@tangle-network/agent-eval 0.175.0 → 0.177.0

package/dist/profile-cell.js

@@ -1,269 +1,2 @@
-import { s as ValidationError } from "./errors-Dngq5h35.js";
-import { n as hashJson } from "./pre-registration-D94b7Of5.js";
-import { createHash } from "node:crypto";
-//#region src/agent-profile-cell.ts
-var AgentProfileCellValidationError = class extends ValidationError {
-	path;
-	constructor(message, path = "") {
-		super(path ? `${message} (at ${path})` : message);
-		this.path = path;
-	}
-};
-const SHA256_HEX = /^[0-9a-f]{64}$/;
-/**
-* A cell id names the digest scheme that produced it. `sha256-rfc8785` is what
-* {@link buildAgentProfileCell} mints; the bare `sha256` form is read-only,
-* carried by cells built under an earlier release, and still verifies.
-*/
-const CELL_ID = /^agent-profile-cell:sha256(?:-rfc8785)?:[0-9a-f]{64}$/;
-const CELL_ID_PREFIX = "agent-profile-cell:sha256-rfc8785:";
-const LEGACY_CELL_ID_PREFIX = "agent-profile-cell:sha256:";
-async function buildAgentProfileCell(input) {
-	const material = await normalizeAgentProfileCellInput(input);
-	const cellId = `${CELL_ID_PREFIX}${await hashJson(material)}`;
-	return {
-		...material,
-		cellId
-	};
-}
-function agentProfileCellHashMaterial(cell) {
-	const { cellId: _cellId, ...material } = cell;
-	return normalizeAgentProfileCell(material);
-}
-/**
-* Verify an `AgentProfileCell`'s `cellId` matches the sha256 of its hash-material
-* fields, confirming the record has not been tampered with. The id names its own
-* digest scheme, so a cell minted by an earlier release verifies under that scheme.
-*/
-async function verifyAgentProfileCell(cell) {
-	validateAgentProfileCell(cell);
-	const material = agentProfileCellHashMaterial(cell);
-	if (cell.cellId.startsWith(CELL_ID_PREFIX)) return cell.cellId === `${CELL_ID_PREFIX}${await hashJson(material)}`;
-	return cell.cellId === `${LEGACY_CELL_ID_PREFIX}${legacyCellDigest(material)}`;
-}
-/**
-* Key-sorted `JSON.stringify` digest. Private and read-only: it verifies a cell
-* id minted before the RFC 8785 scheme, and no path that MINTS an id calls it.
-*/
-function legacyCellDigest(value) {
-	return createHash("sha256").update(JSON.stringify(sortKeysDeep(value)), "utf8").digest("hex");
-}
-function sortKeysDeep(value) {
-	if (value === null || typeof value !== "object") return value;
-	if (Array.isArray(value)) return value.map(sortKeysDeep);
-	const out = {};
-	for (const key of Object.keys(value).sort()) out[key] = sortKeysDeep(value[key]);
-	return out;
-}
-function validateAgentProfileCell(input) {
-	if (input === null || typeof input !== "object") throw new AgentProfileCellValidationError("expected object");
-	const obj = input;
-	expectLiteral(obj.schemaVersion, "agent-profile-cell/v1", "schemaVersion");
-	if (typeof obj.cellId !== "string" || !CELL_ID.test(obj.cellId)) throw new AgentProfileCellValidationError("cellId must match agent-profile-cell:sha256:<64 lowercase hex chars>", "cellId");
-	expectString(obj.profileId, "profileId");
-	validateSource(obj.sourceProfile, "sourceProfile");
-	if (obj.harness !== void 0) validateHarness(obj.harness, "harness");
-	if (obj.model !== void 0) expectString(obj.model, "model");
-	if (obj.promptHash !== void 0) expectString(obj.promptHash, "promptHash");
-	if (obj.dimensions !== void 0) validateDimensions(obj.dimensions, "dimensions");
-	return input;
-}
-function requireAgentProfileCell(record) {
-	if (!record.agentProfile) throw new AgentProfileCellValidationError(`run "${record.runId}" is missing agentProfile; profile-cell grouping requires explicit profile identity`, "agentProfile");
-	return validateAgentProfileCell(record.agentProfile);
-}
-function agentProfileCellKey(record) {
-	return requireAgentProfileCell(record).cellId;
-}
-async function assertRunAgentProfileCell(record) {
-	const profile = requireAgentProfileCell(record);
-	if (!await verifyAgentProfileCell(profile)) throw new AgentProfileCellValidationError(`run "${record.runId}" has an agentProfile.cellId that does not match its content`, "agentProfile.cellId");
-	if (profile.model !== void 0 && profile.model !== record.model) throw new AgentProfileCellValidationError(`run "${record.runId}" agentProfile.model "${profile.model}" does not match model "${record.model}"`, "agentProfile.model");
-	if (profile.promptHash !== void 0 && profile.promptHash !== record.promptHash) throw new AgentProfileCellValidationError(`run "${record.runId}" agentProfile.promptHash "${profile.promptHash}" does not match promptHash "${record.promptHash}"`, "agentProfile.promptHash");
-	return profile;
-}
-function groupRunsByAgentProfileCell(records) {
-	const groups = /* @__PURE__ */ new Map();
-	for (const record of records) {
-		const key = agentProfileCellKey(record);
-		const bucket = groups.get(key);
-		if (bucket) bucket.push(record);
-		else groups.set(key, [record]);
-	}
-	return groups;
-}
-async function normalizeAgentProfileCellInput(input) {
-	return normalizeAgentProfileCell({
-		schemaVersion: "agent-profile-cell/v1",
-		profileId: input.profileId,
-		sourceProfile: await normalizeSourceInput(input.sourceProfile),
-		harness: input.harness,
-		model: input.model,
-		promptHash: input.promptHash,
-		dimensions: input.dimensions
-	});
-}
-function normalizeAgentProfileCell(input) {
-	return compactObject({
-		schemaVersion: "agent-profile-cell/v1",
-		profileId: requireNonEmpty(input.profileId, "profileId"),
-		sourceProfile: normalizeSource(input.sourceProfile),
-		harness: input.harness ? normalizeHarness(input.harness, "harness") : void 0,
-		model: optionalNonEmpty(input.model, "model"),
-		promptHash: optionalNonEmpty(input.promptHash, "promptHash"),
-		dimensions: input.dimensions ? nonEmptyRecord(normalizeDimensions(input.dimensions)) : void 0
-	});
-}
-async function normalizeSourceInput(input) {
-	const kind = requireNonEmpty(input.kind, "sourceProfile.kind");
-	if (input.hash !== void 0 && input.profile !== void 0) throw new AgentProfileCellValidationError("sourceProfile must provide either hash or profile, not both", "sourceProfile");
-	if (input.hash !== void 0) return {
-		kind,
-		hash: requireSha256Hex(input.hash, "sourceProfile.hash")
-	};
-	if (input.profile === void 0) throw new AgentProfileCellValidationError("sourceProfile must provide hash or profile", "sourceProfile");
-	assertJson(input.profile, "sourceProfile.profile");
-	return {
-		kind,
-		hash: await hashJson(input.profile)
-	};
-}
-function normalizeSource(input) {
-	return {
-		kind: requireNonEmpty(input.kind, "sourceProfile.kind"),
-		hash: requireSha256Hex(input.hash, "sourceProfile.hash")
-	};
-}
-function normalizeHarness(input, path) {
-	return compactObject({
-		id: requireNonEmpty(input.id, `${path}.id`),
-		version: optionalNonEmpty(input.version, `${path}.version`),
-		hash: optionalNonEmpty(input.hash, `${path}.hash`)
-	});
-}
-function normalizeDimensions(input) {
-	const out = {};
-	for (const key of Object.keys(input).sort()) {
-		const value = input[key];
-		requireNonEmpty(key, "dimensions.<key>");
-		if (value !== null && typeof value !== "string" && typeof value !== "number" && typeof value !== "boolean") throw new AgentProfileCellValidationError("expected primitive dimension value", `dimensions.${key}`);
-		if (typeof value === "number" && !Number.isFinite(value)) throw new AgentProfileCellValidationError("expected finite number", `dimensions.${key}`);
-		out[key] = value;
-	}
-	return out;
-}
-function compactObject(input) {
-	const out = {};
-	for (const [key, value] of Object.entries(input)) if (value !== void 0) out[key] = value;
-	return out;
-}
-function nonEmptyRecord(input) {
-	return Object.keys(input).length > 0 ? input : void 0;
-}
-function validateSource(value, path) {
-	if (value === null || typeof value !== "object" || Array.isArray(value)) throw new AgentProfileCellValidationError("expected object", path);
-	const rec = value;
-	expectString(rec.kind, `${path}.kind`);
-	requireSha256Hex(rec.hash, `${path}.hash`);
-}
-function validateHarness(value, path) {
-	if (value === null || typeof value !== "object" || Array.isArray(value)) throw new AgentProfileCellValidationError("expected object", path);
-	const rec = value;
-	expectString(rec.id, `${path}.id`);
-	if (rec.version !== void 0) expectString(rec.version, `${path}.version`);
-	if (rec.hash !== void 0) expectString(rec.hash, `${path}.hash`);
-}
-function validateDimensions(value, path) {
-	if (value === null || typeof value !== "object" || Array.isArray(value)) throw new AgentProfileCellValidationError("expected object", path);
-	normalizeDimensions(value);
-}
-function assertJson(value, path) {
-	if (value === null) return;
-	const type = typeof value;
-	if (type === "string" || type === "boolean") return;
-	if (type === "number") {
-		if (!Number.isFinite(value)) throw new AgentProfileCellValidationError("expected finite number", path);
-		return;
-	}
-	if (Array.isArray(value)) {
-		value.forEach((item, index) => {
-			assertJson(item, `${path}[${index}]`);
-		});
-		return;
-	}
-	if (type === "object") {
-		for (const [key, nested] of Object.entries(value)) {
-			requireNonEmpty(key, `${path}.<key>`);
-			assertJson(nested, `${path}.${key}`);
-		}
-		return;
-	}
-	throw new AgentProfileCellValidationError("expected JSON-compatible value", path);
-}
-function expectLiteral(value, expected, path) {
-	if (value !== expected) throw new AgentProfileCellValidationError(`expected ${expected}`, path);
-}
-function expectString(value, path) {
-	if (typeof value !== "string" || value.length === 0) throw new AgentProfileCellValidationError("expected non-empty string", path);
-}
-function requireNonEmpty(value, path) {
-	if (typeof value !== "string" || value.length === 0) throw new AgentProfileCellValidationError("expected non-empty string", path);
-	return value;
-}
-function optionalNonEmpty(value, path) {
-	if (value === void 0) return void 0;
-	return requireNonEmpty(value, path);
-}
-function requireSha256Hex(value, path) {
-	if (typeof value !== "string" || !SHA256_HEX.test(value)) throw new AgentProfileCellValidationError("expected 64 lowercase sha256 hex chars", path);
-	return value;
-}
-/** Canonical `sourceProfile.kind` values. Two products fingerprinting the
-*  same canonical profile MUST use the same kind for their cells to share
-*  `sourceProfile.hash`. Extend rather than create new strings — adding a
-*  new kind is a deliberate cross-product schema change. */
-const AGENT_PROFILE_KINDS = { 
-/** A profile declared via `defineAgentProfile(...)` from
-*  `@tangle-network/agent-interface`. The default kind for router-backed
-*  and sandbox-backed products. */
-AGENT_INTERFACE_PROFILE: "agent-interface-profile" };
-/** Canonicalize an arbitrary value into `AgentProfileJson` by JSON
-*  round-trip. Throws when the value contains anything not representable
-*  as JSON (functions, BigInt, cycles) — non-portable profiles fail loud
-*  rather than silently dropping fields. */
-function toAgentProfileJson(value) {
-	let serialized;
-	try {
-		serialized = JSON.stringify(value);
-	} catch (err) {
-		throw new AgentProfileCellValidationError(`agent profile must be JSON-serializable: ${err instanceof Error ? err.message : String(err)}`, "sourceProfile.profile");
-	}
-	if (serialized === void 0) throw new AgentProfileCellValidationError("agent profile must be JSON-serializable (got undefined after JSON.stringify)", "sourceProfile.profile");
-	return JSON.parse(serialized);
-}
-/** Higher-level helper that hard-codes the canonical
-*  `agent-interface-profile` kind plus the JSON canonicalization. Equivalent
-*  to calling `buildAgentProfileCell` with `profileId = \`${name}@${version}\``
-*  and `sourceProfile = { kind: AGENT_INTERFACE_PROFILE, profile: <round-tripped> }`.
-*
-*  Use this from any product consuming an agent-interface `AgentProfile`; the
-*  manual `buildAgentProfileCell` call is reserved for advanced cases
-*  (custom kinds, pre-computed source hashes, alternate profileId
-*  conventions). */
-async function buildAgentInterfaceProfileCell(profile, input) {
-	if (!profile || typeof profile !== "object") throw new AgentProfileCellValidationError("AgentProfile must be an object", "profile");
-	if (typeof profile.name !== "string" || profile.name.length === 0) throw new AgentProfileCellValidationError("AgentProfile must have a non-empty `name`", "profile.name");
-	if (typeof profile.version !== "string" || profile.version.length === 0) throw new AgentProfileCellValidationError("AgentProfile must have a non-empty `version`", "profile.version");
-	return buildAgentProfileCell({
-		...input,
-		profileId: `${profile.name}@${profile.version}`,
-		sourceProfile: {
-			kind: AGENT_PROFILE_KINDS.AGENT_INTERFACE_PROFILE,
-			profile: toAgentProfileJson(profile)
-		}
-	});
-}
-//#endregion
+import { a as assertRunAgentProfileCell, c as groupRunsByAgentProfileCell, d as validateAgentProfileCell, f as verifyAgentProfileCell, i as agentProfileCellKey, l as requireAgentProfileCell, n as AgentProfileCellValidationError, o as buildAgentInterfaceProfileCell, r as agentProfileCellHashMaterial, s as buildAgentProfileCell, t as AGENT_PROFILE_KINDS, u as toAgentProfileJson } from "./agent-profile-cell-0gSi5ffD.js";
 export { AGENT_PROFILE_KINDS, AgentProfileCellValidationError, agentProfileCellHashMaterial, agentProfileCellKey, assertRunAgentProfileCell, buildAgentInterfaceProfileCell, buildAgentProfileCell, groupRunsByAgentProfileCell, requireAgentProfileCell, toAgentProfileJson, validateAgentProfileCell, verifyAgentProfileCell };
-
-//# sourceMappingURL=profile-cell.js.map

package/dist/{promotion-policy-LY9mVQ7W.js → promotion-policy-DWOm70gx.js}

@@ -1,5 +1,5 @@
 import { r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
-import { o as pairHoldout, r as detectScale, s as decidePairedPromotion } from "./power-preflight-CFXm0Vjo.js";
+import { A as detectScale, N as pairHoldout, P as decidePairedPromotion } from "./campaign-evidence-D8DBLqLI.js";
 //#region src/campaign/gates/promotion-policy.ts
 /**
 * Promotion policy over the evidence VECTOR — the substrate's answer to "never
@@ -183,4 +183,4 @@ function paretoSignificanceGate(options) {
 //#endregion
 export { paretoPolicy as n, paretoSignificanceGate as r, buildEvidenceVector as t };
 
-//# sourceMappingURL=promotion-policy-LY9mVQ7W.js.map
+//# sourceMappingURL=promotion-policy-DWOm70gx.js.map

package/dist/{promotion-policy-LY9mVQ7W.js.map → promotion-policy-DWOm70gx.js.map}

@@ -1 +1 @@
-{"version":3,"file":"promotion-policy-LY9mVQ7W.js","names":[],"sources":["../src/campaign/gates/promotion-policy.ts"],"sourcesContent":["/**\n * Promotion policy over the evidence VECTOR — the substrate's answer to \"never\n * collapse the multi-objective promotion decision into one scalar.\" A\n * `defaultProductionGate` is one opinionated composition; this module factors\n * the decision into two reusable pieces so MANY policies can compete over the\n * SAME evidence (the quant-desk pattern: one evidence bus, plural strategies):\n *\n *   buildEvidenceVector(ctx, objectives, opts) -> EvidenceVector   // the bus\n *   PromotionPolicy = (ev: EvidenceVector) => GateResult           // a strategy\n *   paretoPolicy(ev)                                               // the default strategy\n *   paretoSignificanceGate(options): Gate                          // bus + policy as a Gate\n *\n * The Pareto policy is SYMMETRIC multi-objective: every objective is BOTH a\n * potential gain source AND a safety floor (unlike `defaultProductionGate`,\n * where only `composite` can win and `criticalDimensions` are pure floors). A\n * candidate ships iff it weakly DOMINATES the baseline at the confidence level —\n * no objective credibly worse (CI floor breach) AND at least one objective\n * credibly better (CI gain). Insufficient evidence on ANY axis -> need_more_work\n * (NOT folded into hold: \"gather more reps\" and \"reject\" are different actions).\n *\n * Cost/latency are NOT CI axes here — `GateContext` carries only an aggregate\n * per-side cost, no per-cell observation vector to bootstrap. Treat them as hard\n * constraints (compose with a budget gate via `composeGate`), not faked CIs.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n} from '../../paired-promotion-decision'\nimport type { Direction } from '../../pareto'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance, or when the exact small-sample test proves\n   *  a drop past it. When omitted it auto-scales off observed magnitudes\n   *  (0.05 on [0,1], 5 on 0-100), matching `dimensionRegressions`. */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the good-direction paired bootstrap. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis.\n   *\n   *  DIAGNOSTIC on a pass/fail axis: there the verdict is decided on Tango's\n   *  score interval instead, because a percentile bootstrap over a three-atom\n   *  delta lattice is not a valid interval at the nonzero margin `floorTolerance`\n   *  and `gainThreshold` create. `ci` carries the interval that decided. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low`/`.high` bracket. `'mean'` unless the\n   *  caller asked for the median — on a pass/fail axis the median and its whole\n   *  CI are pinned at 0 by tie domination and can see neither a gain nor a\n   *  regression. `bootstrap.median` still carries the median point estimate. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval the axis verdict was actually decided on, good-direction and\n   *  in the axis's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail axis; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width — no evidence in either direction, so the axis is\n   *  neither improved nor regressed however the point estimate sits. */\n  indeterminate: boolean\n  /** Paired observations contributing to this axis. */\n  n: number\n  minimumRequired: number\n  decisionMethod: PairedDecisionMethod\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. The exact small-sample test may require more\n   *  observations at the selected confidence. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n  /** Paired statistic every axis CI is computed on. Default `'mean'` — see\n   *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const n = paired.before.length\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Axes are decided on the MEAN paired delta — which for a pass/fail axis is\n    // exactly the change in success rate. The median is structurally blind on\n    // the shapes eval data lands in: with most pairs tied its bootstrap CI\n    // collapses to [0,0] and the axis reads 'flat', hiding real gains AND real\n    // regressions — pass/fail axes on ANY encoding ({0,1} and the 0-100 one\n    // `detectScale` exists for), and low-cardinality axes even below half ties.\n    // Orthogonal to `pairedDeltaTest`'s own small-sample switch: that picks the\n    // TEST (bootstrap CI at n ≥ 20, exact sign test below), this picks the\n    // ESTIMATOR the test is applied to. Both are needed — an exact sign test on\n    // a tie-pinned median is still blind.\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    // Both burdens of proof route through the ONE shared rule\n    // (`decidePairedPromotion`), so a pass/fail axis is judged on Tango's score\n    // interval — the only paired-binary construction valid at the nonzero\n    // margins `gainThreshold` / `floorTolerance` create — and a zero-width\n    // interval cannot buy a verdict in either direction.\n    const improvement = decidePairedPromotion(before, after, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: gainThreshold,\n      minPairs: opts.minProductiveRuns,\n    })\n    const regression = decidePairedPromotion(after, before, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: floorTolerance,\n      minPairs: opts.minProductiveRuns,\n    })\n    const bootstrap =\n      improvement.bootstrap ??\n      pairedBootstrap(before, after, {\n        confidence,\n        resamples,\n        statistic: bootstrapStatistic,\n        seed,\n      })\n    // A floor breach fires on EITHER burden of proof, because they cover\n    // different failures and the floor is the anti-Goodhart guard:\n    //   - `improvement.low < -floorTolerance` — the CREDIBLE WORST CASE exceeds\n    //     the tolerance. This is the contract `AxisEvidence.floorTolerance` and\n    //     `paretoPolicy`'s own reason string state, and it is the conservative\n    //     posture a safety axis needs: block unless the data can rule the\n    //     breach out, rather than waiting for the breach to be proven. Read off\n    //     the DECIDING interval, so a pass/fail axis is not screened by a\n    //     bootstrap that is pinned wherever ties dominate.\n    //   - `regression.promote` — a PROVEN drop past the tolerance. Adds the\n    //     small-sample path, where the decision is an exact sign test because\n    //     the bootstrap interval is descriptive only.\n    // The credible-worst-case arm stays on the BOOTSTRAP, deliberately. Reading\n    // it off the score interval instead would change what the floor MEANS on a\n    // pass/fail axis: with every pair concordant the score interval is\n    // ±z²/(n+z²) — ±0.39 at n=6, ±0.16 at n=20 — so a completely unchanged\n    // safety axis would breach a 0.05 floor at any realistic n, and the gate\n    // would refuse everything. That the bootstrap arm is instead fail-OPEN on a\n    // tied pass/fail axis is a real and separate weakness: the honest fix is a\n    // minimum-power requirement on the floor, not a wider interval, because the\n    // data genuinely cannot rule a 5pp drop out at n=20 and a gate that says so\n    // by blocking every candidate is not usable. `regression.promote` — the\n    // PROVEN-drop arm — does route through the shared rule, so a real pass/fail\n    // regression is now caught on an interval valid at the nonzero tolerance.\n    const floorBreached = bootstrap.low < -floorTolerance || regression.promote\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict = !improvement.sufficient\n      ? 'few_runs'\n      : floorBreached\n        ? 'regressed'\n        : improvement.promote\n          ? 'improved'\n          : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: improvement.low, high: improvement.high },\n      decisionStatistic: improvement.statistic,\n      mcnemar: improvement.mcnemar,\n      indeterminate: improvement.indeterminate,\n      n,\n      minimumRequired: improvement.minimumPairs,\n      decisionMethod: improvement.method,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * The default strategy: symmetric multi-objective Pareto significance. Ship iff\n * the candidate weakly dominates the baseline at the confidence level — no axis\n * credibly worse AND ≥1 axis credibly better. Floor breach on any axis → hold\n * (anti-Goodhart, dominates everything). Insufficient evidence on any axis →\n * need_more_work. Statistically equivalent → hold (never ship noise).\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    status:\n      ax.verdict === 'regressed'\n        ? ('fail' as const)\n        : ax.verdict === 'few_runs'\n          ? ('not_evaluated' as const)\n          : ('pass' as const),\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.ci.low,\n      ciHigh: ax.ci.high,\n      decisionStatistic: ax.decisionStatistic,\n      decisionMethod: ax.decisionMethod,\n      mcnemar: ax.mcnemar,\n      indeterminate: ax.indeterminate,\n      bootstrapCiLow: ax.bootstrap.low,\n      bootstrapCiHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const fewRuns = ev.axes.filter((a) => a.verdict === 'few_runs')\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // Floor breach dominates: a credible regression on ANY axis blocks ship even\n    // if another axis improved. This makes the +gain/−safety false positive\n    // structurally impossible whenever the safety dim is an objective.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' regressed: good-direction CI.low ${a.ci.low.toFixed(3)} < -${a.floorTolerance} (n=${a.n})`,\n      )\n    }\n  } else if (fewRuns.length > 0) {\n    // No credible regression on the scored axes, but ≥1 axis lacks the evidence\n    // to claim a gain ⇒ gather more reps, do NOT reject.\n    decision = 'need_more_work'\n    for (const a of fewRuns) {\n      reasons.push(\n        `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`,\n      )\n    }\n  } else if (improved.length > 0) {\n    // Weakly dominates (no axis worse) AND strictly better on ≥1 axis ⇒ a Pareto\n    // improvement at the confidence level.\n    decision = 'ship'\n    reasons.push(\n      `Pareto improvement at the confidence level: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.ci.low > 0 ? a.ci.low.toFixed(3) : a.bootstrap.mean.toFixed(3)} (CI.low ${a.ci.low.toFixed(3)})`,\n        )\n        .join(', ')}; no objective regressed`,\n    )\n  } else {\n    // Enough evidence, nothing credibly better or worse ⇒ statistically\n    // equivalent. Do NOT ship a no-op.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: candidate statistically equivalent to baseline on every objective',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2IA,SAAgB,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;CAChB,IAAI,WAAW,WAAW,GACxB,MAAM,IAAI,MAAM,oDAAoD;CAEtE,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,IAAI,uBAAuB,IAAI;CAChD,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;CAE1D,MAAM,OAAuB,CAAC;CAC9B,KAAK,MAAM,OAAO,YAAY;EAC5B,IAAI;EACJ,IAAI,IAAI,OAAO,SAAS,aACtB,UAAU,MAAM,EAAE;OACb;GACL,MAAM,MAAM,IAAI,OAAO;GACvB,UAAU,MAAM,EAAE,WAAW;EAC/B;EACA,MAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;EAIzE,MAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;EACrE,MAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;EACnE,MAAM,IAAI,OAAO,OAAO;EACxB,MAAM,iBACJ,IAAI,kBAAkB,MAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC9E,MAAM,gBAAgB,IAAI,iBAAiB;EAW3C,MAAM,qBAAqB,KAAK,aAAA;EAMhC,MAAM,cAAc,sBAAsB,QAAQ,OAAO;GACvD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,aAAa,sBAAsB,OAAO,QAAQ;GACtD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,YACJ,YAAY,aACZ,gBAAgB,QAAQ,OAAO;GAC7B;GACA;GACA,WAAW;GACX;EACF,CAAC;EAyBH,MAAM,gBAAgB,UAAU,MAAM,CAAC,kBAAkB,WAAW;EAMpE,MAAM,UAAuB,CAAC,YAAY,aACtC,aACA,gBACE,cACA,YAAY,UACV,aACA;EACR,KAAK,KAAK;GACR,MAAM,IAAI;GACV,QAAQ,IAAI;GACZ,WAAW,IAAI;GACf;GACA;GACA,IAAI;IAAE,KAAK,YAAY;IAAK,MAAM,YAAY;GAAK;GACnD,mBAAmB,YAAY;GAC/B,SAAS,YAAY;GACrB,eAAe,YAAY;GAC3B;GACA,iBAAiB,YAAY;GAC7B,gBAAgB,YAAY;GAC5B;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC;CAEnD,OAAO;EAAE;EAAM,MADF,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;EAC1B,MAAM;GAAE,WAAW,IAAI,KAAK;GAAW,UAAU,IAAI,KAAK;EAAS;CAAE;AAC5F;;;;;;;;AASA,MAAa,gBAAiC,OAAO;CACnD,MAAM,oBAAoB,GAAG,KAAK,KAAK,QAAQ;EAC7C,MAAM,aAAa,GAAG;EACtB,QACE,GAAG,YAAY,cACV,SACD,GAAG,YAAY,aACZ,kBACA;EACT,QAAQ;GACN,WAAW,GAAG;GACd,QAAQ,GAAG;GACX,SAAS,GAAG;GACZ,GAAG,GAAG;GACN,aAAa,GAAG,UAAU;GAC1B,OAAO,GAAG,GAAG;GACb,QAAQ,GAAG,GAAG;GACd,mBAAmB,GAAG;GACtB,gBAAgB,GAAG;GACnB,SAAS,GAAG;GACZ,eAAe,GAAG;GAClB,gBAAgB,GAAG,UAAU;GAC7B,iBAAiB,GAAG,UAAU;GAC9B,YAAY,GAAG,UAAU;GACzB,eAAe,GAAG;GAClB,gBAAgB,GAAG;EACrB;CACF,EAAE;CAEF,MAAM,YAAY,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,WAAW;CACjE,MAAM,UAAU,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAC9D,MAAM,WAAW,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAE/D,IAAI;CACJ,MAAM,UAAoB,CAAC;CAC3B,IAAI,UAAU,SAAS,GAAG;EAIxB,WAAW;EACX,KAAK,MAAM,KAAK,WACd,QAAQ,KACN,cAAc,EAAE,KAAK,qCAAqC,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,MAAM,EAAE,eAAe,MAAM,EAAE,EAAE,EACjH;CAEJ,OAAO,IAAI,QAAQ,SAAS,GAAG;EAG7B,WAAW;EACX,KAAK,MAAM,KAAK,SACd,QAAQ,KACN,cAAc,EAAE,KAAK,eAAe,EAAE,EAAE,2DAC1C;CAEJ,OAAO,IAAI,SAAS,SAAS,GAAG;EAG9B,WAAW;EACX,QAAQ,KACN,+CAA+C,SAC5C,KACE,MACC,IAAI,EAAE,KAAK,KAAK,EAAE,GAAG,MAAM,IAAI,EAAE,GAAG,IAAI,QAAQ,CAAC,IAAI,EAAE,UAAU,KAAK,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,EACpH,CAAC,CACA,KAAK,IAAI,EAAE,yBAChB;CACF,OAAO;EAGL,WAAW;EACX,QAAQ,KACN,0FACF;CACF;CAIA,MAAM,YAAY,GAAG,KAAK,MAAM,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK;CAChF,OAAO;EAAE;EAAU;EAAS;EAAmB,OAAO,WAAW,UAAU;CAAO;AACpF;;;;;;AAiBA,SAAgB,uBACd,SAC4B;CAC5B,IAAI,QAAQ,WAAW,WAAW,GAChC,MAAM,IAAI,MAAM,uDAAuD;CAEzE,MAAM,SAAS,QAAQ,UAAU;CACjC,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB,MAAM,OAAO,KAA6D;GACxE,MAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;GAC/D,OAAO,OAAO,EAAE;EAClB;CACF;AACF"}
+{"version":3,"file":"promotion-policy-DWOm70gx.js","names":[],"sources":["../src/campaign/gates/promotion-policy.ts"],"sourcesContent":["/**\n * Promotion policy over the evidence VECTOR — the substrate's answer to \"never\n * collapse the multi-objective promotion decision into one scalar.\" A\n * `defaultProductionGate` is one opinionated composition; this module factors\n * the decision into two reusable pieces so MANY policies can compete over the\n * SAME evidence (the quant-desk pattern: one evidence bus, plural strategies):\n *\n *   buildEvidenceVector(ctx, objectives, opts) -> EvidenceVector   // the bus\n *   PromotionPolicy = (ev: EvidenceVector) => GateResult           // a strategy\n *   paretoPolicy(ev)                                               // the default strategy\n *   paretoSignificanceGate(options): Gate                          // bus + policy as a Gate\n *\n * The Pareto policy is SYMMETRIC multi-objective: every objective is BOTH a\n * potential gain source AND a safety floor (unlike `defaultProductionGate`,\n * where only `composite` can win and `criticalDimensions` are pure floors). A\n * candidate ships iff it weakly DOMINATES the baseline at the confidence level —\n * no objective credibly worse (CI floor breach) AND at least one objective\n * credibly better (CI gain). Insufficient evidence on ANY axis -> need_more_work\n * (NOT folded into hold: \"gather more reps\" and \"reject\" are different actions).\n *\n * Cost/latency are NOT CI axes here — `GateContext` carries only an aggregate\n * per-side cost, no per-cell observation vector to bootstrap. Treat them as hard\n * constraints (compose with a budget gate via `composeGate`), not faked CIs.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n} from '../../paired-promotion-decision'\nimport type { Direction } from '../../pareto'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance, or when the exact small-sample test proves\n   *  a drop past it. When omitted it auto-scales off observed magnitudes\n   *  (0.05 on [0,1], 5 on 0-100), matching `dimensionRegressions`. */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the good-direction paired bootstrap. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis.\n   *\n   *  DIAGNOSTIC on a pass/fail axis: there the verdict is decided on Tango's\n   *  score interval instead, because a percentile bootstrap over a three-atom\n   *  delta lattice is not a valid interval at the nonzero margin `floorTolerance`\n   *  and `gainThreshold` create. `ci` carries the interval that decided. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low`/`.high` bracket. `'mean'` unless the\n   *  caller asked for the median — on a pass/fail axis the median and its whole\n   *  CI are pinned at 0 by tie domination and can see neither a gain nor a\n   *  regression. `bootstrap.median` still carries the median point estimate. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval the axis verdict was actually decided on, good-direction and\n   *  in the axis's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail axis; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width — no evidence in either direction, so the axis is\n   *  neither improved nor regressed however the point estimate sits. */\n  indeterminate: boolean\n  /** Paired observations contributing to this axis. */\n  n: number\n  minimumRequired: number\n  decisionMethod: PairedDecisionMethod\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. The exact small-sample test may require more\n   *  observations at the selected confidence. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n  /** Paired statistic every axis CI is computed on. Default `'mean'` — see\n   *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const n = paired.before.length\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Axes are decided on the MEAN paired delta — which for a pass/fail axis is\n    // exactly the change in success rate. The median is structurally blind on\n    // the shapes eval data lands in: with most pairs tied its bootstrap CI\n    // collapses to [0,0] and the axis reads 'flat', hiding real gains AND real\n    // regressions — pass/fail axes on ANY encoding ({0,1} and the 0-100 one\n    // `detectScale` exists for), and low-cardinality axes even below half ties.\n    // Orthogonal to `pairedDeltaTest`'s own small-sample switch: that picks the\n    // TEST (bootstrap CI at n ≥ 20, exact sign test below), this picks the\n    // ESTIMATOR the test is applied to. Both are needed — an exact sign test on\n    // a tie-pinned median is still blind.\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    // Both burdens of proof route through the ONE shared rule\n    // (`decidePairedPromotion`), so a pass/fail axis is judged on Tango's score\n    // interval — the only paired-binary construction valid at the nonzero\n    // margins `gainThreshold` / `floorTolerance` create — and a zero-width\n    // interval cannot buy a verdict in either direction.\n    const improvement = decidePairedPromotion(before, after, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: gainThreshold,\n      minPairs: opts.minProductiveRuns,\n    })\n    const regression = decidePairedPromotion(after, before, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: floorTolerance,\n      minPairs: opts.minProductiveRuns,\n    })\n    const bootstrap =\n      improvement.bootstrap ??\n      pairedBootstrap(before, after, {\n        confidence,\n        resamples,\n        statistic: bootstrapStatistic,\n        seed,\n      })\n    // A floor breach fires on EITHER burden of proof, because they cover\n    // different failures and the floor is the anti-Goodhart guard:\n    //   - `improvement.low < -floorTolerance` — the CREDIBLE WORST CASE exceeds\n    //     the tolerance. This is the contract `AxisEvidence.floorTolerance` and\n    //     `paretoPolicy`'s own reason string state, and it is the conservative\n    //     posture a safety axis needs: block unless the data can rule the\n    //     breach out, rather than waiting for the breach to be proven. Read off\n    //     the DECIDING interval, so a pass/fail axis is not screened by a\n    //     bootstrap that is pinned wherever ties dominate.\n    //   - `regression.promote` — a PROVEN drop past the tolerance. Adds the\n    //     small-sample path, where the decision is an exact sign test because\n    //     the bootstrap interval is descriptive only.\n    // The credible-worst-case arm stays on the BOOTSTRAP, deliberately. Reading\n    // it off the score interval instead would change what the floor MEANS on a\n    // pass/fail axis: with every pair concordant the score interval is\n    // ±z²/(n+z²) — ±0.39 at n=6, ±0.16 at n=20 — so a completely unchanged\n    // safety axis would breach a 0.05 floor at any realistic n, and the gate\n    // would refuse everything. That the bootstrap arm is instead fail-OPEN on a\n    // tied pass/fail axis is a real and separate weakness: the honest fix is a\n    // minimum-power requirement on the floor, not a wider interval, because the\n    // data genuinely cannot rule a 5pp drop out at n=20 and a gate that says so\n    // by blocking every candidate is not usable. `regression.promote` — the\n    // PROVEN-drop arm — does route through the shared rule, so a real pass/fail\n    // regression is now caught on an interval valid at the nonzero tolerance.\n    const floorBreached = bootstrap.low < -floorTolerance || regression.promote\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict = !improvement.sufficient\n      ? 'few_runs'\n      : floorBreached\n        ? 'regressed'\n        : improvement.promote\n          ? 'improved'\n          : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: improvement.low, high: improvement.high },\n      decisionStatistic: improvement.statistic,\n      mcnemar: improvement.mcnemar,\n      indeterminate: improvement.indeterminate,\n      n,\n      minimumRequired: improvement.minimumPairs,\n      decisionMethod: improvement.method,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * The default strategy: symmetric multi-objective Pareto significance. Ship iff\n * the candidate weakly dominates the baseline at the confidence level — no axis\n * credibly worse AND ≥1 axis credibly better. Floor breach on any axis → hold\n * (anti-Goodhart, dominates everything). Insufficient evidence on any axis →\n * need_more_work. Statistically equivalent → hold (never ship noise).\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    status:\n      ax.verdict === 'regressed'\n        ? ('fail' as const)\n        : ax.verdict === 'few_runs'\n          ? ('not_evaluated' as const)\n          : ('pass' as const),\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.ci.low,\n      ciHigh: ax.ci.high,\n      decisionStatistic: ax.decisionStatistic,\n      decisionMethod: ax.decisionMethod,\n      mcnemar: ax.mcnemar,\n      indeterminate: ax.indeterminate,\n      bootstrapCiLow: ax.bootstrap.low,\n      bootstrapCiHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const fewRuns = ev.axes.filter((a) => a.verdict === 'few_runs')\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // Floor breach dominates: a credible regression on ANY axis blocks ship even\n    // if another axis improved. This makes the +gain/−safety false positive\n    // structurally impossible whenever the safety dim is an objective.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' regressed: good-direction CI.low ${a.ci.low.toFixed(3)} < -${a.floorTolerance} (n=${a.n})`,\n      )\n    }\n  } else if (fewRuns.length > 0) {\n    // No credible regression on the scored axes, but ≥1 axis lacks the evidence\n    // to claim a gain ⇒ gather more reps, do NOT reject.\n    decision = 'need_more_work'\n    for (const a of fewRuns) {\n      reasons.push(\n        `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`,\n      )\n    }\n  } else if (improved.length > 0) {\n    // Weakly dominates (no axis worse) AND strictly better on ≥1 axis ⇒ a Pareto\n    // improvement at the confidence level.\n    decision = 'ship'\n    reasons.push(\n      `Pareto improvement at the confidence level: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.ci.low > 0 ? a.ci.low.toFixed(3) : a.bootstrap.mean.toFixed(3)} (CI.low ${a.ci.low.toFixed(3)})`,\n        )\n        .join(', ')}; no objective regressed`,\n    )\n  } else {\n    // Enough evidence, nothing credibly better or worse ⇒ statistically\n    // equivalent. Do NOT ship a no-op.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: candidate statistically equivalent to baseline on every objective',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2IA,SAAgB,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;CAChB,IAAI,WAAW,WAAW,GACxB,MAAM,IAAI,MAAM,oDAAoD;CAEtE,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,IAAI,uBAAuB,IAAI;CAChD,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;CAE1D,MAAM,OAAuB,CAAC;CAC9B,KAAK,MAAM,OAAO,YAAY;EAC5B,IAAI;EACJ,IAAI,IAAI,OAAO,SAAS,aACtB,UAAU,MAAM,EAAE;OACb;GACL,MAAM,MAAM,IAAI,OAAO;GACvB,UAAU,MAAM,EAAE,WAAW;EAC/B;EACA,MAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;EAIzE,MAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;EACrE,MAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;EACnE,MAAM,IAAI,OAAO,OAAO;EACxB,MAAM,iBACJ,IAAI,kBAAkB,MAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC9E,MAAM,gBAAgB,IAAI,iBAAiB;EAW3C,MAAM,qBAAqB,KAAK,aAAA;EAMhC,MAAM,cAAc,sBAAsB,QAAQ,OAAO;GACvD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,aAAa,sBAAsB,OAAO,QAAQ;GACtD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,YACJ,YAAY,aACZ,gBAAgB,QAAQ,OAAO;GAC7B;GACA;GACA,WAAW;GACX;EACF,CAAC;EAyBH,MAAM,gBAAgB,UAAU,MAAM,CAAC,kBAAkB,WAAW;EAMpE,MAAM,UAAuB,CAAC,YAAY,aACtC,aACA,gBACE,cACA,YAAY,UACV,aACA;EACR,KAAK,KAAK;GACR,MAAM,IAAI;GACV,QAAQ,IAAI;GACZ,WAAW,IAAI;GACf;GACA;GACA,IAAI;IAAE,KAAK,YAAY;IAAK,MAAM,YAAY;GAAK;GACnD,mBAAmB,YAAY;GAC/B,SAAS,YAAY;GACrB,eAAe,YAAY;GAC3B;GACA,iBAAiB,YAAY;GAC7B,gBAAgB,YAAY;GAC5B;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC;CAEnD,OAAO;EAAE;EAAM,MADF,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;EAC1B,MAAM;GAAE,WAAW,IAAI,KAAK;GAAW,UAAU,IAAI,KAAK;EAAS;CAAE;AAC5F;;;;;;;;AASA,MAAa,gBAAiC,OAAO;CACnD,MAAM,oBAAoB,GAAG,KAAK,KAAK,QAAQ;EAC7C,MAAM,aAAa,GAAG;EACtB,QACE,GAAG,YAAY,cACV,SACD,GAAG,YAAY,aACZ,kBACA;EACT,QAAQ;GACN,WAAW,GAAG;GACd,QAAQ,GAAG;GACX,SAAS,GAAG;GACZ,GAAG,GAAG;GACN,aAAa,GAAG,UAAU;GAC1B,OAAO,GAAG,GAAG;GACb,QAAQ,GAAG,GAAG;GACd,mBAAmB,GAAG;GACtB,gBAAgB,GAAG;GACnB,SAAS,GAAG;GACZ,eAAe,GAAG;GAClB,gBAAgB,GAAG,UAAU;GAC7B,iBAAiB,GAAG,UAAU;GAC9B,YAAY,GAAG,UAAU;GACzB,eAAe,GAAG;GAClB,gBAAgB,GAAG;EACrB;CACF,EAAE;CAEF,MAAM,YAAY,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,WAAW;CACjE,MAAM,UAAU,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAC9D,MAAM,WAAW,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAE/D,IAAI;CACJ,MAAM,UAAoB,CAAC;CAC3B,IAAI,UAAU,SAAS,GAAG;EAIxB,WAAW;EACX,KAAK,MAAM,KAAK,WACd,QAAQ,KACN,cAAc,EAAE,KAAK,qCAAqC,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,MAAM,EAAE,eAAe,MAAM,EAAE,EAAE,EACjH;CAEJ,OAAO,IAAI,QAAQ,SAAS,GAAG;EAG7B,WAAW;EACX,KAAK,MAAM,KAAK,SACd,QAAQ,KACN,cAAc,EAAE,KAAK,eAAe,EAAE,EAAE,2DAC1C;CAEJ,OAAO,IAAI,SAAS,SAAS,GAAG;EAG9B,WAAW;EACX,QAAQ,KACN,+CAA+C,SAC5C,KACE,MACC,IAAI,EAAE,KAAK,KAAK,EAAE,GAAG,MAAM,IAAI,EAAE,GAAG,IAAI,QAAQ,CAAC,IAAI,EAAE,UAAU,KAAK,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,EACpH,CAAC,CACA,KAAK,IAAI,EAAE,yBAChB;CACF,OAAO;EAGL,WAAW;EACX,QAAQ,KACN,0FACF;CACF;CAIA,MAAM,YAAY,GAAG,KAAK,MAAM,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK;CAChF,OAAO;EAAE;EAAU;EAAS;EAAmB,OAAO,WAAW,UAAU;CAAO;AACpF;;;;;;AAiBA,SAAgB,uBACd,SAC4B;CAC5B,IAAI,QAAQ,WAAW,WAAW,GAChC,MAAM,IAAI,MAAM,uDAAuD;CAEzE,MAAM,SAAS,QAAQ,UAAU;CACjC,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB,MAAM,OAAO,KAA6D;GACxE,MAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;GAC/D,OAAO,OAAO,EAAE;EAClB;CACF;AACF"}

package/dist/{release-confidence-BsGEg_xg.js → release-confidence-BcGCclTB.js}

@@ -2,7 +2,7 @@ import { c as VerificationError, s as ValidationError } from "./errors-Dngq5h35.
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { t as FAILURE_CLASSES } from "./schema-CSf6qWgZ.js";
 import { r as observedSplitScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
-import { c as validateRunRecord } from "./run-record-ZIsR9Fif.js";
+import { c as validateRunRecord } from "./run-record-DQpSf7t-.js";
 //#region src/promotion-gate.ts
 /**
 * Bootstrap-CI promotion gate.
@@ -523,4 +523,4 @@ function fmt(x) {
 //#endregion
 export { judgeReplayGate as i, evaluateReleaseConfidence as n, bootstrapCi as r, assertReleaseConfidence as t };
 
-//# sourceMappingURL=release-confidence-BsGEg_xg.js.map
+//# sourceMappingURL=release-confidence-BcGCclTB.js.map