npm - @tangle-network/agent-eval - Versions diffs - 0.182.0 → 0.184.0 - Mend

@tangle-network/agent-eval 0.182.0 → 0.184.0

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 (227) hide show

package/dist/{run-record-DPa2PeHx.js.map → run-record-BhMVsHR1.js.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"run-record-DPa2PeHx.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.ts"],"sourcesContent":["import {\n type PairedBootstrapOptions,\n type PairedBootstrapResult,\n pairedBootstrap,\n pairedSignTest,\n} from './statistics'\n\nexport interface PairedDeltaTestOptions extends PairedBootstrapOptions {\n /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. /\n threshold?: number\n /* Caller-required paired observations. The exact test may impose a higher minimum. /\n minPairs?: number\n}\n\nexport interface PairedDeltaTestResult {\n bootstrap: PairedBootstrapResult\n method: 'bootstrap-ci' \| 'exact-sign'\n /* Exact one-sided p-value below the bootstrap minimum; otherwise null. /\n pValue: number \| null\n /* Effective observation minimum after accounting for confidence. /\n minimumPairs: number\n sufficient: boolean\n /\n The bootstrap interval has zero width (or is non-finite), so it carries no\n * information about how far the estimate could be wrong and cannot support a\n * decision in either direction. See {@link pairedDeltaTest}.\n /\n indeterminate: boolean\n significant: boolean\n}\n\n/* Smallest all-positive sample that can clear a one-sided exact sign test. /\nexport function minimumPairsForPairedDeltaTest(confidence = 0.95): number {\n if (!Number.isFinite(confidence) \|\| confidence <= 0 \|\| confidence >= 1) {\n throw new Error(\n `minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`,\n )\n }\n const oneSidedAlpha = (1 - confidence) / 2\n return Math.ceil(Math.log2(1 / oneSidedAlpha))\n}\n\n/\n Tests whether a paired candidate-minus-baseline delta clears a threshold.\n \n At 20 or more pairs, the percentile bootstrap lower bound carries the\n * decision. Below that point the interval is descriptive only, so the function\n * switches to a pre-registered one-sided exact sign test. The exact path is\n * deliberately conservative: it requires both a point estimate above the\n * threshold and enough consistently positive paired differences.\n \n ## A zero-width interval is never significant\n \n When every paired delta is identical the resample distribution is a point\n * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n\n * identical deltas of g. Neither says the effect is certain — both say the\n * sample carries no information about how far the estimate could be wrong, and\n * `low > threshold` then answers on the point estimate alone. It fails in both\n * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a\n * tie-dominated pass/fail comparison laundered a regression into a\n * noninferiority pass, and `[g, g]` clears every threshold below g with no\n * spread behind it. Under a bounded asymmetric null whose true mean paired\n * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —\n * every sample that misses the drop is exactly that shape, and deciding on\n * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.\n \n So `indeterminate` is reported and `significant` is false whenever the\n * interval has zero width, on BOTH paths: at small n the exact sign test is a\n * test of the MEDIAN and a zero-spread sample is precisely where it stops\n * saying anything about the mean the caller is thresholding.\n \n `threshold` may be negative — that is a noninferiority margin, and it is the\n * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome\n * the percentile bootstrap is not a valid interval at a nonzero margin at all;\n * use {@link decidePairedPromotion}, which routes those to Tango's score\n * interval, rather than thresholding this function's bootstrap directly.\n /\nexport function pairedDeltaTest(\n before: number[],\n after: number[],\n options: PairedDeltaTestOptions = {},\n): PairedDeltaTestResult {\n const threshold = options.threshold ?? 0\n if (!Number.isFinite(threshold)) {\n throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`)\n }\n const confidence = options.confidence ?? 0.95\n const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n const requestedMinimum = options.minPairs ?? exactMinimum\n if (!Number.isInteger(requestedMinimum) \|\| requestedMinimum < 1) {\n throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`)\n }\n const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n const bootstrap = pairedBootstrap(before, after, options)\n const sufficient = bootstrap.n >= minimumPairs\n // A point-mass resample distribution is an absence of evidence, not a\n // certainty, and it is the shape that clears every negative threshold. Both\n // paths refuse it: see the \"zero-width\" section of this function's docs.\n const indeterminate =\n bootstrap.n > 0 &&\n (!Number.isFinite(bootstrap.low) \|\|\n !Number.isFinite(bootstrap.high) \|\|\n bootstrap.low === bootstrap.high)\n\n if (bootstrap.gateEligible) {\n return {\n bootstrap,\n method: 'bootstrap-ci',\n pValue: null,\n minimumPairs,\n sufficient,\n indeterminate,\n significant: sufficient && !indeterminate && bootstrap.low > threshold,\n }\n }\n\n const differences = before.map((value, index) => after[index]! - value - threshold)\n const exact = pairedSignTest(differences, 'greater')\n const estimate = options.statistic === 'mean' ? bootstrap.mean : bootstrap.median\n return {\n bootstrap,\n method: 'exact-sign',\n pValue: exact.pValue,\n minimumPairs,\n sufficient,\n indeterminate,\n significant:\n sufficient &&\n !indeterminate &&\n estimate > threshold &&\n exact.pValue <= (1 - bootstrap.confidence) / 2,\n }\n}\n","/\n @module\n * ONE rule for \"does this paired interval clear a promotion threshold\".\n \n The rule below was derived on `HeldOutGate` (#479) after the same estimator\n * bug shipped twice. It then turned out that a SECOND gate — the composable\n * `heldOutGate`, plus everything else routed through `heldoutSignificance` —\n * still carried the original defect, because the rule had been written into one\n * gate's method body rather than into a shared function. Two copies of a\n * statistical rule is how a defect survives in one of them, so there is now\n * exactly one copy and both gates call it.\n \n Three things the rule does that a bare `pairedBootstrap(...).low > threshold`\n * does not:\n \n 1. Two-point (pass/fail) outcomes decide on Tango's SCORE interval. On a\n * pass/fail eval the paired delta vector is dominated by ties, so the\n * bootstrap of the mean is a resample of a lattice with three atoms and its\n * percentile interval is not valid at a nonzero margin. The score interval\n * (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under\n * each hypothesised margin instead of fixing it at the observed value, which\n * is the only construction that stays a confidence interval as the margin\n * moves off zero — the regime every noninferiority threshold lives in.\n * Measured on the composable gate before this change, at a true risk\n * difference sitting exactly on the production caller's -0.05 margin and a\n * nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.\n * 2. McNemar's exact test holds a VETO at every non-negative threshold.\n * Redundant with the interval by construction and kept anyway, so that\n * swapping the estimator for one without that duality cannot silently\n * reintroduce \"promotes what the exact test refuses\". Witness: n = 6, b = 5,\n * c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs\n * (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE\n * threshold is a noninferiority question, which McNemar's test of \"no\n * difference\" is not the right test for, so the veto does not apply there.\n * 3. A ZERO-WIDTH interval is refused, wherever it sits. At [0, 0] it\n * cannot tell a gain from a regression and clears every negative threshold.\n * Away from zero it fails the opposite way: n identical positive deltas give\n * [g, g], which clears threshold 0 on no spread at all. Both are an absence\n * of evidence. Measured on the composable gate before this change, under a\n * bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %\n * false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.\n \n Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n /\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n BOOTSTRAP_GATE_MIN_N,\n type PairedBootstrapResult,\n pairedBinaryScale,\n pairedDeltaTieFraction,\n pairedRiskDifferenceExact,\n pairedRiskDifferenceScore,\n} from './statistics'\n\n/* Which paired estimator produced the deciding interval. /\nexport type PairedDecisionStatistic =\n \| 'paired_risk_difference'\n \| 'mean_bootstrap'\n \| 'median_bootstrap'\n\n/* Which test carried the decision, given the estimator and the sample size. /\nexport type PairedDecisionMethod = 'score-interval' \| 'bootstrap-ci' \| 'exact-sign'\n\n/* McNemar's exact paired-binary evidence, on the two-point path only. /\nexport interface PairedMcNemarEvidence {\n /* Discordant pairs the treatment won. /\n b: number\n /* Discordant pairs the control won. /\n c: number\n /* b + c — the only pairs carrying information. /\n nDiscordant: number\n /* Two-sided exact p-value. /\n pValue: number\n}\n\nexport interface PairedPromotionDecisionOptions {\n /* Smallest candidate-minus-baseline delta that counts as improvement, in the\n * caller's native units. May be negative (a noninferiority margin). Default 0. /\n threshold?: number\n /* Confidence level. Default 0.95. /\n confidence?: number\n /* Bootstrap resamples, on the paths where a bootstrap decides. Default 2000. /\n resamples?: number\n /* Deterministic bootstrap seed. Omitted ⇒ derived from the deltas. /\n seed?: number\n /* Caller-required paired observations. The test also imposes its own\n * minimum: bootstrap eligibility for a continuous mean, or the exact-test\n * minimum for binary and explicit median targets. This is not a power guarantee. /\n minPairs?: number\n /\n `'mean'` (default) routes by SHAPE: a two-point (pass/fail) outcome on any\n * encoding decides on the score interval, everything else on the mean\n * bootstrap. `'median'` forces the median bootstrap on every input, including\n * shapes where it is structurally blind — kept for callers who want outlier\n * robustness on genuinely continuous outcomes and accept that cost.\n /\n statistic?: 'mean' \| 'median'\n /* Declared binary support {0, binaryScale}, including when all observations\n * are zero. Must be finite and positive; every observation must use that\n * support. Uses the risk-difference mean, so cannot accompany 'median'.\n * Omitted: infer binary support from observed positive values. /\n binaryScale?: number\n}\n\nexport interface PairedPromotionDecision {\n /* Paired observations supplied. /\n n: number\n /* Threshold the interval was judged against, native units. /\n threshold: number\n confidence: number\n statistic: PairedDecisionStatistic\n method: PairedDecisionMethod\n /* Declared or inferred positive level ({0,1} ⇒ 1, {0,100} ⇒ 100),\n * or null when the outcome is not two-point. Non-null is exactly the\n * condition for the `paired_risk_difference` path, and it is the factor\n * `delta` / `low` / `high` were rescaled by. /\n binaryScale: number \| null\n /* Exact-tie fraction over the paired deltas; null when there are no pairs. /\n tieFraction: number \| null\n /* Point estimate of the DECIDING statistic, in the caller's native units. /\n delta: number\n /* Lower bound of the DECIDING interval, native units. /\n low: number\n /* Upper bound of the DECIDING interval, native units. /\n high: number\n /* The bootstrap that decided, or null when the score interval did. Callers\n * that need a bootstrap as a diagnostic on the two-point path compute their\n * own — it is not computed here, so the binary path costs no resamples. /\n bootstrap: PairedBootstrapResult \| null\n /* McNemar's exact evidence, or null off the two-point path. /\n mcnemar: PairedMcNemarEvidence \| null\n /* Exact one-sided sign-test p-value on the small-sample bootstrap path;\n * null otherwise. /\n pValue: number \| null\n /* Effective observation minimum for the target, confidence, and caller floor. /\n minimumPairs: number\n /* n >= minimumPairs. /\n sufficient: boolean\n /* The deciding interval is zero-width at numeric precision or non-finite — no evidence in either\n * direction, so it cannot clear any threshold on evidence. /\n indeterminate: boolean\n /* McNemar's exact test refuses at a non-negative threshold. /\n exactTestVetoes: boolean\n /* The deciding interval clears the threshold, ignoring the other two guards. /\n clearsThreshold: boolean\n /* `sufficient && !indeterminate && clearsThreshold && !exactTestVetoes` —\n * the whole rule. /\n promote: boolean\n /* What `delta` measures, for a reason string. /\n label: 'success-rate' \| 'mean' \| 'median'\n /* Why a zero-width interval is zero-width; empty when it is not. /\n indeterminateCause: string\n /* Sentence naming the test when the exact sign test decided; else empty. /\n methodDetail: string\n}\n\n/* The shape facts that pick the estimator, without computing an interval. /\nexport interface PairedDecisionShape {\n statistic: PairedDecisionStatistic\n /* Declared or inferred positive binary level; null off the binary path. /\n binaryScale: number \| null\n /* Exact-tie fraction over the paired deltas; null when there are no pairs. /\n tieFraction: number \| null\n}\n\n/\n Which estimator {@link decidePairedPromotion} would use on this data, and the\n * shape facts behind it — for callers that must report the shape on a path\n * where no interval is computed at all (an early rejection, or zero pairs).\n * Cheap: no bootstrap, no interval.\n /\nexport function pairedDecisionShape(\n before: number[],\n after: number[],\n statistic: 'mean' \| 'median' = 'mean',\n declaredBinaryScale?: number,\n): PairedDecisionShape {\n if (declaredBinaryScale !== undefined) {\n if (!Number.isFinite(declaredBinaryScale) \|\| declaredBinaryScale <= 0) {\n throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n }\n if (statistic === 'median') {\n throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n }\n for (const [name, arm] of [\n ['before', before],\n ['after', after],\n ] as const) {\n for (let i = 0; i < arm.length; i++) {\n const value = arm[i]!\n if (value !== 0 && value !== declaredBinaryScale) {\n throw new Error(\n `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n )\n }\n }\n }\n }\n const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n if (statistic === 'median') {\n return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n }\n const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n if (binaryScale !== null) {\n return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n }\n return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/\n Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n /\nexport function decidePairedPromotion(\n before: number[],\n after: number[],\n options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n if (before.length !== after.length) {\n throw new Error(\n `decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`,\n )\n }\n const threshold = options.threshold ?? 0\n if (!Number.isFinite(threshold)) {\n throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`)\n }\n const confidence = options.confidence ?? 0.95\n const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n const requestedMinimum = options.minPairs ?? exactMinimum\n if (!Number.isInteger(requestedMinimum) \|\| requestedMinimum < 1) {\n throw new Error(\n `decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`,\n )\n }\n const { binaryScale, tieFraction } = pairedDecisionShape(\n before,\n after,\n options.statistic,\n options.binaryScale,\n )\n const estimatorMinimum =\n binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n const n = before.length\n const sufficient = n >= minimumPairs\n\n let core: {\n statistic: PairedDecisionStatistic\n method: PairedDecisionMethod\n delta: number\n low: number\n high: number\n bootstrap: PairedBootstrapResult \| null\n mcnemar: PairedMcNemarEvidence \| null\n pValue: number \| null\n clearsThreshold: boolean\n label: PairedPromotionDecision['label']\n methodDetail: string\n }\n\n if (binaryScale !== null) {\n // Normalise the two-point encoding to {0,1} so the estimators see the\n // pass/fail structure, then rescale the answer back into the caller's\n // native units — the threshold is read in the units of the scores, so a\n // 0-100 pass/fail dimension must be gated in points, not in a rate.\n const unitControl = before.map((v) => v / binaryScale)\n const unitTreatment = after.map((v) => v / binaryScale)\n // TWO estimators with different jobs, because no single one does both.\n // `exact` is the authority on RD = 0 and supplies the veto; its\n // Clopper-Pearson interval conditions on the discordant count and is\n // therefore NOT a confidence interval at a nonzero margin. `score` is\n // Tango's, which is.\n const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence)\n const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence)\n const low = score.lower binaryScale\n core = {\n statistic: 'paired_risk_difference',\n method: 'score-interval',\n delta: score.riskDifference * binaryScale,\n low,\n high: score.upper * binaryScale,\n bootstrap: null,\n mcnemar: {\n b: exact.b,\n c: exact.c,\n nDiscordant: exact.nDiscordant,\n pValue: exact.pValue,\n },\n pValue: null,\n clearsThreshold: low > threshold,\n label: 'success-rate',\n methodDetail: '',\n }\n } else {\n const bootstrapStatistic = options.statistic === 'median' ? 'median' : 'mean'\n const test = pairedDeltaTest(before, after, {\n confidence,\n resamples: options.resamples,\n statistic: bootstrapStatistic,\n seed: options.seed,\n threshold,\n minPairs: minimumPairs,\n })\n const ci = test.bootstrap\n core = {\n statistic: bootstrapStatistic === 'mean' ? 'mean_bootstrap' : 'median_bootstrap',\n method: test.method,\n delta: bootstrapStatistic === 'mean' ? ci.mean : ci.median,\n low: ci.low,\n high: ci.high,\n bootstrap: ci,\n mcnemar: null,\n pValue: test.pValue,\n clearsThreshold: test.significant,\n label: bootstrapStatistic,\n methodDetail:\n test.method === 'exact-sign'\n ? bootstrapStatistic === 'mean'\n ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n : '',\n }\n }\n\n // Equivalent deltas can differ by rounding after subtraction and averaging.\n // Use the interval's own scale so tiny score units retain the same decision\n // as a rescaled copy of the evidence.\n const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n const intervalTolerance = intervalScale * Number.EPSILON * 8\n const indeterminate =\n !Number.isFinite(core.low) \|\|\n !Number.isFinite(core.high) \|\|\n core.high - core.low <= intervalTolerance\n const indeterminateCause = !indeterminate\n ? ''\n : tieFraction === 1\n ? 'every paired delta is an exact tie'\n : core.mcnemar !== null && core.mcnemar.nDiscordant === 0\n ? 'every pair is concordant (0 discordant pairs)'\n : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`\n // Only at a non-negative threshold: a negative threshold asks a\n // noninferiority question, which McNemar's test of \"no difference\" does not\n // answer.\n const exactTestVetoes =\n core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence)\n\n return {\n n,\n threshold,\n confidence,\n binaryScale,\n tieFraction,\n minimumPairs,\n sufficient,\n indeterminate,\n indeterminateCause,\n exactTestVetoes,\n promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,\n ...core,\n }\n}\n\nfunction fmt(x: number): string {\n return x.toFixed(4)\n}\n","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunOutcome,\n RunRecord,\n RunSplitTag,\n RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\n/*\n A campaign cell carried a judge score without a `dimensions` record.\n \n Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n /\nexport class CampaignJudgeScoreShapeError extends TypeError {\n constructor(judgeName: string) {\n super(\n `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n `({ judgeName, dimension, score }) is a different type with the same name.`,\n )\n this.name = 'CampaignJudgeScoreShapeError'\n }\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/\n Project one campaign cell into the canonical run format.\n \n A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n /\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const cellCostProvenance = campaignCellCostProvenance(cell)\n const costProvenance: CostProvenance =\n cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n ? { kind: 'estimated', usd: options.defaultCostUsd }\n : cellCostProvenance\n const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n // Retain the observed subtotal even when the caller supplies an estimated total.\n ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n })\n}\n\n/\n Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n /\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' \| 'costUsd' \| 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) \|\| cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance \|\| typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') \|\|\n !Number.isFinite(provenance.usd) \|\|\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/\n Produce the only task-quality view used by campaign aggregates and exports.\n \n Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n /\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n if (\n typeof dimensionsShape !== 'object' \|\|\n dimensionsShape === null \|\|\n Array.isArray(dimensionsShape)\n ) {\n throw new CampaignJudgeScoreShapeError(judgeName)\n }\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed \|\| !Number.isFinite(score.composite) \|\| !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' \| ') } : {}),\n },\n }\n}\n\n/* Read the canonical task score without recomputing cell quality. /\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number \| undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/* Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> \| undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,IAAI,gBAAgB,MAClB,OAAO;EAAE,WAAW;EAA0B;EAAa;CAAY;CAEzE,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;AACvE;;;;;;AAOA,SAAgB,sBACd,QACA,OACA,UAA0C,CAAC,GAClB;CACzB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,gDAAgD,OAAO,OAAO,MAAM,MAAM,OAAO,EACnF;CAEF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,wDAAwD,WAAW;CAErF,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,eAAe,+BAA+B,UAAU;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MACR,mEAAmE,kBACrE;CAEF,MAAM,EAAE,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,WACV;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,IAAI;CAcJ,IAAI,gBAAgB,MAAM;EAKxB,MAAM,cAAc,OAAO,KAAK,MAAM,IAAI,WAAW;EACrD,MAAM,gBAAgB,MAAM,KAAK,MAAM,IAAI,WAAW;EAMtD,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,MAAM,MAAM,QAAQ;EAC1B,OAAO;GACL,WAAW;GACX,QAAQ;GACR,OAAO,MAAM,iBAAiB;GAC9B;GACA,MAAM,MAAM,QAAQ;GACpB,WAAW;GACX,SAAS;IACP,GAAG,MAAM;IACT,GAAG,MAAM;IACT,aAAa,MAAM;IACnB,QAAQ,MAAM;GAChB;GACA,QAAQ;GACR,iBAAiB,MAAM;GACvB,OAAO;GACP,cAAc;EAChB;CACF,OAAO;EACL,MAAM,qBAAqB,QAAQ,cAAc,WAAW,WAAW;EACvE,MAAM,OAAO,gBAAgB,QAAQ,OAAO;GAC1C;GACA,WAAW,QAAQ;GACnB,WAAW;GACX,MAAM,QAAQ;GACd;GACA,UAAU;EACZ,CAAC;EACD,MAAM,KAAK,KAAK;EAChB,OAAO;GACL,WAAW,uBAAuB,SAAS,mBAAmB;GAC9D,QAAQ,KAAK;GACb,OAAO,uBAAuB,SAAS,GAAG,OAAO,GAAG;GACpD,KAAK,GAAG;GACR,MAAM,GAAG;GACT,WAAW;GACX,SAAS;GACT,QAAQ,KAAK;GACb,iBAAiB,KAAK;GACtB,OAAO;GACP,cACE,KAAK,WAAW,eACZ,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,MAAM,qBAAqB,CAAC,gBACxB,KACA,gBAAgB,IACd,uCACA,KAAK,YAAY,QAAQ,KAAK,QAAQ,gBAAgB,IACpD,kDACA,OAAO,KAAK,MAAM,8BAA8B,IAAI,KAAK,GAAG;CAIpE,MAAM,kBACJ,KAAK,YAAY,QAAQ,aAAa,KAAK,EAAE,KAAK,QAAQ,SAAS,IAAI;CAEzE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,cAAc,CAAC,iBAAiB,KAAK,mBAAmB,CAAC;EAClE,GAAG;CACL;AACF;AAEA,SAAS,IAAI,GAAmB;CAC9B,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}
1	+ {"version":3,"file":"run-record-BhMVsHR1.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.ts"],"sourcesContent":["import {\n type PairedBootstrapOptions,\n type PairedBootstrapResult,\n pairedBootstrap,\n pairedSignTest,\n} from './statistics'\n\nexport interface PairedDeltaTestOptions extends PairedBootstrapOptions {\n /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. /\n threshold?: number\n /* Caller-required paired observations. The exact test may impose a higher minimum. /\n minPairs?: number\n}\n\nexport interface PairedDeltaTestResult {\n bootstrap: PairedBootstrapResult\n method: 'bootstrap-ci' \| 'exact-sign'\n /* Exact one-sided p-value below the bootstrap minimum; otherwise null. /\n pValue: number \| null\n /* Effective observation minimum after accounting for confidence. /\n minimumPairs: number\n sufficient: boolean\n /\n The bootstrap interval has zero width (or is non-finite), so it carries no\n * information about how far the estimate could be wrong and cannot support a\n * decision in either direction. See {@link pairedDeltaTest}.\n /\n indeterminate: boolean\n significant: boolean\n}\n\n/* Smallest all-positive sample that can clear a one-sided exact sign test. /\nexport function minimumPairsForPairedDeltaTest(confidence = 0.95): number {\n if (!Number.isFinite(confidence) \|\| confidence <= 0 \|\| confidence >= 1) {\n throw new Error(\n `minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`,\n )\n }\n const oneSidedAlpha = (1 - confidence) / 2\n return Math.ceil(Math.log2(1 / oneSidedAlpha))\n}\n\n/\n Tests whether a paired candidate-minus-baseline delta clears a threshold.\n \n At 20 or more pairs, the percentile bootstrap lower bound carries the\n * decision. Below that point the interval is descriptive only, so the function\n * switches to a pre-registered one-sided exact sign test. The exact path is\n * deliberately conservative: it requires both a point estimate above the\n * threshold and enough consistently positive paired differences.\n \n ## A zero-width interval is never significant\n \n When every paired delta is identical the resample distribution is a point\n * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n\n * identical deltas of g. Neither says the effect is certain — both say the\n * sample carries no information about how far the estimate could be wrong, and\n * `low > threshold` then answers on the point estimate alone. It fails in both\n * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a\n * tie-dominated pass/fail comparison laundered a regression into a\n * noninferiority pass, and `[g, g]` clears every threshold below g with no\n * spread behind it. Under a bounded asymmetric null whose true mean paired\n * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —\n * every sample that misses the drop is exactly that shape, and deciding on\n * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.\n \n So `indeterminate` is reported and `significant` is false whenever the\n * interval has zero width, on BOTH paths: at small n the exact sign test is a\n * test of the MEDIAN and a zero-spread sample is precisely where it stops\n * saying anything about the mean the caller is thresholding.\n \n `threshold` may be negative — that is a noninferiority margin, and it is the\n * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome\n * the percentile bootstrap is not a valid interval at a nonzero margin at all;\n * use {@link decidePairedPromotion}, which routes those to Tango's score\n * interval, rather than thresholding this function's bootstrap directly.\n /\nexport function pairedDeltaTest(\n before: number[],\n after: number[],\n options: PairedDeltaTestOptions = {},\n): PairedDeltaTestResult {\n const threshold = options.threshold ?? 0\n if (!Number.isFinite(threshold)) {\n throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`)\n }\n const confidence = options.confidence ?? 0.95\n const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n const requestedMinimum = options.minPairs ?? exactMinimum\n if (!Number.isInteger(requestedMinimum) \|\| requestedMinimum < 1) {\n throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`)\n }\n const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n const bootstrap = pairedBootstrap(before, after, options)\n const sufficient = bootstrap.n >= minimumPairs\n // A point-mass resample distribution is an absence of evidence, not a\n // certainty, and it is the shape that clears every negative threshold. Both\n // paths refuse it: see the \"zero-width\" section of this function's docs.\n const indeterminate =\n bootstrap.n > 0 &&\n (!Number.isFinite(bootstrap.low) \|\|\n !Number.isFinite(bootstrap.high) \|\|\n bootstrap.low === bootstrap.high)\n\n if (bootstrap.gateEligible) {\n return {\n bootstrap,\n method: 'bootstrap-ci',\n pValue: null,\n minimumPairs,\n sufficient,\n indeterminate,\n significant: sufficient && !indeterminate && bootstrap.low > threshold,\n }\n }\n\n const differences = before.map((value, index) => after[index]! - value - threshold)\n const exact = pairedSignTest(differences, 'greater')\n const estimate = options.statistic === 'mean' ? bootstrap.mean : bootstrap.median\n return {\n bootstrap,\n method: 'exact-sign',\n pValue: exact.pValue,\n minimumPairs,\n sufficient,\n indeterminate,\n significant:\n sufficient &&\n !indeterminate &&\n estimate > threshold &&\n exact.pValue <= (1 - bootstrap.confidence) / 2,\n }\n}\n","/\n @module\n * ONE rule for \"does this paired interval clear a promotion threshold\".\n \n The rule below was derived on `HeldOutGate` (#479) after the same estimator\n * bug shipped twice. It then turned out that a SECOND gate — the composable\n * `heldOutGate`, plus everything else routed through `heldoutSignificance` —\n * still carried the original defect, because the rule had been written into one\n * gate's method body rather than into a shared function. Two copies of a\n * statistical rule is how a defect survives in one of them, so there is now\n * exactly one copy and both gates call it.\n \n Three things the rule does that a bare `pairedBootstrap(...).low > threshold`\n * does not:\n \n 1. Two-point (pass/fail) outcomes decide on Tango's SCORE interval. On a\n * pass/fail eval the paired delta vector is dominated by ties, so the\n * bootstrap of the mean is a resample of a lattice with three atoms and its\n * percentile interval is not valid at a nonzero margin. The score interval\n * (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under\n * each hypothesised margin instead of fixing it at the observed value, which\n * is the only construction that stays a confidence interval as the margin\n * moves off zero — the regime every noninferiority threshold lives in.\n * Measured on the composable gate before this change, at a true risk\n * difference sitting exactly on the production caller's -0.05 margin and a\n * nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.\n * 2. McNemar's exact test holds a VETO at every non-negative threshold.\n * Redundant with the interval by construction and kept anyway, so that\n * swapping the estimator for one without that duality cannot silently\n * reintroduce \"promotes what the exact test refuses\". Witness: n = 6, b = 5,\n * c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs\n * (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE\n * threshold is a noninferiority question, which McNemar's test of \"no\n * difference\" is not the right test for, so the veto does not apply there.\n * 3. A ZERO-WIDTH interval is refused, wherever it sits. At [0, 0] it\n * cannot tell a gain from a regression and clears every negative threshold.\n * Away from zero it fails the opposite way: n identical positive deltas give\n * [g, g], which clears threshold 0 on no spread at all. Both are an absence\n * of evidence. Measured on the composable gate before this change, under a\n * bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %\n * false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.\n \n Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n /\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n BOOTSTRAP_GATE_MIN_N,\n type PairedBootstrapResult,\n pairedBinaryScale,\n pairedDeltaTieFraction,\n pairedRiskDifferenceExact,\n pairedRiskDifferenceScore,\n} from './statistics'\n\n/* Which paired estimator produced the deciding interval. /\nexport type PairedDecisionStatistic =\n \| 'paired_risk_difference'\n \| 'mean_bootstrap'\n \| 'median_bootstrap'\n\n/* Which test carried the decision, given the estimator and the sample size. /\nexport type PairedDecisionMethod = 'score-interval' \| 'bootstrap-ci' \| 'exact-sign'\n\n/* McNemar's exact paired-binary evidence, on the two-point path only. /\nexport interface PairedMcNemarEvidence {\n /* Discordant pairs the treatment won. /\n b: number\n /* Discordant pairs the control won. /\n c: number\n /* b + c — the only pairs carrying information. /\n nDiscordant: number\n /* Two-sided exact p-value. /\n pValue: number\n}\n\nexport interface PairedPromotionDecisionOptions {\n /* Smallest candidate-minus-baseline delta that counts as improvement, in the\n * caller's native units. May be negative (a noninferiority margin). Default 0. /\n threshold?: number\n /* Confidence level. Default 0.95. /\n confidence?: number\n /* Bootstrap resamples, on the paths where a bootstrap decides. Default 2000. /\n resamples?: number\n /* Deterministic bootstrap seed. Omitted ⇒ derived from the deltas. /\n seed?: number\n /* Caller-required paired observations. The test also imposes its own\n * minimum: bootstrap eligibility for a continuous mean, or the exact-test\n * minimum for binary and explicit median targets. This is not a power guarantee. /\n minPairs?: number\n /\n `'mean'` (default) routes by SHAPE: a two-point (pass/fail) outcome on any\n * encoding decides on the score interval, everything else on the mean\n * bootstrap. `'median'` forces the median bootstrap on every input, including\n * shapes where it is structurally blind — kept for callers who want outlier\n * robustness on genuinely continuous outcomes and accept that cost.\n /\n statistic?: 'mean' \| 'median'\n /* Declared binary support {0, binaryScale}, including when all observations\n * are zero. Must be finite and positive; every observation must use that\n * support. Uses the risk-difference mean, so cannot accompany 'median'.\n * Omitted: infer binary support from observed positive values. /\n binaryScale?: number\n}\n\nexport interface PairedPromotionDecision {\n /* Paired observations supplied. /\n n: number\n /* Threshold the interval was judged against, native units. /\n threshold: number\n confidence: number\n statistic: PairedDecisionStatistic\n method: PairedDecisionMethod\n /* Declared or inferred positive level ({0,1} ⇒ 1, {0,100} ⇒ 100),\n * or null when the outcome is not two-point. Non-null is exactly the\n * condition for the `paired_risk_difference` path, and it is the factor\n * `delta` / `low` / `high` were rescaled by. /\n binaryScale: number \| null\n /* Exact-tie fraction over the paired deltas; null when there are no pairs. /\n tieFraction: number \| null\n /* Point estimate of the DECIDING statistic, in the caller's native units. /\n delta: number\n /* Lower bound of the DECIDING interval, native units. /\n low: number\n /* Upper bound of the DECIDING interval, native units. /\n high: number\n /* The bootstrap that decided, or null when the score interval did. Callers\n * that need a bootstrap as a diagnostic on the two-point path compute their\n * own — it is not computed here, so the binary path costs no resamples. /\n bootstrap: PairedBootstrapResult \| null\n /* McNemar's exact evidence, or null off the two-point path. /\n mcnemar: PairedMcNemarEvidence \| null\n /* Exact one-sided sign-test p-value on the small-sample bootstrap path;\n * null otherwise. /\n pValue: number \| null\n /* Effective observation minimum for the target, confidence, and caller floor. /\n minimumPairs: number\n /* n >= minimumPairs. /\n sufficient: boolean\n /* The deciding interval is zero-width at numeric precision or non-finite — no evidence in either\n * direction, so it cannot clear any threshold on evidence. /\n indeterminate: boolean\n /* McNemar's exact test refuses at a non-negative threshold. /\n exactTestVetoes: boolean\n /* The deciding interval clears the threshold, ignoring the other two guards. /\n clearsThreshold: boolean\n /* `sufficient && !indeterminate && clearsThreshold && !exactTestVetoes` —\n * the whole rule. /\n promote: boolean\n /* What `delta` measures, for a reason string. /\n label: 'success-rate' \| 'mean' \| 'median'\n /* Why a zero-width interval is zero-width; empty when it is not. /\n indeterminateCause: string\n /* Sentence naming the test when the exact sign test decided; else empty. /\n methodDetail: string\n}\n\n/* The shape facts that pick the estimator, without computing an interval. /\nexport interface PairedDecisionShape {\n statistic: PairedDecisionStatistic\n /* Declared or inferred positive binary level; null off the binary path. /\n binaryScale: number \| null\n /* Exact-tie fraction over the paired deltas; null when there are no pairs. /\n tieFraction: number \| null\n}\n\n/\n Which estimator {@link decidePairedPromotion} would use on this data, and the\n * shape facts behind it — for callers that must report the shape on a path\n * where no interval is computed at all (an early rejection, or zero pairs).\n * Cheap: no bootstrap, no interval.\n /\nexport function pairedDecisionShape(\n before: number[],\n after: number[],\n statistic: 'mean' \| 'median' = 'mean',\n declaredBinaryScale?: number,\n): PairedDecisionShape {\n if (declaredBinaryScale !== undefined) {\n if (!Number.isFinite(declaredBinaryScale) \|\| declaredBinaryScale <= 0) {\n throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n }\n if (statistic === 'median') {\n throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n }\n for (const [name, arm] of [\n ['before', before],\n ['after', after],\n ] as const) {\n for (let i = 0; i < arm.length; i++) {\n const value = arm[i]!\n if (value !== 0 && value !== declaredBinaryScale) {\n throw new Error(\n `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n )\n }\n }\n }\n }\n const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n if (statistic === 'median') {\n return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n }\n const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n if (binaryScale !== null) {\n return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n }\n return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/\n Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n /\nexport function decidePairedPromotion(\n before: number[],\n after: number[],\n options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n if (before.length !== after.length) {\n throw new Error(\n `decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`,\n )\n }\n const threshold = options.threshold ?? 0\n if (!Number.isFinite(threshold)) {\n throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`)\n }\n const confidence = options.confidence ?? 0.95\n const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n const requestedMinimum = options.minPairs ?? exactMinimum\n if (!Number.isInteger(requestedMinimum) \|\| requestedMinimum < 1) {\n throw new Error(\n `decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`,\n )\n }\n const { binaryScale, tieFraction } = pairedDecisionShape(\n before,\n after,\n options.statistic,\n options.binaryScale,\n )\n const estimatorMinimum =\n binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n const n = before.length\n const sufficient = n >= minimumPairs\n\n let core: {\n statistic: PairedDecisionStatistic\n method: PairedDecisionMethod\n delta: number\n low: number\n high: number\n bootstrap: PairedBootstrapResult \| null\n mcnemar: PairedMcNemarEvidence \| null\n pValue: number \| null\n clearsThreshold: boolean\n label: PairedPromotionDecision['label']\n methodDetail: string\n }\n\n if (binaryScale !== null) {\n // Normalise the two-point encoding to {0,1} so the estimators see the\n // pass/fail structure, then rescale the answer back into the caller's\n // native units — the threshold is read in the units of the scores, so a\n // 0-100 pass/fail dimension must be gated in points, not in a rate.\n const unitControl = before.map((v) => v / binaryScale)\n const unitTreatment = after.map((v) => v / binaryScale)\n // TWO estimators with different jobs, because no single one does both.\n // `exact` is the authority on RD = 0 and supplies the veto; its\n // Clopper-Pearson interval conditions on the discordant count and is\n // therefore NOT a confidence interval at a nonzero margin. `score` is\n // Tango's, which is.\n const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence)\n const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence)\n const low = score.lower binaryScale\n core = {\n statistic: 'paired_risk_difference',\n method: 'score-interval',\n delta: score.riskDifference * binaryScale,\n low,\n high: score.upper * binaryScale,\n bootstrap: null,\n mcnemar: {\n b: exact.b,\n c: exact.c,\n nDiscordant: exact.nDiscordant,\n pValue: exact.pValue,\n },\n pValue: null,\n clearsThreshold: low > threshold,\n label: 'success-rate',\n methodDetail: '',\n }\n } else {\n const bootstrapStatistic = options.statistic === 'median' ? 'median' : 'mean'\n const test = pairedDeltaTest(before, after, {\n confidence,\n resamples: options.resamples,\n statistic: bootstrapStatistic,\n seed: options.seed,\n threshold,\n minPairs: minimumPairs,\n })\n const ci = test.bootstrap\n core = {\n statistic: bootstrapStatistic === 'mean' ? 'mean_bootstrap' : 'median_bootstrap',\n method: test.method,\n delta: bootstrapStatistic === 'mean' ? ci.mean : ci.median,\n low: ci.low,\n high: ci.high,\n bootstrap: ci,\n mcnemar: null,\n pValue: test.pValue,\n clearsThreshold: test.significant,\n label: bootstrapStatistic,\n methodDetail:\n test.method === 'exact-sign'\n ? bootstrapStatistic === 'mean'\n ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n : '',\n }\n }\n\n // Equivalent deltas can differ by rounding after subtraction and averaging.\n // Use the interval's own scale so tiny score units retain the same decision\n // as a rescaled copy of the evidence.\n const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n const intervalTolerance = intervalScale * Number.EPSILON * 8\n const indeterminate =\n !Number.isFinite(core.low) \|\|\n !Number.isFinite(core.high) \|\|\n core.high - core.low <= intervalTolerance\n const indeterminateCause = !indeterminate\n ? ''\n : tieFraction === 1\n ? 'every paired delta is an exact tie'\n : core.mcnemar !== null && core.mcnemar.nDiscordant === 0\n ? 'every pair is concordant (0 discordant pairs)'\n : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`\n // Only at a non-negative threshold: a negative threshold asks a\n // noninferiority question, which McNemar's test of \"no difference\" does not\n // answer.\n const exactTestVetoes =\n core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence)\n\n return {\n n,\n threshold,\n confidence,\n binaryScale,\n tieFraction,\n minimumPairs,\n sufficient,\n indeterminate,\n indeterminateCause,\n exactTestVetoes,\n promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,\n ...core,\n }\n}\n\nfunction fmt(x: number): string {\n return x.toFixed(4)\n}\n","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunOutcome,\n RunRecord,\n RunSplitTag,\n RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\n/*\n A campaign cell carried a judge score without a `dimensions` record.\n \n Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n /\nexport class CampaignJudgeScoreShapeError extends TypeError {\n constructor(judgeName: string) {\n super(\n `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n `({ judgeName, dimension, score }) is a different type with the same name.`,\n )\n this.name = 'CampaignJudgeScoreShapeError'\n }\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/\n Project one campaign cell into the canonical run format.\n \n A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n /\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const cellCostProvenance = campaignCellCostProvenance(cell)\n const costProvenance: CostProvenance =\n cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n ? { kind: 'estimated', usd: options.defaultCostUsd }\n : cellCostProvenance\n const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n // Retain the observed subtotal even when the caller supplies an estimated total.\n ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n })\n}\n\n/\n Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n /\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' \| 'costUsd' \| 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) \|\| cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance \|\| typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') \|\|\n !Number.isFinite(provenance.usd) \|\|\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/\n Produce the only task-quality view used by campaign aggregates and exports.\n \n Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n /\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n if (\n typeof dimensionsShape !== 'object' \|\|\n dimensionsShape === null \|\|\n Array.isArray(dimensionsShape)\n ) {\n throw new CampaignJudgeScoreShapeError(judgeName)\n }\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed \|\| !Number.isFinite(score.composite) \|\| !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' \| ') } : {}),\n },\n }\n}\n\n/* Read the canonical task score without recomputing cell quality. /\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number \| undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/* Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> \| undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,IAAI,gBAAgB,MAClB,OAAO;EAAE,WAAW;EAA0B;EAAa;CAAY;CAEzE,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;AACvE;;;;;;AAOA,SAAgB,sBACd,QACA,OACA,UAA0C,CAAC,GAClB;CACzB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,gDAAgD,OAAO,OAAO,MAAM,MAAM,OAAO,EACnF;CAEF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,wDAAwD,WAAW;CAErF,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,eAAe,+BAA+B,UAAU;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MACR,mEAAmE,kBACrE;CAEF,MAAM,EAAE,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,WACV;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,IAAI;CAcJ,IAAI,gBAAgB,MAAM;EAKxB,MAAM,cAAc,OAAO,KAAK,MAAM,IAAI,WAAW;EACrD,MAAM,gBAAgB,MAAM,KAAK,MAAM,IAAI,WAAW;EAMtD,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,MAAM,MAAM,QAAQ;EAC1B,OAAO;GACL,WAAW;GACX,QAAQ;GACR,OAAO,MAAM,iBAAiB;GAC9B;GACA,MAAM,MAAM,QAAQ;GACpB,WAAW;GACX,SAAS;IACP,GAAG,MAAM;IACT,GAAG,MAAM;IACT,aAAa,MAAM;IACnB,QAAQ,MAAM;GAChB;GACA,QAAQ;GACR,iBAAiB,MAAM;GACvB,OAAO;GACP,cAAc;EAChB;CACF,OAAO;EACL,MAAM,qBAAqB,QAAQ,cAAc,WAAW,WAAW;EACvE,MAAM,OAAO,gBAAgB,QAAQ,OAAO;GAC1C;GACA,WAAW,QAAQ;GACnB,WAAW;GACX,MAAM,QAAQ;GACd;GACA,UAAU;EACZ,CAAC;EACD,MAAM,KAAK,KAAK;EAChB,OAAO;GACL,WAAW,uBAAuB,SAAS,mBAAmB;GAC9D,QAAQ,KAAK;GACb,OAAO,uBAAuB,SAAS,GAAG,OAAO,GAAG;GACpD,KAAK,GAAG;GACR,MAAM,GAAG;GACT,WAAW;GACX,SAAS;GACT,QAAQ,KAAK;GACb,iBAAiB,KAAK;GACtB,OAAO;GACP,cACE,KAAK,WAAW,eACZ,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,MAAM,qBAAqB,CAAC,gBACxB,KACA,gBAAgB,IACd,uCACA,KAAK,YAAY,QAAQ,KAAK,QAAQ,gBAAgB,IACpD,kDACA,OAAO,KAAK,MAAM,8BAA8B,IAAI,KAAK,GAAG;CAIpE,MAAM,kBACJ,KAAK,YAAY,QAAQ,aAAa,KAAK,EAAE,KAAK,QAAQ,SAAS,IAAI;CAEzE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,cAAc,CAAC,iBAAiB,KAAK,mBAAmB,CAAC;EAClE,GAAG;CACL;AACF;AAEA,SAAS,IAAI,GAAmB;CAC9B,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}

package/dist/{run-record-BiTWauyO.d.ts → run-record-D9wS35fK.d.ts} RENAMED Viewed

@@ -1,6 +1,6 @@
 import { c as ValidationError } from "./errors-DEE6u6ot.js";
 import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
-import { r as AgentProfileCell } from "./agent-profile-cell-s__adRnK.js";
+import { r as AgentProfileCell } from "./agent-profile-cell-BHcB-mZX.js";
 import { o as FailureClass } from "./schema-CR5cpjQ3.js";
 //#region src/run-record.d.ts
 /** Search/dev/holdout split tag. 'search' is the paper-grade alias for the
@@ -244,4 +244,4 @@ declare function roundTripRunRecord(record: RunRecord): RunRecord;
 declare function modelHasSnapshot(model: string): boolean;
 //#endregion
 export { validateRunRecord as _, RunRecord as a, RunTaskFailure as c, UNKNOWN_MODEL as d, isRunRecord as f, runTaskScore as g, roundTripRunRecord as h, RunOutcome as i, RunTerminalOutcome as l, parseRunRecordSafe as m, RunCostProvenance as n, RunRecordValidationError as o, modelHasSnapshot as p, RunJudgeMetadata as r, RunSplitTag as s, JudgeScoresRecord as t, RunTokenUsage as u };
-//# sourceMappingURL=run-record-BiTWauyO.d.ts.map
+//# sourceMappingURL=run-record-D9wS35fK.d.ts.map

package/dist/{run-record-BiTWauyO.d.ts.map → run-record-D9wS35fK.d.ts.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"run-record-~~BiTWauyO~~.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAgPnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}
1	+ {"version":3,"file":"run-record-D9wS35fK.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAgPnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}

package/dist/{run-record-DualPTn2.js → run-record-tg_5H2IK.js} RENAMED Viewed

@@ -1,5 +1,5 @@
 import { s as ValidationError } from "./errors-Dngq5h35.js";
-import { d as validateAgentProfileCell } from "./agent-profile-cell-Cv6UA-W_.js";
+import { d as validateAgentProfileCell } from "./agent-profile-cell-DcFW4vOc.js";
 import { t as FAILURE_CLASSES } from "./schema-CSf6qWgZ.js";
 import { n as observedScore } from "./reward-nw2xZGZG.js";
 //#region src/run-record.ts
@@ -255,4 +255,4 @@ function validSnapshotDate(year, month, day) {
 //#endregion
 export { parseRunRecordSafe as a, validateRunRecord as c, modelHasSnapshot as i, UNKNOWN_MODEL as n, roundTripRunRecord as o, isRunRecord as r, runTaskScore as s, RunRecordValidationError as t };
-//# sourceMappingURL=run-record-DualPTn2.js.map
+//# sourceMappingURL=run-record-tg_5H2IK.js.map

package/dist/{run-record-DualPTn2.js.map → run-record-tg_5H2IK.js.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"run-record-DualPTn2.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/*\n Paper-grade RunRecord schema + runtime validator.\n \n Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n \n This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n \n Validate at the boundary:\n \n const rec = validateRunRecord(rawJson) // throws on missing\n * const ok = isRunRecord(rawJson) // boolean check\n * const rec = parseRunRecordSafe(rawJson) // { ok, value \| error }\n \n The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n /\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/* Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n * combined train+test pool that the optimizer is allowed to read. /\nexport type RunSplitTag = 'search' \| 'dev' \| 'holdout'\n\n/\n Explicit execution-lifecycle result for a run.\n \n This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n /\nexport type RunTerminalOutcome = 'succeeded' \| 'failed' \| 'cancelled' \| 'incomplete' \| 'unknown'\n\n/* Explicit model value for a row that never produced a served model snapshot. /\nexport const UNKNOWN_MODEL = 'unknown'\n\nexport interface RunTokenUsage {\n input: number\n /* All generated tokens charged as output, including reasoning tokens. /\n output: number\n /* Present only when one or more paid calls did not report token usage.\n * In that case, every numeric field is a known subtotal, not a measured total. /\n tokensKnown?: false\n /* Reasoning-token subset of `output`, when the provider reports it. /\n reasoning?: number\n /* Prompt tokens served from a provider cache. /\n cached?: number\n /* Prompt tokens written into a provider cache. /\n cacheWrite?: number\n}\n\n/* How a run's USD amount was obtained. /\nexport type RunCostProvenance = CostProvenance\n\nexport interface RunJudgeMetadata {\n model: string\n promptVersion: string\n /* [0,1] confidence the judge declared. Constant judge confidence\n * across many runs is a fallback signal (see `canary.ts`). /\n confidence: number\n /* True if the judge degraded to a fallback path (rules-only,\n * prior-call cache, etc.). The canary uses this to alert. /\n fallback: boolean\n}\n\n/\n Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n \n The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n \n `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n \n Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n /\nexport interface JudgeScoresRecord {\n /* Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. /\n perJudge: Record<string, Record<string, number>>\n /* Per-dim mean across judges. Convenience — derivable from `perJudge`. /\n perDimMean: Record<string, number>\n /* Composite mean across successful judges. Mirrors the task score only\n * when `failedJudges` is empty. /\n composite: number\n /* Judges that errored or returned an unparseable verdict. Recorded\n * by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n * not inferred from missing keys in `perJudge`. /\n failedJudges?: string[]\n /* Free-form notes the judges emitted (joined across judges or\n * first-judge only — consumer's choice). /\n notes?: string\n}\n\nexport interface RunOutcome {\n /* Score on the search/optimization split. Optional for holdout-only and\n * execution-only records. /\n searchScore?: number\n /* Score on the held-out split. Optional for search-only and execution-only\n * records. When both scores are absent, the run is explicitly unlabeled. /\n holdoutScore?: number\n /* Bag of any other metric the run produced — judge dimensions,\n * pass/fail counters, latency stats, etc. Numeric only — keeps\n * reporters honest. /\n raw: Record<string, number>\n /* Per-judge / per-dim breakdown. Consumers writing ensemble\n * judgements populate this; substrate primitives like\n * `interRaterReliability` and `corpusInterRaterAgreement` accept\n * these records as input. Optional — single-judge or scalar-only\n * runs leave it unset. /\n judgeScores?: JudgeScoresRecord\n /* Authenticity / realness verdict — did the run build the REAL thing on the\n * intended infra, or fake it (see `./authenticity`)? Optional: only domains\n * with an authenticity config populate it. Carried in the corpus so the\n * flywheel / off-policy learning can optimize for real completion, not gamed\n * pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n * must not count as a real success regardless of `score`. /\n realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/\n Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n \n Hash discipline:\n * - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n * model (after any steering bundle merge).\n * - `configHash` is the sha256 of the effective run config (model,\n * temperature, tools, judges, splits). The pair (promptHash,\n * configHash) uniquely identifies an experiment cell.\n \n Model snapshot discipline:\n * - successful rows MUST encode a snapshot version. Bare aliases like\n * `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n * Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n * - a failed, cancelled, incomplete, or otherwise unknown row may use\n * `UNKNOWN_MODEL` when no served model was observed. This is an explicit\n * absence marker, not a fabricated snapshot.\n /\nexport interface RunRecord {\n /* UUID for the run. /\n runId: string\n /* Logical experiment grouping (a treatment vs a baseline within\n * the same sweep should share `experimentId`). /\n experimentId: string\n /* Stable identifier for the candidate (variant) being run. The\n * promotion gate compares two `candidateId`s on matched items. /\n candidateId: string\n /* RNG seed for the run. Always recorded — silent re-seeding is\n * the most common cause of non-reproducible numbers. /\n seed: number\n /* Model identifier WITH snapshot version. /\n model: string\n /* sha256 of the effective prompt (post-steering). /\n promptHash: string\n /* sha256 of the effective config. /\n configHash: string\n /* Git SHA the harness was run from. /\n commitSha: string\n /* End-to-end wall-clock duration in milliseconds. /\n wallMs: number\n /* Time spent queued before execution started, if known. /\n queueMs?: number\n /* Total USD cost, or null when the producer could not capture one. /\n costUsd: number \| null\n /* Whether `costUsd` came from billing data, a price calculation, or is unavailable. /\n costProvenance: RunCostProvenance\n /* Token usage breakdown. /\n tokenUsage: RunTokenUsage\n /* Root-run or process terminal result. Never inferred from a child span. /\n terminalOutcome: RunTerminalOutcome\n /* Root-run or process failure reason. Valid only for a failed, cancelled,\n * or incomplete terminal result; never populated from a child span. /\n terminalFailureReason?: string\n /* Judge-side metadata, if a judge was used. /\n judgeMetadata?: RunJudgeMetadata\n /* Per-split scores + raw bag. /\n outcome: RunOutcome\n /* Canonical task-failure class drawn from the shared\n * `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n * evidence. Execution errors belong in\n * `outcome.raw.execution_error_count`. /\n failureClass?: FailureClass\n /* Free-form task-failure detail scoped under a non-success\n * `failureClass`. It is invalid without that class. /\n failureMode?: string\n /* Which split this run was drawn from. /\n splitTag: RunSplitTag\n /\n Stable scenario identifier the run observed or was scored against.\n * Comparison primitives match this identity rather than input order.\n /\n scenarioId: string\n /\n Canonical identity for the agent profile cell that produced this row:\n * profile artifact hash plus optional harness/model/prompt/reporting\n * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n * longitudinal reports by the complete source profile, not by a loose\n * candidate label or opaque config hash.\n /\n agentProfile?: AgentProfileCell\n}\n\n/\n Canonical task-result classification.\n \n A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n /\nexport type RunTaskFailure =\n \| { failureClass?: undefined; failureMode?: undefined }\n \| { failureClass: 'success'; failureMode?: undefined }\n \| {\n failureClass: Exclude<FailureClass, 'success'>\n failureMode?: string\n }\n\n/\n Return task quality, preferring held-out evidence when both scores exist.\n \n RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n /\nexport function runTaskScore(record: RunRecord): number \| undefined {\n const score = observedScore(record)\n return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n 'runId',\n 'experimentId',\n 'candidateId',\n 'seed',\n 'model',\n 'promptHash',\n 'configHash',\n 'commitSha',\n 'wallMs',\n 'costUsd',\n 'costProvenance',\n 'tokenUsage',\n 'terminalOutcome',\n 'outcome',\n 'splitTag',\n 'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n 'succeeded',\n 'failed',\n 'cancelled',\n 'incomplete',\n 'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n readonly path: string\n constructor(message: string, path = '') {\n super(path ? `${message} (at ${path})` : message)\n this.path = path\n }\n}\n\n/\n Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n /\nexport function validateRunRecord(input: unknown): RunRecord {\n if (input === null \|\| typeof input !== 'object') {\n throw new RunRecordValidationError('expected object')\n }\n const obj = input as Record<string, unknown>\n\n for (const key of MANDATORY_TOP_LEVEL) {\n if (!(key in obj)) {\n throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n }\n }\n\n expectString(obj.runId, 'runId')\n expectString(obj.experimentId, 'experimentId')\n expectString(obj.candidateId, 'candidateId')\n expectFiniteNumber(obj.seed, 'seed')\n expectString(obj.model, 'model')\n expectString(obj.promptHash, 'promptHash')\n expectString(obj.configHash, 'configHash')\n expectString(obj.commitSha, 'commitSha')\n expectNonNegativeNumber(obj.wallMs, 'wallMs')\n if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n validateCost(obj.costUsd, obj.costProvenance)\n\n // Snapshot discipline: successful rows require a served model snapshot.\n // Non-success rows may carry the explicit absence marker when execution\n // stopped before a model identity was observed.\n if (\n !modelHasSnapshot(obj.model as string) &&\n !(obj.model === UNKNOWN_MODEL && obj.terminalOutcome !== 'succeeded')\n ) {\n throw new RunRecordValidationError(\n `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD', or '${UNKNOWN_MODEL}' for a non-success row without a served model)`,\n 'model',\n )\n }\n\n // Token usage.\n const tu = obj.tokenUsage\n if (tu === null \|\| typeof tu !== 'object') {\n throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n }\n const tuRec = tu as Record<string, unknown>\n expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n if (tuRec.tokensKnown !== undefined && tuRec.tokensKnown !== false) {\n throw new RunRecordValidationError(\n 'tokensKnown must be false when present; omit it when token usage is complete',\n 'tokenUsage.tokensKnown',\n )\n }\n if (tuRec.reasoning !== undefined) {\n expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n throw new RunRecordValidationError(\n 'reasoning tokens must be a subset of output tokens',\n 'tokenUsage.reasoning',\n )\n }\n }\n if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n if (tuRec.cacheWrite !== undefined) {\n expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n }\n\n // Judge metadata, optional.\n if (obj.judgeMetadata !== undefined) {\n const jm = obj.judgeMetadata\n if (jm === null \|\| typeof jm !== 'object') {\n throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n }\n const jmRec = jm as Record<string, unknown>\n expectString(jmRec.model, 'judgeMetadata.model')\n expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n if (typeof jmRec.fallback !== 'boolean') {\n throw new RunRecordValidationError(\n 'judgeMetadata.fallback must be boolean',\n 'judgeMetadata.fallback',\n )\n }\n }\n\n // Outcome.\n const out = obj.outcome\n if (out === null \|\| typeof out !== 'object') {\n throw new RunRecordValidationError('outcome must be an object', 'outcome')\n }\n const outRec = out as Record<string, unknown>\n if (outRec.searchScore !== undefined)\n expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n if (outRec.holdoutScore !== undefined)\n expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n const raw = outRec.raw\n if (raw === null \|\| typeof raw !== 'object') {\n throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n }\n for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n expectFiniteNumber(v, `outcome.raw.${k}`)\n }\n // Realness verdict, optional.\n if (outRec.realness !== undefined) {\n const r = outRec.realness\n if (r === null \|\| typeof r !== 'object') {\n throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n }\n const rr = r as Record<string, unknown>\n expectFiniteNumber(rr.score, 'outcome.realness.score')\n if (typeof rr.gated !== 'boolean') {\n throw new RunRecordValidationError(\n 'outcome.realness.gated must be a boolean',\n 'outcome.realness.gated',\n )\n }\n }\n\n // Per-judge / per-dim breakdown, optional.\n if (outRec.judgeScores !== undefined) {\n validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n }\n\n // Failure mode optional.\n if (\n obj.failureClass !== undefined &&\n (typeof obj.failureClass !== 'string' \|\|\n !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n ) {\n throw new RunRecordValidationError(\n `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n 'failureClass',\n )\n }\n if (obj.failureMode !== undefined) {\n expectString(obj.failureMode, 'failureMode')\n if (obj.failureClass === undefined \|\| obj.failureClass === 'success') {\n throw new RunRecordValidationError(\n 'failureMode requires a non-success failureClass',\n 'failureMode',\n )\n }\n }\n\n if (\n typeof obj.terminalOutcome !== 'string' \|\|\n !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n ) {\n throw new RunRecordValidationError(\n `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n 'terminalOutcome',\n )\n }\n if (obj.terminalFailureReason !== undefined) {\n expectString(obj.terminalFailureReason, 'terminalFailureReason')\n if (\n obj.terminalOutcome !== 'failed' &&\n obj.terminalOutcome !== 'cancelled' &&\n obj.terminalOutcome !== 'incomplete'\n ) {\n throw new RunRecordValidationError(\n 'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n 'terminalFailureReason',\n )\n }\n }\n\n if (obj.agentProfile !== undefined) {\n try {\n const profile = validateAgentProfileCell(obj.agentProfile)\n if (profile.model !== undefined && profile.model !== obj.model) {\n throw new RunRecordValidationError(\n `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n 'agentProfile.model',\n )\n }\n if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n throw new RunRecordValidationError(\n `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n 'agentProfile.promptHash',\n )\n }\n } catch (error) {\n if (error instanceof RunRecordValidationError) throw error\n if (error instanceof Error) {\n throw new RunRecordValidationError(error.message, 'agentProfile')\n }\n throw error\n }\n }\n\n expectString(obj.scenarioId, 'scenarioId')\n\n // Split tag.\n if (typeof obj.splitTag !== 'string' \|\| !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n throw new RunRecordValidationError(\n `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n 'splitTag',\n )\n }\n\n return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n if (provenance === null \|\| typeof provenance !== 'object') {\n throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n }\n const value = provenance as Record<string, unknown>\n if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n throw new RunRecordValidationError(\n 'costProvenance.kind must be observed, estimated, or uncaptured',\n 'costProvenance.kind',\n )\n }\n // A record must never read as a total it cannot support, so an uncaptured\n // cost carries no number at all. A matrix `CellResult` keeps its known\n // subtotal instead, because a cost ceiling must charge the part it can see;\n // converting one to the other drops that subtotal.\n if (value.kind === 'uncaptured') {\n if (value.usd !== null) {\n throw new RunRecordValidationError(\n 'uncaptured costProvenance.usd must be null',\n 'costProvenance.usd',\n )\n }\n if (costUsd !== null) {\n throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n }\n return\n }\n expectNonNegativeNumber(costUsd, 'costUsd')\n expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n if (value.usd !== costUsd) {\n throw new RunRecordValidationError(\n 'costProvenance.usd must equal costUsd',\n 'costProvenance.usd',\n )\n }\n}\n\n/* Boolean validator — convenience for filtering arrays. /\nexport function isRunRecord(input: unknown): input is RunRecord {\n try {\n validateRunRecord(input)\n return true\n } catch {\n return false\n }\n}\n\n/* Non-throwing validator — returns a discriminated union. /\nexport function parseRunRecordSafe(\n input: unknown,\n): { ok: true; value: RunRecord } \| { ok: false; error: RunRecordValidationError } {\n try {\n return { ok: true, value: validateRunRecord(input) }\n } catch (e) {\n if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n throw e\n }\n}\n\n/* Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. /\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n const json = JSON.stringify(record)\n return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n if (typeof value !== 'string' \|\| value.length === 0) {\n throw new RunRecordValidationError(`expected non-empty string`, path)\n }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n if (typeof value !== 'number' \|\| !Number.isFinite(value)) {\n throw new RunRecordValidationError(`expected finite number`, path)\n }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n expectFiniteNumber(value, path)\n if ((value as number) < 0) {\n throw new RunRecordValidationError('expected non-negative number', path)\n }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n if (value === null \|\| typeof value !== 'object') {\n throw new RunRecordValidationError('judgeScores must be an object', path)\n }\n const rec = value as Record<string, unknown>\n\n const perJudge = rec.perJudge\n if (perJudge === null \|\| typeof perJudge !== 'object') {\n throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n }\n for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n if (dims === null \|\| typeof dims !== 'object') {\n throw new RunRecordValidationError(\n 'per-judge entry must be an object of dimension scores',\n `${path}.perJudge.${judgeId}`,\n )\n }\n for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n }\n }\n\n const perDimMean = rec.perDimMean\n if (perDimMean === null \|\| typeof perDimMean !== 'object') {\n throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n }\n for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n }\n\n expectFiniteNumber(rec.composite, `${path}.composite`)\n\n if (rec.failedJudges !== undefined) {\n if (!Array.isArray(rec.failedJudges)) {\n throw new RunRecordValidationError(\n 'failedJudges must be an array of strings',\n `${path}.failedJudges`,\n )\n }\n for (let i = 0; i < rec.failedJudges.length; i++) {\n const id = rec.failedJudges[i]\n if (typeof id !== 'string' \|\| id.length === 0) {\n throw new RunRecordValidationError(\n 'failedJudges entry must be a non-empty string',\n `${path}.failedJudges[${i}]`,\n )\n }\n }\n }\n\n if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n }\n}\n\n/\n Snapshot check for provider model identifiers. Accepts ISO and compact\n * dates, Router's `-MMDD` snapshots, one opaque `@token`, and Vertex-style\n * `:date-token` suffixes. Routing selectors such as `@preset/name` are not\n * immutable model identities.\n /\nexport function modelHasSnapshot(model: string): boolean {\n if (model.length === 0 \|\| model.trim() !== model) return false\n\n const opaqueAt = model.lastIndexOf('@')\n if (opaqueAt > 0) {\n const base = model.slice(0, opaqueAt)\n const token = model.slice(opaqueAt + 1)\n if (!base.includes('@') && /^[A-Za-z0-9](?:[A-Za-z0-9._-][A-Za-z0-9])?$/u.test(token)) {\n return true\n }\n }\n\n const isoDate = model.match(/-(\\d{4})-(\\d{2})-(\\d{2})$/u)\n if (isoDate && validSnapshotDate(isoDate[1]!, isoDate[2]!, isoDate[3]!)) return true\n\n const compactDate = model.match(/-(\\d{4})(\\d{2})(\\d{2})$/u)\n if (compactDate && validSnapshotDate(compactDate[1]!, compactDate[2]!, compactDate[3]!)) {\n return true\n }\n\n const routerDate = model.match(/-(\\d{2})(\\d{2})$/u)\n if (routerDate && validSnapshotDate(undefined, routerDate[1]!, routerDate[2]!)) return true\n\n return /:date-[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(model)\n}\n\nfunction validSnapshotDate(year: string \| undefined, month: string, day: string): boolean {\n const monthNumber = Number(month)\n const dayNumber = Number(day)\n if (!Number.isInteger(monthNumber) \|\| monthNumber < 1 \|\| monthNumber > 12) return false\n\n const yearNumber = year === undefined ? undefined : Number(year)\n const leapYear =\n yearNumber === undefined \|\|\n (yearNumber % 4 === 0 && (yearNumber % 100 !== 0 \|\| yearNumber % 400 === 0))\n const daysInMonth = [31, leapYear ? 29 : 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31]\n return Number.isInteger(dayNumber) && dayNumber >= 1 && dayNumber <= daysInMonth[monthNumber - 1]!\n}\n"],"mappings":";;;;;;AAiDA,MAAa,gBAAgB;;;;;;;;;AAuM7B,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAK5C,IACE,CAAC,iBAAiB,IAAI,KAAe,KACrC,EAAE,IAAI,UAAA,aAA2B,IAAI,oBAAoB,cAEzD,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,4EAA4E,cAAc,kDAC9G,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,gBAAgB,KAAA,KAAa,MAAM,gBAAgB,OAC3D,MAAM,IAAI,yBACR,gFACA,wBACF;CAEF,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAGzC,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,yBACR,kEACA,qBACF;CAMF,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;AAQA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,OAAO,OAAO;CAEzD,MAAM,WAAW,MAAM,YAAY,GAAG;CACtC,IAAI,WAAW,GAAG;EAChB,MAAM,OAAO,MAAM,MAAM,GAAG,QAAQ;EACpC,MAAM,QAAQ,MAAM,MAAM,WAAW,CAAC;EACtC,IAAI,CAAC,KAAK,SAAS,GAAG,KAAK,gDAAgD,KAAK,KAAK,GACnF,OAAO;CAEX;CAEA,MAAM,UAAU,MAAM,MAAM,4BAA4B;CACxD,IAAI,WAAW,kBAAkB,QAAQ,IAAK,QAAQ,IAAK,QAAQ,EAAG,GAAG,OAAO;CAEhF,MAAM,cAAc,MAAM,MAAM,0BAA0B;CAC1D,IAAI,eAAe,kBAAkB,YAAY,IAAK,YAAY,IAAK,YAAY,EAAG,GACpF,OAAO;CAGT,MAAM,aAAa,MAAM,MAAM,mBAAmB;CAClD,IAAI,cAAc,kBAAkB,KAAA,GAAW,WAAW,IAAK,WAAW,EAAG,GAAG,OAAO;CAEvF,OAAO,qDAAqD,KAAK,KAAK;AACxE;AAEA,SAAS,kBAAkB,MAA0B,OAAe,KAAsB;CACxF,MAAM,cAAc,OAAO,KAAK;CAChC,MAAM,YAAY,OAAO,GAAG;CAC5B,IAAI,CAAC,OAAO,UAAU,WAAW,KAAK,cAAc,KAAK,cAAc,IAAI,OAAO;CAElF,MAAM,aAAa,SAAS,KAAA,IAAY,KAAA,IAAY,OAAO,IAAI;CAI/D,MAAM,cAAc;EAAC;EAFnB,eAAe,KAAA,KACd,aAAa,MAAM,MAAM,aAAa,QAAQ,KAAK,aAAa,QAAQ,KACvC,KAAK;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;CAAE;CACnF,OAAO,OAAO,UAAU,SAAS,KAAK,aAAa,KAAK,aAAa,YAAY,cAAc;AACjG"}
1	+ {"version":3,"file":"run-record-tg_5H2IK.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/*\n Paper-grade RunRecord schema + runtime validator.\n \n Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n \n This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n \n Validate at the boundary:\n \n const rec = validateRunRecord(rawJson) // throws on missing\n * const ok = isRunRecord(rawJson) // boolean check\n * const rec = parseRunRecordSafe(rawJson) // { ok, value \| error }\n \n The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n /\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/* Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n * combined train+test pool that the optimizer is allowed to read. /\nexport type RunSplitTag = 'search' \| 'dev' \| 'holdout'\n\n/\n Explicit execution-lifecycle result for a run.\n \n This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n /\nexport type RunTerminalOutcome = 'succeeded' \| 'failed' \| 'cancelled' \| 'incomplete' \| 'unknown'\n\n/* Explicit model value for a row that never produced a served model snapshot. /\nexport const UNKNOWN_MODEL = 'unknown'\n\nexport interface RunTokenUsage {\n input: number\n /* All generated tokens charged as output, including reasoning tokens. /\n output: number\n /* Present only when one or more paid calls did not report token usage.\n * In that case, every numeric field is a known subtotal, not a measured total. /\n tokensKnown?: false\n /* Reasoning-token subset of `output`, when the provider reports it. /\n reasoning?: number\n /* Prompt tokens served from a provider cache. /\n cached?: number\n /* Prompt tokens written into a provider cache. /\n cacheWrite?: number\n}\n\n/* How a run's USD amount was obtained. /\nexport type RunCostProvenance = CostProvenance\n\nexport interface RunJudgeMetadata {\n model: string\n promptVersion: string\n /* [0,1] confidence the judge declared. Constant judge confidence\n * across many runs is a fallback signal (see `canary.ts`). /\n confidence: number\n /* True if the judge degraded to a fallback path (rules-only,\n * prior-call cache, etc.). The canary uses this to alert. /\n fallback: boolean\n}\n\n/\n Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n \n The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n \n `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n \n Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n /\nexport interface JudgeScoresRecord {\n /* Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. /\n perJudge: Record<string, Record<string, number>>\n /* Per-dim mean across judges. Convenience — derivable from `perJudge`. /\n perDimMean: Record<string, number>\n /* Composite mean across successful judges. Mirrors the task score only\n * when `failedJudges` is empty. /\n composite: number\n /* Judges that errored or returned an unparseable verdict. Recorded\n * by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n * not inferred from missing keys in `perJudge`. /\n failedJudges?: string[]\n /* Free-form notes the judges emitted (joined across judges or\n * first-judge only — consumer's choice). /\n notes?: string\n}\n\nexport interface RunOutcome {\n /* Score on the search/optimization split. Optional for holdout-only and\n * execution-only records. /\n searchScore?: number\n /* Score on the held-out split. Optional for search-only and execution-only\n * records. When both scores are absent, the run is explicitly unlabeled. /\n holdoutScore?: number\n /* Bag of any other metric the run produced — judge dimensions,\n * pass/fail counters, latency stats, etc. Numeric only — keeps\n * reporters honest. /\n raw: Record<string, number>\n /* Per-judge / per-dim breakdown. Consumers writing ensemble\n * judgements populate this; substrate primitives like\n * `interRaterReliability` and `corpusInterRaterAgreement` accept\n * these records as input. Optional — single-judge or scalar-only\n * runs leave it unset. /\n judgeScores?: JudgeScoresRecord\n /* Authenticity / realness verdict — did the run build the REAL thing on the\n * intended infra, or fake it (see `./authenticity`)? Optional: only domains\n * with an authenticity config populate it. Carried in the corpus so the\n * flywheel / off-policy learning can optimize for real completion, not gamed\n * pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n * must not count as a real success regardless of `score`. /\n realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/\n Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n \n Hash discipline:\n * - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n * model (after any steering bundle merge).\n * - `configHash` is the sha256 of the effective run config (model,\n * temperature, tools, judges, splits). The pair (promptHash,\n * configHash) uniquely identifies an experiment cell.\n \n Model snapshot discipline:\n * - successful rows MUST encode a snapshot version. Bare aliases like\n * `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n * Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n * - a failed, cancelled, incomplete, or otherwise unknown row may use\n * `UNKNOWN_MODEL` when no served model was observed. This is an explicit\n * absence marker, not a fabricated snapshot.\n /\nexport interface RunRecord {\n /* UUID for the run. /\n runId: string\n /* Logical experiment grouping (a treatment vs a baseline within\n * the same sweep should share `experimentId`). /\n experimentId: string\n /* Stable identifier for the candidate (variant) being run. The\n * promotion gate compares two `candidateId`s on matched items. /\n candidateId: string\n /* RNG seed for the run. Always recorded — silent re-seeding is\n * the most common cause of non-reproducible numbers. /\n seed: number\n /* Model identifier WITH snapshot version. /\n model: string\n /* sha256 of the effective prompt (post-steering). /\n promptHash: string\n /* sha256 of the effective config. /\n configHash: string\n /* Git SHA the harness was run from. /\n commitSha: string\n /* End-to-end wall-clock duration in milliseconds. /\n wallMs: number\n /* Time spent queued before execution started, if known. /\n queueMs?: number\n /* Total USD cost, or null when the producer could not capture one. /\n costUsd: number \| null\n /* Whether `costUsd` came from billing data, a price calculation, or is unavailable. /\n costProvenance: RunCostProvenance\n /* Token usage breakdown. /\n tokenUsage: RunTokenUsage\n /* Root-run or process terminal result. Never inferred from a child span. /\n terminalOutcome: RunTerminalOutcome\n /* Root-run or process failure reason. Valid only for a failed, cancelled,\n * or incomplete terminal result; never populated from a child span. /\n terminalFailureReason?: string\n /* Judge-side metadata, if a judge was used. /\n judgeMetadata?: RunJudgeMetadata\n /* Per-split scores + raw bag. /\n outcome: RunOutcome\n /* Canonical task-failure class drawn from the shared\n * `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n * evidence. Execution errors belong in\n * `outcome.raw.execution_error_count`. /\n failureClass?: FailureClass\n /* Free-form task-failure detail scoped under a non-success\n * `failureClass`. It is invalid without that class. /\n failureMode?: string\n /* Which split this run was drawn from. /\n splitTag: RunSplitTag\n /\n Stable scenario identifier the run observed or was scored against.\n * Comparison primitives match this identity rather than input order.\n /\n scenarioId: string\n /\n Canonical identity for the agent profile cell that produced this row:\n * profile artifact hash plus optional harness/model/prompt/reporting\n * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n * longitudinal reports by the complete source profile, not by a loose\n * candidate label or opaque config hash.\n /\n agentProfile?: AgentProfileCell\n}\n\n/\n Canonical task-result classification.\n \n A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n /\nexport type RunTaskFailure =\n \| { failureClass?: undefined; failureMode?: undefined }\n \| { failureClass: 'success'; failureMode?: undefined }\n \| {\n failureClass: Exclude<FailureClass, 'success'>\n failureMode?: string\n }\n\n/\n Return task quality, preferring held-out evidence when both scores exist.\n \n RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n /\nexport function runTaskScore(record: RunRecord): number \| undefined {\n const score = observedScore(record)\n return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n 'runId',\n 'experimentId',\n 'candidateId',\n 'seed',\n 'model',\n 'promptHash',\n 'configHash',\n 'commitSha',\n 'wallMs',\n 'costUsd',\n 'costProvenance',\n 'tokenUsage',\n 'terminalOutcome',\n 'outcome',\n 'splitTag',\n 'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n 'succeeded',\n 'failed',\n 'cancelled',\n 'incomplete',\n 'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n readonly path: string\n constructor(message: string, path = '') {\n super(path ? `${message} (at ${path})` : message)\n this.path = path\n }\n}\n\n/\n Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n /\nexport function validateRunRecord(input: unknown): RunRecord {\n if (input === null \|\| typeof input !== 'object') {\n throw new RunRecordValidationError('expected object')\n }\n const obj = input as Record<string, unknown>\n\n for (const key of MANDATORY_TOP_LEVEL) {\n if (!(key in obj)) {\n throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n }\n }\n\n expectString(obj.runId, 'runId')\n expectString(obj.experimentId, 'experimentId')\n expectString(obj.candidateId, 'candidateId')\n expectFiniteNumber(obj.seed, 'seed')\n expectString(obj.model, 'model')\n expectString(obj.promptHash, 'promptHash')\n expectString(obj.configHash, 'configHash')\n expectString(obj.commitSha, 'commitSha')\n expectNonNegativeNumber(obj.wallMs, 'wallMs')\n if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n validateCost(obj.costUsd, obj.costProvenance)\n\n // Snapshot discipline: successful rows require a served model snapshot.\n // Non-success rows may carry the explicit absence marker when execution\n // stopped before a model identity was observed.\n if (\n !modelHasSnapshot(obj.model as string) &&\n !(obj.model === UNKNOWN_MODEL && obj.terminalOutcome !== 'succeeded')\n ) {\n throw new RunRecordValidationError(\n `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD', or '${UNKNOWN_MODEL}' for a non-success row without a served model)`,\n 'model',\n )\n }\n\n // Token usage.\n const tu = obj.tokenUsage\n if (tu === null \|\| typeof tu !== 'object') {\n throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n }\n const tuRec = tu as Record<string, unknown>\n expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n if (tuRec.tokensKnown !== undefined && tuRec.tokensKnown !== false) {\n throw new RunRecordValidationError(\n 'tokensKnown must be false when present; omit it when token usage is complete',\n 'tokenUsage.tokensKnown',\n )\n }\n if (tuRec.reasoning !== undefined) {\n expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n throw new RunRecordValidationError(\n 'reasoning tokens must be a subset of output tokens',\n 'tokenUsage.reasoning',\n )\n }\n }\n if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n if (tuRec.cacheWrite !== undefined) {\n expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n }\n\n // Judge metadata, optional.\n if (obj.judgeMetadata !== undefined) {\n const jm = obj.judgeMetadata\n if (jm === null \|\| typeof jm !== 'object') {\n throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n }\n const jmRec = jm as Record<string, unknown>\n expectString(jmRec.model, 'judgeMetadata.model')\n expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n if (typeof jmRec.fallback !== 'boolean') {\n throw new RunRecordValidationError(\n 'judgeMetadata.fallback must be boolean',\n 'judgeMetadata.fallback',\n )\n }\n }\n\n // Outcome.\n const out = obj.outcome\n if (out === null \|\| typeof out !== 'object') {\n throw new RunRecordValidationError('outcome must be an object', 'outcome')\n }\n const outRec = out as Record<string, unknown>\n if (outRec.searchScore !== undefined)\n expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n if (outRec.holdoutScore !== undefined)\n expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n const raw = outRec.raw\n if (raw === null \|\| typeof raw !== 'object') {\n throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n }\n for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n expectFiniteNumber(v, `outcome.raw.${k}`)\n }\n // Realness verdict, optional.\n if (outRec.realness !== undefined) {\n const r = outRec.realness\n if (r === null \|\| typeof r !== 'object') {\n throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n }\n const rr = r as Record<string, unknown>\n expectFiniteNumber(rr.score, 'outcome.realness.score')\n if (typeof rr.gated !== 'boolean') {\n throw new RunRecordValidationError(\n 'outcome.realness.gated must be a boolean',\n 'outcome.realness.gated',\n )\n }\n }\n\n // Per-judge / per-dim breakdown, optional.\n if (outRec.judgeScores !== undefined) {\n validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n }\n\n // Failure mode optional.\n if (\n obj.failureClass !== undefined &&\n (typeof obj.failureClass !== 'string' \|\|\n !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n ) {\n throw new RunRecordValidationError(\n `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n 'failureClass',\n )\n }\n if (obj.failureMode !== undefined) {\n expectString(obj.failureMode, 'failureMode')\n if (obj.failureClass === undefined \|\| obj.failureClass === 'success') {\n throw new RunRecordValidationError(\n 'failureMode requires a non-success failureClass',\n 'failureMode',\n )\n }\n }\n\n if (\n typeof obj.terminalOutcome !== 'string' \|\|\n !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n ) {\n throw new RunRecordValidationError(\n `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n 'terminalOutcome',\n )\n }\n if (obj.terminalFailureReason !== undefined) {\n expectString(obj.terminalFailureReason, 'terminalFailureReason')\n if (\n obj.terminalOutcome !== 'failed' &&\n obj.terminalOutcome !== 'cancelled' &&\n obj.terminalOutcome !== 'incomplete'\n ) {\n throw new RunRecordValidationError(\n 'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n 'terminalFailureReason',\n )\n }\n }\n\n if (obj.agentProfile !== undefined) {\n try {\n const profile = validateAgentProfileCell(obj.agentProfile)\n if (profile.model !== undefined && profile.model !== obj.model) {\n throw new RunRecordValidationError(\n `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n 'agentProfile.model',\n )\n }\n if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n throw new RunRecordValidationError(\n `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n 'agentProfile.promptHash',\n )\n }\n } catch (error) {\n if (error instanceof RunRecordValidationError) throw error\n if (error instanceof Error) {\n throw new RunRecordValidationError(error.message, 'agentProfile')\n }\n throw error\n }\n }\n\n expectString(obj.scenarioId, 'scenarioId')\n\n // Split tag.\n if (typeof obj.splitTag !== 'string' \|\| !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n throw new RunRecordValidationError(\n `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n 'splitTag',\n )\n }\n\n return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n if (provenance === null \|\| typeof provenance !== 'object') {\n throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n }\n const value = provenance as Record<string, unknown>\n if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n throw new RunRecordValidationError(\n 'costProvenance.kind must be observed, estimated, or uncaptured',\n 'costProvenance.kind',\n )\n }\n // A record must never read as a total it cannot support, so an uncaptured\n // cost carries no number at all. A matrix `CellResult` keeps its known\n // subtotal instead, because a cost ceiling must charge the part it can see;\n // converting one to the other drops that subtotal.\n if (value.kind === 'uncaptured') {\n if (value.usd !== null) {\n throw new RunRecordValidationError(\n 'uncaptured costProvenance.usd must be null',\n 'costProvenance.usd',\n )\n }\n if (costUsd !== null) {\n throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n }\n return\n }\n expectNonNegativeNumber(costUsd, 'costUsd')\n expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n if (value.usd !== costUsd) {\n throw new RunRecordValidationError(\n 'costProvenance.usd must equal costUsd',\n 'costProvenance.usd',\n )\n }\n}\n\n/* Boolean validator — convenience for filtering arrays. /\nexport function isRunRecord(input: unknown): input is RunRecord {\n try {\n validateRunRecord(input)\n return true\n } catch {\n return false\n }\n}\n\n/* Non-throwing validator — returns a discriminated union. /\nexport function parseRunRecordSafe(\n input: unknown,\n): { ok: true; value: RunRecord } \| { ok: false; error: RunRecordValidationError } {\n try {\n return { ok: true, value: validateRunRecord(input) }\n } catch (e) {\n if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n throw e\n }\n}\n\n/* Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. /\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n const json = JSON.stringify(record)\n return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n if (typeof value !== 'string' \|\| value.length === 0) {\n throw new RunRecordValidationError(`expected non-empty string`, path)\n }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n if (typeof value !== 'number' \|\| !Number.isFinite(value)) {\n throw new RunRecordValidationError(`expected finite number`, path)\n }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n expectFiniteNumber(value, path)\n if ((value as number) < 0) {\n throw new RunRecordValidationError('expected non-negative number', path)\n }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n if (value === null \|\| typeof value !== 'object') {\n throw new RunRecordValidationError('judgeScores must be an object', path)\n }\n const rec = value as Record<string, unknown>\n\n const perJudge = rec.perJudge\n if (perJudge === null \|\| typeof perJudge !== 'object') {\n throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n }\n for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n if (dims === null \|\| typeof dims !== 'object') {\n throw new RunRecordValidationError(\n 'per-judge entry must be an object of dimension scores',\n `${path}.perJudge.${judgeId}`,\n )\n }\n for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n }\n }\n\n const perDimMean = rec.perDimMean\n if (perDimMean === null \|\| typeof perDimMean !== 'object') {\n throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n }\n for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n }\n\n expectFiniteNumber(rec.composite, `${path}.composite`)\n\n if (rec.failedJudges !== undefined) {\n if (!Array.isArray(rec.failedJudges)) {\n throw new RunRecordValidationError(\n 'failedJudges must be an array of strings',\n `${path}.failedJudges`,\n )\n }\n for (let i = 0; i < rec.failedJudges.length; i++) {\n const id = rec.failedJudges[i]\n if (typeof id !== 'string' \|\| id.length === 0) {\n throw new RunRecordValidationError(\n 'failedJudges entry must be a non-empty string',\n `${path}.failedJudges[${i}]`,\n )\n }\n }\n }\n\n if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n }\n}\n\n/\n Snapshot check for provider model identifiers. Accepts ISO and compact\n * dates, Router's `-MMDD` snapshots, one opaque `@token`, and Vertex-style\n * `:date-token` suffixes. Routing selectors such as `@preset/name` are not\n * immutable model identities.\n /\nexport function modelHasSnapshot(model: string): boolean {\n if (model.length === 0 \|\| model.trim() !== model) return false\n\n const opaqueAt = model.lastIndexOf('@')\n if (opaqueAt > 0) {\n const base = model.slice(0, opaqueAt)\n const token = model.slice(opaqueAt + 1)\n if (!base.includes('@') && /^[A-Za-z0-9](?:[A-Za-z0-9._-][A-Za-z0-9])?$/u.test(token)) {\n return true\n }\n }\n\n const isoDate = model.match(/-(\\d{4})-(\\d{2})-(\\d{2})$/u)\n if (isoDate && validSnapshotDate(isoDate[1]!, isoDate[2]!, isoDate[3]!)) return true\n\n const compactDate = model.match(/-(\\d{4})(\\d{2})(\\d{2})$/u)\n if (compactDate && validSnapshotDate(compactDate[1]!, compactDate[2]!, compactDate[3]!)) {\n return true\n }\n\n const routerDate = model.match(/-(\\d{2})(\\d{2})$/u)\n if (routerDate && validSnapshotDate(undefined, routerDate[1]!, routerDate[2]!)) return true\n\n return /:date-[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(model)\n}\n\nfunction validSnapshotDate(year: string \| undefined, month: string, day: string): boolean {\n const monthNumber = Number(month)\n const dayNumber = Number(day)\n if (!Number.isInteger(monthNumber) \|\| monthNumber < 1 \|\| monthNumber > 12) return false\n\n const yearNumber = year === undefined ? undefined : Number(year)\n const leapYear =\n yearNumber === undefined \|\|\n (yearNumber % 4 === 0 && (yearNumber % 100 !== 0 \|\| yearNumber % 400 === 0))\n const daysInMonth = [31, leapYear ? 29 : 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31]\n return Number.isInteger(dayNumber) && dayNumber >= 1 && dayNumber <= daysInMonth[monthNumber - 1]!\n}\n"],"mappings":";;;;;;AAiDA,MAAa,gBAAgB;;;;;;;;;AAuM7B,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAK5C,IACE,CAAC,iBAAiB,IAAI,KAAe,KACrC,EAAE,IAAI,UAAA,aAA2B,IAAI,oBAAoB,cAEzD,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,4EAA4E,cAAc,kDAC9G,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,gBAAgB,KAAA,KAAa,MAAM,gBAAgB,OAC3D,MAAM,IAAI,yBACR,gFACA,wBACF;CAEF,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAGzC,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,yBACR,kEACA,qBACF;CAMF,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;AAQA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,OAAO,OAAO;CAEzD,MAAM,WAAW,MAAM,YAAY,GAAG;CACtC,IAAI,WAAW,GAAG;EAChB,MAAM,OAAO,MAAM,MAAM,GAAG,QAAQ;EACpC,MAAM,QAAQ,MAAM,MAAM,WAAW,CAAC;EACtC,IAAI,CAAC,KAAK,SAAS,GAAG,KAAK,gDAAgD,KAAK,KAAK,GACnF,OAAO;CAEX;CAEA,MAAM,UAAU,MAAM,MAAM,4BAA4B;CACxD,IAAI,WAAW,kBAAkB,QAAQ,IAAK,QAAQ,IAAK,QAAQ,EAAG,GAAG,OAAO;CAEhF,MAAM,cAAc,MAAM,MAAM,0BAA0B;CAC1D,IAAI,eAAe,kBAAkB,YAAY,IAAK,YAAY,IAAK,YAAY,EAAG,GACpF,OAAO;CAGT,MAAM,aAAa,MAAM,MAAM,mBAAmB;CAClD,IAAI,cAAc,kBAAkB,KAAA,GAAW,WAAW,IAAK,WAAW,EAAG,GAAG,OAAO;CAEvF,OAAO,qDAAqD,KAAK,KAAK;AACxE;AAEA,SAAS,kBAAkB,MAA0B,OAAe,KAAsB;CACxF,MAAM,cAAc,OAAO,KAAK;CAChC,MAAM,YAAY,OAAO,GAAG;CAC5B,IAAI,CAAC,OAAO,UAAU,WAAW,KAAK,cAAc,KAAK,cAAc,IAAI,OAAO;CAElF,MAAM,aAAa,SAAS,KAAA,IAAY,KAAA,IAAY,OAAO,IAAI;CAI/D,MAAM,cAAc;EAAC;EAFnB,eAAe,KAAA,KACd,aAAa,MAAM,MAAM,aAAa,QAAQ,KAAK,aAAa,QAAQ,KACvC,KAAK;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;CAAE;CACnF,OAAO,OAAO,UAAU,SAAS,KAAK,aAAa,KAAK,aAAa,YAAY,cAAc;AACjG"}