@tangle-network/agent-eval 0.216.1 → 0.218.1

package/dist/json-BhdsJaOQ.js

@@ -0,0 +1,111 @@
+import { t as AgentEvalError } from "./errors-BcOs5MKW.js";
+//#region src/llm-client/errors.ts
+var LlmCallError = class extends AgentEvalError {
+	status;
+	body;
+	model;
+	constructor(message, status, body, model) {
+		super("judge", message);
+		this.status = status;
+		this.body = body;
+		this.model = model;
+	}
+};
+/** A provider response completed and incurred measurable usage, but its content
+*  could not satisfy the caller's response contract. The response envelope is
+*  retained so accounting can commit the receipt before the error propagates. */
+var LlmResponseError = class extends AgentEvalError {
+	result;
+	constructor(message, result, options) {
+		super("judge", message, options);
+		this.result = result;
+	}
+};
+//#endregion
+//#region src/llm-client/json.ts
+/**
+* Strip a ```json / ``` code fence if the model emitted one.
+* Idempotent for naked JSON. Some models (claude-code via router, certain
+* deepseek models) wrap output even under json_object.
+*/
+function stripFencedJson(raw) {
+	const trimmed = raw.trim();
+	const m = trimmed.match(/^```(?:json)?\s*\n?([\s\S]*?)\n?```\s*$/);
+	return m ? m[1].trim() : trimmed;
+}
+function extractJsonPayload(raw) {
+	const stripped = stripFencedJson(raw);
+	try {
+		JSON.parse(stripped);
+		return stripped;
+	} catch {
+		if (stripped.startsWith("{") || stripped.startsWith("[")) return stripped;
+	}
+	const starts = [...stripped.matchAll(/[[{]/g)].map((match) => match.index).filter((index) => index != null);
+	for (const start of starts) {
+		const candidate = extractBalancedJson(stripped, start);
+		if (!candidate) continue;
+		try {
+			JSON.parse(candidate);
+			return candidate;
+		} catch {}
+	}
+	return stripped;
+}
+function extractBalancedJson(input, start) {
+	const opener = input[start];
+	const closer = opener === "{" ? "}" : opener === "[" ? "]" : null;
+	if (!closer) return null;
+	const stack = [closer];
+	let isInString = false;
+	let isEscaped = false;
+	for (let i = start + 1; i < input.length; i++) {
+		const char = input[i];
+		if (isEscaped) {
+			isEscaped = false;
+			continue;
+		}
+		if (char === "\\") {
+			isEscaped = isInString;
+			continue;
+		}
+		if (char === "\"") {
+			isInString = !isInString;
+			continue;
+		}
+		if (isInString) continue;
+		if (char === "{") stack.push("}");
+		else if (char === "[") stack.push("]");
+		else if (char === stack[stack.length - 1]) {
+			stack.pop();
+			if (stack.length === 0) return input.slice(start, i + 1);
+		}
+	}
+	return null;
+}
+function parseJsonResult(result, jsonPayloadMode) {
+	try {
+		if (result.finishReason === "length") throw new Error(`LLM returned truncated JSON content (model=${result.model}, finishReason=length)`);
+		return parseJsonSafely(result.content, result.model, jsonPayloadMode);
+	} catch (error) {
+		if (error instanceof LlmResponseError) throw error;
+		const cause = error instanceof Error ? error : new Error(String(error));
+		throw new LlmResponseError(cause.message, result, { cause });
+	}
+}
+function parseJsonSafely(content, model, jsonPayloadMode) {
+	const payload = jsonPayloadMode === "exact" ? content : extractJsonPayload(content);
+	try {
+		return JSON.parse(payload);
+	} catch {
+		throw new Error(`LLM returned non-JSON content (model=${model})`);
+	}
+}
+/** A non-array object: the JSON object shape that model and export payloads are narrowed to. */
+function isRecord(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+//#endregion
+export { LlmCallError as a, stripFencedJson as i, isRecord as n, LlmResponseError as o, parseJsonResult as r, extractJsonPayload as t };
+
+//# sourceMappingURL=json-BhdsJaOQ.js.map

package/dist/json-BhdsJaOQ.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"json-BhdsJaOQ.js","names":[],"sources":["../src/llm-client/errors.ts","../src/llm-client/json.ts"],"sourcesContent":["import { AgentEvalError } from '../errors'\nimport type { LlmCallResult } from './contract'\n\nexport class LlmCallError extends AgentEvalError {\n  constructor(\n    message: string,\n    public readonly status: number,\n    public readonly body: string,\n    public readonly model: string,\n  ) {\n    super('judge', message)\n  }\n}\n\n/** A provider response completed and incurred measurable usage, but its content\n *  could not satisfy the caller's response contract. The response envelope is\n *  retained so accounting can commit the receipt before the error propagates. */\nexport class LlmResponseError extends AgentEvalError {\n  constructor(\n    message: string,\n    public readonly result: LlmCallResult,\n    options?: { cause?: unknown },\n  ) {\n    super('judge', message, options)\n  }\n}\n","import type { LlmCallResult, LlmClientOptions } from './contract'\nimport { LlmResponseError } from './errors'\n\n// ─── Public API ─────────────────────────────────────────────────────────\n\n/**\n * Strip a ```json / ``` code fence if the model emitted one.\n * Idempotent for naked JSON. Some models (claude-code via router, certain\n * deepseek models) wrap output even under json_object.\n */\nexport function stripFencedJson(raw: string): string {\n  const trimmed = raw.trim()\n  const m = trimmed.match(/^```(?:json)?\\s*\\n?([\\s\\S]*?)\\n?```\\s*$/)\n  return m ? m[1]!.trim() : trimmed\n}\n\nexport function extractJsonPayload(raw: string): string {\n  const stripped = stripFencedJson(raw)\n  try {\n    JSON.parse(stripped)\n    return stripped\n  } catch {\n    // A response that declares a JSON root must parse as that complete root.\n    // Scanning onward could turn a truncated object into one of its valid nested\n    // arrays or objects and silently change the response schema.\n    if (stripped.startsWith('{') || stripped.startsWith('[')) return stripped\n  }\n\n  // Only prose-leading responses may contain a recoverable JSON payload.\n  const starts = [...stripped.matchAll(/[[{]/g)]\n    .map((match) => match.index)\n    .filter((index) => index != null)\n  for (const start of starts) {\n    const candidate = extractBalancedJson(stripped, start)\n    if (!candidate) continue\n    try {\n      JSON.parse(candidate)\n      return candidate\n    } catch {\n      // Keep scanning; earlier braces may belong to prose.\n    }\n  }\n\n  return stripped\n}\n\nfunction extractBalancedJson(input: string, start: number): string | null {\n  const opener = input[start]\n  const closer = opener === '{' ? '}' : opener === '[' ? ']' : null\n  if (!closer) return null\n\n  const stack: string[] = [closer]\n  let isInString = false\n  let isEscaped = false\n\n  for (let i = start + 1; i < input.length; i++) {\n    const char = input[i]!\n    if (isEscaped) {\n      isEscaped = false\n      continue\n    }\n    if (char === '\\\\') {\n      isEscaped = isInString\n      continue\n    }\n    if (char === '\"') {\n      isInString = !isInString\n      continue\n    }\n    if (isInString) continue\n\n    if (char === '{') stack.push('}')\n    else if (char === '[') stack.push(']')\n    else if (char === stack[stack.length - 1]) {\n      stack.pop()\n      if (stack.length === 0) return input.slice(start, i + 1)\n    }\n  }\n\n  return null\n}\n\nexport function parseJsonResult<T>(\n  result: LlmCallResult,\n  jsonPayloadMode: NonNullable<LlmClientOptions['jsonPayloadMode']>,\n): T {\n  try {\n    if (result.finishReason === 'length') {\n      throw new Error(\n        `LLM returned truncated JSON content (model=${result.model}, finishReason=length)`,\n      )\n    }\n    return parseJsonSafely<T>(result.content, result.model, jsonPayloadMode)\n  } catch (error) {\n    if (error instanceof LlmResponseError) throw error\n    const cause = error instanceof Error ? error : new Error(String(error))\n    throw new LlmResponseError(cause.message, result, { cause })\n  }\n}\n\nfunction parseJsonSafely<T>(\n  content: string,\n  model: string,\n  jsonPayloadMode: NonNullable<LlmClientOptions['jsonPayloadMode']>,\n): T {\n  const payload = jsonPayloadMode === 'exact' ? content : extractJsonPayload(content)\n  try {\n    return JSON.parse(payload) as T\n  } catch {\n    throw new Error(`LLM returned non-JSON content (model=${model})`)\n  }\n}\n\n/** A non-array object: the JSON object shape that model and export payloads are narrowed to. */\nexport function isRecord(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n"],"mappings":";;AAGA,IAAa,eAAb,cAAkC,eAAe;CAG7B;CACA;CACA;CAJlB,YACE,SACA,QACA,MACA,OACA;EACA,MAAM,SAAS,OAAO;EAJN,KAAA,SAAA;EACA,KAAA,OAAA;EACA,KAAA,QAAA;CAGlB;AACF;;;;AAKA,IAAa,mBAAb,cAAsC,eAAe;CAGjC;CAFlB,YACE,SACA,QACA,SACA;EACA,MAAM,SAAS,SAAS,OAAO;EAHf,KAAA,SAAA;CAIlB;AACF;;;;;;;;ACfA,SAAgB,gBAAgB,KAAqB;CACnD,MAAM,UAAU,IAAI,KAAK;CACzB,MAAM,IAAI,QAAQ,MAAM,yCAAyC;CACjE,OAAO,IAAI,EAAE,EAAE,CAAE,KAAK,IAAI;AAC5B;AAEA,SAAgB,mBAAmB,KAAqB;CACtD,MAAM,WAAW,gBAAgB,GAAG;CACpC,IAAI;EACF,KAAK,MAAM,QAAQ;EACnB,OAAO;CACT,QAAQ;EAIN,IAAI,SAAS,WAAW,GAAG,KAAK,SAAS,WAAW,GAAG,GAAG,OAAO;CACnE;CAGA,MAAM,SAAS,CAAC,GAAG,SAAS,SAAS,OAAO,CAAC,CAAC,CAC3C,KAAK,UAAU,MAAM,KAAK,CAAC,CAC3B,QAAQ,UAAU,SAAS,IAAI;CAClC,KAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,YAAY,oBAAoB,UAAU,KAAK;EACrD,IAAI,CAAC,WAAW;EAChB,IAAI;GACF,KAAK,MAAM,SAAS;GACpB,OAAO;EACT,QAAQ,CAER;CACF;CAEA,OAAO;AACT;AAEA,SAAS,oBAAoB,OAAe,OAA8B;CACxE,MAAM,SAAS,MAAM;CACrB,MAAM,SAAS,WAAW,MAAM,MAAM,WAAW,MAAM,MAAM;CAC7D,IAAI,CAAC,QAAQ,OAAO;CAEpB,MAAM,QAAkB,CAAC,MAAM;CAC/B,IAAI,aAAa;CACjB,IAAI,YAAY;CAEhB,KAAK,IAAI,IAAI,QAAQ,GAAG,IAAI,MAAM,QAAQ,KAAK;EAC7C,MAAM,OAAO,MAAM;EACnB,IAAI,WAAW;GACb,YAAY;GACZ;EACF;EACA,IAAI,SAAS,MAAM;GACjB,YAAY;GACZ;EACF;EACA,IAAI,SAAS,MAAK;GAChB,aAAa,CAAC;GACd;EACF;EACA,IAAI,YAAY;EAEhB,IAAI,SAAS,KAAK,MAAM,KAAK,GAAG;OAC3B,IAAI,SAAS,KAAK,MAAM,KAAK,GAAG;OAChC,IAAI,SAAS,MAAM,MAAM,SAAS,IAAI;GACzC,MAAM,IAAI;GACV,IAAI,MAAM,WAAW,GAAG,OAAO,MAAM,MAAM,OAAO,IAAI,CAAC;EACzD;CACF;CAEA,OAAO;AACT;AAEA,SAAgB,gBACd,QACA,iBACG;CACH,IAAI;EACF,IAAI,OAAO,iBAAiB,UAC1B,MAAM,IAAI,MACR,8CAA8C,OAAO,MAAM,uBAC7D;EAEF,OAAO,gBAAmB,OAAO,SAAS,OAAO,OAAO,eAAe;CACzE,SAAS,OAAO;EACd,IAAI,iBAAiB,kBAAkB,MAAM;EAC7C,MAAM,QAAQ,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;EACtE,MAAM,IAAI,iBAAiB,MAAM,SAAS,QAAQ,EAAE,MAAM,CAAC;CAC7D;AACF;AAEA,SAAS,gBACP,SACA,OACA,iBACG;CACH,MAAM,UAAU,oBAAoB,UAAU,UAAU,mBAAmB,OAAO;CAClF,IAAI;EACF,OAAO,KAAK,MAAM,OAAO;CAC3B,QAAQ;EACN,MAAM,IAAI,MAAM,wCAAwC,MAAM,EAAE;CAClE;AACF;;AAGA,SAAgB,SAAS,OAAkD;CACzE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E"}

package/dist/{judge-calibration-BpCy5XQl.js → judge-calibration-BkXMIWA4.js}

@@ -1,6 +1,6 @@
 import { t as interpolatedQuantile } from "./quantile--0yb07pD.js";
-import { t as mulberry32 } from "./random-ByUSnV9w.js";
-import { a as spearmanR, r as pearsonR } from "./descriptive-DboxPZrG.js";
+import { t as mulberry32 } from "./random-DDDcC7YU.js";
+import { a as spearmanR, r as pearsonR } from "./descriptive-CVNy48rR.js";
 //#region src/judge-calibration.ts
 /**
 * Judge calibration — measure judge quality against human gold + bias.
@@ -340,4 +340,4 @@ function percentileBounds(ciLevel) {
 //#endregion
 export { selfPreference as a, positionalBias as i, calibrateJudgeContinuous as n, verbosityBias as o, continuousAgreement as r, calibrateJudge as t };
 
-//# sourceMappingURL=judge-calibration-BpCy5XQl.js.map
+//# sourceMappingURL=judge-calibration-BkXMIWA4.js.map

package/dist/{judge-calibration-BpCy5XQl.js.map → judge-calibration-BkXMIWA4.js.map}

@@ -1 +1 @@
-{"version":3,"file":"judge-calibration-BpCy5XQl.js","names":["quantile"],"sources":["../src/judge-calibration.ts"],"sourcesContent":["/**\n * Judge calibration — measure judge quality against human gold + bias.\n *\n * Workflow:\n *   1. Build a golden set: {itemId, humanScore}[].\n *   2. Run candidate judges; each produces {itemId, score}.\n *   3. `calibrateJudge(golden, candidate)` reports κ + Pearson + MAE.\n *   4. `calibrateJudgeContinuous(golden, candidate)` adds quadratic-weighted\n *      κ over the un-rounded [0,1] scores plus ICC(2,1), Pearson, Spearman,\n *      and bootstrap CIs — use this for fine-grained judges where rounding\n *      to int discards information (e.g. 0.78 vs 0.81 both round to 1 and\n *      look \"perfectly agreed\" to integer κ).\n *   5. Run bias probes (positional, verbosity, self-preference) to\n *      detect systematic score inflation.\n *   6. For N≥2 judges on the same items, `continuousAgreement(scores)`\n *      reports ICC(2,1) + κ_w + Pearson + Spearman with bootstrap CIs.\n *\n * Returns actionable diagnostics, not a single number. Consumers then\n * decide whether to trust the judge, retrain it, or add a tie-breaker.\n */\n\nimport { interpolatedQuantile as quantile } from './math/quantile'\nimport { pearsonR, spearmanR } from './statistics/descriptive'\nimport { mulberry32 } from './statistics/random'\n\nexport interface GoldenItem {\n  itemId: string\n  humanScore: number\n  /** Optional group used for per-group bias audits (e.g. model-of-output family). */\n  group?: string\n}\n\nexport interface CandidateScore {\n  itemId: string\n  score: number\n  /** Optional — enables positional-bias analysis (did order matter?). */\n  positionOfAInput?: 'first' | 'second'\n}\n\nexport interface CalibrationResult {\n  n: number\n  pearson: number\n  /** Cohen's κ with quadratic weights over integer-rounded scores. */\n  kappa: number\n  /** Mean absolute error vs human. */\n  mae: number\n  /** Worst-5 miscalibrations (largest |judge - human|). */\n  worstItems: Array<{ itemId: string; judge: number; human: number; delta: number }>\n}\n\n/**\n * Measure judge quality against human gold labels: computes Cohen's κ, Pearson correlation, and MAE over matched item ids.\n */\nexport function calibrateJudge(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n): CalibrationResult {\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const common = [...map.values()].filter((v) => Number.isFinite(v.j))\n  const n = common.length\n  if (n < 2) {\n    return { n, pearson: NaN, kappa: NaN, mae: NaN, worstItems: [] }\n  }\n  const humans = common.map((c) => c.h)\n  const judges = common.map((c) => c.j)\n  const pearson = pearsonR(humans, judges)\n  const kappa = weightedKappa(humans.map(Math.round), judges.map(Math.round))\n  const absDiffs = common.map((c) => Math.abs(c.j - c.h))\n  const mae = absDiffs.reduce((a, b) => a + b, 0) / n\n  const worst = [...map.entries()]\n    .filter(([, v]) => Number.isFinite(v.j))\n    .map(([itemId, v]) => ({ itemId, judge: v.j, human: v.h, delta: Math.abs(v.j - v.h) }))\n    .sort((a, b) => b.delta - a.delta)\n    .slice(0, 5)\n  return { n, pearson, kappa, mae, worstItems: worst }\n}\n\nexport interface PositionalBiasResult {\n  /**\n   * Score delta (first-position - second-position) averaged across items\n   * presented in both positions. Non-zero = positional bias.\n   */\n  avgDelta: number\n  n: number\n}\n\n/**\n * Feed the same items to the judge twice with A/B swapped and pass all\n * results here. Items that don't appear in both positions are ignored.\n */\nexport function positionalBias(scores: CandidateScore[]): PositionalBiasResult {\n  const pairs = new Map<string, { first?: number; second?: number }>()\n  for (const s of scores) {\n    const slot = pairs.get(s.itemId) ?? {}\n    if (s.positionOfAInput === 'first') slot.first = s.score\n    else if (s.positionOfAInput === 'second') slot.second = s.score\n    pairs.set(s.itemId, slot)\n  }\n  const deltas: number[] = []\n  for (const { first, second } of pairs.values()) {\n    if (first !== undefined && second !== undefined) deltas.push(first - second)\n  }\n  if (deltas.length === 0) return { avgDelta: 0, n: 0 }\n  return { avgDelta: deltas.reduce((a, b) => a + b, 0) / deltas.length, n: deltas.length }\n}\n\nexport interface VerbosityBiasResult {\n  /** Pearson correlation between output length and score. Strong positive = verbosity bias. */\n  pearson: number\n  n: number\n}\n\nexport function verbosityBias(\n  samples: Array<{ outputLen: number; score: number }>,\n): VerbosityBiasResult {\n  const n = samples.length\n  if (n < 3) return { pearson: NaN, n }\n  return {\n    pearson: pearsonR(\n      samples.map((s) => s.outputLen),\n      samples.map((s) => s.score),\n    ),\n    n,\n  }\n}\n\nexport interface SelfPreferenceResult {\n  /** Mean judge score when judge's family matches output's family. */\n  inFamilyMean: number\n  outOfFamilyMean: number\n  deltaMean: number\n  n: number\n}\n\n/**\n * Pass the same scenarios scored with judge-model X grading outputs from\n * model X (in-family) and model Y (out-of-family). Non-zero delta\n * indicates self-preference.\n */\nexport function selfPreference(\n  samples: Array<{ score: number; inFamily: boolean }>,\n): SelfPreferenceResult {\n  const inF = samples.filter((s) => s.inFamily).map((s) => s.score)\n  const outF = samples.filter((s) => !s.inFamily).map((s) => s.score)\n  if (inF.length === 0 || outF.length === 0)\n    return { inFamilyMean: 0, outOfFamilyMean: 0, deltaMean: 0, n: 0 }\n  const inMean = inF.reduce((a, b) => a + b, 0) / inF.length\n  const outMean = outF.reduce((a, b) => a + b, 0) / outF.length\n  return {\n    inFamilyMean: inMean,\n    outOfFamilyMean: outMean,\n    deltaMean: inMean - outMean,\n    n: samples.length,\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\n/** Quadratic weighted Cohen's κ over bounded integer scores. */\nfunction weightedKappa(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return NaN\n  const min = Math.min(...a, ...b)\n  const max = Math.max(...a, ...b)\n  const K = max - min + 1\n  if (K < 2) return 1\n  const observed: number[][] = Array.from({ length: K }, () => new Array(K).fill(0))\n  const rowMarg = new Array(K).fill(0)\n  const colMarg = new Array(K).fill(0)\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]! - min\n    const bi = b[i]! - min\n    const row = observed[ai]!\n    row[bi] = (row[bi] ?? 0) + 1\n    rowMarg[ai]++\n    colMarg[bi]++\n  }\n  let num = 0\n  let den = 0\n  for (let i = 0; i < K; i++) {\n    for (let j = 0; j < K; j++) {\n      const w = (i - j) ** 2 / (K - 1) ** 2\n      const expected = (rowMarg[i] * colMarg[j]) / a.length\n      num += w * observed[i]![j]!\n      den += w * expected\n    }\n  }\n  if (den === 0) return 1\n  return 1 - num / den\n}\n\n// ── Continuous-value inter-rater agreement ──────────────────────────\n//\n// `weightedKappa` above quantises to integers. For [0,1] judges that is\n// lossy: 0.78 and 0.81 both round to 1 and look identical to integer κ.\n// `continuousAgreement` keeps the original magnitudes:\n//\n//   - κ_w   : Cohen's weighted κ generalised to continuous scores. For two\n//             raters, weighted κ = 1 − Σ w_ij·o_ij / Σ w_ij·e_ij collapses\n//             into the closed form below when each \"category\" is the score\n//             itself (Janson & Olsson 2001; Cohen 1968 quadratic weights).\n//   - ICC   : Intraclass Correlation Coefficient, ICC(2,1) per Shrout &\n//             Fleiss (1979) — two-way random effects, absolute agreement,\n//             single rater. The right reliability coefficient when judges\n//             are a random sample of the judge population.\n//   - r, ρ  : Pearson + Spearman for sanity-check. Pearson catches linear\n//             association regardless of scale; Spearman catches monotone\n//             association regardless of distribution.\n//   - CIs   : Bootstrap percentile CIs over items (rows). Bias is preserved\n//             because we resample the full row, i.e. a judge's vector\n//             stays internally consistent within each bootstrap replicate.\n//\n// References:\n//   - Shrout, P. E. & Fleiss, J. L. (1979). Intraclass correlations: uses\n//     in assessing rater reliability. Psychological Bulletin, 86(2), 420.\n//   - McGraw, K. O. & Wong, S. P. (1996). Forming inferences about some\n//     intraclass correlation coefficients. Psychological Methods, 1, 30.\n//   - Janson, H. & Olsson, U. (2001). A measure of agreement for interval\n//     or nominal multivariate observations. Educ. Psychol. Meas., 61, 277.\n\nexport interface ContinuousAgreement {\n  /** Cohen's κ_w with quadratic weights, computed on raw [0,1] scores. */\n  weightedKappa: number\n  /** ICC(2,1): two-way random effects, absolute agreement, single rater. */\n  icc: number\n  /** Pearson product-moment correlation (averaged over rater pairs if N>2). */\n  pearson: number\n  /** Spearman rank correlation (averaged over rater pairs if N>2). */\n  spearman: number\n  /** 95% bootstrap percentile CIs over items. */\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n  /** Number of complete items (no NaN across raters). */\n  n: number\n  /** Number of raters. */\n  raters: number\n}\n\nexport interface ContinuousAgreementOptions {\n  /** Bootstrap iterations. Default 1000. Set to 0 to skip CIs (CI = [NaN, NaN]). */\n  bootstrap?: number\n  /** κ weighting scheme. Default 'quadratic'. */\n  weights?: 'linear' | 'quadratic'\n  /** PRNG seed for reproducible bootstrap. Default 0xC0FFEE. */\n  seed?: number\n  /** Confidence level for percentile CI. Default 0.95. */\n  ciLevel?: number\n}\n\n/**\n * Inter-rater agreement on continuous (typically [0,1]) scores.\n *\n * `scores` has shape [n_items][n_raters]. Rows with any non-finite entry\n * are dropped. Returns NaN metrics if fewer than 2 raters or 2 complete\n * items remain.\n */\nexport function continuousAgreement(\n  scores: number[][],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousAgreement {\n  const bootstrap = opts.bootstrap ?? 1000\n  const weights = opts.weights ?? 'quadratic'\n  const seed = opts.seed ?? 0xc0ffee\n  const ciLevel = opts.ciLevel ?? 0.95\n\n  const matrix = scores.filter((row) => row.length >= 2 && row.every((v) => Number.isFinite(v)))\n  const raters = matrix[0]?.length ?? 0\n  // All rows must have the same rater count, else drop the offenders.\n  const clean = matrix.filter((row) => row.length === raters)\n  const nClean = clean.length\n\n  if (nClean < 2 || raters < 2) {\n    return {\n      weightedKappa: NaN,\n      icc: NaN,\n      pearson: NaN,\n      spearman: NaN,\n      ci: { icc: [NaN, NaN], weightedKappa: [NaN, NaN] },\n      n: nClean,\n      raters,\n    }\n  }\n\n  const kappa = continuousWeightedKappa(clean, weights)\n  const icc = icc21(clean)\n  const pearson = avgPairwise(clean, pearsonR)\n  const spearman = avgPairwise(clean, spearmanR)\n\n  const ciIcc: [number, number] = [NaN, NaN]\n  const ciKappa: [number, number] = [NaN, NaN]\n  if (bootstrap > 0) {\n    const rng = mulberry32(seed)\n    const iccs: number[] = []\n    const kappas: number[] = []\n    for (let b = 0; b < bootstrap; b++) {\n      const sample: number[][] = new Array(nClean)\n      for (let i = 0; i < nClean; i++) {\n        sample[i] = clean[Math.floor(rng() * nClean)]!\n      }\n      const iccB = icc21(sample)\n      const kB = continuousWeightedKappa(sample, weights)\n      if (Number.isFinite(iccB)) iccs.push(iccB)\n      if (Number.isFinite(kB)) kappas.push(kB)\n    }\n    const [lo, hi] = percentileBounds(ciLevel)\n    if (iccs.length > 0) {\n      iccs.sort((a, b) => a - b)\n      ciIcc[0] = quantile(iccs, lo)\n      ciIcc[1] = quantile(iccs, hi)\n    }\n    if (kappas.length > 0) {\n      kappas.sort((a, b) => a - b)\n      ciKappa[0] = quantile(kappas, lo)\n      ciKappa[1] = quantile(kappas, hi)\n    }\n  }\n\n  return {\n    weightedKappa: kappa,\n    icc,\n    pearson,\n    spearman,\n    ci: { icc: ciIcc, weightedKappa: ciKappa },\n    n: nClean,\n    raters,\n  }\n}\n\nexport interface ContinuousCalibrationResult extends CalibrationResult {\n  /** Cohen's κ_w computed on raw (un-rounded) scores. */\n  weightedKappaContinuous: number\n  /** ICC(2,1) treating golden + candidate as two raters. */\n  icc: number\n  spearman: number\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n}\n\n/**\n * Extends `calibrateJudge` with continuous-value agreement metrics while\n * retaining its base calibration summary.\n */\nexport function calibrateJudgeContinuous(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousCalibrationResult {\n  const base = calibrateJudge(golden, candidate)\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const rows: number[][] = []\n  for (const v of map.values()) {\n    if (Number.isFinite(v.j)) rows.push([v.h, v.j])\n  }\n  const agreement = continuousAgreement(rows, opts)\n  return {\n    ...base,\n    weightedKappaContinuous: agreement.weightedKappa,\n    icc: agreement.icc,\n    spearman: agreement.spearman,\n    ci: agreement.ci,\n  }\n}\n\n// ── Continuous-agreement internals ──────────────────────────────────\n\n/**\n * Quadratic-weighted κ on continuous scores. With weights w(x,y) = (x-y)^2\n * (or |x-y| for linear) the formula collapses to:\n *\n *   κ_w = 1 − E_obs[w] / E_exp[w]\n *\n * where E_obs averages w over paired (a_i, b_i) and E_exp averages w over\n * the independent product distribution (sum_{i,j} w(a_i, b_j) / n^2).\n * The normalisation by (max-min)^2 in the integer version cancels in the\n * ratio, so we don't need it here. Generalises to N raters by averaging κ_w\n * over all rater pairs (mean pairwise weighted agreement).\n */\nfunction continuousWeightedKappa(rows: number[][], scheme: 'linear' | 'quadratic'): number {\n  if (rows.length === 0) return NaN\n  const raters = rows[0]!.length\n  if (raters < 2) return NaN\n  const wFn =\n    scheme === 'linear'\n      ? (x: number, y: number) => Math.abs(x - y)\n      : (x: number, y: number) => (x - y) ** 2\n  let sum = 0\n  let pairs = 0\n  for (let r1 = 0; r1 < raters; r1++) {\n    for (let r2 = r1 + 1; r2 < raters; r2++) {\n      const a = rows.map((row) => row[r1]!)\n      const b = rows.map((row) => row[r2]!)\n      const n = a.length\n      let obs = 0\n      for (let i = 0; i < n; i++) obs += wFn(a[i]!, b[i]!)\n      obs /= n\n      // Expected under independence: average of w over all i,j cross pairs.\n      let exp = 0\n      for (let i = 0; i < n; i++) {\n        for (let j = 0; j < n; j++) exp += wFn(a[i]!, b[j]!)\n      }\n      exp /= n * n\n      if (exp === 0) {\n        // Degenerate: at least one rater is constant. Convention: perfect\n        // agreement iff observed disagreement is also zero, else 0.\n        sum += obs === 0 ? 1 : 0\n      } else {\n        sum += 1 - obs / exp\n      }\n      pairs++\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\n/**\n * ICC(2,1) — two-way random effects, absolute agreement, single rater.\n *\n *   ICC(2,1) = (MSR − MSE) / (MSR + (k−1)·MSE + k·(MSC − MSE)/n)\n *\n * where MSR = between-rows MS, MSC = between-columns MS, MSE = residual MS,\n * n = rows (items), k = columns (raters).\n */\nfunction icc21(rows: number[][]): number {\n  const n = rows.length\n  if (n < 2) return NaN\n  const k = rows[0]!.length\n  if (k < 2) return NaN\n\n  // Row means.\n  const rowMeans = rows.map((row) => row.reduce((s, v) => s + v, 0) / k)\n  // Column means.\n  const colMeans = new Array(k).fill(0)\n  for (let j = 0; j < k; j++) {\n    let s = 0\n    for (let i = 0; i < n; i++) s += rows[i]![j]!\n    colMeans[j] = s / n\n  }\n  // Grand mean.\n  let grand = 0\n  for (let i = 0; i < n; i++) grand += rowMeans[i]!\n  grand /= n\n\n  let ssR = 0\n  for (let i = 0; i < n; i++) ssR += (rowMeans[i]! - grand) ** 2\n  ssR *= k\n  let ssC = 0\n  for (let j = 0; j < k; j++) ssC += (colMeans[j]! - grand) ** 2\n  ssC *= n\n  let ssT = 0\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < k; j++) ssT += (rows[i]![j]! - grand) ** 2\n  }\n  const ssE = ssT - ssR - ssC\n\n  const dfR = n - 1\n  const dfC = k - 1\n  const dfE = (n - 1) * (k - 1)\n  const msR = ssR / dfR\n  const msC = ssC / dfC\n  const msE = dfE > 0 ? ssE / dfE : 0\n\n  const denom = msR + (k - 1) * msE + (k * (msC - msE)) / n\n  if (denom === 0) {\n    // Degenerate (all values equal): convention = 1 if rows identical,\n    // else 0. Identical rows ⇒ msR = 0 too, so msR − msE near 0.\n    return msR === 0 && msE === 0 ? 1 : 0\n  }\n  return (msR - msE) / denom\n}\n\n/** Average pairwise statistic over all rater pairs. */\nfunction avgPairwise(rows: number[][], fn: (a: number[], b: number[]) => number): number {\n  const k = rows[0]?.length ?? 0\n  if (k < 2) return NaN\n  let sum = 0\n  let pairs = 0\n  for (let i = 0; i < k; i++) {\n    for (let j = i + 1; j < k; j++) {\n      const a = rows.map((row) => row[i]!)\n      const b = rows.map((row) => row[j]!)\n      const r = fn(a, b)\n      if (Number.isFinite(r)) {\n        sum += r\n        pairs++\n      }\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\nfunction percentileBounds(ciLevel: number): [number, number] {\n  const tail = (1 - ciLevel) / 2\n  return [tail, 1 - tail]\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAqDA,SAAgB,eACd,QACA,WACmB;CACnB,MAAM,sBAAM,IAAI,IAAsC;CACtD,KAAK,MAAM,KAAK,QAAQ,IAAI,IAAI,EAAE,QAAQ;EAAE,GAAG,EAAE;EAAY,GAAG;CAAI,CAAC;CACrE,KAAK,MAAM,KAAK,WAAW;EACzB,MAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;EAC9B,IAAI,OAAO,MAAM,IAAI,EAAE;CACzB;CACA,MAAM,SAAS,CAAC,GAAG,IAAI,OAAO,CAAC,CAAC,CAAC,QAAQ,MAAM,OAAO,SAAS,EAAE,CAAC,CAAC;CACnE,MAAM,IAAI,OAAO;CACjB,IAAI,IAAI,GACN,OAAO;EAAE;EAAG,SAAS;EAAK,OAAO;EAAK,KAAK;EAAK,YAAY,CAAC;CAAE;CAEjE,MAAM,SAAS,OAAO,KAAK,MAAM,EAAE,CAAC;CACpC,MAAM,SAAS,OAAO,KAAK,MAAM,EAAE,CAAC;CAUpC,OAAO;EAAE;EAAG,SATI,SAAS,QAAQ,MASf;EAAG,OARP,cAAc,OAAO,IAAI,KAAK,KAAK,GAAG,OAAO,IAAI,KAAK,KAAK,CAQhD;EAAG,KAPX,OAAO,KAAK,MAAM,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC,CAClC,CAAC,CAAC,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;EAMjB,YALnB,CAAC,GAAG,IAAI,QAAQ,CAAC,CAAC,CAC7B,QAAQ,GAAG,OAAO,OAAO,SAAS,EAAE,CAAC,CAAC,CAAC,CACvC,KAAK,CAAC,QAAQ,QAAQ;GAAE;GAAQ,OAAO,EAAE;GAAG,OAAO,EAAE;GAAG,OAAO,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC;EAAE,EAAE,CAAC,CACtF,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,CAAC,CACjC,MAAM,GAAG,CACqC;CAAE;AACrD;;;;;AAeA,SAAgB,eAAe,QAAgD;CAC7E,MAAM,wBAAQ,IAAI,IAAiD;CACnE,KAAK,MAAM,KAAK,QAAQ;EACtB,MAAM,OAAO,MAAM,IAAI,EAAE,MAAM,KAAK,CAAC;EACrC,IAAI,EAAE,qBAAqB,SAAS,KAAK,QAAQ,EAAE;OAC9C,IAAI,EAAE,qBAAqB,UAAU,KAAK,SAAS,EAAE;EAC1D,MAAM,IAAI,EAAE,QAAQ,IAAI;CAC1B;CACA,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,EAAE,OAAO,YAAY,MAAM,OAAO,GAC3C,IAAI,UAAU,KAAA,KAAa,WAAW,KAAA,GAAW,OAAO,KAAK,QAAQ,MAAM;CAE7E,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,UAAU;EAAG,GAAG;CAAE;CACpD,OAAO;EAAE,UAAU,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;EAAQ,GAAG,OAAO;CAAO;AACzF;AAQA,SAAgB,cACd,SACqB;CACrB,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI,GAAG,OAAO;EAAE,SAAS;EAAK;CAAE;CACpC,OAAO;EACL,SAAS,SACP,QAAQ,KAAK,MAAM,EAAE,SAAS,GAC9B,QAAQ,KAAK,MAAM,EAAE,KAAK,CAC5B;EACA;CACF;AACF;;;;;;AAeA,SAAgB,eACd,SACsB;CACtB,MAAM,MAAM,QAAQ,QAAQ,MAAM,EAAE,QAAQ,CAAC,CAAC,KAAK,MAAM,EAAE,KAAK;CAChE,MAAM,OAAO,QAAQ,QAAQ,MAAM,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,MAAM,EAAE,KAAK;CAClE,IAAI,IAAI,WAAW,KAAK,KAAK,WAAW,GACtC,OAAO;EAAE,cAAc;EAAG,iBAAiB;EAAG,WAAW;EAAG,GAAG;CAAE;CACnE,MAAM,SAAS,IAAI,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,IAAI;CACpD,MAAM,UAAU,KAAK,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK;CACvD,OAAO;EACL,cAAc;EACd,iBAAiB;EACjB,WAAW,SAAS;EACpB,GAAG,QAAQ;CACb;AACF;;AAKA,SAAS,cAAc,GAAa,GAAqB;CACvD,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,GAAG,OAAO;CACpD,MAAM,MAAM,KAAK,IAAI,GAAG,GAAG,GAAG,CAAC;CAE/B,MAAM,IADM,KAAK,IAAI,GAAG,GAAG,GAAG,CAClB,IAAI,MAAM;CACtB,IAAI,IAAI,GAAG,OAAO;CAClB,MAAM,WAAuB,MAAM,KAAK,EAAE,QAAQ,EAAE,SAAS,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;CACjF,MAAM,UAAU,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACnC,MAAM,UAAU,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACnC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;EACjC,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,MAAM,SAAS;EACrB,IAAI,OAAO,IAAI,OAAO,KAAK;EAC3B,QAAQ,GAAG;EACX,QAAQ,GAAG;CACb;CACA,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,IAAI,MAAM,KAAK,IAAI,MAAM;EACpC,MAAM,WAAY,QAAQ,KAAK,QAAQ,KAAM,EAAE;EAC/C,OAAO,IAAI,SAAS,EAAE,CAAE;EACxB,OAAO,IAAI;CACb;CAEF,IAAI,QAAQ,GAAG,OAAO;CACtB,OAAO,IAAI,MAAM;AACnB;;;;;;;;AAqEA,SAAgB,oBACd,QACA,OAAmC,CAAC,GACf;CACrB,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,UAAU,KAAK,WAAW;CAChC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,UAAU,KAAK,WAAW;CAEhC,MAAM,SAAS,OAAO,QAAQ,QAAQ,IAAI,UAAU,KAAK,IAAI,OAAO,MAAM,OAAO,SAAS,CAAC,CAAC,CAAC;CAC7F,MAAM,SAAS,OAAO,EAAE,EAAE,UAAU;CAEpC,MAAM,QAAQ,OAAO,QAAQ,QAAQ,IAAI,WAAW,MAAM;CAC1D,MAAM,SAAS,MAAM;CAErB,IAAI,SAAS,KAAK,SAAS,GACzB,OAAO;EACL,eAAe;EACf,KAAK;EACL,SAAS;EACT,UAAU;EACV,IAAI;GAAE,KAAK,CAAC,KAAK,GAAG;GAAG,eAAe,CAAC,KAAK,GAAG;EAAE;EACjD,GAAG;EACH;CACF;CAGF,MAAM,QAAQ,wBAAwB,OAAO,OAAO;CACpD,MAAM,MAAM,MAAM,KAAK;CACvB,MAAM,UAAU,YAAY,OAAO,QAAQ;CAC3C,MAAM,WAAW,YAAY,OAAO,SAAS;CAE7C,MAAM,QAA0B,CAAC,KAAK,GAAG;CACzC,MAAM,UAA4B,CAAC,KAAK,GAAG;CAC3C,IAAI,YAAY,GAAG;EACjB,MAAM,MAAM,WAAW,IAAI;EAC3B,MAAM,OAAiB,CAAC;EACxB,MAAM,SAAmB,CAAC;EAC1B,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAClC,MAAM,SAAqB,IAAI,MAAM,MAAM;GAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAC1B,OAAO,KAAK,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM;GAE7C,MAAM,OAAO,MAAM,MAAM;GACzB,MAAM,KAAK,wBAAwB,QAAQ,OAAO;GAClD,IAAI,OAAO,SAAS,IAAI,GAAG,KAAK,KAAK,IAAI;GACzC,IAAI,OAAO,SAAS,EAAE,GAAG,OAAO,KAAK,EAAE;EACzC;EACA,MAAM,CAAC,IAAI,MAAM,iBAAiB,OAAO;EACzC,IAAI,KAAK,SAAS,GAAG;GACnB,KAAK,MAAM,GAAG,MAAM,IAAI,CAAC;GACzB,MAAM,KAAKA,qBAAS,MAAM,EAAE;GAC5B,MAAM,KAAKA,qBAAS,MAAM,EAAE;EAC9B;EACA,IAAI,OAAO,SAAS,GAAG;GACrB,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;GAC3B,QAAQ,KAAKA,qBAAS,QAAQ,EAAE;GAChC,QAAQ,KAAKA,qBAAS,QAAQ,EAAE;EAClC;CACF;CAEA,OAAO;EACL,eAAe;EACf;EACA;EACA;EACA,IAAI;GAAE,KAAK;GAAO,eAAe;EAAQ;EACzC,GAAG;EACH;CACF;AACF;;;;;AAkBA,SAAgB,yBACd,QACA,WACA,OAAmC,CAAC,GACP;CAC7B,MAAM,OAAO,eAAe,QAAQ,SAAS;CAC7C,MAAM,sBAAM,IAAI,IAAsC;CACtD,KAAK,MAAM,KAAK,QAAQ,IAAI,IAAI,EAAE,QAAQ;EAAE,GAAG,EAAE;EAAY,GAAG;CAAI,CAAC;CACrE,KAAK,MAAM,KAAK,WAAW;EACzB,MAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;EAC9B,IAAI,OAAO,MAAM,IAAI,EAAE;CACzB;CACA,MAAM,OAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,IAAI,OAAO,GACzB,IAAI,OAAO,SAAS,EAAE,CAAC,GAAG,KAAK,KAAK,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;CAEhD,MAAM,YAAY,oBAAoB,MAAM,IAAI;CAChD,OAAO;EACL,GAAG;EACH,yBAAyB,UAAU;EACnC,KAAK,UAAU;EACf,UAAU,UAAU;EACpB,IAAI,UAAU;CAChB;AACF;;;;;;;;;;;;;AAgBA,SAAS,wBAAwB,MAAkB,QAAwC;CACzF,IAAI,KAAK,WAAW,GAAG,OAAO;CAC9B,MAAM,SAAS,KAAK,EAAE,CAAE;CACxB,IAAI,SAAS,GAAG,OAAO;CACvB,MAAM,MACJ,WAAW,YACN,GAAW,MAAc,KAAK,IAAI,IAAI,CAAC,KACvC,GAAW,OAAe,IAAI,MAAM;CAC3C,IAAI,MAAM;CACV,IAAI,QAAQ;CACZ,KAAK,IAAI,KAAK,GAAG,KAAK,QAAQ,MAC5B,KAAK,IAAI,KAAK,KAAK,GAAG,KAAK,QAAQ,MAAM;EACvC,MAAM,IAAI,KAAK,KAAK,QAAQ,IAAI,GAAI;EACpC,MAAM,IAAI,KAAK,KAAK,QAAQ,IAAI,GAAI;EACpC,MAAM,IAAI,EAAE;EACZ,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,OAAO,IAAI,EAAE,IAAK,EAAE,EAAG;EACnD,OAAO;EAEP,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,OAAO,IAAI,EAAE,IAAK,EAAE,EAAG;EAErD,OAAO,IAAI;EACX,IAAI,QAAQ,GAGV,OAAO,QAAQ,IAAI,IAAI;OAEvB,OAAO,IAAI,MAAM;EAEnB;CACF;CAEF,OAAO,UAAU,IAAI,MAAM,MAAM;AACnC;;;;;;;;;AAUA,SAAS,MAAM,MAA0B;CACvC,MAAM,IAAI,KAAK;CACf,IAAI,IAAI,GAAG,OAAO;CAClB,MAAM,IAAI,KAAK,EAAE,CAAE;CACnB,IAAI,IAAI,GAAG,OAAO;CAGlB,MAAM,WAAW,KAAK,KAAK,QAAQ,IAAI,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC;CAErE,MAAM,WAAW,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACpC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,IAAI,IAAI;EACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,KAAK,KAAK,EAAE,CAAE;EAC1C,SAAS,KAAK,IAAI;CACpB;CAEA,IAAI,QAAQ;CACZ,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,SAAS,SAAS;CAC9C,SAAS;CAET,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,SAAS,KAAM,UAAU;CAC7D,OAAO;CACP,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,SAAS,KAAM,UAAU;CAC7D,OAAO;CACP,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,KAAK,EAAE,CAAE,KAAM,UAAU;CAE/D,MAAM,MAAM,MAAM,MAAM;CAExB,MAAM,MAAM,IAAI;CAChB,MAAM,MAAM,IAAI;CAChB,MAAM,OAAO,IAAI,MAAM,IAAI;CAC3B,MAAM,MAAM,MAAM;CAClB,MAAM,MAAM,MAAM;CAClB,MAAM,MAAM,MAAM,IAAI,MAAM,MAAM;CAElC,MAAM,QAAQ,OAAO,IAAI,KAAK,MAAO,KAAK,MAAM,OAAQ;CACxD,IAAI,UAAU,GAGZ,OAAO,QAAQ,KAAK,QAAQ,IAAI,IAAI;CAEtC,QAAQ,MAAM,OAAO;AACvB;;AAGA,SAAS,YAAY,MAAkB,IAAkD;CACvF,MAAM,IAAI,KAAK,EAAE,EAAE,UAAU;CAC7B,IAAI,IAAI,GAAG,OAAO;CAClB,IAAI,MAAM;CACV,IAAI,QAAQ;CACZ,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAG9B,MAAM,IAAI,GAFA,KAAK,KAAK,QAAQ,IAAI,EAEnB,GADH,KAAK,KAAK,QAAQ,IAAI,EAChB,CAAC;EACjB,IAAI,OAAO,SAAS,CAAC,GAAG;GACtB,OAAO;GACP;EACF;CACF;CAEF,OAAO,UAAU,IAAI,MAAM,MAAM;AACnC;AAEA,SAAS,iBAAiB,SAAmC;CAC3D,MAAM,QAAQ,IAAI,WAAW;CAC7B,OAAO,CAAC,MAAM,IAAI,IAAI;AACxB"}
+{"version":3,"file":"judge-calibration-BkXMIWA4.js","names":["quantile"],"sources":["../src/judge-calibration.ts"],"sourcesContent":["/**\n * Judge calibration — measure judge quality against human gold + bias.\n *\n * Workflow:\n *   1. Build a golden set: {itemId, humanScore}[].\n *   2. Run candidate judges; each produces {itemId, score}.\n *   3. `calibrateJudge(golden, candidate)` reports κ + Pearson + MAE.\n *   4. `calibrateJudgeContinuous(golden, candidate)` adds quadratic-weighted\n *      κ over the un-rounded [0,1] scores plus ICC(2,1), Pearson, Spearman,\n *      and bootstrap CIs — use this for fine-grained judges where rounding\n *      to int discards information (e.g. 0.78 vs 0.81 both round to 1 and\n *      look \"perfectly agreed\" to integer κ).\n *   5. Run bias probes (positional, verbosity, self-preference) to\n *      detect systematic score inflation.\n *   6. For N≥2 judges on the same items, `continuousAgreement(scores)`\n *      reports ICC(2,1) + κ_w + Pearson + Spearman with bootstrap CIs.\n *\n * Returns actionable diagnostics, not a single number. Consumers then\n * decide whether to trust the judge, retrain it, or add a tie-breaker.\n */\n\nimport { interpolatedQuantile as quantile } from './math/quantile'\nimport { pearsonR, spearmanR } from './statistics/descriptive'\nimport { mulberry32 } from './statistics/random'\n\nexport interface GoldenItem {\n  itemId: string\n  humanScore: number\n  /** Optional group used for per-group bias audits (e.g. model-of-output family). */\n  group?: string\n}\n\nexport interface CandidateScore {\n  itemId: string\n  score: number\n  /** Optional — enables positional-bias analysis (did order matter?). */\n  positionOfAInput?: 'first' | 'second'\n}\n\nexport interface CalibrationResult {\n  n: number\n  pearson: number\n  /** Cohen's κ with quadratic weights over integer-rounded scores. */\n  kappa: number\n  /** Mean absolute error vs human. */\n  mae: number\n  /** Worst-5 miscalibrations (largest |judge - human|). */\n  worstItems: Array<{ itemId: string; judge: number; human: number; delta: number }>\n}\n\n/**\n * Measure judge quality against human gold labels: computes Cohen's κ, Pearson correlation, and MAE over matched item ids.\n */\nexport function calibrateJudge(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n): CalibrationResult {\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const common = [...map.values()].filter((v) => Number.isFinite(v.j))\n  const n = common.length\n  if (n < 2) {\n    return { n, pearson: NaN, kappa: NaN, mae: NaN, worstItems: [] }\n  }\n  const humans = common.map((c) => c.h)\n  const judges = common.map((c) => c.j)\n  const pearson = pearsonR(humans, judges)\n  const kappa = weightedKappa(humans.map(Math.round), judges.map(Math.round))\n  const absDiffs = common.map((c) => Math.abs(c.j - c.h))\n  const mae = absDiffs.reduce((a, b) => a + b, 0) / n\n  const worst = [...map.entries()]\n    .filter(([, v]) => Number.isFinite(v.j))\n    .map(([itemId, v]) => ({ itemId, judge: v.j, human: v.h, delta: Math.abs(v.j - v.h) }))\n    .sort((a, b) => b.delta - a.delta)\n    .slice(0, 5)\n  return { n, pearson, kappa, mae, worstItems: worst }\n}\n\nexport interface PositionalBiasResult {\n  /**\n   * Score delta (first-position - second-position) averaged across items\n   * presented in both positions. Non-zero = positional bias.\n   */\n  avgDelta: number\n  n: number\n}\n\n/**\n * Feed the same items to the judge twice with A/B swapped and pass all\n * results here. Items that don't appear in both positions are ignored.\n */\nexport function positionalBias(scores: CandidateScore[]): PositionalBiasResult {\n  const pairs = new Map<string, { first?: number; second?: number }>()\n  for (const s of scores) {\n    const slot = pairs.get(s.itemId) ?? {}\n    if (s.positionOfAInput === 'first') slot.first = s.score\n    else if (s.positionOfAInput === 'second') slot.second = s.score\n    pairs.set(s.itemId, slot)\n  }\n  const deltas: number[] = []\n  for (const { first, second } of pairs.values()) {\n    if (first !== undefined && second !== undefined) deltas.push(first - second)\n  }\n  if (deltas.length === 0) return { avgDelta: 0, n: 0 }\n  return { avgDelta: deltas.reduce((a, b) => a + b, 0) / deltas.length, n: deltas.length }\n}\n\nexport interface VerbosityBiasResult {\n  /** Pearson correlation between output length and score. Strong positive = verbosity bias. */\n  pearson: number\n  n: number\n}\n\nexport function verbosityBias(\n  samples: Array<{ outputLen: number; score: number }>,\n): VerbosityBiasResult {\n  const n = samples.length\n  if (n < 3) return { pearson: NaN, n }\n  return {\n    pearson: pearsonR(\n      samples.map((s) => s.outputLen),\n      samples.map((s) => s.score),\n    ),\n    n,\n  }\n}\n\nexport interface SelfPreferenceResult {\n  /** Mean judge score when judge's family matches output's family. */\n  inFamilyMean: number\n  outOfFamilyMean: number\n  deltaMean: number\n  n: number\n}\n\n/**\n * Pass the same scenarios scored with judge-model X grading outputs from\n * model X (in-family) and model Y (out-of-family). Non-zero delta\n * indicates self-preference.\n */\nexport function selfPreference(\n  samples: Array<{ score: number; inFamily: boolean }>,\n): SelfPreferenceResult {\n  const inF = samples.filter((s) => s.inFamily).map((s) => s.score)\n  const outF = samples.filter((s) => !s.inFamily).map((s) => s.score)\n  if (inF.length === 0 || outF.length === 0)\n    return { inFamilyMean: 0, outOfFamilyMean: 0, deltaMean: 0, n: 0 }\n  const inMean = inF.reduce((a, b) => a + b, 0) / inF.length\n  const outMean = outF.reduce((a, b) => a + b, 0) / outF.length\n  return {\n    inFamilyMean: inMean,\n    outOfFamilyMean: outMean,\n    deltaMean: inMean - outMean,\n    n: samples.length,\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\n/** Quadratic weighted Cohen's κ over bounded integer scores. */\nfunction weightedKappa(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return NaN\n  const min = Math.min(...a, ...b)\n  const max = Math.max(...a, ...b)\n  const K = max - min + 1\n  if (K < 2) return 1\n  const observed: number[][] = Array.from({ length: K }, () => new Array(K).fill(0))\n  const rowMarg = new Array(K).fill(0)\n  const colMarg = new Array(K).fill(0)\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]! - min\n    const bi = b[i]! - min\n    const row = observed[ai]!\n    row[bi] = (row[bi] ?? 0) + 1\n    rowMarg[ai]++\n    colMarg[bi]++\n  }\n  let num = 0\n  let den = 0\n  for (let i = 0; i < K; i++) {\n    for (let j = 0; j < K; j++) {\n      const w = (i - j) ** 2 / (K - 1) ** 2\n      const expected = (rowMarg[i] * colMarg[j]) / a.length\n      num += w * observed[i]![j]!\n      den += w * expected\n    }\n  }\n  if (den === 0) return 1\n  return 1 - num / den\n}\n\n// ── Continuous-value inter-rater agreement ──────────────────────────\n//\n// `weightedKappa` above quantises to integers. For [0,1] judges that is\n// lossy: 0.78 and 0.81 both round to 1 and look identical to integer κ.\n// `continuousAgreement` keeps the original magnitudes:\n//\n//   - κ_w   : Cohen's weighted κ generalised to continuous scores. For two\n//             raters, weighted κ = 1 − Σ w_ij·o_ij / Σ w_ij·e_ij collapses\n//             into the closed form below when each \"category\" is the score\n//             itself (Janson & Olsson 2001; Cohen 1968 quadratic weights).\n//   - ICC   : Intraclass Correlation Coefficient, ICC(2,1) per Shrout &\n//             Fleiss (1979) — two-way random effects, absolute agreement,\n//             single rater. The right reliability coefficient when judges\n//             are a random sample of the judge population.\n//   - r, ρ  : Pearson + Spearman for sanity-check. Pearson catches linear\n//             association regardless of scale; Spearman catches monotone\n//             association regardless of distribution.\n//   - CIs   : Bootstrap percentile CIs over items (rows). Bias is preserved\n//             because we resample the full row, i.e. a judge's vector\n//             stays internally consistent within each bootstrap replicate.\n//\n// References:\n//   - Shrout, P. E. & Fleiss, J. L. (1979). Intraclass correlations: uses\n//     in assessing rater reliability. Psychological Bulletin, 86(2), 420.\n//   - McGraw, K. O. & Wong, S. P. (1996). Forming inferences about some\n//     intraclass correlation coefficients. Psychological Methods, 1, 30.\n//   - Janson, H. & Olsson, U. (2001). A measure of agreement for interval\n//     or nominal multivariate observations. Educ. Psychol. Meas., 61, 277.\n\nexport interface ContinuousAgreement {\n  /** Cohen's κ_w with quadratic weights, computed on raw [0,1] scores. */\n  weightedKappa: number\n  /** ICC(2,1): two-way random effects, absolute agreement, single rater. */\n  icc: number\n  /** Pearson product-moment correlation (averaged over rater pairs if N>2). */\n  pearson: number\n  /** Spearman rank correlation (averaged over rater pairs if N>2). */\n  spearman: number\n  /** 95% bootstrap percentile CIs over items. */\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n  /** Number of complete items (no NaN across raters). */\n  n: number\n  /** Number of raters. */\n  raters: number\n}\n\nexport interface ContinuousAgreementOptions {\n  /** Bootstrap iterations. Default 1000. Set to 0 to skip CIs (CI = [NaN, NaN]). */\n  bootstrap?: number\n  /** κ weighting scheme. Default 'quadratic'. */\n  weights?: 'linear' | 'quadratic'\n  /** PRNG seed for reproducible bootstrap. Default 0xC0FFEE. */\n  seed?: number\n  /** Confidence level for percentile CI. Default 0.95. */\n  ciLevel?: number\n}\n\n/**\n * Inter-rater agreement on continuous (typically [0,1]) scores.\n *\n * `scores` has shape [n_items][n_raters]. Rows with any non-finite entry\n * are dropped. Returns NaN metrics if fewer than 2 raters or 2 complete\n * items remain.\n */\nexport function continuousAgreement(\n  scores: number[][],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousAgreement {\n  const bootstrap = opts.bootstrap ?? 1000\n  const weights = opts.weights ?? 'quadratic'\n  const seed = opts.seed ?? 0xc0ffee\n  const ciLevel = opts.ciLevel ?? 0.95\n\n  const matrix = scores.filter((row) => row.length >= 2 && row.every((v) => Number.isFinite(v)))\n  const raters = matrix[0]?.length ?? 0\n  // All rows must have the same rater count, else drop the offenders.\n  const clean = matrix.filter((row) => row.length === raters)\n  const nClean = clean.length\n\n  if (nClean < 2 || raters < 2) {\n    return {\n      weightedKappa: NaN,\n      icc: NaN,\n      pearson: NaN,\n      spearman: NaN,\n      ci: { icc: [NaN, NaN], weightedKappa: [NaN, NaN] },\n      n: nClean,\n      raters,\n    }\n  }\n\n  const kappa = continuousWeightedKappa(clean, weights)\n  const icc = icc21(clean)\n  const pearson = avgPairwise(clean, pearsonR)\n  const spearman = avgPairwise(clean, spearmanR)\n\n  const ciIcc: [number, number] = [NaN, NaN]\n  const ciKappa: [number, number] = [NaN, NaN]\n  if (bootstrap > 0) {\n    const rng = mulberry32(seed)\n    const iccs: number[] = []\n    const kappas: number[] = []\n    for (let b = 0; b < bootstrap; b++) {\n      const sample: number[][] = new Array(nClean)\n      for (let i = 0; i < nClean; i++) {\n        sample[i] = clean[Math.floor(rng() * nClean)]!\n      }\n      const iccB = icc21(sample)\n      const kB = continuousWeightedKappa(sample, weights)\n      if (Number.isFinite(iccB)) iccs.push(iccB)\n      if (Number.isFinite(kB)) kappas.push(kB)\n    }\n    const [lo, hi] = percentileBounds(ciLevel)\n    if (iccs.length > 0) {\n      iccs.sort((a, b) => a - b)\n      ciIcc[0] = quantile(iccs, lo)\n      ciIcc[1] = quantile(iccs, hi)\n    }\n    if (kappas.length > 0) {\n      kappas.sort((a, b) => a - b)\n      ciKappa[0] = quantile(kappas, lo)\n      ciKappa[1] = quantile(kappas, hi)\n    }\n  }\n\n  return {\n    weightedKappa: kappa,\n    icc,\n    pearson,\n    spearman,\n    ci: { icc: ciIcc, weightedKappa: ciKappa },\n    n: nClean,\n    raters,\n  }\n}\n\nexport interface ContinuousCalibrationResult extends CalibrationResult {\n  /** Cohen's κ_w computed on raw (un-rounded) scores. */\n  weightedKappaContinuous: number\n  /** ICC(2,1) treating golden + candidate as two raters. */\n  icc: number\n  spearman: number\n  ci: {\n    icc: [number, number]\n    weightedKappa: [number, number]\n  }\n}\n\n/**\n * Extends `calibrateJudge` with continuous-value agreement metrics while\n * retaining its base calibration summary.\n */\nexport function calibrateJudgeContinuous(\n  golden: GoldenItem[],\n  candidate: CandidateScore[],\n  opts: ContinuousAgreementOptions = {},\n): ContinuousCalibrationResult {\n  const base = calibrateJudge(golden, candidate)\n  const map = new Map<string, { h: number; j: number }>()\n  for (const g of golden) map.set(g.itemId, { h: g.humanScore, j: NaN })\n  for (const c of candidate) {\n    const entry = map.get(c.itemId)\n    if (entry) entry.j = c.score\n  }\n  const rows: number[][] = []\n  for (const v of map.values()) {\n    if (Number.isFinite(v.j)) rows.push([v.h, v.j])\n  }\n  const agreement = continuousAgreement(rows, opts)\n  return {\n    ...base,\n    weightedKappaContinuous: agreement.weightedKappa,\n    icc: agreement.icc,\n    spearman: agreement.spearman,\n    ci: agreement.ci,\n  }\n}\n\n// ── Continuous-agreement internals ──────────────────────────────────\n\n/**\n * Quadratic-weighted κ on continuous scores. With weights w(x,y) = (x-y)^2\n * (or |x-y| for linear) the formula collapses to:\n *\n *   κ_w = 1 − E_obs[w] / E_exp[w]\n *\n * where E_obs averages w over paired (a_i, b_i) and E_exp averages w over\n * the independent product distribution (sum_{i,j} w(a_i, b_j) / n^2).\n * The normalisation by (max-min)^2 in the integer version cancels in the\n * ratio, so we don't need it here. Generalises to N raters by averaging κ_w\n * over all rater pairs (mean pairwise weighted agreement).\n */\nfunction continuousWeightedKappa(rows: number[][], scheme: 'linear' | 'quadratic'): number {\n  if (rows.length === 0) return NaN\n  const raters = rows[0]!.length\n  if (raters < 2) return NaN\n  const wFn =\n    scheme === 'linear'\n      ? (x: number, y: number) => Math.abs(x - y)\n      : (x: number, y: number) => (x - y) ** 2\n  let sum = 0\n  let pairs = 0\n  for (let r1 = 0; r1 < raters; r1++) {\n    for (let r2 = r1 + 1; r2 < raters; r2++) {\n      const a = rows.map((row) => row[r1]!)\n      const b = rows.map((row) => row[r2]!)\n      const n = a.length\n      let obs = 0\n      for (let i = 0; i < n; i++) obs += wFn(a[i]!, b[i]!)\n      obs /= n\n      // Expected under independence: average of w over all i,j cross pairs.\n      let exp = 0\n      for (let i = 0; i < n; i++) {\n        for (let j = 0; j < n; j++) exp += wFn(a[i]!, b[j]!)\n      }\n      exp /= n * n\n      if (exp === 0) {\n        // Degenerate: at least one rater is constant. Convention: perfect\n        // agreement iff observed disagreement is also zero, else 0.\n        sum += obs === 0 ? 1 : 0\n      } else {\n        sum += 1 - obs / exp\n      }\n      pairs++\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\n/**\n * ICC(2,1) — two-way random effects, absolute agreement, single rater.\n *\n *   ICC(2,1) = (MSR − MSE) / (MSR + (k−1)·MSE + k·(MSC − MSE)/n)\n *\n * where MSR = between-rows MS, MSC = between-columns MS, MSE = residual MS,\n * n = rows (items), k = columns (raters).\n */\nfunction icc21(rows: number[][]): number {\n  const n = rows.length\n  if (n < 2) return NaN\n  const k = rows[0]!.length\n  if (k < 2) return NaN\n\n  // Row means.\n  const rowMeans = rows.map((row) => row.reduce((s, v) => s + v, 0) / k)\n  // Column means.\n  const colMeans = new Array(k).fill(0)\n  for (let j = 0; j < k; j++) {\n    let s = 0\n    for (let i = 0; i < n; i++) s += rows[i]![j]!\n    colMeans[j] = s / n\n  }\n  // Grand mean.\n  let grand = 0\n  for (let i = 0; i < n; i++) grand += rowMeans[i]!\n  grand /= n\n\n  let ssR = 0\n  for (let i = 0; i < n; i++) ssR += (rowMeans[i]! - grand) ** 2\n  ssR *= k\n  let ssC = 0\n  for (let j = 0; j < k; j++) ssC += (colMeans[j]! - grand) ** 2\n  ssC *= n\n  let ssT = 0\n  for (let i = 0; i < n; i++) {\n    for (let j = 0; j < k; j++) ssT += (rows[i]![j]! - grand) ** 2\n  }\n  const ssE = ssT - ssR - ssC\n\n  const dfR = n - 1\n  const dfC = k - 1\n  const dfE = (n - 1) * (k - 1)\n  const msR = ssR / dfR\n  const msC = ssC / dfC\n  const msE = dfE > 0 ? ssE / dfE : 0\n\n  const denom = msR + (k - 1) * msE + (k * (msC - msE)) / n\n  if (denom === 0) {\n    // Degenerate (all values equal): convention = 1 if rows identical,\n    // else 0. Identical rows ⇒ msR = 0 too, so msR − msE near 0.\n    return msR === 0 && msE === 0 ? 1 : 0\n  }\n  return (msR - msE) / denom\n}\n\n/** Average pairwise statistic over all rater pairs. */\nfunction avgPairwise(rows: number[][], fn: (a: number[], b: number[]) => number): number {\n  const k = rows[0]?.length ?? 0\n  if (k < 2) return NaN\n  let sum = 0\n  let pairs = 0\n  for (let i = 0; i < k; i++) {\n    for (let j = i + 1; j < k; j++) {\n      const a = rows.map((row) => row[i]!)\n      const b = rows.map((row) => row[j]!)\n      const r = fn(a, b)\n      if (Number.isFinite(r)) {\n        sum += r\n        pairs++\n      }\n    }\n  }\n  return pairs === 0 ? NaN : sum / pairs\n}\n\nfunction percentileBounds(ciLevel: number): [number, number] {\n  const tail = (1 - ciLevel) / 2\n  return [tail, 1 - tail]\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAqDA,SAAgB,eACd,QACA,WACmB;CACnB,MAAM,sBAAM,IAAI,IAAsC;CACtD,KAAK,MAAM,KAAK,QAAQ,IAAI,IAAI,EAAE,QAAQ;EAAE,GAAG,EAAE;EAAY,GAAG;CAAI,CAAC;CACrE,KAAK,MAAM,KAAK,WAAW;EACzB,MAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;EAC9B,IAAI,OAAO,MAAM,IAAI,EAAE;CACzB;CACA,MAAM,SAAS,CAAC,GAAG,IAAI,OAAO,CAAC,CAAC,CAAC,QAAQ,MAAM,OAAO,SAAS,EAAE,CAAC,CAAC;CACnE,MAAM,IAAI,OAAO;CACjB,IAAI,IAAI,GACN,OAAO;EAAE;EAAG,SAAS;EAAK,OAAO;EAAK,KAAK;EAAK,YAAY,CAAC;CAAE;CAEjE,MAAM,SAAS,OAAO,KAAK,MAAM,EAAE,CAAC;CACpC,MAAM,SAAS,OAAO,KAAK,MAAM,EAAE,CAAC;CAUpC,OAAO;EAAE;EAAG,SATI,SAAS,QAAQ,MASf;EAAG,OARP,cAAc,OAAO,IAAI,KAAK,KAAK,GAAG,OAAO,IAAI,KAAK,KAAK,CAQhD;EAAG,KAPX,OAAO,KAAK,MAAM,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC,CAClC,CAAC,CAAC,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;EAMjB,YALnB,CAAC,GAAG,IAAI,QAAQ,CAAC,CAAC,CAC7B,QAAQ,GAAG,OAAO,OAAO,SAAS,EAAE,CAAC,CAAC,CAAC,CACvC,KAAK,CAAC,QAAQ,QAAQ;GAAE;GAAQ,OAAO,EAAE;GAAG,OAAO,EAAE;GAAG,OAAO,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC;EAAE,EAAE,CAAC,CACtF,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,CAAC,CACjC,MAAM,GAAG,CACqC;CAAE;AACrD;;;;;AAeA,SAAgB,eAAe,QAAgD;CAC7E,MAAM,wBAAQ,IAAI,IAAiD;CACnE,KAAK,MAAM,KAAK,QAAQ;EACtB,MAAM,OAAO,MAAM,IAAI,EAAE,MAAM,KAAK,CAAC;EACrC,IAAI,EAAE,qBAAqB,SAAS,KAAK,QAAQ,EAAE;OAC9C,IAAI,EAAE,qBAAqB,UAAU,KAAK,SAAS,EAAE;EAC1D,MAAM,IAAI,EAAE,QAAQ,IAAI;CAC1B;CACA,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,EAAE,OAAO,YAAY,MAAM,OAAO,GAC3C,IAAI,UAAU,KAAA,KAAa,WAAW,KAAA,GAAW,OAAO,KAAK,QAAQ,MAAM;CAE7E,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,UAAU;EAAG,GAAG;CAAE;CACpD,OAAO;EAAE,UAAU,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;EAAQ,GAAG,OAAO;CAAO;AACzF;AAQA,SAAgB,cACd,SACqB;CACrB,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI,GAAG,OAAO;EAAE,SAAS;EAAK;CAAE;CACpC,OAAO;EACL,SAAS,SACP,QAAQ,KAAK,MAAM,EAAE,SAAS,GAC9B,QAAQ,KAAK,MAAM,EAAE,KAAK,CAC5B;EACA;CACF;AACF;;;;;;AAeA,SAAgB,eACd,SACsB;CACtB,MAAM,MAAM,QAAQ,QAAQ,MAAM,EAAE,QAAQ,CAAC,CAAC,KAAK,MAAM,EAAE,KAAK;CAChE,MAAM,OAAO,QAAQ,QAAQ,MAAM,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,MAAM,EAAE,KAAK;CAClE,IAAI,IAAI,WAAW,KAAK,KAAK,WAAW,GACtC,OAAO;EAAE,cAAc;EAAG,iBAAiB;EAAG,WAAW;EAAG,GAAG;CAAE;CACnE,MAAM,SAAS,IAAI,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,IAAI;CACpD,MAAM,UAAU,KAAK,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,KAAK;CACvD,OAAO;EACL,cAAc;EACd,iBAAiB;EACjB,WAAW,SAAS;EACpB,GAAG,QAAQ;CACb;AACF;;AAKA,SAAS,cAAc,GAAa,GAAqB;CACvD,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,GAAG,OAAO;CACpD,MAAM,MAAM,KAAK,IAAI,GAAG,GAAG,GAAG,CAAC;CAE/B,MAAM,IADM,KAAK,IAAI,GAAG,GAAG,GAAG,CAClB,IAAI,MAAM;CACtB,IAAI,IAAI,GAAG,OAAO;CAClB,MAAM,WAAuB,MAAM,KAAK,EAAE,QAAQ,EAAE,SAAS,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;CACjF,MAAM,UAAU,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACnC,MAAM,UAAU,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACnC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;EACjC,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,MAAM,SAAS;EACrB,IAAI,OAAO,IAAI,OAAO,KAAK;EAC3B,QAAQ,GAAG;EACX,QAAQ,GAAG;CACb;CACA,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,IAAI,MAAM,KAAK,IAAI,MAAM;EACpC,MAAM,WAAY,QAAQ,KAAK,QAAQ,KAAM,EAAE;EAC/C,OAAO,IAAI,SAAS,EAAE,CAAE;EACxB,OAAO,IAAI;CACb;CAEF,IAAI,QAAQ,GAAG,OAAO;CACtB,OAAO,IAAI,MAAM;AACnB;;;;;;;;AAqEA,SAAgB,oBACd,QACA,OAAmC,CAAC,GACf;CACrB,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,UAAU,KAAK,WAAW;CAChC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,UAAU,KAAK,WAAW;CAEhC,MAAM,SAAS,OAAO,QAAQ,QAAQ,IAAI,UAAU,KAAK,IAAI,OAAO,MAAM,OAAO,SAAS,CAAC,CAAC,CAAC;CAC7F,MAAM,SAAS,OAAO,EAAE,EAAE,UAAU;CAEpC,MAAM,QAAQ,OAAO,QAAQ,QAAQ,IAAI,WAAW,MAAM;CAC1D,MAAM,SAAS,MAAM;CAErB,IAAI,SAAS,KAAK,SAAS,GACzB,OAAO;EACL,eAAe;EACf,KAAK;EACL,SAAS;EACT,UAAU;EACV,IAAI;GAAE,KAAK,CAAC,KAAK,GAAG;GAAG,eAAe,CAAC,KAAK,GAAG;EAAE;EACjD,GAAG;EACH;CACF;CAGF,MAAM,QAAQ,wBAAwB,OAAO,OAAO;CACpD,MAAM,MAAM,MAAM,KAAK;CACvB,MAAM,UAAU,YAAY,OAAO,QAAQ;CAC3C,MAAM,WAAW,YAAY,OAAO,SAAS;CAE7C,MAAM,QAA0B,CAAC,KAAK,GAAG;CACzC,MAAM,UAA4B,CAAC,KAAK,GAAG;CAC3C,IAAI,YAAY,GAAG;EACjB,MAAM,MAAM,WAAW,IAAI;EAC3B,MAAM,OAAiB,CAAC;EACxB,MAAM,SAAmB,CAAC;EAC1B,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAClC,MAAM,SAAqB,IAAI,MAAM,MAAM;GAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAC1B,OAAO,KAAK,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM;GAE7C,MAAM,OAAO,MAAM,MAAM;GACzB,MAAM,KAAK,wBAAwB,QAAQ,OAAO;GAClD,IAAI,OAAO,SAAS,IAAI,GAAG,KAAK,KAAK,IAAI;GACzC,IAAI,OAAO,SAAS,EAAE,GAAG,OAAO,KAAK,EAAE;EACzC;EACA,MAAM,CAAC,IAAI,MAAM,iBAAiB,OAAO;EACzC,IAAI,KAAK,SAAS,GAAG;GACnB,KAAK,MAAM,GAAG,MAAM,IAAI,CAAC;GACzB,MAAM,KAAKA,qBAAS,MAAM,EAAE;GAC5B,MAAM,KAAKA,qBAAS,MAAM,EAAE;EAC9B;EACA,IAAI,OAAO,SAAS,GAAG;GACrB,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;GAC3B,QAAQ,KAAKA,qBAAS,QAAQ,EAAE;GAChC,QAAQ,KAAKA,qBAAS,QAAQ,EAAE;EAClC;CACF;CAEA,OAAO;EACL,eAAe;EACf;EACA;EACA;EACA,IAAI;GAAE,KAAK;GAAO,eAAe;EAAQ;EACzC,GAAG;EACH;CACF;AACF;;;;;AAkBA,SAAgB,yBACd,QACA,WACA,OAAmC,CAAC,GACP;CAC7B,MAAM,OAAO,eAAe,QAAQ,SAAS;CAC7C,MAAM,sBAAM,IAAI,IAAsC;CACtD,KAAK,MAAM,KAAK,QAAQ,IAAI,IAAI,EAAE,QAAQ;EAAE,GAAG,EAAE;EAAY,GAAG;CAAI,CAAC;CACrE,KAAK,MAAM,KAAK,WAAW;EACzB,MAAM,QAAQ,IAAI,IAAI,EAAE,MAAM;EAC9B,IAAI,OAAO,MAAM,IAAI,EAAE;CACzB;CACA,MAAM,OAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,IAAI,OAAO,GACzB,IAAI,OAAO,SAAS,EAAE,CAAC,GAAG,KAAK,KAAK,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;CAEhD,MAAM,YAAY,oBAAoB,MAAM,IAAI;CAChD,OAAO;EACL,GAAG;EACH,yBAAyB,UAAU;EACnC,KAAK,UAAU;EACf,UAAU,UAAU;EACpB,IAAI,UAAU;CAChB;AACF;;;;;;;;;;;;;AAgBA,SAAS,wBAAwB,MAAkB,QAAwC;CACzF,IAAI,KAAK,WAAW,GAAG,OAAO;CAC9B,MAAM,SAAS,KAAK,EAAE,CAAE;CACxB,IAAI,SAAS,GAAG,OAAO;CACvB,MAAM,MACJ,WAAW,YACN,GAAW,MAAc,KAAK,IAAI,IAAI,CAAC,KACvC,GAAW,OAAe,IAAI,MAAM;CAC3C,IAAI,MAAM;CACV,IAAI,QAAQ;CACZ,KAAK,IAAI,KAAK,GAAG,KAAK,QAAQ,MAC5B,KAAK,IAAI,KAAK,KAAK,GAAG,KAAK,QAAQ,MAAM;EACvC,MAAM,IAAI,KAAK,KAAK,QAAQ,IAAI,GAAI;EACpC,MAAM,IAAI,KAAK,KAAK,QAAQ,IAAI,GAAI;EACpC,MAAM,IAAI,EAAE;EACZ,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,OAAO,IAAI,EAAE,IAAK,EAAE,EAAG;EACnD,OAAO;EAEP,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,OAAO,IAAI,EAAE,IAAK,EAAE,EAAG;EAErD,OAAO,IAAI;EACX,IAAI,QAAQ,GAGV,OAAO,QAAQ,IAAI,IAAI;OAEvB,OAAO,IAAI,MAAM;EAEnB;CACF;CAEF,OAAO,UAAU,IAAI,MAAM,MAAM;AACnC;;;;;;;;;AAUA,SAAS,MAAM,MAA0B;CACvC,MAAM,IAAI,KAAK;CACf,IAAI,IAAI,GAAG,OAAO;CAClB,MAAM,IAAI,KAAK,EAAE,CAAE;CACnB,IAAI,IAAI,GAAG,OAAO;CAGlB,MAAM,WAAW,KAAK,KAAK,QAAQ,IAAI,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC;CAErE,MAAM,WAAW,IAAI,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC;CACpC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,IAAI,IAAI;EACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,KAAK,KAAK,EAAE,CAAE;EAC1C,SAAS,KAAK,IAAI;CACpB;CAEA,IAAI,QAAQ;CACZ,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,SAAS,SAAS;CAC9C,SAAS;CAET,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,SAAS,KAAM,UAAU;CAC7D,OAAO;CACP,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,SAAS,KAAM,UAAU;CAC7D,OAAO;CACP,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,QAAQ,KAAK,EAAE,CAAE,KAAM,UAAU;CAE/D,MAAM,MAAM,MAAM,MAAM;CAExB,MAAM,MAAM,IAAI;CAChB,MAAM,MAAM,IAAI;CAChB,MAAM,OAAO,IAAI,MAAM,IAAI;CAC3B,MAAM,MAAM,MAAM;CAClB,MAAM,MAAM,MAAM;CAClB,MAAM,MAAM,MAAM,IAAI,MAAM,MAAM;CAElC,MAAM,QAAQ,OAAO,IAAI,KAAK,MAAO,KAAK,MAAM,OAAQ;CACxD,IAAI,UAAU,GAGZ,OAAO,QAAQ,KAAK,QAAQ,IAAI,IAAI;CAEtC,QAAQ,MAAM,OAAO;AACvB;;AAGA,SAAS,YAAY,MAAkB,IAAkD;CACvF,MAAM,IAAI,KAAK,EAAE,EAAE,UAAU;CAC7B,IAAI,IAAI,GAAG,OAAO;CAClB,IAAI,MAAM;CACV,IAAI,QAAQ;CACZ,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAG9B,MAAM,IAAI,GAFA,KAAK,KAAK,QAAQ,IAAI,EAEnB,GADH,KAAK,KAAK,QAAQ,IAAI,EAChB,CAAC;EACjB,IAAI,OAAO,SAAS,CAAC,GAAG;GACtB,OAAO;GACP;EACF;CACF;CAEF,OAAO,UAAU,IAAI,MAAM,MAAM;AACnC;AAEA,SAAS,iBAAiB,SAAmC;CAC3D,MAAM,QAAQ,IAAI,WAAW;CAC7B,OAAO,CAAC,MAAM,IAAI,IAAI;AACxB"}

package/dist/{judge-gate-iY9s4L44.js → judge-gate-Co5tuh57.js}

@@ -1,5 +1,5 @@
-import { i as JudgeError, s as ValidationError } from "./errors-DVZiqS7K.js";
-import { t as computeInterval } from "./interval-CELpye1n.js";
+import { i as JudgeError, s as ValidationError } from "./errors-BcOs5MKW.js";
+import { t as computeInterval } from "./interval-Dq9gt7jx.js";
 //#region src/series-convergence.ts
 function analyzeSeries(values, options = {}) {
 	const window = options.window ?? 5;
@@ -164,29 +164,6 @@ function snapshotFromSentinelSet(scores, golden, meta, options = {}) {
 	validateSentinelSnapshot(snapshot);
 	return snapshot;
 }
-function inMemorySentinelStore(initial = []) {
-	for (const s of initial) validateSentinelSnapshot(s);
-	const items = initial.map((s) => ({
-		...s,
-		metrics: { ...s.metrics }
-	}));
-	return {
-		async append(snapshot) {
-			validateSentinelSnapshot(snapshot);
-			items.push({
-				...snapshot,
-				metrics: { ...snapshot.metrics }
-			});
-		},
-		async history(judgeId) {
-			const all = items.map((s) => ({
-				...s,
-				metrics: { ...s.metrics }
-			}));
-			return judgeId === void 0 ? all : all.filter((s) => s.judgeId === judgeId);
-		}
-	};
-}
 /**
 * JSONL store — one snapshot per line, appended atomically per call.
 * A corrupt or shape-invalid line is a loud error naming the file and
@@ -653,6 +630,6 @@ function snapshotFromJudgeAgreement(agreement, options = {}) {
 	return snapshot;
 }
 //#endregion
-export { analyzeSeries as _, judgeGateDecision as a, snapshotFromJudgeAgreement as c, inMemorySentinelStore as d, judgeSentinelReport as f, validateSentinelSnapshot as g, snapshotFromSentinelSet as h, assertJudgeMayGate as i, evalHealthStamp as l, snapshotFromCalibration as m, JudgeGateRefusedError as n, measureJudgeAgreement as o, snapshotFromAgreement as p, addCalibrationVerdicts as r, registerCalibrationSet as s, DEFAULT_JUDGE_GATE_POLICY as t, fileSentinelStore as u };
+export { judgeGateDecision as a, snapshotFromJudgeAgreement as c, judgeSentinelReport as d, snapshotFromAgreement as f, analyzeSeries as g, validateSentinelSnapshot as h, assertJudgeMayGate as i, evalHealthStamp as l, snapshotFromSentinelSet as m, JudgeGateRefusedError as n, measureJudgeAgreement as o, snapshotFromCalibration as p, addCalibrationVerdicts as r, registerCalibrationSet as s, DEFAULT_JUDGE_GATE_POLICY as t, fileSentinelStore as u };
 
-//# sourceMappingURL=judge-gate-iY9s4L44.js.map
+//# sourceMappingURL=judge-gate-Co5tuh57.js.map