc,EAAE,KAAK,eAAe,EAAE,EAAE,8DACxC,cAAc,EAAE,KAAK,sCAAsC,EAAE,GAAG,IAAI,IAAI,EAAE,GAAG,KAAK,6DAA6D,EAAE,EAAE,EACzJ;CAEJ,OAAO,IAAI,SAAS,SAAS,GAAG;EAC9B,WAAW;EACX,QAAQ,KACN,6BAA6B,SAC1B,KACE,MACC,IAAI,EAAE,KAAK,KAAK,EAAE,GAAG,MAAM,IAAI,EAAE,GAAG,IAAI,QAAQ,CAAC,IAAI,EAAE,UAAU,KAAK,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,EACpH,CAAC,CACA,KAAK,IAAI,EAAE,iCAChB;CACF,OAAO;EAEL,WAAW;EACX,QAAQ,KACN,8FACF;CACF;CAIA,MAAM,YAAY,GAAG,KAAK,MAAM,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK;CAChF,OAAO;EAAE;EAAU;EAAS;EAAmB,OAAO,WAAW,UAAU;CAAO;AACpF;;;;;;AAiBA,SAAgB,uBACd,SAC4B;CAC5B,IAAI,QAAQ,WAAW,WAAW,GAChC,MAAM,IAAI,MAAM,uDAAuD;CAEzE,MAAM,SAAS,QAAQ,UAAU;CACjC,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB,MAAM,OAAO,KAA6D;GACxE,MAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;GAC/D,OAAO,OAAO,EAAE;EAClB;CACF;AACF"}

package/dist/{proposal-findings-WFQrxnfe.js → proposal-findings-duV5flPv.js}

@@ -1,4 +1,4 @@
-import { S as ProposalFindingSchema } from "./external-optimizer-subprocess-U616hV_l.js";
+import { S as ProposalFindingSchema } from "./external-optimizer-subprocess-wzMaA0oN.js";
 // src/analyst/proposal-findings.ts
 /** True when a finding names a source candidate generation may learn from. */
 function isProposalFinding(finding) {
@@ -22,4 +22,4 @@ function findingLabel(finding, index) {
 }
 export { assertProposalFindings as t };
 
-//# sourceMappingURL=proposal-findings-WFQrxnfe.js.map
+//# sourceMappingURL=proposal-findings-duV5flPv.js.map

package/dist/{proposal-findings-WFQrxnfe.js.map → proposal-findings-duV5flPv.js.map}

@@ -1 +1 @@
-{"version":3,"file":"proposal-findings-WFQrxnfe.js","names":[],"sources":["../src/analyst/proposal-findings.ts"],"sourcesContent":["import { ProposalFindingSchema } from './finding-scope'\nimport type { ProposalFinding } from './types'\n\n/** True when a finding names a source candidate generation may learn from. */\nfunction isProposalFinding(finding: unknown): finding is ProposalFinding {\n  return ProposalFindingSchema.safeParse(finding).success\n}\n\n/**\n * Reject findings whose source has not been explicitly admitted for candidate\n * generation. Search feedback and observed production behavior are allowed;\n * final evaluation data has no allowed origin.\n */\nexport function assertProposalFindings(\n  findings: unknown,\n  context = 'proposal findings',\n): ReadonlyArray<ProposalFinding> {\n  if (!Array.isArray(findings)) {\n    throw new TypeError(`${context}: expected an array`)\n  }\n  const rejected = findings.flatMap((finding, index) =>\n    isProposalFinding(finding) ? [] : [findingLabel(finding, index)],\n  )\n  if (rejected.length > 0) {\n    throw new Error(\n      `${context}: every finding must match AnalystFinding and declare ` +\n        `proposal_origin as search or production. ` +\n        `Rejected findings: [${rejected.join(', ')}].`,\n    )\n  }\n  return findings as ReadonlyArray<ProposalFinding>\n}\n\nfunction findingLabel(finding: unknown, index: number): string {\n  if (typeof finding !== 'object' || finding === null) return `index ${index}`\n  const id = (finding as { finding_id?: unknown }).finding_id\n  return typeof id === 'string' && id.length > 0 ? id : `index ${index}`\n}\n"],"mappings":";;;AAIA,SAAS,kBAAkB,SAA8C;CACvE,OAAO,sBAAsB,UAAU,OAAO,CAAC,CAAC;AAClD;;;;;;AAOA,SAAgB,uBACd,UACA,UAAU,qBACsB;CAChC,IAAI,CAAC,MAAM,QAAQ,QAAQ,GACzB,MAAM,IAAI,UAAU,GAAG,QAAQ,oBAAoB;CAErD,MAAM,WAAW,SAAS,SAAS,SAAS,UAC1C,kBAAkB,OAAO,IAAI,CAAC,IAAI,CAAC,aAAa,SAAS,KAAK,CAAC,CACjE;CACA,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,MACR,GAAG,QAAQ,qHAEc,SAAS,KAAK,IAAI,EAAE,GAC/C;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,OAAuB;CAC7D,IAAI,OAAO,YAAY,YAAY,YAAY,MAAM,OAAO,SAAS;CACrE,MAAM,KAAM,QAAqC;CACjD,OAAO,OAAO,OAAO,YAAY,GAAG,SAAS,IAAI,KAAK,SAAS;AACjE"}
+{"version":3,"file":"proposal-findings-duV5flPv.js","names":[],"sources":["../src/analyst/proposal-findings.ts"],"sourcesContent":["import { ProposalFindingSchema } from './finding-scope'\nimport type { ProposalFinding } from './types'\n\n/** True when a finding names a source candidate generation may learn from. */\nfunction isProposalFinding(finding: unknown): finding is ProposalFinding {\n  return ProposalFindingSchema.safeParse(finding).success\n}\n\n/**\n * Reject findings whose source has not been explicitly admitted for candidate\n * generation. Search feedback and observed production behavior are allowed;\n * final evaluation data has no allowed origin.\n */\nexport function assertProposalFindings(\n  findings: unknown,\n  context = 'proposal findings',\n): ReadonlyArray<ProposalFinding> {\n  if (!Array.isArray(findings)) {\n    throw new TypeError(`${context}: expected an array`)\n  }\n  const rejected = findings.flatMap((finding, index) =>\n    isProposalFinding(finding) ? [] : [findingLabel(finding, index)],\n  )\n  if (rejected.length > 0) {\n    throw new Error(\n      `${context}: every finding must match AnalystFinding and declare ` +\n        `proposal_origin as search or production. ` +\n        `Rejected findings: [${rejected.join(', ')}].`,\n    )\n  }\n  return findings as ReadonlyArray<ProposalFinding>\n}\n\nfunction findingLabel(finding: unknown, index: number): string {\n  if (typeof finding !== 'object' || finding === null) return `index ${index}`\n  const id = (finding as { finding_id?: unknown }).finding_id\n  return typeof id === 'string' && id.length > 0 ? id : `index ${index}`\n}\n"],"mappings":";;;AAIA,SAAS,kBAAkB,SAA8C;CACvE,OAAO,sBAAsB,UAAU,OAAO,CAAC,CAAC;AAClD;;;;;;AAOA,SAAgB,uBACd,UACA,UAAU,qBACsB;CAChC,IAAI,CAAC,MAAM,QAAQ,QAAQ,GACzB,MAAM,IAAI,UAAU,GAAG,QAAQ,oBAAoB;CAErD,MAAM,WAAW,SAAS,SAAS,SAAS,UAC1C,kBAAkB,OAAO,IAAI,CAAC,IAAI,CAAC,aAAa,SAAS,KAAK,CAAC,CACjE;CACA,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,MACR,GAAG,QAAQ,qHAEc,SAAS,KAAK,IAAI,EAAE,GAC/C;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,OAAuB;CAC7D,IAAI,OAAO,YAAY,YAAY,YAAY,MAAM,OAAO,SAAS;CACrE,MAAM,KAAM,QAAqC;CACjD,OAAO,OAAO,OAAO,YAAY,GAAG,SAAS,IAAI,KAAK,SAAS;AACjE"}

package/dist/{raw-provider-sink-DDLkP_eF.js → raw-provider-sink-RbaEv2vB.js}

@@ -1,4 +1,4 @@
-import { a as redact } from "./redact-D6APJS1d.js";
+import { o as redact } from "./redact-zp_ILC4l.js";
 import { promises } from "node:fs";
 import * as path from "node:path";
 // src/trace/raw-provider-sink.ts
@@ -160,4 +160,4 @@ function providerFromBaseUrl(baseUrl) {
 }
 export { providerFromBaseUrl as a, defaultProviderRedactor as i, InMemoryRawProviderSink as n, NoopRawProviderSink as r, FileSystemRawProviderSink as t };
 
-//# sourceMappingURL=raw-provider-sink-DDLkP_eF.js.map
+//# sourceMappingURL=raw-provider-sink-RbaEv2vB.js.map

package/dist/{raw-provider-sink-DDLkP_eF.js.map → raw-provider-sink-RbaEv2vB.js.map}

@@ -1 +1 @@
-{"version":3,"file":"raw-provider-sink-DDLkP_eF.js","names":["fs"],"sources":["../src/trace/raw-provider-sink.ts"],"sourcesContent":["/**\n * RawProviderSink — first-class persistence for the actual HTTP-level\n * request/response bodies of every LLM provider call.\n *\n * Why this is a separate sink from the structured `LlmSpan`:\n *\n *   - `LlmSpan` records the *intent* — model name, messages, output text,\n *     usage. It's what dashboards read; it's NOT enough for forensics.\n *   - When a downstream consumer reports \"the verifier used the wrong route\"\n *     or \"tokens look right but reasoning was missing,\" the only way to\n *     answer is the raw HTTP body. Span fields can lie (a proxy can echo\n *     a different `model` value than what actually answered); the raw\n *     response is ground truth.\n *\n * Default behaviour: opt-in. Pass `rawSink` to `LlmClientOptions` (or the\n * matrix runner / BuilderSession sets it up automatically) and every\n * request, response, and error is recorded — including retries, with the\n * attempt index attached so a flaky call's full event chain is recoverable.\n *\n * Redaction is enforced at sink time. The default redactor runs the redaction\n * core (`./redact`, default profile) over headers, bodies and the error\n * message: credential headers and fields are replaced by name, credential and\n * personal-data values by shape. Override via the sink constructor or the\n * per-call `redactor`. The `redactedFields` array on the persisted event lets\n * a reviewer see what was replaced without exposing the values.\n */\n\nimport { promises as fs } from 'node:fs'\nimport * as path from 'node:path'\nimport { redact } from './redact'\n\nexport type RawProviderDirection = 'request' | 'response' | 'error'\n\nexport interface RawProviderEvent {\n  /** Stable id. Generated by the sink if omitted. */\n  eventId: string\n  /** Trace context populated by `LlmClient` when the call is wrapped in a span. */\n  runId?: string\n  spanId?: string\n  /**\n   * Logical provider name. Free-form so callers can use whatever id matches\n   * their topology (`'openai'`, `'anthropic'`, `'tangle-router'`, …). When\n   * omitted, derived from `baseUrl` in `LlmClientOptions`.\n   */\n  provider: string\n  model: string\n  /** Endpoint path, e.g. `'/v1/chat/completions'`. */\n  endpoint: string\n  /** Base URL used for the call (already-normalised — no trailing slash). */\n  baseUrl: string\n  /** 0-indexed retry attempt. The first attempt is 0; a retried call gets 1, 2, … */\n  attemptIndex: number\n  direction: RawProviderDirection\n  /** Unix ms. */\n  timestamp: number\n  /** Wall-clock duration of the call leg. Set on `response` and `error` events; null on `request`. */\n  durationMs?: number\n  statusCode?: number\n  requestHeaders?: Record<string, string>\n  requestBody?: unknown\n  responseHeaders?: Record<string, string>\n  responseBody?: unknown\n  /** Set on `direction: 'error'` events. */\n  errorMessage?: string\n  /** JSON Pointers of the fields the redactor replaced ('/requestHeaders/Authorization', '/requestBody/apiKey', …). */\n  redactedFields: string[]\n}\n\nexport interface RawProviderSinkFilter {\n  runId?: string\n  spanId?: string\n  direction?: RawProviderDirection\n  attemptIndex?: number\n}\n\nexport interface RawProviderSink {\n  record(event: RawProviderEvent): Promise<void>\n  /** Optional listing — implementations that durably persist (file, db) should support this. */\n  list?(filter?: RawProviderSinkFilter): Promise<RawProviderEvent[]>\n  /** Optional teardown for backed implementations. */\n  close?(): Promise<void>\n}\n\nexport type ProviderRedactor = (event: RawProviderEvent) => RawProviderEvent\n\n/**\n * The default redactor: the redaction core over the parts of an event that\n * carry provider traffic. Every replaced path is added to `redactedFields`.\n */\nexport function defaultProviderRedactor(event: RawProviderEvent): RawProviderEvent {\n  const { requestHeaders, requestBody, responseHeaders, responseBody, errorMessage } = event\n  const { value, report } = redact({\n    requestHeaders,\n    requestBody,\n    responseHeaders,\n    responseBody,\n    errorMessage,\n  })\n  return {\n    ...event,\n    ...value,\n    redactedFields: [\n      ...(event.redactedFields ?? []),\n      ...report.findings.map((finding) => finding.path),\n    ],\n  }\n}\n\n// ── In-memory ────────────────────────────────────────────────────────────\n\nexport interface InMemoryRawProviderSinkOptions {\n  redactor?: ProviderRedactor\n}\n\nexport class InMemoryRawProviderSink implements RawProviderSink {\n  private events: RawProviderEvent[] = []\n  private redactor: ProviderRedactor\n\n  constructor(opts: InMemoryRawProviderSinkOptions = {}) {\n    this.redactor = opts.redactor ?? defaultProviderRedactor\n  }\n\n  async record(event: RawProviderEvent): Promise<void> {\n    this.events.push(this.redactor({ ...event, redactedFields: event.redactedFields ?? [] }))\n  }\n\n  async list(filter: RawProviderSinkFilter = {}): Promise<RawProviderEvent[]> {\n    return this.events.filter(\n      (e) =>\n        (filter.runId === undefined || e.runId === filter.runId) &&\n        (filter.spanId === undefined || e.spanId === filter.spanId) &&\n        (filter.direction === undefined || e.direction === filter.direction) &&\n        (filter.attemptIndex === undefined || e.attemptIndex === filter.attemptIndex),\n    )\n  }\n\n  size(): number {\n    return this.events.length\n  }\n}\n\nexport class NoopRawProviderSink implements RawProviderSink {\n  async record(): Promise<void> {\n    /* no-op */\n  }\n  /**\n   * Returns an empty array. Implemented so `assertRunCaptured` does not\n   * trip the `no_raw_sink` issue when a caller explicitly opts out of\n   * capture by passing this sink — opt-out is a deliberate choice, not a\n   * misconfiguration.\n   */\n  async list(): Promise<RawProviderEvent[]> {\n    return []\n  }\n}\n\n// ── Filesystem (NDJSON) ──────────────────────────────────────────────────\n\nexport interface FileSystemRawProviderSinkOptions {\n  /** Directory the NDJSON file is written into. Created if missing. */\n  dir: string\n  /** File name; default `'raw-provider-events.ndjson'`. */\n  fileName?: string\n  /** Bytes after which the writer rolls over to a new file (default 32 MiB). */\n  rollAtBytes?: number\n  redactor?: ProviderRedactor\n}\n\nexport class FileSystemRawProviderSink implements RawProviderSink {\n  private dir: string\n  private fileName: string\n  private rollAtBytes: number\n  private redactor: ProviderRedactor\n  private bytesWritten = 0\n  private rollIndex = 0\n  private initPromise: Promise<void> | null = null\n\n  constructor(opts: FileSystemRawProviderSinkOptions) {\n    this.dir = opts.dir\n    this.fileName = opts.fileName ?? 'raw-provider-events.ndjson'\n    this.rollAtBytes = opts.rollAtBytes ?? 32 * 1024 * 1024\n    this.redactor = opts.redactor ?? defaultProviderRedactor\n  }\n\n  private async ensureInit(): Promise<void> {\n    if (!this.initPromise) {\n      this.initPromise = fs.mkdir(this.dir, { recursive: true }).then(() => undefined)\n    }\n    await this.initPromise\n  }\n\n  private currentPath(): string {\n    if (this.rollIndex === 0) return path.join(this.dir, this.fileName)\n    return path.join(this.dir, `${this.fileName}.${this.rollIndex}`)\n  }\n\n  async record(event: RawProviderEvent): Promise<void> {\n    await this.ensureInit()\n    const redacted = this.redactor({ ...event, redactedFields: event.redactedFields ?? [] })\n    const line = `${JSON.stringify(redacted)}\\n`\n    if (this.bytesWritten + line.length > this.rollAtBytes && this.bytesWritten > 0) {\n      this.rollIndex += 1\n      this.bytesWritten = 0\n    }\n    await fs.appendFile(this.currentPath(), line, 'utf8')\n    this.bytesWritten += line.length\n  }\n\n  async list(filter: RawProviderSinkFilter = {}): Promise<RawProviderEvent[]> {\n    await this.ensureInit()\n    const out: RawProviderEvent[] = []\n    for (let i = 0; i <= this.rollIndex; i++) {\n      const file =\n        i === 0 ? path.join(this.dir, this.fileName) : path.join(this.dir, `${this.fileName}.${i}`)\n      let body: string\n      try {\n        body = await fs.readFile(file, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') continue\n        throw err\n      }\n      for (const line of body.split('\\n')) {\n        if (!line) continue\n        const event = JSON.parse(line) as RawProviderEvent\n        if (filter.runId !== undefined && event.runId !== filter.runId) continue\n        if (filter.spanId !== undefined && event.spanId !== filter.spanId) continue\n        if (filter.direction !== undefined && event.direction !== filter.direction) continue\n        if (filter.attemptIndex !== undefined && event.attemptIndex !== filter.attemptIndex)\n          continue\n        out.push(event)\n      }\n    }\n    return out\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\n/**\n * Best-effort provider id from a base URL. Falls back to the URL host when\n * none of the well-known patterns match.\n */\nexport function providerFromBaseUrl(baseUrl: string): string {\n  const lower = baseUrl.toLowerCase()\n  if (lower.includes('api.openai.com')) return 'openai'\n  if (lower.includes('api.anthropic.com')) return 'anthropic'\n  if (lower.includes('generativelanguage.googleapis.com')) return 'google'\n  if (lower.includes('api.together.ai') || lower.includes('api.together.xyz')) return 'together'\n  if (lower.includes('api.deepseek.com')) return 'deepseek'\n  if (lower.includes('router.tangle.tools')) return 'tangle-router'\n  if (lower.includes('api.litellm') || lower.includes('litellm')) return 'litellm'\n  try {\n    return new URL(baseUrl).host\n  } catch {\n    return baseUrl\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyFA,SAAgB,wBAAwB,OAA2C;CACjF,MAAM,EAAE,gBAAgB,aAAa,iBAAiB,cAAc,iBAAiB;CACrF,MAAM,EAAE,OAAO,WAAW,OAAO;EAC/B;EACA;EACA;EACA;EACA;CACF,CAAC;CACD,OAAO;EACL,GAAG;EACH,GAAG;EACH,gBAAgB,CACd,GAAI,MAAM,kBAAkB,CAAC,GAC7B,GAAG,OAAO,SAAS,KAAK,YAAY,QAAQ,IAAI,CAClD;CACF;AACF;AAQA,IAAa,0BAAb,MAAgE;CAC9D,SAAqC,CAAC;CACtC;CAEA,YAAY,OAAuC,CAAC,GAAG;EACrD,KAAK,WAAW,KAAK,YAAY;CACnC;CAEA,MAAM,OAAO,OAAwC;EACnD,KAAK,OAAO,KAAK,KAAK,SAAS;GAAE,GAAG;GAAO,gBAAgB,MAAM,kBAAkB,CAAC;EAAE,CAAC,CAAC;CAC1F;CAEA,MAAM,KAAK,SAAgC,CAAC,GAAgC;EAC1E,OAAO,KAAK,OAAO,QAChB,OACE,OAAO,UAAU,KAAA,KAAa,EAAE,UAAU,OAAO,WACjD,OAAO,WAAW,KAAA,KAAa,EAAE,WAAW,OAAO,YACnD,OAAO,cAAc,KAAA,KAAa,EAAE,cAAc,OAAO,eACzD,OAAO,iBAAiB,KAAA,KAAa,EAAE,iBAAiB,OAAO,aACpE;CACF;CAEA,OAAe;EACb,OAAO,KAAK,OAAO;CACrB;AACF;AAEA,IAAa,sBAAb,MAA4D;CAC1D,MAAM,SAAwB,CAE9B;;;;;;;CAOA,MAAM,OAAoC;EACxC,OAAO,CAAC;CACV;AACF;AAcA,IAAa,4BAAb,MAAkE;CAChE;CACA;CACA;CACA;CACA,eAAuB;CACvB,YAAoB;CACpB,cAA4C;CAE5C,YAAY,MAAwC;EAClD,KAAK,MAAM,KAAK;EAChB,KAAK,WAAW,KAAK,YAAY;EACjC,KAAK,cAAc,KAAK,eAAe;EACvC,KAAK,WAAW,KAAK,YAAY;CACnC;CAEA,MAAc,aAA4B;EACxC,IAAI,CAAC,KAAK,aACR,KAAK,cAAcA,SAAG,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC,CAAC,CAAC,WAAW,KAAA,CAAS;EAEjF,MAAM,KAAK;CACb;CAEA,cAA8B;EAC5B,IAAI,KAAK,cAAc,GAAG,OAAO,KAAK,KAAK,KAAK,KAAK,KAAK,QAAQ;EAClE,OAAO,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,SAAS,GAAG,KAAK,WAAW;CACjE;CAEA,MAAM,OAAO,OAAwC;EACnD,MAAM,KAAK,WAAW;EACtB,MAAM,WAAW,KAAK,SAAS;GAAE,GAAG;GAAO,gBAAgB,MAAM,kBAAkB,CAAC;EAAE,CAAC;EACvF,MAAM,OAAO,GAAG,KAAK,UAAU,QAAQ,EAAE;EACzC,IAAI,KAAK,eAAe,KAAK,SAAS,KAAK,eAAe,KAAK,eAAe,GAAG;GAC/E,KAAK,aAAa;GAClB,KAAK,eAAe;EACtB;EACA,MAAMA,SAAG,WAAW,KAAK,YAAY,GAAG,MAAM,MAAM;EACpD,KAAK,gBAAgB,KAAK;CAC5B;CAEA,MAAM,KAAK,SAAgC,CAAC,GAAgC;EAC1E,MAAM,KAAK,WAAW;EACtB,MAAM,MAA0B,CAAC;EACjC,KAAK,IAAI,IAAI,GAAG,KAAK,KAAK,WAAW,KAAK;GACxC,MAAM,OACJ,MAAM,IAAI,KAAK,KAAK,KAAK,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,SAAS,GAAG,GAAG;GAC5F,IAAI;GACJ,IAAI;IACF,OAAO,MAAMA,SAAG,SAAS,MAAM,MAAM;GACvC,SAAS,KAAK;IACZ,IAAK,IAA8B,SAAS,UAAU;IACtD,MAAM;GACR;GACA,KAAK,MAAM,QAAQ,KAAK,MAAM,IAAI,GAAG;IACnC,IAAI,CAAC,MAAM;IACX,MAAM,QAAQ,KAAK,MAAM,IAAI;IAC7B,IAAI,OAAO,UAAU,KAAA,KAAa,MAAM,UAAU,OAAO,OAAO;IAChE,IAAI,OAAO,WAAW,KAAA,KAAa,MAAM,WAAW,OAAO,QAAQ;IACnE,IAAI,OAAO,cAAc,KAAA,KAAa,MAAM,cAAc,OAAO,WAAW;IAC5E,IAAI,OAAO,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,OAAO,cACrE;IACF,IAAI,KAAK,KAAK;GAChB;EACF;EACA,OAAO;CACT;AACF;;;;;AAQA,SAAgB,oBAAoB,SAAyB;CAC3D,MAAM,QAAQ,QAAQ,YAAY;CAClC,IAAI,MAAM,SAAS,gBAAgB,GAAG,OAAO;CAC7C,IAAI,MAAM,SAAS,mBAAmB,GAAG,OAAO;CAChD,IAAI,MAAM,SAAS,mCAAmC,GAAG,OAAO;CAChE,IAAI,MAAM,SAAS,iBAAiB,KAAK,MAAM,SAAS,kBAAkB,GAAG,OAAO;CACpF,IAAI,MAAM,SAAS,kBAAkB,GAAG,OAAO;CAC/C,IAAI,MAAM,SAAS,qBAAqB,GAAG,OAAO;CAClD,IAAI,MAAM,SAAS,aAAa,KAAK,MAAM,SAAS,SAAS,GAAG,OAAO;CACvE,IAAI;EACF,OAAO,IAAI,IAAI,OAAO,CAAC,CAAC;CAC1B,QAAQ;EACN,OAAO;CACT;AACF"}
+{"version":3,"file":"raw-provider-sink-RbaEv2vB.js","names":["fs"],"sources":["../src/trace/raw-provider-sink.ts"],"sourcesContent":["/**\n * RawProviderSink — first-class persistence for the actual HTTP-level\n * request/response bodies of every LLM provider call.\n *\n * Why this is a separate sink from the structured `LlmSpan`:\n *\n *   - `LlmSpan` records the *intent* — model name, messages, output text,\n *     usage. It's what dashboards read; it's NOT enough for forensics.\n *   - When a downstream consumer reports \"the verifier used the wrong route\"\n *     or \"tokens look right but reasoning was missing,\" the only way to\n *     answer is the raw HTTP body. Span fields can lie (a proxy can echo\n *     a different `model` value than what actually answered); the raw\n *     response is ground truth.\n *\n * Default behaviour: opt-in. Pass `rawSink` to `LlmClientOptions` (or the\n * matrix runner / BuilderSession sets it up automatically) and every\n * request, response, and error is recorded — including retries, with the\n * attempt index attached so a flaky call's full event chain is recoverable.\n *\n * Redaction is enforced at sink time. The default redactor runs the redaction\n * core (`./redact`, default profile) over headers, bodies and the error\n * message: credential headers and fields are replaced by name, credential and\n * personal-data values by shape. Override via the sink constructor or the\n * per-call `redactor`. The `redactedFields` array on the persisted event lets\n * a reviewer see what was replaced without exposing the values.\n */\n\nimport { promises as fs } from 'node:fs'\nimport * as path from 'node:path'\nimport { redact } from './redact'\n\nexport type RawProviderDirection = 'request' | 'response' | 'error'\n\nexport interface RawProviderEvent {\n  /** Stable id. Generated by the sink if omitted. */\n  eventId: string\n  /** Trace context populated by `LlmClient` when the call is wrapped in a span. */\n  runId?: string\n  spanId?: string\n  /**\n   * Logical provider name. Free-form so callers can use whatever id matches\n   * their topology (`'openai'`, `'anthropic'`, `'tangle-router'`, …). When\n   * omitted, derived from `baseUrl` in `LlmClientOptions`.\n   */\n  provider: string\n  model: string\n  /** Endpoint path, e.g. `'/v1/chat/completions'`. */\n  endpoint: string\n  /** Base URL used for the call (already-normalised — no trailing slash). */\n  baseUrl: string\n  /** 0-indexed retry attempt. The first attempt is 0; a retried call gets 1, 2, … */\n  attemptIndex: number\n  direction: RawProviderDirection\n  /** Unix ms. */\n  timestamp: number\n  /** Wall-clock duration of the call leg. Set on `response` and `error` events; null on `request`. */\n  durationMs?: number\n  statusCode?: number\n  requestHeaders?: Record<string, string>\n  requestBody?: unknown\n  responseHeaders?: Record<string, string>\n  responseBody?: unknown\n  /** Set on `direction: 'error'` events. */\n  errorMessage?: string\n  /** JSON Pointers of the fields the redactor replaced ('/requestHeaders/Authorization', '/requestBody/apiKey', …). */\n  redactedFields: string[]\n}\n\nexport interface RawProviderSinkFilter {\n  runId?: string\n  spanId?: string\n  direction?: RawProviderDirection\n  attemptIndex?: number\n}\n\nexport interface RawProviderSink {\n  record(event: RawProviderEvent): Promise<void>\n  /** Optional listing — implementations that durably persist (file, db) should support this. */\n  list?(filter?: RawProviderSinkFilter): Promise<RawProviderEvent[]>\n  /** Optional teardown for backed implementations. */\n  close?(): Promise<void>\n}\n\nexport type ProviderRedactor = (event: RawProviderEvent) => RawProviderEvent\n\n/**\n * The default redactor: the redaction core over the parts of an event that\n * carry provider traffic. Every replaced path is added to `redactedFields`.\n */\nexport function defaultProviderRedactor(event: RawProviderEvent): RawProviderEvent {\n  const { requestHeaders, requestBody, responseHeaders, responseBody, errorMessage } = event\n  const { value, report } = redact({\n    requestHeaders,\n    requestBody,\n    responseHeaders,\n    responseBody,\n    errorMessage,\n  })\n  return {\n    ...event,\n    ...value,\n    redactedFields: [\n      ...(event.redactedFields ?? []),\n      ...report.findings.map((finding) => finding.path),\n    ],\n  }\n}\n\n// ── In-memory ────────────────────────────────────────────────────────────\n\nexport interface InMemoryRawProviderSinkOptions {\n  redactor?: ProviderRedactor\n}\n\nexport class InMemoryRawProviderSink implements RawProviderSink {\n  private events: RawProviderEvent[] = []\n  private redactor: ProviderRedactor\n\n  constructor(opts: InMemoryRawProviderSinkOptions = {}) {\n    this.redactor = opts.redactor ?? defaultProviderRedactor\n  }\n\n  async record(event: RawProviderEvent): Promise<void> {\n    this.events.push(this.redactor({ ...event, redactedFields: event.redactedFields ?? [] }))\n  }\n\n  async list(filter: RawProviderSinkFilter = {}): Promise<RawProviderEvent[]> {\n    return this.events.filter(\n      (e) =>\n        (filter.runId === undefined || e.runId === filter.runId) &&\n        (filter.spanId === undefined || e.spanId === filter.spanId) &&\n        (filter.direction === undefined || e.direction === filter.direction) &&\n        (filter.attemptIndex === undefined || e.attemptIndex === filter.attemptIndex),\n    )\n  }\n\n  size(): number {\n    return this.events.length\n  }\n}\n\nexport class NoopRawProviderSink implements RawProviderSink {\n  async record(): Promise<void> {\n    /* no-op */\n  }\n  /**\n   * Returns an empty array. Implemented so `assertRunCaptured` does not\n   * trip the `no_raw_sink` issue when a caller explicitly opts out of\n   * capture by passing this sink — opt-out is a deliberate choice, not a\n   * misconfiguration.\n   */\n  async list(): Promise<RawProviderEvent[]> {\n    return []\n  }\n}\n\n// ── Filesystem (NDJSON) ──────────────────────────────────────────────────\n\nexport interface FileSystemRawProviderSinkOptions {\n  /** Directory the NDJSON file is written into. Created if missing. */\n  dir: string\n  /** File name; default `'raw-provider-events.ndjson'`. */\n  fileName?: string\n  /** Bytes after which the writer rolls over to a new file (default 32 MiB). */\n  rollAtBytes?: number\n  redactor?: ProviderRedactor\n}\n\nexport class FileSystemRawProviderSink implements RawProviderSink {\n  private dir: string\n  private fileName: string\n  private rollAtBytes: number\n  private redactor: ProviderRedactor\n  private bytesWritten = 0\n  private rollIndex = 0\n  private initPromise: Promise<void> | null = null\n\n  constructor(opts: FileSystemRawProviderSinkOptions) {\n    this.dir = opts.dir\n    this.fileName = opts.fileName ?? 'raw-provider-events.ndjson'\n    this.rollAtBytes = opts.rollAtBytes ?? 32 * 1024 * 1024\n    this.redactor = opts.redactor ?? defaultProviderRedactor\n  }\n\n  private async ensureInit(): Promise<void> {\n    if (!this.initPromise) {\n      this.initPromise = fs.mkdir(this.dir, { recursive: true }).then(() => undefined)\n    }\n    await this.initPromise\n  }\n\n  private currentPath(): string {\n    if (this.rollIndex === 0) return path.join(this.dir, this.fileName)\n    return path.join(this.dir, `${this.fileName}.${this.rollIndex}`)\n  }\n\n  async record(event: RawProviderEvent): Promise<void> {\n    await this.ensureInit()\n    const redacted = this.redactor({ ...event, redactedFields: event.redactedFields ?? [] })\n    const line = `${JSON.stringify(redacted)}\\n`\n    if (this.bytesWritten + line.length > this.rollAtBytes && this.bytesWritten > 0) {\n      this.rollIndex += 1\n      this.bytesWritten = 0\n    }\n    await fs.appendFile(this.currentPath(), line, 'utf8')\n    this.bytesWritten += line.length\n  }\n\n  async list(filter: RawProviderSinkFilter = {}): Promise<RawProviderEvent[]> {\n    await this.ensureInit()\n    const out: RawProviderEvent[] = []\n    for (let i = 0; i <= this.rollIndex; i++) {\n      const file =\n        i === 0 ? path.join(this.dir, this.fileName) : path.join(this.dir, `${this.fileName}.${i}`)\n      let body: string\n      try {\n        body = await fs.readFile(file, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') continue\n        throw err\n      }\n      for (const line of body.split('\\n')) {\n        if (!line) continue\n        const event = JSON.parse(line) as RawProviderEvent\n        if (filter.runId !== undefined && event.runId !== filter.runId) continue\n        if (filter.spanId !== undefined && event.spanId !== filter.spanId) continue\n        if (filter.direction !== undefined && event.direction !== filter.direction) continue\n        if (filter.attemptIndex !== undefined && event.attemptIndex !== filter.attemptIndex)\n          continue\n        out.push(event)\n      }\n    }\n    return out\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\n/**\n * Best-effort provider id from a base URL. Falls back to the URL host when\n * none of the well-known patterns match.\n */\nexport function providerFromBaseUrl(baseUrl: string): string {\n  const lower = baseUrl.toLowerCase()\n  if (lower.includes('api.openai.com')) return 'openai'\n  if (lower.includes('api.anthropic.com')) return 'anthropic'\n  if (lower.includes('generativelanguage.googleapis.com')) return 'google'\n  if (lower.includes('api.together.ai') || lower.includes('api.together.xyz')) return 'together'\n  if (lower.includes('api.deepseek.com')) return 'deepseek'\n  if (lower.includes('router.tangle.tools')) return 'tangle-router'\n  if (lower.includes('api.litellm') || lower.includes('litellm')) return 'litellm'\n  try {\n    return new URL(baseUrl).host\n  } catch {\n    return baseUrl\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyFA,SAAgB,wBAAwB,OAA2C;CACjF,MAAM,EAAE,gBAAgB,aAAa,iBAAiB,cAAc,iBAAiB;CACrF,MAAM,EAAE,OAAO,WAAW,OAAO;EAC/B;EACA;EACA;EACA;EACA;CACF,CAAC;CACD,OAAO;EACL,GAAG;EACH,GAAG;EACH,gBAAgB,CACd,GAAI,MAAM,kBAAkB,CAAC,GAC7B,GAAG,OAAO,SAAS,KAAK,YAAY,QAAQ,IAAI,CAClD;CACF;AACF;AAQA,IAAa,0BAAb,MAAgE;CAC9D,SAAqC,CAAC;CACtC;CAEA,YAAY,OAAuC,CAAC,GAAG;EACrD,KAAK,WAAW,KAAK,YAAY;CACnC;CAEA,MAAM,OAAO,OAAwC;EACnD,KAAK,OAAO,KAAK,KAAK,SAAS;GAAE,GAAG;GAAO,gBAAgB,MAAM,kBAAkB,CAAC;EAAE,CAAC,CAAC;CAC1F;CAEA,MAAM,KAAK,SAAgC,CAAC,GAAgC;EAC1E,OAAO,KAAK,OAAO,QAChB,OACE,OAAO,UAAU,KAAA,KAAa,EAAE,UAAU,OAAO,WACjD,OAAO,WAAW,KAAA,KAAa,EAAE,WAAW,OAAO,YACnD,OAAO,cAAc,KAAA,KAAa,EAAE,cAAc,OAAO,eACzD,OAAO,iBAAiB,KAAA,KAAa,EAAE,iBAAiB,OAAO,aACpE;CACF;CAEA,OAAe;EACb,OAAO,KAAK,OAAO;CACrB;AACF;AAEA,IAAa,sBAAb,MAA4D;CAC1D,MAAM,SAAwB,CAE9B;;;;;;;CAOA,MAAM,OAAoC;EACxC,OAAO,CAAC;CACV;AACF;AAcA,IAAa,4BAAb,MAAkE;CAChE;CACA;CACA;CACA;CACA,eAAuB;CACvB,YAAoB;CACpB,cAA4C;CAE5C,YAAY,MAAwC;EAClD,KAAK,MAAM,KAAK;EAChB,KAAK,WAAW,KAAK,YAAY;EACjC,KAAK,cAAc,KAAK,eAAe;EACvC,KAAK,WAAW,KAAK,YAAY;CACnC;CAEA,MAAc,aAA4B;EACxC,IAAI,CAAC,KAAK,aACR,KAAK,cAAcA,SAAG,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC,CAAC,CAAC,WAAW,KAAA,CAAS;EAEjF,MAAM,KAAK;CACb;CAEA,cAA8B;EAC5B,IAAI,KAAK,cAAc,GAAG,OAAO,KAAK,KAAK,KAAK,KAAK,KAAK,QAAQ;EAClE,OAAO,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,SAAS,GAAG,KAAK,WAAW;CACjE;CAEA,MAAM,OAAO,OAAwC;EACnD,MAAM,KAAK,WAAW;EACtB,MAAM,WAAW,KAAK,SAAS;GAAE,GAAG;GAAO,gBAAgB,MAAM,kBAAkB,CAAC;EAAE,CAAC;EACvF,MAAM,OAAO,GAAG,KAAK,UAAU,QAAQ,EAAE;EACzC,IAAI,KAAK,eAAe,KAAK,SAAS,KAAK,eAAe,KAAK,eAAe,GAAG;GAC/E,KAAK,aAAa;GAClB,KAAK,eAAe;EACtB;EACA,MAAMA,SAAG,WAAW,KAAK,YAAY,GAAG,MAAM,MAAM;EACpD,KAAK,gBAAgB,KAAK;CAC5B;CAEA,MAAM,KAAK,SAAgC,CAAC,GAAgC;EAC1E,MAAM,KAAK,WAAW;EACtB,MAAM,MAA0B,CAAC;EACjC,KAAK,IAAI,IAAI,GAAG,KAAK,KAAK,WAAW,KAAK;GACxC,MAAM,OACJ,MAAM,IAAI,KAAK,KAAK,KAAK,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,SAAS,GAAG,GAAG;GAC5F,IAAI;GACJ,IAAI;IACF,OAAO,MAAMA,SAAG,SAAS,MAAM,MAAM;GACvC,SAAS,KAAK;IACZ,IAAK,IAA8B,SAAS,UAAU;IACtD,MAAM;GACR;GACA,KAAK,MAAM,QAAQ,KAAK,MAAM,IAAI,GAAG;IACnC,IAAI,CAAC,MAAM;IACX,MAAM,QAAQ,KAAK,MAAM,IAAI;IAC7B,IAAI,OAAO,UAAU,KAAA,KAAa,MAAM,UAAU,OAAO,OAAO;IAChE,IAAI,OAAO,WAAW,KAAA,KAAa,MAAM,WAAW,OAAO,QAAQ;IACnE,IAAI,OAAO,cAAc,KAAA,KAAa,MAAM,cAAc,OAAO,WAAW;IAC5E,IAAI,OAAO,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,OAAO,cACrE;IACF,IAAI,KAAK,KAAK;GAChB;EACF;EACA,OAAO;CACT;AACF;;;;;AAQA,SAAgB,oBAAoB,SAAyB;CAC3D,MAAM,QAAQ,QAAQ,YAAY;CAClC,IAAI,MAAM,SAAS,gBAAgB,GAAG,OAAO;CAC7C,IAAI,MAAM,SAAS,mBAAmB,GAAG,OAAO;CAChD,IAAI,MAAM,SAAS,mCAAmC,GAAG,OAAO;CAChE,IAAI,MAAM,SAAS,iBAAiB,KAAK,MAAM,SAAS,kBAAkB,GAAG,OAAO;CACpF,IAAI,MAAM,SAAS,kBAAkB,GAAG,OAAO;CAC/C,IAAI,MAAM,SAAS,qBAAqB,GAAG,OAAO;CAClD,IAAI,MAAM,SAAS,aAAa,KAAK,MAAM,SAAS,SAAS,GAAG,OAAO;CACvE,IAAI;EACF,OAAO,IAAI,IAAI,OAAO,CAAC,CAAC;CAC1B,QAAQ;EACN,OAAO;CACT;AACF"}

package/dist/{redact-C_wzC4c_.d.ts → redact-jrT0APTU.d.ts}

@@ -1,39 +1,54 @@
 /**
- * The redaction core: the one place that decides what must not leave a
- * process or a machine. Every trace writer, exporter, uploader and share path
- * calls it.
- *
- * Three profiles, each a superset of the one before:
- *   - `default` removes credentials, found by key name and by value shape, and
- *     personal data (email addresses, card numbers, SSNs, phone numbers).
- *     Content is kept. Strings are capped at 1 MiB.
- *   - `share` also removes phone, IP and postal-address fields, replaces
- *     account, user, tenant, session and request identifiers with pseudonyms
- *     that stay consistent within one call, and caps strings at 64 KiB.
- *   - `strict` also removes prompt, completion, message, tool payload and other
- *     raw-content fields and embedded media, and caps strings at 4 KiB.
- *
- * A string that holds a credential is replaced whole with
- * `[REDACTED:<detector>]`. Replacing only the matched span can leave part of a
- * secret behind when a pattern's boundary is wrong. Personal data inside prose
- * is replaced in place, so the rest of the text stays readable.
- *
- * `assessShareSafety` reads a value without changing it and returns SAFE,
- * SAFE_WITH_WARNINGS, UNSAFE or UNKNOWN. UNKNOWN means the scanner could not
- * read part of the value (a cycle, binary data, nesting past the depth limit);
- * it is never reported as safe. `redactForShare` redacts and then assesses the
- * redacted output, so a detector gap shows up as UNSAFE instead of a leak.
- *
- * Reports and findings carry paths, categories and detector ids, never the
- * values they matched.
+ * Canonical snake_case form of a field name, so camelCase, kebab-case, dotted
+ * and snake_case spellings match one list: `userPassword`, `user-password` and
+ * `user.password` all become `user_password`.
+ */
+declare function normalizeKey(key: string): string;
+/**
+ * The safety category a field name puts its value in, or undefined for an
+ * ordinary field. Token counts and model limits (`inputTokens`, `max_tokens`,
+ * `token_count`) and names such as `author` are not credentials.
+ */
+declare function classifyKey(key: string): SafetyCategory | undefined;
+interface ValueDetectorOptions {
+  /** Conservative scanning treats untyped bare numbers as private; formatted preserves the core's existing policy. */
+  numericPolicy?: 'formatted' | 'conservative';
+}
+interface ValueDetector {
+  readonly id: string;
+  readonly pattern: RegExp;
+  readonly requires?: string;
+  readonly substringPattern?: RegExp;
+  readonly conservativePattern?: RegExp;
+  readonly accept?: (match: string, options?: ValueDetectorOptions) => boolean;
+}
+/** Stateless UTF-16 ranges for String.slice. Named `secret` groups preserve surrounding syntax. */
+declare function matchValueDetector(text: string, detector: ValueDetector, options?: ValueDetectorOptions): Generator<{
+  start: number;
+  end: number;
+}>;
+/**
+ * High-confidence credential shapes. A hit replaces the whole string. Each
+ * pattern is written without the `g` flag so `test` stays stateless.
  *
- * The key classifier and the credential detector list are adapted from
- * agent-inspect (MIT, commit 3c3cbeda). `scripts/redaction-corpus.json` holds
- * the must-flag and must-not-flag cases, including the copied agent-inspect
- * safety corpus with its MIT notice; `pnpm redaction:corpus` runs this module
- * over every case.
+ * The provider-branded prefixes below (`sk-ant-`, `AIza`, `ghp_`, `xoxb-`,
+ * `AKIA`, `hubcap_`, `eyJ`…) have no leading `\b`. A `\b` word-boundary
+ * assertion does not fire between two word characters, so it silently passed
+ * a real token when something else glued a word character directly in front
+ * of it — a compound identifier like `env.GITHUB_TOKEN_ghp_<token>` (a `_`
+ * before `ghp_`) or `BACKUP_AKIA<key>` (a `_` before `AKIA`). These prefixes
+ * are distinctive enough on their own, so scanning without that boundary
+ * catches the same real credential wherever it is glued, at the cost of no
+ * meaningful new false positives (`scripts/redaction-corpus.json` covers
+ * both).
  */
+declare const CREDENTIAL_DETECTORS: readonly ValueDetector[];
+/** Ordered rules. Conservative numeric scanning is explicit; core patterns keep their existing defaults. */
+declare const PERSONAL_DATA_DETECTORS: readonly ValueDetector[];
+/** The first credential detector that matches, or undefined. */
+declare function detectCredential(text: string): string | undefined;
 type RedactionProfile = 'default' | 'share' | 'strict';
+declare const REDACTION_PROFILES: readonly RedactionProfile[];
 /** What a finding is about. `media` is embedded binary content (a `data:` URI) no text detector can read. */
 type SafetyCategory = 'credential' | 'personal-data' | 'identifier' | 'raw-content' | 'media';
 /**
@@ -41,7 +56,7 @@ type SafetyCategory = 'credential' | 'personal-data' | 'identifier' | 'raw-conte
  * redaction behavior (saved optimizer inputs, per-span redaction stamps) keys
  * on it, so bump it whenever a detector, key list, cap or marker changes.
  */
-declare const REDACTION_VERSION = "2.4.0";
+declare const REDACTION_VERSION = "2.5.0";
 interface RedactionFinding {
   /** JSON Pointer (RFC 6901) to the value or key that was changed. */
   path: string;
@@ -137,5 +152,5 @@ declare function redactForShare<T>(value: T, options?: RedactOptions): {
   report: RedactionReport;
   verdict: ShareSafetyVerdict;
 };
-export { shareAllowed as _, RedactionProfile as a, ShareSafetyFinding as c, assessShareSafety as d, combineVerdicts as f, redactText as g, redactForShare as h, RedactionFinding as i, ShareSafetyStatus as l, redact as m, REDACTION_VERSION as n, RedactionReport as o, emptyRedactionReport as p, RedactOptions as r, SafetyCategory as s, AssessOptions as t, ShareSafetyVerdict as u };
-//# sourceMappingURL=redact-C_wzC4c_.d.ts.map
+export { classifyKey as C, normalizeKey as E, ValueDetectorOptions as S, matchValueDetector as T, redactText as _, RedactionFinding as a, PERSONAL_DATA_DETECTORS as b, SafetyCategory as c, ShareSafetyVerdict as d, assessShareSafety as f, redactForShare as g, redact as h, RedactOptions as i, ShareSafetyFinding as l, emptyRedactionReport as m, REDACTION_PROFILES as n, RedactionProfile as o, combineVerdicts as p, REDACTION_VERSION as r, RedactionReport as s, AssessOptions as t, ShareSafetyStatus as u, shareAllowed as v, detectCredential as w, ValueDetector as x, CREDENTIAL_DETECTORS as y };
+//# sourceMappingURL=redact-jrT0APTU.d.ts.map

package/dist/redact-jrT0APTU.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"redact-jrT0APTU.d.ts","names":[],"sources":["../src/trace/redaction/detectors.ts","../src/trace/redact.ts"],"mappings":";;;;;iBASgB,aAAa;;;;;;iBAmNb,YAAY,cAAc;UAWzB;;EAEf;;UAGe;WACN;WACA,SAAS;WACT;WACA,mBAAmB;WACnB,sBAAsB;WACtB,UAAU,eAAe,UAAU;;;iBAI7B,mBACf,cACA,UAAU,eACV,UAAS,uBACR;EAAY;EAAe;;;;;;;;;;;;;;;;;cAiEjB,+BAA+B;;cA8I/B,kCAAkC;;iBAuF/B,iBAAiB;KC1erB;cAEC,6BAA6B;;KAG9B;;;;;;cAOC;UAEI;;EAEf;EACA,UAAU;EACV;EACA;;UAGe;EACf;EACA,SAAS;;EAET;EACA,YAAY;EACZ;EACA,UAAU;;UAGK;;EAEf,UAAU;;EAEV;;;;;;;;;EASA;;EAEA,SAAS;;UAGM;;EAEf,UAAU;;EAEV;;KAGU;UAEK;EACf,UAAU;EACV;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;EACR,SAAS;EACT,UAAU;;EAEV;;;iBAkVc,qBAAqB,UAAS,mBAA+B;;;;;;iBAkC7D,OAAO,GACrB,OAAO,GACP,UAAS;EACN,OAAO;EAAG,QAAQ;;;;;;iBAUP,WAAW,cAAc,UAAS;;;;;;;iBAgKlC,kBAAkB,gBAAgB,UAAS,gBAAqB;;;;;iBAoBhE,gBACd,SAAS,kBACT,mBAAmB,sBACnB,iCACC;;iBAsCa,aAAa,SAAS;;;;;;iBAStB,eAAe,GAC7B,OAAO,GACP,UAAS;EACN,OAAO;EAAG,QAAQ;EAAiB,SAAS"}