npm - @tangle-network/agent-eval - Versions diffs - 1.0.0 → 1.0.2 - Mend

@tangle-network/agent-eval 1.0.0 → 1.0.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (221) hide show

package/dist/{release-confidence-CTUJz5mM.js.map → release-confidence--1iP-d3x.js.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"release-confidence-CTUJz5mM.js","names":[],"sources":["../src/promotion-gate.ts","../src/release-confidence.ts"],"sourcesContent":["/*\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\nimport { mulberry32 } from './statistics'\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. Absent, the seed is derived from the\n * compared arms, so the same inputs reproduce the same verdict. /\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 // A ZERO-WIDTH interval is an absence of evidence, not a certainty. Constant\n // arms make every resample identical, so pass/fail data promotes on nothing:\n // [0,0,0] vs [1,1,1] is six samples, clears the `minTotalSamples` floor, and\n // yields [1, 1] — an \"ADVANCE\" carrying no information about how far the\n // estimate could be wrong. Same rule as `pairedDeltaTest`, one estimator over.\n if (!Number.isFinite(ciLower) \|\| !Number.isFinite(ciUpper) \|\| ciLower === ciUpper) {\n verdict = 'INCONCLUSIVE'\n } else 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/** 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","/\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 { HeldOutGateDecision } from './held-out-gate'\nimport { isRealnessGated, observedSplitScore } from './rollout/reward'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\nimport { summarizeNumberSeries } from './statistics'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/* Severity of an actionable finding attached to a run/trace. /\ntype 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?: HeldOutGateDecision \| 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 Runs excluded from `passRate` because the authenticity gate flagged them as\n * gamed. Surfaced, never silent: a release whose pass rate is computed over a\n * shrunken denominator has to say by how much, or the exclusion is just a\n * different way of hiding the same runs.\n /\n realnessGatedRuns: 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: HeldOutGateDecision \| 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 // Realness-gated runs are EXCLUDED from the pass rate — numerator AND\n // denominator — and the count of what was dropped ships beside the rate as\n // `metrics.realnessGatedRuns`. Counting a gamed run as a pass made faking a\n // success the cheapest way to improve a release scorecard; scoring it 0\n // instead would be the other error, silently deflating the rate with a run\n // the gate says carries no usable verdict at all.\n const honestRuns = runs.filter((run) => !isRealnessGated(run))\n const passOutcomes =\n runs.length > 0\n ? honestRuns.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.usd === null ? [] : [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 realnessGatedRuns: runs.length - honestRuns.length,\n meanScore: meanOrNull(scoreUniverse),\n searchMeanScore,\n holdoutMeanScore,\n overfitGap: diffOrNull(searchMeanScore, holdoutMeanScore),\n meanCostUsd,\n p95WallMs: summarizeNumberSeries(wallTimes, { percentiles: [95] })?.percentiles?.[95] ?? null,\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: `${metrics.terminalFailureRuns} run(s) ended failed, cancelled, or incomplete.`,\n })\n }\n}\n\nfunction checkGeneralization(\n gateDecision: HeldOutGateDecision \| null,\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n issues.push({\n axis: 'generalization',\n severity: 'critical',\n code: 'few_holdout_runs',\n detail: `${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 // Ungated: a failure-mode census counts what the runs REPORTED, which is\n // the only way a gamed run's inflated score is visible at all. It is not a\n // promotion number — `passRate` is, and that one excludes gated runs and\n // publishes the excluded count as `metrics.realnessGatedRuns`.\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\n/\n RAW and split-exact (`observedSplitScore`): this feeds the per-split means,\n * the overfit gap, and the pass threshold — descriptions of what the runs\n * reported. A gamed run inflating them is visible next to `realnessGatedRuns`,\n * and the promotion number (`passRate`) excludes gated runs entirely.\n */\nfunction runSplitScore(run: RunRecord): number \| undefined {\n const score = observedSplitScore(run, run.splitTag === 'holdout' ? 'holdout' : 'search')\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 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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,SAAgB,YACd,UACA,WACA,UAA4B,CAAC,GACZ;CACjB,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,WAAW,QAAQ,mBAAmB;CAC5C,MAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;CAEpE,MAAM,eAAe,KAAK,QAAQ;CAClC,MAAM,gBAAgB,KAAK,SAAS;CACpC,MAAM,QAAQ,gBAAgB;CAE9B,IACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GAErB,OAAO;EACL;EACA;EACA;EACA,SAAS;EACT,SAAS;EACT,YAAY;EACZ;EACA,SAAS;CACX;CAGF,MAAM,SAAmB,IAAI,MAAM,UAAU;CAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,UAAU,GAAG;EACxC,MAAM,YAAY,SAAS,WAAW,GAAG;EACzC,OAAO,KAAK,KAAK,SAAS,IAAI,KAAK,SAAS;CAC9C;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;CACpD,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;CAC5D,MAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ;CAC3C,MAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ;CAExD,IAAI;CAMJ,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,CAAC,OAAO,SAAS,OAAO,KAAK,YAAY,SACxE,UAAU;MACL,IAAI,UAAU,GAAG,UAAU;MAC7B,IAAI,UAAU,GAAG,UAAU;MAC3B,IAAI,SAAS,GAAG,UAAU;MAC1B,UAAU;CAEf,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,IAAI,KAAK;CACzB,OAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;CAC3D,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM;CAC5E,OAAO;AACT;;AAGA,SAAS,SAAS,GAAa,GAAqB;CAClD,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;EAC5B,MAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;EACjC,MAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;EACxC,KAAK,MAAM,QAAQ,OAAO;GACxB,KAAK;GACL,IAAI,KAAK,KAAK,GAAG,QAAQ;EAC3B;CACF;CACA,OAAO,MAAM;AACf;;;;AAiCA,eAAsB,gBACpB,MACkF;CAClF,MAAM,cAAc,KAAK,oBAAoB;CAC7C,MAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;CACnF,MAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;CAMrF,OAAO;EACL,GANS,YAAY,gBAAgB,iBAAiB;GACtD,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,GAAI,KAAK,eAAe,KAAA,IAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;GACvE,GAAI,KAAK,SAAS,KAAA,IAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;EACvD,CAEM;EACJ,iBAAiB,eAAe;EAChC,kBAAkB,gBAAgB;CACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;CACnB,MAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;CAClD,IAAI,OAAO;CACX,eAAe,SAAwB;EACrC,OAAO,MAAM;GACX,MAAM,IAAI;GACV,IAAI,KAAK,QAAQ,QAAQ;GACzB,MAAM,IAAI,MAAM,MAAM,QAAQ,EAAG;GACjC,QAAQ,KAAK,OAAO,SAAS,CAAC,IAAI,IAAI;EACxC;CACF;CACA,MAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,SAAS,OAAO,CAAC,CAAC;CAClF,OAAO;AACT;;AC/DA,MAAM,qBAA4D;CAChE,eAAe;CACf,kBAAkB;CAClB,eAAe;CACf,gBAAgB;CAChB,gBAAgB;CAChB,aAAa;CACb,cAAc;CACd,eAAe;CACf,gBAAgB,OAAO;CACvB,cAAc,OAAO;CACrB,uBAAuB;CACvB,uBAAuB;AACzB;AAEA,SAAgB,0BACd,OAC4B;CAC5B,MAAM,aAAa;EAAE,GAAG;EAAoB,GAAG,MAAM;CAAW;CAChE,MAAM,cAAc,MAAM,eAAe;CACzC,MAAM,OAAO,iBACV,MAAM,QAAQ,CAAC,EAAA,CAAG,IAAI,iBAAiB,GACxC,aACA,MAAM,UACR;CACA,MAAM,SAAS,sBACZ,MAAM,UAAU,CAAC,EAAA,CAAG,IAAI,4BAA4B,GACrD,aACA,MAAM,UACR;CACA,MAAM,YAAY,MAAM,aAAa,CAAC;CACtC,MAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;CAChE,MAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;CAC/E,MAAM,eAAe,UAAU,MAAM,QAAQ;CAC7C,MAAM,gBAAgB,UAAU,MAAM,SAAS;CAC/C,MAAM,YAAY,KAAK,IAAI,aAAa,CAAC,CAAC,OAAO,cAAc;CAC/D,MAAM,cAAc,OAAO,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,OAAO,cAAc;CACpE,MAAM,gBAAgB,KAAK,SAAS,IAAI,YAAY;CACpD,MAAM,cAAc,KAAK,QAAQ,QAAQ,cAAc,GAAG,MAAM,KAAA,CAAS;CACzE,MAAM,eAAe,KAAK,QACvB,QACC,cAAc,GAAG,MAAM,KAAA,KACvB,CAAC,uBAAuB,GAAG,KAC3B,CAAC,wBAAwB,IAAI,eAAe,CAChD,CAAC,CAAC;CAOF,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,gBAAgB,GAAG,CAAC;CAC7D,MAAM,eACJ,KAAK,SAAS,IACV,WAAW,KAAK,QAAQ,eAAe,KAAK,WAAW,qBAAqB,CAAC,IAC7E,OAAO,KAAK,UAAU,iBAAiB,OAAO,WAAW,qBAAqB,CAAC;CACrF,MAAM,kBACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,gBAAgB,IAAI,eAAe,CAAC,IACtD,OAAO,KAAK,UAAU,MAAM,EAAE;CACpC,MAAM,2BAA2B,gBAAgB,QAAQ,YAAY,YAAY,KAAA,CAAS,CAAC,CAAC;CAC5F,MAAM,sBAAsB,gBAAgB,QAAQ,YAAY,YAAY,KAAK,CAAC,CAAC;CACnF,MAAM,kBACJ,gBAAgB,WAAW,KAAK,2BAA2B,IACvD,QACC,gBAAgB,SAAS,uBAAuB,gBAAgB;CACvE,MAAM,aAAa,YAAY,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,cAAc,YAAY,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACxE,MAAM,SAAS,WAAW,MAAM,QAAQ,WAAW,qBAAqB;CACxE,MAAM,kBAAkB,WAAW,YAAY;CAC/C,MAAM,mBAAmB,WAAW,aAAa;CACjD,MAAM,WAAW,KAAK,SAAS,QAC7B,IAAI,eAAe,QAAQ,OAAO,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG,CAChE;CACA,MAAM,aAAa,OAAO,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,OAAO,cAAc;CAC7E,MAAM,cACJ,KAAK,SAAS,IACV,SAAS,WAAW,KAAK,SACvB,WAAW,QAAQ,IACnB,OACF,WAAW,UAAU;CAC3B,MAAM,YACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,IAAI,MAAM,IAC5B,OAAO,KAAK,UAAU,MAAM,UAAU,CAAC,CAAC,OAAO,cAAc;CACnE,MAAM,UAAoC;EACxC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU,gBAAgB,YAAY;EACtC,mBAAmB,KAAK,SAAS,WAAW;EAC5C,WAAW,WAAW,aAAa;EACnC;EACA;EACA,YAAY,WAAW,iBAAiB,gBAAgB;EACxD;EACA,WAAW,sBAAsB,WAAW,EAAE,aAAa,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,cAAc,OAAO;EACzF,YAAY,OAAO;EACnB,iBAAiB,OAAO,QAAQ,QAAQ,IAAI,MAAM,CAAC,CAAC;EACpD,kBAAkB,OAAO,QAAQ,MAAM,EAAE,cAAc,CAAC,CAAC,CAAC;EAC1D,iBAAiB,OAAO,QAAQ,OAAO,EAAE,aAAa,KAAK,CAAC,CAAC,CAAC;EAC9D;EACA,cAAc,aAAa,SAAS;EACpC,oBAAoB,oBAAoB,MAAM,QAAQ,WAAW,qBAAqB;EACtF,0BAA0B,yBAAyB,MAAM;CAC3D;CAEA,MAAM,SAAmC,CAAC;CAC1C,YAAY,OAAO,YAAY,SAAS,MAAM;CAC9C,aAAa,YAAY,SAAS,MAAM;CACxC,iBAAiB,SAAS,MAAM;CAChC,oBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;CAC3E,iBAAiB,YAAY,SAAS,MAAM;CAC5C,gBAAgB,YAAY,SAAS,MAAM;CAE3C,MAAM,OAAO,UAAU,SAAS,YAAY,MAAM;CAClD,MAAM,SAAS,OAAO,MAAM,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;CAEN,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM,cAAc;EAChC;EACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;EACjF;EACA;EACA;EACA,SAAS,MAAM,WAAW;EAC1B,cAAc,MAAM,gBAAgB;EACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;CAC9D;AACF;AAUA,SAAS,gBACP,MACA,aACA,YACa;CACb,IAAI,aAAa,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,WAAW;CACxE,IAAI,YAAY,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,UAAU;CACtE,OAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;CACxB,IAAI,aACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,WAAW;CAC1F,IAAI,YACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,UAAU;CACzF,OAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,6BACP,OACA,OACsB;CACtB,MAAM,QAAQ;CACd,IAAI,OAAO,OAAO,OAAO,aAAa,GACpC,MAAM,IAAI,gBACR,UAAU,MAAM,2DAClB;CAEF,IAAI,MAAM,iBAAiB,KAAA,KAAa,CAAC,gBAAgB,SAAS,MAAM,YAAY,GAClF,MAAM,IAAI,gBACR,UAAU,MAAM,gCAAgC,gBAAgB,KAAK,IAAI,GAC3E;CAEF,OAAO;AACT;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;CACN,IAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GACnF,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,gBAAgB,WAAW,kBACrC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,cAAc,qBAAqB,WAAW,iBAAiB;CACpF,CAAC;CAEH,IAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;AAEL;AAEA,SAAS,aACP,YACA,SACA,QACM;CACN,IAAI,QAAQ,aAAa,WAAW,eAClC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,WAAW,uBAAuB,WAAW,cAAc;CAChF,CAAC;CAEH,IAAI,QAAQ,eAAe,GACzB,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa;CAClC,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,cAAc,MACrD,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,WAAW,WAAW,aAC7D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,EAAE,KAAK,IAAI,WAAW,WAAW,EAAE;CAC7E,CAAC;CAEH,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,iBACP,SACA,QACM;CACN,IAAI,QAAQ,oBAAoB,MAC9B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QACE,QAAQ,2BAA2B,IAC/B,GAAG,QAAQ,yBAAyB,wDACpC;CACR,CAAC;CAEH,IAAI,QAAQ,sBAAsB,GAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,oBAAoB;CACzC,CAAC;AAEL;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;CACN,IAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAChE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,YAAY,wBAAwB,WAAW,eAAe;CACnF,CAAC;CAEH,IAAI,QAAQ,eAAe,QAAQ,QAAQ,aAAa,WAAW,eACjE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,EAAE,KAAK,IAAI,WAAW,aAAa,EAAE;CAC3F,CAAC;CAEH,IAAI,gBAAgB,CAAC,aAAa,SAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM,QAAQ,aAAa,iBAAiB;EAC5C,QAAQ,aAAa;CACvB,CAAC;AAEL;AAEA,SAAS,iBACP,YACA,SACA,QACM;CACN,IAAI,CAAC,WAAW,uBAAuB;CACvC,IAAI,QAAQ,aAAa,QAAQ,iBAC/B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,gBAAgB;CAC1D,CAAC;AAEL;AAEA,SAAS,gBACP,YACA,SACA,QACM;CACN,IAAI,OAAO,SAAS,WAAW,cAAc,KAAK,QAAQ,gBAAgB,MACxE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,gBAAgB,QAAQ,QAAQ,cAAc,WAAW,gBAC1E,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,EAAE,KAAK,IAAI,WAAW,cAAc,EAAE;CACtF,CAAC;CAEH,IAAI,OAAO,SAAS,WAAW,YAAY,KAAK,QAAQ,cAAc,MACpE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cACtE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,UACP,SACA,YACA,QACyB;CACzB,OAAO;EACL,KACE,UACA,QACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC,GACxE,GAAG,QAAQ,cAAc,sBAAsB,QAAQ,YAAY,SACrE;EACA,KACE,WACA,QACA,QAAQ,aAAa,QAAQ,QAAQ,cAAc,OAC/C,OACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS,GAChD,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS,GACtE;EACA,KACE,eACA,QACA,QAAQ,iBACR,eAAe,IAAI,QAAQ,eAAe,EAAE,oBAAoB,QAAQ,oBAAoB,gBAAgB,QAAQ,0BACtH;EACA,KACE,kBACA,QACA,SAAS,QAAQ,YAAY,WAAW,aAAa,GACrD,eAAe,QAAQ,YAAY,cAAc,IAAI,QAAQ,UAAU,GACzE;EACA,KACE,eACA,QACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ,YACjE,mBAAmB,QAAQ,gBAAgB,GAAG,QAAQ,YACxD;EACA,KACE,cACA,QACA,gBAAgB,SAAS,UAAU,GACnC,eAAe,IAAI,QAAQ,WAAW,EAAE,aAAa,IAAI,QAAQ,SAAS,GAC5E;CACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;CACvB,MAAM,MAAM,OAAO,QAAQ,MAAM,EAAE,SAAS,IAAI;CAMhD,OAAO;EAAE;EAAM,QALA,IAAI,MAAM,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;EACiB,OAAO,UAAU,OAAO,OAAO,QAAQ,KAAK;EAAG;CAAO;AAC/E;AAEA,SAAS,oBAAoB,WAAqE;CAChG,MAAM,SAAuC;EAAE,OAAO;EAAG,KAAK;EAAG,MAAM;EAAG,SAAS;CAAE;CACrF,KAAK,MAAM,YAAY,WAAW,OAAO,SAAS,SAAS,QAAQ;CACnE,OAAO;AACT;AAEA,SAAS,aAAa,WAA+D;CACnF,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;EACnE,IAAI,WAAW,IAAI,WAAW,KAAK;CACrC;CACA,OAAO;AACT;AAEA,SAAS,oBACP,MACA,QACA,WACuC;CACvC,MAAM,MAA6C,CAAC;CACpD,KAAK,MAAM,OAAO,MAKhB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,eACJ,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,YACnD,IAAI,eACJ;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;EAChD,MAAM,eACJ,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,YACvD,MAAM,eACN;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,OAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;CACjG,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,SAAS,QAClB,KAAK,MAAM,OAAO,MAAM,OAAO,CAAC,GAAG;EACjC,MAAM,UAAU,IAAI,sBAAsB;EAC1C,IAAI,YAAY,IAAI,YAAY,KAAK;CACvC;CAEF,OAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;CAC5B,MAAM,MAAkC,CAAC;CACzC,KAAK,MAAM,OAAO,MAChB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,YAAY,IAAI,QAAQ,IAAI;EAClC,IAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;CACrE;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OACzC,IAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;CAGrD,OAAO;AACT;AAEA,SAAS,gBAAgB,UAAsD;CAC7E,MAAM,aAAa,SAAS,QAAQ,YAAgC,YAAY,IAAI;CACpF,IAAI,WAAW,WAAW,GAAG,OAAO;CACpC,OAAO,WAAW,OAAO,OAAO,CAAC,CAAC,SAAS,WAAW;AACxD;AAEA,SAAS,eAAe,KAAgB,WAAmC;CACzE,IAAI,uBAAuB,GAAG,GAAG,OAAO;CACxC,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO,UAAU,KAAA,IAAY,OAAO,SAAS;AAC/C;AAEA,SAAS,uBAAuB,KAAyB;CACvD,OAAO,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB;AAChE;AAEA,SAAS,iBAAiB,OAA6B,WAAmC;CACxF,IAAI,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,WAAW,OAAO;CACjF,IAAI,MAAM,OAAO,OAAO,OAAO;CAC/B,IAAI,eAAe,MAAM,KAAK,GAAG,OAAO,MAAM,SAAS;CACvD,OAAO,MAAM,OAAO,OAAO,OAAO;AACpC;AAEA,SAAS,wBACP,SACkD;CAClD,OAAO,YAAY,YAAY,YAAY,eAAe,YAAY;AACxE;AAEA,SAAS,gBAAgB,SAA4D;CACnF,IAAI,YAAY,aAAa,OAAO;CACpC,IAAI,wBAAwB,OAAO,GAAG,OAAO;AAE/C;AAEA,SAAS,UAAU,MAA4B,OAA8B;CAC3E,OAAO,KACJ,QAAQ,QAAQ,IAAI,aAAa,KAAK,CAAC,CACvC,IAAI,aAAa,CAAC,CAClB,OAAO,cAAc;AAC1B;;;;;;;AAQA,SAAS,cAAc,KAAoC;CACzD,MAAM,QAAQ,mBAAmB,KAAK,IAAI,aAAa,YAAY,YAAY,QAAQ;CACvF,OAAO,eAAe,KAAK,IAAI,QAAQ,KAAA;AACzC;AAEA,SAAS,WAAW,IAAsC;CACxD,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,eAAe,OAAiC;CACvD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,WAAW,GAAkB,GAAiC;CACrE,IAAI,MAAM,QAAQ,MAAM,MAAM,OAAO;CACrC,OAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAoB,QAA+B;CACnE,IAAI,QAAQ,MAAM,OAAO;CACzB,IAAI,UAAU,GAAG,OAAO,OAAO,IAAI,IAAI;CACvC,OAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACe;CACf,MAAM,OAAO,OAAO,SAAS,WAAW,cAAc,IAClD,QAAQ,gBAAgB,OACtB,OACA,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IAC1E;CACJ,MAAM,UAAU,OAAO,SAAS,WAAW,YAAY,IACnD,QAAQ,cAAc,OACpB,OACA,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACtE;CACJ,IAAI,SAAS,QAAQ,YAAY,MAAM,OAAO;CAC9C,OAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;CACR,MAAM,SAAS,sBAAsB,OAAO,IAAI;CAChD,MAAM,aAAa,aAAa,QAAQ,cAAc,cAAc,QAAQ,WAAW,eAAe,QAAQ,YAAY,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS;CAC9L,IAAI,OAAO,WAAW,GAAG,OAAO,GAAG,OAAO,IAAI;CAC9C,OAAO,GAAG,OAAO,IAAI,WAAW,WAAW,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;AAC/E;AAEA,SAAS,IAAI,GAA0B;CACrC,IAAI,MAAM,MAAM,OAAO;CACvB,OAAO,EAAE,QAAQ,CAAC;AACpB"}
1	+ {"version":3,"file":"release-confidence--1iP-d3x.js","names":[],"sources":["../src/promotion-gate.ts","../src/release-confidence.ts"],"sourcesContent":["/*\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\nimport { mulberry32 } from './statistics'\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. Absent, the seed is derived from the\n * compared arms, so the same inputs reproduce the same verdict. /\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 // A ZERO-WIDTH interval is an absence of evidence, not a certainty. Constant\n // arms make every resample identical, so pass/fail data promotes on nothing:\n // [0,0,0] vs [1,1,1] is six samples, clears the `minTotalSamples` floor, and\n // yields [1, 1] — an \"ADVANCE\" carrying no information about how far the\n // estimate could be wrong. Same rule as `pairedDeltaTest`, one estimator over.\n if (!Number.isFinite(ciLower) \|\| !Number.isFinite(ciUpper) \|\| ciLower === ciUpper) {\n verdict = 'INCONCLUSIVE'\n } else 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/** 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","/\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 { HeldOutGateDecision } from './held-out-gate'\nimport { isRealnessGated, observedSplitScore } from './rollout/reward'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\nimport { summarizeNumberSeries } from './statistics'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/* Severity of an actionable finding attached to a run/trace. /\ntype 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?: HeldOutGateDecision \| 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 Runs excluded from `passRate` because the authenticity gate flagged them as\n * gamed. Surfaced, never silent: a release whose pass rate is computed over a\n * shrunken denominator has to say by how much, or the exclusion is just a\n * different way of hiding the same runs.\n /\n realnessGatedRuns: 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: HeldOutGateDecision \| 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 // Realness-gated runs are EXCLUDED from the pass rate — numerator AND\n // denominator — and the count of what was dropped ships beside the rate as\n // `metrics.realnessGatedRuns`. Counting a gamed run as a pass made faking a\n // success the cheapest way to improve a release scorecard; scoring it 0\n // instead would be the other error, silently deflating the rate with a run\n // the gate says carries no usable verdict at all.\n const honestRuns = runs.filter((run) => !isRealnessGated(run))\n const passOutcomes =\n runs.length > 0\n ? honestRuns.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.usd === null ? [] : [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 realnessGatedRuns: runs.length - honestRuns.length,\n meanScore: meanOrNull(scoreUniverse),\n searchMeanScore,\n holdoutMeanScore,\n overfitGap: diffOrNull(searchMeanScore, holdoutMeanScore),\n meanCostUsd,\n p95WallMs: summarizeNumberSeries(wallTimes, { percentiles: [95] })?.percentiles?.[95] ?? null,\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: `${metrics.terminalFailureRuns} run(s) ended failed, cancelled, or incomplete.`,\n })\n }\n}\n\nfunction checkGeneralization(\n gateDecision: HeldOutGateDecision \| null,\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n issues.push({\n axis: 'generalization',\n severity: 'critical',\n code: 'few_holdout_runs',\n detail: `${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 // Ungated: a failure-mode census counts what the runs REPORTED, which is\n // the only way a gamed run's inflated score is visible at all. It is not a\n // promotion number — `passRate` is, and that one excludes gated runs and\n // publishes the excluded count as `metrics.realnessGatedRuns`.\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\n/\n RAW and split-exact (`observedSplitScore`): this feeds the per-split means,\n * the overfit gap, and the pass threshold — descriptions of what the runs\n * reported. A gamed run inflating them is visible next to `realnessGatedRuns`,\n * and the promotion number (`passRate`) excludes gated runs entirely.\n */\nfunction runSplitScore(run: RunRecord): number \| undefined {\n const score = observedSplitScore(run, run.splitTag === 'holdout' ? 'holdout' : 'search')\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 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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,SAAgB,YACd,UACA,WACA,UAA4B,CAAC,GACZ;CACjB,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,WAAW,QAAQ,mBAAmB;CAC5C,MAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;CAEpE,MAAM,eAAe,KAAK,QAAQ;CAClC,MAAM,gBAAgB,KAAK,SAAS;CACpC,MAAM,QAAQ,gBAAgB;CAE9B,IACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GAErB,OAAO;EACL;EACA;EACA;EACA,SAAS;EACT,SAAS;EACT,YAAY;EACZ;EACA,SAAS;CACX;CAGF,MAAM,SAAmB,IAAI,MAAM,UAAU;CAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,UAAU,GAAG;EACxC,MAAM,YAAY,SAAS,WAAW,GAAG;EACzC,OAAO,KAAK,KAAK,SAAS,IAAI,KAAK,SAAS;CAC9C;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;CACpD,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;CAC5D,MAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ;CAC3C,MAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ;CAExD,IAAI;CAMJ,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,CAAC,OAAO,SAAS,OAAO,KAAK,YAAY,SACxE,UAAU;MACL,IAAI,UAAU,GAAG,UAAU;MAC7B,IAAI,UAAU,GAAG,UAAU;MAC3B,IAAI,SAAS,GAAG,UAAU;MAC1B,UAAU;CAEf,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,IAAI,KAAK;CACzB,OAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;CAC3D,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM;CAC5E,OAAO;AACT;;AAGA,SAAS,SAAS,GAAa,GAAqB;CAClD,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;EAC5B,MAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;EACjC,MAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;EACxC,KAAK,MAAM,QAAQ,OAAO;GACxB,KAAK;GACL,IAAI,KAAK,KAAK,GAAG,QAAQ;EAC3B;CACF;CACA,OAAO,MAAM;AACf;;;;AAiCA,eAAsB,gBACpB,MACkF;CAClF,MAAM,cAAc,KAAK,oBAAoB;CAC7C,MAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;CACnF,MAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;CAMrF,OAAO;EACL,GANS,YAAY,gBAAgB,iBAAiB;GACtD,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,GAAI,KAAK,eAAe,KAAA,IAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;GACvE,GAAI,KAAK,SAAS,KAAA,IAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;EACvD,CAEM;EACJ,iBAAiB,eAAe;EAChC,kBAAkB,gBAAgB;CACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;CACnB,MAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;CAClD,IAAI,OAAO;CACX,eAAe,SAAwB;EACrC,OAAO,MAAM;GACX,MAAM,IAAI;GACV,IAAI,KAAK,QAAQ,QAAQ;GACzB,MAAM,IAAI,MAAM,MAAM,QAAQ,EAAG;GACjC,QAAQ,KAAK,OAAO,SAAS,CAAC,IAAI,IAAI;EACxC;CACF;CACA,MAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,SAAS,OAAO,CAAC,CAAC;CAClF,OAAO;AACT;;AC/DA,MAAM,qBAA4D;CAChE,eAAe;CACf,kBAAkB;CAClB,eAAe;CACf,gBAAgB;CAChB,gBAAgB;CAChB,aAAa;CACb,cAAc;CACd,eAAe;CACf,gBAAgB,OAAO;CACvB,cAAc,OAAO;CACrB,uBAAuB;CACvB,uBAAuB;AACzB;AAEA,SAAgB,0BACd,OAC4B;CAC5B,MAAM,aAAa;EAAE,GAAG;EAAoB,GAAG,MAAM;CAAW;CAChE,MAAM,cAAc,MAAM,eAAe;CACzC,MAAM,OAAO,iBACV,MAAM,QAAQ,CAAC,EAAA,CAAG,IAAI,iBAAiB,GACxC,aACA,MAAM,UACR;CACA,MAAM,SAAS,sBACZ,MAAM,UAAU,CAAC,EAAA,CAAG,IAAI,4BAA4B,GACrD,aACA,MAAM,UACR;CACA,MAAM,YAAY,MAAM,aAAa,CAAC;CACtC,MAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;CAChE,MAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;CAC/E,MAAM,eAAe,UAAU,MAAM,QAAQ;CAC7C,MAAM,gBAAgB,UAAU,MAAM,SAAS;CAC/C,MAAM,YAAY,KAAK,IAAI,aAAa,CAAC,CAAC,OAAO,cAAc;CAC/D,MAAM,cAAc,OAAO,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,OAAO,cAAc;CACpE,MAAM,gBAAgB,KAAK,SAAS,IAAI,YAAY;CACpD,MAAM,cAAc,KAAK,QAAQ,QAAQ,cAAc,GAAG,MAAM,KAAA,CAAS;CACzE,MAAM,eAAe,KAAK,QACvB,QACC,cAAc,GAAG,MAAM,KAAA,KACvB,CAAC,uBAAuB,GAAG,KAC3B,CAAC,wBAAwB,IAAI,eAAe,CAChD,CAAC,CAAC;CAOF,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,gBAAgB,GAAG,CAAC;CAC7D,MAAM,eACJ,KAAK,SAAS,IACV,WAAW,KAAK,QAAQ,eAAe,KAAK,WAAW,qBAAqB,CAAC,IAC7E,OAAO,KAAK,UAAU,iBAAiB,OAAO,WAAW,qBAAqB,CAAC;CACrF,MAAM,kBACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,gBAAgB,IAAI,eAAe,CAAC,IACtD,OAAO,KAAK,UAAU,MAAM,EAAE;CACpC,MAAM,2BAA2B,gBAAgB,QAAQ,YAAY,YAAY,KAAA,CAAS,CAAC,CAAC;CAC5F,MAAM,sBAAsB,gBAAgB,QAAQ,YAAY,YAAY,KAAK,CAAC,CAAC;CACnF,MAAM,kBACJ,gBAAgB,WAAW,KAAK,2BAA2B,IACvD,QACC,gBAAgB,SAAS,uBAAuB,gBAAgB;CACvE,MAAM,aAAa,YAAY,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,cAAc,YAAY,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACxE,MAAM,SAAS,WAAW,MAAM,QAAQ,WAAW,qBAAqB;CACxE,MAAM,kBAAkB,WAAW,YAAY;CAC/C,MAAM,mBAAmB,WAAW,aAAa;CACjD,MAAM,WAAW,KAAK,SAAS,QAC7B,IAAI,eAAe,QAAQ,OAAO,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG,CAChE;CACA,MAAM,aAAa,OAAO,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,OAAO,cAAc;CAC7E,MAAM,cACJ,KAAK,SAAS,IACV,SAAS,WAAW,KAAK,SACvB,WAAW,QAAQ,IACnB,OACF,WAAW,UAAU;CAC3B,MAAM,YACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,IAAI,MAAM,IAC5B,OAAO,KAAK,UAAU,MAAM,UAAU,CAAC,CAAC,OAAO,cAAc;CACnE,MAAM,UAAoC;EACxC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU,gBAAgB,YAAY;EACtC,mBAAmB,KAAK,SAAS,WAAW;EAC5C,WAAW,WAAW,aAAa;EACnC;EACA;EACA,YAAY,WAAW,iBAAiB,gBAAgB;EACxD;EACA,WAAW,sBAAsB,WAAW,EAAE,aAAa,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,cAAc,OAAO;EACzF,YAAY,OAAO;EACnB,iBAAiB,OAAO,QAAQ,QAAQ,IAAI,MAAM,CAAC,CAAC;EACpD,kBAAkB,OAAO,QAAQ,MAAM,EAAE,cAAc,CAAC,CAAC,CAAC;EAC1D,iBAAiB,OAAO,QAAQ,OAAO,EAAE,aAAa,KAAK,CAAC,CAAC,CAAC;EAC9D;EACA,cAAc,aAAa,SAAS;EACpC,oBAAoB,oBAAoB,MAAM,QAAQ,WAAW,qBAAqB;EACtF,0BAA0B,yBAAyB,MAAM;CAC3D;CAEA,MAAM,SAAmC,CAAC;CAC1C,YAAY,OAAO,YAAY,SAAS,MAAM;CAC9C,aAAa,YAAY,SAAS,MAAM;CACxC,iBAAiB,SAAS,MAAM;CAChC,oBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;CAC3E,iBAAiB,YAAY,SAAS,MAAM;CAC5C,gBAAgB,YAAY,SAAS,MAAM;CAE3C,MAAM,OAAO,UAAU,SAAS,YAAY,MAAM;CAClD,MAAM,SAAS,OAAO,MAAM,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;CAEN,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM,cAAc;EAChC;EACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;EACjF;EACA;EACA;EACA,SAAS,MAAM,WAAW;EAC1B,cAAc,MAAM,gBAAgB;EACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;CAC9D;AACF;AAUA,SAAS,gBACP,MACA,aACA,YACa;CACb,IAAI,aAAa,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,WAAW;CACxE,IAAI,YAAY,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,UAAU;CACtE,OAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;CACxB,IAAI,aACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,WAAW;CAC1F,IAAI,YACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,UAAU;CACzF,OAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,6BACP,OACA,OACsB;CACtB,MAAM,QAAQ;CACd,IAAI,OAAO,OAAO,OAAO,aAAa,GACpC,MAAM,IAAI,gBACR,UAAU,MAAM,2DAClB;CAEF,IAAI,MAAM,iBAAiB,KAAA,KAAa,CAAC,gBAAgB,SAAS,MAAM,YAAY,GAClF,MAAM,IAAI,gBACR,UAAU,MAAM,gCAAgC,gBAAgB,KAAK,IAAI,GAC3E;CAEF,OAAO;AACT;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;CACN,IAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GACnF,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,gBAAgB,WAAW,kBACrC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,cAAc,qBAAqB,WAAW,iBAAiB;CACpF,CAAC;CAEH,IAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;AAEL;AAEA,SAAS,aACP,YACA,SACA,QACM;CACN,IAAI,QAAQ,aAAa,WAAW,eAClC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,WAAW,uBAAuB,WAAW,cAAc;CAChF,CAAC;CAEH,IAAI,QAAQ,eAAe,GACzB,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa;CAClC,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,cAAc,MACrD,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,WAAW,WAAW,aAC7D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,EAAE,KAAK,IAAI,WAAW,WAAW,EAAE;CAC7E,CAAC;CAEH,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,iBACP,SACA,QACM;CACN,IAAI,QAAQ,oBAAoB,MAC9B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QACE,QAAQ,2BAA2B,IAC/B,GAAG,QAAQ,yBAAyB,wDACpC;CACR,CAAC;CAEH,IAAI,QAAQ,sBAAsB,GAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,oBAAoB;CACzC,CAAC;AAEL;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;CACN,IAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAChE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,YAAY,wBAAwB,WAAW,eAAe;CACnF,CAAC;CAEH,IAAI,QAAQ,eAAe,QAAQ,QAAQ,aAAa,WAAW,eACjE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,EAAE,KAAK,IAAI,WAAW,aAAa,EAAE;CAC3F,CAAC;CAEH,IAAI,gBAAgB,CAAC,aAAa,SAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM,QAAQ,aAAa,iBAAiB;EAC5C,QAAQ,aAAa;CACvB,CAAC;AAEL;AAEA,SAAS,iBACP,YACA,SACA,QACM;CACN,IAAI,CAAC,WAAW,uBAAuB;CACvC,IAAI,QAAQ,aAAa,QAAQ,iBAC/B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,gBAAgB;CAC1D,CAAC;AAEL;AAEA,SAAS,gBACP,YACA,SACA,QACM;CACN,IAAI,OAAO,SAAS,WAAW,cAAc,KAAK,QAAQ,gBAAgB,MACxE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,gBAAgB,QAAQ,QAAQ,cAAc,WAAW,gBAC1E,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,EAAE,KAAK,IAAI,WAAW,cAAc,EAAE;CACtF,CAAC;CAEH,IAAI,OAAO,SAAS,WAAW,YAAY,KAAK,QAAQ,cAAc,MACpE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cACtE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,UACP,SACA,YACA,QACyB;CACzB,OAAO;EACL,KACE,UACA,QACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC,GACxE,GAAG,QAAQ,cAAc,sBAAsB,QAAQ,YAAY,SACrE;EACA,KACE,WACA,QACA,QAAQ,aAAa,QAAQ,QAAQ,cAAc,OAC/C,OACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS,GAChD,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS,GACtE;EACA,KACE,eACA,QACA,QAAQ,iBACR,eAAe,IAAI,QAAQ,eAAe,EAAE,oBAAoB,QAAQ,oBAAoB,gBAAgB,QAAQ,0BACtH;EACA,KACE,kBACA,QACA,SAAS,QAAQ,YAAY,WAAW,aAAa,GACrD,eAAe,QAAQ,YAAY,cAAc,IAAI,QAAQ,UAAU,GACzE;EACA,KACE,eACA,QACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ,YACjE,mBAAmB,QAAQ,gBAAgB,GAAG,QAAQ,YACxD;EACA,KACE,cACA,QACA,gBAAgB,SAAS,UAAU,GACnC,eAAe,IAAI,QAAQ,WAAW,EAAE,aAAa,IAAI,QAAQ,SAAS,GAC5E;CACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;CACvB,MAAM,MAAM,OAAO,QAAQ,MAAM,EAAE,SAAS,IAAI;CAMhD,OAAO;EAAE;EAAM,QALA,IAAI,MAAM,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;EACiB,OAAO,UAAU,OAAO,OAAO,QAAQ,KAAK;EAAG;CAAO;AAC/E;AAEA,SAAS,oBAAoB,WAAqE;CAChG,MAAM,SAAuC;EAAE,OAAO;EAAG,KAAK;EAAG,MAAM;EAAG,SAAS;CAAE;CACrF,KAAK,MAAM,YAAY,WAAW,OAAO,SAAS,SAAS,QAAQ;CACnE,OAAO;AACT;AAEA,SAAS,aAAa,WAA+D;CACnF,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;EACnE,IAAI,WAAW,IAAI,WAAW,KAAK;CACrC;CACA,OAAO;AACT;AAEA,SAAS,oBACP,MACA,QACA,WACuC;CACvC,MAAM,MAA6C,CAAC;CACpD,KAAK,MAAM,OAAO,MAKhB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,eACJ,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,YACnD,IAAI,eACJ;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;EAChD,MAAM,eACJ,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,YACvD,MAAM,eACN;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,OAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;CACjG,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,SAAS,QAClB,KAAK,MAAM,OAAO,MAAM,OAAO,CAAC,GAAG;EACjC,MAAM,UAAU,IAAI,sBAAsB;EAC1C,IAAI,YAAY,IAAI,YAAY,KAAK;CACvC;CAEF,OAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;CAC5B,MAAM,MAAkC,CAAC;CACzC,KAAK,MAAM,OAAO,MAChB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,YAAY,IAAI,QAAQ,IAAI;EAClC,IAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;CACrE;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OACzC,IAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;CAGrD,OAAO;AACT;AAEA,SAAS,gBAAgB,UAAsD;CAC7E,MAAM,aAAa,SAAS,QAAQ,YAAgC,YAAY,IAAI;CACpF,IAAI,WAAW,WAAW,GAAG,OAAO;CACpC,OAAO,WAAW,OAAO,OAAO,CAAC,CAAC,SAAS,WAAW;AACxD;AAEA,SAAS,eAAe,KAAgB,WAAmC;CACzE,IAAI,uBAAuB,GAAG,GAAG,OAAO;CACxC,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO,UAAU,KAAA,IAAY,OAAO,SAAS;AAC/C;AAEA,SAAS,uBAAuB,KAAyB;CACvD,OAAO,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB;AAChE;AAEA,SAAS,iBAAiB,OAA6B,WAAmC;CACxF,IAAI,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,WAAW,OAAO;CACjF,IAAI,MAAM,OAAO,OAAO,OAAO;CAC/B,IAAI,eAAe,MAAM,KAAK,GAAG,OAAO,MAAM,SAAS;CACvD,OAAO,MAAM,OAAO,OAAO,OAAO;AACpC;AAEA,SAAS,wBACP,SACkD;CAClD,OAAO,YAAY,YAAY,YAAY,eAAe,YAAY;AACxE;AAEA,SAAS,gBAAgB,SAA4D;CACnF,IAAI,YAAY,aAAa,OAAO;CACpC,IAAI,wBAAwB,OAAO,GAAG,OAAO;AAE/C;AAEA,SAAS,UAAU,MAA4B,OAA8B;CAC3E,OAAO,KACJ,QAAQ,QAAQ,IAAI,aAAa,KAAK,CAAC,CACvC,IAAI,aAAa,CAAC,CAClB,OAAO,cAAc;AAC1B;;;;;;;AAQA,SAAS,cAAc,KAAoC;CACzD,MAAM,QAAQ,mBAAmB,KAAK,IAAI,aAAa,YAAY,YAAY,QAAQ;CACvF,OAAO,eAAe,KAAK,IAAI,QAAQ,KAAA;AACzC;AAEA,SAAS,WAAW,IAAsC;CACxD,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,eAAe,OAAiC;CACvD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,WAAW,GAAkB,GAAiC;CACrE,IAAI,MAAM,QAAQ,MAAM,MAAM,OAAO;CACrC,OAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAoB,QAA+B;CACnE,IAAI,QAAQ,MAAM,OAAO;CACzB,IAAI,UAAU,GAAG,OAAO,OAAO,IAAI,IAAI;CACvC,OAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACe;CACf,MAAM,OAAO,OAAO,SAAS,WAAW,cAAc,IAClD,QAAQ,gBAAgB,OACtB,OACA,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IAC1E;CACJ,MAAM,UAAU,OAAO,SAAS,WAAW,YAAY,IACnD,QAAQ,cAAc,OACpB,OACA,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACtE;CACJ,IAAI,SAAS,QAAQ,YAAY,MAAM,OAAO;CAC9C,OAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;CACR,MAAM,SAAS,sBAAsB,OAAO,IAAI;CAChD,MAAM,aAAa,aAAa,QAAQ,cAAc,cAAc,QAAQ,WAAW,eAAe,QAAQ,YAAY,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS;CAC9L,IAAI,OAAO,WAAW,GAAG,OAAO,GAAG,OAAO,IAAI;CAC9C,OAAO,GAAG,OAAO,IAAI,WAAW,WAAW,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;AAC/E;AAEA,SAAS,IAAI,GAA0B;CACrC,IAAI,MAAM,MAAM,OAAO;CACvB,OAAO,EAAE,QAAQ,CAAC;AACpB"}

package/dist/{render-pfAZ7IR4.js → render-CFrJ7M0o.js} RENAMED Viewed

@@ -1,6 +1,6 @@
 import { a as hashCanonical } from "./canonical-CicJlINA.js";
 import { c as unscreenedRewardFields } from "./reward-P0VD_Aft.js";
-import { o as summarizeNumberSeries } from "./descriptive-cPwITVfy.js";
+import { c as summarizeNumberSeries } from "./descriptive-BTzLL-aj.js";
 import { t as roundDecimal } from "./round-decimal-ZmjAQPo1.js";
 import { S as licenseClassOfName } from "./schema-YUIXT8iW.js";
 import { _ as showMeasured, a as readTerminalRecord, c as parseJsonRecord, d as recordOrNull, f as workerSourceKey, g as isUnavailable, h as SUPERVISOR_RUN_SCHEMA, l as parseJsonl, n as supervisorRunRolloutLinesFromFacts, o as asRecord, p as NO_SOURCE_LIMITS, s as nonEmptyString, u as parseSupervisorTree, v as unavailable, y as expandJournalEventGroup } from "./rollout-nodes-BHR9Q5jP.js";
@@ -682,16 +682,17 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 * analyst reads communication and steer edges from it.
 */
 const COORDINATION_FILE = "coordination-log.jsonl";
+const readFileText = (path) => readFile(path, "utf8");
 /** File text, or null when the file does not exist. Any other read error throws. */
-async function readMaybe(path) {
-	return readFile(path, "utf8").catch((error) => {
+async function readMaybe(path, read = readFileText) {
+	return read(path).catch((error) => {
 		if (typeof error === "object" && error !== null && "code" in error && error.code === "ENOENT") return null;
 		throw error;
 	});
 }
 /** The run's coordination log, or null when Runtime wrote none. */
-function readCoordinationLog(runDir) {
-	return readMaybe(join(runDir, COORDINATION_FILE));
+function readCoordinationLog(runDir, read = readFileText) {
+	return readMaybe(join(runDir, COORDINATION_FILE), read);
 }
 // src/supervisor-run/runtime-source-limits.ts
 function spendRecord(value) {
@@ -1015,10 +1016,10 @@ const HARNESS_SESSION_FILE = {
 	opencode: /(^|\/)(\.local\/share\/opencode\/(opencode\.db|storage\/session\/.+\.json|export\/.+\.json)|\.opencode\/sessions\/.+)$/u
 };
 /** Session files the receipt's transcript blob lists, or undefined when it cannot say. */
-async function receiptSessionFiles(file) {
+async function receiptSessionFiles(file, read) {
 	let descriptor;
 	try {
-		descriptor = JSON.parse(await readFile(file, "utf8"));
+		descriptor = JSON.parse(await read(file));
 	} catch {
 		return;
 	}
@@ -1055,7 +1056,7 @@ function nativeTraceSummary(workers, root) {
 	const gaps = [count("partial") > 0 ? `${count("partial")} only a partial copy` : null, count("no-session-file") > 0 ? `${count("no-session-file")} a receipt with no session file` : null].filter((gap) => gap !== null);
 	return `per node — ${count("complete")} of ${workers.length} workers carry a complete native session receipt (workers[].nativeSession)${gaps.length > 0 ? `; ${gaps.join("; ")}` : ""}; root: ${rootText} (result.json rootHarnessTranscript)`;
 }
-async function workspaceCapturesFromOutput(runDir, ref) {
+async function workspaceCapturesFromOutput(runDir, ref, read) {
 	const outRef = typeof ref === "string" ? ref : null;
 	const blobPath = blobFile(runDir, ref);
 	const missing = (projectionGap) => ({
@@ -1070,7 +1071,7 @@ async function workspaceCapturesFromOutput(runDir, ref) {
 	});
 	if (blobPath === null) return [missing("Runtime output has no valid content address")];
 	if (!await isFile(blobPath)) return [missing("Runtime output blob is missing or not a file")];
-	const raw = await readMaybe(blobPath);
+	const raw = await readMaybe(blobPath, read);
 	if (raw === null) return [missing("Runtime output blob is missing")];
 	let output;
 	try {
@@ -1146,7 +1147,7 @@ async function workspaceCapturesFromOutput(runDir, ref) {
 * Blob files are checked on disk: a receipt that names a blob the directory no longer holds is
 * not a retained transcript.
 */
-async function nodeTranscripts(runDir, events, nodeIds, root) {
+async function nodeTranscripts(runDir, events, nodeIds, read, root) {
 	const facts = /* @__PURE__ */ new Map();
 	const entry = (id) => {
 		let found = facts.get(id);
@@ -1193,14 +1194,14 @@ async function nodeTranscripts(runDir, events, nodeIds, root) {
 				retained.push(file);
 				if (turn) retainedTurns += 1;
 			}
-			workspaceCaptures.push(...await workspaceCapturesFromOutput(runDir, ref));
+			workspaceCaptures.push(...await workspaceCapturesFromOutput(runDir, ref, read));
 		}
 		let nativeSession = null;
 		if (fact.receipt !== null) {
 			if (fact.receipt.status === "available") {
 				const file = blobFile(runDir, fact.receipt.transcriptRef);
 				if (file !== null && await isFile(file)) {
-					const sessionFiles = await receiptSessionFiles(file);
+					const sessionFiles = await receiptSessionFiles(file, read);
 					const coverageComplete = fact.receipt.coverageComplete;
 					nativeSession = {
 						status: "available",
@@ -1230,11 +1231,6 @@ async function nodeTranscripts(runDir, events, nodeIds, root) {
 	return out;
 }
 /**
-* Sources for a run dir whose spawn journal does not exist. Mirrors the
-* absent shape the reader returns for a missing store: every
-* journal-dependent metric downstream reads `unavailable`, never 0.
-*/
-/**
 * The run identity a terminal record names when no journal exists yet: the
 * settle record's `tree.root`, else the failure record's `runId`. Both are
 * Runtime's own `runId`, so the report names the run the record is about
@@ -1290,12 +1286,13 @@ function absentRuntimeSupervisorRun(runDir, resultText, failureText, instanceId)
 * verdict.
 */
 async function readRuntimeSupervisorRun(runDir, opts = {}) {
+	const read = opts.readFile ?? readFileText;
 	const journalPath = join(runDir, JOURNAL_FILE);
-	const rawJournal = opts.strict === true ? await readFile(journalPath, "utf8") : await readMaybe(journalPath);
+	const rawJournal = opts.strict === true ? await read(journalPath) : await readMaybe(journalPath, read);
 	const resultPath = join(runDir, RESULT_FILE);
 	const failurePath = join(runDir, FAILURE_FILE);
-	const resultText = await readMaybe(resultPath);
-	const failureText = await readMaybe(failurePath);
+	const resultText = await readMaybe(resultPath, read);
+	const failureText = await readMaybe(failurePath, read);
 	const failure = parseOptionalRecord(failureText, failurePath);
 	assertFailureRecord(failure, failurePath);
 	const result = parseOptionalRecord(resultText, resultPath);
@@ -1304,7 +1301,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 	assertResultMatchesJournal(normalized.root, result, resultPath);
 	const childSpawns = normalized.events.filter((event) => event.kind === "spawned" && nonEmptyString(event.id) !== null).filter((event) => event.id !== normalized.root);
 	const workerIds = new Set(childSpawns.map((event) => nonEmptyString(event.id)).filter((id) => id !== null));
-	const transcripts = await nodeTranscripts(runDir, normalized.events, /* @__PURE__ */ new Set([normalized.root, ...workerIds]), {
+	const transcripts = await nodeTranscripts(runDir, normalized.events, /* @__PURE__ */ new Set([normalized.root, ...workerIds]), read, {
 		id: normalized.root,
 		receipt: recordOrNull(result?.rootHarnessTranscript)
 	});
@@ -1346,7 +1343,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 		judgeSource: null,
 		patch: null,
 		driverLog: null,
-		coordinationLog: await readCoordinationLog(runDir),
+		coordinationLog: await readCoordinationLog(runDir, read),
 		harnessWorkerTokens: null,
 		harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
 		limits: sourceLimits(normalized.root, normalized.events, workerIds),
@@ -1365,10 +1362,13 @@ async function isRuntimeSupervisorRunDir(runDir) {
 	return false;
 }
 async function isFile(path) {
-	return stat(path).then((entry) => entry.isFile()).catch((error) => {
-		if (typeof error === "object" && error !== null && "code" in error && (error.code === "ENOENT" || error.code === "ENOTDIR")) return false;
+	try {
+		return (await stat(path)).isFile();
+	} catch (error) {
+		const code = error !== null && typeof error === "object" && "code" in error ? error.code : null;
+		if (code === "ENOENT" || code === "ENOTDIR") return false;
 		throw error;
-	});
+	}
 }
 // src/supervisor-run/discovery-rollouts.ts
 /**
@@ -1866,4 +1866,4 @@ function renderSupervisorRollupMarkdown(rollup, title = "Round rollup") {
 }
 export { findSupervisorRunDirs as a, discoveryRunRolloutLines as c, analyzeSupervisorRunSources as d, parsePatch as f, analyzeSupervisorRun as i, isRuntimeSupervisorRunDir as l, renderSupervisorRunHeadline as n, discoveryPlayCheckReward as o, renderSupervisorRunMarkdown as r, discoveryPursuitOf as s, renderSupervisorRollupMarkdown as t, readRuntimeSupervisorRun as u };
-//# sourceMappingURL=render-pfAZ7IR4.js.map
+//# sourceMappingURL=render-CFrJ7M0o.js.map