@tangle-network/agent-eval 0.120.2 → 0.120.4

package/dist/chunk-SQQED7ZH.js

@@ -0,0 +1,998 @@
+import {
+  summarizeBackendIntegrity
+} from "./chunk-SYV364BL.js";
+import {
+  paretoChart
+} from "./chunk-DPZAEKA6.js";
+import {
+  cohensD,
+  pairedBootstrap,
+  pairedMde,
+  pairedTTest,
+  pearsonR,
+  requiredSampleSize,
+  spearmanR
+} from "./chunk-PJQFMIOX.js";
+import {
+  llmSpans
+} from "./chunk-ZET2UAYW.js";
+import {
+  resolveRunCostProvenance
+} from "./chunk-S3UZOQ5Y.js";
+
+// src/contamination-guard.ts
+function checkCanaries(output, scenarios) {
+  const leaks = [];
+  for (const s of scenarios) {
+    if (!s.canary) continue;
+    if (output.includes(s.canary)) {
+      leaks.push({ scenarioId: s.id, canary: s.canary, evidence: excerpt(output, s.canary) });
+    }
+  }
+  return leaks;
+}
+function checkBehavioralCanary(output, scenario) {
+  const pattern = scenario.forbiddenPattern ?? scenario.canary;
+  if (!pattern) return null;
+  const hit = matchForbidden(output, pattern);
+  if (!hit) return null;
+  return {
+    scenarioId: scenario.id,
+    canary: pattern,
+    evidence: excerpt(output, hit)
+  };
+}
+function runBehavioralCanaries(cases) {
+  const leaks = [];
+  for (const c of cases) {
+    const leak = checkBehavioralCanary(c.output, c.scenario);
+    if (leak) leaks.push({ ...leak, runId: c.runId ?? leak.runId });
+  }
+  return leaks;
+}
+function matchForbidden(output, pattern) {
+  const re = tryParseRegex(pattern);
+  if (re) {
+    const m = output.match(re);
+    return m && m[0].length > 0 ? m[0] : null;
+  }
+  return output.includes(pattern) ? pattern : null;
+}
+function tryParseRegex(pattern) {
+  if (pattern.length < 2 || pattern[0] !== "/") return null;
+  const last = pattern.lastIndexOf("/");
+  if (last <= 0) return null;
+  const body = pattern.slice(1, last);
+  const flags = pattern.slice(last + 1);
+  if (!/^[gimsuy]*$/.test(flags)) return null;
+  try {
+    return new RegExp(body, flags);
+  } catch {
+    return null;
+  }
+}
+async function canaryLeakView(store, scenarios) {
+  const targets = scenarios.filter((s) => !!s.canary);
+  if (targets.length === 0) return [];
+  const spans = await llmSpans(store);
+  const leaks = [];
+  for (const span of spans) {
+    const output = span.output ?? "";
+    for (const s of targets) {
+      if (s.canary && output.includes(s.canary)) {
+        leaks.push({
+          scenarioId: s.id,
+          canary: s.canary,
+          runId: span.runId,
+          evidence: excerpt(output, s.canary)
+        });
+      }
+    }
+  }
+  return leaks;
+}
+var HoldoutAuditor = class {
+  scenarios;
+  accessLog = [];
+  constructor(scenarios) {
+    this.scenarios = scenarios;
+  }
+  /** Retrieve a holdout scenario for a declared purpose. Non-'evaluation' throws. */
+  get(scenarioId, purpose) {
+    if (purpose !== "evaluation" && purpose !== "debugging") {
+      throw new Error(
+        `HoldoutAuditor.get: purpose must be 'evaluation' or 'debugging', got ${purpose}`
+      );
+    }
+    const s = this.scenarios.find((x) => x.id === scenarioId);
+    if (!s) throw new Error(`holdout scenario "${scenarioId}" not found`);
+    this.accessLog.push({ scenarioId, purpose, at: Date.now() });
+    return s;
+  }
+  getAccessLog() {
+    return this.accessLog;
+  }
+};
+function excerpt(source, needle) {
+  const at = source.indexOf(needle);
+  if (at < 0) return "";
+  const start = Math.max(0, at - 30);
+  const end = Math.min(source.length, at + needle.length + 30);
+  return (start > 0 ? "\u2026" : "") + source.slice(start, end) + (end < source.length ? "\u2026" : "");
+}
+
+// src/contract/analyze-runs.ts
+function summarizeExecution(opts) {
+  const bins = opts.histogramBins ?? 12;
+  return {
+    execution: computeExecutionInsight(opts.runs, bins),
+    costProvenance: summarizeCostProvenance(opts.runs)
+  };
+}
+async function analyzeRuns(opts) {
+  const runs = opts.runs;
+  const bins = opts.histogramBins ?? 12;
+  const threshold = opts.decisionThreshold ?? 0.02;
+  const split = resolveSplit(runs, opts.split ?? "auto");
+  const compositeWithIds = runs.map((r) => ({ runId: r.runId, score: compositeOf(r, split) })).filter((p) => Number.isFinite(p.score));
+  const composite = distributionOf(
+    compositeWithIds.map((p) => p.score),
+    bins,
+    compositeWithIds
+  );
+  const perDimension = computePerDimension(runs, bins);
+  const { execution, costProvenance: provenance } = summarizeExecution({
+    runs,
+    histogramBins: bins
+  });
+  const knownCostRuns = runs.filter((run) => resolveRunCostProvenance(run).kind !== "uncaptured");
+  const costs = knownCostRuns.map((r) => r.costUsd).filter(Number.isFinite);
+  const costDist = distributionOf(costs, bins);
+  const pareto = paretoChart(knownCostRuns, { split });
+  const degraded = {};
+  if (provenance.uncaptured.n > 0) {
+    degraded.cost = diagnoseCostCoverage(runs, provenance);
+  } else if (costs.length === 0 || costs.every((c) => c === 0)) {
+    degraded.cost = runs.length > 0 && runs.every((run) => run.costProvenance !== void 0) ? `all ${runs.length} explicitly observed or estimated USD values are $0` : diagnoseZeroCost(runs);
+  }
+  if (pareto.points.length < 2) {
+    degraded.pareto = pareto.points.length === 0 ? "no candidates \u2014 Pareto unavailable" : "single candidate \u2014 Pareto is a single point, not a frontier";
+  }
+  const costQuality = {
+    cost: costDist,
+    pareto,
+    provenance,
+    ...degraded.cost || degraded.pareto ? { degraded } : {}
+  };
+  const judges = computeJudgeInsights(runs);
+  const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : void 0;
+  const lift = computeLift(runs, opts.baselineCandidateId, opts.candidateCandidateId, split);
+  const failureClusters = opts.analyst ? await computeFailureClusters(runs, opts.analyst, split) : void 0;
+  const failureModes = computeFailureModes(runs);
+  const contamination = opts.canaryScenarios ? computeContamination(runs, opts.canaryScenarios) : void 0;
+  const outcomeCorrelation = opts.outcomeSignal ? computeOutcomeCorrelation(runs, opts.outcomeSignal, split) : void 0;
+  const release = buildReleaseScorecard(composite, lift, contamination);
+  const priorPeriodComparison = opts.baselineRuns ? computePriorPeriodComparison(runs, opts.baselineRuns, split, opts.baselineLabel) : void 0;
+  const recommendations = buildRecommendations({
+    composite,
+    judges,
+    interRater,
+    lift,
+    failureClusters,
+    failureModes,
+    contamination,
+    outcomeCorrelation,
+    priorPeriodComparison,
+    threshold
+  });
+  return {
+    n: runs.length,
+    execution,
+    composite,
+    perDimension,
+    costQuality,
+    judges,
+    interRater,
+    lift,
+    failureClusters,
+    contamination,
+    outcomeCorrelation,
+    release,
+    ...failureModes ? { failureModes } : {},
+    ...priorPeriodComparison ? { priorPeriodComparison } : {},
+    recommendations
+  };
+}
+function computeExecutionInsight(runs, bins) {
+  const aggregateRows = runs.flatMap((run) => {
+    const usage = aggregateTokenUsage(run);
+    return usage ? [{ usage, costUsd: finiteRaw(run, "aggregate_cost_usd") }] : [];
+  });
+  const aggregateCosts = aggregateRows.flatMap(
+    (row) => row.costUsd !== void 0 ? [row.costUsd] : []
+  );
+  const modelCounts = /* @__PURE__ */ new Map();
+  let failureRuns = 0;
+  let reportedErrorEvents = 0;
+  let errorReportingRuns = 0;
+  let modelCallRuns = 0;
+  let modelCallEvents = 0;
+  let modelCallReportingRuns = 0;
+  for (const run of runs) {
+    modelCounts.set(run.model, (modelCounts.get(run.model) ?? 0) + 1);
+    const modelCalls = run.outcome.raw.llm_span_count;
+    if (Number.isFinite(modelCalls)) {
+      modelCallEvents += modelCalls;
+      modelCallReportingRuns += 1;
+    }
+    const usage = run.tokenUsage;
+    if ((modelCalls ?? 0) > 0 || usage.input > 0 || usage.output > 0 || (usage.cached ?? 0) > 0 || (usage.cacheWrite ?? 0) > 0) {
+      modelCallRuns += 1;
+    }
+    const errorEvents = run.outcome.raw.error_span_count;
+    if (Number.isFinite(errorEvents)) {
+      reportedErrorEvents += errorEvents;
+      errorReportingRuns += 1;
+    }
+    if (run.failureClass !== void 0 && run.failureClass !== "success" || run.failureMode !== void 0 || (errorEvents ?? 0) > 0) {
+      failureRuns += 1;
+    }
+  }
+  return {
+    durationMs: distributionOf(
+      runs.map((run) => run.wallMs),
+      bins
+    ),
+    queueMs: distributionOf(
+      runs.filter((run) => run.queueMs !== void 0).map((run) => run.queueMs),
+      bins
+    ),
+    tokenUsage: summarizeTokenUsage(
+      runs.map((run) => run.tokenUsage),
+      bins
+    ),
+    aggregateUsage: {
+      runs: aggregateRows.length,
+      tokenUsage: summarizeTokenUsage(
+        aggregateRows.map((row) => row.usage),
+        bins
+      ),
+      costUsd: distributionOf(aggregateCosts, bins),
+      totalCostUsd: aggregateCosts.reduce((total, value) => total + value, 0)
+    },
+    models: [...modelCounts.entries()].map(([model, count]) => ({ model, runs: count })).sort((left, right) => right.runs - left.runs || left.model.localeCompare(right.model)),
+    modelCalls: {
+      runs: modelCallRuns,
+      events: modelCallEvents,
+      reportingRuns: modelCallReportingRuns
+    },
+    failures: {
+      runs: failureRuns,
+      fraction: runs.length > 0 ? failureRuns / runs.length : 0,
+      reportedErrorEvents,
+      reportingRuns: errorReportingRuns
+    }
+  };
+}
+function summarizeTokenUsage(usages, bins) {
+  const reasoning = usages.flatMap(
+    (usage) => usage.reasoning !== void 0 ? [usage.reasoning] : []
+  );
+  const cached = usages.flatMap((usage) => usage.cached !== void 0 ? [usage.cached] : []);
+  const cacheWrite = usages.flatMap(
+    (usage) => usage.cacheWrite !== void 0 ? [usage.cacheWrite] : []
+  );
+  return {
+    input: distributionOf(
+      usages.map((usage) => usage.input),
+      bins
+    ),
+    output: distributionOf(
+      usages.map((usage) => usage.output),
+      bins
+    ),
+    reasoning: distributionOf(reasoning, bins),
+    cached: distributionOf(cached, bins),
+    cacheWrite: distributionOf(cacheWrite, bins),
+    totals: {
+      input: usages.reduce((total, usage) => total + usage.input, 0),
+      output: usages.reduce((total, usage) => total + usage.output, 0),
+      reasoning: reasoning.reduce((total, value) => total + value, 0),
+      cached: cached.reduce((total, value) => total + value, 0),
+      cacheWrite: cacheWrite.reduce((total, value) => total + value, 0)
+    }
+  };
+}
+function aggregateTokenUsage(run) {
+  const input = finiteRaw(run, "aggregate_prompt_tokens");
+  const output = finiteRaw(run, "aggregate_completion_tokens");
+  const reasoning = finiteRaw(run, "aggregate_reasoning_tokens");
+  const cached = finiteRaw(run, "aggregate_cached_tokens");
+  const cacheWrite = finiteRaw(run, "aggregate_cache_write_tokens");
+  if (input === void 0 && output === void 0 && reasoning === void 0 && cached === void 0 && cacheWrite === void 0)
+    return void 0;
+  return {
+    input: input ?? 0,
+    output: output ?? 0,
+    ...reasoning !== void 0 ? { reasoning } : {},
+    ...cached !== void 0 ? { cached } : {},
+    ...cacheWrite !== void 0 ? { cacheWrite } : {}
+  };
+}
+function finiteRaw(run, key) {
+  const value = run.outcome.raw[key];
+  return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
+}
+function summarizeCostProvenance(runs) {
+  const summary = {
+    observed: { n: 0, totalUsd: 0 },
+    estimated: { n: 0, totalUsd: 0 },
+    uncaptured: { n: 0 },
+    knownFraction: 0
+  };
+  for (const run of runs) {
+    const cost = resolveRunCostProvenance(run);
+    if (cost.kind === "uncaptured") {
+      summary.uncaptured.n += 1;
+    } else {
+      summary[cost.kind].n += 1;
+      summary[cost.kind].totalUsd += cost.usd;
+    }
+  }
+  const known = summary.observed.n + summary.estimated.n;
+  summary.knownFraction = runs.length > 0 ? known / runs.length : 0;
+  return summary;
+}
+function diagnoseCostCoverage(runs, provenance) {
+  const uncaptured = provenance.uncaptured.n;
+  const known = provenance.observed.n + provenance.estimated.n;
+  const explicitUncaptured = runs.some((run) => run.costProvenance?.kind === "uncaptured");
+  if (uncaptured === runs.length && !explicitUncaptured) return diagnoseZeroCost(runs);
+  if (uncaptured === runs.length) {
+    return `USD cost uncaptured for all ${runs.length} runs \u2014 no observed or estimated USD values; token and wall-time metrics remain available.`;
+  }
+  return `USD cost uncaptured for ${uncaptured}/${runs.length} runs; excluded those rows from cost statistics (${known}/${runs.length} retained: ${provenance.observed.n} observed, ${provenance.estimated.n} estimated).`;
+}
+function diagnoseZeroCost(runs) {
+  const integrity = summarizeBackendIntegrity(runs);
+  const { totalRecords, stubRecords, uncostedRecords } = integrity;
+  if (totalRecords > 0 && stubRecords === totalRecords) {
+    return `no costUsd values recorded \u2014 all ${totalRecords} records are stub-mode (zero token usage). The backend never reported real LLM activity, so cost cannot be computed; verify the backend actually ran before trusting this corpus.`;
+  }
+  if (uncostedRecords > 0) {
+    return `no costUsd values recorded \u2014 ${uncostedRecords}/${totalRecords} records have token usage but $0 cost (unpriced model). Check isModelPriced(model) for the run's model id and add it to FAMILY_PRICING.`;
+  }
+  if (stubRecords > 0) {
+    return `no costUsd values recorded \u2014 ${stubRecords}/${totalRecords} records are stub-mode (zero token usage); the remainder reported neither tokens nor cost. Cost axis carries no signal.`;
+  }
+  return "no costUsd values recorded \u2014 cost axis carries no signal";
+}
+function computeFailureModes(runs) {
+  const counts = /* @__PURE__ */ new Map();
+  for (const r of runs) {
+    const key = r.failureClass ?? r.failureMode;
+    if (key) counts.set(key, (counts.get(key) ?? 0) + 1);
+  }
+  if (counts.size === 0) return void 0;
+  const n = runs.length;
+  return [...counts.entries()].map(([mode, count]) => ({ mode, count, share: n > 0 ? count / n : 0 })).sort((a, b) => b.count - a.count || a.mode.localeCompare(b.mode));
+}
+function computePriorPeriodComparison(current, baseline, split, windowLabel) {
+  if (current.length === 0 || baseline.length === 0) return void 0;
+  const metrics = {};
+  const directions = {};
+  const compositeCurrent = current.map((r) => compositeOf(r, split)).filter(Number.isFinite);
+  const compositeBaseline = baseline.map((r) => compositeOf(r, split)).filter(Number.isFinite);
+  if (compositeCurrent.length > 0 && compositeBaseline.length > 0) {
+    metrics.composite = welchCompare(compositeBaseline, compositeCurrent);
+    directions.composite = "higher-is-better";
+  }
+  const costCurrent = knownCostValues(current);
+  const costBaseline = knownCostValues(baseline);
+  if (costCurrent.length > 0 && costBaseline.length > 0) {
+    metrics.cost = welchCompare(costBaseline, costCurrent);
+    directions.cost = "lower-is-better";
+  }
+  const durCurrent = current.map((r) => r.wallMs).filter(Number.isFinite);
+  const durBaseline = baseline.map((r) => r.wallMs).filter(Number.isFinite);
+  if (durCurrent.length > 0 && durBaseline.length > 0) {
+    metrics.duration = welchCompare(durBaseline, durCurrent);
+    directions.duration = "lower-is-better";
+  }
+  const tokCurrent = current.map((r) => (r.tokenUsage.input ?? 0) + (r.tokenUsage.output ?? 0)).filter(Number.isFinite);
+  const tokBaseline = baseline.map((r) => (r.tokenUsage.input ?? 0) + (r.tokenUsage.output ?? 0)).filter(Number.isFinite);
+  if (tokCurrent.length > 0 && tokBaseline.length > 0) {
+    metrics.tokenUsage = welchCompare(tokBaseline, tokCurrent);
+    directions.tokenUsage = "lower-is-better";
+  }
+  const dimsCurrent = collectPerDimension(current);
+  const dimsBaseline = collectPerDimension(baseline);
+  for (const dim of Object.keys(dimsCurrent)) {
+    const b = dimsBaseline[dim];
+    const c = dimsCurrent[dim];
+    if (!b || b.length === 0 || !c || c.length === 0) continue;
+    metrics[`dim.${dim}`] = welchCompare(b, c);
+    directions[`dim.${dim}`] = "higher-is-better";
+  }
+  const regressedMetrics = [];
+  const improvedMetrics = [];
+  for (const [name, delta] of Object.entries(metrics)) {
+    if (!delta.significant) continue;
+    const dir = directions[name] ?? "higher-is-better";
+    const better = dir === "higher-is-better" ? delta.delta > 0 : delta.delta < 0;
+    if (better) improvedMetrics.push(name);
+    else regressedMetrics.push(name);
+  }
+  return {
+    baselineN: baseline.length,
+    currentN: current.length,
+    ...windowLabel ? { windowLabel } : {},
+    metrics,
+    regressedMetrics,
+    improvedMetrics
+  };
+}
+function knownCostValues(runs) {
+  return runs.filter((run) => resolveRunCostProvenance(run).kind !== "uncaptured").map((run) => run.costUsd).filter(Number.isFinite);
+}
+function collectPerDimension(runs) {
+  const out = {};
+  for (const r of runs) {
+    const perDim = r.outcome.judgeScores?.perDimMean;
+    if (!perDim) continue;
+    for (const [dim, value] of Object.entries(perDim)) {
+      if (!Number.isFinite(value)) continue;
+      if (!out[dim]) out[dim] = [];
+      out[dim].push(value);
+    }
+  }
+  return out;
+}
+function welchCompare(baseline, current) {
+  const baselineMean = mean(baseline);
+  const currentMean = mean(current);
+  const baselineVar = sampleVariance(baseline, baselineMean);
+  const currentVar = sampleVariance(current, currentMean);
+  const baselineN = baseline.length;
+  const currentN = current.length;
+  const delta = currentMean - baselineMean;
+  const se = Math.sqrt(baselineVar / baselineN + currentVar / currentN);
+  const halfWidth = 1.96 * (se > 0 ? se : 0);
+  const ci95 = [delta - halfWidth, delta + halfWidth];
+  const t = se > 0 ? delta / se : 0;
+  const pValue = se > 0 ? 2 * (1 - standardNormalCdf(Math.abs(t))) : 1;
+  const pooledStddev = Math.sqrt(
+    ((baselineN - 1) * baselineVar + (currentN - 1) * currentVar) / Math.max(1, baselineN + currentN - 2)
+  );
+  const cohensD2 = pooledStddev > 0 ? delta / pooledStddev : 0;
+  const significant = pValue < 0.05 && Math.abs(cohensD2) >= 0.2;
+  return {
+    current: currentMean,
+    baseline: baselineMean,
+    delta,
+    ci95,
+    pValue,
+    cohensD: cohensD2,
+    baselineN,
+    currentN,
+    significant
+  };
+}
+function sampleVariance(xs, xsMean) {
+  if (xs.length < 2) return 0;
+  let s = 0;
+  for (const x of xs) s += (x - xsMean) ** 2;
+  return s / (xs.length - 1);
+}
+function standardNormalCdf(z) {
+  const a1 = 0.254829592;
+  const a2 = -0.284496736;
+  const a3 = 1.421413741;
+  const a4 = -1.453152027;
+  const a5 = 1.061405429;
+  const p = 0.3275911;
+  const sign = z < 0 ? -1 : 1;
+  const x = Math.abs(z) / Math.SQRT2;
+  const t = 1 / (1 + p * x);
+  const y = 1 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp(-x * x);
+  return 0.5 * (1 + sign * y);
+}
+function resolveSplit(runs, pref) {
+  if (pref !== "auto") return pref;
+  const hasHoldout = runs.some((r) => Number.isFinite(r.outcome.holdoutScore));
+  return hasHoldout ? "holdout" : "search";
+}
+function compositeOf(run, split) {
+  const primary = split === "holdout" ? run.outcome.holdoutScore : run.outcome.searchScore;
+  if (Number.isFinite(primary)) return primary;
+  const alt = split === "holdout" ? run.outcome.searchScore : run.outcome.holdoutScore;
+  return Number.isFinite(alt) ? alt : Number.NaN;
+}
+function distributionOf(values, bins, withIds) {
+  if (values.length === 0) {
+    return {
+      n: 0,
+      mean: 0,
+      p50: 0,
+      p95: 0,
+      stddev: 0,
+      min: 0,
+      max: 0,
+      histogram: []
+    };
+  }
+  const sorted = [...values].sort((a, b) => a - b);
+  const n = sorted.length;
+  const mean2 = sorted.reduce((s, v) => s + v, 0) / n;
+  const variance = sorted.reduce((s, v) => s + (v - mean2) ** 2, 0) / n;
+  const stddev = Math.sqrt(variance);
+  const tailRuns = withIds ? [...withIds].sort((a, b) => a.score - b.score).slice(0, Math.min(5, withIds.length)) : void 0;
+  return {
+    n,
+    mean: mean2,
+    p50: percentile(sorted, 0.5),
+    p95: percentile(sorted, 0.95),
+    stddev,
+    min: sorted[0],
+    max: sorted[n - 1],
+    histogram: histogram(sorted, bins),
+    ...tailRuns ? { tailRuns } : {}
+  };
+}
+function percentile(sorted, q) {
+  if (sorted.length === 0) return 0;
+  if (sorted.length === 1) return sorted[0];
+  const idx = (sorted.length - 1) * q;
+  const lo = Math.floor(idx);
+  const hi = Math.ceil(idx);
+  if (lo === hi) return sorted[lo];
+  const w = idx - lo;
+  return sorted[lo] * (1 - w) + sorted[hi] * w;
+}
+function histogram(sorted, bins) {
+  if (sorted.length === 0 || bins < 1) return [];
+  const min = sorted[0];
+  const max = sorted[sorted.length - 1];
+  if (min === max) return [{ lo: min, hi: max, count: sorted.length }];
+  const width = (max - min) / bins;
+  const out = [];
+  for (let i = 0; i < bins; i++) {
+    const lo = min + i * width;
+    const hi = i === bins - 1 ? max : lo + width;
+    out.push({ lo, hi, count: 0 });
+  }
+  for (const v of sorted) {
+    const idx = Math.min(bins - 1, Math.floor((v - min) / width));
+    out[idx].count++;
+  }
+  return out;
+}
+function computePerDimension(runs, bins) {
+  const byDim = /* @__PURE__ */ new Map();
+  for (const run of runs) {
+    const scores = run.outcome.judgeScores;
+    if (!scores) continue;
+    for (const [dim, value] of Object.entries(scores.perDimMean ?? {})) {
+      if (!Number.isFinite(value)) continue;
+      const arr = byDim.get(dim) ?? [];
+      arr.push(value);
+      byDim.set(dim, arr);
+    }
+  }
+  const out = {};
+  for (const [dim, values] of byDim) out[dim] = distributionOf(values, bins);
+  return out;
+}
+function computeJudgeInsights(runs) {
+  const out = {};
+  const byJudge = /* @__PURE__ */ new Map();
+  for (const run of runs) {
+    const scores = run.outcome.judgeScores;
+    if (!scores?.perJudge) continue;
+    for (const [judgeId, dims] of Object.entries(scores.perJudge)) {
+      const dimValues = Object.values(dims).filter(Number.isFinite);
+      if (dimValues.length === 0) continue;
+      const judgeMean = dimValues.reduce((s, v) => s + v, 0) / dimValues.length;
+      const arr = byJudge.get(judgeId) ?? [];
+      arr.push(judgeMean);
+      byJudge.set(judgeId, arr);
+    }
+  }
+  for (const [judgeId, values] of byJudge) {
+    out[judgeId] = {
+      n: values.length,
+      meanScore: values.reduce((s, v) => s + v, 0) / values.length
+    };
+  }
+  return out;
+}
+function computeInterRater(ratings) {
+  const byRun = /* @__PURE__ */ new Map();
+  for (const r of ratings) {
+    if (!Number.isFinite(r.score)) continue;
+    const list = byRun.get(r.runId) ?? [];
+    list.push({ rater: r.rater, score: r.score });
+    byRun.set(r.runId, list);
+  }
+  const raters = new Set(ratings.map((r) => r.rater));
+  const jointlyRated = [];
+  for (const [runId, ratersForRun] of byRun) {
+    const seen = new Set(ratersForRun.map((r) => r.rater));
+    let all = true;
+    for (const r of raters) if (!seen.has(r)) all = false;
+    if (all) jointlyRated.push(runId);
+  }
+  if (raters.size < 2 || jointlyRated.length === 0) return void 0;
+  const raterList = [...raters].sort();
+  const perPair = {};
+  for (let i = 0; i < raterList.length; i++) {
+    for (let j = i + 1; j < raterList.length; j++) {
+      const a = raterList[i];
+      const b = raterList[j];
+      const aScores = [];
+      const bScores = [];
+      for (const runId of jointlyRated) {
+        const ratersForRun = byRun.get(runId);
+        const sa = ratersForRun.find((r) => r.rater === a)?.score;
+        const sb = ratersForRun.find((r) => r.rater === b)?.score;
+        if (sa !== void 0 && sb !== void 0) {
+          aScores.push(sa);
+          bScores.push(sb);
+        }
+      }
+      perPair[`${a}::${b}`] = pearsonR(aScores, bScores);
+    }
+  }
+  const pairKappas = Object.values(perPair).filter((v) => Number.isFinite(v));
+  const kappa = pairKappas.length === 0 ? 0 : pairKappas.reduce((s, v) => s + v, 0) / pairKappas.length;
+  const disagreementCases = jointlyRated.map((runId) => {
+    const ratersForRun = byRun.get(runId);
+    const scores = ratersForRun.map((r) => r.score);
+    const range = Math.max(...scores) - Math.min(...scores);
+    return { runId, ratings: ratersForRun, range };
+  }).sort((a, b) => b.range - a.range).slice(0, 20);
+  return {
+    raters: raters.size,
+    jointlyRated: jointlyRated.length,
+    kappa,
+    perPair,
+    disagreementCases
+  };
+}
+function computeLift(runs, baselineId, candidateId, split) {
+  let bId = baselineId;
+  let cId = candidateId;
+  if (!bId || !cId) {
+    const ids = [...new Set(runs.map((r) => r.candidateId))];
+    if (ids.length !== 2) return void 0;
+    const [idA, idB] = ids;
+    const meanA = mean(runs.filter((r) => r.candidateId === idA).map((r) => compositeOf(r, split)));
+    const meanB = mean(runs.filter((r) => r.candidateId === idB).map((r) => compositeOf(r, split)));
+    bId = meanA <= meanB ? idA : idB;
+    cId = meanA <= meanB ? idB : idA;
+  }
+  const baseline = runs.filter((r) => r.candidateId === bId);
+  const candidate = runs.filter((r) => r.candidateId === cId);
+  if (baseline.length === 0 || candidate.length === 0) return void 0;
+  const baselineByKey = new Map(baseline.map((r) => [pairingKey(r), r]));
+  const pairedBaseline = [];
+  const pairedCandidate = [];
+  let usedKeyPairing = false;
+  for (const cand of candidate) {
+    const b = baselineByKey.get(pairingKey(cand));
+    if (b) {
+      const bC = compositeOf(b, split);
+      const cC = compositeOf(cand, split);
+      if (Number.isFinite(bC) && Number.isFinite(cC)) {
+        pairedBaseline.push(bC);
+        pairedCandidate.push(cC);
+        usedKeyPairing = true;
+      }
+    }
+  }
+  if (!usedKeyPairing) {
+    const n = Math.min(baseline.length, candidate.length);
+    for (let i = 0; i < n; i++) {
+      const bC = compositeOf(baseline[i], split);
+      const cC = compositeOf(candidate[i], split);
+      if (Number.isFinite(bC) && Number.isFinite(cC)) {
+        pairedBaseline.push(bC);
+        pairedCandidate.push(cC);
+      }
+    }
+  }
+  if (pairedBaseline.length === 0) return void 0;
+  const baselineMean = mean(pairedBaseline);
+  const candidateMean = mean(pairedCandidate);
+  const delta = candidateMean - baselineMean;
+  const bootstrap = pairedBootstrap(pairedBaseline, pairedCandidate, {
+    confidence: 0.95,
+    resamples: 2e3,
+    statistic: "mean"
+  });
+  const tTest = pairedTTest(pairedBaseline, pairedCandidate);
+  const d = cohensD(pairedBaseline, pairedCandidate);
+  const mde = pairedMde({ nPaired: pairedBaseline.length, power: 0.8, alpha: 0.05 });
+  const requiredN = requiredSampleSize({
+    effect: Math.max(Math.abs(delta), 1e-6),
+    power: 0.8,
+    alpha: 0.05
+  });
+  return {
+    baselineMean,
+    candidateMean,
+    delta,
+    ci95: [bootstrap.low, bootstrap.high],
+    pValue: tTest.p,
+    n: pairedBaseline.length,
+    cohensD: d,
+    mde,
+    requiredN
+  };
+}
+function pairingKey(r) {
+  return `${r.experimentId}::${r.seed}`;
+}
+function mean(arr) {
+  return arr.length === 0 ? 0 : arr.reduce((s, v) => s + v, 0) / arr.length;
+}
+async function computeFailureClusters(runs, analyst, split) {
+  const failed = runs.filter((r) => compositeOf(r, split) < 0.5 || r.failureMode !== void 0);
+  if (failed.length === 0) return { clusters: [], totalFailures: 0 };
+  const clusters = /* @__PURE__ */ new Map();
+  for (const run of failed) {
+    try {
+      const result = await analyst.run(run.runId, { runRecord: run });
+      for (const finding of result.findings) {
+        const key = finding.area || finding.analyst_id || "unclassified";
+        const c = clusters.get(key) ?? { exemplars: [], share: 0 };
+        if (c.exemplars.length < 5) c.exemplars.push(run.runId);
+        clusters.set(key, c);
+      }
+    } catch {
+      const c = clusters.get("analyst-error") ?? { exemplars: [], share: 0 };
+      if (c.exemplars.length < 5) c.exemplars.push(run.runId);
+      clusters.set("analyst-error", c);
+    }
+  }
+  const clusterList = [...clusters.entries()].map(([id, c]) => ({
+    id,
+    name: id,
+    share: c.exemplars.length / failed.length,
+    exemplars: c.exemplars
+  }));
+  clusterList.sort((a, b) => b.share - a.share);
+  return { clusters: clusterList, totalFailures: failed.length };
+}
+function computeContamination(runs, canaries) {
+  let leaks = 0;
+  const details = [];
+  for (const run of runs) {
+    const output = stringifyOutput(run);
+    if (!output) continue;
+    const leaksHere = checkCanaries(output, canaries);
+    for (const leak of leaksHere) {
+      leaks++;
+      details.push({ runId: run.runId, canary: leak.canary, matched: leak.evidence });
+    }
+  }
+  return { leaks, holdoutAuditPassed: leaks === 0, details };
+}
+function stringifyOutput(run) {
+  const metadata = run.metadata;
+  if (typeof metadata?.output === "string") return metadata.output;
+  if (typeof metadata?.text === "string") return metadata.text;
+  return void 0;
+}
+function computeOutcomeCorrelation(runs, outcome, split) {
+  const xs = [];
+  const ys = [];
+  for (const run of runs) {
+    const y = outcome.valueByRunId[run.runId];
+    if (y === void 0 || !Number.isFinite(y)) continue;
+    const x = compositeOf(run, split);
+    if (!Number.isFinite(x)) continue;
+    xs.push(x);
+    ys.push(y);
+  }
+  if (xs.length < 3) return void 0;
+  const p = pearsonR(xs, ys);
+  const s = spearmanR(xs, ys);
+  const meanX = mean(xs);
+  const meanY = mean(ys);
+  let num = 0;
+  let denom = 0;
+  for (let i = 0; i < xs.length; i++) {
+    num += (xs[i] - meanX) * (ys[i] - meanY);
+    denom += (xs[i] - meanX) ** 2;
+  }
+  const slope = denom === 0 ? 0 : num / denom;
+  const intercept = meanY - slope * meanX;
+  const ssTot = ys.reduce((a, y) => a + (y - meanY) ** 2, 0);
+  const ssRes = ys.reduce((a, y, i) => a + (y - (intercept + slope * xs[i])) ** 2, 0);
+  const r2 = ssTot === 0 ? 0 : 1 - ssRes / ssTot;
+  return {
+    metric: outcome.metric,
+    n: xs.length,
+    pearson: p,
+    spearman: s,
+    rewardModel: { intercept, slope, r2 }
+  };
+}
+function buildReleaseScorecard(composite, lift, contamination) {
+  const axes = [];
+  const liftPass = lift === void 0 || lift.ci95[0] > 0 ? "pass" : lift.delta > 0 ? "warn" : "fail";
+  axes.push({
+    name: "quality-lift",
+    status: liftPass,
+    detail: lift ? `delta=${lift.delta.toFixed(3)}, CI95=[${lift.ci95[0].toFixed(3)}, ${lift.ci95[1].toFixed(3)}], n=${lift.n}` : "no baseline/candidate pair available"
+  });
+  const contamPass = contamination === void 0 || contamination.leaks === 0 ? "pass" : "fail";
+  axes.push({
+    name: "contamination",
+    status: contamPass,
+    detail: contamination ? `${contamination.leaks} canary leak(s)` : "no canaries supplied"
+  });
+  axes.push({
+    name: "composite-distribution",
+    status: composite.mean >= 0.5 ? "pass" : composite.mean >= 0.3 ? "warn" : "fail",
+    detail: `mean=${composite.mean.toFixed(3)}, p50=${composite.p50.toFixed(3)}, p95=${composite.p95.toFixed(3)} over n=${composite.n}`
+  });
+  const status = axes.some((a) => a.status === "fail") ? "fail" : axes.some((a) => a.status === "warn") ? "warn" : "pass";
+  return {
+    status,
+    axes,
+    issues: []
+  };
+}
+function buildRecommendations(ctx) {
+  const out = [];
+  if (ctx.priorPeriodComparison) {
+    const ppc = ctx.priorPeriodComparison;
+    const label = ppc.windowLabel ?? "baseline period";
+    for (const name of ppc.regressedMetrics) {
+      const d = ppc.metrics[name];
+      if (!d) continue;
+      out.push({
+        priority: "critical",
+        kind: "investigate",
+        title: `${name} regressed from ${d.baseline.toFixed(3)} \u2192 ${d.current.toFixed(3)} vs ${label}`,
+        detail: `Welch CI95 = [${d.ci95[0].toFixed(3)}, ${d.ci95[1].toFixed(3)}], p=${d.pValue.toFixed(4)}, Cohen's d=${d.cohensD.toFixed(2)} (n_current=${d.currentN}, n_baseline=${d.baselineN}). The regression is statistically significant at p<0.05 with at-least-small effect size.`,
+        evidencePath: `priorPeriodComparison.metrics.${name}`
+      });
+    }
+    for (const name of ppc.improvedMetrics) {
+      const d = ppc.metrics[name];
+      if (!d) continue;
+      out.push({
+        priority: "low",
+        kind: "ship",
+        title: `${name} improved from ${d.baseline.toFixed(3)} \u2192 ${d.current.toFixed(3)} vs ${label}`,
+        detail: `Welch CI95 = [${d.ci95[0].toFixed(3)}, ${d.ci95[1].toFixed(3)}], p=${d.pValue.toFixed(4)}, Cohen's d=${d.cohensD.toFixed(2)} (n_current=${d.currentN}, n_baseline=${d.baselineN}). Statistically significant improvement worth flagging.`,
+        evidencePath: `priorPeriodComparison.metrics.${name}`
+      });
+    }
+  }
+  if (ctx.composite.n > 0) {
+    if (ctx.composite.mean < 0.3) {
+      const tail = ctx.composite.tailRuns ?? [];
+      const names = tail.slice(0, 5).map((t) => `${t.runId}=${t.score.toFixed(3)}`).join(", ");
+      out.push({
+        priority: "critical",
+        kind: "investigate",
+        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below the 0.3 floor \u2014 the agent is broken on this corpus`,
+        detail: tail.length > 0 ? `Worst ${tail.length} run${tail.length === 1 ? "" : "s"} to inspect first: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.` : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,
+        evidencePath: "composite.tailRuns"
+      });
+    } else if (ctx.composite.mean < 0.5) {
+      const tail = ctx.composite.tailRuns ?? [];
+      const names = tail.slice(0, 3).map((t) => `${t.runId}=${t.score.toFixed(3)}`).join(", ");
+      out.push({
+        priority: "high",
+        kind: "investigate",
+        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below 0.5 \u2014 investigate the lower tail before claiming the agent is healthy`,
+        detail: tail.length > 0 ? `Worst ${tail.length} run${tail.length === 1 ? "" : "s"}: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.` : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,
+        evidencePath: "composite.tailRuns"
+      });
+    }
+  }
+  if (ctx.failureModes && ctx.failureModes.length > 0) {
+    const top = ctx.failureModes[0];
+    if (top.count >= 3 && top.share >= 0.15) {
+      out.push({
+        priority: top.share >= 0.25 ? "high" : "medium",
+        kind: "investigate",
+        title: `'${top.mode}' is the dominant failure mode \u2014 ${top.count} runs (${(top.share * 100).toFixed(0)}% of the corpus)`,
+        detail: `The mean composite can look acceptable while one named failure dominates the lower tail. ${top.count} of ${ctx.composite.n} runs failed with '${top.mode}'${ctx.failureModes.length > 1 ? ` (next: '${ctx.failureModes[1].mode}' \xD7${ctx.failureModes[1].count})` : ""}. Fix this cause first.`,
+        evidencePath: "failureModes"
+      });
+    }
+  }
+  if (Object.keys(ctx.judges).length === 0 && ctx.composite.n > 0) {
+    out.push({
+      priority: "medium",
+      kind: "expand-corpus",
+      title: "No judge scores recorded \u2014 per-dimension + calibration insights unavailable",
+      detail: "Records have no `outcome.judgeScores`. To unlock perDimension, judges, and calibration, attach a Judge run during your eval pass and populate `outcome.judgeScores.perJudge[judgeName][dimension] = score`. See `docs/insight-report.md` for the expected shape.",
+      evidencePath: "judges"
+    });
+  }
+  if (ctx.lift) {
+    const decisive = ctx.lift.ci95[0] > ctx.threshold;
+    const inconclusive = ctx.lift.ci95[0] <= ctx.threshold && ctx.lift.ci95[1] > ctx.threshold;
+    if (decisive) {
+      out.push({
+        priority: "critical",
+        kind: "ship",
+        title: `Ship \u2014 lift ${ctx.lift.delta.toFixed(3)} (95% CI ${ctx.lift.ci95[0].toFixed(3)}..${ctx.lift.ci95[1].toFixed(3)})`,
+        detail: `Holdout lift exceeds threshold ${ctx.threshold} with 95% bootstrap confidence (n=${ctx.lift.n}, p=${ctx.lift.pValue.toFixed(4)}, d=${ctx.lift.cohensD.toFixed(2)}).`,
+        evidencePath: "lift"
+      });
+    } else if (inconclusive) {
+      out.push({
+        priority: "high",
+        kind: "expand-corpus",
+        title: `Inconclusive \u2014 need ~${ctx.lift.requiredN} paired runs (have ${ctx.lift.n}) at current effect size`,
+        detail: `CI straddles threshold. Current MDE at 80% power is ${ctx.lift.mde.toFixed(3)}; observed delta is ${ctx.lift.delta.toFixed(3)}.`,
+        evidencePath: "lift"
+      });
+    } else {
+      out.push({
+        priority: "critical",
+        kind: "hold",
+        title: `Hold \u2014 lift CI lower bound ${ctx.lift.ci95[0].toFixed(3)} is at or below threshold ${ctx.threshold}`,
+        detail: `Bootstrap CI provides no statistical evidence the candidate is better. Consider tightening the mutation or expanding the holdout.`,
+        evidencePath: "lift"
+      });
+    }
+  }
+  if (ctx.contamination && ctx.contamination.leaks > 0) {
+    out.push({
+      priority: "critical",
+      kind: "fix",
+      title: `${ctx.contamination.leaks} canary leak${ctx.contamination.leaks === 1 ? "" : "s"} detected`,
+      detail: `Holdout integrity is compromised. The lift number is unreliable until you investigate.`,
+      evidencePath: "contamination"
+    });
+  }
+  if (ctx.interRater && ctx.interRater.kappa < 0.5) {
+    out.push({
+      priority: "high",
+      kind: "recalibrate",
+      title: `Inter-rater agreement \u03BA=${ctx.interRater.kappa.toFixed(2)} is below 0.5`,
+      detail: `Raters disagree on what 'good' looks like. Top disagreement cases listed in interRater.disagreementCases \u2014 consider a triage meeting or refining the rubric.`,
+      evidencePath: "interRater"
+    });
+  }
+  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {
+    const top = ctx.failureClusters.clusters[0];
+    out.push({
+      priority: "high",
+      kind: "investigate",
+      title: `Top failure cluster: ${top.name} (${(top.share * 100).toFixed(0)}% of failures)`,
+      detail: `${ctx.failureClusters.totalFailures} runs failed. The largest cluster groups ${top.exemplars.length} exemplars under '${top.name}'.`,
+      evidencePath: "failureClusters.clusters[0]"
+    });
+  }
+  if (ctx.outcomeCorrelation && Math.abs(ctx.outcomeCorrelation.spearman) < 0.3) {
+    out.push({
+      priority: "medium",
+      kind: "recalibrate",
+      title: `Judge scores decoupled from ${ctx.outcomeCorrelation.metric} (Spearman \u03C1=${ctx.outcomeCorrelation.spearman.toFixed(2)})`,
+      detail: `Your judges score what they were trained to score, but it isn't predicting downstream ${ctx.outcomeCorrelation.metric}. Consider retraining the judge against ${ctx.outcomeCorrelation.metric} as the gold signal.`,
+      evidencePath: "outcomeCorrelation"
+    });
+  }
+  return out;
+}
+
+export {
+  checkCanaries,
+  checkBehavioralCanary,
+  runBehavioralCanaries,
+  canaryLeakView,
+  HoldoutAuditor,
+  summarizeExecution,
+  analyzeRuns
+};
+//# sourceMappingURL=chunk-SQQED7ZH.js.map