@tangle-network/agent-eval 0.102.1 → 0.103.1

package/dist/chunk-RQNOLV3I.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/product-benchmark/index.ts","../src/product-benchmark/export.ts"],"sourcesContent":["import { existsSync, readFileSync, statSync } from 'node:fs'\nimport { dirname, join } from 'node:path'\nimport { ValidationError } from '../errors'\n\nexport const productBenchmarkSplits = ['practice', 'dev', 'holdout', 'safety', 'sentinel'] as const\n\nexport type ProductBenchmarkSplit = (typeof productBenchmarkSplits)[number]\n\nexport interface ProductBenchmarkRepoRef {\n  readonly url: string\n  readonly commit: string\n  readonly branch: string\n}\n\nexport interface ProductBenchmarkSubstrateVersions {\n  readonly agentEval: string\n  readonly agentRuntime: string\n  readonly agentInterface: string\n  readonly sandbox: string\n  readonly agentBench?: string\n}\n\nexport interface ProductBenchmarkProfileRef {\n  readonly id: string\n  readonly profileHash: string\n  readonly agentProfilePath: string\n}\n\nexport interface ProductBenchmarkArm {\n  readonly id: string\n  readonly profileId: string\n  readonly mutableSurfaces: readonly string[]\n  readonly policyAxes: Record<string, unknown>\n}\n\nexport interface ProductBenchmarkScenario {\n  readonly id: string\n  readonly split: ProductBenchmarkSplit\n  readonly tags: readonly string[]\n  readonly sourceAllowedForSynthesis: boolean\n}\n\nexport interface ProductBenchmarkBudgets {\n  readonly maxUsd: number\n  readonly maxCells: number\n  readonly maxWallMs: number\n}\n\nexport interface ProductBenchmarkManifest {\n  readonly schemaVersion: 1\n  readonly projectId: string\n  readonly benchmarkId: string\n  readonly repo: ProductBenchmarkRepoRef\n  readonly substrate: ProductBenchmarkSubstrateVersions\n  readonly profiles: readonly ProductBenchmarkProfileRef[]\n  readonly arms: readonly ProductBenchmarkArm[]\n  readonly scenarios: readonly ProductBenchmarkScenario[]\n  readonly budgets: ProductBenchmarkBudgets\n  readonly expectedArtifactDir: string\n}\n\nexport interface AgentProfileRuntimeReceipt {\n  readonly model: string\n  readonly harness: string\n  readonly backend: string\n  readonly reasoningEffort?: string\n}\n\nexport type RuntimeResolution = AgentProfileRuntimeReceipt\n\nexport interface ProductBenchmarkRecord {\n  readonly schemaVersion: 1\n  readonly projectId: string\n  readonly benchmarkId: string\n  readonly runId: string\n  readonly scenarioId: string\n  readonly split: ProductBenchmarkSplit\n  readonly armId: string\n  readonly rep: number\n  readonly agentProfile: {\n    readonly id: string\n    readonly hash: string\n    readonly path: string\n    readonly declared: RuntimeResolution\n    readonly resolved: RuntimeResolution\n  }\n  readonly model: {\n    readonly provider: string\n    readonly id: string\n  }\n  readonly backend: {\n    readonly kind: string\n    readonly version: string\n  }\n  readonly outcome: {\n    readonly pass: boolean\n    readonly score: number\n    readonly dimensions: Record<string, number>\n    readonly failureMode: string | null\n  }\n  readonly usage: {\n    readonly inputTokens: number\n    readonly outputTokens: number\n    readonly costUsd: number\n    readonly wallMs: number\n    readonly toolCalls: number\n  }\n  readonly integrity: {\n    readonly realBackend: boolean\n    readonly rawCapture: boolean\n    readonly traceCapture: boolean\n    readonly noStubRows: boolean\n    readonly priced: boolean\n    readonly profileMaterialized: boolean\n  }\n  readonly artifacts: {\n    readonly records: string\n    readonly traces: string\n    readonly raws: string\n    readonly scores: string\n    readonly workspace: string\n  }\n}\n\nexport interface ProductBenchmarkRunInput {\n  readonly manifestPath: string\n  readonly recordsPath: string\n  readonly artifactRoot?: string\n  readonly checkArtifacts?: boolean\n}\n\nexport interface ProductBenchmarkValidationReport {\n  readonly manifestPath: string\n  readonly recordsPath: string\n  readonly records: number\n  /** Manifest repo fields that are empty or the `'unknown'` export sentinel. */\n  readonly repoFailures: readonly string[]\n  /** Manifest substrate versions that are empty or the `'unknown'` export\n   *  sentinel — a bundle without substrate identity is not reproducible. */\n  readonly substrateFailures: readonly string[]\n  readonly projects: readonly string[]\n  readonly benchmarks: readonly string[]\n  readonly arms: readonly string[]\n  readonly scenarios: readonly string[]\n  readonly passed: number\n  readonly failed: number\n  readonly inputTokens: number\n  readonly outputTokens: number\n  readonly costUsd: number\n  readonly wallMs: number\n  readonly integrityFailures: readonly string[]\n  readonly missingArtifacts: readonly string[]\n}\n\nexport interface ProductBenchmarkArtifactPaths {\n  readonly manifestPath: string\n  readonly recordsPath: string\n}\n\nfunction isObject(value: unknown): value is Record<string, unknown> {\n  return Boolean(value) && typeof value === 'object' && !Array.isArray(value)\n}\n\nfunction fail(path: string, message: string): never {\n  throw new ValidationError(`${path}: ${message}`)\n}\n\nfunction wrapValidationError(path: string, err: unknown): never {\n  if (err instanceof ValidationError) {\n    throw new ValidationError(`${path}: ${err.message}`, { cause: err })\n  }\n  throw new ValidationError(`${path}: ${err instanceof Error ? err.message : String(err)}`, {\n    cause: err,\n  })\n}\n\nfunction expectObject(value: unknown, path: string): Record<string, unknown> {\n  if (!isObject(value)) fail(path, 'must be an object')\n  return value\n}\n\nfunction expectString(value: unknown, path: string): string {\n  if (typeof value !== 'string' || value.trim().length === 0)\n    fail(path, 'must be a non-empty string')\n  return value\n}\n\nfunction expectBoolean(value: unknown, path: string): boolean {\n  if (typeof value !== 'boolean') fail(path, 'must be a boolean')\n  return value\n}\n\nfunction expectNumber(\n  value: unknown,\n  path: string,\n  opts: { readonly min?: number; readonly max?: number; readonly integer?: boolean } = {},\n): number {\n  if (typeof value !== 'number' || !Number.isFinite(value)) fail(path, 'must be a finite number')\n  if (opts.integer && !Number.isInteger(value)) fail(path, 'must be an integer')\n  if (opts.min !== undefined && value < opts.min) fail(path, `must be >= ${opts.min}`)\n  if (opts.max !== undefined && value > opts.max) fail(path, `must be <= ${opts.max}`)\n  return value\n}\n\nfunction expectStringArray(value: unknown, path: string): string[] {\n  if (!Array.isArray(value)) fail(path, 'must be an array')\n  return value.map((entry, index) => expectString(entry, `${path}[${index}]`))\n}\n\nfunction expectObjectArray(value: unknown, path: string): Record<string, unknown>[] {\n  if (!Array.isArray(value)) fail(path, 'must be an array')\n  return value.map((entry, index) => expectObject(entry, `${path}[${index}]`))\n}\n\nfunction expectSplit(value: unknown, path: string): ProductBenchmarkSplit {\n  const split = expectString(value, path)\n  if (!productBenchmarkSplits.includes(split as ProductBenchmarkSplit)) {\n    fail(path, `must be one of ${productBenchmarkSplits.join(', ')}`)\n  }\n  return split as ProductBenchmarkSplit\n}\n\nfunction optionalString(value: unknown, path: string): string | undefined {\n  if (value === undefined) return undefined\n  return expectString(value, path)\n}\n\nfunction expectDimensions(value: unknown, path: string): Record<string, number> {\n  const obj = expectObject(value, path)\n  const out: Record<string, number> = {}\n  for (const [key, raw] of Object.entries(obj)) out[key] = expectNumber(raw, `${path}.${key}`)\n  return out\n}\n\nfunction expectRuntimeResolution(value: unknown, path: string): RuntimeResolution {\n  const obj = expectObject(value, path)\n  return {\n    model: expectString(obj.model, `${path}.model`),\n    harness: expectString(obj.harness, `${path}.harness`),\n    backend: expectString(obj.backend, `${path}.backend`),\n    ...(obj.reasoningEffort !== undefined\n      ? { reasoningEffort: optionalString(obj.reasoningEffort, `${path}.reasoningEffort`) }\n      : {}),\n  }\n}\n\nexport function validateProductBenchmarkManifest(value: unknown): ProductBenchmarkManifest {\n  const obj = expectObject(value, 'manifest')\n  if (obj.schemaVersion !== 1) fail('manifest.schemaVersion', 'must be 1')\n  const repo = expectObject(obj.repo, 'manifest.repo')\n  const substrate = expectObject(obj.substrate, 'manifest.substrate')\n  const profiles = expectObjectArray(obj.profiles, 'manifest.profiles').map((profile, index) => ({\n    id: expectString(profile.id, `manifest.profiles[${index}].id`),\n    profileHash: expectString(profile.profileHash, `manifest.profiles[${index}].profileHash`),\n    agentProfilePath: expectString(\n      profile.agentProfilePath,\n      `manifest.profiles[${index}].agentProfilePath`,\n    ),\n  }))\n  const arms = expectObjectArray(obj.arms, 'manifest.arms').map((arm, index) => ({\n    id: expectString(arm.id, `manifest.arms[${index}].id`),\n    profileId: expectString(arm.profileId, `manifest.arms[${index}].profileId`),\n    mutableSurfaces: expectStringArray(\n      arm.mutableSurfaces,\n      `manifest.arms[${index}].mutableSurfaces`,\n    ),\n    policyAxes: expectObject(arm.policyAxes, `manifest.arms[${index}].policyAxes`),\n  }))\n  const scenarios = expectObjectArray(obj.scenarios, 'manifest.scenarios').map(\n    (scenario, index) => ({\n      id: expectString(scenario.id, `manifest.scenarios[${index}].id`),\n      split: expectSplit(scenario.split, `manifest.scenarios[${index}].split`),\n      tags: expectStringArray(scenario.tags, `manifest.scenarios[${index}].tags`),\n      sourceAllowedForSynthesis: expectBoolean(\n        scenario.sourceAllowedForSynthesis,\n        `manifest.scenarios[${index}].sourceAllowedForSynthesis`,\n      ),\n    }),\n  )\n  const budgets = expectObject(obj.budgets, 'manifest.budgets')\n  if (profiles.length === 0) fail('manifest.profiles', 'must contain at least one profile')\n  if (arms.length === 0) fail('manifest.arms', 'must contain at least one arm')\n  if (scenarios.length === 0) fail('manifest.scenarios', 'must contain at least one scenario')\n  assertUnique(\n    profiles.map((profile) => profile.id),\n    'manifest.profiles.id',\n  )\n  assertUnique(\n    arms.map((arm) => arm.id),\n    'manifest.arms.id',\n  )\n  assertUnique(\n    scenarios.map((scenario) => scenario.id),\n    'manifest.scenarios.id',\n  )\n  const profileIds = new Set(profiles.map((profile) => profile.id))\n  for (const arm of arms) {\n    if (!profileIds.has(arm.profileId))\n      fail(`manifest.arms.${arm.id}.profileId`, `unknown profile ${arm.profileId}`)\n  }\n  return {\n    schemaVersion: 1,\n    projectId: expectString(obj.projectId, 'manifest.projectId'),\n    benchmarkId: expectString(obj.benchmarkId, 'manifest.benchmarkId'),\n    repo: {\n      url: expectString(repo.url, 'manifest.repo.url'),\n      commit: expectString(repo.commit, 'manifest.repo.commit'),\n      branch: expectString(repo.branch, 'manifest.repo.branch'),\n    },\n    substrate: {\n      agentEval: expectString(substrate.agentEval, 'manifest.substrate.agentEval'),\n      agentRuntime: expectString(substrate.agentRuntime, 'manifest.substrate.agentRuntime'),\n      agentInterface: expectString(substrate.agentInterface, 'manifest.substrate.agentInterface'),\n      sandbox: expectString(substrate.sandbox, 'manifest.substrate.sandbox'),\n      ...(substrate.agentBench !== undefined\n        ? { agentBench: expectString(substrate.agentBench, 'manifest.substrate.agentBench') }\n        : {}),\n    },\n    profiles,\n    arms,\n    scenarios,\n    budgets: {\n      maxUsd: expectNumber(budgets.maxUsd, 'manifest.budgets.maxUsd', { min: 0 }),\n      maxCells: expectNumber(budgets.maxCells, 'manifest.budgets.maxCells', {\n        min: 0,\n        integer: true,\n      }),\n      maxWallMs: expectNumber(budgets.maxWallMs, 'manifest.budgets.maxWallMs', {\n        min: 0,\n        integer: true,\n      }),\n    },\n    expectedArtifactDir: expectString(obj.expectedArtifactDir, 'manifest.expectedArtifactDir'),\n  }\n}\n\nexport function validateProductBenchmarkRecord(value: unknown): ProductBenchmarkRecord {\n  const obj = expectObject(value, 'record')\n  if (obj.schemaVersion !== 1) fail('record.schemaVersion', 'must be 1')\n  const agentProfile = expectObject(obj.agentProfile, 'record.agentProfile')\n  const model = expectObject(obj.model, 'record.model')\n  const backend = expectObject(obj.backend, 'record.backend')\n  const outcome = expectObject(obj.outcome, 'record.outcome')\n  const usage = expectObject(obj.usage, 'record.usage')\n  const integrity = expectObject(obj.integrity, 'record.integrity')\n  const artifacts = expectObject(obj.artifacts, 'record.artifacts')\n  const record: ProductBenchmarkRecord = {\n    schemaVersion: 1,\n    projectId: expectString(obj.projectId, 'record.projectId'),\n    benchmarkId: expectString(obj.benchmarkId, 'record.benchmarkId'),\n    runId: expectString(obj.runId, 'record.runId'),\n    scenarioId: expectString(obj.scenarioId, 'record.scenarioId'),\n    split: expectSplit(obj.split, 'record.split'),\n    armId: expectString(obj.armId, 'record.armId'),\n    rep: expectNumber(obj.rep, 'record.rep', { min: 1, integer: true }),\n    agentProfile: {\n      id: expectString(agentProfile.id, 'record.agentProfile.id'),\n      hash: expectString(agentProfile.hash, 'record.agentProfile.hash'),\n      path: expectString(agentProfile.path, 'record.agentProfile.path'),\n      declared: expectRuntimeResolution(agentProfile.declared, 'record.agentProfile.declared'),\n      resolved: expectRuntimeResolution(agentProfile.resolved, 'record.agentProfile.resolved'),\n    },\n    model: {\n      provider: expectString(model.provider, 'record.model.provider'),\n      id: expectString(model.id, 'record.model.id'),\n    },\n    backend: {\n      kind: expectString(backend.kind, 'record.backend.kind'),\n      version: expectString(backend.version, 'record.backend.version'),\n    },\n    outcome: {\n      pass: expectBoolean(outcome.pass, 'record.outcome.pass'),\n      score: expectNumber(outcome.score, 'record.outcome.score', { min: 0, max: 1 }),\n      dimensions: expectDimensions(outcome.dimensions, 'record.outcome.dimensions'),\n      failureMode:\n        outcome.failureMode === null\n          ? null\n          : expectString(outcome.failureMode, 'record.outcome.failureMode'),\n    },\n    usage: {\n      inputTokens: expectNumber(usage.inputTokens, 'record.usage.inputTokens', {\n        min: 0,\n        integer: true,\n      }),\n      outputTokens: expectNumber(usage.outputTokens, 'record.usage.outputTokens', {\n        min: 0,\n        integer: true,\n      }),\n      costUsd: expectNumber(usage.costUsd, 'record.usage.costUsd', { min: 0 }),\n      wallMs: expectNumber(usage.wallMs, 'record.usage.wallMs', { min: 0, integer: true }),\n      toolCalls: expectNumber(usage.toolCalls, 'record.usage.toolCalls', { min: 0, integer: true }),\n    },\n    integrity: {\n      realBackend: expectBoolean(integrity.realBackend, 'record.integrity.realBackend'),\n      rawCapture: expectBoolean(integrity.rawCapture, 'record.integrity.rawCapture'),\n      traceCapture: expectBoolean(integrity.traceCapture, 'record.integrity.traceCapture'),\n      noStubRows: expectBoolean(integrity.noStubRows, 'record.integrity.noStubRows'),\n      priced: expectBoolean(integrity.priced, 'record.integrity.priced'),\n      profileMaterialized: expectBoolean(\n        integrity.profileMaterialized,\n        'record.integrity.profileMaterialized',\n      ),\n    },\n    artifacts: {\n      records: expectString(artifacts.records, 'record.artifacts.records'),\n      traces: expectString(artifacts.traces, 'record.artifacts.traces'),\n      raws: expectString(artifacts.raws, 'record.artifacts.raws'),\n      scores: expectString(artifacts.scores, 'record.artifacts.scores'),\n      workspace: expectString(artifacts.workspace, 'record.artifacts.workspace'),\n    },\n  }\n  if (record.integrity.realBackend && record.usage.inputTokens + record.usage.outputTokens === 0) {\n    fail('record.usage', 'realBackend rows must carry non-zero token usage')\n  }\n  return record\n}\n\nexport function productBenchmarkIntegrityFailures(record: ProductBenchmarkRecord): string[] {\n  const failures: string[] = []\n  for (const [key, value] of Object.entries(record.integrity)) {\n    if (!value) failures.push(`${record.runId}:${key}=false`)\n  }\n  return failures\n}\n\nexport function readProductBenchmarkRecords(path: string): ProductBenchmarkRecord[] {\n  const lines = readFileSync(path, 'utf8')\n    .split('\\n')\n    .map((line) => line.trim())\n    .filter(Boolean)\n  const records: ProductBenchmarkRecord[] = []\n  for (const [index, line] of lines.entries()) {\n    try {\n      records.push(validateProductBenchmarkRecord(JSON.parse(line)))\n    } catch (err) {\n      wrapValidationError(`${path}:${index + 1}`, err)\n    }\n  }\n  return records\n}\n\nexport function readProductBenchmarkManifest(path: string): ProductBenchmarkManifest {\n  try {\n    return validateProductBenchmarkManifest(JSON.parse(readFileSync(path, 'utf8')))\n  } catch (err) {\n    wrapValidationError(path, err)\n  }\n}\n\nexport function validateProductBenchmarkRun(\n  input: ProductBenchmarkRunInput,\n): ProductBenchmarkValidationReport {\n  const manifest = readProductBenchmarkManifest(input.manifestPath)\n  const records = readProductBenchmarkRecords(input.recordsPath)\n  const manifestProjectBench = `${manifest.projectId}/${manifest.benchmarkId}`\n  const integrityFailures = records.flatMap(productBenchmarkIntegrityFailures)\n  const missingArtifacts =\n    input.checkArtifacts === false\n      ? []\n      : records.flatMap((record) =>\n          missingArtifactsForRecord(record, input.artifactRoot ?? dirname(input.recordsPath)),\n        )\n  for (const [index, record] of records.entries()) {\n    const recordProjectBench = `${record.projectId}/${record.benchmarkId}`\n    if (recordProjectBench !== manifestProjectBench) {\n      fail(\n        `records[${index}]`,\n        `project/benchmark ${recordProjectBench} does not match manifest ${manifestProjectBench}`,\n      )\n    }\n    if (!manifest.arms.some((arm) => arm.id === record.armId))\n      fail(`records[${index}].armId`, `unknown arm ${record.armId}`)\n    if (!manifest.scenarios.some((scenario) => scenario.id === record.scenarioId)) {\n      fail(`records[${index}].scenarioId`, `unknown scenario ${record.scenarioId}`)\n    }\n  }\n  const repoFailures = (['url', 'commit', 'branch'] as const)\n    .filter((key) => manifest.repo[key].trim().length === 0 || manifest.repo[key] === 'unknown')\n    .map((key) => `manifest.repo.${key}`)\n  const substrateFailures = (['agentEval', 'agentRuntime', 'agentInterface', 'sandbox'] as const)\n    .filter(\n      (key) => manifest.substrate[key].trim().length === 0 || manifest.substrate[key] === 'unknown',\n    )\n    .map((key) => `manifest.substrate.${key}`)\n  return {\n    manifestPath: input.manifestPath,\n    recordsPath: input.recordsPath,\n    records: records.length,\n    repoFailures,\n    substrateFailures,\n    projects: sortedUnique(records.map((record) => record.projectId)),\n    benchmarks: sortedUnique(records.map((record) => record.benchmarkId)),\n    arms: sortedUnique(records.map((record) => record.armId)),\n    scenarios: sortedUnique(records.map((record) => record.scenarioId)),\n    passed: records.filter((record) => record.outcome.pass).length,\n    failed: records.filter((record) => !record.outcome.pass).length,\n    inputTokens: sum(records, (record) => record.usage.inputTokens),\n    outputTokens: sum(records, (record) => record.usage.outputTokens),\n    costUsd: sum(records, (record) => record.usage.costUsd),\n    wallMs: sum(records, (record) => record.usage.wallMs),\n    integrityFailures,\n    missingArtifacts,\n  }\n}\n\nfunction missingArtifactsForRecord(record: ProductBenchmarkRecord, artifactRoot: string): string[] {\n  const missing: string[] = []\n  for (const [key, value] of Object.entries(record.artifacts)) {\n    const path = resolveArtifactPath(value, artifactRoot)\n    if (!existsSync(path)) missing.push(`${record.runId}:${key}:${value}`)\n  }\n  return missing\n}\n\nfunction resolveArtifactPath(value: string, artifactRoot: string): string {\n  return value.startsWith('/') ? value : join(artifactRoot, value)\n}\n\nfunction assertUnique(values: readonly string[], path: string): void {\n  const seen = new Set<string>()\n  for (const value of values) {\n    if (seen.has(value)) fail(path, `duplicate ${value}`)\n    seen.add(value)\n  }\n}\n\nfunction sortedUnique(values: readonly string[]): string[] {\n  return [...new Set(values)].sort()\n}\n\nfunction sum<T>(items: readonly T[], fn: (item: T) => number): number {\n  return items.reduce((total, item) => total + fn(item), 0)\n}\n\nexport function findProductBenchmarkArtifacts(\n  runDir: string,\n): ProductBenchmarkArtifactPaths | null {\n  const manifestPath = join(runDir, 'product-benchmark-manifest.json')\n  const recordsPath = join(runDir, 'product-benchmark-records.jsonl')\n  if (\n    existsSync(manifestPath) &&\n    statSync(manifestPath).isFile() &&\n    existsSync(recordsPath) &&\n    statSync(recordsPath).isFile()\n  ) {\n    return { manifestPath, recordsPath }\n  }\n  return null\n}\n\n/**\n * Fail-loud gate over a bundle directory: locates the manifest + records,\n * runs `validateProductBenchmarkRun`, and throws with every repo,\n * integrity, and artifact failure listed. Returns the report when clean.\n */\nexport function assertProductBenchmarkRun(runDir: string): ProductBenchmarkValidationReport {\n  const artifacts = findProductBenchmarkArtifacts(runDir)\n  if (!artifacts) {\n    fail(runDir, 'missing product-benchmark-manifest.json or product-benchmark-records.jsonl')\n  }\n  const report = validateProductBenchmarkRun({ ...artifacts, artifactRoot: runDir })\n  const failures = [\n    ...report.repoFailures,\n    ...report.substrateFailures,\n    ...report.integrityFailures,\n    ...report.missingArtifacts,\n  ]\n  if (failures.length > 0) {\n    fail(runDir, `product benchmark validation failed:\\n${failures.join('\\n')}`)\n  }\n  return report\n}\n\nexport type {\n  ProductBenchmarkExportOptions,\n  ProductBenchmarkExportResult,\n  ProductBenchmarkSingleRunExportOptions,\n} from './export'\nexport {\n  buildProductBenchmarkManifest,\n  exportProductBenchmark,\n  exportProductBenchmarkRuns,\n  productBenchmarkMutableSurfaces,\n  productBenchmarkRepoIdentity,\n  runRecordToProductBenchmarkRecord,\n} from './export'\n","/**\n * Export side of the product benchmark bundle contract: convert product\n * eval run directories (`records.jsonl` of `RunRecord` rows + trace/raw\n * artifacts) into a portable `product-benchmark-manifest.json` +\n * `product-benchmark-records.jsonl` bundle that\n * `validateProductBenchmarkRun` accepts.\n *\n * Product-specific policy (safety-split detection, tool-call recovery,\n * profile id fallback, artifact materialization) enters through explicit\n * options; everything else is the shared union of the tax/legal/creative\n * exporters. Scenario catalogs, smoke runners, and CLIs stay in the\n * products.\n *\n * Input rows are checked structurally, not with `validateRunRecord`:\n * product harnesses record bare model aliases and partial provenance, and\n * the bundle contract's own validators re-check every field that matters\n * on the way out.\n */\n\nimport { spawnSync } from 'node:child_process'\nimport { createHash } from 'node:crypto'\nimport { cpSync, existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { basename, dirname, isAbsolute, join, relative, resolve } from 'node:path'\nimport { ValidationError } from '../errors'\nimport type { RunRecord } from '../run-record'\nimport type {\n  ProductBenchmarkManifest,\n  ProductBenchmarkRecord,\n  ProductBenchmarkSplit,\n  RuntimeResolution,\n} from './index'\nimport { validateProductBenchmarkManifest, validateProductBenchmarkRecord } from './index'\n\n/** Full mutable-surface superset a product arm may declare. */\nexport const productBenchmarkMutableSurfaces = [\n  'prompt',\n  'resources.files',\n  'tools',\n  'mcp',\n  'hooks',\n  'subagents',\n] as const\n\nexport interface ProductBenchmarkExportOptions {\n  /** Source eval run directories, each containing a `records.jsonl` of RunRecord rows. */\n  readonly runDirs: readonly string[]\n  /** Destination directory for the bundle (manifest + records + materialized source runs). */\n  readonly outDir: string\n  readonly projectId: string\n  readonly benchmarkId: string\n  /** Repo-relative path of the product's canonical agent profile source. */\n  readonly agentProfilePath: string\n  /** Pass threshold applied when a row carries no explicit `outcome.raw.pass`. Default 0.7. */\n  readonly passThreshold?: number\n  /**\n   * First scenario tag. Defaults to `projectId` with a trailing `-agent`\n   * stripped (`tax-agent` → `tax`), matching the product exporters.\n   */\n  readonly scenarioTagPrefix?: string\n  /** Profile id used when a row has no `agentProfile.profileId`. Defaults to the row's arm id. */\n  readonly fallbackProfileId?: string\n  /** Arm mutable surfaces recorded in the manifest. Defaults to the full superset. */\n  readonly mutableSurfaces?: readonly string[]\n  /**\n   * Copy each run dir into `<outDir>/source-runs/` and record\n   * bundle-relative artifact paths (portable, self-contained). When false,\n   * artifacts keep absolute paths into the original run dirs. Default true.\n   */\n  readonly materializeSourceRuns?: boolean\n  /**\n   * Override split classification for a row. Return undefined to fall back\n   * to the default (`outcome.raw.safety === 1` → safety, then splitTag).\n   */\n  readonly classifySplit?: (record: RunRecord) => ProductBenchmarkSplit | undefined\n  /** Recovers a tool-call count when the row's raw bag carries none (e.g. from turn artifacts). */\n  readonly toolCallFallback?: (record: RunRecord, runDir: string) => number\n  /** Backend version recorded per row. Defaults to the cwd package.json's `@tangle-network/sandbox` range. */\n  readonly backendVersion?: string\n  /**\n   * Explicit substrate versions for the manifest, merged over what the cwd\n   * package.json / node_modules resolve. Use when a substrate package is not\n   * installed where the export runs — the validator refuses an `'unknown'`\n   * version, so provide the real one rather than shipping the sentinel.\n   */\n  readonly substrate?: Partial<ProductBenchmarkManifest['substrate']>\n}\n\nexport interface ProductBenchmarkSingleRunExportOptions\n  extends Omit<ProductBenchmarkExportOptions, 'runDirs'> {\n  readonly runDir: string\n}\n\nexport interface ProductBenchmarkExportResult {\n  readonly manifestPath: string\n  readonly recordsPath: string\n  readonly records: number\n}\n\ninterface ResolvedExportOptions {\n  readonly projectId: string\n  readonly benchmarkId: string\n  readonly agentProfilePath: string\n  readonly passThreshold: number\n  readonly scenarioTagPrefix: string\n  readonly fallbackProfileId?: string\n  readonly mutableSurfaces: readonly string[]\n  readonly classifySplit?: (record: RunRecord) => ProductBenchmarkSplit | undefined\n  readonly toolCallFallback?: (record: RunRecord, runDir: string) => number\n  readonly backendVersion: string\n}\n\n// ── Provenance helpers ───────────────────────────────────────────────\n\nfunction sha256(value: string): string {\n  return `sha256:${createHash('sha256').update(value).digest('hex')}`\n}\n\nfunction safePathPart(value: string): string {\n  return (\n    value\n      .replace(/[^a-zA-Z0-9_.-]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 80) || 'artifact'\n  )\n}\n\nfunction git(args: readonly string[], fallback: string): string {\n  const res = spawnSync('git', args, { cwd: process.cwd(), encoding: 'utf8' })\n  const value = res.status === 0 ? res.stdout.trim() : ''\n  return value.length > 0 ? value : fallback\n}\n\nfunction knownEnv(name: string): string | undefined {\n  const value = process.env[name]?.trim()\n  return value && value !== 'unknown' ? value : undefined\n}\n\nfunction ciBranchName(): string | undefined {\n  const direct =\n    knownEnv('GITHUB_HEAD_REF') ?? knownEnv('GITHUB_REF_NAME') ?? knownEnv('VERCEL_GIT_COMMIT_REF')\n  if (direct) return direct\n  const ref = knownEnv('GITHUB_REF')\n  return ref?.startsWith('refs/heads/') ? ref.slice('refs/heads/'.length) : undefined\n}\n\nfunction repoBranch(): string {\n  const ci = ciBranchName()\n  if (ci) return ci\n  const branch = git(['branch', '--show-current'], '')\n  if (branch) return branch\n  const name = git(['name-rev', '--name-only', '--exclude=tags/*', 'HEAD'], '')\n  if (name && name !== 'undefined') return name\n  return `detached:${git(['rev-parse', '--short', 'HEAD'], 'unknown')}`\n}\n\n/** Repo identity from the exporting process's cwd. `'unknown'` values are flagged by `validateProductBenchmarkRun`. */\nexport function productBenchmarkRepoIdentity(): ProductBenchmarkManifest['repo'] {\n  return {\n    url: git(['config', '--get', 'remote.origin.url'], 'unknown'),\n    commit: git(['rev-parse', 'HEAD'], 'unknown'),\n    branch: repoBranch(),\n  }\n}\n\n/** The declared range from the cwd package.json when present, else the\n *  INSTALLED version read from node_modules (covers transitive installs —\n *  a product that gets sandbox via agent-runtime still records real\n *  provenance), else the `'unknown'` sentinel the validator refuses. */\nfunction packageVersion(name: string): string {\n  const pkg = JSON.parse(readFileSync(resolve('package.json'), 'utf8')) as {\n    name?: string\n    version?: string\n    dependencies?: Record<string, string>\n    devDependencies?: Record<string, string>\n  }\n  if (pkg.name === name && pkg.version) return pkg.version\n  const declared = pkg.dependencies?.[name] ?? pkg.devDependencies?.[name]\n  if (declared) return declared\n  const installed = resolve('node_modules', name, 'package.json')\n  if (existsSync(installed)) {\n    const installedPkg = JSON.parse(readFileSync(installed, 'utf8')) as { version?: string }\n    if (installedPkg.version) return installedPkg.version\n  }\n  return 'unknown'\n}\n\n// ── Run record ingestion ─────────────────────────────────────────────\n\nfunction readRunRecords(path: string): RunRecord[] {\n  const lines = readFileSync(path, 'utf8')\n    .split('\\n')\n    .map((line) => line.trim())\n    .filter(Boolean)\n  return lines.map((line, index) => {\n    let parsed: unknown\n    try {\n      parsed = JSON.parse(line)\n    } catch (err) {\n      throw new ValidationError(\n        `${path}:${index + 1}: ${err instanceof Error ? err.message : String(err)}`,\n        { cause: err },\n      )\n    }\n    return expectRunRecordShape(parsed, `${path}:${index + 1}`)\n  })\n}\n\nfunction expectRunRecordShape(value: unknown, path: string): RunRecord {\n  if (!value || typeof value !== 'object' || Array.isArray(value)) {\n    throw new ValidationError(`${path}: run record must be an object`)\n  }\n  const obj = value as Record<string, unknown>\n  for (const key of ['runId', 'model'] as const) {\n    if (typeof obj[key] !== 'string' || obj[key].length === 0) {\n      throw new ValidationError(`${path}: run record ${key} must be a non-empty string`)\n    }\n  }\n  for (const key of ['costUsd', 'wallMs'] as const) {\n    if (typeof obj[key] !== 'number' || !Number.isFinite(obj[key])) {\n      throw new ValidationError(`${path}: run record ${key} must be a finite number`)\n    }\n  }\n  const tokenUsage = obj.tokenUsage as Record<string, unknown> | undefined\n  if (\n    !tokenUsage ||\n    typeof tokenUsage.input !== 'number' ||\n    typeof tokenUsage.output !== 'number'\n  ) {\n    throw new ValidationError(`${path}: run record tokenUsage.input/output must be numbers`)\n  }\n  const outcome = obj.outcome as Record<string, unknown> | undefined\n  if (!outcome || !outcome.raw || typeof outcome.raw !== 'object') {\n    throw new ValidationError(`${path}: run record outcome.raw must be an object`)\n  }\n  return value as RunRecord\n}\n\n// ── Row mapping ──────────────────────────────────────────────────────\n\nfunction clamp01(value: number): number {\n  return Math.max(0, Math.min(1, value))\n}\n\nfunction splitOf(record: RunRecord, opts: ResolvedExportOptions): ProductBenchmarkSplit {\n  const custom = opts.classifySplit?.(record)\n  if (custom) return custom\n  if ((record.outcome.raw as Record<string, number>).safety === 1) return 'safety'\n  if (record.splitTag === 'holdout') return 'holdout'\n  if (record.splitTag === 'dev') return 'dev'\n  return 'practice'\n}\n\nfunction scoreOf(record: RunRecord): number {\n  const score = record.outcome.holdoutScore ?? record.outcome.searchScore\n  if (typeof score === 'number' && Number.isFinite(score)) return clamp01(score)\n  const rawScore = record.outcome.raw.score ?? record.outcome.raw.composite\n  return typeof rawScore === 'number' && Number.isFinite(rawScore) ? clamp01(rawScore) : 0\n}\n\nfunction rawPassOf(record: RunRecord): boolean | null {\n  const rawPass = (record.outcome.raw as Record<string, unknown>).pass\n  if (typeof rawPass === 'boolean') return rawPass\n  if (typeof rawPass === 'number' && Number.isFinite(rawPass)) return rawPass >= 1\n  return null\n}\n\nfunction passOf(record: RunRecord, score: number, threshold: number): boolean {\n  const rawPass = rawPassOf(record)\n  if (rawPass !== null) return rawPass && !record.failureMode\n  return score >= threshold && !record.failureMode\n}\n\n/** A failed row always carries a failure mode; synthesized when the harness left it empty. */\nfunction failureModeOf(record: RunRecord, score: number, threshold: number): string | null {\n  if (record.failureMode) return record.failureMode\n  const belowThreshold = `quality-below-threshold: ${Math.round(score * 100)}% < ${Math.round(threshold * 100)}%`\n  if (rawPassOf(record) === false) {\n    return score < threshold ? belowThreshold : 'product-pass-failed'\n  }\n  if (passOf(record, score, threshold)) return null\n  return belowThreshold\n}\n\nfunction numericDimensions(record: RunRecord): Record<string, number> {\n  const dimensions: Record<string, number> = {}\n  for (const [key, value] of Object.entries(record.outcome.raw ?? {})) {\n    if (typeof value === 'number' && Number.isFinite(value)) dimensions[key] = value\n  }\n  if (Object.keys(dimensions).length === 0) dimensions.score = scoreOf(record)\n  return dimensions\n}\n\nfunction armIdOf(record: RunRecord): string {\n  if (record.candidateId) return record.candidateId\n  const variant = record.agentProfile?.dimensions?.variant\n  if (typeof variant === 'string' && variant.length > 0) return variant\n  const variantId = record.agentProfile?.dimensions?.variantId\n  if (typeof variantId === 'string' && variantId.length > 0) return variantId\n  return 'production-profile'\n}\n\nfunction backendOf(record: RunRecord): string {\n  const backend = record.agentProfile?.dimensions?.backend\n  return typeof backend === 'string' && backend.length > 0 ? backend : 'unknown'\n}\n\nfunction modelProvider(record: RunRecord): string {\n  const backend = backendOf(record)\n  if (backend === 'cli-bridge') return 'cli-bridge'\n  if (backend === 'sandbox') return 'router'\n  if (record.model.startsWith('router/')) return 'router'\n  return backend === 'unknown' ? 'router' : backend\n}\n\nfunction runtimeResolution(record: RunRecord, opts: ResolvedExportOptions): RuntimeResolution {\n  const reasoningEffort = record.agentProfile?.dimensions?.reasoningLevel\n  return {\n    model: record.agentProfile?.model ?? record.model,\n    harness: record.agentProfile?.harness?.id ?? `${opts.projectId}-canonical-eval`,\n    backend: backendOf(record),\n    ...(typeof reasoningEffort === 'string' && reasoningEffort.length > 0\n      ? { reasoningEffort }\n      : {}),\n  }\n}\n\nfunction sourceProfileHash(record: RunRecord): string {\n  return (\n    record.agentProfile?.sourceProfile?.hash ??\n    record.agentProfile?.cellId ??\n    sha256(\n      JSON.stringify({\n        candidateId: record.candidateId,\n        promptHash: record.promptHash,\n        configHash: record.configHash,\n      }),\n    )\n  )\n}\n\nfunction materializedProfileHash(record: RunRecord, runtime: RuntimeResolution): string {\n  return sha256(\n    JSON.stringify({ sourceProfileHash: sourceProfileHash(record), model: runtime.model }),\n  )\n}\n\nfunction profileIdOf(\n  record: RunRecord,\n  armId: string,\n  runtime: RuntimeResolution,\n  opts: ResolvedExportOptions,\n): string {\n  const base = record.agentProfile?.profileId ?? opts.fallbackProfileId ?? armId\n  const withArm = base.endsWith(`:${armId}`) ? base : `${base}:${armId}`\n  return `${withArm}:${runtime.model}`\n}\n\nfunction toolCallsOf(record: RunRecord, runDir: string, opts: ResolvedExportOptions): number {\n  const raw = record.outcome.raw as Record<string, unknown>\n  const primary = Number(raw.tool_call_count ?? raw.tool_calls ?? raw.toolCalls ?? 0)\n  if (Number.isFinite(primary) && primary > 0) return primary\n  const fallback = opts.toolCallFallback?.(record, runDir)\n  if (fallback !== undefined && Number.isFinite(fallback) && fallback > 0) return fallback\n  return Number.isFinite(primary) ? primary : 0\n}\n\n// ── Artifact layout ──────────────────────────────────────────────────\n\nconst TRACE_CANDIDATES = ['traces.jsonl', 'traces', 'trace', 'trace-store'] as const\nconst RAW_CANDIDATES = ['raws.jsonl', 'raw-events'] as const\nconst SCORE_CANDIDATES = ['scores.json', 'manifest.json'] as const\n\nfunction firstExisting(runDir: string, candidates: readonly string[]): string {\n  return candidates.find((candidate) => existsSync(join(runDir, candidate))) ?? candidates[0]!\n}\n\nfunction relativeArtifacts(runDir: string): ProductBenchmarkRecord['artifacts'] {\n  return {\n    records: 'records.jsonl',\n    traces: firstExisting(runDir, TRACE_CANDIDATES),\n    raws: firstExisting(runDir, RAW_CANDIDATES),\n    scores: firstExisting(runDir, SCORE_CANDIDATES),\n    workspace: existsSync(join(runDir, 'workspace')) ? 'workspace' : '.',\n  }\n}\n\nfunction absoluteArtifacts(runDir: string): ProductBenchmarkRecord['artifacts'] {\n  const rel = relativeArtifacts(runDir)\n  return {\n    records: resolve(runDir, rel.records),\n    traces: resolve(runDir, rel.traces),\n    raws: resolve(runDir, rel.raws),\n    scores: resolve(runDir, rel.scores),\n    workspace: resolve(runDir, rel.workspace),\n  }\n}\n\nfunction materializeRunArtifacts(\n  runDir: string,\n  outDir: string,\n  index: number,\n): ProductBenchmarkRecord['artifacts'] {\n  if (resolve(runDir) === resolve(outDir)) return relativeArtifacts(runDir)\n  const rel = relative(runDir, outDir)\n  if (rel && !rel.startsWith('..') && !isAbsolute(rel)) {\n    throw new ValidationError(`outDir must not be nested inside runDir: ${outDir}`)\n  }\n  const id = `${String(index + 1).padStart(2, '0')}-${safePathPart(basename(runDir))}-${sha256(runDir).slice(7, 19)}`\n  const destRel = join('source-runs', id)\n  const destAbs = join(outDir, destRel)\n  mkdirSync(dirname(destAbs), { recursive: true })\n  cpSync(runDir, destAbs, { recursive: true, force: true })\n  const source = relativeArtifacts(runDir)\n  return {\n    records: join(destRel, source.records),\n    traces: join(destRel, source.traces),\n    raws: join(destRel, source.raws),\n    scores: join(destRel, source.scores),\n    workspace: source.workspace === '.' ? destRel : join(destRel, source.workspace),\n  }\n}\n\nfunction existsArtifact(artifactRoot: string, value: string): boolean {\n  return existsSync(resolve(artifactRoot, value))\n}\n\n// ── Record + manifest construction ───────────────────────────────────\n\n/** Map one RunRecord row to a validated product benchmark record. */\nexport function runRecordToProductBenchmarkRecord(\n  record: RunRecord,\n  runDir: string,\n  artifactRoot: string,\n  artifacts: ProductBenchmarkRecord['artifacts'],\n  options: ProductBenchmarkExportOptions | ProductBenchmarkSingleRunExportOptions,\n): ProductBenchmarkRecord {\n  const opts = resolveOptions(options)\n  const runtime = runtimeResolution(record, opts)\n  const score = scoreOf(record)\n  const armId = armIdOf(record)\n  const inputTokens = record.tokenUsage.input\n  const outputTokens = record.tokenUsage.output\n  const toolCallCount = toolCallsOf(record, runDir, opts)\n  const dimensions = numericDimensions(record)\n  if (!('tool_calls' in dimensions)) dimensions.tool_calls = toolCallCount\n  const product: ProductBenchmarkRecord = {\n    schemaVersion: 1,\n    projectId: opts.projectId,\n    benchmarkId: opts.benchmarkId,\n    runId: record.runId,\n    scenarioId: record.scenarioId ?? record.experimentId,\n    split: splitOf(record, opts),\n    armId,\n    rep: Number(record.seed ?? 0) + 1,\n    agentProfile: {\n      id: profileIdOf(record, armId, runtime, opts),\n      hash: materializedProfileHash(record, runtime),\n      path: opts.agentProfilePath,\n      declared: runtime,\n      resolved: runtime,\n    },\n    model: { provider: modelProvider(record), id: record.model },\n    backend: { kind: backendOf(record), version: opts.backendVersion },\n    outcome: {\n      pass: passOf(record, score, opts.passThreshold),\n      score,\n      dimensions,\n      failureMode: failureModeOf(record, score, opts.passThreshold),\n    },\n    usage: {\n      inputTokens,\n      outputTokens,\n      costUsd: record.costUsd,\n      // Rounded: the bundle contract requires integer milliseconds.\n      wallMs: Math.round(record.wallMs),\n      toolCalls: toolCallCount,\n    },\n    integrity: {\n      realBackend: inputTokens + outputTokens > 0,\n      rawCapture: existsArtifact(artifactRoot, artifacts.raws),\n      traceCapture: existsArtifact(artifactRoot, artifacts.traces),\n      noStubRows: inputTokens + outputTokens > 0,\n      priced: record.costUsd > 0,\n      profileMaterialized: Boolean(record.agentProfile?.cellId),\n    },\n    artifacts,\n  }\n  return validateProductBenchmarkRecord(product)\n}\n\n/** One record per arm id, failing loud when records sharing an arm id\n *  disagree on any attribute the manifest arm carries (profile, model,\n *  backend, harness, reasoning effort) — an id-only dedup would silently\n *  keep one policy and misrepresent the rest. */\nfunction uniqueArmRecords(records: readonly ProductBenchmarkRecord[]): ProductBenchmarkRecord[] {\n  const byArm = new Map<string, ProductBenchmarkRecord>()\n  for (const record of records) {\n    const prior = byArm.get(record.armId)\n    if (!prior) {\n      byArm.set(record.armId, record)\n      continue\n    }\n    const fields = [\n      ['profileId', prior.agentProfile.id, record.agentProfile.id],\n      ['model', prior.agentProfile.resolved.model, record.agentProfile.resolved.model],\n      ['backend', prior.agentProfile.resolved.backend, record.agentProfile.resolved.backend],\n      ['harness', prior.agentProfile.resolved.harness, record.agentProfile.resolved.harness],\n      [\n        'reasoningEffort',\n        prior.agentProfile.resolved.reasoningEffort,\n        record.agentProfile.resolved.reasoningEffort,\n      ],\n    ] as const\n    for (const [name, a, b] of fields) {\n      if (a !== b) {\n        throw new ValidationError(\n          `records for arm '${record.armId}' disagree on ${name} (${String(a)} vs ${String(b)}) — one arm id must map to one policy`,\n        )\n      }\n    }\n  }\n  return [...byArm.values()]\n}\n\n/** One record per scenario id, failing loud on a divergent split — the\n *  manifest scenario carries one split, so two records disagreeing would be\n *  silently misfiled. */\nfunction uniqueScenarioRecords(\n  records: readonly ProductBenchmarkRecord[],\n): ProductBenchmarkRecord[] {\n  const byScenario = new Map<string, ProductBenchmarkRecord>()\n  for (const record of records) {\n    const prior = byScenario.get(record.scenarioId)\n    if (!prior) {\n      byScenario.set(record.scenarioId, record)\n      continue\n    }\n    if (prior.split !== record.split) {\n      throw new ValidationError(\n        `records for scenario '${record.scenarioId}' disagree on split (${prior.split} vs ${record.split})`,\n      )\n    }\n  }\n  return [...byScenario.values()]\n}\n\n/** Derive the bundle manifest from already-normalized records. */\nexport function buildProductBenchmarkManifest(\n  records: readonly ProductBenchmarkRecord[],\n  options: Pick<\n    ProductBenchmarkExportOptions,\n    'outDir' | 'projectId' | 'benchmarkId' | 'scenarioTagPrefix' | 'mutableSurfaces' | 'substrate'\n  >,\n): ProductBenchmarkManifest {\n  if (records.length === 0) {\n    throw new ValidationError('cannot build a product benchmark manifest from zero records')\n  }\n  const scenarioTagPrefix = options.scenarioTagPrefix ?? defaultScenarioTagPrefix(options.projectId)\n  const mutableSurfaces = options.mutableSurfaces ?? productBenchmarkMutableSurfaces\n  const byProfile = new Map<string, ProductBenchmarkManifest['profiles'][number]>()\n  for (const record of records) {\n    byProfile.set(record.agentProfile.id, {\n      id: record.agentProfile.id,\n      profileHash: record.agentProfile.hash,\n      agentProfilePath: record.agentProfile.path,\n    })\n  }\n  const manifest: ProductBenchmarkManifest = {\n    schemaVersion: 1,\n    projectId: options.projectId,\n    benchmarkId: options.benchmarkId,\n    repo: productBenchmarkRepoIdentity(),\n    substrate: {\n      agentEval: packageVersion('@tangle-network/agent-eval'),\n      agentRuntime: packageVersion('@tangle-network/agent-runtime'),\n      agentInterface: packageVersion('@tangle-network/agent-interface'),\n      sandbox: packageVersion('@tangle-network/sandbox'),\n      ...options.substrate,\n    },\n    profiles: [...byProfile.values()],\n    arms: uniqueArmRecords(records).map((record) => ({\n      id: record.armId,\n      profileId: record.agentProfile.id,\n      mutableSurfaces: [...mutableSurfaces],\n      policyAxes: {\n        carrier: record.armId.includes('policy') ? 'resource-file' : 'profile',\n        model: record.agentProfile.resolved.model,\n        backend: record.agentProfile.resolved.backend,\n        harness: record.agentProfile.resolved.harness,\n        ...(record.agentProfile.resolved.reasoningEffort !== undefined\n          ? { reasoningEffort: record.agentProfile.resolved.reasoningEffort }\n          : {}),\n      },\n    })),\n    scenarios: uniqueScenarioRecords(records).map((record) => ({\n      id: record.scenarioId,\n      split: record.split,\n      tags: [scenarioTagPrefix, options.benchmarkId, record.split],\n      sourceAllowedForSynthesis: false,\n    })),\n    budgets: {\n      maxUsd: records.reduce((sum, record) => sum + record.usage.costUsd, 0),\n      maxCells: records.length,\n      maxWallMs: records.reduce((sum, record) => sum + record.usage.wallMs, 0),\n    },\n    expectedArtifactDir: resolve(options.outDir),\n  }\n  return validateProductBenchmarkManifest(manifest)\n}\n\n// ── Entry points ─────────────────────────────────────────────────────\n\n/** Single-run convenience wrapper over `exportProductBenchmarkRuns`. */\nexport function exportProductBenchmark(\n  options: ProductBenchmarkSingleRunExportOptions,\n): ProductBenchmarkExportResult {\n  const { runDir, ...rest } = options\n  return exportProductBenchmarkRuns({ ...rest, runDirs: [runDir] })\n}\n\n/**\n * Export one or more product eval run dirs into a validated product\n * benchmark bundle at `outDir`. Both the manifest and every record are\n * run through the contract validators before anything is written.\n */\nexport function exportProductBenchmarkRuns(\n  options: ProductBenchmarkExportOptions,\n): ProductBenchmarkExportResult {\n  const opts = resolveOptions(options)\n  const outDir = resolve(options.outDir)\n  const runDirs = options.runDirs.map((runDir) => resolve(runDir))\n  if (runDirs.length === 0) throw new ValidationError('export requires at least one run directory')\n  const materialize = options.materializeSourceRuns !== false\n  const rows = runDirs.flatMap((runDir) => {\n    const sourceRecordsPath = join(runDir, 'records.jsonl')\n    if (!existsSync(sourceRecordsPath)) throw new ValidationError(`missing ${sourceRecordsPath}`)\n    const records = readRunRecords(sourceRecordsPath)\n    if (records.length === 0) throw new ValidationError(`${sourceRecordsPath} is empty`)\n    return records.map((record) => ({ record, runDir }))\n  })\n  mkdirSync(outDir, { recursive: true })\n  const artifactsByRunDir = new Map<string, ProductBenchmarkRecord['artifacts']>()\n  for (const [index, runDir] of runDirs.entries()) {\n    artifactsByRunDir.set(\n      runDir,\n      materialize ? materializeRunArtifacts(runDir, outDir, index) : absoluteArtifacts(runDir),\n    )\n  }\n  const artifactRoot = materialize ? outDir : '/'\n  const normalized = rows.map(({ record, runDir }) =>\n    runRecordToProductBenchmarkRecord(\n      record,\n      runDir,\n      artifactRoot,\n      artifactsByRunDir.get(runDir)!,\n      { ...options, runDirs, backendVersion: opts.backendVersion },\n    ),\n  )\n  const manifest = buildProductBenchmarkManifest(normalized, {\n    outDir,\n    projectId: opts.projectId,\n    benchmarkId: opts.benchmarkId,\n    scenarioTagPrefix: opts.scenarioTagPrefix,\n    mutableSurfaces: opts.mutableSurfaces,\n    ...(options.substrate !== undefined ? { substrate: options.substrate } : {}),\n  })\n  const manifestPath = join(outDir, 'product-benchmark-manifest.json')\n  const recordsPath = join(outDir, 'product-benchmark-records.jsonl')\n  writeFileSync(manifestPath, `${JSON.stringify(manifest, null, 2)}\\n`)\n  writeFileSync(recordsPath, `${normalized.map((record) => JSON.stringify(record)).join('\\n')}\\n`)\n  return { manifestPath, recordsPath, records: normalized.length }\n}\n\nfunction defaultScenarioTagPrefix(projectId: string): string {\n  return projectId.replace(/-agent$/, '') || projectId\n}\n\nfunction resolveOptions(\n  options: ProductBenchmarkExportOptions | ProductBenchmarkSingleRunExportOptions,\n): ResolvedExportOptions {\n  return {\n    projectId: options.projectId,\n    benchmarkId: options.benchmarkId,\n    agentProfilePath: options.agentProfilePath,\n    passThreshold: options.passThreshold ?? 0.7,\n    scenarioTagPrefix: options.scenarioTagPrefix ?? defaultScenarioTagPrefix(options.projectId),\n    ...(options.fallbackProfileId !== undefined\n      ? { fallbackProfileId: options.fallbackProfileId }\n      : {}),\n    mutableSurfaces: options.mutableSurfaces ?? productBenchmarkMutableSurfaces,\n    ...(options.classifySplit !== undefined ? { classifySplit: options.classifySplit } : {}),\n    ...(options.toolCallFallback !== undefined\n      ? { toolCallFallback: options.toolCallFallback }\n      : {}),\n    backendVersion: options.backendVersion ?? packageVersion('@tangle-network/sandbox'),\n  }\n}\n"],"mappings":";;;;;AAAA,SAAS,cAAAA,aAAY,gBAAAC,eAAc,gBAAgB;AACnD,SAAS,WAAAC,UAAS,QAAAC,aAAY;;;ACkB9B,SAAS,iBAAiB;AAC1B,SAAS,kBAAkB;AAC3B,SAAS,QAAQ,YAAY,WAAW,cAAc,qBAAqB;AAC3E,SAAS,UAAU,SAAS,YAAY,MAAM,UAAU,eAAe;AAYhE,IAAM,kCAAkC;AAAA,EAC7C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAwEA,SAAS,OAAO,OAAuB;AACrC,SAAO,UAAU,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,CAAC;AACnE;AAEA,SAAS,aAAa,OAAuB;AAC3C,SACE,MACG,QAAQ,qBAAqB,GAAG,EAChC,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEA,SAAS,IAAI,MAAyB,UAA0B;AAC9D,QAAM,MAAM,UAAU,OAAO,MAAM,EAAE,KAAK,QAAQ,IAAI,GAAG,UAAU,OAAO,CAAC;AAC3E,QAAM,QAAQ,IAAI,WAAW,IAAI,IAAI,OAAO,KAAK,IAAI;AACrD,SAAO,MAAM,SAAS,IAAI,QAAQ;AACpC;AAEA,SAAS,SAAS,MAAkC;AAClD,QAAM,QAAQ,QAAQ,IAAI,IAAI,GAAG,KAAK;AACtC,SAAO,SAAS,UAAU,YAAY,QAAQ;AAChD;AAEA,SAAS,eAAmC;AAC1C,QAAM,SACJ,SAAS,iBAAiB,KAAK,SAAS,iBAAiB,KAAK,SAAS,uBAAuB;AAChG,MAAI,OAAQ,QAAO;AACnB,QAAM,MAAM,SAAS,YAAY;AACjC,SAAO,KAAK,WAAW,aAAa,IAAI,IAAI,MAAM,cAAc,MAAM,IAAI;AAC5E;AAEA,SAAS,aAAqB;AAC5B,QAAM,KAAK,aAAa;AACxB,MAAI,GAAI,QAAO;AACf,QAAM,SAAS,IAAI,CAAC,UAAU,gBAAgB,GAAG,EAAE;AACnD,MAAI,OAAQ,QAAO;AACnB,QAAM,OAAO,IAAI,CAAC,YAAY,eAAe,oBAAoB,MAAM,GAAG,EAAE;AAC5E,MAAI,QAAQ,SAAS,YAAa,QAAO;AACzC,SAAO,YAAY,IAAI,CAAC,aAAa,WAAW,MAAM,GAAG,SAAS,CAAC;AACrE;AAGO,SAAS,+BAAiE;AAC/E,SAAO;AAAA,IACL,KAAK,IAAI,CAAC,UAAU,SAAS,mBAAmB,GAAG,SAAS;AAAA,IAC5D,QAAQ,IAAI,CAAC,aAAa,MAAM,GAAG,SAAS;AAAA,IAC5C,QAAQ,WAAW;AAAA,EACrB;AACF;AAMA,SAAS,eAAe,MAAsB;AAC5C,QAAM,MAAM,KAAK,MAAM,aAAa,QAAQ,cAAc,GAAG,MAAM,CAAC;AAMpE,MAAI,IAAI,SAAS,QAAQ,IAAI,QAAS,QAAO,IAAI;AACjD,QAAM,WAAW,IAAI,eAAe,IAAI,KAAK,IAAI,kBAAkB,IAAI;AACvE,MAAI,SAAU,QAAO;AACrB,QAAM,YAAY,QAAQ,gBAAgB,MAAM,cAAc;AAC9D,MAAI,WAAW,SAAS,GAAG;AACzB,UAAM,eAAe,KAAK,MAAM,aAAa,WAAW,MAAM,CAAC;AAC/D,QAAI,aAAa,QAAS,QAAO,aAAa;AAAA,EAChD;AACA,SAAO;AACT;AAIA,SAAS,eAAe,MAA2B;AACjD,QAAM,QAAQ,aAAa,MAAM,MAAM,EACpC,MAAM,IAAI,EACV,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,OAAO,OAAO;AACjB,SAAO,MAAM,IAAI,CAAC,MAAM,UAAU;AAChC,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,IAAI;AAAA,IAC1B,SAAS,KAAK;AACZ,YAAM,IAAI;AAAA,QACR,GAAG,IAAI,IAAI,QAAQ,CAAC,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACzE,EAAE,OAAO,IAAI;AAAA,MACf;AAAA,IACF;AACA,WAAO,qBAAqB,QAAQ,GAAG,IAAI,IAAI,QAAQ,CAAC,EAAE;AAAA,EAC5D,CAAC;AACH;AAEA,SAAS,qBAAqB,OAAgB,MAAyB;AACrE,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AAC/D,UAAM,IAAI,gBAAgB,GAAG,IAAI,gCAAgC;AAAA,EACnE;AACA,QAAM,MAAM;AACZ,aAAW,OAAO,CAAC,SAAS,OAAO,GAAY;AAC7C,QAAI,OAAO,IAAI,GAAG,MAAM,YAAY,IAAI,GAAG,EAAE,WAAW,GAAG;AACzD,YAAM,IAAI,gBAAgB,GAAG,IAAI,gBAAgB,GAAG,6BAA6B;AAAA,IACnF;AAAA,EACF;AACA,aAAW,OAAO,CAAC,WAAW,QAAQ,GAAY;AAChD,QAAI,OAAO,IAAI,GAAG,MAAM,YAAY,CAAC,OAAO,SAAS,IAAI,GAAG,CAAC,GAAG;AAC9D,YAAM,IAAI,gBAAgB,GAAG,IAAI,gBAAgB,GAAG,0BAA0B;AAAA,IAChF;AAAA,EACF;AACA,QAAM,aAAa,IAAI;AACvB,MACE,CAAC,cACD,OAAO,WAAW,UAAU,YAC5B,OAAO,WAAW,WAAW,UAC7B;AACA,UAAM,IAAI,gBAAgB,GAAG,IAAI,sDAAsD;AAAA,EACzF;AACA,QAAM,UAAU,IAAI;AACpB,MAAI,CAAC,WAAW,CAAC,QAAQ,OAAO,OAAO,QAAQ,QAAQ,UAAU;AAC/D,UAAM,IAAI,gBAAgB,GAAG,IAAI,4CAA4C;AAAA,EAC/E;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,OAAuB;AACtC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,CAAC;AACvC;AAEA,SAAS,QAAQ,QAAmB,MAAoD;AACtF,QAAM,SAAS,KAAK,gBAAgB,MAAM;AAC1C,MAAI,OAAQ,QAAO;AACnB,MAAK,OAAO,QAAQ,IAA+B,WAAW,EAAG,QAAO;AACxE,MAAI,OAAO,aAAa,UAAW,QAAO;AAC1C,MAAI,OAAO,aAAa,MAAO,QAAO;AACtC,SAAO;AACT;AAEA,SAAS,QAAQ,QAA2B;AAC1C,QAAM,QAAQ,OAAO,QAAQ,gBAAgB,OAAO,QAAQ;AAC5D,MAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,EAAG,QAAO,QAAQ,KAAK;AAC7E,QAAM,WAAW,OAAO,QAAQ,IAAI,SAAS,OAAO,QAAQ,IAAI;AAChE,SAAO,OAAO,aAAa,YAAY,OAAO,SAAS,QAAQ,IAAI,QAAQ,QAAQ,IAAI;AACzF;AAEA,SAAS,UAAU,QAAmC;AACpD,QAAM,UAAW,OAAO,QAAQ,IAAgC;AAChE,MAAI,OAAO,YAAY,UAAW,QAAO;AACzC,MAAI,OAAO,YAAY,YAAY,OAAO,SAAS,OAAO,EAAG,QAAO,WAAW;AAC/E,SAAO;AACT;AAEA,SAAS,OAAO,QAAmB,OAAe,WAA4B;AAC5E,QAAM,UAAU,UAAU,MAAM;AAChC,MAAI,YAAY,KAAM,QAAO,WAAW,CAAC,OAAO;AAChD,SAAO,SAAS,aAAa,CAAC,OAAO;AACvC;AAGA,SAAS,cAAc,QAAmB,OAAe,WAAkC;AACzF,MAAI,OAAO,YAAa,QAAO,OAAO;AACtC,QAAM,iBAAiB,4BAA4B,KAAK,MAAM,QAAQ,GAAG,CAAC,OAAO,KAAK,MAAM,YAAY,GAAG,CAAC;AAC5G,MAAI,UAAU,MAAM,MAAM,OAAO;AAC/B,WAAO,QAAQ,YAAY,iBAAiB;AAAA,EAC9C;AACA,MAAI,OAAO,QAAQ,OAAO,SAAS,EAAG,QAAO;AAC7C,SAAO;AACT;AAEA,SAAS,kBAAkB,QAA2C;AACpE,QAAM,aAAqC,CAAC;AAC5C,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,QAAQ,OAAO,CAAC,CAAC,GAAG;AACnE,QAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,EAAG,YAAW,GAAG,IAAI;AAAA,EAC7E;AACA,MAAI,OAAO,KAAK,UAAU,EAAE,WAAW,EAAG,YAAW,QAAQ,QAAQ,MAAM;AAC3E,SAAO;AACT;AAEA,SAAS,QAAQ,QAA2B;AAC1C,MAAI,OAAO,YAAa,QAAO,OAAO;AACtC,QAAM,UAAU,OAAO,cAAc,YAAY;AACjD,MAAI,OAAO,YAAY,YAAY,QAAQ,SAAS,EAAG,QAAO;AAC9D,QAAM,YAAY,OAAO,cAAc,YAAY;AACnD,MAAI,OAAO,cAAc,YAAY,UAAU,SAAS,EAAG,QAAO;AAClE,SAAO;AACT;AAEA,SAAS,UAAU,QAA2B;AAC5C,QAAM,UAAU,OAAO,cAAc,YAAY;AACjD,SAAO,OAAO,YAAY,YAAY,QAAQ,SAAS,IAAI,UAAU;AACvE;AAEA,SAAS,cAAc,QAA2B;AAChD,QAAM,UAAU,UAAU,MAAM;AAChC,MAAI,YAAY,aAAc,QAAO;AACrC,MAAI,YAAY,UAAW,QAAO;AAClC,MAAI,OAAO,MAAM,WAAW,SAAS,EAAG,QAAO;AAC/C,SAAO,YAAY,YAAY,WAAW;AAC5C;AAEA,SAAS,kBAAkB,QAAmB,MAAgD;AAC5F,QAAM,kBAAkB,OAAO,cAAc,YAAY;AACzD,SAAO;AAAA,IACL,OAAO,OAAO,cAAc,SAAS,OAAO;AAAA,IAC5C,SAAS,OAAO,cAAc,SAAS,MAAM,GAAG,KAAK,SAAS;AAAA,IAC9D,SAAS,UAAU,MAAM;AAAA,IACzB,GAAI,OAAO,oBAAoB,YAAY,gBAAgB,SAAS,IAChE,EAAE,gBAAgB,IAClB,CAAC;AAAA,EACP;AACF;AAEA,SAAS,kBAAkB,QAA2B;AACpD,SACE,OAAO,cAAc,eAAe,QACpC,OAAO,cAAc,UACrB;AAAA,IACE,KAAK,UAAU;AAAA,MACb,aAAa,OAAO;AAAA,MACpB,YAAY,OAAO;AAAA,MACnB,YAAY,OAAO;AAAA,IACrB,CAAC;AAAA,EACH;AAEJ;AAEA,SAAS,wBAAwB,QAAmB,SAAoC;AACtF,SAAO;AAAA,IACL,KAAK,UAAU,EAAE,mBAAmB,kBAAkB,MAAM,GAAG,OAAO,QAAQ,MAAM,CAAC;AAAA,EACvF;AACF;AAEA,SAAS,YACP,QACA,OACA,SACA,MACQ;AACR,QAAM,OAAO,OAAO,cAAc,aAAa,KAAK,qBAAqB;AACzE,QAAM,UAAU,KAAK,SAAS,IAAI,KAAK,EAAE,IAAI,OAAO,GAAG,IAAI,IAAI,KAAK;AACpE,SAAO,GAAG,OAAO,IAAI,QAAQ,KAAK;AACpC;AAEA,SAAS,YAAY,QAAmB,QAAgB,MAAqC;AAC3F,QAAM,MAAM,OAAO,QAAQ;AAC3B,QAAM,UAAU,OAAO,IAAI,mBAAmB,IAAI,cAAc,IAAI,aAAa,CAAC;AAClF,MAAI,OAAO,SAAS,OAAO,KAAK,UAAU,EAAG,QAAO;AACpD,QAAM,WAAW,KAAK,mBAAmB,QAAQ,MAAM;AACvD,MAAI,aAAa,UAAa,OAAO,SAAS,QAAQ,KAAK,WAAW,EAAG,QAAO;AAChF,SAAO,OAAO,SAAS,OAAO,IAAI,UAAU;AAC9C;AAIA,IAAM,mBAAmB,CAAC,gBAAgB,UAAU,SAAS,aAAa;AAC1E,IAAM,iBAAiB,CAAC,cAAc,YAAY;AAClD,IAAM,mBAAmB,CAAC,eAAe,eAAe;AAExD,SAAS,cAAc,QAAgB,YAAuC;AAC5E,SAAO,WAAW,KAAK,CAAC,cAAc,WAAW,KAAK,QAAQ,SAAS,CAAC,CAAC,KAAK,WAAW,CAAC;AAC5F;AAEA,SAAS,kBAAkB,QAAqD;AAC9E,SAAO;AAAA,IACL,SAAS;AAAA,IACT,QAAQ,cAAc,QAAQ,gBAAgB;AAAA,IAC9C,MAAM,cAAc,QAAQ,cAAc;AAAA,IAC1C,QAAQ,cAAc,QAAQ,gBAAgB;AAAA,IAC9C,WAAW,WAAW,KAAK,QAAQ,WAAW,CAAC,IAAI,cAAc;AAAA,EACnE;AACF;AAEA,SAAS,kBAAkB,QAAqD;AAC9E,QAAM,MAAM,kBAAkB,MAAM;AACpC,SAAO;AAAA,IACL,SAAS,QAAQ,QAAQ,IAAI,OAAO;AAAA,IACpC,QAAQ,QAAQ,QAAQ,IAAI,MAAM;AAAA,IAClC,MAAM,QAAQ,QAAQ,IAAI,IAAI;AAAA,IAC9B,QAAQ,QAAQ,QAAQ,IAAI,MAAM;AAAA,IAClC,WAAW,QAAQ,QAAQ,IAAI,SAAS;AAAA,EAC1C;AACF;AAEA,SAAS,wBACP,QACA,QACA,OACqC;AACrC,MAAI,QAAQ,MAAM,MAAM,QAAQ,MAAM,EAAG,QAAO,kBAAkB,MAAM;AACxE,QAAM,MAAM,SAAS,QAAQ,MAAM;AACnC,MAAI,OAAO,CAAC,IAAI,WAAW,IAAI,KAAK,CAAC,WAAW,GAAG,GAAG;AACpD,UAAM,IAAI,gBAAgB,4CAA4C,MAAM,EAAE;AAAA,EAChF;AACA,QAAM,KAAK,GAAG,OAAO,QAAQ,CAAC,EAAE,SAAS,GAAG,GAAG,CAAC,IAAI,aAAa,SAAS,MAAM,CAAC,CAAC,IAAI,OAAO,MAAM,EAAE,MAAM,GAAG,EAAE,CAAC;AACjH,QAAM,UAAU,KAAK,eAAe,EAAE;AACtC,QAAM,UAAU,KAAK,QAAQ,OAAO;AACpC,YAAU,QAAQ,OAAO,GAAG,EAAE,WAAW,KAAK,CAAC;AAC/C,SAAO,QAAQ,SAAS,EAAE,WAAW,MAAM,OAAO,KAAK,CAAC;AACxD,QAAM,SAAS,kBAAkB,MAAM;AACvC,SAAO;AAAA,IACL,SAAS,KAAK,SAAS,OAAO,OAAO;AAAA,IACrC,QAAQ,KAAK,SAAS,OAAO,MAAM;AAAA,IACnC,MAAM,KAAK,SAAS,OAAO,IAAI;AAAA,IAC/B,QAAQ,KAAK,SAAS,OAAO,MAAM;AAAA,IACnC,WAAW,OAAO,cAAc,MAAM,UAAU,KAAK,SAAS,OAAO,SAAS;AAAA,EAChF;AACF;AAEA,SAAS,eAAe,cAAsB,OAAwB;AACpE,SAAO,WAAW,QAAQ,cAAc,KAAK,CAAC;AAChD;AAKO,SAAS,kCACd,QACA,QACA,cACA,WACA,SACwB;AACxB,QAAM,OAAO,eAAe,OAAO;AACnC,QAAM,UAAU,kBAAkB,QAAQ,IAAI;AAC9C,QAAM,QAAQ,QAAQ,MAAM;AAC5B,QAAM,QAAQ,QAAQ,MAAM;AAC5B,QAAM,cAAc,OAAO,WAAW;AACtC,QAAM,eAAe,OAAO,WAAW;AACvC,QAAM,gBAAgB,YAAY,QAAQ,QAAQ,IAAI;AACtD,QAAM,aAAa,kBAAkB,MAAM;AAC3C,MAAI,EAAE,gBAAgB,YAAa,YAAW,aAAa;AAC3D,QAAM,UAAkC;AAAA,IACtC,eAAe;AAAA,IACf,WAAW,KAAK;AAAA,IAChB,aAAa,KAAK;AAAA,IAClB,OAAO,OAAO;AAAA,IACd,YAAY,OAAO,cAAc,OAAO;AAAA,IACxC,OAAO,QAAQ,QAAQ,IAAI;AAAA,IAC3B;AAAA,IACA,KAAK,OAAO,OAAO,QAAQ,CAAC,IAAI;AAAA,IAChC,cAAc;AAAA,MACZ,IAAI,YAAY,QAAQ,OAAO,SAAS,IAAI;AAAA,MAC5C,MAAM,wBAAwB,QAAQ,OAAO;AAAA,MAC7C,MAAM,KAAK;AAAA,MACX,UAAU;AAAA,MACV,UAAU;AAAA,IACZ;AAAA,IACA,OAAO,EAAE,UAAU,cAAc,MAAM,GAAG,IAAI,OAAO,MAAM;AAAA,IAC3D,SAAS,EAAE,MAAM,UAAU,MAAM,GAAG,SAAS,KAAK,eAAe;AAAA,IACjE,SAAS;AAAA,MACP,MAAM,OAAO,QAAQ,OAAO,KAAK,aAAa;AAAA,MAC9C;AAAA,MACA;AAAA,MACA,aAAa,cAAc,QAAQ,OAAO,KAAK,aAAa;AAAA,IAC9D;AAAA,IACA,OAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,SAAS,OAAO;AAAA;AAAA,MAEhB,QAAQ,KAAK,MAAM,OAAO,MAAM;AAAA,MAChC,WAAW;AAAA,IACb;AAAA,IACA,WAAW;AAAA,MACT,aAAa,cAAc,eAAe;AAAA,MAC1C,YAAY,eAAe,cAAc,UAAU,IAAI;AAAA,MACvD,cAAc,eAAe,cAAc,UAAU,MAAM;AAAA,MAC3D,YAAY,cAAc,eAAe;AAAA,MACzC,QAAQ,OAAO,UAAU;AAAA,MACzB,qBAAqB,QAAQ,OAAO,cAAc,MAAM;AAAA,IAC1D;AAAA,IACA;AAAA,EACF;AACA,SAAO,+BAA+B,OAAO;AAC/C;AAMA,SAAS,iBAAiB,SAAsE;AAC9F,QAAM,QAAQ,oBAAI,IAAoC;AACtD,aAAW,UAAU,SAAS;AAC5B,UAAM,QAAQ,MAAM,IAAI,OAAO,KAAK;AACpC,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,OAAO,OAAO,MAAM;AAC9B;AAAA,IACF;AACA,UAAM,SAAS;AAAA,MACb,CAAC,aAAa,MAAM,aAAa,IAAI,OAAO,aAAa,EAAE;AAAA,MAC3D,CAAC,SAAS,MAAM,aAAa,SAAS,OAAO,OAAO,aAAa,SAAS,KAAK;AAAA,MAC/E,CAAC,WAAW,MAAM,aAAa,SAAS,SAAS,OAAO,aAAa,SAAS,OAAO;AAAA,MACrF,CAAC,WAAW,MAAM,aAAa,SAAS,SAAS,OAAO,aAAa,SAAS,OAAO;AAAA,MACrF;AAAA,QACE;AAAA,QACA,MAAM,aAAa,SAAS;AAAA,QAC5B,OAAO,aAAa,SAAS;AAAA,MAC/B;AAAA,IACF;AACA,eAAW,CAAC,MAAM,GAAG,CAAC,KAAK,QAAQ;AACjC,UAAI,MAAM,GAAG;AACX,cAAM,IAAI;AAAA,UACR,oBAAoB,OAAO,KAAK,iBAAiB,IAAI,KAAK,OAAO,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC;AAAA,QACrF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO,CAAC,GAAG,MAAM,OAAO,CAAC;AAC3B;AAKA,SAAS,sBACP,SAC0B;AAC1B,QAAM,aAAa,oBAAI,IAAoC;AAC3D,aAAW,UAAU,SAAS;AAC5B,UAAM,QAAQ,WAAW,IAAI,OAAO,UAAU;AAC9C,QAAI,CAAC,OAAO;AACV,iBAAW,IAAI,OAAO,YAAY,MAAM;AACxC;AAAA,IACF;AACA,QAAI,MAAM,UAAU,OAAO,OAAO;AAChC,YAAM,IAAI;AAAA,QACR,yBAAyB,OAAO,UAAU,wBAAwB,MAAM,KAAK,OAAO,OAAO,KAAK;AAAA,MAClG;AAAA,IACF;AAAA,EACF;AACA,SAAO,CAAC,GAAG,WAAW,OAAO,CAAC;AAChC;AAGO,SAAS,8BACd,SACA,SAI0B;AAC1B,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI,gBAAgB,6DAA6D;AAAA,EACzF;AACA,QAAM,oBAAoB,QAAQ,qBAAqB,yBAAyB,QAAQ,SAAS;AACjG,QAAM,kBAAkB,QAAQ,mBAAmB;AACnD,QAAM,YAAY,oBAAI,IAA0D;AAChF,aAAW,UAAU,SAAS;AAC5B,cAAU,IAAI,OAAO,aAAa,IAAI;AAAA,MACpC,IAAI,OAAO,aAAa;AAAA,MACxB,aAAa,OAAO,aAAa;AAAA,MACjC,kBAAkB,OAAO,aAAa;AAAA,IACxC,CAAC;AAAA,EACH;AACA,QAAM,WAAqC;AAAA,IACzC,eAAe;AAAA,IACf,WAAW,QAAQ;AAAA,IACnB,aAAa,QAAQ;AAAA,IACrB,MAAM,6BAA6B;AAAA,IACnC,WAAW;AAAA,MACT,WAAW,eAAe,4BAA4B;AAAA,MACtD,cAAc,eAAe,+BAA+B;AAAA,MAC5D,gBAAgB,eAAe,iCAAiC;AAAA,MAChE,SAAS,eAAe,yBAAyB;AAAA,MACjD,GAAG,QAAQ;AAAA,IACb;AAAA,IACA,UAAU,CAAC,GAAG,UAAU,OAAO,CAAC;AAAA,IAChC,MAAM,iBAAiB,OAAO,EAAE,IAAI,CAAC,YAAY;AAAA,MAC/C,IAAI,OAAO;AAAA,MACX,WAAW,OAAO,aAAa;AAAA,MAC/B,iBAAiB,CAAC,GAAG,eAAe;AAAA,MACpC,YAAY;AAAA,QACV,SAAS,OAAO,MAAM,SAAS,QAAQ,IAAI,kBAAkB;AAAA,QAC7D,OAAO,OAAO,aAAa,SAAS;AAAA,QACpC,SAAS,OAAO,aAAa,SAAS;AAAA,QACtC,SAAS,OAAO,aAAa,SAAS;AAAA,QACtC,GAAI,OAAO,aAAa,SAAS,oBAAoB,SACjD,EAAE,iBAAiB,OAAO,aAAa,SAAS,gBAAgB,IAChE,CAAC;AAAA,MACP;AAAA,IACF,EAAE;AAAA,IACF,WAAW,sBAAsB,OAAO,EAAE,IAAI,CAAC,YAAY;AAAA,MACzD,IAAI,OAAO;AAAA,MACX,OAAO,OAAO;AAAA,MACd,MAAM,CAAC,mBAAmB,QAAQ,aAAa,OAAO,KAAK;AAAA,MAC3D,2BAA2B;AAAA,IAC7B,EAAE;AAAA,IACF,SAAS;AAAA,MACP,QAAQ,QAAQ,OAAO,CAACC,MAAK,WAAWA,OAAM,OAAO,MAAM,SAAS,CAAC;AAAA,MACrE,UAAU,QAAQ;AAAA,MAClB,WAAW,QAAQ,OAAO,CAACA,MAAK,WAAWA,OAAM,OAAO,MAAM,QAAQ,CAAC;AAAA,IACzE;AAAA,IACA,qBAAqB,QAAQ,QAAQ,MAAM;AAAA,EAC7C;AACA,SAAO,iCAAiC,QAAQ;AAClD;AAKO,SAAS,uBACd,SAC8B;AAC9B,QAAM,EAAE,QAAQ,GAAG,KAAK,IAAI;AAC5B,SAAO,2BAA2B,EAAE,GAAG,MAAM,SAAS,CAAC,MAAM,EAAE,CAAC;AAClE;AAOO,SAAS,2BACd,SAC8B;AAC9B,QAAM,OAAO,eAAe,OAAO;AACnC,QAAM,SAAS,QAAQ,QAAQ,MAAM;AACrC,QAAM,UAAU,QAAQ,QAAQ,IAAI,CAAC,WAAW,QAAQ,MAAM,CAAC;AAC/D,MAAI,QAAQ,WAAW,EAAG,OAAM,IAAI,gBAAgB,4CAA4C;AAChG,QAAM,cAAc,QAAQ,0BAA0B;AACtD,QAAM,OAAO,QAAQ,QAAQ,CAAC,WAAW;AACvC,UAAM,oBAAoB,KAAK,QAAQ,eAAe;AACtD,QAAI,CAAC,WAAW,iBAAiB,EAAG,OAAM,IAAI,gBAAgB,WAAW,iBAAiB,EAAE;AAC5F,UAAM,UAAU,eAAe,iBAAiB;AAChD,QAAI,QAAQ,WAAW,EAAG,OAAM,IAAI,gBAAgB,GAAG,iBAAiB,WAAW;AACnF,WAAO,QAAQ,IAAI,CAAC,YAAY,EAAE,QAAQ,OAAO,EAAE;AAAA,EACrD,CAAC;AACD,YAAU,QAAQ,EAAE,WAAW,KAAK,CAAC;AACrC,QAAM,oBAAoB,oBAAI,IAAiD;AAC/E,aAAW,CAAC,OAAO,MAAM,KAAK,QAAQ,QAAQ,GAAG;AAC/C,sBAAkB;AAAA,MAChB;AAAA,MACA,cAAc,wBAAwB,QAAQ,QAAQ,KAAK,IAAI,kBAAkB,MAAM;AAAA,IACzF;AAAA,EACF;AACA,QAAM,eAAe,cAAc,SAAS;AAC5C,QAAM,aAAa,KAAK;AAAA,IAAI,CAAC,EAAE,QAAQ,OAAO,MAC5C;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,IAAI,MAAM;AAAA,MAC5B,EAAE,GAAG,SAAS,SAAS,gBAAgB,KAAK,eAAe;AAAA,IAC7D;AAAA,EACF;AACA,QAAM,WAAW,8BAA8B,YAAY;AAAA,IACzD;AAAA,IACA,WAAW,KAAK;AAAA,IAChB,aAAa,KAAK;AAAA,IAClB,mBAAmB,KAAK;AAAA,IACxB,iBAAiB,KAAK;AAAA,IACtB,GAAI,QAAQ,cAAc,SAAY,EAAE,WAAW,QAAQ,UAAU,IAAI,CAAC;AAAA,EAC5E,CAAC;AACD,QAAM,eAAe,KAAK,QAAQ,iCAAiC;AACnE,QAAM,cAAc,KAAK,QAAQ,iCAAiC;AAClE,gBAAc,cAAc,GAAG,KAAK,UAAU,UAAU,MAAM,CAAC,CAAC;AAAA,CAAI;AACpE,gBAAc,aAAa,GAAG,WAAW,IAAI,CAAC,WAAW,KAAK,UAAU,MAAM,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,CAAI;AAC/F,SAAO,EAAE,cAAc,aAAa,SAAS,WAAW,OAAO;AACjE;AAEA,SAAS,yBAAyB,WAA2B;AAC3D,SAAO,UAAU,QAAQ,WAAW,EAAE,KAAK;AAC7C;AAEA,SAAS,eACP,SACuB;AACvB,SAAO;AAAA,IACL,WAAW,QAAQ;AAAA,IACnB,aAAa,QAAQ;AAAA,IACrB,kBAAkB,QAAQ;AAAA,IAC1B,eAAe,QAAQ,iBAAiB;AAAA,IACxC,mBAAmB,QAAQ,qBAAqB,yBAAyB,QAAQ,SAAS;AAAA,IAC1F,GAAI,QAAQ,sBAAsB,SAC9B,EAAE,mBAAmB,QAAQ,kBAAkB,IAC/C,CAAC;AAAA,IACL,iBAAiB,QAAQ,mBAAmB;AAAA,IAC5C,GAAI,QAAQ,kBAAkB,SAAY,EAAE,eAAe,QAAQ,cAAc,IAAI,CAAC;AAAA,IACtF,GAAI,QAAQ,qBAAqB,SAC7B,EAAE,kBAAkB,QAAQ,iBAAiB,IAC7C,CAAC;AAAA,IACL,gBAAgB,QAAQ,kBAAkB,eAAe,yBAAyB;AAAA,EACpF;AACF;;;ADprBO,IAAM,yBAAyB,CAAC,YAAY,OAAO,WAAW,UAAU,UAAU;AA2JzF,SAAS,SAAS,OAAkD;AAClE,SAAO,QAAQ,KAAK,KAAK,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEA,SAAS,KAAK,MAAc,SAAwB;AAClD,QAAM,IAAI,gBAAgB,GAAG,IAAI,KAAK,OAAO,EAAE;AACjD;AAEA,SAAS,oBAAoB,MAAc,KAAqB;AAC9D,MAAI,eAAe,iBAAiB;AAClC,UAAM,IAAI,gBAAgB,GAAG,IAAI,KAAK,IAAI,OAAO,IAAI,EAAE,OAAO,IAAI,CAAC;AAAA,EACrE;AACA,QAAM,IAAI,gBAAgB,GAAG,IAAI,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,IAAI;AAAA,IACxF,OAAO;AAAA,EACT,CAAC;AACH;AAEA,SAAS,aAAa,OAAgB,MAAuC;AAC3E,MAAI,CAAC,SAAS,KAAK,EAAG,MAAK,MAAM,mBAAmB;AACpD,SAAO;AACT;AAEA,SAAS,aAAa,OAAgB,MAAsB;AAC1D,MAAI,OAAO,UAAU,YAAY,MAAM,KAAK,EAAE,WAAW;AACvD,SAAK,MAAM,4BAA4B;AACzC,SAAO;AACT;AAEA,SAAS,cAAc,OAAgB,MAAuB;AAC5D,MAAI,OAAO,UAAU,UAAW,MAAK,MAAM,mBAAmB;AAC9D,SAAO;AACT;AAEA,SAAS,aACP,OACA,MACA,OAAqF,CAAC,GAC9E;AACR,MAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG,MAAK,MAAM,yBAAyB;AAC9F,MAAI,KAAK,WAAW,CAAC,OAAO,UAAU,KAAK,EAAG,MAAK,MAAM,oBAAoB;AAC7E,MAAI,KAAK,QAAQ,UAAa,QAAQ,KAAK,IAAK,MAAK,MAAM,cAAc,KAAK,GAAG,EAAE;AACnF,MAAI,KAAK,QAAQ,UAAa,QAAQ,KAAK,IAAK,MAAK,MAAM,cAAc,KAAK,GAAG,EAAE;AACnF,SAAO;AACT;AAEA,SAAS,kBAAkB,OAAgB,MAAwB;AACjE,MAAI,CAAC,MAAM,QAAQ,KAAK,EAAG,MAAK,MAAM,kBAAkB;AACxD,SAAO,MAAM,IAAI,CAAC,OAAO,UAAU,aAAa,OAAO,GAAG,IAAI,IAAI,KAAK,GAAG,CAAC;AAC7E;AAEA,SAAS,kBAAkB,OAAgB,MAAyC;AAClF,MAAI,CAAC,MAAM,QAAQ,KAAK,EAAG,MAAK,MAAM,kBAAkB;AACxD,SAAO,MAAM,IAAI,CAAC,OAAO,UAAU,aAAa,OAAO,GAAG,IAAI,IAAI,KAAK,GAAG,CAAC;AAC7E;AAEA,SAAS,YAAY,OAAgB,MAAqC;AACxE,QAAM,QAAQ,aAAa,OAAO,IAAI;AACtC,MAAI,CAAC,uBAAuB,SAAS,KAA8B,GAAG;AACpE,SAAK,MAAM,kBAAkB,uBAAuB,KAAK,IAAI,CAAC,EAAE;AAAA,EAClE;AACA,SAAO;AACT;AAEA,SAAS,eAAe,OAAgB,MAAkC;AACxE,MAAI,UAAU,OAAW,QAAO;AAChC,SAAO,aAAa,OAAO,IAAI;AACjC;AAEA,SAAS,iBAAiB,OAAgB,MAAsC;AAC9E,QAAM,MAAM,aAAa,OAAO,IAAI;AACpC,QAAM,MAA8B,CAAC;AACrC,aAAW,CAAC,KAAK,GAAG,KAAK,OAAO,QAAQ,GAAG,EAAG,KAAI,GAAG,IAAI,aAAa,KAAK,GAAG,IAAI,IAAI,GAAG,EAAE;AAC3F,SAAO;AACT;AAEA,SAAS,wBAAwB,OAAgB,MAAiC;AAChF,QAAM,MAAM,aAAa,OAAO,IAAI;AACpC,SAAO;AAAA,IACL,OAAO,aAAa,IAAI,OAAO,GAAG,IAAI,QAAQ;AAAA,IAC9C,SAAS,aAAa,IAAI,SAAS,GAAG,IAAI,UAAU;AAAA,IACpD,SAAS,aAAa,IAAI,SAAS,GAAG,IAAI,UAAU;AAAA,IACpD,GAAI,IAAI,oBAAoB,SACxB,EAAE,iBAAiB,eAAe,IAAI,iBAAiB,GAAG,IAAI,kBAAkB,EAAE,IAClF,CAAC;AAAA,EACP;AACF;AAEO,SAAS,iCAAiC,OAA0C;AACzF,QAAM,MAAM,aAAa,OAAO,UAAU;AAC1C,MAAI,IAAI,kBAAkB,EAAG,MAAK,0BAA0B,WAAW;AACvE,QAAM,OAAO,aAAa,IAAI,MAAM,eAAe;AACnD,QAAM,YAAY,aAAa,IAAI,WAAW,oBAAoB;AAClE,QAAM,WAAW,kBAAkB,IAAI,UAAU,mBAAmB,EAAE,IAAI,CAAC,SAAS,WAAW;AAAA,IAC7F,IAAI,aAAa,QAAQ,IAAI,qBAAqB,KAAK,MAAM;AAAA,IAC7D,aAAa,aAAa,QAAQ,aAAa,qBAAqB,KAAK,eAAe;AAAA,IACxF,kBAAkB;AAAA,MAChB,QAAQ;AAAA,MACR,qBAAqB,KAAK;AAAA,IAC5B;AAAA,EACF,EAAE;AACF,QAAM,OAAO,kBAAkB,IAAI,MAAM,eAAe,EAAE,IAAI,CAAC,KAAK,WAAW;AAAA,IAC7E,IAAI,aAAa,IAAI,IAAI,iBAAiB,KAAK,MAAM;AAAA,IACrD,WAAW,aAAa,IAAI,WAAW,iBAAiB,KAAK,aAAa;AAAA,IAC1E,iBAAiB;AAAA,MACf,IAAI;AAAA,MACJ,iBAAiB,KAAK;AAAA,IACxB;AAAA,IACA,YAAY,aAAa,IAAI,YAAY,iBAAiB,KAAK,cAAc;AAAA,EAC/E,EAAE;AACF,QAAM,YAAY,kBAAkB,IAAI,WAAW,oBAAoB,EAAE;AAAA,IACvE,CAAC,UAAU,WAAW;AAAA,MACpB,IAAI,aAAa,SAAS,IAAI,sBAAsB,KAAK,MAAM;AAAA,MAC/D,OAAO,YAAY,SAAS,OAAO,sBAAsB,KAAK,SAAS;AAAA,MACvE,MAAM,kBAAkB,SAAS,MAAM,sBAAsB,KAAK,QAAQ;AAAA,MAC1E,2BAA2B;AAAA,QACzB,SAAS;AAAA,QACT,sBAAsB,KAAK;AAAA,MAC7B;AAAA,IACF;AAAA,EACF;AACA,QAAM,UAAU,aAAa,IAAI,SAAS,kBAAkB;AAC5D,MAAI,SAAS,WAAW,EAAG,MAAK,qBAAqB,mCAAmC;AACxF,MAAI,KAAK,WAAW,EAAG,MAAK,iBAAiB,+BAA+B;AAC5E,MAAI,UAAU,WAAW,EAAG,MAAK,sBAAsB,oCAAoC;AAC3F;AAAA,IACE,SAAS,IAAI,CAAC,YAAY,QAAQ,EAAE;AAAA,IACpC;AAAA,EACF;AACA;AAAA,IACE,KAAK,IAAI,CAAC,QAAQ,IAAI,EAAE;AAAA,IACxB;AAAA,EACF;AACA;AAAA,IACE,UAAU,IAAI,CAAC,aAAa,SAAS,EAAE;AAAA,IACvC;AAAA,EACF;AACA,QAAM,aAAa,IAAI,IAAI,SAAS,IAAI,CAAC,YAAY,QAAQ,EAAE,CAAC;AAChE,aAAW,OAAO,MAAM;AACtB,QAAI,CAAC,WAAW,IAAI,IAAI,SAAS;AAC/B,WAAK,iBAAiB,IAAI,EAAE,cAAc,mBAAmB,IAAI,SAAS,EAAE;AAAA,EAChF;AACA,SAAO;AAAA,IACL,eAAe;AAAA,IACf,WAAW,aAAa,IAAI,WAAW,oBAAoB;AAAA,IAC3D,aAAa,aAAa,IAAI,aAAa,sBAAsB;AAAA,IACjE,MAAM;AAAA,MACJ,KAAK,aAAa,KAAK,KAAK,mBAAmB;AAAA,MAC/C,QAAQ,aAAa,KAAK,QAAQ,sBAAsB;AAAA,MACxD,QAAQ,aAAa,KAAK,QAAQ,sBAAsB;AAAA,IAC1D;AAAA,IACA,WAAW;AAAA,MACT,WAAW,aAAa,UAAU,WAAW,8BAA8B;AAAA,MAC3E,cAAc,aAAa,UAAU,cAAc,iCAAiC;AAAA,MACpF,gBAAgB,aAAa,UAAU,gBAAgB,mCAAmC;AAAA,MAC1F,SAAS,aAAa,UAAU,SAAS,4BAA4B;AAAA,MACrE,GAAI,UAAU,eAAe,SACzB,EAAE,YAAY,aAAa,UAAU,YAAY,+BAA+B,EAAE,IAClF,CAAC;AAAA,IACP;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS;AAAA,MACP,QAAQ,aAAa,QAAQ,QAAQ,2BAA2B,EAAE,KAAK,EAAE,CAAC;AAAA,MAC1E,UAAU,aAAa,QAAQ,UAAU,6BAA6B;AAAA,QACpE,KAAK;AAAA,QACL,SAAS;AAAA,MACX,CAAC;AAAA,MACD,WAAW,aAAa,QAAQ,WAAW,8BAA8B;AAAA,QACvE,KAAK;AAAA,QACL,SAAS;AAAA,MACX,CAAC;AAAA,IACH;AAAA,IACA,qBAAqB,aAAa,IAAI,qBAAqB,8BAA8B;AAAA,EAC3F;AACF;AAEO,SAAS,+BAA+B,OAAwC;AACrF,QAAM,MAAM,aAAa,OAAO,QAAQ;AACxC,MAAI,IAAI,kBAAkB,EAAG,MAAK,wBAAwB,WAAW;AACrE,QAAM,eAAe,aAAa,IAAI,cAAc,qBAAqB;AACzE,QAAM,QAAQ,aAAa,IAAI,OAAO,cAAc;AACpD,QAAM,UAAU,aAAa,IAAI,SAAS,gBAAgB;AAC1D,QAAM,UAAU,aAAa,IAAI,SAAS,gBAAgB;AAC1D,QAAM,QAAQ,aAAa,IAAI,OAAO,cAAc;AACpD,QAAM,YAAY,aAAa,IAAI,WAAW,kBAAkB;AAChE,QAAM,YAAY,aAAa,IAAI,WAAW,kBAAkB;AAChE,QAAM,SAAiC;AAAA,IACrC,eAAe;AAAA,IACf,WAAW,aAAa,IAAI,WAAW,kBAAkB;AAAA,IACzD,aAAa,aAAa,IAAI,aAAa,oBAAoB;AAAA,IAC/D,OAAO,aAAa,IAAI,OAAO,cAAc;AAAA,IAC7C,YAAY,aAAa,IAAI,YAAY,mBAAmB;AAAA,IAC5D,OAAO,YAAY,IAAI,OAAO,cAAc;AAAA,IAC5C,OAAO,aAAa,IAAI,OAAO,cAAc;AAAA,IAC7C,KAAK,aAAa,IAAI,KAAK,cAAc,EAAE,KAAK,GAAG,SAAS,KAAK,CAAC;AAAA,IAClE,cAAc;AAAA,MACZ,IAAI,aAAa,aAAa,IAAI,wBAAwB;AAAA,MAC1D,MAAM,aAAa,aAAa,MAAM,0BAA0B;AAAA,MAChE,MAAM,aAAa,aAAa,MAAM,0BAA0B;AAAA,MAChE,UAAU,wBAAwB,aAAa,UAAU,8BAA8B;AAAA,MACvF,UAAU,wBAAwB,aAAa,UAAU,8BAA8B;AAAA,IACzF;AAAA,IACA,OAAO;AAAA,MACL,UAAU,aAAa,MAAM,UAAU,uBAAuB;AAAA,MAC9D,IAAI,aAAa,MAAM,IAAI,iBAAiB;AAAA,IAC9C;AAAA,IACA,SAAS;AAAA,MACP,MAAM,aAAa,QAAQ,MAAM,qBAAqB;AAAA,MACtD,SAAS,aAAa,QAAQ,SAAS,wBAAwB;AAAA,IACjE;AAAA,IACA,SAAS;AAAA,MACP,MAAM,cAAc,QAAQ,MAAM,qBAAqB;AAAA,MACvD,OAAO,aAAa,QAAQ,OAAO,wBAAwB,EAAE,KAAK,GAAG,KAAK,EAAE,CAAC;AAAA,MAC7E,YAAY,iBAAiB,QAAQ,YAAY,2BAA2B;AAAA,MAC5E,aACE,QAAQ,gBAAgB,OACpB,OACA,aAAa,QAAQ,aAAa,4BAA4B;AAAA,IACtE;AAAA,IACA,OAAO;AAAA,MACL,aAAa,aAAa,MAAM,aAAa,4BAA4B;AAAA,QACvE,KAAK;AAAA,QACL,SAAS;AAAA,MACX,CAAC;AAAA,MACD,cAAc,aAAa,MAAM,cAAc,6BAA6B;AAAA,QAC1E,KAAK;AAAA,QACL,SAAS;AAAA,MACX,CAAC;AAAA,MACD,SAAS,aAAa,MAAM,SAAS,wBAAwB,EAAE,KAAK,EAAE,CAAC;AAAA,MACvE,QAAQ,aAAa,MAAM,QAAQ,uBAAuB,EAAE,KAAK,GAAG,SAAS,KAAK,CAAC;AAAA,MACnF,WAAW,aAAa,MAAM,WAAW,0BAA0B,EAAE,KAAK,GAAG,SAAS,KAAK,CAAC;AAAA,IAC9F;AAAA,IACA,WAAW;AAAA,MACT,aAAa,cAAc,UAAU,aAAa,8BAA8B;AAAA,MAChF,YAAY,cAAc,UAAU,YAAY,6BAA6B;AAAA,MAC7E,cAAc,cAAc,UAAU,cAAc,+BAA+B;AAAA,MACnF,YAAY,cAAc,UAAU,YAAY,6BAA6B;AAAA,MAC7E,QAAQ,cAAc,UAAU,QAAQ,yBAAyB;AAAA,MACjE,qBAAqB;AAAA,QACnB,UAAU;AAAA,QACV;AAAA,MACF;AAAA,IACF;AAAA,IACA,WAAW;AAAA,MACT,SAAS,aAAa,UAAU,SAAS,0BAA0B;AAAA,MACnE,QAAQ,aAAa,UAAU,QAAQ,yBAAyB;AAAA,MAChE,MAAM,aAAa,UAAU,MAAM,uBAAuB;AAAA,MAC1D,QAAQ,aAAa,UAAU,QAAQ,yBAAyB;AAAA,MAChE,WAAW,aAAa,UAAU,WAAW,4BAA4B;AAAA,IAC3E;AAAA,EACF;AACA,MAAI,OAAO,UAAU,eAAe,OAAO,MAAM,cAAc,OAAO,MAAM,iBAAiB,GAAG;AAC9F,SAAK,gBAAgB,kDAAkD;AAAA,EACzE;AACA,SAAO;AACT;AAEO,SAAS,kCAAkC,QAA0C;AAC1F,QAAM,WAAqB,CAAC;AAC5B,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,SAAS,GAAG;AAC3D,QAAI,CAAC,MAAO,UAAS,KAAK,GAAG,OAAO,KAAK,IAAI,GAAG,QAAQ;AAAA,EAC1D;AACA,SAAO;AACT;AAEO,SAAS,4BAA4B,MAAwC;AAClF,QAAM,QAAQC,cAAa,MAAM,MAAM,EACpC,MAAM,IAAI,EACV,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,OAAO,OAAO;AACjB,QAAM,UAAoC,CAAC;AAC3C,aAAW,CAAC,OAAO,IAAI,KAAK,MAAM,QAAQ,GAAG;AAC3C,QAAI;AACF,cAAQ,KAAK,+BAA+B,KAAK,MAAM,IAAI,CAAC,CAAC;AAAA,IAC/D,SAAS,KAAK;AACZ,0BAAoB,GAAG,IAAI,IAAI,QAAQ,CAAC,IAAI,GAAG;AAAA,IACjD;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,6BAA6B,MAAwC;AACnF,MAAI;AACF,WAAO,iCAAiC,KAAK,MAAMA,cAAa,MAAM,MAAM,CAAC,CAAC;AAAA,EAChF,SAAS,KAAK;AACZ,wBAAoB,MAAM,GAAG;AAAA,EAC/B;AACF;AAEO,SAAS,4BACd,OACkC;AAClC,QAAM,WAAW,6BAA6B,MAAM,YAAY;AAChE,QAAM,UAAU,4BAA4B,MAAM,WAAW;AAC7D,QAAM,uBAAuB,GAAG,SAAS,SAAS,IAAI,SAAS,WAAW;AAC1E,QAAM,oBAAoB,QAAQ,QAAQ,iCAAiC;AAC3E,QAAM,mBACJ,MAAM,mBAAmB,QACrB,CAAC,IACD,QAAQ;AAAA,IAAQ,CAAC,WACf,0BAA0B,QAAQ,MAAM,gBAAgBC,SAAQ,MAAM,WAAW,CAAC;AAAA,EACpF;AACN,aAAW,CAAC,OAAO,MAAM,KAAK,QAAQ,QAAQ,GAAG;AAC/C,UAAM,qBAAqB,GAAG,OAAO,SAAS,IAAI,OAAO,WAAW;AACpE,QAAI,uBAAuB,sBAAsB;AAC/C;AAAA,QACE,WAAW,KAAK;AAAA,QAChB,qBAAqB,kBAAkB,4BAA4B,oBAAoB;AAAA,MACzF;AAAA,IACF;AACA,QAAI,CAAC,SAAS,KAAK,KAAK,CAAC,QAAQ,IAAI,OAAO,OAAO,KAAK;AACtD,WAAK,WAAW,KAAK,WAAW,eAAe,OAAO,KAAK,EAAE;AAC/D,QAAI,CAAC,SAAS,UAAU,KAAK,CAAC,aAAa,SAAS,OAAO,OAAO,UAAU,GAAG;AAC7E,WAAK,WAAW,KAAK,gBAAgB,oBAAoB,OAAO,UAAU,EAAE;AAAA,IAC9E;AAAA,EACF;AACA,QAAM,eAAgB,CAAC,OAAO,UAAU,QAAQ,EAC7C,OAAO,CAAC,QAAQ,SAAS,KAAK,GAAG,EAAE,KAAK,EAAE,WAAW,KAAK,SAAS,KAAK,GAAG,MAAM,SAAS,EAC1F,IAAI,CAAC,QAAQ,iBAAiB,GAAG,EAAE;AACtC,QAAM,oBAAqB,CAAC,aAAa,gBAAgB,kBAAkB,SAAS,EACjF;AAAA,IACC,CAAC,QAAQ,SAAS,UAAU,GAAG,EAAE,KAAK,EAAE,WAAW,KAAK,SAAS,UAAU,GAAG,MAAM;AAAA,EACtF,EACC,IAAI,CAAC,QAAQ,sBAAsB,GAAG,EAAE;AAC3C,SAAO;AAAA,IACL,cAAc,MAAM;AAAA,IACpB,aAAa,MAAM;AAAA,IACnB,SAAS,QAAQ;AAAA,IACjB;AAAA,IACA;AAAA,IACA,UAAU,aAAa,QAAQ,IAAI,CAAC,WAAW,OAAO,SAAS,CAAC;AAAA,IAChE,YAAY,aAAa,QAAQ,IAAI,CAAC,WAAW,OAAO,WAAW,CAAC;AAAA,IACpE,MAAM,aAAa,QAAQ,IAAI,CAAC,WAAW,OAAO,KAAK,CAAC;AAAA,IACxD,WAAW,aAAa,QAAQ,IAAI,CAAC,WAAW,OAAO,UAAU,CAAC;AAAA,IAClE,QAAQ,QAAQ,OAAO,CAAC,WAAW,OAAO,QAAQ,IAAI,EAAE;AAAA,IACxD,QAAQ,QAAQ,OAAO,CAAC,WAAW,CAAC,OAAO,QAAQ,IAAI,EAAE;AAAA,IACzD,aAAa,IAAI,SAAS,CAAC,WAAW,OAAO,MAAM,WAAW;AAAA,IAC9D,cAAc,IAAI,SAAS,CAAC,WAAW,OAAO,MAAM,YAAY;AAAA,IAChE,SAAS,IAAI,SAAS,CAAC,WAAW,OAAO,MAAM,OAAO;AAAA,IACtD,QAAQ,IAAI,SAAS,CAAC,WAAW,OAAO,MAAM,MAAM;AAAA,IACpD;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,0BAA0B,QAAgC,cAAgC;AACjG,QAAM,UAAoB,CAAC;AAC3B,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,SAAS,GAAG;AAC3D,UAAM,OAAO,oBAAoB,OAAO,YAAY;AACpD,QAAI,CAACC,YAAW,IAAI,EAAG,SAAQ,KAAK,GAAG,OAAO,KAAK,IAAI,GAAG,IAAI,KAAK,EAAE;AAAA,EACvE;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,OAAe,cAA8B;AACxE,SAAO,MAAM,WAAW,GAAG,IAAI,QAAQC,MAAK,cAAc,KAAK;AACjE;AAEA,SAAS,aAAa,QAA2B,MAAoB;AACnE,QAAM,OAAO,oBAAI,IAAY;AAC7B,aAAW,SAAS,QAAQ;AAC1B,QAAI,KAAK,IAAI,KAAK,EAAG,MAAK,MAAM,aAAa,KAAK,EAAE;AACpD,SAAK,IAAI,KAAK;AAAA,EAChB;AACF;AAEA,SAAS,aAAa,QAAqC;AACzD,SAAO,CAAC,GAAG,IAAI,IAAI,MAAM,CAAC,EAAE,KAAK;AACnC;AAEA,SAAS,IAAO,OAAqB,IAAiC;AACpE,SAAO,MAAM,OAAO,CAAC,OAAO,SAAS,QAAQ,GAAG,IAAI,GAAG,CAAC;AAC1D;AAEO,SAAS,8BACd,QACsC;AACtC,QAAM,eAAeA,MAAK,QAAQ,iCAAiC;AACnE,QAAM,cAAcA,MAAK,QAAQ,iCAAiC;AAClE,MACED,YAAW,YAAY,KACvB,SAAS,YAAY,EAAE,OAAO,KAC9BA,YAAW,WAAW,KACtB,SAAS,WAAW,EAAE,OAAO,GAC7B;AACA,WAAO,EAAE,cAAc,YAAY;AAAA,EACrC;AACA,SAAO;AACT;AAOO,SAAS,0BAA0B,QAAkD;AAC1F,QAAM,YAAY,8BAA8B,MAAM;AACtD,MAAI,CAAC,WAAW;AACd,SAAK,QAAQ,4EAA4E;AAAA,EAC3F;AACA,QAAM,SAAS,4BAA4B,EAAE,GAAG,WAAW,cAAc,OAAO,CAAC;AACjF,QAAM,WAAW;AAAA,IACf,GAAG,OAAO;AAAA,IACV,GAAG,OAAO;AAAA,IACV,GAAG,OAAO;AAAA,IACV,GAAG,OAAO;AAAA,EACZ;AACA,MAAI,SAAS,SAAS,GAAG;AACvB,SAAK,QAAQ;AAAA,EAAyC,SAAS,KAAK,IAAI,CAAC,EAAE;AAAA,EAC7E;AACA,SAAO;AACT;","names":["existsSync","readFileSync","dirname","join","sum","readFileSync","dirname","existsSync","join"]}

package/dist/{chunk-JU6ZX3CX.js → chunk-SMCACT4Z.js}

@@ -4,7 +4,7 @@ import {
 import {
   aggregateRunScore,
   clamp01
-} from "./chunk-IN3SHQML.js";
+} from "./chunk-LMSQ6EFA.js";
 import {
   LlmClient,
   callLlmJson
@@ -790,4 +790,4 @@ export {
   runSemanticConceptJudge,
   createSemanticConceptJudge
 };
-//# sourceMappingURL=chunk-JU6ZX3CX.js.map
+//# sourceMappingURL=chunk-SMCACT4Z.js.map

package/dist/{chunk-6Q2DYRWV.js → chunk-TCPP4Z5Q.js}

@@ -1,9 +1,9 @@
 import {
   summarizeBackendIntegrity
-} from "./chunk-QIT2XZ4E.js";
+} from "./chunk-VNM52AGA.js";
 import {
   paretoChart
-} from "./chunk-6GT4NI4V.js";
+} from "./chunk-CLS3374R.js";
 import {
   cohensD,
   pairedBootstrap,
@@ -12,10 +12,10 @@ import {
   pearsonR,
   requiredSampleSize,
   spearmanR
-} from "./chunk-HRGUJTER.js";
+} from "./chunk-HYGRFL7C.js";
 import {
   llmSpans
-} from "./chunk-UFSG7ACU.js";
+} from "./chunk-JZXGWLK5.js";
 
 // src/contamination-guard.ts
 function checkCanaries(output, scenarios) {
@@ -822,4 +822,4 @@ export {
   HoldoutAuditor,
   analyzeRuns
 };
-//# sourceMappingURL=chunk-6Q2DYRWV.js.map
+//# sourceMappingURL=chunk-TCPP4Z5Q.js.map

package/dist/{chunk-VDGPPGE3.js → chunk-UMEAR2FI.js}

@@ -1,6 +1,6 @@
 import {
   summaryTable
-} from "./chunk-6GT4NI4V.js";
+} from "./chunk-CLS3374R.js";
 import {
   VerificationError
 } from "./chunk-3BFEG2F6.js";
@@ -574,4 +574,4 @@ export {
   judgeReplayGate,
   renderReleaseReport
 };
-//# sourceMappingURL=chunk-VDGPPGE3.js.map
+//# sourceMappingURL=chunk-UMEAR2FI.js.map

package/dist/{chunk-VDGPPGE3.js.map → chunk-UMEAR2FI.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"sourcesContent":["/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n *   - Dataset manifests prove corpus/version coverage.\n *   - RunRecord rows prove reproducible search/holdout outcomes.\n *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n *   - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport type { RunRecord, RunSplitTag } from './run-record'\n\n/** Severity of an actionable finding attached to a run/trace. */\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\n/** Actionable side-info — a diagnosed finding the loop can act on. */\nexport interface ActionableSideInfo {\n  /** Stable expectation/check id when available. */\n  expectationId?: string\n  /** Human-readable diagnosis of what happened. */\n  message: string\n  severity?: AsiSeverity\n  /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n  evidence?: string\n  /** Prompt/tool/context surface likely responsible. */\n  responsibleSurface?: string\n  /** Suggested fix in natural language. */\n  suggestion?: string\n  /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n  matched?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n  | 'corpus'\n  | 'quality'\n  | 'generalization'\n  | 'diagnostics'\n  | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n  scenarioId: string\n  candidateId?: string\n  split?: RunSplitTag\n  score?: number\n  ok?: boolean\n  turnCount?: number\n  costUsd?: number\n  durationMs?: number\n  failureMode?: string\n  asi?: ActionableSideInfo[]\n  metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n  /** Require a Dataset manifest or explicit scenarios. Default true. */\n  requireCorpus?: boolean\n  minScenarioCount?: number\n  minSearchRuns?: number\n  minHoldoutRuns?: number\n  /** Require at least one holdout scenario/run. Default true. */\n  requireHoldout?: boolean\n  minPassRate?: number\n  minMeanScore?: number\n  /** Search mean may exceed holdout mean by at most this much. */\n  maxOverfitGap?: number\n  maxMeanCostUsd?: number\n  maxP95WallMs?: number\n  /** Low-score/failed rows must carry ASI. Default true. */\n  requireAsiForFailures?: boolean\n  /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n  failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n  target: string\n  candidateId?: string\n  baselineId?: string\n  dataset?: DatasetManifest\n  scenarios?: readonly DatasetScenario[]\n  runs?: readonly RunRecord[]\n  traces?: readonly ReleaseTraceEvidence[]\n  gateDecision?: GateDecision | null\n  thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n  name: ReleaseConfidenceAxisName\n  status: ReleaseConfidenceStatus\n  score: number\n  detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n  axis: ReleaseConfidenceAxisName\n  severity: 'critical' | 'warning'\n  code: string\n  detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n  scenarioCount: number\n  searchRuns: number\n  holdoutRuns: number\n  passRate: number\n  meanScore: number\n  searchMeanScore: number\n  holdoutMeanScore: number\n  overfitGap: number\n  meanCostUsd: number\n  p95WallMs: number\n  failedRows: number\n  failuresWithAsi: number\n  singleShotTraces: number\n  multiShotTraces: number\n  splitCounts: Record<DatasetSplit, number>\n  domainCounts: Record<string, number>\n  failureModeCounts: Record<string, number>\n  responsibleSurfaceCounts: Record<string, number>\n}\n\nexport interface ReleaseConfidenceScorecard {\n  target: string\n  candidateId: string | null\n  baselineId: string | null\n  status: ReleaseConfidenceStatus\n  promote: boolean\n  axes: ReleaseConfidenceAxis[]\n  issues: ReleaseConfidenceIssue[]\n  metrics: ReleaseConfidenceMetrics\n  dataset: DatasetManifest | null\n  gateDecision: GateDecision | null\n  summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n  requireCorpus: true,\n  minScenarioCount: 1,\n  minSearchRuns: 1,\n  minHoldoutRuns: 1,\n  requireHoldout: true,\n  minPassRate: 0.8,\n  minMeanScore: 0.7,\n  maxOverfitGap: 0.15,\n  maxMeanCostUsd: Number.POSITIVE_INFINITY,\n  maxP95WallMs: Number.POSITIVE_INFINITY,\n  requireAsiForFailures: true,\n  failureScoreThreshold: 0.5,\n}\n\nexport function evaluateReleaseConfidence(\n  input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n  const candidateId = input.candidateId ?? null\n  const runs = filterCandidate(input.runs ?? [], candidateId, input.baselineId)\n  const traces = filterTraceCandidate(input.traces ?? [], candidateId, input.baselineId)\n  const scenarios = input.scenarios ?? []\n  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n  const searchScores = scoresFor(runs, 'search')\n  const holdoutScores = scoresFor(runs, 'holdout')\n  const allScores = [...searchScores, ...holdoutScores]\n  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n  const scoreUniverse = allScores.length > 0 ? allScores : traceScores\n  const searchRuns = runs.filter((r) => r.splitTag === 'search').length\n  const holdoutRuns = runs.filter((r) => r.splitTag === 'holdout').length\n  const searchMeanScore = mean(searchScores)\n  const holdoutMeanScore = mean(holdoutScores)\n  const metrics: ReleaseConfidenceMetrics = {\n    scenarioCount,\n    searchRuns,\n    holdoutRuns,\n    passRate: passRate(runs, traces, thresholds.failureScoreThreshold),\n    meanScore: mean(scoreUniverse),\n    searchMeanScore,\n    holdoutMeanScore,\n    overfitGap: safeDiff(searchMeanScore, holdoutMeanScore),\n    meanCostUsd: mean([\n      ...runs.map((r) => r.costUsd),\n      ...traces.map((t) => t.costUsd).filter(isFiniteNumber),\n    ]),\n    p95WallMs: percentile(\n      [...runs.map((r) => r.wallMs), ...traces.map((t) => t.durationMs).filter(isFiniteNumber)],\n      0.95,\n    ),\n    failedRows: failedRows(runs, traces, thresholds.failureScoreThreshold).length,\n    failuresWithAsi: failedRows(runs, traces, thresholds.failureScoreThreshold).filter(\n      (row) => row.hasAsi,\n    ).length,\n    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n    splitCounts,\n    domainCounts: countDomains(scenarios),\n    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),\n    responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n  }\n\n  const issues: ReleaseConfidenceIssue[] = []\n  checkCorpus(input, thresholds, metrics, issues)\n  checkQuality(thresholds, metrics, issues)\n  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n  checkDiagnostics(thresholds, metrics, issues)\n  checkEfficiency(thresholds, metrics, issues)\n\n  const axes = buildAxes(metrics, thresholds, input.gateDecision ?? null, issues)\n  const status = issues.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : issues.length > 0\n      ? 'warn'\n      : 'pass'\n\n  return {\n    target: input.target,\n    candidateId,\n    baselineId: input.baselineId ?? null,\n    status,\n    promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n    axes,\n    issues,\n    metrics,\n    dataset: input.dataset ?? null,\n    gateDecision: input.gateDecision ?? null,\n    summary: renderSummary(input.target, status, metrics, issues),\n  }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n  const scorecard = evaluateReleaseConfidence(input)\n  if (scorecard.status === 'fail') {\n    throw new VerificationError(scorecard.summary)\n  }\n  return scorecard\n}\n\nfunction filterCandidate(\n  runs: readonly RunRecord[],\n  candidateId: string | null,\n  baselineId?: string,\n): RunRecord[] {\n  if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n  return [...runs]\n}\n\nfunction filterTraceCandidate(\n  traces: readonly ReleaseTraceEvidence[],\n  candidateId: string | null,\n  baselineId?: string,\n): ReleaseTraceEvidence[] {\n  if (candidateId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n  if (baselineId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n  return [...traces]\n}\n\nfunction checkCorpus(\n  input: ReleaseConfidenceInput,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_corpus',\n      detail: 'No Dataset manifest or scenarios supplied.',\n    })\n  }\n  if (metrics.scenarioCount < thresholds.minScenarioCount) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'few_scenarios',\n      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n    })\n  }\n  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_holdout_split',\n      detail: 'Corpus has no holdout scenarios.',\n    })\n  }\n}\n\nfunction checkQuality(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.searchRuns < thresholds.minSearchRuns) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'few_search_runs',\n      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n    })\n  }\n  if (metrics.passRate < thresholds.minPassRate) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_pass_rate',\n      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n    })\n  }\n  if (metrics.meanScore < thresholds.minMeanScore) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_mean_score',\n      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n    })\n  }\n}\n\nfunction checkGeneralization(\n  gateDecision: GateDecision | null,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'few_holdout_runs',\n      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n    })\n  }\n  if (Number.isFinite(metrics.overfitGap) && metrics.overfitGap > thresholds.maxOverfitGap) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'overfit_gap',\n      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n    })\n  }\n  if (gateDecision && !gateDecision.promote) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n      detail: gateDecision.reason,\n    })\n  }\n}\n\nfunction checkDiagnostics(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (!thresholds.requireAsiForFailures) return\n  if (metrics.failedRows > metrics.failuresWithAsi) {\n    issues.push({\n      axis: 'diagnostics',\n      severity: 'critical',\n      code: 'missing_failure_asi',\n      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n    })\n  }\n}\n\nfunction checkEfficiency(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'cost_budget',\n      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n    })\n  }\n  if (metrics.p95WallMs > thresholds.maxP95WallMs) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'latency_budget',\n      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n    })\n  }\n}\n\nfunction buildAxes(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  gateDecision: GateDecision | null,\n  issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n  return [\n    axis(\n      'corpus',\n      issues,\n      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n    ),\n    axis(\n      'quality',\n      issues,\n      Math.min(metrics.passRate, metrics.meanScore),\n      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n    ),\n    axis(\n      'generalization',\n      issues,\n      gateDecision && !gateDecision.promote\n        ? 0\n        : gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n    ),\n    axis(\n      'diagnostics',\n      issues,\n      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n    ),\n    axis(\n      'efficiency',\n      issues,\n      efficiencyScore(metrics, thresholds),\n      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n    ),\n  ]\n}\n\nfunction axis(\n  name: ReleaseConfidenceAxisName,\n  issues: ReleaseConfidenceIssue[],\n  score: number,\n  detail: string,\n): ReleaseConfidenceAxis {\n  const own = issues.filter((i) => i.axis === name)\n  const status = own.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : own.length > 0\n      ? 'warn'\n      : 'pass'\n  return { name, status, score: bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n  const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n  for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n  return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const scenario of scenarios) {\n    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n    out[domain] = (out[domain] ?? 0) + 1\n  }\n  return out\n}\n\nfunction countFailureModes(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const mode = run.failureMode ?? 'low_score'\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      const mode = trace.failureMode ?? (trace.ok === false ? 'not_ok' : 'low_score')\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const trace of traces) {\n    for (const asi of trace.asi ?? []) {\n      const surface = asi.responsibleSurface ?? 'unknown'\n      out[surface] = (out[surface] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction failedRows(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Array<{ hasAsi: boolean }> {\n  const out: Array<{ hasAsi: boolean }> = []\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const asiMetric = run.outcome.raw.asi\n      out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n    }\n  }\n  return out\n}\n\nfunction passRate(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): number {\n  const outcomes = [\n    ...runs.map((run) => {\n      const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n      return !run.failureMode && score !== undefined && score >= threshold\n    }),\n    ...traces.map(\n      (trace) => trace.ok !== false && (trace.score === undefined || trace.score >= threshold),\n    ),\n  ]\n  if (outcomes.length === 0) return 0\n  return outcomes.filter(Boolean).length / outcomes.length\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n  return runs\n    .filter((run) => run.splitTag === split)\n    .map((run) => (split === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore))\n    .filter(isFiniteNumber)\n}\n\nfunction mean(xs: readonly number[]): number {\n  if (xs.length === 0) return Number.NaN\n  return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentile(xs: readonly number[], p: number): number {\n  if (xs.length === 0) return Number.NaN\n  const sorted = [...xs].sort((a, b) => a - b)\n  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction safeDiff(a: number, b: number): number {\n  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN\n  return a - b\n}\n\nfunction gapScore(gap: number, maxGap: number): number {\n  if (!Number.isFinite(gap)) return 0\n  if (maxGap <= 0) return gap <= 0 ? 1 : 0\n  return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n): number {\n  const cost =\n    Number.isFinite(thresholds.maxMeanCostUsd) && Number.isFinite(metrics.meanCostUsd)\n      ? bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n      : 1\n  const latency =\n    Number.isFinite(thresholds.maxP95WallMs) && Number.isFinite(metrics.p95WallMs)\n      ? bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n      : 1\n  return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n  target: string,\n  status: ReleaseConfidenceStatus,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): string {\n  const prefix = `release confidence ${status}: ${target}`\n  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n  if (issues.length === 0) return `${prefix}; ${metricText}`\n  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n *   - Bootstrap mean baseline vs candidate (1k resamples).\n *   - Compute the delta distribution; pass if the lower CI bound > 0.\n *   - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)\n *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)\n *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n  baselineMean: number\n  candidateMean: number\n  /** candidateMean - baselineMean, point estimate. */\n  delta: number\n  /** Lower bound of the (1 - alpha) CI on the delta. */\n  ciLower: number\n  /** Upper bound of the (1 - alpha) CI on the delta. */\n  ciUpper: number\n  /** Number of bootstrap resamples used. */\n  iterations: number\n  alpha: number\n  verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n  /** Confidence level alpha (default 0.05 → 95% CI). */\n  alpha?: number\n  /** Number of resamples (default 1000). */\n  iterations?: number\n  /**\n   * Minimum total samples (baseline + candidate) below which we always\n   * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n   * Default 6 (combined).\n   */\n  minTotalSamples?: number\n  /** RNG seed for reproducibility. Default: Math.random. */\n  seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n  baseline: number[],\n  candidate: number[],\n  options: BootstrapOptions = {},\n): BootstrapResult {\n  const alpha = options.alpha ?? 0.05\n  const iterations = options.iterations ?? 1000\n  const minTotal = options.minTotalSamples ?? 6\n  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n  const baselineMean = mean(baseline)\n  const candidateMean = mean(candidate)\n  const delta = candidateMean - baselineMean\n\n  if (\n    baseline.length + candidate.length < minTotal ||\n    baseline.length === 0 ||\n    candidate.length === 0\n  ) {\n    return {\n      baselineMean,\n      candidateMean,\n      delta,\n      ciLower: -Infinity,\n      ciUpper: Infinity,\n      iterations: 0,\n      alpha,\n      verdict: 'INCONCLUSIVE',\n    }\n  }\n\n  const deltas: number[] = new Array(iterations)\n  for (let i = 0; i < iterations; i++) {\n    const bResample = resample(baseline, rng)\n    const cResample = resample(candidate, rng)\n    deltas[i] = mean(cResample) - mean(bResample)\n  }\n  deltas.sort((a, b) => a - b)\n  const lowerIdx = Math.floor((alpha / 2) * iterations)\n  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n  const ciLower = deltas[Math.max(0, lowerIdx)]!\n  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n  let verdict: Verdict\n  if (ciLower > 0) verdict = 'ADVANCE'\n  else if (ciUpper < 0) verdict = 'REVERT'\n  else if (delta >= 0) verdict = 'KEEP'\n  else verdict = 'INCONCLUSIVE'\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ciLower,\n    ciUpper,\n    iterations,\n    alpha,\n    verdict,\n  }\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  let s = 0\n  for (const x of xs) s += x\n  return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n  const out = new Array(xs.length)\n  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n  return out\n}\n\n/** Mulberry32 — fast deterministic PRNG. Stable across runs given the same seed. */\nfunction mulberry32(seed: number): () => number {\n  let t = seed >>> 0\n  return () => {\n    t += 0x6d2b79f5\n    let r = t\n    r = Math.imul(r ^ (r >>> 15), r | 1)\n    r ^= r + Math.imul(r ^ (r >>> 7), r | 61)\n    return ((r ^ (r >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n  let h = 2166136261\n  for (const x of [...a, ...b]) {\n    const view = new Float64Array([x])\n    const bytes = new Uint8Array(view.buffer)\n    for (const byte of bytes) {\n      h ^= byte\n      h = Math.imul(h, 16777619)\n    }\n  }\n  return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n  baselineOutputs: TOutput[]\n  candidateOutputs: TOutput[]\n  /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n  judge: (output: TOutput) => Promise<number> | number\n  alpha?: number\n  iterations?: number\n  /** RNG seed for reproducibility. */\n  seed?: number\n  /** Maximum concurrent judge calls. Default 4. */\n  judgeConcurrency?: number\n}\n\nexport async function judgeReplayGate<TOutput>(\n  args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n  const concurrency = args.judgeConcurrency ?? 4\n  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n  const ci = bootstrapCi(baselineScores, candidateScores, {\n    ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n    ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n    ...(args.seed !== undefined ? { seed: args.seed } : {}),\n  })\n  return {\n    ...ci,\n    baselineSamples: baselineScores.length,\n    candidateSamples: candidateScores.length,\n  }\n}\n\nasync function scoreAll<TOutput>(\n  outputs: TOutput[],\n  judge: (output: TOutput) => Promise<number> | number,\n  concurrency: number,\n): Promise<number[]> {\n  const results: number[] = new Array(outputs.length)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= outputs.length) return\n      const v = await judge(outputs[i]!)\n      results[i] = Number.isFinite(v) ? v : 0\n    }\n  }\n  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n  return results\n}\n","import type { ReleaseConfidenceScorecard } from './release-confidence'\nimport type { RunRecord } from './run-record'\nimport { summaryTable } from './summary-report'\n\nexport interface RenderReleaseReportOptions {\n  title?: string\n  runs?: readonly RunRecord[]\n  comparator?: string\n  traceAnalystFindings?: readonly string[]\n  nextActions?: readonly string[]\n}\n\nexport function renderReleaseReport(\n  scorecard: ReleaseConfidenceScorecard,\n  options: RenderReleaseReportOptions = {},\n): string {\n  const title = options.title ?? `Release Report: ${scorecard.target}`\n  const lines: string[] = []\n  lines.push(`# ${title}`)\n  lines.push('')\n  lines.push(`Status: **${scorecard.status.toUpperCase()}**`)\n  lines.push(`Promote: **${scorecard.promote ? 'yes' : 'no'}**`)\n  if (scorecard.candidateId) lines.push(`Candidate: \\`${scorecard.candidateId}\\``)\n  if (scorecard.baselineId) lines.push(`Baseline: \\`${scorecard.baselineId}\\``)\n  lines.push('')\n  lines.push(scorecard.summary)\n  lines.push('')\n\n  lines.push('## Metrics')\n  lines.push('')\n  lines.push('| Metric | Value |')\n  lines.push('|---|---:|')\n  lines.push(`| Scenarios | ${scorecard.metrics.scenarioCount} |`)\n  lines.push(`| Search runs | ${scorecard.metrics.searchRuns} |`)\n  lines.push(`| Holdout runs | ${scorecard.metrics.holdoutRuns} |`)\n  lines.push(`| Pass rate | ${pct(scorecard.metrics.passRate)} |`)\n  lines.push(`| Mean score | ${num(scorecard.metrics.meanScore)} |`)\n  lines.push(`| Search mean | ${num(scorecard.metrics.searchMeanScore)} |`)\n  lines.push(`| Holdout mean | ${num(scorecard.metrics.holdoutMeanScore)} |`)\n  lines.push(`| Overfit gap | ${num(scorecard.metrics.overfitGap)} |`)\n  lines.push(`| Mean cost | $${num(scorecard.metrics.meanCostUsd)} |`)\n  lines.push(`| p95 wall time | ${Math.round(scorecard.metrics.p95WallMs)} ms |`)\n  lines.push('')\n\n  if (scorecard.issues.length > 0) {\n    lines.push('## Issues')\n    lines.push('')\n    for (const issue of scorecard.issues) {\n      lines.push(`- **${issue.severity}** \\`${issue.code}\\` (${issue.axis}): ${issue.detail}`)\n    }\n    lines.push('')\n  }\n\n  const surfaces = entries(scorecard.metrics.responsibleSurfaceCounts)\n  if (surfaces.length > 0) {\n    lines.push('## Responsible Surfaces')\n    lines.push('')\n    for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`)\n    lines.push('')\n  }\n\n  const failures = entries(scorecard.metrics.failureModeCounts)\n  if (failures.length > 0) {\n    lines.push('## Failure Modes')\n    lines.push('')\n    for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`)\n    lines.push('')\n  }\n\n  if (options.runs && options.runs.length > 0) {\n    lines.push('## Run Summary')\n    lines.push('')\n    lines.push(\n      summaryTable([...options.runs], {\n        comparator: options.comparator ?? scorecard.baselineId ?? undefined,\n        split: 'holdout',\n      }).markdown,\n    )\n    lines.push('')\n  }\n\n  if (options.traceAnalystFindings && options.traceAnalystFindings.length > 0) {\n    lines.push('## TraceAnalyst Findings')\n    lines.push('')\n    for (const finding of options.traceAnalystFindings) lines.push(`- ${finding}`)\n    lines.push('')\n  }\n\n  const nextActions = options.nextActions ?? defaultNextActions(scorecard)\n  if (nextActions.length > 0) {\n    lines.push('## Next Actions')\n    lines.push('')\n    for (const action of nextActions) lines.push(`- ${action}`)\n    lines.push('')\n  }\n\n  return `${lines.join('\\n').trimEnd()}\\n`\n}\n\nfunction defaultNextActions(scorecard: ReleaseConfidenceScorecard): string[] {\n  if (scorecard.promote) return ['Promote the candidate and keep canaries enabled.']\n  return scorecard.issues\n    .filter((issue) => issue.severity === 'critical')\n    .map((issue) => `Resolve ${issue.code}: ${issue.detail}`)\n}\n\nfunction entries(values: Record<string, number>): Array<[string, number]> {\n  return Object.entries(values)\n    .filter(([, count]) => count > 0)\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n}\n\nfunction pct(value: number): string {\n  return Number.isFinite(value) ? `${(value * 100).toFixed(1)}%` : 'n/a'\n}\n\nfunction num(value: number): string {\n  return Number.isFinite(value) ? value.toFixed(3) : 'n/a'\n}\n"],"mappings":";;;;;;;;AA+IA,IAAM,qBAA4D;AAAA,EAChE,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,aAAa;AAAA,EACb,cAAc;AAAA,EACd,eAAe;AAAA,EACf,gBAAgB,OAAO;AAAA,EACvB,cAAc,OAAO;AAAA,EACrB,uBAAuB;AAAA,EACvB,uBAAuB;AACzB;AAEO,SAAS,0BACd,OAC4B;AAC5B,QAAM,aAAa,EAAE,GAAG,oBAAoB,GAAG,MAAM,WAAW;AAChE,QAAM,cAAc,MAAM,eAAe;AACzC,QAAM,OAAO,gBAAgB,MAAM,QAAQ,CAAC,GAAG,aAAa,MAAM,UAAU;AAC5E,QAAM,SAAS,qBAAqB,MAAM,UAAU,CAAC,GAAG,aAAa,MAAM,UAAU;AACrF,QAAM,YAAY,MAAM,aAAa,CAAC;AACtC,QAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;AAChE,QAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;AAC/E,QAAM,eAAe,UAAU,MAAM,QAAQ;AAC7C,QAAM,gBAAgB,UAAU,MAAM,SAAS;AAC/C,QAAM,YAAY,CAAC,GAAG,cAAc,GAAG,aAAa;AACpD,QAAM,cAAc,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,OAAO,cAAc;AACpE,QAAM,gBAAgB,UAAU,SAAS,IAAI,YAAY;AACzD,QAAM,aAAa,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ,EAAE;AAC/D,QAAM,cAAc,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,SAAS,EAAE;AACjE,QAAM,kBAAkB,KAAK,YAAY;AACzC,QAAM,mBAAmB,KAAK,aAAa;AAC3C,QAAM,UAAoC;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA,UAAU,SAAS,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACjE,WAAW,KAAK,aAAa;AAAA,IAC7B;AAAA,IACA;AAAA,IACA,YAAY,SAAS,iBAAiB,gBAAgB;AAAA,IACtD,aAAa,KAAK;AAAA,MAChB,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,OAAO;AAAA,MAC5B,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,cAAc;AAAA,IACvD,CAAC;AAAA,IACD,WAAW;AAAA,MACT,CAAC,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,OAAO,cAAc,CAAC;AAAA,MACxF;AAAA,IACF;AAAA,IACA,YAAY,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,IACvE,iBAAiB,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,MAC1E,CAAC,QAAQ,IAAI;AAAA,IACf,EAAE;AAAA,IACF,kBAAkB,OAAO,OAAO,CAAC,MAAM,EAAE,cAAc,CAAC,EAAE;AAAA,IAC1D,iBAAiB,OAAO,OAAO,CAAC,OAAO,EAAE,aAAa,KAAK,CAAC,EAAE;AAAA,IAC9D;AAAA,IACA,cAAc,aAAa,SAAS;AAAA,IACpC,mBAAmB,kBAAkB,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACnF,0BAA0B,yBAAyB,MAAM;AAAA,EAC3D;AAEA,QAAM,SAAmC,CAAC;AAC1C,cAAY,OAAO,YAAY,SAAS,MAAM;AAC9C,eAAa,YAAY,SAAS,MAAM;AACxC,sBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;AAC3E,mBAAiB,YAAY,SAAS,MAAM;AAC5C,kBAAgB,YAAY,SAAS,MAAM;AAE3C,QAAM,OAAO,UAAU,SAAS,YAAY,MAAM,gBAAgB,MAAM,MAAM;AAC9E,QAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;AAEN,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd;AAAA,IACA,YAAY,MAAM,cAAc;AAAA,IAChC;AAAA,IACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;AAAA,IACjF;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,MAAM,WAAW;AAAA,IAC1B,cAAc,MAAM,gBAAgB;AAAA,IACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;AAAA,EAC9D;AACF;AAEO,SAAS,wBAAwB,OAA2D;AACjG,QAAM,YAAY,0BAA0B,KAAK;AACjD,MAAI,UAAU,WAAW,QAAQ;AAC/B,UAAM,IAAI,kBAAkB,UAAU,OAAO;AAAA,EAC/C;AACA,SAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;AACb,MAAI,YAAa,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,WAAW;AACxE,MAAI,WAAY,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAU;AACtE,SAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;AACxB,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,WAAW;AAC1F,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,UAAU;AACzF,SAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;AACN,MAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GAAG;AACtF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,gBAAgB,WAAW,kBAAkB;AACvD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,sBAAsB,WAAW,gBAAgB;AAAA,IACnF,CAAC;AAAA,EACH;AACA,MAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAAG;AAClE,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,aAAa,WAAW,eAAe;AACjD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,UAAU,wBAAwB,WAAW,aAAa;AAAA,IAC/E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,WAAW,WAAW,aAAa;AAC7C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,CAAC,MAAM,IAAI,WAAW,WAAW,CAAC;AAAA,IAC5E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;AACN,MAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAAgB;AAChF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,WAAW,yBAAyB,WAAW,cAAc;AAAA,IAClF,CAAC;AAAA,EACH;AACA,MAAI,OAAO,SAAS,QAAQ,UAAU,KAAK,QAAQ,aAAa,WAAW,eAAe;AACxF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,CAAC,MAAM,IAAI,WAAW,aAAa,CAAC;AAAA,IAC1F,CAAC;AAAA,EACH;AACA,MAAI,gBAAgB,CAAC,aAAa,SAAS;AACzC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM,QAAQ,aAAa,iBAAiB,QAAQ;AAAA,MACpD,QAAQ,aAAa;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,YACA,SACA,QACM;AACN,MAAI,CAAC,WAAW,sBAAuB;AACvC,MAAI,QAAQ,aAAa,QAAQ,iBAAiB;AAChD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,eAAe;AAAA,IACzD,CAAC;AAAA,EACH;AACF;AAEA,SAAS,gBACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,cAAc,WAAW,gBAAgB;AACnD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,CAAC,MAAM,IAAI,WAAW,cAAc,CAAC;AAAA,IACrF,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,UACP,SACA,YACA,cACA,QACyB;AACzB,SAAO;AAAA,IACL;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC;AAAA,MACxE,GAAG,QAAQ,aAAa,uBAAuB,QAAQ,YAAY,OAAO;AAAA,IAC5E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS;AAAA,MAC5C,YAAY,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IACvE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,CAAC,aAAa,UAC1B,IACA,SAAS,QAAQ,YAAY,WAAW,aAAa;AAAA,MACzD,eAAe,QAAQ,WAAW,eAAe,IAAI,QAAQ,UAAU,CAAC;AAAA,IAC1E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ;AAAA,MACjE,mBAAmB,QAAQ,eAAe,IAAI,QAAQ,UAAU;AAAA,IAClE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,SAAS,UAAU;AAAA,MACnC,eAAe,IAAI,QAAQ,WAAW,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IAC7E;AAAA,EACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;AACvB,QAAM,MAAM,OAAO,OAAO,CAAC,MAAM,EAAE,SAAS,IAAI;AAChD,QAAM,SAAS,IAAI,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;AACN,SAAO,EAAE,MAAM,QAAQ,OAAO,QAAQ,KAAK,GAAG,OAAO;AACvD;AAEA,SAAS,oBAAoB,WAAqE;AAChG,QAAM,SAAuC,EAAE,OAAO,GAAG,KAAK,GAAG,MAAM,GAAG,SAAS,EAAE;AACrF,aAAW,YAAY,UAAW,QAAO,SAAS,SAAS,OAAO;AAClE,SAAO;AACT;AAEA,SAAS,aAAa,WAA+D;AACnF,QAAM,MAA8B,CAAC;AACrC,aAAW,YAAY,WAAW;AAChC,UAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;AACnE,QAAI,MAAM,KAAK,IAAI,MAAM,KAAK,KAAK;AAAA,EACrC;AACA,SAAO;AACT;AAEA,SAAS,kBACP,MACA,QACA,WACwB;AACxB,QAAM,MAA8B,CAAC;AACrC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,OAAO,IAAI,eAAe;AAChC,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,YAAM,OAAO,MAAM,gBAAgB,MAAM,OAAO,QAAQ,WAAW;AACnE,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;AACjG,QAAM,MAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,eAAW,OAAO,MAAM,OAAO,CAAC,GAAG;AACjC,YAAM,UAAU,IAAI,sBAAsB;AAC1C,UAAI,OAAO,KAAK,IAAI,OAAO,KAAK,KAAK;AAAA,IACvC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;AAC5B,QAAM,MAAkC,CAAC;AACzC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,YAAY,IAAI,QAAQ,IAAI;AAClC,UAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,UAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;AAAA,IACnD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,SACP,MACA,QACA,WACQ;AACR,QAAM,WAAW;AAAA,IACf,GAAG,KAAK,IAAI,CAAC,QAAQ;AACnB,YAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,aAAO,CAAC,IAAI,eAAe,UAAU,UAAa,SAAS;AAAA,IAC7D,CAAC;AAAA,IACD,GAAG,OAAO;AAAA,MACR,CAAC,UAAU,MAAM,OAAO,UAAU,MAAM,UAAU,UAAa,MAAM,SAAS;AAAA,IAChF;AAAA,EACF;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,SAAO,SAAS,OAAO,OAAO,EAAE,SAAS,SAAS;AACpD;AAEA,SAAS,UAAU,MAA4B,OAA8B;AAC3E,SAAO,KACJ,OAAO,CAAC,QAAQ,IAAI,aAAa,KAAK,EACtC,IAAI,CAAC,QAAS,UAAU,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ,WAAY,EACvF,OAAO,cAAc;AAC1B;AAEA,SAAS,KAAK,IAA+B;AAC3C,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,WAAW,IAAuB,GAAmB;AAC5D,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,SAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC,CAAC;AAC1F;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,SAAS,GAAW,GAAmB;AAC9C,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO;AAC9D,SAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAa,QAAwB;AACrD,MAAI,CAAC,OAAO,SAAS,GAAG,EAAG,QAAO;AAClC,MAAI,UAAU,EAAG,QAAO,OAAO,IAAI,IAAI;AACvC,SAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACQ;AACR,QAAM,OACJ,OAAO,SAAS,WAAW,cAAc,KAAK,OAAO,SAAS,QAAQ,WAAW,IAC7E,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IACxE;AACN,QAAM,UACJ,OAAO,SAAS,WAAW,YAAY,KAAK,OAAO,SAAS,QAAQ,SAAS,IACzE,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACpE;AACN,SAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;AACR,QAAM,SAAS,sBAAsB,MAAM,KAAK,MAAM;AACtD,QAAM,aAAa,aAAa,QAAQ,aAAa,eAAe,QAAQ,UAAU,gBAAgB,QAAQ,WAAW,aAAa,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAC/L,MAAI,OAAO,WAAW,EAAG,QAAO,GAAG,MAAM,KAAK,UAAU;AACxD,SAAO,GAAG,MAAM,KAAK,UAAU,YAAY,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAChF;AAEA,SAAS,IAAI,GAAmB;AAC9B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO,CAAC;AACxC,SAAO,EAAE,QAAQ,CAAC;AACpB;;;AC3iBO,SAAS,YACd,UACA,WACA,UAA4B,CAAC,GACZ;AACjB,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,WAAW,QAAQ,mBAAmB;AAC5C,QAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;AAEpE,QAAM,eAAeA,MAAK,QAAQ;AAClC,QAAM,gBAAgBA,MAAK,SAAS;AACpC,QAAM,QAAQ,gBAAgB;AAE9B,MACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GACrB;AACA,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,SAAS;AAAA,MACT,YAAY;AAAA,MACZ;AAAA,MACA,SAAS;AAAA,IACX;AAAA,EACF;AAEA,QAAM,SAAmB,IAAI,MAAM,UAAU;AAC7C,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,YAAY,SAAS,UAAU,GAAG;AACxC,UAAM,YAAY,SAAS,WAAW,GAAG;AACzC,WAAO,CAAC,IAAIA,MAAK,SAAS,IAAIA,MAAK,SAAS;AAAA,EAC9C;AACA,SAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,QAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;AACpD,QAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;AAC5D,QAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ,CAAC;AAC5C,QAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ,CAAC;AAEzD,MAAI;AACJ,MAAI,UAAU,EAAG,WAAU;AAAA,WAClB,UAAU,EAAG,WAAU;AAAA,WACvB,SAAS,EAAG,WAAU;AAAA,MAC1B,WAAU;AAEf,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAASA,MAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,MAAI,IAAI;AACR,aAAW,KAAK,GAAI,MAAK;AACzB,SAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;AAC3D,QAAM,MAAM,IAAI,MAAM,GAAG,MAAM;AAC/B,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,KAAI,CAAC,IAAI,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC7E,SAAO;AACT;AAGA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAGA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,IAAI;AACR,aAAW,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;AAC5B,UAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;AACjC,UAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;AACxC,eAAW,QAAQ,OAAO;AACxB,WAAK;AACL,UAAI,KAAK,KAAK,GAAG,QAAQ;AAAA,IAC3B;AAAA,EACF;AACA,SAAO,MAAM;AACf;AA8BA,eAAsB,gBACpB,MACkF;AAClF,QAAM,cAAc,KAAK,oBAAoB;AAC7C,QAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;AACnF,QAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;AACrF,QAAM,KAAK,YAAY,gBAAgB,iBAAiB;AAAA,IACtD,GAAI,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;AAAA,IACxD,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,IACvE,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,EACvD,CAAC;AACD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,iBAAiB,eAAe;AAAA,IAChC,kBAAkB,gBAAgB;AAAA,EACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;AACnB,QAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;AAClD,MAAI,OAAO;AACX,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,QAAQ,OAAQ;AACzB,YAAM,IAAI,MAAM,MAAM,QAAQ,CAAC,CAAE;AACjC,cAAQ,CAAC,IAAI,OAAO,SAAS,CAAC,IAAI,IAAI;AAAA,IACxC;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,GAAG,MAAM,OAAO,CAAC,CAAC;AAClF,SAAO;AACT;;;ACvNO,SAAS,oBACd,WACA,UAAsC,CAAC,GAC/B;AACR,QAAM,QAAQ,QAAQ,SAAS,mBAAmB,UAAU,MAAM;AAClE,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,KAAK,KAAK,EAAE;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,aAAa,UAAU,OAAO,YAAY,CAAC,IAAI;AAC1D,QAAM,KAAK,cAAc,UAAU,UAAU,QAAQ,IAAI,IAAI;AAC7D,MAAI,UAAU,YAAa,OAAM,KAAK,gBAAgB,UAAU,WAAW,IAAI;AAC/E,MAAI,UAAU,WAAY,OAAM,KAAK,eAAe,UAAU,UAAU,IAAI;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,UAAU,OAAO;AAC5B,QAAM,KAAK,EAAE;AAEb,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,oBAAoB;AAC/B,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,iBAAiB,UAAU,QAAQ,aAAa,IAAI;AAC/D,QAAM,KAAK,mBAAmB,UAAU,QAAQ,UAAU,IAAI;AAC9D,QAAM,KAAK,oBAAoB,UAAU,QAAQ,WAAW,IAAI;AAChE,QAAM,KAAK,iBAAiB,IAAI,UAAU,QAAQ,QAAQ,CAAC,IAAI;AAC/D,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,SAAS,CAAC,IAAI;AACjE,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,eAAe,CAAC,IAAI;AACxE,QAAM,KAAK,oBAAoB,IAAI,UAAU,QAAQ,gBAAgB,CAAC,IAAI;AAC1E,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,UAAU,CAAC,IAAI;AACnE,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,WAAW,CAAC,IAAI;AACnE,QAAM,KAAK,qBAAqB,KAAK,MAAM,UAAU,QAAQ,SAAS,CAAC,OAAO;AAC9E,QAAM,KAAK,EAAE;AAEb,MAAI,UAAU,OAAO,SAAS,GAAG;AAC/B,UAAM,KAAK,WAAW;AACtB,UAAM,KAAK,EAAE;AACb,eAAW,SAAS,UAAU,QAAQ;AACpC,YAAM,KAAK,OAAO,MAAM,QAAQ,QAAQ,MAAM,IAAI,OAAO,MAAM,IAAI,MAAM,MAAM,MAAM,EAAE;AAAA,IACzF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,wBAAwB;AACnE,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,yBAAyB;AACpC,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,SAAS,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,OAAO,KAAK,KAAK,EAAE;AAC5E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,iBAAiB;AAC5D,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,kBAAkB;AAC7B,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,MAAM,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,IAAI,KAAK,KAAK,EAAE;AACtE,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,QAAQ,QAAQ,KAAK,SAAS,GAAG;AAC3C,UAAM,KAAK,gBAAgB;AAC3B,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,aAAa,CAAC,GAAG,QAAQ,IAAI,GAAG;AAAA,QAC9B,YAAY,QAAQ,cAAc,UAAU,cAAc;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC,EAAE;AAAA,IACL;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,wBAAwB,QAAQ,qBAAqB,SAAS,GAAG;AAC3E,UAAM,KAAK,0BAA0B;AACrC,UAAM,KAAK,EAAE;AACb,eAAW,WAAW,QAAQ,qBAAsB,OAAM,KAAK,KAAK,OAAO,EAAE;AAC7E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,cAAc,QAAQ,eAAe,mBAAmB,SAAS;AACvE,MAAI,YAAY,SAAS,GAAG;AAC1B,UAAM,KAAK,iBAAiB;AAC5B,UAAM,KAAK,EAAE;AACb,eAAW,UAAU,YAAa,OAAM,KAAK,KAAK,MAAM,EAAE;AAC1D,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,CAAC;AAAA;AACtC;AAEA,SAAS,mBAAmB,WAAiD;AAC3E,MAAI,UAAU,QAAS,QAAO,CAAC,kDAAkD;AACjF,SAAO,UAAU,OACd,OAAO,CAAC,UAAU,MAAM,aAAa,UAAU,EAC/C,IAAI,CAAC,UAAU,WAAW,MAAM,IAAI,KAAK,MAAM,MAAM,EAAE;AAC5D;AAEA,SAAS,QAAQ,QAAyD;AACxE,SAAO,OAAO,QAAQ,MAAM,EACzB,OAAO,CAAC,CAAC,EAAE,KAAK,MAAM,QAAQ,CAAC,EAC/B,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC;AAC3D;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC,MAAM;AACnE;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,MAAM,QAAQ,CAAC,IAAI;AACrD;","names":["mean"]}
+{"version":3,"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"sourcesContent":["/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n *   - Dataset manifests prove corpus/version coverage.\n *   - RunRecord rows prove reproducible search/holdout outcomes.\n *   - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n *   - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport type { RunRecord, RunSplitTag } from './run-record'\n\n/** Severity of an actionable finding attached to a run/trace. */\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\n/** Actionable side-info — a diagnosed finding the loop can act on. */\nexport interface ActionableSideInfo {\n  /** Stable expectation/check id when available. */\n  expectationId?: string\n  /** Human-readable diagnosis of what happened. */\n  message: string\n  severity?: AsiSeverity\n  /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n  evidence?: string\n  /** Prompt/tool/context surface likely responsible. */\n  responsibleSurface?: string\n  /** Suggested fix in natural language. */\n  suggestion?: string\n  /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n  matched?: boolean\n  metadata?: Record<string, unknown>\n}\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n  | 'corpus'\n  | 'quality'\n  | 'generalization'\n  | 'diagnostics'\n  | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n  scenarioId: string\n  candidateId?: string\n  split?: RunSplitTag\n  score?: number\n  ok?: boolean\n  turnCount?: number\n  costUsd?: number\n  durationMs?: number\n  failureMode?: string\n  asi?: ActionableSideInfo[]\n  metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n  /** Require a Dataset manifest or explicit scenarios. Default true. */\n  requireCorpus?: boolean\n  minScenarioCount?: number\n  minSearchRuns?: number\n  minHoldoutRuns?: number\n  /** Require at least one holdout scenario/run. Default true. */\n  requireHoldout?: boolean\n  minPassRate?: number\n  minMeanScore?: number\n  /** Search mean may exceed holdout mean by at most this much. */\n  maxOverfitGap?: number\n  maxMeanCostUsd?: number\n  maxP95WallMs?: number\n  /** Low-score/failed rows must carry ASI. Default true. */\n  requireAsiForFailures?: boolean\n  /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n  failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n  target: string\n  candidateId?: string\n  baselineId?: string\n  dataset?: DatasetManifest\n  scenarios?: readonly DatasetScenario[]\n  runs?: readonly RunRecord[]\n  traces?: readonly ReleaseTraceEvidence[]\n  gateDecision?: GateDecision | null\n  thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n  name: ReleaseConfidenceAxisName\n  status: ReleaseConfidenceStatus\n  score: number\n  detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n  axis: ReleaseConfidenceAxisName\n  severity: 'critical' | 'warning'\n  code: string\n  detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n  scenarioCount: number\n  searchRuns: number\n  holdoutRuns: number\n  passRate: number\n  meanScore: number\n  searchMeanScore: number\n  holdoutMeanScore: number\n  overfitGap: number\n  meanCostUsd: number\n  p95WallMs: number\n  failedRows: number\n  failuresWithAsi: number\n  singleShotTraces: number\n  multiShotTraces: number\n  splitCounts: Record<DatasetSplit, number>\n  domainCounts: Record<string, number>\n  failureModeCounts: Record<string, number>\n  responsibleSurfaceCounts: Record<string, number>\n}\n\nexport interface ReleaseConfidenceScorecard {\n  target: string\n  candidateId: string | null\n  baselineId: string | null\n  status: ReleaseConfidenceStatus\n  promote: boolean\n  axes: ReleaseConfidenceAxis[]\n  issues: ReleaseConfidenceIssue[]\n  metrics: ReleaseConfidenceMetrics\n  dataset: DatasetManifest | null\n  gateDecision: GateDecision | null\n  summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n  requireCorpus: true,\n  minScenarioCount: 1,\n  minSearchRuns: 1,\n  minHoldoutRuns: 1,\n  requireHoldout: true,\n  minPassRate: 0.8,\n  minMeanScore: 0.7,\n  maxOverfitGap: 0.15,\n  maxMeanCostUsd: Number.POSITIVE_INFINITY,\n  maxP95WallMs: Number.POSITIVE_INFINITY,\n  requireAsiForFailures: true,\n  failureScoreThreshold: 0.5,\n}\n\nexport function evaluateReleaseConfidence(\n  input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n  const candidateId = input.candidateId ?? null\n  const runs = filterCandidate(input.runs ?? [], candidateId, input.baselineId)\n  const traces = filterTraceCandidate(input.traces ?? [], candidateId, input.baselineId)\n  const scenarios = input.scenarios ?? []\n  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n  const searchScores = scoresFor(runs, 'search')\n  const holdoutScores = scoresFor(runs, 'holdout')\n  const allScores = [...searchScores, ...holdoutScores]\n  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n  const scoreUniverse = allScores.length > 0 ? allScores : traceScores\n  const searchRuns = runs.filter((r) => r.splitTag === 'search').length\n  const holdoutRuns = runs.filter((r) => r.splitTag === 'holdout').length\n  const searchMeanScore = mean(searchScores)\n  const holdoutMeanScore = mean(holdoutScores)\n  const metrics: ReleaseConfidenceMetrics = {\n    scenarioCount,\n    searchRuns,\n    holdoutRuns,\n    passRate: passRate(runs, traces, thresholds.failureScoreThreshold),\n    meanScore: mean(scoreUniverse),\n    searchMeanScore,\n    holdoutMeanScore,\n    overfitGap: safeDiff(searchMeanScore, holdoutMeanScore),\n    meanCostUsd: mean([\n      ...runs.map((r) => r.costUsd),\n      ...traces.map((t) => t.costUsd).filter(isFiniteNumber),\n    ]),\n    p95WallMs: percentile(\n      [...runs.map((r) => r.wallMs), ...traces.map((t) => t.durationMs).filter(isFiniteNumber)],\n      0.95,\n    ),\n    failedRows: failedRows(runs, traces, thresholds.failureScoreThreshold).length,\n    failuresWithAsi: failedRows(runs, traces, thresholds.failureScoreThreshold).filter(\n      (row) => row.hasAsi,\n    ).length,\n    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n    splitCounts,\n    domainCounts: countDomains(scenarios),\n    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),\n    responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n  }\n\n  const issues: ReleaseConfidenceIssue[] = []\n  checkCorpus(input, thresholds, metrics, issues)\n  checkQuality(thresholds, metrics, issues)\n  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n  checkDiagnostics(thresholds, metrics, issues)\n  checkEfficiency(thresholds, metrics, issues)\n\n  const axes = buildAxes(metrics, thresholds, input.gateDecision ?? null, issues)\n  const status = issues.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : issues.length > 0\n      ? 'warn'\n      : 'pass'\n\n  return {\n    target: input.target,\n    candidateId,\n    baselineId: input.baselineId ?? null,\n    status,\n    promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n    axes,\n    issues,\n    metrics,\n    dataset: input.dataset ?? null,\n    gateDecision: input.gateDecision ?? null,\n    summary: renderSummary(input.target, status, metrics, issues),\n  }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n  const scorecard = evaluateReleaseConfidence(input)\n  if (scorecard.status === 'fail') {\n    throw new VerificationError(scorecard.summary)\n  }\n  return scorecard\n}\n\nfunction filterCandidate(\n  runs: readonly RunRecord[],\n  candidateId: string | null,\n  baselineId?: string,\n): RunRecord[] {\n  if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n  return [...runs]\n}\n\nfunction filterTraceCandidate(\n  traces: readonly ReleaseTraceEvidence[],\n  candidateId: string | null,\n  baselineId?: string,\n): ReleaseTraceEvidence[] {\n  if (candidateId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n  if (baselineId)\n    return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n  return [...traces]\n}\n\nfunction checkCorpus(\n  input: ReleaseConfidenceInput,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_corpus',\n      detail: 'No Dataset manifest or scenarios supplied.',\n    })\n  }\n  if (metrics.scenarioCount < thresholds.minScenarioCount) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'few_scenarios',\n      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n    })\n  }\n  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n    issues.push({\n      axis: 'corpus',\n      severity: 'critical',\n      code: 'missing_holdout_split',\n      detail: 'Corpus has no holdout scenarios.',\n    })\n  }\n}\n\nfunction checkQuality(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.searchRuns < thresholds.minSearchRuns) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'few_search_runs',\n      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n    })\n  }\n  if (metrics.passRate < thresholds.minPassRate) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_pass_rate',\n      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n    })\n  }\n  if (metrics.meanScore < thresholds.minMeanScore) {\n    issues.push({\n      axis: 'quality',\n      severity: 'critical',\n      code: 'low_mean_score',\n      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n    })\n  }\n}\n\nfunction checkGeneralization(\n  gateDecision: GateDecision | null,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'few_holdout_runs',\n      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n    })\n  }\n  if (Number.isFinite(metrics.overfitGap) && metrics.overfitGap > thresholds.maxOverfitGap) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: 'overfit_gap',\n      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n    })\n  }\n  if (gateDecision && !gateDecision.promote) {\n    issues.push({\n      axis: 'generalization',\n      severity: 'critical',\n      code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n      detail: gateDecision.reason,\n    })\n  }\n}\n\nfunction checkDiagnostics(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (!thresholds.requireAsiForFailures) return\n  if (metrics.failedRows > metrics.failuresWithAsi) {\n    issues.push({\n      axis: 'diagnostics',\n      severity: 'critical',\n      code: 'missing_failure_asi',\n      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n    })\n  }\n}\n\nfunction checkEfficiency(\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): void {\n  if (metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'cost_budget',\n      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n    })\n  }\n  if (metrics.p95WallMs > thresholds.maxP95WallMs) {\n    issues.push({\n      axis: 'efficiency',\n      severity: 'critical',\n      code: 'latency_budget',\n      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n    })\n  }\n}\n\nfunction buildAxes(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n  gateDecision: GateDecision | null,\n  issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n  return [\n    axis(\n      'corpus',\n      issues,\n      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n    ),\n    axis(\n      'quality',\n      issues,\n      Math.min(metrics.passRate, metrics.meanScore),\n      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n    ),\n    axis(\n      'generalization',\n      issues,\n      gateDecision && !gateDecision.promote\n        ? 0\n        : gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n    ),\n    axis(\n      'diagnostics',\n      issues,\n      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n    ),\n    axis(\n      'efficiency',\n      issues,\n      efficiencyScore(metrics, thresholds),\n      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n    ),\n  ]\n}\n\nfunction axis(\n  name: ReleaseConfidenceAxisName,\n  issues: ReleaseConfidenceIssue[],\n  score: number,\n  detail: string,\n): ReleaseConfidenceAxis {\n  const own = issues.filter((i) => i.axis === name)\n  const status = own.some((i) => i.severity === 'critical')\n    ? 'fail'\n    : own.length > 0\n      ? 'warn'\n      : 'pass'\n  return { name, status, score: bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n  const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n  for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n  return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const scenario of scenarios) {\n    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n    out[domain] = (out[domain] ?? 0) + 1\n  }\n  return out\n}\n\nfunction countFailureModes(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const mode = run.failureMode ?? 'low_score'\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      const mode = trace.failureMode ?? (trace.ok === false ? 'not_ok' : 'low_score')\n      out[mode] = (out[mode] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n  const out: Record<string, number> = {}\n  for (const trace of traces) {\n    for (const asi of trace.asi ?? []) {\n      const surface = asi.responsibleSurface ?? 'unknown'\n      out[surface] = (out[surface] ?? 0) + 1\n    }\n  }\n  return out\n}\n\nfunction failedRows(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): Array<{ hasAsi: boolean }> {\n  const out: Array<{ hasAsi: boolean }> = []\n  for (const run of runs) {\n    const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (run.failureMode || (score !== undefined && score < threshold)) {\n      const asiMetric = run.outcome.raw.asi\n      out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n    }\n  }\n  for (const trace of traces) {\n    if (\n      trace.failureMode ||\n      trace.ok === false ||\n      (trace.score !== undefined && trace.score < threshold)\n    ) {\n      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n    }\n  }\n  return out\n}\n\nfunction passRate(\n  runs: readonly RunRecord[],\n  traces: readonly ReleaseTraceEvidence[],\n  threshold: number,\n): number {\n  const outcomes = [\n    ...runs.map((run) => {\n      const score = run.outcome.holdoutScore ?? run.outcome.searchScore\n      return !run.failureMode && score !== undefined && score >= threshold\n    }),\n    ...traces.map(\n      (trace) => trace.ok !== false && (trace.score === undefined || trace.score >= threshold),\n    ),\n  ]\n  if (outcomes.length === 0) return 0\n  return outcomes.filter(Boolean).length / outcomes.length\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n  return runs\n    .filter((run) => run.splitTag === split)\n    .map((run) => (split === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore))\n    .filter(isFiniteNumber)\n}\n\nfunction mean(xs: readonly number[]): number {\n  if (xs.length === 0) return Number.NaN\n  return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentile(xs: readonly number[], p: number): number {\n  if (xs.length === 0) return Number.NaN\n  const sorted = [...xs].sort((a, b) => a - b)\n  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction safeDiff(a: number, b: number): number {\n  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN\n  return a - b\n}\n\nfunction gapScore(gap: number, maxGap: number): number {\n  if (!Number.isFinite(gap)) return 0\n  if (maxGap <= 0) return gap <= 0 ? 1 : 0\n  return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n  metrics: ReleaseConfidenceMetrics,\n  thresholds: Required<ReleaseConfidenceThresholds>,\n): number {\n  const cost =\n    Number.isFinite(thresholds.maxMeanCostUsd) && Number.isFinite(metrics.meanCostUsd)\n      ? bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n      : 1\n  const latency =\n    Number.isFinite(thresholds.maxP95WallMs) && Number.isFinite(metrics.p95WallMs)\n      ? bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n      : 1\n  return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n  target: string,\n  status: ReleaseConfidenceStatus,\n  metrics: ReleaseConfidenceMetrics,\n  issues: ReleaseConfidenceIssue[],\n): string {\n  const prefix = `release confidence ${status}: ${target}`\n  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n  if (issues.length === 0) return `${prefix}; ${metricText}`\n  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number): string {\n  if (!Number.isFinite(x)) return String(x)\n  return x.toFixed(4)\n}\n","/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n *   - Bootstrap mean baseline vs candidate (1k resamples).\n *   - Compute the delta distribution; pass if the lower CI bound > 0.\n *   - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n *   - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n *   - KEEP:    overlap, but candidate point estimate >= baseline (neutral)\n *   - REVERT:  candidate's CI upper bound < baseline mean (real regression)\n *   - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n  baselineMean: number\n  candidateMean: number\n  /** candidateMean - baselineMean, point estimate. */\n  delta: number\n  /** Lower bound of the (1 - alpha) CI on the delta. */\n  ciLower: number\n  /** Upper bound of the (1 - alpha) CI on the delta. */\n  ciUpper: number\n  /** Number of bootstrap resamples used. */\n  iterations: number\n  alpha: number\n  verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n  /** Confidence level alpha (default 0.05 → 95% CI). */\n  alpha?: number\n  /** Number of resamples (default 1000). */\n  iterations?: number\n  /**\n   * Minimum total samples (baseline + candidate) below which we always\n   * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n   * Default 6 (combined).\n   */\n  minTotalSamples?: number\n  /** RNG seed for reproducibility. Default: Math.random. */\n  seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n  baseline: number[],\n  candidate: number[],\n  options: BootstrapOptions = {},\n): BootstrapResult {\n  const alpha = options.alpha ?? 0.05\n  const iterations = options.iterations ?? 1000\n  const minTotal = options.minTotalSamples ?? 6\n  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n  const baselineMean = mean(baseline)\n  const candidateMean = mean(candidate)\n  const delta = candidateMean - baselineMean\n\n  if (\n    baseline.length + candidate.length < minTotal ||\n    baseline.length === 0 ||\n    candidate.length === 0\n  ) {\n    return {\n      baselineMean,\n      candidateMean,\n      delta,\n      ciLower: -Infinity,\n      ciUpper: Infinity,\n      iterations: 0,\n      alpha,\n      verdict: 'INCONCLUSIVE',\n    }\n  }\n\n  const deltas: number[] = new Array(iterations)\n  for (let i = 0; i < iterations; i++) {\n    const bResample = resample(baseline, rng)\n    const cResample = resample(candidate, rng)\n    deltas[i] = mean(cResample) - mean(bResample)\n  }\n  deltas.sort((a, b) => a - b)\n  const lowerIdx = Math.floor((alpha / 2) * iterations)\n  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n  const ciLower = deltas[Math.max(0, lowerIdx)]!\n  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n  let verdict: Verdict\n  if (ciLower > 0) verdict = 'ADVANCE'\n  else if (ciUpper < 0) verdict = 'REVERT'\n  else if (delta >= 0) verdict = 'KEEP'\n  else verdict = 'INCONCLUSIVE'\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ciLower,\n    ciUpper,\n    iterations,\n    alpha,\n    verdict,\n  }\n}\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  let s = 0\n  for (const x of xs) s += x\n  return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n  const out = new Array(xs.length)\n  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n  return out\n}\n\n/** Mulberry32 — fast deterministic PRNG. Stable across runs given the same seed. */\nfunction mulberry32(seed: number): () => number {\n  let t = seed >>> 0\n  return () => {\n    t += 0x6d2b79f5\n    let r = t\n    r = Math.imul(r ^ (r >>> 15), r | 1)\n    r ^= r + Math.imul(r ^ (r >>> 7), r | 61)\n    return ((r ^ (r >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n  let h = 2166136261\n  for (const x of [...a, ...b]) {\n    const view = new Float64Array([x])\n    const bytes = new Uint8Array(view.buffer)\n    for (const byte of bytes) {\n      h ^= byte\n      h = Math.imul(h, 16777619)\n    }\n  }\n  return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n  baselineOutputs: TOutput[]\n  candidateOutputs: TOutput[]\n  /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n  judge: (output: TOutput) => Promise<number> | number\n  alpha?: number\n  iterations?: number\n  /** RNG seed for reproducibility. */\n  seed?: number\n  /** Maximum concurrent judge calls. Default 4. */\n  judgeConcurrency?: number\n}\n\n/**\n * Confirm a candidate's win with a stronger judge: score baseline and candidate outputs independently, then bootstrap a CI to verify the lift generalises beyond the inner loop.\n */\nexport async function judgeReplayGate<TOutput>(\n  args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n  const concurrency = args.judgeConcurrency ?? 4\n  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n  const ci = bootstrapCi(baselineScores, candidateScores, {\n    ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n    ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n    ...(args.seed !== undefined ? { seed: args.seed } : {}),\n  })\n  return {\n    ...ci,\n    baselineSamples: baselineScores.length,\n    candidateSamples: candidateScores.length,\n  }\n}\n\nasync function scoreAll<TOutput>(\n  outputs: TOutput[],\n  judge: (output: TOutput) => Promise<number> | number,\n  concurrency: number,\n): Promise<number[]> {\n  const results: number[] = new Array(outputs.length)\n  let next = 0\n  async function worker(): Promise<void> {\n    while (true) {\n      const i = next++\n      if (i >= outputs.length) return\n      const v = await judge(outputs[i]!)\n      results[i] = Number.isFinite(v) ? v : 0\n    }\n  }\n  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n  return results\n}\n","import type { ReleaseConfidenceScorecard } from './release-confidence'\nimport type { RunRecord } from './run-record'\nimport { summaryTable } from './summary-report'\n\nexport interface RenderReleaseReportOptions {\n  title?: string\n  runs?: readonly RunRecord[]\n  comparator?: string\n  traceAnalystFindings?: readonly string[]\n  nextActions?: readonly string[]\n}\n\nexport function renderReleaseReport(\n  scorecard: ReleaseConfidenceScorecard,\n  options: RenderReleaseReportOptions = {},\n): string {\n  const title = options.title ?? `Release Report: ${scorecard.target}`\n  const lines: string[] = []\n  lines.push(`# ${title}`)\n  lines.push('')\n  lines.push(`Status: **${scorecard.status.toUpperCase()}**`)\n  lines.push(`Promote: **${scorecard.promote ? 'yes' : 'no'}**`)\n  if (scorecard.candidateId) lines.push(`Candidate: \\`${scorecard.candidateId}\\``)\n  if (scorecard.baselineId) lines.push(`Baseline: \\`${scorecard.baselineId}\\``)\n  lines.push('')\n  lines.push(scorecard.summary)\n  lines.push('')\n\n  lines.push('## Metrics')\n  lines.push('')\n  lines.push('| Metric | Value |')\n  lines.push('|---|---:|')\n  lines.push(`| Scenarios | ${scorecard.metrics.scenarioCount} |`)\n  lines.push(`| Search runs | ${scorecard.metrics.searchRuns} |`)\n  lines.push(`| Holdout runs | ${scorecard.metrics.holdoutRuns} |`)\n  lines.push(`| Pass rate | ${pct(scorecard.metrics.passRate)} |`)\n  lines.push(`| Mean score | ${num(scorecard.metrics.meanScore)} |`)\n  lines.push(`| Search mean | ${num(scorecard.metrics.searchMeanScore)} |`)\n  lines.push(`| Holdout mean | ${num(scorecard.metrics.holdoutMeanScore)} |`)\n  lines.push(`| Overfit gap | ${num(scorecard.metrics.overfitGap)} |`)\n  lines.push(`| Mean cost | $${num(scorecard.metrics.meanCostUsd)} |`)\n  lines.push(`| p95 wall time | ${Math.round(scorecard.metrics.p95WallMs)} ms |`)\n  lines.push('')\n\n  if (scorecard.issues.length > 0) {\n    lines.push('## Issues')\n    lines.push('')\n    for (const issue of scorecard.issues) {\n      lines.push(`- **${issue.severity}** \\`${issue.code}\\` (${issue.axis}): ${issue.detail}`)\n    }\n    lines.push('')\n  }\n\n  const surfaces = entries(scorecard.metrics.responsibleSurfaceCounts)\n  if (surfaces.length > 0) {\n    lines.push('## Responsible Surfaces')\n    lines.push('')\n    for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`)\n    lines.push('')\n  }\n\n  const failures = entries(scorecard.metrics.failureModeCounts)\n  if (failures.length > 0) {\n    lines.push('## Failure Modes')\n    lines.push('')\n    for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`)\n    lines.push('')\n  }\n\n  if (options.runs && options.runs.length > 0) {\n    lines.push('## Run Summary')\n    lines.push('')\n    lines.push(\n      summaryTable([...options.runs], {\n        comparator: options.comparator ?? scorecard.baselineId ?? undefined,\n        split: 'holdout',\n      }).markdown,\n    )\n    lines.push('')\n  }\n\n  if (options.traceAnalystFindings && options.traceAnalystFindings.length > 0) {\n    lines.push('## TraceAnalyst Findings')\n    lines.push('')\n    for (const finding of options.traceAnalystFindings) lines.push(`- ${finding}`)\n    lines.push('')\n  }\n\n  const nextActions = options.nextActions ?? defaultNextActions(scorecard)\n  if (nextActions.length > 0) {\n    lines.push('## Next Actions')\n    lines.push('')\n    for (const action of nextActions) lines.push(`- ${action}`)\n    lines.push('')\n  }\n\n  return `${lines.join('\\n').trimEnd()}\\n`\n}\n\nfunction defaultNextActions(scorecard: ReleaseConfidenceScorecard): string[] {\n  if (scorecard.promote) return ['Promote the candidate and keep canaries enabled.']\n  return scorecard.issues\n    .filter((issue) => issue.severity === 'critical')\n    .map((issue) => `Resolve ${issue.code}: ${issue.detail}`)\n}\n\nfunction entries(values: Record<string, number>): Array<[string, number]> {\n  return Object.entries(values)\n    .filter(([, count]) => count > 0)\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n}\n\nfunction pct(value: number): string {\n  return Number.isFinite(value) ? `${(value * 100).toFixed(1)}%` : 'n/a'\n}\n\nfunction num(value: number): string {\n  return Number.isFinite(value) ? value.toFixed(3) : 'n/a'\n}\n"],"mappings":";;;;;;;;AA+IA,IAAM,qBAA4D;AAAA,EAChE,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,aAAa;AAAA,EACb,cAAc;AAAA,EACd,eAAe;AAAA,EACf,gBAAgB,OAAO;AAAA,EACvB,cAAc,OAAO;AAAA,EACrB,uBAAuB;AAAA,EACvB,uBAAuB;AACzB;AAEO,SAAS,0BACd,OAC4B;AAC5B,QAAM,aAAa,EAAE,GAAG,oBAAoB,GAAG,MAAM,WAAW;AAChE,QAAM,cAAc,MAAM,eAAe;AACzC,QAAM,OAAO,gBAAgB,MAAM,QAAQ,CAAC,GAAG,aAAa,MAAM,UAAU;AAC5E,QAAM,SAAS,qBAAqB,MAAM,UAAU,CAAC,GAAG,aAAa,MAAM,UAAU;AACrF,QAAM,YAAY,MAAM,aAAa,CAAC;AACtC,QAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;AAChE,QAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;AAC/E,QAAM,eAAe,UAAU,MAAM,QAAQ;AAC7C,QAAM,gBAAgB,UAAU,MAAM,SAAS;AAC/C,QAAM,YAAY,CAAC,GAAG,cAAc,GAAG,aAAa;AACpD,QAAM,cAAc,OAAO,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,OAAO,cAAc;AACpE,QAAM,gBAAgB,UAAU,SAAS,IAAI,YAAY;AACzD,QAAM,aAAa,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,QAAQ,EAAE;AAC/D,QAAM,cAAc,KAAK,OAAO,CAAC,MAAM,EAAE,aAAa,SAAS,EAAE;AACjE,QAAM,kBAAkB,KAAK,YAAY;AACzC,QAAM,mBAAmB,KAAK,aAAa;AAC3C,QAAM,UAAoC;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA,UAAU,SAAS,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACjE,WAAW,KAAK,aAAa;AAAA,IAC7B;AAAA,IACA;AAAA,IACA,YAAY,SAAS,iBAAiB,gBAAgB;AAAA,IACtD,aAAa,KAAK;AAAA,MAChB,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,OAAO;AAAA,MAC5B,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,cAAc;AAAA,IACvD,CAAC;AAAA,IACD,WAAW;AAAA,MACT,CAAC,GAAG,KAAK,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,GAAG,OAAO,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,OAAO,cAAc,CAAC;AAAA,MACxF;AAAA,IACF;AAAA,IACA,YAAY,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,IACvE,iBAAiB,WAAW,MAAM,QAAQ,WAAW,qBAAqB,EAAE;AAAA,MAC1E,CAAC,QAAQ,IAAI;AAAA,IACf,EAAE;AAAA,IACF,kBAAkB,OAAO,OAAO,CAAC,MAAM,EAAE,cAAc,CAAC,EAAE;AAAA,IAC1D,iBAAiB,OAAO,OAAO,CAAC,OAAO,EAAE,aAAa,KAAK,CAAC,EAAE;AAAA,IAC9D;AAAA,IACA,cAAc,aAAa,SAAS;AAAA,IACpC,mBAAmB,kBAAkB,MAAM,QAAQ,WAAW,qBAAqB;AAAA,IACnF,0BAA0B,yBAAyB,MAAM;AAAA,EAC3D;AAEA,QAAM,SAAmC,CAAC;AAC1C,cAAY,OAAO,YAAY,SAAS,MAAM;AAC9C,eAAa,YAAY,SAAS,MAAM;AACxC,sBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;AAC3E,mBAAiB,YAAY,SAAS,MAAM;AAC5C,kBAAgB,YAAY,SAAS,MAAM;AAE3C,QAAM,OAAO,UAAU,SAAS,YAAY,MAAM,gBAAgB,MAAM,MAAM;AAC9E,QAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;AAEN,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd;AAAA,IACA,YAAY,MAAM,cAAc;AAAA,IAChC;AAAA,IACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;AAAA,IACjF;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,MAAM,WAAW;AAAA,IAC1B,cAAc,MAAM,gBAAgB;AAAA,IACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;AAAA,EAC9D;AACF;AAEO,SAAS,wBAAwB,OAA2D;AACjG,QAAM,YAAY,0BAA0B,KAAK;AACjD,MAAI,UAAU,WAAW,QAAQ;AAC/B,UAAM,IAAI,kBAAkB,UAAU,OAAO;AAAA,EAC/C;AACA,SAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;AACb,MAAI,YAAa,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,WAAW;AACxE,MAAI,WAAY,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAU;AACtE,SAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;AACxB,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,WAAW;AAC1F,MAAI;AACF,WAAO,OAAO,OAAO,CAAC,MAAM,EAAE,gBAAgB,UAAa,EAAE,gBAAgB,UAAU;AACzF,SAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;AACN,MAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GAAG;AACtF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,gBAAgB,WAAW,kBAAkB;AACvD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,sBAAsB,WAAW,gBAAgB;AAAA,IACnF,CAAC;AAAA,EACH;AACA,MAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAAG;AAClE,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,aAAa,WAAW,eAAe;AACjD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,UAAU,wBAAwB,WAAW,aAAa;AAAA,IAC/E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,WAAW,WAAW,aAAa;AAC7C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,CAAC,MAAM,IAAI,WAAW,WAAW,CAAC;AAAA,IAC5E,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;AACN,MAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAAgB;AAChF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,WAAW,yBAAyB,WAAW,cAAc;AAAA,IAClF,CAAC;AAAA,EACH;AACA,MAAI,OAAO,SAAS,QAAQ,UAAU,KAAK,QAAQ,aAAa,WAAW,eAAe;AACxF,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,CAAC,MAAM,IAAI,WAAW,aAAa,CAAC;AAAA,IAC1F,CAAC;AAAA,EACH;AACA,MAAI,gBAAgB,CAAC,aAAa,SAAS;AACzC,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM,QAAQ,aAAa,iBAAiB,QAAQ;AAAA,MACpD,QAAQ,aAAa;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,YACA,SACA,QACM;AACN,MAAI,CAAC,WAAW,sBAAuB;AACvC,MAAI,QAAQ,aAAa,QAAQ,iBAAiB;AAChD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,eAAe;AAAA,IACzD,CAAC;AAAA,EACH;AACF;AAEA,SAAS,gBACP,YACA,SACA,QACM;AACN,MAAI,QAAQ,cAAc,WAAW,gBAAgB;AACnD,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,CAAC,MAAM,IAAI,WAAW,cAAc,CAAC;AAAA,IACrF,CAAC;AAAA,EACH;AACA,MAAI,QAAQ,YAAY,WAAW,cAAc;AAC/C,WAAO,KAAK;AAAA,MACV,MAAM;AAAA,MACN,UAAU;AAAA,MACV,MAAM;AAAA,MACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,CAAC,MAAM,IAAI,WAAW,YAAY,CAAC;AAAA,IAC/E,CAAC;AAAA,EACH;AACF;AAEA,SAAS,UACP,SACA,YACA,cACA,QACyB;AACzB,SAAO;AAAA,IACL;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC;AAAA,MACxE,GAAG,QAAQ,aAAa,uBAAuB,QAAQ,YAAY,OAAO;AAAA,IAC5E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS;AAAA,MAC5C,YAAY,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IACvE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,CAAC,aAAa,UAC1B,IACA,SAAS,QAAQ,YAAY,WAAW,aAAa;AAAA,MACzD,eAAe,QAAQ,WAAW,eAAe,IAAI,QAAQ,UAAU,CAAC;AAAA,IAC1E;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ;AAAA,MACjE,mBAAmB,QAAQ,eAAe,IAAI,QAAQ,UAAU;AAAA,IAClE;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA,gBAAgB,SAAS,UAAU;AAAA,MACnC,eAAe,IAAI,QAAQ,WAAW,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAAA,IAC7E;AAAA,EACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;AACvB,QAAM,MAAM,OAAO,OAAO,CAAC,MAAM,EAAE,SAAS,IAAI;AAChD,QAAM,SAAS,IAAI,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;AACN,SAAO,EAAE,MAAM,QAAQ,OAAO,QAAQ,KAAK,GAAG,OAAO;AACvD;AAEA,SAAS,oBAAoB,WAAqE;AAChG,QAAM,SAAuC,EAAE,OAAO,GAAG,KAAK,GAAG,MAAM,GAAG,SAAS,EAAE;AACrF,aAAW,YAAY,UAAW,QAAO,SAAS,SAAS,OAAO;AAClE,SAAO;AACT;AAEA,SAAS,aAAa,WAA+D;AACnF,QAAM,MAA8B,CAAC;AACrC,aAAW,YAAY,WAAW;AAChC,UAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;AACnE,QAAI,MAAM,KAAK,IAAI,MAAM,KAAK,KAAK;AAAA,EACrC;AACA,SAAO;AACT;AAEA,SAAS,kBACP,MACA,QACA,WACwB;AACxB,QAAM,MAA8B,CAAC;AACrC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,OAAO,IAAI,eAAe;AAChC,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,YAAM,OAAO,MAAM,gBAAgB,MAAM,OAAO,QAAQ,WAAW;AACnE,UAAI,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;AACjG,QAAM,MAA8B,CAAC;AACrC,aAAW,SAAS,QAAQ;AAC1B,eAAW,OAAO,MAAM,OAAO,CAAC,GAAG;AACjC,YAAM,UAAU,IAAI,sBAAsB;AAC1C,UAAI,OAAO,KAAK,IAAI,OAAO,KAAK,KAAK;AAAA,IACvC;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;AAC5B,QAAM,MAAkC,CAAC;AACzC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,QAAI,IAAI,eAAgB,UAAU,UAAa,QAAQ,WAAY;AACjE,YAAM,YAAY,IAAI,QAAQ,IAAI;AAClC,UAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AACA,aAAW,SAAS,QAAQ;AAC1B,QACE,MAAM,eACN,MAAM,OAAO,SACZ,MAAM,UAAU,UAAa,MAAM,QAAQ,WAC5C;AACA,UAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;AAAA,IACnD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,SACP,MACA,QACA,WACQ;AACR,QAAM,WAAW;AAAA,IACf,GAAG,KAAK,IAAI,CAAC,QAAQ;AACnB,YAAM,QAAQ,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACtD,aAAO,CAAC,IAAI,eAAe,UAAU,UAAa,SAAS;AAAA,IAC7D,CAAC;AAAA,IACD,GAAG,OAAO;AAAA,MACR,CAAC,UAAU,MAAM,OAAO,UAAU,MAAM,UAAU,UAAa,MAAM,SAAS;AAAA,IAChF;AAAA,EACF;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,SAAO,SAAS,OAAO,OAAO,EAAE,SAAS,SAAS;AACpD;AAEA,SAAS,UAAU,MAA4B,OAA8B;AAC3E,SAAO,KACJ,OAAO,CAAC,QAAQ,IAAI,aAAa,KAAK,EACtC,IAAI,CAAC,QAAS,UAAU,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ,WAAY,EACvF,OAAO,cAAc;AAC1B;AAEA,SAAS,KAAK,IAA+B;AAC3C,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,WAAW,IAAuB,GAAmB;AAC5D,MAAI,GAAG,WAAW,EAAG,QAAO,OAAO;AACnC,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,SAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC,CAAC;AAC1F;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,SAAS,GAAW,GAAmB;AAC9C,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO;AAC9D,SAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAa,QAAwB;AACrD,MAAI,CAAC,OAAO,SAAS,GAAG,EAAG,QAAO;AAClC,MAAI,UAAU,EAAG,QAAO,OAAO,IAAI,IAAI;AACvC,SAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACQ;AACR,QAAM,OACJ,OAAO,SAAS,WAAW,cAAc,KAAK,OAAO,SAAS,QAAQ,WAAW,IAC7E,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IACxE;AACN,QAAM,UACJ,OAAO,SAAS,WAAW,YAAY,KAAK,OAAO,SAAS,QAAQ,SAAS,IACzE,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACpE;AACN,SAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;AACR,QAAM,SAAS,sBAAsB,MAAM,KAAK,MAAM;AACtD,QAAM,aAAa,aAAa,QAAQ,aAAa,eAAe,QAAQ,UAAU,gBAAgB,QAAQ,WAAW,aAAa,IAAI,QAAQ,QAAQ,CAAC,cAAc,IAAI,QAAQ,SAAS,CAAC;AAC/L,MAAI,OAAO,WAAW,EAAG,QAAO,GAAG,MAAM,KAAK,UAAU;AACxD,SAAO,GAAG,MAAM,KAAK,UAAU,YAAY,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAChF;AAEA,SAAS,IAAI,GAAmB;AAC9B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO,OAAO,CAAC;AACxC,SAAO,EAAE,QAAQ,CAAC;AACpB;;;AC3iBO,SAAS,YACd,UACA,WACA,UAA4B,CAAC,GACZ;AACjB,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,WAAW,QAAQ,mBAAmB;AAC5C,QAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;AAEpE,QAAM,eAAeA,MAAK,QAAQ;AAClC,QAAM,gBAAgBA,MAAK,SAAS;AACpC,QAAM,QAAQ,gBAAgB;AAE9B,MACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GACrB;AACA,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,SAAS;AAAA,MACT,YAAY;AAAA,MACZ;AAAA,MACA,SAAS;AAAA,IACX;AAAA,EACF;AAEA,QAAM,SAAmB,IAAI,MAAM,UAAU;AAC7C,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,YAAY,SAAS,UAAU,GAAG;AACxC,UAAM,YAAY,SAAS,WAAW,GAAG;AACzC,WAAO,CAAC,IAAIA,MAAK,SAAS,IAAIA,MAAK,SAAS;AAAA,EAC9C;AACA,SAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,QAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;AACpD,QAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;AAC5D,QAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ,CAAC;AAC5C,QAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ,CAAC;AAEzD,MAAI;AACJ,MAAI,UAAU,EAAG,WAAU;AAAA,WAClB,UAAU,EAAG,WAAU;AAAA,WACvB,SAAS,EAAG,WAAU;AAAA,MAC1B,WAAU;AAEf,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAASA,MAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,MAAI,IAAI;AACR,aAAW,KAAK,GAAI,MAAK;AACzB,SAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;AAC3D,QAAM,MAAM,IAAI,MAAM,GAAG,MAAM;AAC/B,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,KAAI,CAAC,IAAI,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM,CAAC;AAC7E,SAAO;AACT;AAGA,SAAS,WAAW,MAA4B;AAC9C,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,SAAK;AACL,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAGA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,IAAI;AACR,aAAW,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;AAC5B,UAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;AACjC,UAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;AACxC,eAAW,QAAQ,OAAO;AACxB,WAAK;AACL,UAAI,KAAK,KAAK,GAAG,QAAQ;AAAA,IAC3B;AAAA,EACF;AACA,SAAO,MAAM;AACf;AAiCA,eAAsB,gBACpB,MACkF;AAClF,QAAM,cAAc,KAAK,oBAAoB;AAC7C,QAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;AACnF,QAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;AACrF,QAAM,KAAK,YAAY,gBAAgB,iBAAiB;AAAA,IACtD,GAAI,KAAK,UAAU,SAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;AAAA,IACxD,GAAI,KAAK,eAAe,SAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;AAAA,IACvE,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,EACvD,CAAC;AACD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,iBAAiB,eAAe;AAAA,IAChC,kBAAkB,gBAAgB;AAAA,EACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;AACnB,QAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;AAClD,MAAI,OAAO;AACX,iBAAe,SAAwB;AACrC,WAAO,MAAM;AACX,YAAM,IAAI;AACV,UAAI,KAAK,QAAQ,OAAQ;AACzB,YAAM,IAAI,MAAM,MAAM,QAAQ,CAAC,CAAE;AACjC,cAAQ,CAAC,IAAI,OAAO,SAAS,CAAC,IAAI,IAAI;AAAA,IACxC;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,GAAG,MAAM,OAAO,CAAC,CAAC;AAClF,SAAO;AACT;;;AC1NO,SAAS,oBACd,WACA,UAAsC,CAAC,GAC/B;AACR,QAAM,QAAQ,QAAQ,SAAS,mBAAmB,UAAU,MAAM;AAClE,QAAM,QAAkB,CAAC;AACzB,QAAM,KAAK,KAAK,KAAK,EAAE;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,aAAa,UAAU,OAAO,YAAY,CAAC,IAAI;AAC1D,QAAM,KAAK,cAAc,UAAU,UAAU,QAAQ,IAAI,IAAI;AAC7D,MAAI,UAAU,YAAa,OAAM,KAAK,gBAAgB,UAAU,WAAW,IAAI;AAC/E,MAAI,UAAU,WAAY,OAAM,KAAK,eAAe,UAAU,UAAU,IAAI;AAC5E,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,UAAU,OAAO;AAC5B,QAAM,KAAK,EAAE;AAEb,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,oBAAoB;AAC/B,QAAM,KAAK,YAAY;AACvB,QAAM,KAAK,iBAAiB,UAAU,QAAQ,aAAa,IAAI;AAC/D,QAAM,KAAK,mBAAmB,UAAU,QAAQ,UAAU,IAAI;AAC9D,QAAM,KAAK,oBAAoB,UAAU,QAAQ,WAAW,IAAI;AAChE,QAAM,KAAK,iBAAiB,IAAI,UAAU,QAAQ,QAAQ,CAAC,IAAI;AAC/D,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,SAAS,CAAC,IAAI;AACjE,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,eAAe,CAAC,IAAI;AACxE,QAAM,KAAK,oBAAoB,IAAI,UAAU,QAAQ,gBAAgB,CAAC,IAAI;AAC1E,QAAM,KAAK,mBAAmB,IAAI,UAAU,QAAQ,UAAU,CAAC,IAAI;AACnE,QAAM,KAAK,kBAAkB,IAAI,UAAU,QAAQ,WAAW,CAAC,IAAI;AACnE,QAAM,KAAK,qBAAqB,KAAK,MAAM,UAAU,QAAQ,SAAS,CAAC,OAAO;AAC9E,QAAM,KAAK,EAAE;AAEb,MAAI,UAAU,OAAO,SAAS,GAAG;AAC/B,UAAM,KAAK,WAAW;AACtB,UAAM,KAAK,EAAE;AACb,eAAW,SAAS,UAAU,QAAQ;AACpC,YAAM,KAAK,OAAO,MAAM,QAAQ,QAAQ,MAAM,IAAI,OAAO,MAAM,IAAI,MAAM,MAAM,MAAM,EAAE;AAAA,IACzF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,wBAAwB;AACnE,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,yBAAyB;AACpC,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,SAAS,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,OAAO,KAAK,KAAK,EAAE;AAC5E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,WAAW,QAAQ,UAAU,QAAQ,iBAAiB;AAC5D,MAAI,SAAS,SAAS,GAAG;AACvB,UAAM,KAAK,kBAAkB;AAC7B,UAAM,KAAK,EAAE;AACb,eAAW,CAAC,MAAM,KAAK,KAAK,SAAU,OAAM,KAAK,KAAK,IAAI,KAAK,KAAK,EAAE;AACtE,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,QAAQ,QAAQ,KAAK,SAAS,GAAG;AAC3C,UAAM,KAAK,gBAAgB;AAC3B,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,aAAa,CAAC,GAAG,QAAQ,IAAI,GAAG;AAAA,QAC9B,YAAY,QAAQ,cAAc,UAAU,cAAc;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC,EAAE;AAAA,IACL;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,QAAQ,wBAAwB,QAAQ,qBAAqB,SAAS,GAAG;AAC3E,UAAM,KAAK,0BAA0B;AACrC,UAAM,KAAK,EAAE;AACb,eAAW,WAAW,QAAQ,qBAAsB,OAAM,KAAK,KAAK,OAAO,EAAE;AAC7E,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,QAAM,cAAc,QAAQ,eAAe,mBAAmB,SAAS;AACvE,MAAI,YAAY,SAAS,GAAG;AAC1B,UAAM,KAAK,iBAAiB;AAC5B,UAAM,KAAK,EAAE;AACb,eAAW,UAAU,YAAa,OAAM,KAAK,KAAK,MAAM,EAAE;AAC1D,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,EAAE,QAAQ,CAAC;AAAA;AACtC;AAEA,SAAS,mBAAmB,WAAiD;AAC3E,MAAI,UAAU,QAAS,QAAO,CAAC,kDAAkD;AACjF,SAAO,UAAU,OACd,OAAO,CAAC,UAAU,MAAM,aAAa,UAAU,EAC/C,IAAI,CAAC,UAAU,WAAW,MAAM,IAAI,KAAK,MAAM,MAAM,EAAE;AAC5D;AAEA,SAAS,QAAQ,QAAyD;AACxE,SAAO,OAAO,QAAQ,MAAM,EACzB,OAAO,CAAC,CAAC,EAAE,KAAK,MAAM,QAAQ,CAAC,EAC/B,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC;AAC3D;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC,MAAM;AACnE;AAEA,SAAS,IAAI,OAAuB;AAClC,SAAO,OAAO,SAAS,KAAK,IAAI,MAAM,QAAQ,CAAC,IAAI;AACrD;","names":["mean"]}

package/dist/{chunk-QIT2XZ4E.js → chunk-VNM52AGA.js}

@@ -1,6 +1,6 @@
 import {
   confidenceInterval
-} from "./chunk-HRGUJTER.js";
+} from "./chunk-HYGRFL7C.js";
 import {
   AgentEvalError
 } from "./chunk-3BFEG2F6.js";
@@ -140,7 +140,7 @@ function canonicalizeAt(value, path) {
       break;
   }
   const obj = value;
-  if (typeof obj["toJSON"] === "function") {
+  if (typeof obj.toJSON === "function") {
     return canonicalizeAt(obj.toJSON(), path);
   }
   if (Array.isArray(obj)) {
@@ -787,4 +787,4 @@ export {
   runCampaign,
   planCampaignRun
 };
-//# sourceMappingURL=chunk-QIT2XZ4E.js.map
+//# sourceMappingURL=chunk-VNM52AGA.js.map