@tangle-network/agent-eval 0.24.0 → 0.25.0

package/dist/{chunk-OHEPNJQN.js → chunk-JLZQWFV3.js}

@@ -336,6 +336,69 @@ function classifyFailure(ctx, rules = DEFAULT_RULES) {
   return { failureClass: "unknown", reason: "no rule matched; run failed for unclassified reason" };
 }
 
+// src/pipelines/failure-cluster.ts
+async function failureClusterView(store, options = {}) {
+  const rules = options.rules ?? DEFAULT_RULES;
+  const minSize = options.minClusterSize ?? 1;
+  const runs = await store.listRuns();
+  const clusters = /* @__PURE__ */ new Map();
+  let totalFailures = 0;
+  for (const run of runs) {
+    if (run.status === "completed" && run.outcome?.pass !== false) continue;
+    totalFailures++;
+    const spans = await store.spans({ runId: run.runId });
+    const events = await store.events({ runId: run.runId });
+    const cls = classifyFailure({ run, spans, events }, rules);
+    let toolName;
+    let argPrefix;
+    let dimension;
+    if (cls.triggerSpanId) {
+      const trig = spans.find((s) => s.spanId === cls.triggerSpanId);
+      if (trig?.kind === "tool") {
+        toolName = trig.toolName;
+        argPrefix = argHash(trig.args).slice(0, 16);
+      } else if (trig?.kind === "judge") {
+        dimension = trig.dimension;
+      }
+    }
+    if (!toolName) {
+      const ts = await toolSpans(store, run.runId);
+      const errored = ts.filter((t) => t.status === "error").pop();
+      if (errored) {
+        toolName = errored.toolName;
+        argPrefix = argHash(errored.args).slice(0, 16);
+      }
+    }
+    if (!dimension) {
+      const judge = spans.find((s) => s.kind === "judge" && typeof s.dimension === "string");
+      if (judge?.kind === "judge") dimension = judge.dimension;
+    }
+    const key = `${cls.failureClass}|${toolName ?? ""}|${argPrefix ?? ""}|${dimension ?? ""}`;
+    let cluster = clusters.get(key);
+    if (!cluster) {
+      cluster = {
+        failureClass: cls.failureClass,
+        toolName,
+        argPrefix,
+        dimension,
+        runCount: 0,
+        scenarioIds: [],
+        exampleRunId: run.runId,
+        exampleError: firstErrorMessage(spans) ?? cls.reason
+      };
+      clusters.set(key, cluster);
+    }
+    cluster.runCount++;
+    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId);
+  }
+  const arr = [...clusters.values()].filter((c) => c.runCount >= minSize).sort((a, b) => b.runCount - a.runCount);
+  return { clusters: arr, totalFailures, totalRuns: runs.length };
+}
+function firstErrorMessage(spans) {
+  const errored = spans.find((s) => s.status === "error");
+  return errored?.error;
+}
+
 // src/tool-use-metrics.ts
 async function computeToolUseMetrics(store, runId, options = {}) {
   const tools = await toolSpans(store, runId);
@@ -546,9 +609,10 @@ function normalCdf(x) {
 export {
   DEFAULT_RULES,
   classifyFailure,
+  failureClusterView,
   computeToolUseMetrics,
   compareToBaseline,
   iqr,
   welchsTTest
 };
-//# sourceMappingURL=chunk-OHEPNJQN.js.map
+//# sourceMappingURL=chunk-JLZQWFV3.js.map

package/dist/chunk-JLZQWFV3.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 — legacy clusters without this field\n   * 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;;;ACtaA,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;;;AC9EA,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/cli.js

@@ -5,7 +5,7 @@ import {
   runRpcBatch,
   runRpcOnce,
   startServer
-} from "./chunk-SY6WAAAD.js";
+} from "./chunk-5LBB5B3Z.js";
 import "./chunk-4S4BM3QQ.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-NG236HPC.js";

package/dist/index.d.ts

@@ -6,12 +6,16 @@ import { A as AgentEvalError } from './errors-BZ9sTdz7.js';
 export { a as AgentEvalErrorCode, C as CaptureIntegrityError, b as ConfigError, J as JudgeError, N as NotFoundError, R as ReplayError, V as ValidationError, c as VerificationError } from './errors-BZ9sTdz7.js';
 import { b as FeedbackLabel, n as FeedbackTrajectoryStore, l as FeedbackTrajectory } from './feedback-trajectory-DfFdrraJ.js';
 export { F as FeedbackArtifactType, a as FeedbackAttempt, c as FeedbackLabelKind, d as FeedbackLabelSource, e as FeedbackOptimizerRow, f as FeedbackOutcome, g as FeedbackReplayAdapter, h as FeedbackReplayResult, i as FeedbackSeverity, j as FeedbackSplitPolicy, k as FeedbackTask, m as FeedbackTrajectoryFilter, o as FileSystemFeedbackTrajectoryStore, I as InMemoryFeedbackTrajectoryStore, P as PreferenceMemoryEntry, p as ProposedSideEffect, q as assignFeedbackSplit, r as controlRunToFeedbackTrajectory, s as createFeedbackTrajectory, t as feedbackTrajectoriesToDatasetScenarios, u as feedbackTrajectoriesToOptimizerRows, v as feedbackTrajectoryToDatasetScenario, w as feedbackTrajectoryToOptimizerRow, x as parseFeedbackTrajectoriesJsonl, y as renderPreferenceMemoryMarkdown, z as replayFeedbackTrajectories, A as replayFeedbackTrajectory, B as serializeFeedbackTrajectoriesJsonl, C as summarizePreferenceMemory, D as withAssignedFeedbackSplit } from './feedback-trajectory-DfFdrraJ.js';
-import { A as ActionableSideInfo, a4 as Objective, a5 as ParetoResult, E as EvolvableVariant, V as VariantAggregate, t as TrialResult, o as MutateAdapter, T as TrialCache } from './summary-report-BXGs_9V0.js';
-export { a as AsiSeverity, D as DEFAULT_MUTATION_PRIMITIVES, a6 as Direction, C as GainDistributionBin, F as GainDistributionFigureSpec, H as GainDistributionOptions, a3 as GateDecision, a7 as GateEvidence, G as GenerationReport, a8 as HeldOutGate, a9 as HeldOutGateConfig, aa as HeldOutGateRejectionCode, I as InMemoryTrialCache, M as MultiShotGateConfig, b as MultiShotGateResult, c as MultiShotMutateAdapter, d as MultiShotOptimizationConfig, e as MultiShotOptimizationResult, f as MultiShotRun, g as MultiShotRunInput, h as MultiShotRunner, i as MultiShotScore, j as MultiShotScorer, k as MultiShotSplit, l as MultiShotTrace, m as MultiShotTrialResult, n as MultiShotVariant, J as ParetoFigureSpec, K as ParetoPoint, P as PromptEvolutionConfig, p as PromptEvolutionEvent, q as PromptEvolutionResult, L as RESEARCH_REPORT_HARD_PAIR_FLOOR, R as ReflectionContext, r as ReflectionProposal, N as ResearchReport, O as ResearchReportCandidate, Q as ResearchReportDecision, U as ResearchReportMethodology, W as ResearchReportOptions, X as ResearchReportRecommendation, S as ScenarioAggregate, s as ScoreAdapter, Y as SummaryTable, Z as SummaryTableOptions, _ as SummaryTableRow, u as TrialTrace, v as buildReflectionPrompt, ab as crowdingDistance, w as defaultMultiShotObjectives, ac as dominates, $ as gainHistogram, a0 as paretoChart, ad as paretoFrontier, ae as paretoFrontierWithCrowding, x as parseReflectionResponse, a1 as researchReport, y as runMultiShotOptimization, z as runPromptEvolution, af as scalarScore, a2 as summaryTable, B as trialTraceFromMultiShotTrial } from './summary-report-BXGs_9V0.js';
+import { A as ActionableSideInfo, h as MultiShotRunner, j as MultiShotScorer, c as MultiShotMutateAdapter, a4 as HeldOutGateConfig, E as EvolvableVariant, m as MultiShotTrialResult, e as MultiShotOptimizationResult, a3 as GateDecision, a5 as Objective, a6 as ParetoResult, V as VariantAggregate, t as TrialResult, o as MutateAdapter, T as TrialCache } from './summary-report-C7VPYEj2.js';
+export { a as AsiSeverity, D as DEFAULT_MUTATION_PRIMITIVES, a7 as Direction, C as GainDistributionBin, F as GainDistributionFigureSpec, H as GainDistributionOptions, a8 as GateEvidence, G as GenerationReport, a9 as HeldOutGate, aa as HeldOutGateRejectionCode, I as InMemoryTrialCache, M as MultiShotGateConfig, b as MultiShotGateResult, d as MultiShotOptimizationConfig, f as MultiShotRun, g as MultiShotRunInput, i as MultiShotScore, k as MultiShotSplit, l as MultiShotTrace, n as MultiShotVariant, J as ParetoFigureSpec, K as ParetoPoint, P as PromptEvolutionConfig, p as PromptEvolutionEvent, q as PromptEvolutionResult, L as RESEARCH_REPORT_HARD_PAIR_FLOOR, R as ReflectionContext, r as ReflectionProposal, N as ResearchReport, O as ResearchReportCandidate, Q as ResearchReportDecision, U as ResearchReportMethodology, W as ResearchReportOptions, X as ResearchReportRecommendation, S as ScenarioAggregate, s as ScoreAdapter, Y as SummaryTable, Z as SummaryTableOptions, _ as SummaryTableRow, u as TrialTrace, v as buildReflectionPrompt, ab as crowdingDistance, w as defaultMultiShotObjectives, ac as dominates, $ as gainHistogram, a0 as paretoChart, ad as paretoFrontier, ae as paretoFrontierWithCrowding, x as parseReflectionResponse, a1 as researchReport, y as runMultiShotOptimization, z as runPromptEvolution, af as scalarScore, a2 as summaryTable, B as trialTraceFromMultiShotTrial } from './summary-report-C7VPYEj2.js';
 export { DataAcquisitionPlan, KnowledgeAcquisitionMode, KnowledgeBundle, KnowledgeFallbackPolicy, KnowledgeFreshness, KnowledgeImportance, KnowledgeReadinessReport, KnowledgeRecommendedAction, KnowledgeRequirement, KnowledgeRequirementCategory, KnowledgeResponsibleSurface, KnowledgeSensitivity, ScoreKnowledgeReadinessOptions, UserQuestion, acquisitionPlansForKnowledgeGaps, blockingKnowledgeEval, knowledgeReadinessTracePayload, scoreKnowledgeReadiness, userQuestionsForKnowledgeGaps } from './knowledge/index.js';
-import { i as ReleaseConfidenceThresholds, g as ReleaseConfidenceScorecard } from './release-report-TDPn1cxq.js';
-export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, P as PairedBootstrapOptions, b as PairedBootstrapResult, R as ReleaseConfidenceAxis, c as ReleaseConfidenceAxisName, d as ReleaseConfidenceInput, e as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, h as ReleaseConfidenceStatus, j as ReleaseTraceEvidence, k as RenderReleaseReportOptions, V as Verdict, l as assertReleaseConfidence, m as bhAdjust, n as bootstrapCi, o as evaluateReleaseConfidence, p as judgeReplayGate, q as pairedBootstrap, r as pairedWilcoxon, s as releaseTraceEvidenceFromMultiShotTrials, t as renderReleaseReport } from './release-report-TDPn1cxq.js';
-import { R as Run$1, S as Span, b as TraceEvent, A as Artifact$1, B as BudgetLedgerEntry, T as TraceStore, g as BudgetSpec, h as RunFilter, L as LlmSpan } from './store-Db2Bv8Cf.js';
+import { i as ReleaseConfidenceThresholds, g as ReleaseConfidenceScorecard } from './release-report-BNgMdqPF.js';
+export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, P as PairedBootstrapOptions, b as PairedBootstrapResult, R as ReleaseConfidenceAxis, c as ReleaseConfidenceAxisName, d as ReleaseConfidenceInput, e as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, h as ReleaseConfidenceStatus, j as ReleaseTraceEvidence, k as RenderReleaseReportOptions, V as Verdict, l as assertReleaseConfidence, m as bhAdjust, n as bootstrapCi, o as evaluateReleaseConfidence, p as judgeReplayGate, q as pairedBootstrap, r as pairedWilcoxon, s as releaseTraceEvidenceFromMultiShotTrials, t as renderReleaseReport } from './release-report-BNgMdqPF.js';
+import { a as FailureCluster } from './failure-cluster-C2EGSDiT.js';
+export { D as DEFAULT_FAILURE_RULES, b as FailureClassification, c as FailureContext, d as FailureRule, e as classifyFailure } from './failure-cluster-C2EGSDiT.js';
+import { a as RunSplitTag, R as RunRecord } from './run-record-CqzahIbx.js';
+export { c as RunJudgeMetadata, d as RunOutcome, e as RunRecordValidationError, b as RunTokenUsage, i as isRunRecord, p as parseRunRecordSafe, r as roundTripRunRecord, v as validateRunRecord } from './run-record-CqzahIbx.js';
+import { T as TraceStore, R as Run$1, S as Span, b as TraceEvent, A as Artifact$1, B as BudgetLedgerEntry, g as BudgetSpec, h as RunFilter, L as LlmSpan } from './store-Db2Bv8Cf.js';
 export { i as EventFilter, E as EventKind, j as FAILURE_CLASSES, F as FailureClass, k as FileSystemTraceStore, l as FileSystemTraceStoreOptions, G as GenericSpan, I as InMemoryTraceStore, J as JudgeSpan, M as Message, e as RetrievalSpan, m as RunLayer, n as RunStatus, f as SandboxSpan, o as SpanBase, p as SpanFilter, d as SpanKind, q as SpanStatus, r as TRACE_SCHEMA_VERSION, a as ToolSpan, s as isJudgeSpan, t as isLlmSpan, u as isRetrievalSpan, v as isSandboxSpan, w as isToolSpan } from './store-Db2Bv8Cf.js';
 import { S as SandboxDriver, H as HarnessConfig, a as SandboxHarnessResult } from './test-graded-scenario-B2kWEdh9.js';
 export { D as DockerSandboxDriver, c as SandboxHarness, d as SandboxResult, e as SubprocessSandboxDriver, f as SubprocessSandboxDriverOptions, g as TestGradedRunOptions, b as TestGradedRunResult, T as TestGradedScenario, h as TestOutputParser, i as composeParsers, j as jestTestParser, p as pytestTestParser, r as runTestGradedScenario, v as vitestTestParser } from './test-graded-scenario-B2kWEdh9.js';
@@ -20,7 +24,6 @@ export { R as RunCompleteHook, a as RunCompleteHookContext, S as SpanHandle, b a
 export { F as FileSystemRawProviderSink, d as FileSystemRawProviderSinkOptions, I as InMemoryRawProviderSink, e as InMemoryRawProviderSinkOptions, N as NoopRawProviderSink, P as ProviderRedactor, f as RawProviderDirection, c as RawProviderEvent, R as RawProviderSink, g as RawProviderSinkFilter, h as RunIntegrityError, a as RunIntegrityExpectations, i as RunIntegrityIssue, j as RunIntegrityIssueCode, b as RunIntegrityReport, k as assertRunCaptured, l as defaultProviderRedactor, p as providerFromBaseUrl, t as throwIfRunIncomplete } from './integrity-DK2EBVZC.js';
 export { D as DEFAULT_REDACTION_RULES, O as OTEL_AGENT_EVAL_SCOPE, a as OtlpExport, b as OtlpResourceSpans, c as OtlpSpan, R as REDACTION_VERSION, d as RedactionReport, e as RedactionRule, f as ReplayCache, g as ReplayCacheEntry, h as ReplayCacheMissError, i as ReplayCacheStats, j as ReplayFetchOptions, k as createReplayFetch, l as exportRunAsOtlp, m as iterateRawCalls, r as redactString, n as redactValue } from './replay-BL96gCEP.js';
 export { a as aggregateLlm, b as argHash, g as groupBy, j as judgeSpans, l as llmSpans, r as runFailureClass, c as runsForScenario, t as toolSpans } from './query-DODUYdPg.js';
-export { D as DEFAULT_FAILURE_RULES, b as FailureClassification, c as FailureContext, d as FailureRule, e as classifyFailure } from './failure-cluster-C2EGSDiT.js';
 import { a as BaselineReport } from './baseline-4R5deP0N.js';
 export { B as BaselineOptions, M as MetricSamples, b as MetricVerdict, T as ToolStats, d as ToolUseMetrics, e as ToolUseOptions, f as compareToBaseline, c as computeToolUseMetrics, i as iqr, w as welchsTTest } from './baseline-4R5deP0N.js';
 import { T as Trajectory, a as TrajectoryStep } from './trajectory-CnoBo-JY.js';
@@ -31,14 +34,136 @@ export { C as CalibrationResult, a as CandidateScore, D as DEFAULT_RED_TEAM_CORP
 import { a as PrmGrader } from './rubric-D5tjHNJQ.js';
 import { L as Layer, S as Severity, a as LayerResult, b as VerifyContext } from './multi-layer-verifier-LkP3LVKj.js';
 export { F as Finding, c as LayerStatus, M as MultiLayerVerifier, V as VerificationReport, d as VerifyOptions, g as gradeSemanticStatus } from './multi-layer-verifier-LkP3LVKj.js';
-import { L as LlmClientOptions } from './researcher-CUOiGcGv.js';
-export { C as CallbackResearcher, a as CallbackResearcherOptions, b as CampaignFactoryParams, c as CampaignIntegrityPolicy, d as CampaignRunContext, e as CampaignRunOutcome, f as CampaignRunner, g as CampaignScenario, h as CampaignVariant, E as EvalCampaignOptions, i as EvalCampaignResult, j as ExperimentPlan, k as ExperimentResult, F as FailedRun, l as FailureMode, m as LlmCallError, n as LlmCallRequest, o as LlmCallResult, p as LlmClient, q as LlmMessage, s as LlmRouteAssertionError, t as LlmRouteRequirements, u as LlmUsage, N as NoopResearcher, R as Researcher, S as SteeringChange, v as assertLlmRoute, w as callLlm, x as callLlmJson, y as probeLlm, r as runEvalCampaign, z as stripFencedJson } from './researcher-CUOiGcGv.js';
+import { L as LlmClientOptions } from './researcher-BPT8x_NT.js';
+export { C as CallbackResearcher, a as CallbackResearcherOptions, b as CampaignFactoryParams, c as CampaignIntegrityPolicy, d as CampaignRunContext, e as CampaignRunOutcome, f as CampaignRunner, g as CampaignScenario, h as CampaignVariant, E as EvalCampaignOptions, i as EvalCampaignResult, j as ExperimentPlan, k as ExperimentResult, F as FailedRun, l as FailureMode, m as LlmCallError, n as LlmCallRequest, o as LlmCallResult, p as LlmClient, q as LlmMessage, s as LlmRouteAssertionError, t as LlmRouteRequirements, u as LlmUsage, N as NoopResearcher, R as Researcher, S as SteeringChange, v as assertLlmRoute, w as callLlm, x as callLlmJson, y as probeLlm, r as runEvalCampaign, z as stripFencedJson } from './researcher-BPT8x_NT.js';
 export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, c as BenchmarkEvaluation, d as benchmarkDeterministicSplit, i as benchmarks } from './index--fVrWDiR.js';
-import { R as RunRecord } from './run-record-CqzahIbx.js';
-export { c as RunJudgeMetadata, d as RunOutcome, e as RunRecordValidationError, a as RunSplitTag, b as RunTokenUsage, i as isRunRecord, p as parseRunRecordSafe, r as roundTripRunRecord, v as validateRunRecord } from './run-record-CqzahIbx.js';
 export { I as InterimReleaseConfidence, a as InterimReleaseConfidenceInput, P as PairedEvalueOptions, b as PairedEvalueSequence, c as PairedEvalueStep, S as SequentialDecision, e as evaluateInterimReleaseConfidence, p as pairedEvalueSequence } from './sequential-Dgz1n51-.js';
 import './outcome-store-D6KWmYvj.js';
 
+/**
+ * Automated pull request opener for the production loop.
+ *
+ * `runProductionLoop` produces a `promotedPrompt` string and a release
+ * scorecard. To close the eval → prod → eval cycle the framework needs
+ * to land that prompt as a reviewable code change. This module does
+ * exactly that:
+ *
+ *   1. Stage a branch off `baseBranch`.
+ *   2. Write each `fileChange` into the worktree.
+ *   3. Commit + push.
+ *   4. Open a PR via the GitHub API.
+ *
+ * Two transports ship in core:
+ *
+ *   - `ghCliClient(opts)` — shells out to the `gh` CLI. No extra deps,
+ *     re-uses the developer machine's `gh auth` state, works with both
+ *     github.com and GitHub Enterprise. This is the recommended default.
+ *   - `httpGithubClient(opts)` — direct `fetch` against `api.github.com`
+ *     with a bearer token. Useful in CI where `gh` may not be installed.
+ *
+ * Both implement the small `AutoPrClient` interface, so tests substitute
+ * a fake without spinning a process or network.
+ *
+ * @experimental — added in 0.25.0. Surface may evolve as consumers wire
+ * it into CI workflows.
+ */
+interface FileChange {
+    /** Repo-relative path. Forward slashes; no `..`. */
+    path: string;
+    /** New file contents. UTF-8. */
+    contents: string;
+    /** Optional explanatory comment shown in the commit body. */
+    rationale?: string;
+}
+interface RepoRef {
+    owner: string;
+    name: string;
+}
+interface ProposeAutomatedPullRequestInput {
+    repo: RepoRef;
+    /** Branch to base the PR on. Default `'main'`. */
+    baseBranch?: string;
+    /** New branch name. Use a prefix + a short stable id; no spaces. */
+    branchName: string;
+    fileChanges: FileChange[];
+    title: string;
+    body: string;
+    /** Optional GitHub usernames to request review from. */
+    reviewers?: string[];
+    /** Optional labels to apply. */
+    labels?: string[];
+    /** Commit author name. Default: derived from the GitHub client. */
+    authorName?: string;
+    /** Commit author email. Default: derived from the GitHub client. */
+    authorEmail?: string;
+    /** Dry-run — do not push or open a PR; just return the would-be plan. */
+    dryRun?: boolean;
+}
+interface ProposeAutomatedPullRequestResult {
+    prUrl: string;
+    branchName: string;
+    headSha: string;
+    dryRun: boolean;
+}
+/** Pluggable transport for the auto-PR pipeline. */
+interface AutoPrClient {
+    /**
+     * Create a branch from `baseBranch`, write file changes, commit, push,
+     * and open a PR. Returns the PR's HTML url and head SHA.
+     *
+     * Implementations must be idempotent on `branchName`: if the branch
+     * already exists with the same head SHA as the would-be commit, return
+     * the existing PR rather than failing. This makes the production loop
+     * safe to retry on transient errors.
+     */
+    proposeChange(input: ProposeAutomatedPullRequestInput): Promise<ProposeAutomatedPullRequestResult>;
+}
+declare function proposeAutomatedPullRequest(client: AutoPrClient, input: ProposeAutomatedPullRequestInput): Promise<ProposeAutomatedPullRequestResult>;
+interface HttpGithubClientOptions {
+    /** Personal access token, GitHub App token, or `GITHUB_TOKEN` from Actions. */
+    token: string;
+    /** Override for GitHub Enterprise. Default `'https://api.github.com'`. */
+    apiBase?: string;
+    /** Test seam — defaults to global `fetch`. */
+    fetchImpl?: typeof fetch;
+    /** Test seam — clock for commit timestamps. */
+    now?: () => Date;
+}
+/**
+ * Direct REST-API GitHub client. No external deps.
+ *
+ * Idempotency strategy: before creating refs/commits/PRs, check whether
+ * the branch already exists at the desired tree. If so, return the
+ * existing PR (or open one if missing). Errors from concurrent runs
+ * (`Reference already exists`) are caught and treated as success.
+ */
+declare function httpGithubClient(opts: HttpGithubClientOptions): AutoPrClient;
+interface GhCliClientOptions {
+    /** Override the CLI binary (`gh`). For testing. */
+    bin?: string;
+    /** Working directory containing a clone of `repo`. Default: process cwd. */
+    cwd?: string;
+    /** Test seam: process spawner. Default: node:child_process spawn. */
+    exec?: (bin: string, args: string[], opts: {
+        cwd: string;
+        stdin?: string;
+    }) => Promise<{
+        stdout: string;
+        stderr: string;
+        exitCode: number;
+    }>;
+}
+/**
+ * `gh` CLI transport. Requires:
+ *   - `gh` installed and authenticated (`gh auth status`).
+ *   - A local clone of the repo with a clean working tree.
+ *   - `git` on PATH.
+ *
+ * Uses `gh api` for repo metadata and `gh pr create` for the PR. The
+ * actual commit lands via `git`, which keeps `gh`'s footprint minimal.
+ */
+declare function ghCliClient(opts?: GhCliClientOptions): AutoPrClient;
+
 interface Scenario {
     id: string;
     persona: string;
@@ -595,6 +720,181 @@ declare class MetricsCollector {
     getConvergenceCurve(): number[];
 }
 
+/**
+ * ProductionLoop — the substrate that closes eval → prod → eval.
+ *
+ * Static prompts decay. Yesterday's regulation flips today; yesterday's
+ * tool quirk becomes today's incident. A production agent that ships a
+ * static prompt and never re-trains is on a clock.
+ *
+ * `runProductionLoop` is the orchestration layer over the eval substrate:
+ *
+ *   1. Ingest production traces + user feedback (via the wire HTTP
+ *      ingestion endpoints, or directly through any `TraceStore` and
+ *      `FeedbackTrajectoryStore` implementation).
+ *   2. Cluster the failures (`failureClusterView`) and prioritize by
+ *      size × severity.
+ *   3. If any cluster crosses the consumer's threshold, run a
+ *      `runMultiShotOptimization` round seeded by the current production
+ *      prompt against holdout-shape scenarios derived from the offending
+ *      cluster.
+ *   4. Gate the promoted prompt with `evaluateReleaseConfidence`. Fail
+ *      closed.
+ *   5. If the gate passes and an `AutoPrClient` is wired, open a PR with
+ *      the new prompt. Otherwise return the proposed change.
+ *
+ * One call = one cycle. Cron / GitHub Actions are the caller's job. The
+ * primitive is idempotent + replayable: re-running with the same
+ * `runId` will produce the same plan.
+ *
+ * @experimental — added in 0.25.0. Surface may evolve as the 5 product
+ * agents wire it in.
+ */
+
+interface FailureClusterConfig {
+    /** Minimum runs in a cluster before it triggers an evolve round. Default 5. */
+    minClusterSize?: number;
+    /**
+     * Severity threshold. A cluster is "actionable" when its size
+     * normalized by total runs exceeds this. Default 0.05 (5% of all runs).
+     */
+    minSeverityRatio?: number;
+    /**
+     * Maximum number of clusters to react to in one cycle. Acting on too
+     * many at once obscures attribution. Default 1 — the worst cluster.
+     */
+    maxClustersPerCycle?: number;
+}
+interface ProductionEvolveConfig<P = string> {
+    /** How to run a candidate prompt against a scenario. */
+    runner: MultiShotRunner<P>;
+    /** How to score the trajectory. Usually a calibrated judge. */
+    scorer: MultiShotScorer<P>;
+    /** How to mutate. Addendum-style mutators (append vs. rewrite) work best. */
+    mutator: MultiShotMutateAdapter<P>;
+    /** The current production prompt. Acts as the baseline + seed. */
+    baselinePrompt: P;
+    /** Stable id for the baseline variant. Default `'baseline'`. */
+    baselineId?: string;
+    /** Scenarios resembling production load. Used as the holdout split. */
+    holdoutScenarios: Scenario[];
+    /** Scenarios used during search. Default: derived from `holdoutScenarios` via deterministic split. */
+    searchScenarios?: Scenario[];
+    /** Gate config for the held-out promotion check. */
+    gate: HeldOutGateConfig;
+    /** Reps per (variant × scenario) cell. Default 3. */
+    reps?: number;
+    /** Number of mutation generations. Default 3. */
+    generations?: number;
+    /** Population size per generation. Default 4. */
+    populationSize?: number;
+    /** Concurrent score() calls. Default 1. */
+    scoreConcurrency?: number;
+    /**
+     * Optional bridge from a scored trial into a paper-grade RunRecord.
+     * If omitted, the loop synthesises a minimal record sufficient for
+     * `HeldOutGate` and `evaluateReleaseConfidence`.
+     */
+    toRunRecord?: (input: {
+        variant: EvolvableVariant<P>;
+        scenarioId: string;
+        rep: number;
+        split: RunSplitTag;
+        seed: number;
+        trial: MultiShotTrialResult;
+    }) => RunRecord;
+}
+interface ProductionShipConfig {
+    repo: RepoRef;
+    /** Branch name prefix. Final branch = `${branchPrefix}/${runId}`. */
+    branchPrefix: string;
+    /** Path (repo-relative) of the file holding the production prompt. */
+    promptFilePath: string;
+    /** Base branch for the PR. Default `'main'`. */
+    baseBranch?: string;
+    reviewers?: string[];
+    labels?: string[];
+    /** Required: the auto-PR transport. Use `ghCliClient()` or `httpGithubClient()`. */
+    client: AutoPrClient;
+    /** Skip the actual push + PR call — for sanity-checking the plan. Default false. */
+    dryRun?: boolean;
+    /** Render PR body from the loop's findings. Optional override. */
+    renderBody?: (ctx: ProductionLoopRenderContext) => string;
+    /** Render the file contents from the new prompt. Default: serialize as the file. */
+    renderPromptFile?: (newPrompt: string, oldFileContents: string | null) => string;
+    /** Read the current prompt file contents for diff context. Optional. */
+    readCurrentPromptFile?: () => Promise<string | null>;
+}
+interface ProductionLoopCronConfig {
+    cadence: 'weekly' | 'daily' | 'hourly';
+    /** Optional jitter (seconds) the consumer's scheduler should add. Surface-only. */
+    jitterSec?: number;
+}
+interface RunProductionLoopOptions<P = string> {
+    /** Stable id; deterministic outputs when reused. */
+    runId: string;
+    /** Human label — surfaces in PR titles and reports. */
+    target: string;
+    traceStore: TraceStore;
+    feedbackStore: FeedbackTrajectoryStore;
+    cluster: FailureClusterConfig;
+    evolve: ProductionEvolveConfig<P>;
+    /** When omitted, the loop returns the proposed prompt without opening a PR. */
+    ship?: ProductionShipConfig;
+    /** Surface-only — encodes scheduler expectations into the artifact. */
+    cron?: ProductionLoopCronConfig;
+    /** Release confidence thresholds. Default: library defaults. */
+    releaseThresholds?: ReleaseConfidenceThresholds;
+    /** Now() seam for reproducibility in tests. */
+    now?: () => Date;
+}
+type ProductionLoopDecision = 'no_actionable_failures' | 'evolve_yielded_no_improvement' | 'gate_failed' | 'proposed_change' | 'pr_opened';
+interface ProductionLoopRenderContext {
+    runId: string;
+    target: string;
+    decision: ProductionLoopDecision;
+    /** Clusters seen in production this cycle, sorted by severity. */
+    clusters: FailureCluster[];
+    /** The cluster the loop acted on (if any). */
+    actedOnCluster: FailureCluster | null;
+    /** Production runs observed this cycle. */
+    observedRunCount: number;
+    /** Feedback trajectories observed this cycle. */
+    observedFeedbackCount: number;
+    /** Evolve result (if evolve ran). */
+    evolution: MultiShotOptimizationResult<unknown> | null;
+    /** Release gate verdict (if evolve ran). */
+    release: ReleaseConfidenceScorecard | null;
+    /** Held-out gate decision (if a candidate was paired against the baseline). */
+    gate: GateDecision | null;
+    /** The baseline (current production) prompt as a string. */
+    baselinePromptString: string;
+    /** The proposed new prompt as a string. Empty if no change was proposed. */
+    promotedPromptString: string;
+}
+interface ProductionLoopResult {
+    runId: string;
+    target: string;
+    decision: ProductionLoopDecision;
+    startedAt: string;
+    finishedAt: string;
+    observedRunCount: number;
+    observedFeedbackCount: number;
+    clusters: FailureCluster[];
+    actedOnCluster: FailureCluster | null;
+    evolution: MultiShotOptimizationResult<unknown> | null;
+    release: ReleaseConfidenceScorecard | null;
+    gate: GateDecision | null;
+    /** Baseline prompt as it entered the cycle. */
+    baselinePrompt: unknown;
+    /** Promoted prompt — equals baseline when no change is proposed. */
+    promotedPrompt: unknown;
+    /** PR artifact when `ship` was wired and gate passed. */
+    pullRequest: ProposeAutomatedPullRequestResult | null;
+    cron: ProductionLoopCronConfig | null;
+}
+declare function runProductionLoop<P = string>(opts: RunProductionLoopOptions<P>): Promise<ProductionLoopResult>;
+
 /**
  * ScenarioRegistry — manages scenario discovery and filtering.
  *
@@ -4617,4 +4917,4 @@ interface ReferenceReplaySteeringRowsOptions<Input = unknown> {
 declare function referenceReplayRunsToSteeringRows<Input = unknown>(runs: ReferenceReplayRun<Input>[], options?: ReferenceReplaySteeringRowsOptions<Input>): SteeringOptimizationRow[];
 declare function referenceReplayScenarioToRunScore(scenarioScore: ReferenceReplayScenarioScore, durationMs?: number): RunScore;
 
-export { ActionableSideInfo, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AlignmentOp, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, Artifact$1 as Artifact, type ArtifactCheck, type Artifact as ArtifactCheckArtifact, type ArtifactResult, type ArtifactValidator, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, BaselineReport, BehaviorAssertion, type BenchmarkReport, BenchmarkRunner, type BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateScenario, type CausalAttributionReport, type CheckResult, type CodeMutationOutcome, type CodeMutationRunner, type CollectedArtifacts, type CommandRunner, type CompletionCriterion, type CompositePolicy, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractMetric, type ContractReport, ControlEvalResult, ConvergenceTracker, type CostEntry, CostLedger, type CostLedgerGeneration, type CostLedgerSnapshot, type CostSummary, CostTracker, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateCompositeMutatorOpts, type CreateDefaultReviewerOptions, type CreateSandboxCodeMutatorOpts, type CreateSandboxPoolOpts, type CrossTraceDiff, type CrossTraceDiffOptions, D1ExperimentStore, type D1ExperimentStoreOptions, type D1Like, type D1PreparedStatementLike, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATORS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DirEntry, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type ErrorCountPattern, type EvalResult, type EvolutionRound, EvolvableVariant, type ExecutorConfig, type Expectation, type Experiment, type Run as ExperimentRun, type ExperimentStore, ExperimentTracker, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, FeedbackLabel, type FeedbackPattern, FeedbackTrajectory, FeedbackTrajectoryStore, FileSystemExperimentStore, type FileSystemExperimentStoreOptions, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GoldenSeverity, type GoldenSpec, type HarnessAdapter, HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, HoldoutAuditor, type HostedJudgeConfig, type HostedJudgeDimension, type HostedJudgeRequest, type HostedJudgeResponse, type HostedRunCriticConfig, type HostedRunScoreRequest, type HostedRunScoreResponse, type HypothesisManifest, type HypothesisResult, INTENT_MATCH_JUDGE_VERSION, type ImageData, InMemoryExperimentStore, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntegrationGateSurface, type IntegrationInvokeFailureInput, type IntegrationManifestGateInput, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JsonlTrialCache, type JudgeConfig, type JudgeFleetOptions, type JudgeFn, type JudgeInput, type JudgeReplayResult, type JudgeRubric, JudgeRunner, type JudgeScore, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LineageKind, type LineageKindResolver, type LineageNode, LineageRecorder, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, LockedJsonlAppender, MODEL_PRICING, type MatchResult, type MatcherResult, type MeasurementPolicy, type MergeOptions, MetricsCollector, type MuffledFinder, type MuffledFinding, type MultiToolchainLayerConfig, MutateAdapter, type MutationAttempt, type MutationChannel, MutationTelemetry, type Mutator, Mutex, Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, ParetoResult, type PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, ProductClient, type ProductClientConfig, type PromptHandle, PromptRegistry, TrialResult as PromptTrialResult, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RouteMap, type RubricDimension, Run$1 as Run, type RunCommandInput, type RunCommandResult, type RunConfig, RunCritic, type RunCriticOptions, type RunDiff, RunFilter, RunRecord, type RunScore, type RunScoreWeights, type RunTrace, SEMANTIC_CONCEPT_JUDGE_VERSION, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, type Scenario, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoredTarget, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SeriesConvergenceOptions, type SeriesConvergenceResult, Severity, type SignedManifest, type SignedManifestAlgo, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, Span, type SteeringBundle, type SteeringDelta, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type StepAttribution, type SynthesisReason, type SynthesisTarget, type TestResult, type ThresholdContract, TokenCounter, type TokenSpec, TraceEmitter, TraceEvent, TraceStore, Trajectory, TrajectoryStep, type TrialAttempt, TrialCache, TrialTelemetry, type Turn, type TurnMetrics, type TurnResult, UNIVERSAL_FINDERS, type ValidationContext, type ValidationIssue, type ValidationResult, VariantAggregate, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, aggregateRunScore, analyzeAntiSlop, analyzeSeries, attributeCounterfactuals, benjaminiHochberg, bisect, bonferroni, buildReviewerPrompt, byteLengthRange, canaryLeakView, canonicalize, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, clamp01, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commitBisect, compareReferenceReplay, compilerJudge, composeValidators, confidenceInterval, containsAll, createAntiSlopJudge, createCompositeMutator, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxCodeMutator, createSandboxPool, createSemanticConceptJudge, crossTraceDiff, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultReferenceReplayMatcher, deployGateLayer, distillPlaybook, estimateCost, estimateTokens, evaluateContract, evaluateHypothesis, evaluateOracles, executeScenario, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fileContains, fileExists, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, formatBenchmarkReport, formatDriverReport, formatFindings, precision as goldenPrecision, hashContent, hashJson, htmlContainsElement, inMemoryReferenceReplayStore, integrationAsi, integrationGateEvals, integrationInvokeFailedPayload, integrationManifestResolvedPayload, integrationManifestValidatedPayload, interRaterReliability, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, keyPreserved, linterJudge, loadScorerFromGrader, localCommandRunner, lowercaseMutator, mannWhitneyU, matchGoldens, mergeLayerResults, mergeSteeringBundle, multiToolchainLayer, normalizeScores, notBlocked, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, partialCredit, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, printDriverSummary, promptBisect, proposeSynthesisTargets, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, renderSteeringText, replayScorerOverCorpus, replayTraceThroughJudge, requiredSampleSize, resetLockedAppendersForTesting, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runE2EWorkflow, runExpectations, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runReferenceReplay, runSelfPlay, runSemanticConceptJudge, scanForMuffledGates, scoreContinuity, scoreReferenceReplay, securityJudge, selectHarnessVariant, sentenceReorderMutator, signManifest, statusAdvanced, summarizeHarnessResults, testJudge, textInSnapshot, toLangfuseEnvelope, toPrometheusText, typoMutator, urlContains, verifyManifest, visualDiff, viteDeployRunner, weightedMean, weightedRecall, whitespaceCollapseMutator, wilcoxonSignedRank, wranglerDeployRunner };
+export { ActionableSideInfo, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AlignmentOp, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, Artifact$1 as Artifact, type ArtifactCheck, type Artifact as ArtifactCheckArtifact, type ArtifactResult, type ArtifactValidator, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, BaselineReport, BehaviorAssertion, type BenchmarkReport, BenchmarkRunner, type BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateScenario, type CausalAttributionReport, type CheckResult, type CodeMutationOutcome, type CodeMutationRunner, type CollectedArtifacts, type CommandRunner, type CompletionCriterion, type CompositePolicy, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractMetric, type ContractReport, ControlEvalResult, ConvergenceTracker, type CostEntry, CostLedger, type CostLedgerGeneration, type CostLedgerSnapshot, type CostSummary, CostTracker, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateCompositeMutatorOpts, type CreateDefaultReviewerOptions, type CreateSandboxCodeMutatorOpts, type CreateSandboxPoolOpts, type CrossTraceDiff, type CrossTraceDiffOptions, D1ExperimentStore, type D1ExperimentStoreOptions, type D1Like, type D1PreparedStatementLike, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATORS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DirEntry, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type ErrorCountPattern, type EvalResult, type EvolutionRound, EvolvableVariant, type ExecutorConfig, type Expectation, type Experiment, type Run as ExperimentRun, type ExperimentStore, ExperimentTracker, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, type FailureClusterConfig, FeedbackLabel, type FeedbackPattern, FeedbackTrajectory, FeedbackTrajectoryStore, type FileChange, FileSystemExperimentStore, type FileSystemExperimentStoreOptions, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, GateDecision, type GhCliClientOptions, type GoldenSeverity, type GoldenSpec, type HarnessAdapter, HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, HeldOutGateConfig, HoldoutAuditor, type HostedJudgeConfig, type HostedJudgeDimension, type HostedJudgeRequest, type HostedJudgeResponse, type HostedRunCriticConfig, type HostedRunScoreRequest, type HostedRunScoreResponse, type HttpGithubClientOptions, type HypothesisManifest, type HypothesisResult, INTENT_MATCH_JUDGE_VERSION, type ImageData, InMemoryExperimentStore, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntegrationGateSurface, type IntegrationInvokeFailureInput, type IntegrationManifestGateInput, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JsonlTrialCache, type JudgeConfig, type JudgeFleetOptions, type JudgeFn, type JudgeInput, type JudgeReplayResult, type JudgeRubric, JudgeRunner, type JudgeScore, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LineageKind, type LineageKindResolver, type LineageNode, LineageRecorder, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, LockedJsonlAppender, MODEL_PRICING, type MatchResult, type MatcherResult, type MeasurementPolicy, type MergeOptions, MetricsCollector, type MuffledFinder, type MuffledFinding, MultiShotMutateAdapter, MultiShotOptimizationResult, MultiShotRunner, MultiShotScorer, MultiShotTrialResult, type MultiToolchainLayerConfig, MutateAdapter, type MutationAttempt, type MutationChannel, MutationTelemetry, type Mutator, Mutex, Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, ParetoResult, type PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, ProductClient, type ProductClientConfig, type ProductionEvolveConfig, type ProductionLoopCronConfig, type ProductionLoopDecision, type ProductionLoopRenderContext, type ProductionLoopResult, type ProductionShipConfig, type PromptHandle, PromptRegistry, TrialResult as PromptTrialResult, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type RepoRef, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RouteMap, type RubricDimension, Run$1 as Run, type RunCommandInput, type RunCommandResult, type RunConfig, RunCritic, type RunCriticOptions, type RunDiff, RunFilter, type RunProductionLoopOptions, RunRecord, type RunScore, type RunScoreWeights, RunSplitTag, type RunTrace, SEMANTIC_CONCEPT_JUDGE_VERSION, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, type Scenario, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoredTarget, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SeriesConvergenceOptions, type SeriesConvergenceResult, Severity, type SignedManifest, type SignedManifestAlgo, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, Span, type SteeringBundle, type SteeringDelta, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type StepAttribution, type SynthesisReason, type SynthesisTarget, type TestResult, type ThresholdContract, TokenCounter, type TokenSpec, TraceEmitter, TraceEvent, TraceStore, Trajectory, TrajectoryStep, type TrialAttempt, TrialCache, TrialTelemetry, type Turn, type TurnMetrics, type TurnResult, UNIVERSAL_FINDERS, type ValidationContext, type ValidationIssue, type ValidationResult, VariantAggregate, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, aggregateRunScore, analyzeAntiSlop, analyzeSeries, attributeCounterfactuals, benjaminiHochberg, bisect, bonferroni, buildReviewerPrompt, byteLengthRange, canaryLeakView, canonicalize, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, clamp01, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commitBisect, compareReferenceReplay, compilerJudge, composeValidators, confidenceInterval, containsAll, createAntiSlopJudge, createCompositeMutator, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxCodeMutator, createSandboxPool, createSemanticConceptJudge, crossTraceDiff, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultReferenceReplayMatcher, deployGateLayer, distillPlaybook, estimateCost, estimateTokens, evaluateContract, evaluateHypothesis, evaluateOracles, executeScenario, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fileContains, fileExists, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, formatBenchmarkReport, formatDriverReport, formatFindings, ghCliClient, precision as goldenPrecision, hashContent, hashJson, htmlContainsElement, httpGithubClient, inMemoryReferenceReplayStore, integrationAsi, integrationGateEvals, integrationInvokeFailedPayload, integrationManifestResolvedPayload, integrationManifestValidatedPayload, interRaterReliability, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, keyPreserved, linterJudge, loadScorerFromGrader, localCommandRunner, lowercaseMutator, mannWhitneyU, matchGoldens, mergeLayerResults, mergeSteeringBundle, multiToolchainLayer, normalizeScores, notBlocked, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, partialCredit, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, printDriverSummary, promptBisect, proposeAutomatedPullRequest, proposeSynthesisTargets, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, renderSteeringText, replayScorerOverCorpus, replayTraceThroughJudge, requiredSampleSize, resetLockedAppendersForTesting, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runE2EWorkflow, runExpectations, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runProductionLoop, runReferenceReplay, runSelfPlay, runSemanticConceptJudge, scanForMuffledGates, scoreContinuity, scoreReferenceReplay, securityJudge, selectHarnessVariant, sentenceReorderMutator, signManifest, statusAdvanced, summarizeHarnessResults, testJudge, textInSnapshot, toLangfuseEnvelope, toPrometheusText, typoMutator, urlContains, verifyManifest, visualDiff, viteDeployRunner, weightedMean, weightedRecall, whitespaceCollapseMutator, wilcoxonSignedRank, wranglerDeployRunner };