@tangle-network/agent-eval 0.179.0 → 0.181.0

package/dist/rubric-predictive-validity-2D5Gw9z9.js

@@ -1,131 +0,0 @@
-import { i as makeRng } from "./internal-BMFSR8Ns.js";
-import { a as spearmanR, r as pearsonR } from "./descriptive-1V17A-qa.js";
-//#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 outcomes = await input.outcomes.list();
-	const outcomesByRun = /* @__PURE__ */ new Map();
-	for (const o of outcomes) {
-		const arr = outcomesByRun.get(o.runId) ?? [];
-		arr.push(o);
-		outcomesByRun.set(o.runId, arr);
-	}
-	const observedRubrics = /* @__PURE__ */ new Set();
-	for (const r of input.runs) for (const k of Object.keys(r.outcome.raw)) observedRubrics.add(k);
-	const rubrics = input.rubrics ?? [...observedRubrics];
-	const buckets = [];
-	for (const r of rubrics) for (const o of input.outcomeMetrics) buckets.push({
-		rubric: r,
-		outcome: o,
-		xs: [],
-		ys: []
-	});
-	let joined = 0;
-	let skipped = 0;
-	for (const run of input.runs) {
-		const os = outcomesByRun.get(run.runId);
-		if (!os || os.length === 0) {
-			skipped++;
-			continue;
-		}
-		let joinedThisRun = false;
-		for (const r of rubrics) {
-			const x = run.outcome.raw[r];
-			if (typeof x !== "number" || !Number.isFinite(x)) continue;
-			for (const o of input.outcomeMetrics) {
-				const values = os.map((row) => row.metrics[o]).filter((v) => typeof v === "number" && Number.isFinite(v));
-				if (values.length === 0) continue;
-				const y = reduce(values, os, o, reduction);
-				if (y === null) continue;
-				const bucket = buckets.find((b) => b.rubric === r && b.outcome === o);
-				bucket.xs.push(x);
-				bucket.ys.push(y);
-				joinedThisRun = true;
-			}
-		}
-		if (joinedThisRun) joined++;
-	}
-	const pairs = [];
-	for (const b of buckets) {
-		if (b.xs.length < minSamples) continue;
-		const pearson = pearsonR(b.xs, b.ys);
-		const spearman = spearmanR(b.xs, b.ys);
-		const ci = bootstrapCi(b.xs, b.ys, resamples, input.seed);
-		const verdict = Math.abs(spearman) >= .7 ? "load_bearing" : Math.abs(spearman) >= .4 ? "informative" : "decorative";
-		pairs.push({
-			rubric: b.rubric,
-			outcome: b.outcome,
-			n: b.xs.length,
-			pearson,
-			spearman,
-			ci95: ci,
-			verdict
-		});
-	}
-	const byRubric = /* @__PURE__ */ new Map();
-	for (const p of pairs) {
-		const arr = byRubric.get(p.rubric) ?? [];
-		arr.push(p);
-		byRubric.set(p.rubric, arr);
-	}
-	const ranked = [...byRubric.entries()].map(([rubric, ps]) => {
-		const best = ps.reduce((a, b) => Math.abs(b.spearman) > Math.abs(a.spearman) ? b : a);
-		return {
-			rubric,
-			bestOutcome: best.outcome,
-			spearman: best.spearman,
-			pearson: best.pearson,
-			n: best.n,
-			verdict: best.verdict
-		};
-	}).sort((a, b) => Math.abs(b.spearman) - Math.abs(a.spearman));
-	const rubricsWithoutData = rubrics.filter((r) => !byRubric.has(r));
-	return {
-		pairs,
-		ranked,
-		joinedSamples: joined,
-		skippedRuns: skipped,
-		rubricsWithoutData
-	};
-}
-function reduce(values, outcomes, metric, kind) {
-	if (values.length === 0) return null;
-	if (kind === "mean") return values.reduce((s, v) => s + v, 0) / values.length;
-	if (kind === "max") return Math.max(...values);
-	return [...outcomes].filter((o) => typeof o.metrics[metric] === "number").sort((a, b) => b.capturedAt - a.capturedAt)[0]?.metrics[metric] ?? null;
-}
-function bootstrapCi(xs, ys, iterations, seed) {
-	const n = xs.length;
-	if (n < 3) return {
-		low: NaN,
-		high: NaN
-	};
-	const rng = makeRng(seed, xs, ys);
-	const samples = [];
-	for (let b = 0; b < iterations; b++) {
-		const rx = new Array(n);
-		const ry = new Array(n);
-		for (let i = 0; i < n; i++) {
-			const idx = Math.floor(rng() * n);
-			rx[i] = xs[idx];
-			ry[i] = ys[idx];
-		}
-		const r = pearsonR(rx, ry);
-		if (Number.isFinite(r)) samples.push(r);
-	}
-	samples.sort((a, b) => a - b);
-	if (samples.length === 0) return {
-		low: NaN,
-		high: NaN
-	};
-	return {
-		low: samples[Math.floor(.025 * samples.length)],
-		high: samples[Math.min(samples.length - 1, Math.floor(.975 * samples.length))]
-	};
-}
-//#endregion
-export { rubricPredictiveValidity as t };
-
-//# sourceMappingURL=rubric-predictive-validity-2D5Gw9z9.js.map

package/dist/rubric-predictive-validity-2D5Gw9z9.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"rubric-predictive-validity-2D5Gw9z9.js","names":[],"sources":["../src/meta-eval/rubric-predictive-validity.ts"],"sourcesContent":["/**\n * Rubric predictive validity — does our eval rubric predict deployment\n * outcomes?\n *\n * `correlationStudy` (already in this package) joins a `TraceStore` to an\n * `OutcomeStore` and computes Pearson + Spearman + bootstrap CI for each\n * (eval-metric, outcome-metric) pair. That answers \"does X correlate with\n * Y at all.\" `rubricPredictiveValidity` is the campaign-shaped wrapper\n * around it: take a sequence of `RunRecord`s (the canonical campaign\n * artifact) and a `DeploymentOutcomeStore`, join on `runId`, return a\n * ranked verdict on every rubric whose dimension scores were captured in\n * `outcome.raw`.\n *\n * The point — quoting the methodology doc — is that **without this loop\n * every rubric is faith-based**. Once it's wired, you know which rubrics\n * have earned their promotion power and which ones are decoration.\n *\n *   const validity = await rubricPredictiveValidity({\n *     runs: lastQuarter,\n *     outcomes: shipFlagOutcomeStore,\n *     outcomeMetrics: ['revenue_lift', 'retention_30d', 'csat'],\n *     rubrics: ['anti_slop', 'semantic_concept', 'tool_recovery'],\n *   })\n *   for (const r of validity.ranked) {\n *     console.log(`${r.rubric} → ${r.bestOutcome}: ρ=${r.spearman.toFixed(2)}`)\n *   }\n *\n * The function is intentionally read-only. Use the verdict to deprecate\n * decorative rubrics, re-weight composite scores, or trigger a\n * recalibration sweep when predictive validity drops below a threshold.\n */\n\nimport type { RunRecord } from '../run-record'\nimport { pearsonR, spearmanR } from '../statistics'\nimport { makeRng } from '../statistics/internal'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface RubricPredictiveValidityInput {\n  /**\n   * Canonical campaign output. Each record's `outcome.raw[<rubricId>]`\n   * provides the eval score; missing keys are silently skipped per pair.\n   */\n  runs: RunRecord[]\n  outcomes: OutcomeStore\n  /**\n   * Outcome metric names to evaluate against. Each must appear in at\n   * least one `DeploymentOutcome.metrics` keyspace; pairs with too few\n   * joined samples are excluded from the result.\n   */\n  outcomeMetrics: string[]\n  /**\n   * Rubric ids to evaluate. Must appear as keys in `RunRecord.outcome.raw`.\n   * If omitted, every numeric key in `outcome.raw` across the run set is\n   * treated as a rubric.\n   */\n  rubrics?: string[]\n  /** Minimum joined-sample count before a pair is reported. Default 8. */\n  minSamples?: number\n  /** Bootstrap resamples for CI. Default 500. */\n  bootstrapResamples?: number\n  /** Seed for the bootstrap. Absent, the seed is derived from the paired\n   *  observations, so the same input reproduces the same interval. */\n  seed?: number\n  /**\n   * Reduction when multiple outcomes attach to one runId. Default `'latest'`\n   * (most recently captured).\n   */\n  reduction?: 'latest' | 'mean' | 'max'\n}\n\nexport interface RubricOutcomePair {\n  rubric: string\n  outcome: string\n  n: number\n  pearson: number\n  spearman: number\n  ci95: { low: number; high: number }\n  /**\n   * Verdict bucket. `load_bearing` ≥ 0.7, `informative` ≥ 0.4,\n   * `decorative` < 0.4 in absolute correlation. A negative correlation\n   * with a desired outcome is also `decorative` — actively misleading\n   * is worse than uninformative.\n   */\n  verdict: 'load_bearing' | 'informative' | 'decorative'\n}\n\nexport interface RubricRanking {\n  rubric: string\n  /** Outcome metric this rubric correlated best with. */\n  bestOutcome: string\n  spearman: number\n  pearson: number\n  n: number\n  verdict: RubricOutcomePair['verdict']\n}\n\nexport interface RubricPredictiveValidityReport {\n  pairs: RubricOutcomePair[]\n  /** Per-rubric best pair, sorted descending by |spearman|. */\n  ranked: RubricRanking[]\n  joinedSamples: number\n  skippedRuns: number\n  /** Rubrics that were declared but never produced a usable score. */\n  rubricsWithoutData: string[]\n}\n\nexport async function rubricPredictiveValidity(\n  input: RubricPredictiveValidityInput,\n): Promise<RubricPredictiveValidityReport> {\n  const minSamples = input.minSamples ?? 8\n  const reduction = input.reduction ?? 'latest'\n  const resamples = input.bootstrapResamples ?? 500\n\n  const outcomes = await input.outcomes.list()\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  // Discover rubrics: caller-declared OR every numeric key in outcome.raw\n  // observed across runs.\n  const observedRubrics = new Set<string>()\n  for (const r of input.runs) {\n    for (const k of Object.keys(r.outcome.raw)) observedRubrics.add(k)\n  }\n  const rubrics = input.rubrics ?? [...observedRubrics]\n\n  // Collect aligned (x, y) pairs per (rubric, outcome).\n  type Bucket = { rubric: string; outcome: string; xs: number[]; ys: number[] }\n  const buckets: Bucket[] = []\n  for (const r of rubrics) {\n    for (const o of input.outcomeMetrics) {\n      buckets.push({ rubric: r, outcome: o, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of input.runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    let joinedThisRun = false\n    for (const r of rubrics) {\n      const x = run.outcome.raw[r]\n      if (typeof x !== 'number' || !Number.isFinite(x)) continue\n      for (const o of input.outcomeMetrics) {\n        const values = os\n          .map((row) => row.metrics[o])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, os, o, reduction)\n        if (y === null) continue\n        const bucket = buckets.find((b) => b.rubric === r && b.outcome === o)!\n        bucket.xs.push(x)\n        bucket.ys.push(y)\n        joinedThisRun = true\n      }\n    }\n    if (joinedThisRun) joined++\n  }\n\n  const pairs: RubricOutcomePair[] = []\n  for (const b of buckets) {\n    if (b.xs.length < minSamples) continue\n    const pearson = pearsonR(b.xs, b.ys)\n    const spearman = spearmanR(b.xs, b.ys)\n    const ci = bootstrapCi(b.xs, b.ys, resamples, input.seed)\n    const verdict: RubricOutcomePair['verdict'] =\n      Math.abs(spearman) >= 0.7\n        ? 'load_bearing'\n        : Math.abs(spearman) >= 0.4\n          ? 'informative'\n          : 'decorative'\n    pairs.push({\n      rubric: b.rubric,\n      outcome: b.outcome,\n      n: b.xs.length,\n      pearson,\n      spearman,\n      ci95: ci,\n      verdict,\n    })\n  }\n\n  const byRubric = new Map<string, RubricOutcomePair[]>()\n  for (const p of pairs) {\n    const arr = byRubric.get(p.rubric) ?? []\n    arr.push(p)\n    byRubric.set(p.rubric, arr)\n  }\n  const ranked: RubricRanking[] = [...byRubric.entries()]\n    .map(([rubric, ps]) => {\n      const best = ps.reduce((a, b) => (Math.abs(b.spearman) > Math.abs(a.spearman) ? b : a))\n      return {\n        rubric,\n        bestOutcome: best.outcome,\n        spearman: best.spearman,\n        pearson: best.pearson,\n        n: best.n,\n        verdict: best.verdict,\n      }\n    })\n    .sort((a, b) => Math.abs(b.spearman) - Math.abs(a.spearman))\n\n  const rubricsWithoutData = rubrics.filter((r) => !byRubric.has(r))\n\n  return { pairs, ranked, joinedSamples: joined, skippedRuns: skipped, rubricsWithoutData }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  outcomes: DeploymentOutcome[],\n  metric: string,\n  kind: 'latest' | 'mean' | 'max',\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((s, v) => s + v, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest'\n  const sorted = [...outcomes]\n    .filter((o) => typeof o.metrics[metric] === 'number')\n    .sort((a, b) => b.capturedAt - a.capturedAt)\n  return sorted[0]?.metrics[metric] ?? null\n}\n\nfunction bootstrapCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n  seed: number | undefined,\n): { low: number; high: number } {\n  const n = xs.length\n  if (n < 3) return { low: Number.NaN, high: Number.NaN }\n  const rng = makeRng(seed, xs, ys)\n  const samples: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx = new Array<number>(n)\n    const ry = new Array<number>(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(rng() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) samples.push(r)\n  }\n  samples.sort((a, b) => a - b)\n  if (samples.length === 0) return { low: Number.NaN, high: Number.NaN }\n  return {\n    low: samples[Math.floor(0.025 * samples.length)]!,\n    high: samples[Math.min(samples.length - 1, Math.floor(0.975 * samples.length))]!,\n  }\n}\n"],"mappings":";;;AA0GA,eAAsB,yBACpB,OACyC;CACzC,MAAM,aAAa,MAAM,cAAc;CACvC,MAAM,YAAY,MAAM,aAAa;CACrC,MAAM,YAAY,MAAM,sBAAsB;CAE9C,MAAM,WAAW,MAAM,MAAM,SAAS,KAAK;CAC3C,MAAM,gCAAgB,IAAI,IAAiC;CAC3D,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;EAC3C,IAAI,KAAK,CAAC;EACV,cAAc,IAAI,EAAE,OAAO,GAAG;CAChC;CAIA,MAAM,kCAAkB,IAAI,IAAY;CACxC,KAAK,MAAM,KAAK,MAAM,MACpB,KAAK,MAAM,KAAK,OAAO,KAAK,EAAE,QAAQ,GAAG,GAAG,gBAAgB,IAAI,CAAC;CAEnE,MAAM,UAAU,MAAM,WAAW,CAAC,GAAG,eAAe;CAIpD,MAAM,UAAoB,CAAC;CAC3B,KAAK,MAAM,KAAK,SACd,KAAK,MAAM,KAAK,MAAM,gBACpB,QAAQ,KAAK;EAAE,QAAQ;EAAG,SAAS;EAAG,IAAI,CAAC;EAAG,IAAI,CAAC;CAAE,CAAC;CAI1D,IAAI,SAAS;CACb,IAAI,UAAU;CACd,KAAK,MAAM,OAAO,MAAM,MAAM;EAC5B,MAAM,KAAK,cAAc,IAAI,IAAI,KAAK;EACtC,IAAI,CAAC,MAAM,GAAG,WAAW,GAAG;GAC1B;GACA;EACF;EACA,IAAI,gBAAgB;EACpB,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,IAAI,IAAI,QAAQ,IAAI;GAC1B,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;GAClD,KAAK,MAAM,KAAK,MAAM,gBAAgB;IACpC,MAAM,SAAS,GACZ,KAAK,QAAQ,IAAI,QAAQ,EAAE,CAAC,CAC5B,QAAQ,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;IACzE,IAAI,OAAO,WAAW,GAAG;IACzB,MAAM,IAAI,OAAO,QAAQ,IAAI,GAAG,SAAS;IACzC,IAAI,MAAM,MAAM;IAChB,MAAM,SAAS,QAAQ,MAAM,MAAM,EAAE,WAAW,KAAK,EAAE,YAAY,CAAC;IACpE,OAAO,GAAG,KAAK,CAAC;IAChB,OAAO,GAAG,KAAK,CAAC;IAChB,gBAAgB;GAClB;EACF;EACA,IAAI,eAAe;CACrB;CAEA,MAAM,QAA6B,CAAC;CACpC,KAAK,MAAM,KAAK,SAAS;EACvB,IAAI,EAAE,GAAG,SAAS,YAAY;EAC9B,MAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;EACnC,MAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;EACrC,MAAM,KAAK,YAAY,EAAE,IAAI,EAAE,IAAI,WAAW,MAAM,IAAI;EACxD,MAAM,UACJ,KAAK,IAAI,QAAQ,KAAK,KAClB,iBACA,KAAK,IAAI,QAAQ,KAAK,KACpB,gBACA;EACR,MAAM,KAAK;GACT,QAAQ,EAAE;GACV,SAAS,EAAE;GACX,GAAG,EAAE,GAAG;GACR;GACA;GACA,MAAM;GACN;EACF,CAAC;CACH;CAEA,MAAM,2BAAW,IAAI,IAAiC;CACtD,KAAK,MAAM,KAAK,OAAO;EACrB,MAAM,MAAM,SAAS,IAAI,EAAE,MAAM,KAAK,CAAC;EACvC,IAAI,KAAK,CAAC;EACV,SAAS,IAAI,EAAE,QAAQ,GAAG;CAC5B;CACA,MAAM,SAA0B,CAAC,GAAG,SAAS,QAAQ,CAAC,CAAC,CACpD,KAAK,CAAC,QAAQ,QAAQ;EACrB,MAAM,OAAO,GAAG,QAAQ,GAAG,MAAO,KAAK,IAAI,EAAE,QAAQ,IAAI,KAAK,IAAI,EAAE,QAAQ,IAAI,IAAI,CAAE;EACtF,OAAO;GACL;GACA,aAAa,KAAK;GAClB,UAAU,KAAK;GACf,SAAS,KAAK;GACd,GAAG,KAAK;GACR,SAAS,KAAK;EAChB;CACF,CAAC,CAAC,CACD,MAAM,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,IAAI,KAAK,IAAI,EAAE,QAAQ,CAAC;CAE7D,MAAM,qBAAqB,QAAQ,QAAQ,MAAM,CAAC,SAAS,IAAI,CAAC,CAAC;CAEjE,OAAO;EAAE;EAAO;EAAQ,eAAe;EAAQ,aAAa;EAAS;CAAmB;AAC1F;AAIA,SAAS,OACP,QACA,UACA,QACA,MACe;CACf,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,SAAS,QAAQ,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;CACvE,IAAI,SAAS,OAAO,OAAO,KAAK,IAAI,GAAG,MAAM;CAK7C,OAHe,CAAC,GAAG,QAAQ,CAAC,CACzB,QAAQ,MAAM,OAAO,EAAE,QAAQ,YAAY,QAAQ,CAAC,CACpD,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UACvB,CAAC,CAAC,EAAE,EAAE,QAAQ,WAAW;AACvC;AAEA,SAAS,YACP,IACA,IACA,YACA,MAC+B;CAC/B,MAAM,IAAI,GAAG;CACb,IAAI,IAAI,GAAG,OAAO;EAAE,KAAK;EAAY,MAAM;CAAW;CACtD,MAAM,MAAM,QAAQ,MAAM,IAAI,EAAE;CAChC,MAAM,UAAoB,CAAC;CAC3B,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,KAAK,IAAI,MAAc,CAAC;EAC9B,MAAM,KAAK,IAAI,MAAc,CAAC;EAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,MAAM,KAAK,MAAM,IAAI,IAAI,CAAC;GAChC,GAAG,KAAK,GAAG;GACX,GAAG,KAAK,GAAG;EACb;EACA,MAAM,IAAI,SAAS,IAAI,EAAE;EACzB,IAAI,OAAO,SAAS,CAAC,GAAG,QAAQ,KAAK,CAAC;CACxC;CACA,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAC5B,IAAI,QAAQ,WAAW,GAAG,OAAO;EAAE,KAAK;EAAY,MAAM;CAAW;CACrE,OAAO;EACL,KAAK,QAAQ,KAAK,MAAM,OAAQ,QAAQ,MAAM;EAC9C,MAAM,QAAQ,KAAK,IAAI,QAAQ,SAAS,GAAG,KAAK,MAAM,OAAQ,QAAQ,MAAM,CAAC;CAC/E;AACF"}

package/dist/rubric-predictive-validity-Dl1dvKCv.d.ts

@@ -1,75 +0,0 @@
-import { a as RunRecord } from "./run-record-DTv1MdjK.js";
-import { o as OutcomeStore } from "./outcome-store-BYHIuO0e.js";
-//#region src/meta-eval/rubric-predictive-validity.d.ts
-interface RubricPredictiveValidityInput {
-  /**
-   * Canonical campaign output. Each record's `outcome.raw[<rubricId>]`
-   * provides the eval score; missing keys are silently skipped per pair.
-   */
-  runs: RunRecord[];
-  outcomes: OutcomeStore;
-  /**
-   * Outcome metric names to evaluate against. Each must appear in at
-   * least one `DeploymentOutcome.metrics` keyspace; pairs with too few
-   * joined samples are excluded from the result.
-   */
-  outcomeMetrics: string[];
-  /**
-   * Rubric ids to evaluate. Must appear as keys in `RunRecord.outcome.raw`.
-   * If omitted, every numeric key in `outcome.raw` across the run set is
-   * treated as a rubric.
-   */
-  rubrics?: string[];
-  /** Minimum joined-sample count before a pair is reported. Default 8. */
-  minSamples?: number;
-  /** Bootstrap resamples for CI. Default 500. */
-  bootstrapResamples?: number;
-  /** Seed for the bootstrap. Absent, the seed is derived from the paired
-   *  observations, so the same input reproduces the same interval. */
-  seed?: number;
-  /**
-   * Reduction when multiple outcomes attach to one runId. Default `'latest'`
-   * (most recently captured).
-   */
-  reduction?: 'latest' | 'mean' | 'max';
-}
-interface RubricOutcomePair {
-  rubric: string;
-  outcome: string;
-  n: number;
-  pearson: number;
-  spearman: number;
-  ci95: {
-    low: number;
-    high: number;
-  };
-  /**
-   * Verdict bucket. `load_bearing` ≥ 0.7, `informative` ≥ 0.4,
-   * `decorative` < 0.4 in absolute correlation. A negative correlation
-   * with a desired outcome is also `decorative` — actively misleading
-   * is worse than uninformative.
-   */
-  verdict: 'load_bearing' | 'informative' | 'decorative';
-}
-interface RubricRanking {
-  rubric: string;
-  /** Outcome metric this rubric correlated best with. */
-  bestOutcome: string;
-  spearman: number;
-  pearson: number;
-  n: number;
-  verdict: RubricOutcomePair['verdict'];
-}
-interface RubricPredictiveValidityReport {
-  pairs: RubricOutcomePair[];
-  /** Per-rubric best pair, sorted descending by |spearman|. */
-  ranked: RubricRanking[];
-  joinedSamples: number;
-  skippedRuns: number;
-  /** Rubrics that were declared but never produced a usable score. */
-  rubricsWithoutData: string[];
-}
-declare function rubricPredictiveValidity(input: RubricPredictiveValidityInput): Promise<RubricPredictiveValidityReport>;
-//#endregion
-export { rubricPredictiveValidity as a, RubricRanking as i, RubricPredictiveValidityInput as n, RubricPredictiveValidityReport as r, RubricOutcomePair as t };
-//# sourceMappingURL=rubric-predictive-validity-Dl1dvKCv.d.ts.map

package/dist/rubric-predictive-validity-Dl1dvKCv.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"rubric-predictive-validity-Dl1dvKCv.d.ts","names":[],"sources":["../src/meta-eval/rubric-predictive-validity.ts"],"mappings":";;;UAqCiB;;;;;EAKf,MAAM;EACN,UAAU;;;;;;EAMV;;;;;;EAMA;;EAEA;;EAEA;;;EAGA;;;;;EAKA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;IAAQ;IAAa;;;;;;;;EAOrB;;UAGe;EACf;;EAEA;EACA;EACA;EACA;EACA,SAAS;;UAGM;EACf,OAAO;;EAEP,QAAQ;EACR;EACA;;EAEA;;iBAGoB,yBACpB,OAAO,gCACN,QAAQ"}

package/dist/run-record-CR63CpHK.js

@@ -1,216 +0,0 @@
-import { c as validateRunRecord } from "./run-record-DQpSf7t-.js";
-//#region src/campaign/run-record.ts
-/**
-* A campaign cell carried a judge score without a `dimensions` record.
-*
-* Two exported types share the name `JudgeScore`: the campaign verdict
-* (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)
-* and the root export's flat per-dimension row (`{ judgeName, dimension, score }`
-* from `@tangle-network/agent-eval`). Campaign aggregation accepts only the
-* campaign shape; the flat shape previously crashed here with an opaque
-* TypeError deep inside aggregation.
-*/
-var CampaignJudgeScoreShapeError = class extends TypeError {
-	constructor(judgeName) {
-		super(`campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. Use the campaign JudgeScore ({ dimensions, composite, notes }) from '@tangle-network/agent-eval/campaign'; the root export's JudgeScore ({ judgeName, dimension, score }) is a different type with the same name.`);
-		this.name = "CampaignJudgeScoreShapeError";
-	}
-};
-/**
-* Project one campaign cell into the canonical run format.
-*
-* A dispatch error establishes terminal execution failure. A judge error only
-* establishes that quality measurement failed after dispatch completed.
-* Failures without a stage remain unknown. No failure becomes a zero-quality
-* label.
-*/
-function campaignCellToRunRecord(cell, options) {
-	const quality = projectCampaignCellQuality(cell);
-	const execution = campaignCellExecutionEvidence(cell);
-	const judgeErrorCount = Math.max(quality.raw.judge_error_count ?? 0, execution.judgeErrorCount ?? 0);
-	const cellCostProvenance = campaignCellCostProvenance(cell);
-	const costProvenance = cellCostProvenance.kind === "uncaptured" && options.defaultCostUsd !== void 0 ? {
-		kind: "estimated",
-		usd: options.defaultCostUsd
-	} : cellCostProvenance;
-	const costUsd = costProvenance.kind === "uncaptured" ? null : costProvenance.usd;
-	const raw = {
-		...finiteMetrics(options.raw),
-		...quality.raw,
-		rep: cell.rep,
-		duration_ms: cell.durationMs,
-		...costUsd === null ? {} : { cost_usd: costUsd },
-		...cellCostProvenance.kind === "uncaptured" ? { cost_known_subtotal_usd: cell.costUsd } : {},
-		cost_observed: costProvenance.kind === "observed" ? 1 : 0,
-		cost_estimated: costProvenance.kind === "estimated" ? 1 : 0,
-		cost_uncaptured: costProvenance.kind === "uncaptured" ? 1 : 0,
-		tokens_input: cell.tokenUsage.input,
-		tokens_output: cell.tokenUsage.output,
-		tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,
-		latency_ms: cell.durationMs,
-		...execution.executionErrorCount === void 0 ? {} : { execution_error_count: execution.executionErrorCount },
-		...judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {},
-		...execution.unclassifiedErrorCount === void 0 ? {} : { unclassified_error_count: execution.unclassifiedErrorCount }
-	};
-	if (typeof cell.generation === "number") raw.generation = cell.generation;
-	if (cell.tokenUsage.reasoning !== void 0) raw.tokens_reasoning = cell.tokenUsage.reasoning;
-	if (cell.tokenUsage.cached !== void 0) raw.tokens_cached = cell.tokenUsage.cached;
-	if (cell.tokenUsage.cacheWrite !== void 0) raw.tokens_cache_write = cell.tokenUsage.cacheWrite;
-	if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd;
-	if (costUsd !== null && quality.score !== void 0 && quality.score > .01) raw.cost_per_quality = costUsd / quality.score;
-	const outcome = {
-		raw,
-		...quality.judgeScores ? { judgeScores: quality.judgeScores } : {}
-	};
-	if (quality.score !== void 0) if (options.splitTag === "holdout") outcome.holdoutScore = quality.score;
-	else outcome.searchScore = quality.score;
-	return validateRunRecord({
-		runId: options.runId,
-		experimentId: options.experimentId,
-		candidateId: options.candidateId,
-		seed: options.seed ?? cell.seed,
-		model: options.model,
-		promptHash: options.promptHash,
-		configHash: options.configHash,
-		commitSha: options.commitSha,
-		wallMs: cell.durationMs,
-		costUsd,
-		costProvenance,
-		tokenUsage: { ...cell.tokenUsage },
-		terminalOutcome: execution.terminalOutcome,
-		...execution.terminalFailureReason ? { terminalFailureReason: execution.terminalFailureReason } : {},
-		outcome,
-		splitTag: options.splitTag,
-		scenarioId: options.scenarioId ?? cell.scenarioId,
-		...options.agentProfile ? { agentProfile: options.agentProfile } : {}
-	});
-}
-/**
-* Validate the cost fields that cross campaign cache and RunRecord boundaries.
-* `costUsd` is a known subtotal for uncaptured cells, but it must equal the
-* authoritative total whenever that total is observed or estimated.
-*/
-function campaignCellCostProvenance(cell) {
-	if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`);
-	const provenance = cell.costProvenance;
-	if (!provenance || typeof provenance !== "object") throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`);
-	if (provenance.kind === "uncaptured") {
-		if (provenance.usd !== null) throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`);
-		return {
-			kind: "uncaptured",
-			usd: null
-		};
-	}
-	if (provenance.kind !== "observed" && provenance.kind !== "estimated" || !Number.isFinite(provenance.usd) || provenance.usd < 0) throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`);
-	if (provenance.usd !== cell.costUsd) throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`);
-	return {
-		kind: provenance.kind,
-		usd: provenance.usd
-	};
-}
-function campaignCellExecutionEvidence(cell) {
-	if (cell.errorStage === "dispatch") return {
-		terminalOutcome: "failed",
-		executionErrorCount: 1,
-		...cell.error ? { terminalFailureReason: cell.error } : {}
-	};
-	if (cell.errorStage === "judge") return {
-		terminalOutcome: "succeeded",
-		executionErrorCount: 0,
-		judgeErrorCount: 1
-	};
-	if (!cell.error) return {
-		terminalOutcome: "succeeded",
-		executionErrorCount: 0
-	};
-	return {
-		terminalOutcome: "unknown",
-		unclassifiedErrorCount: 1
-	};
-}
-/**
-* Produce the only task-quality view used by campaign aggregates and exports.
-*
-* Successful judge results remain available for diagnosis after another judge
-* fails, but a task score exists only for an error-free cell whose reported
-* judge values are all finite.
-*/
-function projectCampaignCellQuality(cell) {
-	if (cell.errorStage === "dispatch") return {
-		successfulJudgeScores: {},
-		failedJudges: [],
-		raw: {}
-	};
-	const perJudge = {};
-	const successfulJudgeScores = {};
-	const dimensionValues = /* @__PURE__ */ new Map();
-	const composites = [];
-	const notes = [];
-	const failedJudges = new Set(cell.errorStage === "judge" ? [cell.errorJudge ?? "unknown-judge"] : []);
-	const raw = {};
-	for (const [judgeName, score] of Object.entries(cell.judgeScores)) {
-		const dimensionsShape = score.dimensions;
-		if (typeof dimensionsShape !== "object" || dimensionsShape === null || Array.isArray(dimensionsShape)) throw new CampaignJudgeScoreShapeError(judgeName);
-		const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite);
-		if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {
-			failedJudges.add(judgeName);
-			continue;
-		}
-		composites.push(score.composite);
-		successfulJudgeScores[judgeName] = score;
-		const dimensions = { ...score.dimensions };
-		perJudge[judgeName] = dimensions;
-		for (const [dimension, value] of Object.entries(dimensions)) {
-			raw[`${judgeName}.${dimension}`] = value;
-			const values = dimensionValues.get(dimension) ?? [];
-			values.push(value);
-			dimensionValues.set(dimension, values);
-		}
-		if (score.notes) notes.push(`${judgeName}: ${score.notes}`);
-		for (const failedJudge of score.failedJudges ?? []) failedJudges.add(`${judgeName}/${failedJudge}`);
-	}
-	if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size;
-	const sortedFailedJudges = [...failedJudges].sort();
-	if (composites.length === 0) return {
-		successfulJudgeScores,
-		failedJudges: sortedFailedJudges,
-		raw
-	};
-	const composite = mean(composites);
-	const perDimMean = Object.fromEntries([...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]));
-	const complete = cell.error === void 0 && cell.errorStage === void 0 && failedJudges.size === 0;
-	if (complete) raw.composite = composite;
-	return {
-		...complete ? { score: composite } : {},
-		raw,
-		successfulJudgeScores,
-		failedJudges: sortedFailedJudges,
-		judgeScores: {
-			perJudge,
-			perDimMean,
-			composite,
-			...sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {},
-			...notes.length > 0 ? { notes: notes.join(" | ") } : {}
-		}
-	};
-}
-/** Read the canonical task score without recomputing cell quality. */
-function campaignCellTaskScore(cell) {
-	return projectCampaignCellQuality(cell).score;
-}
-/** Read canonical successful judge dimensions without recomputing cell quality. */
-function campaignCellJudgeDimensions(cell) {
-	return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {};
-}
-function finiteMetrics(metrics) {
-	const finite = {};
-	for (const [key, value] of Object.entries(metrics ?? {})) if (Number.isFinite(value)) finite[key] = value;
-	return finite;
-}
-function mean(values) {
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
-}
-//#endregion
-export { campaignCellToRunRecord as a, campaignCellTaskScore as i, campaignCellExecutionEvidence as n, projectCampaignCellQuality as o, campaignCellJudgeDimensions as r, campaignCellCostProvenance as t };
-
-//# sourceMappingURL=run-record-CR63CpHK.js.map

package/dist/run-record-CR63CpHK.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"run-record-CR63CpHK.js","names":[],"sources":["../src/campaign/run-record.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;;;;;;;;AA8CA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}

package/dist/sequential-B5gXgcyp.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"sequential-B5gXgcyp.js","names":[],"sources":["../src/campaign/gates/sequential.ts"],"sourcesContent":["/**\n * Anytime-valid sequential promotion gate — an e-process (betting\n * test-martingale, see `eProcess` in `statistics.ts`) over paired\n * per-scenario deltas, so a campaign stops the MOMENT evidence decides\n * instead of burning a fixed-n budget. Decisions remain valid at any\n * data-dependent stopping time (Ville's inequality), which is exactly what\n * fixed-n machinery cannot offer: peeking at a bootstrap CI after every\n * observation and stopping on the first significant peek inflates type-I\n * error far beyond alpha.\n *\n * REPLACES, never layers on, a fixed-n gate. Running `heldoutSignificance`\n * or `paretoSignificanceGate` repeatedly on a growing sample and stopping\n * early is optional stopping no matter how it is dressed up; this gate is\n * the valid way to stop early. Use one or the other per evidence stream.\n *\n * Pre-registration binding: anytime validity holds only for the\n * PRE-REGISTERED statistic. When a `SignedManifest` is bound, the gate takes\n * alpha from `manifest.alpha`, the observation budget from\n * `manifest.preRegisteredN`, orients deltas by `manifest.direction`, and\n * shifts the null boundary by `manifest.minEffect` — re-deciding the same\n * stream under different parameters after seeing data would reopen optional\n * stopping under a fancier name. The manifest's content hash is verified at\n * construction (sync, same `sha256-content` scheme as `signManifest`).\n *\n * Non-iid caveat (stated honestly): the supermartingale guarantee needs each\n * delta's conditional mean under H0 to stay ≤ the null boundary given the\n * past — exchangeable scenario deltas suffice. Scenario streams ordered by\n * difficulty or by scenario family violate this; `decide(ctx)` therefore\n * shuffles the paired deltas with a SEEDED permutation by default (the\n * permutation is data-independent, so bet predictability is preserved).\n * Stratified betting (per-stratum λ) is future work, not implemented here.\n */\n\nimport { manifestContentDigest, type SignedManifest } from '../../pre-registration'\nimport { type EProcessState, eProcess, mulberry32 } from '../../statistics'\nimport type { Gate, GateContext, GateResult, GenerationRecord, Scenario } from '../types'\nimport { pairHoldout } from './statistical-heldout'\n\nexport type SequentialDecision = 'promote' | 'continue' | 'undecided-at-maxN'\n\nexport interface SequentialObservation {\n  decision: SequentialDecision\n  /** Current e-value (the betting wealth) against H0. */\n  eValue: number\n  /** Paired deltas consumed so far. */\n  n: number\n  /** Names the decision basis. For 'undecided-at-maxN' it states explicitly\n   *  that exhausting the budget is NOT evidence of no effect. */\n  reason: string\n}\n\nexport interface SequentialPairedGateOptions {\n  /** Type-I budget. With `preRegistration` bound this MUST match\n   *  `manifest.alpha` (conflict throws). Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a promote may fire. The stopping rule is\n   *  \"first n ≥ minN with e-value ≥ 1/alpha\" — still a valid stopping time.\n   *  Default 5. */\n  minN?: number\n  /** Pre-registered observation budget. Required unless `preRegistration`\n   *  supplies it via `preRegisteredN` (conflict throws). */\n  maxN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |delta| in the judge's native scale; deltas are mapped to\n   *  x = (d/scale + 1)/2 ∈ [0,1]. A delta outside ±scale throws (use\n   *  `detectScale` to pick 1 vs 100 BEFORE streaming). Default 1. */\n  scale?: number\n  /** Seed for the data-independent shuffle of paired deltas in `decide(ctx)`\n   *  (exchangeability guard). Default 1337. */\n  shuffleSeed?: number\n  /** Bind the pre-registered hypothesis. Verified (content hash) at\n   *  construction; alpha/maxN/direction/minEffect come FROM the manifest. */\n  preRegistration?: SignedManifest\n  /** Override the gate name in reports. */\n  name?: string\n  /** Continue the observe-stream from a `state()` snapshot taken before a\n   *  restart. The gate's parameters still resolve from this options object\n   *  (and the manifest); the snapshot must have been recorded under the same\n   *  alpha, maxBet, and null boundary, and its decision must be one this\n   *  configuration could have reached at its n, or construction throws.\n   *  `decide(ctx)` is unaffected — it always runs on a fresh stream. */\n  resume?: SequentialStreamState\n}\n\n/** Snapshot of one observe-stream: the e-process state plus the gate's own\n *  decision, which applies `minN` and `maxN` on top of the core latch. A\n *  gate rebuilt with `resume` from this value continues exactly where the\n *  snapshot was taken. */\nexport type SequentialStreamState = EProcessState & { decision: SequentialDecision }\n\nexport interface SequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>\n  extends Gate<TArtifact, TScenario> {\n  /** Streaming entry point: feed one paired per-scenario delta\n   *  (candidate − baseline, native scale). Each gate instance carries ONE\n   *  observe-stream; `decide(ctx)` runs on its own fresh stream and never\n   *  consumes or advances this one. 'promote' is sticky; observing past the\n   *  pre-registered maxN throws (extending a finished stream after seeing\n   *  the result reopens optional stopping — start a NEW pre-registered\n   *  test). */\n  observe(delta: number): SequentialObservation\n  /** Read-only snapshot of the observe-stream. Pass it back as\n   *  `resume` to continue the stream after a restart. */\n  state(): SequentialStreamState\n}\n\ninterface ResolvedConfig {\n  alpha: number\n  minN: number\n  maxN: number\n  maxBet: number\n  scale: number\n  shuffleSeed: number\n  /** 'decrease' negates deltas so \"better\" is always positive. */\n  direction: 'increase' | 'decrease'\n  /** H0 boundary in x-space: 1/2 + minEffect/(2·scale). */\n  nullMean: number\n  minEffect: number\n}\n\n/** Sync twin of `verifyManifest`, so gate construction stays synchronous and\n *  fails loud before any observation is consumed. The digest scheme belongs to\n *  `pre-registration`; this gate only compares. */\nfunction verifyManifestSync(m: SignedManifest): boolean {\n  return manifestContentDigest(m) === m.contentHash\n}\n\nfunction resolveConfig(opts: SequentialPairedGateOptions): ResolvedConfig {\n  const m = opts.preRegistration\n  let alpha: number\n  let maxN: number\n  let direction: 'increase' | 'decrease'\n  let minEffect: number\n  if (m) {\n    if (!verifyManifestSync(m)) {\n      throw new Error(\n        `sequentialPairedGate: pre-registration manifest '${m.id}' content hash mismatch (tampered)`,\n      )\n    }\n    if (opts.alpha !== undefined && opts.alpha !== m.alpha) {\n      throw new Error(\n        `sequentialPairedGate: alpha ${opts.alpha} conflicts with pre-registered alpha ${m.alpha} — ` +\n          `the registered statistic is the only one anytime validity covers`,\n      )\n    }\n    if (opts.maxN !== undefined && opts.maxN !== m.preRegisteredN) {\n      throw new Error(\n        `sequentialPairedGate: maxN ${opts.maxN} conflicts with pre-registered N ${m.preRegisteredN}`,\n      )\n    }\n    alpha = m.alpha\n    maxN = m.preRegisteredN\n    direction = m.direction\n    minEffect = m.minEffect\n  } else {\n    alpha = opts.alpha ?? 0.05\n    if (opts.maxN === undefined) {\n      throw new Error(\n        'sequentialPairedGate: maxN is required (or bind a preRegistration manifest whose ' +\n          'preRegisteredN is the budget) — an unbounded stream has no pre-registered budget',\n      )\n    }\n    maxN = opts.maxN\n    direction = 'increase'\n    minEffect = 0\n  }\n  const scale = opts.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialPairedGate: scale must be > 0, got ${scale}`)\n  }\n  if (!Number.isInteger(maxN) || maxN < 1) {\n    throw new Error(`sequentialPairedGate: maxN must be a positive integer, got ${maxN}`)\n  }\n  const minN = opts.minN ?? 5\n  if (!Number.isInteger(minN) || minN < 1 || minN > maxN) {\n    throw new Error(\n      `sequentialPairedGate: minN must be an integer in [1, maxN=${maxN}], got ${minN}`,\n    )\n  }\n  if (!Number.isFinite(minEffect) || minEffect < 0 || minEffect >= scale) {\n    throw new Error(\n      `sequentialPairedGate: minEffect must be in [0, scale=${scale}), got ${minEffect}`,\n    )\n  }\n  // H0: mean delta ≤ minEffect. In x = (d/scale + 1)/2 space that is\n  // E[x] ≤ 1/2 + minEffect/(2·scale) — a shifted null boundary, NOT a clamp\n  // of the observations (clamping is mean-distorting for skewed deltas).\n  const nullMean = 0.5 + minEffect / (2 * scale)\n  return {\n    alpha,\n    minN,\n    maxN,\n    maxBet: opts.maxBet ?? 0.5,\n    scale,\n    shuffleSeed: opts.shuffleSeed ?? 1337,\n    direction,\n    nullMean,\n    minEffect,\n  }\n}\n\ninterface SequentialStream {\n  observe(delta: number): SequentialObservation\n  state(): SequentialStreamState\n}\n\nconst SEQUENTIAL_DECISIONS: readonly SequentialDecision[] = [\n  'promote',\n  'continue',\n  'undecided-at-maxN',\n]\n\n/** Refuse a snapshot whose gate decision this configuration could not have\n *  reached at the snapshot's n. The e-process parameters are checked by\n *  `eProcess` itself; this covers the gate's own stopping rule. */\nfunction assertResumableStream(resume: SequentialStreamState, cfg: ResolvedConfig): void {\n  const fail = (detail: string): never => {\n    throw new Error(`sequentialPairedGate: cannot resume — ${detail}`)\n  }\n  if (!SEQUENTIAL_DECISIONS.includes(resume.decision)) {\n    fail(`unknown decision '${String(resume.decision)}'`)\n  }\n  if (resume.n > cfg.maxN) {\n    fail(`snapshot n=${resume.n} exceeds the pre-registered maxN=${cfg.maxN}`)\n  }\n  const threshold = 1 / cfg.alpha\n  switch (resume.decision) {\n    case 'promote':\n      if (resume.n < cfg.minN) {\n        fail(`decision 'promote' at n=${resume.n} is below minN=${cfg.minN}`)\n      }\n      break\n    case 'undecided-at-maxN':\n      if (resume.n !== cfg.maxN) {\n        fail(`decision 'undecided-at-maxN' at n=${resume.n} does not match maxN=${cfg.maxN}`)\n      }\n      break\n    case 'continue':\n      if (resume.n >= cfg.maxN) {\n        fail(`decision 'continue' at n=${resume.n} with maxN=${cfg.maxN}; the stream is finished`)\n      }\n      if (resume.n >= cfg.minN && resume.wealth >= threshold) {\n        fail(\n          `decision 'continue' at n=${resume.n} ≥ minN=${cfg.minN} with e-value ` +\n            `${resume.wealth} ≥ 1/α=${threshold}; this stream would have promoted`,\n        )\n      }\n      break\n  }\n}\n\nfunction makeStream(cfg: ResolvedConfig, resume?: SequentialStreamState): SequentialStream {\n  if (resume !== undefined) assertResumableStream(resume, cfg)\n  const proc = eProcess({\n    alpha: cfg.alpha,\n    maxBet: cfg.maxBet,\n    nullMean: cfg.nullMean,\n    ...(resume === undefined ? {} : { resume }),\n  })\n  const threshold = 1 / cfg.alpha\n  let terminal: SequentialDecision | undefined =\n    resume === undefined || resume.decision === 'continue' ? undefined : resume.decision\n  return {\n    observe(delta: number): SequentialObservation {\n      if (terminal === 'undecided-at-maxN') {\n        throw new Error(\n          `sequentialPairedGate: pre-registered maxN=${cfg.maxN} exhausted — extending the ` +\n            'stream after seeing the result reopens optional stopping; start a NEW ' +\n            'pre-registered test',\n        )\n      }\n      if (!Number.isFinite(delta) || Math.abs(delta) > cfg.scale) {\n        throw new Error(\n          `sequentialPairedGate: delta ${delta} outside ±scale=${cfg.scale} — pass the judge's ` +\n            `native scale explicitly (detectScale helps pick 1 vs 100)`,\n        )\n      }\n      const d = cfg.direction === 'decrease' ? -delta : delta\n      const step = proc.update((d / cfg.scale + 1) / 2)\n      // Stopping rule: FIRST n ≥ minN with current wealth ≥ 1/alpha — a valid\n      // stopping time (measurable w.r.t. the past), sticky once latched. The\n      // core's own `decided` latch ignores minN, so the gate re-derives the\n      // decision from current wealth here.\n      if (terminal === undefined && step.n >= cfg.minN && step.wealth >= threshold) {\n        terminal = 'promote'\n      }\n      if (terminal === 'promote') {\n        return {\n          decision: 'promote',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} at n=${step.n} ` +\n            `(minN=${cfg.minN}): the paired improvement exceeds ${cfg.minEffect} at anytime-valid ` +\n            `level α=${cfg.alpha}`,\n        }\n      }\n      if (step.n >= cfg.maxN) {\n        terminal = 'undecided-at-maxN'\n        return {\n          decision: 'undecided-at-maxN',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `undecided at pre-registered maxN=${cfg.maxN} (e-value ${step.wealth.toFixed(2)} < ` +\n            `1/α=${threshold.toFixed(2)}). This is NOT evidence of no effect — the effect may be ` +\n            'real but smaller than this budget can detect; re-register with a larger N to test that',\n        }\n      }\n      return {\n        decision: 'continue',\n        eValue: step.wealth,\n        n: step.n,\n        reason: `e-value ${step.wealth.toFixed(2)} < 1/α=${threshold.toFixed(2)} at n=${step.n}/${cfg.maxN} — keep observing`,\n      }\n    },\n    state() {\n      return { ...proc.state(), decision: terminal ?? 'continue' }\n    },\n  }\n}\n\n/** Data-independent in-place Fisher–Yates with a seeded PRNG — the permutation\n *  depends only on the seed, never the values, so bet predictability survives. */\nfunction seededShuffle<T>(items: T[], seed: number): T[] {\n  const rng = mulberry32(seed)\n  for (let i = items.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    const tmp = items[i]!\n    items[i] = items[j]!\n    items[j] = tmp\n  }\n  return items\n}\n\n/**\n * Anytime-valid sequential paired gate. Conforms to the existing `Gate`\n * contract (`decide(ctx)` consumes candidate vs baseline judge scores via\n * `pairHoldout` — same pairing granularity as the fixed-n gates: full cellId,\n * never scenarioId) and adds a streaming `observe(delta)` entry for campaigns\n * that score cells incrementally and want to stop mid-stream.\n *\n * Decision mapping onto the substrate's five-valued `GateDecision`:\n *   - 'promote'            → 'ship'\n *   - 'continue'           → 'need_more_work' (stream ended before maxN with\n *                            the e-value undecided — more reps could decide)\n *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT\n *                            evidence of no effect (never a silent default)\n */\nexport function sequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: SequentialPairedGateOptions,\n): SequentialPairedGate<TArtifact, TScenario> {\n  const cfg = resolveConfig(options)\n  const name = options.name ?? 'sequentialPairedGate'\n  const manifest = options.preRegistration\n  const observeStream = makeStream(cfg, options.resume)\n\n  return {\n    name,\n    observe(delta: number): SequentialObservation {\n      return observeStream.observe(delta)\n    },\n    state() {\n      return observeStream.state()\n    },\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      if (!ctx.baselineJudgeScores) {\n        throw new Error(\n          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      const deltas = paired.after.map((a, i) => a - paired.before[i]!)\n      seededShuffle(deltas, cfg.shuffleSeed)\n\n      const stream = makeStream(cfg)\n      let last: SequentialObservation | undefined\n      for (const d of deltas) {\n        last = stream.observe(d)\n        if (last.decision !== 'continue') break\n      }\n\n      const detail = {\n        ...stream.state(),\n        minN: cfg.minN,\n        maxN: cfg.maxN,\n        scale: cfg.scale,\n        shuffleSeed: cfg.shuffleSeed,\n        direction: cfg.direction,\n        minEffect: cfg.minEffect,\n        pairedN: deltas.length,\n        ...(manifest ? { preRegisteredId: manifest.id, metric: manifest.metric } : {}),\n      }\n      const meanDelta =\n        deltas.length === 0 ? undefined : deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n      if (last === undefined) {\n        return {\n          decision: 'need_more_work',\n          reasons: [`${name}: no paired holdout observations — nothing to test`],\n          contributingGates: [{ name, status: 'not_evaluated', detail }],\n        }\n      }\n      const decision =\n        last.decision === 'promote'\n          ? 'ship'\n          : last.decision === 'continue'\n            ? 'need_more_work'\n            : 'hold'\n      return {\n        decision,\n        reasons: [`${name}: ${last.reason}`],\n        contributingGates: [\n          {\n            name,\n            status:\n              decision === 'ship'\n                ? 'pass'\n                : decision === 'need_more_work'\n                  ? 'not_evaluated'\n                  : 'fail',\n            detail,\n          },\n        ],\n        delta: meanDelta,\n      }\n    },\n  }\n}\n\nexport interface SequentialDecideOptions {\n  /** Type-I budget for the early-stop evidence. Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a stop may fire. Default 5. */\n  minN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |per-scenario composite delta|. Default 1. */\n  scale?: number\n}\n\nexport interface SequentialDecideFn {\n  (args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n  /** Read-only snapshot of the accumulated e-process (observability + tests). */\n  state(): EProcessState\n}\n\n/**\n * `SurfaceProposer.decide` adapter — stops the optimization loop the moment\n * the e-process decides the loop has produced a real improvement, instead of\n * always running `maxGenerations`.\n *\n * Stream: for each generation g ≥ 1, the per-scenario composite deltas of\n * generation g's top candidate vs the generation-0 top candidate (the\n * incumbent the loop set out to beat), paired by scenarioId. H0: no proposed\n * surface improves any scenario's expected composite over the incumbent —\n * under it every delta has conditional mean ≤ 0 and the e-process is valid.\n * Once wealth ≥ 1/alpha the loop stops and hands the winner to the promotion\n * gate (which re-scores on HELD-OUT data — this adapter only spends the\n * exploration budget, it never promotes).\n *\n * Honesty caveats: (1) the incumbent's scores are measured once and shared\n * across all generations' deltas, so type-I control is exact only insofar as\n * those scores approximate the incumbent's true per-scenario means (more reps\n * → tighter); (2) an UNDECIDED process never stops the loop — absence of a\n * crossing is NOT evidence of no effect, so the loop simply runs its normal\n * course. Calling the adapter repeatedly with a growing history consumes each\n * generation exactly once (re-feeding an already-seen record would double-count\n * evidence).\n */\nexport function sequentialDecide(options: SequentialDecideOptions = {}): SequentialDecideFn {\n  const alpha = options.alpha ?? 0.05\n  const minN = options.minN ?? 5\n  const scale = options.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialDecide: scale must be > 0, got ${scale}`)\n  }\n  const proc = eProcess({ alpha, maxBet: options.maxBet ?? 0.5, nullMean: 0.5 })\n  const threshold = 1 / alpha\n  let processedGenerations = 0\n  let reference: Map<string, number> | undefined\n  let stopped: { stop: true; reason: string } | undefined\n\n  const topCandidate = (record: GenerationRecord) => {\n    if (record.candidates.length === 0) {\n      throw new Error(\n        `sequentialDecide: generation ${record.generationIndex} has no candidates — ` +\n          'cannot extract a top candidate',\n      )\n    }\n    return record.candidates.reduce((best, candidate) => {\n      if (best.composite === null) return candidate\n      if (candidate.composite === null) return best\n      return candidate.composite > best.composite ? candidate : best\n    })\n  }\n\n  const decide: SequentialDecideFn = ({ history }) => {\n    if (stopped) return stopped\n    if (history.length === 0) return { stop: false }\n    if (reference === undefined) {\n      reference = new Map(\n        topCandidate(history[0]!).scenarios.map((s) => [s.scenarioId, s.composite]),\n      )\n    }\n    // Consume only generations not yet seen; start at 1 — generation 0 IS the\n    // reference and would contribute all-zero deltas (pure wealth dilution).\n    for (let g = Math.max(1, processedGenerations); g < history.length; g++) {\n      const top = topCandidate(history[g]!)\n      const byScenario = new Map(top.scenarios.map((s) => [s.scenarioId, s.composite]))\n      for (const [scenarioId, refComposite] of reference) {\n        const candComposite = byScenario.get(scenarioId)\n        if (candComposite === undefined) {\n          throw new Error(\n            `sequentialDecide: generation ${history[g]!.generationIndex} top candidate is missing ` +\n              `scenario '${scenarioId}' — generations must score the same scenario set to pair`,\n          )\n        }\n        const delta = candComposite - refComposite\n        if (!Number.isFinite(delta) || Math.abs(delta) > scale) {\n          throw new Error(\n            `sequentialDecide: paired delta ${delta} outside ±scale=${scale} on scenario ` +\n              `'${scenarioId}' — pass the composite scale explicitly`,\n          )\n        }\n        const step = proc.update((delta / scale + 1) / 2)\n        if (step.n >= minN && step.wealth >= threshold) {\n          stopped = {\n            stop: true,\n            reason:\n              `sequential e-process decided at generation ${history[g]!.generationIndex}: ` +\n              `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} ` +\n              'paired deltas vs the generation-0 incumbent — the improvement is real at ' +\n              `α=${alpha}; stop exploring and promote via the gate`,\n          }\n          processedGenerations = history.length\n          return stopped\n        }\n      }\n    }\n    processedGenerations = history.length\n    return { stop: false }\n  }\n  decide.state = () => proc.state()\n  return decide\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2HA,SAAS,mBAAmB,GAA4B;CACtD,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;AAEA,SAAS,cAAc,MAAmD;CACxE,MAAM,IAAI,KAAK;CACf,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,GAAG;EACL,IAAI,CAAC,mBAAmB,CAAC,GACvB,MAAM,IAAI,MACR,oDAAoD,EAAE,GAAG,mCAC3D;EAEF,IAAI,KAAK,UAAU,KAAA,KAAa,KAAK,UAAU,EAAE,OAC/C,MAAM,IAAI,MACR,+BAA+B,KAAK,MAAM,uCAAuC,EAAE,MAAM,oEAE3F;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,KAAK,SAAS,EAAE,gBAC7C,MAAM,IAAI,MACR,8BAA8B,KAAK,KAAK,mCAAmC,EAAE,gBAC/E;EAEF,QAAQ,EAAE;EACV,OAAO,EAAE;EACT,YAAY,EAAE;EACd,YAAY,EAAE;CAChB,OAAO;EACL,QAAQ,KAAK,SAAS;EACtB,IAAI,KAAK,SAAS,KAAA,GAChB,MAAM,IAAI,MACR,mKAEF;EAEF,OAAO,KAAK;EACZ,YAAY;EACZ,YAAY;CACd;CACA,MAAM,QAAQ,KAAK,SAAS;CAC5B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,gDAAgD,OAAO;CAEzE,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,GACpC,MAAM,IAAI,MAAM,8DAA8D,MAAM;CAEtF,MAAM,OAAO,KAAK,QAAQ;CAC1B,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,MAChD,MAAM,IAAI,MACR,6DAA6D,KAAK,SAAS,MAC7E;CAEF,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,KAAK,aAAa,OAC/D,MAAM,IAAI,MACR,wDAAwD,MAAM,SAAS,WACzE;CAKF,MAAM,WAAW,KAAM,aAAa,IAAI;CACxC,OAAO;EACL;EACA;EACA;EACA,QAAQ,KAAK,UAAU;EACvB;EACA,aAAa,KAAK,eAAe;EACjC;EACA;EACA;CACF;AACF;AAOA,MAAM,uBAAsD;CAC1D;CACA;CACA;AACF;;;;AAKA,SAAS,sBAAsB,QAA+B,KAA2B;CACvF,MAAM,QAAQ,WAA0B;EACtC,MAAM,IAAI,MAAM,yCAAyC,QAAQ;CACnE;CACA,IAAI,CAAC,qBAAqB,SAAS,OAAO,QAAQ,GAChD,KAAK,qBAAqB,OAAO,OAAO,QAAQ,EAAE,EAAE;CAEtD,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,cAAc,OAAO,EAAE,mCAAmC,IAAI,MAAM;CAE3E,MAAM,YAAY,IAAI,IAAI;CAC1B,QAAQ,OAAO,UAAf;EACE,KAAK;GACH,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,2BAA2B,OAAO,EAAE,iBAAiB,IAAI,MAAM;GAEtE;EACF,KAAK;GACH,IAAI,OAAO,MAAM,IAAI,MACnB,KAAK,qCAAqC,OAAO,EAAE,uBAAuB,IAAI,MAAM;GAEtF;EACF,KAAK;GACH,IAAI,OAAO,KAAK,IAAI,MAClB,KAAK,4BAA4B,OAAO,EAAE,aAAa,IAAI,KAAK,yBAAyB;GAE3F,IAAI,OAAO,KAAK,IAAI,QAAQ,OAAO,UAAU,WAC3C,KACE,4BAA4B,OAAO,EAAE,UAAU,IAAI,KAAK,gBACnD,OAAO,OAAO,SAAS,UAAU,kCACxC;GAEF;CACJ;AACF;AAEA,SAAS,WAAW,KAAqB,QAAkD;CACzF,IAAI,WAAW,KAAA,GAAW,sBAAsB,QAAQ,GAAG;CAC3D,MAAM,OAAO,SAAS;EACpB,OAAO,IAAI;EACX,QAAQ,IAAI;EACZ,UAAU,IAAI;EACd,GAAI,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO;CAC3C,CAAC;CACD,MAAM,YAAY,IAAI,IAAI;CAC1B,IAAI,WACF,WAAW,KAAA,KAAa,OAAO,aAAa,aAAa,KAAA,IAAY,OAAO;CAC9E,OAAO;EACL,QAAQ,OAAsC;GAC5C,IAAI,aAAa,qBACf,MAAM,IAAI,MACR,6CAA6C,IAAI,KAAK,qHAGxD;GAEF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,OACnD,MAAM,IAAI,MACR,+BAA+B,MAAM,kBAAkB,IAAI,MAAM,8EAEnE;GAEF,MAAM,IAAI,IAAI,cAAc,aAAa,CAAC,QAAQ;GAClD,MAAM,OAAO,KAAK,QAAQ,IAAI,IAAI,QAAQ,KAAK,CAAC;GAKhD,IAAI,aAAa,KAAA,KAAa,KAAK,KAAK,IAAI,QAAQ,KAAK,UAAU,WACjE,WAAW;GAEb,IAAI,aAAa,WACf,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QACE,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,SACtE,IAAI,KAAK,oCAAoC,IAAI,UAAU,4BACzD,IAAI;GACnB;GAEF,IAAI,KAAK,KAAK,IAAI,MAAM;IACtB,WAAW;IACX,OAAO;KACL,UAAU;KACV,QAAQ,KAAK;KACb,GAAG,KAAK;KACR,QACE,oCAAoC,IAAI,KAAK,YAAY,KAAK,OAAO,QAAQ,CAAC,EAAE,SACzE,UAAU,QAAQ,CAAC,EAAE;IAEhC;GACF;GACA,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QAAQ,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,GAAG,IAAI,KAAK;GACrG;EACF;EACA,QAAQ;GACN,OAAO;IAAE,GAAG,KAAK,MAAM;IAAG,UAAU,YAAY;GAAW;EAC7D;CACF;AACF;;;AAIA,SAAS,cAAiB,OAAY,MAAmB;CACvD,MAAM,MAAM,WAAW,IAAI;CAC3B,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,IAAI,GAAG,KAAK;EACzC,MAAM,IAAI,KAAK,MAAM,IAAI,KAAK,IAAI,EAAE;EACpC,MAAM,MAAM,MAAM;EAClB,MAAM,KAAK,MAAM;EACjB,MAAM,KAAK;CACb;CACA,OAAO;AACT;;;;;;;;;;;;;;;AAgBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CAEpD,OAAO;EACL;EACA,QAAQ,OAAsC;GAC5C,OAAO,cAAc,QAAQ,KAAK;EACpC;EACA,QAAQ;GACN,OAAO,cAAc,MAAM;EAC7B;EACA,MAAM,OAAO,KAA6D;GACxE,IAAI,CAAC,IAAI,qBACP,MAAM,IAAI,MACR,GAAG,KAAK,sJAEV;GAEF,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;GAC1D,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,MAAM,SAAS,OAAO,MAAM,KAAK,GAAG,MAAM,IAAI,OAAO,OAAO,EAAG;GAC/D,cAAc,QAAQ,IAAI,WAAW;GAErC,MAAM,SAAS,WAAW,GAAG;GAC7B,IAAI;GACJ,KAAK,MAAM,KAAK,QAAQ;IACtB,OAAO,OAAO,QAAQ,CAAC;IACvB,IAAI,KAAK,aAAa,YAAY;GACpC;GAEA,MAAM,SAAS;IACb,GAAG,OAAO,MAAM;IAChB,MAAM,IAAI;IACV,MAAM,IAAI;IACV,OAAO,IAAI;IACX,aAAa,IAAI;IACjB,WAAW,IAAI;IACf,WAAW,IAAI;IACf,SAAS,OAAO;IAChB,GAAI,WAAW;KAAE,iBAAiB,SAAS;KAAI,QAAQ,SAAS;IAAO,IAAI,CAAC;GAC9E;GACA,MAAM,YACJ,OAAO,WAAW,IAAI,KAAA,IAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;GAE/E,IAAI,SAAS,KAAA,GACX,OAAO;IACL,UAAU;IACV,SAAS,CAAC,GAAG,KAAK,mDAAmD;IACrE,mBAAmB,CAAC;KAAE;KAAM,QAAQ;KAAiB;IAAO,CAAC;GAC/D;GAEF,MAAM,WACJ,KAAK,aAAa,YACd,SACA,KAAK,aAAa,aAChB,mBACA;GACR,OAAO;IACL;IACA,SAAS,CAAC,GAAG,KAAK,IAAI,KAAK,QAAQ;IACnC,mBAAmB,CACjB;KACE;KACA,QACE,aAAa,SACT,SACA,aAAa,mBACX,kBACA;KACR;IACF,CACF;IACA,OAAO;GACT;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAmC,CAAC,GAAuB;CAC1F,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,4CAA4C,OAAO;CAErE,MAAM,OAAO,SAAS;EAAE;EAAO,QAAQ,QAAQ,UAAU;EAAK,UAAU;CAAI,CAAC;CAC7E,MAAM,YAAY,IAAI;CACtB,IAAI,uBAAuB;CAC3B,IAAI;CACJ,IAAI;CAEJ,MAAM,gBAAgB,WAA6B;EACjD,IAAI,OAAO,WAAW,WAAW,GAC/B,MAAM,IAAI,MACR,gCAAgC,OAAO,gBAAgB,oDAEzD;EAEF,OAAO,OAAO,WAAW,QAAQ,MAAM,cAAc;GACnD,IAAI,KAAK,cAAc,MAAM,OAAO;GACpC,IAAI,UAAU,cAAc,MAAM,OAAO;GACzC,OAAO,UAAU,YAAY,KAAK,YAAY,YAAY;EAC5D,CAAC;CACH;CAEA,MAAM,UAA8B,EAAE,cAAc;EAClD,IAAI,SAAS,OAAO;EACpB,IAAI,QAAQ,WAAW,GAAG,OAAO,EAAE,MAAM,MAAM;EAC/C,IAAI,cAAc,KAAA,GAChB,YAAY,IAAI,IACd,aAAa,QAAQ,EAAG,CAAC,CAAC,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAC5E;EAIF,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,oBAAoB,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvE,MAAM,MAAM,aAAa,QAAQ,EAAG;GACpC,MAAM,aAAa,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;GAChF,KAAK,MAAM,CAAC,YAAY,iBAAiB,WAAW;IAClD,MAAM,gBAAgB,WAAW,IAAI,UAAU;IAC/C,IAAI,kBAAkB,KAAA,GACpB,MAAM,IAAI,MACR,gCAAgC,QAAQ,EAAE,CAAE,gBAAgB,sCAC7C,WAAW,yDAC5B;IAEF,MAAM,QAAQ,gBAAgB;IAC9B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,OAC/C,MAAM,IAAI,MACR,kCAAkC,MAAM,kBAAkB,MAAM,gBAC1D,WAAW,wCACnB;IAEF,MAAM,OAAO,KAAK,QAAQ,QAAQ,QAAQ,KAAK,CAAC;IAChD,IAAI,KAAK,KAAK,QAAQ,KAAK,UAAU,WAAW;KAC9C,UAAU;MACR,MAAM;MACN,QACE,8CAA8C,QAAQ,EAAE,CAAE,gBAAgB,YAC/D,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,WAAW,KAAK,EAAE,8EAE7E,MAAM;KACf;KACA,uBAAuB,QAAQ;KAC/B,OAAO;IACT;GACF;EACF;EACA,uBAAuB,QAAQ;EAC/B,OAAO,EAAE,MAAM,MAAM;CACvB;CACA,OAAO,cAAc,KAAK,MAAM;CAChC,OAAO;AACT"}

package/dist/statistical-heldout-0La5ZTlv.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"statistical-heldout-0La5ZTlv.d.ts","names":[],"sources":["../src/statistics/paired-binary.ts","../src/statistics/sequential-eprocess.ts","../src/attestation.ts","../src/experiment/evidence-receipt.ts","../src/experiment/campaign-evidence.ts","../src/pareto.ts","../src/campaign/gates/promotion-policy.ts","../src/campaign/gates/power-preflight.ts","../src/paired-arms.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts"],"mappings":";;;;;UAgBiB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,OAAO,mBAAmB,WAAW,sBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2CzD,sBAAsB,QAAQ;;UAU7B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;iBAkBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+Bc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;;;;;;;;;;;;;;;;;iBA6Ea,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;UCjgB7B;;;EAGf;;;;EAIA;;;;EAIA;;;;;;;;EAQA,SAAS;;UAGM;;EAEf;;EAEA;;EAEA;;;;;;;;;;UAWe,sBAAsB;EACrC;EACA;EACA;;EAEA;;EAEA;;EAEA;;;EAGA;;UAGe;;;EAGf,OAAO,YAAY;EACnB,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8BK,SAAS,OAAM,kBAAuB;;;;;;;;;;;;;;;;;;;;;;;cCrEzC;UAEI;;EAEf,eAAe;;EAEf;;;EAGA;;EAEA;;EAEA;;;EAGA;;UAGe;;EAEf;EACA,YAAY;EACZ,kBAAkB;;;;;;EAMlB;;UAGe;EACf;;EAEA;;EAEA;;;;;;;;iBAiBc,OAAO,iBAAiB,YAAY,wBAAwB;;;;;;;;;;;iBAqB5D,kBACd,iBACA,UAAU,iBACT;;;cCjFU;;cAGA;KAQD,gCAAgC;cAE/B;UAOI;WACN,MAAM;;WAEN;;UAGM;WACN,sBAAsB;;WAEtB;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA,WAAW;;WAEX;;WAEA;;UAGM;WACN,SAAS;WACT,aAAa;;UAGP;WACN;WACA;;UAGM,mCAAmC,KAAK;;;;;;;iBAQzC,sBACd,OAAO,4BACP,YAAY,wBACX;;;;;;;iBA8Ba,sBAAsB,SAAS,kBAAkB;;;;;;iBA6CjD,sBAAsB,SAAS;;;;KC9JnC,0BAA0B,KACpC;EAOI,YAAY,KAAK;;;iBAGP,8BAA8B,UAAU,UAAU,GAAG;EACnE,UAAU,eAAe,GAAG;EAC5B,SAAS;EACT,SAAS;IAAA;;;;;;;;;;;;;;;;;;KCPC;UAEK,UAAU;;EAEzB;EACA,WAAW;EACX,QAAQ,WAAW;;UAGJ,aAAa;EAC5B,UAAU;EACV,WAAW;;EAEX,cAAc;IAAQ,WAAW;IAAG,WAAW;;;;iBAIjC,UAAU,GAAG,GAAG,GAAG,GAAG,GAAG,YAAY,UAAU;;;;;;iBAmB/C,eAAe,GAAG,YAAY,KAAK,YAAY,UAAU,OAAO,aAAa;;;;;KCVjF;EAAoB;;EAAwB;EAAmB;;UAE1D;;EAEf;EACA,QAAQ;;;EAGR,WAAW;;;;EAIX;;;;;EAKA;;;KAIU;UAEK;EACf;EACA,QAAQ;EACR,WAAW;;;;;;;;EAQX,WAAW;;;;;EAKX;;;EAGA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;;EAGT;;EAEA;EACA;EACA,gBAAgB;EAChB;EACA;EACA,SAAS;;UAGM;;EAEf,MAAM;;;EAGN;;;EAGA;IAAQ;IAAmB;;;;;;KAMjB,mBAAmB,IAAI,mBAAmB;UAErC;;;;EAIf;;EAEA;;EAEA;;EAEA;;;EAGA;;;;;;;;iBASc,oBAAoB,WAAW,kBAAkB,UAC/D,KAAK,YAAY,WAAW,YAC5B,YAAY,sBACZ,OAAM,6BACL;;;;;;;;cAwIU,cAAc;UAiFV,sCAAsC;;EAErD,YAAY;;;EAGZ,SAAS;;EAET;;;;;;;iBAQc,uBAAuB,qBAAqB,kBAAkB,WAAW,UACvF,SAAS,gCACR,KAAK,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;UCjWF;;EAEf;;;EAGA;;EAEA;;EAEA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;;EAGA;EACA;EACA;;;EAGA;;EAEA;;;;;;iBAec,eAAe,MAAM,wBAAwB;;;;;UC1C5C;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;;;;;;;;;;;;;;;UClWc;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;;;;;;;;;;;;KAeU;UAEK,uBAAuB;;EAEtC;;;;;;;;;EASA,OAAO;;UAGQ;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;;;;;;;iBAkBoB,SAAS,GAAG,KAAK,IAAI;;;;;;;iBA+B3B,sBAAsB,UAAU;;;;;;iBAe1B,aAAa,GAAG,qBAAqB,QAAQ;;;;;iBAS7C,eAAe,GAAG,iBAAiB;;;;;iBAQnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KChIC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;EAGA,kBAAkB;;EAElB;;;;;;;EAOA,SAAS;;;;;;KAOC,wBAAwB;EAAkB,UAAU;;UAE/C,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;EAQxB,QAAQ,gBAAgB;;;EAGxB,SAAS;;;;;;;;;;;;;;;;iBAsPK,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UAsFlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UC/axD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBAcc,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAsDc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6Bc,oBACd,QAAQ,eACR,OAAM,6BACL;UAkEc;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;;EAKA;EACA;EACA;;;;;iBAMc,YAAY;;;;;;;;;;;;;;;;;;;iBAsBZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;IAED"}