@tangle-network/agent-eval 0.216.1 → 0.218.1

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

@@ -0,0 +1 @@
+{"version":3,"file":"define-agent-eval-BGdFKlrc.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 { MutableSurface, 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\ntype MethodRunArgs<TScenario extends Scenario, TArtifact> = {\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}\n\ntype FinalComparison<TScenario extends Scenario, TArtifact> = Awaited<\n  ReturnType<typeof runFinalComparison<TScenario, TArtifact>>\n>\n\ntype HoldoutReport = ReturnType<typeof pairedCompositeSummary>\n\nconst METHOD_FINAL_PHASE = 'selfImprove.method-final'\n\n/** Complete methods own search and selection; this path only measures their final choice. */\nexport async function runSelfImproveMethod<TScenario extends Scenario, TArtifact>(\n  args: MethodRunArgs<TScenario, TArtifact>,\n): Promise<SelfImproveMethodResult<TScenario, TArtifact>> {\n  const { opts, holdout, costLedger, startedAt } = args\n  const budget = opts.budget ?? {}\n  assertAutoPrTarget(opts)\n  const evidenceContext = captureEvidenceContext(opts, budget)\n  const baselineSurface = structuredClone(opts.baselineSurface)\n  const { selected, cost: methodCost, history } = await searchWithMethod(args, baselineSurface)\n  const totalCost = (): ComparisonCost => methodTotalCost(opts.method.name, methodCost, costLedger)\n  const comparison = await measureMethodWinner(args, baselineSurface, selected, totalCost)\n  const evidence = evidenceContext\n    ? methodEvidenceReceipts(comparison, baselineSurface, selected, evidenceContext)\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 analyzeMethodHoldout(args, comparison, baselineSurface, selected)\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 provenance = recordMethodProvenance({\n    args,\n    comparison,\n    baselineSurface,\n    selected,\n    report,\n    lift,\n    optimization,\n    cost,\n    durationMs,\n    receipts,\n  })\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  await shipMethodSearchLedger(opts, selected)\n  return result\n}\n\n/** An auto-PR needs a target repository before any spend. */\nfunction assertAutoPrTarget(opts: { autoOnPromote?: string; ghOwner?: string; ghRepo?: string }) {\n  if (opts.autoOnPromote === 'pr' && (!opts.ghOwner || !opts.ghRepo)) {\n    throw new Error(\"selfImprove: autoOnPromote='pr' requires ghOwner + ghRepo\")\n  }\n}\n\n/** Detach the caller's evidence context; receipts need a measured holdout. */\nfunction captureEvidenceContext<TScenario extends Scenario, TArtifact>(\n  opts: MethodRunArgs<TScenario, TArtifact>['opts'],\n  budget: NonNullable<SelfImproveOptions<TScenario, TArtifact>['budget']>,\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  return evidenceContext\n}\n\n/** Let the method search and select a winner on the train and selection splits. */\nfunction searchWithMethod<TScenario extends Scenario, TArtifact>(\n  args: MethodRunArgs<TScenario, TArtifact>,\n  baselineSurface: MutableSurface,\n) {\n  const { opts, train, selection, costLedger, storage, runDir } = args\n  const budget = opts.budget ?? {}\n  return 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: [opts.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}\n\n/** Spend on the final holdout comparison, read back from the shared ledger. */\nfunction methodFinalCost(costLedger: CostLedgerHandle): ComparisonCost {\n  return combineComparisonCosts(\n    ['holdout.baseline', 'holdout.winner', 'holdout.neutralized', 'promotion.gate'].map(\n      (phase) => ({\n        label: phase,\n        cost: costFromLedgerSummary(\n          costLedger.summary({ phase: `${METHOD_FINAL_PHASE}.${phase}` }),\n        ),\n      }),\n    ),\n  )\n}\n\n/** The method's own spend plus the final comparison's. */\nfunction methodTotalCost(\n  methodName: string,\n  methodCost: ComparisonCost,\n  costLedger: CostLedgerHandle,\n): ComparisonCost {\n  return combineComparisonCosts([\n    { label: methodName, cost: methodCost },\n    { label: 'final comparison', cost: methodFinalCost(costLedger) },\n  ])\n}\n\n/** Measure baseline and winner on the holdout; the gate holds when spend breaks the budget. */\nfunction measureMethodWinner<TScenario extends Scenario, TArtifact>(\n  args: MethodRunArgs<TScenario, TArtifact>,\n  baselineSurface: MutableSurface,\n  selected: OptimizationMethodResult,\n  totalCost: () => ComparisonCost,\n) {\n  const { opts, holdout, costLedger, storage, runDir } = args\n  const budget = opts.budget ?? {}\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  return 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: [opts.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: METHOD_FINAL_PHASE,\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}\n\n/** Evidence receipts for the baseline and winner holdout campaigns. */\nfunction methodEvidenceReceipts<TScenario extends Scenario, TArtifact>(\n  comparison: FinalComparison<TScenario, TArtifact>,\n  baselineSurface: MutableSurface,\n  selected: OptimizationMethodResult,\n  evidenceContext: NonNullable<SelfImproveOptions<TScenario, TArtifact>['evidence']>,\n) {\n  return {\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}\n\n/** Analyze the measured holdout runs; a winner that is the baseline adds no second candidate. */\nfunction analyzeMethodHoldout<TScenario extends Scenario, TArtifact>(\n  args: MethodRunArgs<TScenario, TArtifact>,\n  comparison: FinalComparison<TScenario, TArtifact>,\n  baselineSurface: MutableSurface,\n  selected: OptimizationMethodResult,\n) {\n  const { opts, holdout, runDir } = args\n  return 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 ? evaluationUnitMap(opts.claim, holdout) : undefined,\n    decisionThreshold: opts.claim?.minimumEffect ?? 0.05,\n    ...(comparison.winnerOnHoldout === comparison.baselineOnHoldout\n      ? {}\n      : { candidateCandidateId: 'winner' }),\n  })\n}\n\n/** Build, persist and return the digest-sealed provenance record. */\nfunction recordMethodProvenance<TScenario extends Scenario, TArtifact>(input: {\n  args: MethodRunArgs<TScenario, TArtifact>\n  comparison: FinalComparison<TScenario, TArtifact>\n  baselineSurface: MutableSurface\n  selected: OptimizationMethodResult\n  report: HoldoutReport | null\n  lift: number | undefined\n  optimization: SelfImproveMethodProvenance['optimizationMethod']\n  cost: ComparisonCost\n  durationMs: number\n  receipts: ReturnType<CostLedgerHandle['list']>\n}): SelfImproveMethodProvenance {\n  const { args, comparison, baselineSurface, selected, report, lift, optimization } = input\n  const { opts, train, selection, holdout, storage, runDir, startedAt } = args\n  const budget = opts.budget ?? {}\n  const baseline = report?.baseline ?? null\n  const winner = report?.winner ?? null\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(input.receipts),\n      ...(report ? { holdoutObservations: report.observations } : {}),\n      ...(comparison.neutralizedOnHoldout\n        ? { neutralizedCampaignDigest: campaignMeasurementDigest(comparison.neutralizedOnHoldout) }\n        : {}),\n    },\n    gate: comparison.gateResult,\n    ...(comparison.holdout === 'deferred' ? { holdout: 'deferred' as const } : {}),\n    ...(baseline && winner\n      ? {\n          baselineHoldoutComposite: baseline.compositeMean,\n          winnerHoldoutComposite: winner.compositeMean,\n          heldOutLift: lift,\n        }\n      : {}),\n    cost: input.cost,\n    totalCostUsd: input.cost.totalCostUsd,\n    totalDurationMs: input.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  return provenance\n}\n\n/** Ship the method's search ledger to the hosted tenant, when one is configured. */\nasync function shipMethodSearchLedger<TScenario extends Scenario, TArtifact>(\n  opts: MethodRunArgs<TScenario, TArtifact>['opts'],\n  selected: OptimizationMethodResult,\n): Promise<void> {\n  if (!opts.hostedTenant) return\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","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;;;ACqBA,MAAM,qBAAqB;;AAG3B,eAAsB,qBACpB,MACwD;CACxD,MAAM,EAAE,MAAM,SAAS,YAAY,cAAc;CACjD,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,mBAAmB,IAAI;CACvB,MAAM,kBAAkB,uBAAuB,MAAM,MAAM;CAC3D,MAAM,kBAAkB,gBAAgB,KAAK,eAAe;CAC5D,MAAM,EAAE,UAAU,MAAM,YAAY,YAAY,MAAM,iBAAiB,MAAM,eAAe;CAC5F,MAAM,kBAAkC,gBAAgB,KAAK,OAAO,MAAM,YAAY,UAAU;CAChG,MAAM,aAAa,MAAM,oBAAoB,MAAM,iBAAiB,UAAU,SAAS;CACvF,MAAM,WAAW,kBACb,uBAAuB,YAAY,iBAAiB,UAAU,eAAe,IAC7E,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,qBAAqB,MAAM,YAAY,iBAAiB,QAAQ;CAC1E,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,aAAa,uBAAuB;EACxC;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF,CAAC;CACD,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,MAAM,uBAAuB,MAAM,QAAQ;CAC3C,OAAO;AACT;;AAGA,SAAS,mBAAmB,MAAqE;CAC/F,IAAI,KAAK,kBAAkB,SAAS,CAAC,KAAK,WAAW,CAAC,KAAK,SACzD,MAAM,IAAI,MAAM,2DAA2D;AAE/E;;AAGA,SAAS,uBACP,MACA,QACA;CACA,MAAM,kBAAkB,KAAK,aAAa,KAAA,IAAY,KAAA,IAAY,gBAAgB,KAAK,QAAQ;CAC/F,IAAI,mBAAmB,OAAO,YAAY,YACxC,MAAM,IAAI,MAAM,yDAAyD;CAC3E,OAAO;AACT;;AAGA,SAAS,iBACP,MACA,iBACA;CACA,MAAM,EAAE,MAAM,OAAO,WAAW,YAAY,SAAS,WAAW;CAChE,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,OAAO,0BAA0B;EAC/B,QAAQ,KAAK;EACb;EACA,qBAAqB,KAAK;EAC1B,2BAA2B,KAAK;EAChC,OAAO;GACL;GACA,gBAAgB;GAChB,oBAAoB;GACpB,qBAAqB,KAAK;GAC1B,QAAQ,CAAC,KAAK,KAAK;GACnB,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;AACH;;AAGA,SAAS,gBAAgB,YAA8C;CACrE,OAAO,uBACL;EAAC;EAAoB;EAAkB;EAAuB;CAAgB,CAAC,CAAC,KAC7E,WAAW;EACV,OAAO;EACP,MAAM,sBACJ,WAAW,QAAQ,EAAE,OAAO,GAAG,mBAAmB,GAAG,QAAQ,CAAC,CAChE;CACF,EACF,CACF;AACF;;AAGA,SAAS,gBACP,YACA,YACA,YACgB;CAChB,OAAO,uBAAuB,CAC5B;EAAE,OAAO;EAAY,MAAM;CAAW,GACtC;EAAE,OAAO;EAAoB,MAAM,gBAAgB,UAAU;CAAE,CACjE,CAAC;AACH;;AAGA,SAAS,oBACP,MACA,iBACA,UACA,WACA;CACA,MAAM,EAAE,MAAM,SAAS,YAAY,SAAS,WAAW;CACvD,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,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,OAAO,mBAAmB;EACxB,OAAO,KAAK;EACZ,eAAe,KAAK;EACpB;EACA,eAAe,SAAS;EACxB,WAAW;EACX,qBAAqB,KAAK;EAC1B,aAAa,KAAK;EAClB,QAAQ,CAAC,KAAK,KAAK;EACnB,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;AACH;;AAGA,SAAS,uBACP,YACA,iBACA,UACA,iBACA;CACA,OAAO;EACL,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;AACF;;AAGA,SAAS,qBACP,MACA,YACA,iBACA,UACA;CACA,MAAM,EAAE,MAAM,SAAS,WAAW;CAClC,OAAO,YAAY;EACjB,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,QAAQ,kBAAkB,KAAK,OAAO,OAAO,IAAI,KAAA;EACnF,mBAAmB,KAAK,OAAO,iBAAiB;EAChD,GAAI,WAAW,oBAAoB,WAAW,oBAC1C,CAAC,IACD,EAAE,sBAAsB,SAAS;CACvC,CAAC;AACH;;AAGA,SAAS,uBAA8D,OAWvC;CAC9B,MAAM,EAAE,MAAM,YAAY,iBAAiB,UAAU,QAAQ,MAAM,iBAAiB;CACpF,MAAM,EAAE,MAAM,OAAO,WAAW,SAAS,SAAS,QAAQ,cAAc;CACxE,MAAM,SAAS,KAAK,UAAU,CAAC;CAC/B,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,SAAS,QAAQ,UAAU;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,MAAM,QAAQ;GAClD,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,YAAY,aAAa,EAAE,SAAS,WAAoB,IAAI,CAAC;EAC5E,GAAI,YAAY,SACZ;GACE,0BAA0B,SAAS;GACnC,wBAAwB,OAAO;GAC/B,aAAa;EACf,IACA,CAAC;EACL,MAAM,MAAM;EACZ,cAAc,MAAM,KAAK;EACzB,iBAAiB,MAAM;CACzB;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,OAAO;AACT;;AAGA,eAAe,uBACb,MACA,UACe;CACf,IAAI,CAAC,KAAK,cAAc;CACxB,MAAM,UAAU,SAAS;CACzB,IAAI,SAAS;EACX,MAAM,OAAO,cAAc,QAAQ,OAAO,GAAG;EAC7C,MAAM,qBAEF,iBAAiB;GACf,QAAQ,KAAK;GACb,QAAQ;IAAE;IAAM,UAAU,QAAQ,QAAQ;GAAS;GACnD,SAAS;EACX,CAAC,GACH,IACF;CACF,OACE,QAAQ,KACN,qCAAqC,KAAK,OAAO,KAAK,mDACxD;AAEJ;;;;AC9FA,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/{define-agent-eval-WSck0CZJ.d.ts → define-agent-eval-BlH10zEw.d.ts}

@@ -1,12 +1,12 @@
 import { a as CostLedgerHandle, c as CostLedgerSummary, u as CostReceipt } from "./cost-ledger-Jk62OBfF.js";
-import { b as ProposalFinding } from "./types-C1tnkEgM.js";
-import { B as PowerPreflight, dt as EvidenceReceipt, it as CampaignEvidenceContext } from "./statistical-heldout-CYpGaoDH.js";
+import { v as ProposalFinding } from "./types-CW8ewATZ.js";
+import { R as PowerPreflight, ct as EvidenceReceipt, nt as CampaignEvidenceContext } from "./statistical-heldout-BHfLZHrm.js";
 import { nt as EvaluationClaim } from "./search-ledger-types-p2xKfiDJ.js";
-import { Ci as ComparisonCost, Ea as FinalEvidenceUse, Ta as FinalEvidencePolicy, bi as OptimizationMethodResult, ca as openAutoPr, ea as RunImprovementLoopResult, fa as CampaignCellRetryPolicy, gi as OptimizationMethod, la as RunEvalOptions, na as PremeasuredOptimizationBaseline, pa as RunCampaignOptions, ra as RunOptimizationOptions, ur as LoopProvenanceRecord, xr as CampaignComparisonUnits, yi as OptimizationMethodProvenance } from "./index-CyszSg4O.js";
-import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-Cq8mSbq8.js";
-import { Mt as SearchHistoryCoverageRow, Pt as SearchHistoryReceipt, Zt as CampaignStorage, kt as SearchHistoryAdmissionOptions } from "./search-kernel-DyNEa6IU.js";
-import { r as HostedTenant } from "./client-BzL9niqi.js";
-import { d as InsightReport } from "./analyze-runs-D68fd-su.js";
+import { $n as CampaignComparisonUnits, Di as PremeasuredOptimizationBaseline, Jr as OptimizationMethod, Li as CampaignCellRetryPolicy, Ni as openAutoPr, Oi as RunOptimizationOptions, Pi as RunEvalOptions, Qr as OptimizationMethodResult, Ri as RunCampaignOptions, Ti as RunImprovementLoopResult, Xi as FinalEvidenceUse, Yi as FinalEvidencePolicy, Yn as LoopProvenanceRecord, Zr as OptimizationMethodProvenance, ti as ComparisonCost } from "./index-B16FNdW5.js";
+import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-DtkO1G9L.js";
+import { At as SearchHistoryReceipt, Ot as SearchHistoryCoverageRow, Tt as SearchHistoryAdmissionOptions, Wt as CampaignStorage } from "./search-kernel-p9JB3SmL.js";
+import { n as HostedTenant } from "./client-hwo3cJjR.js";
+import { d as InsightReport } from "./analyze-runs-DjlZOVXx.js";
 //#region src/campaign/presets/run-final-comparison.d.ts
 interface FinalComparisonOptions<TScenario extends Scenario, TArtifact> extends Omit<RunCampaignOptions<TScenario, TArtifact>, 'dispatch'> {
   baselineSurface: MutableSurface;
@@ -512,4 +512,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
 declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
 //#endregion
 export { SelfImproveMethodResult as _, DefinedAgentEval as a, SelfImproveMethodOptions as c, SelfImproveProposerOptions as d, SelfImproveProposerResult as f, SelfImproveMethodProvenance as g, selfImprove as h, DefineAgentEvalOptions as i, SelfImproveOptions as l, SelfImproveRunError as m, AgentEvalEvaluateOptions as n, defineAgentEval as o, SelfImproveResult as p, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveProgressEvent as u };
-//# sourceMappingURL=define-agent-eval-WSck0CZJ.d.ts.map
+//# sourceMappingURL=define-agent-eval-BlH10zEw.d.ts.map

package/dist/define-agent-eval-BlH10zEw.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"define-agent-eval-BlH10zEw.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaU,uBAAuB,kBAAkB,UAAU,mBACnD,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}

package/dist/{descriptive-DboxPZrG.js → descriptive-CVNy48rR.js}

@@ -1,4 +1,4 @@
-import { i as makeRng } from "./internal-Epl-YBZW.js";
+import { i as makeRng } from "./internal-BTK91Bqt.js";
 //#region src/statistics/descriptive.ts
 /**
 * Descriptive statistics: means, bootstrap spread, correlation, and the
@@ -161,4 +161,4 @@ function weightedComposite(input) {
 //#endregion
 export { spearmanR as a, weightedMean as c, ranks as i, partialCredit as n, summarizeNumberSeries as o, pearsonR as r, weightedComposite as s, confidenceInterval as t };
 
-//# sourceMappingURL=descriptive-DboxPZrG.js.map
+//# sourceMappingURL=descriptive-CVNy48rR.js.map

package/dist/{descriptive-DboxPZrG.js.map → descriptive-CVNy48rR.js.map}

@@ -1 +1 @@
-{"version":3,"file":"descriptive-DboxPZrG.js","names":[],"sources":["../src/statistics/descriptive.ts"],"sourcesContent":["/**\n * Descriptive statistics: means, bootstrap spread, correlation, and the\n * weighted judge-dimension composite. Nothing here is a significance test.\n */\n\nimport { makeRng } from './internal'\n\n/** Weighted mean — falls back to uniform weights when omitted */\nexport function weightedMean(scores: { score: number; weight?: number }[]): number {\n  if (scores.length === 0) return 0\n  let totalWeight = 0\n  let weightedSum = 0\n  for (const { score, weight } of scores) {\n    const w = weight ?? 1\n    weightedSum += score * w\n    totalWeight += w\n  }\n  return totalWeight > 0 ? weightedSum / totalWeight : 0\n}\n\n/**\n * Percentile bootstrap confidence interval on the mean of `scores`.\n *\n * Descriptive spread. It is not a significance test, and at small n its bounds\n * are anti-conservative in the same way {@link pairedBootstrap}'s are — see\n * {@link BOOTSTRAP_GATE_MIN_N}. With no `seed` the resampling is seeded from\n * the scores themselves, so the interval is reproducible either way.\n */\nexport function confidenceInterval(\n  scores: number[],\n  confidence = 0.95,\n  opts: { seed?: number; resamples?: number } = {},\n): { mean: number; lower: number; upper: number } {\n  if (scores.length === 0) return { mean: 0, lower: 0, upper: 0 }\n  if (scores.length === 1) return { mean: scores[0]!, lower: scores[0]!, upper: scores[0]! }\n\n  const n = scores.length\n  const mean = scores.reduce((a, b) => a + b, 0) / n\n\n  const B = opts.resamples ?? 1000\n  const rng = makeRng(opts.seed, scores)\n  const bootstrapMeans: number[] = []\n\n  for (let i = 0; i < B; i++) {\n    let sum = 0\n    for (let j = 0; j < n; j++) {\n      sum += scores[Math.floor(rng() * n)]!\n    }\n    bootstrapMeans.push(sum / n)\n  }\n\n  bootstrapMeans.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowerIdx = Math.floor((alpha / 2) * B)\n  const upperIdx = Math.floor((1 - alpha / 2) * B) - 1\n\n  return {\n    mean,\n    lower: bootstrapMeans[lowerIdx]!,\n    upper: bootstrapMeans[Math.min(upperIdx, B - 1)]!,\n  }\n}\n\n/** Partial credit: returns 0-1 ratio of current toward target */\nexport function partialCredit(current: number, target: number): number {\n  if (target <= 0) return 1\n  return Math.min(1, Math.max(0, current / target))\n}\n\n/** Distribution summary of a number series: count, extremes, quantiles, sum. */\nexport interface SeriesDistribution {\n  readonly n: number\n  readonly min: number\n  readonly p50: number\n  readonly p90: number\n  readonly max: number\n  readonly sum: number\n}\n\n/**\n * Fold a number series into its distribution summary. Quantiles use the\n * nearest-rank definition — the `ceil(q·n)`-th order statistic — so every\n * reported quantile is a value from the series. Returns `null` for an empty\n * series: an empty series has no distribution, and a zero-filled summary\n * would read as a measured all-zero series.\n */\nexport function summarizeNumberSeries(values: readonly number[]): SeriesDistribution | null {\n  if (values.length === 0) return null\n  const sorted = [...values].sort((a, b) => a - b)\n  const rank = (q: number): number =>\n    sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(q * sorted.length) - 1))] as number\n  return {\n    n: sorted.length,\n    min: sorted[0] as number,\n    p50: rank(0.5),\n    p90: rank(0.9),\n    max: sorted[sorted.length - 1] as number,\n    sum: sorted.reduce((a, b) => a + b, 0),\n  }\n}\n\n// ── Correlation (Pearson / Spearman) ─────────────────────────────────\n//\n// The single source for linear (Pearson) and rank (Spearman) correlation.\n// Edge-case contract is explicit so every caller agrees on what a\n// degenerate input means:\n//   - length mismatch or n < 2  → NaN   (correlation is undefined; not 0)\n//   - both series constant      → 1     (degenerate perfect agreement)\n//   - exactly one series constant → 0   (no covariation to detect)\n// Returning NaN for n < 2 (rather than 0) keeps \"not enough data\" distinct\n// from \"measured zero correlation\", which a 0 would silently conflate.\n\n/**\n * Average-rank-with-ties transform (1-indexed). Tied values receive the mean\n * of the ranks they span, the standard correction for Spearman's ρ.\n */\nexport function ranks(xs: number[]): number[] {\n  const indexed = xs.map((v, i) => ({ v, i })).sort((a, b) => a.v - b.v)\n  const r = new Array<number>(xs.length)\n  let i = 0\n  while (i < indexed.length) {\n    let j = i\n    while (j + 1 < indexed.length && indexed[j + 1]!.v === indexed[i]!.v) j++\n    const avg = (i + j) / 2 + 1\n    for (let k = i; k <= j; k++) r[indexed[k]!.i] = avg\n    i = j + 1\n  }\n  return r\n}\n\n/**\n * Pearson product-moment correlation coefficient r ∈ [-1, 1] between two\n * equal-length series. See the edge-case contract above: NaN for n < 2 or\n * unequal lengths, 1 when both series are constant, 0 when exactly one is.\n */\nexport function pearsonR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return Number.NaN\n  const n = a.length\n  const meanA = a.reduce((s, v) => s + v, 0) / n\n  const meanB = b.reduce((s, v) => s + v, 0) / n\n  let num = 0\n  let varA = 0\n  let varB = 0\n  for (let i = 0; i < n; i++) {\n    const da = a[i]! - meanA\n    const db = b[i]! - meanB\n    num += da * db\n    varA += da * da\n    varB += db * db\n  }\n  if (varA === 0 || varB === 0) return varA === 0 && varB === 0 ? 1 : 0\n  return num / Math.sqrt(varA * varB)\n}\n\n/**\n * Spearman's rank correlation ρ — Pearson over the average-rank-with-ties\n * transform of each series. Same edge-case contract as {@link pearsonR}.\n */\nexport function spearmanR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return Number.NaN\n  return pearsonR(ranks(a), ranks(b))\n}\n\nexport interface WeightedCompositeInput {\n  /** Per-dimension scores (typically 0..1). */\n  dims: Record<string, number>\n  /** Weight per dimension. Every weighted dimension MUST be present in\n   *  `dims` — a weight for an absent dimension is a config error and throws,\n   *  because silently dropping it would renormalise the composite onto a\n   *  different denominator than intended. */\n  weights: Record<string, number>\n  /** Optional pass threshold; when set, the result reports `pass`. */\n  threshold?: number\n}\n\nexport interface WeightedCompositeResult {\n  composite: number\n  pass?: boolean\n}\n\n/**\n * Weighted composite over judge dimensions: `Σ(score_d · w_d) / Σ(w_d)` across\n * the weighted dimensions. The canonical replacement for the per-consumer\n * hand-rolled composite math (tax/legal/creative/gtm each ship a copy).\n *\n * Fail-loud: throws if a weighted dimension is missing from `dims`, if any\n * weight is negative, or if the weights sum to 0 — none of which can produce\n * a meaningful composite.\n */\nexport function weightedComposite(input: WeightedCompositeInput): WeightedCompositeResult {\n  const entries = Object.entries(input.weights)\n  if (entries.length === 0) {\n    throw new Error('weightedComposite: `weights` is empty — nothing to combine')\n  }\n  let weightedSum = 0\n  let weightTotal = 0\n  for (const [dim, weight] of entries) {\n    if (weight < 0) {\n      throw new Error(`weightedComposite: weight for '${dim}' is negative (${weight})`)\n    }\n    if (!(dim in input.dims)) {\n      throw new Error(\n        `weightedComposite: weighted dimension '${dim}' is absent from \\`dims\\` — ` +\n          'refusing to renormalise onto a different denominator',\n      )\n    }\n    weightedSum += input.dims[dim]! * weight\n    weightTotal += weight\n  }\n  if (weightTotal === 0) {\n    throw new Error('weightedComposite: weights sum to 0 — composite is undefined')\n  }\n  const composite = weightedSum / weightTotal\n  return input.threshold === undefined\n    ? { composite }\n    : { composite, pass: composite >= input.threshold }\n}\n"],"mappings":";;;;;;;AAQA,SAAgB,aAAa,QAAsD;CACjF,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,cAAc;CAClB,IAAI,cAAc;CAClB,KAAK,MAAM,EAAE,OAAO,YAAY,QAAQ;EACtC,MAAM,IAAI,UAAU;EACpB,eAAe,QAAQ;EACvB,eAAe;CACjB;CACA,OAAO,cAAc,IAAI,cAAc,cAAc;AACvD;;;;;;;;;AAUA,SAAgB,mBACd,QACA,aAAa,KACb,OAA8C,CAAC,GACC;CAChD,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,MAAM;EAAG,OAAO;EAAG,OAAO;CAAE;CAC9D,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,MAAM,OAAO;EAAK,OAAO,OAAO;EAAK,OAAO,OAAO;CAAI;CAEzF,MAAM,IAAI,OAAO;CACjB,MAAM,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAEjD,MAAM,IAAI,KAAK,aAAa;CAC5B,MAAM,MAAM,QAAQ,KAAK,MAAM,MAAM;CACrC,MAAM,iBAA2B,CAAC;CAElC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,OAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEpC,eAAe,KAAK,MAAM,CAAC;CAC7B;CAEA,eAAe,MAAM,GAAG,MAAM,IAAI,CAAC;CAEnC,MAAM,QAAQ,IAAI;CAClB,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,CAAC;CAC3C,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,CAAC,IAAI;CAEnD,OAAO;EACL;EACA,OAAO,eAAe;EACtB,OAAO,eAAe,KAAK,IAAI,UAAU,IAAI,CAAC;CAChD;AACF;;AAGA,SAAgB,cAAc,SAAiB,QAAwB;CACrE,IAAI,UAAU,GAAG,OAAO;CACxB,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC;AAClD;;;;;;;;AAmBA,SAAgB,sBAAsB,QAAsD;CAC1F,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,SAAS,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC/C,MAAM,QAAQ,MACZ,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC;CAClF,OAAO;EACL,GAAG,OAAO;EACV,KAAK,OAAO;EACZ,KAAK,KAAK,EAAG;EACb,KAAK,KAAK,EAAG;EACb,KAAK,OAAO,OAAO,SAAS;EAC5B,KAAK,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC;CACvC;AACF;;;;;AAiBA,SAAgB,MAAM,IAAwB;CAC5C,MAAM,UAAU,GAAG,KAAK,GAAG,OAAO;EAAE;EAAG;CAAE,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;CACrE,MAAM,IAAI,IAAI,MAAc,GAAG,MAAM;CACrC,IAAI,IAAI;CACR,OAAO,IAAI,QAAQ,QAAQ;EACzB,IAAI,IAAI;EACR,OAAO,IAAI,IAAI,QAAQ,UAAU,QAAQ,IAAI,EAAE,CAAE,MAAM,QAAQ,EAAE,CAAE,GAAG;EACtE,MAAM,OAAO,IAAI,KAAK,IAAI;EAC1B,KAAK,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,EAAE,QAAQ,EAAE,CAAE,KAAK;EAChD,IAAI,IAAI;CACV;CACA,OAAO;AACT;;;;;;AAOA,SAAgB,SAAS,GAAa,GAAqB;CACzD,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,GAAG,OAAO;CAClD,MAAM,IAAI,EAAE;CACZ,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAC7C,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAC7C,IAAI,MAAM;CACV,IAAI,OAAO;CACX,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,KAAK,EAAE,KAAM;EACnB,OAAO,KAAK;EACZ,QAAQ,KAAK;EACb,QAAQ,KAAK;CACf;CACA,IAAI,SAAS,KAAK,SAAS,GAAG,OAAO,SAAS,KAAK,SAAS,IAAI,IAAI;CACpE,OAAO,MAAM,KAAK,KAAK,OAAO,IAAI;AACpC;;;;;AAMA,SAAgB,UAAU,GAAa,GAAqB;CAC1D,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,GAAG,OAAO;CAClD,OAAO,SAAS,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC;AACpC;;;;;;;;;;AA4BA,SAAgB,kBAAkB,OAAwD;CACxF,MAAM,UAAU,OAAO,QAAQ,MAAM,OAAO;CAC5C,IAAI,QAAQ,WAAW,GACrB,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IAAI,cAAc;CAClB,IAAI,cAAc;CAClB,KAAK,MAAM,CAAC,KAAK,WAAW,SAAS;EACnC,IAAI,SAAS,GACX,MAAM,IAAI,MAAM,kCAAkC,IAAI,iBAAiB,OAAO,EAAE;EAElF,IAAI,EAAE,OAAO,MAAM,OACjB,MAAM,IAAI,MACR,0CAA0C,IAAI,iFAEhD;EAEF,eAAe,MAAM,KAAK,OAAQ;EAClC,eAAe;CACjB;CACA,IAAI,gBAAgB,GAClB,MAAM,IAAI,MAAM,8DAA8D;CAEhF,MAAM,YAAY,cAAc;CAChC,OAAO,MAAM,cAAc,KAAA,IACvB,EAAE,UAAU,IACZ;EAAE;EAAW,MAAM,aAAa,MAAM;CAAU;AACtD"}
+{"version":3,"file":"descriptive-CVNy48rR.js","names":[],"sources":["../src/statistics/descriptive.ts"],"sourcesContent":["/**\n * Descriptive statistics: means, bootstrap spread, correlation, and the\n * weighted judge-dimension composite. Nothing here is a significance test.\n */\n\nimport { makeRng } from './internal'\n\n/** Weighted mean — falls back to uniform weights when omitted */\nexport function weightedMean(scores: { score: number; weight?: number }[]): number {\n  if (scores.length === 0) return 0\n  let totalWeight = 0\n  let weightedSum = 0\n  for (const { score, weight } of scores) {\n    const w = weight ?? 1\n    weightedSum += score * w\n    totalWeight += w\n  }\n  return totalWeight > 0 ? weightedSum / totalWeight : 0\n}\n\n/**\n * Percentile bootstrap confidence interval on the mean of `scores`.\n *\n * Descriptive spread. It is not a significance test, and at small n its bounds\n * are anti-conservative in the same way {@link pairedBootstrap}'s are — see\n * {@link BOOTSTRAP_GATE_MIN_N}. With no `seed` the resampling is seeded from\n * the scores themselves, so the interval is reproducible either way.\n */\nexport function confidenceInterval(\n  scores: number[],\n  confidence = 0.95,\n  opts: { seed?: number; resamples?: number } = {},\n): { mean: number; lower: number; upper: number } {\n  if (scores.length === 0) return { mean: 0, lower: 0, upper: 0 }\n  if (scores.length === 1) return { mean: scores[0]!, lower: scores[0]!, upper: scores[0]! }\n\n  const n = scores.length\n  const mean = scores.reduce((a, b) => a + b, 0) / n\n\n  const B = opts.resamples ?? 1000\n  const rng = makeRng(opts.seed, scores)\n  const bootstrapMeans: number[] = []\n\n  for (let i = 0; i < B; i++) {\n    let sum = 0\n    for (let j = 0; j < n; j++) {\n      sum += scores[Math.floor(rng() * n)]!\n    }\n    bootstrapMeans.push(sum / n)\n  }\n\n  bootstrapMeans.sort((a, b) => a - b)\n\n  const alpha = 1 - confidence\n  const lowerIdx = Math.floor((alpha / 2) * B)\n  const upperIdx = Math.floor((1 - alpha / 2) * B) - 1\n\n  return {\n    mean,\n    lower: bootstrapMeans[lowerIdx]!,\n    upper: bootstrapMeans[Math.min(upperIdx, B - 1)]!,\n  }\n}\n\n/** Partial credit: returns 0-1 ratio of current toward target */\nexport function partialCredit(current: number, target: number): number {\n  if (target <= 0) return 1\n  return Math.min(1, Math.max(0, current / target))\n}\n\n/** Distribution summary of a number series: count, extremes, quantiles, sum. */\nexport interface SeriesDistribution {\n  readonly n: number\n  readonly min: number\n  readonly p50: number\n  readonly p90: number\n  readonly max: number\n  readonly sum: number\n}\n\n/**\n * Fold a number series into its distribution summary. Quantiles use the\n * nearest-rank definition — the `ceil(q·n)`-th order statistic — so every\n * reported quantile is a value from the series. Returns `null` for an empty\n * series: an empty series has no distribution, and a zero-filled summary\n * would read as a measured all-zero series.\n */\nexport function summarizeNumberSeries(values: readonly number[]): SeriesDistribution | null {\n  if (values.length === 0) return null\n  const sorted = [...values].sort((a, b) => a - b)\n  const rank = (q: number): number =>\n    sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(q * sorted.length) - 1))] as number\n  return {\n    n: sorted.length,\n    min: sorted[0] as number,\n    p50: rank(0.5),\n    p90: rank(0.9),\n    max: sorted[sorted.length - 1] as number,\n    sum: sorted.reduce((a, b) => a + b, 0),\n  }\n}\n\n// ── Correlation (Pearson / Spearman) ─────────────────────────────────\n//\n// The single source for linear (Pearson) and rank (Spearman) correlation.\n// Edge-case contract is explicit so every caller agrees on what a\n// degenerate input means:\n//   - length mismatch or n < 2  → NaN   (correlation is undefined; not 0)\n//   - both series constant      → 1     (degenerate perfect agreement)\n//   - exactly one series constant → 0   (no covariation to detect)\n// Returning NaN for n < 2 (rather than 0) keeps \"not enough data\" distinct\n// from \"measured zero correlation\", which a 0 would silently conflate.\n\n/**\n * Average-rank-with-ties transform (1-indexed). Tied values receive the mean\n * of the ranks they span, the standard correction for Spearman's ρ.\n */\nexport function ranks(xs: number[]): number[] {\n  const indexed = xs.map((v, i) => ({ v, i })).sort((a, b) => a.v - b.v)\n  const r = new Array<number>(xs.length)\n  let i = 0\n  while (i < indexed.length) {\n    let j = i\n    while (j + 1 < indexed.length && indexed[j + 1]!.v === indexed[i]!.v) j++\n    const avg = (i + j) / 2 + 1\n    for (let k = i; k <= j; k++) r[indexed[k]!.i] = avg\n    i = j + 1\n  }\n  return r\n}\n\n/**\n * Pearson product-moment correlation coefficient r ∈ [-1, 1] between two\n * equal-length series. See the edge-case contract above: NaN for n < 2 or\n * unequal lengths, 1 when both series are constant, 0 when exactly one is.\n */\nexport function pearsonR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return Number.NaN\n  const n = a.length\n  const meanA = a.reduce((s, v) => s + v, 0) / n\n  const meanB = b.reduce((s, v) => s + v, 0) / n\n  let num = 0\n  let varA = 0\n  let varB = 0\n  for (let i = 0; i < n; i++) {\n    const da = a[i]! - meanA\n    const db = b[i]! - meanB\n    num += da * db\n    varA += da * da\n    varB += db * db\n  }\n  if (varA === 0 || varB === 0) return varA === 0 && varB === 0 ? 1 : 0\n  return num / Math.sqrt(varA * varB)\n}\n\n/**\n * Spearman's rank correlation ρ — Pearson over the average-rank-with-ties\n * transform of each series. Same edge-case contract as {@link pearsonR}.\n */\nexport function spearmanR(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length < 2) return Number.NaN\n  return pearsonR(ranks(a), ranks(b))\n}\n\nexport interface WeightedCompositeInput {\n  /** Per-dimension scores (typically 0..1). */\n  dims: Record<string, number>\n  /** Weight per dimension. Every weighted dimension MUST be present in\n   *  `dims` — a weight for an absent dimension is a config error and throws,\n   *  because silently dropping it would renormalise the composite onto a\n   *  different denominator than intended. */\n  weights: Record<string, number>\n  /** Optional pass threshold; when set, the result reports `pass`. */\n  threshold?: number\n}\n\nexport interface WeightedCompositeResult {\n  composite: number\n  pass?: boolean\n}\n\n/**\n * Weighted composite over judge dimensions: `Σ(score_d · w_d) / Σ(w_d)` across\n * the weighted dimensions. The canonical replacement for the per-consumer\n * hand-rolled composite math (tax/legal/creative/gtm each ship a copy).\n *\n * Fail-loud: throws if a weighted dimension is missing from `dims`, if any\n * weight is negative, or if the weights sum to 0 — none of which can produce\n * a meaningful composite.\n */\nexport function weightedComposite(input: WeightedCompositeInput): WeightedCompositeResult {\n  const entries = Object.entries(input.weights)\n  if (entries.length === 0) {\n    throw new Error('weightedComposite: `weights` is empty — nothing to combine')\n  }\n  let weightedSum = 0\n  let weightTotal = 0\n  for (const [dim, weight] of entries) {\n    if (weight < 0) {\n      throw new Error(`weightedComposite: weight for '${dim}' is negative (${weight})`)\n    }\n    if (!(dim in input.dims)) {\n      throw new Error(\n        `weightedComposite: weighted dimension '${dim}' is absent from \\`dims\\` — ` +\n          'refusing to renormalise onto a different denominator',\n      )\n    }\n    weightedSum += input.dims[dim]! * weight\n    weightTotal += weight\n  }\n  if (weightTotal === 0) {\n    throw new Error('weightedComposite: weights sum to 0 — composite is undefined')\n  }\n  const composite = weightedSum / weightTotal\n  return input.threshold === undefined\n    ? { composite }\n    : { composite, pass: composite >= input.threshold }\n}\n"],"mappings":";;;;;;;AAQA,SAAgB,aAAa,QAAsD;CACjF,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,cAAc;CAClB,IAAI,cAAc;CAClB,KAAK,MAAM,EAAE,OAAO,YAAY,QAAQ;EACtC,MAAM,IAAI,UAAU;EACpB,eAAe,QAAQ;EACvB,eAAe;CACjB;CACA,OAAO,cAAc,IAAI,cAAc,cAAc;AACvD;;;;;;;;;AAUA,SAAgB,mBACd,QACA,aAAa,KACb,OAA8C,CAAC,GACC;CAChD,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,MAAM;EAAG,OAAO;EAAG,OAAO;CAAE;CAC9D,IAAI,OAAO,WAAW,GAAG,OAAO;EAAE,MAAM,OAAO;EAAK,OAAO,OAAO;EAAK,OAAO,OAAO;CAAI;CAEzF,MAAM,IAAI,OAAO;CACjB,MAAM,OAAO,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAEjD,MAAM,IAAI,KAAK,aAAa;CAC5B,MAAM,MAAM,QAAQ,KAAK,MAAM,MAAM;CACrC,MAAM,iBAA2B,CAAC;CAElC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,IAAI,MAAM;EACV,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KACrB,OAAO,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC;EAEpC,eAAe,KAAK,MAAM,CAAC;CAC7B;CAEA,eAAe,MAAM,GAAG,MAAM,IAAI,CAAC;CAEnC,MAAM,QAAQ,IAAI;CAClB,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,CAAC;CAC3C,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,CAAC,IAAI;CAEnD,OAAO;EACL;EACA,OAAO,eAAe;EACtB,OAAO,eAAe,KAAK,IAAI,UAAU,IAAI,CAAC;CAChD;AACF;;AAGA,SAAgB,cAAc,SAAiB,QAAwB;CACrE,IAAI,UAAU,GAAG,OAAO;CACxB,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC;AAClD;;;;;;;;AAmBA,SAAgB,sBAAsB,QAAsD;CAC1F,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,SAAS,CAAC,GAAG,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC/C,MAAM,QAAQ,MACZ,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC;CAClF,OAAO;EACL,GAAG,OAAO;EACV,KAAK,OAAO;EACZ,KAAK,KAAK,EAAG;EACb,KAAK,KAAK,EAAG;EACb,KAAK,OAAO,OAAO,SAAS;EAC5B,KAAK,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC;CACvC;AACF;;;;;AAiBA,SAAgB,MAAM,IAAwB;CAC5C,MAAM,UAAU,GAAG,KAAK,GAAG,OAAO;EAAE;EAAG;CAAE,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;CACrE,MAAM,IAAI,IAAI,MAAc,GAAG,MAAM;CACrC,IAAI,IAAI;CACR,OAAO,IAAI,QAAQ,QAAQ;EACzB,IAAI,IAAI;EACR,OAAO,IAAI,IAAI,QAAQ,UAAU,QAAQ,IAAI,EAAE,CAAE,MAAM,QAAQ,EAAE,CAAE,GAAG;EACtE,MAAM,OAAO,IAAI,KAAK,IAAI;EAC1B,KAAK,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,EAAE,QAAQ,EAAE,CAAE,KAAK;EAChD,IAAI,IAAI;CACV;CACA,OAAO;AACT;;;;;;AAOA,SAAgB,SAAS,GAAa,GAAqB;CACzD,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,GAAG,OAAO;CAClD,MAAM,IAAI,EAAE;CACZ,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAC7C,MAAM,QAAQ,EAAE,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAC7C,IAAI,MAAM;CACV,IAAI,OAAO;CACX,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,KAAK,EAAE,KAAM;EACnB,MAAM,KAAK,EAAE,KAAM;EACnB,OAAO,KAAK;EACZ,QAAQ,KAAK;EACb,QAAQ,KAAK;CACf;CACA,IAAI,SAAS,KAAK,SAAS,GAAG,OAAO,SAAS,KAAK,SAAS,IAAI,IAAI;CACpE,OAAO,MAAM,KAAK,KAAK,OAAO,IAAI;AACpC;;;;;AAMA,SAAgB,UAAU,GAAa,GAAqB;CAC1D,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,SAAS,GAAG,OAAO;CAClD,OAAO,SAAS,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC;AACpC;;;;;;;;;;AA4BA,SAAgB,kBAAkB,OAAwD;CACxF,MAAM,UAAU,OAAO,QAAQ,MAAM,OAAO;CAC5C,IAAI,QAAQ,WAAW,GACrB,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IAAI,cAAc;CAClB,IAAI,cAAc;CAClB,KAAK,MAAM,CAAC,KAAK,WAAW,SAAS;EACnC,IAAI,SAAS,GACX,MAAM,IAAI,MAAM,kCAAkC,IAAI,iBAAiB,OAAO,EAAE;EAElF,IAAI,EAAE,OAAO,MAAM,OACjB,MAAM,IAAI,MACR,0CAA0C,IAAI,iFAEhD;EAEF,eAAe,MAAM,KAAK,OAAQ;EAClC,eAAe;CACjB;CACA,IAAI,gBAAgB,GAClB,MAAM,IAAI,MAAM,8DAA8D;CAEhF,MAAM,YAAY,cAAc;CAChC,OAAO,MAAM,cAAc,KAAA,IACvB,EAAE,UAAU,IACZ;EAAE;EAAW,MAAM,aAAa,MAAM;CAAU;AACtD"}

package/dist/diagnosis/index.d.ts

@@ -1,8 +1,8 @@
-import { o as FailureBlame, s as FailureClassification } from "../failure-cluster-CbgEDcFJ.js";
-import { r as ExecutionReport } from "../analyze-runs-D68fd-su.js";
-import { l as PrimeRawUsage } from "../prime-protocol-HiB9HX7E.js";
+import { a as FailureClassification, i as FailureBlame } from "../failure-cluster-DN0t63Z0.js";
+import { r as ExecutionReport } from "../analyze-runs-DjlZOVXx.js";
+import { l as PrimeRawUsage } from "../prime-protocol-BWG2DQIk.js";
 import { i as nodeHttpPrimeBridgeTransport, n as PrimeBridgeTransportRequest, r as PrimeBridgeTransportResult, t as PrimeBridgeTransport } from "../prime-bridge-transport-6feEglLf.js";
-import { a as isContentAttribute, i as ingestSpans, n as IngestReport, r as epochMillis, t as DiagnosisSpan } from "../spans-BHAUXMGN.js";
+import { a as isContentAttribute, i as ingestSpans, n as IngestReport, r as epochMillis, t as DiagnosisSpan } from "../spans-CVngGij3.js";
 import { SpanKind, TraceValidation } from "@tangle-network/agent-trace-contract";
 //#region src/diagnosis/findings.d.ts
 /**
@@ -13,7 +13,6 @@ import { SpanKind, TraceValidation } from "@tangle-network/agent-trace-contract"
  * objects, so a document that passes {@link validateDiagnosisFindings} passes
  * that schema.
  */
-export declare const DIAGNOSIS_FINDINGS_SCHEMA_ID = "https://tangle.tools/schemas/diagnosis-findings-v1.json";
 type DiagnosisSeverity = 'critical' | 'high' | 'medium' | 'low';
 /** `observed` is read directly off the trace; `inferred` is a model's reading of it. */
 type DiagnosisConfidence = 'observed' | 'inferred';
@@ -64,15 +63,6 @@ interface DiagnosisFindingsDocument {
   };
   findings: DiagnosisFinding[];
 }
-/**
- * Structural check against the v1 schema plus the two rules the schema states
- * in prose: an unavailable capability carries a reason, and a measure carries a
- * denominator. When `spanIds` is given, every evidence id must resolve to it.
- * Returns the list of defects; empty means valid.
- */
-export declare function validateDiagnosisFindings(document: unknown, options?: {
-  spanIds?: ReadonlySet<string>;
-}): string[];
 //#endregion
 //#region src/diagnosis/first-failure.d.ts
 type FirstFailureStage = 'error-span' | 'failed-outcome';

package/dist/diagnosis/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/diagnosis/findings.ts","../../src/diagnosis/first-failure.ts","../../src/diagnosis/deterministic.ts","../../src/diagnosis/model-pass.ts","../../src/diagnosis/engine.ts"],"mappings":";;;;;;;;;;;;;;;qBASa;KAGD;;KAGA;UAEK;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA,UAAU;EACV;EACA;EACA,UAAU;;EAEV;EACA;EACA,YAAY;EACZ;;UAGe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,QAAQ;EACR;;UAGe;EACf;EACA;IACE;IACA;IACA;IACA;;EAEF;IACE,cAAc,eAAe;IAC7B,SAAS;IACT,YAAY;;EAEd,UAAU;;;;;;;;wBAiBI,0BACd,mBACA;EAAW,UAAU;;;;KCpCX;UAEK;EACf;EACA;EACA,OAAO;EACP;EACA;EACA,MAAM;;EAEN;;EAEA,gBAAgB;EAChB,OAAO;;EAEP;EACA;;UAGe;EACf;EACA;EACA,OAAO;;EAEP;EACA;EACA;;UAGe;EACf;EACA;EACA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA;;KAGU,eACR,oBACA,wBACA,mBACA;;wBAGY,iBAAiB,gBAAgB,kBAAkB;;;UCjClD;EACf;EACA;EACA,aAAa;EACb;EACA;EACA;;EAEA,eAAe;;EAEf,WAAW;EACX;;UAGe;EACf,YAAY;EACZ,cAAc,eAAe;EAC7B,SAAS;EACT,OAAO;EACP,UAAU;EACV;;;;UCpEe;;EAEf,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;IAAY;IAA4B;;;UAGzB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;IAAa;IAAiB;IAAkB;;EAChD,OAAO;EACP;;;;UC/Be;;EAEf;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;IAAa,MAAM;MAAQ;QAAe;MAA6B;;KAG7D;EACN;EAAuB;EAAiB,SAAS;;EACjD;EAAe,OAAO;EAAuB,SAAS;;UAE3C;EACf;EACA;EACA;EACA;;UAGe;;EAEf,UAAU;;EAEV,WAAW;;EAEX,OAAO;;EAEP,WAAW;IAAQ;IAAiB;;EACpC,OAAO;EACP,aAAa;EACb,QAAQ;;EAER,UAAU;EACV;IACE;IACA;IACA,MAAM;IACN;IACA,OAAO;;EAET;IAAU;IAA8B;;;wBAOpB,cACpB,2BACA,SAAS,kBACT,SAAS,mBACR,QAAQ"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/diagnosis/findings.ts","../../src/diagnosis/first-failure.ts","../../src/diagnosis/deterministic.ts","../../src/diagnosis/model-pass.ts","../../src/diagnosis/engine.ts"],"mappings":";;;;;;;;;;;;;;;KAWY;;KAGA;UAEK;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA,UAAU;EACV;EACA;EACA,UAAU;;EAEV;EACA;EACA,YAAY;EACZ;;UAGe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,QAAQ;EACR;;UAGe;EACf;EACA;IACE;IACA;IACA;IACA;;EAEF;IACE,cAAc,eAAe;IAC7B,SAAS;IACT,YAAY;;EAEd,UAAU;;;;KChBA;UAEK;EACf;EACA;EACA,OAAO;EACP;EACA;EACA,MAAM;;EAEN;;EAEA,gBAAgB;EAChB,OAAO;;EAEP;EACA;;UAGe;EACf;EACA;EACA,OAAO;;EAEP;EACA;EACA;;UAGe;EACf;EACA;EACA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA;;KAGU,eACR,oBACA,wBACA,mBACA;;wBAGY,iBAAiB,gBAAgB,kBAAkB;;;UCjClD;EACf;EACA;EACA,aAAa;EACb;EACA;EACA;;EAEA,eAAe;;EAEf,WAAW;EACX;;UAGe;EACf,YAAY;EACZ,cAAc,eAAe;EAC7B,SAAS;EACT,OAAO;EACP,UAAU;EACV;;;;UCpEe;;EAEf,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;IAAY;IAA4B;;;UAGzB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;IAAa;IAAiB;IAAkB;;EAChD,OAAO;EACP;;;;UC/Be;;EAEf;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;IAAa,MAAM;MAAQ;QAAe;MAA6B;;KAG7D;EACN;EAAuB;EAAiB,SAAS;;EACjD;EAAe,OAAO;EAAuB,SAAS;;UAE3C;EACf;EACA;EACA;EACA;;UAGe;;EAEf,UAAU;;EAEV,WAAW;;EAEX,OAAO;;EAEP,WAAW;IAAQ;IAAiB;;EACpC,OAAO;EACP,aAAa;EACb,QAAQ;;EAER,UAAU;EACV;IACE;IACA;IACA,MAAM;IACN;IACA,OAAO;;EAET;IAAU;IAA8B;;;wBAOpB,cACpB,2BACA,SAAS,kBACT,SAAS,mBACR,QAAQ"}

package/dist/diagnosis/index.js

@@ -1,12 +1,15 @@
-import { r as summarizeExecution } from "../analyze-runs-C84irm4i.js";
-import { a as emptyRedactionReport, c as redactText, o as redact } from "../redact-B5xXKqQ3.js";
+import { r as summarizeExecution } from "../analyze-runs-DqPJaJ4X.js";
+import { a as redact, i as emptyRedactionReport, s as redactText } from "../redact-mtkTiJVk.js";
 import { t as packageVersion } from "../package-version-D7lQHt_-.js";
-import { o as classifyFailureReason } from "../failure-taxonomy-CjabTATj.js";
-import { c as primeProtocolSha256, d as runPrimeExchange, n as buildPrimePrompt } from "../prime-protocol-CHKgA3a6.js";
+import { r as classifyFailureReason } from "../failure-taxonomy-Bx5oWqaW.js";
+import { c as runPrimeExchange, n as buildPrimePrompt, o as primeProtocolSha256 } from "../prime-protocol-BFgGrEMP.js";
 import { t as nodeHttpPrimeBridgeTransport } from "../prime-bridge-transport-Z3cpKGoV.js";
-import { t as fromOtelSpans } from "../otel-spans-B9tUXRHF.js";
+import { t as fromOtelSpans } from "../otel-spans-CdN909k1.js";
 import { createHash, createHmac, randomBytes } from "node:crypto";
 import { MODEL_ATTR_KEYS, SPAN_KINDS, TOOL_NAME_ATTR_KEYS, resolveSpanKind, validateTraceSpans } from "@tangle-network/agent-trace-contract";
+//#region src/diagnosis/first-failure.ts
+/** Span ids listed beside a result; the count carries the rest. */
+const MAX_FIRST_FAILURE_IDS = 8;
 const OUTCOME_ATTR = "agent.outcome";
 /** Rank the first failure of one trace. Throws when the spans belong to more than one trace. */
 function rankFirstFailure(spans) {
@@ -73,7 +76,7 @@ function rank(traceId, stage, candidates) {
 		message: first.statusMessage,
 		classification,
 		blame: classification?.blame ?? "unknown",
-		later: later.slice(0, 8).map((span) => span.spanId),
+		later: later.slice(0, MAX_FIRST_FAILURE_IDS).map((span) => span.spanId),
 		laterCount: later.length
 	};
 }
@@ -83,7 +86,7 @@ function ambiguous(traceId, stage, candidates, reason) {
 		status: "ambiguous",
 		traceId,
 		stage,
-		candidates: ids.slice(0, 8),
+		candidates: ids.slice(0, MAX_FIRST_FAILURE_IDS),
 		candidateCount: ids.length,
 		reason
 	};
@@ -311,6 +314,28 @@ function firstString(attributes, keys) {
 function stringOr(value, fallback) {
 	return typeof value === "string" && value.length > 0 ? value : fallback;
 }
+//#endregion
+//#region src/diagnosis/deterministic.ts
+/**
+* The deterministic pass: everything the engine can say without a model.
+*
+* It produces the capability table (the trace contract's own validator, so the
+* reasons are the trace's reasons), execution facts with token and cost
+* accounting (agent-eval's `fromOtelSpans` + `summarizeExecution`), and
+* `observed` findings computed directly from span data. Every observed finding
+* carries a measure with its denominator and cites the spans it counted.
+*/
+/** Evidence ids cited per finding; the measure carries the full count. */
+const MAX_EVIDENCE_PER_FINDING = 8;
+/**
+* Identical tool calls with no different tool call between them that count as
+* a repeated-call run. Model turns between the calls do not break the run: an
+* agent that calls, reads, and calls again with the same input is the waste
+* this detects.
+*/
+const REPEAT_RUN_LENGTH = 3;
+/** A tool is reported only once it has at least this many calls. */
+const MIN_TOOL_CALLS = 3;
 /** The analysis each capability gates, as the report names it. */
 const CAPABILITY_ANALYSES = {
 	"token-accounting": "token accounting",
@@ -397,7 +422,7 @@ function toolFailureFindings(tools) {
 	const findings = [];
 	for (const [tool, calls] of byTool) {
 		const failed = calls.filter((span) => span.status === "ERROR");
-		if (calls.length < 3 || failed.length < 2) continue;
+		if (calls.length < MIN_TOOL_CALLS || failed.length < 2) continue;
 		const rate = failed.length / calls.length;
 		const severity = rate >= .5 && failed.length >= 5 ? "high" : rate >= .2 ? "medium" : "low";
 		findings.push({
@@ -411,7 +436,7 @@ function toolFailureFindings(tools) {
 				unit: "%",
 				denominator: `${calls.length} calls to ${tool}`
 			},
-			evidence: failed.slice(0, 8).map((span) => span.spanId),
+			evidence: failed.slice(0, MAX_EVIDENCE_PER_FINDING).map((span) => span.spanId),
 			confidence: "observed",
 			recommendation: `Read the failed ${tool} calls' status messages and fix the most common cause, or give the agent the precondition it keeps missing.`
 		});
@@ -429,22 +454,22 @@ function repeatedCallFindings(traces) {
 			const last = current[current.length - 1];
 			if (last && last.inputDigest === span.inputDigest && (last.toolName ?? last.name) === (span.toolName ?? span.name)) current.push(span);
 			else {
-				if (current.length >= 3) runs.push(current);
+				if (current.length >= REPEAT_RUN_LENGTH) runs.push(current);
 				current = [span];
 			}
 		}
-		if (current.length >= 3) runs.push(current);
+		if (current.length >= REPEAT_RUN_LENGTH) runs.push(current);
 	}
 	if (runs.length === 0) return [];
 	const wasted = runs.reduce((sum, run) => sum + run.length - 1, 0);
 	const tracesAffected = new Set(runs.map((run) => run[0].traceId)).size;
 	runs.sort((a, b) => b.length - a.length);
-	const evidence = runs.flatMap((run) => run.slice(0, 2).map((span) => span.spanId)).slice(0, 8);
+	const evidence = runs.flatMap((run) => run.slice(0, 2).map((span) => span.spanId)).slice(0, MAX_EVIDENCE_PER_FINDING);
 	const rate = wasted / digestCalls;
 	return [{
 		id: stableId("repeated-calls", String(runs.length)),
 		severity: rate >= .1 ? "high" : rate >= .03 ? "medium" : "low",
-		claim: `The agent made the same tool call with the same input 3 or more times with no other tool call in between, ${runs.length} times across ${tracesAffected} runs; the longest streak was ${runs[0].length} calls.`,
+		claim: `The agent made the same tool call with the same input ${REPEAT_RUN_LENGTH} or more times with no other tool call in between, ${runs.length} times across ${tracesAffected} runs; the longest streak was ${runs[0].length} calls.`,
 		consequence: `${wasted} calls repeated the previous tool call with the same input, which is work the agent paid for twice without new information.`,
 		measure: {
 			name: "repeated identical tool calls",
@@ -477,7 +502,7 @@ function failedRunFindings(traces) {
 			unit: "%",
 			denominator: `${traces.size} runs`
 		},
-		evidence: failedRoots.slice(0, 8).map((span) => span.spanId),
+		evidence: failedRoots.slice(0, MAX_EVIDENCE_PER_FINDING).map((span) => span.spanId),
 		confidence: "observed",
 		recommendation: "Group the failed runs by their root status message and fix the largest group first."
 	}];
@@ -556,15 +581,6 @@ function round(value) {
 }
 //#endregion
 //#region src/diagnosis/findings.ts
-/**
-* The diagnosis findings document: what the engine returns and what the
-* diagnosis report renders. It mirrors the kit's JSON Schema
-* (`https://tangle.tools/schemas/diagnosis-findings-v1.json`, owned by the
-* traces repository's diagnosis kit) field for field, including its closed
-* objects, so a document that passes {@link validateDiagnosisFindings} passes
-* that schema.
-*/
-const DIAGNOSIS_FINDINGS_SCHEMA_ID = "https://tangle.tools/schemas/diagnosis-findings-v1.json";
 const SEVERITIES$1 = /* @__PURE__ */ new Set([
 	"critical",
 	"high",
@@ -1102,6 +1118,29 @@ function shortString(value) {
 	if (!trimmed) return void 0;
 	return trimmed.length <= MAX_STRING ? trimmed : `${trimmed.slice(0, MAX_STRING)}…`;
 }
+//#endregion
+//#region src/diagnosis/engine.ts
+/**
+* The agent-failure diagnosis engine: flat spans plus context in, a findings
+* document out.
+*
+* It is a plain function. Its only side effect is the model call made through
+* the caller's transport, and in deterministic mode it makes none. It keeps no
+* storage, files no issues, and retains nothing: where the spans came from and
+* where the findings go belong to the caller. That is what lets one engine
+* serve both the internal daily mining run and a customer diagnosis while their
+* data paths stay separate.
+*
+* Order of operations, each step reading only the previous step's output:
+*   1. secret filter and content drop over every input (spans, focus, topology);
+*   2. the deterministic pass: capability table, execution facts, observed findings;
+*   3. in model mode, the prime-protocol pass over the highest-signal runs;
+*   4. evidence resolution: every cited id must be a span id of the input that
+*      no other run shares; findings left without evidence are rejected.
+*/
+const DEFAULT_MAX_MODEL_TRACES = 6;
+/** Runs with fewer spans carry too little for the model to read. */
+const MIN_MODEL_SPANS = 3;
 async function diagnoseSpans(spans, context, options) {
 	if (context.subject !== "internal" && context.subject !== "customer") throw new TypeError(`diagnoseSpans: subject must be internal or customer, got ${String(context.subject)}`);
 	if (context.subject === "customer" && context.topology !== void 0) throw new TypeError("diagnoseSpans: topology analysis is internal only; a customer context must not carry one");
@@ -1124,17 +1163,17 @@ async function diagnoseSpans(spans, context, options) {
 		reason: options.reason ?? "deterministic mode: no model call was made"
 	});
 	else {
-		const maxTraces = options.model.maxTraces ?? 6;
+		const maxTraces = options.model.maxTraces ?? DEFAULT_MAX_MODEL_TRACES;
 		const ranked = rankTraces(traces);
 		const selected = ranked.slice(0, maxTraces);
 		const short = traces.size - ranked.length;
 		if (ranked.length === 0) skipped.push({
 			analysis: "model reading of the runs",
-			reason: `no run has the 3 spans the model needs to read it; ${traces.size} runs were too short`
+			reason: `no run has the ${MIN_MODEL_SPANS} spans the model needs to read it; ${traces.size} runs were too short`
 		});
 		else if (short > 0) skipped.push({
 			analysis: `model reading of ${short} of ${traces.size} runs`,
-			reason: `those runs have fewer than 3 spans, too few for the model to read`
+			reason: `those runs have fewer than ${MIN_MODEL_SPANS} spans, too few for the model to read`
 		});
 		if (ranked.length > selected.length) skipped.push({
 			analysis: `model reading of ${ranked.length - selected.length} of ${ranked.length} runs`,
@@ -1222,7 +1261,7 @@ async function diagnoseSpans(spans, context, options) {
 }
 /** Runs with more errors first, then larger runs; runs under MIN_MODEL_SPANS are left out. */
 function rankTraces(traces) {
-	return [...traces.entries()].filter(([, spans]) => spans.length >= 3).map(([traceId, spans]) => ({
+	return [...traces.entries()].filter(([, spans]) => spans.length >= MIN_MODEL_SPANS).map(([traceId, spans]) => ({
 		traceId,
 		spans,
 		errors: spans.filter((span) => span.status === "ERROR").length
@@ -1266,6 +1305,6 @@ function totalUsage(runs, pricing) {
 	};
 }
 //#endregion
-export { DIAGNOSIS_FINDINGS_SCHEMA_ID, diagnoseSpans, epochMillis, ingestSpans, isContentAttribute, nodeHttpPrimeBridgeTransport, rankFirstFailure, validateDiagnosisFindings };
+export { diagnoseSpans, epochMillis, ingestSpans, isContentAttribute, nodeHttpPrimeBridgeTransport, rankFirstFailure };
 
 //# sourceMappingURL=index.js.map