@tangle-network/agent-eval 0.19.0 → 0.19.1

package/dist/index.js

@@ -417,7 +417,7 @@ function confidenceInterval(scores2, confidence = 0.95) {
   if (scores2.length === 0) return { mean: 0, lower: 0, upper: 0 };
   if (scores2.length === 1) return { mean: scores2[0], lower: scores2[0], upper: scores2[0] };
   const n = scores2.length;
-  const mean9 = scores2.reduce((a, b) => a + b, 0) / n;
+  const mean10 = scores2.reduce((a, b) => a + b, 0) / n;
   const B = 1e3;
   const bootstrapMeans = [];
   for (let i = 0; i < B; i++) {
@@ -432,7 +432,7 @@ function confidenceInterval(scores2, confidence = 0.95) {
   const lowerIdx = Math.floor(alpha / 2 * B);
   const upperIdx = Math.floor((1 - alpha / 2) * B) - 1;
   return {
-    mean: mean9,
+    mean: mean10,
     lower: bootstrapMeans[lowerIdx],
     upper: bootstrapMeans[Math.min(upperIdx, B - 1)]
   };
@@ -520,11 +520,11 @@ function pairedTTest(before, after) {
   const n = before.length;
   if (n < 2) return { t: 0, df: 0, p: 1 };
   const diffs = before.map((b, i) => after[i] - b);
-  const mean9 = diffs.reduce((a, b) => a + b, 0) / n;
-  const variance2 = diffs.reduce((acc, d) => acc + (d - mean9) ** 2, 0) / (n - 1);
+  const mean10 = diffs.reduce((a, b) => a + b, 0) / n;
+  const variance2 = diffs.reduce((acc, d) => acc + (d - mean10) ** 2, 0) / (n - 1);
   const se = Math.sqrt(variance2 / n);
-  if (se === 0) return { t: mean9 === 0 ? 0 : Infinity, df: n - 1, p: mean9 === 0 ? 1 : 0 };
-  const t = mean9 / se;
+  if (se === 0) return { t: mean10 === 0 ? 0 : Infinity, df: n - 1, p: mean10 === 0 ? 1 : 0 };
+  const t = mean10 / se;
   const df = n - 1;
   const p = 2 * (1 - studentTCdf(Math.abs(t), df));
   return { t, df, p };
@@ -548,9 +548,9 @@ function wilcoxonSignedRank(before, after) {
   }
   let wPlus = 0;
   for (let k = 0; k < n; k++) if (diffs[k] > 0) wPlus += ranks3[k];
-  const mean9 = n * (n + 1) / 4;
+  const mean10 = n * (n + 1) / 4;
   const variance2 = n * (n + 1) * (2 * n + 1) / 24;
-  const z = (wPlus - mean9) / Math.sqrt(variance2);
+  const z = (wPlus - mean10) / Math.sqrt(variance2);
   const p = 2 * (1 - normalCdf(Math.abs(z)));
   return { w: wPlus, p };
 }
@@ -6457,10 +6457,10 @@ function analyzeSeries(values, options = {}) {
     return { state: "insufficient-data", windowMean: 0, windowCv: 0, tailRun: 0, stable: false };
   }
   const tail = values.slice(-window);
-  const mean9 = tail.reduce((a, b) => a + b, 0) / tail.length;
-  const variance2 = tail.reduce((acc, v) => acc + (v - mean9) ** 2, 0) / tail.length;
+  const mean10 = tail.reduce((a, b) => a + b, 0) / tail.length;
+  const variance2 = tail.reduce((acc, v) => acc + (v - mean10) ** 2, 0) / tail.length;
   const stdDev = Math.sqrt(variance2);
-  const refMean = Math.abs(mean9) > 1e-9 ? Math.abs(mean9) : 1;
+  const refMean = Math.abs(mean10) > 1e-9 ? Math.abs(mean10) : 1;
   const cv = stdDev / refMean;
   const stable = tail.length >= window && cv <= stableCv;
   let tailRun = 0;
@@ -6481,7 +6481,7 @@ function analyzeSeries(values, options = {}) {
   } else {
     state = "noisy";
   }
-  return { state, windowMean: mean9, windowCv: cv, tailRun, stable };
+  return { state, windowMean: mean10, windowCv: cv, tailRun, stable };
 }
 
 // src/state-continuity.ts
@@ -6938,9 +6938,9 @@ function benjaminiHochberg(pValues, fdr = 0.05) {
   for (let k = n - 1; k >= 0; k--) {
     const rank = k + 1;
     const raw = indexed[k].p * n / rank;
-    const bounded = Math.min(minRight, raw);
-    minRight = bounded;
-    q[indexed[k].i] = Math.min(1, bounded);
+    const bounded2 = Math.min(minRight, raw);
+    minRight = bounded2;
+    q[indexed[k].i] = Math.min(1, bounded2);
   }
   const significant = q.map((v) => v < fdr);
   return { qValues: q, significant };
@@ -7470,12 +7470,12 @@ async function paraphraseRobustness(prompt, mutators, scoreFn, options = {}) {
     variantScores.push({ mutator: id, score, mutated });
     all.push(score);
   }
-  const mean9 = all.reduce((a, b) => a + b, 0) / all.length;
-  const variance2 = all.reduce((a, v) => a + (v - mean9) ** 2, 0) / all.length;
+  const mean10 = all.reduce((a, b) => a + b, 0) / all.length;
+  const variance2 = all.reduce((a, v) => a + (v - mean10) ** 2, 0) / all.length;
   const stdDev = Math.sqrt(variance2);
-  const ref = Math.abs(mean9) > 1e-9 ? Math.abs(mean9) : 1;
+  const ref = Math.abs(mean10) > 1e-9 ? Math.abs(mean10) : 1;
   const robustness = Math.max(0, 1 - stdDev / ref);
-  return { originalScore, variantScores, meanScore: mean9, stdDev, robustness };
+  return { originalScore, variantScores, meanScore: mean10, stdDev, robustness };
 }
 var lowercaseMutator = (p) => p.toLowerCase();
 var sentenceReorderMutator = (p, seed) => {
@@ -8396,8 +8396,8 @@ async function prmBestOfN(store, grader, runIds) {
   if (runIds.length === 0) throw new Error("prmBestOfN: at least 1 candidate required");
   const graded = await Promise.all(runIds.map((id) => grader.grade(store, id)));
   const ranked = [...graded].sort((a, b) => b.aggregateScore - a.aggregateScore);
-  const mean9 = graded.reduce((a, g) => a + g.aggregateScore, 0) / graded.length;
-  const variance2 = graded.reduce((a, g) => a + (g.aggregateScore - mean9) ** 2, 0) / graded.length;
+  const mean10 = graded.reduce((a, g) => a + g.aggregateScore, 0) / graded.length;
+  const variance2 = graded.reduce((a, g) => a + (g.aggregateScore - mean10) ** 2, 0) / graded.length;
   return { winner: ranked[0], ranked, stdDev: Math.sqrt(variance2) };
 }
 async function prmEnsembleBestOfN(store, graders, runIds) {
@@ -8419,8 +8419,8 @@ async function prmEnsembleBestOfN(store, graders, runIds) {
   const ranked = [...byRun.values()].sort(
     (a, b) => (bordaScores.get(b.runId) ?? 0) - (bordaScores.get(a.runId) ?? 0)
   );
-  const mean9 = ranked.reduce((a, g) => a + g.aggregateScore, 0) / ranked.length;
-  const variance2 = ranked.reduce((a, g) => a + (g.aggregateScore - mean9) ** 2, 0) / ranked.length;
+  const mean10 = ranked.reduce((a, g) => a + g.aggregateScore, 0) / ranked.length;
+  const variance2 = ranked.reduce((a, g) => a + (g.aggregateScore - mean10) ** 2, 0) / ranked.length;
   return { winner: ranked[0], ranked, stdDev: Math.sqrt(variance2) };
 }
 
@@ -8950,8 +8950,8 @@ async function proposeSynthesisTargets(dataset, traceStore, options = {}) {
     const sRuns = runs.filter((r) => r.scenarioId === s.id);
     const scores2 = sRuns.map((r) => r.outcome?.score).filter((x) => typeof x === "number");
     if (scores2.length < 3) continue;
-    const mean9 = scores2.reduce((a, b) => a + b, 0) / scores2.length;
-    const variance2 = scores2.reduce((a, b) => a + (b - mean9) ** 2, 0) / scores2.length;
+    const mean10 = scores2.reduce((a, b) => a + b, 0) / scores2.length;
+    const variance2 = scores2.reduce((a, b) => a + (b - mean10) ** 2, 0) / scores2.length;
     if (variance2 > varianceThreshold) {
       targets.push({
         reason: "high-variance",
@@ -12813,6 +12813,289 @@ function traceExcerpt(trace) {
   return void 0;
 }
 
+// src/release-confidence.ts
+var DEFAULT_THRESHOLDS = {
+  requireCorpus: true,
+  minScenarioCount: 1,
+  minSearchRuns: 1,
+  minHoldoutRuns: 1,
+  requireHoldout: true,
+  minPassRate: 0.8,
+  minMeanScore: 0.7,
+  maxOverfitGap: 0.15,
+  maxMeanCostUsd: Number.POSITIVE_INFINITY,
+  maxP95WallMs: Number.POSITIVE_INFINITY,
+  requireAsiForFailures: true,
+  failureScoreThreshold: 0.5
+};
+function releaseTraceEvidenceFromMultiShotTrials(trials) {
+  return trials.map((trial) => ({
+    scenarioId: trial.scenarioId,
+    candidateId: trial.variantId,
+    split: trial.split === "holdout" ? "holdout" : trial.split === "dev" ? "dev" : "search",
+    score: trial.score,
+    ok: trial.ok,
+    turnCount: Array.isArray(trial.trace?.turns) ? trial.trace.turns.length : void 0,
+    costUsd: trial.cost,
+    durationMs: trial.durationMs,
+    failureMode: trial.error ? "runtime_error" : void 0,
+    asi: trial.asi,
+    metadata: trial.metadata
+  }));
+}
+function evaluateReleaseConfidence(input) {
+  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds };
+  const candidateId = input.candidateId ?? null;
+  const runs = filterCandidate(input.runs ?? [], candidateId, input.baselineId);
+  const traces = filterTraceCandidate(input.traces ?? [], candidateId, input.baselineId);
+  const scenarios = input.scenarios ?? [];
+  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length;
+  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios);
+  const searchScores = scoresFor(runs, "search");
+  const holdoutScores = scoresFor(runs, "holdout");
+  const allScores = [...searchScores, ...holdoutScores];
+  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber);
+  const scoreUniverse = allScores.length > 0 ? allScores : traceScores;
+  const searchRuns = runs.filter((r) => r.splitTag === "search").length;
+  const holdoutRuns = runs.filter((r) => r.splitTag === "holdout").length;
+  const searchMeanScore = mean8(searchScores);
+  const holdoutMeanScore = mean8(holdoutScores);
+  const metrics = {
+    scenarioCount,
+    searchRuns,
+    holdoutRuns,
+    passRate: passRate(runs, traces, thresholds.failureScoreThreshold),
+    meanScore: mean8(scoreUniverse),
+    searchMeanScore,
+    holdoutMeanScore,
+    overfitGap: safeDiff2(searchMeanScore, holdoutMeanScore),
+    meanCostUsd: mean8([...runs.map((r) => r.costUsd), ...traces.map((t) => t.costUsd).filter(isFiniteNumber)]),
+    p95WallMs: percentile([...runs.map((r) => r.wallMs), ...traces.map((t) => t.durationMs).filter(isFiniteNumber)], 0.95),
+    failedRows: failedRows(runs, traces, thresholds.failureScoreThreshold).length,
+    failuresWithAsi: failedRows(runs, traces, thresholds.failureScoreThreshold).filter((row) => row.hasAsi).length,
+    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,
+    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,
+    splitCounts,
+    domainCounts: countDomains(scenarios),
+    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),
+    responsibleSurfaceCounts: countResponsibleSurfaces(traces)
+  };
+  const issues = [];
+  checkCorpus(input, thresholds, metrics, issues);
+  checkQuality(thresholds, metrics, issues);
+  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues);
+  checkDiagnostics(thresholds, metrics, issues);
+  checkEfficiency(thresholds, metrics, issues);
+  const axes = buildAxes(metrics, thresholds, input.gateDecision ?? null, issues);
+  const status = issues.some((i) => i.severity === "critical") ? "fail" : issues.length > 0 ? "warn" : "pass";
+  return {
+    target: input.target,
+    candidateId,
+    baselineId: input.baselineId ?? null,
+    status,
+    promote: status === "pass" && (input.gateDecision ? input.gateDecision.promote : true),
+    axes,
+    issues,
+    metrics,
+    dataset: input.dataset ?? null,
+    gateDecision: input.gateDecision ?? null,
+    summary: renderSummary(input.target, status, metrics, issues)
+  };
+}
+function assertReleaseConfidence(input) {
+  const scorecard = evaluateReleaseConfidence(input);
+  if (scorecard.status === "fail") {
+    throw new Error(scorecard.summary);
+  }
+  return scorecard;
+}
+function filterCandidate(runs, candidateId, baselineId) {
+  if (candidateId) return runs.filter((r) => r.candidateId === candidateId);
+  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId);
+  return [...runs];
+}
+function filterTraceCandidate(traces, candidateId, baselineId) {
+  if (candidateId) return traces.filter((t) => t.candidateId === void 0 || t.candidateId === candidateId);
+  if (baselineId) return traces.filter((t) => t.candidateId === void 0 || t.candidateId !== baselineId);
+  return [...traces];
+}
+function checkCorpus(input, thresholds, metrics, issues) {
+  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {
+    issues.push({ axis: "corpus", severity: "critical", code: "missing_corpus", detail: "No Dataset manifest or scenarios supplied." });
+  }
+  if (metrics.scenarioCount < thresholds.minScenarioCount) {
+    issues.push({ axis: "corpus", severity: "critical", code: "few_scenarios", detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.` });
+  }
+  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {
+    issues.push({ axis: "corpus", severity: "critical", code: "missing_holdout_split", detail: "Corpus has no holdout scenarios." });
+  }
+}
+function checkQuality(thresholds, metrics, issues) {
+  if (metrics.searchRuns < thresholds.minSearchRuns) {
+    issues.push({ axis: "quality", severity: "critical", code: "few_search_runs", detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.` });
+  }
+  if (metrics.passRate < thresholds.minPassRate) {
+    issues.push({ axis: "quality", severity: "critical", code: "low_pass_rate", detail: `passRate ${fmt3(metrics.passRate)} < ${fmt3(thresholds.minPassRate)}.` });
+  }
+  if (metrics.meanScore < thresholds.minMeanScore) {
+    issues.push({ axis: "quality", severity: "critical", code: "low_mean_score", detail: `meanScore ${fmt3(metrics.meanScore)} < ${fmt3(thresholds.minMeanScore)}.` });
+  }
+}
+function checkGeneralization(gateDecision, thresholds, metrics, issues) {
+  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {
+    issues.push({ axis: "generalization", severity: "critical", code: "few_holdout_runs", detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.` });
+  }
+  if (Number.isFinite(metrics.overfitGap) && metrics.overfitGap > thresholds.maxOverfitGap) {
+    issues.push({ axis: "generalization", severity: "critical", code: "overfit_gap", detail: `search-holdout gap ${fmt3(metrics.overfitGap)} > ${fmt3(thresholds.maxOverfitGap)}.` });
+  }
+  if (gateDecision && !gateDecision.promote) {
+    issues.push({ axis: "generalization", severity: "critical", code: `gate_${gateDecision.rejectionCode ?? "reject"}`, detail: gateDecision.reason });
+  }
+}
+function checkDiagnostics(thresholds, metrics, issues) {
+  if (!thresholds.requireAsiForFailures) return;
+  if (metrics.failedRows > metrics.failuresWithAsi) {
+    issues.push({
+      axis: "diagnostics",
+      severity: "critical",
+      code: "missing_failure_asi",
+      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`
+    });
+  }
+}
+function checkEfficiency(thresholds, metrics, issues) {
+  if (metrics.meanCostUsd > thresholds.maxMeanCostUsd) {
+    issues.push({ axis: "efficiency", severity: "critical", code: "cost_budget", detail: `meanCostUsd ${fmt3(metrics.meanCostUsd)} > ${fmt3(thresholds.maxMeanCostUsd)}.` });
+  }
+  if (metrics.p95WallMs > thresholds.maxP95WallMs) {
+    issues.push({ axis: "efficiency", severity: "critical", code: "latency_budget", detail: `p95WallMs ${fmt3(metrics.p95WallMs)} > ${fmt3(thresholds.maxP95WallMs)}.` });
+  }
+}
+function buildAxes(metrics, thresholds, gateDecision, issues) {
+  return [
+    axis("corpus", issues, bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)), `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`),
+    axis("quality", issues, Math.min(metrics.passRate, metrics.meanScore), `passRate=${fmt3(metrics.passRate)} meanScore=${fmt3(metrics.meanScore)}`),
+    axis("generalization", issues, gateDecision && !gateDecision.promote ? 0 : gapScore(metrics.overfitGap, thresholds.maxOverfitGap), `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt3(metrics.overfitGap)}`),
+    axis("diagnostics", issues, metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows, `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`),
+    axis("efficiency", issues, efficiencyScore(metrics, thresholds), `meanCostUsd=${fmt3(metrics.meanCostUsd)} p95WallMs=${fmt3(metrics.p95WallMs)}`)
+  ];
+}
+function axis(name, issues, score, detail) {
+  const own = issues.filter((i) => i.axis === name);
+  const status = own.some((i) => i.severity === "critical") ? "fail" : own.length > 0 ? "warn" : "pass";
+  return { name, status, score: bounded(score), detail };
+}
+function countScenarioSplits(scenarios) {
+  const counts = { train: 0, dev: 0, test: 0, holdout: 0 };
+  for (const scenario of scenarios) counts[scenario.split ?? "train"]++;
+  return counts;
+}
+function countDomains(scenarios) {
+  const out = {};
+  for (const scenario of scenarios) {
+    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? "uncategorized";
+    out[domain] = (out[domain] ?? 0) + 1;
+  }
+  return out;
+}
+function countFailureModes(runs, traces, threshold) {
+  const out = {};
+  for (const run of runs) {
+    const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
+    if (run.failureMode || score !== void 0 && score < threshold) {
+      const mode = run.failureMode ?? "low_score";
+      out[mode] = (out[mode] ?? 0) + 1;
+    }
+  }
+  for (const trace of traces) {
+    if (trace.failureMode || trace.ok === false || trace.score !== void 0 && trace.score < threshold) {
+      const mode = trace.failureMode ?? (trace.ok === false ? "not_ok" : "low_score");
+      out[mode] = (out[mode] ?? 0) + 1;
+    }
+  }
+  return out;
+}
+function countResponsibleSurfaces(traces) {
+  const out = {};
+  for (const trace of traces) {
+    for (const asi of trace.asi ?? []) {
+      const surface = asi.responsibleSurface ?? "unknown";
+      out[surface] = (out[surface] ?? 0) + 1;
+    }
+  }
+  return out;
+}
+function failedRows(runs, traces, threshold) {
+  const out = [];
+  for (const run of runs) {
+    const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
+    if (run.failureMode || score !== void 0 && score < threshold) {
+      const asiMetric = run.outcome.raw.asi;
+      out.push({ hasAsi: typeof asiMetric === "number" && asiMetric > 0 });
+    }
+  }
+  for (const trace of traces) {
+    if (trace.failureMode || trace.ok === false || trace.score !== void 0 && trace.score < threshold) {
+      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 });
+    }
+  }
+  return out;
+}
+function passRate(runs, traces, threshold) {
+  const outcomes = [
+    ...runs.map((run) => {
+      const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
+      return !run.failureMode && score !== void 0 && score >= threshold;
+    }),
+    ...traces.map((trace) => trace.ok !== false && (trace.score === void 0 || trace.score >= threshold))
+  ];
+  if (outcomes.length === 0) return 0;
+  return outcomes.filter(Boolean).length / outcomes.length;
+}
+function scoresFor(runs, split) {
+  return runs.filter((run) => run.splitTag === split).map((run) => split === "holdout" ? run.outcome.holdoutScore : run.outcome.searchScore).filter(isFiniteNumber);
+}
+function mean8(xs) {
+  if (xs.length === 0) return Number.NaN;
+  return xs.reduce((sum2, x) => sum2 + x, 0) / xs.length;
+}
+function percentile(xs, p) {
+  if (xs.length === 0) return Number.NaN;
+  const sorted = [...xs].sort((a, b) => a - b);
+  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))];
+}
+function isFiniteNumber(value) {
+  return typeof value === "number" && Number.isFinite(value);
+}
+function safeDiff2(a, b) {
+  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN;
+  return a - b;
+}
+function gapScore(gap, maxGap) {
+  if (!Number.isFinite(gap)) return 0;
+  if (maxGap <= 0) return gap <= 0 ? 1 : 0;
+  return bounded(1 - Math.max(0, gap) / maxGap);
+}
+function efficiencyScore(metrics, thresholds) {
+  const cost = Number.isFinite(thresholds.maxMeanCostUsd) && Number.isFinite(metrics.meanCostUsd) ? bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12)) : 1;
+  const latency = Number.isFinite(thresholds.maxP95WallMs) && Number.isFinite(metrics.p95WallMs) ? bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12)) : 1;
+  return Math.min(cost, latency);
+}
+function bounded(x) {
+  if (!Number.isFinite(x)) return 0;
+  return Math.max(0, Math.min(1, x));
+}
+function renderSummary(target, status, metrics, issues) {
+  const prefix = `release confidence ${status}: ${target}`;
+  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt3(metrics.passRate)} meanScore=${fmt3(metrics.meanScore)}`;
+  if (issues.length === 0) return `${prefix}; ${metricText}`;
+  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(",")}`;
+}
+function fmt3(x) {
+  if (!Number.isFinite(x)) return String(x);
+  return x.toFixed(4);
+}
+
 // src/jsonl-trial-cache.ts
 import { appendFileSync as appendFileSync4, existsSync as existsSync6, mkdirSync as mkdirSync4, readFileSync as readFileSync5 } from "fs";
 import { dirname as dirname4 } from "path";
@@ -13458,9 +13741,9 @@ function passOrthogonality(input) {
       sims.push(cosineSimilarity(vectors[i], vectors[j]));
     }
   }
-  const mean9 = sims.length === 0 ? 0 : sims.reduce((a, b) => a + b, 0) / sims.length;
+  const mean10 = sims.length === 0 ? 0 : sims.reduce((a, b) => a + b, 0) / sims.length;
   return {
-    orthogonality: Math.max(0, Math.min(1, 1 - mean9)),
+    orthogonality: Math.max(0, Math.min(1, 1 - mean10)),
     passCount: passes.length,
     similarities: sims
   };
@@ -13506,8 +13789,8 @@ function bootstrapCi(baseline, candidate, options = {}) {
   const iterations = options.iterations ?? 1e3;
   const minTotal = options.minTotalSamples ?? 6;
   const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate));
-  const baselineMean = mean8(baseline);
-  const candidateMean = mean8(candidate);
+  const baselineMean = mean9(baseline);
+  const candidateMean = mean9(candidate);
   const delta = candidateMean - baselineMean;
   if (baseline.length + candidate.length < minTotal || baseline.length === 0 || candidate.length === 0) {
     return {
@@ -13525,7 +13808,7 @@ function bootstrapCi(baseline, candidate, options = {}) {
   for (let i = 0; i < iterations; i++) {
     const bResample = resample(baseline, rng);
     const cResample = resample(candidate, rng);
-    deltas[i] = mean8(cResample) - mean8(bResample);
+    deltas[i] = mean9(cResample) - mean9(bResample);
   }
   deltas.sort((a, b) => a - b);
   const lowerIdx = Math.floor(alpha / 2 * iterations);
@@ -13548,7 +13831,7 @@ function bootstrapCi(baseline, candidate, options = {}) {
     verdict
   };
 }
-function mean8(xs) {
+function mean9(xs) {
   if (xs.length === 0) return 0;
   let s = 0;
   for (const x of xs) s += x;
@@ -13872,6 +14155,7 @@ export {
   analyzeAntiSlop,
   analyzeSeries,
   argHash,
+  assertReleaseConfidence,
   assignFeedbackSplit,
   attributeCounterfactuals,
   deterministicSplit as benchmarkDeterministicSplit,
@@ -13942,6 +14226,7 @@ export {
   evaluateContract,
   evaluateHypothesis,
   evaluateOracles,
+  evaluateReleaseConfidence,
   executeScenario,
   expectAgent,
   exportRewardModel,
@@ -14041,6 +14326,7 @@ export {
   regexMatch,
   regexMatches,
   regressionView,
+  releaseTraceEvidenceFromMultiShotTrials,
   renderMarkdown,
   renderMarkdownReport,
   renderPlaybookMarkdown,