@tangle-network/agent-eval 0.150.1 → 0.161.0

package/dist/matrix-DrVnRp4G.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"matrix-DrVnRp4G.d.ts","names":[],"sources":["../src/multishot/judges.ts","../src/multishot/types.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;cAyBa;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;EAEA;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;EACA;;EAEA;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAkIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UClM/B;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;EAAO;EAAgB;EAAiB,SAAS;MAC9C;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;UCpH3D,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,iBAAiB;;;;EAIjB,kBAAkB;EAClB;EACA;EACA,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UCtFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,iBAAiB;;EAEjB,kBAAkB;;;;;;;;;;;;;;;EAelB,UAAU,cAAc;;EAExB;EACA;;;;;UAMe;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}

package/dist/pre-registration-DakwTRXk.js

@@ -1,96 +0,0 @@
-//#region src/pre-registration.ts
-/**
-* Deterministic JSON canonicalization — sort object keys recursively.
-*
-* Two semantically-equal objects produce byte-identical canonicalized output;
-* this is what makes a content-hash stable across encoders, key insertion
-* orders, and runtime versions. Exported for any consumer that needs the same
-* canonicalization guarantee outside the manifest-signing path (e.g., signing
-* an artifact bundle, hashing a dataset version, etc.).
-*/
-function canonicalize(v) {
-	if (v === null || typeof v !== "object") return v;
-	if (Array.isArray(v)) return v.map(canonicalize);
-	const keys = Object.keys(v).sort();
-	const out = {};
-	for (const k of keys) out[k] = canonicalize(v[k]);
-	return out;
-}
-/**
-* SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.
-*
-* The same primitive `signManifest` and `verifyManifest` are built on, exposed
-* directly so consumers signing arbitrary structured content (artifact bundles,
-* production packets, dataset manifests, etc.) don't have to re-derive
-* canonicalize+sha256 from scratch.
-*
-* Stable across:
-*   - object key insertion order (canonicalization sorts keys recursively)
-*   - encoder choice (UTF-8 via TextEncoder, fixed)
-*   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)
-*
-* Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,
-* which takes a string input and returns a truncated 12-char prompt id.
-* Use `hashJson` when you mean "canonicalize then hash."
-*
-* @example
-*   const hash = await hashJson({ id: '1', kind: 'spec' })
-*   // 'a3f1...' (64 hex chars)
-*/
-async function hashJson(obj) {
-	const canonical = canonicalize(obj);
-	const bytes = new TextEncoder().encode(JSON.stringify(canonical));
-	const digest = await globalThis.crypto.subtle.digest("SHA-256", bytes);
-	return Array.from(new Uint8Array(digest)).map((b) => b.toString(16).padStart(2, "0")).join("");
-}
-/**
-* Sign a manifest with a SHA-256 content hash.
-*
-* The hash covers the canonicalized manifest with the `contentHash`
-* and `algo` fields stripped; this lets verifiers re-sign the rest and
-* compare. Returned manifest always carries `algo: 'sha256-content'`
-* so downstream consumers can identify the scheme; manifests without
-* `algo` still verify because it is stripped before hashing on both sides.
-*/
-async function signManifest(m) {
-	const hash = await hashJson(m);
-	return {
-		...m,
-		contentHash: hash,
-		algo: "sha256-content"
-	};
-}
-/**
-* Verify that a signed manifest has not been tampered with.
-*
-* Strips `contentHash` and `algo` before re-signing so manifests without
-* `algo` verify identically to ones that carry it.
-*/
-async function verifyManifest(m) {
-	const { contentHash, algo: _algo, ...rest } = m;
-	return (await signManifest(rest)).contentHash === contentHash;
-}
-/**
-* Evaluate a pre-registered hypothesis against observed results.
-* Mechanical — no re-interpretation permitted.
-*/
-async function evaluateHypothesis(manifest, observed) {
-	if (!await verifyManifest(manifest)) throw new Error("evaluateHypothesis: manifest content hash mismatch (tampered)");
-	const reasons = [];
-	if (!(manifest.direction === "increase" ? observed.effect > 0 : observed.effect < 0)) reasons.push("wrong_direction");
-	if (Math.abs(observed.effect) < manifest.minEffect) reasons.push("effect_too_small");
-	if (observed.pValue >= manifest.alpha) reasons.push("not_significant");
-	if (observed.n < manifest.preRegisteredN) reasons.push("undersampled");
-	return {
-		manifest,
-		observedN: observed.n,
-		observedEffect: observed.effect,
-		observedPValue: observed.pValue,
-		confirmed: reasons.length === 0,
-		rejectionReasons: reasons
-	};
-}
-//#endregion
-export { verifyManifest as a, signManifest as i, evaluateHypothesis as n, hashJson as r, canonicalize as t };
-
-//# sourceMappingURL=pre-registration-DakwTRXk.js.map

package/dist/pre-registration-DakwTRXk.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"pre-registration-DakwTRXk.js","names":[],"sources":["../src/pre-registration.ts"],"sourcesContent":["/**\n * Pre-registered hypotheses — declare what you're testing BEFORE the\n * run, check it AFTER. Prevents p-hacking, optional stopping, and the\n * \"we ran until it looked good\" failure mode.\n *\n * Manifest is a plain JSON-friendly object. Sign it with a content hash\n * + timestamp; the registered record becomes immutable. Post-run,\n * evaluate the manifest against observed results — the library refuses\n * to let you re-interpret a different metric as the declared one.\n */\n\nexport interface HypothesisManifest {\n  id: string\n  /** Human prose — goes into the audit trail. */\n  hypothesis: string\n  /** Metric the hypothesis claims to move. */\n  metric: string\n  /** 'increase' = candidate should score higher than baseline; 'decrease' = lower. */\n  direction: 'increase' | 'decrease'\n  /** Minimum effect size to count (same units as the metric). */\n  minEffect: number\n  /** Alpha threshold. */\n  alpha: number\n  /** Target statistical power at which sample size was pre-computed. */\n  power: number\n  /** Declared N per arm before running. */\n  preRegisteredN: number\n  /** ISO8601 timestamp the manifest was registered. */\n  registeredAt: string\n  /** Optional identifiers to tie into the trace corpus. */\n  baselineLabel?: string\n  candidateLabel?: string\n}\n\n/**\n * Identifier for the hashing scheme used to produce `contentHash`.\n *\n * `'sha256-content'` — sha256 hex over the canonicalized manifest with\n * the `contentHash` and `algo` fields stripped. Held as a string union\n * so future schemes can be added without breaking parsers; SignedManifest\n * values without `algo` deserialize cleanly because the field is optional.\n */\nexport type SignedManifestAlgo = 'sha256-content'\n\nexport interface SignedManifest extends HypothesisManifest {\n  /** sha256 hex of canonicalized manifest (everything except contentHash and algo). */\n  contentHash: string\n  /**\n   * Algorithm string describing how `contentHash` was produced.\n   *\n   * Optional on the type so serialized manifests without it still parse,\n   * but ALWAYS populated by {@link signManifest}. Consumers that want to\n   * enforce a known algorithm should reject manifests where this field\n   * is missing or unrecognized.\n   */\n  algo?: SignedManifestAlgo\n}\n\nexport interface HypothesisResult {\n  manifest: SignedManifest\n  observedN: number\n  observedEffect: number\n  observedPValue: number\n  /** True iff the observed effect hits the pre-declared direction with\n   *  magnitude ≥ minEffect AND p < alpha. */\n  confirmed: boolean\n  /** Enumerated reasons the hypothesis was rejected (each a machine-tag). */\n  rejectionReasons: Array<\n    'wrong_direction' | 'effect_too_small' | 'not_significant' | 'undersampled'\n  >\n  notes?: string\n}\n\n/**\n * Deterministic JSON canonicalization — sort object keys recursively.\n *\n * Two semantically-equal objects produce byte-identical canonicalized output;\n * this is what makes a content-hash stable across encoders, key insertion\n * orders, and runtime versions. Exported for any consumer that needs the same\n * canonicalization guarantee outside the manifest-signing path (e.g., signing\n * an artifact bundle, hashing a dataset version, etc.).\n */\nexport function canonicalize(v: unknown): unknown {\n  if (v === null || typeof v !== 'object') return v\n  if (Array.isArray(v)) return v.map(canonicalize)\n  const keys = Object.keys(v as Record<string, unknown>).sort()\n  const out: Record<string, unknown> = {}\n  for (const k of keys) out[k] = canonicalize((v as Record<string, unknown>)[k])\n  return out\n}\n\n/**\n * SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.\n *\n * The same primitive `signManifest` and `verifyManifest` are built on, exposed\n * directly so consumers signing arbitrary structured content (artifact bundles,\n * production packets, dataset manifests, etc.) don't have to re-derive\n * canonicalize+sha256 from scratch.\n *\n * Stable across:\n *   - object key insertion order (canonicalization sorts keys recursively)\n *   - encoder choice (UTF-8 via TextEncoder, fixed)\n *   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)\n *\n * Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,\n * which takes a string input and returns a truncated 12-char prompt id.\n * Use `hashJson` when you mean \"canonicalize then hash.\"\n *\n * @example\n *   const hash = await hashJson({ id: '1', kind: 'spec' })\n *   // 'a3f1...' (64 hex chars)\n */\nexport async function hashJson<T>(obj: T): Promise<string> {\n  const canonical = canonicalize(obj)\n  const bytes = new TextEncoder().encode(JSON.stringify(canonical))\n  const digest = await globalThis.crypto.subtle.digest('SHA-256', bytes)\n  return Array.from(new Uint8Array(digest))\n    .map((b) => b.toString(16).padStart(2, '0'))\n    .join('')\n}\n\n/**\n * Sign a manifest with a SHA-256 content hash.\n *\n * The hash covers the canonicalized manifest with the `contentHash`\n * and `algo` fields stripped; this lets verifiers re-sign the rest and\n * compare. Returned manifest always carries `algo: 'sha256-content'`\n * so downstream consumers can identify the scheme; manifests without\n * `algo` still verify because it is stripped before hashing on both sides.\n */\nexport async function signManifest(m: HypothesisManifest): Promise<SignedManifest> {\n  const hash = await hashJson(m)\n  return { ...m, contentHash: hash, algo: 'sha256-content' }\n}\n\n/**\n * Verify that a signed manifest has not been tampered with.\n *\n * Strips `contentHash` and `algo` before re-signing so manifests without\n * `algo` verify identically to ones that carry it.\n */\nexport async function verifyManifest(m: SignedManifest): Promise<boolean> {\n  const { contentHash, algo: _algo, ...rest } = m\n  void _algo\n  const resigned = await signManifest(rest)\n  return resigned.contentHash === contentHash\n}\n\n/**\n * Evaluate a pre-registered hypothesis against observed results.\n * Mechanical — no re-interpretation permitted.\n */\nexport async function evaluateHypothesis(\n  manifest: SignedManifest,\n  observed: { n: number; effect: number; pValue: number },\n): Promise<HypothesisResult> {\n  if (!(await verifyManifest(manifest))) {\n    throw new Error('evaluateHypothesis: manifest content hash mismatch (tampered)')\n  }\n  const reasons: HypothesisResult['rejectionReasons'] = []\n  const directionOk = manifest.direction === 'increase' ? observed.effect > 0 : observed.effect < 0\n  if (!directionOk) reasons.push('wrong_direction')\n  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push('effect_too_small')\n  if (observed.pValue >= manifest.alpha) reasons.push('not_significant')\n  if (observed.n < manifest.preRegisteredN) reasons.push('undersampled')\n  return {\n    manifest,\n    observedN: observed.n,\n    observedEffect: observed.effect,\n    observedPValue: observed.pValue,\n    confirmed: reasons.length === 0,\n    rejectionReasons: reasons,\n  }\n}\n"],"mappings":";;;;;;;;;;AAkFA,SAAgB,aAAa,GAAqB;CAChD,IAAI,MAAM,QAAQ,OAAO,MAAM,UAAU,OAAO;CAChD,IAAI,MAAM,QAAQ,CAAC,GAAG,OAAO,EAAE,IAAI,YAAY;CAC/C,MAAM,OAAO,OAAO,KAAK,CAA4B,CAAC,CAAC,KAAK;CAC5D,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,KAAK,MAAM,IAAI,KAAK,aAAc,EAA8B,EAAE;CAC7E,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;AAuBA,eAAsB,SAAY,KAAyB;CACzD,MAAM,YAAY,aAAa,GAAG;CAClC,MAAM,QAAQ,IAAI,YAAY,CAAC,CAAC,OAAO,KAAK,UAAU,SAAS,CAAC;CAChE,MAAM,SAAS,MAAM,WAAW,OAAO,OAAO,OAAO,WAAW,KAAK;CACrE,OAAO,MAAM,KAAK,IAAI,WAAW,MAAM,CAAC,CAAC,CACtC,KAAK,MAAM,EAAE,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG,CAAC,CAAC,CAC3C,KAAK,EAAE;AACZ;;;;;;;;;;AAWA,eAAsB,aAAa,GAAgD;CACjF,MAAM,OAAO,MAAM,SAAS,CAAC;CAC7B,OAAO;EAAE,GAAG;EAAG,aAAa;EAAM,MAAM;CAAiB;AAC3D;;;;;;;AAQA,eAAsB,eAAe,GAAqC;CACxE,MAAM,EAAE,aAAa,MAAM,OAAO,GAAG,SAAS;CAG9C,QAAO,MADgB,aAAa,IAAI,EAAA,CACxB,gBAAgB;AAClC;;;;;AAMA,eAAsB,mBACpB,UACA,UAC2B;CAC3B,IAAI,CAAE,MAAM,eAAe,QAAQ,GACjC,MAAM,IAAI,MAAM,+DAA+D;CAEjF,MAAM,UAAgD,CAAC;CAEvD,IAAI,EADgB,SAAS,cAAc,aAAa,SAAS,SAAS,IAAI,SAAS,SAAS,IAC9E,QAAQ,KAAK,iBAAiB;CAChD,IAAI,KAAK,IAAI,SAAS,MAAM,IAAI,SAAS,WAAW,QAAQ,KAAK,kBAAkB;CACnF,IAAI,SAAS,UAAU,SAAS,OAAO,QAAQ,KAAK,iBAAiB;CACrE,IAAI,SAAS,IAAI,SAAS,gBAAgB,QAAQ,KAAK,cAAc;CACrE,OAAO;EACL;EACA,WAAW,SAAS;EACpB,gBAAgB,SAAS;EACzB,gBAAgB,SAAS;EACzB,WAAW,QAAQ,WAAW;EAC9B,kBAAkB;CACpB;AACF"}

package/dist/pre-registration-zFSLEiFU.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"pre-registration-zFSLEiFU.d.ts","names":[],"sources":["../src/statistics/paired-binary.ts","../src/statistics/sequential-eprocess.ts","../src/paired-arms.ts","../src/pre-registration.ts"],"mappings":";;;;UAgBiB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,OAAO,mBAAmB,WAAW,sBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2CzD,sBAAsB,QAAQ;;UAU7B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;iBAkBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+Bc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;;;;;;;;;;;;;;;;;iBA6Ea,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;UCjgB7B;;;EAGf;;;;EAIA;;;;EAIA;;UAGe;;EAEf;;EAEA;;EAEA;;UAGe,sBAAsB;EACrC;EACA;EACA;;EAEA;;EAEA;;UAGe;;;EAGf,OAAO,YAAY;EACnB,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8BK,SAAS,OAAM,kBAAuB;;;;;UC/BrC;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;;;;;;;;;;UC1Wc;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;;;;;;;;KAWU;UAEK,uBAAuB;;EAEtC;;;;;;;;;EASA,OAAO;;UAGQ;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;iBAYc,aAAa;;;;;;;;;;;;;;;;;;;;;;iBA8BP,SAAS,GAAG,KAAK,IAAI;;;;;;;;;;iBAkBrB,aAAa,GAAG,qBAAqB,QAAQ;;;;;;;iBAW7C,eAAe,GAAG,iBAAiB;;;;;iBAWnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ"}

package/dist/produced-state-jfk8Du3b.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"produced-state-jfk8Du3b.js","names":[],"sources":["../src/agent-profile.ts","../src/completion-verifier.ts","../src/produced-state.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { type HarnessType, harnessSupportsModel } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport type { AgentProfile, HarnessType } from '@tangle-network/agent-interface'\n\n/**\n * The agentic coding harnesses an eval sweeps by default — the ones we care about\n * ranking. This is the SINGLE source of that list; consumers import it instead of\n * re-declaring their own (a re-declared list is how the fleet drifts). Pass an\n * explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known\n * harness) to widen beyond these.\n */\nexport const CODING_HARNESSES: readonly HarnessType[] = [\n  'opencode',\n  'claude-code',\n  'codex',\n  'kimi-code',\n]\n\nexport interface ProfileAxisSpec {\n  /** The domain profile to sweep. Its prompt/tools/skills are held fixed; only the\n   *  harness and model vary. `model.default` is the fallback model. */\n  base: AgentProfile\n  /** Harnesses to cross. Default: {@link CODING_HARNESSES}. */\n  harnesses?: readonly HarnessType[]\n  /** Models to cross. Default: `[base.model.default]` — one model, i.e. today's\n   *  single-model behaviour, so omitting this never changes an existing run. */\n  models?: readonly string[]\n  /** Force every (harness, model) pair verbatim, even ones the harness can't run —\n   *  for deliberately testing failure modes. Default (false): SNAP instead — a\n   *  vendor-locked harness runs only the swept models in its family, or its native\n   *  default when it supports none, so no harness is dropped and none gets a\n   *  guaranteed-failing foreign-model cell. */\n  keepIncompatible?: boolean\n}\n\n/** Model sentinel for a vendor-locked harness that supports none of the swept models:\n *  it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness\n *  resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi\n *  model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship\n *  table that would rot as router catalogs change. */\nexport const HARNESS_NATIVE_MODEL = 'default'\n\n/**\n * Expand a base profile across the harness × model matrix into the `AgentProfile[]`\n * that `runProfileMatrix` / `selfImprove` score — the ONE place \"which harnesses ×\n * which models do we evaluate\" lives, so no product hand-rolls its own harness list\n * or column→profile mapping (the pattern that let those copies drift and silently\n * break the harness pivot).\n *\n * Each cell clones `base`, sets the canonical top-level `harness` and `model.default`,\n * and stamps `metadata.harness` + `metadata.harnessModel` for matrix grouping. Both\n * metadata fields are hash-bearing, so every cell gets a distinct `agentProfileId` row\n * and results join back by harness/model via {@link harnessAxisOf} with no\n * hand-recomputed key. A vendor-locked harness snaps to its family's swept models — or\n * its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none — so every\n * requested harness runs; `keepIncompatible` forces every pair verbatim.\n *\n * Omit `harnesses`/`models` to sweep the full default set — the \"turn it on for\n * everything we care about\" switch, identical in shape whether one harness or all.\n */\nexport function expandProfileAxes(spec: ProfileAxisSpec): AgentProfile[] {\n  const harnesses = spec.harnesses ?? CODING_HARNESSES\n  if (harnesses.length === 0) throw new ValidationError('expandProfileAxes: no harnesses to sweep')\n  const baseModel = spec.base.model?.default\n  const models = spec.models ?? (baseModel ? [baseModel] : [])\n  if (models.length === 0) {\n    throw new ValidationError(\n      'expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied',\n    )\n  }\n  const out: AgentProfile[] = []\n  const seen = new Set<string>()\n  for (const harness of harnesses) {\n    // A universal (router-backed) harness — opencode/pi/claudish — runs every swept\n    // model. A vendor-locked harness — codex/claude-code/kimi-code — runs only the\n    // swept models in its own family; when it supports NONE of them it snaps to its\n    // native default (the `HARNESS_NATIVE_MODEL` sentinel it resolves at runtime)\n    // rather than being dropped, so every requested harness still appears in the\n    // sweep on a model it can actually run — e.g. sweeping `deepseek/x` puts opencode\n    // on deepseek and kimi-code on its own Kimi model, a real head-to-head.\n    // `keepIncompatible` forces every (harness, model) pair verbatim (failure-mode runs).\n    const supported = spec.keepIncompatible\n      ? models\n      : models.filter((model) => harnessSupportsModel(harness, model))\n    const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL]\n    for (const model of effective) {\n      const profile: AgentProfile = {\n        ...spec.base,\n        harness,\n        name: `${spec.base.name ?? 'agent'}/${harness}/${model}`,\n        model: { ...spec.base.model, default: model },\n        metadata: { ...(spec.base.metadata ?? {}), harness, harnessModel: model },\n      }\n      const id = agentProfileId(profile)\n      if (seen.has(id)) continue\n      seen.add(id)\n      out.push(profile)\n    }\n  }\n  if (out.length === 0) {\n    // Unreachable in normal use — snapping guarantees ≥1 cell per harness — but keep a\n    // fail-closed guard so a future refactor can't silently produce an empty sweep.\n    throw new ValidationError(\n      `expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(', ')}], models=[${models.join(', ')}]).`,\n    )\n  }\n  return out\n}\n\n/**\n * Read the (harness, model) a matrix cell ran under, off a profile or a result row's\n * profile — the join-back for a `byHarness` pivot. Returns undefined when the profile\n * wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by\n * this instead of recomputing an id (recomputing the wrong key is what broke the pivot\n * in the hand-rolled copies).\n */\nexport function harnessAxisOf(\n  profile: Pick<AgentProfile, 'metadata'>,\n): { harness: HarnessType; model: string } | undefined {\n  const m = profile.metadata as Record<string, unknown> | undefined\n  const harness = m?.harness\n  const model = m?.harnessModel\n  if (typeof harness === 'string' && typeof model === 'string') {\n    return { harness: harness as HarnessType, model }\n  }\n  return undefined\n}\n\n/**\n * Collision-resistant, path-safe, human-readable profile id for eval artifacts.\n * Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix\n * keys, and directory names where two profiles must not collapse onto one row.\n * The suffix is the first 64 bits of the behaviour hash, enough for ordinary\n * eval matrices while keeping filenames readable.\n */\nexport function agentProfileId(profile: AgentProfile): string {\n  const label = pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? 'profile'\n  return `${label}-${agentProfileHash(profile).slice(0, 16)}`\n}\n\n/**\n * Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete\n * model id because run records reject bare/missing model aliases.\n */\nexport function agentProfileModelId(profile: AgentProfile): string {\n  const model = profile.model?.default?.trim()\n  if (!model) {\n    const label = agentProfileDisplayLabel(profile) ?? 'unnamed profile'\n    throw new ValidationError(\n      `AgentProfile \"${label}\" has no model.default — cannot record eval run`,\n    )\n  }\n  return model\n}\n\nfunction agentProfileDisplayLabel(profile: AgentProfile): string | undefined {\n  return profile.name?.trim() || profile.version?.trim() || undefined\n}\n\nfunction pathSafeProfileLabel(label: string | undefined): string | undefined {\n  const safe = label\n    ?.trim()\n    .replace(/[^A-Za-z0-9._-]+/g, '-')\n    .replace(/-+/g, '-')\n    .replace(/^-|-$/g, '')\n  return safe || undefined\n}\n\nfunction compact<T extends Record<string, unknown>>(input: T): Partial<T> {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as Partial<T>\n}\n\n/**\n * Deterministic behaviour identity for the canonical\n * `@tangle-network/agent-interface` AgentProfile.\n *\n * `name` and `description` are labels and do not affect the hash. Profile\n * `version`, prompt, model hints, tools, resources, hooks, modes, permissions,\n * and extensions do affect the hash. Resource array order is hash-bearing\n * because mount order can change agent behaviour. Undefined fields are treated\n * as absent; explicit `null` fields remain hash-bearing.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  const model = agentProfileModelId(profile)\n  const behaviour = {\n    ...profile,\n    name: undefined,\n    description: undefined,\n    tags: profile.tags ? [...profile.tags].sort() : undefined,\n    model: compact({ ...profile.model, default: model }),\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n","/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n */\n\nimport { randomUUID } from 'node:crypto'\nimport type { ChatClient, ChatRequest } from './analyst/chat-client'\nimport type { Artifact } from './artifact-validator'\nimport { CostLedger, type CostLedgerHandle } from './cost-ledger'\nimport { assertServedModel, ModelSubstitutionError } from './integrity/served-model'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport {\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport { packageVersion } from './package-version'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport {\n  certificationEvidenceDigest,\n  type DefaultVerdict,\n  type VerdictCertification,\n} from './verdict'\nimport type { CheckerIdentity, VerificationStrategySource } from './verification-strategy'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `Artifact` shape. */\n  artifacts: Artifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, when the matched item carries no content\n   * to assess, or when the correctness check itself failed (`unmeasured`).\n   */\n  correct: boolean | null\n  /** structurallyPresent && !unmeasured && correct !== false. */\n  satisfied: boolean\n  /**\n   * Set when the correctness check itself errored (LLM call failure or an\n   * unparseable response after retry). The requirement's fulfilment is\n   * UNKNOWN — `correct` stays null, `satisfied` is false, and\n   * `completionVerdict` excludes the row from `completionRate`'s\n   * denominator. Never folded into a zero: a synthetic zero is\n   * indistinguishable from a real failure (see `JudgeParseError`).\n   */\n  unmeasured?: true\n  /** Why the correctness check could not be measured (present iff `unmeasured`). */\n  unmeasuredReason?: string\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\n/** Extends the substrate verdict spine: `valid` = `fullyComplete` and\n *  `score` = `completionRate` — derived in `completionVerdict()`, the one\n *  place those equalities hold by construction. */\nexport interface CompletionVerdict extends DefaultVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */\n  completionRate: number\n  /** Every measurable requirement satisfied (false when anything is unmeasured). */\n  fullyComplete: boolean\n  /** Requirements whose correctness check errored — reported, never scored as zero. */\n  unmeasuredCount: number\n}\n\n/**\n * Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n */\nexport function completionVerdict(input: {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** What certified the correctness stage, when anything did. Omitted =\n   *  an uncertified verdict — the honest default for a bare checker. */\n  certification?: VerdictCertification\n}): CompletionVerdict {\n  if (input.requirements.length === 0) {\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n    )\n  }\n  const measurable = input.requirements.filter((r) => !r.unmeasured)\n  const unmeasuredCount = input.requirements.length - measurable.length\n  if (measurable.length === 0) {\n    // Every check errored: this is an infrastructure failure, not a scored\n    // run. A 0-rate verdict here would be a fabricated measurement.\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n    )\n  }\n  const satisfiedCount = measurable.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / measurable.length\n  // A run with unmeasured rows can still report a rate over what WAS\n  // measured, but must not claim full completion.\n  const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    unmeasuredCount,\n    valid: fullyComplete,\n    score: completionRate,\n    ...(input.certification === undefined ? {} : { certification: input.certification }),\n  }\n}\n\n/**\n * What a correctness checker declares about itself so the completion\n * verdict can carry a certification: which strategy member it discharges,\n * its exact identity, and the steps its answers rest on unverified.\n */\nexport interface CorrectnessCheckerAttestation {\n  strategy: VerificationStrategySource\n  checker: CheckerIdentity\n  assumptions: string[]\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n *\n * `attestation` is optional metadata on the function value: a checker that\n * carries one yields certified completion verdicts; a bare function yields\n * the same verdict uncertified. A plain arrow function remains a valid\n * checker.\n */\nexport interface CorrectnessChecker {\n  (\n    requirement: CompletionRequirement,\n    content: string,\n  ): Promise<{ correct: boolean; reason: string }>\n  attestation?: CorrectnessCheckerAttestation\n}\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/** artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n  'generated',\n  'generate',\n  'view',\n  'render',\n  'rendered',\n  'persisted',\n  'persist',\n  'artifact',\n  'file',\n  'document',\n  'note',\n  'proposal',\n  'deliverable',\n  'output',\n  'created',\n  'create',\n  'produce',\n  'produced',\n  'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    // Match against the artifact CONTENT too, not just its path + kind — a\n    // generated view / note whose path is generic still satisfies a requirement\n    // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n    // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n    let score = tokenRecall(\n      reqText,\n      `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n    )\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    // A proposal needs an assessable BODY to be a deliverable. A bare title is\n    // not completion: correctness cannot be judged on it, so a title-only match\n    // would auto-pass the oracle (structurallyPresent && correct===null →\n    // satisfied) with no verifiable content. Tool calls are the only\n    // legitimately content-less deliverable (`toolCallCandidates`).\n    const body = (p.content ?? '').trim()\n    if (body.length < MIN_CONTENT_CHARS) continue\n    // Match against the body as well as the (often short) title — a refusal /\n    // flag / analysis proposal whose title is a label still satisfies a\n    // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n    // + the requirement's distinctive tokens keep an off-topic proposal out;\n    // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n    // polarity/fulfilment, so a negation that merely contains the tokens fails.\n    // Structural and correctness must use different evidence or the two-stage\n    // check collapses to one lexical gate.\n    const score = tokenRecall(reqText, `${p.title} ${body}`)\n    if (score < MATCH_THRESHOLD) continue\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n    let unmeasuredReason: string | undefined\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        try {\n          const r = await checkCorrectness(req, match.content)\n          correct = r.correct\n          evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n        } catch (err) {\n          // The CHECKER failed, not the requirement. Recording this as a\n          // zero would fabricate a model failure out of an infrastructure\n          // one; the requirement is unmeasured and leaves the denominator.\n          unmeasuredReason =\n            err instanceof JudgeParseError\n              ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n              : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n          evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n        }\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const unmeasured = unmeasuredReason !== undefined\n    const satisfied = structurallyPresent && !unmeasured && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n      evidence,\n    })\n  }\n\n  // The verdict blends the deterministic structural stage with the injected\n  // correctness stage, so the certification names the checker's own member\n  // and carries the structural stage's lexical nature as an assumption. A\n  // checker without an attestation yields an uncertified verdict.\n  const attestation = checkCorrectness.attestation\n  const certification: VerdictCertification | undefined = attestation\n    ? {\n        strategy: attestation.strategy,\n        checker: attestation.checker,\n        assumptions: [\n          'structural matching is lexical token recall over produced items — the correctness stage only sees items it matched',\n          ...attestation.assumptions,\n        ],\n        evidenceDigest: certificationEvidenceDigest({ taskId: gold.taskId, requirements }),\n      }\n    : undefined\n\n  return completionVerdict({\n    taskId: gold.taskId,\n    requirements,\n    ...(certification === undefined ? {} : { certification }),\n  })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Optional ledger for direct use. */\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n  /**\n   * Checker LLM calls per requirement before giving up (parse failures and\n   * call errors both consume attempts). The failure then surfaces as an\n   * `unmeasured` requirement, never a zero.\n   */\n  maxAttempts?: number\n  /**\n   * Forensic capture of every checker request/response/error — without it a\n   * checker failure is unauditable (the agent-turn raws never contain the\n   * checker's own calls). Same sink contract as `LlmClient`.\n   */\n  rawSink?: RawProviderSink\n}\n\n/**\n * Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const readVerdict = (candidate: unknown): { correct: boolean; reason: string } | null => {\n    if (candidate === null || typeof candidate !== 'object') return null\n    const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n    if (typeof correct !== 'boolean') return null\n    return { correct, reason: typeof reason === 'string' ? reason : '' }\n  }\n\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (match) {\n    try {\n      const strict = readVerdict(JSON.parse(match[0]))\n      if (strict) return strict\n    } catch {\n      // fall through to truncation recovery\n    }\n  }\n  // The strict path needs a closing `}`; a cap-hit response has none. Take\n  // everything from the first `{` and auto-close it.\n  const start = raw.indexOf('{')\n  if (start !== -1) {\n    const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n    if (recovered) return recovered\n  }\n  throw new JudgeParseError('correctness-checker', raw)\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  chat: ChatClient,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  const maxAttempts = opts.maxAttempts ?? 2\n  const costLedger = opts.costLedger ?? new CostLedger()\n  const sink = opts.rawSink\n  const record = async (event: RawProviderEvent): Promise<void> => {\n    // Forensic capture is best-effort; the verdict is the system of record.\n    try {\n      await sink?.record(event)\n    } catch {\n      // Intentionally swallowed.\n    }\n  }\n  const checker: CorrectnessChecker = async (requirement, content) => {\n    const request = {\n      model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user' as const,\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    } satisfies ChatRequest\n    let lastErr: unknown\n    for (let attempt = 0; attempt < maxAttempts; attempt++) {\n      const started = Date.now()\n      await record({\n        eventId: randomUUID(),\n        provider: chat.transport,\n        model,\n        endpoint: '/chat',\n        baseUrl: '',\n        attemptIndex: attempt,\n        direction: 'request',\n        timestamp: started,\n        requestBody: request,\n        redactedFields: [],\n      })\n      try {\n        const paid = await costLedger.runPaidCall({\n          channel: 'verifier',\n          phase: opts.costPhase ?? 'completion.correctness',\n          actor: 'correctness-checker',\n          model,\n          maximumCharge:\n            chat.maximumAttempts === undefined\n              ? undefined\n              : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),\n          tags: {\n            ...opts.costTags,\n            requirementId: requirement.reqId,\n            attempt: String(attempt),\n          },\n          signal: opts.signal,\n          execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),\n          receipt: costReceiptFromLlm,\n          receiptFromError: costReceiptFromLlmError,\n        })\n        if (!paid.succeeded) throw paid.error\n        const resp = paid.value\n        // Hold the transport to its own word: a verdict produced by a model\n        // other than the one requested is not this model's verdict. A\n        // transport that echoes no id cannot be made to prove identity here —\n        // callers needing that proof enable it at the client\n        // (`LlmClientOptions.assertServedModel`) or gate on `assertModelsServed`.\n        assertServedModel(model, resp.servedModel, {\n          allowUnreported: true,\n          context: `correctness checker for requirement ${requirement.reqId}`,\n        })\n        const raw = resp.content\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'response',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          responseBody: resp,\n          redactedFields: [],\n        })\n        return parseCorrectnessResponse(raw)\n      } catch (err) {\n        lastErr = err\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'error',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          errorMessage: err instanceof Error ? err.message : String(err),\n          redactedFields: [],\n        })\n        // Substitution is a verdict about identity, not a transient fault:\n        // another attempt reaches the same wrong model and spends more tokens.\n        if (err instanceof ModelSubstitutionError) throw err\n      }\n    }\n    throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n  }\n  checker.attestation = {\n    strategy: 'judge',\n    checker: { name: 'llm-correctness-checker', version: model },\n    assumptions: [\n      'served-model identity is accepted when the transport reports none (assertServedModel allowUnreported)',\n    ],\n  }\n  return checker\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n *\n * Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  const checker: CorrectnessChecker = async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n  // 'schema': token recall checks shape, not meaning — the member whose\n  // documented failure mode (\"a well-formed wrong answer passes\") is\n  // exactly this checker's polarity blindness.\n  checker.attestation = {\n    strategy: 'schema',\n    checker: { name: 'token-recall-checker', version: packageVersion() },\n    assumptions: ['polarity-blind: a negation that recalls the requirement tokens passes'],\n  }\n  return checker\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n  // body of the proposal (e.g. a submit_proposal `description`). When present,\n  // the completion oracle correctness-checks it like artifact content; absent,\n  // the proposal is graded presence-only.\n  content?: string\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({\n        id: p.proposalId,\n        title: p.title,\n        status: p.status ?? 'pending',\n        ...(p.content !== undefined ? { content: p.content } : {}),\n      })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAeA,MAAa,mBAA2C;CACtD;CACA;CACA;CACA;AACF;;;;;;AAwBA,MAAa,uBAAuB;;;;;;;;;;;;;;;;;;;AAoBpC,SAAgB,kBAAkB,MAAuC;CACvE,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,IAAI,gBAAgB,0CAA0C;CAChG,MAAM,YAAY,KAAK,KAAK,OAAO;CACnC,MAAM,SAAS,KAAK,WAAW,YAAY,CAAC,SAAS,IAAI,CAAC;CAC1D,IAAI,OAAO,WAAW,GACpB,MAAM,IAAI,gBACR,kGACF;CAEF,MAAM,MAAsB,CAAC;CAC7B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,WAAW,WAAW;EAS/B,MAAM,YAAY,KAAK,mBACnB,SACA,OAAO,QAAQ,UAAU,qBAAqB,SAAS,KAAK,CAAC;EACjE,MAAM,YAAY,UAAU,SAAS,IAAI,YAAY,CAAC,oBAAoB;EAC1E,KAAK,MAAM,SAAS,WAAW;GAC7B,MAAM,UAAwB;IAC5B,GAAG,KAAK;IACR;IACA,MAAM,GAAG,KAAK,KAAK,QAAQ,QAAQ,GAAG,QAAQ,GAAG;IACjD,OAAO;KAAE,GAAG,KAAK,KAAK;KAAO,SAAS;IAAM;IAC5C,UAAU;KAAE,GAAI,KAAK,KAAK,YAAY,CAAC;KAAI;KAAS,cAAc;IAAM;GAC1E;GACA,MAAM,KAAK,eAAe,OAAO;GACjC,IAAI,KAAK,IAAI,EAAE,GAAG;GAClB,KAAK,IAAI,EAAE;GACX,IAAI,KAAK,OAAO;EAClB;CACF;CACA,IAAI,IAAI,WAAW,GAGjB,MAAM,IAAI,gBACR,uDAAuD,UAAU,KAAK,IAAI,EAAE,aAAa,OAAO,KAAK,IAAI,EAAE,IAC7G;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,cACd,SACqD;CACrD,MAAM,IAAI,QAAQ;CAClB,MAAM,UAAU,GAAG;CACnB,MAAM,QAAQ,GAAG;CACjB,IAAI,OAAO,YAAY,YAAY,OAAO,UAAU,UAClD,OAAO;EAAW;EAAwB;CAAM;AAGpD;;;;;;;;AASA,SAAgB,eAAe,SAA+B;CAE5D,OAAO,GADO,qBAAqB,yBAAyB,OAAO,CAAC,KAAK,UACzD,GAAG,iBAAiB,OAAO,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1D;;;;;AAMA,SAAgB,oBAAoB,SAA+B;CACjE,MAAM,QAAQ,QAAQ,OAAO,SAAS,KAAK;CAC3C,IAAI,CAAC,OAEH,MAAM,IAAI,gBACR,iBAFY,yBAAyB,OAAO,KAAK,kBAE1B,gDACzB;CAEF,OAAO;AACT;AAEA,SAAS,yBAAyB,SAA2C;CAC3E,OAAO,QAAQ,MAAM,KAAK,KAAK,QAAQ,SAAS,KAAK,KAAK,KAAA;AAC5D;AAEA,SAAS,qBAAqB,OAA+C;CAM3E,OALa,OACT,KAAK,CAAC,CACP,QAAQ,qBAAqB,GAAG,CAAC,CACjC,QAAQ,OAAO,GAAG,CAAC,CACnB,QAAQ,UAAU,EAAE,KACR,KAAA;AACjB;AAEA,SAAS,QAA2C,OAAsB;CACxE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,iBAAiB,SAA+B;CAC9D,MAAM,QAAQ,oBAAoB,OAAO;CACzC,MAAM,YAAY;EAChB,GAAG;EACH,MAAM,KAAA;EACN,aAAa,KAAA;EACb,MAAM,QAAQ,OAAO,CAAC,GAAG,QAAQ,IAAI,CAAC,CAAC,KAAK,IAAI,KAAA;EAChD,OAAO,QAAQ;GAAE,GAAG,QAAQ;GAAO,SAAS;EAAM,CAAC;CACrD;CACA,OAAO,WAAW,QAAQ,CAAC,CACxB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,CAAC,CAC/C,OAAO,KAAK;AACjB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACxEA,SAAgB,kBAAkB,OAMZ;CACpB,IAAI,MAAM,aAAa,WAAW,GAChC,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,+DAC3C;CAEF,MAAM,aAAa,MAAM,aAAa,QAAQ,MAAM,CAAC,EAAE,UAAU;CACjE,MAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;CAC/D,IAAI,WAAW,WAAW,GAGxB,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,yCAAyC,MAAM,aAAa,OAAO,8BAA8B,MAAM,aAAa,EAAE,EAAE,oBAAoB,iBAAiB,EACxM;CAEF,MAAM,iBAAiB,WAAW,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC;CAC7D,MAAM,iBAAiB,iBAAiB,WAAW;CAGnD,MAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;CAC7E,OAAO;EACL,QAAQ,MAAM;EACd,cAAc,MAAM;EACpB;EACA;EACA;EACA,OAAO;EACP,OAAO;EACP,GAAI,MAAM,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,MAAM,cAAc;CACpF;AACF;AA+BA,MAAM,4BAAY,IAAI,IAAI;CACxB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,MAAM,6CAA6B,IAAI,IAAI;CACzC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;AACxB,MAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;CAC/D,OAAO,IAAI,IACT,EACG,YAAY,CAAC,CACb,MAAM,YAAY,CAAC,CACnB,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAC1E;AACF;;;;;;;AAQA,SAAS,YAAY,iBAAyB,eAA+B;CAC3E,MAAM,MAAM,OAAO,iBAAiB,0BAA0B;CAC9D,IAAI,IAAI,SAAS,GAAG,OAAO;CAC3B,MAAM,OAAO,OAAO,aAAa;CACjC,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK,IAAI,KAAK,IAAI,CAAC,GAAG;CACtC,OAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,GAAG,MAAM;EAC1B,KAAK,EAAE,WAAW,GAAA,CAAI,KAAK,CAAC,CAAC,SAAS,mBAAmB;EAKzD,IAAI,QAAQ,YACV,SACA,GAAG,EAAE,QAAQ,GAAG,GAAG,EAAE,KAAK,IAAI,EAAE,WAAW,GAAA,CAAI,MAAM,GAAG,GAAI,GAC9D;EACA,IAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAC9E,QAAQ,KAAK,IAAI,OAAO,CAAC;EAE3B,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY;GACrB;GACA,UAAU,aAAa,EAAE,QAAQ,EAAE,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,CAAC;CACD,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,WAAW;EAEzB,IAAI,EAAE,WAAW,YAAY;EAM7B,MAAM,QAAQ,EAAE,WAAW,GAAA,CAAI,KAAK;EACpC,IAAI,KAAK,SAAS,mBAAmB;EASrC,MAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,MAAM,GAAG,MAAM;EACvD,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY,EAAE;GACvB;GACA,UAAU,sBAAsB,EAAE,MAAM,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS;EACX,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,MAAM,MAAM;EAC7B,MAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;EACzC,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,QAAQ;GACjB;GACA,UAAU,cAAc,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACxE,SAAS;EACX,CAAC;CACH,CAAC;CACD,OAAO;AACT;;;;;;;;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;CAC5B,IAAI,KAAK,aAAa,WAAW,GAC/B,MAAM,IAAI,MACR,2BAA2B,KAAK,OAAO,4CACzC;CAMF,MAAM,aAA0B,CAAC;CACjC,KAAK,aAAa,SAAS,KAAK,MAAM;EACpC,MAAM,KAAK,IAAI,eAAe;EAC9B,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,eAAe,OAAO,OAC/B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;CAElE,CAAC;CACD,WAAW,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;CAE3C,MAAM,2BAAW,IAAI,IAAuB;CAC5C,MAAM,4BAAY,IAAI,IAAY;CAClC,KAAK,MAAM,KAAK,YAAY;EAC1B,IAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,GAAG;EAC1D,SAAS,IAAI,EAAE,UAAU,CAAC;EAC1B,UAAU,IAAI,EAAE,OAAO;CACzB;CAEA,MAAM,eAAmC,CAAC;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;EACjD,MAAM,MAAM,KAAK,aAAa;EAC9B,MAAM,QAAQ,SAAS,IAAI,CAAC;EAC5B,MAAM,WAAqB,CAAC;EAC5B,IAAI,UAA0B;EAC9B,IAAI;EAEJ,IAAI,OAAO;GACT,SAAS,KAAK,MAAM,QAAQ;GAC5B,IAAI,MAAM,YAAY,MACpB,IAAI;IACF,MAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;IACnD,UAAU,EAAE;IACZ,SAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,OAAO,KAAK,EAAE,QAAQ;GAC3E,SAAS,KAAK;IAIZ,mBACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,MACjE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7E,SAAS,KAAK,6BAA6B,kBAAkB;GAC/D;QAEA,SAAS,KAAK,6DAA6D;EAE/E,OAAO;GACL,MAAM,KAAK,IAAI,eAAe;GAC9B,MAAM,OAAO,OAAO,QAAQ,gCAAgC;GAC5D,SAAS,KAAK,eAAe,KAAK,0BAA0B;EAC9D;EAEA,MAAM,sBAAsB,UAAU,KAAA;EACtC,MAAM,aAAa,qBAAqB,KAAA;EACxC,MAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;EACpE,aAAa,KAAK;GAChB,OAAO,IAAI;GACX,OAAO,IAAI;GACX;GACA;GACA;GACA,GAAI,aAAa;IAAE,YAAY;IAAe;GAAiB,IAAI,CAAC;GACpE;EACF,CAAC;CACH;CAMA,MAAM,cAAc,iBAAiB;CACrC,MAAM,gBAAkD,cACpD;EACE,UAAU,YAAY;EACtB,SAAS,YAAY;EACrB,aAAa,CACX,sHACA,GAAG,YAAY,WACjB;EACA,gBAAgB,4BAA4B;GAAE,QAAQ,KAAK;GAAQ;EAAa,CAAC;CACnF,IACA,KAAA;CAEJ,OAAO,kBAAkB;EACvB,QAAQ,KAAK;EACb;EACA,GAAI,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,cAAc;CACzD,CAAC;AACH;;;;;;;;AAgCA,SAAgB,yBAAyB,KAAmD;CAC1F,MAAM,eAAe,cAAoE;EACvF,IAAI,cAAc,QAAQ,OAAO,cAAc,UAAU,OAAO;EAChE,MAAM,EAAE,SAAS,WAAW;EAC5B,IAAI,OAAO,YAAY,WAAW,OAAO;EACzC,OAAO;GAAE;GAAS,QAAQ,OAAO,WAAW,WAAW,SAAS;EAAG;CACrE;CAEA,MAAM,QAAQ,IAAI,MAAM,aAAa;CACrC,IAAI,OACF,IAAI;EACF,MAAM,SAAS,YAAY,KAAK,MAAM,MAAM,EAAE,CAAC;EAC/C,IAAI,QAAQ,OAAO;CACrB,QAAQ,CAER;CAIF,MAAM,QAAQ,IAAI,QAAQ,GAAG;CAC7B,IAAI,UAAU,IAAI;EAChB,MAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;EACpE,IAAI,WAAW,OAAO;CACxB;CACA,MAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;;;;;;;AAQA,SAAgB,4BACd,MACA,OAAkC,CAAC,GACf;CACpB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,aAAa,KAAK,cAAc,IAAI,WAAW;CACrD,MAAM,OAAO,KAAK;CAClB,MAAM,SAAS,OAAO,UAA2C;EAE/D,IAAI;GACF,MAAM,MAAM,OAAO,KAAK;EAC1B,QAAQ,CAER;CACF;CACA,MAAM,UAA8B,OAAO,aAAa,YAAY;EAClE,MAAM,UAAU;GACd;GACA,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IACE,MAAM;IACN,SAAS,gBAAgB,YAAY,MAAM,IACzC,YAAY,WAAW,aAAa,YAAY,SAAS,MAAM,GAChE,wBAAwB,QAAQ,MAAM,GAAG,eAAe;GAC3D,CACF;GACA,aAAa;GACb,WAAW;EACb;EACA,IAAI;EACJ,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;GACtD,MAAM,UAAU,KAAK,IAAI;GACzB,MAAM,OAAO;IACX,SAAS,WAAW;IACpB,UAAU,KAAK;IACf;IACA,UAAU;IACV,SAAS;IACT,cAAc;IACd,WAAW;IACX,WAAW;IACX,aAAa;IACb,gBAAgB,CAAC;GACnB,CAAC;GACD,IAAI;IACF,MAAM,OAAO,MAAM,WAAW,YAAY;KACxC,SAAS;KACT,OAAO,KAAK,aAAa;KACzB,OAAO;KACP;KACA,eACE,KAAK,oBAAoB,KAAA,IACrB,KAAA,IACA,2BAA2B,SAAS,EAAE,iBAAiB,KAAK,gBAAgB,CAAC;KACnF,MAAM;MACJ,GAAG,KAAK;MACR,eAAe,YAAY;MAC3B,SAAS,OAAO,OAAO;KACzB;KACA,QAAQ,KAAK;KACb,UAAU,QAAQ,WAAW,KAAK,KAAK,SAAS;MAAE;MAAQ,gBAAgB;KAAO,CAAC;KAClF,SAAS;KACT,kBAAkB;IACpB,CAAC;IACD,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;IAChC,MAAM,OAAO,KAAK;IAMlB,kBAAkB,OAAO,KAAK,aAAa;KACzC,iBAAiB;KACjB,SAAS,uCAAuC,YAAY;IAC9D,CAAC;IACD,MAAM,MAAM,KAAK;IACjB,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc;KACd,gBAAgB,CAAC;IACnB,CAAC;IACD,OAAO,yBAAyB,GAAG;GACrC,SAAS,KAAK;IACZ,UAAU;IACV,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAGD,IAAI,eAAe,wBAAwB,MAAM;GACnD;EACF;EACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;CACtE;CACA,QAAQ,cAAc;EACpB,UAAU;EACV,SAAS;GAAE,MAAM;GAA2B,SAAS;EAAM;EAC3D,aAAa,CACX,uGACF;CACF;CACA,OAAO;AACT;;;ACpmBA,SAAS,aAAa,UAAsC;CAC1D,IAAI,CAAC,UAAU,OAAO;CACtB,IAAI,SAAS,SAAS,MAAM,GAAG,OAAO;CACtC,IAAI,SAAS,WAAW,OAAO,GAAG,OAAO;CACzC,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,qBAAqB,QAAoD;CACvF,MAAM,YAAwB,CAAC;CAC/B,MAAM,YAAgC,CAAC;CACvC,MAAM,YAAsB,CAAC;CAC7B,MAAM,4BAAY,IAAI,IAAY;CAElC,KAAK,MAAM,MAAM,QACf,IAAI,GAAG,SAAS,aAAa;EAC3B,MAAM,OAAQ,GAAyB;EACvC,IAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;GAChC,UAAU,IAAI,IAAI;GAClB,UAAU,KAAK,IAAI;EACrB;CACF,OAAO,IAAI,GAAG,SAAS,YAAY;EACjC,MAAM,IAAI;EACV,UAAU,KAAK;GACb,MAAM,aAAa,EAAE,QAAQ;GAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;GAC3B,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,OAAO,IAAI,GAAG,SAAS,oBAAoB;EACzC,MAAM,IAAI;EACV,UAAU,KAAK;GACb,IAAI,EAAE;GACN,OAAO,EAAE;GACT,QAAQ,EAAE,UAAU;GACpB,GAAI,EAAE,YAAY,KAAA,IAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;EAC1D,CAAC;CACH;CAGF,OAAO;EAAE;EAAW;EAAW;CAAU;AAC3C"}

package/dist/provenance-oA4-zUqm.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"provenance-oA4-zUqm.d.ts","names":[],"sources":["../src/llm-judge.ts","../src/campaign/auto-pr.ts","../src/campaign/storage.ts","../src/campaign/plan-campaign-run.ts","../src/campaign/run-campaign.ts","../src/campaign/cell-schedule.ts","../src/campaign/cell-cache.ts","../src/campaign/external-optimizer-observations.ts","../src/campaign/gepa-candidate-population.ts","../src/campaign/search-ledger.ts","../src/campaign/search-history-receipt.ts","../src/campaign/presets/compare-optimization-methods.ts","../src/canary.ts","../src/red-team.ts","../src/campaign/presets/run-eval.ts","../src/campaign/presets/run-optimization.ts","../src/campaign/presets/run-improvement-loop.ts","../src/campaign/provenance.ts"],"mappings":";;;;;;;;;;;;;KA6CY,6BAA6B;UAExB,gBAAgB,WAAW,kBAAkB,WAAW;;;;EAIvE,MAAM;;;EAGN,aAAa;;EAEb;;EAEA;EACA;EACA;;;EAGA,UAAU;;;;;EAKV;;EAEA,aAAa,UAAU;;;EAGvB,cAAc;IAAS,UAAU;IAAW,UAAU;;;EAEtD,aAAa;EACb;IAAmB;IAAc,QAAQ,EAAE;;;;;;;;;;;;;iBAoB7B,SAAS,qBAAqB,kBAAkB,WAAW,UACzE,cACA,gBACA,MAAM,gBAAgB,WAAW,aAChC,YAAY,WAAW;;;UClFT,kBAAkB,WAAW,kBAAkB;;EAE9D,QAAQ,eAAe,WAAW;;;EAGlC,MAAM;;;EAGN;;EAEA;EACA;;EAEA;;EAEA;;;EAGA;;EAEA,UAAU;IAAqB;IAAgB;IAAgB;;;UAGhD;EACf;EACA;EACA;EACA;;;;;iBAMc,WAAW,WAAW,kBAAkB,UACtD,SAAS,kBAAkB,WAAW,aACrC;;;;;;;;;;;;;;;;;;UCjCc;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;EAEL,MAAM,cAAc,kBAAkB;;;EAGtC,OAAO,cAAc,iBAAiB;;;;;;;;;iBAUxB,qBAAqB;;;;iBA0CrB,2BAA2B;;iBAmC3B,oBAAoB;EAClC,SAAS;EACT;EACA;;EAEA;IACE;;;UC9Ga;EACf;EACA;EACA;EACA;EACA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB,kBAAkB,UAAU;EAClE,WAAW;EACX,WAAW,WAAW,WAAW;EACjC;EACA,SAAS,YAAY,WAAW;EAChC;EACA;EACA;;EAEA;EACA;;EAEA;EACA,UAAU;;EAEV,aAAa;;EAEb,WAAW,SAAS;;;;;;iBAON,gBAAgB,kBAAkB,UAAU,WAC1D,MAAM,uBAAuB,WAAW,aACvC;;;UCvBc,mBAAmB,kBAAkB,UAAU;EAC9D,WAAW;EACX,UAAU,WAAW,WAAW;;EAEhC,SAAS;;;;;;EAMT;EACA,SAAS,YAAY,WAAW;;EAEhC;;;EAGA;;;EAGA;;;;;;EAMA,cAAc;IAAS,UAAU;IAAW;;;;;;EAK5C;;;;;;EAMA;;;;EAIA,eAAe;EACf;EACA;;EAEA;;;EAGA,aAAa;;EAEb;;EAEA,WAAW,SAAS;;EAEpB;;;;;;;;EAQA;;;;;;;;;;EAUA;;;;;EAKA;;;;EAIA;;;EAGA;;;;EAIA;;;;;;;;;;;EAWA;;EAEA,YAAY;;EAEZ,oBAAoB,gBAAgB,gBAAgB;;;;;;EAMpD,UAAU;;;;;;;;;;;;EAYV,iBAAiB;IACf,UAAU;IACV;IACA;;;;;;;UAQa,2BAA2B;EAC1C;EACA;EACA;EACA;IACE;IACA;IACA;MACE;MACA;MACA;;;EAGJ,MAAM,mBAAmB;EACzB,MAAM;;;;;iBAMc,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,eAAe,WAAW;;;iBC5IrB,cAAc,gBAAgB;;;KCkElC;KASA,UAAU;EAChB;EAAe,MAAM,mBAAmB;;EACxC;EAAgB,QAAQ;;iBAEd,eAAe,WAAW;EACxC,SAAS;EACT;EACA;EACA;IACE,UAAU;;;UCzHG;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;;WAEN,WAAW;;WAEX;WACA;WACA;;WAEA;aACE;aACA;;;UAII;WACN,SAAS;WACT,uBAAuB;;WAEvB,qBAAqB;;;;;;;;;;iBAWhB,yCAAyC;EACvD,SAAS;EACT,UAAU;IACR;;;UCrDa;WACN;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;;UAGM;WACN;WACA;;UAGM;;WAEN;WACA,WAAW;WACX;WACA;;WAEA;;WAEA;WACA,0BAA0B;WAC1B;;UAGM;WACN,SAAS;WACT;WACA;WACA,qBAAqB;;;;;;;;;;iBAWhB,oCAAoC;EAClD,SAAS;EACT,UAAU;IACR;;;cCpBS;KAED,mBAAmB;KAEnB;;;UAYK;EACf;EACA;EACA,QAAQ;EACR;;;;UAKe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,MAAM;EACN,UAAU;;UAGK;;EAEf;EACA;EACA;EACA;EACA,gBAAgB;;KAGN;UAOK;EACf;EACA,QAAQ;EACR,WAAW;;;EAGX;;UAGe;EACf;EACA,MAAM;;UAGS;EACf;;;EAGA;;UAGe;;EAEf,gBAAgB;;EAEhB,OAAO;;EAEP,YAAY;;KAGF;EAEN;EACA;EACA;EACA;;EAGA;EACA;;KAGM;EAEN;EACA;EACA;;EAGA;;EAEA;EACA;;UAGW;EACf,QAAQ;EACR,MAAM;;UAGS;EACf;EACA;;KAGU;EAEN;EACA;EACA,SAAS;;EAGT;EACA;EACA,SAAS;EACT,SAAS;;EAGT;EACA,SAAS;EACT,OAAO;IAAwB;;;KAGzB;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;;;;UAKW;EACf;EACA;EACA;EACA,QAAQ;EACR,UAAU;;UAGF;EACR;EACA;EACA,WAAW;;UAGI,2BAA2B;EAC1C;EACA,MAAM;;UAGS,uCAAuC;EACtD;EACA;EACA;EACA;EACA,SAAS;EACT,UAAU;;UAGK,uCAAuC;EACtD;EACA;EACA;EACA,QAAQ;;UAGO,iCAAiC;EAChD;EACA;EACA;EACA;EACA;IACE;IACA,QAAQ;;EAEV;IACE,OAAO;IACP,OAAO;IACP,WAAW;;EAEb,SAAS;EACT,YAAY;EACZ,iBAAiB;;UAGF,qCAAqC;EACpD;EACA;EACA,eAAe;EACf;IAEM;IACA,OAAO;IACP,QAAQ;;IAGR;IACA,QAAQ;;EAEd;IACM;;IACA;IAAmB,SAAS;;IAC5B;IAAkB,SAAS;;EACjC,YAAY;;UAGG,oCAAoC;EACnD;EACA;EACA;IACM;;IAEA;IACA,QAAQ;;;UAIC,6BAA6B;EAC5C;EACA;IAEM;IACA;;IAGA;IACA,QAAQ;;;KAIJ,oBACR,qBACA,iCACA,iCACA,2BACA,+BACA,8BACA;UAEa;EACf,eAAe;EACf;EACA;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGW;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAY;IAAgB;IAAgB;;EAC5C;IAAqB;IAAmB;IAAiB;;EACzD;IAAa;IAAkB;IAAkB;;EACjD;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;EACA,YAAY;EACZ,UAAU;;UAGK;EACf,SAAS;EACT,MAAM;EACN,YAAY;EACZ,sBAAsB;EACtB,UAAU;EACV,YAAY;EACZ,WAAW;EACX,YAAY;EACZ,OAAO;;UAGQ;EACf,OAAO;;EAEP;EACA,QAAQ;;;;iBA+ZM,0BAA0B,iBAAiB;;;;;;;;;;;;;;;KAsB/C;UAEK;EACf;EACA;EACA,cAAc;;UAGC;WACN;WACA;;WAEA;EACT,OAAO,OAAO,oBAAoB,QAAQ;EAC1C,UAAU,QAAQ;;EAElB,eAAe,QAAQ;;;EAGvB,kBAAkB,QAAQ;;;;;;;EAO1B,oBAAoB,QAAQ;;;;iBAKd,iBAAiB,SAAS,0BAA0B;;cA8BvD,4BAA4B;WAC9B;WACA;WACA;mBACQ;mBACA;EAMjB,YAAY,cAAc,oBAAoB,cAAa;EAerD,UAAU,QAAQ;EAIlB,OAAO,OAAO,oBAAoB,QAAQ;EAU1C,eAAe,QAAQ;EAIvB,kBAAkB,QAAQ;EAI1B,oBAAoB,QAAQ;;;;cC74BvB;cACA;;UAGI;WACN;WACA,UAAU;WACV,QAAQ;WACR;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;;;;;;;;;;UAWM;WACN,sBAAsB;WACtB;WACA,wBAAwB;WACxB,eAAe;;WAEf;;WAEA;WACA,QAAQ;;WAER,aAAa;WACb,SAAS;WACT;WACA;;UAGM;WACN;WACA;;WAEA,QAAQ;;WAER,QAAQ;;KAGP;UAEK;WACN;WACA;WACA;WACA,UAAU;;UAGJ;WACN,QAAQ;WACR;WACA,oBAAoB;;cAGlB,mCAAmC;WACrC;WACA;EAET,YAAY,oBAAoB;;;iBAUlB,2BACd,OAAO,kCACN;;iBA2Ba,2BAA2B,SAAS,uBAAuB;;iBA6C3D,iCACd,SAAS,sBACT,QAAQ;;iBAeM,4BACd,oBACA,SAAS,2CACA,WAAW;;iBAcN,yBACd,oBACA,SAAS,mCACR;;;;KC/JS,6BAA6B,kBAAkB,UAAU,aAAa,KAChF,mBAAmB,WAAW;;UAKf;;EAEf;;EAEA,gBAAgB;EAChB;EACA;;UAGe;EACf;;EAEA;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;;EAEf;;EAEA;;EAEA;EACA;;EAEA;EACA;EACA;;UAGe;;EAEf,QAAQ;;EAER,SAAS;;EAET,UAAU;;EAEV,SAAS;;EAET;;EAEA;EACA;;EAEA;EACA;;EAEA;;EAEA;;;;;;EAMA;EACA;EACA,aAAa;;EAEb,eAAe;;EAEf,0BAA0B;;EAE1B,kBAAkB;;EAElB,oBAAoB;;;UAIL,wBAAwB,kBAAkB,UAAU;;WAE1D,iBAAiB;;WAEjB,yBAAyB;;WAEzB,6BAA6B;;WAE7B,sBACP,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;;WAEJ,iBAAiB,YAAY,WAAW;;WAExC;WACA;;WAEA,YAAY,SAAS,6BAA6B,WAAW;;WAE7D,YAAY;;UAGN;;EAEf,eAAe;;EAEf,MAAM;;EAEN;;EAEA,aAAa;;EAEb,gBAAgB;;;UAID,mBAAmB,kBAAkB,WAAW,UAAU;;EAEzE;EACA,WACE,OAAO,wBAAwB,WAAW,eACvC,QAAQ;;UAGE;EACf;;EAEA;;EAEA;;EAEA;;;EAGA;IAAU;IAAa;;;EAEvB,kBAAkB;;EAElB;;EAEA,aAAa;;EAEb,gBAAgB;IACd;IACA;IACA;IACA;;EAEF,eAAe;;EAEf;;UAGe;;EAEf;EACA;;EAEA;EACA;EACA;;EAEA;;UAGe;;EAEf,QAAQ;EACR,MAAM;;EAEN,UAAU;EACV;;EAEA,kBAAkB;;EAElB,UAAU;;EAEV,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,eAAe;;UAGA,kCAAkC,kBAAkB,UAAU,mBACrE,KAAK,mBAAmB,WAAW;EAC3C,SAAS,mBAAmB,WAAW;EACvC,iBAAiB;;EAEjB,gBAAgB;;EAEhB,oBAAoB;;EAEpB,eAAe;;EAEf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;EACb,QAAQ,YAAY,WAAW;;;EAG/B;;EAEA,yBAAyB,6BAA6B,WAAW;;EAEjE;;;EAGA;;EAEA;;;;;EAKA,sBAAsB;;;;;iBAMF,2BAA2B,kBAAkB,UAAU,WAC3E,MAAM,kCAAkC,WAAW,aAClD,QAAQ;;iBAqxBK,sBAAsB,SAAS,oBAAoB;;iBAYnD,kCACd,SAAS,mBACT,mBAAmB,gBAClB;;iBA0Ba,uBACd,SAAS;EAAgB;EAAe,MAAM;KAC7C;;;KCxkCS;KAEA;UAEK;EACf,MAAM;EACN,UAAU;EACV;;;EAGA,UAAU;;UAGK;EACf,QAAQ;;EAER,QAAQ,OAAO;;EAEf,aAAa;;UAGE;EACf,MAAM;EACN;EACA;EACA;;UAGe;;;;;;;;;;EAUf;IACE;IACA;;IAEA;;;;;;;;;;;;;EAcF;IACE;IACA;IACA;IACA;;;;;;;;;;EAWF;IACE,WAAW,KAAK;IAChB;IACA;IACA;IACA;;;;;;;;iBASY,YAAY,MAAM,aAAa,OAAM,gBAAqB;;;KClG9D;UAUK;EACf,UAAU;;EAEV;;;;;EAKA;;EAEA;;EAEA;;UAGe,oBAAoB;EACnC,SAAS;;UAGM;EACf;EACA,UAAU;EACV;EACA;EACA;;UAGe;EACf,UAAU;EACV,oBAAoB,OAAO;EAC3B;;;cA0DW,yBAAyB;iBA6FtB,eAAe,aAAY,gBAAqB;;;;;iBAkBhD,mBACd,gBACA,qBACA,QAAQ,cACP;;iBA4Fa,cAAc,UAAU,mBAAmB;;;UCzT1C,eAAe,kBAAkB,UAAU,mBAClD,KAAK,mBAAmB,WAAW;EAC3C;;;;;iBAMoB,QAAQ,kBAAkB,UAAU,WACxD,MAAM,eAAe,WAAW,aAC/B,QAAQ,eAAe,WAAW;;;UC0BpB,gCAAgC,WAAW,kBAAkB;;EAE5E;;EAEA,UAAU,eAAe,WAAW;;UAGrB,2BAA2B,kBAAkB,UAAU,mBAC9D,KAAK,mBAAmB,WAAW;;EAE3C,iBAAiB;;;;;;;;;EASjB,sBAAsB,gCAAgC,WAAW;;EAEjE,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;;EAEb,UAAU,gBAAgB;EAC1B;EACA;;;EAGA;;;EAGA;;EAEA,WAAW,cAAc;;;;;;;;;;;EAWzB,qBAAqB;IACnB;IACA;IACA,YAAY;MACV;MACA,UAAU,eAAe,WAAW;MACpC;;IAEF,SAAS;;IAET,aAAa;IACb;QACI,QAAQ,cAAc;;;;;;;;;;;;;;;;;;EAkB5B,oBAAoB,UAAU,eAAe,WAAW;;KAG9C,uBACV,kBAAkB,UAClB,aACE,2BAA2B,WAAW;UAEzB,sBAAsB,WAAW,kBAAkB;EAClE,aAAa;IACX,QAAQ;IACR,UAAU;MACR;MACA,SAAS;MACT,UAAU,eAAe,WAAW;;;;EAIxC,iBAAiB;EACjB,eAAe;EACf;;;;EAIA;;;;EAIA;EACA,kBAAkB,eAAe,WAAW;;EAE5C,MAAM;;;;;;EAMN,gBAAgB;;;;;iBAMI,gBAAgB,kBAAkB,UAAU,WAChE,MAAM,uBAAuB,WAAW,aACvC,QAAQ,sBAAsB,WAAW;;;KCpJhC,0BACV,kBAAkB,UAClB,aACE,uBAAuB,WAAW;;;EAGpC,kBAAkB;;;;;;;;;EASlB;;;;EAIA,MAAM,KAAK,WAAW;;;;;EAKtB;;EAEA;EACA;;;;;;;;EAQA,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;UAGlE,yBAAyB,WAAW,kBAAkB,kBAC7D,sBAAsB,WAAW;EACzC,mBAAmB,eAAe,WAAW;EAC7C,iBAAiB,eAAe,WAAW;EAC3C,uBAAuB,eAAe,WAAW;EACjD,qBAAqB;EACrB,YAAY,QAAQ,WAAW,KAAK,WAAW;;;;EAI/C;;;;;EAKA;EACA,WAAW,kBAAkB;;;;;iBAMT,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,0BAA0B,WAAW,aAC1C,QAAQ,yBAAyB,WAAW;;;UC1B9B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,UAAU,YAAY;;EAEtB;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;EACA;EACA;EACA;;UAGe;EACf;IACE;IACA;;EAEF;IACE;IACA;IACA;IACA;MACE;MACA;MACA;MACA;;;EAGJ;;UAGe;EACf;EACA,MAAM;EACN;EACA,aAAa;;;;;;;UAQE;EACf;;EAEA;EACA;EACA;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;;EAEA,YAAY;;EAEZ,qBAAqB;;EAErB,UAAU;;EAEV;;EAEA;IACE,UAAU;IACV;IACA;IACA,mBAAmB;;;;;;EAMrB;;EAEA;;EAEA;;;EAGA;;EAEA,SAAS;EACT;EACA;;UAGe,wBAAwB,WAAW,kBAAkB;EACpE;EACA;EACA;EACA,iBAAiB;EACjB,eAAe;EACf;EACA;;EAEA,wBAAwB,eAAe,WAAW;;EAElD,aAAa;IACX;IACA,YAAY;IACZ;;;IAGA,UAAU;MACR;MACA,SAAS;MACT,UAAU,eAAe,WAAW;;;EAGxC,MAAM;;;;EAIN;EACA,mBAAmB,eAAe,WAAW;EAC7C,iBAAiB,eAAe,WAAW;EAC3C,qBAAqB;EACrB,uBAAuB,eAAe,WAAW;;EAEjD,cAAc,cAAc;EAC5B;EACA;EACA,qBAAqB;;UAGN,6BAA6B,WAAW,kBAAkB;EACzE;EACA;EACA;EACA,iBAAiB;EACjB,QAAQ,yBAAyB,WAAW;EAC5C,cAAc,cAAc;EAC5B;EACA;;;iBAIc,6BAA6B,WAAW,kBAAkB,UACxE,OAAO,6BAA6B,WAAW,aAC9C,wBAAwB,WAAW;;iBA4CtB,0BAA0B,WAAW,kBAAkB,UACrE,MAAM,wBAAwB,WAAW,aACxC;;iBAoRa,0BAA0B,WAAW,kBAAkB,UACrE,UAAU,eAAe,WAAW;;iBAqCtB,2BAA2B,QAAQ,uBAAuB;;iBAa1D,gBAAgB;;;;;;;;;;;;iBA2KhB,oBACd,QAAQ,sBACR;EAAQ;IACP;;iBAiJa,qBAAqB;;;;iBAMrB,oBAAoB;UAInB;EACf,QAAQ;EACR,OAAO;;EAEP;EACA;;UAGe,uBAAuB,WAAW,kBAAkB,kBAC3D,wBAAwB,WAAW;;EAE3C,SAAS;;;EAGT,eAAe;;;;;;;;;;;;iBA+EK,mBAAmB,WAAW,kBAAkB,UACpE,MAAM,uBAAuB,WAAW,aACvC,QAAQ"}

package/dist/query-CJ_DX8vl.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"query-CJ_DX8vl.d.ts","names":[],"sources":["../src/trace/query.ts"],"mappings":";;;iBAasB,gBAAgB,OAAO,YAAY,qBAAqB,QAAQ;iBAIhE,SAAS,OAAO,YAAY,iBAAiB,QAAQ;iBAKrD,UACpB,OAAO,YACP,gBACA,oBACC,QAAQ;;iBAMW,WAAW,OAAO,YAAY,iBAAiB,QAAQ;;iBAM7D,QAAQ,GAAG,2BAA2B,OAAO,KAAK,MAAM,GAAG,MAAM,IAAI,IAAI,GAAG;;iBAe5E,QAAQ;;iBAKR,oBAAoB,MAAM;;iBAuB1B,aAAa,OAAO;EAClC;EACA;EACA;EACA;EACA;EACA;;;iBAuBc,gBAAgB,KAAK,MAAM"}

package/dist/query-Di7eEQ79.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"query-Di7eEQ79.js","names":[],"sources":["../src/trace/query.ts"],"sourcesContent":["/**\n * Typed query helpers over TraceStore.\n *\n * Not a full SQL engine — a minimal, composable set of operators that\n * cover the canned-pipeline use cases. For ad-hoc analytics, persist to\n * NDJSON and point DuckDB at it; the schema is stable so external SQL\n * tooling works out of the box.\n */\n\nimport type { FailureClass, JudgeSpan, LlmSpan, Run, ToolSpan } from './schema'\nimport { isJudgeSpan, isLlmSpan, isToolSpan } from './schema'\nimport type { TraceStore } from './store'\n\nexport async function runsForScenario(store: TraceStore, scenarioId: string): Promise<Run[]> {\n  return store.listRuns({ scenarioId })\n}\n\nexport async function llmSpans(store: TraceStore, runId?: string): Promise<LlmSpan[]> {\n  const spans = await store.spans({ runId, kind: 'llm' })\n  return spans.filter(isLlmSpan)\n}\n\nexport async function toolSpans(\n  store: TraceStore,\n  runId?: string,\n  toolName?: string,\n): Promise<ToolSpan[]> {\n  const spans = await store.spans({ runId, kind: 'tool', toolName })\n  return spans.filter(isToolSpan)\n}\n\n/** Query judge-kind spans from the trace store, optionally scoped to a single run. */\nexport async function judgeSpans(store: TraceStore, runId?: string): Promise<JudgeSpan[]> {\n  const spans = await store.spans({ runId, kind: 'judge' })\n  return spans.filter(isJudgeSpan)\n}\n\n/** Group spans by any key selector. */\nexport function groupBy<T, K extends string | number>(items: T[], key: (t: T) => K): Map<K, T[]> {\n  const map = new Map<K, T[]>()\n  for (const item of items) {\n    const k = key(item)\n    let bucket = map.get(k)\n    if (!bucket) {\n      bucket = []\n      map.set(k, bucket)\n    }\n    bucket.push(item)\n  }\n  return map\n}\n\n/** Hash tool arguments to an orderless-key-stable string for de-duplication. */\nexport function argHash(args: unknown): string {\n  return stableStringify(args)\n}\n\n/** Whether argument-based comparisons are valid for this tool call. */\nexport function hasCapturedToolArgs(span: ToolSpan): boolean {\n  return span.argsCaptured !== false\n}\n\nfunction stableStringify(value: unknown): string {\n  // Must ALWAYS return a string: JSON.stringify(undefined) — and stringify of\n  // functions/symbols — yields the JS value `undefined`, which would make\n  // argHash return a non-string and silently break the de-dup keys behind\n  // stuck-loop and failure-cluster.\n  if (value === undefined) return 'undefined'\n  if (value === null || typeof value !== 'object') return JSON.stringify(value) ?? 'undefined'\n  if (Array.isArray(value)) return `[${value.map(stableStringify).join(',')}]`\n  // Drop undefined-valued keys to match JSON.stringify semantics, so\n  // `{a:1}` and `{a:1,b:undefined}` hash identically.\n  const obj = value as Record<string, unknown>\n  const keys = Object.keys(obj)\n    .filter((k) => obj[k] !== undefined)\n    .sort()\n  const parts = keys.map((k) => `${JSON.stringify(k)}:${stableStringify(obj[k])}`)\n  return `{${parts.join(',')}}`\n}\n\n/** Sum an LLM-span array into aggregate token + cost. */\nexport function aggregateLlm(spans: LlmSpan[]): {\n  inputTokens: number\n  outputTokens: number\n  cachedTokens: number\n  cacheWriteTokens: number\n  reasoningTokens: number\n  costUsd: number\n} {\n  return spans.reduce(\n    (acc, s) => ({\n      inputTokens: acc.inputTokens + (s.inputTokens ?? 0),\n      outputTokens: acc.outputTokens + (s.outputTokens ?? 0),\n      cachedTokens: acc.cachedTokens + (s.cachedTokens ?? 0),\n      cacheWriteTokens: acc.cacheWriteTokens + (s.cacheWriteTokens ?? 0),\n      reasoningTokens: acc.reasoningTokens + (s.reasoningTokens ?? 0),\n      costUsd: acc.costUsd + (s.costUsd ?? 0),\n    }),\n    {\n      inputTokens: 0,\n      outputTokens: 0,\n      cachedTokens: 0,\n      cacheWriteTokens: 0,\n      reasoningTokens: 0,\n      costUsd: 0,\n    },\n  )\n}\n\n/** Pick the outcome's failure class when present, else derive 'success' from run status. */\nexport function runFailureClass(run: Run): FailureClass {\n  if (run.outcome?.failureClass) return run.outcome.failureClass\n  if (run.status === 'completed' && run.outcome?.pass !== false) return 'success'\n  if (run.status === 'aborted') return 'budget_exceeded'\n  return 'unknown'\n}\n"],"mappings":";;AAaA,eAAsB,gBAAgB,OAAmB,YAAoC;CAC3F,OAAO,MAAM,SAAS,EAAE,WAAW,CAAC;AACtC;AAEA,eAAsB,SAAS,OAAmB,OAAoC;CAEpF,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;CAAM,CAAC,EAAA,CACzC,OAAO,SAAS;AAC/B;AAEA,eAAsB,UACpB,OACA,OACA,UACqB;CAErB,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;EAAQ;CAAS,CAAC,EAAA,CACpD,OAAO,UAAU;AAChC;;AAGA,eAAsB,WAAW,OAAmB,OAAsC;CAExF,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;CAAQ,CAAC,EAAA,CAC3C,OAAO,WAAW;AACjC;;AAGA,SAAgB,QAAsC,OAAY,KAA+B;CAC/F,MAAM,sBAAM,IAAI,IAAY;CAC5B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,IAAI,IAAI,IAAI;EAClB,IAAI,SAAS,IAAI,IAAI,CAAC;EACtB,IAAI,CAAC,QAAQ;GACX,SAAS,CAAC;GACV,IAAI,IAAI,GAAG,MAAM;EACnB;EACA,OAAO,KAAK,IAAI;CAClB;CACA,OAAO;AACT;;AAGA,SAAgB,QAAQ,MAAuB;CAC7C,OAAO,gBAAgB,IAAI;AAC7B;;AAGA,SAAgB,oBAAoB,MAAyB;CAC3D,OAAO,KAAK,iBAAiB;AAC/B;AAEA,SAAS,gBAAgB,OAAwB;CAK/C,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO,KAAK,UAAU,KAAK,KAAK;CACjF,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,IAAI,MAAM,IAAI,eAAe,CAAC,CAAC,KAAK,GAAG,EAAE;CAG1E,MAAM,MAAM;CAKZ,OAAO,IAJM,OAAO,KAAK,GAAG,CAAC,CAC1B,QAAQ,MAAM,IAAI,OAAO,KAAA,CAAS,CAAC,CACnC,KACc,CAAC,CAAC,KAAK,MAAM,GAAG,KAAK,UAAU,CAAC,EAAE,GAAG,gBAAgB,IAAI,EAAE,GAC7D,CAAC,CAAC,KAAK,GAAG,EAAE;AAC7B;;AAGA,SAAgB,aAAa,OAO3B;CACA,OAAO,MAAM,QACV,KAAK,OAAO;EACX,aAAa,IAAI,eAAe,EAAE,eAAe;EACjD,cAAc,IAAI,gBAAgB,EAAE,gBAAgB;EACpD,cAAc,IAAI,gBAAgB,EAAE,gBAAgB;EACpD,kBAAkB,IAAI,oBAAoB,EAAE,oBAAoB;EAChE,iBAAiB,IAAI,mBAAmB,EAAE,mBAAmB;EAC7D,SAAS,IAAI,WAAW,EAAE,WAAW;CACvC,IACA;EACE,aAAa;EACb,cAAc;EACd,cAAc;EACd,kBAAkB;EAClB,iBAAiB;EACjB,SAAS;CACX,CACF;AACF;;AAGA,SAAgB,gBAAgB,KAAwB;CACtD,IAAI,IAAI,SAAS,cAAc,OAAO,IAAI,QAAQ;CAClD,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO,OAAO;CACtE,IAAI,IAAI,WAAW,WAAW,OAAO;CACrC,OAAO;AACT"}

package/dist/release-confidence-4XrqlpFD.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"release-confidence-4XrqlpFD.d.ts","names":[],"sources":["../src/statistics/rank-tests.ts","../src/promotion-gate.ts","../src/release-confidence.ts"],"mappings":";;;;;;KA4BY;;;;;KAMA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;;cAIW;;cAEA;;cAEA;;cAEA;UAEI;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;KCjLS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UAsHc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;;KCtL5C;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB"}

package/dist/release-confidence-BknrpBnO.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"release-confidence-BknrpBnO.js","names":[],"sources":["../src/promotion-gate.ts","../src/release-confidence.ts"],"sourcesContent":["/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n *   - Bootstrap mean baseline vs candidate (1k resamples).\n *   - Compute the delta distribution; pass if the lower CI bound > 0.\n *   - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)\n *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)\n *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n  baselineMean: number\n  candidateMean: number\n  /** candidateMean - baselineMean, point estimate. */\n  delta: number\n  /** Lower bound of the (1 - alpha) CI on the delta. */\n  ciLower: number\n  /** Upper bound of the (1 - alpha) CI on the delta. */\n  ciUpper: number\n  /** Number of bootstrap resamples used. */\n  iterations: number\n  alpha: number\n  verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n  /** Confidence level alpha (default 0.05 → 95% CI). */\n  alpha?: number\n  /** Number of resamples (default 1000). */\n  iterations?: number\n  /**\n   * Minimum total samples (baseline + candidate) below which we always\n   * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n   * Default 6 (combined).\n   */\n  minTotalSamples?: number\n  /** RNG seed for reproducibility. Default: Math.random. */\n  seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n  baseline: number[],\n  candidate: number[],\n  options: BootstrapOptions = {},\n): BootstrapResult {\n  const alpha = options.alpha ?? 0.05\n  const iterations = options.iterations ?? 1000\n  const minTotal = options.minTotalSamples ?? 6\n  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n  const baselineMean = mean(baseline)\n  const candidateMean = mean(candidate)\n  const delta = candidateMean - baselineMean\n\n  if (\n    baseline.length + candidate.length < minTotal ||\n    baseline.length === 0 ||\n    candidate.length === 0\n  ) {\n    return {\n      baselineMean,\n      candidateMean,\n      delta,\n      ciLower: -Infinity,\n      ciUpper: Infinity,\n      iterations: 0,\n      alpha,\n      verdict: 'INCONCLUSIVE',\n    }\n  }\n\n  const deltas: number[] = new Array(iterations)\n  for (let i = 0; i < iterations; i++) {\n    const bResample = resample(baseline, rng)\n    const cResample = resample(candidate, rng)\n    deltas[i] = mean(cResample) - mean(bResample)\n  }\n  deltas.sort((a, b) => a - b)\n  const lowerIdx = Math.floor((alpha / 2) * iterations)\n  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n  const ciLower = deltas[Math.max(0, lowerIdx)]!\n  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n  let verdict: Verdict\n  // A ZERO-WIDTH interval is an absence of evidence, not a certainty. Constant\n  // arms make every resample identical, so pass/fail data promotes on nothing:\n  // [0,0,0] vs [1,1,1] is six samples, clears the `minTotalSamples` floor, and\n  // yields [1, 1] — an \"ADVANCE\" carrying no information about how far the\n  // estimate could be wrong. Same rule as `pairedDeltaTest`, one estimator over.\n  if (!Number.isFinite(ciLower) || !Number.isFinite(ciUpper) || ciLower === ciUpper) {\n    verdict = 'INCONCLUSIVE'\n  } else if (ciLower > 0) verdict = 'ADVANCE'\n  else if (ciUpper < 0) verdict = 'REVERT'\n  else if (delta >= 0) verdict = 'KEEP'\n  else verdict = 'INCONCLUSIVE'\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ciLower,\n    ciUpper,\n    iterations,\n    alpha,\n    verdict,\n  }\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  let s = 0\n  for (const x of xs) s += x\n  return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n  const out = new Array(xs.length)\n  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n  return out\n}\n\n/** Mulberry32 — fast deterministic PRNG. Stable across runs given the same seed. */\nfunction mulberry32(seed: number): () => number {\n  let t = seed >>> 0\n  return () => {\n    t += 0x6d2b79f5\n    let r = t\n    r = Math.imul(r ^ (r >>> 15), r | 1)\n    r ^= r + Math.imul(r ^ (r >>> 7), r | 61)\n    return ((r ^ (r >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n  let h = 2166136261\n  for (const x of [...a, ...b]) {\n    const view = new Float64Array([x])\n    const bytes = new Uint8Array(view.buffer)\n    for (const byte of bytes) {\n      h ^= byte\n      h = Math.imul(h, 16777619)\n    }\n  }\n  return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n  baselineOutputs: TOutput[]\n  candidateOutputs: TOutput[]\n  /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n  judge: (output: TOutput) => Promise<number> | number\n  alpha?: number\n  iterations?: number\n  /** RNG seed for reproducibility. */\n  seed?: number\n  /** Maximum concurrent judge calls. Default 4. */\n  judgeConcurrency?: number\n}\n\n/**\n * Confirm a candidate's win with a stronger judge: score baseline and candidate outputs independently, then bootstrap a CI to verify the lift generalises beyond the inner loop.\n */\nexport async function judgeReplayGate<TOutput>(\n  args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n  const concurrency = args.judgeConcurrency ?? 4\n  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n  const ci = bootstrapCi(baselineScores, candidateScores, {\n    ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n    ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n    ...(args.seed !== undefined ? { seed: args.seed } : {}),\n  })\n  return {\n    ...ci,\n    baselineSamples: baselineScores.length,\n    candidateSamples: candidateScores.length,\n  }\n}\n\nasync function scoreAll<TOutput>(\n  outputs: TOutput[],\n  judge: (output: TOutput) => Promise<number> | number,\n  concurrency: number,\n): Promise<number[]> {\n  const results: number[] = new Array(outputs.length)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= outputs.length) return\n      const v = await judge(outputs[i]!)\n      results[i] = Number.isFinite(v) ? v : 0\n    }\n  }\n  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n  return results\n}\n","/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n *   - Dataset manifests prove corpus/version coverage.\n *   - RunRecord rows prove reproducible search/holdout outcomes.\n *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n *   - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { ValidationError, VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport { isRealnessGated, observedSplitScore } from './rollout/reward'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Severity of an actionable finding attached to a run/trace. */\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\n/** Actionable side-info — a diagnosed finding the loop can act on. */\nexport interface ActionableSideInfo {\n  /** Stable expectation/check id when available. */\n  expectationId?: string\n  /** Human-readable diagnosis of what happened. */\n  message: string\n  severity?: AsiSeverity\n  /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n  evidence?: string\n  /** Prompt/tool/context surface likely responsible. */\n  responsibleSurface?: string\n  /** Suggested fix in natural language. */\n  suggestion?: string\n  /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n  matched?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n  | 'corpus'\n  | 'quality'\n  | 'reliability'\n  | 'generalization'\n  | 'diagnostics'\n  | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n  scenarioId: string\n  candidateId?: string\n  split?: RunSplitTag\n  score?: number\n  ok?: boolean\n  turnCount?: number\n  costUsd?: number\n  durationMs?: number\n  /** Canonical task-failure class. Free-form detail belongs in ASI. */\n  failureClass?: FailureClass\n  asi?: ActionableSideInfo[]\n  metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n  /** Require a Dataset manifest or explicit scenarios. Default true. */\n  requireCorpus?: boolean\n  minScenarioCount?: number\n  minSearchRuns?: number\n  minHoldoutRuns?: number\n  /** Require at least one holdout scenario/run. Default true. */\n  requireHoldout?: boolean\n  minPassRate?: number\n  minMeanScore?: number\n  /** Search mean may exceed holdout mean by at most this much. */\n  maxOverfitGap?: number\n  maxMeanCostUsd?: number\n  maxP95WallMs?: number\n  /** Low-score/failed rows must carry ASI. Default true. */\n  requireAsiForFailures?: boolean\n  /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n  failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n  target: string\n  candidateId?: string\n  baselineId?: string\n  dataset?: DatasetManifest\n  scenarios?: readonly DatasetScenario[]\n  runs?: readonly RunRecord[]\n  traces?: readonly ReleaseTraceEvidence[]\n  gateDecision?: GateDecision | null\n  thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n  name: ReleaseConfidenceAxisName\n  status: ReleaseConfidenceStatus\n  score: number | null\n  detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n  axis: ReleaseConfidenceAxisName\n  severity: 'critical' | 'warning'\n  code: string\n  detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n  scenarioCount: number\n  /** Search rows with a finite search score. */\n  searchRuns: number\n  /** Holdout rows with a finite holdout score. */\n  holdoutRuns: number\n  /** Runs with neither a split-matched score nor an explicit task failure. */\n  unscoredRuns: number\n  /** Run rows, or trace rows when no runs exist, with no classified terminal result. */\n  unclassifiedTerminalRuns: number\n  /** Run rows, or trace rows when no runs exist, that ended unsuccessfully. */\n  terminalFailureRuns: number\n  /** Success fraction when every run or fallback trace row has a classified result. */\n  reliabilityRate: number | null\n  passRate: number | null\n  meanScore: number | null\n  searchMeanScore: number | null\n  holdoutMeanScore: number | null\n  overfitGap: number | null\n  meanCostUsd: number | null\n  p95WallMs: number | null\n  failedRows: number\n  failuresWithAsi: number\n  singleShotTraces: number\n  multiShotTraces: number\n  splitCounts: Record<DatasetSplit, number>\n  domainCounts: Record<string, number>\n  failureClassCounts: Partial<Record<FailureClass, number>>\n  responsibleSurfaceCounts: Record<string, number>\n  /**\n   * Runs excluded from `passRate` because the authenticity gate flagged them as\n   * gamed. Surfaced, never silent: a release whose pass rate is computed over a\n   * shrunken denominator has to say by how much, or the exclusion is just a\n   * different way of hiding the same runs.\n   */\n  realnessGatedRuns: number\n}\n\nexport interface ReleaseConfidenceScorecard {\n  target: string\n  candidateId: string | null\n  baselineId: string | null\n  status: ReleaseConfidenceStatus\n  promote: boolean\n  axes: ReleaseConfidenceAxis[]\n  issues: ReleaseConfidenceIssue[]\n  metrics: ReleaseConfidenceMetrics\n  dataset: DatasetManifest | null\n  gateDecision: GateDecision | null\n  summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n  requireCorpus: true,\n  minScenarioCount: 1,\n  minSearchRuns: 1,\n  minHoldoutRuns: 1,\n  requireHoldout: true,\n  minPassRate: 0.8,\n  minMeanScore: 0.7,\n  maxOverfitGap: 0.15,\n  maxMeanCostUsd: Number.POSITIVE_INFINITY,\n  maxP95WallMs: Number.POSITIVE_INFINITY,\n  requireAsiForFailures: true,\n  failureScoreThreshold: 0.5,\n}\n\nexport function evaluateReleaseConfidence(\n  input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n  const candidateId = input.candidateId ?? null\n  const runs = filterCandidate(\n    (input.runs ?? []).map(validateRunRecord),\n    candidateId,\n    input.baselineId,\n  )\n  const traces = filterTraceCandidate(\n    (input.traces ?? []).map(validateReleaseTraceEvidence),\n    candidateId,\n    input.baselineId,\n  )\n  const scenarios = input.scenarios ?? []\n  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n  const searchScores = scoresFor(runs, 'search')\n  const holdoutScores = scoresFor(runs, 'holdout')\n  const runScores = runs.map(runSplitScore).filter(isFiniteNumber)\n  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n  const scoreUniverse = runs.length > 0 ? runScores : traceScores\n  const qualityRuns = runs.filter((run) => runSplitScore(run) !== undefined)\n  const unscoredRuns = runs.filter(\n    (run) =>\n      runSplitScore(run) === undefined &&\n      !hasExplicitTaskFailure(run) &&\n      !isFailedTerminalOutcome(run.terminalOutcome),\n  ).length\n  // Realness-gated runs are EXCLUDED from the pass rate — numerator AND\n  // denominator — and the count of what was dropped ships beside the rate as\n  // `metrics.realnessGatedRuns`. Counting a gamed run as a pass made faking a\n  // success the cheapest way to improve a release scorecard; scoring it 0\n  // instead would be the other error, silently deflating the rate with a run\n  // the gate says carries no usable verdict at all.\n  const honestRuns = runs.filter((run) => !isRealnessGated(run))\n  const passOutcomes =\n    runs.length > 0\n      ? honestRuns.map((run) => runPassOutcome(run, thresholds.failureScoreThreshold))\n      : traces.map((trace) => tracePassOutcome(trace, thresholds.failureScoreThreshold))\n  const reliabilityRows =\n    runs.length > 0\n      ? runs.map((run) => terminalSuccess(run.terminalOutcome))\n      : traces.map((trace) => trace.ok)\n  const unclassifiedTerminalRuns = reliabilityRows.filter((outcome) => outcome === undefined).length\n  const terminalFailureRuns = reliabilityRows.filter((outcome) => outcome === false).length\n  const reliabilityRate =\n    reliabilityRows.length === 0 || unclassifiedTerminalRuns > 0\n      ? null\n      : (reliabilityRows.length - terminalFailureRuns) / reliabilityRows.length\n  const searchRuns = qualityRuns.filter((r) => r.splitTag === 'search').length\n  const holdoutRuns = qualityRuns.filter((r) => r.splitTag === 'holdout').length\n  const failed = failedRows(runs, traces, thresholds.failureScoreThreshold)\n  const searchMeanScore = meanOrNull(searchScores)\n  const holdoutMeanScore = meanOrNull(holdoutScores)\n  const runCosts = runs.flatMap((run) =>\n    run.costProvenance.kind === 'uncaptured' ? [] : [run.costProvenance.usd],\n  )\n  const traceCosts = traces.map((trace) => trace.costUsd).filter(isFiniteNumber)\n  const meanCostUsd =\n    runs.length > 0\n      ? runCosts.length === runs.length\n        ? meanOrNull(runCosts)\n        : null\n      : meanOrNull(traceCosts)\n  const wallTimes =\n    runs.length > 0\n      ? runs.map((run) => run.wallMs)\n      : traces.map((trace) => trace.durationMs).filter(isFiniteNumber)\n  const metrics: ReleaseConfidenceMetrics = {\n    scenarioCount,\n    searchRuns,\n    holdoutRuns,\n    unscoredRuns,\n    unclassifiedTerminalRuns,\n    terminalFailureRuns,\n    reliabilityRate,\n    passRate: passOutcomeRate(passOutcomes),\n    realnessGatedRuns: runs.length - honestRuns.length,\n    meanScore: meanOrNull(scoreUniverse),\n    searchMeanScore,\n    holdoutMeanScore,\n    overfitGap: diffOrNull(searchMeanScore, holdoutMeanScore),\n    meanCostUsd,\n    p95WallMs: percentileOrNull(wallTimes, 0.95),\n    failedRows: failed.length,\n    failuresWithAsi: failed.filter((row) => row.hasAsi).length,\n    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n    splitCounts,\n    domainCounts: countDomains(scenarios),\n    failureClassCounts: countFailureClasses(runs, traces, thresholds.failureScoreThreshold),\n    responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n  }\n\n  const issues: ReleaseConfidenceIssue[] = []\n  checkCorpus(input, thresholds, metrics, issues)\n  checkQuality(thresholds, metrics, issues)\n  checkReliability(metrics, issues)\n  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n  checkDiagnostics(thresholds, metrics, issues)\n  checkEfficiency(thresholds, metrics, issues)\n\n  const axes = buildAxes(metrics, thresholds, issues)\n  const status = issues.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : issues.length > 0\n      ? 'warn'\n      : 'pass'\n\n  return {\n    target: input.target,\n    candidateId,\n    baselineId: input.baselineId ?? null,\n    status,\n    promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n    axes,\n    issues,\n    metrics,\n    dataset: input.dataset ?? null,\n    gateDecision: input.gateDecision ?? null,\n    summary: renderSummary(input.target, status, metrics, issues),\n  }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n  const scorecard = evaluateReleaseConfidence(input)\n  if (scorecard.status === 'fail') {\n    throw new VerificationError(scorecard.summary)\n  }\n  return scorecard\n}\n\nfunction filterCandidate(\n  runs: readonly RunRecord[],\n  candidateId: string | null,\n  baselineId?: string,\n): RunRecord[] {\n  if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n  return [...runs]\n}\n\nfunction filterTraceCandidate(\n  traces: readonly ReleaseTraceEvidence[],\n  candidateId: string | null,\n  baselineId?: string,\n): ReleaseTraceEvidence[] {\n  if (candidateId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n  if (baselineId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n  return [...traces]\n}\n\nfunction validateReleaseTraceEvidence(\n  trace: ReleaseTraceEvidence,\n  index: number,\n): ReleaseTraceEvidence {\n  const value = trace as unknown as Record<string, unknown>\n  if (Object.hasOwn(value, 'failureMode')) {\n    throw new ValidationError(\n      `traces[${index}].failureMode is not supported; use canonical failureClass`,\n    )\n  }\n  if (trace.failureClass !== undefined && !FAILURE_CLASSES.includes(trace.failureClass)) {\n    throw new ValidationError(\n      `traces[${index}].failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n    )\n  }\n  return trace\n}\n\nfunction checkCorpus(\n  input: ReleaseConfidenceInput,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_corpus',\n      detail: 'No Dataset manifest or scenarios supplied.',\n    })\n  }\n  if (metrics.scenarioCount < thresholds.minScenarioCount) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'few_scenarios',\n      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n    })\n  }\n  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_holdout_split',\n      detail: 'Corpus has no holdout scenarios.',\n    })\n  }\n}\n\nfunction checkQuality(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.searchRuns < thresholds.minSearchRuns) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'few_search_runs',\n      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n    })\n  }\n  if (metrics.unscoredRuns > 0) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'unscored_runs',\n      detail: `${metrics.unscoredRuns} supplied run(s) have no task result.`,\n    })\n  }\n  if (metrics.passRate === null || metrics.meanScore === null) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'missing_quality_scores',\n      detail: 'No task-quality scores are available for pass-rate and mean-score checks.',\n    })\n  }\n  if (metrics.passRate !== null && metrics.passRate < thresholds.minPassRate) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_pass_rate',\n      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n    })\n  }\n  if (metrics.meanScore !== null && metrics.meanScore < thresholds.minMeanScore) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_mean_score',\n      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n    })\n  }\n}\n\nfunction checkReliability(\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.reliabilityRate === null) {\n    issues.push({\n      axis: 'reliability',\n      severity: 'critical',\n      code: 'missing_reliability_evidence',\n      detail:\n        metrics.unclassifiedTerminalRuns > 0\n          ? `${metrics.unclassifiedTerminalRuns} supplied run(s) have no classified terminal result.`\n          : 'No classified terminal results are available.',\n    })\n  }\n  if (metrics.terminalFailureRuns > 0) {\n    issues.push({\n      axis: 'reliability',\n      severity: 'critical',\n      code: 'terminal_run_failures',\n      detail: `${metrics.terminalFailureRuns} run(s) ended failed, cancelled, or incomplete.`,\n    })\n  }\n}\n\nfunction checkGeneralization(\n  gateDecision: GateDecision | null,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'few_holdout_runs',\n      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n    })\n  }\n  if (metrics.overfitGap !== null && metrics.overfitGap > thresholds.maxOverfitGap) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'overfit_gap',\n      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n    })\n  }\n  if (gateDecision && !gateDecision.promote) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n      detail: gateDecision.reason,\n    })\n  }\n}\n\nfunction checkDiagnostics(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (!thresholds.requireAsiForFailures) return\n  if (metrics.failedRows > metrics.failuresWithAsi) {\n    issues.push({\n      axis: 'diagnostics',\n      severity: 'critical',\n      code: 'missing_failure_asi',\n      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n    })\n  }\n}\n\nfunction checkEfficiency(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (Number.isFinite(thresholds.maxMeanCostUsd) && metrics.meanCostUsd === null) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'missing_cost',\n      detail: 'A finite cost limit was configured but no cost evidence is available.',\n    })\n  } else if (metrics.meanCostUsd !== null && metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'cost_budget',\n      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n    })\n  }\n  if (Number.isFinite(thresholds.maxP95WallMs) && metrics.p95WallMs === null) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'missing_latency',\n      detail: 'A finite latency limit was configured but no latency evidence is available.',\n    })\n  } else if (metrics.p95WallMs !== null && metrics.p95WallMs > thresholds.maxP95WallMs) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'latency_budget',\n      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n    })\n  }\n}\n\nfunction buildAxes(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n  return [\n    axis(\n      'corpus',\n      issues,\n      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n    ),\n    axis(\n      'quality',\n      issues,\n      metrics.passRate === null || metrics.meanScore === null\n        ? null\n        : Math.min(metrics.passRate, metrics.meanScore),\n      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n    ),\n    axis(\n      'reliability',\n      issues,\n      metrics.reliabilityRate,\n      `successRate=${fmt(metrics.reliabilityRate)} terminalFailures=${metrics.terminalFailureRuns} unclassified=${metrics.unclassifiedTerminalRuns}`,\n    ),\n    axis(\n      'generalization',\n      issues,\n      gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n    ),\n    axis(\n      'diagnostics',\n      issues,\n      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n    ),\n    axis(\n      'efficiency',\n      issues,\n      efficiencyScore(metrics, thresholds),\n      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n    ),\n  ]\n}\n\nfunction axis(\n  name: ReleaseConfidenceAxisName,\n  issues: ReleaseConfidenceIssue[],\n  score: number | null,\n  detail: string,\n): ReleaseConfidenceAxis {\n  const own = issues.filter((i) => i.axis === name)\n  const status = own.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : own.length > 0\n      ? 'warn'\n      : 'pass'\n  return { name, status, score: score === null ? null : bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n  const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n  for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n  return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const scenario of scenarios) {\n    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n    out[domain] = (out[domain] ?? 0) + 1\n  }\n  return out\n}\n\nfunction countFailureClasses(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Partial<Record<FailureClass, number>> {\n  const out: Partial<Record<FailureClass, number>> = {}\n  for (const run of runs) {\n    // Ungated: a failure-mode census counts what the runs REPORTED, which is\n    // the only way a gamed run's inflated score is visible at all. It is not a\n    // promotion number — `passRate` is, and that one excludes gated runs and\n    // publishes the excluded count as `metrics.realnessGatedRuns`.\n    if (runPassOutcome(run, threshold) === false) {\n      const failureClass =\n        run.failureClass !== undefined && run.failureClass !== 'success'\n          ? run.failureClass\n          : 'unknown'\n      out[failureClass] = (out[failureClass] ?? 0) + 1\n    }\n  }\n  for (const trace of traces) {\n    if (tracePassOutcome(trace, threshold) === false) {\n      const failureClass =\n        trace.failureClass !== undefined && trace.failureClass !== 'success'\n          ? trace.failureClass\n          : 'unknown'\n      out[failureClass] = (out[failureClass] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const trace of traces) {\n    for (const asi of trace.asi ?? []) {\n      const surface = asi.responsibleSurface ?? 'unknown'\n      out[surface] = (out[surface] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction failedRows(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Array<{ hasAsi: boolean }> {\n  const out: Array<{ hasAsi: boolean }> = []\n  for (const run of runs) {\n    if (runPassOutcome(run, threshold) === false) {\n      const asiMetric = run.outcome.raw.asi\n      out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n    }\n  }\n  for (const trace of traces) {\n    if (tracePassOutcome(trace, threshold) === false) {\n      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n    }\n  }\n  return out\n}\n\nfunction passOutcomeRate(outcomes: readonly (boolean | null)[]): number | null {\n  const classified = outcomes.filter((outcome): outcome is boolean => outcome !== null)\n  if (classified.length === 0) return null\n  return classified.filter(Boolean).length / classified.length\n}\n\nfunction runPassOutcome(run: RunRecord, threshold: number): boolean | null {\n  if (hasExplicitTaskFailure(run)) return false\n  const score = runSplitScore(run)\n  return score === undefined ? null : score >= threshold\n}\n\nfunction hasExplicitTaskFailure(run: RunRecord): boolean {\n  return run.failureClass !== undefined && run.failureClass !== 'success'\n}\n\nfunction tracePassOutcome(trace: ReleaseTraceEvidence, threshold: number): boolean | null {\n  if (trace.failureClass !== undefined && trace.failureClass !== 'success') return false\n  if (trace.ok === false) return false\n  if (isFiniteNumber(trace.score)) return trace.score >= threshold\n  return trace.ok === true ? true : null\n}\n\nfunction isFailedTerminalOutcome(\n  outcome: RunRecord['terminalOutcome'],\n): outcome is 'failed' | 'cancelled' | 'incomplete' {\n  return outcome === 'failed' || outcome === 'cancelled' || outcome === 'incomplete'\n}\n\nfunction terminalSuccess(outcome: RunRecord['terminalOutcome']): boolean | undefined {\n  if (outcome === 'succeeded') return true\n  if (isFailedTerminalOutcome(outcome)) return false\n  return undefined\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n  return runs\n    .filter((run) => run.splitTag === split)\n    .map(runSplitScore)\n    .filter(isFiniteNumber)\n}\n\n/**\n * RAW and split-exact (`observedSplitScore`): this feeds the per-split means,\n * the overfit gap, and the pass threshold — descriptions of what the runs\n * reported. A gamed run inflating them is visible next to `realnessGatedRuns`,\n * and the promotion number (`passRate`) excludes gated runs entirely.\n */\nfunction runSplitScore(run: RunRecord): number | undefined {\n  const score = observedSplitScore(run, run.splitTag === 'holdout' ? 'holdout' : 'search')\n  return isFiniteNumber(score) ? score : undefined\n}\n\nfunction meanOrNull(xs: readonly number[]): number | null {\n  if (xs.length === 0) return null\n  return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentileOrNull(xs: readonly number[], p: number): number | null {\n  if (xs.length === 0) return null\n  const sorted = [...xs].sort((a, b) => a - b)\n  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction diffOrNull(a: number | null, b: number | null): number | null {\n  if (a === null || b === null) return null\n  return a - b\n}\n\nfunction gapScore(gap: number | null, maxGap: number): number | null {\n  if (gap === null) return null\n  if (maxGap <= 0) return gap <= 0 ? 1 : 0\n  return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n): number | null {\n  const cost = Number.isFinite(thresholds.maxMeanCostUsd)\n    ? metrics.meanCostUsd === null\n      ? null\n      : bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n    : 1\n  const latency = Number.isFinite(thresholds.maxP95WallMs)\n    ? metrics.p95WallMs === null\n      ? null\n      : bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n    : 1\n  if (cost === null || latency === null) return null\n  return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n  target: string,\n  status: ReleaseConfidenceStatus,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): string {\n  const prefix = `release confidence ${status}: ${target}`\n  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n  if (issues.length === 0) return `${prefix}; ${metricText}`\n  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number | null): string {\n  if (x === null) return 'n/a'\n  return x.toFixed(4)\n}\n"],"mappings":";;;;;;;;;;;;AAgEA,SAAgB,YACd,UACA,WACA,UAA4B,CAAC,GACZ;CACjB,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,WAAW,QAAQ,mBAAmB;CAC5C,MAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;CAEpE,MAAM,eAAe,KAAK,QAAQ;CAClC,MAAM,gBAAgB,KAAK,SAAS;CACpC,MAAM,QAAQ,gBAAgB;CAE9B,IACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GAErB,OAAO;EACL;EACA;EACA;EACA,SAAS;EACT,SAAS;EACT,YAAY;EACZ;EACA,SAAS;CACX;CAGF,MAAM,SAAmB,IAAI,MAAM,UAAU;CAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,UAAU,GAAG;EAExC,OAAO,KAAK,KADM,SAAS,WAAW,GACb,CAAC,IAAI,KAAK,SAAS;CAC9C;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;CACpD,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;CAC5D,MAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ;CAC3C,MAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ;CAExD,IAAI;CAMJ,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,CAAC,OAAO,SAAS,OAAO,KAAK,YAAY,SACxE,UAAU;MACL,IAAI,UAAU,GAAG,UAAU;MAC7B,IAAI,UAAU,GAAG,UAAU;MAC3B,IAAI,SAAS,GAAG,UAAU;MAC1B,UAAU;CAEf,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,IAAI,KAAK;CACzB,OAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;CAC3D,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM;CAC5E,OAAO;AACT;;AAGA,SAAS,WAAW,MAA4B;CAC9C,IAAI,IAAI,SAAS;CACjB,aAAa;EACX,KAAK;EACL,IAAI,IAAI;EACR,IAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;EACnC,KAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;EACxC,SAAS,IAAK,MAAM,QAAS,KAAK;CACpC;AACF;;AAGA,SAAS,SAAS,GAAa,GAAqB;CAClD,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;EAC5B,MAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;EACjC,MAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;EACxC,KAAK,MAAM,QAAQ,OAAO;GACxB,KAAK;GACL,IAAI,KAAK,KAAK,GAAG,QAAQ;EAC3B;CACF;CACA,OAAO,MAAM;AACf;;;;AAiCA,eAAsB,gBACpB,MACkF;CAClF,MAAM,cAAc,KAAK,oBAAoB;CAC7C,MAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;CACnF,MAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;CAMrF,OAAO;EACL,GANS,YAAY,gBAAgB,iBAAiB;GACtD,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,GAAI,KAAK,eAAe,KAAA,IAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;GACvE,GAAI,KAAK,SAAS,KAAA,IAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;EACvD,CAEM;EACJ,iBAAiB,eAAe;EAChC,kBAAkB,gBAAgB;CACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;CACnB,MAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;CAClD,IAAI,OAAO;CACX,eAAe,SAAwB;EACrC,OAAO,MAAM;GACX,MAAM,IAAI;GACV,IAAI,KAAK,QAAQ,QAAQ;GACzB,MAAM,IAAI,MAAM,MAAM,QAAQ,EAAG;GACjC,QAAQ,KAAK,OAAO,SAAS,CAAC,IAAI,IAAI;EACxC;CACF;CACA,MAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,SAAS,OAAO,CAAC,CAAC;CAClF,OAAO;AACT;;;ACzEA,MAAM,qBAA4D;CAChE,eAAe;CACf,kBAAkB;CAClB,eAAe;CACf,gBAAgB;CAChB,gBAAgB;CAChB,aAAa;CACb,cAAc;CACd,eAAe;CACf,gBAAgB,OAAO;CACvB,cAAc,OAAO;CACrB,uBAAuB;CACvB,uBAAuB;AACzB;AAEA,SAAgB,0BACd,OAC4B;CAC5B,MAAM,aAAa;EAAE,GAAG;EAAoB,GAAG,MAAM;CAAW;CAChE,MAAM,cAAc,MAAM,eAAe;CACzC,MAAM,OAAO,iBACV,MAAM,QAAQ,CAAC,EAAA,CAAG,IAAI,iBAAiB,GACxC,aACA,MAAM,UACR;CACA,MAAM,SAAS,sBACZ,MAAM,UAAU,CAAC,EAAA,CAAG,IAAI,4BAA4B,GACrD,aACA,MAAM,UACR;CACA,MAAM,YAAY,MAAM,aAAa,CAAC;CACtC,MAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;CAChE,MAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;CAC/E,MAAM,eAAe,UAAU,MAAM,QAAQ;CAC7C,MAAM,gBAAgB,UAAU,MAAM,SAAS;CAC/C,MAAM,YAAY,KAAK,IAAI,aAAa,CAAC,CAAC,OAAO,cAAc;CAC/D,MAAM,cAAc,OAAO,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,OAAO,cAAc;CACpE,MAAM,gBAAgB,KAAK,SAAS,IAAI,YAAY;CACpD,MAAM,cAAc,KAAK,QAAQ,QAAQ,cAAc,GAAG,MAAM,KAAA,CAAS;CACzE,MAAM,eAAe,KAAK,QACvB,QACC,cAAc,GAAG,MAAM,KAAA,KACvB,CAAC,uBAAuB,GAAG,KAC3B,CAAC,wBAAwB,IAAI,eAAe,CAChD,CAAC,CAAC;CAOF,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,gBAAgB,GAAG,CAAC;CAC7D,MAAM,eACJ,KAAK,SAAS,IACV,WAAW,KAAK,QAAQ,eAAe,KAAK,WAAW,qBAAqB,CAAC,IAC7E,OAAO,KAAK,UAAU,iBAAiB,OAAO,WAAW,qBAAqB,CAAC;CACrF,MAAM,kBACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,gBAAgB,IAAI,eAAe,CAAC,IACtD,OAAO,KAAK,UAAU,MAAM,EAAE;CACpC,MAAM,2BAA2B,gBAAgB,QAAQ,YAAY,YAAY,KAAA,CAAS,CAAC,CAAC;CAC5F,MAAM,sBAAsB,gBAAgB,QAAQ,YAAY,YAAY,KAAK,CAAC,CAAC;CACnF,MAAM,kBACJ,gBAAgB,WAAW,KAAK,2BAA2B,IACvD,QACC,gBAAgB,SAAS,uBAAuB,gBAAgB;CACvE,MAAM,aAAa,YAAY,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,cAAc,YAAY,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACxE,MAAM,SAAS,WAAW,MAAM,QAAQ,WAAW,qBAAqB;CACxE,MAAM,kBAAkB,WAAW,YAAY;CAC/C,MAAM,mBAAmB,WAAW,aAAa;CACjD,MAAM,WAAW,KAAK,SAAS,QAC7B,IAAI,eAAe,SAAS,eAAe,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG,CACzE;CACA,MAAM,aAAa,OAAO,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,OAAO,cAAc;CAC7E,MAAM,cACJ,KAAK,SAAS,IACV,SAAS,WAAW,KAAK,SACvB,WAAW,QAAQ,IACnB,OACF,WAAW,UAAU;CAC3B,MAAM,YACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,IAAI,MAAM,IAC5B,OAAO,KAAK,UAAU,MAAM,UAAU,CAAC,CAAC,OAAO,cAAc;CACnE,MAAM,UAAoC;EACxC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU,gBAAgB,YAAY;EACtC,mBAAmB,KAAK,SAAS,WAAW;EAC5C,WAAW,WAAW,aAAa;EACnC;EACA;EACA,YAAY,WAAW,iBAAiB,gBAAgB;EACxD;EACA,WAAW,iBAAiB,WAAW,GAAI;EAC3C,YAAY,OAAO;EACnB,iBAAiB,OAAO,QAAQ,QAAQ,IAAI,MAAM,CAAC,CAAC;EACpD,kBAAkB,OAAO,QAAQ,MAAM,EAAE,cAAc,CAAC,CAAC,CAAC;EAC1D,iBAAiB,OAAO,QAAQ,OAAO,EAAE,aAAa,KAAK,CAAC,CAAC,CAAC;EAC9D;EACA,cAAc,aAAa,SAAS;EACpC,oBAAoB,oBAAoB,MAAM,QAAQ,WAAW,qBAAqB;EACtF,0BAA0B,yBAAyB,MAAM;CAC3D;CAEA,MAAM,SAAmC,CAAC;CAC1C,YAAY,OAAO,YAAY,SAAS,MAAM;CAC9C,aAAa,YAAY,SAAS,MAAM;CACxC,iBAAiB,SAAS,MAAM;CAChC,oBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;CAC3E,iBAAiB,YAAY,SAAS,MAAM;CAC5C,gBAAgB,YAAY,SAAS,MAAM;CAE3C,MAAM,OAAO,UAAU,SAAS,YAAY,MAAM;CAClD,MAAM,SAAS,OAAO,MAAM,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;CAEN,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM,cAAc;EAChC;EACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;EACjF;EACA;EACA;EACA,SAAS,MAAM,WAAW;EAC1B,cAAc,MAAM,gBAAgB;EACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;CAC9D;AACF;AAEA,SAAgB,wBAAwB,OAA2D;CACjG,MAAM,YAAY,0BAA0B,KAAK;CACjD,IAAI,UAAU,WAAW,QACvB,MAAM,IAAI,kBAAkB,UAAU,OAAO;CAE/C,OAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;CACb,IAAI,aAAa,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,WAAW;CACxE,IAAI,YAAY,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,UAAU;CACtE,OAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;CACxB,IAAI,aACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,WAAW;CAC1F,IAAI,YACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,UAAU;CACzF,OAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,6BACP,OACA,OACsB;CACtB,MAAM,QAAQ;CACd,IAAI,OAAO,OAAO,OAAO,aAAa,GACpC,MAAM,IAAI,gBACR,UAAU,MAAM,2DAClB;CAEF,IAAI,MAAM,iBAAiB,KAAA,KAAa,CAAC,gBAAgB,SAAS,MAAM,YAAY,GAClF,MAAM,IAAI,gBACR,UAAU,MAAM,gCAAgC,gBAAgB,KAAK,IAAI,GAC3E;CAEF,OAAO;AACT;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;CACN,IAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GACnF,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,gBAAgB,WAAW,kBACrC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,cAAc,qBAAqB,WAAW,iBAAiB;CACpF,CAAC;CAEH,IAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;AAEL;AAEA,SAAS,aACP,YACA,SACA,QACM;CACN,IAAI,QAAQ,aAAa,WAAW,eAClC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,WAAW,uBAAuB,WAAW,cAAc;CAChF,CAAC;CAEH,IAAI,QAAQ,eAAe,GACzB,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa;CAClC,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,cAAc,MACrD,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,WAAW,WAAW,aAC7D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,EAAE,KAAK,IAAI,WAAW,WAAW,EAAE;CAC7E,CAAC;CAEH,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,iBACP,SACA,QACM;CACN,IAAI,QAAQ,oBAAoB,MAC9B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QACE,QAAQ,2BAA2B,IAC/B,GAAG,QAAQ,yBAAyB,wDACpC;CACR,CAAC;CAEH,IAAI,QAAQ,sBAAsB,GAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,oBAAoB;CACzC,CAAC;AAEL;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;CACN,IAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAChE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,YAAY,wBAAwB,WAAW,eAAe;CACnF,CAAC;CAEH,IAAI,QAAQ,eAAe,QAAQ,QAAQ,aAAa,WAAW,eACjE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,EAAE,KAAK,IAAI,WAAW,aAAa,EAAE;CAC3F,CAAC;CAEH,IAAI,gBAAgB,CAAC,aAAa,SAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM,QAAQ,aAAa,iBAAiB;EAC5C,QAAQ,aAAa;CACvB,CAAC;AAEL;AAEA,SAAS,iBACP,YACA,SACA,QACM;CACN,IAAI,CAAC,WAAW,uBAAuB;CACvC,IAAI,QAAQ,aAAa,QAAQ,iBAC/B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,gBAAgB;CAC1D,CAAC;AAEL;AAEA,SAAS,gBACP,YACA,SACA,QACM;CACN,IAAI,OAAO,SAAS,WAAW,cAAc,KAAK,QAAQ,gBAAgB,MACxE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,gBAAgB,QAAQ,QAAQ,cAAc,WAAW,gBAC1E,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,EAAE,KAAK,IAAI,WAAW,cAAc,EAAE;CACtF,CAAC;CAEH,IAAI,OAAO,SAAS,WAAW,YAAY,KAAK,QAAQ,cAAc,MACpE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cACtE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,UACP,SACA,YACA,QACyB;CACzB,OAAO;EACL,KACE,UACA,QACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC,GACxE,GAAG,QAAQ,cAAc,sBAAsB,QAAQ,YAAY,SACrE;EACA,KACE,WACA,QACA,QAAQ,aAAa,QAAQ,QAAQ,cAAc,OAC/C,OACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS,GAChD,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS,GACtE;EACA,KACE,eACA,QACA,QAAQ,iBACR,eAAe,IAAI,QAAQ,eAAe,EAAE,oBAAoB,QAAQ,oBAAoB,gBAAgB,QAAQ,0BACtH;EACA,KACE,kBACA,QACA,SAAS,QAAQ,YAAY,WAAW,aAAa,GACrD,eAAe,QAAQ,YAAY,cAAc,IAAI,QAAQ,UAAU,GACzE;EACA,KACE,eACA,QACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ,YACjE,mBAAmB,QAAQ,gBAAgB,GAAG,QAAQ,YACxD;EACA,KACE,cACA,QACA,gBAAgB,SAAS,UAAU,GACnC,eAAe,IAAI,QAAQ,WAAW,EAAE,aAAa,IAAI,QAAQ,SAAS,GAC5E;CACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;CACvB,MAAM,MAAM,OAAO,QAAQ,MAAM,EAAE,SAAS,IAAI;CAMhD,OAAO;EAAE;EAAM,QALA,IAAI,MAAM,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;EACiB,OAAO,UAAU,OAAO,OAAO,QAAQ,KAAK;EAAG;CAAO;AAC/E;AAEA,SAAS,oBAAoB,WAAqE;CAChG,MAAM,SAAuC;EAAE,OAAO;EAAG,KAAK;EAAG,MAAM;EAAG,SAAS;CAAE;CACrF,KAAK,MAAM,YAAY,WAAW,OAAO,SAAS,SAAS,QAAQ;CACnE,OAAO;AACT;AAEA,SAAS,aAAa,WAA+D;CACnF,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;EACnE,IAAI,WAAW,IAAI,WAAW,KAAK;CACrC;CACA,OAAO;AACT;AAEA,SAAS,oBACP,MACA,QACA,WACuC;CACvC,MAAM,MAA6C,CAAC;CACpD,KAAK,MAAM,OAAO,MAKhB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,eACJ,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,YACnD,IAAI,eACJ;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;EAChD,MAAM,eACJ,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,YACvD,MAAM,eACN;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,OAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;CACjG,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,SAAS,QAClB,KAAK,MAAM,OAAO,MAAM,OAAO,CAAC,GAAG;EACjC,MAAM,UAAU,IAAI,sBAAsB;EAC1C,IAAI,YAAY,IAAI,YAAY,KAAK;CACvC;CAEF,OAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;CAC5B,MAAM,MAAkC,CAAC;CACzC,KAAK,MAAM,OAAO,MAChB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,YAAY,IAAI,QAAQ,IAAI;EAClC,IAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;CACrE;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OACzC,IAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;CAGrD,OAAO;AACT;AAEA,SAAS,gBAAgB,UAAsD;CAC7E,MAAM,aAAa,SAAS,QAAQ,YAAgC,YAAY,IAAI;CACpF,IAAI,WAAW,WAAW,GAAG,OAAO;CACpC,OAAO,WAAW,OAAO,OAAO,CAAC,CAAC,SAAS,WAAW;AACxD;AAEA,SAAS,eAAe,KAAgB,WAAmC;CACzE,IAAI,uBAAuB,GAAG,GAAG,OAAO;CACxC,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO,UAAU,KAAA,IAAY,OAAO,SAAS;AAC/C;AAEA,SAAS,uBAAuB,KAAyB;CACvD,OAAO,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB;AAChE;AAEA,SAAS,iBAAiB,OAA6B,WAAmC;CACxF,IAAI,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,WAAW,OAAO;CACjF,IAAI,MAAM,OAAO,OAAO,OAAO;CAC/B,IAAI,eAAe,MAAM,KAAK,GAAG,OAAO,MAAM,SAAS;CACvD,OAAO,MAAM,OAAO,OAAO,OAAO;AACpC;AAEA,SAAS,wBACP,SACkD;CAClD,OAAO,YAAY,YAAY,YAAY,eAAe,YAAY;AACxE;AAEA,SAAS,gBAAgB,SAA4D;CACnF,IAAI,YAAY,aAAa,OAAO;CACpC,IAAI,wBAAwB,OAAO,GAAG,OAAO;AAE/C;AAEA,SAAS,UAAU,MAA4B,OAA8B;CAC3E,OAAO,KACJ,QAAQ,QAAQ,IAAI,aAAa,KAAK,CAAC,CACvC,IAAI,aAAa,CAAC,CAClB,OAAO,cAAc;AAC1B;;;;;;;AAQA,SAAS,cAAc,KAAoC;CACzD,MAAM,QAAQ,mBAAmB,KAAK,IAAI,aAAa,YAAY,YAAY,QAAQ;CACvF,OAAO,eAAe,KAAK,IAAI,QAAQ,KAAA;AACzC;AAEA,SAAS,WAAW,IAAsC;CACxD,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,iBAAiB,IAAuB,GAA0B;CACzE,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,MAAM,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,OAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC;AACzF;AAEA,SAAS,eAAe,OAAiC;CACvD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,WAAW,GAAkB,GAAiC;CACrE,IAAI,MAAM,QAAQ,MAAM,MAAM,OAAO;CACrC,OAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAoB,QAA+B;CACnE,IAAI,QAAQ,MAAM,OAAO;CACzB,IAAI,UAAU,GAAG,OAAO,OAAO,IAAI,IAAI;CACvC,OAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACe;CACf,MAAM,OAAO,OAAO,SAAS,WAAW,cAAc,IAClD,QAAQ,gBAAgB,OACtB,OACA,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IAC1E;CACJ,MAAM,UAAU,OAAO,SAAS,WAAW,YAAY,IACnD,QAAQ,cAAc,OACpB,OACA,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACtE;CACJ,IAAI,SAAS,QAAQ,YAAY,MAAM,OAAO;CAC9C,OAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;CACR,MAAM,SAAS,sBAAsB,OAAO,IAAI;CAChD,MAAM,aAAa,aAAa,QAAQ,cAAc,cAAc,QAAQ,WAAW,eAAe,QAAQ,YAAY,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS;CAC9L,IAAI,OAAO,WAAW,GAAG,OAAO,GAAG,OAAO,IAAI;CAC9C,OAAO,GAAG,OAAO,IAAI,WAAW,WAAW,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;AAC/E;AAEA,SAAS,IAAI,GAA0B;CACrC,IAAI,MAAM,MAAM,OAAO;CACvB,OAAO,EAAE,QAAQ,CAAC;AACpB"}

package/dist/researcher-DJnoUE8c.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"researcher-DJnoUE8c.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;UAyEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;;EAMT,SAAS;;UAGD;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ"}

package/dist/rubric-predictive-validity-C7LnNvF2.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"rubric-predictive-validity-C7LnNvF2.d.ts","names":[],"sources":["../src/meta-eval/rubric-predictive-validity.ts"],"mappings":";;;UAoCiB;;;;;EAKf,MAAM;EACN,UAAU;;;;;;EAMV;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;;;;EAKA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;IAAQ;IAAa;;;;;;;;EAOrB;;UAGe;EACf;;EAEA;EACA;EACA;EACA;EACA,SAAS;;UAGM;EACf,OAAO;;EAEP,QAAQ;EACR;EACA;;EAEA;;iBAGoB,yBACpB,OAAO,gCACN,QAAQ"}

package/dist/rubric-predictive-validity-Cwwyd7ah.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"rubric-predictive-validity-Cwwyd7ah.js","names":[],"sources":["../src/meta-eval/rubric-predictive-validity.ts"],"sourcesContent":["/**\n * Rubric predictive validity — does our eval rubric predict deployment\n * outcomes?\n *\n * `correlationStudy` (already in this package) joins a `TraceStore` to an\n * `OutcomeStore` and computes Pearson + Spearman + bootstrap CI for each\n * (eval-metric, outcome-metric) pair. That answers \"does X correlate with\n * Y at all.\" `rubricPredictiveValidity` is the campaign-shaped wrapper\n * around it: take a sequence of `RunRecord`s (the canonical campaign\n * artifact) and a `DeploymentOutcomeStore`, join on `runId`, return a\n * ranked verdict on every rubric whose dimension scores were captured in\n * `outcome.raw`.\n *\n * The point — quoting the methodology doc — is that **without this loop\n * every rubric is faith-based**. Once it's wired, you know which rubrics\n * have earned their promotion power and which ones are decoration.\n *\n *   const validity = await rubricPredictiveValidity({\n *     runs: lastQuarter,\n *     outcomes: shipFlagOutcomeStore,\n *     outcomeMetrics: ['revenue_lift', 'retention_30d', 'csat'],\n *     rubrics: ['anti_slop', 'semantic_concept', 'tool_recovery'],\n *   })\n *   for (const r of validity.ranked) {\n *     console.log(`${r.rubric} → ${r.bestOutcome}: ρ=${r.spearman.toFixed(2)}`)\n *   }\n *\n * The function is intentionally read-only. Use the verdict to deprecate\n * decorative rubrics, re-weight composite scores, or trigger a\n * recalibration sweep when predictive validity drops below a threshold.\n */\n\nimport type { RunRecord } from '../run-record'\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface RubricPredictiveValidityInput {\n  /**\n   * Canonical campaign output. Each record's `outcome.raw[<rubricId>]`\n   * provides the eval score; missing keys are silently skipped per pair.\n   */\n  runs: RunRecord[]\n  outcomes: OutcomeStore\n  /**\n   * Outcome metric names to evaluate against. Each must appear in at\n   * least one `DeploymentOutcome.metrics` keyspace; pairs with too few\n   * joined samples are excluded from the result.\n   */\n  outcomeMetrics: string[]\n  /**\n   * Rubric ids to evaluate. Must appear as keys in `RunRecord.outcome.raw`.\n   * If omitted, every numeric key in `outcome.raw` across the run set is\n   * treated as a rubric.\n   */\n  rubrics?: string[]\n  /** Minimum joined-sample count before a pair is reported. Default 8. */\n  minSamples?: number\n  /** Bootstrap resamples for CI. Default 500. */\n  bootstrapResamples?: number\n  /** Random seed for the bootstrap (mulberry32). Default unset (Math.random). */\n  seed?: number\n  /**\n   * Reduction when multiple outcomes attach to one runId. Default `'latest'`\n   * (most recently captured).\n   */\n  reduction?: 'latest' | 'mean' | 'max'\n}\n\nexport interface RubricOutcomePair {\n  rubric: string\n  outcome: string\n  n: number\n  pearson: number\n  spearman: number\n  ci95: { low: number; high: number }\n  /**\n   * Verdict bucket. `load_bearing` ≥ 0.7, `informative` ≥ 0.4,\n   * `decorative` < 0.4 in absolute correlation. A negative correlation\n   * with a desired outcome is also `decorative` — actively misleading\n   * is worse than uninformative.\n   */\n  verdict: 'load_bearing' | 'informative' | 'decorative'\n}\n\nexport interface RubricRanking {\n  rubric: string\n  /** Outcome metric this rubric correlated best with. */\n  bestOutcome: string\n  spearman: number\n  pearson: number\n  n: number\n  verdict: RubricOutcomePair['verdict']\n}\n\nexport interface RubricPredictiveValidityReport {\n  pairs: RubricOutcomePair[]\n  /** Per-rubric best pair, sorted descending by |spearman|. */\n  ranked: RubricRanking[]\n  joinedSamples: number\n  skippedRuns: number\n  /** Rubrics that were declared but never produced a usable score. */\n  rubricsWithoutData: string[]\n}\n\nexport async function rubricPredictiveValidity(\n  input: RubricPredictiveValidityInput,\n): Promise<RubricPredictiveValidityReport> {\n  const minSamples = input.minSamples ?? 8\n  const reduction = input.reduction ?? 'latest'\n  const resamples = input.bootstrapResamples ?? 500\n  const rng = makeRng(input.seed)\n\n  const outcomes = await input.outcomes.list()\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  // Discover rubrics: caller-declared OR every numeric key in outcome.raw\n  // observed across runs.\n  const observedRubrics = new Set<string>()\n  for (const r of input.runs) {\n    for (const k of Object.keys(r.outcome.raw)) observedRubrics.add(k)\n  }\n  const rubrics = input.rubrics ?? [...observedRubrics]\n\n  // Collect aligned (x, y) pairs per (rubric, outcome).\n  type Bucket = { rubric: string; outcome: string; xs: number[]; ys: number[] }\n  const buckets: Bucket[] = []\n  for (const r of rubrics) {\n    for (const o of input.outcomeMetrics) {\n      buckets.push({ rubric: r, outcome: o, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of input.runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    let joinedThisRun = false\n    for (const r of rubrics) {\n      const x = run.outcome.raw[r]\n      if (typeof x !== 'number' || !Number.isFinite(x)) continue\n      for (const o of input.outcomeMetrics) {\n        const values = os\n          .map((row) => row.metrics[o])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, os, o, reduction)\n        if (y === null) continue\n        const bucket = buckets.find((b) => b.rubric === r && b.outcome === o)!\n        bucket.xs.push(x)\n        bucket.ys.push(y)\n        joinedThisRun = true\n      }\n    }\n    if (joinedThisRun) joined++\n  }\n\n  const pairs: RubricOutcomePair[] = []\n  for (const b of buckets) {\n    if (b.xs.length < minSamples) continue\n    const pearson = pearsonR(b.xs, b.ys)\n    const spearman = spearmanR(b.xs, b.ys)\n    const ci = bootstrapCi(b.xs, b.ys, resamples, rng)\n    const verdict: RubricOutcomePair['verdict'] =\n      Math.abs(spearman) >= 0.7\n        ? 'load_bearing'\n        : Math.abs(spearman) >= 0.4\n          ? 'informative'\n          : 'decorative'\n    pairs.push({\n      rubric: b.rubric,\n      outcome: b.outcome,\n      n: b.xs.length,\n      pearson,\n      spearman,\n      ci95: ci,\n      verdict,\n    })\n  }\n\n  const byRubric = new Map<string, RubricOutcomePair[]>()\n  for (const p of pairs) {\n    const arr = byRubric.get(p.rubric) ?? []\n    arr.push(p)\n    byRubric.set(p.rubric, arr)\n  }\n  const ranked: RubricRanking[] = [...byRubric.entries()]\n    .map(([rubric, ps]) => {\n      const best = ps.reduce((a, b) => (Math.abs(b.spearman) > Math.abs(a.spearman) ? b : a))\n      return {\n        rubric,\n        bestOutcome: best.outcome,\n        spearman: best.spearman,\n        pearson: best.pearson,\n        n: best.n,\n        verdict: best.verdict,\n      }\n    })\n    .sort((a, b) => Math.abs(b.spearman) - Math.abs(a.spearman))\n\n  const rubricsWithoutData = rubrics.filter((r) => !byRubric.has(r))\n\n  return { pairs, ranked, joinedSamples: joined, skippedRuns: skipped, rubricsWithoutData }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  outcomes: DeploymentOutcome[],\n  metric: string,\n  kind: 'latest' | 'mean' | 'max',\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((s, v) => s + v, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest'\n  const sorted = [...outcomes]\n    .filter((o) => typeof o.metrics[metric] === 'number')\n    .sort((a, b) => b.capturedAt - a.capturedAt)\n  return sorted[0]?.metrics[metric] ?? null\n}\n\nfunction bootstrapCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n  rng: () => number,\n): { low: number; high: number } {\n  const n = xs.length\n  if (n < 3) return { low: Number.NaN, high: Number.NaN }\n  const samples: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx = new Array<number>(n)\n    const ry = new Array<number>(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(rng() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) samples.push(r)\n  }\n  samples.sort((a, b) => a - b)\n  if (samples.length === 0) return { low: Number.NaN, high: Number.NaN }\n  return {\n    low: samples[Math.floor(0.025 * samples.length)]!,\n    high: samples[Math.min(samples.length - 1, Math.floor(0.975 * samples.length))]!,\n  }\n}\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n"],"mappings":";;AAwGA,eAAsB,yBACpB,OACyC;CACzC,MAAM,aAAa,MAAM,cAAc;CACvC,MAAM,YAAY,MAAM,aAAa;CACrC,MAAM,YAAY,MAAM,sBAAsB;CAC9C,MAAM,MAAM,QAAQ,MAAM,IAAI;CAE9B,MAAM,WAAW,MAAM,MAAM,SAAS,KAAK;CAC3C,MAAM,gCAAgB,IAAI,IAAiC;CAC3D,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;EAC3C,IAAI,KAAK,CAAC;EACV,cAAc,IAAI,EAAE,OAAO,GAAG;CAChC;CAIA,MAAM,kCAAkB,IAAI,IAAY;CACxC,KAAK,MAAM,KAAK,MAAM,MACpB,KAAK,MAAM,KAAK,OAAO,KAAK,EAAE,QAAQ,GAAG,GAAG,gBAAgB,IAAI,CAAC;CAEnE,MAAM,UAAU,MAAM,WAAW,CAAC,GAAG,eAAe;CAIpD,MAAM,UAAoB,CAAC;CAC3B,KAAK,MAAM,KAAK,SACd,KAAK,MAAM,KAAK,MAAM,gBACpB,QAAQ,KAAK;EAAE,QAAQ;EAAG,SAAS;EAAG,IAAI,CAAC;EAAG,IAAI,CAAC;CAAE,CAAC;CAI1D,IAAI,SAAS;CACb,IAAI,UAAU;CACd,KAAK,MAAM,OAAO,MAAM,MAAM;EAC5B,MAAM,KAAK,cAAc,IAAI,IAAI,KAAK;EACtC,IAAI,CAAC,MAAM,GAAG,WAAW,GAAG;GAC1B;GACA;EACF;EACA,IAAI,gBAAgB;EACpB,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,IAAI,IAAI,QAAQ,IAAI;GAC1B,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;GAClD,KAAK,MAAM,KAAK,MAAM,gBAAgB;IACpC,MAAM,SAAS,GACZ,KAAK,QAAQ,IAAI,QAAQ,EAAE,CAAC,CAC5B,QAAQ,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;IACzE,IAAI,OAAO,WAAW,GAAG;IACzB,MAAM,IAAI,OAAO,QAAQ,IAAI,GAAG,SAAS;IACzC,IAAI,MAAM,MAAM;IAChB,MAAM,SAAS,QAAQ,MAAM,MAAM,EAAE,WAAW,KAAK,EAAE,YAAY,CAAC;IACpE,OAAO,GAAG,KAAK,CAAC;IAChB,OAAO,GAAG,KAAK,CAAC;IAChB,gBAAgB;GAClB;EACF;EACA,IAAI,eAAe;CACrB;CAEA,MAAM,QAA6B,CAAC;CACpC,KAAK,MAAM,KAAK,SAAS;EACvB,IAAI,EAAE,GAAG,SAAS,YAAY;EAC9B,MAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;EACnC,MAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;EACrC,MAAM,KAAK,YAAY,EAAE,IAAI,EAAE,IAAI,WAAW,GAAG;EACjD,MAAM,UACJ,KAAK,IAAI,QAAQ,KAAK,KAClB,iBACA,KAAK,IAAI,QAAQ,KAAK,KACpB,gBACA;EACR,MAAM,KAAK;GACT,QAAQ,EAAE;GACV,SAAS,EAAE;GACX,GAAG,EAAE,GAAG;GACR;GACA;GACA,MAAM;GACN;EACF,CAAC;CACH;CAEA,MAAM,2BAAW,IAAI,IAAiC;CACtD,KAAK,MAAM,KAAK,OAAO;EACrB,MAAM,MAAM,SAAS,IAAI,EAAE,MAAM,KAAK,CAAC;EACvC,IAAI,KAAK,CAAC;EACV,SAAS,IAAI,EAAE,QAAQ,GAAG;CAC5B;CACA,MAAM,SAA0B,CAAC,GAAG,SAAS,QAAQ,CAAC,CAAC,CACpD,KAAK,CAAC,QAAQ,QAAQ;EACrB,MAAM,OAAO,GAAG,QAAQ,GAAG,MAAO,KAAK,IAAI,EAAE,QAAQ,IAAI,KAAK,IAAI,EAAE,QAAQ,IAAI,IAAI,CAAE;EACtF,OAAO;GACL;GACA,aAAa,KAAK;GAClB,UAAU,KAAK;GACf,SAAS,KAAK;GACd,GAAG,KAAK;GACR,SAAS,KAAK;EAChB;CACF,CAAC,CAAC,CACD,MAAM,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,IAAI,KAAK,IAAI,EAAE,QAAQ,CAAC;CAE7D,MAAM,qBAAqB,QAAQ,QAAQ,MAAM,CAAC,SAAS,IAAI,CAAC,CAAC;CAEjE,OAAO;EAAE;EAAO;EAAQ,eAAe;EAAQ,aAAa;EAAS;CAAmB;AAC1F;AAIA,SAAS,OACP,QACA,UACA,QACA,MACe;CACf,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,SAAS,QAAQ,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;CACvE,IAAI,SAAS,OAAO,OAAO,KAAK,IAAI,GAAG,MAAM;CAK7C,OAHe,CAAC,GAAG,QAAQ,CAAC,CACzB,QAAQ,MAAM,OAAO,EAAE,QAAQ,YAAY,QAAQ,CAAC,CACpD,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UACvB,CAAC,CAAC,EAAE,EAAE,QAAQ,WAAW;AACvC;AAEA,SAAS,YACP,IACA,IACA,YACA,KAC+B;CAC/B,MAAM,IAAI,GAAG;CACb,IAAI,IAAI,GAAG,OAAO;EAAE,KAAK;EAAY,MAAM;CAAW;CACtD,MAAM,UAAoB,CAAC;CAC3B,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,KAAK,IAAI,MAAc,CAAC;EAC9B,MAAM,KAAK,IAAI,MAAc,CAAC;EAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,MAAM,KAAK,MAAM,IAAI,IAAI,CAAC;GAChC,GAAG,KAAK,GAAG;GACX,GAAG,KAAK,GAAG;EACb;EACA,MAAM,IAAI,SAAS,IAAI,EAAE;EACzB,IAAI,OAAO,SAAS,CAAC,GAAG,QAAQ,KAAK,CAAC;CACxC;CACA,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAC5B,IAAI,QAAQ,WAAW,GAAG,OAAO;EAAE,KAAK;EAAY,MAAM;CAAW;CACrE,OAAO;EACL,KAAK,QAAQ,KAAK,MAAM,OAAQ,QAAQ,MAAM;EAC9C,MAAM,QAAQ,KAAK,IAAI,QAAQ,SAAS,GAAG,KAAK,MAAM,OAAQ,QAAQ,MAAM,CAAC;CAC/E;AACF;AAEA,SAAS,QAAQ,MAA6B;CAC5C,IAAI,SAAS,KAAA,GAAW,OAAO,KAAK;CACpC,IAAI,IAAI,SAAS;CACjB,aAAa;EACX,IAAK,IAAI,eAAgB;EACzB,IAAI,IAAI;EACR,IAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;EACnC,KAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;EACxC,SAAS,IAAK,MAAM,QAAS,KAAK;CACpC;AACF"}