@tangle-network/agent-eval 0.127.0 → 0.128.1

package/dist/{chunk-IMGJZBFG.js → chunk-VLOATJQ2.js}

@@ -12,7 +12,7 @@ import {
 import {
   canonicalize,
   hashJson
-} from "./chunk-VSMTAMNK.js";
+} from "./chunk-2JX3CFMB.js";
 
 // src/summary-report.ts
 function summaryTable(runs, opts = {}) {
@@ -905,4 +905,4 @@ export {
   RESEARCH_REPORT_HARD_PAIR_FLOOR,
   researchReport
 };
-//# sourceMappingURL=chunk-IMGJZBFG.js.map
+//# sourceMappingURL=chunk-VLOATJQ2.js.map

package/dist/{chunk-Y5CYOL42.js → chunk-XDWDC2MP.js}

@@ -1,7 +1,14 @@
 import {
   summaryTable
-} from "./chunk-IMGJZBFG.js";
+} from "./chunk-VLOATJQ2.js";
 import {
+  validateRunRecord
+} from "./chunk-2JX3CFMB.js";
+import {
+  FAILURE_CLASSES
+} from "./chunk-MA6HLL3S.js";
+import {
+  ValidationError,
   VerificationError
 } from "./chunk-ONWEPEDO.js";
 
@@ -23,8 +30,16 @@ var DEFAULT_THRESHOLDS = {
 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 runs = filterCandidate(
+    (input.runs ?? []).map(validateRunRecord),
+    candidateId,
+    input.baselineId
+  );
+  const traces = filterTraceCandidate(
+    (input.traces ?? []).map(validateReleaseTraceEvidence),
+    candidateId,
+    input.baselineId
+  );
   const scenarios = input.scenarios ?? [];
   const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length;
   const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios);
@@ -74,7 +89,7 @@ function evaluateReleaseConfidence(input) {
     multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,
     splitCounts,
     domainCounts: countDomains(scenarios),
-    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),
+    failureClassCounts: countFailureClasses(runs, traces, thresholds.failureScoreThreshold),
     responsibleSurfaceCounts: countResponsibleSurfaces(traces)
   };
   const issues = [];
@@ -119,6 +134,20 @@ function filterTraceCandidate(traces, candidateId, baselineId) {
     return traces.filter((t) => t.candidateId === void 0 || t.candidateId !== baselineId);
   return [...traces];
 }
+function validateReleaseTraceEvidence(trace, index) {
+  const value = trace;
+  if (Object.hasOwn(value, "failureMode")) {
+    throw new ValidationError(
+      `traces[${index}].failureMode is not supported; use canonical failureClass`
+    );
+  }
+  if (trace.failureClass !== void 0 && !FAILURE_CLASSES.includes(trace.failureClass)) {
+    throw new ValidationError(
+      `traces[${index}].failureClass must be one of ${FAILURE_CLASSES.join(", ")}`
+    );
+  }
+  return trace;
+}
 function checkCorpus(input, thresholds, metrics, issues) {
   if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {
     issues.push({
@@ -332,18 +361,18 @@ function countDomains(scenarios) {
   }
   return out;
 }
-function countFailureModes(runs, traces, threshold) {
+function countFailureClasses(runs, traces, threshold) {
   const out = {};
   for (const run of runs) {
     if (runPassOutcome(run, threshold) === false) {
-      const mode = run.failureMode ?? run.failureClass ?? "low_score";
-      out[mode] = (out[mode] ?? 0) + 1;
+      const failureClass = run.failureClass !== void 0 && run.failureClass !== "success" ? run.failureClass : "unknown";
+      out[failureClass] = (out[failureClass] ?? 0) + 1;
     }
   }
   for (const trace of traces) {
     if (tracePassOutcome(trace, threshold) === false) {
-      const mode = trace.failureMode ?? (trace.ok === false ? "not_ok" : "low_score");
-      out[mode] = (out[mode] ?? 0) + 1;
+      const failureClass = trace.failureClass !== void 0 && trace.failureClass !== "success" ? trace.failureClass : "unknown";
+      out[failureClass] = (out[failureClass] ?? 0) + 1;
     }
   }
   return out;
@@ -384,10 +413,10 @@ function runPassOutcome(run, threshold) {
   return score === void 0 ? null : score >= threshold;
 }
 function hasExplicitTaskFailure(run) {
-  return run.failureClass !== void 0 && run.failureClass !== "success" || run.failureMode !== void 0;
+  return run.failureClass !== void 0 && run.failureClass !== "success";
 }
 function tracePassOutcome(trace, threshold) {
-  if (trace.failureMode !== void 0) return false;
+  if (trace.failureClass !== void 0 && trace.failureClass !== "success") return false;
   if (trace.ok === false) return false;
   if (isFiniteNumber(trace.score)) return trace.score >= threshold;
   return trace.ok === true ? true : null;
@@ -605,9 +634,9 @@ function renderReleaseReport(scorecard, options = {}) {
     for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`);
     lines.push("");
   }
-  const failures = entries(scorecard.metrics.failureModeCounts);
+  const failures = entries(scorecard.metrics.failureClassCounts);
   if (failures.length > 0) {
-    lines.push("## Failure Modes");
+    lines.push("## Failure Classes");
     lines.push("");
     for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`);
     lines.push("");
@@ -663,4 +692,4 @@ export {
   judgeReplayGate,
   renderReleaseReport
 };
-//# sourceMappingURL=chunk-Y5CYOL42.js.map
+//# sourceMappingURL=chunk-XDWDC2MP.js.map

package/dist/chunk-XDWDC2MP.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"sourcesContent":["/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n *   - Dataset manifests prove corpus/version coverage.\n *   - RunRecord rows prove reproducible search/holdout outcomes.\n *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n *   - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { ValidationError, VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Severity of an actionable finding attached to a run/trace. */\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\n/** Actionable side-info — a diagnosed finding the loop can act on. */\nexport interface ActionableSideInfo {\n  /** Stable expectation/check id when available. */\n  expectationId?: string\n  /** Human-readable diagnosis of what happened. */\n  message: string\n  severity?: AsiSeverity\n  /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n  evidence?: string\n  /** Prompt/tool/context surface likely responsible. */\n  responsibleSurface?: string\n  /** Suggested fix in natural language. */\n  suggestion?: string\n  /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n  matched?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n  | 'corpus'\n  | 'quality'\n  | 'reliability'\n  | 'generalization'\n  | 'diagnostics'\n  | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n  scenarioId: string\n  candidateId?: string\n  split?: RunSplitTag\n  score?: number\n  ok?: boolean\n  turnCount?: number\n  costUsd?: number\n  durationMs?: number\n  /** Canonical task-failure class. Free-form detail belongs in ASI. */\n  failureClass?: FailureClass\n  asi?: ActionableSideInfo[]\n  metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n  /** Require a Dataset manifest or explicit scenarios. Default true. */\n  requireCorpus?: boolean\n  minScenarioCount?: number\n  minSearchRuns?: number\n  minHoldoutRuns?: number\n  /** Require at least one holdout scenario/run. Default true. */\n  requireHoldout?: boolean\n  minPassRate?: number\n  minMeanScore?: number\n  /** Search mean may exceed holdout mean by at most this much. */\n  maxOverfitGap?: number\n  maxMeanCostUsd?: number\n  maxP95WallMs?: number\n  /** Low-score/failed rows must carry ASI. Default true. */\n  requireAsiForFailures?: boolean\n  /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n  failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n  target: string\n  candidateId?: string\n  baselineId?: string\n  dataset?: DatasetManifest\n  scenarios?: readonly DatasetScenario[]\n  runs?: readonly RunRecord[]\n  traces?: readonly ReleaseTraceEvidence[]\n  gateDecision?: GateDecision | null\n  thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n  name: ReleaseConfidenceAxisName\n  status: ReleaseConfidenceStatus\n  score: number | null\n  detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n  axis: ReleaseConfidenceAxisName\n  severity: 'critical' | 'warning'\n  code: string\n  detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n  scenarioCount: number\n  /** Search rows with a finite search score. */\n  searchRuns: number\n  /** Holdout rows with a finite holdout score. */\n  holdoutRuns: number\n  /** Runs with neither a split-matched score nor an explicit task failure. */\n  unscoredRuns: number\n  /** Run rows, or trace rows when no runs exist, with no classified terminal result. */\n  unclassifiedTerminalRuns: number\n  /** Run rows, or trace rows when no runs exist, that ended unsuccessfully. */\n  terminalFailureRuns: number\n  /** Success fraction when every run or fallback trace row has a classified result. */\n  reliabilityRate: number | null\n  passRate: number | null\n  meanScore: number | null\n  searchMeanScore: number | null\n  holdoutMeanScore: number | null\n  overfitGap: number | null\n  meanCostUsd: number | null\n  p95WallMs: number | null\n  failedRows: number\n  failuresWithAsi: number\n  singleShotTraces: number\n  multiShotTraces: number\n  splitCounts: Record<DatasetSplit, number>\n  domainCounts: Record<string, number>\n  failureClassCounts: Partial<Record<FailureClass, number>>\n  responsibleSurfaceCounts: Record<string, number>\n}\n\nexport interface ReleaseConfidenceScorecard {\n  target: string\n  candidateId: string | null\n  baselineId: string | null\n  status: ReleaseConfidenceStatus\n  promote: boolean\n  axes: ReleaseConfidenceAxis[]\n  issues: ReleaseConfidenceIssue[]\n  metrics: ReleaseConfidenceMetrics\n  dataset: DatasetManifest | null\n  gateDecision: GateDecision | null\n  summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n  requireCorpus: true,\n  minScenarioCount: 1,\n  minSearchRuns: 1,\n  minHoldoutRuns: 1,\n  requireHoldout: true,\n  minPassRate: 0.8,\n  minMeanScore: 0.7,\n  maxOverfitGap: 0.15,\n  maxMeanCostUsd: Number.POSITIVE_INFINITY,\n  maxP95WallMs: Number.POSITIVE_INFINITY,\n  requireAsiForFailures: true,\n  failureScoreThreshold: 0.5,\n}\n\nexport function evaluateReleaseConfidence(\n  input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n  const candidateId = input.candidateId ?? null\n  const runs = filterCandidate(\n    (input.runs ?? []).map(validateRunRecord),\n    candidateId,\n    input.baselineId,\n  )\n  const traces = filterTraceCandidate(\n    (input.traces ?? []).map(validateReleaseTraceEvidence),\n    candidateId,\n    input.baselineId,\n  )\n  const scenarios = input.scenarios ?? []\n  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n  const searchScores = scoresFor(runs, 'search')\n  const holdoutScores = scoresFor(runs, 'holdout')\n  const runScores = runs.map(runSplitScore).filter(isFiniteNumber)\n  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n  const scoreUniverse = runs.length > 0 ? runScores : traceScores\n  const qualityRuns = runs.filter((run) => runSplitScore(run) !== undefined)\n  const unscoredRuns = runs.filter(\n    (run) =>\n      runSplitScore(run) === undefined &&\n      !hasExplicitTaskFailure(run) &&\n      !isFailedTerminalOutcome(run.terminalOutcome),\n  ).length\n  const passOutcomes =\n    runs.length > 0\n      ? runs.map((run) => runPassOutcome(run, thresholds.failureScoreThreshold))\n      : traces.map((trace) => tracePassOutcome(trace, thresholds.failureScoreThreshold))\n  const reliabilityRows =\n    runs.length > 0\n      ? runs.map((run) => terminalSuccess(run.terminalOutcome))\n      : traces.map((trace) => trace.ok)\n  const unclassifiedTerminalRuns = reliabilityRows.filter((outcome) => outcome === undefined).length\n  const terminalFailureRuns = reliabilityRows.filter((outcome) => outcome === false).length\n  const reliabilityRate =\n    reliabilityRows.length === 0 || unclassifiedTerminalRuns > 0\n      ? null\n      : (reliabilityRows.length - terminalFailureRuns) / reliabilityRows.length\n  const searchRuns = qualityRuns.filter((r) => r.splitTag === 'search').length\n  const holdoutRuns = qualityRuns.filter((r) => r.splitTag === 'holdout').length\n  const failed = failedRows(runs, traces, thresholds.failureScoreThreshold)\n  const searchMeanScore = meanOrNull(searchScores)\n  const holdoutMeanScore = meanOrNull(holdoutScores)\n  const runCosts = runs.flatMap((run) =>\n    run.costProvenance.kind === 'uncaptured' ? [] : [run.costProvenance.usd],\n  )\n  const traceCosts = traces.map((trace) => trace.costUsd).filter(isFiniteNumber)\n  const meanCostUsd =\n    runs.length > 0\n      ? runCosts.length === runs.length\n        ? meanOrNull(runCosts)\n        : null\n      : meanOrNull(traceCosts)\n  const wallTimes =\n    runs.length > 0\n      ? runs.map((run) => run.wallMs)\n      : traces.map((trace) => trace.durationMs).filter(isFiniteNumber)\n  const metrics: ReleaseConfidenceMetrics = {\n    scenarioCount,\n    searchRuns,\n    holdoutRuns,\n    unscoredRuns,\n    unclassifiedTerminalRuns,\n    terminalFailureRuns,\n    reliabilityRate,\n    passRate: passOutcomeRate(passOutcomes),\n    meanScore: meanOrNull(scoreUniverse),\n    searchMeanScore,\n    holdoutMeanScore,\n    overfitGap: diffOrNull(searchMeanScore, holdoutMeanScore),\n    meanCostUsd,\n    p95WallMs: percentileOrNull(wallTimes, 0.95),\n    failedRows: failed.length,\n    failuresWithAsi: failed.filter((row) => row.hasAsi).length,\n    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n    splitCounts,\n    domainCounts: countDomains(scenarios),\n    failureClassCounts: countFailureClasses(runs, traces, thresholds.failureScoreThreshold),\n    responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n  }\n\n  const issues: ReleaseConfidenceIssue[] = []\n  checkCorpus(input, thresholds, metrics, issues)\n  checkQuality(thresholds, metrics, issues)\n  checkReliability(metrics, issues)\n  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n  checkDiagnostics(thresholds, metrics, issues)\n  checkEfficiency(thresholds, metrics, issues)\n\n  const axes = buildAxes(metrics, thresholds, issues)\n  const status = issues.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : issues.length > 0\n      ? 'warn'\n      : 'pass'\n\n  return {\n    target: input.target,\n    candidateId,\n    baselineId: input.baselineId ?? null,\n    status,\n    promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n    axes,\n    issues,\n    metrics,\n    dataset: input.dataset ?? null,\n    gateDecision: input.gateDecision ?? null,\n    summary: renderSummary(input.target, status, metrics, issues),\n  }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n  const scorecard = evaluateReleaseConfidence(input)\n  if (scorecard.status === 'fail') {\n    throw new VerificationError(scorecard.summary)\n  }\n  return scorecard\n}\n\nfunction filterCandidate(\n  runs: readonly RunRecord[],\n  candidateId: string | null,\n  baselineId?: string,\n): RunRecord[] {\n  if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n  return [...runs]\n}\n\nfunction filterTraceCandidate(\n  traces: readonly ReleaseTraceEvidence[],\n  candidateId: string | null,\n  baselineId?: string,\n): ReleaseTraceEvidence[] {\n  if (candidateId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n  if (baselineId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n  return [...traces]\n}\n\nfunction validateReleaseTraceEvidence(\n  trace: ReleaseTraceEvidence,\n  index: number,\n): ReleaseTraceEvidence {\n  const value = trace as unknown as Record<string, unknown>\n  if (Object.hasOwn(value, 'failureMode')) {\n    throw new ValidationError(\n      `traces[${index}].failureMode is not supported; use canonical failureClass`,\n    )\n  }\n  if (trace.failureClass !== undefined && !FAILURE_CLASSES.includes(trace.failureClass)) {\n    throw new ValidationError(\n      `traces[${index}].failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n    )\n  }\n  return trace\n}\n\nfunction checkCorpus(\n  input: ReleaseConfidenceInput,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_corpus',\n      detail: 'No Dataset manifest or scenarios supplied.',\n    })\n  }\n  if (metrics.scenarioCount < thresholds.minScenarioCount) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'few_scenarios',\n      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n    })\n  }\n  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_holdout_split',\n      detail: 'Corpus has no holdout scenarios.',\n    })\n  }\n}\n\nfunction checkQuality(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.searchRuns < thresholds.minSearchRuns) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'few_search_runs',\n      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n    })\n  }\n  if (metrics.unscoredRuns > 0) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'unscored_runs',\n      detail: `${metrics.unscoredRuns} supplied run(s) have no task result.`,\n    })\n  }\n  if (metrics.passRate === null || metrics.meanScore === null) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'missing_quality_scores',\n      detail: 'No task-quality scores are available for pass-rate and mean-score checks.',\n    })\n  }\n  if (metrics.passRate !== null && metrics.passRate < thresholds.minPassRate) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_pass_rate',\n      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n    })\n  }\n  if (metrics.meanScore !== null && metrics.meanScore < thresholds.minMeanScore) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_mean_score',\n      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n    })\n  }\n}\n\nfunction checkReliability(\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.reliabilityRate === null) {\n    issues.push({\n      axis: 'reliability',\n      severity: 'critical',\n      code: 'missing_reliability_evidence',\n      detail:\n        metrics.unclassifiedTerminalRuns > 0\n          ? `${metrics.unclassifiedTerminalRuns} supplied run(s) have no classified terminal result.`\n          : 'No classified terminal results are available.',\n    })\n  }\n  if (metrics.terminalFailureRuns > 0) {\n    issues.push({\n      axis: 'reliability',\n      severity: 'critical',\n      code: 'terminal_run_failures',\n      detail: `${metrics.terminalFailureRuns} run(s) ended failed, cancelled, or incomplete.`,\n    })\n  }\n}\n\nfunction checkGeneralization(\n  gateDecision: GateDecision | null,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'few_holdout_runs',\n      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n    })\n  }\n  if (metrics.overfitGap !== null && metrics.overfitGap > thresholds.maxOverfitGap) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'overfit_gap',\n      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n    })\n  }\n  if (gateDecision && !gateDecision.promote) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n      detail: gateDecision.reason,\n    })\n  }\n}\n\nfunction checkDiagnostics(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (!thresholds.requireAsiForFailures) return\n  if (metrics.failedRows > metrics.failuresWithAsi) {\n    issues.push({\n      axis: 'diagnostics',\n      severity: 'critical',\n      code: 'missing_failure_asi',\n      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n    })\n  }\n}\n\nfunction checkEfficiency(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (Number.isFinite(thresholds.maxMeanCostUsd) && metrics.meanCostUsd === null) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'missing_cost',\n      detail: 'A finite cost limit was configured but no cost evidence is available.',\n    })\n  } else if (metrics.meanCostUsd !== null && metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'cost_budget',\n      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n    })\n  }\n  if (Number.isFinite(thresholds.maxP95WallMs) && metrics.p95WallMs === null) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'missing_latency',\n      detail: 'A finite latency limit was configured but no latency evidence is available.',\n    })\n  } else if (metrics.p95WallMs !== null && metrics.p95WallMs > thresholds.maxP95WallMs) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'latency_budget',\n      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n    })\n  }\n}\n\nfunction buildAxes(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n  return [\n    axis(\n      'corpus',\n      issues,\n      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n    ),\n    axis(\n      'quality',\n      issues,\n      metrics.passRate === null || metrics.meanScore === null\n        ? null\n        : Math.min(metrics.passRate, metrics.meanScore),\n      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n    ),\n    axis(\n      'reliability',\n      issues,\n      metrics.reliabilityRate,\n      `successRate=${fmt(metrics.reliabilityRate)} terminalFailures=${metrics.terminalFailureRuns} unclassified=${metrics.unclassifiedTerminalRuns}`,\n    ),\n    axis(\n      'generalization',\n      issues,\n      gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n    ),\n    axis(\n      'diagnostics',\n      issues,\n      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n    ),\n    axis(\n      'efficiency',\n      issues,\n      efficiencyScore(metrics, thresholds),\n      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n    ),\n  ]\n}\n\nfunction axis(\n  name: ReleaseConfidenceAxisName,\n  issues: ReleaseConfidenceIssue[],\n  score: number | null,\n  detail: string,\n): ReleaseConfidenceAxis {\n  const own = issues.filter((i) => i.axis === name)\n  const status = own.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : own.length > 0\n      ? 'warn'\n      : 'pass'\n  return { name, status, score: score === null ? null : bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n  const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n  for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n  return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const scenario of scenarios) {\n    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n    out[domain] = (out[domain] ?? 0) + 1\n  }\n  return out\n}\n\nfunction countFailureClasses(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Partial<Record<FailureClass, number>> {\n  const out: Partial<Record<FailureClass, number>> = {}\n  for (const run of runs) {\n    if (runPassOutcome(run, threshold) === false) {\n      const failureClass =\n        run.failureClass !== undefined && run.failureClass !== 'success'\n          ? run.failureClass\n          : 'unknown'\n      out[failureClass] = (out[failureClass] ?? 0) + 1\n    }\n  }\n  for (const trace of traces) {\n    if (tracePassOutcome(trace, threshold) === false) {\n      const failureClass =\n        trace.failureClass !== undefined && trace.failureClass !== 'success'\n          ? trace.failureClass\n          : 'unknown'\n      out[failureClass] = (out[failureClass] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const trace of traces) {\n    for (const asi of trace.asi ?? []) {\n      const surface = asi.responsibleSurface ?? 'unknown'\n      out[surface] = (out[surface] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction failedRows(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Array<{ hasAsi: boolean }> {\n  const out: Array<{ hasAsi: boolean }> = []\n  for (const run of runs) {\n    if (runPassOutcome(run, threshold) === false) {\n      const asiMetric = run.outcome.raw.asi\n      out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n    }\n  }\n  for (const trace of traces) {\n    if (tracePassOutcome(trace, threshold) === false) {\n      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n    }\n  }\n  return out\n}\n\nfunction passOutcomeRate(outcomes: readonly (boolean | null)[]): number | null {\n  const classified = outcomes.filter((outcome): outcome is boolean => outcome !== null)\n  if (classified.length === 0) return null\n  return classified.filter(Boolean).length / classified.length\n}\n\nfunction runPassOutcome(run: RunRecord, threshold: number): boolean | null {\n  if (hasExplicitTaskFailure(run)) return false\n  const score = runSplitScore(run)\n  return score === undefined ? null : score >= threshold\n}\n\nfunction hasExplicitTaskFailure(run: RunRecord): boolean {\n  return run.failureClass !== undefined && run.failureClass !== 'success'\n}\n\nfunction tracePassOutcome(trace: ReleaseTraceEvidence, threshold: number): boolean | null {\n  if (trace.failureClass !== undefined && trace.failureClass !== 'success') return false\n  if (trace.ok === false) return false\n  if (isFiniteNumber(trace.score)) return trace.score >= threshold\n  return trace.ok === true ? true : null\n}\n\nfunction isFailedTerminalOutcome(\n  outcome: RunRecord['terminalOutcome'],\n): outcome is 'failed' | 'cancelled' | 'incomplete' {\n  return outcome === 'failed' || outcome === 'cancelled' || outcome === 'incomplete'\n}\n\nfunction terminalSuccess(outcome: RunRecord['terminalOutcome']): boolean | undefined {\n  if (outcome === 'succeeded') return true\n  if (isFailedTerminalOutcome(outcome)) return false\n  return undefined\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n  return runs\n    .filter((run) => run.splitTag === split)\n    .map(runSplitScore)\n    .filter(isFiniteNumber)\n}\n\nfunction runSplitScore(run: RunRecord): number | undefined {\n  const score = run.splitTag === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore\n  return isFiniteNumber(score) ? score : undefined\n}\n\nfunction meanOrNull(xs: readonly number[]): number | null {\n  if (xs.length === 0) return null\n  return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentileOrNull(xs: readonly number[], p: number): number | null {\n  if (xs.length === 0) return null\n  const sorted = [...xs].sort((a, b) => a - b)\n  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction diffOrNull(a: number | null, b: number | null): number | null {\n  if (a === null || b === null) return null\n  return a - b\n}\n\nfunction gapScore(gap: number | null, maxGap: number): number | null {\n  if (gap === null) return null\n  if (maxGap <= 0) return gap <= 0 ? 1 : 0\n  return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n): number | null {\n  const cost = Number.isFinite(thresholds.maxMeanCostUsd)\n    ? metrics.meanCostUsd === null\n      ? null\n      : bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n    : 1\n  const latency = Number.isFinite(thresholds.maxP95WallMs)\n    ? metrics.p95WallMs === null\n      ? null\n      : bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n    : 1\n  if (cost === null || latency === null) return null\n  return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n  target: string,\n  status: ReleaseConfidenceStatus,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): string {\n  const prefix = `release confidence ${status}: ${target}`\n  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n  if (issues.length === 0) return `${prefix}; ${metricText}`\n  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number | null): string {\n  if (x === null) return 'n/a'\n  return x.toFixed(4)\n}\n","/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n *   - Bootstrap mean baseline vs candidate (1k resamples).\n *   - Compute the delta distribution; pass if the lower CI bound > 0.\n *   - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)\n *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)\n *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n  baselineMean: number\n  candidateMean: number\n  /** candidateMean - baselineMean, point estimate. */\n  delta: number\n  /** Lower bound of the (1 - alpha) CI on the delta. */\n  ciLower: number\n  /** Upper bound of the (1 - alpha) CI on the delta. */\n  ciUpper: number\n  /** Number of bootstrap resamples used. */\n  iterations: number\n  alpha: number\n  verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n  /** Confidence level alpha (default 0.05 → 95% CI). */\n  alpha?: number\n  /** Number of resamples (default 1000). */\n  iterations?: number\n  /**\n   * Minimum total samples (baseline + candidate) below which we always\n   * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n   * Default 6 (combined).\n   */\n  minTotalSamples?: number\n  /** RNG seed for reproducibility. Default: Math.random. */\n  seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n  baseline: number[],\n  candidate: number[],\n  options: BootstrapOptions = {},\n): BootstrapResult {\n  const alpha = options.alpha ?? 0.05\n  const iterations = options.iterations ?? 1000\n  const minTotal = options.minTotalSamples ?? 6\n  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n  const baselineMean = mean(baseline)\n  const candidateMean = mean(candidate)\n  const delta = candidateMean - baselineMean\n\n  if (\n    baseline.length + candidate.length < minTotal ||\n    baseline.length === 0 ||\n    candidate.length === 0\n  ) {\n    return {\n      baselineMean,\n      candidateMean,\n      delta,\n      ciLower: -Infinity,\n      ciUpper: Infinity,\n      iterations: 0,\n      alpha,\n      verdict: 'INCONCLUSIVE',\n    }\n  }\n\n  const deltas: number[] = new Array(iterations)\n  for (let i = 0; i < iterations; i++) {\n    const bResample = resample(baseline, rng)\n    const cResample = resample(candidate, rng)\n    deltas[i] = mean(cResample) - mean(bResample)\n  }\n  deltas.sort((a, b) => a - b)\n  const lowerIdx = Math.floor((alpha / 2) * iterations)\n  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n  const ciLower = deltas[Math.max(0, lowerIdx)]!\n  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n  let verdict: Verdict\n  if (ciLower > 0) verdict = 'ADVANCE'\n  else if (ciUpper < 0) verdict = 'REVERT'\n  else if (delta >= 0) verdict = 'KEEP'\n  else verdict = 'INCONCLUSIVE'\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ciLower,\n    ciUpper,\n    iterations,\n    alpha,\n    verdict,\n  }\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  let s = 0\n  for (const x of xs) s += x\n  return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n  const out = new Array(xs.length)\n  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n  return out\n}\n\n/** Mulberry32 — fast deterministic PRNG. Stable across runs given the same seed. */\nfunction mulberry32(seed: number): () => number {\n  let t = seed >>> 0\n  return () => {\n    t += 0x6d2b79f5\n    let r = t\n    r = Math.imul(r ^ (r >>> 15), r | 1)\n    r ^= r + Math.imul(r ^ (r >>> 7), r | 61)\n    return ((r ^ (r >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n  let h = 2166136261\n  for (const x of [...a, ...b]) {\n    const view = new Float64Array([x])\n    const bytes = new Uint8Array(view.buffer)\n    for (const byte of bytes) {\n      h ^= byte\n      h = Math.imul(h, 16777619)\n    }\n  }\n  return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n  baselineOutputs: TOutput[]\n  candidateOutputs: TOutput[]\n  /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n  judge: (output: TOutput) => Promise<number> | number\n  alpha?: number\n  iterations?: number\n  /** RNG seed for reproducibility. */\n  seed?: number\n  /** Maximum concurrent judge calls. Default 4. */\n  judgeConcurrency?: number\n}\n\n/**\n * Confirm a candidate's win with a stronger judge: score baseline and candidate outputs independently, then bootstrap a CI to verify the lift generalises beyond the inner loop.\n */\nexport async function judgeReplayGate<TOutput>(\n  args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n  const concurrency = args.judgeConcurrency ?? 4\n  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n  const ci = bootstrapCi(baselineScores, candidateScores, {\n    ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n    ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n    ...(args.seed !== undefined ? { seed: args.seed } : {}),\n  })\n  return {\n    ...ci,\n    baselineSamples: baselineScores.length,\n    candidateSamples: candidateScores.length,\n  }\n}\n\nasync function scoreAll<TOutput>(\n  outputs: TOutput[],\n  judge: (output: TOutput) => Promise<number> | number,\n  concurrency: number,\n): Promise<number[]> {\n  const results: number[] = new Array(outputs.length)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= outputs.length) return\n      const v = await judge(outputs[i]!)\n      results[i] = Number.isFinite(v) ? v : 0\n    }\n  }\n  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n  return results\n}\n","import type { ReleaseConfidenceScorecard } from './release-confidence'\nimport type { RunRecord } from './run-record'\nimport { summaryTable } from './summary-report'\n\nexport interface RenderReleaseReportOptions {\n  title?: string\n  runs?: readonly RunRecord[]\n  comparator?: string\n  traceAnalystFindings?: readonly string[]\n  nextActions?: readonly string[]\n}\n\nexport function renderReleaseReport(\n  scorecard: ReleaseConfidenceScorecard,\n  options: RenderReleaseReportOptions = {},\n): string {\n  const title = options.title ?? `Release Report: ${scorecard.target}`\n  const lines: string[] = []\n  lines.push(`# ${title}`)\n  lines.push('')\n  lines.push(`Status: **${scorecard.status.toUpperCase()}**`)\n  lines.push(`Promote: **${scorecard.promote ? 'yes' : 'no'}**`)\n  if (scorecard.candidateId) lines.push(`Candidate: \\`${scorecard.candidateId}\\``)\n  if (scorecard.baselineId) lines.push(`Baseline: \\`${scorecard.baselineId}\\``)\n  lines.push('')\n  lines.push(scorecard.summary)\n  lines.push('')\n\n  lines.push('## Metrics')\n  lines.push('')\n  lines.push('| Metric | Value |')\n  lines.push('|---|---:|')\n  lines.push(`| Scenarios | ${scorecard.metrics.scenarioCount} |`)\n  lines.push(`| Search runs | ${scorecard.metrics.searchRuns} |`)\n  lines.push(`| Holdout runs | ${scorecard.metrics.holdoutRuns} |`)\n  lines.push(`| Unscored runs | ${scorecard.metrics.unscoredRuns} |`)\n  lines.push(`| Terminal failures | ${scorecard.metrics.terminalFailureRuns} |`)\n  lines.push(`| Pass rate | ${pct(scorecard.metrics.passRate)} |`)\n  lines.push(`| Mean score | ${num(scorecard.metrics.meanScore)} |`)\n  lines.push(`| Search mean | ${num(scorecard.metrics.searchMeanScore)} |`)\n  lines.push(`| Holdout mean | ${num(scorecard.metrics.holdoutMeanScore)} |`)\n  lines.push(`| Overfit gap | ${num(scorecard.metrics.overfitGap)} |`)\n  lines.push(`| Mean cost | $${num(scorecard.metrics.meanCostUsd)} |`)\n  lines.push(`| p95 wall time | ${duration(scorecard.metrics.p95WallMs)} |`)\n  lines.push('')\n\n  if (scorecard.issues.length > 0) {\n    lines.push('## Issues')\n    lines.push('')\n    for (const issue of scorecard.issues) {\n      lines.push(`- **${issue.severity}** \\`${issue.code}\\` (${issue.axis}): ${issue.detail}`)\n    }\n    lines.push('')\n  }\n\n  const surfaces = entries(scorecard.metrics.responsibleSurfaceCounts)\n  if (surfaces.length > 0) {\n    lines.push('## Responsible Surfaces')\n    lines.push('')\n    for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`)\n    lines.push('')\n  }\n\n  const failures = entries(scorecard.metrics.failureClassCounts)\n  if (failures.length > 0) {\n    lines.push('## Failure Classes')\n    lines.push('')\n    for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`)\n    lines.push('')\n  }\n\n  if (options.runs && options.runs.length > 0) {\n    lines.push('## Run Summary')\n    lines.push('')\n    lines.push(\n      summaryTable([...options.runs], {\n        comparator: options.comparator ?? scorecard.baselineId ?? undefined,\n        split: 'holdout',\n      }).markdown,\n    )\n    lines.push('')\n  }\n\n  if (options.traceAnalystFindings && options.traceAnalystFindings.length > 0) {\n    lines.push('## TraceAnalyst Findings')\n    lines.push('')\n    for (const finding of options.traceAnalystFindings) lines.push(`- ${finding}`)\n    lines.push('')\n  }\n\n  const nextActions = options.nextActions ?? defaultNextActions(scorecard)\n  if (nextActions.length > 0) {\n    lines.push('## Next Actions')\n    lines.push('')\n    for (const action of nextActions) lines.push(`- ${action}`)\n    lines.push('')\n  }\n\n  return `${lines.join('\\n').trimEnd()}\\n`\n}\n\nfunction defaultNextActions(scorecard: ReleaseConfidenceScorecard): string[] {\n  if (scorecard.promote) return ['Promote the candidate and keep canaries enabled.']\n  return scorecard.issues\n    .filter((issue) => issue.severity === 'critical')\n    .map((issue) => `Resolve ${issue.code}: ${issue.detail}`)\n}\n\nfunction entries(values: Record<string, number>): Array<[string, number]> {\n  return Object.entries(values)\n    .filter(([, count]) => count > 0)\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n}\n\nfunction pct(value: number | null): string {\n  return value === null ? 'n/a' : `${(value * 100).toFixed(1)}%`\n}\n\nfunction num(value: number | null): string {\n  return value === null ? 'n/a' : value.toFixed(3)\n}\n\nfunction duration(value: number | null): string {\n  return value === null ? 'n/a' : `${Math.round(value)} ms`\n}\n"],"mappings":";;;;;;;;;;;;;;;AA4JA,IAAM,qBAA4D;AAAA,EAChE,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,aAAa;AAAA,EACb,cAAc;AAAA,EACd,eAAe;AAAA,EACf,gBAAgB,OAAO;AAAA,EACvB,cAAc,OAAO;AAAA,EACrB,uBAAuB;AAAA,EACvB,uBAAuB;AACzB;AAEO,SAAS,0BACd,OAC4B;AAC5B,QAAM,aAAa,EAAE,GAAG,oBAAoB,GAAG,MAAM,WAAW;AAChE,QAAM,cAAc,MAAM,eAAe;AACzC,QAAM,OAAO;AAAA,KACV,MAAM,QAAQ,CAAC,GAAG,IAAI,iBAAiB;AAAA,IACxC;AAAA,IACA,MAAM;AAAA,EACR;AACA,QAAM,SAAS;AAAA,KACZ,MAAM,UAAU,CAAC,GAAG,IAAI,4BAA4B;AAAA,IACrD;AAAA,IACA,MAAM;AAAA,EACR;AACA,QAAM,YAAY,MAAM,aAAa,CAAC;AACtC,QAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;AAChE,QAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;AAC/E,QAAM,eAAe,UAAU,MAAM,QAAQ;AAC7C,QAAM,gBAAgB,UAAU,MAAM,SAAS;AAC/C,QAAM,YAAY,KAAK,IAAI,aAAa,EAAE,OAAO,cAAc;AAC/D,QAAM,cAAc,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,OAAO,cAAc;AACpE,QAAM,gBAAgB,KAAK,SAAS,IAAI,YAAY;AACpD,QAAM,cAAc,KAAK,OAAO,CAAC,QAAQ,cAAc,GAAG,MAAM,MAAS;AACzE,QAAM,eAAe,KAAK;AAAA,IACxB,CAAC,QACC,cAAc,GAAG,MAAM,UACvB,CAAC,uBAAuB,GAAG,KAC3B,CAAC,wBAAwB,IAAI,eAAe;AAAA,EAChD,EAAE;AACF,QAAM,eACJ,KAAK,SAAS,IACV,KAAK,IAAI,CAAC,QAAQ,eAAe,KAAK,WAAW,qBAAqB,CAAC,IACvE,OAAO,IAAI,CAAC,UAAU,iBAAiB,OAAO,WAAW,qBAAqB,CAAC;AACrF,QAAM,kBACJ,KAAK,SAAS,IACV,KAAK,IAAI,CAAC,QAAQ,gBAAgB,IAAI,eAAe,CAAC,IACtD,OAAO,IAAI,CAAC,UAAU,MAAM,EAAE;AACpC,QAAM,2BAA2B,gBAAgB,OAAO,CAAC,YAAY,YAAY,MAAS,EAAE;AAC5F,QAAM,sBAAsB,gBAAgB,OAAO,CAAC,YAAY,YAAY,KAAK,EAAE;AACnF,QAAM,kBACJ,gBAAgB,WAAW,KAAK,2BAA2B,IACvD,QACC,gBAAgB,SAAS,uBAAuB,gBAAgB;AACvE,QAAM,aAAa,YAAY,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ,EAAE;AACtE,QAAM,cAAc,YAAY,OAAO,CAAC,MAAM,EAAE,aAAa,SAAS,EAAE;AACxE,QAAM,SAAS,WAAW,MAAM,QAAQ,WAAW,qBAAqB;AACxE,QAAM,kBAAkB,WAAW,YAAY;AAC/C,QAAM,mBAAmB,WAAW,aAAa;AACjD,QAAM,WAAW,KAAK;AAAA,IAAQ,CAAC,QAC7B,IAAI,eAAe,SAAS,eAAe,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG;AAAA,EACzE;AACA,QAAM,aAAa,OAAO,IAAI,CAAC,UAAU,MAAM,OAAO,EAAE,OAAO,cAAc;AAC7E,QAAM,cACJ,KAAK,SAAS,IACV,SAAS,WAAW,KAAK,SACvB,WAAW,QAAQ,IACnB,OACF,WAAW,UAAU;AAC3B,QAAM,YACJ,KAAK,SAAS,IACV,KAAK,IAAI,CAAC,QAAQ,IAAI,MAAM,IAC5B,OAAO,IAAI,CAAC,UAAU,MAAM,UAAU,EAAE,OAAO,cAAc;AACnE,QAAM,UAAoC;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,UAAU,gBAAgB,YAAY;AAAA,IACtC,WAAW,WAAW,aAAa;AAAA,IACnC;AAAA,IACA;AAAA,IACA,YAAY,WAAW,iBAAiB,gBAAgB;AAAA,IACxD;AAAA,IACA,WAAW,iBAAiB,WAAW,IAAI;AAAA,IAC3C,YAAY,OAAO;AAAA,IACnB,iBAAiB,OAAO,OAAO,CAAC,QAAQ,IAAI,MAAM,EAAE;AAAA,IACpD,kBAAkB,OAAO,OAAO,CAAC,MAAM,EAAE,cAAc,CAAC,EAAE;AAAA,IAC1D,iBAAiB,OAAO,OAAO,CAAC,OAAO,EAAE,aAAa,KAAK,CAAC,EAAE;AAAA,IAC9D;AAAA,IACA,cAAc,aAAa,SAAS;AAAA,IACpC,oBAAoB,oBAAoB,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACtF,0BAA0B,yBAAyB,MAAM;AAAA,EAC3D;AAEA,QAAM,SAAmC,CAAC;AAC1C,cAAY,OAAO,YAAY,SAAS,MAAM;AAC9C,eAAa,YAAY,SAAS,MAAM;AACxC,mBAAiB,SAAS,MAAM;AAChC,sBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;AAC3E,mBAAiB,YAAY,SAAS,MAAM;AAC5C,kBAAgB,YAAY,SAAS,MAAM;AAE3C,QAAM,OAAO,UAAU,SAAS,YAAY,MAAM;AAClD,QAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;AAEN,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd;AAAA,IACA,YAAY,MAAM,cAAc;AAAA,IAChC;AAAA,IACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;AAAA,IACjF;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,MAAM,WAAW;AAAA,IAC1B,cAAc,MAAM,gBAAgB;AAAA,IACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;AAAA,EAC9D;AACF;AAEO,SAAS,wBAAwB,OAA2D;AACjG,QAAM,YAAY,0BAA0B,KAAK;AACjD,MAAI,UAAU,WAAW,QAAQ;AAC/B,UAAM,IAAI,kBAAkB,UAAU,OAAO;AAAA,EAC/C;AACA,SAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;AACb,MAAI,YAAa,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,WAAW;AACxE,MAAI,WAAY,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAU;AACtE,SAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;AACxB,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,WAAW;AAC1F,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,UAAU;AACzF,SAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,6BACP,OACA,OACsB;AACtB,QAAM,QAAQ;AACd,MAAI,OAAO,OAAO,OAAO,aAAa,GAAG;AACvC,UAAM,IAAI;AAAA,MACR,UAAU,KAAK;AAAA,IACjB;AAAA,EACF;AACA,MAAI,MAAM,iBAAiB,UAAa,CAAC,gBAAgB,SAAS,MAAM,YAAY,GAAG;AACrF,UAAM,IAAI;AAAA,MACR,UAAU,KAAK,iCAAiC,gBAAgB,KAAK,IAAI,CAAC;AAAA,IAC5E;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;AACN,MAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GAAG;AACtF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,gBAAgB,WAAW,kBAAkB;AACvD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,sBAAsB,WAAW,gBAAgB;AAAA,IACnF,CAAC;AAAA,EACH;AACA,MAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAAG;AAClE,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,aAAa,WAAW,eAAe;AACjD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,UAAU,wBAAwB,WAAW,aAAa;AAAA,IAC/E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,eAAe,GAAG;AAC5B,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,YAAY;AAAA,IACjC,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,aAAa,QAAQ,QAAQ,cAAc,MAAM;AAC3D,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,aAAa,QAAQ,QAAQ,WAAW,WAAW,aAAa;AAC1E,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,CAAC,MAAM,IAAI,WAAW,WAAW,CAAC;AAAA,IAC5E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAAc;AAC7E,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,SACA,QACM;AACN,MAAI,QAAQ,oBAAoB,MAAM;AACpC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QACE,QAAQ,2BAA2B,IAC/B,GAAG,QAAQ,wBAAwB,yDACnC;AAAA,IACR,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,sBAAsB,GAAG;AACnC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,mBAAmB;AAAA,IACxC,CAAC;AAAA,EACH;AACF;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;AACN,MAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAAgB;AAChF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,WAAW,yBAAyB,WAAW,cAAc;AAAA,IAClF,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,eAAe,QAAQ,QAAQ,aAAa,WAAW,eAAe;AAChF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,CAAC,MAAM,IAAI,WAAW,aAAa,CAAC;AAAA,IAC1F,CAAC;AAAA,EACH;AACA,MAAI,gBAAgB,CAAC,aAAa,SAAS;AACzC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM,QAAQ,aAAa,iBAAiB,QAAQ;AAAA,MACpD,QAAQ,aAAa;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,YACA,SACA,QACM;AACN,MAAI,CAAC,WAAW,sBAAuB;AACvC,MAAI,QAAQ,aAAa,QAAQ,iBAAiB;AAChD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,eAAe;AAAA,IACzD,CAAC;AAAA,EACH;AACF;AAEA,SAAS,gBACP,YACA,SACA,QACM;AACN,MAAI,OAAO,SAAS,WAAW,cAAc,KAAK,QAAQ,gBAAgB,MAAM;AAC9E,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH,WAAW,QAAQ,gBAAgB,QAAQ,QAAQ,cAAc,WAAW,gBAAgB;AAC1F,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,CAAC,MAAM,IAAI,WAAW,cAAc,CAAC;AAAA,IACrF,CAAC;AAAA,EACH;AACA,MAAI,OAAO,SAAS,WAAW,YAAY,KAAK,QAAQ,cAAc,MAAM;AAC1E,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH,WAAW,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAAc;AACpF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,UACP,SACA,YACA,QACyB;AACzB,SAAO;AAAA,IACL;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC;AAAA,MACxE,GAAG,QAAQ,aAAa,uBAAuB,QAAQ,YAAY,OAAO;AAAA,IAC5E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,aAAa,QAAQ,QAAQ,cAAc,OAC/C,OACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS;AAAA,MAChD,YAAY,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IACvE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,MACR,eAAe,IAAI,QAAQ,eAAe,CAAC,qBAAqB,QAAQ,mBAAmB,iBAAiB,QAAQ,wBAAwB;AAAA,IAC9I;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,SAAS,QAAQ,YAAY,WAAW,aAAa;AAAA,MACrD,eAAe,QAAQ,WAAW,eAAe,IAAI,QAAQ,UAAU,CAAC;AAAA,IAC1E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ;AAAA,MACjE,mBAAmB,QAAQ,eAAe,IAAI,QAAQ,UAAU;AAAA,IAClE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,SAAS,UAAU;AAAA,MACnC,eAAe,IAAI,QAAQ,WAAW,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IAC7E;AAAA,EACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;AACvB,QAAM,MAAM,OAAO,OAAO,CAAC,MAAM,EAAE,SAAS,IAAI;AAChD,QAAM,SAAS,IAAI,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;AACN,SAAO,EAAE,MAAM,QAAQ,OAAO,UAAU,OAAO,OAAO,QAAQ,KAAK,GAAG,OAAO;AAC/E;AAEA,SAAS,oBAAoB,WAAqE;AAChG,QAAM,SAAuC,EAAE,OAAO,GAAG,KAAK,GAAG,MAAM,GAAG,SAAS,EAAE;AACrF,aAAW,YAAY,UAAW,QAAO,SAAS,SAAS,OAAO;AAClE,SAAO;AACT;AAEA,SAAS,aAAa,WAA+D;AACnF,QAAM,MAA8B,CAAC;AACrC,aAAW,YAAY,WAAW;AAChC,UAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;AACnE,QAAI,MAAM,KAAK,IAAI,MAAM,KAAK,KAAK;AAAA,EACrC;AACA,SAAO;AACT;AAEA,SAAS,oBACP,MACA,QACA,WACuC;AACvC,QAAM,MAA6C,CAAC;AACpD,aAAW,OAAO,MAAM;AACtB,QAAI,eAAe,KAAK,SAAS,MAAM,OAAO;AAC5C,YAAM,eACJ,IAAI,iBAAiB,UAAa,IAAI,iBAAiB,YACnD,IAAI,eACJ;AACN,UAAI,YAAY,KAAK,IAAI,YAAY,KAAK,KAAK;AAAA,IACjD;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;AAChD,YAAM,eACJ,MAAM,iBAAiB,UAAa,MAAM,iBAAiB,YACvD,MAAM,eACN;AACN,UAAI,YAAY,KAAK,IAAI,YAAY,KAAK,KAAK;AAAA,IACjD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;AACjG,QAAM,MAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,eAAW,OAAO,MAAM,OAAO,CAAC,GAAG;AACjC,YAAM,UAAU,IAAI,sBAAsB;AAC1C,UAAI,OAAO,KAAK,IAAI,OAAO,KAAK,KAAK;AAAA,IACvC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;AAC5B,QAAM,MAAkC,CAAC;AACzC,aAAW,OAAO,MAAM;AACtB,QAAI,eAAe,KAAK,SAAS,MAAM,OAAO;AAC5C,YAAM,YAAY,IAAI,QAAQ,IAAI;AAClC,UAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;AAChD,UAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;AAAA,IACnD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,gBAAgB,UAAsD;AAC7E,QAAM,aAAa,SAAS,OAAO,CAAC,YAAgC,YAAY,IAAI;AACpF,MAAI,WAAW,WAAW,EAAG,QAAO;AACpC,SAAO,WAAW,OAAO,OAAO,EAAE,SAAS,WAAW;AACxD;AAEA,SAAS,eAAe,KAAgB,WAAmC;AACzE,MAAI,uBAAuB,GAAG,EAAG,QAAO;AACxC,QAAM,QAAQ,cAAc,GAAG;AAC/B,SAAO,UAAU,SAAY,OAAO,SAAS;AAC/C;AAEA,SAAS,uBAAuB,KAAyB;AACvD,SAAO,IAAI,iBAAiB,UAAa,IAAI,iBAAiB;AAChE;AAEA,SAAS,iBAAiB,OAA6B,WAAmC;AACxF,MAAI,MAAM,iBAAiB,UAAa,MAAM,iBAAiB,UAAW,QAAO;AACjF,MAAI,MAAM,OAAO,MAAO,QAAO;AAC/B,MAAI,eAAe,MAAM,KAAK,EAAG,QAAO,MAAM,SAAS;AACvD,SAAO,MAAM,OAAO,OAAO,OAAO;AACpC;AAEA,SAAS,wBACP,SACkD;AAClD,SAAO,YAAY,YAAY,YAAY,eAAe,YAAY;AACxE;AAEA,SAAS,gBAAgB,SAA4D;AACnF,MAAI,YAAY,YAAa,QAAO;AACpC,MAAI,wBAAwB,OAAO,EAAG,QAAO;AAC7C,SAAO;AACT;AAEA,SAAS,UAAU,MAA4B,OAA8B;AAC3E,SAAO,KACJ,OAAO,CAAC,QAAQ,IAAI,aAAa,KAAK,EACtC,IAAI,aAAa,EACjB,OAAO,cAAc;AAC1B;AAEA,SAAS,cAAc,KAAoC;AACzD,QAAM,QAAQ,IAAI,aAAa,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ;AAClF,SAAO,eAAe,KAAK,IAAI,QAAQ;AACzC;AAEA,SAAS,WAAW,IAAsC;AACxD,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,iBAAiB,IAAuB,GAA0B;AACzE,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,SAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC,CAAC;AAC1F;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,WAAW,GAAkB,GAAiC;AACrE,MAAI,MAAM,QAAQ,MAAM,KAAM,QAAO;AACrC,SAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAoB,QAA+B;AACnE,MAAI,QAAQ,KAAM,QAAO;AACzB,MAAI,UAAU,EAAG,QAAO,OAAO,IAAI,IAAI;AACvC,SAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACe;AACf,QAAM,OAAO,OAAO,SAAS,WAAW,cAAc,IAClD,QAAQ,gBAAgB,OACtB,OACA,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IAC1E;AACJ,QAAM,UAAU,OAAO,SAAS,WAAW,YAAY,IACnD,QAAQ,cAAc,OACpB,OACA,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACtE;AACJ,MAAI,SAAS,QAAQ,YAAY,KAAM,QAAO;AAC9C,SAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;AACR,QAAM,SAAS,sBAAsB,MAAM,KAAK,MAAM;AACtD,QAAM,aAAa,aAAa,QAAQ,aAAa,eAAe,QAAQ,UAAU,gBAAgB,QAAQ,WAAW,aAAa,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAC/L,MAAI,OAAO,WAAW,EAAG,QAAO,GAAG,MAAM,KAAK,UAAU;AACxD,SAAO,GAAG,MAAM,KAAK,UAAU,YAAY,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAChF;AAEA,SAAS,IAAI,GAA0B;AACrC,MAAI,MAAM,KAAM,QAAO;AACvB,SAAO,EAAE,QAAQ,CAAC;AACpB;;;ACpsBO,SAAS,YACd,UACA,WACA,UAA4B,CAAC,GACZ;AACjB,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,WAAW,QAAQ,mBAAmB;AAC5C,QAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;AAEpE,QAAM,eAAe,KAAK,QAAQ;AAClC,QAAM,gBAAgB,KAAK,SAAS;AACpC,QAAM,QAAQ,gBAAgB;AAE9B,MACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GACrB;AACA,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,SAAS;AAAA,MACT,YAAY;AAAA,MACZ;AAAA,MACA,SAAS;AAAA,IACX;AAAA,EACF;AAEA,QAAM,SAAmB,IAAI,MAAM,UAAU;AAC7C,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,YAAY,SAAS,UAAU,GAAG;AACxC,UAAM,YAAY,SAAS,WAAW,GAAG;AACzC,WAAO,CAAC,IAAI,KAAK,SAAS,IAAI,KAAK,SAAS;AAAA,EAC9C;AACA,SAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,QAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;AACpD,QAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;AAC5D,QAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ,CAAC;AAC5C,QAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ,CAAC;AAEzD,MAAI;AACJ,MAAI,UAAU,EAAG,WAAU;AAAA,WAClB,UAAU,EAAG,WAAU;AAAA,WACvB,SAAS,EAAG,WAAU;AAAA,MAC1B,WAAU;AAEf,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,KAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,MAAI,IAAI;AACR,aAAW,KAAK,GAAI,MAAK;AACzB,SAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;AAC3D,QAAM,MAAM,IAAI,MAAM,GAAG,MAAM;AAC/B,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,KAAI,CAAC,IAAI,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC7E,SAAO;AACT;AAGA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAGA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,IAAI;AACR,aAAW,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;AAC5B,UAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;AACjC,UAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;AACxC,eAAW,QAAQ,OAAO;AACxB,WAAK;AACL,UAAI,KAAK,KAAK,GAAG,QAAQ;AAAA,IAC3B;AAAA,EACF;AACA,SAAO,MAAM;AACf;AAiCA,eAAsB,gBACpB,MACkF;AAClF,QAAM,cAAc,KAAK,oBAAoB;AAC7C,QAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;AACnF,QAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;AACrF,QAAM,KAAK,YAAY,gBAAgB,iBAAiB;AAAA,IACtD,GAAI,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;AAAA,IACxD,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,IACvE,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,EACvD,CAAC;AACD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,iBAAiB,eAAe;AAAA,IAChC,kBAAkB,gBAAgB;AAAA,EACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;AACnB,QAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;AAClD,MAAI,OAAO;AACX,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,QAAQ,OAAQ;AACzB,YAAM,IAAI,MAAM,MAAM,QAAQ,CAAC,CAAE;AACjC,cAAQ,CAAC,IAAI,OAAO,SAAS,CAAC,IAAI,IAAI;AAAA,IACxC;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,GAAG,MAAM,OAAO,CAAC,CAAC;AAClF,SAAO;AACT;;;AC1NO,SAAS,oBACd,WACA,UAAsC,CAAC,GAC/B;AACR,QAAM,QAAQ,QAAQ,SAAS,mBAAmB,UAAU,MAAM;AAClE,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,KAAK,KAAK,EAAE;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,aAAa,UAAU,OAAO,YAAY,CAAC,IAAI;AAC1D,QAAM,KAAK,cAAc,UAAU,UAAU,QAAQ,IAAI,IAAI;AAC7D,MAAI,UAAU,YAAa,OAAM,KAAK,gBAAgB,UAAU,WAAW,IAAI;AAC/E,MAAI,UAAU,WAAY,OAAM,KAAK,eAAe,UAAU,UAAU,IAAI;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,UAAU,OAAO;AAC5B,QAAM,KAAK,EAAE;AAEb,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,oBAAoB;AAC/B,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,iBAAiB,UAAU,QAAQ,aAAa,IAAI;AAC/D,QAAM,KAAK,mBAAmB,UAAU,QAAQ,UAAU,IAAI;AAC9D,QAAM,KAAK,oBAAoB,UAAU,QAAQ,WAAW,IAAI;AAChE,QAAM,KAAK,qBAAqB,UAAU,QAAQ,YAAY,IAAI;AAClE,QAAM,KAAK,yBAAyB,UAAU,QAAQ,mBAAmB,IAAI;AAC7E,QAAM,KAAK,iBAAiB,IAAI,UAAU,QAAQ,QAAQ,CAAC,IAAI;AAC/D,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,SAAS,CAAC,IAAI;AACjE,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,eAAe,CAAC,IAAI;AACxE,QAAM,KAAK,oBAAoB,IAAI,UAAU,QAAQ,gBAAgB,CAAC,IAAI;AAC1E,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,UAAU,CAAC,IAAI;AACnE,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,WAAW,CAAC,IAAI;AACnE,QAAM,KAAK,qBAAqB,SAAS,UAAU,QAAQ,SAAS,CAAC,IAAI;AACzE,QAAM,KAAK,EAAE;AAEb,MAAI,UAAU,OAAO,SAAS,GAAG;AAC/B,UAAM,KAAK,WAAW;AACtB,UAAM,KAAK,EAAE;AACb,eAAW,SAAS,UAAU,QAAQ;AACpC,YAAM,KAAK,OAAO,MAAM,QAAQ,QAAQ,MAAM,IAAI,OAAO,MAAM,IAAI,MAAM,MAAM,MAAM,EAAE;AAAA,IACzF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,wBAAwB;AACnE,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,yBAAyB;AACpC,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,SAAS,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,OAAO,KAAK,KAAK,EAAE;AAC5E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,kBAAkB;AAC7D,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,oBAAoB;AAC/B,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,MAAM,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,IAAI,KAAK,KAAK,EAAE;AACtE,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,QAAQ,QAAQ,KAAK,SAAS,GAAG;AAC3C,UAAM,KAAK,gBAAgB;AAC3B,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,aAAa,CAAC,GAAG,QAAQ,IAAI,GAAG;AAAA,QAC9B,YAAY,QAAQ,cAAc,UAAU,cAAc;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC,EAAE;AAAA,IACL;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,wBAAwB,QAAQ,qBAAqB,SAAS,GAAG;AAC3E,UAAM,KAAK,0BAA0B;AACrC,UAAM,KAAK,EAAE;AACb,eAAW,WAAW,QAAQ,qBAAsB,OAAM,KAAK,KAAK,OAAO,EAAE;AAC7E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,cAAc,QAAQ,eAAe,mBAAmB,SAAS;AACvE,MAAI,YAAY,SAAS,GAAG;AAC1B,UAAM,KAAK,iBAAiB;AAC5B,UAAM,KAAK,EAAE;AACb,eAAW,UAAU,YAAa,OAAM,KAAK,KAAK,MAAM,EAAE;AAC1D,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,CAAC;AAAA;AACtC;AAEA,SAAS,mBAAmB,WAAiD;AAC3E,MAAI,UAAU,QAAS,QAAO,CAAC,kDAAkD;AACjF,SAAO,UAAU,OACd,OAAO,CAAC,UAAU,MAAM,aAAa,UAAU,EAC/C,IAAI,CAAC,UAAU,WAAW,MAAM,IAAI,KAAK,MAAM,MAAM,EAAE;AAC5D;AAEA,SAAS,QAAQ,QAAyD;AACxE,SAAO,OAAO,QAAQ,MAAM,EACzB,OAAO,CAAC,CAAC,EAAE,KAAK,MAAM,QAAQ,CAAC,EAC/B,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC;AAC3D;AAEA,SAAS,IAAI,OAA8B;AACzC,SAAO,UAAU,OAAO,QAAQ,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC;AAC7D;AAEA,SAAS,IAAI,OAA8B;AACzC,SAAO,UAAU,OAAO,QAAQ,MAAM,QAAQ,CAAC;AACjD;AAEA,SAAS,SAAS,OAA8B;AAC9C,SAAO,UAAU,OAAO,QAAQ,GAAG,KAAK,MAAM,KAAK,CAAC;AACtD;","names":[]}

package/dist/{chunk-ARCFIY7U.js → chunk-XPRT64IE.js}

@@ -1,7 +1,7 @@
 import {
   fsCampaignStorage,
   runCampaign
-} from "./chunk-WED5KLZQ.js";
+} from "./chunk-EZJEIH2R.js";
 import {
   __export
 } from "./chunk-PZ5AY32C.js";
@@ -763,4 +763,4 @@ export {
   retrievalMetricsAtCutoff,
   benchmarks_exports
 };
-//# sourceMappingURL=chunk-ARCFIY7U.js.map
+//# sourceMappingURL=chunk-XPRT64IE.js.map

package/dist/contract/index.d.ts

@@ -517,15 +517,12 @@ interface RunRecord {
     /** Per-split scores + raw bag. */
     outcome: RunOutcome;
     /** Canonical task-failure class drawn from the shared
-     *  `FAILURE_CLASSES` taxonomy. This is the aggregation key that makes
-     *  "which failure dominates across the whole fleet" answerable in ONE
-     *  vocabulary — every agent classifies against the same enum. Producers
-     *  set it only from task-result evidence. Execution errors belong in
-     *  `outcome.raw.execution_error_count`, even when the run later fails. */
+     *  `FAILURE_CLASSES` taxonomy. Producers set it only from task-result
+     *  evidence. Execution errors belong in
+     *  `outcome.raw.execution_error_count`. */
     failureClass?: FailureClass;
-    /** Free-form task-failure detail, scoped UNDER `failureClass`
-     *  (e.g. failureClass='tool_recovery_failure', failureMode='forge_build_unsatisfied').
-     *  Do not populate this from a child execution error alone. */
+    /** Free-form task-failure detail scoped under a non-success
+     *  `failureClass`. It is invalid without that class. */
     failureMode?: string;
     /** Which split this run was drawn from. */
     splitTag: RunSplitTag;
@@ -3195,10 +3192,10 @@ interface InsightReport {
      *  Surfaced metrics: composite, cost, duration, tokenUsage, plus any
      *  per-dimension judge metric present in both windows. */
     priorPeriodComparison?: PriorPeriodComparison;
-    /** Model-free task-failure breakdown from `RunRecord.failureClass` or
-     *  `failureMode`, ranked by count descending. Tags from successful or
-     *  otherwise non-failed tasks are excluded. */
-    failureModes?: FailureModeTally[];
+    /** Model-free task-failure breakdown from `RunRecord.failureClass`, ranked
+     *  by count descending. Domain-specific `failureMode` detail is retained on
+     *  each record but never creates a second aggregation vocabulary. */
+    failureClasses?: FailureClassTally[];
     /** Top-N actionable recommendations, ranked by priority. The packet's
      *  human-readable layer; the numeric sections are the evidence. */
     recommendations: Recommendation[];
@@ -3406,15 +3403,13 @@ interface FailureClusterInsight {
     }>;
     totalFailures: number;
 }
-/** Model-free failure breakdown over the structured `RunRecord.failureMode`
- *  enum. Unlike `failureClusters` (semantic, requires an LLM analyst), this
- *  is computed directly from the tags the harness already recorded — so a
- *  customer ingesting one batch with no judge/analyst still learns which
- *  named failure dominates. */
-interface FailureModeTally {
-    /** The `failureMode` tag. */
-    mode: string;
-    /** Number of runs carrying this tag. */
+/** Model-free task-failure breakdown over canonical `RunRecord.failureClass`
+ *  values. Unlike semantic failure clusters, this is computed directly from
+ *  run records and does not require a model analyst. */
+interface FailureClassTally {
+    /** Canonical task-failure class. */
+    failureClass: FailureClass;
+    /** Number of failed runs carrying this class. */
     count: number;
     /** Share of the whole corpus, 0..1. */
     share: number;
@@ -3939,6 +3934,13 @@ interface SelfImproveOptions<TScenario extends Scenario$1, TArtifact> {
      * baseline-only run (set `budget.generations = 0`).
      */
     agent: (surface: MutableSurface, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>;
+    /**
+     * Snapshot-bearing model identity for agents that do not report a paid-call
+     * receipt through `ctx.cost.runPaidCall()`.
+     *
+     * Omit this when every cell reports its concrete model in a receipt.
+     */
+    model?: string;
     /** Scenarios to evaluate against. Train/holdout split is computed from
      *  these unless `budget.holdoutScenarios` is set explicitly. */
     scenarios: TScenario[];
@@ -5774,6 +5776,7 @@ declare function fromFeedbackTable(opts: FromFeedbackTableOptions): FromFeedback
  *   - `runId` (the group key)
  *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`
  *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`
+ *   - task failure class and detail from explicit `tangle.task.*` attributes
  *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`
  *   - token usage from model-call input, output, cache-read, and cache-write
  *     attributes without double-counting aggregate parent spans
@@ -5803,4 +5806,4 @@ interface FromOtelSpansOptions {
 }
 declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
 
-export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeSurface, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationProposer, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario$1 as Scenario, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeSurface, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationProposer, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario$1 as Scenario, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };

package/dist/contract/index.js

@@ -10,11 +10,11 @@ import {
 } from "../chunk-SZLVEKMJ.js";
 import {
   createHostedClient
-} from "../chunk-IG35YDN3.js";
+} from "../chunk-DRYIUNWY.js";
 import {
   analyzeRuns,
   summarizeExecution
-} from "../chunk-FA3MFYNA.js";
+} from "../chunk-NACAGYSY.js";
 import {
   REFERENCE_EQUIVALENCE_INPUT_LIMITS,
   REFERENCE_EQUIVALENCE_JUDGE_VERSION,
@@ -40,7 +40,7 @@ import {
   skillOptOptimizationMethod,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-65QETEKG.js";
+} from "../chunk-NKAGIDE2.js";
 import {
   campaignSplitDigest,
   createRunCostLedger,
@@ -48,7 +48,7 @@ import {
   inMemoryCampaignStorage,
   resolveRunDir,
   runCampaign
-} from "../chunk-WED5KLZQ.js";
+} from "../chunk-EZJEIH2R.js";
 import {
   buildDefaultAnalystRegistry,
   createChatClient
@@ -59,15 +59,15 @@ import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
 } from "../chunk-3RF76KTD.js";
-import "../chunk-424XHVKV.js";
+import "../chunk-NYLOYM6N.js";
 import {
   campaignCellExecutionEvidence,
   campaignCellJudgeDimensions,
   campaignCellTaskScore,
   campaignCellToRunRecord
-} from "../chunk-ZPUEEM5U.js";
+} from "../chunk-2MKQIFS4.js";
 import "../chunk-PBE2LOSS.js";
-import "../chunk-IMGJZBFG.js";
+import "../chunk-VLOATJQ2.js";
 import "../chunk-DPUHNQLN.js";
 import {
   pairedBootstrap
@@ -75,10 +75,11 @@ import {
 import "../chunk-WS3NZZQQ.js";
 import "../chunk-VI2UW6B6.js";
 import {
+  readTaskFailureLabels,
   recordAggregateMeasurements,
   summarizeExecutionMeasurements,
   summarizeTraceErrors
-} from "../chunk-IN455XKS.js";
+} from "../chunk-7ZZMD7UK.js";
 import "../chunk-PXE2VKMX.js";
 import "../chunk-ZET2UAYW.js";
 import "../chunk-GGE4NNQT.js";
@@ -88,10 +89,10 @@ import {
 } from "../chunk-P6FYH6K4.js";
 import "../chunk-PC4UYEBM.js";
 import {
+  modelHasSnapshot,
   parseRunRecordSafe
-} from "../chunk-NL33OXEP.js";
+} from "../chunk-2JX3CFMB.js";
 import "../chunk-MA6HLL3S.js";
-import "../chunk-VSMTAMNK.js";
 import {
   ValidationError
 } from "../chunk-ONWEPEDO.js";
@@ -408,14 +409,16 @@ async function runSelfImprove(opts, costLedger, startedAt, runDir, storage) {
         "baseline",
         runDir,
         opts.baselineSurface,
-        reportSplit
+        reportSplit,
+        opts.model
       ),
       ...cellsToRunRecords(
         reportWinnerCampaign.cells,
         "winner",
         runDir,
         result.winnerSurface,
-        reportSplit
+        reportSplit,
+        opts.model
       )
     ],
     baselineCandidateId: "baseline",
@@ -553,23 +556,34 @@ function hashString(s) {
   }
   return h.toString(16).padStart(8, "0");
 }
-function cellsToRunRecords(cells, candidateId, runId, surface, splitTag) {
+function cellsToRunRecords(cells, candidateId, runId, surface, splitTag, fallbackModel) {
   const promptHash = surfaceContentHash(surface);
   const configHash = surfaceContentHash(candidateId);
-  return cells.map(
-    (cell) => campaignCellToRunRecord(cell, {
+  return cells.map((cell) => {
+    const model = cell.resolvedModel ?? fallbackModel;
+    if (!model) {
+      throw new ValidationError(
+        `selfImprove.model is required when cell ${cell.cellId} has no paid-call model receipt`
+      );
+    }
+    if (!modelHasSnapshot(model)) {
+      throw new ValidationError(
+        `selfImprove model "${model}" lacks a snapshot version for cell ${cell.cellId}`
+      );
+    }
+    return campaignCellToRunRecord(cell, {
       runId: `${runId}::${candidateId}::${cell.cellId}`,
       experimentId: runId,
       candidateId,
       // scenarioId is explicit; seed keeps repeated runs distinct.
       seed: cell.rep * 1e6 + hashString(cell.scenarioId).slice(0, 6).split("").reduce((a, c) => a * 31 + c.charCodeAt(0) >>> 0, 0),
-      model: "campaign-cell",
+      model,
       promptHash,
       configHash,
       commitSha: "cell",
       splitTag
-    })
-  );
+    });
+  });
 }
 
 // src/contract/define-agent-eval.ts
@@ -1782,6 +1796,10 @@ function fromOtelSpans(opts) {
     const promptHash = readAttrString(groupSpans2, PROMPT_HASH_KEYS) ?? "sha256:unknown";
     const configHash = readAttrString(groupSpans2, CONFIG_HASH_KEYS) ?? "sha256:unknown";
     const score = resolveTaskScore(groupKey, groupSpans2, opts.scoreForRun);
+    const taskFailure = readTaskFailureLabels(
+      groupSpans2.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span)),
+      `fromOtelSpans: run '${groupKey}'`
+    );
     const rawNumeric = collectNumericAttrs(groupSpans2);
     const errorSummary = summarizeTraceErrors(
       groupSpans2.map((span) => ({
@@ -1837,6 +1855,7 @@ function fromOtelSpans(opts) {
       terminalOutcome,
       ...failedRoot ? { terminalFailureReason: failedRoot.status?.message ?? failedRoot.name } : {},
       outcome,
+      ...taskFailure,
       splitTag: defaultSplit,
       scenarioId
     });