@tangle-network/agent-eval 0.180.0 → 0.182.0

package/dist/rubric-predictive-validity-CCK-1B7w.js

@@ -0,0 +1,178 @@
+import { i as makeRng } from "./internal-BMFSR8Ns.js";
+import { a as spearmanR, r as pearsonR } from "./descriptive-1V17A-qa.js";
+//#region src/meta-eval/outcome-observations.ts
+/** Reduce observations of one named metric; unrelated keys cannot supply a value. */
+function reduceOutcomeMetric(outcomes, metric, reduction) {
+	const observations = outcomes.flatMap((outcome) => {
+		const value = outcome.metrics[metric];
+		return Number.isFinite(outcome.capturedAt) && typeof value === "number" && Number.isFinite(value) ? [{
+			capturedAt: outcome.capturedAt,
+			value
+		}] : [];
+	});
+	if (observations.length === 0) return null;
+	if (reduction === "mean") return observations.reduce((sum, observation) => sum + observation.value / observations.length, 0);
+	if (reduction === "max") return Math.max(...observations.map((observation) => observation.value));
+	return observations.reduce((a, b) => b.capturedAt > a.capturedAt ? b : a).value;
+}
+function hasVariation(values) {
+	return values.some((value) => value !== values[0]);
+}
+function correlationSummary(xs, ys, iterations, seed) {
+	const rng = makeRng(seed, xs, ys);
+	const pearsons = [];
+	const spearmans = [];
+	for (let b = 0; b < iterations; b++) {
+		const rx = [];
+		const ry = [];
+		for (let i = 0; i < xs.length; i++) {
+			const index = Math.floor(rng() * xs.length);
+			rx.push(xs[index]);
+			ry.push(ys[index]);
+		}
+		if (!hasVariation(rx) || !hasVariation(ry)) continue;
+		const pearson = pearsonR(rx, ry);
+		const spearman = spearmanR(rx, ry);
+		if (Number.isFinite(pearson)) pearsons.push(pearson);
+		if (Number.isFinite(spearman)) spearmans.push(spearman);
+	}
+	return {
+		pearson: pearsonR(xs, ys),
+		spearman: spearmanR(xs, ys),
+		pearsonCi95: interval(pearsons),
+		spearmanCi95: interval(spearmans)
+	};
+}
+function validateObservationOptions(reduction, iterations, seed) {
+	if (![
+		"latest",
+		"mean",
+		"max"
+	].includes(reduction)) throw new Error("outcome reduction must be latest, mean, or max");
+	if (!Number.isSafeInteger(iterations) || iterations < 1) throw new Error("bootstrap iterations must be a positive safe integer");
+	if (seed !== void 0 && !Number.isFinite(seed)) throw new Error("bootstrap seed must be finite");
+}
+function assertUniqueObservationIds(ids, name) {
+	if (ids.some((id) => typeof id !== "string" || id.trim().length === 0)) throw new Error(`${name} must be a nonempty string`);
+	if (new Set(ids).size !== ids.length) throw new Error(`duplicate ${name} in outcome study`);
+}
+function validateOutcomeMetricSpecifications(metrics) {
+	if (!Array.isArray(metrics) || metrics.length === 0) throw new Error("outcomeMetrics must declare at least one outcome metric and direction");
+	for (const metric of metrics) {
+		if (!metric || typeof metric.id !== "string" || metric.id.trim().length === 0) throw new Error("each outcome metric must have a nonempty id and explicit direction");
+		if (metric.direction !== "higher-is-better" && metric.direction !== "lower-is-better") throw new Error(`outcome metric ${metric.id} must declare higher-is-better or lower-is-better`);
+	}
+	assertUniqueObservationIds(metrics.map((metric) => metric.id), "outcome metric");
+}
+function interval(values) {
+	if (values.length === 0) return null;
+	values.sort((a, b) => a - b);
+	return {
+		lower: Math.max(-1, values[Math.floor(.025 * values.length)]),
+		upper: Math.min(1, values[Math.min(values.length - 1, Math.floor(.975 * values.length))])
+	};
+}
+//#endregion
+//#region src/meta-eval/rubric-predictive-validity.ts
+async function rubricPredictiveValidity(input) {
+	const minSamples = input.minSamples ?? 8;
+	const reduction = input.reduction ?? "latest";
+	const resamples = input.bootstrapResamples ?? 500;
+	const seed = input.seed;
+	if (!Number.isSafeInteger(minSamples) || minSamples < 3) throw new Error("minSamples must be a safe integer at least 3");
+	validateObservationOptions(reduction, resamples, seed);
+	validateOutcomeMetricSpecifications(input.outcomeMetrics);
+	assertUniqueObservationIds(input.runs.map((run) => run.runId), "runId");
+	const outcomeMetrics = input.outcomeMetrics.map((metric) => ({ ...metric }));
+	const runs = input.runs.map((run) => ({
+		runId: run.runId,
+		scores: { ...run.outcome.raw }
+	}));
+	const declaredRubrics = input.rubrics === void 0 ? void 0 : [...input.rubrics];
+	if (declaredRubrics !== void 0) assertUniqueObservationIds(declaredRubrics, "rubric");
+	const outcomes = await input.outcomes.list();
+	const outcomesByRun = /* @__PURE__ */ new Map();
+	for (const outcome of outcomes) {
+		const rows = outcomesByRun.get(outcome.runId) ?? [];
+		rows.push(outcome);
+		outcomesByRun.set(outcome.runId, rows);
+	}
+	const observedRubrics = /* @__PURE__ */ new Set();
+	for (const run of runs) for (const [rubric, value] of Object.entries(run.scores)) if (typeof value === "number" && Number.isFinite(value)) observedRubrics.add(rubric);
+	const rubrics = declaredRubrics ?? [...observedRubrics];
+	const buckets = rubrics.flatMap((rubric) => outcomeMetrics.map((outcome) => ({
+		rubric,
+		outcome,
+		xs: [],
+		ys: []
+	})));
+	let joined = 0;
+	for (const run of runs) {
+		const rows = outcomesByRun.get(run.runId) ?? [];
+		let joinedThisRun = false;
+		for (const bucket of buckets) {
+			const x = run.scores[bucket.rubric];
+			if (typeof x !== "number" || !Number.isFinite(x)) continue;
+			const y = reduceOutcomeMetric(rows, bucket.outcome.id, reduction);
+			if (y === null) continue;
+			bucket.xs.push(x);
+			bucket.ys.push(y);
+			joinedThisRun = true;
+		}
+		if (joinedThisRun) joined++;
+	}
+	const pairs = [];
+	const excludedPairs = [];
+	for (const bucket of buckets) {
+		const identity = {
+			rubric: bucket.rubric,
+			outcome: bucket.outcome.id,
+			outcomeDirection: bucket.outcome.direction,
+			n: bucket.xs.length
+		};
+		const reason = bucket.xs.length < minSamples ? "insufficient_samples" : !hasVariation(bucket.xs) ? "constant_rubric" : !hasVariation(bucket.ys) ? "constant_outcome" : null;
+		if (reason !== null) {
+			excludedPairs.push({
+				...identity,
+				reason
+			});
+			continue;
+		}
+		const summary = correlationSummary(bucket.xs, bucket.ys, resamples, seed);
+		const sign = bucket.outcome.direction === "higher-is-better" ? 1 : -1;
+		const alignedSpearman = summary.spearman * sign;
+		const alignedSpearmanCi95 = summary.spearmanCi95 === null ? null : sign === 1 ? summary.spearmanCi95 : {
+			lower: -summary.spearmanCi95.upper,
+			upper: -summary.spearmanCi95.lower
+		};
+		pairs.push({
+			...identity,
+			...summary,
+			alignedPearson: summary.pearson * sign,
+			alignedSpearman,
+			alignedSpearmanCi95,
+			verdict: alignedSpearman >= .4 ? "aligned" : alignedSpearman <= -.4 ? "inverse" : "weak"
+		});
+	}
+	const bestByRubric = /* @__PURE__ */ new Map();
+	for (const pair of pairs) {
+		const best = bestByRubric.get(pair.rubric);
+		if (!best || pair.alignedSpearman > best.alignedSpearman) bestByRubric.set(pair.rubric, pair);
+	}
+	return {
+		outcomeMetrics,
+		pairs,
+		excludedPairs,
+		ranked: [...bestByRubric.values()].map(({ outcome, ...pair }) => ({
+			...pair,
+			bestOutcome: outcome
+		})).sort((a, b) => b.alignedSpearman - a.alignedSpearman),
+		joinedSamples: joined,
+		skippedRuns: runs.length - joined,
+		rubricsWithoutData: rubrics.filter((rubric) => !observedRubrics.has(rubric))
+	};
+}
+//#endregion
+export { reduceOutcomeMetric as a, hasVariation as i, assertUniqueObservationIds as n, validateObservationOptions as o, correlationSummary as r, validateOutcomeMetricSpecifications as s, rubricPredictiveValidity as t };
+
+//# sourceMappingURL=rubric-predictive-validity-CCK-1B7w.js.map

package/dist/rubric-predictive-validity-CCK-1B7w.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"rubric-predictive-validity-CCK-1B7w.js","names":[],"sources":["../src/meta-eval/outcome-observations.ts","../src/meta-eval/rubric-predictive-validity.ts"],"sourcesContent":["import { pearsonR, spearmanR } from '../statistics'\nimport { makeRng } from '../statistics/internal'\nimport type { DeploymentOutcome } from './outcome-store'\n\nexport type OutcomeReduction = 'latest' | 'mean' | 'max'\n\nexport interface CorrelationInterval {\n  lower: number\n  upper: number\n}\n\n/** Reduce observations of one named metric; unrelated keys cannot supply a value. */\nexport function reduceOutcomeMetric(\n  outcomes: readonly DeploymentOutcome[],\n  metric: string,\n  reduction: OutcomeReduction,\n): number | null {\n  const observations = outcomes.flatMap((outcome) => {\n    const value = outcome.metrics[metric]\n    return Number.isFinite(outcome.capturedAt) &&\n      typeof value === 'number' &&\n      Number.isFinite(value)\n      ? [{ capturedAt: outcome.capturedAt, value }]\n      : []\n  })\n  if (observations.length === 0) return null\n  if (reduction === 'mean') {\n    return observations.reduce(\n      (sum, observation) => sum + observation.value / observations.length,\n      0,\n    )\n  }\n  if (reduction === 'max') return Math.max(...observations.map((observation) => observation.value))\n  return observations.reduce((a, b) => (b.capturedAt > a.capturedAt ? b : a)).value\n}\n\nexport function hasVariation(values: readonly number[]): boolean {\n  return values.some((value) => value !== values[0])\n}\n\nexport function correlationSummary(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n  seed: number | undefined,\n): {\n  pearson: number\n  spearman: number\n  pearsonCi95: CorrelationInterval | null\n  spearmanCi95: CorrelationInterval | null\n} {\n  const rng = makeRng(seed, xs, ys)\n  const pearsons: number[] = []\n  const spearmans: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx: number[] = []\n    const ry: number[] = []\n    for (let i = 0; i < xs.length; i++) {\n      const index = Math.floor(rng() * xs.length)\n      rx.push(xs[index]!)\n      ry.push(ys[index]!)\n    }\n    // A constant resample has no estimable correlation, even when both sides agree.\n    if (!hasVariation(rx) || !hasVariation(ry)) continue\n    const pearson = pearsonR(rx, ry)\n    const spearman = spearmanR(rx, ry)\n    if (Number.isFinite(pearson)) pearsons.push(pearson)\n    if (Number.isFinite(spearman)) spearmans.push(spearman)\n  }\n  return {\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n    pearsonCi95: interval(pearsons),\n    spearmanCi95: interval(spearmans),\n  }\n}\n\nexport function validateObservationOptions(\n  reduction: OutcomeReduction,\n  iterations: number,\n  seed: number | undefined,\n): void {\n  if (!['latest', 'mean', 'max'].includes(reduction)) {\n    throw new Error('outcome reduction must be latest, mean, or max')\n  }\n  if (!Number.isSafeInteger(iterations) || iterations < 1) {\n    throw new Error('bootstrap iterations must be a positive safe integer')\n  }\n  if (seed !== undefined && !Number.isFinite(seed)) {\n    throw new Error('bootstrap seed must be finite')\n  }\n}\n\nexport function assertUniqueObservationIds(ids: readonly string[], name: string): void {\n  if (ids.some((id) => typeof id !== 'string' || id.trim().length === 0)) {\n    throw new Error(`${name} must be a nonempty string`)\n  }\n  if (new Set(ids).size !== ids.length) throw new Error(`duplicate ${name} in outcome study`)\n}\n\nexport function validateOutcomeMetricSpecifications(\n  metrics: readonly { id: string; direction: string }[],\n): void {\n  if (!Array.isArray(metrics) || metrics.length === 0) {\n    throw new Error('outcomeMetrics must declare at least one outcome metric and direction')\n  }\n  for (const metric of metrics) {\n    if (!metric || typeof metric.id !== 'string' || metric.id.trim().length === 0) {\n      throw new Error('each outcome metric must have a nonempty id and explicit direction')\n    }\n    if (metric.direction !== 'higher-is-better' && metric.direction !== 'lower-is-better') {\n      throw new Error(\n        `outcome metric ${metric.id} must declare higher-is-better or lower-is-better`,\n      )\n    }\n  }\n  assertUniqueObservationIds(\n    metrics.map((metric) => metric.id),\n    'outcome metric',\n  )\n}\n\nfunction interval(values: number[]): CorrelationInterval | null {\n  if (values.length === 0) return null\n  values.sort((a, b) => a - b)\n  return {\n    lower: Math.max(-1, values[Math.floor(0.025 * values.length)]!),\n    upper: Math.min(1, values[Math.min(values.length - 1, Math.floor(0.975 * values.length))]!),\n  }\n}\n","/**\n * Join rubric scores to deployment outcomes and measure their association.\n * Higher rubric scores always mean better evaluated behavior.\n * Outcome directions are explicit because success rate and failure rate have opposite meanings.\n * These descriptive associations neither establish causation nor validate a change to rubric weights.\n */\n\nimport type { RunRecord } from '../run-record'\nimport {\n  assertUniqueObservationIds,\n  type CorrelationInterval,\n  correlationSummary,\n  hasVariation,\n  reduceOutcomeMetric,\n  validateObservationOptions,\n  validateOutcomeMetricSpecifications,\n} from './outcome-observations'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface OutcomeMetricSpec {\n  /** Exact key in DeploymentOutcome.metrics. */\n  id: string\n  direction: 'higher-is-better' | 'lower-is-better'\n}\n\nexport interface RubricPredictiveValidityInput {\n  /** One record per independent run; rubric scores come from outcome.raw. */\n  runs: RunRecord[]\n  outcomes: OutcomeStore\n  /** Declare desired directions before inspecting associations. */\n  outcomeMetrics: readonly OutcomeMetricSpec[]\n  /** Higher is better for each rubric. Omit to discover finite numeric outcome.raw keys. */\n  rubrics?: readonly string[]\n  /** Minimum joined runs for an estimate; an integer at least 3. Default 8. */\n  minSamples?: number\n  /** Bootstrap resamples for both correlation intervals. Default 500. */\n  bootstrapResamples?: number\n  /** Omit to derive a reproducible seed from the paired observations. */\n  seed?: number\n  /** Reduce finite observations of each named outcome within a run. Default latest. */\n  reduction?: 'latest' | 'mean' | 'max'\n}\n\nexport interface RubricOutcomePair {\n  rubric: string\n  outcome: string\n  outcomeDirection: OutcomeMetricSpec['direction']\n  n: number\n  /** Raw association with the recorded outcome, before direction alignment. */\n  pearson: number\n  spearman: number\n  pearsonCi95: CorrelationInterval | null\n  spearmanCi95: CorrelationInterval | null\n  /** Positive values associate higher rubric scores with better outcomes. */\n  alignedPearson: number\n  alignedSpearman: number\n  alignedSpearmanCi95: CorrelationInterval | null\n  /** Descriptive buckets at aligned Spearman +/-0.4; no causal or release authority. */\n  verdict: 'aligned' | 'inverse' | 'weak'\n}\n\nexport interface RubricRanking extends Omit<RubricOutcomePair, 'outcome'> {\n  /** Outcome with the greatest direction-aligned Spearman for this rubric. */\n  bestOutcome: string\n}\n\nexport interface RubricOutcomeExclusion {\n  rubric: string\n  outcome: string\n  outcomeDirection: OutcomeMetricSpec['direction']\n  /** Finite joined observations, including measured zeros. */\n  n: number\n  reason: 'insufficient_samples' | 'constant_rubric' | 'constant_outcome'\n}\n\nexport interface RubricPredictiveValidityReport {\n  outcomeMetrics: OutcomeMetricSpec[]\n  pairs: RubricOutcomePair[]\n  /** All declared pairs lacking an estimate, with their usable observation count. */\n  excludedPairs: RubricOutcomeExclusion[]\n  /** Exploratory ordering by aligned Spearman; never use outcome selection as confirmatory evidence. */\n  ranked: RubricRanking[]\n  /** Runs contributing at least one finite pair, including pairs below minSamples. */\n  joinedSamples: number\n  /** Runs contributing no finite pair; joinedSamples + skippedRuns equals the input run count. */\n  skippedRuns: number\n  /** Declared rubrics with no finite score, distinct from too few outcomes or constant observations. */\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 seed = input.seed\n  if (!Number.isSafeInteger(minSamples) || minSamples < 3) {\n    throw new Error('minSamples must be a safe integer at least 3')\n  }\n  validateObservationOptions(reduction, resamples, seed)\n  validateOutcomeMetricSpecifications(input.outcomeMetrics)\n  assertUniqueObservationIds(\n    input.runs.map((run) => run.runId),\n    'runId',\n  )\n\n  const outcomeMetrics = input.outcomeMetrics.map((metric) => ({ ...metric }))\n  const runs = input.runs.map((run) => ({ runId: run.runId, scores: { ...run.outcome.raw } }))\n  const declaredRubrics = input.rubrics === undefined ? undefined : [...input.rubrics]\n  if (declaredRubrics !== undefined) assertUniqueObservationIds(declaredRubrics, 'rubric')\n\n  const outcomes = await input.outcomes.list()\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const outcome of outcomes) {\n    const rows = outcomesByRun.get(outcome.runId) ?? []\n    rows.push(outcome)\n    outcomesByRun.set(outcome.runId, rows)\n  }\n\n  const observedRubrics = new Set<string>()\n  for (const run of runs) {\n    for (const [rubric, value] of Object.entries(run.scores)) {\n      if (typeof value === 'number' && Number.isFinite(value)) observedRubrics.add(rubric)\n    }\n  }\n  const rubrics = declaredRubrics ?? [...observedRubrics]\n  const buckets = rubrics.flatMap((rubric) =>\n    outcomeMetrics.map((outcome) => ({\n      rubric,\n      outcome,\n      xs: [] as number[],\n      ys: [] as number[],\n    })),\n  )\n\n  let joined = 0\n  for (const run of runs) {\n    const rows = outcomesByRun.get(run.runId) ?? []\n    let joinedThisRun = false\n    for (const bucket of buckets) {\n      const x = run.scores[bucket.rubric]\n      if (typeof x !== 'number' || !Number.isFinite(x)) continue\n      const y = reduceOutcomeMetric(rows, bucket.outcome.id, reduction)\n      if (y === null) continue\n      bucket.xs.push(x)\n      bucket.ys.push(y)\n      joinedThisRun = true\n    }\n    if (joinedThisRun) joined++\n  }\n\n  const pairs: RubricOutcomePair[] = []\n  const excludedPairs: RubricOutcomeExclusion[] = []\n  for (const bucket of buckets) {\n    const identity = {\n      rubric: bucket.rubric,\n      outcome: bucket.outcome.id,\n      outcomeDirection: bucket.outcome.direction,\n      n: bucket.xs.length,\n    }\n    const reason =\n      bucket.xs.length < minSamples\n        ? 'insufficient_samples'\n        : !hasVariation(bucket.xs)\n          ? 'constant_rubric'\n          : !hasVariation(bucket.ys)\n            ? 'constant_outcome'\n            : null\n    if (reason !== null) {\n      excludedPairs.push({ ...identity, reason })\n      continue\n    }\n    const summary = correlationSummary(bucket.xs, bucket.ys, resamples, seed)\n    const sign = bucket.outcome.direction === 'higher-is-better' ? 1 : -1\n    const alignedSpearman = summary.spearman * sign\n    const alignedSpearmanCi95 =\n      summary.spearmanCi95 === null\n        ? null\n        : sign === 1\n          ? summary.spearmanCi95\n          : { lower: -summary.spearmanCi95.upper, upper: -summary.spearmanCi95.lower }\n    pairs.push({\n      ...identity,\n      ...summary,\n      alignedPearson: summary.pearson * sign,\n      alignedSpearman,\n      alignedSpearmanCi95,\n      verdict: alignedSpearman >= 0.4 ? 'aligned' : alignedSpearman <= -0.4 ? 'inverse' : 'weak',\n    })\n  }\n\n  const bestByRubric = new Map<string, RubricOutcomePair>()\n  for (const pair of pairs) {\n    const best = bestByRubric.get(pair.rubric)\n    if (!best || pair.alignedSpearman > best.alignedSpearman) bestByRubric.set(pair.rubric, pair)\n  }\n  const ranked = [...bestByRubric.values()]\n    .map(({ outcome, ...pair }) => ({ ...pair, bestOutcome: outcome }))\n    .sort((a, b) => b.alignedSpearman - a.alignedSpearman)\n\n  return {\n    outcomeMetrics,\n    pairs,\n    excludedPairs,\n    ranked,\n    joinedSamples: joined,\n    skippedRuns: runs.length - joined,\n    rubricsWithoutData: rubrics.filter((rubric) => !observedRubrics.has(rubric)),\n  }\n}\n"],"mappings":";;;;AAYA,SAAgB,oBACd,UACA,QACA,WACe;CACf,MAAM,eAAe,SAAS,SAAS,YAAY;EACjD,MAAM,QAAQ,QAAQ,QAAQ;EAC9B,OAAO,OAAO,SAAS,QAAQ,UAAU,KACvC,OAAO,UAAU,YACjB,OAAO,SAAS,KAAK,IACnB,CAAC;GAAE,YAAY,QAAQ;GAAY;EAAM,CAAC,IAC1C,CAAC;CACP,CAAC;CACD,IAAI,aAAa,WAAW,GAAG,OAAO;CACtC,IAAI,cAAc,QAChB,OAAO,aAAa,QACjB,KAAK,gBAAgB,MAAM,YAAY,QAAQ,aAAa,QAC7D,CACF;CAEF,IAAI,cAAc,OAAO,OAAO,KAAK,IAAI,GAAG,aAAa,KAAK,gBAAgB,YAAY,KAAK,CAAC;CAChG,OAAO,aAAa,QAAQ,GAAG,MAAO,EAAE,aAAa,EAAE,aAAa,IAAI,CAAE,CAAC,CAAC;AAC9E;AAEA,SAAgB,aAAa,QAAoC;CAC/D,OAAO,OAAO,MAAM,UAAU,UAAU,OAAO,EAAE;AACnD;AAEA,SAAgB,mBACd,IACA,IACA,YACA,MAMA;CACA,MAAM,MAAM,QAAQ,MAAM,IAAI,EAAE;CAChC,MAAM,WAAqB,CAAC;CAC5B,MAAM,YAAsB,CAAC;CAC7B,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,KAAe,CAAC;EACtB,MAAM,KAAe,CAAC;EACtB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK;GAClC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM;GAC1C,GAAG,KAAK,GAAG,MAAO;GAClB,GAAG,KAAK,GAAG,MAAO;EACpB;EAEA,IAAI,CAAC,aAAa,EAAE,KAAK,CAAC,aAAa,EAAE,GAAG;EAC5C,MAAM,UAAU,SAAS,IAAI,EAAE;EAC/B,MAAM,WAAW,UAAU,IAAI,EAAE;EACjC,IAAI,OAAO,SAAS,OAAO,GAAG,SAAS,KAAK,OAAO;EACnD,IAAI,OAAO,SAAS,QAAQ,GAAG,UAAU,KAAK,QAAQ;CACxD;CACA,OAAO;EACL,SAAS,SAAS,IAAI,EAAE;EACxB,UAAU,UAAU,IAAI,EAAE;EAC1B,aAAa,SAAS,QAAQ;EAC9B,cAAc,SAAS,SAAS;CAClC;AACF;AAEA,SAAgB,2BACd,WACA,YACA,MACM;CACN,IAAI,CAAC;EAAC;EAAU;EAAQ;CAAK,CAAC,CAAC,SAAS,SAAS,GAC/C,MAAM,IAAI,MAAM,gDAAgD;CAElE,IAAI,CAAC,OAAO,cAAc,UAAU,KAAK,aAAa,GACpD,MAAM,IAAI,MAAM,sDAAsD;CAExE,IAAI,SAAS,KAAA,KAAa,CAAC,OAAO,SAAS,IAAI,GAC7C,MAAM,IAAI,MAAM,+BAA+B;AAEnD;AAEA,SAAgB,2BAA2B,KAAwB,MAAoB;CACrF,IAAI,IAAI,MAAM,OAAO,OAAO,OAAO,YAAY,GAAG,KAAK,CAAC,CAAC,WAAW,CAAC,GACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2BAA2B;CAErD,IAAI,IAAI,IAAI,GAAG,CAAC,CAAC,SAAS,IAAI,QAAQ,MAAM,IAAI,MAAM,aAAa,KAAK,kBAAkB;AAC5F;AAEA,SAAgB,oCACd,SACM;CACN,IAAI,CAAC,MAAM,QAAQ,OAAO,KAAK,QAAQ,WAAW,GAChD,MAAM,IAAI,MAAM,uEAAuE;CAEzF,KAAK,MAAM,UAAU,SAAS;EAC5B,IAAI,CAAC,UAAU,OAAO,OAAO,OAAO,YAAY,OAAO,GAAG,KAAK,CAAC,CAAC,WAAW,GAC1E,MAAM,IAAI,MAAM,oEAAoE;EAEtF,IAAI,OAAO,cAAc,sBAAsB,OAAO,cAAc,mBAClE,MAAM,IAAI,MACR,kBAAkB,OAAO,GAAG,kDAC9B;CAEJ;CACA,2BACE,QAAQ,KAAK,WAAW,OAAO,EAAE,GACjC,gBACF;AACF;AAEA,SAAS,SAAS,QAA8C;CAC9D,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,OAAO;EACL,OAAO,KAAK,IAAI,IAAI,OAAO,KAAK,MAAM,OAAQ,OAAO,MAAM,EAAG;EAC9D,OAAO,KAAK,IAAI,GAAG,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,MAAM,OAAQ,OAAO,MAAM,CAAC,EAAG;CAC5F;AACF;;;ACvCA,eAAsB,yBACpB,OACyC;CACzC,MAAM,aAAa,MAAM,cAAc;CACvC,MAAM,YAAY,MAAM,aAAa;CACrC,MAAM,YAAY,MAAM,sBAAsB;CAC9C,MAAM,OAAO,MAAM;CACnB,IAAI,CAAC,OAAO,cAAc,UAAU,KAAK,aAAa,GACpD,MAAM,IAAI,MAAM,8CAA8C;CAEhE,2BAA2B,WAAW,WAAW,IAAI;CACrD,oCAAoC,MAAM,cAAc;CACxD,2BACE,MAAM,KAAK,KAAK,QAAQ,IAAI,KAAK,GACjC,OACF;CAEA,MAAM,iBAAiB,MAAM,eAAe,KAAK,YAAY,EAAE,GAAG,OAAO,EAAE;CAC3E,MAAM,OAAO,MAAM,KAAK,KAAK,SAAS;EAAE,OAAO,IAAI;EAAO,QAAQ,EAAE,GAAG,IAAI,QAAQ,IAAI;CAAE,EAAE;CAC3F,MAAM,kBAAkB,MAAM,YAAY,KAAA,IAAY,KAAA,IAAY,CAAC,GAAG,MAAM,OAAO;CACnF,IAAI,oBAAoB,KAAA,GAAW,2BAA2B,iBAAiB,QAAQ;CAEvF,MAAM,WAAW,MAAM,MAAM,SAAS,KAAK;CAC3C,MAAM,gCAAgB,IAAI,IAAiC;CAC3D,KAAK,MAAM,WAAW,UAAU;EAC9B,MAAM,OAAO,cAAc,IAAI,QAAQ,KAAK,KAAK,CAAC;EAClD,KAAK,KAAK,OAAO;EACjB,cAAc,IAAI,QAAQ,OAAO,IAAI;CACvC;CAEA,MAAM,kCAAkB,IAAI,IAAY;CACxC,KAAK,MAAM,OAAO,MAChB,KAAK,MAAM,CAAC,QAAQ,UAAU,OAAO,QAAQ,IAAI,MAAM,GACrD,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,GAAG,gBAAgB,IAAI,MAAM;CAGvF,MAAM,UAAU,mBAAmB,CAAC,GAAG,eAAe;CACtD,MAAM,UAAU,QAAQ,SAAS,WAC/B,eAAe,KAAK,aAAa;EAC/B;EACA;EACA,IAAI,CAAC;EACL,IAAI,CAAC;CACP,EAAE,CACJ;CAEA,IAAI,SAAS;CACb,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,OAAO,cAAc,IAAI,IAAI,KAAK,KAAK,CAAC;EAC9C,IAAI,gBAAgB;EACpB,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,IAAI,IAAI,OAAO,OAAO;GAC5B,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;GAClD,MAAM,IAAI,oBAAoB,MAAM,OAAO,QAAQ,IAAI,SAAS;GAChE,IAAI,MAAM,MAAM;GAChB,OAAO,GAAG,KAAK,CAAC;GAChB,OAAO,GAAG,KAAK,CAAC;GAChB,gBAAgB;EAClB;EACA,IAAI,eAAe;CACrB;CAEA,MAAM,QAA6B,CAAC;CACpC,MAAM,gBAA0C,CAAC;CACjD,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,WAAW;GACf,QAAQ,OAAO;GACf,SAAS,OAAO,QAAQ;GACxB,kBAAkB,OAAO,QAAQ;GACjC,GAAG,OAAO,GAAG;EACf;EACA,MAAM,SACJ,OAAO,GAAG,SAAS,aACf,yBACA,CAAC,aAAa,OAAO,EAAE,IACrB,oBACA,CAAC,aAAa,OAAO,EAAE,IACrB,qBACA;EACV,IAAI,WAAW,MAAM;GACnB,cAAc,KAAK;IAAE,GAAG;IAAU;GAAO,CAAC;GAC1C;EACF;EACA,MAAM,UAAU,mBAAmB,OAAO,IAAI,OAAO,IAAI,WAAW,IAAI;EACxE,MAAM,OAAO,OAAO,QAAQ,cAAc,qBAAqB,IAAI;EACnE,MAAM,kBAAkB,QAAQ,WAAW;EAC3C,MAAM,sBACJ,QAAQ,iBAAiB,OACrB,OACA,SAAS,IACP,QAAQ,eACR;GAAE,OAAO,CAAC,QAAQ,aAAa;GAAO,OAAO,CAAC,QAAQ,aAAa;EAAM;EACjF,MAAM,KAAK;GACT,GAAG;GACH,GAAG;GACH,gBAAgB,QAAQ,UAAU;GAClC;GACA;GACA,SAAS,mBAAmB,KAAM,YAAY,mBAAmB,MAAO,YAAY;EACtF,CAAC;CACH;CAEA,MAAM,+BAAe,IAAI,IAA+B;CACxD,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,OAAO,aAAa,IAAI,KAAK,MAAM;EACzC,IAAI,CAAC,QAAQ,KAAK,kBAAkB,KAAK,iBAAiB,aAAa,IAAI,KAAK,QAAQ,IAAI;CAC9F;CAKA,OAAO;EACL;EACA;EACA;EACA,QARa,CAAC,GAAG,aAAa,OAAO,CAAC,CAAC,CACtC,KAAK,EAAE,SAAS,GAAG,YAAY;GAAE,GAAG;GAAM,aAAa;EAAQ,EAAE,CAAC,CAClE,MAAM,GAAG,MAAM,EAAE,kBAAkB,EAAE,eAMjC;EACL,eAAe;EACf,aAAa,KAAK,SAAS;EAC3B,oBAAoB,QAAQ,QAAQ,WAAW,CAAC,gBAAgB,IAAI,MAAM,CAAC;CAC7E;AACF"}

package/dist/{run-record-DTv1MdjK.d.ts → run-record-BiTWauyO.d.ts}

@@ -1,6 +1,6 @@
 import { c as ValidationError } from "./errors-DEE6u6ot.js";
 import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
-import { r as AgentProfileCell } from "./agent-profile-cell-CTOZJUuE.js";
+import { r as AgentProfileCell } from "./agent-profile-cell-s__adRnK.js";
 import { o as FailureClass } from "./schema-CR5cpjQ3.js";
 //#region src/run-record.d.ts
 /** Search/dev/holdout split tag. 'search' is the paper-grade alias for the
@@ -244,4 +244,4 @@ declare function roundTripRunRecord(record: RunRecord): RunRecord;
 declare function modelHasSnapshot(model: string): boolean;
 //#endregion
 export { validateRunRecord as _, RunRecord as a, RunTaskFailure as c, UNKNOWN_MODEL as d, isRunRecord as f, runTaskScore as g, roundTripRunRecord as h, RunOutcome as i, RunTerminalOutcome as l, parseRunRecordSafe as m, RunCostProvenance as n, RunRecordValidationError as o, modelHasSnapshot as p, RunJudgeMetadata as r, RunSplitTag as s, JudgeScoresRecord as t, RunTokenUsage as u };
-//# sourceMappingURL=run-record-DTv1MdjK.d.ts.map
+//# sourceMappingURL=run-record-BiTWauyO.d.ts.map

package/dist/{run-record-DTv1MdjK.d.ts.map → run-record-BiTWauyO.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"run-record-DTv1MdjK.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAgPnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}
+{"version":3,"file":"run-record-BiTWauyO.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAgPnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}