@tangle-network/agent-eval 1.0.12 → 1.0.14

package/dist/index.js

@@ -3,43 +3,44 @@ import { i as JudgeError, n as CaptureIntegrityError, o as NotFoundError, r as C
 import { a as hashCanonical, r as canonicalString } from "./canonical-CicJlINA.js";
 import { a as agentProfileModelId, i as agentProfileId, n as HARNESS_NATIVE_MODEL, o as expandProfileAxes, r as agentProfileHash, s as harnessAxisOf, t as CODING_HARNESSES } from "./agent-profile-Dngk3LZ0.js";
 import { a as groupRunsByAgentProfileCell, c as verifyAgentProfileCell, i as buildAgentProfileCell, l as hashJson, o as toAgentProfileJson, r as agentProfileCellKey, s as validateAgentProfileCell, t as AGENT_PROFILE_KINDS, u as manifestContentDigest } from "./agent-profile-cell-CT5JYueM.js";
-import { r as observedSplitScore, t as isRealnessGated } from "./reward-P0VD_Aft.js";
+import { r as observedSplitScore, t as isRealnessGated } from "./reward-DqMfaEax.js";
 import { i as isToolSpan, n as isJudgeSpan, r as isLlmSpan, t as FAILURE_CLASSES } from "./schema-D-xu3MU_.js";
-import { a as parseRunRecordSafe, c as validateRunRecord, i as modelHasSnapshot, n as UNKNOWN_MODEL, o as runTaskScore, r as isRunRecord, t as RunRecordValidationError } from "./run-record-DEjx43cI.js";
-import { a as cliffsDelta, c as pairedCohensDz, i as summaryTable, n as gainHistogram, o as cohensD, s as interpretCliffs } from "./summary-report-B4BLQb88.js";
+import { a as parseRunRecordSafe, c as validateRunRecord, i as modelHasSnapshot, n as UNKNOWN_MODEL, o as runTaskScore, r as isRunRecord, t as RunRecordValidationError } from "./run-record-CpLS4Uid.js";
+import { a as cliffsDelta, c as pairedCohensDz, i as summaryTable, n as gainHistogram, o as cohensD, s as interpretCliffs } from "./summary-report-C3IJvp0i.js";
 import { t as interpolatedQuantile } from "./quantile-CncXP-Ss.js";
-import { a as wilcoxonSignedRank, i as mannWhitneyU, n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-C0Vd9bGG.js";
-import { c as medianInPlace, i as assertFiniteSample, s as makeRng, t as mean$1 } from "./mean-DDFhslBt.js";
+import { a as wilcoxonSignedRank, i as mannWhitneyU, n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-3-tsXvvC.js";
+import { a as assertFiniteSample, l as makeRng, t as mean$1, u as medianInPlace } from "./mean-D1d6gedE.js";
 import { t as mulberry32 } from "./random-DrzcgC7F.js";
-import { a as ranks, c as summarizeNumberSeries, i as pearsonR, l as weightedComposite, n as meanConfidenceInterval, o as spearmanR, r as partialCredit, s as studentTCritical, t as confidenceInterval, u as weightedMean } from "./descriptive-BTzLL-aj.js";
-import { n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "./judge-calibration-D93UCm1k.js";
-import { n as corpusInterRaterAgreementFromJudgeScores, r as interRaterReliability, t as corpusInterRaterAgreement } from "./agreement-irr-w-t3XUTD.js";
-import { a as welchsTTest, c as fisherExactTwoSided, d as studentTTwoSidedPValue, f as twoByTwoPValue, i as iqr, l as linearRegressionSlope, n as checkCanaries, o as chiSquared, s as chiSquaredPValue, t as analyzeRuns, u as normalTwoSidedPValue } from "./analyze-runs-dGYmYNtb.js";
-import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, o as passAtK, r as pairedRiskDifference, s as wilson, t as mcnemar } from "./paired-binary-B477lksS.js";
+import { a as ranks, c as summarizeNumberSeries, i as pearsonR, l as weightedComposite, n as meanConfidenceInterval, o as spearmanR, r as partialCredit, s as studentTCritical, t as confidenceInterval, u as weightedMean } from "./descriptive-B3Eb7tX0.js";
+import { n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "./judge-calibration-DKR6Xi-w.js";
+import { n as corpusInterRaterAgreementFromJudgeScores, r as interRaterReliability, t as corpusInterRaterAgreement } from "./agreement-irr-DXcXEt4I.js";
+import { i as rocAuc, n as diffFindings, r as createChatTraceEngine, t as FindingsStore } from "./findings-store-DWiBvJVM.js";
+import { a as welchsTTest, c as fisherExactTwoSided, d as studentTTwoSidedPValue, f as twoByTwoPValue, i as iqr, l as linearRegressionSlope, n as checkCanaries, o as chiSquared, s as chiSquaredPValue, t as analyzeRuns, u as normalTwoSidedPValue } from "./analyze-runs-C7vWBDhl.js";
 import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-DAd4xaDC.js";
-import { a as pairedSignTest, o as pairedTTest, r as pairedBootstrap, t as BOOTSTRAP_GATE_MIN_N } from "./paired-tests-kASEKpc-.js";
-import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "./power-and-mde-CRxepy8y.js";
+import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, o as passAtK, r as pairedRiskDifference, s as wilson, t as mcnemar } from "./paired-binary-CMzmSSd5.js";
+import { a as pairedSignTest, o as pairedTTest, r as pairedBootstrap, t as BOOTSTRAP_GATE_MIN_N } from "./paired-tests-Buh6feJ4.js";
+import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "./power-and-mde-g3vOn0XK.js";
 import { t as eProcess } from "./sequential-eprocess-C1YLwFaz.js";
-import { i as pairedDeltaTest, n as pairedDecisionShape, r as minimumPairsForPairedDeltaTest, t as decidePairedPromotion } from "./paired-promotion-decision-5Pl_qJLX.js";
+import { i as pairedDeltaTest, n as pairedDecisionShape, r as minimumPairsForPairedDeltaTest, t as decidePairedPromotion } from "./paired-promotion-decision-CiueaTTj.js";
 import { c as runsForScenario, i as judgeSpans, t as argHash } from "./query-CGXzO9d4.js";
 import { n as contentHash, r as fileVerdictCache, t as canonicalJson } from "./verdict-cache-BVH-fl18.js";
-import { B as DEFAULT_RED_TEAM_CORPUS, H as redTeamReport, U as scoreRedTeamOutput, V as redTeamDataset, W as runCanaries, Z as runCampaign, a as clamp01, i as aggregateRunScore, r as transientDispatchFailure, t as llmJudge } from "./llm-judge-CyDyqb6V.js";
+import { B as DEFAULT_RED_TEAM_CORPUS, H as redTeamReport, U as scoreRedTeamOutput, V as redTeamDataset, W as runCanaries, Z as runCampaign, a as clamp01, i as aggregateRunScore, r as transientDispatchFailure, t as llmJudge } from "./llm-judge-CHLvjbVR.js";
 import { a as resolveModelPricing, i as isModelPriced, n as estimateCost, r as estimateTokens, t as MODEL_PRICING } from "./metrics-DNCAUtuB.js";
 import { c as costForTokenPricing, i as CostCeilingReachedError, l as costForUsage, n as CostAccountingIncompleteError, t as CostLedger } from "./cost-ledger-CS8R3IZj.js";
-import { C as assertRealBackend, S as BackendIntegrityError, _ as parseReflectionResponse, a as ATTESTATION_ALGORITHM, g as buildReflectionPrompt, o as attest, s as verifyAttestation, w as summarizeBackendIntegrity } from "./campaign-evidence-BICB3a9Z.js";
+import { C as assertRealBackend, S as BackendIntegrityError, _ as parseReflectionResponse, a as ATTESTATION_ALGORITHM, g as buildReflectionPrompt, o as attest, s as verifyAttestation, w as summarizeBackendIntegrity } from "./campaign-evidence-CcA7a67m.js";
 import { s as surfaceContentHash } from "./json-recovery-DK4MDm5r.js";
 import { c as redactText, l as shareAllowed, n as REDACTION_VERSION, o as redact, r as assessShareSafety, s as redactForShare } from "./redact-zp_ILC4l.js";
-import { a as makeProposalFinding, i as makeFinding, r as computeFindingId } from "./types-BK0blWhe.js";
-import { S as requiredPairsForPairedPromotion, n as paretoFrontier, t as dominates, x as pairedPromotionPower } from "./pareto-DY_x3HHg.js";
+import { a as makeProposalFinding, i as makeFinding, r as computeFindingId } from "./types-CL1_N7AZ.js";
+import { S as requiredPairsForPairedPromotion, n as paretoFrontier, t as dominates, x as pairedPromotionPower } from "./pareto-B34N-vnn.js";
 import { a as LlmCallError, i as stripFencedJson, n as isRecord, o as LlmResponseError } from "./json-DtvZNAU9.js";
-import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-DyzJ0wOo.js";
+import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-CjF2nTA3.js";
 import { n as equivalenceVerdict, t as certificationEvidenceDigest } from "./verdict-C8TcdwfD.js";
-import { c as createBoundedTraceAnalysisStore, p as DEFAULT_TRACE_ANALYST_BUDGETS, t as createTraceAnalyst } from "./kind-factory-VKysp1ds.js";
-import { C as assertUniqueFindingIds, D as validateAnalystReviewDecisions, E as snapshotAnalystRun, S as analystRunDigest, T as readAnalystReview, c as FAILURE_MODE_KIND_SPEC, i as DEFAULT_TRACE_ANALYST_KINDS, n as AnalystRegistry, t as buildDefaultAnalystRegistry, w as completedAnalystReviewQuality, x as analystFindingDigest } from "./default-registry-gmDeN0KT.js";
+import { c as createBoundedTraceAnalysisStore, p as DEFAULT_TRACE_ANALYST_BUDGETS, t as createTraceAnalyst } from "./kind-factory-Cfz8DUCc.js";
+import { C as assertUniqueFindingIds, D as validateAnalystReviewDecisions, E as snapshotAnalystRun, S as analystRunDigest, T as readAnalystReview, c as FAILURE_MODE_KIND_SPEC, i as DEFAULT_TRACE_ANALYST_KINDS, n as AnalystRegistry, t as buildDefaultAnalystRegistry, w as completedAnalystReviewQuality, x as analystFindingDigest } from "./default-registry-eCkZFfuF.js";
 import { OUTPUT_VALUE } from "./trace-attributes.js";
 import { n as extractUsageFromSse, t as extractUsage } from "./extract-usage-ISFqhfDz.js";
 import { n as InMemoryRawProviderSink, r as NoopRawProviderSink, t as FileSystemRawProviderSink } from "./raw-provider-sink-RbaEv2vB.js";
-import { _ as exportRunAsOtlp, c as describeTraceInsightScope, g as analyzeTraces, h as tokenizeDomainWords, l as domainEvidencePattern, m as scoreTraceInsightReadiness, n as toolSpansToTraceAnalysisStore, o as buildTraceInsightContext, s as buildTraceInsightPrompt, u as inferDomainKeywords, v as captureFetchToRawSink } from "./store-tool-spans-GGaCijFy.js";
+import { _ as exportRunAsOtlp, c as describeTraceInsightScope, g as analyzeTraces, h as tokenizeDomainWords, l as domainEvidencePattern, m as scoreTraceInsightReadiness, n as toolSpansToTraceAnalysisStore, o as buildTraceInsightContext, s as buildTraceInsightPrompt, u as inferDomainKeywords, v as captureFetchToRawSink } from "./store-tool-spans-BrSWHB9O.js";
 import { t as TraceEmitter } from "./emitter-DhQ65mza.js";
 import { n as assertRunCaptured, t as RunIntegrityError } from "./integrity-DOHt1mwU.js";
 import { n as InMemoryTraceStore, t as FileSystemTraceStore } from "./store-ZykIyQmy.js";
@@ -47,112 +48,77 @@ import { t as packageVersion$1 } from "./package-version-CJl_pY9W.js";
 import { t as buildTrajectory } from "./trajectory-O3GDHmev.js";
 import { i as failureBlame, n as classifyFailure, r as classifyFailureReason } from "./failure-taxonomy-9eQfXKoW.js";
 import { n as runBoundedProcess, t as DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS } from "./bounded-process-BZrWXIKW.js";
-import { t as runEvalCampaign } from "./eval-campaign-B-dQl_lA.js";
+import { t as runEvalCampaign } from "./eval-campaign-CZSCjxVb.js";
 import { _ as assertCrossFamily, a as maximumChargeForLlmRequest, d as assertServedModel, g as CrossFamilyError, h as describeServedModelCheck, i as costReceiptFromLlmError, l as ModelSubstitutionError, m as checkServedModel, p as assertServedModels, r as costReceiptFromLlm, s as isTransientLlmError, u as assertCrossFamilyServed, v as judgeFamily } from "./llm-client-BdSiZ4qz.js";
 import { n as createLlmCorrectnessChecker, r as verifyCompletion, t as completionVerdict } from "./completion-verifier-ClmJV_Yc.js";
-import "./rollout-BC4u-d_I.js";
-import { t as mintRolloutRows } from "./mint-D9ZysYtY.js";
+import "./rollout-CbOGxuHl.js";
+import { t as mintRolloutRows } from "./mint-BH_3Ak2x.js";
 import { n as pairedEvalueSequence, t as evaluateInterimReleaseConfidence } from "./sequential-Bjs5roiK.js";
 import { n as paidJsonChat } from "./chat-json-call-C473CwTt.js";
-import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-Xt3znknf.js";
-import { n as diffFindings, r as createChatTraceEngine, t as FindingsStore } from "./findings-store-Bp-0Guuv.js";
-import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "./tool-waste-CRJqOASv.js";
-import { _ as analyzeSeries, a as judgeGateDecision, c as snapshotFromJudgeAgreement, g as withJudgeRetry, i as assertJudgeMayGate, n as JudgeGateRefusedError, o as measureJudgeAgreement, r as addCalibrationVerdicts, s as registerCalibrationSet, t as DEFAULT_JUDGE_GATE_POLICY } from "./judge-gate-DMlX38B7.js";
-import { a as spanRecordsToTraceAnalysisStore, i as otlpTextToTraceAnalysisStore, r as OtlpFileTraceStore } from "./store-otlp-DmLm65XW.js";
-import { n as bootstrapCi, t as evaluateReleaseConfidence } from "./release-confidence--1iP-d3x.js";
+import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-_MEe2jIO.js";
+import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "./tool-waste-Czt3aWID.js";
+import { _ as analyzeSeries, a as judgeGateDecision, c as snapshotFromJudgeAgreement, g as withJudgeRetry, i as assertJudgeMayGate, n as JudgeGateRefusedError, o as measureJudgeAgreement, r as addCalibrationVerdicts, s as registerCalibrationSet, t as DEFAULT_JUDGE_GATE_POLICY } from "./judge-gate-CIlBrn6q.js";
+import { a as spanRecordsToTraceAnalysisStore, i as otlpTextToTraceAnalysisStore, r as OtlpFileTraceStore } from "./store-otlp-BfPRa2p8.js";
+import { n as bootstrapCi, t as evaluateReleaseConfidence } from "./release-confidence-nfQNh3eb.js";
 import { t as createChatClient } from "./chat-client-D0me8iwq.js";
 import { createHash } from "node:crypto";
 import { accessSync, appendFileSync, constants, cpSync, existsSync, mkdirSync, promises, readFileSync, readdirSync, statSync, writeFileSync } from "node:fs";
 import { basename, delimiter, dirname, extname, isAbsolute, join, relative, resolve } from "node:path";
 import { spawnSync } from "node:child_process";
 import { ATTR, INPUT_TOKEN_ATTR_KEYS, MODEL_ATTR_KEYS, OUTPUT_TOKEN_ATTR_KEYS, SPAN_KINDS, TOOL_NAME_ATTR_KEYS, firstNumberAttr, firstStringAttr, resolveSpanKind } from "@tangle-network/agent-trace-contract";
-// src/statistics/comparison-effects.ts
-/** Independent means, oriented b − a: Welch 95% interval and Hedges' g with a normal 95% interval. */
-function meanEffect(a, b) {
-	const n = [a.length, b.length];
-	const comparison = welchsTTest(a, b);
-	const diff = a.length && b.length ? comparison.delta : null;
-	if (comparison.status !== "ok") return {
-		n,
-		diff,
-		ci: comparison.status === "zero-variance" ? [comparison.delta, comparison.delta] : null,
-		g: null,
-		gCi: null
-	};
-	const g = (1 - 3 / (4 * (a.length + b.length) - 9)) * comparison.cohensD;
-	const gSe = Math.sqrt((a.length + b.length) / (a.length * b.length) + g ** 2 / (2 * (a.length + b.length)));
-	return {
-		n,
-		diff,
-		ci: comparison.ci95,
-		g,
-		gCi: [g - 1.96 * gSe, g + 1.96 * gSe]
-	};
+// src/statistics/distribution-summary.ts
+/** Nearest-rank percentile on a sorted, non-empty sample (0-100 scale). */
+function nearestRank(sorted, percentile) {
+	return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(percentile / 100 * sorted.length) - 1))];
 }
-function binary(values, name) {
-	if (values.some((value) => value !== 0 && value !== 1)) throw new ValidationError(`${name}: expected binary 0/1 values`);
+/**
+* Nearest-rank percentile on an arbitrary sample, `percentile` on a 0-100
+* scale — the same rule `summarizeDistribution` applies, exposed for a single
+* percentile the summary struct does not carry (e.g. p99.9). Returns NaN for
+* an empty sample; callers that need a numeric floor guard it.
+*/
+function percentileNearestRank(values, percentile) {
+	if (!Array.isArray(values)) throw new ValidationError("percentileNearestRank: values must be an array");
+	if (values.length === 0) return NaN;
+	return nearestRank([...values].sort((a, b) => a - b), percentile);
 }
-/** Independent binary samples, oriented b − a, with Newcombe's hybrid Wilson 95% interval. */
-function rateEffect(a, b) {
-	binary(a, "rateEffect.a");
-	binary(b, "rateEffect.b");
-	const n = [a.length, b.length];
-	if (!a.length || !b.length) return {
-		n,
-		diff: null,
-		ci: null,
-		g: null,
-		gCi: null
-	};
-	const s1 = a.reduce((sum, value) => sum + value, 0), s2 = b.reduce((sum, value) => sum + value, 0);
-	const { lower: l1, upper: u1 } = wilson(s1, a.length), { lower: l2, upper: u2 } = wilson(s2, b.length);
-	const p1 = s1 / a.length, p2 = s2 / b.length, diff = p2 - p1;
-	return {
+function summarizeDistribution(values, options = {}) {
+	if (values.some((value) => !Number.isFinite(value))) throw new ValidationError("summarizeDistribution: samples must be finite");
+	const finite = options.lowerBound === void 0 ? values : values.filter((value) => value >= options.lowerBound);
+	if (finite.length === 0) return {
+		min: NaN,
+		p50: NaN,
+		p90: NaN,
+		p95: NaN,
+		p99: NaN,
+		max: NaN,
+		mean: NaN,
+		n: 0,
+		values: []
+	};
+	const sorted = [...finite].sort((a, b) => a - b);
+	const n = sorted.length;
+	const mean = sorted.reduce((sum, value) => sum + value / n, 0);
+	const sumSquares = sorted.map((value) => value - mean).reduce((sum, value) => sum + value ** 2, 0);
+	const summary = {
+		min: sorted[0],
+		p50: nearestRank(sorted, 50),
+		p90: nearestRank(sorted, 90),
+		p95: nearestRank(sorted, 95),
+		p99: nearestRank(sorted, 99),
+		max: sorted[n - 1],
+		mean,
 		n,
-		rates: [p1, p2],
-		diff,
-		ci: [diff - Math.sqrt((p2 - l2) ** 2 + (u1 - p1) ** 2), diff + Math.sqrt((u2 - p2) ** 2 + (p1 - l1) ** 2)],
-		g: null,
-		gCi: null
-	};
-}
-/** Mann–Whitney ROC area: ties count half; no observed member of either class means unknown. */
-function rocAuc(scores, labels) {
-	assertFiniteSample("rocAuc", "scores", scores);
-	binary(labels, "rocAuc.labels");
-	if (scores.length !== labels.length) throw new ValidationError("rocAuc: scores and labels must have equal lengths");
-	const ranked = ranks(scores);
-	const positive = labels.reduce((sum, label) => sum + label, 0);
-	const sum = labels.reduce((total, label, index) => total + label * ranked[index], 0);
-	const negative = labels.length - positive;
-	return positive && negative ? (sum - positive * (positive + 1) / 2) / (positive * negative) : null;
-}
-/** L2 logistic fit by fixed-step gradient descent; the returned linear score preserves ROC ordering. */
-function fitLogistic(rows, labels, { lambda = 1, steps = 600, rate = .3 } = {}) {
-	binary(labels, "fitLogistic.labels");
-	const width = rows[0]?.length ?? 0;
-	if (!rows.length || rows.length !== labels.length || rows.some((row) => row.length !== width)) throw new ValidationError("fitLogistic: nonempty equal-width rows and one label per row required");
-	rows.forEach((row) => assertFiniteSample("fitLogistic", "row", row));
-	if (!(Number.isFinite(lambda) && lambda >= 0 && Number.isInteger(steps) && steps >= 0 && Number.isFinite(rate) && rate > 0)) throw new ValidationError("fitLogistic: invalid optimizer bounds");
-	const weights = new Array(width).fill(0);
-	let bias = 0;
-	const score = (row) => bias + row.reduce((sum, value, index) => sum + value * weights[index], 0);
-	for (let step = 0; step < steps; step++) {
-		const gradient = new Array(width).fill(0);
-		let gradientBias = 0;
-		rows.forEach((row, index) => {
-			const error = 1 / (1 + Math.exp(-score(row))) - labels[index];
-			gradientBias += error;
-			row.forEach((value, column) => {
-				gradient[column] = gradient[column] + error * value;
-			});
-		});
-		bias -= rate * gradientBias / rows.length;
-		weights.forEach((weight, column) => {
-			weights[column] = weight - rate * (gradient[column] / rows.length + lambda * weight / rows.length);
-		});
+		values: sorted
+	};
+	if (n >= 2) {
+		const stddev = Math.sqrt(sumSquares / (n - 1));
+		const margin = studentTCritical(n - 1, .95) * stddev / Math.sqrt(n);
+		summary.stddev = stddev;
+		summary.ci95Low = mean - margin;
+		summary.ci95High = mean + margin;
 	}
-	return score;
+	return summary;
 }
 // src/statistics/samples.ts
 /** Interpolated sample median; an empty population has no measured median. */
@@ -228,96 +194,6 @@ function summarizeSample(values, options = {}) {
 	].some((value) => value !== null && !Number.isFinite(value))) throw new ValidationError("summarizeSample: sample arithmetic exceeded finite range");
 	return summary;
 }
-// src/statistics/two-sample.ts
-/** Welch's t and Satterthwaite df for unequal variances; nulls when unestimable. */
-function welch(baseline, candidate) {
-	const sdA = baseline.stddev;
-	const sdB = candidate.stddev;
-	const meanA = baseline.mean;
-	const meanB = candidate.mean;
-	if (baseline.n < 2 || candidate.n < 2 || sdA === void 0 || sdB === void 0 || meanA === void 0 || meanB === void 0) return {
-		t: null,
-		df: null
-	};
-	const va = sdA ** 2 / baseline.n;
-	const vb = sdB ** 2 / candidate.n;
-	const denom = Math.sqrt(va + vb);
-	if (denom === 0) return {
-		t: null,
-		df: null
-	};
-	return {
-		t: (meanA - meanB) / denom,
-		df: (va + vb) ** 2 / (va ** 2 / (baseline.n - 1) + vb ** 2 / (candidate.n - 1))
-	};
-}
-/** Evidence that two independent windows differ; conservative on thin data. */
-function twoSampleEvidence(baseline, candidate, alpha = .05) {
-	const { t, df } = welch(baseline, candidate);
-	const pValue = t !== null && df !== null ? studentTTwoSidedPValue(t, df) : null;
-	const ciDisjoint = baseline.ci95High !== void 0 && baseline.ci95Low !== void 0 && candidate.ci95High !== void 0 && candidate.ci95Low !== void 0 && (baseline.ci95Low > candidate.ci95High || candidate.ci95Low > baseline.ci95High);
-	return {
-		t,
-		df,
-		pValue,
-		ciDisjoint,
-		significant: ciDisjoint || pValue !== null && pValue < alpha
-	};
-}
-// src/statistics/distribution-summary.ts
-/** Nearest-rank percentile on a sorted, non-empty sample (0-100 scale). */
-function nearestRank(sorted, percentile) {
-	return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(percentile / 100 * sorted.length) - 1))];
-}
-/**
-* Nearest-rank percentile on an arbitrary sample, `percentile` on a 0-100
-* scale — the same rule `summarizeDistribution` applies, exposed for a single
-* percentile the summary struct does not carry (e.g. p99.9). Returns NaN for
-* an empty sample; callers that need a numeric floor guard it.
-*/
-function percentileNearestRank(values, percentile) {
-	if (!Array.isArray(values)) throw new ValidationError("percentileNearestRank: values must be an array");
-	if (values.length === 0) return NaN;
-	return nearestRank([...values].sort((a, b) => a - b), percentile);
-}
-function summarizeDistribution(values, options = {}) {
-	if (values.some((value) => !Number.isFinite(value))) throw new ValidationError("summarizeDistribution: samples must be finite");
-	const finite = options.lowerBound === void 0 ? values : values.filter((value) => value >= options.lowerBound);
-	if (finite.length === 0) return {
-		min: NaN,
-		p50: NaN,
-		p90: NaN,
-		p95: NaN,
-		p99: NaN,
-		max: NaN,
-		mean: NaN,
-		n: 0,
-		values: []
-	};
-	const sorted = [...finite].sort((a, b) => a - b);
-	const n = sorted.length;
-	const mean = sorted.reduce((sum, value) => sum + value / n, 0);
-	const sumSquares = sorted.map((value) => value - mean).reduce((sum, value) => sum + value ** 2, 0);
-	const summary = {
-		min: sorted[0],
-		p50: nearestRank(sorted, 50),
-		p90: nearestRank(sorted, 90),
-		p95: nearestRank(sorted, 95),
-		p99: nearestRank(sorted, 99),
-		max: sorted[n - 1],
-		mean,
-		n,
-		values: sorted
-	};
-	if (n >= 2) {
-		const stddev = Math.sqrt(sumSquares / (n - 1));
-		const margin = studentTCritical(n - 1, .95) * stddev / Math.sqrt(n);
-		summary.stddev = stddev;
-		summary.ci95Low = mean - margin;
-		summary.ci95High = mean + margin;
-	}
-	return summary;
-}
 // src/profile/baselines.ts
 /**
 * Strong, generically-useful baseline ROLES — the top zone of an `AgentProfile`
@@ -9043,6 +8919,6 @@ function assertHasField(name, obj, field) {
 		actual: `keys: ${Object.keys(obj).join(", ")}`
 	};
 }
-export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, AgentEvalError, AnalystRegistry, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BudgetBreachError, BudgetGuard, CODING_HARNESSES, ConfigError, CostAccountingIncompleteError, CostCeilingReachedError, CostLedger, CostTracker, CrossFamilyError, DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, ERROR_COUNT_PATTERNS, EquivalenceProtocolError, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARNESS_NATIVE_MODEL, HeldOutGate, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, JudgeError, JudgeGateRefusedError, LlmCallError, LlmResponseError, MODEL_PRICING, ModelSubstitutionError, MultiLayerVerifier, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, OtlpFileTraceStore, PairwiseSteeringOptimizer, ProductClient, PromptRegistry, REDACTION_VERSION, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, TraceEmitter, UNKNOWN_MODEL, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, acquisitionPlansForKnowledgeGaps, addCalibrationVerdicts, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCapabilityHeadroom, assertContains, assertCrossFamily, assertCrossFamilyServed, assertEventCount, assertExists, assertHasField, assertHttpOk, assertJudgeMayGate, assertLatencyUnder, assertNoHiddenLeak, assertNotContains, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assertStatus, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, chiSquared, chiSquaredPValue, clamp01, classifyFailure, classifyFailureReason, cliffsDelta, cohensD, comparePairedArms, compileTraceContractSpec, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeServedModelCheck, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, expandProfileAxes, explainTraceContract, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, fisherExactTwoSided, fitLogistic, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, holm, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isJudgeSpan, isLlmSpan, isModelPriced, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, knowledgeReadinessTracePayload, leaderboard, linearRegressionSlope, lintTraceContractSpec, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, meanConfidenceInterval, meanEffect, measureJudgeAgreement, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, mulberry32, normalTwoSidedPValue, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedEvalueSequence, pairedMde, pairedPromotionPower, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, passAtK, pearsonR, percentileNearestRank, preflightModels, productBenchmarkRepoIdentity, profile_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, ranks, rateEffect, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redact, redactForShare, redactText, registerCalibrationSet, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredPairsForPairedPromotion, requiredSampleSize, resolveModelPricing, resolveSeat, rocAuc, runAgentControlLoop, runBoundedProcess, runCampaign, runCanaries, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runSemanticConceptJudge, runTaskScore, runsForScenario, sampleMedian, scoreKnowledgeReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, seatPresets, selfImprove, shareAllowed, snapshotFromJudgeAgreement, spanRecordsToTraceAnalysisStore, spearmanR, stripFencedJson, studentTCritical, studentTTwoSidedPValue, subjectiveEval, summarizeBackendIntegrity, summarizeDistribution, summarizeNumberSeries, summarizePreferenceMemory, summarizeSample, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, twoByTwoPValue, twoSampleEvidence, userQuestionsForKnowledgeGaps, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, welch, welchsTTest, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withJudgeRetry, wranglerDeployRunner };
+export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, AgentEvalError, AnalystRegistry, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BudgetBreachError, BudgetGuard, CODING_HARNESSES, ConfigError, CostAccountingIncompleteError, CostCeilingReachedError, CostLedger, CostTracker, CrossFamilyError, DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, ERROR_COUNT_PATTERNS, EquivalenceProtocolError, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARNESS_NATIVE_MODEL, HeldOutGate, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, JudgeError, JudgeGateRefusedError, LlmCallError, LlmResponseError, MODEL_PRICING, ModelSubstitutionError, MultiLayerVerifier, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, OtlpFileTraceStore, PairwiseSteeringOptimizer, ProductClient, PromptRegistry, REDACTION_VERSION, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, TraceEmitter, UNKNOWN_MODEL, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, acquisitionPlansForKnowledgeGaps, addCalibrationVerdicts, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCapabilityHeadroom, assertContains, assertCrossFamily, assertCrossFamilyServed, assertEventCount, assertExists, assertHasField, assertHttpOk, assertJudgeMayGate, assertLatencyUnder, assertNoHiddenLeak, assertNotContains, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assertStatus, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, chiSquared, chiSquaredPValue, clamp01, classifyFailure, classifyFailureReason, cliffsDelta, cohensD, comparePairedArms, compileTraceContractSpec, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeServedModelCheck, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, expandProfileAxes, explainTraceContract, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, fisherExactTwoSided, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, holm, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isJudgeSpan, isLlmSpan, isModelPriced, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, knowledgeReadinessTracePayload, leaderboard, linearRegressionSlope, lintTraceContractSpec, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, meanConfidenceInterval, measureJudgeAgreement, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, mulberry32, normalTwoSidedPValue, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedEvalueSequence, pairedMde, pairedPromotionPower, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, passAtK, pearsonR, percentileNearestRank, preflightModels, productBenchmarkRepoIdentity, profile_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, ranks, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redact, redactForShare, redactText, registerCalibrationSet, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredPairsForPairedPromotion, requiredSampleSize, resolveModelPricing, resolveSeat, rocAuc, runAgentControlLoop, runBoundedProcess, runCampaign, runCanaries, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runSemanticConceptJudge, runTaskScore, runsForScenario, sampleMedian, scoreKnowledgeReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, seatPresets, selfImprove, shareAllowed, snapshotFromJudgeAgreement, spanRecordsToTraceAnalysisStore, spearmanR, stripFencedJson, studentTCritical, studentTTwoSidedPValue, subjectiveEval, summarizeBackendIntegrity, summarizeDistribution, summarizeNumberSeries, summarizePreferenceMemory, summarizeSample, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, twoByTwoPValue, userQuestionsForKnowledgeGaps, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, welchsTTest, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withJudgeRetry, wranglerDeployRunner };
 
 //# sourceMappingURL=index.js.map