@tangle-network/agent-eval 0.217.1 → 0.218.1

package/dist/define-agent-eval-BDxNlP3D.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"define-agent-eval-BDxNlP3D.js","names":[],"sources":["../src/contract/self-improve-reporting.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"sourcesContent":["import { campaignCellToRunRecord } from '../campaign/run-record'\nimport { pairedCampaignComposites } from '../campaign/score-utils'\nimport { surfaceContentHash } from '../campaign/surface-identity'\nimport type {\n  CampaignCellResult,\n  CampaignResult,\n  MutableSurface,\n  Scenario,\n} from '../campaign/types'\nimport { ValidationError } from '../errors'\nimport { modelHasSnapshot, type RunRecord, type RunSplitTag } from '../run-record'\n\nexport function pairedCompositeSummary<TArtifact, TScenario extends Scenario>(\n  baseline: CampaignResult<TArtifact, TScenario>,\n  winner: CampaignResult<TArtifact, TScenario>,\n  independentUnitByScenarioId?: ReadonlyMap<string, string>,\n) {\n  const scores = pairedCampaignComposites(baseline, winner, independentUnitByScenarioId)\n  const perScenario = (campaign: CampaignResult<TArtifact, TScenario>) =>\n    Object.fromEntries(\n      Object.entries(campaign.aggregates.byScenario).map(([id, aggregate]) => [\n        id,\n        aggregate.meanComposite,\n      ]),\n    )\n  return {\n    baseline: { compositeMean: scores.beforeMean, perScenario: perScenario(baseline) },\n    winner: { compositeMean: scores.afterMean, perScenario: perScenario(winner) },\n    baselineComposites: scores.before,\n    observations: scores.observations,\n  }\n}\n\n/** 32-bit FNV-1a over raw UTF-16 code units, rendered as hex for a cell key.\n *\n *  Frozen: the key names a persisted cell, so a change orphans every cell\n *  already written. This is a cell-key function, not a general-purpose hash. */\nfunction hashString(s: string): string {\n  let h = 2166136261 >>> 0\n  for (let i = 0; i < s.length; i++) {\n    h ^= s.charCodeAt(i)\n    h = Math.imul(h, 16777619) >>> 0\n  }\n  return h.toString(16).padStart(8, '0')\n}\n\n/**\n * Adapt campaign cells into the `RunRecord` shape `analyzeRuns()` consumes.\n * Each cell becomes one run; `candidateId` is the caller-supplied label so\n * baseline + winner pair cleanly on `(experimentId, scenarioId, seed)`.\n *\n * `promptHash` identifies the executed surface; `configHash` identifies the candidate label.\n */\nexport function cellsToRunRecords<TArtifact>(\n  cells: ReadonlyArray<CampaignCellResult<TArtifact>>,\n  candidateId: 'baseline' | 'winner',\n  runId: string,\n  surface: MutableSurface,\n  splitTag: RunSplitTag,\n  fallbackModel?: string,\n): RunRecord[] {\n  const promptHash = surfaceContentHash(surface)\n  const configHash = surfaceContentHash(candidateId)\n  return cells.map((cell) => {\n    const receiptModels = cell.resolvedModels ?? (cell.resolvedModel ? [cell.resolvedModel] : [])\n    if (receiptModels.length > 1) {\n      throw new ValidationError(\n        `selfImprove cell ${cell.cellId} used multiple agent models: ${receiptModels.join(', ')}`,\n      )\n    }\n    const model = receiptModels[0] ?? fallbackModel\n    if (!model) {\n      throw new ValidationError(\n        `selfImprove.model is required when cell ${cell.cellId} has no paid-call model receipt`,\n      )\n    }\n    if (!modelHasSnapshot(model)) {\n      throw new ValidationError(\n        `selfImprove model \"${model}\" lacks a snapshot version for cell ${cell.cellId}`,\n      )\n    }\n    return campaignCellToRunRecord(cell, {\n      runId: `${runId}::${candidateId}::${cell.cellId}`,\n      experimentId: runId,\n      candidateId,\n      // scenarioId is explicit; seed keeps repeated runs distinct.\n      seed:\n        cell.rep * 1_000_000 +\n        hashString(cell.scenarioId)\n          .slice(0, 6)\n          .split('')\n          .reduce((a, c) => (a * 31 + c.charCodeAt(0)) >>> 0, 0),\n      model,\n      promptHash,\n      configHash,\n      commitSha: 'cell',\n      splitTag,\n    })\n  })\n}\n","import { fileURLToPath } from 'node:url'\nimport { openAutoPr } from '../campaign/auto-pr'\nimport { campaignSplitDigest } from '../campaign/coverage'\nimport { evaluationUnitMap } from '../campaign/final-evidence'\nimport { defaultProductionGate } from '../campaign/gates/default-production-gate'\nimport {\n  type ComparisonCost,\n  combineComparisonCosts,\n  costFromLedgerSummary,\n} from '../campaign/optimization-cost'\nimport type { OptimizationMethodResult } from '../campaign/optimization-method'\nimport { executeOptimizationMethod } from '../campaign/optimization-method'\nimport { runFinalComparison } from '../campaign/presets/run-final-comparison'\nimport { campaignMeasurementDigest, canonicalDigest } from '../campaign/provenance'\nimport type { CampaignComparisonUnits } from '../campaign/score-utils'\nimport type { SearchHistoryCoverageRow } from '../campaign/search-history-receipt'\nimport type { CampaignStorage } from '../campaign/storage'\nimport { surfaceContentHash, surfaceHash } from '../campaign/surface-identity'\nimport type { Scenario } from '../campaign/types'\nimport type { CostLedgerHandle, CostLedgerSummary } from '../cost-ledger'\nimport { createCampaignEvidenceReceipt } from '../experiment/campaign-evidence'\nimport type { EvaluationClaim } from '../experiment/claim'\nimport type { EvidenceReceipt } from '../experiment/evidence-receipt'\nimport { shipBestEffort, shipSearchLedger } from '../hosted/search-shipper'\nimport { analyzeRuns } from './analyze-runs'\nimport type {\n  SelfImproveMethodOptions,\n  SelfImproveOptions,\n  SelfImproveProposerResult,\n} from './self-improve'\nimport { cellsToRunRecords, pairedCompositeSummary } from './self-improve-reporting'\n\nexport interface SelfImproveMethodProvenance {\n  schema: 'tangle.method-improvement'\n  recordDigest: `sha256:${string}`\n  runId: string\n  runDir: string\n  timestamp: string\n  baselineContentHash: string\n  winnerContentHash: string\n  diff: string\n  optimizationMethod: NonNullable<SelfImproveProposerResult<Scenario, unknown>['optimization']>\n  claim?: EvaluationClaim\n  evidence: {\n    trainSplitDigest: `sha256:${string}`\n    selectionSplitDigest: `sha256:${string}`\n    holdoutSplitDigest: `sha256:${string}`\n    baselineCampaignDigest: `sha256:${string}`\n    winnerCampaignDigest: `sha256:${string}`\n    costReceiptsDigest: `sha256:${string}`\n    neutralizedCampaignDigest?: `sha256:${string}`\n    holdoutObservations?: CampaignComparisonUnits\n  }\n  gate: Awaited<ReturnType<typeof runFinalComparison>>['gateResult']\n  holdout?: 'deferred'\n  baselineHoldoutComposite?: number\n  winnerHoldoutComposite?: number\n  heldOutLift?: number\n  cost: ComparisonCost\n  totalCostUsd: number\n  totalDurationMs: number\n}\n\nexport interface SelfImproveMethodResult<TScenario extends Scenario, TArtifact>\n  extends Omit<\n    SelfImproveProposerResult<TScenario, TArtifact>,\n    | 'mode'\n    | 'baseline'\n    | 'winner'\n    | 'provenance'\n    | 'generationsExplored'\n    | 'cost'\n    | 'raw'\n    | 'optimization'\n    | 'insight'\n  > {\n  evidence?: { baseline: EvidenceReceipt; winner: EvidenceReceipt }\n  searchHistoryCoverage: SearchHistoryCoverageRow\n  mode: 'method'\n  /** No final measurement exists when holdout is deferred. */\n  baseline: SelfImproveProposerResult<TScenario, TArtifact>['baseline'] | null\n  winner: Omit<SelfImproveProposerResult<TScenario, TArtifact>['winner'], 'compositeMean'> & {\n    compositeMean: number | null\n  }\n  provenance: SelfImproveMethodProvenance\n  optimization: NonNullable<SelfImproveProposerResult<TScenario, TArtifact>['optimization']>\n  /** Includes the method's reported cost and final measurements without double counting receipts. */\n  cost: ComparisonCost\n  /** Only calls actually recorded by the shared ledger have token and channel breakdowns. */\n  ledgerCost: CostLedgerSummary\n  insight?: SelfImproveProposerResult<TScenario, TArtifact>['insight']\n  raw: Awaited<ReturnType<typeof runFinalComparison<TScenario, TArtifact>>> & {\n    kind: 'method'\n    baselineSurface: SelfImproveProposerResult<TScenario, TArtifact>['raw']['baselineSurface']\n    winnerSurface: SelfImproveProposerResult<TScenario, TArtifact>['raw']['winnerSurface']\n    winnerSurfaceHash: string\n    method: OptimizationMethodResult\n    cost: ComparisonCost\n    prResult?: ReturnType<typeof openAutoPr>\n  }\n}\n\n/** Complete methods own search and selection; this path only measures their final choice. */\nexport async function runSelfImproveMethod<TScenario extends Scenario, TArtifact>(args: {\n  opts: SelfImproveOptions<TScenario, TArtifact> &\n    Pick<SelfImproveMethodOptions<TScenario, TArtifact>, 'method'>\n  train: TScenario[]\n  selection: TScenario[]\n  holdout: TScenario[]\n  costLedger: CostLedgerHandle\n  storage: CampaignStorage\n  runDir: string\n  startedAt: number\n}): Promise<SelfImproveMethodResult<TScenario, TArtifact>> {\n  const { opts, train, selection, holdout, costLedger, storage, runDir, startedAt } = args\n  const budget = opts.budget ?? {}\n  if (opts.autoOnPromote === 'pr' && (!opts.ghOwner || !opts.ghRepo)) {\n    throw new Error(\"selfImprove: autoOnPromote='pr' requires ghOwner + ghRepo\")\n  }\n  const evidenceContext = opts.evidence === undefined ? undefined : structuredClone(opts.evidence)\n  if (evidenceContext && budget.holdout === 'deferred')\n    throw new Error('selfImprove: evidence receipts require measured holdout')\n  const baselineSurface = structuredClone(opts.baselineSurface)\n  const judge = opts.judge\n  const {\n    selected,\n    cost: methodCost,\n    history,\n  } = await executeOptimizationMethod({\n    method: opts.method,\n    storage,\n    searchHistoryPolicy: opts.searchHistoryPolicy,\n    searchHistoryVerification: opts.searchHistoryVerification,\n    input: {\n      baselineSurface,\n      trainScenarios: train,\n      selectionScenarios: selection,\n      dispatchWithSurface: opts.agent,\n      judges: [judge],\n      runDir: `${runDir}/optimization/${opts.method.name.replace(/[^a-zA-Z0-9._-]/g, '_')}`,\n      seed: 42,\n      runOptions: {\n        storage,\n        maxConcurrency: budget.maxConcurrency ?? 2,\n        reps: budget.reps,\n        dispatchRef: opts.dispatchRef,\n        dispatchTimeoutMs: opts.dispatchTimeoutMs,\n        cellRetry: opts.cellRetry,\n        cellPlacement: opts.cellPlacement,\n        labeledStore: opts.labeledStore,\n        captureSource: opts.captureSource,\n        expectUsage: opts.expectUsage ?? 'assert',\n      },\n      costLedger,\n    },\n  })\n  const finalPhase = 'selfImprove.method-final'\n  const finalCost = (): ComparisonCost =>\n    combineComparisonCosts(\n      ['holdout.baseline', 'holdout.winner', 'holdout.neutralized', 'promotion.gate'].map(\n        (phase) => ({\n          label: phase,\n          cost: costFromLedgerSummary(costLedger.summary({ phase: `${finalPhase}.${phase}` })),\n        }),\n      ),\n    )\n  const totalCost = (): ComparisonCost =>\n    combineComparisonCosts([\n      { label: opts.method.name, cost: methodCost },\n      { label: 'final comparison', cost: finalCost() },\n    ])\n  const gate =\n    opts.gate ??\n    defaultProductionGate<TArtifact, TScenario>({\n      holdoutScenarios: holdout,\n      deltaThreshold: opts.claim?.minimumEffect ?? 0.05,\n      independentUnitByScenarioId: opts.claim ? evaluationUnitMap(opts.claim, holdout) : undefined,\n    })\n  const comparison = await runFinalComparison({\n    claim: opts.claim,\n    finalEvidence: opts.finalEvidence,\n    baselineSurface,\n    winnerSurface: selected.winnerSurface,\n    scenarios: holdout,\n    dispatchWithSurface: opts.agent,\n    dispatchRef: opts.dispatchRef,\n    judges: [judge],\n    holdout: budget.holdout,\n    neutralize: opts.neutralize,\n    gate: {\n      ...gate,\n      async decide(context) {\n        const cost = totalCost()\n        if (\n          budget.dollars !== undefined &&\n          (!cost.accountingComplete || cost.totalCostUsd > budget.dollars)\n        ) {\n          return {\n            decision: 'hold' as const,\n            reasons: ['complete method spend cannot satisfy the declared run budget'],\n            contributingGates: [\n              {\n                name: 'budget',\n                status: 'fail' as const,\n                detail: { cost, budgetUsd: budget.dollars },\n              },\n            ],\n          }\n        }\n        return gate.decide(context)\n      },\n    },\n    runDir,\n    storage,\n    costLedger,\n    costPhase: finalPhase,\n    reps: budget.reps,\n    maxConcurrency: budget.maxConcurrency ?? 2,\n    dispatchTimeoutMs: opts.dispatchTimeoutMs,\n    cellRetry: opts.cellRetry,\n    cellPlacement: opts.cellPlacement,\n    expectUsage: opts.expectUsage ?? 'assert',\n    label: 'selfImprove',\n  })\n  const evidence = evidenceContext\n    ? {\n        baseline: createCampaignEvidenceReceipt({\n          campaign: comparison.baselineOnHoldout,\n          surface: baselineSurface,\n          context: evidenceContext,\n        }),\n        winner: createCampaignEvidenceReceipt({\n          campaign: comparison.winnerOnHoldout,\n          surface: selected.winnerSurface,\n          context: evidenceContext,\n        }),\n      }\n    : undefined\n  const deferred = comparison.holdout === 'deferred'\n  const report = deferred\n    ? null\n    : pairedCompositeSummary(\n        comparison.baselineOnHoldout,\n        comparison.winnerOnHoldout,\n        opts.claim ? evaluationUnitMap(opts.claim, holdout) : undefined,\n      )\n  const baseline = report?.baseline ?? null\n  const winner = report?.winner ?? null\n  const lift = baseline && winner ? winner.compositeMean - baseline.compositeMean : undefined\n  const insight = deferred\n    ? undefined\n    : await analyzeRuns({\n        runs: [\n          ...cellsToRunRecords(\n            comparison.baselineOnHoldout.cells,\n            'baseline',\n            runDir,\n            baselineSurface,\n            'holdout',\n            opts.model,\n          ),\n          ...(comparison.winnerOnHoldout === comparison.baselineOnHoldout\n            ? []\n            : cellsToRunRecords(\n                comparison.winnerOnHoldout.cells,\n                'winner',\n                runDir,\n                selected.winnerSurface,\n                'holdout',\n                opts.model,\n              )),\n        ],\n        baselineCandidateId: 'baseline',\n        independentUnitByScenarioId: opts.claim\n          ? evaluationUnitMap(opts.claim, holdout)\n          : undefined,\n        decisionThreshold: opts.claim?.minimumEffect ?? 0.05,\n        ...(comparison.winnerOnHoldout === comparison.baselineOnHoldout\n          ? {}\n          : { candidateCandidateId: 'winner' }),\n      })\n  const cost = totalCost()\n  const optimization = {\n    name: opts.method.name,\n    cost: methodCost,\n    ...(selected.durationMs === undefined ? {} : { durationMs: selected.durationMs }),\n    ...(selected.provenance === undefined ? {} : { provenance: selected.provenance }),\n  }\n  const durationMs = Date.now() - startedAt\n  const receipts = costLedger.list()\n  const record = {\n    schema: 'tangle.method-improvement' as const,\n    runId: `${runDir}#${startedAt}`,\n    runDir,\n    timestamp: new Date(startedAt).toISOString(),\n    baselineContentHash: surfaceContentHash(baselineSurface),\n    winnerContentHash: surfaceContentHash(selected.winnerSurface),\n    diff: comparison.promotedDiff,\n    optimizationMethod: optimization,\n    ...(opts.claim ? { claim: opts.claim } : {}),\n    evidence: {\n      trainSplitDigest: campaignSplitDigest(train, budget.reps ?? 1),\n      selectionSplitDigest: campaignSplitDigest(selection, budget.reps ?? 1),\n      holdoutSplitDigest: campaignSplitDigest(holdout, budget.reps ?? 1),\n      baselineCampaignDigest: campaignMeasurementDigest(comparison.baselineOnHoldout),\n      winnerCampaignDigest: campaignMeasurementDigest(comparison.winnerOnHoldout),\n      costReceiptsDigest: canonicalDigest(receipts),\n      ...(report ? { holdoutObservations: report.observations } : {}),\n      ...(comparison.neutralizedOnHoldout\n        ? { neutralizedCampaignDigest: campaignMeasurementDigest(comparison.neutralizedOnHoldout) }\n        : {}),\n    },\n    gate: comparison.gateResult,\n    ...(deferred ? { holdout: 'deferred' as const } : {}),\n    ...(baseline && winner\n      ? {\n          baselineHoldoutComposite: baseline.compositeMean,\n          winnerHoldoutComposite: winner.compositeMean,\n          heldOutLift: lift,\n        }\n      : {}),\n    cost,\n    totalCostUsd: cost.totalCostUsd,\n    totalDurationMs: durationMs,\n  }\n  // The persisted JSON and returned record must have the same canonical content.\n  const serialized: Omit<SelfImproveMethodProvenance, 'recordDigest'> = JSON.parse(\n    JSON.stringify(record),\n  )\n  const provenance: SelfImproveMethodProvenance = {\n    ...serialized,\n    recordDigest: canonicalDigest(serialized),\n  }\n  storage.ensureDir(runDir)\n  storage.write(`${runDir}/method-provenance.json`, JSON.stringify(provenance, null, 2))\n  opts.onProvenance?.(provenance)\n  opts.onProgress?.({\n    kind: 'gate.decided',\n    decision: comparison.gateResult.decision,\n    ...(lift === undefined ? {} : { lift }),\n  })\n  const prResult =\n    opts.autoOnPromote === 'pr' && comparison.gateResult.decision === 'ship'\n      ? openAutoPr({\n          result: comparison.winnerOnHoldout,\n          gate: comparison.gateResult,\n          promotedDiff: comparison.promotedDiff,\n          ghOwner: opts.ghOwner!,\n          ghRepo: opts.ghRepo!,\n        })\n      : undefined\n  const result: SelfImproveMethodResult<TScenario, TArtifact> = {\n    ...(evidence ? { evidence } : {}),\n    mode: 'method',\n    ...(opts.claim ? { claim: opts.claim } : {}),\n    ...(comparison.finalEvidence ? { finalEvidence: comparison.finalEvidence } : {}),\n    searchHistoryCoverage: history,\n    baseline,\n    winner: {\n      surface: selected.winnerSurface,\n      compositeMean: winner?.compositeMean ?? null,\n      perScenario: winner?.perScenario ?? {},\n      label: opts.method.name,\n    },\n    ...(lift === undefined ? {} : { lift }),\n    diff: comparison.promotedDiff,\n    gateDecision: comparison.gateResult.decision,\n    provenance,\n    optimization,\n    cost,\n    ledgerCost: costLedger.summary(),\n    totalCostUsd: cost.totalCostUsd,\n    durationMs,\n    receipts,\n    ...(insight ? { insight } : {}),\n    ...(selected.searchHistory ? { searchHistory: selected.searchHistory } : {}),\n    raw: {\n      ...comparison,\n      kind: 'method',\n      baselineSurface,\n      winnerSurface: selected.winnerSurface,\n      winnerSurfaceHash: surfaceHash(selected.winnerSurface),\n      method: selected,\n      cost,\n      ...(prResult ? { prResult } : {}),\n    },\n  }\n  if (opts.hostedTenant) {\n    const receipt = selected.searchHistory\n    if (receipt) {\n      const path = fileURLToPath(receipt.ledger.uri)\n      await shipBestEffort(\n        () =>\n          shipSearchLedger({\n            tenant: opts.hostedTenant!,\n            ledger: { path, searchId: receipt.summary.searchId },\n            runKind: 'optimization',\n          }),\n        path,\n      )\n    } else {\n      console.warn(\n        `[agent-eval] hostedTenant: method ${opts.method.name} recorded no search ledger, so nothing was shipped`,\n      )\n    }\n  }\n  return result\n}\n","import { defineEvaluationClaim, type EvaluationClaim } from '../experiment/claim'\n/**\n * Run one complete improvement job.\n *\n * A caller-owned `proposer` can generate candidates across local generations.\n * An external `method`, such as official GEPA or SkillOpt, owns its complete\n * search and returns one candidate. Both paths remeasure the selected candidate\n * against cases that candidate generation never receives.\n */\n\nimport type { ProposalFinding } from '../analyst/types'\nimport {\n  assertIndependentEvaluationSplit,\n  captureFinalEvidencePolicy,\n  evaluationUnitMap,\n  type FinalEvidencePolicy,\n  type FinalEvidenceUse,\n  reserveFinalEvidence,\n} from '../campaign/final-evidence'\nimport { defaultProductionGate } from '../campaign/gates/default-production-gate'\nimport { type PowerPreflight, powerPreflight } from '../campaign/gates/power-preflight'\nimport { captureJudge } from '../campaign/judge-snapshot'\nimport type {\n  OptimizationMethod,\n  OptimizationMethodProvenance,\n  OptimizationMethodResult,\n} from '../campaign/optimization-method'\nimport {\n  type RunImprovementLoopResult,\n  runImprovementLoop,\n} from '../campaign/presets/run-improvement-loop'\nimport type {\n  PremeasuredOptimizationBaseline,\n  RunOptimizationOptions,\n} from '../campaign/presets/run-optimization'\nimport {\n  emitLoopProvenance,\n  type LoopProvenanceRecord,\n  loopProvenanceArgsFromResult,\n} from '../campaign/provenance'\nimport type { CampaignCellRetryPolicy } from '../campaign/run-campaign'\nimport { resolveRunDir } from '../campaign/run-dir'\nimport type { SearchHistoryReceipt } from '../campaign/search-history-receipt'\nimport {\n  assertSearchHistoryAdmissionOptions,\n  type SearchHistoryAdmissionOptions,\n} from '../campaign/search-history-receipt'\nimport {\n  type CampaignStorage,\n  createRunCostLedger,\n  fsCampaignStorage,\n  inMemoryCampaignStorage,\n} from '../campaign/storage'\nimport type {\n  DispatchContext,\n  Gate,\n  JudgeConfig,\n  LabeledScenarioStore,\n  MutableSurface,\n  Scenario,\n  SurfaceProposer,\n} from '../campaign/types'\nimport type { CostLedgerHandle, CostLedgerSummary, CostReceipt } from '../cost-ledger'\nimport type { CampaignEvidenceContext } from '../experiment/campaign-evidence'\nimport { createHostedClient, type HostedTenant } from '../hosted/client'\nimport { shipBestEffort, startSearchShipper } from '../hosted/search-shipper'\nimport type { RunSplitTag } from '../run-record'\nimport { analyzeRuns } from './analyze-runs'\nimport type { InsightReport } from './insight-report'\nimport {\n  runSelfImproveMethod,\n  type SelfImproveMethodProvenance,\n  type SelfImproveMethodResult,\n} from './self-improve-method'\nimport { cellsToRunRecords, pairedCompositeSummary } from './self-improve-reporting'\n\nexport type { SelfImproveMethodProvenance, SelfImproveMethodResult } from './self-improve-method'\n\nexport interface SelfImproveBudget {\n  /** Hard spend cap across the full run. Each paid call reserves its enforced\n   *  maximum before dispatch, so completed spend cannot cross this amount. */\n  dollars?: number\n  /** Proposer generations. Default: 3. External methods own their rounds and\n   *  require this value to be omitted or set to 1. Set 0 only for a\n   *  proposer-free baseline run. */\n  generations?: number\n  /** Candidates the proposer emits per generation. Default 2. */\n  populationSize?: number\n  /** Max concurrent cells across the loop. Default 2. */\n  maxConcurrency?: number\n  /** Candidate campaigns scored in parallel. Default 1. Total concurrent\n   *  cells are bounded by `candidateConcurrency * maxConcurrency`. */\n  candidateConcurrency?: number\n  /** Fraction of `scenarios` held out from training, used for the gate.\n   *  Default 0.25. Ignored when `holdoutScenarios` is set explicitly. */\n  holdoutFraction?: number\n  /** Fraction of the non-final cases reserved for method selection.\n   *  Default 0.25. Used only with `method` and ignored when\n   *  `selectionScenarios` is supplied explicitly. */\n  selectionFraction?: number\n  /** Explicit held-out scenarios; overrides `holdoutFraction`. */\n  holdoutScenarios?: Scenario[]\n  /** Holdout policy. Default `'measured'`: split, re-score baseline vs winner\n   *  on the held-out set, gate on that comparison. `'deferred'`: run the\n   *  improvement-set campaigns + search promotion, dispatch ZERO holdout cells,\n   *  force the gate to `'hold'`, return `lift: undefined`, and record\n   *  `holdout: 'deferred'` in the provenance record — for callers that measure\n   *  the held-out comparison in a separate later run instead of faking a\n   *  static holdout scenario and recording a meaningless lift. Unless\n   *  `holdoutScenarios` reserves an explicit set, ALL scenarios train. */\n  holdout?: 'measured' | 'deferred'\n  /** Repeated executions per scenario. A claim's independent-unit count stays unchanged. Default 1. */\n  reps?: number\n  /** DEPTH dial forwarded to the proposer's `propose()` as\n   *  `ctx.maxImprovementShots` — max iterations an agentic candidate generator\n   *  may take per candidate (verify-in-session retries). Unset ⇒ the\n   *  proposer's own default. */\n  maxImprovementShots?: number\n}\n\nexport type SelfImproveProgressEvent =\n  | { kind: 'baseline.started'; scenarios: number }\n  | { kind: 'baseline.completed'; compositeMean: number; durationMs: number }\n  | { kind: 'generation.started'; index: number; populationSize: number }\n  | { kind: 'generation.completed'; index: number; bestComposite: number; durationMs: number }\n  // `lift` is absent when `budget.holdout === 'deferred'` — no held-out\n  // measurement ran, and the search-split delta must not masquerade as one.\n  | { kind: 'gate.decided'; decision: string; lift?: number }\n  | { kind: 'power.estimated'; n: number; sd: number; mde: number; underpowered: boolean }\n\nexport interface SelfImproveOptions<TScenario extends Scenario, TArtifact>\n  extends SearchHistoryAdmissionOptions {\n  /**\n   * Your agent — a function that takes the current `MutableSurface`\n   * (typically a system prompt the loop is optimizing) plus the\n   * scenario + cell ctx, and returns the artifact your judge scores.\n   *\n   * Same shape as `RunOptimizationOptions.dispatchWithSurface`. Wrap a\n   * plain `Dispatch` if you don't have a surface seam:\n   *\n   *   agent: (_surface, scenario, ctx) => yourPlainDispatch(scenario, ctx)\n   *\n   * That mode evaluates without mutating any surface — useful as a\n   * baseline-only run (set `budget.generations = 0`).\n   */\n  agent: (surface: MutableSurface, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>\n\n  /**\n   * Snapshot-bearing model identity for agents that do not report a paid-call\n   * receipt through `ctx.cost.runPaidCall()`.\n   *\n   * Omit this when every cell reports its concrete model in a receipt.\n   */\n  model?: string\n  /** Version of execution behavior outside the candidate surface, used by measurement caches. */\n  dispatchRef?: string\n\n  /** Scenarios to evaluate against. Train/holdout split is computed from\n   *  these unless `budget.holdoutScenarios` is set explicitly. */\n  scenarios: TScenario[]\n\n  /** Judge that scores artifacts. Bring your own, or wrap `llmJudge`. */\n  judge: JudgeConfig<TArtifact, TScenario>\n\n  /** Starting surface — system prompt, JSON config, anything `MutableSurface`\n   *  accepts. The proposer mutates this each generation. */\n  baselineSurface: MutableSurface\n\n  /** Budget + loop shape. All fields optional. */\n  budget?: SelfImproveBudget\n\n  /**\n   * Complete prior measurement of `baselineSurface` over the TRAIN split.\n   * Forwarded to the loop body, which validates its surface hash, scenario\n   * split, seed (42), reps, evaluator manifest, and coverage, then skips the baseline search\n   * campaign entirely — no baseline dispatch, no resumability lookup. The\n   * train split is `scenarios` minus the holdout split, so premeasure with\n   * exactly that scenario set (explicit `budget.holdoutScenarios`, or\n   * `budget.holdout: 'deferred'` with no reserved set, makes the train split\n   * deterministic). Prior spend stays in the imported campaign aggregates and\n   * is not re-added to this run's cost ledger. Premeasure with\n   * `dispatchRef: surfaceDispatchRef(baselineSurface, dispatchRef)` and the same judge revision.\n   */\n  premeasuredBaseline?: PremeasuredOptimizationBaseline<TArtifact, TScenario>\n\n  /**\n   * Candidate generator for this local generation loop.\n   * Required when `budget.generations` is greater than zero.\n   */\n  proposer?: SurfaceProposer<ProposalFinding>\n\n  /**\n   * Complete optimization method, such as official GEPA or SkillOpt.\n   * The method receives disjoint train and selection cases and never receives\n   * the final comparison cases. Mutually exclusive with `proposer`.\n   */\n  method?: OptimizationMethod<TScenario, TArtifact>\n\n  /** Explicit method-selection cases. They must also appear in `scenarios`\n   *  and must not overlap the final comparison cases. */\n  selectionScenarios?: TScenario[]\n\n  /** Custom gate. Default is `defaultProductionGate` with\n   *  `deltaThreshold: 0.05` on the held-out split. */\n  gate?: Gate<TArtifact, TScenario>\n\n  /** Placebo control. When supplied AND the winner differs from baseline, the\n   *  loop scores a THIRD held-out arm: the winner surface with its content\n   *  footprint-matched-blanked by this fn (typically via `neutralizeText`). Its\n   *  scores reach the gate as `ctx.neutralizedJudgeScores`, letting a\n   *  `neutralizationGate` reject a win whose lift survives blanking the content\n   *  (decorative — driven by footprint, not content). Costs one extra held-out\n   *  campaign; omit to skip. Compose `neutralizationGate` into `gate` to act on it. */\n  neutralize?: (winnerSurface: MutableSurface, baselineSurface: MutableSurface) => MutableSurface\n\n  /** Storage backend. A filesystem run directory uses `fsCampaignStorage()`;\n   *  a `mem://` directory uses in-memory storage. External methods default to\n   *  a filesystem directory because their official state must survive. */\n  storage?: CampaignStorage\n\n  /** Run directory. Proposer mode defaults to\n   *  `mem://selfImprove-<timestamp>`. External method mode defaults to\n   *  `.agent-eval/runs/self-improve-<timestamp>`. */\n  runDir?: string\n\n  /** Fires once the durable provenance record is written.\n   *  Receives the structured record for inline assertions / custom routing. */\n  onProvenance?: (record: LoopProvenanceRecord | SelfImproveMethodProvenance) => void\n\n  /** Distributed execution seam — same as `RunCampaignOptions.cellPlacement`.\n   *  Returns an opaque placement key the substrate forwards to your agent\n   *  as `ctx.placement`. Combined with `httpDispatch` from\n   *  `/adapters/http`, fans cells across regions. */\n  cellPlacement?: (input: {\n    scenario: TScenario\n    rep: number\n    generation?: number\n  }) => string | undefined\n\n  /** Per-cell agent dispatch deadline, applied to baseline, candidate, and\n   *  held-out campaigns. Default 600_000 ms. Set 0 to disable. */\n  dispatchTimeoutMs?: number\n\n  /** Bounded in-run retry of failed cells — same as\n   *  `RunCampaignOptions.cellRetry`, applied to baseline, candidate, and\n   *  held-out campaigns. Pair with `transientDispatchFailure()` so a\n   *  transport hiccup (a router 503, a dropped stream) is re-dispatched in\n   *  the same slot instead of leaving holdout coverage incomplete. Absent by\n   *  default: a failed cell is final. */\n  cellRetry?: CampaignCellRetryPolicy\n\n  /** Streaming hook — fires on baseline + each generation + gate decision.\n   *  Consumer routes events wherever (UI, dashboard, logs). */\n  onProgress?: (event: SelfImproveProgressEvent) => void\n\n  /** Auto-promotion behavior on a ship decision. Default `'none'` — we\n   *  return the winner; you ship it however you ship. `'pr'` opens a\n   *  GitHub PR via `openAutoPr`; requires `ghOwner` + `ghRepo`. */\n  autoOnPromote?: 'pr' | 'none'\n  ghOwner?: string\n  ghRepo?: string\n\n  /**\n   * Opt-in: ship this run's search ledger to a hosted store (ours, your\n   * self-hosted one, or any implementation of `docs/hosted-ingest-spec.md`).\n   * In proposer mode it requires `searchLedger`, and the ledger ships while\n   * the loop runs; in method mode the ledger the method recorded ships when\n   * the method returns. A failed ship is logged and never fails the loop: the\n   * local ledger is the source of truth, and `agent-eval search ship <ledger>`\n   * resumes from the store's head.\n   *\n   * For our orchestrator: `{ endpoint: 'https://orchestrator.tangle.tools', apiKey, tenantId }`.\n   * For your self-hosted one: see `examples/hosted-ingest-server/`.\n   */\n  hostedTenant?: HostedTenant\n\n  /** Capture every search artifact and judge score to this store.\n   *  The store is output only and is never exposed to candidate generation.\n   *  Pass `'off'` to disable. Default: off. */\n  labeledStore?: LabeledScenarioStore | 'off'\n\n  /** Capture-source tag for `labeledStore`. Default `'eval-run'`. */\n  captureSource?: 'production-trace' | 'eval-run' | 'manual' | 'red-team' | 'synthetic'\n\n  /**\n   * Per-cell backend-integrity expectation — the fail-loud guard. A cell that\n   * produced an artifact but reported `costUsd === 0` AND zero tokens is a\n   * stub. Modes: `'assert'` throws on the first such cell, `'warn'` logs it,\n   * `'off'` skips the check (offline/replay). Default `'assert'` — `selfImprove`\n   * is the real-run path, so a stub fails loud rather than scoring a clean 0.\n   */\n  expectUsage?: 'assert' | 'warn' | 'off'\n\n  /**\n   * Per-generation findings producer. Runs once on the baseline campaign (as\n   * `generation: -1`) before generation 0 proposes — so single-generation runs\n   * propose with trace context — and again after each generation is scored;\n   * whatever it returns REPLACES the proposer's `findings` for the next\n   * `propose()`. Plug a trace-analyst registry / HALO here. When absent,\n   * findings stay `opts.findings`.\n   */\n  analyzeGeneration?: RunOptimizationOptions<TScenario, TArtifact>['analyzeGeneration']\n\n  /** Static findings forwarded to the proposer's `propose()` as `ctx.findings`\n   *  (a findings-grounded proposer consumes them). Default: none. */\n  findings?: ProposalFinding[]\n\n  /** Which measured surface each proposal extends and which one the run\n   *  keeps. Default `incumbent()`: the hill climb. Pass\n   *  `crowdedFrontierParent({ seed })` to draw parents from the Pareto\n   *  frontier. Proposer mode only. See `RunOptimizationOptions.policy`. */\n  policy?: RunOptimizationOptions<TScenario, TArtifact>['policy']\n\n  /** Where the run's search ledger goes and the identities it records. The\n   *  ledger is always written; see `RunOptimizationOptions.searchLedger`. */\n  searchLedger?: RunOptimizationOptions<TScenario, TArtifact>['searchLedger']\n  /** Complete-method final measurement receipts; authority and environment remain caller-owned. */\n  evidence?: CampaignEvidenceContext\n  /** Declare the population and independent units without requiring fresh data. */\n  claim?: EvaluationClaim\n  /** Opt into durable fresh-evidence consumption for this final comparison. */\n  finalEvidence?: FinalEvidencePolicy\n}\n\nexport interface SelfImproveProposerResult<TScenario extends Scenario, TArtifact> {\n  mode: 'proposer'\n  claim?: EvaluationClaim\n  finalEvidence?: FinalEvidenceUse\n  /** Composite mean across all scenarios, baseline run. When\n   *  `budget.holdout === 'deferred'` this is measured on the improvement\n   *  (search) split — no holdout campaign ran. */\n  baseline: {\n    compositeMean: number\n    perScenario: Record<string, number>\n  }\n  /** Composite mean on the held-out set, winner run. When\n   *  `budget.holdout === 'deferred'` this is the winner's improvement-set\n   *  (search) measurement — no holdout campaign ran. */\n  winner: {\n    compositeMean: number\n    perScenario: Record<string, number>\n    surface: MutableSurface\n    /** Proposer label for the promoted change. Absent ⇒ winner == baseline or\n     *  a bare-surface mutator. */\n    label?: string\n    /** Proposer rationale — the \"because Z\" that motivated the promoted change.\n     *  Threaded from the proposer's `ProposedCandidate` through the loop.\n     *  Absent ⇒ winner == baseline. */\n    rationale?: string\n  }\n  /** `winner.compositeMean - baselineOnHoldout.compositeMean`. Positive\n   *  means the gate observed improvement. Absent iff\n   *  `budget.holdout === 'deferred'` — no held-out measurement ran, so there\n   *  is no lift to report (never a fabricated 0). */\n  lift?: number\n  /** The explicit baseline→winner unified diff. Always present (empty string\n   *  when winner == baseline). */\n  diff: string\n  /** Durable, queryable provenance record: candidate→cell→gate→promote chain +\n   *  rationale + diff + backend provenance. The artifact the hosted ingest\n   *  path stores; the +lift RECOMPUTES from `record.heldOutLift`. */\n  provenance: LoopProvenanceRecord\n  /** `defaultProductionGate.decide()` result. */\n  gateDecision: 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling'\n  /** Number of generations actually explored (may be less than the\n   *  budget if the proposer gave up early). */\n  generationsExplored: number\n  /** Wall-clock total. */\n  durationMs: number\n  /** Total newly observed cost across the full run. */\n  totalCostUsd: number\n  /** Canonical run-wide spend summary. */\n  cost: CostLedgerSummary\n  /** Run-wide receipts across proposal, search, holdout, judging, analysis,\n   *  and promotion work, with phase and actor attribution. */\n  receipts: CostReceipt[]\n  /** Bounded proof envelope over this run's search ledger: always present in\n   *  proposer mode, and in method mode when the method recorded a search. */\n  searchHistory?: SearchHistoryReceipt\n  /** Exact external method and source identity, when `method` was used. */\n  optimization?: {\n    name: string\n    cost: OptimizationMethodResult['cost']\n    durationMs?: number\n    provenance?: OptimizationMethodProvenance\n  }\n  /**\n   * Rigor packet: distributional summary, paired-bootstrap lift CI,\n   * judge stats, contamination check, recommendations. Wired through\n   * `analyzeRuns()` on the baseline + winner cells of the campaign.\n   * Hosted-tier dashboards render this as the v3-vs-v4 decision view.\n   */\n  insight: InsightReport\n  /** Approximate detectable lift from baseline holdout observations.\n   *  Declared independent units determine n and baseline variance.\n   *  Absent with fewer than three observations or a deferred holdout. */\n  power?: PowerPreflight\n  /**\n   * Raw substrate result for advanced inspection — full per-generation\n   * candidates, full campaign artifacts, all judge scores. Useful for\n   * debugging or reporting beyond the summary.\n   */\n  raw: RunImprovementLoopResult<TArtifact, TScenario>\n}\n\nexport type SelfImproveResult<TScenario extends Scenario, TArtifact> =\n  | SelfImproveProposerResult<TScenario, TArtifact>\n  | SelfImproveMethodResult<TScenario, TArtifact>\n\nexport type SelfImproveMethodOptions<TScenario extends Scenario, TArtifact> = Omit<\n  SelfImproveOptions<TScenario, TArtifact>,\n  'proposer' | 'onProvenance'\n> & {\n  method: OptimizationMethod<TScenario, TArtifact>\n  proposer?: never\n  onProvenance?: (record: SelfImproveMethodProvenance) => void\n}\n\nexport type SelfImproveProposerOptions<TScenario extends Scenario, TArtifact> = Omit<\n  SelfImproveOptions<TScenario, TArtifact>,\n  'method' | 'onProvenance'\n> & {\n  method?: never\n  onProvenance?: (record: LoopProvenanceRecord) => void\n}\n\n/** Failed self-improvement run with an immutable receipt snapshot. */\nexport class SelfImproveRunError extends Error {\n  readonly cost: CostLedgerSummary\n  readonly receipts: CostReceipt[]\n\n  constructor(cause: unknown, ledger: CostLedgerHandle) {\n    const original = cause instanceof Error ? cause : new Error(String(cause))\n    super(original.message, { cause: original })\n    this.name = 'SelfImproveRunError'\n    this.cost = ledger.summary()\n    this.receipts = ledger.list()\n  }\n}\n\nfunction assertSelfImproveSearchMode<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n): void {\n  assertSearchHistoryAdmissionOptions(opts)\n  if (\n    !opts.method &&\n    (opts.searchHistoryPolicy !== undefined ||\n      opts.searchHistoryVerification !== undefined ||\n      opts.evidence !== undefined)\n  )\n    throw new Error('selfImprove: search history admission requires method')\n  if (opts.method && opts.proposer) {\n    throw new Error('selfImprove: method and proposer are mutually exclusive')\n  }\n  if (!opts.method) {\n    if (opts.selectionScenarios !== undefined) {\n      throw new Error('selfImprove: selectionScenarios requires method')\n    }\n    if (opts.hostedTenant && !opts.searchLedger) {\n      throw new Error(\n        'selfImprove: hostedTenant ships the search ledger; pass searchLedger (docs/search-ledger.md)',\n      )\n    }\n    return\n  }\n  if (\n    typeof opts.method.name !== 'string' ||\n    !opts.method.name.trim() ||\n    opts.method.name.trim() !== opts.method.name ||\n    typeof opts.method.optimize !== 'function'\n  ) {\n    throw new Error('selfImprove: method must have a trimmed name and optimize(input)')\n  }\n  const budget = opts.budget\n  if (budget?.generations !== undefined) {\n    throw new Error(\n      'selfImprove: method owns its rounds; budget.generations applies only to proposer mode',\n    )\n  }\n  if (budget?.populationSize !== undefined) {\n    throw new Error(\n      'selfImprove: method owns its candidates; budget.populationSize applies only to proposer mode',\n    )\n  }\n  if (\n    budget?.candidateConcurrency !== undefined ||\n    budget?.maxImprovementShots !== undefined ||\n    opts.analyzeGeneration !== undefined ||\n    opts.findings !== undefined ||\n    opts.policy !== undefined ||\n    opts.premeasuredBaseline !== undefined ||\n    opts.searchLedger !== undefined\n  ) {\n    throw new Error(\n      'selfImprove: candidateConcurrency, maxImprovementShots, analyzeGeneration, findings, policy, premeasuredBaseline, and searchLedger apply only to proposer mode',\n    )\n  }\n}\n\nfunction splitMethodPartitions<TScenario extends Scenario>(\n  searchScenarios: TScenario[],\n  explicitSelection: TScenario[] | undefined,\n  fraction: number,\n  unitByScenario?: ReadonlyMap<string, string>,\n): { train: TScenario[]; selection: TScenario[] } {\n  if (!Number.isFinite(fraction) || fraction <= 0 || fraction >= 1) {\n    throw new Error('selfImprove: budget.selectionFraction must be in (0, 1)')\n  }\n  const byId = new Map<string, TScenario>()\n  for (const scenario of searchScenarios) {\n    if (byId.has(scenario.id)) {\n      throw new Error(`selfImprove: duplicate scenario id '${scenario.id}'`)\n    }\n    byId.set(scenario.id, scenario)\n  }\n  if (explicitSelection) {\n    if (explicitSelection.length === 0) {\n      throw new Error('selfImprove: selectionScenarios must not be empty')\n    }\n    const selectionIds = new Set<string>()\n    for (const scenario of explicitSelection) {\n      if (!byId.has(scenario.id)) {\n        throw new Error(\n          `selfImprove: selection scenario '${scenario.id}' is absent from the non-final cases`,\n        )\n      }\n      if (selectionIds.has(scenario.id)) {\n        throw new Error(`selfImprove: duplicate selection scenario id '${scenario.id}'`)\n      }\n      selectionIds.add(scenario.id)\n    }\n    const train = searchScenarios.filter((scenario) => !selectionIds.has(scenario.id))\n    if (train.length === 0) {\n      throw new Error('selfImprove: method train split is empty')\n    }\n    if (unitByScenario) {\n      const selectedUnits = new Set([...selectionIds].map((id) => unitByScenario.get(id)))\n      if (train.some((scenario) => selectedUnits.has(unitByScenario.get(scenario.id)))) {\n        throw new Error('selfImprove: training and selection share independent units')\n      }\n    }\n    return {\n      train,\n      selection: explicitSelection.map((scenario) => byId.get(scenario.id)!),\n    }\n  }\n  if (searchScenarios.length < 2) {\n    throw new Error('selfImprove: method requires at least two non-final scenarios')\n  }\n  if (unitByScenario) {\n    const split = splitTrainHoldout(searchScenarios, fraction, unitByScenario)\n    return { train: split.train, selection: split.holdout }\n  }\n  const sorted = [...searchScenarios].sort(\n    (a, b) => stableScenarioHash(a.id) - stableScenarioHash(b.id),\n  )\n  const count = Math.max(1, Math.min(sorted.length - 1, Math.round(sorted.length * fraction)))\n  return {\n    selection: sorted.slice(0, count),\n    train: sorted.slice(count),\n  }\n}\n\n/** 32-bit FNV-1a over raw UTF-16 code units, read as an unsigned int and used\n *  only to order scenarios deterministically.\n *\n *  Frozen: the order it produces assigns scenarios to partitions, so a change\n *  re-partitions decisions already recorded. It masks no byte, unlike\n *  `fnv1a32` in `src/partition-held-out.ts`, so the two disagree on any\n *  non-ASCII id and cannot be shared. */\nfunction stableScenarioHash(value: string): number {\n  let hash = 2166136261 >>> 0\n  for (let index = 0; index < value.length; index++) {\n    hash ^= value.charCodeAt(index)\n    hash = Math.imul(hash, 16777619) >>> 0\n  }\n  return hash\n}\n\n/** Deterministic split by scenario identity, or by source identity for new-unit claims. */\nfunction splitTrainHoldout<TScenario extends Scenario>(\n  scenarios: TScenario[],\n  fraction: number,\n  unitByScenario?: ReadonlyMap<string, string>,\n): { train: TScenario[]; holdout: TScenario[] } {\n  if (unitByScenario) {\n    const units = [...new Set(scenarios.map((scenario) => unitByScenario.get(scenario.id)!))].sort(\n      (a, b) => stableScenarioHash(a) - stableScenarioHash(b) || (a < b ? -1 : a > b ? 1 : 0),\n    )\n    if (units.length < 2)\n      throw new Error('selfImprove: splitting needs at least two independent units')\n    const count = Math.max(1, Math.min(units.length - 1, Math.round(units.length * fraction)))\n    const finalUnits = new Set(units.slice(0, count))\n    return {\n      holdout: scenarios.filter((scenario) => finalUnits.has(unitByScenario.get(scenario.id)!)),\n      train: scenarios.filter((scenario) => !finalUnits.has(unitByScenario.get(scenario.id)!)),\n    }\n  }\n  const sorted = [...scenarios].sort((a, b) => stableScenarioHash(a.id) - stableScenarioHash(b.id))\n  const nHoldout = Math.max(1, Math.min(sorted.length - 1, Math.round(sorted.length * fraction)))\n  return {\n    holdout: sorted.slice(0, nHoldout),\n    train: sorted.slice(nHoldout),\n  }\n}\n\n/**\n * Latest search campaign measured for the winner surface; the baseline search\n * campaign when the winner IS the baseline. Used by the deferred-holdout\n * summary, where no holdout campaign exists to summarize.\n */\nfunction winnerSearchCampaign<TScenario extends Scenario, TArtifact>(\n  result: RunImprovementLoopResult<TArtifact, TScenario>,\n): RunImprovementLoopResult<TArtifact, TScenario>['baselineCampaign'] {\n  for (let i = result.generations.length - 1; i >= 0; i--) {\n    const measured = result.generations[i]?.surfaces.find(\n      (s) => s.surfaceHash === result.winnerSurfaceHash,\n    )\n    if (measured) return measured.campaign\n  }\n  return result.baselineCampaign\n}\n\n/**\n * One-shot self-improvement loop. See module docstring for defaults +\n * extension points.\n *\n * @example Minimum:\n *\n *   const result = await selfImprove({\n *     agent: (surface, scenario, ctx) => myAgent(surface, scenario, ctx.signal),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *     proposer,\n *   })\n *   console.log(`lift: ${result.lift.toFixed(3)} (${result.gateDecision})`)\n *\n * @example Distributed (workers in three regions):\n *\n *   await selfImprove({\n *     agent: httpDispatch({ resolveUrl: ({ placement }) => REGION_URLS[placement!] }),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *     cellPlacement: ({ scenario }) => scenario.region,\n *     budget: { maxConcurrency: 12 },\n *   })\n */\nexport function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveMethodOptions<TScenario, TArtifact>,\n): Promise<SelfImproveMethodResult<TScenario, TArtifact>>\nexport function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveProposerOptions<TScenario, TArtifact>,\n): Promise<SelfImproveProposerResult<TScenario, TArtifact>>\nexport function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n): Promise<SelfImproveResult<TScenario, TArtifact>>\nexport async function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts:\n    | SelfImproveOptions<TScenario, TArtifact>\n    | SelfImproveMethodOptions<TScenario, TArtifact>\n    | SelfImproveProposerOptions<TScenario, TArtifact>,\n): Promise<SelfImproveResult<TScenario, TArtifact>> {\n  opts = {\n    ...opts,\n    judge: captureJudge(opts.judge),\n    ...(opts.claim || opts.finalEvidence\n      ? {\n          claim: opts.claim && defineEvaluationClaim(opts.claim),\n          finalEvidence: opts.finalEvidence && captureFinalEvidencePolicy(opts.finalEvidence),\n          scenarios: structuredClone(opts.scenarios),\n          selectionScenarios: opts.selectionScenarios && structuredClone(opts.selectionScenarios),\n          budget: opts.budget && structuredClone(opts.budget),\n        }\n      : {}),\n  }\n  const startedAt = Date.now()\n  const requestedRunDir =\n    opts.runDir ??\n    (opts.method ? `.agent-eval/runs/self-improve-${startedAt}` : `mem://selfImprove-${startedAt}`)\n  const runDir = resolveRunDir(requestedRunDir)\n  const storage =\n    opts.storage ?? (runDir.startsWith('mem://') ? inMemoryCampaignStorage() : fsCampaignStorage())\n  const costLedger = createRunCostLedger({\n    storage,\n    runDir,\n    costCeilingUsd: opts.budget?.dollars,\n  })\n  try {\n    return await runSelfImprove(\n      opts as SelfImproveOptions<TScenario, TArtifact>,\n      costLedger,\n      startedAt,\n      runDir,\n      storage,\n    )\n  } catch (error) {\n    throw new SelfImproveRunError(error, costLedger)\n  }\n}\n\nasync function runSelfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n  costLedger: CostLedgerHandle,\n  startedAt: number,\n  runDir: string,\n  storage: CampaignStorage,\n): Promise<SelfImproveResult<TScenario, TArtifact>> {\n  const budget = opts.budget ?? {}\n  assertSelfImproveSearchMode(opts)\n  const maxConcurrency = budget.maxConcurrency ?? 2\n  const holdoutFraction = budget.holdoutFraction ?? 0.25\n  const holdoutMode = budget.holdout ?? 'measured'\n  const holdoutDeferred = holdoutMode === 'deferred'\n  const expectUsage = opts.expectUsage ?? 'assert'\n  const unitByScenario = opts.claim ? evaluationUnitMap(opts.claim, opts.scenarios) : undefined\n  const splitUnitByScenario =\n    opts.claim?.generalization === 'new-units' ? unitByScenario : undefined\n\n  // Deferred holdout without an explicitly reserved set trains on EVERYTHING:\n  // there is no held-out measurement in this run, so carving out a fraction\n  // would waste scenarios. An explicit `holdoutScenarios` set stays reserved\n  // (excluded from training) even when deferred, for the later measured run.\n  const explicitHoldout = budget.holdoutScenarios\n  const { train, holdout } = explicitHoldout\n    ? {\n        train: opts.scenarios.filter((s) => !explicitHoldout.some((h) => h.id === s.id)),\n        holdout: explicitHoldout as TScenario[],\n      }\n    : holdoutDeferred\n      ? { train: opts.scenarios, holdout: [] as TScenario[] }\n      : splitTrainHoldout(opts.scenarios, holdoutFraction, splitUnitByScenario)\n\n  if (train.length === 0) {\n    throw new Error(\n      'selfImprove: train split is empty. Reduce holdoutFraction or pass more scenarios.',\n    )\n  }\n  if (holdout.length === 0 && !holdoutDeferred) {\n    throw new Error('selfImprove: holdout split is empty. Pass more scenarios.')\n  }\n\n  if (opts.claim?.generalization === 'new-units') {\n    assertIndependentEvaluationSplit(opts.claim, holdout, train)\n  }\n  if (opts.finalEvidence) {\n    if (holdoutDeferred) throw new Error('final evidence requires measured holdout')\n    if (opts.method) await reserveFinalEvidence(opts.finalEvidence, opts.claim, holdout, train)\n  }\n\n  if (opts.method) {\n    const partitions = splitMethodPartitions(\n      train,\n      opts.selectionScenarios,\n      budget.selectionFraction ?? 0.25,\n      splitUnitByScenario,\n    )\n    return runSelfImproveMethod({\n      opts: { ...opts, method: opts.method },\n      train: partitions.train,\n      selection: partitions.selection,\n      holdout,\n      costLedger,\n      storage,\n      runDir,\n      startedAt,\n    })\n  }\n  const generations = budget.generations ?? 3\n  const populationSize = budget.populationSize ?? 2\n  if (generations > 0 && !opts.proposer) {\n    throw new Error(\n      'selfImprove: method or proposer is required when budget.generations is greater than zero',\n    )\n  }\n  const proposer: SurfaceProposer<ProposalFinding> = opts.proposer ?? {\n    kind: 'baseline-only',\n    propose: async () => [],\n  }\n\n  const gate: Gate<TArtifact, TScenario> =\n    opts.gate ??\n    defaultProductionGate<TArtifact, TScenario>({\n      holdoutScenarios: holdout,\n      deltaThreshold: opts.claim?.minimumEffect ?? 0.05,\n      independentUnitByScenarioId: opts.claim ? evaluationUnitMap(opts.claim, holdout) : undefined,\n    })\n\n  if (opts.onProgress) {\n    opts.onProgress({ kind: 'baseline.started', scenarios: opts.scenarios.length })\n  }\n\n  const shipper =\n    opts.hostedTenant && opts.searchLedger\n      ? startSearchShipper({\n          tenant: opts.hostedTenant,\n          ledger: opts.searchLedger.ledger,\n          runKind: 'optimization',\n        })\n      : undefined\n  const result = await runImprovementLoop<TScenario, TArtifact>({\n    scenarios: train,\n    baselineSurface: opts.baselineSurface,\n    premeasuredBaseline: opts.premeasuredBaseline,\n    dispatchWithSurface: opts.agent,\n    dispatchRef: opts.dispatchRef,\n    proposer,\n    judges: [opts.judge],\n    populationSize,\n    maxGenerations: generations,\n    candidateConcurrency: budget.candidateConcurrency,\n    reps: budget.reps,\n    maxImprovementShots: budget.maxImprovementShots,\n    holdoutScenarios: holdout,\n    claim: opts.claim,\n    finalEvidence: opts.finalEvidence,\n    holdout: holdoutMode,\n    gate,\n    neutralize: opts.neutralize,\n    autoOnPromote: opts.autoOnPromote ?? 'none',\n    ghOwner: opts.ghOwner,\n    ghRepo: opts.ghRepo,\n    storage,\n    runDir,\n    maxConcurrency,\n    cellPlacement: opts.cellPlacement,\n    dispatchTimeoutMs: opts.dispatchTimeoutMs,\n    cellRetry: opts.cellRetry,\n    costLedger,\n    expectUsage,\n    labeledStore: opts.labeledStore,\n    captureSource: opts.captureSource,\n    analyzeGeneration: opts.analyzeGeneration,\n    findings: opts.findings,\n    policy: opts.policy,\n    searchLedger: opts.searchLedger,\n  }).finally(() => shipper && shipBestEffort(() => shipper.stop(), opts.searchLedger!.ledger.path))\n\n  // Deferred holdout ran zero holdout cells, so the summary stats come from\n  // the improvement-set (search) campaigns — labeled as such on the result\n  // type — and `lift` is omitted rather than fabricated from empty campaigns.\n  const reportSplit: RunSplitTag = holdoutDeferred ? 'search' : 'holdout'\n  const reportBaselineCampaign = holdoutDeferred\n    ? result.baselineCampaign\n    : result.baselineOnHoldout\n  const reportWinnerCampaign = holdoutDeferred\n    ? winnerSearchCampaign(result)\n    : result.winnerOnHoldout\n  const reportUnitMap = opts.claim\n    ? evaluationUnitMap(opts.claim, holdoutDeferred ? train : holdout)\n    : undefined\n  const report = pairedCompositeSummary(reportBaselineCampaign, reportWinnerCampaign, reportUnitMap)\n  const baseline = report.baseline\n  const winnerStats = report.winner\n\n  // Repetitions refine a declared unit's mean without increasing the sample size.\n  let power: PowerPreflight | undefined\n  const baselineHoldoutComposites = holdoutDeferred ? [] : report.baselineComposites\n  if (baselineHoldoutComposites.length >= 3) {\n    // The shared judge's systematic bias remains outside this variance estimate.\n    power = powerPreflight({\n      baselineComposites: baselineHoldoutComposites,\n      deltaThreshold: opts.claim?.minimumEffect ?? 0.05,\n      sharedScorerChannel: true,\n    })\n    if (opts.onProgress) {\n      opts.onProgress({\n        kind: 'power.estimated',\n        n: power.n,\n        sd: power.sd,\n        mde: power.mde,\n        underpowered: power.underpowered,\n      })\n    }\n    if (power.underpowered && generations > 0) {\n      console.warn(`[selfImprove] ${power.recommendation}`)\n    }\n  }\n\n  if (opts.onProgress) {\n    opts.onProgress({\n      kind: 'baseline.completed',\n      compositeMean: baseline.compositeMean,\n      durationMs: Date.now() - startedAt,\n    })\n    opts.onProgress({\n      kind: 'gate.decided',\n      decision: result.gateResult.decision,\n      // Deferred holdout has no held-out measurement: in that mode the summary\n      // stats are search-split numbers, and emitting their delta as `lift`\n      // would misreport a train-split delta as a held-out one. Omit instead.\n      ...(holdoutDeferred ? {} : { lift: winnerStats.compositeMean - baseline.compositeMean }),\n    })\n  }\n\n  const cost = result.cost\n  const totalCost = cost.totalCostUsd\n\n  // Rigor packet: feed baseline + winner cells through analyzeRuns().\n  // The two candidates (`baseline` / `winner`) give the lift section a\n  // clean paired comparison; per-judge / per-dimension / cost-quality\n  // sections populate from the cells' judgeScores.\n  const insight = await analyzeRuns({\n    runs: [\n      ...cellsToRunRecords(\n        reportBaselineCampaign.cells,\n        'baseline',\n        runDir,\n        opts.baselineSurface,\n        reportSplit,\n        opts.model,\n      ),\n      ...(reportWinnerCampaign === reportBaselineCampaign\n        ? []\n        : cellsToRunRecords(\n            reportWinnerCampaign.cells,\n            'winner',\n            runDir,\n            result.winnerSurface,\n            reportSplit,\n            opts.model,\n          )),\n    ],\n    baselineCandidateId: 'baseline',\n    independentUnitByScenarioId: reportUnitMap,\n    decisionThreshold: opts.claim?.minimumEffect ?? 0.05,\n    ...(reportWinnerCampaign === reportBaselineCampaign ? {} : { candidateCandidateId: 'winner' }),\n  })\n\n  // ── Durable provenance: candidate→cell→gate→promote chain + rationale +\n  // diff + backend provenance. Always emitted; the +lift recomputes from it.\n  const durationMs = Date.now() - startedAt\n  const { record: provenance } = await emitLoopProvenance<TArtifact, TScenario>({\n    ...loopProvenanceArgsFromResult({\n      runId: `${runDir}#${startedAt}`,\n      runDir,\n      timestamp: new Date(startedAt).toISOString(),\n      baselineSurface: opts.baselineSurface,\n      result,\n      costReceipts: costLedger.list(),\n      totalCostUsd: totalCost,\n      totalDurationMs: durationMs,\n      independentUnitByScenarioId: holdoutDeferred ? undefined : reportUnitMap,\n    }),\n    storage,\n    hostedClient: opts.hostedTenant ? createHostedClient(opts.hostedTenant) : undefined,\n  })\n  if (opts.onProvenance) opts.onProvenance(provenance)\n\n  const summary: SelfImproveProposerResult<TScenario, TArtifact> = {\n    ...(opts.claim ? { claim: opts.claim } : {}),\n    ...(result.finalEvidence ? { finalEvidence: result.finalEvidence } : {}),\n    mode: 'proposer',\n    baseline,\n    winner: {\n      ...winnerStats,\n      surface: result.winnerSurface,\n      ...(result.winnerLabel ? { label: result.winnerLabel } : {}),\n      ...(result.winnerRationale ? { rationale: result.winnerRationale } : {}),\n    },\n    ...(holdoutDeferred ? {} : { lift: winnerStats.compositeMean - baseline.compositeMean }),\n    diff: result.promotedDiff,\n    provenance,\n    gateDecision: result.gateResult.decision,\n    generationsExplored: result.generations.length,\n    durationMs,\n    totalCostUsd: totalCost,\n    cost,\n    receipts: costLedger.list(),\n    searchHistory: result.searchHistory,\n    insight,\n    ...(power ? { power } : {}),\n    raw: result,\n  }\n\n  return summary\n}\n","import type { RunEvalOptions } from '../campaign/presets/run-eval'\nimport { runEval } from '../campaign/presets/run-eval'\nimport { inMemoryCampaignStorage } from '../campaign/storage'\nimport { surfaceDispatchRef } from '../campaign/surface-identity'\nimport type {\n  CampaignResult,\n  DispatchContext,\n  JudgeConfig,\n  MutableSurface,\n  Scenario,\n} from '../campaign/types'\nimport type { HostedTenant } from '../hosted/client'\nimport {\n  type SelfImproveBudget,\n  type SelfImproveOptions,\n  type SelfImproveResult,\n  selfImprove,\n} from './self-improve'\n\nexport type AgentEvalAgent<TScenario extends Scenario, TArtifact> = (\n  surface: MutableSurface,\n  scenario: TScenario,\n  ctx: DispatchContext,\n) => Promise<TArtifact>\n\nexport type DefineAgentEvalOptions<TScenario extends Scenario, TArtifact> = SelfImproveOptions<\n  TScenario,\n  TArtifact\n>\n\nexport interface AgentEvalEvaluateOptions<TScenario extends Scenario, TArtifact>\n  extends Omit<\n    RunEvalOptions<TScenario, TArtifact>,\n    'dispatch' | 'judges' | 'runDir' | 'scenarios'\n  > {\n  /** Scenario set to evaluate. Defaults to the scenarios passed to `defineAgentEval`. */\n  scenarios?: TScenario[]\n  /** Surface to evaluate. Defaults to the baseline surface passed to `defineAgentEval`. */\n  surface?: MutableSurface\n  /** Agent to evaluate. Defaults to the agent passed to `defineAgentEval`. */\n  agent?: AgentEvalAgent<TScenario, TArtifact>\n  /** Single judge override. Ignored when `judges` is set. */\n  judge?: JudgeConfig<TArtifact, TScenario>\n  /** Full judge list override. Defaults to the single judge passed to `defineAgentEval`. */\n  judges?: JudgeConfig<TArtifact, TScenario>[]\n  /** Logical or filesystem run directory. Defaults to an in-memory run. */\n  runDir?: string\n}\n\nexport type AgentEvalImproveOptions<TScenario extends Scenario, TArtifact> = Omit<\n  Partial<SelfImproveOptions<TScenario, TArtifact>>,\n  'budget' | 'hostedTenant'\n> & {\n  budget?: Partial<SelfImproveBudget>\n  hostedTenant?: Partial<HostedTenant>\n}\n\nexport interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {\n  /** The default scenarios used by `evaluate()` and `improve()`. */\n  readonly scenarios: readonly TScenario[]\n  /** The default baseline surface used by `evaluate()` and `improve()`. */\n  readonly baselineSurface: MutableSurface\n  /**\n   * Run one scored evaluation. Use this for a baseline score or to score one\n   * candidate surface without running an improvement loop.\n   */\n  evaluate(\n    opts?: AgentEvalEvaluateOptions<TScenario, TArtifact>,\n  ): Promise<CampaignResult<TArtifact, TScenario>>\n  /**\n   * Run the closed improvement loop. Per-call overrides replace the definition\n   * except for nested config objects (`budget`, `hostedTenant`), which\n   * are merged field-by-field so callers can override one knob without\n   * repeating secrets or budget defaults.\n   */\n  improve(\n    opts?: AgentEvalImproveOptions<TScenario, TArtifact>,\n  ): Promise<SelfImproveResult<TScenario, TArtifact>>\n}\n\n/**\n * Define an agent eval once, then either score a surface with `evaluate()` or\n * run the closed loop with `improve()`.\n *\n * This is a DX wrapper only: it delegates to `runEval()` and `selfImprove()` and\n * returns their native result shapes.\n */\nexport function defineAgentEval<TScenario extends Scenario, TArtifact>(\n  defaults: DefineAgentEvalOptions<TScenario, TArtifact>,\n): DefinedAgentEval<TScenario, TArtifact> {\n  const defaultEvaluateOptions = evaluateDefaults(defaults)\n\n  return {\n    scenarios: defaults.scenarios,\n    baselineSurface: defaults.baselineSurface,\n\n    async evaluate(opts = {}) {\n      const { agent, judge, judges, runDir, scenarios, surface, ...campaignOpts } = opts\n      const selectedAgent = agent ?? defaults.agent\n      const selectedSurface = surface ?? defaults.baselineSurface\n      const selectedRunDir = runDir ?? defaults.runDir ?? `mem://defineAgentEval-${Date.now()}`\n      const selectedStorage =\n        campaignOpts.storage ??\n        defaultEvaluateOptions.storage ??\n        (selectedRunDir.startsWith('mem://') ? inMemoryCampaignStorage() : undefined)\n      const evalOptions: RunEvalOptions<TScenario, TArtifact> = {\n        ...defaultEvaluateOptions,\n        ...campaignOpts,\n        ...(selectedStorage ? { storage: selectedStorage } : {}),\n        runDir: selectedRunDir,\n        scenarios: scenarios ?? defaults.scenarios,\n        dispatch: (scenario, ctx) => selectedAgent(selectedSurface, scenario, ctx),\n        dispatchRef: surfaceDispatchRef(\n          selectedSurface,\n          campaignOpts.dispatchRef ?? defaults.dispatchRef,\n        ),\n        judges: evaluateJudges(judges, judge ?? defaults.judge),\n      }\n      if (evalOptions.reps !== undefined)\n        evalOptions.reps = requirePositiveInteger(evalOptions.reps, 'reps')\n      return runEval<TScenario, TArtifact>(evalOptions)\n    },\n\n    async improve(opts = {}) {\n      const {\n        budget: budgetOverride,\n        hostedTenant: hostedTenantOverride,\n        ...topLevelOverrides\n      } = opts\n      const merged = mergeDefined(defaults, topLevelOverrides)\n      const budget = mergeBudget(defaults.budget, budgetOverride)\n      const hostedTenant = mergeHostedTenant(defaults.hostedTenant, hostedTenantOverride)\n      return selfImprove<TScenario, TArtifact>({\n        ...merged,\n        ...(budget ? { budget } : {}),\n        ...(hostedTenant ? { hostedTenant } : {}),\n      })\n    },\n  }\n}\n\ntype SharedEvaluateDefaults<TScenario extends Scenario, TArtifact> = Omit<\n  RunEvalOptions<TScenario, TArtifact>,\n  'dispatch' | 'judges' | 'runDir' | 'scenarios'\n>\n\nfunction evaluateDefaults<TScenario extends Scenario, TArtifact>(\n  defaults: DefineAgentEvalOptions<TScenario, TArtifact>,\n): SharedEvaluateDefaults<TScenario, TArtifact> {\n  const out: SharedEvaluateDefaults<TScenario, TArtifact> = {}\n  if (defaults.storage) out.storage = defaults.storage\n  if (defaults.labeledStore) out.labeledStore = defaults.labeledStore\n  if (defaults.captureSource) out.captureSource = defaults.captureSource\n  if (defaults.cellPlacement) out.cellPlacement = defaults.cellPlacement\n  if (defaults.expectUsage) out.expectUsage = defaults.expectUsage\n  if (defaults.budget?.dollars !== undefined) out.costCeiling = defaults.budget.dollars\n  if (defaults.budget?.maxConcurrency !== undefined)\n    out.maxConcurrency = defaults.budget.maxConcurrency\n  if (defaults.budget?.reps !== undefined)\n    out.reps = requirePositiveInteger(defaults.budget.reps, 'budget.reps')\n  return out\n}\n\nfunction mergeBudget(\n  defaults: SelfImproveBudget | undefined,\n  overrides: Partial<SelfImproveBudget> | undefined,\n): SelfImproveBudget | undefined {\n  const merged = mergeOptionalObject(defaults, overrides)\n  if (merged?.reps !== undefined) merged.reps = requirePositiveInteger(merged.reps, 'budget.reps')\n  return merged\n}\n\nfunction mergeHostedTenant(\n  defaults: HostedTenant | undefined,\n  overrides: Partial<HostedTenant> | undefined,\n): HostedTenant | undefined {\n  const merged = mergeOptionalObject(defaults, overrides)\n  if (!merged) return undefined\n  if (!merged.endpoint?.trim() || !merged.apiKey?.trim() || !merged.tenantId?.trim()) {\n    throw new Error(\n      'defineAgentEval.improve: hostedTenant requires endpoint, apiKey, and tenantId after merging defaults and overrides',\n    )\n  }\n  return merged\n}\n\nfunction mergeDefined<T extends object>(defaults: T, overrides: Partial<T> | undefined): T {\n  if (!overrides) return defaults\n  const merged = { ...defaults } as Record<string, unknown>\n  for (const [key, value] of Object.entries(overrides)) {\n    if (value !== undefined) merged[key] = value\n  }\n  return merged as T\n}\n\nfunction mergeOptionalObject<T extends object>(\n  defaults: T | undefined,\n  overrides: Partial<T> | undefined,\n): T | undefined {\n  if (!defaults && !overrides) return undefined\n  return mergeDefined(defaults ?? ({} as T), overrides)\n}\n\nfunction evaluateJudges<TArtifact, TScenario extends Scenario>(\n  judges: JudgeConfig<TArtifact, TScenario>[] | undefined,\n  defaultJudge: JudgeConfig<TArtifact, TScenario>,\n): JudgeConfig<TArtifact, TScenario>[] {\n  if (judges !== undefined) {\n    if (judges.length === 0) {\n      throw new Error('defineAgentEval.evaluate: judges must not be empty')\n    }\n    return judges\n  }\n  return [defaultJudge]\n}\n\nfunction requirePositiveInteger(value: number, field: string): number {\n  if (!Number.isInteger(value) || value < 1) {\n    throw new Error(`defineAgentEval: ${field} must be a positive integer`)\n  }\n  return value\n}\n"],"mappings":";;;;;;;;;;;AAYA,SAAgB,uBACd,UACA,QACA,6BACA;CACA,MAAM,SAAS,yBAAyB,UAAU,QAAQ,2BAA2B;CACrF,MAAM,eAAe,aACnB,OAAO,YACL,OAAO,QAAQ,SAAS,WAAW,UAAU,CAAC,CAAC,KAAK,CAAC,IAAI,eAAe,CACtE,IACA,UAAU,aACZ,CAAC,CACH;CACF,OAAO;EACL,UAAU;GAAE,eAAe,OAAO;GAAY,aAAa,YAAY,QAAQ;EAAE;EACjF,QAAQ;GAAE,eAAe,OAAO;GAAW,aAAa,YAAY,MAAM;EAAE;EAC5E,oBAAoB,OAAO;EAC3B,cAAc,OAAO;CACvB;AACF;;;;;AAMA,SAAS,WAAW,GAAmB;CACrC,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;EACjC,KAAK,EAAE,WAAW,CAAC;EACnB,IAAI,KAAK,KAAK,GAAG,QAAQ,MAAM;CACjC;CACA,OAAO,EAAE,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG;AACvC;;;;;;;;AASA,SAAgB,kBACd,OACA,aACA,OACA,SACA,UACA,eACa;CACb,MAAM,aAAa,mBAAmB,OAAO;CAC7C,MAAM,aAAa,mBAAmB,WAAW;CACjD,OAAO,MAAM,KAAK,SAAS;EACzB,MAAM,gBAAgB,KAAK,mBAAmB,KAAK,gBAAgB,CAAC,KAAK,aAAa,IAAI,CAAC;EAC3F,IAAI,cAAc,SAAS,GACzB,MAAM,IAAI,gBACR,oBAAoB,KAAK,OAAO,+BAA+B,cAAc,KAAK,IAAI,GACxF;EAEF,MAAM,QAAQ,cAAc,MAAM;EAClC,IAAI,CAAC,OACH,MAAM,IAAI,gBACR,2CAA2C,KAAK,OAAO,gCACzD;EAEF,IAAI,CAAC,iBAAiB,KAAK,GACzB,MAAM,IAAI,gBACR,sBAAsB,MAAM,sCAAsC,KAAK,QACzE;EAEF,OAAO,wBAAwB,MAAM;GACnC,OAAO,GAAG,MAAM,IAAI,YAAY,IAAI,KAAK;GACzC,cAAc;GACd;GAEA,MACE,KAAK,MAAM,MACX,WAAW,KAAK,UAAU,CAAC,CACxB,MAAM,GAAG,CAAC,CAAC,CACX,MAAM,EAAE,CAAC,CACT,QAAQ,GAAG,MAAO,IAAI,KAAK,EAAE,WAAW,CAAC,MAAO,GAAG,CAAC;GACzD;GACA;GACA;GACA,WAAW;GACX;EACF,CAAC;CACH,CAAC;AACH;;;;ACIA,eAAsB,qBAA4D,MAUvB;CACzD,MAAM,EAAE,MAAM,OAAO,WAAW,SAAS,YAAY,SAAS,QAAQ,cAAc;CACpF,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,IAAI,KAAK,kBAAkB,SAAS,CAAC,KAAK,WAAW,CAAC,KAAK,SACzD,MAAM,IAAI,MAAM,2DAA2D;CAE7E,MAAM,kBAAkB,KAAK,aAAa,KAAA,IAAY,KAAA,IAAY,gBAAgB,KAAK,QAAQ;CAC/F,IAAI,mBAAmB,OAAO,YAAY,YACxC,MAAM,IAAI,MAAM,yDAAyD;CAC3E,MAAM,kBAAkB,gBAAgB,KAAK,eAAe;CAC5D,MAAM,QAAQ,KAAK;CACnB,MAAM,EACJ,UACA,MAAM,YACN,YACE,MAAM,0BAA0B;EAClC,QAAQ,KAAK;EACb;EACA,qBAAqB,KAAK;EAC1B,2BAA2B,KAAK;EAChC,OAAO;GACL;GACA,gBAAgB;GAChB,oBAAoB;GACpB,qBAAqB,KAAK;GAC1B,QAAQ,CAAC,KAAK;GACd,QAAQ,GAAG,OAAO,gBAAgB,KAAK,OAAO,KAAK,QAAQ,oBAAoB,GAAG;GAClF,MAAM;GACN,YAAY;IACV;IACA,gBAAgB,OAAO,kBAAkB;IACzC,MAAM,OAAO;IACb,aAAa,KAAK;IAClB,mBAAmB,KAAK;IACxB,WAAW,KAAK;IAChB,eAAe,KAAK;IACpB,cAAc,KAAK;IACnB,eAAe,KAAK;IACpB,aAAa,KAAK,eAAe;GACnC;GACA;EACF;CACF,CAAC;CACD,MAAM,aAAa;CACnB,MAAM,kBACJ,uBACE;EAAC;EAAoB;EAAkB;EAAuB;CAAgB,CAAC,CAAC,KAC7E,WAAW;EACV,OAAO;EACP,MAAM,sBAAsB,WAAW,QAAQ,EAAE,OAAO,GAAG,WAAW,GAAG,QAAQ,CAAC,CAAC;CACrF,EACF,CACF;CACF,MAAM,kBACJ,uBAAuB,CACrB;EAAE,OAAO,KAAK,OAAO;EAAM,MAAM;CAAW,GAC5C;EAAE,OAAO;EAAoB,MAAM,UAAU;CAAE,CACjD,CAAC;CACH,MAAM,OACJ,KAAK,QACL,sBAA4C;EAC1C,kBAAkB;EAClB,gBAAgB,KAAK,OAAO,iBAAiB;EAC7C,6BAA6B,KAAK,QAAQ,kBAAkB,KAAK,OAAO,OAAO,IAAI,KAAA;CACrF,CAAC;CACH,MAAM,aAAa,MAAM,mBAAmB;EAC1C,OAAO,KAAK;EACZ,eAAe,KAAK;EACpB;EACA,eAAe,SAAS;EACxB,WAAW;EACX,qBAAqB,KAAK;EAC1B,aAAa,KAAK;EAClB,QAAQ,CAAC,KAAK;EACd,SAAS,OAAO;EAChB,YAAY,KAAK;EACjB,MAAM;GACJ,GAAG;GACH,MAAM,OAAO,SAAS;IACpB,MAAM,OAAO,UAAU;IACvB,IACE,OAAO,YAAY,KAAA,MAClB,CAAC,KAAK,sBAAsB,KAAK,eAAe,OAAO,UAExD,OAAO;KACL,UAAU;KACV,SAAS,CAAC,8DAA8D;KACxE,mBAAmB,CACjB;MACE,MAAM;MACN,QAAQ;MACR,QAAQ;OAAE;OAAM,WAAW,OAAO;MAAQ;KAC5C,CACF;IACF;IAEF,OAAO,KAAK,OAAO,OAAO;GAC5B;EACF;EACA;EACA;EACA;EACA,WAAW;EACX,MAAM,OAAO;EACb,gBAAgB,OAAO,kBAAkB;EACzC,mBAAmB,KAAK;EACxB,WAAW,KAAK;EAChB,eAAe,KAAK;EACpB,aAAa,KAAK,eAAe;EACjC,OAAO;CACT,CAAC;CACD,MAAM,WAAW,kBACb;EACE,UAAU,8BAA8B;GACtC,UAAU,WAAW;GACrB,SAAS;GACT,SAAS;EACX,CAAC;EACD,QAAQ,8BAA8B;GACpC,UAAU,WAAW;GACrB,SAAS,SAAS;GAClB,SAAS;EACX,CAAC;CACH,IACA,KAAA;CACJ,MAAM,WAAW,WAAW,YAAY;CACxC,MAAM,SAAS,WACX,OACA,uBACE,WAAW,mBACX,WAAW,iBACX,KAAK,QAAQ,kBAAkB,KAAK,OAAO,OAAO,IAAI,KAAA,CACxD;CACJ,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,SAAS,QAAQ,UAAU;CACjC,MAAM,OAAO,YAAY,SAAS,OAAO,gBAAgB,SAAS,gBAAgB,KAAA;CAClF,MAAM,UAAU,WACZ,KAAA,IACA,MAAM,YAAY;EAChB,MAAM,CACJ,GAAG,kBACD,WAAW,kBAAkB,OAC7B,YACA,QACA,iBACA,WACA,KAAK,KACP,GACA,GAAI,WAAW,oBAAoB,WAAW,oBAC1C,CAAC,IACD,kBACE,WAAW,gBAAgB,OAC3B,UACA,QACA,SAAS,eACT,WACA,KAAK,KACP,CACN;EACA,qBAAqB;EACrB,6BAA6B,KAAK,QAC9B,kBAAkB,KAAK,OAAO,OAAO,IACrC,KAAA;EACJ,mBAAmB,KAAK,OAAO,iBAAiB;EAChD,GAAI,WAAW,oBAAoB,WAAW,oBAC1C,CAAC,IACD,EAAE,sBAAsB,SAAS;CACvC,CAAC;CACL,MAAM,OAAO,UAAU;CACvB,MAAM,eAAe;EACnB,MAAM,KAAK,OAAO;EAClB,MAAM;EACN,GAAI,SAAS,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,YAAY,SAAS,WAAW;EAC/E,GAAI,SAAS,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,YAAY,SAAS,WAAW;CACjF;CACA,MAAM,aAAa,KAAK,IAAI,IAAI;CAChC,MAAM,WAAW,WAAW,KAAK;CACjC,MAAM,SAAS;EACb,QAAQ;EACR,OAAO,GAAG,OAAO,GAAG;EACpB;EACA,WAAW,IAAI,KAAK,SAAS,CAAC,CAAC,YAAY;EAC3C,qBAAqB,mBAAmB,eAAe;EACvD,mBAAmB,mBAAmB,SAAS,aAAa;EAC5D,MAAM,WAAW;EACjB,oBAAoB;EACpB,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;EAC1C,UAAU;GACR,kBAAkB,oBAAoB,OAAO,OAAO,QAAQ,CAAC;GAC7D,sBAAsB,oBAAoB,WAAW,OAAO,QAAQ,CAAC;GACrE,oBAAoB,oBAAoB,SAAS,OAAO,QAAQ,CAAC;GACjE,wBAAwB,0BAA0B,WAAW,iBAAiB;GAC9E,sBAAsB,0BAA0B,WAAW,eAAe;GAC1E,oBAAoB,gBAAgB,QAAQ;GAC5C,GAAI,SAAS,EAAE,qBAAqB,OAAO,aAAa,IAAI,CAAC;GAC7D,GAAI,WAAW,uBACX,EAAE,2BAA2B,0BAA0B,WAAW,oBAAoB,EAAE,IACxF,CAAC;EACP;EACA,MAAM,WAAW;EACjB,GAAI,WAAW,EAAE,SAAS,WAAoB,IAAI,CAAC;EACnD,GAAI,YAAY,SACZ;GACE,0BAA0B,SAAS;GACnC,wBAAwB,OAAO;GAC/B,aAAa;EACf,IACA,CAAC;EACL;EACA,cAAc,KAAK;EACnB,iBAAiB;CACnB;CAEA,MAAM,aAAgE,KAAK,MACzE,KAAK,UAAU,MAAM,CACvB;CACA,MAAM,aAA0C;EAC9C,GAAG;EACH,cAAc,gBAAgB,UAAU;CAC1C;CACA,QAAQ,UAAU,MAAM;CACxB,QAAQ,MAAM,GAAG,OAAO,0BAA0B,KAAK,UAAU,YAAY,MAAM,CAAC,CAAC;CACrF,KAAK,eAAe,UAAU;CAC9B,KAAK,aAAa;EAChB,MAAM;EACN,UAAU,WAAW,WAAW;EAChC,GAAI,SAAS,KAAA,IAAY,CAAC,IAAI,EAAE,KAAK;CACvC,CAAC;CACD,MAAM,WACJ,KAAK,kBAAkB,QAAQ,WAAW,WAAW,aAAa,SAC9D,WAAW;EACT,QAAQ,WAAW;EACnB,MAAM,WAAW;EACjB,cAAc,WAAW;EACzB,SAAS,KAAK;EACd,QAAQ,KAAK;CACf,CAAC,IACD,KAAA;CACN,MAAM,SAAwD;EAC5D,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;EAC/B,MAAM;EACN,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;EAC1C,GAAI,WAAW,gBAAgB,EAAE,eAAe,WAAW,cAAc,IAAI,CAAC;EAC9E,uBAAuB;EACvB;EACA,QAAQ;GACN,SAAS,SAAS;GAClB,eAAe,QAAQ,iBAAiB;GACxC,aAAa,QAAQ,eAAe,CAAC;GACrC,OAAO,KAAK,OAAO;EACrB;EACA,GAAI,SAAS,KAAA,IAAY,CAAC,IAAI,EAAE,KAAK;EACrC,MAAM,WAAW;EACjB,cAAc,WAAW,WAAW;EACpC;EACA;EACA;EACA,YAAY,WAAW,QAAQ;EAC/B,cAAc,KAAK;EACnB;EACA;EACA,GAAI,UAAU,EAAE,QAAQ,IAAI,CAAC;EAC7B,GAAI,SAAS,gBAAgB,EAAE,eAAe,SAAS,cAAc,IAAI,CAAC;EAC1E,KAAK;GACH,GAAG;GACH,MAAM;GACN;GACA,eAAe,SAAS;GACxB,mBAAmB,YAAY,SAAS,aAAa;GACrD,QAAQ;GACR;GACA,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;EACjC;CACF;CACA,IAAI,KAAK,cAAc;EACrB,MAAM,UAAU,SAAS;EACzB,IAAI,SAAS;GACX,MAAM,OAAO,cAAc,QAAQ,OAAO,GAAG;GAC7C,MAAM,qBAEF,iBAAiB;IACf,QAAQ,KAAK;IACb,QAAQ;KAAE;KAAM,UAAU,QAAQ,QAAQ;IAAS;IACnD,SAAS;GACX,CAAC,GACH,IACF;EACF,OACE,QAAQ,KACN,qCAAqC,KAAK,OAAO,KAAK,mDACxD;CAEJ;CACA,OAAO;AACT;;;;ACoBA,IAAa,sBAAb,cAAyC,MAAM;CAC7C;CACA;CAEA,YAAY,OAAgB,QAA0B;EACpD,MAAM,WAAW,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;EACzE,MAAM,SAAS,SAAS,EAAE,OAAO,SAAS,CAAC;EAC3C,KAAK,OAAO;EACZ,KAAK,OAAO,OAAO,QAAQ;EAC3B,KAAK,WAAW,OAAO,KAAK;CAC9B;AACF;AAEA,SAAS,4BACP,MACM;CACN,oCAAoC,IAAI;CACxC,IACE,CAAC,KAAK,WACL,KAAK,wBAAwB,KAAA,KAC5B,KAAK,8BAA8B,KAAA,KACnC,KAAK,aAAa,KAAA,IAEpB,MAAM,IAAI,MAAM,uDAAuD;CACzE,IAAI,KAAK,UAAU,KAAK,UACtB,MAAM,IAAI,MAAM,yDAAyD;CAE3E,IAAI,CAAC,KAAK,QAAQ;EAChB,IAAI,KAAK,uBAAuB,KAAA,GAC9B,MAAM,IAAI,MAAM,iDAAiD;EAEnE,IAAI,KAAK,gBAAgB,CAAC,KAAK,cAC7B,MAAM,IAAI,MACR,8FACF;EAEF;CACF;CACA,IACE,OAAO,KAAK,OAAO,SAAS,YAC5B,CAAC,KAAK,OAAO,KAAK,KAAK,KACvB,KAAK,OAAO,KAAK,KAAK,MAAM,KAAK,OAAO,QACxC,OAAO,KAAK,OAAO,aAAa,YAEhC,MAAM,IAAI,MAAM,kEAAkE;CAEpF,MAAM,SAAS,KAAK;CACpB,IAAI,QAAQ,gBAAgB,KAAA,GAC1B,MAAM,IAAI,MACR,uFACF;CAEF,IAAI,QAAQ,mBAAmB,KAAA,GAC7B,MAAM,IAAI,MACR,8FACF;CAEF,IACE,QAAQ,yBAAyB,KAAA,KACjC,QAAQ,wBAAwB,KAAA,KAChC,KAAK,sBAAsB,KAAA,KAC3B,KAAK,aAAa,KAAA,KAClB,KAAK,WAAW,KAAA,KAChB,KAAK,wBAAwB,KAAA,KAC7B,KAAK,iBAAiB,KAAA,GAEtB,MAAM,IAAI,MACR,gKACF;AAEJ;AAEA,SAAS,sBACP,iBACA,mBACA,UACA,gBACgD;CAChD,IAAI,CAAC,OAAO,SAAS,QAAQ,KAAK,YAAY,KAAK,YAAY,GAC7D,MAAM,IAAI,MAAM,yDAAyD;CAE3E,MAAM,uBAAO,IAAI,IAAuB;CACxC,KAAK,MAAM,YAAY,iBAAiB;EACtC,IAAI,KAAK,IAAI,SAAS,EAAE,GACtB,MAAM,IAAI,MAAM,uCAAuC,SAAS,GAAG,EAAE;EAEvE,KAAK,IAAI,SAAS,IAAI,QAAQ;CAChC;CACA,IAAI,mBAAmB;EACrB,IAAI,kBAAkB,WAAW,GAC/B,MAAM,IAAI,MAAM,mDAAmD;EAErE,MAAM,+BAAe,IAAI,IAAY;EACrC,KAAK,MAAM,YAAY,mBAAmB;GACxC,IAAI,CAAC,KAAK,IAAI,SAAS,EAAE,GACvB,MAAM,IAAI,MACR,oCAAoC,SAAS,GAAG,qCAClD;GAEF,IAAI,aAAa,IAAI,SAAS,EAAE,GAC9B,MAAM,IAAI,MAAM,iDAAiD,SAAS,GAAG,EAAE;GAEjF,aAAa,IAAI,SAAS,EAAE;EAC9B;EACA,MAAM,QAAQ,gBAAgB,QAAQ,aAAa,CAAC,aAAa,IAAI,SAAS,EAAE,CAAC;EACjF,IAAI,MAAM,WAAW,GACnB,MAAM,IAAI,MAAM,0CAA0C;EAE5D,IAAI,gBAAgB;GAClB,MAAM,gBAAgB,IAAI,IAAI,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK,OAAO,eAAe,IAAI,EAAE,CAAC,CAAC;GACnF,IAAI,MAAM,MAAM,aAAa,cAAc,IAAI,eAAe,IAAI,SAAS,EAAE,CAAC,CAAC,GAC7E,MAAM,IAAI,MAAM,6DAA6D;EAEjF;EACA,OAAO;GACL;GACA,WAAW,kBAAkB,KAAK,aAAa,KAAK,IAAI,SAAS,EAAE,CAAE;EACvE;CACF;CACA,IAAI,gBAAgB,SAAS,GAC3B,MAAM,IAAI,MAAM,+DAA+D;CAEjF,IAAI,gBAAgB;EAClB,MAAM,QAAQ,kBAAkB,iBAAiB,UAAU,cAAc;EACzE,OAAO;GAAE,OAAO,MAAM;GAAO,WAAW,MAAM;EAAQ;CACxD;CACA,MAAM,SAAS,CAAC,GAAG,eAAe,CAAC,CAAC,MACjC,GAAG,MAAM,mBAAmB,EAAE,EAAE,IAAI,mBAAmB,EAAE,EAAE,CAC9D;CACA,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC,CAAC;CAC3F,OAAO;EACL,WAAW,OAAO,MAAM,GAAG,KAAK;EAChC,OAAO,OAAO,MAAM,KAAK;CAC3B;AACF;;;;;;;;AASA,SAAS,mBAAmB,OAAuB;CACjD,IAAI,OAAO;CACX,KAAK,IAAI,QAAQ,GAAG,QAAQ,MAAM,QAAQ,SAAS;EACjD,QAAQ,MAAM,WAAW,KAAK;EAC9B,OAAO,KAAK,KAAK,MAAM,QAAQ,MAAM;CACvC;CACA,OAAO;AACT;;AAGA,SAAS,kBACP,WACA,UACA,gBAC8C;CAC9C,IAAI,gBAAgB;EAClB,MAAM,QAAQ,CAAC,GAAG,IAAI,IAAI,UAAU,KAAK,aAAa,eAAe,IAAI,SAAS,EAAE,CAAE,CAAC,CAAC,CAAC,CAAC,MACvF,GAAG,MAAM,mBAAmB,CAAC,IAAI,mBAAmB,CAAC,MAAM,IAAI,IAAI,KAAK,IAAI,IAAI,IAAI,EACvF;EACA,IAAI,MAAM,SAAS,GACjB,MAAM,IAAI,MAAM,6DAA6D;EAC/E,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,SAAS,GAAG,KAAK,MAAM,MAAM,SAAS,QAAQ,CAAC,CAAC;EACzF,MAAM,aAAa,IAAI,IAAI,MAAM,MAAM,GAAG,KAAK,CAAC;EAChD,OAAO;GACL,SAAS,UAAU,QAAQ,aAAa,WAAW,IAAI,eAAe,IAAI,SAAS,EAAE,CAAE,CAAC;GACxF,OAAO,UAAU,QAAQ,aAAa,CAAC,WAAW,IAAI,eAAe,IAAI,SAAS,EAAE,CAAE,CAAC;EACzF;CACF;CACA,MAAM,SAAS,CAAC,GAAG,SAAS,CAAC,CAAC,MAAM,GAAG,MAAM,mBAAmB,EAAE,EAAE,IAAI,mBAAmB,EAAE,EAAE,CAAC;CAChG,MAAM,WAAW,KAAK,IAAI,GAAG,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC,CAAC;CAC9F,OAAO;EACL,SAAS,OAAO,MAAM,GAAG,QAAQ;EACjC,OAAO,OAAO,MAAM,QAAQ;CAC9B;AACF;;;;;;AAOA,SAAS,qBACP,QACoE;CACpE,KAAK,IAAI,IAAI,OAAO,YAAY,SAAS,GAAG,KAAK,GAAG,KAAK;EACvD,MAAM,WAAW,OAAO,YAAY,EAAE,EAAE,SAAS,MAC9C,MAAM,EAAE,gBAAgB,OAAO,iBAClC;EACA,IAAI,UAAU,OAAO,SAAS;CAChC;CACA,OAAO,OAAO;AAChB;AAqCA,eAAsB,YACpB,MAIkD;CAClD,OAAO;EACL,GAAG;EACH,OAAO,aAAa,KAAK,KAAK;EAC9B,GAAI,KAAK,SAAS,KAAK,gBACnB;GACE,OAAO,KAAK,SAAS,sBAAsB,KAAK,KAAK;GACrD,eAAe,KAAK,iBAAiB,2BAA2B,KAAK,aAAa;GAClF,WAAW,gBAAgB,KAAK,SAAS;GACzC,oBAAoB,KAAK,sBAAsB,gBAAgB,KAAK,kBAAkB;GACtF,QAAQ,KAAK,UAAU,gBAAgB,KAAK,MAAM;EACpD,IACA,CAAC;CACP;CACA,MAAM,YAAY,KAAK,IAAI;CAC3B,MAAM,kBACJ,KAAK,WACJ,KAAK,SAAS,iCAAiC,cAAc,qBAAqB;CACrF,MAAM,SAAS,cAAc,eAAe;CAC5C,MAAM,UACJ,KAAK,YAAY,OAAO,WAAW,QAAQ,IAAI,wBAAwB,IAAI,kBAAkB;CAC/F,MAAM,aAAa,oBAAoB;EACrC;EACA;EACA,gBAAgB,KAAK,QAAQ;CAC/B,CAAC;CACD,IAAI;EACF,OAAO,MAAM,eACX,MACA,YACA,WACA,QACA,OACF;CACF,SAAS,OAAO;EACd,MAAM,IAAI,oBAAoB,OAAO,UAAU;CACjD;AACF;AAEA,eAAe,eACb,MACA,YACA,WACA,QACA,SACkD;CAClD,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,4BAA4B,IAAI;CAChC,MAAM,iBAAiB,OAAO,kBAAkB;CAChD,MAAM,kBAAkB,OAAO,mBAAmB;CAClD,MAAM,cAAc,OAAO,WAAW;CACtC,MAAM,kBAAkB,gBAAgB;CACxC,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,iBAAiB,KAAK,QAAQ,kBAAkB,KAAK,OAAO,KAAK,SAAS,IAAI,KAAA;CACpF,MAAM,sBACJ,KAAK,OAAO,mBAAmB,cAAc,iBAAiB,KAAA;CAMhE,MAAM,kBAAkB,OAAO;CAC/B,MAAM,EAAE,OAAO,YAAY,kBACvB;EACE,OAAO,KAAK,UAAU,QAAQ,MAAM,CAAC,gBAAgB,MAAM,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC;EAC/E,SAAS;CACX,IACA,kBACE;EAAE,OAAO,KAAK;EAAW,SAAS,CAAC;CAAiB,IACpD,kBAAkB,KAAK,WAAW,iBAAiB,mBAAmB;CAE5E,IAAI,MAAM,WAAW,GACnB,MAAM,IAAI,MACR,mFACF;CAEF,IAAI,QAAQ,WAAW,KAAK,CAAC,iBAC3B,MAAM,IAAI,MAAM,2DAA2D;CAG7E,IAAI,KAAK,OAAO,mBAAmB,aACjC,iCAAiC,KAAK,OAAO,SAAS,KAAK;CAE7D,IAAI,KAAK,eAAe;EACtB,IAAI,iBAAiB,MAAM,IAAI,MAAM,0CAA0C;EAC/E,IAAI,KAAK,QAAQ,MAAM,qBAAqB,KAAK,eAAe,KAAK,OAAO,SAAS,KAAK;CAC5F;CAEA,IAAI,KAAK,QAAQ;EACf,MAAM,aAAa,sBACjB,OACA,KAAK,oBACL,OAAO,qBAAqB,KAC5B,mBACF;EACA,OAAO,qBAAqB;GAC1B,MAAM;IAAE,GAAG;IAAM,QAAQ,KAAK;GAAO;GACrC,OAAO,WAAW;GAClB,WAAW,WAAW;GACtB;GACA;GACA;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,cAAc,OAAO,eAAe;CAC1C,MAAM,iBAAiB,OAAO,kBAAkB;CAChD,IAAI,cAAc,KAAK,CAAC,KAAK,UAC3B,MAAM,IAAI,MACR,0FACF;CAEF,MAAM,WAA6C,KAAK,YAAY;EAClE,MAAM;EACN,SAAS,YAAY,CAAC;CACxB;CAEA,MAAM,OACJ,KAAK,QACL,sBAA4C;EAC1C,kBAAkB;EAClB,gBAAgB,KAAK,OAAO,iBAAiB;EAC7C,6BAA6B,KAAK,QAAQ,kBAAkB,KAAK,OAAO,OAAO,IAAI,KAAA;CACrF,CAAC;CAEH,IAAI,KAAK,YACP,KAAK,WAAW;EAAE,MAAM;EAAoB,WAAW,KAAK,UAAU;CAAO,CAAC;CAGhF,MAAM,UACJ,KAAK,gBAAgB,KAAK,eACtB,mBAAmB;EACjB,QAAQ,KAAK;EACb,QAAQ,KAAK,aAAa;EAC1B,SAAS;CACX,CAAC,IACD,KAAA;CACN,MAAM,SAAS,MAAM,mBAAyC;EAC5D,WAAW;EACX,iBAAiB,KAAK;EACtB,qBAAqB,KAAK;EAC1B,qBAAqB,KAAK;EAC1B,aAAa,KAAK;EAClB;EACA,QAAQ,CAAC,KAAK,KAAK;EACnB;EACA,gBAAgB;EAChB,sBAAsB,OAAO;EAC7B,MAAM,OAAO;EACb,qBAAqB,OAAO;EAC5B,kBAAkB;EAClB,OAAO,KAAK;EACZ,eAAe,KAAK;EACpB,SAAS;EACT;EACA,YAAY,KAAK;EACjB,eAAe,KAAK,iBAAiB;EACrC,SAAS,KAAK;EACd,QAAQ,KAAK;EACb;EACA;EACA;EACA,eAAe,KAAK;EACpB,mBAAmB,KAAK;EACxB,WAAW,KAAK;EAChB;EACA;EACA,cAAc,KAAK;EACnB,eAAe,KAAK;EACpB,mBAAmB,KAAK;EACxB,UAAU,KAAK;EACf,QAAQ,KAAK;EACb,cAAc,KAAK;CACrB,CAAC,CAAC,CAAC,cAAc,WAAW,qBAAqB,QAAQ,KAAK,GAAG,KAAK,aAAc,OAAO,IAAI,CAAC;CAKhG,MAAM,cAA2B,kBAAkB,WAAW;CAC9D,MAAM,yBAAyB,kBAC3B,OAAO,mBACP,OAAO;CACX,MAAM,uBAAuB,kBACzB,qBAAqB,MAAM,IAC3B,OAAO;CACX,MAAM,gBAAgB,KAAK,QACvB,kBAAkB,KAAK,OAAO,kBAAkB,QAAQ,OAAO,IAC/D,KAAA;CACJ,MAAM,SAAS,uBAAuB,wBAAwB,sBAAsB,aAAa;CACjG,MAAM,WAAW,OAAO;CACxB,MAAM,cAAc,OAAO;CAG3B,IAAI;CACJ,MAAM,4BAA4B,kBAAkB,CAAC,IAAI,OAAO;CAChE,IAAI,0BAA0B,UAAU,GAAG;EAEzC,QAAQ,eAAe;GACrB,oBAAoB;GACpB,gBAAgB,KAAK,OAAO,iBAAiB;GAC7C,qBAAqB;EACvB,CAAC;EACD,IAAI,KAAK,YACP,KAAK,WAAW;GACd,MAAM;GACN,GAAG,MAAM;GACT,IAAI,MAAM;GACV,KAAK,MAAM;GACX,cAAc,MAAM;EACtB,CAAC;EAEH,IAAI,MAAM,gBAAgB,cAAc,GACtC,QAAQ,KAAK,iBAAiB,MAAM,gBAAgB;CAExD;CAEA,IAAI,KAAK,YAAY;EACnB,KAAK,WAAW;GACd,MAAM;GACN,eAAe,SAAS;GACxB,YAAY,KAAK,IAAI,IAAI;EAC3B,CAAC;EACD,KAAK,WAAW;GACd,MAAM;GACN,UAAU,OAAO,WAAW;GAI5B,GAAI,kBAAkB,CAAC,IAAI,EAAE,MAAM,YAAY,gBAAgB,SAAS,cAAc;EACxF,CAAC;CACH;CAEA,MAAM,OAAO,OAAO;CACpB,MAAM,YAAY,KAAK;CAMvB,MAAM,UAAU,MAAM,YAAY;EAChC,MAAM,CACJ,GAAG,kBACD,uBAAuB,OACvB,YACA,QACA,KAAK,iBACL,aACA,KAAK,KACP,GACA,GAAI,yBAAyB,yBACzB,CAAC,IACD,kBACE,qBAAqB,OACrB,UACA,QACA,OAAO,eACP,aACA,KAAK,KACP,CACN;EACA,qBAAqB;EACrB,6BAA6B;EAC7B,mBAAmB,KAAK,OAAO,iBAAiB;EAChD,GAAI,yBAAyB,yBAAyB,CAAC,IAAI,EAAE,sBAAsB,SAAS;CAC9F,CAAC;CAID,MAAM,aAAa,KAAK,IAAI,IAAI;CAChC,MAAM,EAAE,QAAQ,eAAe,MAAM,mBAAyC;EAC5E,GAAG,6BAA6B;GAC9B,OAAO,GAAG,OAAO,GAAG;GACpB;GACA,WAAW,IAAI,KAAK,SAAS,CAAC,CAAC,YAAY;GAC3C,iBAAiB,KAAK;GACtB;GACA,cAAc,WAAW,KAAK;GAC9B,cAAc;GACd,iBAAiB;GACjB,6BAA6B,kBAAkB,KAAA,IAAY;EAC7D,CAAC;EACD;EACA,cAAc,KAAK,eAAe,mBAAmB,KAAK,YAAY,IAAI,KAAA;CAC5E,CAAC;CACD,IAAI,KAAK,cAAc,KAAK,aAAa,UAAU;CA4BnD,OAAO;EAzBL,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;EAC1C,GAAI,OAAO,gBAAgB,EAAE,eAAe,OAAO,cAAc,IAAI,CAAC;EACtE,MAAM;EACN;EACA,QAAQ;GACN,GAAG;GACH,SAAS,OAAO;GAChB,GAAI,OAAO,cAAc,EAAE,OAAO,OAAO,YAAY,IAAI,CAAC;GAC1D,GAAI,OAAO,kBAAkB,EAAE,WAAW,OAAO,gBAAgB,IAAI,CAAC;EACxE;EACA,GAAI,kBAAkB,CAAC,IAAI,EAAE,MAAM,YAAY,gBAAgB,SAAS,cAAc;EACtF,MAAM,OAAO;EACb;EACA,cAAc,OAAO,WAAW;EAChC,qBAAqB,OAAO,YAAY;EACxC;EACA,cAAc;EACd;EACA,UAAU,WAAW,KAAK;EAC1B,eAAe,OAAO;EACtB;EACA,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;EACzB,KAAK;CAGM;AACf;;;;;;;;;;AC13BA,SAAgB,gBACd,UACwC;CACxC,MAAM,yBAAyB,iBAAiB,QAAQ;CAExD,OAAO;EACL,WAAW,SAAS;EACpB,iBAAiB,SAAS;EAE1B,MAAM,SAAS,OAAO,CAAC,GAAG;GACxB,MAAM,EAAE,OAAO,OAAO,QAAQ,QAAQ,WAAW,SAAS,GAAG,iBAAiB;GAC9E,MAAM,gBAAgB,SAAS,SAAS;GACxC,MAAM,kBAAkB,WAAW,SAAS;GAC5C,MAAM,iBAAiB,UAAU,SAAS,UAAU,yBAAyB,KAAK,IAAI;GACtF,MAAM,kBACJ,aAAa,WACb,uBAAuB,YACtB,eAAe,WAAW,QAAQ,IAAI,wBAAwB,IAAI,KAAA;GACrE,MAAM,cAAoD;IACxD,GAAG;IACH,GAAG;IACH,GAAI,kBAAkB,EAAE,SAAS,gBAAgB,IAAI,CAAC;IACtD,QAAQ;IACR,WAAW,aAAa,SAAS;IACjC,WAAW,UAAU,QAAQ,cAAc,iBAAiB,UAAU,GAAG;IACzE,aAAa,mBACX,iBACA,aAAa,eAAe,SAAS,WACvC;IACA,QAAQ,eAAe,QAAQ,SAAS,SAAS,KAAK;GACxD;GACA,IAAI,YAAY,SAAS,KAAA,GACvB,YAAY,OAAO,uBAAuB,YAAY,MAAM,MAAM;GACpE,OAAO,QAA8B,WAAW;EAClD;EAEA,MAAM,QAAQ,OAAO,CAAC,GAAG;GACvB,MAAM,EACJ,QAAQ,gBACR,cAAc,sBACd,GAAG,sBACD;GACJ,MAAM,SAAS,aAAa,UAAU,iBAAiB;GACvD,MAAM,SAAS,YAAY,SAAS,QAAQ,cAAc;GAC1D,MAAM,eAAe,kBAAkB,SAAS,cAAc,oBAAoB;GAClF,OAAO,YAAkC;IACvC,GAAG;IACH,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;IAC3B,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;GACzC,CAAC;EACH;CACF;AACF;AAOA,SAAS,iBACP,UAC8C;CAC9C,MAAM,MAAoD,CAAC;CAC3D,IAAI,SAAS,SAAS,IAAI,UAAU,SAAS;CAC7C,IAAI,SAAS,cAAc,IAAI,eAAe,SAAS;CACvD,IAAI,SAAS,eAAe,IAAI,gBAAgB,SAAS;CACzD,IAAI,SAAS,eAAe,IAAI,gBAAgB,SAAS;CACzD,IAAI,SAAS,aAAa,IAAI,cAAc,SAAS;CACrD,IAAI,SAAS,QAAQ,YAAY,KAAA,GAAW,IAAI,cAAc,SAAS,OAAO;CAC9E,IAAI,SAAS,QAAQ,mBAAmB,KAAA,GACtC,IAAI,iBAAiB,SAAS,OAAO;CACvC,IAAI,SAAS,QAAQ,SAAS,KAAA,GAC5B,IAAI,OAAO,uBAAuB,SAAS,OAAO,MAAM,aAAa;CACvE,OAAO;AACT;AAEA,SAAS,YACP,UACA,WAC+B;CAC/B,MAAM,SAAS,oBAAoB,UAAU,SAAS;CACtD,IAAI,QAAQ,SAAS,KAAA,GAAW,OAAO,OAAO,uBAAuB,OAAO,MAAM,aAAa;CAC/F,OAAO;AACT;AAEA,SAAS,kBACP,UACA,WAC0B;CAC1B,MAAM,SAAS,oBAAoB,UAAU,SAAS;CACtD,IAAI,CAAC,QAAQ,OAAO,KAAA;CACpB,IAAI,CAAC,OAAO,UAAU,KAAK,KAAK,CAAC,OAAO,QAAQ,KAAK,KAAK,CAAC,OAAO,UAAU,KAAK,GAC/E,MAAM,IAAI,MACR,oHACF;CAEF,OAAO;AACT;AAEA,SAAS,aAA+B,UAAa,WAAsC;CACzF,IAAI,CAAC,WAAW,OAAO;CACvB,MAAM,SAAS,EAAE,GAAG,SAAS;CAC7B,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,SAAS,GACjD,IAAI,UAAU,KAAA,GAAW,OAAO,OAAO;CAEzC,OAAO;AACT;AAEA,SAAS,oBACP,UACA,WACe;CACf,IAAI,CAAC,YAAY,CAAC,WAAW,OAAO,KAAA;CACpC,OAAO,aAAa,YAAa,CAAC,GAAS,SAAS;AACtD;AAEA,SAAS,eACP,QACA,cACqC;CACrC,IAAI,WAAW,KAAA,GAAW;EACxB,IAAI,OAAO,WAAW,GACpB,MAAM,IAAI,MAAM,oDAAoD;EAEtE,OAAO;CACT;CACA,OAAO,CAAC,YAAY;AACtB;AAEA,SAAS,uBAAuB,OAAe,OAAuB;CACpE,IAAI,CAAC,OAAO,UAAU,KAAK,KAAK,QAAQ,GACtC,MAAM,IAAI,MAAM,oBAAoB,MAAM,4BAA4B;CAExE,OAAO;AACT"}

package/dist/eval-campaign-BLEfBGbA.js

@@ -1,343 +0,0 @@
-import { c as verifyAgentProfileCell, i as buildAgentProfileCell, l as hashJson } from "./agent-profile-cell-B1V1gUZs.js";
-import { c as validateRunRecord } from "./run-record-DhEJgoPz.js";
-import { t as researchReport } from "./summary-report-BsXJoYX5.js";
-import { t as FileSystemRawProviderSink } from "./raw-provider-sink-75w4FWBj.js";
-import { t as TraceEmitter } from "./emitter-DbMXUIVP.js";
-import { n as assertRunCaptured, t as RunIntegrityError } from "./integrity-W4CADcFc.js";
-//#region src/eval-campaign.ts
-/**
-* EvalCampaign — opinionated matrix runner that wires the four
-* capture-integrity directives by construction.
-*
-* The canonical benchmark shape — matrix runner → for each
-* (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the
-* run → analyze — has a bug class at the integration boundary: raw
-* events not captured, route silently wrong, integrity not asserted,
-* analyst never run. The directives in `SKILL.md § Capture integrity`
-* are the mitigations.
-*
-* `EvalCampaign` is the structural fix — consumers don't wire the
-* integrity surface themselves; the campaign owns it. Specifically:
-*
-*   - constructs a per-run `TraceStore` and `RawProviderSink` via factories
-*   - builds the run's `ChatClient` through `chatFactory`, handing it the
-*     run's raw sink and trace context — a transport built any other way has
-*     no raw HTTP envelope, and `assertRunCaptured` says so
-*   - calls `assertRunCaptured` after every `endRun` and routes failures
-*     through a configurable policy (`throw` / `mark_failed` / `log`)
-*   - assembles per-run `RunRecord`s and runs `researchReport` at the end
-*     so the campaign artifact is launch-decision-grade by default
-*   - embeds the campaign fingerprint (a SHA-256 over the canonicalised
-*     run set) and optional `preregistrationHash` in the report
-*
-* The runner contract is intentionally narrow: produce a `CampaignRunOutcome`
-* given a fully-wired `CampaignRunContext`. Everything orchestration-shaped
-* lives in the campaign. This is the inversion-of-control point — consumers
-* stop writing matrix runners and start writing scenario-runners.
-*
-* Out of scope for v1 (tracked in `docs/research-report-methodology.md`):
-*
-*   - Distributed/cluster execution (concurrency is local async)
-*   - Adaptive sampling / sequential interim looks
-*   - Resume from partial state across crashes
-*   - LLM-call retry beyond what the caller's transport already does
-*/
-const DEFAULT_INTEGRITY = {
-	llmSpansMin: 1,
-	requireRawCoverageOfLlmSpans: true,
-	requireOutcome: true
-};
-async function runEvalCampaign(opts) {
-	if (opts.variants.length === 0) throw new Error("runEvalCampaign: variants must be non-empty.");
-	if (opts.scenarios.length === 0) throw new Error("runEvalCampaign: scenarios must be non-empty.");
-	const variantIds = /* @__PURE__ */ new Set();
-	for (const v of opts.variants) {
-		if (variantIds.has(v.id)) throw new Error(`runEvalCampaign: duplicate variant id "${v.id}".`);
-		variantIds.add(v.id);
-	}
-	const scenarioIds = /* @__PURE__ */ new Set();
-	for (const s of opts.scenarios) {
-		if (scenarioIds.has(s.scenarioId)) throw new Error(`runEvalCampaign: duplicate scenarioId "${s.scenarioId}".`);
-		scenarioIds.add(s.scenarioId);
-	}
-	if (opts.report?.comparator && !variantIds.has(opts.report.comparator)) throw new Error(`runEvalCampaign: report.comparator "${opts.report.comparator}" is not a configured variantId.`);
-	if (!opts.commitSha) throw new Error("runEvalCampaign: commitSha is required (every RunRecord needs it).");
-	const seeds = opts.seeds ?? [
-		0,
-		1,
-		2
-	];
-	const splitTag = opts.splitTag ?? "holdout";
-	const concurrency = Math.max(1, opts.concurrency ?? 1);
-	const integrity = {
-		...DEFAULT_INTEGRITY,
-		...opts.integrity ?? {}
-	};
-	const onIntegrityFailure = opts.onIntegrityFailure ?? "mark_failed";
-	const now = opts.now ?? (() => Date.now());
-	const executionRef = opts.executionRef ?? null;
-	const preregistrationHash = opts.preregistrationHash ?? null;
-	const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir);
-	const campaignFingerprint = await hashJson({
-		campaignId: opts.campaignId,
-		variants: opts.variants.map((v) => v.id).sort(),
-		scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),
-		seeds: [...seeds].sort((a, b) => a - b),
-		splitTag,
-		comparator: opts.report?.comparator ?? null,
-		executionRef,
-		preregistrationHash
-	});
-	const cells = [];
-	for (const variant of opts.variants) for (const scenario of opts.scenarios) for (const seed of seeds) cells.push({
-		variant,
-		scenario,
-		seed
-	});
-	const startedAt = new Date(now()).toISOString();
-	const runs = [];
-	const integrityReports = [];
-	const failedRuns = [];
-	let cursor = 0;
-	let aborting = false;
-	const genuineErrors = [];
-	const openRuns = /* @__PURE__ */ new Map();
-	async function worker() {
-		while (!aborting) {
-			const i = cursor++;
-			if (i >= cells.length) return;
-			const cell = cells[i];
-			try {
-				const result = await runOneCell(cell);
-				runs.push(result.record);
-				integrityReports.push(result.integrity);
-			} catch (err) {
-				if (err instanceof CellExecutionError) {
-					failedRuns.push(err.failed);
-					if (err.integrity) integrityReports.push(err.integrity);
-				} else {
-					genuineErrors.push(err);
-					aborting = true;
-					return;
-				}
-			}
-		}
-	}
-	async function runOneCell(cell) {
-		const runId = (opts.runId ?? defaultRunId)({
-			campaignId: opts.campaignId,
-			runId: "",
-			variantId: cell.variant.id,
-			scenarioId: cell.scenario.scenarioId,
-			seed: cell.seed
-		});
-		const factoryParams = {
-			campaignId: opts.campaignId,
-			runId,
-			variantId: cell.variant.id,
-			scenarioId: cell.scenario.scenarioId,
-			seed: cell.seed
-		};
-		const store = opts.storeFactory(factoryParams);
-		const rawSink = rawSinkFactory(factoryParams);
-		const emitter = new TraceEmitter(store, {
-			runId,
-			now: opts.now,
-			onRunComplete: opts.onRunComplete
-		});
-		openRuns.set(runId, emitter);
-		const chat = opts.chatFactory({
-			rawSink,
-			runId
-		});
-		const ctx = {
-			runId,
-			experimentId: opts.campaignId,
-			variant: cell.variant.payload,
-			variantId: cell.variant.id,
-			scenarioId: cell.scenario.scenarioId,
-			scenarioTags: cell.scenario.tags ?? {},
-			seed: cell.seed,
-			splitTag,
-			emitter,
-			store,
-			rawSink,
-			chat
-		};
-		try {
-			const wallStart = now();
-			let outcome;
-			try {
-				outcome = await opts.runner(ctx);
-			} catch (err) {
-				const message = err instanceof Error ? err.message : String(err);
-				await finalizeAbort(emitter, runId, message);
-				throw new CellExecutionError({
-					runId,
-					variantId: cell.variant.id,
-					scenarioId: cell.scenario.scenarioId,
-					seed: cell.seed,
-					reason: "runner_threw",
-					error: message
-				});
-			}
-			const wallMs = now() - wallStart;
-			const integrityReport = await assertRunCaptured(store, runId, {
-				...integrity,
-				rawSink
-			});
-			if (!integrityReport.ok) switch (onIntegrityFailure) {
-				case "throw": throw new RunIntegrityError(integrityReport);
-				case "mark_failed": throw new CellExecutionError({
-					runId,
-					variantId: cell.variant.id,
-					scenarioId: cell.scenario.scenarioId,
-					seed: cell.seed,
-					reason: "integrity_failed",
-					error: integrityReport.issues.map((i) => i.code).join(", ")
-				}, integrityReport);
-			}
-			const recordOutcome = { raw: outcome.raw ?? {} };
-			if (splitTag === "holdout") recordOutcome.holdoutScore = outcome.score;
-			else recordOutcome.searchScore = outcome.score;
-			if (outcome.judgeScores !== void 0) recordOutcome.judgeScores = outcome.judgeScores;
-			const record = {
-				runId,
-				experimentId: opts.campaignId,
-				candidateId: cell.variant.id,
-				seed: cell.seed,
-				model: outcome.model,
-				promptHash: outcome.promptHash,
-				configHash: outcome.configHash,
-				commitSha: opts.commitSha,
-				wallMs,
-				costUsd: outcome.costUsd,
-				costProvenance: outcome.costProvenance,
-				tokenUsage: outcome.tokenUsage,
-				terminalOutcome: "succeeded",
-				judgeMetadata: outcome.judgeMetadata,
-				outcome: recordOutcome,
-				...outcome.failureClass ? { failureClass: outcome.failureClass } : {},
-				failureMode: outcome.failureMode,
-				splitTag,
-				scenarioId: cell.scenario.scenarioId
-			};
-			const profileSource = outcome.agentProfile ?? (typeof opts.agentProfile === "function" ? await opts.agentProfile({
-				campaignId: opts.campaignId,
-				runId,
-				variantId: cell.variant.id,
-				scenarioId: cell.scenario.scenarioId,
-				seed: cell.seed,
-				variant: cell.variant.payload,
-				scenarioTags: cell.scenario.tags ?? {}
-			}) : opts.agentProfile);
-			if (profileSource !== void 0) {
-				const agentProfile = await resolveAgentProfileCell(profileSource);
-				assertAgentProfileMatchesRun(agentProfile, outcome.model, outcome.promptHash);
-				record.agentProfile = agentProfile;
-			}
-			return {
-				record: validateRunRecord(record),
-				integrity: integrityReport
-			};
-		} finally {
-			openRuns.delete(runId);
-		}
-	}
-	const workers = Array.from({ length: Math.min(concurrency, cells.length) }, () => worker());
-	await Promise.allSettled(workers);
-	for (const [runId, emitter] of openRuns) await finalizeAbort(emitter, runId, "campaign aborted: genuine error in a sibling run");
-	openRuns.clear();
-	if (genuineErrors.length > 0) throw genuineErrors.length === 1 ? genuineErrors[0] : new AggregateError(genuineErrors, `runEvalCampaign: ${genuineErrors.length} runs failed with genuine (non-run-level) errors`);
-	let report;
-	if (opts.report) {
-		const reportOpts = {
-			...opts.report,
-			comparator: opts.report.comparator,
-			split: splitTag === "dev" ? "search" : splitTag,
-			generatedAt: new Date(now()).toISOString(),
-			preregistrationHash: preregistrationHash ?? void 0
-		};
-		report = await researchReport(runs, reportOpts);
-	}
-	const endedAt = new Date(now()).toISOString();
-	return {
-		campaignId: opts.campaignId,
-		campaignFingerprint,
-		preregistrationHash,
-		runs,
-		integrityReports,
-		failedRuns,
-		report,
-		startedAt,
-		endedAt
-	};
-}
-var CellExecutionError = class extends Error {
-	failed;
-	integrity;
-	constructor(failed, integrity) {
-		super(`cell ${failed.variantId}/${failed.scenarioId}@${failed.seed} failed: ${failed.reason}`);
-		this.failed = failed;
-		this.integrity = integrity;
-	}
-};
-/**
-* Abort a run whose owning work threw or was orphaned by a sibling's genuine
-* error, finalizing the emitter so hooks fire and the store records a terminal
-* status. Safe to call unconditionally: it only aborts runs that are still
-* `running`. Two no-op cases are intentional and benign:
-*
-*   - the run was never started (absent from the store) — nothing to finalize
-*   - the run is already terminal (`completed` / `failed` / `aborted`) —
-*     finalizing again would overwrite the real outcome (e.g. flip a passed run
-*     to `aborted` with `{ pass: false, notes: reason }`), so we leave it alone
-*
-* Any OTHER failure of the store read or the abort write (disk full, FS fault,
-* backend down) is a genuine diagnostic and propagates rather than being
-* swallowed.
-*/
-async function finalizeAbort(emitter, runId, reason) {
-	const existing = await emitter.traceStore.getRun(runId);
-	if (existing === void 0) return;
-	if (existing.status !== "running") return;
-	const droppedBefore = emitter.captureStats().dropped;
-	await emitter.abortRun(reason);
-	const after = emitter.captureStats();
-	if (after.dropped > droppedBefore) throw new Error(`finalizeAbort: run ${runId} could not be finalized: ${after.lastError}`);
-}
-function defaultRawSinkFactory(workDir) {
-	return (params) => {
-		if (!workDir) throw new Error("runEvalCampaign: rawSinkFactory not supplied and workDir not set. Pass either to enable raw provider capture, or pass `new NoopRawProviderSink()` via rawSinkFactory to opt out explicitly.");
-		return new FileSystemRawProviderSink({ dir: `${workDir}/raw-events/${params.runId}` });
-	};
-}
-async function resolveAgentProfileCell(input) {
-	if (isAgentProfileCell(input)) {
-		if (!await verifyAgentProfileCell(input)) throw new Error(`runEvalCampaign: agentProfile.cellId does not match its content`);
-		return input;
-	}
-	return buildAgentProfileCell(input);
-}
-function isAgentProfileCell(input) {
-	return "schemaVersion" in input && "cellId" in input;
-}
-function assertAgentProfileMatchesRun(profile, model, promptHash) {
-	if (profile.model !== void 0 && profile.model !== model) throw new Error(`runEvalCampaign: agentProfile.model "${profile.model}" does not match outcome.model "${model}"`);
-	if (profile.promptHash !== void 0 && profile.promptHash !== promptHash) throw new Error(`runEvalCampaign: agentProfile.promptHash "${profile.promptHash}" does not match outcome.promptHash "${promptHash}"`);
-}
-function defaultRunId(params) {
-	const base = `${params.campaignId}::${params.variantId}::${params.scenarioId}::${params.seed}`;
-	let h1 = 2166136261;
-	let h2 = 305419896;
-	for (let i = 0; i < base.length; i++) {
-		const c = base.charCodeAt(i);
-		h1 = Math.imul(h1 ^ c, 16777619) >>> 0;
-		h2 = Math.imul(h2 ^ c, 2654435761) >>> 0;
-	}
-	return `run-${h1.toString(16).padStart(8, "0")}${h2.toString(16).padStart(8, "0")}`;
-}
-//#endregion
-export { runEvalCampaign as t };
-
-//# sourceMappingURL=eval-campaign-BLEfBGbA.js.map

package/dist/eval-campaign-BLEfBGbA.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"eval-campaign-BLEfBGbA.js","names":[],"sources":["../src/eval-campaign.ts"],"sourcesContent":["/**\n * EvalCampaign — opinionated matrix runner that wires the four\n * capture-integrity directives by construction.\n *\n * The canonical benchmark shape — matrix runner → for each\n * (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the\n * run → analyze — has a bug class at the integration boundary: raw\n * events not captured, route silently wrong, integrity not asserted,\n * analyst never run. The directives in `SKILL.md § Capture integrity`\n * are the mitigations.\n *\n * `EvalCampaign` is the structural fix — consumers don't wire the\n * integrity surface themselves; the campaign owns it. Specifically:\n *\n *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories\n *   - builds the run's `ChatClient` through `chatFactory`, handing it the\n *     run's raw sink and trace context — a transport built any other way has\n *     no raw HTTP envelope, and `assertRunCaptured` says so\n *   - calls `assertRunCaptured` after every `endRun` and routes failures\n *     through a configurable policy (`throw` / `mark_failed` / `log`)\n *   - assembles per-run `RunRecord`s and runs `researchReport` at the end\n *     so the campaign artifact is launch-decision-grade by default\n *   - embeds the campaign fingerprint (a SHA-256 over the canonicalised\n *     run set) and optional `preregistrationHash` in the report\n *\n * The runner contract is intentionally narrow: produce a `CampaignRunOutcome`\n * given a fully-wired `CampaignRunContext`. Everything orchestration-shaped\n * lives in the campaign. This is the inversion-of-control point — consumers\n * stop writing matrix runners and start writing scenario-runners.\n *\n * Out of scope for v1 (tracked in `docs/research-report-methodology.md`):\n *\n *   - Distributed/cluster execution (concurrency is local async)\n *   - Adaptive sampling / sequential interim looks\n *   - Resume from partial state across crashes\n *   - LLM-call retry beyond what the caller's transport already does\n */\n\nimport {\n  type AgentProfileCell,\n  type AgentProfileCellInput,\n  buildAgentProfileCell,\n  verifyAgentProfileCell,\n} from './agent-profile-cell'\nimport type { ChatClient } from './analyst/chat-client'\nimport { hashJson } from './pre-registration'\nimport type {\n  JudgeScoresRecord,\n  RunCostProvenance,\n  RunJudgeMetadata,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTaskFailure,\n  RunTokenUsage,\n} from './run-record'\nimport { validateRunRecord } from './run-record'\nimport { type ResearchReport, type ResearchReportOptions, researchReport } from './summary-report'\nimport type { RunCompleteHook } from './trace/emitter'\nimport { TraceEmitter } from './trace/emitter'\nimport {\n  assertRunCaptured,\n  RunIntegrityError,\n  type RunIntegrityExpectations,\n  type RunIntegrityReport,\n} from './trace/integrity'\nimport { FileSystemRawProviderSink, type RawProviderSink } from './trace/raw-provider-sink'\nimport type { TraceStore } from './trace/store'\n\n// ── Public types ─────────────────────────────────────────────────────────\n\nexport interface CampaignVariant<V> {\n  id: string\n  payload: V\n}\n\nexport interface CampaignScenario {\n  scenarioId: string\n  /** Free-form metadata propagated to runs and reports. */\n  tags?: Record<string, string>\n}\n\nexport interface CampaignRunContext<V> {\n  /** Stable run id. The campaign generates this; the runner does not. */\n  runId: string\n  /** Logical experiment id (campaignId by default; overridable per-run via opts). */\n  experimentId: string\n  variant: V\n  variantId: string\n  scenarioId: string\n  scenarioTags: Record<string, string>\n  seed: number\n  splitTag: RunSplitTag\n  /**\n   * The TraceEmitter for this run, with `onRunComplete` hooks pre-wired\n   * (analyst auto-execution if configured, plus integrity check). The\n   * runner MUST call `emitter.startRun` before doing any work and either\n   * `emitter.endRun` or `emitter.abortRun` before returning.\n   */\n  emitter: TraceEmitter\n  store: TraceStore\n  rawSink: RawProviderSink\n  /**\n   * The run's model transport, built by `chatFactory` with this run's\n   * `rawSink` and `runId` already bound.\n   */\n  chat: ChatClient\n}\n\n/** What the campaign binds into the run's transport. */\nexport interface CampaignChatWiring {\n  /**\n   * Raw provider sink for this run. Bind it into the transport: the campaign's\n   * integrity check requires every LLM span to carry a matching raw request\n   * event, so a transport built without it fails `assertRunCaptured`.\n   */\n  rawSink: RawProviderSink\n  runId: string\n}\n\ninterface CampaignRunOutcomeFields {\n  /** Did the run pass? Mirrors `RunOutcome.pass` semantics. */\n  pass: boolean\n  /** Score for the run on its split. Maps to `searchScore` or `holdoutScore`. */\n  score: number\n  /** Cost in USD, or null when the runner could not capture it. */\n  costUsd: number | null\n  /** Source of the cost amount. */\n  costProvenance: RunCostProvenance\n  tokenUsage: RunTokenUsage\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** sha256 of the effective prompt sent to the model. */\n  promptHash: string\n  /** sha256 of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Optional extra numeric metrics to land in `outcome.raw`. */\n  raw?: Record<string, number>\n  /** Optional judge metadata when a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /**\n   * Optional per-judge / per-dim breakdown for ensemble-judged runs.\n   * Propagated to `outcome.judgeScores` on the resulting `RunRecord`.\n   * Single-judge or scalar-only runs leave this unset.\n   */\n  judgeScores?: JudgeScoresRecord\n  /**\n   * Agent profile cell observed by the runner. When supplied, it overrides\n   * `EvalCampaignOptions.agentProfile` for this run and must match the\n   * outcome's `model` and `promptHash`.\n   */\n  agentProfile?: AgentProfileCell | AgentProfileCellInput\n}\n\n/** Campaign result with the same task-failure invariant as `RunRecord`. */\nexport type CampaignRunOutcome = CampaignRunOutcomeFields & RunTaskFailure\n\nexport type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>\n\nexport type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log'\n\nexport interface EvalCampaignOptions<V> {\n  /**\n   * Stable id for the campaign. Used as the default `experimentId` on\n   * every run, and folded into the campaign fingerprint.\n   */\n  campaignId: string\n  variants: CampaignVariant<V>[]\n  scenarios: CampaignScenario[]\n  /** Default `[0, 1, 2]`. */\n  seeds?: number[]\n  /** Default `'holdout'` — the split that anchors a launch decision. */\n  splitTag?: RunSplitTag\n  /** Git SHA the campaign is run against. Mandatory; `RunRecord` rejects unset. */\n  commitSha: string\n  /**\n   * Build the model transport for one run. agent-eval executes no paid model:\n   * the caller owns the transport and the credential never enters this\n   * package. The campaign calls this once per run and passes the run's raw\n   * provider sink and `runId`, so a transport that binds them captures the\n   * raw HTTP envelope `assertRunCaptured` checks for.\n   */\n  chatFactory: (wiring: CampaignChatWiring) => ChatClient\n  /**\n   * Caller-declared identity of the execution route, folded into the campaign\n   * fingerprint so two campaigns run against different endpoints do not share\n   * one identity. agent-eval no longer knows the endpoint; the owner of\n   * execution names it.\n   */\n  executionRef?: string\n  /**\n   * Per-run TraceStore factory. Common shape: a fresh store per run keyed\n   * on `runId`. Implementations that share a store across the campaign\n   * are valid — the campaign only writes through `emitter`.\n   */\n  storeFactory: (params: CampaignFactoryParams) => TraceStore\n  /**\n   * Per-run RawProviderSink factory. Defaults to `FileSystemRawProviderSink`\n   * rooted at `${workDir}/raw-events/${runId}` if `workDir` is supplied;\n   * otherwise required. Forensic capture is non-negotiable in a campaign\n   * run — pass `NoopRawProviderSink` explicitly if you want to opt out.\n   */\n  rawSinkFactory?: (params: CampaignFactoryParams) => RawProviderSink\n  /**\n   * Filesystem root for default `rawSinkFactory`. Ignored if\n   * `rawSinkFactory` is supplied.\n   */\n  workDir?: string\n  /**\n   * Extra `onRunComplete` hooks the campaign appends (after its own\n   * integrity-check hook). Pass `traceAnalystOnRunComplete(...)` here.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /**\n   * Per-run integrity expectations. Defaults to:\n   *   `{ llmSpansMin: 1, requireRawCoverageOfLlmSpans: true, requireOutcome: true }`.\n   * Override (e.g. `{ llmSpansMin: 0 }`) for runs that don't call LLMs.\n   */\n  integrity?: RunIntegrityExpectations\n  /** Behaviour when integrity fails. Default `'mark_failed'`. */\n  onIntegrityFailure?: CampaignIntegrityPolicy\n  /**\n   * Per-run runner. Receives a fully-wired context; produces an outcome\n   * the campaign converts into a `RunRecord`.\n   */\n  runner: CampaignRunner<V>\n  /**\n   * If set, the campaign computes `researchReport` at the end. `comparator`\n   * is a `variantId`. Other fields are forwarded verbatim.\n   */\n  report?: { comparator?: string } & Omit<\n    ResearchReportOptions,\n    'comparator' | 'preregistrationHash' | 'generatedAt'\n  >\n  /**\n   * Hash of a signed `HypothesisManifest` (see `pre-registration.ts`).\n   * Embedded in the campaign fingerprint and the research report.\n   */\n  preregistrationHash?: string\n  /** Local concurrency. Default `1` (sequential). */\n  concurrency?: number\n  /**\n   * Override the time source. Tests pass a mock to make wallMs deterministic.\n   */\n  now?: () => number\n  /** Override the runId generator. Tests pin this. */\n  runId?: (params: CampaignFactoryParams) => string\n  /**\n   * Agent profile cell for campaign runs. Static profiles can pass an object;\n   * routers or variant-specific harnesses can pass a factory. The campaign\n   * stamps the built cell onto every `RunRecord` and rejects profile/model or\n   * profile/prompt contradictions.\n   */\n  agentProfile?:\n    | AgentProfileCell\n    | AgentProfileCellInput\n    | ((\n        params: CampaignFactoryParams & {\n          variant: V\n          scenarioTags: Record<string, string>\n        },\n      ) =>\n        | AgentProfileCell\n        | AgentProfileCellInput\n        | Promise<AgentProfileCell | AgentProfileCellInput>)\n}\n\nexport interface CampaignFactoryParams {\n  campaignId: string\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n}\n\nexport interface FailedRun {\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n  reason: string\n  error?: string\n}\n\nexport interface EvalCampaignResult {\n  campaignId: string\n  /** SHA-256 over canonicalised `(variantIds, scenarioIds, seeds, comparator, splitTag, executionRef, preregistrationHash)`. */\n  campaignFingerprint: string\n  preregistrationHash: string | null\n  /** Successful runs only. Failed runs land in `failedRuns`. */\n  runs: RunRecord[]\n  /** Integrity reports for every successful run. */\n  integrityReports: RunIntegrityReport[]\n  failedRuns: FailedRun[]\n  /** Computed when `report` is set on options. */\n  report?: ResearchReport\n  startedAt: string\n  endedAt: string\n}\n\n// ── Implementation ───────────────────────────────────────────────────────\n\nconst DEFAULT_INTEGRITY: RunIntegrityExpectations = {\n  llmSpansMin: 1,\n  requireRawCoverageOfLlmSpans: true,\n  requireOutcome: true,\n}\n\nexport async function runEvalCampaign<V>(\n  opts: EvalCampaignOptions<V>,\n): Promise<EvalCampaignResult> {\n  // ── Preflight ──────────────────────────────────────────────────────\n  if (opts.variants.length === 0) {\n    throw new Error('runEvalCampaign: variants must be non-empty.')\n  }\n  if (opts.scenarios.length === 0) {\n    throw new Error('runEvalCampaign: scenarios must be non-empty.')\n  }\n  const variantIds = new Set<string>()\n  for (const v of opts.variants) {\n    if (variantIds.has(v.id)) {\n      throw new Error(`runEvalCampaign: duplicate variant id \"${v.id}\".`)\n    }\n    variantIds.add(v.id)\n  }\n  const scenarioIds = new Set<string>()\n  for (const s of opts.scenarios) {\n    if (scenarioIds.has(s.scenarioId)) {\n      throw new Error(`runEvalCampaign: duplicate scenarioId \"${s.scenarioId}\".`)\n    }\n    scenarioIds.add(s.scenarioId)\n  }\n  if (opts.report?.comparator && !variantIds.has(opts.report.comparator)) {\n    throw new Error(\n      `runEvalCampaign: report.comparator \"${opts.report.comparator}\" is not a configured variantId.`,\n    )\n  }\n  if (!opts.commitSha) {\n    throw new Error('runEvalCampaign: commitSha is required (every RunRecord needs it).')\n  }\n\n  const seeds = opts.seeds ?? [0, 1, 2]\n  const splitTag: RunSplitTag = opts.splitTag ?? 'holdout'\n  const concurrency = Math.max(1, opts.concurrency ?? 1)\n  const integrity = { ...DEFAULT_INTEGRITY, ...(opts.integrity ?? {}) }\n  const onIntegrityFailure: CampaignIntegrityPolicy = opts.onIntegrityFailure ?? 'mark_failed'\n  const now = opts.now ?? (() => Date.now())\n  const executionRef = opts.executionRef ?? null\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir)\n\n  // ── Fingerprint ────────────────────────────────────────────────────\n  const campaignFingerprint = await hashJson({\n    campaignId: opts.campaignId,\n    variants: opts.variants.map((v) => v.id).sort(),\n    scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),\n    seeds: [...seeds].sort((a, b) => a - b),\n    splitTag,\n    comparator: opts.report?.comparator ?? null,\n    executionRef,\n    preregistrationHash,\n  })\n\n  // ── Plan the matrix ────────────────────────────────────────────────\n  type Cell = { variant: CampaignVariant<V>; scenario: CampaignScenario; seed: number }\n  const cells: Cell[] = []\n  for (const variant of opts.variants) {\n    for (const scenario of opts.scenarios) {\n      for (const seed of seeds) {\n        cells.push({ variant, scenario, seed })\n      }\n    }\n  }\n\n  const startedAt = new Date(now()).toISOString()\n  const runs: RunRecord[] = []\n  const integrityReports: RunIntegrityReport[] = []\n  const failedRuns: FailedRun[] = []\n\n  // ── Execute (bounded-concurrency worker pool) ──────────────────────\n  // A genuine (non-CellExecutionError) error from any worker is a bug, not a\n  // run-level failure. We must NOT let `Promise.all` reject mid-flight and\n  // orphan the other workers' in-progress runs (emitters never finalized,\n  // sinks/handles leak, partial work silently discarded). Instead: capture the\n  // first genuine error, stop dispatching new cells so in-flight workers wind\n  // down, finalize every still-open run, then re-throw the aggregated error.\n  let cursor = 0\n  let aborting = false\n  const genuineErrors: unknown[] = []\n  // Emitters for runs that are currently mid-flight, keyed by runId. A worker\n  // registers its emitter before invoking the runner and removes it once the\n  // run is finalized (via endRun/abortRun, success or failure). Anything left\n  // here after the pool settles is an orphan we must abort.\n  const openRuns = new Map<string, TraceEmitter>()\n\n  async function worker(): Promise<void> {\n    while (!aborting) {\n      const i = cursor++\n      if (i >= cells.length) return\n      const cell = cells[i]!\n      try {\n        const result = await runOneCell(cell)\n        runs.push(result.record)\n        integrityReports.push(result.integrity)\n      } catch (err) {\n        if (err instanceof CellExecutionError) {\n          failedRuns.push(err.failed)\n          if (err.integrity) integrityReports.push(err.integrity)\n        } else {\n          // Genuine bug — not a runner failure, not an integrity failure.\n          // Capture it and stop dispatching so peers wind down gracefully;\n          // re-thrown after the pool settles. Do not surface here — that would\n          // reject Promise.all and orphan the other workers' runs.\n          genuineErrors.push(err)\n          aborting = true\n          return\n        }\n      }\n    }\n  }\n\n  async function runOneCell(\n    cell: Cell,\n  ): Promise<{ record: RunRecord; integrity: RunIntegrityReport }> {\n    const runId = (opts.runId ?? defaultRunId)({\n      campaignId: opts.campaignId,\n      runId: '', // unused by default generator\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    })\n    const factoryParams: CampaignFactoryParams = {\n      campaignId: opts.campaignId,\n      runId,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    }\n    const store = opts.storeFactory(factoryParams)\n    const rawSink = rawSinkFactory(factoryParams)\n\n    const emitter = new TraceEmitter(store, {\n      runId,\n      now: opts.now,\n      onRunComplete: opts.onRunComplete,\n    })\n    // Track this run as open so a genuine error elsewhere in the pool can\n    // finalize it instead of orphaning it. Removed in the finally below.\n    openRuns.set(runId, emitter)\n\n    const chat = opts.chatFactory({ rawSink, runId })\n\n    const ctx: CampaignRunContext<V> = {\n      runId,\n      experimentId: opts.campaignId,\n      variant: cell.variant.payload,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      scenarioTags: cell.scenario.tags ?? {},\n      seed: cell.seed,\n      splitTag,\n      emitter,\n      store,\n      rawSink,\n      chat,\n    }\n\n    try {\n      const wallStart = now()\n      let outcome: CampaignRunOutcome\n      try {\n        outcome = await opts.runner(ctx)\n      } catch (err) {\n        const message = err instanceof Error ? err.message : String(err)\n        // The runner threw mid-execution. Abort the run so the emitter\n        // finalizes. The only benign abortRun failure is \"the runner never\n        // started the run\" (nothing to finalize). A store-write failure (disk\n        // full, FS error) is a genuine diagnostic — surface it rather than\n        // masking it as a plain runner failure.\n        await finalizeAbort(emitter, runId, message)\n        throw new CellExecutionError({\n          runId,\n          variantId: cell.variant.id,\n          scenarioId: cell.scenario.scenarioId,\n          seed: cell.seed,\n          reason: 'runner_threw',\n          error: message,\n        })\n      }\n      const wallMs = now() - wallStart\n\n      const integrityReport = await assertRunCaptured(store, runId, { ...integrity, rawSink })\n      if (!integrityReport.ok) {\n        switch (onIntegrityFailure) {\n          case 'throw':\n            throw new RunIntegrityError(integrityReport)\n          case 'mark_failed':\n            throw new CellExecutionError(\n              {\n                runId,\n                variantId: cell.variant.id,\n                scenarioId: cell.scenario.scenarioId,\n                seed: cell.seed,\n                reason: 'integrity_failed',\n                error: integrityReport.issues.map((i) => i.code).join(', '),\n              },\n              integrityReport,\n            )\n          case 'log':\n            // Caller wants the run admitted with a flagged report; fall through.\n            break\n        }\n      }\n\n      const recordOutcome: RunOutcome = {\n        raw: outcome.raw ?? {},\n      }\n      if (splitTag === 'holdout') recordOutcome.holdoutScore = outcome.score\n      else recordOutcome.searchScore = outcome.score\n      if (outcome.judgeScores !== undefined) recordOutcome.judgeScores = outcome.judgeScores\n\n      const record: RunRecord = {\n        runId,\n        experimentId: opts.campaignId,\n        candidateId: cell.variant.id,\n        seed: cell.seed,\n        model: outcome.model,\n        promptHash: outcome.promptHash,\n        configHash: outcome.configHash,\n        commitSha: opts.commitSha,\n        wallMs,\n        costUsd: outcome.costUsd,\n        costProvenance: outcome.costProvenance,\n        tokenUsage: outcome.tokenUsage,\n        terminalOutcome: 'succeeded',\n        judgeMetadata: outcome.judgeMetadata,\n        outcome: recordOutcome,\n        ...(outcome.failureClass ? { failureClass: outcome.failureClass } : {}),\n        failureMode: outcome.failureMode,\n        splitTag,\n        scenarioId: cell.scenario.scenarioId,\n      }\n      const profileSource =\n        outcome.agentProfile ??\n        (typeof opts.agentProfile === 'function'\n          ? await opts.agentProfile({\n              campaignId: opts.campaignId,\n              runId,\n              variantId: cell.variant.id,\n              scenarioId: cell.scenario.scenarioId,\n              seed: cell.seed,\n              variant: cell.variant.payload,\n              scenarioTags: cell.scenario.tags ?? {},\n            })\n          : opts.agentProfile)\n      if (profileSource !== undefined) {\n        const agentProfile = await resolveAgentProfileCell(profileSource)\n        assertAgentProfileMatchesRun(agentProfile, outcome.model, outcome.promptHash)\n        record.agentProfile = agentProfile\n      }\n      return { record: validateRunRecord(record), integrity: integrityReport }\n    } finally {\n      // This run's worker has finished with it (success, run-level failure, or\n      // genuine error). It is no longer the pool's job to finalize — drop it so\n      // the post-settle sweep doesn't double-abort a finalized run.\n      openRuns.delete(runId)\n    }\n  }\n\n  const workers = Array.from({ length: Math.min(concurrency, cells.length) }, () => worker())\n  // allSettled (not all): a genuine error in one worker must not reject the\n  // pool mid-flight and orphan the others. Each worker captures its own\n  // genuine error into `genuineErrors` and returns; we re-throw below.\n  await Promise.allSettled(workers)\n\n  // Finalize any run still open after the pool wound down. With the\n  // stop-dispatch flag these are the runs that were mid-flight in peer workers\n  // when the first genuine error fired — abort them so their emitters finalize\n  // (hooks fire, store records a terminal status) instead of leaking.\n  for (const [runId, emitter] of openRuns) {\n    await finalizeAbort(emitter, runId, 'campaign aborted: genuine error in a sibling run')\n  }\n  openRuns.clear()\n\n  if (genuineErrors.length > 0) {\n    throw genuineErrors.length === 1\n      ? genuineErrors[0]\n      : new AggregateError(\n          genuineErrors,\n          `runEvalCampaign: ${genuineErrors.length} runs failed with genuine (non-run-level) errors`,\n        )\n  }\n\n  // ── Optional research report ───────────────────────────────────────\n  let report: ResearchReport | undefined\n  if (opts.report) {\n    const reportOpts: ResearchReportOptions = {\n      ...opts.report,\n      comparator: opts.report.comparator,\n      split: splitTag === 'dev' ? 'search' : splitTag,\n      generatedAt: new Date(now()).toISOString(),\n      preregistrationHash: preregistrationHash ?? undefined,\n    }\n    report = await researchReport(runs, reportOpts)\n  }\n\n  const endedAt = new Date(now()).toISOString()\n\n  return {\n    campaignId: opts.campaignId,\n    campaignFingerprint,\n    preregistrationHash,\n    runs,\n    integrityReports,\n    failedRuns,\n    report,\n    startedAt,\n    endedAt,\n  }\n}\n\n// ── Internal ─────────────────────────────────────────────────────────────\n\nclass CellExecutionError extends Error {\n  readonly failed: FailedRun\n  readonly integrity?: RunIntegrityReport\n  constructor(failed: FailedRun, integrity?: RunIntegrityReport) {\n    super(`cell ${failed.variantId}/${failed.scenarioId}@${failed.seed} failed: ${failed.reason}`)\n    this.failed = failed\n    this.integrity = integrity\n  }\n}\n\n/**\n * Abort a run whose owning work threw or was orphaned by a sibling's genuine\n * error, finalizing the emitter so hooks fire and the store records a terminal\n * status. Safe to call unconditionally: it only aborts runs that are still\n * `running`. Two no-op cases are intentional and benign:\n *\n *   - the run was never started (absent from the store) — nothing to finalize\n *   - the run is already terminal (`completed` / `failed` / `aborted`) —\n *     finalizing again would overwrite the real outcome (e.g. flip a passed run\n *     to `aborted` with `{ pass: false, notes: reason }`), so we leave it alone\n *\n * Any OTHER failure of the store read or the abort write (disk full, FS fault,\n * backend down) is a genuine diagnostic and propagates rather than being\n * swallowed.\n */\nasync function finalizeAbort(\n  emitter: TraceEmitter,\n  runId: string,\n  reason: string,\n): Promise<void> {\n  const existing = await emitter.traceStore.getRun(runId)\n  if (existing === undefined) return // run never started; nothing to abort\n  if (existing.status !== 'running') return // already finalized; never overwrite a real outcome\n  // The emitter never throws on a failed store write; it counts it. A failed\n  // abort write is a genuine diagnostic here, so read the count and surface it.\n  const droppedBefore = emitter.captureStats().dropped\n  await emitter.abortRun(reason)\n  const after = emitter.captureStats()\n  if (after.dropped > droppedBefore) {\n    throw new Error(`finalizeAbort: run ${runId} could not be finalized: ${after.lastError}`)\n  }\n}\n\nfunction defaultRawSinkFactory(workDir: string | undefined) {\n  return (params: CampaignFactoryParams): RawProviderSink => {\n    if (!workDir) {\n      throw new Error(\n        'runEvalCampaign: rawSinkFactory not supplied and workDir not set. Pass either to enable raw provider capture, or pass `new NoopRawProviderSink()` via rawSinkFactory to opt out explicitly.',\n      )\n    }\n    return new FileSystemRawProviderSink({\n      dir: `${workDir}/raw-events/${params.runId}`,\n    })\n  }\n}\n\nasync function resolveAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  if (isAgentProfileCell(input)) {\n    if (!(await verifyAgentProfileCell(input))) {\n      throw new Error(`runEvalCampaign: agentProfile.cellId does not match its content`)\n    }\n    return input\n  }\n  return buildAgentProfileCell(input)\n}\n\nfunction isAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): input is AgentProfileCell {\n  return 'schemaVersion' in input && 'cellId' in input\n}\n\nfunction assertAgentProfileMatchesRun(\n  profile: AgentProfileCell,\n  model: string,\n  promptHash: string,\n): void {\n  if (profile.model !== undefined && profile.model !== model) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.model \"${profile.model}\" does not match outcome.model \"${model}\"`,\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== promptHash) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.promptHash \"${profile.promptHash}\" does not match outcome.promptHash \"${promptHash}\"`,\n    )\n  }\n}\n\nfunction defaultRunId(params: CampaignFactoryParams): string {\n  // Stable across re-runs: fingerprint of (campaignId, variantId, scenarioId, seed).\n  // Caller can override via opts.runId for non-deterministic IDs.\n  const base = `${params.campaignId}::${params.variantId}::${params.scenarioId}::${params.seed}`\n  // TWO 32-bit FNV-1a accumulators with different offsets, concatenated for a\n  // 64-bit-wide id. Not crypto-grade; stability and uniqueness are what this\n  // needs. Frozen: the value names a persisted cell, and the second\n  // accumulator's offset is part of the identity, so no single-accumulator\n  // helper can replace it.\n  let h1 = 0x811c9dc5\n  let h2 = 0x12345678\n  for (let i = 0; i < base.length; i++) {\n    const c = base.charCodeAt(i)\n    h1 = Math.imul(h1 ^ c, 0x01000193) >>> 0\n    h2 = Math.imul(h2 ^ c, 0x9e3779b1) >>> 0\n  }\n  return `run-${h1.toString(16).padStart(8, '0')}${h2.toString(16).padStart(8, '0')}`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8SA,MAAM,oBAA8C;CAClD,aAAa;CACb,8BAA8B;CAC9B,gBAAgB;AAClB;AAEA,eAAsB,gBACpB,MAC6B;CAE7B,IAAI,KAAK,SAAS,WAAW,GAC3B,MAAM,IAAI,MAAM,8CAA8C;CAEhE,IAAI,KAAK,UAAU,WAAW,GAC5B,MAAM,IAAI,MAAM,+CAA+C;CAEjE,MAAM,6BAAa,IAAI,IAAY;CACnC,KAAK,MAAM,KAAK,KAAK,UAAU;EAC7B,IAAI,WAAW,IAAI,EAAE,EAAE,GACrB,MAAM,IAAI,MAAM,0CAA0C,EAAE,GAAG,GAAG;EAEpE,WAAW,IAAI,EAAE,EAAE;CACrB;CACA,MAAM,8BAAc,IAAI,IAAY;CACpC,KAAK,MAAM,KAAK,KAAK,WAAW;EAC9B,IAAI,YAAY,IAAI,EAAE,UAAU,GAC9B,MAAM,IAAI,MAAM,0CAA0C,EAAE,WAAW,GAAG;EAE5E,YAAY,IAAI,EAAE,UAAU;CAC9B;CACA,IAAI,KAAK,QAAQ,cAAc,CAAC,WAAW,IAAI,KAAK,OAAO,UAAU,GACnE,MAAM,IAAI,MACR,uCAAuC,KAAK,OAAO,WAAW,iCAChE;CAEF,IAAI,CAAC,KAAK,WACR,MAAM,IAAI,MAAM,oEAAoE;CAGtF,MAAM,QAAQ,KAAK,SAAS;EAAC;EAAG;EAAG;CAAC;CACpC,MAAM,WAAwB,KAAK,YAAY;CAC/C,MAAM,cAAc,KAAK,IAAI,GAAG,KAAK,eAAe,CAAC;CACrD,MAAM,YAAY;EAAE,GAAG;EAAmB,GAAI,KAAK,aAAa,CAAC;CAAG;CACpE,MAAM,qBAA8C,KAAK,sBAAsB;CAC/E,MAAM,MAAM,KAAK,cAAc,KAAK,IAAI;CACxC,MAAM,eAAe,KAAK,gBAAgB;CAC1C,MAAM,sBAAsB,KAAK,uBAAuB;CAExD,MAAM,iBAAiB,KAAK,kBAAkB,sBAAsB,KAAK,OAAO;CAGhF,MAAM,sBAAsB,MAAM,SAAS;EACzC,YAAY,KAAK;EACjB,UAAU,KAAK,SAAS,KAAK,MAAM,EAAE,EAAE,CAAC,CAAC,KAAK;EAC9C,WAAW,KAAK,UAAU,KAAK,MAAM,EAAE,UAAU,CAAC,CAAC,KAAK;EACxD,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;EACtC;EACA,YAAY,KAAK,QAAQ,cAAc;EACvC;EACA;CACF,CAAC;CAID,MAAM,QAAgB,CAAC;CACvB,KAAK,MAAM,WAAW,KAAK,UACzB,KAAK,MAAM,YAAY,KAAK,WAC1B,KAAK,MAAM,QAAQ,OACjB,MAAM,KAAK;EAAE;EAAS;EAAU;CAAK,CAAC;CAK5C,MAAM,YAAY,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;CAC9C,MAAM,OAAoB,CAAC;CAC3B,MAAM,mBAAyC,CAAC;CAChD,MAAM,aAA0B,CAAC;CASjC,IAAI,SAAS;CACb,IAAI,WAAW;CACf,MAAM,gBAA2B,CAAC;CAKlC,MAAM,2BAAW,IAAI,IAA0B;CAE/C,eAAe,SAAwB;EACrC,OAAO,CAAC,UAAU;GAChB,MAAM,IAAI;GACV,IAAI,KAAK,MAAM,QAAQ;GACvB,MAAM,OAAO,MAAM;GACnB,IAAI;IACF,MAAM,SAAS,MAAM,WAAW,IAAI;IACpC,KAAK,KAAK,OAAO,MAAM;IACvB,iBAAiB,KAAK,OAAO,SAAS;GACxC,SAAS,KAAK;IACZ,IAAI,eAAe,oBAAoB;KACrC,WAAW,KAAK,IAAI,MAAM;KAC1B,IAAI,IAAI,WAAW,iBAAiB,KAAK,IAAI,SAAS;IACxD,OAAO;KAKL,cAAc,KAAK,GAAG;KACtB,WAAW;KACX;IACF;GACF;EACF;CACF;CAEA,eAAe,WACb,MAC+D;EAC/D,MAAM,SAAS,KAAK,SAAS,aAAA,CAAc;GACzC,YAAY,KAAK;GACjB,OAAO;GACP,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,MAAM,KAAK;EACb,CAAC;EACD,MAAM,gBAAuC;GAC3C,YAAY,KAAK;GACjB;GACA,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,MAAM,KAAK;EACb;EACA,MAAM,QAAQ,KAAK,aAAa,aAAa;EAC7C,MAAM,UAAU,eAAe,aAAa;EAE5C,MAAM,UAAU,IAAI,aAAa,OAAO;GACtC;GACA,KAAK,KAAK;GACV,eAAe,KAAK;EACtB,CAAC;EAGD,SAAS,IAAI,OAAO,OAAO;EAE3B,MAAM,OAAO,KAAK,YAAY;GAAE;GAAS;EAAM,CAAC;EAEhD,MAAM,MAA6B;GACjC;GACA,cAAc,KAAK;GACnB,SAAS,KAAK,QAAQ;GACtB,WAAW,KAAK,QAAQ;GACxB,YAAY,KAAK,SAAS;GAC1B,cAAc,KAAK,SAAS,QAAQ,CAAC;GACrC,MAAM,KAAK;GACX;GACA;GACA;GACA;GACA;EACF;EAEA,IAAI;GACF,MAAM,YAAY,IAAI;GACtB,IAAI;GACJ,IAAI;IACF,UAAU,MAAM,KAAK,OAAO,GAAG;GACjC,SAAS,KAAK;IACZ,MAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAM/D,MAAM,cAAc,SAAS,OAAO,OAAO;IAC3C,MAAM,IAAI,mBAAmB;KAC3B;KACA,WAAW,KAAK,QAAQ;KACxB,YAAY,KAAK,SAAS;KAC1B,MAAM,KAAK;KACX,QAAQ;KACR,OAAO;IACT,CAAC;GACH;GACA,MAAM,SAAS,IAAI,IAAI;GAEvB,MAAM,kBAAkB,MAAM,kBAAkB,OAAO,OAAO;IAAE,GAAG;IAAW;GAAQ,CAAC;GACvF,IAAI,CAAC,gBAAgB,IACnB,QAAQ,oBAAR;IACE,KAAK,SACH,MAAM,IAAI,kBAAkB,eAAe;IAC7C,KAAK,eACH,MAAM,IAAI,mBACR;KACE;KACA,WAAW,KAAK,QAAQ;KACxB,YAAY,KAAK,SAAS;KAC1B,MAAM,KAAK;KACX,QAAQ;KACR,OAAO,gBAAgB,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,IAAI;IAC5D,GACA,eACF;GAIJ;GAGF,MAAM,gBAA4B,EAChC,KAAK,QAAQ,OAAO,CAAC,EACvB;GACA,IAAI,aAAa,WAAW,cAAc,eAAe,QAAQ;QAC5D,cAAc,cAAc,QAAQ;GACzC,IAAI,QAAQ,gBAAgB,KAAA,GAAW,cAAc,cAAc,QAAQ;GAE3E,MAAM,SAAoB;IACxB;IACA,cAAc,KAAK;IACnB,aAAa,KAAK,QAAQ;IAC1B,MAAM,KAAK;IACX,OAAO,QAAQ;IACf,YAAY,QAAQ;IACpB,YAAY,QAAQ;IACpB,WAAW,KAAK;IAChB;IACA,SAAS,QAAQ;IACjB,gBAAgB,QAAQ;IACxB,YAAY,QAAQ;IACpB,iBAAiB;IACjB,eAAe,QAAQ;IACvB,SAAS;IACT,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;IACrE,aAAa,QAAQ;IACrB;IACA,YAAY,KAAK,SAAS;GAC5B;GACA,MAAM,gBACJ,QAAQ,iBACP,OAAO,KAAK,iBAAiB,aAC1B,MAAM,KAAK,aAAa;IACtB,YAAY,KAAK;IACjB;IACA,WAAW,KAAK,QAAQ;IACxB,YAAY,KAAK,SAAS;IAC1B,MAAM,KAAK;IACX,SAAS,KAAK,QAAQ;IACtB,cAAc,KAAK,SAAS,QAAQ,CAAC;GACvC,CAAC,IACD,KAAK;GACX,IAAI,kBAAkB,KAAA,GAAW;IAC/B,MAAM,eAAe,MAAM,wBAAwB,aAAa;IAChE,6BAA6B,cAAc,QAAQ,OAAO,QAAQ,UAAU;IAC5E,OAAO,eAAe;GACxB;GACA,OAAO;IAAE,QAAQ,kBAAkB,MAAM;IAAG,WAAW;GAAgB;EACzE,UAAU;GAIR,SAAS,OAAO,KAAK;EACvB;CACF;CAEA,MAAM,UAAU,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,aAAa,MAAM,MAAM,EAAE,SAAS,OAAO,CAAC;CAI1F,MAAM,QAAQ,WAAW,OAAO;CAMhC,KAAK,MAAM,CAAC,OAAO,YAAY,UAC7B,MAAM,cAAc,SAAS,OAAO,kDAAkD;CAExF,SAAS,MAAM;CAEf,IAAI,cAAc,SAAS,GACzB,MAAM,cAAc,WAAW,IAC3B,cAAc,KACd,IAAI,eACF,eACA,oBAAoB,cAAc,OAAO,iDAC3C;CAIN,IAAI;CACJ,IAAI,KAAK,QAAQ;EACf,MAAM,aAAoC;GACxC,GAAG,KAAK;GACR,YAAY,KAAK,OAAO;GACxB,OAAO,aAAa,QAAQ,WAAW;GACvC,aAAa,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;GACzC,qBAAqB,uBAAuB,KAAA;EAC9C;EACA,SAAS,MAAM,eAAe,MAAM,UAAU;CAChD;CAEA,MAAM,UAAU,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,YAAY;CAE5C,OAAO;EACL,YAAY,KAAK;EACjB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAIA,IAAM,qBAAN,cAAiC,MAAM;CACrC;CACA;CACA,YAAY,QAAmB,WAAgC;EAC7D,MAAM,QAAQ,OAAO,UAAU,GAAG,OAAO,WAAW,GAAG,OAAO,KAAK,WAAW,OAAO,QAAQ;EAC7F,KAAK,SAAS;EACd,KAAK,YAAY;CACnB;AACF;;;;;;;;;;;;;;;;AAiBA,eAAe,cACb,SACA,OACA,QACe;CACf,MAAM,WAAW,MAAM,QAAQ,WAAW,OAAO,KAAK;CACtD,IAAI,aAAa,KAAA,GAAW;CAC5B,IAAI,SAAS,WAAW,WAAW;CAGnC,MAAM,gBAAgB,QAAQ,aAAa,CAAC,CAAC;CAC7C,MAAM,QAAQ,SAAS,MAAM;CAC7B,MAAM,QAAQ,QAAQ,aAAa;CACnC,IAAI,MAAM,UAAU,eAClB,MAAM,IAAI,MAAM,sBAAsB,MAAM,2BAA2B,MAAM,WAAW;AAE5F;AAEA,SAAS,sBAAsB,SAA6B;CAC1D,QAAQ,WAAmD;EACzD,IAAI,CAAC,SACH,MAAM,IAAI,MACR,6LACF;EAEF,OAAO,IAAI,0BAA0B,EACnC,KAAK,GAAG,QAAQ,cAAc,OAAO,QACvC,CAAC;CACH;AACF;AAEA,eAAe,wBACb,OAC2B;CAC3B,IAAI,mBAAmB,KAAK,GAAG;EAC7B,IAAI,CAAE,MAAM,uBAAuB,KAAK,GACtC,MAAM,IAAI,MAAM,iEAAiE;EAEnF,OAAO;CACT;CACA,OAAO,sBAAsB,KAAK;AACpC;AAEA,SAAS,mBACP,OAC2B;CAC3B,OAAO,mBAAmB,SAAS,YAAY;AACjD;AAEA,SAAS,6BACP,SACA,OACA,YACM;CACN,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,OACnD,MAAM,IAAI,MACR,wCAAwC,QAAQ,MAAM,kCAAkC,MAAM,EAChG;CAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,YAC7D,MAAM,IAAI,MACR,6CAA6C,QAAQ,WAAW,uCAAuC,WAAW,EACpH;AAEJ;AAEA,SAAS,aAAa,QAAuC;CAG3D,MAAM,OAAO,GAAG,OAAO,WAAW,IAAI,OAAO,UAAU,IAAI,OAAO,WAAW,IAAI,OAAO;CAMxF,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,KAAK,WAAW,CAAC;EAC3B,KAAK,KAAK,KAAK,KAAK,GAAG,QAAU,MAAM;EACvC,KAAK,KAAK,KAAK,KAAK,GAAG,UAAU,MAAM;CACzC;CACA,OAAO,OAAO,GAAG,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG,IAAI,GAAG,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG;AAClF"}