@tangle-network/agent-eval 0.180.0 → 0.181.0

package/dist/{sequential-B5gXgcyp.js → sequential-DAsyV2T9.js}

@@ -1,7 +1,7 @@
-import { h as manifestContentDigest } from "./agent-profile-cell-0gSi5ffD.js";
+import { h as manifestContentDigest } from "./agent-profile-cell-Cv6UA-W_.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { t as eProcess } from "./sequential-eprocess-D1jKoihe.js";
-import { N as pairHoldout } from "./campaign-evidence-D8DBLqLI.js";
+import { at as pairHoldout, tt as aggregatePairedHoldout } from "./campaign-evidence-B8oF9xQ6.js";
 //#region src/campaign/gates/sequential.ts
 /**
 * Anytime-valid sequential promotion gate — an e-process (betting
@@ -25,15 +25,17 @@ import { N as pairHoldout } from "./campaign-evidence-D8DBLqLI.js";
 * shifts the null boundary by `manifest.minEffect` — re-deciding the same
 * stream under different parameters after seeing data would reopen optional
 * stopping under a fancier name. The manifest's content hash is verified at
-* construction (sync, same `sha256-content` scheme as `signManifest`).
+* construction with the canonical RFC 8785 digest used by signManifest.
 *
-* Non-iid caveat (stated honestly): the supermartingale guarantee needs each
-* delta's conditional mean under H0 to stay ≤ the null boundary given the
-* past — exchangeable scenario deltas suffice. Scenario streams ordered by
-* difficulty or by scenario family violate this; `decide(ctx)` therefore
-* shuffles the paired deltas with a SEEDED permutation by default (the
-* permutation is data-independent, so bet predictability is preserved).
-* Stratified betting (per-stratum λ) is future work, not implemented here.
+* The supermartingale guarantee requires E[delta_t | past] ≤ the null
+* boundary under H0. Independent sampling with that mean bound suffices.
+* Exchangeability or shuffling alone does not: one fair sign repeated many
+* times is exchangeable and has marginal mean zero, but its conditional mean
+* becomes the revealed sign. A seeded shuffle cannot remove that dependence.
+* `decide(ctx)` pairs cells, then averages repetitions within each scenario
+* before testing. Pass independentUnitByScenarioId for scenarios sharing a
+* latent source or incident. Units must still satisfy the conditional-mean null.
+* Direct `observe()` callers own this sampling and aggregation contract.
 */
 /** Sync twin of `verifyManifest`, so gate construction stays synchronous and
 *  fails loud before any observation is consumed. The digest scheme belongs to
@@ -170,15 +172,16 @@ function seededShuffle(items, seed) {
 }
 /**
 * Anytime-valid sequential paired gate. Conforms to the existing `Gate`
-* contract (`decide(ctx)` consumes candidate vs baseline judge scores via
-* `pairHoldout` — same pairing granularity as the fixed-n gates: full cellId,
-* never scenarioId) and adds a streaming `observe(delta)` entry for campaigns
-* that score cells incrementally and want to stop mid-stream.
+* contract: decide(ctx) pairs candidate and baseline by full cellId, then
+* averages deltas within each configured unit (scenario by default). The
+* e-process consumes one observation per unit, with equal unit weights.
+* The streaming observe(delta) entry consumes caller-aggregated independent
+* units. Repetitions improve a unit's precision and do not increase n.
 *
 * Decision mapping onto the substrate's five-valued `GateDecision`:
 *   - 'promote'            → 'ship'
 *   - 'continue'           → 'need_more_work' (stream ended before maxN with
-*                            the e-value undecided — more reps could decide)
+*                            the e-value undecided — more independent units needed)
 *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT
 *                            evidence of no effect (never a silent default)
 */
@@ -187,6 +190,7 @@ function sequentialPairedGate(options) {
 	const name = options.name ?? "sequentialPairedGate";
 	const manifest = options.preRegistration;
 	const observeStream = makeStream(cfg, options.resume);
+	const independentUnitByScenarioId = options.independentUnitByScenarioId === void 0 ? void 0 : new Map(options.independentUnitByScenarioId);
 	return {
 		name,
 		observe(delta) {
@@ -198,8 +202,18 @@ function sequentialPairedGate(options) {
 		async decide(ctx) {
 			if (!ctx.baselineJudgeScores) throw new Error(`${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own scores would compare the candidate against itself (delta 0, silent no-op)`);
 			const scenarioIds = new Set(ctx.scenarios.map((s) => s.id));
+			const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]));
+			for (const id of scenarioIds) {
+				const unit = unitMap.get(id);
+				if (typeof unit !== "string" || unit.length === 0 || unit.trim() !== unit) throw new Error(`${name}: missing independent unit for scenario '${id}'`);
+			}
 			const paired = pairHoldout(ctx.judgeScores, ctx.baselineJudgeScores, scenarioIds, (s) => s.composite);
-			const deltas = paired.after.map((a, i) => a - paired.before[i]);
+			for (let i = 0; i < paired.cellIds.length; i++) {
+				const delta = paired.after[i] - paired.before[i];
+				if (Math.abs(delta) > cfg.scale) throw new Error(`${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`);
+			}
+			const observations = aggregatePairedHoldout(paired, unitMap);
+			const deltas = observations.after.map((after, i) => after - observations.before[i]);
 			seededShuffle(deltas, cfg.shuffleSeed);
 			const stream = makeStream(cfg);
 			let last;
@@ -216,6 +230,9 @@ function sequentialPairedGate(options) {
 				direction: cfg.direction,
 				minEffect: cfg.minEffect,
 				pairedN: deltas.length,
+				pairedCellN: paired.cellIds.length,
+				observationUnit: independentUnitByScenarioId === void 0 ? "scenario" : "registered",
+				unitIds: observations.unitIds,
 				...manifest ? {
 					preRegisteredId: manifest.id,
 					metric: manifest.metric
@@ -246,9 +263,8 @@ function sequentialPairedGate(options) {
 	};
 }
 /**
-* `SurfaceProposer.decide` adapter — stops the optimization loop the moment
-* the e-process decides the loop has produced a real improvement, instead of
-* always running `maxGenerations`.
+* SurfaceProposer.decide adapter that stops exploration at an e-value threshold.
+* The selected candidate still requires an independent held-out decision.
 *
 * Stream: for each generation g ≥ 1, the per-scenario composite deltas of
 * generation g's top candidate vs the generation-0 top candidate (the
@@ -259,10 +275,11 @@ function sequentialPairedGate(options) {
 * gate (which re-scores on HELD-OUT data — this adapter only spends the
 * exploration budget, it never promotes).
 *
-* Honesty caveats: (1) the incumbent's scores are measured once and shared
-* across all generations' deltas, so type-I control is exact only insofar as
-* those scores approximate the incumbent's true per-scenario means (more reps
-* → tighter); (2) an UNDECIDED process never stops the loop — absence of a
+* This is an exploration stopping heuristic. Selection of the best measured
+* candidate and reuse of measured incumbent scores generally violate the
+* conditional-mean null, even when the candidates have no true improvement.
+* More repetitions reduce noise but do not establish exact type-I control.
+* An UNDECIDED process never stops the loop — absence of a
 * crossing is NOT evidence of no effect, so the loop simply runs its normal
 * course. Calling the adapter repeatedly with a growing history consumes each
 * generation exactly once (re-feeding an already-seen record would double-count
@@ -306,7 +323,7 @@ function sequentialDecide(options = {}) {
 				if (step.n >= minN && step.wealth >= threshold) {
 					stopped = {
 						stop: true,
-						reason: `sequential e-process decided at generation ${history[g].generationIndex}: e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} paired deltas vs the generation-0 incumbent — the improvement is real at α=${alpha}; stop exploring and promote via the gate`
+						reason: `sequential e-process decided at generation ${history[g].generationIndex}: e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} paired deltas vs the generation-0 incumbent; stop exploring and require an independent held-out promotion decision`
 					};
 					processedGenerations = history.length;
 					return stopped;
@@ -322,4 +339,4 @@ function sequentialDecide(options = {}) {
 //#endregion
 export { sequentialPairedGate as n, sequentialDecide as t };
 
-//# sourceMappingURL=sequential-B5gXgcyp.js.map
+//# sourceMappingURL=sequential-DAsyV2T9.js.map

package/dist/sequential-DAsyV2T9.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"sequential-DAsyV2T9.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 with the canonical RFC 8785 digest used by signManifest.\n *\n * The supermartingale guarantee requires E[delta_t | past] ≤ the null\n * boundary under H0. Independent sampling with that mean bound suffices.\n * Exchangeability or shuffling alone does not: one fair sign repeated many\n * times is exchangeable and has marginal mean zero, but its conditional mean\n * becomes the revealed sign. A seeded shuffle cannot remove that dependence.\n * `decide(ctx)` pairs cells, then averages repetitions within each scenario\n * before testing. Pass independentUnitByScenarioId for scenarios sharing a\n * latent source or incident. Units must still satisfy the conditional-mean null.\n * Direct `observe()` callers own this sampling and aggregation contract.\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 { aggregatePairedHoldout, 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 independent paired units 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 independent-unit budget (scenarios by default in decide()). 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 reproducible ordering of scenario deltas in decide().\n   *  Shuffling does not establish the conditional-mean null. Default 1337. */\n  shuffleSeed?: number\n  /** Independent unit for every scenario passed to decide(). Full cell pairs\n   *  are averaged within each unit before testing. Defaults to scenario IDs.\n   *  The mapping is copied at construction. Direct observe() is unchanged. */\n  independentUnitByScenarioId?: ReadonlyMap<string, string>\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 independent-unit delta\n   *  (candidate − baseline, native scale). Aggregate correlated repetitions\n   *  before calling; the caller must justify E[delta_t | past] under H0.\n   *  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) pairs candidate and baseline by full cellId, then\n * averages deltas within each configured unit (scenario by default). The\n * e-process consumes one observation per unit, with equal unit weights.\n * The streaming observe(delta) entry consumes caller-aggregated independent\n * units. Repetitions improve a unit's precision and do not increase n.\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 independent units needed)\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  const independentUnitByScenarioId =\n    options.independentUnitByScenarioId === undefined\n      ? undefined\n      : new Map(options.independentUnitByScenarioId)\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          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]))\n      for (const id of scenarioIds) {\n        const unit = unitMap.get(id)\n        if (typeof unit !== 'string' || unit.length === 0 || unit.trim() !== unit) {\n          throw new Error(`${name}: missing independent unit for scenario '${id}'`)\n        }\n      }\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      for (let i = 0; i < paired.cellIds.length; i++) {\n        const delta = paired.after[i]! - paired.before[i]!\n        if (Math.abs(delta) > cfg.scale) {\n          throw new Error(\n            `${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`,\n          )\n        }\n      }\n      const observations = aggregatePairedHoldout(paired, unitMap)\n      const deltas = observations.after.map((after, i) => after - observations.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        pairedCellN: paired.cellIds.length,\n        observationUnit: independentUnitByScenarioId === undefined ? 'scenario' : 'registered',\n        unitIds: observations.unitIds,\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 that stops exploration at an e-value threshold.\n * The selected candidate still requires an independent held-out decision.\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 * This is an exploration stopping heuristic. Selection of the best measured\n * candidate and reuse of measured incumbent scores generally violate the\n * conditional-mean null, even when the candidates have no true improvement.\n * More repetitions reduce noise but do not establish exact type-I control.\n * 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; stop exploring and require ' +\n              'an independent held-out promotion decision',\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":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmIA,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;;;;;;;;;;;;;;;;AAiBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CACpD,MAAM,8BACJ,QAAQ,gCAAgC,KAAA,IACpC,KAAA,IACA,IAAI,IAAI,QAAQ,2BAA2B;CAEjD,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,UAAU,+BAA+B,IAAI,IAAI,CAAC,GAAG,WAAW,CAAC,CAAC,KAAK,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC;GAC7F,KAAK,MAAM,MAAM,aAAa;IAC5B,MAAM,OAAO,QAAQ,IAAI,EAAE;IAC3B,IAAI,OAAO,SAAS,YAAY,KAAK,WAAW,KAAK,KAAK,KAAK,MAAM,MACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2CAA2C,GAAG,EAAE;GAE5E;GACA,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;IAC9C,MAAM,QAAQ,OAAO,MAAM,KAAM,OAAO,OAAO;IAC/C,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,OACxB,MAAM,IAAI,MACR,GAAG,KAAK,UAAU,OAAO,QAAQ,GAAG,UAAU,MAAM,kBAAkB,IAAI,OAC5E;GAEJ;GACA,MAAM,eAAe,uBAAuB,QAAQ,OAAO;GAC3D,MAAM,SAAS,aAAa,MAAM,KAAK,OAAO,MAAM,QAAQ,aAAa,OAAO,EAAG;GACnF,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,aAAa,OAAO,QAAQ;IAC5B,iBAAiB,gCAAgC,KAAA,IAAY,aAAa;IAC1E,SAAS,aAAa;IACtB,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;KAGtF;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/{series-convergence-DeG33RpC.d.ts → series-convergence-BnMs_uAr.d.ts}

@@ -1,5 +1,5 @@
-import { s as JudgeScore } from "./types-gvRsyJLh.js";
-import { i as ContinuousAgreementOptions, r as ContinuousAgreement } from "./judge-calibration-C5CbMYce.js";
+import { u as JudgeScore } from "./types-CBbLtr2J.js";
+import { i as ContinuousAgreementOptions, r as ContinuousAgreement } from "./judge-calibration-DFtEMlde.js";
 //#region src/statistics/agreement-irr.d.ts
 /**
  * Inter-rater reliability — Krippendorff's α under the squared-difference
@@ -123,4 +123,4 @@ interface SeriesConvergenceResult {
 declare function analyzeSeries(values: number[], options?: SeriesConvergenceOptions): SeriesConvergenceResult;
 //#endregion
 export { CorpusAgreementPerDimension as a, corpusInterRaterAgreement as c, CorpusAgreementOptions as i, corpusInterRaterAgreementFromJudgeScores as l, SeriesConvergenceResult as n, CorpusAgreementReport as o, analyzeSeries as r, CorpusScoreRecord as s, SeriesConvergenceOptions as t, interRaterReliability as u };
-//# sourceMappingURL=series-convergence-DeG33RpC.d.ts.map
+//# sourceMappingURL=series-convergence-BnMs_uAr.d.ts.map

package/dist/{series-convergence-DeG33RpC.d.ts.map → series-convergence-BnMs_uAr.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"series-convergence-DeG33RpC.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../src/series-convergence.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;;;;;;;;;;;;UCvRc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR"}
+{"version":3,"file":"series-convergence-BnMs_uAr.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../src/series-convergence.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;;;;;;;;;;;;UCvRc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR"}

package/dist/{skillopt-optimization-method-C3oYul8v.js → skillopt-optimization-method-CL_0aArC.js}

@@ -1,10 +1,10 @@
+import { at as pairHoldout, it as heldoutSignificance, x as campaignBreakdown, z as surfaceContentHash } from "./campaign-evidence-B8oF9xQ6.js";
 import { n as contentHash, t as canonicalJson } from "./verdict-cache-B3eCVQtY.js";
-import { A as openSearchLedger, G as runCampaign, b as recordCandidatePopulationSearch, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, t as llmJudge } from "./llm-judge-v80Kmu9g.js";
-import { H as surfaceContentHash, M as heldoutSignificance, N as pairHoldout, x as campaignBreakdown } from "./campaign-evidence-D8DBLqLI.js";
-import { S as fsCampaignStorage, _ as removeCredentialEnvironment, b as safePathComponent, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as isRecord, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, x as createRunCostLedger, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-q3VzlGAO.js";
+import { A as openSearchLedger, X as runCampaign, b as recordCandidatePopulationSearch, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, t as llmJudge } from "./llm-judge-DEFZeSiu.js";
+import { S as fsCampaignStorage, _ as removeCredentialEnvironment, b as safePathComponent, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as isRecord, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, x as createRunCostLedger, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-D4dzUBZI.js";
 import { l as deepFreezeCanonicalJson } from "./types-DQ0e2E7y.js";
 import { g as assertServedModelPolicy } from "./llm-client-CGlSi8sb.js";
-import { n as acquireSingleRunLock, t as startExternalOptimizerCallback } from "./external-optimizer-process-Dlz8YxrT.js";
+import { n as acquireSingleRunLock, t as startExternalOptimizerCallback } from "./external-optimizer-process-QDRURJAM.js";
 import { createHash, randomBytes } from "node:crypto";
 import { z } from "zod";
 import { mkdir } from "node:fs/promises";
@@ -2140,4 +2140,4 @@ function skillOptOptimizationMethod(config) {
 //#endregion
 export { externalTextOptimizationMethod as a, REFERENCE_EQUIVALENCE_INPUT_LIMITS as c, runReferenceEquivalenceJudge as d, composeGate as i, REFERENCE_EQUIVALENCE_JUDGE_VERSION as l, gepaOptimizationMethod as n, decodeExternalTextCandidate as o, heldOutGate as r, readExternalOptimizerObservationArtifact as s, skillOptOptimizationMethod as t, createReferenceEquivalenceJudge as u };
 
-//# sourceMappingURL=skillopt-optimization-method-C3oYul8v.js.map
+//# sourceMappingURL=skillopt-optimization-method-CL_0aArC.js.map