@tangle-network/agent-eval 0.183.0 → 0.185.0

package/dist/{campaign-evidence-PYPUftRy.js → campaign-evidence-BsrCmBSE.js}

@@ -1,264 +1,13 @@
 import { s as ValidationError, t as AgentEvalError } from "./errors-Dngq5h35.js";
-import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-DPyQ_rpt.js";
-import { r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
-import { n as campaignCellExecutionEvidence, o as projectCampaignCellQuality, s as decidePairedPromotion } from "./run-record-DPa2PeHx.js";
-import { n as contentHash } from "./verdict-cache-B3eCVQtY.js";
-import { t as FileLedgerJournal } from "./journal-Cs9f7385.js";
+import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-BsZVgBhW.js";
+import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-CeWT_W7y.js";
+import { n as contentHash } from "./verdict-cache-BQEsVy9n.js";
+import { n as campaignCellExecutionEvidence, o as projectCampaignCellQuality } from "./run-record-CNNzWGpB.js";
+import { t as FileLedgerJournal } from "./journal-BuRLm6ho.js";
 import { g as readField } from "./ast-CP9ae9B0.js";
 import { createHash } from "node:crypto";
 import { z } from "zod";
 import { join } from "node:path";
-//#region src/campaign/gates/statistical-heldout.ts
-/**
-* Held-out inference pairs execution cells and judge identities before taking
-* means within registered independent units. Repetitions can improve a unit's
-* precision without increasing n. Ungrouped inference concerns independently
-* sampled execution cells conditional on a fixed scenario roster.
-*
-* The shared paired decision rule selects the estimator and statistical test.
-* Dimension reports retain missing coverage so required safety checks cannot
-* pass through absent evidence. Thresholds use the judge's native score scale.
-*/
-/** Tie fraction at/above which a gate annotates its verdict with the tie share.
-*  Tie-domination of the median bites structurally at >= 0.5 (the median is then
-*  0 by construction); 0.4 is a softer warn threshold that flags a run APPROACHING
-*  that regime, so an operator sees it before the median goes fully blind. */
-const TIE_WARN_FRACTION = .4;
-/** Campaign cell IDs append a numeric repetition after the scenario's full ID. */
-function scenarioIdFromCellId(cellId) {
-	const separator = cellId.lastIndexOf(":");
-	const repetition = cellId.slice(separator + 1);
-	if (separator < 1 || cellId.trim() !== cellId || !/^(0|[1-9]\d*)$/.test(repetition) || !Number.isSafeInteger(Number(repetition))) throw new Error(`pairHoldout: malformed cellId '${cellId}'; expected scenarioId:rep`);
-	return cellId.slice(0, separator);
-}
-/** Preserve cell pairing before taking equal-weight independent-unit means. */
-function aggregatePairedHoldout(paired, independentUnitByScenarioId) {
-	if (paired.before.length !== paired.after.length || paired.before.length !== paired.cellIds.length) throw new Error("aggregatePairedHoldout: scores and cellIds must have the same length");
-	if (new Set(paired.cellIds).size !== paired.cellIds.length) throw new Error("aggregatePairedHoldout: duplicate cellIds cannot count as new observations");
-	if (paired.before.some((value) => !Number.isFinite(value)) || paired.after.some((value) => !Number.isFinite(value))) throw new Error("aggregatePairedHoldout: paired scores must be finite");
-	if (independentUnitByScenarioId === void 0) return {
-		before: [...paired.before],
-		after: [...paired.after],
-		unitIds: [...paired.cellIds]
-	};
-	const scenarioIds = paired.cellIds.map(scenarioIdFromCellId);
-	const groups = /* @__PURE__ */ new Map();
-	for (let i = 0; i < paired.cellIds.length; i++) {
-		const scenarioId = scenarioIds[i];
-		const unitId = independentUnitByScenarioId.get(scenarioId);
-		if (typeof unitId !== "string" || unitId.length === 0 || unitId.trim() !== unitId) throw new Error(`aggregatePairedHoldout: missing independent unit for scenario '${scenarioId}'`);
-		const group = groups.get(unitId) ?? {
-			before: 0,
-			after: 0,
-			n: 0
-		};
-		group.before += paired.before[i];
-		group.after += paired.after[i];
-		group.n += 1;
-		groups.set(unitId, group);
-	}
-	const unitIds = [...groups.keys()].sort();
-	return {
-		before: unitIds.map((id) => groups.get(id).before / groups.get(id).n),
-		after: unitIds.map((id) => groups.get(id).after / groups.get(id).n),
-		unitIds
-	};
-}
-/**
-* Pair candidate vs baseline holdout observations by FULL cellId. `select`
-* pulls the scalar from a cell's judge reports (composite, or a named
-* dimension); a cell contributes the mean of `select` across its judges. Cells
-* whose scenario is not in `scenarioIds`, or where `select` is undefined for
-* every judge on both sides, are skipped. The selected judge IDs must agree
-* within each pair. Throws when the two maps disagree on holdout cell IDs — a
-* load-bearing invariant: the baseline + winner holdout campaigns run the same
-* scenarios with the same seed base, so their cellIds MUST align; a mismatch
-* means a silent pairing bug, not a soft fallback.
-*/
-function pairHoldout(candidate, baseline, scenarioIds, select) {
-	const cellValues = (byCell, cellId) => {
-		const scores = byCell.get(cellId);
-		const values = /* @__PURE__ */ new Map();
-		if (!scores) return values;
-		for (const [judgeId, s] of Object.entries(scores)) {
-			if (s.failed === true) throw new Error(`pairHoldout: cell '${cellId}' contains a failed judge score`);
-			const v = select(s);
-			if (typeof v === "number" && !Number.isFinite(v)) throw new Error(`pairHoldout: cell '${cellId}' contains a non-finite selected score`);
-			if (typeof v === "number") values.set(judgeId, v);
-		}
-		return values;
-	};
-	const inScope = (cellId) => scenarioIds.has(scenarioIdFromCellId(cellId));
-	const candCells = [...candidate.keys()].filter(inScope).sort();
-	const baseCells = [...baseline.keys()].filter(inScope).sort();
-	if (candCells.length !== baseCells.length || candCells.some((c, i) => c !== baseCells[i])) throw new Error(`pairHoldout: candidate/baseline holdout cells do not align — candidate=[${candCells.join(",")}] baseline=[${baseCells.join(",")}]. Both holdout campaigns must run the same scenarios with the same seed base.`);
-	const before = [];
-	const after = [];
-	const cellIds = [];
-	for (const cellId of candCells) {
-		const b = cellValues(baseline, cellId);
-		const a = cellValues(candidate, cellId);
-		if (b.size === 0 && a.size === 0) continue;
-		if (b.size === 0 || a.size === 0) throw new Error(`pairHoldout: cell '${cellId}' has a selected score on only one arm`);
-		if (b.size !== a.size || [...b.keys()].some((id) => !a.has(id))) throw new Error(`pairHoldout: cell '${cellId}' selected judge IDs do not align`);
-		const judgeIds = [...b.keys()].sort();
-		before.push(meanSelectedScores(judgeIds.map((id) => b.get(id))));
-		after.push(meanSelectedScores(judgeIds.map((id) => a.get(id))));
-		cellIds.push(cellId);
-	}
-	return {
-		before,
-		after,
-		cellIds
-	};
-}
-function meanSelectedScores(values) {
-	const first = values[0];
-	return values.every((value) => value === first) ? first : values.reduce((sum, value) => sum + value, 0) / values.length;
-}
-/**
-* Significance of the held-out composite lift: ship only when the lower bound
-* of the interval the outcome's shape ADMITS exceeds `deltaThreshold` (default
-* 0 ⇒ "confidently positive"). Interpret `deltaThreshold` in the judge's native
-* scale.
-*
-* The decision is delegated whole to {@link decidePairedPromotion}, the one
-* copy of the rule (`src/paired-promotion-decision.ts`), which `HeldOutGate`
-* also calls. That module's header carries the measurements; the short version
-* is three guards a bare `bootstrap.low > threshold` does not have:
-*
-*   - a two-point (pass/fail) outcome decides on Tango's SCORE interval, the
-*     only paired-binary construction that stays valid at a nonzero margin;
-*   - McNemar's exact test VETOES at any non-negative threshold;
-*   - a ZERO-WIDTH interval is refused rather than promoted, in either
-*     direction — [0,0] clears every negative threshold and [g,g] clears every
-*     threshold below g, and both are an absence of evidence, not a result.
-*
-* Measured on this function before those guards landed, at a nominal 5 %:
-* 14.60 % false promotion at n = 40 on a paired-binary noninferiority boundary,
-* and 88.50 % at n = 6 under a bounded asymmetric null whose true mean paired
-* delta is exactly 0.
-*
-* Continuous mean targets require bootstrap eligibility. Explicit median
-* targets can use the exact sign test at its confidence-dependent minimum.
-*/
-function heldoutSignificance(paired, opts = {}) {
-	const deltaThreshold = opts.deltaThreshold ?? 0;
-	const confidence = opts.confidence ?? .95;
-	const resamples = opts.resamples ?? 2e3;
-	const seed = opts.seed ?? 1337;
-	const statistic = opts.statistic ?? "mean";
-	const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId);
-	const decision = decidePairedPromotion(observations.before, observations.after, {
-		confidence,
-		resamples,
-		statistic,
-		seed,
-		threshold: deltaThreshold,
-		minPairs: opts.minProductiveRuns
-	});
-	const bootstrap = decision.bootstrap ?? pairedBootstrap(observations.before, observations.after, {
-		confidence,
-		resamples,
-		statistic,
-		seed
-	});
-	const medianBootstrap = statistic === "median" ? bootstrap : pairedBootstrap(observations.before, observations.after, {
-		confidence,
-		resamples,
-		statistic: "median",
-		seed
-	});
-	const n = observations.before.length;
-	let ties = 0;
-	for (let i = 0; i < n; i += 1) {
-		const after = observations.after[i];
-		const before = observations.before[i];
-		if (Math.abs(after - before) < 1e-9) ties += 1;
-	}
-	const tieFraction = n === 0 ? 0 : ties / n;
-	return {
-		paired,
-		bootstrap,
-		medianBootstrap,
-		decision,
-		decisionStatistic: decision.statistic,
-		mcnemar: decision.mcnemar,
-		tieFraction,
-		n,
-		pairedCellN: paired.cellIds.length,
-		observationUnit: opts.independentUnitByScenarioId === void 0 ? "cell" : "registered",
-		unitIds: observations.unitIds,
-		minimumRequired: decision.minimumPairs,
-		decisionMethod: decision.method,
-		pValue: decision.pValue,
-		significant: decision.promote,
-		fewRuns: !decision.sufficient
-	};
-}
-/** Detect the native scale of a set of scores: 0-100 when any magnitude clears
-*  1.5, else [0,1]. Used to auto-scale the regression tolerance so a default
-*  expressed for [0,1] is not silently a no-op on a 0-100 dimension. */
-function detectScale(values) {
-	return values.some((v) => Math.abs(v) > 1.5) ? 100 : 1;
-}
-/**
-* Report required-dimension evidence after full pairing and optional unit means.
-* A bootstrap floor breach or a shared paired test supporting a drop marks
-* regression. These two criteria are distinct; `ci` records the shared
-* estimator and `bootstrap` records the floor interval. Missing observations
-* and insufficient n remain explicit for the caller's evidence policy.
-* The default tolerance is 0.05 on [0,1] and 5 on a detected 0-100 scale.
-*/
-function dimensionRegressions(candidate, baseline, scenarioIds, criticalDimensions, opts = {}) {
-	const out = [];
-	const expectedCellIds = [...baseline.keys()].filter((cellId) => scenarioIds.has(scenarioIdFromCellId(cellId))).sort();
-	for (const dim of criticalDimensions) {
-		const paired = pairHoldout(candidate, baseline, scenarioIds, (s) => s.dimensions[dim]);
-		if (paired.before.length === 0) continue;
-		const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId);
-		const measuredCells = new Set(paired.cellIds);
-		const measuredScenarios = new Set(paired.cellIds.map(scenarioIdFromCellId));
-		const tolerance = opts.tolerance ?? .05 * detectScale([...paired.before, ...paired.after]);
-		const bootstrapStatistic = opts.statistic ?? "mean";
-		const shared = {
-			confidence: opts.confidence ?? .95,
-			resamples: opts.resamples ?? 2e3,
-			statistic: bootstrapStatistic,
-			seed: opts.seed ?? 1337,
-			minPairs: opts.minProductiveRuns
-		};
-		const guard = decidePairedPromotion(observations.before, observations.after, shared);
-		const regression = decidePairedPromotion(observations.after, observations.before, {
-			...shared,
-			threshold: tolerance
-		});
-		const bootstrap = guard.bootstrap ?? pairedBootstrap(observations.before, observations.after, shared);
-		out.push({
-			dimension: dim,
-			bootstrap,
-			bootstrapStatistic,
-			ci: {
-				low: guard.low,
-				high: guard.high
-			},
-			decisionStatistic: guard.statistic,
-			mcnemar: guard.mcnemar,
-			indeterminate: guard.indeterminate,
-			regressed: bootstrap.low < -tolerance || regression.promote,
-			tolerance,
-			n: observations.before.length,
-			pairedCellN: paired.cellIds.length,
-			observationUnit: opts.independentUnitByScenarioId === void 0 ? "cell" : "registered",
-			minimumRequired: guard.minimumPairs,
-			fewRuns: !guard.sufficient,
-			missingCellIds: expectedCellIds.filter((cellId) => !measuredCells.has(cellId)),
-			missingScenarioIds: [...scenarioIds].filter((id) => !measuredScenarios.has(id)).sort()
-		});
-	}
-	return out;
-}
-//#endregion
 //#region src/campaign/coverage.ts
 /** Reject campaign designs whose denominator cannot be identified exactly. */
 function assertCampaignDesign(scenarios, reps) {
@@ -2122,6 +1871,6 @@ function createCampaignEvidenceReceipt(input) {
 	});
 }
 //#endregion
-export { formatCoverageFailures as $, FinalEvidenceConflictError as A, surfaceDispatchRef as B, campaignMeanCompositeOrNull as C, parseReflectionResponse as D, buildReflectionPrompt as E, assertCodeSurfaceIdentity as F, summarizeBackendIntegrity as G, surfaceHashMatches as H, codeSurfaceIdentityMaterial as I, assertCompleteCampaign as J, assertCampaignDesign as K, componentSurfaceIdentityMaterial as L, openFinalEvidenceLedger as M, defineEvaluationClaim as N, recoverTruncatedJson as O, summarizeEvaluationUnits as P, campaignSplitDigestFromIdentities as Q, renderSurfaceDiff as R, campaignMeanComposite as S, pairedCampaignComposites as T, BackendIntegrityError as U, surfaceHash as V, assertRealBackend as W, campaignScenarioIdentity as X, campaignCoverage as Y, campaignSplitDigest as Z, provenanceRecordPath as _, createEvidenceReceipt as a, pairHoldout as at, assertFiniteRankKey as b, ATTESTATION_ALGORITHM as c, buildLoopProvenanceRecord as d, TIE_WARN_FRACTION as et, campaignMeasurementDigest as f, loopProvenanceSpans as g, loopProvenanceArgsFromResult as h, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS as i, heldoutSignificance as it, FinalEvidenceError as j, powerPreflight as k, attest as l, emitLoopProvenance as m, EVIDENCE_AUTHORITY_KINDS as n, detectScale as nt, isIndependentEvidence as o, canonicalDigest as p, assertCampaignSplitIdentity as q, EVIDENCE_RECEIPT_VERSION as r, dimensionRegressions as rt, verifyEvidenceReceipt as s, createCampaignEvidenceReceipt as t, aggregatePairedHoldout as tt, verifyAttestation as u, provenanceSpansPath as v, compareRankKeys as w, campaignBreakdown as x, verifyLoopProvenanceRecord as y, surfaceContentHash as z };
+export { formatCoverageFailures as $, FinalEvidenceConflictError as A, surfaceDispatchRef as B, campaignMeanCompositeOrNull as C, parseReflectionResponse as D, buildReflectionPrompt as E, assertCodeSurfaceIdentity as F, summarizeBackendIntegrity as G, surfaceHashMatches as H, codeSurfaceIdentityMaterial as I, assertCompleteCampaign as J, assertCampaignDesign as K, componentSurfaceIdentityMaterial as L, openFinalEvidenceLedger as M, defineEvaluationClaim as N, recoverTruncatedJson as O, summarizeEvaluationUnits as P, campaignSplitDigestFromIdentities as Q, renderSurfaceDiff as R, campaignMeanComposite as S, pairedCampaignComposites as T, BackendIntegrityError as U, surfaceHash as V, assertRealBackend as W, campaignScenarioIdentity as X, campaignCoverage as Y, campaignSplitDigest as Z, provenanceRecordPath as _, createEvidenceReceipt as a, assertFiniteRankKey as b, ATTESTATION_ALGORITHM as c, buildLoopProvenanceRecord as d, campaignMeasurementDigest as f, loopProvenanceSpans as g, loopProvenanceArgsFromResult as h, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS as i, FinalEvidenceError as j, powerPreflight as k, attest as l, emitLoopProvenance as m, EVIDENCE_AUTHORITY_KINDS as n, isIndependentEvidence as o, canonicalDigest as p, assertCampaignSplitIdentity as q, EVIDENCE_RECEIPT_VERSION as r, verifyEvidenceReceipt as s, createCampaignEvidenceReceipt as t, verifyAttestation as u, provenanceSpansPath as v, compareRankKeys as w, campaignBreakdown as x, verifyLoopProvenanceRecord as y, surfaceContentHash as z };
 
-//# sourceMappingURL=campaign-evidence-PYPUftRy.js.map
+//# sourceMappingURL=campaign-evidence-BsrCmBSE.js.map