@tangle-network/agent-eval 0.53.0 → 0.54.0

package/dist/adapters/http.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-Dbj5gu8n.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-BgrxOJSf.js';
 
 /**
  * # `@tangle-network/agent-eval/adapters/http` — distributed Dispatch over HTTP.

package/dist/adapters/langchain.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-Dbj5gu8n.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-BgrxOJSf.js';
 
 /**
  * # `@tangle-network/agent-eval/adapters/langchain` — wrap any LangChain

package/dist/adapters/otel.d.ts

@@ -1,10 +1,11 @@
-import { T as TraceSpanEvent, H as HostedClient } from '../index-C7RhhEME.js';
-import '../types-Dbj5gu8n.js';
-import '../summary-report-B7gNRX-r.js';
-import '../run-record-BGY6bHRh.js';
+import { T as TraceSpanEvent, H as HostedClient } from '../index-D2nT6_KT.js';
+import '../types-BgrxOJSf.js';
+import '../summary-report-DLxh4yWk.js';
+import '../run-record-etiCMsUq.js';
 import '../errors-mje_cKOs.js';
-import '../failure-cluster-Cw65_5FY.js';
-import '../store-Db2Bv8Cf.js';
+import '../schema-m0gsnbt3.js';
+import '../failure-cluster-CL7IVgkJ.js';
+import '../store-CKUAgsJz.js';
 import '../judge-calibration-DilmB3Ml.js';
 
 /**

package/dist/{baseline-4R5deP0N.d.ts → baseline-DE36-Np7.d.ts}

@@ -1,4 +1,4 @@
-import { T as TraceStore } from './store-Db2Bv8Cf.js';
+import { T as TraceStore } from './store-CKUAgsJz.js';
 
 /**
  * Tool-use metrics — derived purely from trace data.

package/dist/benchmarks/index.d.ts

@@ -1,3 +1,4 @@
-export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, c as BenchmarkEvaluation, d as deterministicSplit, e as routing } from '../index-0pu_fBwZ.js';
-import '../run-record-BGY6bHRh.js';
+export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, c as BenchmarkEvaluation, d as deterministicSplit, e as routing } from '../index-wlaiph9Y.js';
+import '../run-record-etiCMsUq.js';
 import '../errors-mje_cKOs.js';
+import '../schema-m0gsnbt3.js';

package/dist/builder-eval/index.d.ts

@@ -1,6 +1,7 @@
-import { S as SandboxDriver, H as HarnessConfig, a as SandboxHarnessResult, T as TestGradedScenario, b as TestGradedRunResult } from '../test-graded-scenario-B2kWEdh9.js';
-import { T as TraceEmitter } from '../emitter-DP_cSSiw.js';
-import { T as TraceStore, R as Run } from '../store-Db2Bv8Cf.js';
+import { S as SandboxDriver, H as HarnessConfig, a as SandboxHarnessResult, T as TestGradedScenario, b as TestGradedRunResult } from '../test-graded-scenario-BdVaPyHT.js';
+import { T as TraceEmitter } from '../emitter-DEZwY14K.js';
+import { R as Run } from '../schema-m0gsnbt3.js';
+import { T as TraceStore } from '../store-CKUAgsJz.js';
 
 /**
  * BuilderSession — ties a builder-of-builders workflow together.

package/dist/campaign/index.d.ts

@@ -1,13 +1,14 @@
-export { C as CampaignStorage, D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, G as GepaDriverConstraints, a as GepaDriverOptions, H as HeldOutGateOptions, O as OpenAutoPrOptions, b as OpenAutoPrResult, R as RunCampaignOptions, c as RunEvalOptions, d as RunImprovementLoopOptions, e as RunImprovementLoopResult, f as RunOptimizationOptions, g as RunOptimizationResult, h as composeGate, i as countSentenceEdits, j as defaultProductionGate, k as evolutionaryDriver, l as extractH2Sections, m as fsCampaignStorage, n as gepaDriver, o as heldOutGate, p as inMemoryCampaignStorage, q as openAutoPr, r as runCampaign, s as runEval, t as runImprovementLoop, u as runOptimization, v as surfaceHash } from '../run-improvement-loop-Cc7oZlRP.js';
-import { L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, C as CodeSurface } from '../types-Dbj5gu8n.js';
-export { f as CampaignAggregates, g as CampaignArtifactWriter, h as CampaignCellResult, i as CampaignCostMeter, j as CampaignResult, k as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, l as GateContext, m as GateDecision, n as GateResult, o as GenerationCandidate, p as GenerationRecord, I as ImprovementDriver, q as JudgeAggregate, a as JudgeConfig, r as JudgeDimension, J as JudgeScore, s as LabeledScenarioSource, M as MutableSurface, t as Mutator, O as OptimizerConfig, P as ProposeContext, R as RedactionStatus, S as Scenario, u as ScenarioAggregate, v as SessionScript, T as TraceSpan } from '../types-Dbj5gu8n.js';
+export { C as CampaignStorage, D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, G as GepaDriverConstraints, a as GepaDriverOptions, H as HeldOutGateOptions, O as OpenAutoPrOptions, b as OpenAutoPrResult, R as RunCampaignOptions, c as RunEvalOptions, d as RunImprovementLoopOptions, e as RunImprovementLoopResult, f as RunOptimizationOptions, g as RunOptimizationResult, h as composeGate, i as countSentenceEdits, j as defaultProductionGate, k as evolutionaryDriver, l as extractH2Sections, m as fsCampaignStorage, n as gepaDriver, o as heldOutGate, p as inMemoryCampaignStorage, q as openAutoPr, r as runCampaign, s as runEval, t as runImprovementLoop, u as runOptimization, v as surfaceHash } from '../run-improvement-loop-BhfdjrMY.js';
+import { L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, f as LabelTrust, C as CodeSurface } from '../types-BgrxOJSf.js';
+export { g as CampaignAggregates, h as CampaignArtifactWriter, i as CampaignCellResult, j as CampaignCostMeter, k as CampaignResult, l as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, m as GateContext, n as GateDecision, o as GateResult, p as GenerationCandidate, q as GenerationRecord, I as ImprovementDriver, r as JudgeAggregate, a as JudgeConfig, s as JudgeDimension, J as JudgeScore, t as LabeledScenarioSource, M as MutableSurface, u as Mutator, O as OptimizerConfig, P as ProposeContext, R as RedactionStatus, S as Scenario, v as ScenarioAggregate, w as SessionScript, T as TraceSpan, x as labelTrustRank } from '../types-BgrxOJSf.js';
 import '../llm-client-BXVRUZyX.js';
 import '../errors-mje_cKOs.js';
 import '../raw-provider-sink-C46HDghv.js';
-import '../red-team-30II1T4o.js';
+import '../red-team-CrC5MZYd.js';
 import '../dataset-BlwAtYYf.js';
-import '../store-Db2Bv8Cf.js';
-import '../run-record-BGY6bHRh.js';
+import '../store-CKUAgsJz.js';
+import '../schema-m0gsnbt3.js';
+import '../run-record-etiCMsUq.js';
 
 /**
  * @experimental
@@ -62,6 +63,7 @@ declare class FsLabeledScenarioStore implements LabeledScenarioStore {
         train: number;
         test: number;
         bySource: Record<string, number>;
+        byTrust: Record<LabelTrust, number>;
     }>;
     private assertProvenance;
     private assertRateLimit;
@@ -124,4 +126,4 @@ declare function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAd
  *  as a ref under the adapter's worktree dir. */
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 
-export { CodeSurface, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type GitWorktreeAdapterOptions, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type Worktree, type WorktreeAdapter, WorktreeAdapterError, gitWorktreeAdapter, resolveWorktreePath };
+export { CodeSurface, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type GitWorktreeAdapterOptions, LabelTrust, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type Worktree, type WorktreeAdapter, WorktreeAdapterError, gitWorktreeAdapter, resolveWorktreePath };

package/dist/campaign/index.js

@@ -30,6 +30,18 @@ import "../chunk-NSBPE2FW.js";
 import { createHash } from "crypto";
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
 import { join } from "path";
+
+// src/campaign/types.ts
+var LABEL_TRUST_RANK = {
+  unverified: 0,
+  "verified-signal": 1,
+  "human-rated": 2
+};
+function labelTrustRank(trust) {
+  return LABEL_TRUST_RANK[trust ?? "unverified"];
+}
+
+// src/campaign/labeled-store/fs-adapter.ts
 var LabeledScenarioStoreError = class extends Error {
   constructor(code, message) {
     super(message);
@@ -97,6 +109,11 @@ var FsLabeledScenarioStore = class {
   }
   async size() {
     const bySource = {};
+    const byTrust = {
+      unverified: 0,
+      "verified-signal": 0,
+      "human-rated": 0
+    };
     let total = 0;
     for (const source of ALL_SOURCES) {
       const path = this.pathForSource(source);
@@ -104,11 +121,19 @@ var FsLabeledScenarioStore = class {
         bySource[source] = 0;
         continue;
       }
-      const count = readFileSync(path, "utf8").split("\n").filter(Boolean).length;
-      bySource[source] = count;
-      total += count;
+      const lines = readFileSync(path, "utf8").split("\n").filter(Boolean);
+      bySource[source] = lines.length;
+      total += lines.length;
+      for (const line of lines) {
+        let trust = "unverified";
+        try {
+          trust = JSON.parse(line).labelTrust ?? "unverified";
+        } catch {
+        }
+        byTrust[trust] += 1;
+      }
     }
-    return { train: total, test: total, bySource };
+    return { train: total, test: total, bySource, byTrust };
   }
   assertProvenance(write) {
     if (!write.source) {
@@ -207,6 +232,9 @@ function matchesFilter(record, args, source) {
     if (f.minComposite !== void 0 && max < f.minComposite) return false;
     if (f.maxComposite !== void 0 && max > f.maxComposite) return false;
   }
+  if (f.minTrust !== void 0 && labelTrustRank(record.labelTrust) < labelTrustRank(f.minTrust)) {
+    return false;
+  }
   return true;
 }
 function sha256(input) {
@@ -294,6 +322,7 @@ export {
   gitWorktreeAdapter,
   heldOutGate,
   inMemoryCampaignStorage,
+  labelTrustRank,
   openAutoPr,
   resolveWorktreePath,
   runCampaign,

package/dist/campaign/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n} from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{ train: number; test: number; bySource: Record<string, number> }> {\n    const bySource: Record<string, number> = {}\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const count = readFileSync(path, 'utf8').split('\\n').filter(Boolean).length\n      bySource[source] = count\n      total += count\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAmBd,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAAmF;AACvF,UAAM,WAAmC,CAAC;AAC1C,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO,EAAE;AACrE,eAAS,MAAM,IAAI;AACnB,eAAS;AAAA,IACX;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,SAAS;AAAA,EAC/C;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AC1QA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}
+{"version":3,"sources":["../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/types.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n  LabelTrust,\n} from '../types'\nimport { labelTrustRank } from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{\n    train: number\n    test: number\n    bySource: Record<string, number>\n    byTrust: Record<LabelTrust, number>\n  }> {\n    const bySource: Record<string, number> = {}\n    const byTrust: Record<LabelTrust, number> = {\n      unverified: 0,\n      'verified-signal': 0,\n      'human-rated': 0,\n    }\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      bySource[source] = lines.length\n      total += lines.length\n      for (const line of lines) {\n        let trust: LabelTrust = 'unverified'\n        try {\n          trust = (JSON.parse(line) as LabeledScenarioRecord).labelTrust ?? 'unverified'\n        } catch {\n          // A malformed line counts as unverified — never silently gold.\n        }\n        byTrust[trust] += 1\n      }\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource, byTrust }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  if (f.minTrust !== undefined && labelTrustRank(record.labelTrust) < labelTrustRank(f.minTrust)) {\n    return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * Pass A substrate types — `runCampaign` is the one primitive every\n * eval flow composes from. Three contracts in this file:\n *\n *   - `Scenario`            input set\n *   - `DispatchFn`          how to run one scenario → artifact\n *   - `CampaignResult`      defined output schema (the contract downstream tools depend on)\n *\n * Three more lifted from earlier substrate work (re-exported):\n *\n *   - `JudgeConfig`         pluggable dimensional scorer (0.38)\n *   - `Mutator`             optimization-loop surface mutator\n *   - `Gate`                promotion gate (`HeldOutGate` and friends adapt to this)\n *\n * No new architecture vs 0.38 — Pass A formalizes the shapes so consumers\n * can build dashboards / CI gates / regression diffs against a stable schema.\n */\n\n/** @experimental Stable identifier + kind tag for any scenario. Consumers\n *  extend with their per-domain payload (persona, task, requirement, ...). */\nexport interface Scenario {\n  id: string\n  kind: string\n  tags?: string[]\n}\n\n/** @experimental Context handed to every dispatch invocation. Scoped — every\n *  trace/span carries the cellId, every artifact write lands under the cell's\n *  artifact root, the cost meter accumulates per cell. */\nexport interface DispatchContext {\n  cellId: string\n  rep: number\n  generation?: number\n  seed: number\n  signal: AbortSignal\n  trace: CampaignTraceWriter\n  artifacts: CampaignArtifactWriter\n  cost: CampaignCostMeter\n  /** Populated when this run is part of a multi-cycle improvement loop. */\n  cycleId?: string\n  /** Populated when the substrate resumed from a prior cache hit. */\n  resumedFrom?: string\n  /**\n   * Opaque placement key supplied by `RunCampaignOptions.cellPlacement`.\n   * The substrate forwards it through unchanged; placement-aware Dispatch\n   * implementations (e.g. `httpDispatch` from `/adapters/http`) read it to\n   * route the cell to the right worker / region / sandbox. `undefined`\n   * when no placement strategy is configured.\n   */\n  placement?: string\n}\n\n/** @experimental One function: scenario + ctx → artifact. Dispatcher chooses\n *  whether to call `runMultishot`, `runLoop`, raw `streamPrompt`, anything. */\nexport type DispatchFn<TScenario extends Scenario, TArtifact> = (\n  scenario: TScenario,\n  ctx: DispatchContext,\n) => Promise<TArtifact>\n\n// ── Sessions ──────────────────────────────────────────────────────────\n\n/** @experimental One session within a multi-session journey. Dispatch is\n *  invoked once per session in order; state from prior session's artifact\n *  is exposed via `ctx.priorSessionArtifact`. */\nexport interface SessionScript<TScenario, TArtifact> {\n  id: string\n  intent: string\n  maxTurns?: number\n  /** When true, knowledge accumulated this session persists to next. */\n  affectsKnowledge?: boolean\n  /** Optional per-session persona evolution — called after the session\n   *  resolves. Returns the persona shape used by the NEXT session. */\n  evolveAfterSession?: (artifact: TArtifact, sessionIndex: number, scenario: TScenario) => TScenario\n}\n\n// ── Judges (re-export 0.38 shape) ─────────────────────────────────────\n\nexport interface JudgeDimension {\n  /** JSON field name + score key. */\n  key: string\n  /** Description shown in the judge's user prompt. */\n  description: string\n}\n\n/** @experimental Pluggable dimensional scorer. `score` is the contract:\n *  given an artifact + scenario, return a `JudgeScore`. This is deliberately a\n *  function, not a fixed LLM-prompt shape — real consumers judge with\n *  ensembles, deterministic checks, or a single LLM call, and the substrate\n *  must not constrain that. The `llmJudge()` helper builds a `score` that does\n *  one LLM call for the common case. `appliesTo` lets a judge run only on\n *  scenarios that match (e.g. a legal-citation judge only on legal scenarios). */\nexport interface JudgeConfig<TArtifact, TScenario extends Scenario = Scenario> {\n  name: string\n  dimensions: JudgeDimension[]\n  /** Score one artifact. Throw on failure — a thrown judge is recorded as a\n   *  failed cell, never silently folded into a zero. */\n  score(input: {\n    artifact: TArtifact\n    scenario: TScenario\n    signal: AbortSignal\n  }): JudgeScore | Promise<JudgeScore>\n  appliesTo?: (scenario: TScenario) => boolean\n}\n\nexport interface JudgeScore {\n  dimensions: Record<string, number>\n  composite: number\n  notes: string\n}\n\n// ── Optimization (population + generations + mutator) ─────────────────\n\n/** @experimental A tier-4 code surface — a candidate change to the agent's\n *  IMPLEMENTATION, not its prompt. Produced by autoresearch (reads codebase +\n *  trace findings → opens a worktree). Measured by checking out `worktreeRef`\n *  and running the worker against the changed code. See the improvement-tier\n *  table in `docs/design/loop-taxonomy.md`. */\nexport interface CodeSurface {\n  kind: 'code'\n  /** Worktree path or git ref holding the candidate code change. The\n   *  consumer's `dispatchWithSurface` checks this out before running. */\n  worktreeRef: string\n  /** Base ref the change is measured against. Default: the repo's main. */\n  baseRef?: string\n  /** Human summary of what changed — rendered into the auto-PR body. */\n  summary?: string\n}\n\n/** @experimental The mutable surface a driver proposes. Tiers (see\n *  `docs/design/loop-taxonomy.md`):\n *   - `string`      — tiers 1-2: system-prompt addendum / serialized tool\n *                     config. Cheap, reversible, text-diffable.\n *   - `CodeSurface` — tier 4: an implementation change behind a worktree ref.\n *  Tier 3 (knowledge) is owned by agent-knowledge and rides its own adapter,\n *  not this type. */\nexport type MutableSurface = string | CodeSurface\n\n/** @experimental Stateless surface mutation — given findings + current\n *  surface, return N candidate surfaces. Pure transform, no generation\n *  awareness. Reflective-mutation, `runMultiShotOptimization`, `AxGEPA`\n *  conform. Wrapped by `evolutionaryDriver` to become an `ImprovementDriver`. */\nexport interface Mutator<TFindings = unknown> {\n  kind: string\n  mutate(args: {\n    findings: TFindings[]\n    currentSurface: MutableSurface\n    populationSize: number\n    signal: AbortSignal\n  }): Promise<MutableSurface[]>\n}\n\n/** @experimental Everything a driver's `propose()` may read to plan the next\n *  batch of candidates. The first six fields are always present; the rest are\n *  optional context the loop supplies when available, so cheap drivers\n *  (`evolutionaryDriver`) can ignore them while a code-tier agentic generator\n *  consumes the research report + dataset to drive a coding harness.\n *  See `docs/design/self-improvement-engine.md`. */\nexport interface ProposeContext<TFindings = unknown> {\n  currentSurface: MutableSurface\n  history: GenerationRecord[]\n  findings: TFindings[]\n  /** BREADTH: how many candidate surfaces to return this generation. */\n  populationSize: number\n  generation: number\n  signal: AbortSignal\n  /** The Phase-2 research report (analyst findings + diff), produced AFTER the\n   *  trace analysts run. Opaque to the substrate — the driver that consumes it\n   *  types it. See the phase diagram in self-improvement-engine.md. */\n  report?: unknown\n  /** Handle to all captured data — the driver samples traces / artifacts /\n   *  rewards here to ground its proposals. */\n  dataset?: LabeledScenarioStore\n  /** DEPTH: max iterations the agentic generator may take per candidate.\n   *  1 = single-shot; >1 = it may iterate on its own change before handing it\n   *  back to be measured. */\n  maxImprovementShots?: number\n}\n\n/** @experimental A surface-improvement strategy — the DRIVER of the\n *  improvement loop. Given the current best surface, the history of what's\n *  been tried + scored, and any external findings, propose the next batch of\n *  candidate surfaces to measure. Optionally decide to stop early.\n *\n *  The evolutionary mutator (`evolutionaryDriver`, here) and agent-runtime's\n *  `improvementDriver` (with reflective / agentic generators) both conform —\n *  drivers of the SAME loop, not separate loops. The loop body\n *  (`runOptimization`) and the gated promotion shell (`runImprovementLoop`)\n *  are driver-agnostic. */\nexport interface ImprovementDriver<TFindings = unknown> {\n  kind: string\n  /** Plan: propose N candidate surfaces for the next generation. */\n  propose(ctx: ProposeContext<TFindings>): Promise<MutableSurface[]>\n  /** Decide: stop early when the driver judges the search converged or\n   *  exhausted. Default (omitted) runs all `maxGenerations`. */\n  decide?(args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n}\n\nexport interface OptimizerConfig {\n  driver: ImprovementDriver\n  populationSize: number\n  maxGenerations: number\n  surfaceExtractor: (profile: unknown) => MutableSurface\n}\n\n// ── Gates ─────────────────────────────────────────────────────────────\n\n/** @experimental Five-valued verdict taxonomy (MOSS-paper alignment). */\nexport type GateDecision = 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling'\n\nexport interface GateContext<TArtifact, TScenario extends Scenario> {\n  candidateArtifacts: Map<string, TArtifact>\n  baselineArtifacts?: Map<string, TArtifact>\n  /** Candidate (winner) judge scores, keyed by cellId. */\n  judgeScores: Map<string, Record<string, JudgeScore>>\n  /** Baseline judge scores, keyed by cellId. SEPARATE from `judgeScores` —\n   *  baseline + candidate share cellIds (same scenarios), so a single map\n   *  cannot represent both. A gate computing a holdout delta MUST read\n   *  candidate from `judgeScores` and baseline from here. */\n  baselineJudgeScores?: Map<string, Record<string, JudgeScore>>\n  scenarios: TScenario[]\n  cost: { candidate: number; baseline: number }\n  signal: AbortSignal\n}\n\nexport interface GateResult {\n  decision: GateDecision\n  reasons: string[]\n  contributingGates: Array<{ name: string; passed: boolean; detail: unknown }>\n  delta?: number\n}\n\n/** @experimental Composable promotion gate. */\nexport interface Gate<TArtifact = unknown, TScenario extends Scenario = Scenario> {\n  name: string\n  decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult>\n}\n\n// ── Tracing / artifacts / cost ────────────────────────────────────────\n\n/** @experimental Scoped trace writer handed to each dispatch — every span\n *  auto-tagged with the cellId so traces filter cleanly. */\nexport interface CampaignTraceWriter {\n  span(name: string, attributes?: Record<string, unknown>): TraceSpan\n  flush(): Promise<void>\n}\n\nexport interface TraceSpan {\n  end(attributes?: Record<string, unknown>): void\n  setAttribute(key: string, value: unknown): void\n}\n\n/** @experimental Scoped artifact writer — `write(path, content)` lands under\n *  `<runDir>/<cellId>/<path>`. */\nexport interface CampaignArtifactWriter {\n  write(path: string, content: string | Uint8Array): Promise<string>\n  writeJson(path: string, value: unknown): Promise<string>\n}\n\n/** @experimental Cell-scoped cost meter. Substrate auto-tracks LLM costs\n *  via the cost-ledger backend hooks; consumers can record additional\n *  spend (sandbox time, tool costs) via `observe`. */\nexport interface CampaignCostMeter {\n  observe(amountUsd: number, source: string): void\n  current(): number\n}\n\n// ── LabeledScenarioStore ──────────────────────────────────────────────\n\n/** @experimental Source tag — required on every store write. Used by the\n *  default training-source filter (production-trace samples NOT used as\n *  training scenarios unless explicitly opted in). */\nexport type LabeledScenarioSource =\n  | 'production-trace'\n  | 'eval-run'\n  | 'manual'\n  | 'red-team'\n  | 'synthetic'\n\nexport type RedactionStatus = 'raw' | 'redacted-pii' | 'redacted-secrets' | 'fully-redacted'\n\n/** How much a label can be trusted to evaluate against — the gold-admission\n *  gate. Strictly ordered: a record qualifies for a `minTrust` filter when its\n *  trust rank is >= the requested rank.\n *\n *   - `unverified`      — label is a heuristic (e.g. raw outcome success/fail).\n *                          Fine as corpus; MUST NOT enter a gold set that lift\n *                          numbers are computed against.\n *   - `verified-signal` — an external signal confirmed the outcome (PR merged,\n *                          tests green, user did not retry, downstream check).\n *   - `human-rated`     — a human explicitly rated or corrected the artifact.\n *\n *  Absent on a write ⇒ treated as `unverified` (fail-closed: a writer must\n *  explicitly assert trust to make a record gold-eligible — it never happens\n *  by accident). */\nexport type LabelTrust = 'unverified' | 'verified-signal' | 'human-rated'\n\nconst LABEL_TRUST_RANK: Record<LabelTrust, number> = {\n  unverified: 0,\n  'verified-signal': 1,\n  'human-rated': 2,\n}\n\n/** Ordinal rank for a label-trust tier; absent ⇒ `unverified` (rank 0). */\nexport function labelTrustRank(trust: LabelTrust | undefined): number {\n  return LABEL_TRUST_RANK[trust ?? 'unverified']\n}\n\n/** @experimental Required-provenance write. The store rejects writes that\n *  lack provenance — a default-on flywheel without provenance is the\n *  data-poisoning vector flagged in the alignment review. */\nexport interface LabeledScenarioWrite<TScenario extends Scenario = Scenario, TArtifact = unknown> {\n  scenario: TScenario\n  artifact: TArtifact\n  judgeScores: Record<string, JudgeScore>\n  source: LabeledScenarioSource\n  sourceVersionHash: string\n  capturedAt: string\n  redactionStatus: RedactionStatus\n  /** Gold-admission trust tier. Absent ⇒ `unverified` (fail-closed): the\n   *  record is corpus, never gold. A writer must explicitly assert\n   *  `verified-signal` or `human-rated` to make it eligible for a gold\n   *  sample. See {@link LabelTrust}. */\n  labelTrust?: LabelTrust\n  /** Optional per-source rate-limit bucket key (e.g., the tenant id). */\n  rateLimitBucket?: string\n}\n\nexport interface LabeledScenarioRecord<TScenario extends Scenario = Scenario, TArtifact = unknown>\n  extends LabeledScenarioWrite<TScenario, TArtifact> {\n  /** Stable hash of (scenario.id, source, capturedAt, sourceVersionHash). */\n  recordHash: string\n  /** Substrate-assigned split — train if captured before the campaign's\n   *  `temporalCutoff`, test if after. Explicit override allowed via filter. */\n  split: 'train' | 'test'\n}\n\nexport interface LabeledScenarioSampleArgs {\n  count: number\n  /** REQUIRED — substrate refuses to sample without an explicit split. */\n  split: 'train' | 'test'\n  /** REQUIRED — only records captured before this timestamp are returned.\n   *  Enforces temporal split discipline (test scenarios captured AFTER train\n   *  cannot enter the training pool). */\n  capturedBefore: string\n  filter?: {\n    kind?: string\n    source?: LabeledScenarioSource | LabeledScenarioSource[]\n    minComposite?: number\n    maxComposite?: number\n    /** Gold gate: only records whose trust rank is >= this tier are\n     *  returned. `sample({ split: 'test', minTrust: 'verified-signal' })` is\n     *  the canonical \"give me the gold set\" call. Absent ⇒ no trust gate\n     *  (corpus-level read). */\n    minTrust?: LabelTrust\n  }\n}\n\nexport interface LabeledScenarioStore {\n  observe(write: LabeledScenarioWrite): Promise<void>\n  sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]>\n  size(): Promise<{\n    train: number\n    test: number\n    bySource: Record<string, number>\n    /** Count by trust tier — tells the flywheel how much gold it has\n     *  accumulated vs. raw corpus. */\n    byTrust: Record<LabelTrust, number>\n  }>\n}\n\n// ── The CampaignResult schema (the downstream-tools contract) ─────────\n\nexport interface CampaignCellResult<TArtifact> {\n  cellId: string\n  scenarioId: string\n  rep: number\n  generation?: number\n  artifact: TArtifact\n  judgeScores: Record<string, JudgeScore>\n  costUsd: number\n  durationMs: number\n  seed: number\n  cached: boolean\n  error?: string\n}\n\nexport interface JudgeAggregate {\n  mean: number\n  stdev: number\n  ci95: [number, number]\n  n: number\n}\n\nexport interface ScenarioAggregate {\n  meanComposite: number\n  ci95: [number, number]\n  n: number\n}\n\nexport interface GenerationRecord {\n  generationIndex: number\n  candidates: GenerationCandidate[]\n  promoted: string[]\n}\n\n/** One scored candidate surface in a generation. `dimensions` + `scenarios`\n *  let a reflective `ImprovementDriver` ground its next proposal on WHICH\n *  dimensions the candidate is weakest on and WHICH scenarios it best/worst\n *  handled — the evidence a blind `Mutator` cannot see. */\nexport interface GenerationCandidate {\n  surfaceHash: string\n  composite: number\n  ci95: [number, number]\n  /** Mean score per judge dimension across all cells (scenarios × reps ×\n   *  judges that reported the dimension). */\n  dimensions: Record<string, number>\n  /** Per-scenario composite (mean over reps + judges). */\n  scenarios: Array<{ scenarioId: string; composite: number }>\n}\n\nexport interface CampaignAggregates {\n  byJudge: Record<string, JudgeAggregate>\n  byScenario: Record<string, ScenarioAggregate>\n  totalCostUsd: number\n  cellsExecuted: number\n  cellsSkipped: number\n  cellsCached: number\n  cellsFailed: number\n}\n\nexport interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scenario> {\n  /** sha256(scenarios, judges, dispatch source ref, optimizer config, seed). Stable identity for reruns. */\n  manifestHash: string\n  seed: number\n  startedAt: string\n  endedAt: string\n  durationMs: number\n  cells: Array<CampaignCellResult<TArtifact>>\n  aggregates: CampaignAggregates\n  optimization?: {\n    generations: GenerationRecord[]\n    winnerSurfaceHash?: string\n  }\n  gate?: GateResult\n  prUrl?: string\n  runDir: string\n  artifactsByPath: Record<string, string>\n  /** Substrate strips the input scenarios to id+kind for the result manifest;\n   *  consumers needing full payload look it up via the original input. The\n   *  type parameter `TScenario` is propagated for downstream consumers that\n   *  want narrowed types when extending `CampaignResult`. */\n  scenarios: Array<Pick<TScenario, 'id' | 'kind'>>\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;;;AC2QrB,IAAM,mBAA+C;AAAA,EACnD,YAAY;AAAA,EACZ,mBAAmB;AAAA,EACnB,eAAe;AACjB;AAGO,SAAS,eAAe,OAAuC;AACpE,SAAO,iBAAiB,SAAS,YAAY;AAC/C;;;AD/PO,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAKH;AACD,UAAM,WAAmC,CAAC;AAC1C,UAAM,UAAsC;AAAA,MAC1C,YAAY;AAAA,MACZ,mBAAmB;AAAA,MACnB,eAAe;AAAA,IACjB;AACA,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,eAAS,MAAM,IAAI,MAAM;AACzB,eAAS,MAAM;AACf,iBAAW,QAAQ,OAAO;AACxB,YAAI,QAAoB;AACxB,YAAI;AACF,kBAAS,KAAK,MAAM,IAAI,EAA4B,cAAc;AAAA,QACpE,QAAQ;AAAA,QAER;AACA,gBAAQ,KAAK,KAAK;AAAA,MACpB;AAAA,IACF;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,UAAU,QAAQ;AAAA,EACxD;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,MAAI,EAAE,aAAa,UAAa,eAAe,OAAO,UAAU,IAAI,eAAe,EAAE,QAAQ,GAAG;AAC9F,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AElSA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}

package/dist/{chunk-L7XMNXLO.js → chunk-J4DIMSRK.js}

@@ -4,7 +4,7 @@ import {
 } from "./chunk-NCRFYPS3.js";
 import {
   validateRunRecord
-} from "./chunk-BWZEGTES.js";
+} from "./chunk-NCK5QLGT.js";
 import {
   TraceEmitter
 } from "./chunk-TVVP3ZZQ.js";
@@ -610,4 +610,4 @@ export {
   runProposeReviewAsControlLoop,
   controlFailureClassFromVerification
 };
-//# sourceMappingURL=chunk-L7XMNXLO.js.map
+//# sourceMappingURL=chunk-J4DIMSRK.js.map

package/dist/{chunk-BWZEGTES.js → chunk-NCK5QLGT.js}

@@ -538,4 +538,4 @@ export {
   parseRunRecordSafe,
   roundTripRunRecord
 };
-//# sourceMappingURL=chunk-BWZEGTES.js.map
+//# sourceMappingURL=chunk-NCK5QLGT.js.map

package/dist/chunk-NCK5QLGT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/agent-profile-cell.ts","../src/run-record.ts"],"sourcesContent":["import { ValidationError } from './errors'\nimport { hashJson } from './pre-registration'\n\nexport type AgentProfileCellSchemaVersion = 'agent-profile-cell/v1'\n\nexport type AgentProfileJson =\n  | string\n  | number\n  | boolean\n  | null\n  | AgentProfileJson[]\n  | { [key: string]: AgentProfileJson }\n\nexport type AgentProfileDimensionValue = string | number | boolean | null\n\nexport interface AgentProfileSource {\n  /** Runtime/profile contract being fingerprinted, e.g. `sandbox-agent-profile`. */\n  kind: string\n  /** sha256 over the canonical source profile object. */\n  hash: string\n}\n\nexport interface AgentProfileSourceInput {\n  kind: string\n  /** Precomputed sha256 for callers that already sign their profile artifact. */\n  hash?: string\n  /** Full canonical runtime profile; hashed and then discarded from the cell. */\n  profile?: AgentProfileJson\n}\n\nexport interface AgentProfileHarness {\n  id: string\n  version?: string\n  hash?: string\n}\n\nexport interface AgentProfileCellInput {\n  profileId: string\n  sourceProfile: AgentProfileSourceInput\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport interface AgentProfileCell {\n  schemaVersion: AgentProfileCellSchemaVersion\n  cellId: string\n  profileId: string\n  sourceProfile: AgentProfileSource\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport class AgentProfileCellValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\nconst SHA256_HEX = /^[0-9a-f]{64}$/\nconst CELL_ID = /^agent-profile-cell:sha256:[0-9a-f]{64}$/\n\nexport async function buildAgentProfileCell(\n  input: AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  const material = await normalizeAgentProfileCellInput(input)\n  const cellId = `agent-profile-cell:sha256:${await hashJson(material)}`\n  return { ...material, cellId }\n}\n\nexport function agentProfileCellHashMaterial(\n  cell: AgentProfileCell,\n): Omit<AgentProfileCell, 'cellId'> {\n  const { cellId: _cellId, ...material } = cell\n  void _cellId\n  return normalizeAgentProfileCell(material)\n}\n\nexport async function verifyAgentProfileCell(cell: AgentProfileCell): Promise<boolean> {\n  validateAgentProfileCell(cell)\n  return (\n    cell.cellId ===\n    `agent-profile-cell:sha256:${await hashJson(agentProfileCellHashMaterial(cell))}`\n  )\n}\n\nexport function validateAgentProfileCell(input: unknown): AgentProfileCell {\n  if (input === null || typeof input !== 'object') {\n    throw new AgentProfileCellValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n  expectLiteral(obj.schemaVersion, 'agent-profile-cell/v1', 'schemaVersion')\n  if (typeof obj.cellId !== 'string' || !CELL_ID.test(obj.cellId)) {\n    throw new AgentProfileCellValidationError(\n      'cellId must match agent-profile-cell:sha256:<64 lowercase hex chars>',\n      'cellId',\n    )\n  }\n  expectString(obj.profileId, 'profileId')\n  validateSource(obj.sourceProfile, 'sourceProfile')\n  if (obj.harness !== undefined) validateHarness(obj.harness, 'harness')\n  if (obj.model !== undefined) expectString(obj.model, 'model')\n  if (obj.promptHash !== undefined) expectString(obj.promptHash, 'promptHash')\n  if (obj.dimensions !== undefined) validateDimensions(obj.dimensions, 'dimensions')\n  return input as AgentProfileCell\n}\n\nexport function requireAgentProfileCell(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): AgentProfileCell {\n  if (!record.agentProfile) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" is missing agentProfile; profile-cell grouping requires explicit profile identity`,\n      'agentProfile',\n    )\n  }\n  return validateAgentProfileCell(record.agentProfile)\n}\n\nexport function agentProfileCellKey(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): string {\n  return requireAgentProfileCell(record).cellId\n}\n\nexport async function assertRunAgentProfileCell(record: {\n  runId: string\n  model: string\n  promptHash: string\n  agentProfile?: AgentProfileCell\n}): Promise<AgentProfileCell> {\n  const profile = requireAgentProfileCell(record)\n  if (!(await verifyAgentProfileCell(profile))) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" has an agentProfile.cellId that does not match its content`,\n      'agentProfile.cellId',\n    )\n  }\n  if (profile.model !== undefined && profile.model !== record.model) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.model \"${profile.model}\" does not match model \"${record.model}\"`,\n      'agentProfile.model',\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== record.promptHash) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${record.promptHash}\"`,\n      'agentProfile.promptHash',\n    )\n  }\n  return profile\n}\n\nexport function groupRunsByAgentProfileCell<\n  T extends { runId: string; agentProfile?: AgentProfileCell },\n>(records: readonly T[]): Map<string, T[]> {\n  const groups = new Map<string, T[]>()\n  for (const record of records) {\n    const key = agentProfileCellKey(record)\n    const bucket = groups.get(key)\n    if (bucket) bucket.push(record)\n    else groups.set(key, [record])\n  }\n  return groups\n}\n\nasync function normalizeAgentProfileCellInput(\n  input: AgentProfileCellInput,\n): Promise<Omit<AgentProfileCell, 'cellId'>> {\n  return normalizeAgentProfileCell({\n    schemaVersion: 'agent-profile-cell/v1',\n    profileId: input.profileId,\n    sourceProfile: await normalizeSourceInput(input.sourceProfile),\n    harness: input.harness,\n    model: input.model,\n    promptHash: input.promptHash,\n    dimensions: input.dimensions,\n  })\n}\n\nfunction normalizeAgentProfileCell(\n  input: Omit<AgentProfileCell, 'cellId'>,\n): Omit<AgentProfileCell, 'cellId'> {\n  return compactObject({\n    schemaVersion: 'agent-profile-cell/v1' as const,\n    profileId: requireNonEmpty(input.profileId, 'profileId'),\n    sourceProfile: normalizeSource(input.sourceProfile),\n    harness: input.harness ? normalizeHarness(input.harness, 'harness') : undefined,\n    model: optionalNonEmpty(input.model, 'model'),\n    promptHash: optionalNonEmpty(input.promptHash, 'promptHash'),\n    dimensions: input.dimensions\n      ? nonEmptyRecord(normalizeDimensions(input.dimensions))\n      : undefined,\n  })\n}\n\nasync function normalizeSourceInput(input: AgentProfileSourceInput): Promise<AgentProfileSource> {\n  const kind = requireNonEmpty(input.kind, 'sourceProfile.kind')\n  if (input.hash !== undefined && input.profile !== undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide either hash or profile, not both',\n      'sourceProfile',\n    )\n  }\n  if (input.hash !== undefined) {\n    return { kind, hash: requireSha256Hex(input.hash, 'sourceProfile.hash') }\n  }\n  if (input.profile === undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide hash or profile',\n      'sourceProfile',\n    )\n  }\n  assertJson(input.profile, 'sourceProfile.profile')\n  return { kind, hash: await hashJson(input.profile) }\n}\n\nfunction normalizeSource(input: AgentProfileSource): AgentProfileSource {\n  return {\n    kind: requireNonEmpty(input.kind, 'sourceProfile.kind'),\n    hash: requireSha256Hex(input.hash, 'sourceProfile.hash'),\n  }\n}\n\nfunction normalizeHarness(input: AgentProfileHarness, path: string): AgentProfileHarness {\n  return compactObject({\n    id: requireNonEmpty(input.id, `${path}.id`),\n    version: optionalNonEmpty(input.version, `${path}.version`),\n    hash: optionalNonEmpty(input.hash, `${path}.hash`),\n  })\n}\n\nfunction normalizeDimensions(\n  input: Record<string, AgentProfileDimensionValue>,\n): Record<string, AgentProfileDimensionValue> {\n  const out: Record<string, AgentProfileDimensionValue> = {}\n  for (const key of Object.keys(input).sort()) {\n    const value = input[key]\n    requireNonEmpty(key, 'dimensions.<key>')\n    if (\n      value !== null &&\n      typeof value !== 'string' &&\n      typeof value !== 'number' &&\n      typeof value !== 'boolean'\n    ) {\n      throw new AgentProfileCellValidationError(\n        'expected primitive dimension value',\n        `dimensions.${key}`,\n      )\n    }\n    if (typeof value === 'number' && !Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', `dimensions.${key}`)\n    }\n    out[key] = value\n  }\n  return out\n}\n\nfunction compactObject<T extends Record<string, unknown>>(input: T): T {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as T\n}\n\nfunction nonEmptyRecord<T extends Record<string, unknown>>(input: T): T | undefined {\n  return Object.keys(input).length > 0 ? input : undefined\n}\n\nfunction validateSource(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.kind, `${path}.kind`)\n  requireSha256Hex(rec.hash, `${path}.hash`)\n}\n\nfunction validateHarness(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.id, `${path}.id`)\n  if (rec.version !== undefined) expectString(rec.version, `${path}.version`)\n  if (rec.hash !== undefined) expectString(rec.hash, `${path}.hash`)\n}\n\nfunction validateDimensions(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  normalizeDimensions(value as Record<string, AgentProfileDimensionValue>)\n}\n\nfunction assertJson(value: AgentProfileJson, path: string): void {\n  if (value === null) return\n  const type = typeof value\n  if (type === 'string' || type === 'boolean') return\n  if (type === 'number') {\n    if (!Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', path)\n    }\n    return\n  }\n  if (Array.isArray(value)) {\n    value.forEach((item, index) => {\n      assertJson(item, `${path}[${index}]`)\n    })\n    return\n  }\n  if (type === 'object') {\n    for (const [key, nested] of Object.entries(value)) {\n      requireNonEmpty(key, `${path}.<key>`)\n      assertJson(nested, `${path}.${key}`)\n    }\n    return\n  }\n  throw new AgentProfileCellValidationError('expected JSON-compatible value', path)\n}\n\nfunction expectLiteral(value: unknown, expected: string, path: string): void {\n  if (value !== expected) {\n    throw new AgentProfileCellValidationError(`expected ${expected}`, path)\n  }\n}\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n}\n\nfunction requireNonEmpty(value: string, path: string): string {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n  return value\n}\n\nfunction optionalNonEmpty(value: string | undefined, path: string): string | undefined {\n  if (value === undefined) return undefined\n  return requireNonEmpty(value, path)\n}\n\nfunction requireSha256Hex(value: unknown, path: string): string {\n  if (typeof value !== 'string' || !SHA256_HEX.test(value)) {\n    throw new AgentProfileCellValidationError('expected 64 lowercase sha256 hex chars', path)\n  }\n  return value\n}\n\n// ── Consumer helpers ─────────────────────────────────────────────────\n//\n// Two pieces of boilerplate every product consuming `buildAgentProfileCell`\n// has been duplicating (gtm-agent #137, blueprint-agent #1756/#1757):\n//\n//   1. A `JSON.parse(JSON.stringify(value))` helper that canonicalizes an\n//      arbitrary sandbox-SDK `AgentProfile` into the recursive\n//      `AgentProfileJson` shape, with a fail-loud error when the profile\n//      is not JSON-serializable.\n//\n//   2. The magic string `'sandbox-agent-profile'` for `sourceProfile.kind`.\n//\n// Both belong here so the cross-product cell join (same canonical profile\n// hashes to the same `sourceProfile.hash` across products) is enforced by\n// the type system, not by every consumer remembering to do it right.\n// See blueprint-agent issue tangle-network/agent-eval#82.\n\n/** Canonical `sourceProfile.kind` values. Two products fingerprinting the\n *  same canonical profile MUST use the same kind for their cells to share\n *  `sourceProfile.hash`. Extend rather than create new strings — adding a\n *  new kind is a deliberate cross-product schema change. */\nexport const AGENT_PROFILE_KINDS = {\n  /** A profile declared via `defineAgentProfile(...)` from\n   *  `@tangle-network/sandbox`. The default kind for sandbox-hosted\n   *  products (gtm-agent, blueprint-agent, sandbox, evals). */\n  SANDBOX_AGENT_PROFILE: 'sandbox-agent-profile',\n} as const\n\nexport type AgentProfileKind = (typeof AGENT_PROFILE_KINDS)[keyof typeof AGENT_PROFILE_KINDS]\n\n/** Canonicalize an arbitrary value into `AgentProfileJson` by JSON\n *  round-trip. Throws when the value contains anything not representable\n *  as JSON (functions, BigInt, cycles) — non-portable profiles fail loud\n *  rather than silently dropping fields. */\nexport function toAgentProfileJson(value: unknown): AgentProfileJson {\n  let serialized: string | undefined\n  try {\n    serialized = JSON.stringify(value)\n  } catch (err) {\n    throw new AgentProfileCellValidationError(\n      `agent profile must be JSON-serializable: ${err instanceof Error ? err.message : String(err)}`,\n      'sourceProfile.profile',\n    )\n  }\n  if (serialized === undefined) {\n    throw new AgentProfileCellValidationError(\n      'agent profile must be JSON-serializable (got undefined after JSON.stringify)',\n      'sourceProfile.profile',\n    )\n  }\n  return JSON.parse(serialized) as AgentProfileJson\n}\n\n/** Minimal shape required of any sandbox-SDK `AgentProfile` — anything\n *  with a non-empty `name` and `version` plus JSON-serializable contents.\n *  Compatible with `defineAgentProfile(...)` output from\n *  `@tangle-network/sandbox`; products that have not yet declared a real\n *  profile can pass a `{ name, version, ...metadata }` stub. */\nexport interface SandboxAgentProfileLike {\n  name: string\n  version: string\n  [key: string]: unknown\n}\n\n/** Higher-level helper that hard-codes the canonical\n *  `sandbox-agent-profile` kind plus the JSON canonicalization. Equivalent\n *  to calling `buildAgentProfileCell` with `profileId = \\`${name}@${version}\\``\n *  and `sourceProfile = { kind: SANDBOX_AGENT_PROFILE, profile: <round-tripped> }`.\n *\n *  Use this from any product consuming a sandbox-SDK `AgentProfile`; the\n *  manual `buildAgentProfileCell` call is reserved for advanced cases\n *  (custom kinds, pre-computed source hashes, alternate profileId\n *  conventions). */\nexport async function buildSandboxAgentProfileCell(\n  profile: SandboxAgentProfileLike,\n  input: Omit<AgentProfileCellInput, 'profileId' | 'sourceProfile'>,\n): Promise<AgentProfileCell> {\n  if (!profile || typeof profile !== 'object') {\n    throw new AgentProfileCellValidationError('sandbox AgentProfile must be an object', 'profile')\n  }\n  if (typeof profile.name !== 'string' || profile.name.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'sandbox AgentProfile must have a non-empty `name`',\n      'profile.name',\n    )\n  }\n  if (typeof profile.version !== 'string' || profile.version.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'sandbox AgentProfile must have a non-empty `version`',\n      'profile.version',\n    )\n  }\n  return buildAgentProfileCell({\n    ...input,\n    profileId: `${profile.name}@${profile.version}`,\n    sourceProfile: {\n      kind: AGENT_PROFILE_KINDS.SANDBOX_AGENT_PROFILE,\n      profile: toAgentProfileJson(profile),\n    },\n  })\n}\n","/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag and either a `searchScore` or a `holdoutScore`.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n *   const rec = validateRunRecord(rawJson)         // throws on missing\n *   const ok  = isRunRecord(rawJson)               // boolean check\n *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport { ValidationError } from './errors'\nimport type { FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n *  combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\nexport interface RunTokenUsage {\n  input: number\n  output: number\n  cached?: number\n}\n\nexport interface RunJudgeMetadata {\n  model: string\n  promptVersion: string\n  /** [0,1] confidence the judge declared. Constant judge confidence\n   *  across many runs is a fallback signal (see `canary.ts`). */\n  confidence: number\n  /** True if the judge degraded to a fallback path (rules-only,\n   *  prior-call cache, etc.). The canary uses this to alert. */\n  fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n  /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n  perJudge: Record<string, Record<string, number>>\n  /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n  perDimMean: Record<string, number>\n  /** Composite mean across all dims and judges. Mirrors the score\n   *  the gate sees on `outcome.searchScore` / `holdoutScore`. */\n  composite: number\n  /** Judges that errored or returned an unparseable verdict. Recorded\n   *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n   *  not inferred from missing keys in `perJudge`. */\n  failedJudges?: string[]\n  /** Free-form notes the judges emitted (joined across judges or\n   *  first-judge only — consumer's choice). */\n  notes?: string\n}\n\nexport interface RunOutcome {\n  /** Score on the search/optimization split. Optional because a\n   *  holdout-only evaluation only fills `holdoutScore`. */\n  searchScore?: number\n  /** Score on the held-out split. Optional because a search-only run\n   *  only fills `searchScore`. At least one must be present. */\n  holdoutScore?: number\n  /** Bag of any other metric the run produced — judge dimensions,\n   *  pass/fail counters, latency stats, etc. Numeric only — keeps\n   *  reporters honest. */\n  raw: Record<string, number>\n  /** Per-judge / per-dim breakdown. Consumers writing ensemble\n   *  judgements populate this; substrate primitives like\n   *  `interRaterReliability` and `corpusInterRaterAgreement` accept\n   *  these records as input. Optional — single-judge or scalar-only\n   *  runs leave it unset. */\n  judgeScores?: JudgeScoresRecord\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n *     model (after any steering bundle merge).\n *   - `configHash` is the sha256 of the effective run config (model,\n *     temperature, tools, judges, splits). The pair (promptHash,\n *     configHash) uniquely identifies an experimental cell.\n *\n * Model snapshot discipline:\n *   - `model` MUST encode a snapshot version. Bare aliases like\n *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n */\nexport interface RunRecord {\n  /** UUID for the run. */\n  runId: string\n  /** Logical experiment grouping (a treatment vs a baseline within\n   *  the same sweep should share `experimentId`). */\n  experimentId: string\n  /** Stable identifier for the candidate (variant) being run. The\n   *  promotion gate compares two `candidateId`s on matched items. */\n  candidateId: string\n  /** RNG seed for the run. Always recorded — silent re-seeding is\n   *  the most common cause of non-reproducible numbers. */\n  seed: number\n  /** Model identifier WITH snapshot version. */\n  model: string\n  /** sha256 of the effective prompt (post-steering). */\n  promptHash: string\n  /** sha256 of the effective config. */\n  configHash: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** End-to-end wall-clock duration in milliseconds. */\n  wallMs: number\n  /** Time spent queued before execution started, if known. */\n  queueMs?: number\n  /** Total USD cost. Mandatory — runs without a cost number are\n   *  unbounded by definition and must not be admitted into the gate. */\n  costUsd: number\n  /** Token usage breakdown. */\n  tokenUsage: RunTokenUsage\n  /** Judge-side metadata, if a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /** Per-split scores + raw bag. */\n  outcome: RunOutcome\n  /** Canonical, cross-agent failure class drawn from the shared\n   *  `FAILURE_CLASSES` taxonomy. This is the aggregation key that makes\n   *  \"which failure dominates across the whole fleet\" answerable in ONE\n   *  vocabulary — every agent classifies against the same enum. Producers\n   *  set it via the substrate classifier; leave unset only when the failure\n   *  genuinely can't be classified. */\n  failureClass?: FailureClass\n  /** Free-form domain-specific failure detail, scoped UNDER `failureClass`\n   *  (e.g. failureClass='tool_recovery_failure', failureMode='forge_build_unsatisfied').\n   *  The within-agent drill-down; `failureClass` is the cross-agent key. */\n  failureMode?: string\n  /** Which split this run was drawn from. */\n  splitTag: RunSplitTag\n  /**\n   * Stable scenario identifier the run was scored against. Optional for\n   * backwards compatibility, but **strongly recommended**: every primitive\n   * that pairs runs by scenario (preferences, paired stats, BT tournament)\n   * keys on this. The campaign artifact populates it canonically; legacy\n   * runs without it fall back to inference from `outcome.raw.scenario_id`\n   * or `experimentId`.\n   */\n  scenarioId?: string\n  /**\n   * Canonical identity for the agent profile cell that produced this row:\n   * profile artifact hash plus optional harness/model/prompt/reporting\n   * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n   * longitudinal reports by the complete source profile, not by a loose\n   * candidate label or opaque config hash.\n   */\n  agentProfile?: AgentProfileCell\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n  'runId',\n  'experimentId',\n  'candidateId',\n  'seed',\n  'model',\n  'promptHash',\n  'configHash',\n  'commitSha',\n  'wallMs',\n  'costUsd',\n  'tokenUsage',\n  'outcome',\n  'splitTag',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\n\nexport class RunRecordValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n\n  for (const key of MANDATORY_TOP_LEVEL) {\n    if (!(key in obj)) {\n      throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n    }\n  }\n\n  expectString(obj.runId, 'runId')\n  expectString(obj.experimentId, 'experimentId')\n  expectString(obj.candidateId, 'candidateId')\n  expectFiniteNumber(obj.seed, 'seed')\n  expectString(obj.model, 'model')\n  expectString(obj.promptHash, 'promptHash')\n  expectString(obj.configHash, 'configHash')\n  expectString(obj.commitSha, 'commitSha')\n  expectFiniteNumber(obj.wallMs, 'wallMs')\n  if (obj.queueMs !== undefined) expectFiniteNumber(obj.queueMs, 'queueMs')\n  expectFiniteNumber(obj.costUsd, 'costUsd')\n\n  // Snapshot discipline: bare model aliases are not paper-grade.\n  if (!modelHasSnapshot(obj.model as string)) {\n    throw new RunRecordValidationError(\n      `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,\n      'model',\n    )\n  }\n\n  // Token usage.\n  const tu = obj.tokenUsage\n  if (tu === null || typeof tu !== 'object') {\n    throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n  }\n  const tuRec = tu as Record<string, unknown>\n  expectFiniteNumber(tuRec.input, 'tokenUsage.input')\n  expectFiniteNumber(tuRec.output, 'tokenUsage.output')\n  if (tuRec.cached !== undefined) expectFiniteNumber(tuRec.cached, 'tokenUsage.cached')\n\n  // Judge metadata, optional.\n  if (obj.judgeMetadata !== undefined) {\n    const jm = obj.judgeMetadata\n    if (jm === null || typeof jm !== 'object') {\n      throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n    }\n    const jmRec = jm as Record<string, unknown>\n    expectString(jmRec.model, 'judgeMetadata.model')\n    expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n    expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n    if (typeof jmRec.fallback !== 'boolean') {\n      throw new RunRecordValidationError(\n        'judgeMetadata.fallback must be boolean',\n        'judgeMetadata.fallback',\n      )\n    }\n  }\n\n  // Outcome.\n  const out = obj.outcome\n  if (out === null || typeof out !== 'object') {\n    throw new RunRecordValidationError('outcome must be an object', 'outcome')\n  }\n  const outRec = out as Record<string, unknown>\n  if (outRec.searchScore !== undefined)\n    expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n  if (outRec.holdoutScore !== undefined)\n    expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n  if (outRec.searchScore === undefined && outRec.holdoutScore === undefined) {\n    throw new RunRecordValidationError(\n      'outcome must define searchScore or holdoutScore (or both)',\n      'outcome',\n    )\n  }\n  const raw = outRec.raw\n  if (raw === null || typeof raw !== 'object') {\n    throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n  }\n  for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n    expectFiniteNumber(v, `outcome.raw.${k}`)\n  }\n\n  // Per-judge / per-dim breakdown, optional.\n  if (outRec.judgeScores !== undefined) {\n    validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n  }\n\n  // Failure mode optional.\n  if (obj.failureMode !== undefined) expectString(obj.failureMode, 'failureMode')\n\n  if (obj.agentProfile !== undefined) {\n    try {\n      const profile = validateAgentProfileCell(obj.agentProfile)\n      if (profile.model !== undefined && profile.model !== obj.model) {\n        throw new RunRecordValidationError(\n          `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n          'agentProfile.model',\n        )\n      }\n      if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n        throw new RunRecordValidationError(\n          `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n          'agentProfile.promptHash',\n        )\n      }\n    } catch (error) {\n      if (error instanceof RunRecordValidationError) throw error\n      if (error instanceof Error) {\n        throw new RunRecordValidationError(error.message, 'agentProfile')\n      }\n      throw error\n    }\n  }\n\n  // Split tag.\n  if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n    throw new RunRecordValidationError(\n      `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n      'splitTag',\n    )\n  }\n\n  return input as RunRecord\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n  try {\n    validateRunRecord(input)\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n  input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n  try {\n    return { ok: true, value: validateRunRecord(input) }\n  } catch (e) {\n    if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n    throw e\n  }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n  const json = JSON.stringify(record)\n  return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new RunRecordValidationError(`expected non-empty string`, path)\n  }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value)) {\n    throw new RunRecordValidationError(`expected finite number`, path)\n  }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object') {\n    throw new RunRecordValidationError('judgeScores must be an object', path)\n  }\n  const rec = value as Record<string, unknown>\n\n  const perJudge = rec.perJudge\n  if (perJudge === null || typeof perJudge !== 'object') {\n    throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n  }\n  for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n    if (dims === null || typeof dims !== 'object') {\n      throw new RunRecordValidationError(\n        'per-judge entry must be an object of dimension scores',\n        `${path}.perJudge.${judgeId}`,\n      )\n    }\n    for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n    }\n  }\n\n  const perDimMean = rec.perDimMean\n  if (perDimMean === null || typeof perDimMean !== 'object') {\n    throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n  }\n  for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n  }\n\n  expectFiniteNumber(rec.composite, `${path}.composite`)\n\n  if (rec.failedJudges !== undefined) {\n    if (!Array.isArray(rec.failedJudges)) {\n      throw new RunRecordValidationError(\n        'failedJudges must be an array of strings',\n        `${path}.failedJudges`,\n      )\n    }\n    for (let i = 0; i < rec.failedJudges.length; i++) {\n      const id = rec.failedJudges[i]\n      if (typeof id !== 'string' || id.length === 0) {\n        throw new RunRecordValidationError(\n          'failedJudges entry must be a non-empty string',\n          `${path}.failedJudges[${i}]`,\n        )\n      }\n    }\n  }\n\n  if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n    throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n  }\n}\n\n/**\n * Heuristic snapshot check. Accepts:\n *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)\n *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)\n *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)\n *   - explicit `:date-...` Vertex-style tags\n *\n * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap\n * silently as providers ship new snapshots.\n */\nfunction modelHasSnapshot(model: string): boolean {\n  if (model.includes('@')) return true\n  if (/-\\d{8}$/.test(model)) return true\n  if (/-\\d{4}-\\d{2}-\\d{2}$/.test(model)) return true\n  if (/:date-/.test(model)) return true\n  return false\n}\n"],"mappings":";;;;;;;;AAwDO,IAAM,kCAAN,cAA8C,gBAAgB;AAAA,EAC1D;AAAA,EACT,YAAY,SAAiB,OAAO,IAAI;AACtC,UAAM,OAAO,GAAG,OAAO,QAAQ,IAAI,MAAM,OAAO;AAChD,SAAK,OAAO;AAAA,EACd;AACF;AAEA,IAAM,aAAa;AACnB,IAAM,UAAU;AAEhB,eAAsB,sBACpB,OAC2B;AAC3B,QAAM,WAAW,MAAM,+BAA+B,KAAK;AAC3D,QAAM,SAAS,6BAA6B,MAAM,SAAS,QAAQ,CAAC;AACpE,SAAO,EAAE,GAAG,UAAU,OAAO;AAC/B;AAEO,SAAS,6BACd,MACkC;AAClC,QAAM,EAAE,QAAQ,SAAS,GAAG,SAAS,IAAI;AACzC,OAAK;AACL,SAAO,0BAA0B,QAAQ;AAC3C;AAEA,eAAsB,uBAAuB,MAA0C;AACrF,2BAAyB,IAAI;AAC7B,SACE,KAAK,WACL,6BAA6B,MAAM,SAAS,6BAA6B,IAAI,CAAC,CAAC;AAEnF;AAEO,SAAS,yBAAyB,OAAkC;AACzE,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,gCAAgC,iBAAiB;AAAA,EAC7D;AACA,QAAM,MAAM;AACZ,gBAAc,IAAI,eAAe,yBAAyB,eAAe;AACzE,MAAI,OAAO,IAAI,WAAW,YAAY,CAAC,QAAQ,KAAK,IAAI,MAAM,GAAG;AAC/D,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,eAAa,IAAI,WAAW,WAAW;AACvC,iBAAe,IAAI,eAAe,eAAe;AACjD,MAAI,IAAI,YAAY,OAAW,iBAAgB,IAAI,SAAS,SAAS;AACrE,MAAI,IAAI,UAAU,OAAW,cAAa,IAAI,OAAO,OAAO;AAC5D,MAAI,IAAI,eAAe,OAAW,cAAa,IAAI,YAAY,YAAY;AAC3E,MAAI,IAAI,eAAe,OAAW,oBAAmB,IAAI,YAAY,YAAY;AACjF,SAAO;AACT;AAEO,SAAS,wBAAwB,QAGnB;AACnB,MAAI,CAAC,OAAO,cAAc;AACxB,UAAM,IAAI;AAAA,MACR,QAAQ,OAAO,KAAK;AAAA,MACpB;AAAA,IACF;AAAA,EACF;AACA,SAAO,yBAAyB,OAAO,YAAY;AACrD;AAEO,SAAS,oBAAoB,QAGzB;AACT,SAAO,wBAAwB,MAAM,EAAE;AACzC;AAEA,eAAsB,0BAA0B,QAKlB;AAC5B,QAAM,UAAU,wBAAwB,MAAM;AAC9C,MAAI,CAAE,MAAM,uBAAuB,OAAO,GAAI;AAC5C,UAAM,IAAI;AAAA,MACR,QAAQ,OAAO,KAAK;AAAA,MACpB;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,UAAU,UAAa,QAAQ,UAAU,OAAO,OAAO;AACjE,UAAM,IAAI;AAAA,MACR,QAAQ,OAAO,KAAK,yBAAyB,QAAQ,KAAK,2BAA2B,OAAO,KAAK;AAAA,MACjG;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,eAAe,UAAa,QAAQ,eAAe,OAAO,YAAY;AAChF,UAAM,IAAI;AAAA,MACR,QAAQ,OAAO,KAAK,8BAA8B,QAAQ,UAAU,gCAAgC,OAAO,UAAU;AAAA,MACrH;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,4BAEd,SAAyC;AACzC,QAAM,SAAS,oBAAI,IAAiB;AACpC,aAAW,UAAU,SAAS;AAC5B,UAAM,MAAM,oBAAoB,MAAM;AACtC,UAAM,SAAS,OAAO,IAAI,GAAG;AAC7B,QAAI,OAAQ,QAAO,KAAK,MAAM;AAAA,QACzB,QAAO,IAAI,KAAK,CAAC,MAAM,CAAC;AAAA,EAC/B;AACA,SAAO;AACT;AAEA,eAAe,+BACb,OAC2C;AAC3C,SAAO,0BAA0B;AAAA,IAC/B,eAAe;AAAA,IACf,WAAW,MAAM;AAAA,IACjB,eAAe,MAAM,qBAAqB,MAAM,aAAa;AAAA,IAC7D,SAAS,MAAM;AAAA,IACf,OAAO,MAAM;AAAA,IACb,YAAY,MAAM;AAAA,IAClB,YAAY,MAAM;AAAA,EACpB,CAAC;AACH;AAEA,SAAS,0BACP,OACkC;AAClC,SAAO,cAAc;AAAA,IACnB,eAAe;AAAA,IACf,WAAW,gBAAgB,MAAM,WAAW,WAAW;AAAA,IACvD,eAAe,gBAAgB,MAAM,aAAa;AAAA,IAClD,SAAS,MAAM,UAAU,iBAAiB,MAAM,SAAS,SAAS,IAAI;AAAA,IACtE,OAAO,iBAAiB,MAAM,OAAO,OAAO;AAAA,IAC5C,YAAY,iBAAiB,MAAM,YAAY,YAAY;AAAA,IAC3D,YAAY,MAAM,aACd,eAAe,oBAAoB,MAAM,UAAU,CAAC,IACpD;AAAA,EACN,CAAC;AACH;AAEA,eAAe,qBAAqB,OAA6D;AAC/F,QAAM,OAAO,gBAAgB,MAAM,MAAM,oBAAoB;AAC7D,MAAI,MAAM,SAAS,UAAa,MAAM,YAAY,QAAW;AAC3D,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,SAAS,QAAW;AAC5B,WAAO,EAAE,MAAM,MAAM,iBAAiB,MAAM,MAAM,oBAAoB,EAAE;AAAA,EAC1E;AACA,MAAI,MAAM,YAAY,QAAW;AAC/B,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,aAAW,MAAM,SAAS,uBAAuB;AACjD,SAAO,EAAE,MAAM,MAAM,MAAM,SAAS,MAAM,OAAO,EAAE;AACrD;AAEA,SAAS,gBAAgB,OAA+C;AACtE,SAAO;AAAA,IACL,MAAM,gBAAgB,MAAM,MAAM,oBAAoB;AAAA,IACtD,MAAM,iBAAiB,MAAM,MAAM,oBAAoB;AAAA,EACzD;AACF;AAEA,SAAS,iBAAiB,OAA4B,MAAmC;AACvF,SAAO,cAAc;AAAA,IACnB,IAAI,gBAAgB,MAAM,IAAI,GAAG,IAAI,KAAK;AAAA,IAC1C,SAAS,iBAAiB,MAAM,SAAS,GAAG,IAAI,UAAU;AAAA,IAC1D,MAAM,iBAAiB,MAAM,MAAM,GAAG,IAAI,OAAO;AAAA,EACnD,CAAC;AACH;AAEA,SAAS,oBACP,OAC4C;AAC5C,QAAM,MAAkD,CAAC;AACzD,aAAW,OAAO,OAAO,KAAK,KAAK,EAAE,KAAK,GAAG;AAC3C,UAAM,QAAQ,MAAM,GAAG;AACvB,oBAAgB,KAAK,kBAAkB;AACvC,QACE,UAAU,QACV,OAAO,UAAU,YACjB,OAAO,UAAU,YACjB,OAAO,UAAU,WACjB;AACA,YAAM,IAAI;AAAA,QACR;AAAA,QACA,cAAc,GAAG;AAAA,MACnB;AAAA,IACF;AACA,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;AACxD,YAAM,IAAI,gCAAgC,0BAA0B,cAAc,GAAG,EAAE;AAAA,IACzF;AACA,QAAI,GAAG,IAAI;AAAA,EACb;AACA,SAAO;AACT;AAEA,SAAS,cAAiD,OAAa;AACrE,QAAM,MAA+B,CAAC;AACtC,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,KAAK,GAAG;AAChD,QAAI,UAAU,OAAW,KAAI,GAAG,IAAI;AAAA,EACtC;AACA,SAAO;AACT;AAEA,SAAS,eAAkD,OAAyB;AAClF,SAAO,OAAO,KAAK,KAAK,EAAE,SAAS,IAAI,QAAQ;AACjD;AAEA,SAAS,eAAe,OAAgB,MAAoB;AAC1D,MAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AACvE,UAAM,IAAI,gCAAgC,mBAAmB,IAAI;AAAA,EACnE;AACA,QAAM,MAAM;AACZ,eAAa,IAAI,MAAM,GAAG,IAAI,OAAO;AACrC,mBAAiB,IAAI,MAAM,GAAG,IAAI,OAAO;AAC3C;AAEA,SAAS,gBAAgB,OAAgB,MAAoB;AAC3D,MAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AACvE,UAAM,IAAI,gCAAgC,mBAAmB,IAAI;AAAA,EACnE;AACA,QAAM,MAAM;AACZ,eAAa,IAAI,IAAI,GAAG,IAAI,KAAK;AACjC,MAAI,IAAI,YAAY,OAAW,cAAa,IAAI,SAAS,GAAG,IAAI,UAAU;AAC1E,MAAI,IAAI,SAAS,OAAW,cAAa,IAAI,MAAM,GAAG,IAAI,OAAO;AACnE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;AAC9D,MAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AACvE,UAAM,IAAI,gCAAgC,mBAAmB,IAAI;AAAA,EACnE;AACA,sBAAoB,KAAmD;AACzE;AAEA,SAAS,WAAW,OAAyB,MAAoB;AAC/D,MAAI,UAAU,KAAM;AACpB,QAAM,OAAO,OAAO;AACpB,MAAI,SAAS,YAAY,SAAS,UAAW;AAC7C,MAAI,SAAS,UAAU;AACrB,QAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,YAAM,IAAI,gCAAgC,0BAA0B,IAAI;AAAA,IAC1E;AACA;AAAA,EACF;AACA,MAAI,MAAM,QAAQ,KAAK,GAAG;AACxB,UAAM,QAAQ,CAAC,MAAM,UAAU;AAC7B,iBAAW,MAAM,GAAG,IAAI,IAAI,KAAK,GAAG;AAAA,IACtC,CAAC;AACD;AAAA,EACF;AACA,MAAI,SAAS,UAAU;AACrB,eAAW,CAAC,KAAK,MAAM,KAAK,OAAO,QAAQ,KAAK,GAAG;AACjD,sBAAgB,KAAK,GAAG,IAAI,QAAQ;AACpC,iBAAW,QAAQ,GAAG,IAAI,IAAI,GAAG,EAAE;AAAA,IACrC;AACA;AAAA,EACF;AACA,QAAM,IAAI,gCAAgC,kCAAkC,IAAI;AAClF;AAEA,SAAS,cAAc,OAAgB,UAAkB,MAAoB;AAC3E,MAAI,UAAU,UAAU;AACtB,UAAM,IAAI,gCAAgC,YAAY,QAAQ,IAAI,IAAI;AAAA,EACxE;AACF;AAEA,SAAS,aAAa,OAAgB,MAAoB;AACxD,MAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAAG;AACnD,UAAM,IAAI,gCAAgC,6BAA6B,IAAI;AAAA,EAC7E;AACF;AAEA,SAAS,gBAAgB,OAAe,MAAsB;AAC5D,MAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAAG;AACnD,UAAM,IAAI,gCAAgC,6BAA6B,IAAI;AAAA,EAC7E;AACA,SAAO;AACT;AAEA,SAAS,iBAAiB,OAA2B,MAAkC;AACrF,MAAI,UAAU,OAAW,QAAO;AAChC,SAAO,gBAAgB,OAAO,IAAI;AACpC;AAEA,SAAS,iBAAiB,OAAgB,MAAsB;AAC9D,MAAI,OAAO,UAAU,YAAY,CAAC,WAAW,KAAK,KAAK,GAAG;AACxD,UAAM,IAAI,gCAAgC,0CAA0C,IAAI;AAAA,EAC1F;AACA,SAAO;AACT;AAuBO,IAAM,sBAAsB;AAAA;AAAA;AAAA;AAAA,EAIjC,uBAAuB;AACzB;AAQO,SAAS,mBAAmB,OAAkC;AACnE,MAAI;AACJ,MAAI;AACF,iBAAa,KAAK,UAAU,KAAK;AAAA,EACnC,SAAS,KAAK;AACZ,UAAM,IAAI;AAAA,MACR,4CAA4C,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,MAC5F;AAAA,IACF;AAAA,EACF;AACA,MAAI,eAAe,QAAW;AAC5B,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,SAAO,KAAK,MAAM,UAAU;AAC9B;AAsBA,eAAsB,6BACpB,SACA,OAC2B;AAC3B,MAAI,CAAC,WAAW,OAAO,YAAY,UAAU;AAC3C,UAAM,IAAI,gCAAgC,0CAA0C,SAAS;AAAA,EAC/F;AACA,MAAI,OAAO,QAAQ,SAAS,YAAY,QAAQ,KAAK,WAAW,GAAG;AACjE,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,OAAO,QAAQ,YAAY,YAAY,QAAQ,QAAQ,WAAW,GAAG;AACvE,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,SAAO,sBAAsB;AAAA,IAC3B,GAAG;AAAA,IACH,WAAW,GAAG,QAAQ,IAAI,IAAI,QAAQ,OAAO;AAAA,IAC7C,eAAe;AAAA,MACb,MAAM,oBAAoB;AAAA,MAC1B,SAAS,mBAAmB,OAAO;AAAA,IACrC;AAAA,EACF,CAAC;AACH;;;AC/QA,IAAM,sBAAsB;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,aAAyC,CAAC,UAAU,OAAO,SAAS;AAEnE,IAAM,2BAAN,cAAuC,gBAAgB;AAAA,EACnD;AAAA,EACT,YAAY,SAAiB,OAAO,IAAI;AACtC,UAAM,OAAO,GAAG,OAAO,QAAQ,IAAI,MAAM,OAAO;AAChD,SAAK,OAAO;AAAA,EACd;AACF;AAOO,SAAS,kBAAkB,OAA2B;AAC3D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iBAAiB;AAAA,EACtD;AACA,QAAM,MAAM;AAEZ,aAAW,OAAO,qBAAqB;AACrC,QAAI,EAAE,OAAO,MAAM;AACjB,YAAM,IAAI,yBAAyB,4BAA4B,GAAG,GAAG;AAAA,IACvE;AAAA,EACF;AAEA,EAAAA,cAAa,IAAI,OAAO,OAAO;AAC/B,EAAAA,cAAa,IAAI,cAAc,cAAc;AAC7C,EAAAA,cAAa,IAAI,aAAa,aAAa;AAC3C,qBAAmB,IAAI,MAAM,MAAM;AACnC,EAAAA,cAAa,IAAI,OAAO,OAAO;AAC/B,EAAAA,cAAa,IAAI,YAAY,YAAY;AACzC,EAAAA,cAAa,IAAI,YAAY,YAAY;AACzC,EAAAA,cAAa,IAAI,WAAW,WAAW;AACvC,qBAAmB,IAAI,QAAQ,QAAQ;AACvC,MAAI,IAAI,YAAY,OAAW,oBAAmB,IAAI,SAAS,SAAS;AACxE,qBAAmB,IAAI,SAAS,SAAS;AAGzC,MAAI,CAAC,iBAAiB,IAAI,KAAe,GAAG;AAC1C,UAAM,IAAI;AAAA,MACR,UAAU,IAAI,KAAK;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AAGA,QAAM,KAAK,IAAI;AACf,MAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,UAAM,IAAI,yBAAyB,gCAAgC,YAAY;AAAA,EACjF;AACA,QAAM,QAAQ;AACd,qBAAmB,MAAM,OAAO,kBAAkB;AAClD,qBAAmB,MAAM,QAAQ,mBAAmB;AACpD,MAAI,MAAM,WAAW,OAAW,oBAAmB,MAAM,QAAQ,mBAAmB;AAGpF,MAAI,IAAI,kBAAkB,QAAW;AACnC,UAAM,KAAK,IAAI;AACf,QAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,YAAM,IAAI,yBAAyB,mCAAmC,eAAe;AAAA,IACvF;AACA,UAAM,QAAQ;AACd,IAAAA,cAAa,MAAM,OAAO,qBAAqB;AAC/C,IAAAA,cAAa,MAAM,eAAe,6BAA6B;AAC/D,uBAAmB,MAAM,YAAY,0BAA0B;AAC/D,QAAI,OAAO,MAAM,aAAa,WAAW;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAGA,QAAM,MAAM,IAAI;AAChB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,6BAA6B,SAAS;AAAA,EAC3E;AACA,QAAM,SAAS;AACf,MAAI,OAAO,gBAAgB;AACzB,uBAAmB,OAAO,aAAa,qBAAqB;AAC9D,MAAI,OAAO,iBAAiB;AAC1B,uBAAmB,OAAO,cAAc,sBAAsB;AAChE,MAAI,OAAO,gBAAgB,UAAa,OAAO,iBAAiB,QAAW;AACzE,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,QAAM,MAAM,OAAO;AACnB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,iCAAiC,aAAa;AAAA,EACnF;AACA,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,GAA8B,GAAG;AACnE,uBAAmB,GAAG,eAAe,CAAC,EAAE;AAAA,EAC1C;AAGA,MAAI,OAAO,gBAAgB,QAAW;AACpC,wBAAoB,OAAO,aAAa,qBAAqB;AAAA,EAC/D;AAGA,MAAI,IAAI,gBAAgB,OAAW,CAAAA,cAAa,IAAI,aAAa,aAAa;AAE9E,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI;AACF,YAAM,UAAU,yBAAyB,IAAI,YAAY;AACzD,UAAI,QAAQ,UAAU,UAAa,QAAQ,UAAU,IAAI,OAAO;AAC9D,cAAM,IAAI;AAAA,UACR,uBAAuB,QAAQ,KAAK,2BAA2B,IAAI,KAAK;AAAA,UACxE;AAAA,QACF;AAAA,MACF;AACA,UAAI,QAAQ,eAAe,UAAa,QAAQ,eAAe,IAAI,YAAY;AAC7E,cAAM,IAAI;AAAA,UACR,4BAA4B,QAAQ,UAAU,gCAAgC,IAAI,UAAU;AAAA,UAC5F;AAAA,QACF;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,yBAA0B,OAAM;AACrD,UAAI,iBAAiB,OAAO;AAC1B,cAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;AAAA,MAClE;AACA,YAAM;AAAA,IACR;AAAA,EACF;AAGA,MAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GAAG;AACzF,UAAM,IAAI;AAAA,MACR,2BAA2B,WAAW,KAAK,IAAI,CAAC,SAAS,OAAO,IAAI,QAAQ,CAAC;AAAA,MAC7E;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAGO,SAAS,YAAY,OAAoC;AAC9D,MAAI;AACF,sBAAkB,KAAK;AACvB,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGO,SAAS,mBACd,OACiF;AACjF,MAAI;AACF,WAAO,EAAE,IAAI,MAAM,OAAO,kBAAkB,KAAK,EAAE;AAAA,EACrD,SAAS,GAAG;AACV,QAAI,aAAa,yBAA0B,QAAO,EAAE,IAAI,OAAO,OAAO,EAAE;AACxE,UAAM;AAAA,EACR;AACF;AAGO,SAAS,mBAAmB,QAA8B;AAC/D,QAAM,OAAO,KAAK,UAAU,MAAM;AAClC,SAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAASA,cAAa,OAAgB,MAAoB;AACxD,MAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAAG;AACnD,UAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAAA,EACtE;AACF;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;AAC9D,MAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;AACxD,UAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAAA,EACnE;AACF;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;AAC/D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iCAAiC,IAAI;AAAA,EAC1E;AACA,QAAM,MAAM;AAEZ,QAAM,WAAW,IAAI;AACrB,MAAI,aAAa,QAAQ,OAAO,aAAa,UAAU;AACrD,UAAM,IAAI,yBAAyB,8BAA8B,GAAG,IAAI,WAAW;AAAA,EACrF;AACA,aAAW,CAAC,SAAS,IAAI,KAAK,OAAO,QAAQ,QAAmC,GAAG;AACjF,QAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI,aAAa,OAAO;AAAA,MAC7B;AAAA,IACF;AACA,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,IAA+B,GAAG;AAC1E,yBAAmB,OAAO,GAAG,IAAI,aAAa,OAAO,IAAI,GAAG,EAAE;AAAA,IAChE;AAAA,EACF;AAEA,QAAM,aAAa,IAAI;AACvB,MAAI,eAAe,QAAQ,OAAO,eAAe,UAAU;AACzD,UAAM,IAAI,yBAAyB,gCAAgC,GAAG,IAAI,aAAa;AAAA,EACzF;AACA,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,UAAqC,GAAG;AAC/E,uBAAmB,MAAM,GAAG,IAAI,eAAe,GAAG,EAAE;AAAA,EACtD;AAEA,qBAAmB,IAAI,WAAW,GAAG,IAAI,YAAY;AAErD,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GAAG;AACpC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI;AAAA,MACT;AAAA,IACF;AACA,aAAS,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;AAChD,YAAM,KAAK,IAAI,aAAa,CAAC;AAC7B,UAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAAG;AAC7C,cAAM,IAAI;AAAA,UACR;AAAA,UACA,GAAG,IAAI,iBAAiB,CAAC;AAAA,QAC3B;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,MAAI,IAAI,UAAU,UAAa,OAAO,IAAI,UAAU,UAAU;AAC5D,UAAM,IAAI,yBAAyB,0BAA0B,GAAG,IAAI,QAAQ;AAAA,EAC9E;AACF;AAYA,SAAS,iBAAiB,OAAwB;AAChD,MAAI,MAAM,SAAS,GAAG,EAAG,QAAO;AAChC,MAAI,UAAU,KAAK,KAAK,EAAG,QAAO;AAClC,MAAI,sBAAsB,KAAK,KAAK,EAAG,QAAO;AAC9C,MAAI,SAAS,KAAK,KAAK,EAAG,QAAO;AACjC,SAAO;AACT;","names":["expectString"]}

package/dist/{chunk-5KSDYBYH.js → chunk-YXTT6GSZ.js}

@@ -2,7 +2,7 @@ import {
   buildAgentProfileCell,
   validateRunRecord,
   verifyAgentProfileCell
-} from "./chunk-BWZEGTES.js";
+} from "./chunk-NCK5QLGT.js";
 import {
   researchReport
 } from "./chunk-EGIPWXHL.js";
@@ -328,4 +328,4 @@ function defaultRunId(params) {
 export {
   runEvalCampaign
 };
-//# sourceMappingURL=chunk-5KSDYBYH.js.map
+//# sourceMappingURL=chunk-YXTT6GSZ.js.map

package/dist/contract/index.d.ts

@@ -1,20 +1,21 @@
-import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, I as ImprovementDriver, G as Gate } from '../types-Dbj5gu8n.js';
-export { f as CampaignAggregates, g as CampaignArtifactWriter, h as CampaignCellResult, i as CampaignCostMeter, j as CampaignResult, k as CampaignTraceWriter, C as CodeSurface, D as Dispatch, l as GateContext, m as GateDecision, n as GateResult, o as GenerationCandidate, p as GenerationRecord, r as JudgeDimension, J as JudgeScore, t as Mutator, O as OptimizerConfig, v as SessionScript } from '../types-Dbj5gu8n.js';
-import { C as CampaignStorage, e as RunImprovementLoopResult } from '../run-improvement-loop-Cc7oZlRP.js';
-export { D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, a as GepaDriverOptions, H as HeldOutGateOptions, R as RunCampaignOptions, c as RunEvalOptions, d as RunImprovementLoopOptions, h as composeGate, j as defaultProductionGate, k as evolutionaryDriver, m as fsCampaignStorage, n as gepaDriver, o as heldOutGate, p as inMemoryCampaignStorage, r as runCampaign, s as runEval, t as runImprovementLoop } from '../run-improvement-loop-Cc7oZlRP.js';
+import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, I as ImprovementDriver, G as Gate } from '../types-BgrxOJSf.js';
+export { g as CampaignAggregates, h as CampaignArtifactWriter, i as CampaignCellResult, j as CampaignCostMeter, k as CampaignResult, l as CampaignTraceWriter, C as CodeSurface, D as Dispatch, m as GateContext, n as GateDecision, o as GateResult, p as GenerationCandidate, q as GenerationRecord, s as JudgeDimension, J as JudgeScore, u as Mutator, O as OptimizerConfig, w as SessionScript } from '../types-BgrxOJSf.js';
+import { C as CampaignStorage, e as RunImprovementLoopResult } from '../run-improvement-loop-BhfdjrMY.js';
+export { D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, a as GepaDriverOptions, H as HeldOutGateOptions, R as RunCampaignOptions, c as RunEvalOptions, d as RunImprovementLoopOptions, h as composeGate, j as defaultProductionGate, k as evolutionaryDriver, m as fsCampaignStorage, n as gepaDriver, o as heldOutGate, p as inMemoryCampaignStorage, r as runCampaign, s as runEval, t as runImprovementLoop } from '../run-improvement-loop-BhfdjrMY.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, b as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, O as OutcomeStore } from '../outcome-store-D6KWmYvj.js';
-import { a as HostedTenant, I as InsightReport, T as TraceSpanEvent } from '../index-C7RhhEME.js';
-export { F as FailureClusterInsight, b as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, c as ReleaseSummary, S as ScalarDistribution } from '../index-C7RhhEME.js';
-import { A as AnalystRegistry } from '../registry-8KAs18kY.js';
+import { a as HostedTenant, I as InsightReport, T as TraceSpanEvent } from '../index-D2nT6_KT.js';
+export { F as FailureClusterInsight, b as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, c as ReleaseSummary, S as ScalarDistribution } from '../index-D2nT6_KT.js';
+import { A as AnalystRegistry } from '../registry-BSWy0rvH.js';
 import { a as DatasetScenario } from '../dataset-BlwAtYYf.js';
-import { R as RunRecord, a as RunSplitTag } from '../run-record-BGY6bHRh.js';
+import { R as RunRecord, a as RunSplitTag } from '../run-record-etiCMsUq.js';
 import '../llm-client-BXVRUZyX.js';
 import '../errors-mje_cKOs.js';
 import '../raw-provider-sink-C46HDghv.js';
-import '../red-team-30II1T4o.js';
-import '../store-Db2Bv8Cf.js';
-import '../summary-report-B7gNRX-r.js';
-import '../failure-cluster-Cw65_5FY.js';
+import '../red-team-CrC5MZYd.js';
+import '../store-CKUAgsJz.js';
+import '../schema-m0gsnbt3.js';
+import '../summary-report-DLxh4yWk.js';
+import '../failure-cluster-CL7IVgkJ.js';
 import '../judge-calibration-DilmB3Ml.js';
 import '../store-CJbzDxZ2.js';
 import '../types-DhqpAi_z.js';

package/dist/contract/index.js

@@ -75,6 +75,7 @@ async function analyzeRuns(opts) {
   const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : void 0;
   const lift = computeLift(runs, opts.baselineCandidateId, opts.candidateCandidateId, split);
   const failureClusters = opts.analyst ? await computeFailureClusters(runs, opts.analyst, split) : void 0;
+  const failureModes = computeFailureModes(runs);
   const contamination = opts.canaryScenarios ? computeContamination(runs, opts.canaryScenarios) : void 0;
   const outcomeCorrelation = opts.outcomeSignal ? computeOutcomeCorrelation(runs, opts.outcomeSignal, split) : void 0;
   const release = buildReleaseScorecard(composite, lift, contamination);
@@ -85,6 +86,7 @@ async function analyzeRuns(opts) {
     interRater,
     lift,
     failureClusters,
+    failureModes,
     contamination,
     outcomeCorrelation,
     priorPeriodComparison,
@@ -102,10 +104,21 @@ async function analyzeRuns(opts) {
     contamination,
     outcomeCorrelation,
     release,
+    ...failureModes ? { failureModes } : {},
     ...priorPeriodComparison ? { priorPeriodComparison } : {},
     recommendations
   };
 }
+function computeFailureModes(runs) {
+  const counts = /* @__PURE__ */ new Map();
+  for (const r of runs) {
+    const key = r.failureClass ?? r.failureMode;
+    if (key) counts.set(key, (counts.get(key) ?? 0) + 1);
+  }
+  if (counts.size === 0) return void 0;
+  const n = runs.length;
+  return [...counts.entries()].map(([mode, count]) => ({ mode, count, share: n > 0 ? count / n : 0 })).sort((a, b) => b.count - a.count || a.mode.localeCompare(b.mode));
+}
 function computePriorPeriodComparison(current, baseline, split, windowLabel) {
   if (current.length === 0 || baseline.length === 0) return void 0;
   const metrics = {};
@@ -661,6 +674,18 @@ function buildRecommendations(ctx) {
       });
     }
   }
+  if (ctx.failureModes && ctx.failureModes.length > 0) {
+    const top = ctx.failureModes[0];
+    if (top.count >= 3 && top.share >= 0.15) {
+      out.push({
+        priority: top.share >= 0.25 ? "high" : "medium",
+        kind: "investigate",
+        title: `'${top.mode}' is the dominant failure mode \u2014 ${top.count} runs (${(top.share * 100).toFixed(0)}% of the corpus)`,
+        detail: `The mean composite can look acceptable while one named failure dominates the lower tail. ${top.count} of ${ctx.composite.n} runs failed with '${top.mode}'${ctx.failureModes.length > 1 ? ` (next: '${ctx.failureModes[1].mode}' \xD7${ctx.failureModes[1].count})` : ""}. Fix this cause first.`,
+        evidencePath: "failureModes"
+      });
+    }
+  }
   if (Object.keys(ctx.judges).length === 0 && ctx.composite.n > 0) {
     out.push({
       priority: "medium",