@tangle-network/agent-eval 0.27.0 → 0.27.2

package/dist/chunk-K33INZHH.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/failure-taxonomy.ts","../src/pipelines/failure-cluster.ts","../src/tool-use-metrics.ts","../src/baseline.ts"],"sourcesContent":["/**\n * Failure taxonomy — canonical classes + a default classifier.\n *\n * Every failed run should end up in a named class. The classifier here\n * is rule-based (fast, deterministic); an LLM fallback can be added by\n * the consumer for novel cases and trained into the rule base over time.\n *\n * Consumers call `classifyFailure(run, spans, events)` and persist the\n * returned class as `Run.outcome.failureClass`.\n */\n\nimport type { FailureClass, Run, Span, TraceEvent } from './trace/schema'\nimport { FAILURE_CLASSES } from './trace/schema'\n\nexport { FAILURE_CLASSES, type FailureClass }\n\nexport interface FailureContext {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n}\n\nexport interface FailureClassification {\n  failureClass: FailureClass\n  reason: string\n  triggerSpanId?: string\n  triggerEventId?: string\n}\n\n/** Ordered rules — first match wins. */\nexport interface FailureRule {\n  id: string\n  match: (ctx: FailureContext) => {\n    failureClass: FailureClass\n    reason: string\n    triggerSpanId?: string\n    triggerEventId?: string\n  } | null\n}\n\nexport const DEFAULT_RULES: FailureRule[] = [\n  // Outcome already named? Respect it.\n  {\n    id: 'explicit-outcome',\n    match: ({ run }) => {\n      const fc = run.outcome?.failureClass\n      if (fc && fc !== 'unknown')\n        return { failureClass: fc, reason: 'outcome.failureClass set explicitly' }\n      return null\n    },\n  },\n  {\n    id: 'knowledge-readiness-blocked',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'readiness_scored' &&\n          e.payload.passed === false,\n      )\n      return event\n        ? {\n            failureClass: 'knowledge_readiness_blocked',\n            reason: 'knowledge readiness report blocked execution',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'bad-integration-manifest',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_manifest_validated' && e.payload.valid === false) ||\n            (e.payload.kind === 'integration_invoke_failed' &&\n              e.payload.code === 'manifest_invalid')),\n      )\n      return event\n        ? {\n            failureClass: 'bad_integration_manifest',\n            reason: 'integration manifest validation failed before launch',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-integration-connection',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_manifest_resolved' &&\n          hasResolutionStatus(e.payload, 'missing_connection'),\n      )\n      return event\n        ? {\n            failureClass: 'missing_integration_connection',\n            reason: 'required integration connection was missing',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-integration-scope',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_manifest_resolved' && hasMissingScopes(e.payload)) ||\n            (e.payload.kind === 'integration_invoke_failed' && e.payload.code === 'scope_denied')),\n      )\n      return event\n        ? {\n            failureClass: 'missing_integration_scope',\n            reason: 'integration grant or connection lacks required scopes',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-approval-required',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_invoke' && e.payload.status === 'approval_required') ||\n            (e.payload.kind === 'integration_invoke_failed' &&\n              e.payload.code === 'approval_required') ||\n            e.payload.kind === 'integration_approval_required'),\n      )\n      return event\n        ? {\n            failureClass: 'integration_approval_required',\n            reason: 'integration write paused for user approval',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-auth-expired',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          (e.payload.code === 'auth_expired' ||\n            e.payload.code === 'connection_not_active' ||\n            e.payload.code === 'capability_expired' ||\n            e.payload.status === 'expired'),\n      )\n      return event\n        ? {\n            failureClass: 'integration_auth_expired',\n            reason: 'integration connection or capability expired',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'unsafe-integration-write-denied',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          (e.payload.code === 'unsafe_write_denied' ||\n            e.payload.code === 'policy_denied' ||\n            e.payload.code === 'action_denied'),\n      )\n      return event\n        ? {\n            failureClass: 'unsafe_integration_write_denied',\n            reason: 'integration write was denied by policy or capability scope',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-provider-failure',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          ![\n            'scope_denied',\n            'approval_required',\n            'auth_expired',\n            'connection_not_active',\n            'capability_expired',\n            'unsafe_write_denied',\n            'policy_denied',\n            'action_denied',\n            'manifest_invalid',\n          ].includes(String(e.payload.code)),\n      )\n      return event\n        ? {\n            failureClass: 'integration_provider_failure',\n            reason: 'integration provider invocation failed',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-credentials',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.category === 'credential_or_secret',\n      )\n      return event\n        ? {\n            failureClass: 'missing_credentials',\n            reason: 'required credential or secret was missing',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'bad-retrieval',\n    match: ({ run, spans }) => {\n      if (run.outcome?.pass !== false) return null\n      const retrieval = spans.find(\n        (s) =>\n          s.kind === 'retrieval' && (s.hits.length === 0 || s.hits.every((hit) => hit.score <= 0)),\n      )\n      return retrieval\n        ? {\n            failureClass: 'bad_retrieval',\n            reason: 'retrieval returned no useful hits for a failed run',\n            triggerSpanId: retrieval.spanId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'insufficient-evidence',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.reason === 'insufficient_evidence',\n      )\n      return event\n        ? {\n            failureClass: 'insufficient_evidence',\n            reason: 'task proceeded with insufficient supporting evidence',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'contradictory-evidence',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.reason === 'contradictory_evidence',\n      )\n      return event\n        ? {\n            failureClass: 'contradictory_evidence',\n            reason: 'supporting evidence contradicted itself',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  // Budget breach events\n  {\n    id: 'budget-breach',\n    match: ({ events }) => {\n      const breach = events.find((e) => e.kind === 'budget_breach')\n      return breach\n        ? {\n            failureClass: 'budget_exceeded',\n            reason: `budget breached on ${breach.payload.dimension ?? 'unknown dimension'}`,\n            triggerEventId: breach.eventId,\n          }\n        : null\n    },\n  },\n  // Policy violations\n  {\n    id: 'policy-violation',\n    match: ({ events }) => {\n      const e = events.find((x) => x.kind === 'policy_violation')\n      return e\n        ? {\n            failureClass: 'policy_violation',\n            reason: 'policy_violation event emitted',\n            triggerEventId: e.eventId,\n          }\n        : null\n    },\n  },\n  // Sandbox non-zero exit code\n  {\n    id: 'sandbox-failure',\n    match: ({ spans }) => {\n      const s = spans.find(\n        (x) => x.kind === 'sandbox' && typeof x.exitCode === 'number' && x.exitCode !== 0,\n      )\n      if (!s) return null\n      return {\n        failureClass: 'sandbox_failure',\n        reason: `sandbox exited ${(s as Extract<Span, { kind: 'sandbox' }>).exitCode}`,\n        triggerSpanId: s.spanId,\n      }\n    },\n  },\n  // Timeout: run aborted by external signal\n  {\n    id: 'timeout',\n    match: ({ run, events }) => {\n      if (run.status !== 'aborted') return null\n      const hasTimeout = events.some(\n        (e) =>\n          e.kind === 'error' &&\n          String(e.payload.reason ?? '')\n            .toLowerCase()\n            .includes('timeout'),\n      )\n      const note = (run.outcome?.notes ?? '').toLowerCase()\n      if (hasTimeout || note.includes('timeout') || note.includes('deadline')) {\n        return { failureClass: 'timeout', reason: 'timeout signal observed' }\n      }\n      return null\n    },\n  },\n  // Tool recovery failure: many consecutive tool errors on the same tool\n  {\n    id: 'tool-recovery-failure',\n    match: ({ spans }) => {\n      const tools = spans.filter((s) => s.kind === 'tool')\n      const byTool = new Map<string, Span[]>()\n      for (const t of tools) {\n        const name = (t as Extract<Span, { kind: 'tool' }>).toolName\n        const arr = byTool.get(name) ?? []\n        arr.push(t)\n        byTool.set(name, arr)\n      }\n      for (const [name, arr] of byTool) {\n        const errs = arr.filter((s) => s.status === 'error')\n        if (errs.length >= 3 && errs.length === arr.length) {\n          return {\n            failureClass: 'tool_recovery_failure',\n            reason: `${errs.length} consecutive errors on tool \"${name}\"`,\n            triggerSpanId: errs[errs.length - 1]!.spanId,\n          }\n        }\n      }\n      return null\n    },\n  },\n  // Tool selection error: the run failed and agent called zero tools despite having them\n  {\n    id: 'tool-selection-error',\n    match: ({ run, spans }) => {\n      if (run.outcome?.pass !== false) return null\n      const hasToolsAvailable = spans.some(\n        (s) =>\n          s.kind === 'agent' &&\n          (s.attributes?.toolsAvailable as number | undefined) !== undefined &&\n          (s.attributes?.toolsAvailable as number) > 0,\n      )\n      const tools = spans.filter((s) => s.kind === 'tool')\n      if (hasToolsAvailable && tools.length === 0) {\n        return {\n          failureClass: 'tool_selection_error',\n          reason: 'tools were available but none were called',\n        }\n      }\n      return null\n    },\n  },\n  // Format drift: scored by a judge with dimension='format' below threshold\n  {\n    id: 'format-drift',\n    match: ({ spans }) => {\n      const judge = spans.find(\n        (s) =>\n          s.kind === 'judge' &&\n          (s as Extract<Span, { kind: 'judge' }>).dimension === 'format' &&\n          (s as Extract<Span, { kind: 'judge' }>).score < 0.5,\n      )\n      return judge\n        ? {\n            failureClass: 'format_drift',\n            reason: 'format judge scored below 0.5',\n            triggerSpanId: judge.spanId,\n          }\n        : null\n    },\n  },\n]\n\nfunction hasResolutionStatus(payload: Record<string, unknown>, status: string): boolean {\n  if (status === 'missing_connection' && stringArray(payload.missingConnections).length > 0)\n    return true\n  return resolutionItems(payload).some((item) => item.status === status)\n}\n\nfunction hasMissingScopes(payload: Record<string, unknown>): boolean {\n  if (stringArray(payload.missingScopes).length > 0) return true\n  return resolutionItems(payload).some(\n    (item) => Array.isArray(item.missingScopes) && item.missingScopes.length > 0,\n  )\n}\n\nfunction resolutionItems(payload: Record<string, unknown>): Array<Record<string, unknown>> {\n  return [\n    ...records(payload.missing),\n    ...records(payload.optionalMissing),\n    ...records(payload.ready),\n  ]\n}\n\nfunction records(value: unknown): Array<Record<string, unknown>> {\n  if (!Array.isArray(value)) return []\n  return value.filter(\n    (item): item is Record<string, unknown> =>\n      Boolean(item) && typeof item === 'object' && !Array.isArray(item),\n  )\n}\n\nfunction stringArray(value: unknown): string[] {\n  return Array.isArray(value)\n    ? value.filter((item): item is string => typeof item === 'string')\n    : []\n}\n\n/** Classify the failure mode of a run using an ordered rule list. */\nexport function classifyFailure(\n  ctx: FailureContext,\n  rules: FailureRule[] = DEFAULT_RULES,\n): FailureClassification {\n  if (ctx.run.outcome?.pass !== false && ctx.run.status === 'completed') {\n    return { failureClass: 'success', reason: 'run completed with pass=true (or no explicit fail)' }\n  }\n  for (const rule of rules) {\n    const hit = rule.match(ctx)\n    if (hit) return hit\n  }\n  return { failureClass: 'unknown', reason: 'no rule matched; run failed for unclassified reason' }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * Tool-use metrics — derived purely from trace data.\n *\n * No scoring assumptions: consumers supply optional ground-truth tool\n * selections per turn + optional \"information used downstream\" signals.\n * Without those, we still compute descriptive metrics (error rate,\n * retry rate, duplicate-call rate) that are useful on their own.\n */\n\nimport { argHash, groupBy, toolSpans } from './trace/query'\nimport type { Span } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface ToolUseMetrics {\n  runId: string\n  totalCalls: number\n  byTool: Record<string, ToolStats>\n  errorRate: number\n  /** Ratio of calls with identical (toolName, argHash) already seen earlier in the same run. */\n  duplicateRate: number\n  /** Ratio of error calls followed by ≥1 retry on same tool. */\n  retryRate: number\n  /** Optional: of the calls agent made, fraction the evaluator marked as \"correct selection\". */\n  selectionAccuracy?: number\n}\n\nexport interface ToolStats {\n  calls: number\n  errors: number\n  avgLatencyMs: number\n  duplicates: number\n}\n\nexport interface ToolUseOptions {\n  /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */\n  selectionLabels?: Record<string, boolean>\n}\n\nexport async function computeToolUseMetrics(\n  store: TraceStore,\n  runId: string,\n  options: ToolUseOptions = {},\n): Promise<ToolUseMetrics> {\n  const tools = await toolSpans(store, runId)\n  if (tools.length === 0) {\n    return { runId, totalCalls: 0, byTool: {}, errorRate: 0, duplicateRate: 0, retryRate: 0 }\n  }\n\n  const byTool: Record<string, ToolStats> = {}\n  let totalErrors = 0\n  let totalDuplicates = 0\n  const sortedTools = [...tools].sort((a, b) => a.startedAt - b.startedAt)\n  const seenSignatures = new Set<string>()\n\n  // duplicate detection + per-tool aggregation\n  for (const t of sortedTools) {\n    const stat = (byTool[t.toolName] ??= { calls: 0, errors: 0, avgLatencyMs: 0, duplicates: 0 })\n    stat.calls += 1\n    if (t.status === 'error') {\n      stat.errors += 1\n      totalErrors += 1\n    }\n    if (typeof t.latencyMs === 'number') stat.avgLatencyMs += t.latencyMs\n    const sig = `${t.toolName}|${argHash(t.args)}`\n    if (seenSignatures.has(sig)) {\n      stat.duplicates += 1\n      totalDuplicates += 1\n    }\n    seenSignatures.add(sig)\n  }\n\n  for (const stat of Object.values(byTool)) {\n    stat.avgLatencyMs = stat.calls > 0 ? stat.avgLatencyMs / stat.calls : 0\n  }\n\n  // retry detection: per-tool chronological adjacency where error → next same-tool call\n  let retryOpportunities = 0\n  let retriesFollowed = 0\n  for (const [, arr] of groupBy(sortedTools, (t) => t.toolName)) {\n    for (let i = 0; i < arr.length; i++) {\n      if (arr[i]!.status !== 'error') continue\n      retryOpportunities += 1\n      if (arr[i + 1]) retriesFollowed += 1\n    }\n  }\n  const retryRate = retryOpportunities > 0 ? retriesFollowed / retryOpportunities : 0\n\n  let selectionAccuracy: number | undefined\n  if (options.selectionLabels) {\n    const labeled = sortedTools.filter((t) => t.spanId in options.selectionLabels!)\n    if (labeled.length > 0) {\n      selectionAccuracy =\n        labeled.filter((t) => options.selectionLabels![t.spanId]).length / labeled.length\n    }\n  }\n\n  return {\n    runId,\n    totalCalls: sortedTools.length,\n    byTool,\n    errorRate: totalErrors / sortedTools.length,\n    duplicateRate: totalDuplicates / sortedTools.length,\n    retryRate,\n    selectionAccuracy,\n  }\n}\n\nexport type { Span }\n","/**\n * Baseline regression detection.\n *\n * Lifted from ADC baseline.ts. Every promotion-blocking signal boils down\n * to: \"is this run measurably worse than baseline?\" — with enough\n * statistical rigor to distinguish noise from drift.\n *\n * Uses:\n *   - Welch's t-test (unequal variance) for per-metric mean comparison\n *   - Cohen's d for effect size magnitude\n *   - IQR for stability flag (unstable samples can't be trusted for comparisons)\n *\n * Returns a structured verdict: improved | regressed | stable | unstable.\n */\n\nimport { cohensD } from './statistics'\n\nexport interface MetricSamples {\n  /** Stable metric key (e.g. \"overallScore\", \"firstTokenMs\"). */\n  metric: string\n  /** Whether higher values are better. */\n  higherIsBetter: boolean\n  baseline: number[]\n  candidate: number[]\n}\n\nexport interface MetricVerdict {\n  metric: string\n  baselineMean: number\n  candidateMean: number\n  delta: number\n  cohensD: number\n  welchT: number\n  welchDf: number\n  welchP: number\n  stable: boolean\n  /** IQR of the combined samples — used as a rough stability indicator. */\n  iqr: number\n  verdict: 'improved' | 'regressed' | 'stable' | 'unstable'\n}\n\nexport interface BaselineReport {\n  metrics: MetricVerdict[]\n  /** True if any critical metric regressed. */\n  hasRegression: boolean\n  /** True if any metric is unstable (too noisy to judge). */\n  hasUnstable: boolean\n}\n\nexport interface BaselineOptions {\n  /** Effect size threshold for meaningful delta (default 0.5 — medium effect). */\n  effectThreshold?: number\n  /** p-value threshold for statistical significance (default 0.05). */\n  alpha?: number\n  /** IQR/mean ratio above which samples are flagged unstable (default 0.30). */\n  unstableCvThreshold?: number\n}\n\n/**\n * Compare candidate samples against baseline per metric. Verdict logic:\n *   - unstable: IQR/|mean| > threshold on either set — not enough signal\n *   - improved: meaningful effect in the \"better\" direction AND p < alpha\n *   - regressed: meaningful effect in the \"worse\" direction AND p < alpha\n *   - stable: otherwise (no significant change)\n */\nexport function compareToBaseline(\n  samples: MetricSamples[],\n  options: BaselineOptions = {},\n): BaselineReport {\n  const effectThreshold = options.effectThreshold ?? 0.5\n  const alpha = options.alpha ?? 0.05\n  const cvThreshold = options.unstableCvThreshold ?? 0.3\n\n  const metrics: MetricVerdict[] = samples.map((s) => {\n    if (s.baseline.length < 2 || s.candidate.length < 2) {\n      throw new Error(`compareToBaseline: need ≥2 samples per side for \"${s.metric}\"`)\n    }\n    const bMean = mean(s.baseline)\n    const cMean = mean(s.candidate)\n    const delta = cMean - bMean\n    const d = cohensD(s.baseline, s.candidate) // positive = candidate higher\n    const { t, df, p } = welchsTTest(s.baseline, s.candidate)\n    // Stability is per-side: a comparison is trustworthy only when BOTH\n    // samples are internally consistent. Combining the sides would flag\n    // large-but-real deltas as \"unstable\" which is exactly what we want\n    // to detect.\n    const baselineIqr = iqr(s.baseline)\n    const candidateIqr = iqr(s.candidate)\n    const baselineStable = baselineIqr / Math.max(Math.abs(bMean), 1e-9) <= cvThreshold\n    const candidateStable = candidateIqr / Math.max(Math.abs(cMean), 1e-9) <= cvThreshold\n    const stable = baselineStable && candidateStable\n    const reportedIqr = Math.max(baselineIqr, candidateIqr)\n\n    let verdict: MetricVerdict['verdict']\n    if (!stable) {\n      verdict = 'unstable'\n    } else if (p < alpha && Math.abs(d) >= effectThreshold) {\n      const candidateIsBetter = s.higherIsBetter ? delta > 0 : delta < 0\n      verdict = candidateIsBetter ? 'improved' : 'regressed'\n    } else {\n      verdict = 'stable'\n    }\n\n    return {\n      metric: s.metric,\n      baselineMean: bMean,\n      candidateMean: cMean,\n      delta,\n      cohensD: d,\n      welchT: t,\n      welchDf: df,\n      welchP: p,\n      stable,\n      iqr: reportedIqr,\n      verdict,\n    }\n  })\n\n  return {\n    metrics,\n    hasRegression: metrics.some((m) => m.verdict === 'regressed'),\n    hasUnstable: metrics.some((m) => m.verdict === 'unstable'),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((a, b) => a + b, 0) / xs.length\n}\n\n/** Inter-quartile range; 0 when the sample has no spread. */\nexport function iqr(xs: number[]): number {\n  if (xs.length === 0) return 0\n  const sorted = [...xs].sort((a, b) => a - b)\n  const q = (p: number) => {\n    const idx = p * (sorted.length - 1)\n    const lo = Math.floor(idx)\n    const hi = Math.ceil(idx)\n    return sorted[lo]! + (sorted[hi]! - sorted[lo]!) * (idx - lo)\n  }\n  return q(0.75) - q(0.25)\n}\n\n/**\n * Welch's t-test — unequal-variance two-sample t. Uses the same Student-t\n * CDF as `pairedTTest` (via incomplete beta); falls back to normal tail\n * when df is large.\n */\nexport function welchsTTest(a: number[], b: number[]): { t: number; df: number; p: number } {\n  if (a.length < 2 || b.length < 2) return { t: 0, df: 0, p: 1 }\n  const mA = mean(a)\n  const mB = mean(b)\n  const vA = variance(a, mA)\n  const vB = variance(b, mB)\n  const seSquared = vA / a.length + vB / b.length\n  if (seSquared === 0) return { t: mA === mB ? 0 : Infinity, df: 0, p: mA === mB ? 1 : 0 }\n  const t = (mB - mA) / Math.sqrt(seSquared)\n  const df =\n    (seSquared * seSquared) /\n    ((vA / a.length) ** 2 / (a.length - 1) + (vB / b.length) ** 2 / (b.length - 1))\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\nfunction variance(xs: number[], m: number): number {\n  return xs.reduce((acc, x) => acc + (x - m) ** 2, 0) / (xs.length - 1)\n}\n\n// Re-used from statistics.ts via small local copy to avoid exporting internals.\nfunction studentTCdf(t: number, df: number): number {\n  if (df <= 0) return 0.5\n  if (df > 100) return normalCdf(t)\n  const x = df / (df + t * t)\n  const ib = incompleteBeta(x, df / 2, 0.5)\n  return t >= 0 ? 1 - 0.5 * ib : 0.5 * ib\n}\n\nfunction incompleteBeta(x: number, a: number, b: number): number {\n  if (x <= 0) return 0\n  if (x >= 1) return 1\n  const lnBeta = lnGamma(a) + lnGamma(b) - lnGamma(a + b)\n  const front = Math.exp(Math.log(x) * a + Math.log(1 - x) * b - lnBeta) / a\n  let c = 1\n  let d = 1 - ((a + b) * x) / (a + 1)\n  if (Math.abs(d) < 1e-30) d = 1e-30\n  d = 1 / d\n  let f = d\n  for (let m = 1; m <= 200; m++) {\n    const m2 = 2 * m\n    let num = (m * (b - m) * x) / ((a + m2 - 1) * (a + m2))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    f *= d * c\n    num = -((a + m) * (a + b + m) * x) / ((a + m2) * (a + m2 + 1))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    const delta = d * c\n    f *= delta\n    if (Math.abs(delta - 1) < 3e-7) break\n  }\n  return front * f\n}\n\nfunction lnGamma(z: number): number {\n  const coefs = [\n    0.99999999999980993, 676.5203681218851, -1259.1392167224028, 771.32342877765313,\n    -176.61502916214059, 12.507343278686905, -0.13857109526572012, 9.9843695780195716e-6,\n    1.5056327351493116e-7,\n  ]\n  if (z < 0.5) return Math.log(Math.PI / Math.sin(Math.PI * z)) - lnGamma(1 - z)\n  z -= 1\n  let x = coefs[0]!\n  for (let i = 1; i < 9; i++) x += coefs[i]! / (z + i)\n  const t = z + 7.5\n  return 0.5 * Math.log(2 * Math.PI) + (z + 0.5) * Math.log(t) - t + Math.log(x)\n}\n\nfunction normalCdf(x: number): number {\n  const a1 = 0.254829592\n  const a2 = -0.284496736\n  const a3 = 1.421413741\n  const a4 = -1.453152027\n  const a5 = 1.061405429\n  const p = 0.3275911\n  const sign = x < 0 ? -1 : 1\n  const absX = Math.abs(x)\n  const t = 1 / (1 + p * absX)\n  const y = 1 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp((-absX * absX) / 2)\n  return 0.5 * (1 + sign * y)\n}\n"],"mappings":";;;;;;;;;;AAwCO,IAAM,gBAA+B;AAAA;AAAA,EAE1C;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,IAAI,MAAM;AAClB,YAAM,KAAK,IAAI,SAAS;AACxB,UAAI,MAAM,OAAO;AACf,eAAO,EAAE,cAAc,IAAI,QAAQ,sCAAsC;AAC3E,aAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,sBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,oCAAoC,EAAE,QAAQ,UAAU,SAC1E,EAAE,QAAQ,SAAS,+BAClB,EAAE,QAAQ,SAAS;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mCACnB,oBAAoB,EAAE,SAAS,oBAAoB;AAAA,MACvD;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,mCAAmC,iBAAiB,EAAE,OAAO,KAC/E,EAAE,QAAQ,SAAS,+BAA+B,EAAE,QAAQ,SAAS;AAAA,MAC5E;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,wBAAwB,EAAE,QAAQ,WAAW,uBAC/D,EAAE,QAAQ,SAAS,+BAClB,EAAE,QAAQ,SAAS,uBACrB,EAAE,QAAQ,SAAS;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,gCAClB,EAAE,QAAQ,SAAS,kBAClB,EAAE,QAAQ,SAAS,2BACnB,EAAE,QAAQ,SAAS,wBACnB,EAAE,QAAQ,WAAW;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,gCAClB,EAAE,QAAQ,SAAS,yBAClB,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,SAAS;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,+BACnB,CAAC;AAAA,UACC;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF,EAAE,SAAS,OAAO,EAAE,QAAQ,IAAI,CAAC;AAAA,MACrC;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,aAAa;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,MAAM,MAAM;AACzB,UAAI,IAAI,SAAS,SAAS,MAAO,QAAO;AACxC,YAAM,YAAY,MAAM;AAAA,QACtB,CAAC,MACC,EAAE,SAAS,gBAAgB,EAAE,KAAK,WAAW,KAAK,EAAE,KAAK,MAAM,CAAC,QAAQ,IAAI,SAAS,CAAC;AAAA,MAC1F;AACA,aAAO,YACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,eAAe,UAAU;AAAA,MAC3B,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,SAAS,eAAe;AAC5D,aAAO,SACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ,sBAAsB,OAAO,QAAQ,aAAa,mBAAmB;AAAA,QAC7E,gBAAgB,OAAO;AAAA,MACzB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,IAAI,OAAO,KAAK,CAAC,MAAM,EAAE,SAAS,kBAAkB;AAC1D,aAAO,IACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,EAAE;AAAA,MACpB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,IAAI,MAAM;AAAA,QACd,CAAC,MAAM,EAAE,SAAS,aAAa,OAAO,EAAE,aAAa,YAAY,EAAE,aAAa;AAAA,MAClF;AACA,UAAI,CAAC,EAAG,QAAO;AACf,aAAO;AAAA,QACL,cAAc;AAAA,QACd,QAAQ,kBAAmB,EAAyC,QAAQ;AAAA,QAC5E,eAAe,EAAE;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,OAAO,MAAM;AAC1B,UAAI,IAAI,WAAW,UAAW,QAAO;AACrC,YAAM,aAAa,OAAO;AAAA,QACxB,CAAC,MACC,EAAE,SAAS,WACX,OAAO,EAAE,QAAQ,UAAU,EAAE,EAC1B,YAAY,EACZ,SAAS,SAAS;AAAA,MACzB;AACA,YAAM,QAAQ,IAAI,SAAS,SAAS,IAAI,YAAY;AACpD,UAAI,cAAc,KAAK,SAAS,SAAS,KAAK,KAAK,SAAS,UAAU,GAAG;AACvE,eAAO,EAAE,cAAc,WAAW,QAAQ,0BAA0B;AAAA,MACtE;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM;AACnD,YAAM,SAAS,oBAAI,IAAoB;AACvC,iBAAW,KAAK,OAAO;AACrB,cAAM,OAAQ,EAAsC;AACpD,cAAM,MAAM,OAAO,IAAI,IAAI,KAAK,CAAC;AACjC,YAAI,KAAK,CAAC;AACV,eAAO,IAAI,MAAM,GAAG;AAAA,MACtB;AACA,iBAAW,CAAC,MAAM,GAAG,KAAK,QAAQ;AAChC,cAAM,OAAO,IAAI,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnD,YAAI,KAAK,UAAU,KAAK,KAAK,WAAW,IAAI,QAAQ;AAClD,iBAAO;AAAA,YACL,cAAc;AAAA,YACd,QAAQ,GAAG,KAAK,MAAM,gCAAgC,IAAI;AAAA,YAC1D,eAAe,KAAK,KAAK,SAAS,CAAC,EAAG;AAAA,UACxC;AAAA,QACF;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,MAAM,MAAM;AACzB,UAAI,IAAI,SAAS,SAAS,MAAO,QAAO;AACxC,YAAM,oBAAoB,MAAM;AAAA,QAC9B,CAAC,MACC,EAAE,SAAS,WACV,EAAE,YAAY,mBAA0C,UACxD,EAAE,YAAY,iBAA4B;AAAA,MAC/C;AACA,YAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM;AACnD,UAAI,qBAAqB,MAAM,WAAW,GAAG;AAC3C,eAAO;AAAA,UACL,cAAc;AAAA,UACd,QAAQ;AAAA,QACV;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,QAAQ,MAAM;AAAA,QAClB,CAAC,MACC,EAAE,SAAS,WACV,EAAuC,cAAc,YACrD,EAAuC,QAAQ;AAAA,MACpD;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,eAAe,MAAM;AAAA,MACvB,IACA;AAAA,IACN;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,SAAkC,QAAyB;AACtF,MAAI,WAAW,wBAAwB,YAAY,QAAQ,kBAAkB,EAAE,SAAS;AACtF,WAAO;AACT,SAAO,gBAAgB,OAAO,EAAE,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM;AACvE;AAEA,SAAS,iBAAiB,SAA2C;AACnE,MAAI,YAAY,QAAQ,aAAa,EAAE,SAAS,EAAG,QAAO;AAC1D,SAAO,gBAAgB,OAAO,EAAE;AAAA,IAC9B,CAAC,SAAS,MAAM,QAAQ,KAAK,aAAa,KAAK,KAAK,cAAc,SAAS;AAAA,EAC7E;AACF;AAEA,SAAS,gBAAgB,SAAkE;AACzF,SAAO;AAAA,IACL,GAAG,QAAQ,QAAQ,OAAO;AAAA,IAC1B,GAAG,QAAQ,QAAQ,eAAe;AAAA,IAClC,GAAG,QAAQ,QAAQ,KAAK;AAAA,EAC1B;AACF;AAEA,SAAS,QAAQ,OAAgD;AAC/D,MAAI,CAAC,MAAM,QAAQ,KAAK,EAAG,QAAO,CAAC;AACnC,SAAO,MAAM;AAAA,IACX,CAAC,SACC,QAAQ,IAAI,KAAK,OAAO,SAAS,YAAY,CAAC,MAAM,QAAQ,IAAI;AAAA,EACpE;AACF;AAEA,SAAS,YAAY,OAA0B;AAC7C,SAAO,MAAM,QAAQ,KAAK,IACtB,MAAM,OAAO,CAAC,SAAyB,OAAO,SAAS,QAAQ,IAC/D,CAAC;AACP;AAGO,SAAS,gBACd,KACA,QAAuB,eACA;AACvB,MAAI,IAAI,IAAI,SAAS,SAAS,SAAS,IAAI,IAAI,WAAW,aAAa;AACrE,WAAO,EAAE,cAAc,WAAW,QAAQ,qDAAqD;AAAA,EACjG;AACA,aAAW,QAAQ,OAAO;AACxB,UAAM,MAAM,KAAK,MAAM,GAAG;AAC1B,QAAI,IAAK,QAAO;AAAA,EAClB;AACA,SAAO,EAAE,cAAc,WAAW,QAAQ,sDAAsD;AAClG;;;ACvaA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;AAC/B,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,UAAU,QAAQ,kBAAkB;AAC1C,QAAM,OAAO,MAAM,MAAM,SAAS;AAGlC,QAAM,WAAW,oBAAI,IAAyB;AAC9C,MAAI,gBAAgB;AAEpB,aAAW,OAAO,MAAM;AACtB,QAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,MAAO;AAC/D;AACA,UAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACpD,UAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;AACtD,UAAM,MAAM,gBAAgB,EAAE,KAAK,OAAO,OAAO,GAAG,KAAK;AAEzD,QAAI;AACJ,QAAI;AACJ,QAAI;AACJ,QAAI,IAAI,eAAe;AACrB,YAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,IAAI,aAAa;AAC7D,UAAI,MAAM,SAAS,QAAQ;AACzB,mBAAW,KAAK;AAChB,oBAAY,QAAQ,KAAK,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC5C,WAAW,MAAM,SAAS,SAAS;AACjC,oBAAY,KAAK;AAAA,MACnB;AAAA,IACF;AAEA,QAAI,CAAC,UAAU;AACb,YAAM,KAAK,MAAM,UAAU,OAAO,IAAI,KAAK;AAC3C,YAAM,UAAU,GAAG,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO,EAAE,IAAI;AAC3D,UAAI,SAAS;AACX,mBAAW,QAAQ;AACnB,oBAAY,QAAQ,QAAQ,IAAI,EAAE,MAAM,GAAG,EAAE;AAAA,MAC/C;AAAA,IACF;AAIA,QAAI,CAAC,WAAW;AACd,YAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;AACrF,UAAI,OAAO,SAAS,QAAS,aAAY,MAAM;AAAA,IACjD;AAEA,UAAM,MAAM,GAAG,IAAI,YAAY,IAAI,YAAY,EAAE,IAAI,aAAa,EAAE,IAAI,aAAa,EAAE;AACvF,QAAI,UAAU,SAAS,IAAI,GAAG;AAC9B,QAAI,CAAC,SAAS;AACZ,gBAAU;AAAA,QACR,cAAc,IAAI;AAAA,QAClB;AAAA,QACA;AAAA,QACA;AAAA,QACA,UAAU;AAAA,QACV,aAAa,CAAC;AAAA,QACd,cAAc,IAAI;AAAA,QAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;AAAA,MAChD;AACA,eAAS,IAAI,KAAK,OAAO;AAAA,IAC3B;AACA,YAAQ;AACR,QAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,EAAG,SAAQ,YAAY,KAAK,IAAI,UAAU;AAAA,EAC5F;AAEA,QAAM,MAAM,CAAC,GAAG,SAAS,OAAO,CAAC,EAC9B,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,WAAW,EAAE,QAAQ;AAEzC,SAAO,EAAE,UAAU,KAAK,eAAe,WAAW,KAAK,OAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;AAC5D,QAAM,UAAU,MAAM,KAAK,CAAC,MAAM,EAAE,WAAW,OAAO;AACtD,SAAO,SAAS;AAClB;;;AC7EA,eAAsB,sBACpB,OACA,OACA,UAA0B,CAAC,GACF;AACzB,QAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,MAAI,MAAM,WAAW,GAAG;AACtB,WAAO,EAAE,OAAO,YAAY,GAAG,QAAQ,CAAC,GAAG,WAAW,GAAG,eAAe,GAAG,WAAW,EAAE;AAAA,EAC1F;AAEA,QAAM,SAAoC,CAAC;AAC3C,MAAI,cAAc;AAClB,MAAI,kBAAkB;AACtB,QAAM,cAAc,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACvE,QAAM,iBAAiB,oBAAI,IAAY;AAGvC,aAAW,KAAK,aAAa;AAC3B,UAAM,OAAQ,OAAO,EAAE,QAAQ,MAAM,EAAE,OAAO,GAAG,QAAQ,GAAG,cAAc,GAAG,YAAY,EAAE;AAC3F,SAAK,SAAS;AACd,QAAI,EAAE,WAAW,SAAS;AACxB,WAAK,UAAU;AACf,qBAAe;AAAA,IACjB;AACA,QAAI,OAAO,EAAE,cAAc,SAAU,MAAK,gBAAgB,EAAE;AAC5D,UAAM,MAAM,GAAG,EAAE,QAAQ,IAAI,QAAQ,EAAE,IAAI,CAAC;AAC5C,QAAI,eAAe,IAAI,GAAG,GAAG;AAC3B,WAAK,cAAc;AACnB,yBAAmB;AAAA,IACrB;AACA,mBAAe,IAAI,GAAG;AAAA,EACxB;AAEA,aAAW,QAAQ,OAAO,OAAO,MAAM,GAAG;AACxC,SAAK,eAAe,KAAK,QAAQ,IAAI,KAAK,eAAe,KAAK,QAAQ;AAAA,EACxE;AAGA,MAAI,qBAAqB;AACzB,MAAI,kBAAkB;AACtB,aAAW,CAAC,EAAE,GAAG,KAAK,QAAQ,aAAa,CAAC,MAAM,EAAE,QAAQ,GAAG;AAC7D,aAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,UAAI,IAAI,CAAC,EAAG,WAAW,QAAS;AAChC,4BAAsB;AACtB,UAAI,IAAI,IAAI,CAAC,EAAG,oBAAmB;AAAA,IACrC;AAAA,EACF;AACA,QAAM,YAAY,qBAAqB,IAAI,kBAAkB,qBAAqB;AAElF,MAAI;AACJ,MAAI,QAAQ,iBAAiB;AAC3B,UAAM,UAAU,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,eAAgB;AAC9E,QAAI,QAAQ,SAAS,GAAG;AACtB,0BACE,QAAQ,OAAO,CAAC,MAAM,QAAQ,gBAAiB,EAAE,MAAM,CAAC,EAAE,SAAS,QAAQ;AAAA,IAC/E;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,YAAY;AAAA,IACxB;AAAA,IACA,WAAW,cAAc,YAAY;AAAA,IACrC,eAAe,kBAAkB,YAAY;AAAA,IAC7C;AAAA,IACA;AAAA,EACF;AACF;;;ACxCO,SAAS,kBACd,SACA,UAA2B,CAAC,GACZ;AAChB,QAAM,kBAAkB,QAAQ,mBAAmB;AACnD,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,cAAc,QAAQ,uBAAuB;AAEnD,QAAM,UAA2B,QAAQ,IAAI,CAAC,MAAM;AAClD,QAAI,EAAE,SAAS,SAAS,KAAK,EAAE,UAAU,SAAS,GAAG;AACnD,YAAM,IAAI,MAAM,yDAAoD,EAAE,MAAM,GAAG;AAAA,IACjF;AACA,UAAM,QAAQ,KAAK,EAAE,QAAQ;AAC7B,UAAM,QAAQ,KAAK,EAAE,SAAS;AAC9B,UAAM,QAAQ,QAAQ;AACtB,UAAM,IAAI,QAAQ,EAAE,UAAU,EAAE,SAAS;AACzC,UAAM,EAAE,GAAG,IAAI,EAAE,IAAI,YAAY,EAAE,UAAU,EAAE,SAAS;AAKxD,UAAM,cAAc,IAAI,EAAE,QAAQ;AAClC,UAAM,eAAe,IAAI,EAAE,SAAS;AACpC,UAAM,iBAAiB,cAAc,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI,KAAK;AACxE,UAAM,kBAAkB,eAAe,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI,KAAK;AAC1E,UAAM,SAAS,kBAAkB;AACjC,UAAM,cAAc,KAAK,IAAI,aAAa,YAAY;AAEtD,QAAI;AACJ,QAAI,CAAC,QAAQ;AACX,gBAAU;AAAA,IACZ,WAAW,IAAI,SAAS,KAAK,IAAI,CAAC,KAAK,iBAAiB;AACtD,YAAM,oBAAoB,EAAE,iBAAiB,QAAQ,IAAI,QAAQ;AACjE,gBAAU,oBAAoB,aAAa;AAAA,IAC7C,OAAO;AACL,gBAAU;AAAA,IACZ;AAEA,WAAO;AAAA,MACL,QAAQ,EAAE;AAAA,MACV,cAAc;AAAA,MACd,eAAe;AAAA,MACf;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,SAAS;AAAA,MACT,QAAQ;AAAA,MACR;AAAA,MACA,KAAK;AAAA,MACL;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA,eAAe,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,WAAW;AAAA,IAC5D,aAAa,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,UAAU;AAAA,EAC3D;AACF;AAEA,SAAS,KAAK,IAAsB;AAClC,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAGO,SAAS,IAAI,IAAsB;AACxC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,IAAI,CAAC,MAAc;AACvB,UAAM,MAAM,KAAK,OAAO,SAAS;AACjC,UAAM,KAAK,KAAK,MAAM,GAAG;AACzB,UAAM,KAAK,KAAK,KAAK,GAAG;AACxB,WAAO,OAAO,EAAE,KAAM,OAAO,EAAE,IAAK,OAAO,EAAE,MAAO,MAAM;AAAA,EAC5D;AACA,SAAO,EAAE,IAAI,IAAI,EAAE,IAAI;AACzB;AAOO,SAAS,YAAY,GAAa,GAAmD;AAC1F,MAAI,EAAE,SAAS,KAAK,EAAE,SAAS,EAAG,QAAO,EAAE,GAAG,GAAG,IAAI,GAAG,GAAG,EAAE;AAC7D,QAAM,KAAK,KAAK,CAAC;AACjB,QAAM,KAAK,KAAK,CAAC;AACjB,QAAM,KAAK,SAAS,GAAG,EAAE;AACzB,QAAM,KAAK,SAAS,GAAG,EAAE;AACzB,QAAM,YAAY,KAAK,EAAE,SAAS,KAAK,EAAE;AACzC,MAAI,cAAc,EAAG,QAAO,EAAE,GAAG,OAAO,KAAK,IAAI,UAAU,IAAI,GAAG,GAAG,OAAO,KAAK,IAAI,EAAE;AACvF,QAAM,KAAK,KAAK,MAAM,KAAK,KAAK,SAAS;AACzC,QAAM,KACH,YAAY,cACX,KAAK,EAAE,WAAW,KAAK,EAAE,SAAS,MAAM,KAAK,EAAE,WAAW,KAAK,EAAE,SAAS;AAC9E,QAAM,IAAI,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;AAC9C,SAAO,EAAE,GAAG,IAAI,EAAE;AACpB;AAEA,SAAS,SAAS,IAAc,GAAmB;AACjD,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,MAAM,GAAG,CAAC,KAAK,GAAG,SAAS;AACrE;AAGA,SAAS,YAAY,GAAW,IAAoB;AAClD,MAAI,MAAM,EAAG,QAAO;AACpB,MAAI,KAAK,IAAK,QAAO,UAAU,CAAC;AAChC,QAAM,IAAI,MAAM,KAAK,IAAI;AACzB,QAAM,KAAK,eAAe,GAAG,KAAK,GAAG,GAAG;AACxC,SAAO,KAAK,IAAI,IAAI,MAAM,KAAK,MAAM;AACvC;AAEA,SAAS,eAAe,GAAW,GAAW,GAAmB;AAC/D,MAAI,KAAK,EAAG,QAAO;AACnB,MAAI,KAAK,EAAG,QAAO;AACnB,QAAM,SAAS,QAAQ,CAAC,IAAI,QAAQ,CAAC,IAAI,QAAQ,IAAI,CAAC;AACtD,QAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,IAAI,IAAI,MAAM,IAAI;AACzE,MAAI,IAAI;AACR,MAAI,IAAI,KAAM,IAAI,KAAK,KAAM,IAAI;AACjC,MAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,MAAI,IAAI;AACR,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,KAAK,KAAK,KAAK;AAC7B,UAAM,KAAK,IAAI;AACf,QAAI,MAAO,KAAK,IAAI,KAAK,MAAO,IAAI,KAAK,MAAM,IAAI;AACnD,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,SAAK,IAAI;AACT,UAAM,GAAG,IAAI,MAAM,IAAI,IAAI,KAAK,OAAO,IAAI,OAAO,IAAI,KAAK;AAC3D,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,UAAM,QAAQ,IAAI;AAClB,SAAK;AACL,QAAI,KAAK,IAAI,QAAQ,CAAC,IAAI,KAAM;AAAA,EAClC;AACA,SAAO,QAAQ;AACjB;AAEA,SAAS,QAAQ,GAAmB;AAClC,QAAM,QAAQ;AAAA,IACZ;AAAA,IAAqB;AAAA,IAAmB;AAAA,IAAqB;AAAA,IAC7D;AAAA,IAAqB;AAAA,IAAoB;AAAA,IAAsB;AAAA,IAC/D;AAAA,EACF;AACA,MAAI,IAAI,IAAK,QAAO,KAAK,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,QAAQ,IAAI,CAAC;AAC7E,OAAK;AACL,MAAI,IAAI,MAAM,CAAC;AACf,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,MAAK,MAAM,CAAC,KAAM,IAAI;AAClD,QAAM,IAAI,IAAI;AACd,SAAO,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,KAAK,IAAI,OAAO,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,CAAC;AAC/E;AAEA,SAAS,UAAU,GAAmB;AACpC,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,IAAI;AACV,QAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,QAAM,OAAO,KAAK,IAAI,CAAC;AACvB,QAAM,IAAI,KAAK,IAAI,IAAI;AACvB,QAAM,IAAI,QAAQ,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,KAAK,IAAK,CAAC,OAAO,OAAQ,CAAC;AAC9F,SAAO,OAAO,IAAI,OAAO;AAC3B;","names":[]}

package/dist/{chunk-NU65VQ7M.js → chunk-MAZ26DC7.js}

@@ -96,4 +96,4 @@ export {
   pairedEvalueSequence,
   evaluateInterimReleaseConfidence
 };
-//# sourceMappingURL=chunk-NU65VQ7M.js.map
+//# sourceMappingURL=chunk-MAZ26DC7.js.map

package/dist/chunk-MAZ26DC7.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/sequential.ts"],"sourcesContent":["/**\n * Always-valid sequential evaluation.\n *\n * `researchReport` assumes a single pre-specified analysis. Real\n * consumers run campaigns weekly / nightly / per-PR; each new run silently\n * inflates the false-discovery rate, because the BH-FDR guarantee is for\n * the *first* look, not the 47th. Without time-uniform inference,\n * launch-decision teams either (a) don't peek, which forfeits the cost\n * advantage of stop-when-decisive, or (b) peek and pretend they didn't,\n * which forfeits scientific validity.\n *\n * This module ships **e-value-based confidence sequences** for paired\n * bounded outcomes. The methodology is the predictable plug-in betting\n * martingale of Waudby-Smith & Ramdas (2024) — provably valid at *any*\n * stopping time. Concretely:\n *\n *   For paired deltas D_1, D_2, … ∈ [-c, c] with the null H_0: E[D] ≤ 0,\n *   a betting fraction λ_i is chosen using only D_{1..i-1} (predictable\n *   plug-in), and the running e-value is\n *\n *     E_t = ∏_{i=1}^{t} (1 + λ_i · D_i)\n *\n *   E_t is a non-negative martingale under H_0 with E[E_t] ≤ 1, so by\n *   Ville's inequality, P(∃ t : E_t ≥ 1/α) ≤ α — we can reject the null\n *   at any time without inflating the type-I error.\n *\n * Combined with `runEvalCampaign`, every consumer running rolling\n * campaigns gains the ability to ship the moment evidence is decisive,\n * stop-early on dead-on-arrival variants, and accumulate evidence across\n * partial runs without spending the FDR budget. No new sweep is wasted.\n *\n * References:\n *   - Howard, S. R., Ramdas, A., McAuliffe, J., Sekhon, J. (2021).\n *     Time-uniform, nonparametric, nonasymptotic confidence sequences.\n *     Annals of Statistics, 49(2), 1055–1080.\n *   - Waudby-Smith, I., Ramdas, A. (2024). Estimating means of bounded\n *     random variables by betting. JRSS B, 86(1), 1–27.\n */\n\nexport type SequentialDecision = 'promote_now' | 'continue' | 'reject_now' | 'equivalent'\n\nexport interface PairedEvalueOptions {\n  /**\n   * Bound on |delta|. Default 1 (matching most score scales). Must satisfy\n   * c > 0; deltas outside [-c, c] are clipped with a warning attached to\n   * the return value.\n   */\n  bound?: number\n  /** Target Type-I error. Default 0.05. */\n  alpha?: number\n  /**\n   * Region of Practical Equivalence on the *mean* paired delta. When\n   * supplied, the verdict can return `'equivalent'` once the running\n   * confidence sequence on the mean is fully contained in [low, high].\n   */\n  rope?: { low: number; high: number }\n  /** Initial bet shrinkage (0 < scale ≤ 1). Default 0.5 — empirically robust. */\n  initialBetShrinkage?: number\n}\n\nexport interface PairedEvalueStep {\n  /** 1-indexed observation count. */\n  t: number\n  delta: number\n  /** Running e-value E_t = ∏ (1 + λ_i · D_i). */\n  evalue: number\n  /** Time-uniform p-value at stopping time t. */\n  pValue: number\n  /** Lower bound of the empirical Bernstein confidence sequence at level 1-α. */\n  csLow: number\n  csHigh: number\n  /** Verdict at this stopping time. */\n  decision: SequentialDecision\n}\n\nexport interface PairedEvalueSequence {\n  steps: PairedEvalueStep[]\n  /** The decision at the final step. */\n  finalDecision: SequentialDecision\n  /** Index (1-based) at which a non-`continue` decision first fired, or null. */\n  decisionFiredAt: number | null\n  /** True if any deltas were clipped to [-bound, bound]. */\n  clipped: boolean\n}\n\n/**\n * Run the paired e-value sequence over an in-order delta stream.\n *\n * Use for *streaming* / interim analyses: pass the deltas you have so\n * far, get the verdict at every prefix length. The decision is\n * monotone-stable in the sense that once `'reject_now'` or `'promote_now'`\n * fires, the verdict at later steps remains decisive (the e-value is a\n * non-negative martingale; once it crosses the threshold, it's crossed).\n */\nexport function pairedEvalueSequence(\n  deltas: number[],\n  opts: PairedEvalueOptions = {},\n): PairedEvalueSequence {\n  const c = opts.bound ?? 1\n  const alpha = opts.alpha ?? 0.05\n  const initialShrink = opts.initialBetShrinkage ?? 0.5\n  const rope = opts.rope ?? null\n  if (c <= 0) throw new Error('pairedEvalueSequence: bound must be > 0')\n  if (alpha <= 0 || alpha >= 1) throw new Error('pairedEvalueSequence: alpha must be in (0,1)')\n  if (rope && !(Number.isFinite(rope.low) && Number.isFinite(rope.high) && rope.low <= rope.high)) {\n    throw new Error('pairedEvalueSequence: rope must satisfy low ≤ high')\n  }\n\n  const steps: PairedEvalueStep[] = []\n  let clipped = false\n  let evalue = 1\n  let decisionFiredAt: number | null = null\n\n  // Running statistics (using only D_{1..i-1} for the bet → predictable plug-in).\n  let sum = 0\n  let sumSq = 0\n  let count = 0\n\n  for (let i = 0; i < deltas.length; i++) {\n    let d = deltas[i]!\n    if (d < -c || d > c) {\n      d = Math.max(-c, Math.min(c, d))\n      clipped = true\n    }\n\n    // Predictable plug-in bet (positive λ tests for E[D] > 0; we run a two-sided\n    // test by tracking the symmetric e-value via |bet|).\n    // λ_i ∝ mean / (variance + bound^2). Shrink early to avoid overbetting.\n    const muHat = count === 0 ? 0 : sum / count\n    const varHat = count === 0 ? c * c : Math.max(1e-12, sumSq / count - muHat * muHat)\n    const t = i + 1\n    const shrink = initialShrink * Math.min(1, count / 32) // anneal toward 1\n    let lambda = (muHat / (varHat + c * c)) * shrink\n    // Clip to ensure 1 + λ·D > 0 for all |D| ≤ c (so the e-value stays non-negative).\n    const lambdaMax = 0.99 / c\n    if (lambda > lambdaMax) lambda = lambdaMax\n    if (lambda < -lambdaMax) lambda = -lambdaMax\n\n    evalue = evalue * (1 + lambda * d)\n    if (!Number.isFinite(evalue) || evalue < 0) evalue = 0\n\n    sum += d\n    sumSq += d * d\n    count += 1\n\n    const pValue = Math.min(1, 1 / Math.max(evalue, 1e-300))\n\n    // Empirical Bernstein confidence sequence on the mean. Howard et al.\n    // (2021), Theorem 4.4 with σ̂² the running sample variance and a\n    // calibration constant tuned for two-sided coverage at level 1 - α.\n    const cs = empiricalBernsteinCs(sum, sumSq, count, c, alpha)\n\n    let decision: SequentialDecision = 'continue'\n    if (rope && cs.low >= rope.low && cs.high <= rope.high) decision = 'equivalent'\n    else if (evalue >= 2 / alpha && muHat > 0) decision = 'promote_now'\n    else if (evalue >= 2 / alpha && muHat < 0) decision = 'reject_now'\n    else if (rope && cs.high < rope.low) decision = 'reject_now'\n\n    if (decision !== 'continue' && decisionFiredAt === null) decisionFiredAt = t\n\n    steps.push({ t, delta: d, evalue, pValue, csLow: cs.low, csHigh: cs.high, decision })\n  }\n\n  const finalDecision = steps.length === 0 ? 'continue' : steps[steps.length - 1]!.decision\n  return { steps, finalDecision, decisionFiredAt, clipped }\n}\n\nexport interface InterimReleaseConfidenceInput {\n  /**\n   * One delta series per candidate (paired deltas vs comparator). Order\n   * within a series is the order the campaigns were run.\n   */\n  deltaSeries: Array<{ candidateId: string; deltas: number[] }>\n  alpha?: number\n  bound?: number\n  rope?: { low: number; high: number }\n}\n\nexport interface InterimReleaseConfidence {\n  candidates: Array<{\n    candidateId: string\n    decision: SequentialDecision\n    decisionFiredAt: number | null\n    finalEvalue: number\n    finalPValue: number\n    pairs: number\n    csLow: number\n    csHigh: number\n  }>\n  /**\n   * Campaign-level recommendation: pick the strongest 'promote_now', else\n   * 'continue' if any candidate is still live, else 'reject_now' if every\n   * candidate is dead, else 'equivalent'.\n   */\n  recommendation: { decision: SequentialDecision; candidateId: string | null }\n}\n\n/**\n * Run interim sequential analyses across many candidates at once,\n * preserving the time-uniform α guarantee for each candidate's series and\n * synthesising a campaign-level recommendation. Designed to be called on\n * every campaign tick — the recommendation is anytime-valid.\n */\nexport function evaluateInterimReleaseConfidence(\n  input: InterimReleaseConfidenceInput,\n): InterimReleaseConfidence {\n  const candidates = input.deltaSeries.map((s) => {\n    const seq = pairedEvalueSequence(s.deltas, {\n      alpha: input.alpha,\n      bound: input.bound,\n      rope: input.rope,\n    })\n    const last = seq.steps[seq.steps.length - 1]\n    return {\n      candidateId: s.candidateId,\n      decision: seq.finalDecision,\n      decisionFiredAt: seq.decisionFiredAt,\n      finalEvalue: last?.evalue ?? 1,\n      finalPValue: last?.pValue ?? 1,\n      pairs: seq.steps.length,\n      csLow: last?.csLow ?? Number.NEGATIVE_INFINITY,\n      csHigh: last?.csHigh ?? Number.POSITIVE_INFINITY,\n    }\n  })\n\n  const promote = candidates.find((c) => c.decision === 'promote_now')\n  if (promote)\n    return {\n      candidates,\n      recommendation: { decision: 'promote_now', candidateId: promote.candidateId },\n    }\n  const live = candidates.find((c) => c.decision === 'continue')\n  if (live) return { candidates, recommendation: { decision: 'continue', candidateId: null } }\n  const equiv = candidates.find((c) => c.decision === 'equivalent')\n  if (equiv)\n    return {\n      candidates,\n      recommendation: { decision: 'equivalent', candidateId: equiv.candidateId },\n    }\n  return { candidates, recommendation: { decision: 'reject_now', candidateId: null } }\n}\n\n// ── Internals ────────────────────────────────────────────────────────────\n\n/**\n * Empirical Bernstein confidence sequence on the mean of bounded variables.\n * Adapted from Howard et al. (2021) §4.4. Provides a time-uniform CI on\n * the running mean; valid at every stopping time.\n */\nfunction empiricalBernsteinCs(\n  sum: number,\n  sumSq: number,\n  n: number,\n  bound: number,\n  alpha: number,\n): { low: number; high: number } {\n  if (n === 0) return { low: -bound, high: bound }\n  const mean = sum / n\n  const variance = Math.max(0, sumSq / n - mean * mean)\n  // Iterated-log calibration constant. The 1.7 exponent matches the\n  // recommended choice in Howard et al. for two-sided coverage at level\n  // 1 - α with mild log-corrections; tightening further requires a\n  // tuned mixture and is out of scope.\n  const psi = Math.log(2 / alpha) + 1.7 * Math.log(Math.log(Math.max(Math.E, n)) + 1)\n  const radius = Math.sqrt((2 * variance * psi) / n) + (3 * bound * psi) / n\n  return { low: mean - radius, high: mean + radius }\n}\n"],"mappings":";AA8FO,SAAS,qBACd,QACA,OAA4B,CAAC,GACP;AACtB,QAAM,IAAI,KAAK,SAAS;AACxB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,gBAAgB,KAAK,uBAAuB;AAClD,QAAM,OAAO,KAAK,QAAQ;AAC1B,MAAI,KAAK,EAAG,OAAM,IAAI,MAAM,yCAAyC;AACrE,MAAI,SAAS,KAAK,SAAS,EAAG,OAAM,IAAI,MAAM,8CAA8C;AAC5F,MAAI,QAAQ,EAAE,OAAO,SAAS,KAAK,GAAG,KAAK,OAAO,SAAS,KAAK,IAAI,KAAK,KAAK,OAAO,KAAK,OAAO;AAC/F,UAAM,IAAI,MAAM,yDAAoD;AAAA,EACtE;AAEA,QAAM,QAA4B,CAAC;AACnC,MAAI,UAAU;AACd,MAAI,SAAS;AACb,MAAI,kBAAiC;AAGrC,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AAEZ,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,QAAI,IAAI,OAAO,CAAC;AAChB,QAAI,IAAI,CAAC,KAAK,IAAI,GAAG;AACnB,UAAI,KAAK,IAAI,CAAC,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AAC/B,gBAAU;AAAA,IACZ;AAKA,UAAM,QAAQ,UAAU,IAAI,IAAI,MAAM;AACtC,UAAM,SAAS,UAAU,IAAI,IAAI,IAAI,KAAK,IAAI,OAAO,QAAQ,QAAQ,QAAQ,KAAK;AAClF,UAAM,IAAI,IAAI;AACd,UAAM,SAAS,gBAAgB,KAAK,IAAI,GAAG,QAAQ,EAAE;AACrD,QAAI,SAAU,SAAS,SAAS,IAAI,KAAM;AAE1C,UAAM,YAAY,OAAO;AACzB,QAAI,SAAS,UAAW,UAAS;AACjC,QAAI,SAAS,CAAC,UAAW,UAAS,CAAC;AAEnC,aAAS,UAAU,IAAI,SAAS;AAChC,QAAI,CAAC,OAAO,SAAS,MAAM,KAAK,SAAS,EAAG,UAAS;AAErD,WAAO;AACP,aAAS,IAAI;AACb,aAAS;AAET,UAAM,SAAS,KAAK,IAAI,GAAG,IAAI,KAAK,IAAI,QAAQ,MAAM,CAAC;AAKvD,UAAM,KAAK,qBAAqB,KAAK,OAAO,OAAO,GAAG,KAAK;AAE3D,QAAI,WAA+B;AACnC,QAAI,QAAQ,GAAG,OAAO,KAAK,OAAO,GAAG,QAAQ,KAAK,KAAM,YAAW;AAAA,aAC1D,UAAU,IAAI,SAAS,QAAQ,EAAG,YAAW;AAAA,aAC7C,UAAU,IAAI,SAAS,QAAQ,EAAG,YAAW;AAAA,aAC7C,QAAQ,GAAG,OAAO,KAAK,IAAK,YAAW;AAEhD,QAAI,aAAa,cAAc,oBAAoB,KAAM,mBAAkB;AAE3E,UAAM,KAAK,EAAE,GAAG,OAAO,GAAG,QAAQ,QAAQ,OAAO,GAAG,KAAK,QAAQ,GAAG,MAAM,SAAS,CAAC;AAAA,EACtF;AAEA,QAAM,gBAAgB,MAAM,WAAW,IAAI,aAAa,MAAM,MAAM,SAAS,CAAC,EAAG;AACjF,SAAO,EAAE,OAAO,eAAe,iBAAiB,QAAQ;AAC1D;AAsCO,SAAS,iCACd,OAC0B;AAC1B,QAAM,aAAa,MAAM,YAAY,IAAI,CAAC,MAAM;AAC9C,UAAM,MAAM,qBAAqB,EAAE,QAAQ;AAAA,MACzC,OAAO,MAAM;AAAA,MACb,OAAO,MAAM;AAAA,MACb,MAAM,MAAM;AAAA,IACd,CAAC;AACD,UAAM,OAAO,IAAI,MAAM,IAAI,MAAM,SAAS,CAAC;AAC3C,WAAO;AAAA,MACL,aAAa,EAAE;AAAA,MACf,UAAU,IAAI;AAAA,MACd,iBAAiB,IAAI;AAAA,MACrB,aAAa,MAAM,UAAU;AAAA,MAC7B,aAAa,MAAM,UAAU;AAAA,MAC7B,OAAO,IAAI,MAAM;AAAA,MACjB,OAAO,MAAM,SAAS,OAAO;AAAA,MAC7B,QAAQ,MAAM,UAAU,OAAO;AAAA,IACjC;AAAA,EACF,CAAC;AAED,QAAM,UAAU,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,aAAa;AACnE,MAAI;AACF,WAAO;AAAA,MACL;AAAA,MACA,gBAAgB,EAAE,UAAU,eAAe,aAAa,QAAQ,YAAY;AAAA,IAC9E;AACF,QAAM,OAAO,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,UAAU;AAC7D,MAAI,KAAM,QAAO,EAAE,YAAY,gBAAgB,EAAE,UAAU,YAAY,aAAa,KAAK,EAAE;AAC3F,QAAM,QAAQ,WAAW,KAAK,CAAC,MAAM,EAAE,aAAa,YAAY;AAChE,MAAI;AACF,WAAO;AAAA,MACL;AAAA,MACA,gBAAgB,EAAE,UAAU,cAAc,aAAa,MAAM,YAAY;AAAA,IAC3E;AACF,SAAO,EAAE,YAAY,gBAAgB,EAAE,UAAU,cAAc,aAAa,KAAK,EAAE;AACrF;AASA,SAAS,qBACP,KACA,OACA,GACA,OACA,OAC+B;AAC/B,MAAI,MAAM,EAAG,QAAO,EAAE,KAAK,CAAC,OAAO,MAAM,MAAM;AAC/C,QAAM,OAAO,MAAM;AACnB,QAAM,WAAW,KAAK,IAAI,GAAG,QAAQ,IAAI,OAAO,IAAI;AAKpD,QAAM,MAAM,KAAK,IAAI,IAAI,KAAK,IAAI,MAAM,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,CAAC,CAAC,IAAI,CAAC;AAClF,QAAM,SAAS,KAAK,KAAM,IAAI,WAAW,MAAO,CAAC,IAAK,IAAI,QAAQ,MAAO;AACzE,SAAO,EAAE,KAAK,OAAO,QAAQ,MAAM,OAAO,OAAO;AACnD;","names":[]}

package/dist/{chunk-LSH4MMOZ.js → chunk-NCRFYPS3.js}

@@ -835,4 +835,4 @@ export {
   subjectiveEval,
   allCriticalPassed
 };
-//# sourceMappingURL=chunk-LSH4MMOZ.js.map
+//# sourceMappingURL=chunk-NCRFYPS3.js.map

package/dist/chunk-NCRFYPS3.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/control-runtime.ts"],"sourcesContent":["/**\n * Policy-based agent control runtime.\n *\n * This is the minimal reusable loop behind driver-agent patterns:\n *\n *   observe state -> validate -> decide next action -> act -> observe -> ...\n *\n * It deliberately does not model named \"topologies\". Direct execution,\n * critic/revise, driver intervention, specialist calls, and human escalation\n * are all just actions chosen by the control policy.\n */\n\nimport { type SpanHandle, TraceEmitter } from './trace/emitter'\nimport type { FailureClass } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport type ControlSeverity = 'info' | 'warning' | 'error' | 'critical'\nexport type ControlActionFailureMode = 'continue' | 'stop'\n\nexport interface ControlEvalResult {\n  /** Stable validator or judge id. */\n  id: string\n  /** Whether this check passed. */\n  passed: boolean\n  /** Optional normalized score. 1 = best, 0 = worst. */\n  score?: number\n  /** Objective validators should usually be \"error\" or \"critical\" when failed. */\n  severity?: ControlSeverity\n  /** Human-readable result. */\n  detail?: string\n  /** Small evidence string or pointer. Avoid large payloads. */\n  evidence?: string\n  /** True when the result came from deterministic state, not LLM judgment. */\n  objective?: boolean\n  /** Structured details for downstream control policies and reports. */\n  metadata?: Record<string, unknown>\n}\n\nexport interface ControlBudget {\n  maxSteps: number\n  maxWallMs?: number\n  maxCostUsd?: number\n}\n\nexport interface ControlStopPolicies<TState, TAction> {\n  /**\n   * Stop after N consecutive steps with no state fingerprint change and\n   * less than `minScoreDelta` score movement. Disabled when omitted.\n   */\n  maxNoProgressSteps?: number\n  /**\n   * Stop after the same action fingerprint is selected N consecutive\n   * times. Disabled when omitted.\n   */\n  maxRepeatedActions?: number\n  /** Minimum score movement that counts as progress. Default 0.001. */\n  minScoreDelta?: number\n  /** Override the default JSON/string fingerprint for state comparisons. */\n  stateFingerprint?: (state: TState) => string\n  /** Override the default JSON/string fingerprint for repeated-action checks. */\n  actionFingerprint?: (action: TAction) => string\n}\n\nexport interface ControlContext<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  state: TState\n  evals: TEval[]\n  history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  budget: ControlBudget\n  stepIndex: number\n  wallMs: number\n  spentCostUsd: number\n  remainingCostUsd?: number\n  abortSignal: AbortSignal\n  emitter?: TraceEmitter\n}\n\nexport type ControlDecision<TAction> =\n  | {\n      type: 'continue'\n      action: TAction\n      reason?: string\n    }\n  | {\n      type: 'stop'\n      reason: string\n      pass?: boolean\n      score?: number\n    }\n\nexport interface StopDecision {\n  stop: boolean\n  pass: boolean\n  reason: string\n  score?: number\n  failureClass?: FailureClass\n}\n\nexport interface ControlActionOutcome<TActionResult> {\n  ok: boolean\n  result?: TActionResult\n  error?: string\n  costUsd?: number\n  durationMs: number\n}\n\nexport interface ControlRuntimeError {\n  phase: 'observe' | 'validate' | 'decide' | 'act' | 'stop-policy' | 'on-step' | 'trace'\n  stepIndex: number\n  message: string\n}\n\nexport interface ControlStep<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  index: number\n  decision: ControlDecision<TAction>\n  beforeState: TState\n  afterState: TState\n  evalsBefore: TEval[]\n  evalsAfter: TEval[]\n  actionOutcome?: ControlActionOutcome<TActionResult>\n  startedAt: string\n  endedAt: string\n}\n\nexport interface ControlRunResult<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  pass: boolean\n  completed: boolean\n  reason: string\n  score?: number\n  steps: ControlStep<TState, TAction, TActionResult, TEval>[]\n  finalState: TState | undefined\n  finalEvals: TEval[]\n  wallMs: number\n  spentCostUsd: number\n  /** null when the run executed without a TraceEmitter wired (no run record was persisted). */\n  runId: string | null\n  failureClass?: FailureClass\n  runtimeErrors: ControlRuntimeError[]\n  stoppedBy: 'policy' | 'stop-policy' | 'budget' | 'abort' | 'runtime-error'\n}\n\nexport interface ControlRuntimeConfig<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  budget?: Partial<ControlBudget>\n  signal?: AbortSignal\n  /** Defaults to `continue`: action failures are recorded, then the policy gets another chance. */\n  actionFailure?: ControlActionFailureMode\n  /**\n   * Extract cost from an action result. Used for `maxCostUsd` budget\n   * enforcement and trace budget ledger emission.\n   */\n  getActionCostUsd?: (ctx: {\n    action: TAction\n    result: TActionResult\n    state: TState\n    evals: TEval[]\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  }) => number | undefined\n\n  /** Read typed task/product state. Prefer structured state over transcript-only context. */\n  observe: (ctx: {\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TState> | TState\n\n  /** Objective validators first, subjective judges only where objective state is insufficient. */\n  validate: (ctx: {\n    intent: string\n    state: TState\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TEval[]> | TEval[]\n\n  /** Choose the next control action. Can call a worker, ask user, run critic, inspect state, or stop. */\n  decide: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<ControlDecision<TAction>> | ControlDecision<TAction>\n\n  /** Execute the action selected by the policy. */\n  act: (\n    action: TAction,\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<TActionResult> | TActionResult\n\n  /** Final stopping policy. Called before decide and after each action. */\n  shouldStop?: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<StopDecision> | StopDecision\n\n  /** Optional hook for tracing or live progress updates. */\n  onStep?: (step: ControlStep<TState, TAction, TActionResult, TEval>) => Promise<void> | void\n\n  /** Optional generic stuck-loop policies. Custom `shouldStop` still runs first. */\n  stopPolicies?: ControlStopPolicies<TState, TAction>\n\n  /** Optional trace sink. Emits one run plus one span per control step. */\n  store?: TraceStore\n  scenarioId?: string\n  projectId?: string\n  variantId?: string\n}\n\nconst DEFAULT_BUDGET: ControlBudget = {\n  maxSteps: 8,\n  maxWallMs: 5 * 60 * 1000,\n}\n\nexport async function runAgentControlLoop<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n>(\n  config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  const budget = normalizeBudget(config.budget)\n  const actionFailure = config.actionFailure ?? 'continue'\n  const controller = new AbortController()\n  const upstreamAbort = () => controller.abort(config.signal?.reason)\n  if (config.signal) {\n    if (config.signal.aborted) controller.abort(config.signal.reason)\n    else config.signal.addEventListener('abort', upstreamAbort, { once: true })\n  }\n\n  const started = Date.now()\n  const wallTimer = budget.maxWallMs\n    ? setTimeout(\n        () => controller.abort(new Error('control runtime wall timeout')),\n        budget.maxWallMs,\n      )\n    : undefined\n  const history: ControlStep<TState, TAction, TActionResult, TEval>[] = []\n  const emitter = config.store ? new TraceEmitter(config.store) : undefined\n  let spentCostUsd = 0\n  const runtimeErrors: ControlRuntimeError[] = []\n  let lastStateFingerprint: string | undefined\n  let lastActionFingerprint: string | undefined\n  let noProgressStreak = 0\n  let repeatedActionStreak = 0\n\n  try {\n    if (emitter) {\n      await runTrace(runtimeErrors, 0, () =>\n        emitter.startRun({\n          scenarioId: config.scenarioId ?? 'agent-control-loop',\n          projectId: config.projectId,\n          variantId: config.variantId,\n          layer: 'meta',\n          tags: {\n            intent: config.intent.slice(0, 120),\n            maxSteps: String(budget.maxSteps),\n            ...(budget.maxCostUsd !== undefined ? { maxCostUsd: String(budget.maxCostUsd) } : {}),\n          },\n        }),\n      )\n    }\n\n    let state: TState\n    let evals: TEval[]\n    try {\n      state = await config.observe({ history, abortSignal: controller.signal })\n    } catch (err) {\n      const error = runtimeError('observe', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: undefined,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    try {\n      evals = await config.validate({\n        intent: config.intent,\n        state,\n        history,\n        abortSignal: controller.signal,\n      })\n      await recordEvalSpans(emitter, evals, 'initial', runtimeErrors, 0)\n    } catch (err) {\n      const error = runtimeError('validate', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: state,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    lastStateFingerprint = fingerprintState(state, config.stopPolicies)\n\n    for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {\n      if (controller.signal.aborted) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: abortReason(controller.signal),\n          score: undefined,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'timeout',\n          runtimeErrors,\n          stoppedBy: 'abort',\n        })\n      }\n\n      const budgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (budgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: budgetStop.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const ctx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let stop: StopDecision\n      try {\n        stop = config.shouldStop ? await config.shouldStop(ctx) : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (stop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: stop.pass,\n          completed: true,\n          reason: stop.reason,\n          score: stop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: stop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      let decision: ControlDecision<TAction>\n      try {\n        decision = await config.decide(ctx)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('decide', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (decision.type === 'stop') {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: decision.pass ?? false,\n          completed: true,\n          reason: decision.reason,\n          score: decision.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: decision.pass === false ? 'unknown' : undefined,\n          runtimeErrors,\n          stoppedBy: 'policy',\n        })\n      }\n\n      const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies)\n      repeatedActionStreak =\n        actionFingerprint === lastActionFingerprint ? repeatedActionStreak + 1 : 1\n      lastActionFingerprint = actionFingerprint\n      const repeatedActionStop = repeatedActionStopDecision(\n        config.stopPolicies,\n        repeatedActionStreak,\n      )\n      if (repeatedActionStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: repeatedActionStop.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const beforeState = state\n      const evalsBefore = evals\n      const scoreBefore = averageScore(evals)\n      const actionStarted = Date.now()\n      const stepHandle = emitter\n        ? await runTrace(runtimeErrors, stepIndex, () =>\n            emitter.tool({\n              name: `control-step-${stepIndex}`,\n              toolName: 'agent-control-action',\n              args: decision.action,\n              attributes: {\n                decision: decision.reason ?? 'continue',\n                repeatedActionStreak,\n              },\n            }),\n          )\n        : undefined\n      let actionOutcome: ControlActionOutcome<TActionResult>\n      try {\n        const result = await config.act(decision.action, ctx)\n        const rawCostUsd = config.getActionCostUsd?.({\n          action: decision.action,\n          result,\n          state,\n          evals,\n          history,\n        })\n        const costUsd = normalizeActionCostUsd(rawCostUsd, runtimeErrors, stepIndex)\n        if (costUsd !== undefined && Number.isFinite(costUsd) && costUsd > 0) {\n          spentCostUsd += costUsd\n          await recordCostBudget(\n            emitter,\n            budget,\n            spentCostUsd,\n            stepHandle,\n            runtimeErrors,\n            stepIndex,\n          )\n        }\n        actionOutcome = {\n          ok: true,\n          result,\n          ...(costUsd !== undefined ? { costUsd } : {}),\n          durationMs: Date.now() - actionStarted,\n        }\n      } catch (err) {\n        runtimeErrors.push(runtimeError('act', stepIndex, err))\n        actionOutcome = {\n          ok: false,\n          error: runtimeErrors[runtimeErrors.length - 1]!.message,\n          durationMs: Date.now() - actionStarted,\n        }\n        if (actionFailure === 'stop') {\n          await runTrace(runtimeErrors, stepIndex, () =>\n            stepHandle?.fail(actionOutcome.error ?? 'action failed'),\n          )\n          const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n            index: stepIndex,\n            decision,\n            beforeState,\n            afterState: state,\n            evalsBefore,\n            evalsAfter: evals,\n            actionOutcome,\n            startedAt: new Date(actionStarted).toISOString(),\n            endedAt: new Date().toISOString(),\n          }\n          history.push(step)\n          await runOnStep(config.onStep, step, runtimeErrors)\n          return finish(emitter, {\n            intent: config.intent,\n            pass: false,\n            completed: false,\n            reason: actionOutcome.error ?? 'action failed',\n            score: averageScore(evals),\n            steps: history,\n            finalState: state,\n            finalEvals: evals,\n            wallMs: Date.now() - started,\n            spentCostUsd,\n            runId: emitter?.runId ?? null,\n            failureClass: 'unknown',\n            runtimeErrors,\n            stoppedBy: 'runtime-error',\n          })\n        }\n      }\n\n      try {\n        state = await config.observe({ history, abortSignal: controller.signal })\n      } catch (err) {\n        runtimeErrors.push(runtimeError('observe', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: beforeState,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: beforeState,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      try {\n        evals = await config.validate({\n          intent: config.intent,\n          state,\n          history,\n          abortSignal: controller.signal,\n        })\n        await recordEvalSpans(\n          emitter,\n          evals,\n          `step-${stepIndex}`,\n          runtimeErrors,\n          stepIndex,\n          stepHandle?.span.spanId,\n        )\n      } catch (err) {\n        runtimeErrors.push(runtimeError('validate', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: state,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      const scoreAfter = averageScore(evals)\n      const stateFingerprint = fingerprintState(state, config.stopPolicies)\n      const noProgressStop = noProgressStopDecision({\n        policies: config.stopPolicies,\n        lastStateFingerprint,\n        stateFingerprint,\n        scoreBefore,\n        scoreAfter,\n        currentStreak: noProgressStreak,\n      })\n      noProgressStreak = noProgressStop.streak\n      lastStateFingerprint = stateFingerprint\n\n      const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n        index: stepIndex,\n        decision,\n        beforeState,\n        afterState: state,\n        evalsBefore,\n        evalsAfter: evals,\n        actionOutcome,\n        startedAt: new Date(actionStarted).toISOString(),\n        endedAt: new Date().toISOString(),\n      }\n      history.push(step)\n      if (actionOutcome.ok) {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.end({\n            attributes: {\n              actionCostUsd: actionOutcome.costUsd ?? null,\n              spentCostUsd,\n              scoreBefore: scoreBefore ?? null,\n              scoreAfter: scoreAfter ?? null,\n              noProgressStreak,\n            },\n          }),\n        )\n      } else {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(actionOutcome.error ?? 'action failed', {\n            attributes: {\n              spentCostUsd,\n              noProgressStreak,\n            },\n          }),\n        )\n      }\n      await runOnStep(config.onStep, step, runtimeErrors)\n\n      if (noProgressStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: noProgressStop.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const postStepBudgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (postStepBudgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: postStepBudgetStop.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const postStepCtx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex + 1,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let postStepStop: StopDecision\n      try {\n        postStepStop = config.shouldStop\n          ? await config.shouldStop(postStepCtx)\n          : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex + 1, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (postStepStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: postStepStop.pass,\n          completed: true,\n          reason: postStepStop.reason,\n          score: postStepStop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: postStepStop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n    }\n\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: `budget exhausted: maxSteps=${budget.maxSteps}`,\n      steps: history,\n      finalState: state,\n      finalEvals: evals,\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'budget_exceeded',\n      runtimeErrors,\n      stoppedBy: 'budget',\n    })\n  } catch (err) {\n    runtimeErrors.push(runtimeError('act', history.length, err))\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n      steps: history,\n      finalState: undefined,\n      finalEvals: [],\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'unknown',\n      runtimeErrors,\n      stoppedBy: 'runtime-error',\n    })\n  } finally {\n    if (wallTimer) clearTimeout(wallTimer)\n    if (config.signal) config.signal.removeEventListener('abort', upstreamAbort)\n  }\n}\n\nexport function stopOnNoProgress<TState, TAction>(\n  maxNoProgressSteps: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxNoProgressSteps'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxNoProgressSteps }\n}\n\nexport function stopOnRepeatedAction<TState, TAction>(\n  maxRepeatedActions: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxRepeatedActions'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxRepeatedActions }\n}\n\nexport function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: true }\n}\n\nexport function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: false }\n}\n\nfunction normalizeBudget(input: Partial<ControlBudget> | undefined): ControlBudget {\n  const raw = { ...DEFAULT_BUDGET, ...input } as Record<string, unknown>\n  if (!Number.isInteger(raw.maxSteps) || (raw.maxSteps as number) < 1) {\n    throw new RangeError(\n      `ControlRuntime budget.maxSteps must be an integer >= 1, got ${String(raw.maxSteps)}`,\n    )\n  }\n  const budget: ControlBudget = { maxSteps: raw.maxSteps as number }\n  if (raw.maxWallMs !== undefined) {\n    if (\n      typeof raw.maxWallMs !== 'number' ||\n      !Number.isFinite(raw.maxWallMs) ||\n      raw.maxWallMs <= 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxWallMs must be a positive finite number, got ${String(raw.maxWallMs)}`,\n      )\n    }\n    budget.maxWallMs = raw.maxWallMs\n  }\n  if (raw.maxCostUsd !== undefined) {\n    if (\n      typeof raw.maxCostUsd !== 'number' ||\n      !Number.isFinite(raw.maxCostUsd) ||\n      raw.maxCostUsd < 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxCostUsd must be a nonnegative finite number, got ${String(raw.maxCostUsd)}`,\n      )\n    }\n    budget.maxCostUsd = raw.maxCostUsd\n  }\n  return budget\n}\n\nfunction normalizeActionCostUsd(\n  costUsd: number | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): number | undefined {\n  if (costUsd === undefined) return undefined\n  if (!Number.isFinite(costUsd) || costUsd < 0) {\n    runtimeErrors.push(\n      runtimeError('act', stepIndex, new Error(`invalid action costUsd: ${String(costUsd)}`)),\n    )\n    return undefined\n  }\n  return costUsd\n}\n\nexport function allCriticalPassed(evals: ControlEvalResult[]): boolean {\n  return evals.every(\n    (result) => result.passed || (result.severity !== 'critical' && result.severity !== 'error'),\n  )\n}\n\nfunction makeContext<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  intent: string,\n  state: TState,\n  evals: TEval[],\n  history: ControlStep<TState, TAction, TActionResult, TEval>[],\n  budget: ControlBudget,\n  stepIndex: number,\n  started: number,\n  spentCostUsd: number,\n  abortSignal: AbortSignal,\n  emitter?: TraceEmitter,\n): ControlContext<TState, TAction, TActionResult, TEval> {\n  return {\n    intent,\n    state,\n    evals,\n    history,\n    budget,\n    stepIndex,\n    wallMs: Date.now() - started,\n    spentCostUsd,\n    remainingCostUsd:\n      budget.maxCostUsd === undefined ? undefined : Math.max(0, budget.maxCostUsd - spentCostUsd),\n    abortSignal,\n    emitter,\n  }\n}\n\nfunction defaultStopDecision(evals: ControlEvalResult[]): StopDecision {\n  if (!evals.length) return { stop: false, pass: false, reason: 'no evals yet' }\n  const pass = allCriticalPassed(evals)\n  return pass\n    ? { stop: true, pass: true, reason: 'all critical evals passed', score: averageScore(evals) }\n    : {\n        stop: false,\n        pass: false,\n        reason: 'critical evals still failing',\n        score: averageScore(evals),\n      }\n}\n\nfunction averageScore(evals: ControlEvalResult[]): number | undefined {\n  const scored = evals\n    .map((result) => result.score)\n    .filter((score): score is number => typeof score === 'number')\n  if (!scored.length) return undefined\n  return Math.round((scored.reduce((sum, score) => sum + score, 0) / scored.length) * 1000) / 1000\n}\n\nfunction budgetStopDecision(\n  budget: ControlBudget,\n  spentCostUsd: number,\n): { stop: boolean; reason: string } {\n  if (budget.maxCostUsd !== undefined && spentCostUsd >= budget.maxCostUsd) {\n    return {\n      stop: true,\n      reason: `budget exhausted: maxCostUsd=${budget.maxCostUsd}`,\n    }\n  }\n  return { stop: false, reason: '' }\n}\n\nasync function recordCostBudget(\n  emitter: TraceEmitter | undefined,\n  budget: ControlBudget,\n  spentCostUsd: number,\n  handle: SpanHandle | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): Promise<void> {\n  if (!emitter || budget.maxCostUsd === undefined) return\n  const maxCostUsd = budget.maxCostUsd\n  await runTrace(runtimeErrors, stepIndex, () =>\n    emitter.recordBudget({\n      dimension: 'usd',\n      limit: maxCostUsd,\n      consumed: spentCostUsd,\n      remaining: Math.max(0, maxCostUsd - spentCostUsd),\n      breached: spentCostUsd >= maxCostUsd,\n      spanId: handle?.span.spanId,\n    }),\n  )\n}\n\nasync function recordEvalSpans(\n  emitter: TraceEmitter | undefined,\n  evals: ControlEvalResult[],\n  phase: string,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  targetSpanId?: string,\n): Promise<void> {\n  if (!emitter) return\n  for (const result of evals) {\n    await runTrace(runtimeErrors, stepIndex, () =>\n      emitter.recordJudge({\n        judgeId: result.objective ? 'objective-validator' : 'subjective-judge',\n        targetSpanId: targetSpanId ?? emitter.runId,\n        name: `control-eval/${result.id}`,\n        dimension: result.id,\n        score: typeof result.score === 'number' ? result.score : result.passed ? 1 : 0,\n        rationale: result.detail,\n        evidence: result.evidence,\n        attributes: {\n          phase,\n          passed: result.passed,\n          severity: result.severity,\n          objective: result.objective,\n        },\n      }),\n    )\n  }\n}\n\nasync function runOnStep<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  onStep: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>['onStep'] | undefined,\n  step: ControlStep<TState, TAction, TActionResult, TEval>,\n  runtimeErrors: ControlRuntimeError[],\n): Promise<void> {\n  if (!onStep) return\n  try {\n    await onStep(step)\n  } catch (err) {\n    runtimeErrors.push(runtimeError('on-step', step.index, err))\n  }\n}\n\nasync function runTrace<T>(\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  write: () => Promise<T | undefined> | T | undefined,\n): Promise<T | undefined> {\n  try {\n    return await write()\n  } catch (err) {\n    runtimeErrors.push(runtimeError('trace', stepIndex, err))\n    return undefined\n  }\n}\n\nfunction noProgressStopDecision<TState, TAction>(args: {\n  policies: ControlStopPolicies<TState, TAction> | undefined\n  lastStateFingerprint: string | undefined\n  stateFingerprint: string\n  scoreBefore: number | undefined\n  scoreAfter: number | undefined\n  currentStreak: number\n}): { stop: boolean; reason: string; streak: number } {\n  const max = args.policies?.maxNoProgressSteps\n  if (!max || max <= 0) return { stop: false, reason: '', streak: 0 }\n  const minScoreDelta = args.policies?.minScoreDelta ?? 0.001\n  const scoreDelta = Math.abs((args.scoreAfter ?? 0) - (args.scoreBefore ?? 0))\n  const stateUnchanged =\n    args.lastStateFingerprint !== undefined && args.lastStateFingerprint === args.stateFingerprint\n  const scoreFlat = scoreDelta < minScoreDelta\n  const streak = stateUnchanged && scoreFlat ? args.currentStreak + 1 : 0\n  return streak >= max\n    ? { stop: true, reason: `stuck: no state/score progress for ${streak} step(s)`, streak }\n    : { stop: false, reason: '', streak }\n}\n\nfunction repeatedActionStopDecision<TState, TAction>(\n  policies: ControlStopPolicies<TState, TAction> | undefined,\n  streak: number,\n): { stop: boolean; reason: string } {\n  const max = policies?.maxRepeatedActions\n  if (!max || max <= 0 || streak < max) return { stop: false, reason: '' }\n  return {\n    stop: true,\n    reason: `stuck: repeated same action for ${streak} step(s)`,\n  }\n}\n\nfunction fingerprintState<TState, TAction>(\n  state: TState,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.stateFingerprint) return policies.stateFingerprint(state)\n  return stableFingerprint(state)\n}\n\nfunction fingerprintAction<TState, TAction>(\n  action: TAction,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.actionFingerprint) return policies.actionFingerprint(action)\n  return stableFingerprint(action)\n}\n\nfunction stableFingerprint(value: unknown): string {\n  if (typeof value === 'string') return value\n  if (typeof value === 'number' || typeof value === 'boolean' || value == null) return String(value)\n  try {\n    return JSON.stringify(sortForFingerprint(value))\n  } catch {\n    return String(value)\n  }\n}\n\nfunction sortForFingerprint(value: unknown): unknown {\n  if (Array.isArray(value)) return value.map(sortForFingerprint)\n  if (!value || typeof value !== 'object') return value\n  const record = value as Record<string, unknown>\n  const sorted: Record<string, unknown> = {}\n  for (const key of Object.keys(record).sort()) {\n    sorted[key] = sortForFingerprint(record[key])\n  }\n  return sorted\n}\n\nfunction abortReason(signal: AbortSignal): string {\n  const reason = signal.reason\n  if (reason instanceof Error) return reason.message\n  return reason ? String(reason) : 'aborted'\n}\n\nfunction runtimeError(\n  phase: ControlRuntimeError['phase'],\n  stepIndex: number,\n  err: unknown,\n): ControlRuntimeError {\n  const message = err instanceof Error ? err.message : String(err)\n  return { phase, stepIndex, message }\n}\n\nasync function finish<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  emitter: TraceEmitter | undefined,\n  result: ControlRunResult<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  await runTrace(result.runtimeErrors, result.steps.length, () =>\n    emitter?.endRun({\n      pass: result.pass,\n      score: result.score ?? averageScore(result.finalEvals),\n      failureClass: result.failureClass,\n      notes: result.reason,\n    }),\n  )\n  return result\n}\n"],"mappings":";;;;;AA+NA,IAAM,iBAAgC;AAAA,EACpC,UAAU;AAAA,EACV,WAAW,IAAI,KAAK;AACtB;AAEA,eAAsB,oBAMpB,QACkE;AAClE,QAAM,SAAS,gBAAgB,OAAO,MAAM;AAC5C,QAAM,gBAAgB,OAAO,iBAAiB;AAC9C,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,gBAAgB,MAAM,WAAW,MAAM,OAAO,QAAQ,MAAM;AAClE,MAAI,OAAO,QAAQ;AACjB,QAAI,OAAO,OAAO,QAAS,YAAW,MAAM,OAAO,OAAO,MAAM;AAAA,QAC3D,QAAO,OAAO,iBAAiB,SAAS,eAAe,EAAE,MAAM,KAAK,CAAC;AAAA,EAC5E;AAEA,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,YAAY,OAAO,YACrB;AAAA,IACE,MAAM,WAAW,MAAM,IAAI,MAAM,8BAA8B,CAAC;AAAA,IAChE,OAAO;AAAA,EACT,IACA;AACJ,QAAM,UAAgE,CAAC;AACvE,QAAM,UAAU,OAAO,QAAQ,IAAI,aAAa,OAAO,KAAK,IAAI;AAChE,MAAI,eAAe;AACnB,QAAM,gBAAuC,CAAC;AAC9C,MAAI;AACJ,MAAI;AACJ,MAAI,mBAAmB;AACvB,MAAI,uBAAuB;AAE3B,MAAI;AACF,QAAI,SAAS;AACX,YAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAG,MAC/B,QAAQ,SAAS;AAAA,UACf,YAAY,OAAO,cAAc;AAAA,UACjC,WAAW,OAAO;AAAA,UAClB,WAAW,OAAO;AAAA,UAClB,OAAO;AAAA,UACP,MAAM;AAAA,YACJ,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,YAClC,UAAU,OAAO,OAAO,QAAQ;AAAA,YAChC,GAAI,OAAO,eAAe,SAAY,EAAE,YAAY,OAAO,OAAO,UAAU,EAAE,IAAI,CAAC;AAAA,UACrF;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAEA,QAAI;AACJ,QAAI;AACJ,QAAI;AACF,cAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,IAC1E,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,WAAW,GAAG,GAAG;AAC5C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AACA,QAAI;AACF,cAAQ,MAAM,OAAO,SAAS;AAAA,QAC5B,QAAQ,OAAO;AAAA,QACf;AAAA,QACA;AAAA,QACA,aAAa,WAAW;AAAA,MAC1B,CAAC;AACD,YAAM,gBAAgB,SAAS,OAAO,WAAW,eAAe,CAAC;AAAA,IACnE,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,YAAY,GAAG,GAAG;AAC7C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AACA,2BAAuB,iBAAiB,OAAO,OAAO,YAAY;AAElE,aAAS,YAAY,GAAG,YAAY,OAAO,UAAU,aAAa;AAChE,UAAI,WAAW,OAAO,SAAS;AAC7B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,YAAY,WAAW,MAAM;AAAA,UACrC,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,aAAa,mBAAmB,QAAQ,YAAY;AAC1D,UAAI,WAAW,MAAM;AACnB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,WAAW;AAAA,UACnB,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,MAAM;AAAA,QACV,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,OAAO,aAAa,MAAM,OAAO,WAAW,GAAG,IAAI,oBAAoB,KAAK;AAAA,MACrF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,WAAW,GAAG,CAAC;AAC9D,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,KAAK,MAAM;AACb,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,KAAK;AAAA,UACb,OAAO,KAAK;AAAA,UACZ,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,KAAK;AAAA,UACnB;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,UAAI;AACJ,UAAI;AACF,mBAAW,MAAM,OAAO,OAAO,GAAG;AAAA,MACpC,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,UAAU,WAAW,GAAG,CAAC;AACzD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,SAAS,SAAS,QAAQ;AAC5B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,SAAS,QAAQ;AAAA,UACvB,WAAW;AAAA,UACX,QAAQ,SAAS;AAAA,UACjB,OAAO,SAAS;AAAA,UAChB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,SAAS,SAAS,QAAQ,YAAY;AAAA,UACpD;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,oBAAoB,kBAAkB,SAAS,QAAQ,OAAO,YAAY;AAChF,6BACE,sBAAsB,wBAAwB,uBAAuB,IAAI;AAC3E,8BAAwB;AACxB,YAAM,qBAAqB;AAAA,QACzB,OAAO;AAAA,QACP;AAAA,MACF;AACA,UAAI,mBAAmB,MAAM;AAC3B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,mBAAmB;AAAA,UAC3B,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AACpB,YAAM,cAAc;AACpB,YAAM,cAAc,aAAa,KAAK;AACtC,YAAM,gBAAgB,KAAK,IAAI;AAC/B,YAAM,aAAa,UACf,MAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAW,MACvC,QAAQ,KAAK;AAAA,UACX,MAAM,gBAAgB,SAAS;AAAA,UAC/B,UAAU;AAAA,UACV,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,YACV,UAAU,SAAS,UAAU;AAAA,YAC7B;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,IACA;AACJ,UAAI;AACJ,UAAI;AACF,cAAM,SAAS,MAAM,OAAO,IAAI,SAAS,QAAQ,GAAG;AACpD,cAAM,aAAa,OAAO,mBAAmB;AAAA,UAC3C,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF,CAAC;AACD,cAAM,UAAU,uBAAuB,YAAY,eAAe,SAAS;AAC3E,YAAI,YAAY,UAAa,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AACpE,0BAAgB;AAChB,gBAAM;AAAA,YACJ;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,UACF;AAAA,QACF;AACA,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ;AAAA,UACA,GAAI,YAAY,SAAY,EAAE,QAAQ,IAAI,CAAC;AAAA,UAC3C,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,OAAO,WAAW,GAAG,CAAC;AACtD,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ,OAAO,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UAChD,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AACA,YAAI,kBAAkB,QAAQ;AAC5B,gBAAM;AAAA,YAAS;AAAA,YAAe;AAAA,YAAW,MACvC,YAAY,KAAK,cAAc,SAAS,eAAe;AAAA,UACzD;AACA,gBAAMA,QAA2D;AAAA,YAC/D,OAAO;AAAA,YACP;AAAA,YACA;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,YAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,UAClC;AACA,kBAAQ,KAAKA,KAAI;AACjB,gBAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,iBAAO,OAAO,SAAS;AAAA,YACrB,QAAQ,OAAO;AAAA,YACf,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,cAAc,SAAS;AAAA,YAC/B,OAAO,aAAa,KAAK;AAAA,YACzB,OAAO;AAAA,YACP,YAAY;AAAA,YACZ,YAAY;AAAA,YACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,YACrB;AAAA,YACA,OAAO,SAAS,SAAS;AAAA,YACzB,cAAc;AAAA,YACd;AAAA,YACA,WAAW;AAAA,UACb,CAAC;AAAA,QACH;AAAA,MACF;AAEA,UAAI;AACF,gBAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,MAC1E,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,WAAW,WAAW,GAAG,CAAC;AAC1D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI;AACF,gBAAQ,MAAM,OAAO,SAAS;AAAA,UAC5B,QAAQ,OAAO;AAAA,UACf;AAAA,UACA;AAAA,UACA,aAAa,WAAW;AAAA,QAC1B,CAAC;AACD,cAAM;AAAA,UACJ;AAAA,UACA;AAAA,UACA,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA,YAAY,KAAK;AAAA,QACnB;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,YAAY,WAAW,GAAG,CAAC;AAC3D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,YAAM,aAAa,aAAa,KAAK;AACrC,YAAM,mBAAmB,iBAAiB,OAAO,OAAO,YAAY;AACpE,YAAM,iBAAiB,uBAAuB;AAAA,QAC5C,UAAU,OAAO;AAAA,QACjB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,eAAe;AAAA,MACjB,CAAC;AACD,yBAAmB,eAAe;AAClC,6BAAuB;AAEvB,YAAM,OAA2D;AAAA,QAC/D,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,QAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC;AACA,cAAQ,KAAK,IAAI;AACjB,UAAI,cAAc,IAAI;AACpB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,IAAI;AAAA,YACd,YAAY;AAAA,cACV,eAAe,cAAc,WAAW;AAAA,cACxC;AAAA,cACA,aAAa,eAAe;AAAA,cAC5B,YAAY,cAAc;AAAA,cAC1B;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF,OAAO;AACL,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,SAAS,iBAAiB;AAAA,YACvD,YAAY;AAAA,cACV;AAAA,cACA;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,UAAU,OAAO,QAAQ,MAAM,aAAa;AAElD,UAAI,eAAe,MAAM;AACvB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,eAAe;AAAA,UACvB,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,qBAAqB,mBAAmB,QAAQ,YAAY;AAClE,UAAI,mBAAmB,MAAM;AAC3B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,mBAAmB;AAAA,UAC3B,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AAAA,QAClB,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,uBAAe,OAAO,aAClB,MAAM,OAAO,WAAW,WAAW,IACnC,oBAAoB,KAAK;AAAA,MAC/B,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,YAAY,GAAG,GAAG,CAAC;AAClE,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,aAAa,MAAM;AACrB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,aAAa;AAAA,UACnB,WAAW;AAAA,UACX,QAAQ,aAAa;AAAA,UACrB,OAAO,aAAa;AAAA,UACpB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,aAAa;AAAA,UAC3B;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAAA,IACF;AAEA,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,8BAA8B,OAAO,QAAQ;AAAA,MACrD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,OAAO,QAAQ,QAAQ,GAAG,CAAC;AAC3D,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,MACjD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY,CAAC;AAAA,MACb,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,UAAE;AACA,QAAI,UAAW,cAAa,SAAS;AACrC,QAAI,OAAO,OAAQ,QAAO,OAAO,oBAAoB,SAAS,aAAa;AAAA,EAC7E;AACF;AAEO,SAAS,iBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,qBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,cAAc,OAAgE;AAC5F,SAAO,EAAE,GAAG,OAAO,WAAW,KAAK;AACrC;AAEO,SAAS,eAAe,OAAgE;AAC7F,SAAO,EAAE,GAAG,OAAO,WAAW,MAAM;AACtC;AAEA,SAAS,gBAAgB,OAA0D;AACjF,QAAM,MAAM,EAAE,GAAG,gBAAgB,GAAG,MAAM;AAC1C,MAAI,CAAC,OAAO,UAAU,IAAI,QAAQ,KAAM,IAAI,WAAsB,GAAG;AACnE,UAAM,IAAI;AAAA,MACR,+DAA+D,OAAO,IAAI,QAAQ,CAAC;AAAA,IACrF;AAAA,EACF;AACA,QAAM,SAAwB,EAAE,UAAU,IAAI,SAAmB;AACjE,MAAI,IAAI,cAAc,QAAW;AAC/B,QACE,OAAO,IAAI,cAAc,YACzB,CAAC,OAAO,SAAS,IAAI,SAAS,KAC9B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,yEAAyE,OAAO,IAAI,SAAS,CAAC;AAAA,MAChG;AAAA,IACF;AACA,WAAO,YAAY,IAAI;AAAA,EACzB;AACA,MAAI,IAAI,eAAe,QAAW;AAChC,QACE,OAAO,IAAI,eAAe,YAC1B,CAAC,OAAO,SAAS,IAAI,UAAU,KAC/B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,6EAA6E,OAAO,IAAI,UAAU,CAAC;AAAA,MACrG;AAAA,IACF;AACA,WAAO,aAAa,IAAI;AAAA,EAC1B;AACA,SAAO;AACT;AAEA,SAAS,uBACP,SACA,eACA,WACoB;AACpB,MAAI,YAAY,OAAW,QAAO;AAClC,MAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AAC5C,kBAAc;AAAA,MACZ,aAAa,OAAO,WAAW,IAAI,MAAM,2BAA2B,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,IACxF;AACA,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEO,SAAS,kBAAkB,OAAqC;AACrE,SAAO,MAAM;AAAA,IACX,CAAC,WAAW,OAAO,UAAW,OAAO,aAAa,cAAc,OAAO,aAAa;AAAA,EACtF;AACF;AAEA,SAAS,YACP,QACA,OACA,OACA,SACA,QACA,WACA,SACA,cACA,aACA,SACuD;AACvD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,KAAK,IAAI,IAAI;AAAA,IACrB;AAAA,IACA,kBACE,OAAO,eAAe,SAAY,SAAY,KAAK,IAAI,GAAG,OAAO,aAAa,YAAY;AAAA,IAC5F;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,OAA0C;AACrE,MAAI,CAAC,MAAM,OAAQ,QAAO,EAAE,MAAM,OAAO,MAAM,OAAO,QAAQ,eAAe;AAC7E,QAAM,OAAO,kBAAkB,KAAK;AACpC,SAAO,OACH,EAAE,MAAM,MAAM,MAAM,MAAM,QAAQ,6BAA6B,OAAO,aAAa,KAAK,EAAE,IAC1F;AAAA,IACE,MAAM;AAAA,IACN,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,OAAO,aAAa,KAAK;AAAA,EAC3B;AACN;AAEA,SAAS,aAAa,OAAgD;AACpE,QAAM,SAAS,MACZ,IAAI,CAAC,WAAW,OAAO,KAAK,EAC5B,OAAO,CAAC,UAA2B,OAAO,UAAU,QAAQ;AAC/D,MAAI,CAAC,OAAO,OAAQ,QAAO;AAC3B,SAAO,KAAK,MAAO,OAAO,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO,SAAU,GAAI,IAAI;AAC9F;AAEA,SAAS,mBACP,QACA,cACmC;AACnC,MAAI,OAAO,eAAe,UAAa,gBAAgB,OAAO,YAAY;AACxE,WAAO;AAAA,MACL,MAAM;AAAA,MACN,QAAQ,gCAAgC,OAAO,UAAU;AAAA,IAC3D;AAAA,EACF;AACA,SAAO,EAAE,MAAM,OAAO,QAAQ,GAAG;AACnC;AAEA,eAAe,iBACb,SACA,QACA,cACA,QACA,eACA,WACe;AACf,MAAI,CAAC,WAAW,OAAO,eAAe,OAAW;AACjD,QAAM,aAAa,OAAO;AAC1B,QAAM;AAAA,IAAS;AAAA,IAAe;AAAA,IAAW,MACvC,QAAQ,aAAa;AAAA,MACnB,WAAW;AAAA,MACX,OAAO;AAAA,MACP,UAAU;AAAA,MACV,WAAW,KAAK,IAAI,GAAG,aAAa,YAAY;AAAA,MAChD,UAAU,gBAAgB;AAAA,MAC1B,QAAQ,QAAQ,KAAK;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,eAAe,gBACb,SACA,OACA,OACA,eACA,WACA,cACe;AACf,MAAI,CAAC,QAAS;AACd,aAAW,UAAU,OAAO;AAC1B,UAAM;AAAA,MAAS;AAAA,MAAe;AAAA,MAAW,MACvC,QAAQ,YAAY;AAAA,QAClB,SAAS,OAAO,YAAY,wBAAwB;AAAA,QACpD,cAAc,gBAAgB,QAAQ;AAAA,QACtC,MAAM,gBAAgB,OAAO,EAAE;AAAA,QAC/B,WAAW,OAAO;AAAA,QAClB,OAAO,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ,OAAO,SAAS,IAAI;AAAA,QAC7E,WAAW,OAAO;AAAA,QAClB,UAAU,OAAO;AAAA,QACjB,YAAY;AAAA,UACV;AAAA,UACA,QAAQ,OAAO;AAAA,UACf,UAAU,OAAO;AAAA,UACjB,WAAW,OAAO;AAAA,QACpB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,eAAe,UACb,QACA,MACA,eACe;AACf,MAAI,CAAC,OAAQ;AACb,MAAI;AACF,UAAM,OAAO,IAAI;AAAA,EACnB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,WAAW,KAAK,OAAO,GAAG,CAAC;AAAA,EAC7D;AACF;AAEA,eAAe,SACb,eACA,WACA,OACwB;AACxB,MAAI;AACF,WAAO,MAAM,MAAM;AAAA,EACrB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,SAAS,WAAW,GAAG,CAAC;AACxD,WAAO;AAAA,EACT;AACF;AAEA,SAAS,uBAAwC,MAOK;AACpD,QAAM,MAAM,KAAK,UAAU;AAC3B,MAAI,CAAC,OAAO,OAAO,EAAG,QAAO,EAAE,MAAM,OAAO,QAAQ,IAAI,QAAQ,EAAE;AAClE,QAAM,gBAAgB,KAAK,UAAU,iBAAiB;AACtD,QAAM,aAAa,KAAK,KAAK,KAAK,cAAc,MAAM,KAAK,eAAe,EAAE;AAC5E,QAAM,iBACJ,KAAK,yBAAyB,UAAa,KAAK,yBAAyB,KAAK;AAChF,QAAM,YAAY,aAAa;AAC/B,QAAM,SAAS,kBAAkB,YAAY,KAAK,gBAAgB,IAAI;AACtE,SAAO,UAAU,MACb,EAAE,MAAM,MAAM,QAAQ,sCAAsC,MAAM,YAAY,OAAO,IACrF,EAAE,MAAM,OAAO,QAAQ,IAAI,OAAO;AACxC;AAEA,SAAS,2BACP,UACA,QACmC;AACnC,QAAM,MAAM,UAAU;AACtB,MAAI,CAAC,OAAO,OAAO,KAAK,SAAS,IAAK,QAAO,EAAE,MAAM,OAAO,QAAQ,GAAG;AACvE,SAAO;AAAA,IACL,MAAM;AAAA,IACN,QAAQ,mCAAmC,MAAM;AAAA,EACnD;AACF;AAEA,SAAS,iBACP,OACA,UACQ;AACR,MAAI,UAAU,iBAAkB,QAAO,SAAS,iBAAiB,KAAK;AACtE,SAAO,kBAAkB,KAAK;AAChC;AAEA,SAAS,kBACP,QACA,UACQ;AACR,MAAI,UAAU,kBAAmB,QAAO,SAAS,kBAAkB,MAAM;AACzE,SAAO,kBAAkB,MAAM;AACjC;AAEA,SAAS,kBAAkB,OAAwB;AACjD,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,MAAI,OAAO,UAAU,YAAY,OAAO,UAAU,aAAa,SAAS,KAAM,QAAO,OAAO,KAAK;AACjG,MAAI;AACF,WAAO,KAAK,UAAU,mBAAmB,KAAK,CAAC;AAAA,EACjD,QAAQ;AACN,WAAO,OAAO,KAAK;AAAA,EACrB;AACF;AAEA,SAAS,mBAAmB,OAAyB;AACnD,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,kBAAkB;AAC7D,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,SAAS;AACf,QAAM,SAAkC,CAAC;AACzC,aAAW,OAAO,OAAO,KAAK,MAAM,EAAE,KAAK,GAAG;AAC5C,WAAO,GAAG,IAAI,mBAAmB,OAAO,GAAG,CAAC;AAAA,EAC9C;AACA,SAAO;AACT;AAEA,SAAS,YAAY,QAA6B;AAChD,QAAM,SAAS,OAAO;AACtB,MAAI,kBAAkB,MAAO,QAAO,OAAO;AAC3C,SAAO,SAAS,OAAO,MAAM,IAAI;AACnC;AAEA,SAAS,aACP,OACA,WACA,KACqB;AACrB,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,SAAO,EAAE,OAAO,WAAW,QAAQ;AACrC;AAEA,eAAe,OACb,SACA,QACkE;AAClE,QAAM;AAAA,IAAS,OAAO;AAAA,IAAe,OAAO,MAAM;AAAA,IAAQ,MACxD,SAAS,OAAO;AAAA,MACd,MAAM,OAAO;AAAA,MACb,OAAO,OAAO,SAAS,aAAa,OAAO,UAAU;AAAA,MACrD,cAAc,OAAO;AAAA,MACrB,OAAO,OAAO;AAAA,IAChB,CAAC;AAAA,EACH;AACA,SAAO;AACT;","names":["step"]}

package/dist/{chunk-ZN274SWR.js → chunk-PALJO75S.js}

@@ -1,7 +1,7 @@
 import {
   objectiveEval,
   runAgentControlLoop
-} from "./chunk-LSH4MMOZ.js";
+} from "./chunk-NCRFYPS3.js";
 import {
   validateRunRecord
 } from "./chunk-NLMNWKVM.js";
@@ -610,4 +610,4 @@ export {
   runProposeReviewAsControlLoop,
   controlFailureClassFromVerification
 };
-//# sourceMappingURL=chunk-ZN274SWR.js.map
+//# sourceMappingURL=chunk-PALJO75S.js.map

package/dist/{chunk-OWLAAMME.js → chunk-QHF6EQKK.js}

@@ -85,7 +85,8 @@ var SubprocessSandboxDriver = class {
         () => {
           try {
             child.kill("SIGKILL");
-          } catch {
+          } catch (err) {
+            console.warn("[sandbox-harness] SIGKILL on timeout failed:", err);
           }
         },
         config.timeoutMs ?? 10 * 6e4
@@ -247,4 +248,4 @@ export {
   SandboxHarness,
   runTestGradedScenario
 };
-//# sourceMappingURL=chunk-OWLAAMME.js.map
+//# sourceMappingURL=chunk-QHF6EQKK.js.map

package/dist/chunk-QHF6EQKK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/sandbox-harness.ts","../src/test-graded-scenario.ts"],"sourcesContent":["/**\n * SandboxHarness — executes a scenario in an isolated environment and\n * emits a rich SandboxSpan into the trace.\n *\n * Two built-in drivers:\n *   - `SubprocessSandboxDriver` — spawn in a local cwd with env vars.\n *     Fast, no dependencies, fine for unit tests and most CI gates.\n *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;\n *     shells out to `docker run`. Stronger isolation, slower startup.\n *\n * Consumers implement `SandboxDriver` for custom backends (Firecracker,\n * Cloudflare sandbox product, etc.). The harness doesn't care which.\n */\n\nimport { ConfigError } from './errors'\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  /** Absolute cwd for the subprocess driver. Ignored by docker driver. */\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /** Image for the docker driver. */\n  image?: string\n  /** Extra env vars (validated; shell-escaped). */\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\n// ── Parsers ──────────────────────────────────────────────────────────\n\n/** Vitest default summary line: \"Tests  X passed | Y failed\". */\nexport const vitestTestParser: TestOutputParser = {\n  id: 'vitest',\n  parse(stdout) {\n    const m = stdout.match(/Tests\\s+(\\d+)\\s+(passed|failed)(?:\\s*\\|\\s*(\\d+)\\s+(passed|failed))?/i)\n    if (!m) return undefined\n    let passed = 0\n    let failed = 0\n    const a = parseInt(m[1]!, 10)\n    const aLabel = m[2]!.toLowerCase()\n    if (aLabel === 'passed') passed += a\n    else failed += a\n    if (m[3] && m[4]) {\n      const b = parseInt(m[3], 10)\n      if (m[4].toLowerCase() === 'passed') passed += b\n      else failed += b\n    }\n    return { testsTotal: passed + failed, testsPassed: passed }\n  },\n}\n\n/** Pytest default: \"collected N items\" + \" X passed, Y failed\". */\nexport const pytestTestParser: TestOutputParser = {\n  id: 'pytest',\n  parse(stdout) {\n    const total = stdout.match(/collected\\s+(\\d+)\\s+items?/i)\n    const passed = stdout.match(/(\\d+)\\s+passed/)\n    if (!total || !passed) return undefined\n    return { testsTotal: parseInt(total[1]!, 10), testsPassed: parseInt(passed[1]!, 10) }\n  },\n}\n\n/** Jest: \"Tests: X passed, Y total\" (and optional failed). */\nexport const jestTestParser: TestOutputParser = {\n  id: 'jest',\n  parse(stdout) {\n    const m = stdout.match(/Tests:\\s+(?:(\\d+)\\s+failed[^,]*,\\s*)?(\\d+)\\s+passed,\\s+(\\d+)\\s+total/i)\n    if (!m) return undefined\n    return { testsTotal: parseInt(m[3]!, 10), testsPassed: parseInt(m[2]!, 10) }\n  },\n}\n\n/** Composite parser — tries a list of parsers in order. */\nexport function composeParsers(...parsers: TestOutputParser[]): TestOutputParser {\n  return {\n    id: parsers.map((p) => p.id).join('|'),\n    parse(stdout, stderr, exitCode) {\n      for (const p of parsers) {\n        const res = p.parse(stdout, stderr, exitCode)\n        if (res) return res\n      }\n      return undefined\n    },\n  }\n}\n\n// ── Drivers ──────────────────────────────────────────────────────────\n\nexport interface SubprocessSandboxDriverOptions {\n  /**\n   * Default cwd for all `exec` calls. Used when the per-call `HarnessConfig`\n   * does not set its own `cwd`. Lets callers bind the driver to a working\n   * directory once instead of spreading cwd into every harness config —\n   * useful when the harness config is constructed far from the call site\n   * (e.g. starter-foundry's promoter passes a static HarnessConfig per\n   * family taxonomy but needs a per-run composed-scaffold cwd).\n   */\n  cwd?: string\n  /**\n   * Default env merged into every `exec` call's env (per-call `HarnessConfig.env`\n   * still wins on key collision). Same ergonomic rationale as `cwd` above.\n   */\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  private defaultCwd?: string\n  private defaultEnv?: Record<string, string>\n\n  constructor(options: SubprocessSandboxDriverOptions = {}) {\n    this.defaultCwd = options.cwd\n    this.defaultEnv = options.env\n  }\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    const { spawn } = await import('node:child_process')\n    const start = Date.now()\n    // Per-call config wins; fall back to constructor defaults. Honoring\n    // the constructor `cwd` keeps the subprocess from inheriting Node's\n    // cwd when only the constructor arg is supplied.\n    const effectiveCwd = config.cwd ?? this.defaultCwd\n    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    return await new Promise<SandboxResult>((resolve) => {\n      const child = spawn(command, {\n        shell: true,\n        cwd: effectiveCwd,\n        env: effectiveEnv,\n      })\n      let stdout = ''\n      let stderr = ''\n      child.stdout?.on('data', (d) => {\n        stdout += String(d)\n      })\n      child.stderr?.on('data', (d) => {\n        stderr += String(d)\n      })\n      const timeout = setTimeout(\n        () => {\n          try {\n            child.kill('SIGKILL')\n          } catch (err) {\n            console.warn('[sandbox-harness] SIGKILL on timeout failed:', err)\n          }\n        },\n        config.timeoutMs ?? 10 * 60_000,\n      )\n      child.on('close', (code) => {\n        clearTimeout(timeout)\n        const wallMs = Date.now() - start\n        const parsed =\n          phase === 'test' && config.testParser\n            ? config.testParser.parse(stdout, stderr, code ?? 1)\n            : undefined\n        resolve({\n          phase,\n          exitCode: code ?? 1,\n          stdout,\n          stderr,\n          wallMs,\n          testsTotal: parsed?.testsTotal,\n          testsPassed: parsed?.testsPassed,\n        })\n      })\n      child.on('error', (err) => {\n        clearTimeout(timeout)\n        const wallMs = Date.now() - start\n        resolve({ phase, exitCode: 127, stdout, stderr: stderr + String(err), wallMs })\n      })\n    })\n  }\n}\n\nexport class DockerSandboxDriver implements SandboxDriver {\n  id = 'docker'\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    if (!config.image) throw new ConfigError('DockerSandboxDriver requires config.image')\n    const sub = new SubprocessSandboxDriver()\n    const envArgs = Object.entries(config.env ?? {})\n      .map(([k, v]) => `-e ${shellQuote(k)}=${shellQuote(v)}`)\n      .join(' ')\n    const wrapped = `docker run --rm ${envArgs} ${shellQuote(config.image)} sh -c ${shellQuote(command)}`\n    return sub.exec(phase, wrapped, { ...config, env: undefined })\n  }\n}\n\nfunction shellQuote(v: string): string {\n  if (/^[A-Za-z0-9_\\-/.@:=]+$/.test(v)) return v\n  return `'${v.replace(/'/g, `'\\\\''`)}'`\n}\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      if (config.setupCommand) {\n        result.setup = await this.driver.exec('setup', config.setupCommand, config)\n        result.totalWallMs += result.setup.wallMs\n        if (result.setup.exitCode !== 0) {\n          await handle.fail(`setup failed (exit ${result.setup.exitCode})`, {\n            exitCode: result.setup.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.runCommand) {\n        result.run = await this.driver.exec('run', config.runCommand, config)\n        result.totalWallMs += result.run.wallMs\n        if (result.run.exitCode !== 0) {\n          await handle.fail(`run failed (exit ${result.run.exitCode})`, {\n            exitCode: result.run.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.testCommand) {\n        result.test = await this.driver.exec('test', config.testCommand, config)\n        result.totalWallMs += result.test.wallMs\n        const passed = result.test.exitCode === 0\n        result.passed = passed\n        if (result.test.testsTotal !== undefined && result.test.testsTotal > 0) {\n          result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal\n        } else {\n          result.score = passed ? 1 : 0\n        }\n        await handle.end({\n          exitCode: result.test.exitCode,\n          testsTotal: result.test.testsTotal,\n          testsPassed: result.test.testsPassed,\n          wallMs: result.totalWallMs,\n          status: passed ? 'ok' : 'error',\n        } as Partial<SandboxSpan>)\n      } else {\n        result.passed = true\n        result.score = 1\n        await handle.end({ wallMs: result.totalWallMs } as Partial<SandboxSpan>)\n      }\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n","/**\n * TestGradedScenario — a scenario whose score comes from a test suite.\n *\n * This is the SWE-bench pattern generalized. The scenario ships:\n *   - fixture data (setup instructions)\n *   - a test command the harness runs\n *   - optional assertion overrides\n *\n * The runner emits a run, delegates to SandboxHarness, records the\n * outcome, and returns a structured verdict. Consumers bind their own\n * agent execution to this contract.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from './sandbox-harness'\nimport { SandboxHarness } from './sandbox-harness'\nimport { TraceEmitter } from './trace/emitter'\nimport type { FailureClass, Run } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TestGradedScenario {\n  id: string\n  description?: string\n  harness: HarnessConfig\n  /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */\n  passThreshold?: number\n  /** Provenance for dataset tracking. */\n  datasetVersion?: string\n  /** Free-form tags (difficulty, category, etc.). */\n  tags?: Record<string, string>\n}\n\nexport interface TestGradedRunOptions {\n  variantId?: string\n  driver?: SandboxDriver\n  /** Metadata recorded on the Run (codeSha, promptSha, modelFingerprint, seed). */\n  provenance?: Pick<Run, 'codeSha' | 'promptSha' | 'modelFingerprint' | 'seed' | 'envFingerprint'>\n}\n\nexport interface TestGradedRunResult {\n  runId: string\n  scenario: TestGradedScenario\n  harness: SandboxHarnessResult\n  pass: boolean\n  score: number\n  failureClass?: FailureClass\n}\n\nexport async function runTestGradedScenario(\n  scenario: TestGradedScenario,\n  store: TraceStore,\n  options: TestGradedRunOptions = {},\n): Promise<TestGradedRunResult> {\n  const emitter = new TraceEmitter(store)\n  await emitter.startRun({\n    scenarioId: scenario.id,\n    variantId: options.variantId,\n    datasetVersion: scenario.datasetVersion,\n    tags: scenario.tags,\n    ...options.provenance,\n  })\n  const harness = new SandboxHarness(options.driver)\n  const result = await harness.run(scenario.harness, emitter)\n  const threshold = scenario.passThreshold ?? 1.0\n  const pass = result.passed && result.score >= threshold\n  const setupFailed = result.setup !== undefined && result.setup.exitCode !== 0\n  const runFailed = result.run !== undefined && result.run.exitCode !== 0\n  const testFailed = result.test !== undefined && result.test.exitCode !== 0\n  const failureClass: FailureClass | undefined = pass\n    ? 'success'\n    : setupFailed || runFailed\n      ? 'sandbox_failure'\n      : testFailed\n        ? 'format_drift'\n        : 'unknown'\n  await emitter.endRun({\n    pass,\n    score: result.score,\n    failureClass,\n    notes: pass ? undefined : reasonForFailure(result),\n  })\n  return {\n    runId: emitter.runId,\n    scenario,\n    harness: result,\n    pass,\n    score: result.score,\n    failureClass,\n  }\n}\n\nfunction reasonForFailure(result: SandboxHarnessResult): string {\n  if (result.setup && result.setup.exitCode !== 0)\n    return `setup failed: exit ${result.setup.exitCode}`\n  if (result.run && result.run.exitCode !== 0) return `run failed: exit ${result.run.exitCode}`\n  if (result.test) {\n    if (result.test.testsTotal !== undefined) {\n      return `tests: ${result.test.testsPassed ?? 0}/${result.test.testsTotal}`\n    }\n    return `test exit ${result.test.exitCode}`\n  }\n  return 'no test command'\n}\n"],"mappings":";;;;;;;;AAoEO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,sEAAsE;AAC7F,QAAI,CAAC,EAAG,QAAO;AACf,QAAI,SAAS;AACb,QAAI,SAAS;AACb,UAAM,IAAI,SAAS,EAAE,CAAC,GAAI,EAAE;AAC5B,UAAM,SAAS,EAAE,CAAC,EAAG,YAAY;AACjC,QAAI,WAAW,SAAU,WAAU;AAAA,QAC9B,WAAU;AACf,QAAI,EAAE,CAAC,KAAK,EAAE,CAAC,GAAG;AAChB,YAAM,IAAI,SAAS,EAAE,CAAC,GAAG,EAAE;AAC3B,UAAI,EAAE,CAAC,EAAE,YAAY,MAAM,SAAU,WAAU;AAAA,UAC1C,WAAU;AAAA,IACjB;AACA,WAAO,EAAE,YAAY,SAAS,QAAQ,aAAa,OAAO;AAAA,EAC5D;AACF;AAGO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,QAAQ,OAAO,MAAM,6BAA6B;AACxD,UAAM,SAAS,OAAO,MAAM,gBAAgB;AAC5C,QAAI,CAAC,SAAS,CAAC,OAAQ,QAAO;AAC9B,WAAO,EAAE,YAAY,SAAS,MAAM,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,OAAO,CAAC,GAAI,EAAE,EAAE;AAAA,EACtF;AACF;AAGO,IAAM,iBAAmC;AAAA,EAC9C,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,uEAAuE;AAC9F,QAAI,CAAC,EAAG,QAAO;AACf,WAAO,EAAE,YAAY,SAAS,EAAE,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,EAAE,CAAC,GAAI,EAAE,EAAE;AAAA,EAC7E;AACF;AAGO,SAAS,kBAAkB,SAA+C;AAC/E,SAAO;AAAA,IACL,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK,GAAG;AAAA,IACrC,MAAM,QAAQ,QAAQ,UAAU;AAC9B,iBAAW,KAAK,SAAS;AACvB,cAAM,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ;AAC5C,YAAI,IAAK,QAAO;AAAA,MAClB;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAqBO,IAAM,0BAAN,MAAuD;AAAA,EAC5D,KAAK;AAAA,EACG;AAAA,EACA;AAAA,EAER,YAAY,UAA0C,CAAC,GAAG;AACxD,SAAK,aAAa,QAAQ;AAC1B,SAAK,aAAa,QAAQ;AAAA,EAC5B;AAAA,EAEA,MAAM,KACJ,OACA,SACA,QACwB;AACxB,UAAM,EAAE,MAAM,IAAI,MAAM,OAAO,eAAoB;AACnD,UAAM,QAAQ,KAAK,IAAI;AAIvB,UAAM,eAAe,OAAO,OAAO,KAAK;AACxC,UAAM,eAAe,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,cAAc,CAAC,GAAI,GAAI,OAAO,OAAO,CAAC,EAAG;AACzF,WAAO,MAAM,IAAI,QAAuB,CAAC,YAAY;AACnD,YAAM,QAAQ,MAAM,SAAS;AAAA,QAC3B,OAAO;AAAA,QACP,KAAK;AAAA,QACL,KAAK;AAAA,MACP,CAAC;AACD,UAAI,SAAS;AACb,UAAI,SAAS;AACb,YAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM;AAC9B,kBAAU,OAAO,CAAC;AAAA,MACpB,CAAC;AACD,YAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM;AAC9B,kBAAU,OAAO,CAAC;AAAA,MACpB,CAAC;AACD,YAAM,UAAU;AAAA,QACd,MAAM;AACJ,cAAI;AACF,kBAAM,KAAK,SAAS;AAAA,UACtB,SAAS,KAAK;AACZ,oBAAQ,KAAK,gDAAgD,GAAG;AAAA,UAClE;AAAA,QACF;AAAA,QACA,OAAO,aAAa,KAAK;AAAA,MAC3B;AACA,YAAM,GAAG,SAAS,CAAC,SAAS;AAC1B,qBAAa,OAAO;AACpB,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,cAAM,SACJ,UAAU,UAAU,OAAO,aACvB,OAAO,WAAW,MAAM,QAAQ,QAAQ,QAAQ,CAAC,IACjD;AACN,gBAAQ;AAAA,UACN;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB;AAAA,UACA;AAAA,UACA;AAAA,UACA,YAAY,QAAQ;AAAA,UACpB,aAAa,QAAQ;AAAA,QACvB,CAAC;AAAA,MACH,CAAC;AACD,YAAM,GAAG,SAAS,CAAC,QAAQ;AACzB,qBAAa,OAAO;AACpB,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,gBAAQ,EAAE,OAAO,UAAU,KAAK,QAAQ,QAAQ,SAAS,OAAO,GAAG,GAAG,OAAO,CAAC;AAAA,MAChF,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACF;AAEO,IAAM,sBAAN,MAAmD;AAAA,EACxD,KAAK;AAAA,EAEL,MAAM,KACJ,OACA,SACA,QACwB;AACxB,QAAI,CAAC,OAAO,MAAO,OAAM,IAAI,YAAY,2CAA2C;AACpF,UAAM,MAAM,IAAI,wBAAwB;AACxC,UAAM,UAAU,OAAO,QAAQ,OAAO,OAAO,CAAC,CAAC,EAC5C,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,IAAI,WAAW,CAAC,CAAC,EAAE,EACtD,KAAK,GAAG;AACX,UAAM,UAAU,mBAAmB,OAAO,IAAI,WAAW,OAAO,KAAK,CAAC,UAAU,WAAW,OAAO,CAAC;AACnG,WAAO,IAAI,KAAK,OAAO,SAAS,EAAE,GAAG,QAAQ,KAAK,OAAU,CAAC;AAAA,EAC/D;AACF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,yBAAyB,KAAK,CAAC,EAAG,QAAO;AAC7C,SAAO,IAAI,EAAE,QAAQ,MAAM,OAAO,CAAC;AACrC;AAcO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACR,YAAY,SAAwB,IAAI,wBAAwB,GAAG;AACjE,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,MAAM,IAAI,QAAuB,SAAsD;AACrF,UAAM,SAAS,MAAM,QAAQ,QAAQ;AAAA,MACnC,MAAM,WAAW,KAAK,OAAO,EAAE;AAAA,MAC/B,OAAO,OAAO;AAAA,MACd,SAAS,CAAC,OAAO,cAAc,OAAO,YAAY,OAAO,WAAW,EACjE,OAAO,OAAO,EACd,KAAK,MAAM;AAAA,IAChB,CAAC;AACD,UAAM,SAA+B,EAAE,QAAQ,OAAO,aAAa,GAAG,OAAO,EAAE;AAC/E,QAAI;AACF,UAAI,OAAO,cAAc;AACvB,eAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;AAC1E,eAAO,eAAe,OAAO,MAAM;AACnC,YAAI,OAAO,MAAM,aAAa,GAAG;AAC/B,gBAAM,OAAO,KAAK,sBAAsB,OAAO,MAAM,QAAQ,KAAK;AAAA,YAChE,UAAU,OAAO,MAAM;AAAA,YACvB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,YAAY;AACrB,eAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;AACpE,eAAO,eAAe,OAAO,IAAI;AACjC,YAAI,OAAO,IAAI,aAAa,GAAG;AAC7B,gBAAM,OAAO,KAAK,oBAAoB,OAAO,IAAI,QAAQ,KAAK;AAAA,YAC5D,UAAU,OAAO,IAAI;AAAA,YACrB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,aAAa;AACtB,eAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;AACvE,eAAO,eAAe,OAAO,KAAK;AAClC,cAAM,SAAS,OAAO,KAAK,aAAa;AACxC,eAAO,SAAS;AAChB,YAAI,OAAO,KAAK,eAAe,UAAa,OAAO,KAAK,aAAa,GAAG;AACtE,iBAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;AAAA,QAC9D,OAAO;AACL,iBAAO,QAAQ,SAAS,IAAI;AAAA,QAC9B;AACA,cAAM,OAAO,IAAI;AAAA,UACf,UAAU,OAAO,KAAK;AAAA,UACtB,YAAY,OAAO,KAAK;AAAA,UACxB,aAAa,OAAO,KAAK;AAAA,UACzB,QAAQ,OAAO;AAAA,UACf,QAAQ,SAAS,OAAO;AAAA,QAC1B,CAAyB;AAAA,MAC3B,OAAO;AACL,eAAO,SAAS;AAChB,eAAO,QAAQ;AACf,cAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;AAAA,MACzE;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;AAC1D,YAAM;AAAA,IACR;AACA,WAAO;AAAA,EACT;AACF;;;AC5QA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;AAC9B,QAAM,UAAU,IAAI,aAAa,KAAK;AACtC,QAAM,QAAQ,SAAS;AAAA,IACrB,YAAY,SAAS;AAAA,IACrB,WAAW,QAAQ;AAAA,IACnB,gBAAgB,SAAS;AAAA,IACzB,MAAM,SAAS;AAAA,IACf,GAAG,QAAQ;AAAA,EACb,CAAC;AACD,QAAM,UAAU,IAAI,eAAe,QAAQ,MAAM;AACjD,QAAM,SAAS,MAAM,QAAQ,IAAI,SAAS,SAAS,OAAO;AAC1D,QAAM,YAAY,SAAS,iBAAiB;AAC5C,QAAM,OAAO,OAAO,UAAU,OAAO,SAAS;AAC9C,QAAM,cAAc,OAAO,UAAU,UAAa,OAAO,MAAM,aAAa;AAC5E,QAAM,YAAY,OAAO,QAAQ,UAAa,OAAO,IAAI,aAAa;AACtE,QAAM,aAAa,OAAO,SAAS,UAAa,OAAO,KAAK,aAAa;AACzE,QAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;AACR,QAAM,QAAQ,OAAO;AAAA,IACnB;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,IACA,OAAO,OAAO,SAAY,iBAAiB,MAAM;AAAA,EACnD,CAAC;AACD,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf;AAAA,IACA,SAAS;AAAA,IACT;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;AAC9D,MAAI,OAAO,SAAS,OAAO,MAAM,aAAa;AAC5C,WAAO,sBAAsB,OAAO,MAAM,QAAQ;AACpD,MAAI,OAAO,OAAO,OAAO,IAAI,aAAa,EAAG,QAAO,oBAAoB,OAAO,IAAI,QAAQ;AAC3F,MAAI,OAAO,MAAM;AACf,QAAI,OAAO,KAAK,eAAe,QAAW;AACxC,aAAO,UAAU,OAAO,KAAK,eAAe,CAAC,IAAI,OAAO,KAAK,UAAU;AAAA,IACzE;AACA,WAAO,aAAa,OAAO,KAAK,QAAQ;AAAA,EAC1C;AACA,SAAO;AACT;","names":[]}