@tangle-network/agent-eval 0.126.6 → 0.127.0

package/dist/{chunk-CGG5SLH3.js → chunk-ARCFIY7U.js}

@@ -1,7 +1,7 @@
 import {
   fsCampaignStorage,
   runCampaign
-} from "./chunk-ZVCHKKOP.js";
+} from "./chunk-WED5KLZQ.js";
 import {
   __export
 } from "./chunk-PZ5AY32C.js";
@@ -763,4 +763,4 @@ export {
   retrievalMetricsAtCutoff,
   benchmarks_exports
 };
-//# sourceMappingURL=chunk-CGG5SLH3.js.map
+//# sourceMappingURL=chunk-ARCFIY7U.js.map

package/dist/{chunk-P22LJ3Y2.js → chunk-CLELFRQG.js}

@@ -1,9 +1,9 @@
 import {
   assertLlmRoute
-} from "./chunk-J5SQWP6Y.js";
+} from "./chunk-PBE2LOSS.js";
 import {
   researchReport
-} from "./chunk-DPZAEKA6.js";
+} from "./chunk-IMGJZBFG.js";
 import {
   RunIntegrityError,
   assertRunCaptured
@@ -18,7 +18,7 @@ import {
   buildAgentProfileCell,
   validateRunRecord,
   verifyAgentProfileCell
-} from "./chunk-IILEIWGW.js";
+} from "./chunk-NL33OXEP.js";
 import {
   canonicalize,
   hashJson
@@ -223,7 +223,9 @@ async function runEvalCampaign(opts) {
         commitSha: opts.commitSha,
         wallMs,
         costUsd: outcome.costUsd,
+        costProvenance: outcome.costProvenance,
         tokenUsage: outcome.tokenUsage,
+        terminalOutcome: "succeeded",
         judgeMetadata: outcome.judgeMetadata,
         outcome: recordOutcome,
         ...outcome.failureClass ? { failureClass: outcome.failureClass } : {},
@@ -352,4 +354,4 @@ function defaultRunId(params) {
 export {
   runEvalCampaign
 };
-//# sourceMappingURL=chunk-P22LJ3Y2.js.map
+//# sourceMappingURL=chunk-CLELFRQG.js.map

package/dist/chunk-CLELFRQG.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/eval-campaign.ts"],"sourcesContent":["/**\n * EvalCampaign — opinionated matrix runner that wires the four\n * capture-integrity directives by construction.\n *\n * The canonical benchmark shape — matrix runner → for each\n * (variant, scenario, seed) → start a TraceEmitter → call LLMs → end the\n * run → analyze — has a bug class at the integration boundary: raw\n * events not captured, route silently wrong, integrity not asserted,\n * analyst never run. The directives in `SKILL.md § Capture integrity`\n * are the mitigations.\n *\n * `EvalCampaign` is the structural fix — consumers don't wire the\n * integrity surface themselves; the campaign owns it. Specifically:\n *\n *   - calls `assertLlmRoute` once at preflight before any work runs\n *   - constructs a per-run `TraceStore` and `RawProviderSink` via factories\n *   - constructs the `TraceEmitter` with `onRunComplete: [analyst hook]`\n *   - hands the runner an `LlmClientOptions` pre-wired with the sink and\n *     trace context — the runner can't accidentally call an LLM without\n *     capturing the raw HTTP envelope\n *   - calls `assertRunCaptured` after every `endRun` and routes failures\n *     through a configurable policy (`throw` / `mark_failed` / `log`)\n *   - assembles per-run `RunRecord`s and runs `researchReport` at the end\n *     so the campaign artifact is launch-decision-grade by default\n *   - embeds the campaign fingerprint (a SHA-256 over the canonicalised\n *     run set) and optional `preregistrationHash` in the report\n *\n * The runner contract is intentionally narrow: produce a `CampaignRunOutcome`\n * given a fully-wired `CampaignRunContext`. Everything orchestration-shaped\n * lives in the campaign. This is the inversion-of-control point — consumers\n * stop writing matrix runners and start writing scenario-runners.\n *\n * Out of scope for v1 (tracked in `docs/research-report-methodology.md`):\n *\n *   - Distributed/cluster execution (concurrency is local async)\n *   - Adaptive sampling / sequential interim looks\n *   - Resume from partial state across crashes\n *   - LLM-call retry beyond what `LlmClient` already does\n */\n\nimport {\n  type AgentProfileCell,\n  type AgentProfileCellInput,\n  buildAgentProfileCell,\n  verifyAgentProfileCell,\n} from './agent-profile-cell'\nimport { assertLlmRoute, type LlmClientOptions, type LlmRouteRequirements } from './llm-client'\nimport { canonicalize, hashJson } from './pre-registration'\nimport type {\n  JudgeScoresRecord,\n  RunCostProvenance,\n  RunJudgeMetadata,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTokenUsage,\n} from './run-record'\nimport { validateRunRecord } from './run-record'\nimport { type ResearchReport, type ResearchReportOptions, researchReport } from './summary-report'\nimport type { RunCompleteHook } from './trace/emitter'\nimport { TraceEmitter } from './trace/emitter'\nimport {\n  assertRunCaptured,\n  RunIntegrityError,\n  type RunIntegrityExpectations,\n  type RunIntegrityReport,\n} from './trace/integrity'\nimport { FileSystemRawProviderSink, type RawProviderSink } from './trace/raw-provider-sink'\nimport type { FailureClass } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\n// ── Public types ─────────────────────────────────────────────────────────\n\nexport interface CampaignVariant<V> {\n  id: string\n  payload: V\n}\n\nexport interface CampaignScenario {\n  scenarioId: string\n  /** Free-form metadata propagated to runs and reports. */\n  tags?: Record<string, string>\n}\n\nexport interface CampaignRunContext<V> {\n  /** Stable run id. The campaign generates this; the runner does not. */\n  runId: string\n  /** Logical experiment id (campaignId by default; overridable per-run via opts). */\n  experimentId: string\n  variant: V\n  variantId: string\n  scenarioId: string\n  scenarioTags: Record<string, string>\n  seed: number\n  splitTag: RunSplitTag\n  /**\n   * The TraceEmitter for this run, with `onRunComplete` hooks pre-wired\n   * (analyst auto-execution if configured, plus integrity check). The\n   * runner MUST call `emitter.startRun` before doing any work and either\n   * `emitter.endRun` or `emitter.abortRun` before returning.\n   */\n  emitter: TraceEmitter\n  store: TraceStore\n  rawSink: RawProviderSink\n  /**\n   * Pre-wired LLM client options — `rawSink` and `traceContext` are populated\n   * so any `callLlm(req, ctx.llmOpts)` automatically captures raw HTTP. The\n   * runner can spread additional fields if needed.\n   */\n  llmOpts: LlmClientOptions\n}\n\nexport interface CampaignRunOutcome {\n  /** Did the run pass? Mirrors `RunOutcome.pass` semantics. */\n  pass: boolean\n  /** Score for the run on its split. Maps to `searchScore` or `holdoutScore`. */\n  score: number\n  /** Cost in USD, or null when the runner could not capture it. */\n  costUsd: number | null\n  /** Source of the cost amount. */\n  costProvenance: RunCostProvenance\n  tokenUsage: RunTokenUsage\n  /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */\n  model: string\n  /** sha256 of the effective prompt sent to the model. */\n  promptHash: string\n  /** sha256 of the effective config (model, temperature, tools, judges, splits). */\n  configHash: string\n  /** Optional extra numeric metrics to land in `outcome.raw`. */\n  raw?: Record<string, number>\n  /** Canonical cross-agent failure class from the shared `FAILURE_CLASSES`\n   *  taxonomy. Propagated to `RunRecord.failureClass` so campaign runs\n   *  aggregate failures in the same vocabulary as every other producer. */\n  failureClass?: FailureClass\n  /** Optional free-form failure detail, scoped under `failureClass`. */\n  failureMode?: string\n  /** Optional judge metadata when a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /**\n   * Optional per-judge / per-dim breakdown for ensemble-judged runs.\n   * Propagated to `outcome.judgeScores` on the resulting `RunRecord`.\n   * Single-judge or scalar-only runs leave this unset.\n   */\n  judgeScores?: JudgeScoresRecord\n  /**\n   * Agent profile cell observed by the runner. When supplied, it overrides\n   * `EvalCampaignOptions.agentProfile` for this run and must match the\n   * outcome's `model` and `promptHash`.\n   */\n  agentProfile?: AgentProfileCell | AgentProfileCellInput\n}\n\nexport type CampaignRunner<V> = (ctx: CampaignRunContext<V>) => Promise<CampaignRunOutcome>\n\nexport type CampaignIntegrityPolicy = 'throw' | 'mark_failed' | 'log'\n\nexport interface EvalCampaignOptions<V> {\n  /**\n   * Stable id for the campaign. Used as the default `experimentId` on\n   * every run, and folded into the campaign fingerprint.\n   */\n  campaignId: string\n  variants: CampaignVariant<V>[]\n  scenarios: CampaignScenario[]\n  /** Default `[0, 1, 2]`. */\n  seeds?: number[]\n  /** Default `'holdout'` — the split that anchors a launch decision. */\n  splitTag?: RunSplitTag\n  /** Git SHA the campaign is run against. Mandatory; `RunRecord` rejects unset. */\n  commitSha: string\n  /**\n   * LLM client config. Augmented per-run with `rawSink` and `traceContext`\n   * before being passed to the runner. The campaign asserts this config\n   * matches `routeRequirements` once at preflight.\n   */\n  llmOpts: LlmClientOptions\n  /**\n   * Default `{ requireExplicitBaseUrl: true, requireAuth: true }` — fail\n   * loud if the campaign would silently fall back to the public router or\n   * run unauthenticated. Override with an empty object to disable.\n   */\n  routeRequirements?: LlmRouteRequirements\n  /**\n   * Per-run TraceStore factory. Common shape: a fresh store per run keyed\n   * on `runId`. Implementations that share a store across the campaign\n   * are valid — the campaign only writes through `emitter`.\n   */\n  storeFactory: (params: CampaignFactoryParams) => TraceStore\n  /**\n   * Per-run RawProviderSink factory. Defaults to `FileSystemRawProviderSink`\n   * rooted at `${workDir}/raw-events/${runId}` if `workDir` is supplied;\n   * otherwise required. Forensic capture is non-negotiable in a campaign\n   * run — pass `NoopRawProviderSink` explicitly if you want to opt out.\n   */\n  rawSinkFactory?: (params: CampaignFactoryParams) => RawProviderSink\n  /**\n   * Filesystem root for default `rawSinkFactory`. Ignored if\n   * `rawSinkFactory` is supplied.\n   */\n  workDir?: string\n  /**\n   * Extra `onRunComplete` hooks the campaign appends (after its own\n   * integrity-check hook). Pass `traceAnalystOnRunComplete(...)` here.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /**\n   * Per-run integrity expectations. Defaults to:\n   *   `{ llmSpansMin: 1, requireRawCoverageOfLlmSpans: true, requireOutcome: true }`.\n   * Override (e.g. `{ llmSpansMin: 0 }`) for runs that don't call LLMs.\n   */\n  integrity?: RunIntegrityExpectations\n  /** Behaviour when integrity fails. Default `'mark_failed'`. */\n  onIntegrityFailure?: CampaignIntegrityPolicy\n  /**\n   * Per-run runner. Receives a fully-wired context; produces an outcome\n   * the campaign converts into a `RunRecord`.\n   */\n  runner: CampaignRunner<V>\n  /**\n   * If set, the campaign computes `researchReport` at the end. `comparator`\n   * is a `variantId`. Other fields are forwarded verbatim.\n   */\n  report?: { comparator?: string } & Omit<\n    ResearchReportOptions,\n    'comparator' | 'preregistrationHash' | 'generatedAt'\n  >\n  /**\n   * Hash of a signed `HypothesisManifest` (see `pre-registration.ts`).\n   * Embedded in the campaign fingerprint and the research report.\n   */\n  preregistrationHash?: string\n  /** Local concurrency. Default `1` (sequential). */\n  concurrency?: number\n  /**\n   * Override the time source. Tests pass a mock to make wallMs deterministic.\n   */\n  now?: () => number\n  /** Override the runId generator. Tests pin this. */\n  runId?: (params: CampaignFactoryParams) => string\n  /**\n   * Agent profile cell for campaign runs. Static profiles can pass an object;\n   * routers or variant-specific harnesses can pass a factory. The campaign\n   * stamps the built cell onto every `RunRecord` and rejects profile/model or\n   * profile/prompt contradictions.\n   */\n  agentProfile?:\n    | AgentProfileCell\n    | AgentProfileCellInput\n    | ((\n        params: CampaignFactoryParams & {\n          variant: V\n          scenarioTags: Record<string, string>\n        },\n      ) =>\n        | AgentProfileCell\n        | AgentProfileCellInput\n        | Promise<AgentProfileCell | AgentProfileCellInput>)\n}\n\nexport interface CampaignFactoryParams {\n  campaignId: string\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n}\n\nexport interface FailedRun {\n  runId: string\n  variantId: string\n  scenarioId: string\n  seed: number\n  reason: string\n  error?: string\n}\n\nexport interface EvalCampaignResult {\n  campaignId: string\n  /** SHA-256 over canonicalised `(variantIds, scenarioIds, seeds, comparator, splitTag, baseUrl, provider, preregistrationHash)`. */\n  campaignFingerprint: string\n  preregistrationHash: string | null\n  /** Successful runs only. Failed runs land in `failedRuns`. */\n  runs: RunRecord[]\n  /** Integrity reports for every successful run. */\n  integrityReports: RunIntegrityReport[]\n  failedRuns: FailedRun[]\n  /** Computed when `report` is set on options. */\n  report?: ResearchReport\n  startedAt: string\n  endedAt: string\n}\n\n// ── Implementation ───────────────────────────────────────────────────────\n\nconst DEFAULT_INTEGRITY: RunIntegrityExpectations = {\n  llmSpansMin: 1,\n  requireRawCoverageOfLlmSpans: true,\n  requireOutcome: true,\n}\n\nconst DEFAULT_ROUTE: LlmRouteRequirements = {\n  requireExplicitBaseUrl: true,\n  requireAuth: true,\n}\n\nexport async function runEvalCampaign<V>(\n  opts: EvalCampaignOptions<V>,\n): Promise<EvalCampaignResult> {\n  // ── Preflight ──────────────────────────────────────────────────────\n  assertLlmRoute(opts.llmOpts, opts.routeRequirements ?? DEFAULT_ROUTE)\n\n  if (opts.variants.length === 0) {\n    throw new Error('runEvalCampaign: variants must be non-empty.')\n  }\n  if (opts.scenarios.length === 0) {\n    throw new Error('runEvalCampaign: scenarios must be non-empty.')\n  }\n  const variantIds = new Set<string>()\n  for (const v of opts.variants) {\n    if (variantIds.has(v.id)) {\n      throw new Error(`runEvalCampaign: duplicate variant id \"${v.id}\".`)\n    }\n    variantIds.add(v.id)\n  }\n  const scenarioIds = new Set<string>()\n  for (const s of opts.scenarios) {\n    if (scenarioIds.has(s.scenarioId)) {\n      throw new Error(`runEvalCampaign: duplicate scenarioId \"${s.scenarioId}\".`)\n    }\n    scenarioIds.add(s.scenarioId)\n  }\n  if (opts.report?.comparator && !variantIds.has(opts.report.comparator)) {\n    throw new Error(\n      `runEvalCampaign: report.comparator \"${opts.report.comparator}\" is not a configured variantId.`,\n    )\n  }\n  if (!opts.commitSha) {\n    throw new Error('runEvalCampaign: commitSha is required (every RunRecord needs it).')\n  }\n\n  const seeds = opts.seeds ?? [0, 1, 2]\n  const splitTag: RunSplitTag = opts.splitTag ?? 'holdout'\n  const concurrency = Math.max(1, opts.concurrency ?? 1)\n  const integrity = { ...DEFAULT_INTEGRITY, ...(opts.integrity ?? {}) }\n  const onIntegrityFailure: CampaignIntegrityPolicy = opts.onIntegrityFailure ?? 'mark_failed'\n  const now = opts.now ?? (() => Date.now())\n  const baseUrl = (opts.llmOpts.baseUrl ?? '').replace(/\\/+$/, '')\n  const provider = opts.llmOpts.provider ?? null\n  const preregistrationHash = opts.preregistrationHash ?? null\n\n  const rawSinkFactory = opts.rawSinkFactory ?? defaultRawSinkFactory(opts.workDir)\n\n  // ── Fingerprint ────────────────────────────────────────────────────\n  const campaignFingerprint = await hashJson(\n    canonicalize({\n      campaignId: opts.campaignId,\n      variants: opts.variants.map((v) => v.id).sort(),\n      scenarios: opts.scenarios.map((s) => s.scenarioId).sort(),\n      seeds: [...seeds].sort((a, b) => a - b),\n      splitTag,\n      comparator: opts.report?.comparator ?? null,\n      baseUrl,\n      provider,\n      preregistrationHash,\n    }),\n  )\n\n  // ── Plan the matrix ────────────────────────────────────────────────\n  type Cell = { variant: CampaignVariant<V>; scenario: CampaignScenario; seed: number }\n  const cells: Cell[] = []\n  for (const variant of opts.variants) {\n    for (const scenario of opts.scenarios) {\n      for (const seed of seeds) {\n        cells.push({ variant, scenario, seed })\n      }\n    }\n  }\n\n  const startedAt = new Date(now()).toISOString()\n  const runs: RunRecord[] = []\n  const integrityReports: RunIntegrityReport[] = []\n  const failedRuns: FailedRun[] = []\n\n  // ── Execute (bounded-concurrency worker pool) ──────────────────────\n  // A genuine (non-CellExecutionError) error from any worker is a bug, not a\n  // run-level failure. We must NOT let `Promise.all` reject mid-flight and\n  // orphan the other workers' in-progress runs (emitters never finalized,\n  // sinks/handles leak, partial work silently discarded). Instead: capture the\n  // first genuine error, stop dispatching new cells so in-flight workers wind\n  // down, finalize every still-open run, then re-throw the aggregated error.\n  let cursor = 0\n  let aborting = false\n  const genuineErrors: unknown[] = []\n  // Emitters for runs that are currently mid-flight, keyed by runId. A worker\n  // registers its emitter before invoking the runner and removes it once the\n  // run is finalized (via endRun/abortRun, success or failure). Anything left\n  // here after the pool settles is an orphan we must abort.\n  const openRuns = new Map<string, TraceEmitter>()\n\n  async function worker(): Promise<void> {\n    while (!aborting) {\n      const i = cursor++\n      if (i >= cells.length) return\n      const cell = cells[i]!\n      try {\n        const result = await runOneCell(cell)\n        runs.push(result.record)\n        integrityReports.push(result.integrity)\n      } catch (err) {\n        if (err instanceof CellExecutionError) {\n          failedRuns.push(err.failed)\n          if (err.integrity) integrityReports.push(err.integrity)\n        } else {\n          // Genuine bug — not a runner failure, not an integrity failure.\n          // Capture it and stop dispatching so peers wind down gracefully;\n          // re-thrown after the pool settles. Do not surface here — that would\n          // reject Promise.all and orphan the other workers' runs.\n          genuineErrors.push(err)\n          aborting = true\n          return\n        }\n      }\n    }\n  }\n\n  async function runOneCell(\n    cell: Cell,\n  ): Promise<{ record: RunRecord; integrity: RunIntegrityReport }> {\n    const runId = (opts.runId ?? defaultRunId)({\n      campaignId: opts.campaignId,\n      runId: '', // unused by default generator\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    })\n    const factoryParams: CampaignFactoryParams = {\n      campaignId: opts.campaignId,\n      runId,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      seed: cell.seed,\n    }\n    const store = opts.storeFactory(factoryParams)\n    const rawSink = rawSinkFactory(factoryParams)\n\n    const emitter = new TraceEmitter(store, {\n      runId,\n      now: opts.now,\n      onRunComplete: opts.onRunComplete,\n    })\n    // Track this run as open so a genuine error elsewhere in the pool can\n    // finalize it instead of orphaning it. Removed in the finally below.\n    openRuns.set(runId, emitter)\n\n    const llmOpts: LlmClientOptions = {\n      ...opts.llmOpts,\n      rawSink,\n      traceContext: { runId },\n    }\n\n    const ctx: CampaignRunContext<V> = {\n      runId,\n      experimentId: opts.campaignId,\n      variant: cell.variant.payload,\n      variantId: cell.variant.id,\n      scenarioId: cell.scenario.scenarioId,\n      scenarioTags: cell.scenario.tags ?? {},\n      seed: cell.seed,\n      splitTag,\n      emitter,\n      store,\n      rawSink,\n      llmOpts,\n    }\n\n    try {\n      const wallStart = now()\n      let outcome: CampaignRunOutcome\n      try {\n        outcome = await opts.runner(ctx)\n      } catch (err) {\n        const message = err instanceof Error ? err.message : String(err)\n        // The runner threw mid-execution. Abort the run so the emitter\n        // finalizes. The only benign abortRun failure is \"the runner never\n        // started the run\" (nothing to finalize). A store-write failure (disk\n        // full, FS error) is a genuine diagnostic — surface it rather than\n        // masking it as a plain runner failure.\n        await finalizeAbort(emitter, runId, message)\n        throw new CellExecutionError({\n          runId,\n          variantId: cell.variant.id,\n          scenarioId: cell.scenario.scenarioId,\n          seed: cell.seed,\n          reason: 'runner_threw',\n          error: message,\n        })\n      }\n      const wallMs = now() - wallStart\n\n      const integrityReport = await assertRunCaptured(store, runId, { ...integrity, rawSink })\n      if (!integrityReport.ok) {\n        switch (onIntegrityFailure) {\n          case 'throw':\n            throw new RunIntegrityError(integrityReport)\n          case 'mark_failed':\n            throw new CellExecutionError(\n              {\n                runId,\n                variantId: cell.variant.id,\n                scenarioId: cell.scenario.scenarioId,\n                seed: cell.seed,\n                reason: 'integrity_failed',\n                error: integrityReport.issues.map((i) => i.code).join(', '),\n              },\n              integrityReport,\n            )\n          case 'log':\n            // Caller wants the run admitted with a flagged report; fall through.\n            break\n        }\n      }\n\n      const recordOutcome: RunOutcome = {\n        raw: outcome.raw ?? {},\n      }\n      if (splitTag === 'holdout') recordOutcome.holdoutScore = outcome.score\n      else recordOutcome.searchScore = outcome.score\n      if (outcome.judgeScores !== undefined) recordOutcome.judgeScores = outcome.judgeScores\n\n      const record: RunRecord = {\n        runId,\n        experimentId: opts.campaignId,\n        candidateId: cell.variant.id,\n        seed: cell.seed,\n        model: outcome.model,\n        promptHash: outcome.promptHash,\n        configHash: outcome.configHash,\n        commitSha: opts.commitSha,\n        wallMs,\n        costUsd: outcome.costUsd,\n        costProvenance: outcome.costProvenance,\n        tokenUsage: outcome.tokenUsage,\n        terminalOutcome: 'succeeded',\n        judgeMetadata: outcome.judgeMetadata,\n        outcome: recordOutcome,\n        ...(outcome.failureClass ? { failureClass: outcome.failureClass } : {}),\n        failureMode: outcome.failureMode,\n        splitTag,\n        scenarioId: cell.scenario.scenarioId,\n      }\n      const profileSource =\n        outcome.agentProfile ??\n        (typeof opts.agentProfile === 'function'\n          ? await opts.agentProfile({\n              campaignId: opts.campaignId,\n              runId,\n              variantId: cell.variant.id,\n              scenarioId: cell.scenario.scenarioId,\n              seed: cell.seed,\n              variant: cell.variant.payload,\n              scenarioTags: cell.scenario.tags ?? {},\n            })\n          : opts.agentProfile)\n      if (profileSource !== undefined) {\n        const agentProfile = await resolveAgentProfileCell(profileSource)\n        assertAgentProfileMatchesRun(agentProfile, outcome.model, outcome.promptHash)\n        record.agentProfile = agentProfile\n      }\n      return { record: validateRunRecord(record), integrity: integrityReport }\n    } finally {\n      // This run's worker has finished with it (success, run-level failure, or\n      // genuine error). It is no longer the pool's job to finalize — drop it so\n      // the post-settle sweep doesn't double-abort a finalized run.\n      openRuns.delete(runId)\n    }\n  }\n\n  const workers = Array.from({ length: Math.min(concurrency, cells.length) }, () => worker())\n  // allSettled (not all): a genuine error in one worker must not reject the\n  // pool mid-flight and orphan the others. Each worker captures its own\n  // genuine error into `genuineErrors` and returns; we re-throw below.\n  await Promise.allSettled(workers)\n\n  // Finalize any run still open after the pool wound down. With the\n  // stop-dispatch flag these are the runs that were mid-flight in peer workers\n  // when the first genuine error fired — abort them so their emitters finalize\n  // (hooks fire, store records a terminal status) instead of leaking.\n  for (const [runId, emitter] of openRuns) {\n    await finalizeAbort(emitter, runId, 'campaign aborted: genuine error in a sibling run')\n  }\n  openRuns.clear()\n\n  if (genuineErrors.length > 0) {\n    throw genuineErrors.length === 1\n      ? genuineErrors[0]\n      : new AggregateError(\n          genuineErrors,\n          `runEvalCampaign: ${genuineErrors.length} runs failed with genuine (non-run-level) errors`,\n        )\n  }\n\n  // ── Optional research report ───────────────────────────────────────\n  let report: ResearchReport | undefined\n  if (opts.report) {\n    const reportOpts: ResearchReportOptions = {\n      ...opts.report,\n      comparator: opts.report.comparator,\n      split: splitTag === 'dev' ? 'search' : splitTag,\n      generatedAt: new Date(now()).toISOString(),\n      preregistrationHash: preregistrationHash ?? undefined,\n    }\n    report = await researchReport(runs, reportOpts)\n  }\n\n  const endedAt = new Date(now()).toISOString()\n\n  return {\n    campaignId: opts.campaignId,\n    campaignFingerprint,\n    preregistrationHash,\n    runs,\n    integrityReports,\n    failedRuns,\n    report,\n    startedAt,\n    endedAt,\n  }\n}\n\n// ── Internal ─────────────────────────────────────────────────────────────\n\nclass CellExecutionError extends Error {\n  readonly failed: FailedRun\n  readonly integrity?: RunIntegrityReport\n  constructor(failed: FailedRun, integrity?: RunIntegrityReport) {\n    super(`cell ${failed.variantId}/${failed.scenarioId}@${failed.seed} failed: ${failed.reason}`)\n    this.failed = failed\n    this.integrity = integrity\n  }\n}\n\n/**\n * Abort a run whose owning work threw or was orphaned by a sibling's genuine\n * error, finalizing the emitter so hooks fire and the store records a terminal\n * status. Safe to call unconditionally: it only aborts runs that are still\n * `running`. Two no-op cases are intentional and benign:\n *\n *   - the run was never started (absent from the store) — nothing to finalize\n *   - the run is already terminal (`completed` / `failed` / `aborted`) —\n *     finalizing again would overwrite the real outcome (e.g. flip a passed run\n *     to `aborted` with `{ pass: false, notes: reason }`), so we leave it alone\n *\n * Any OTHER failure of the store read or the abort write (disk full, FS fault,\n * backend down) is a genuine diagnostic and propagates rather than being\n * swallowed.\n */\nexport async function finalizeAbort(\n  emitter: TraceEmitter,\n  runId: string,\n  reason: string,\n): Promise<void> {\n  const existing = await emitter.traceStore.getRun(runId)\n  if (existing === undefined) return // run never started; nothing to abort\n  if (existing.status !== 'running') return // already finalized; never overwrite a real outcome\n  await emitter.abortRun(reason)\n}\n\nfunction defaultRawSinkFactory(workDir: string | undefined) {\n  return (params: CampaignFactoryParams): RawProviderSink => {\n    if (!workDir) {\n      throw new Error(\n        'runEvalCampaign: rawSinkFactory not supplied and workDir not set. Pass either to enable raw provider capture, or pass `new NoopRawProviderSink()` via rawSinkFactory to opt out explicitly.',\n      )\n    }\n    return new FileSystemRawProviderSink({\n      dir: `${workDir}/raw-events/${params.runId}`,\n    })\n  }\n}\n\nasync function resolveAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  if (isAgentProfileCell(input)) {\n    if (!(await verifyAgentProfileCell(input))) {\n      throw new Error(`runEvalCampaign: agentProfile.cellId does not match its content`)\n    }\n    return input\n  }\n  return buildAgentProfileCell(input)\n}\n\nfunction isAgentProfileCell(\n  input: AgentProfileCell | AgentProfileCellInput,\n): input is AgentProfileCell {\n  return 'schemaVersion' in input && 'cellId' in input\n}\n\nfunction assertAgentProfileMatchesRun(\n  profile: AgentProfileCell,\n  model: string,\n  promptHash: string,\n): void {\n  if (profile.model !== undefined && profile.model !== model) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.model \"${profile.model}\" does not match outcome.model \"${model}\"`,\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== promptHash) {\n    throw new Error(\n      `runEvalCampaign: agentProfile.promptHash \"${profile.promptHash}\" does not match outcome.promptHash \"${promptHash}\"`,\n    )\n  }\n}\n\nfunction defaultRunId(params: CampaignFactoryParams): string {\n  // Stable across re-runs: fingerprint of (campaignId, variantId, scenarioId, seed).\n  // Caller can override via opts.runId for non-deterministic IDs.\n  const base = `${params.campaignId}::${params.variantId}::${params.scenarioId}::${params.seed}`\n  // Lightweight hex: we don't need crypto-grade here, just stability + uniqueness.\n  let h1 = 0x811c9dc5\n  let h2 = 0x12345678\n  for (let i = 0; i < base.length; i++) {\n    const c = base.charCodeAt(i)\n    h1 = Math.imul(h1 ^ c, 0x01000193) >>> 0\n    h2 = Math.imul(h2 ^ c, 0x9e3779b1) >>> 0\n  }\n  return `run-${h1.toString(16).padStart(8, '0')}${h2.toString(16).padStart(8, '0')}`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAsSA,IAAM,oBAA8C;AAAA,EAClD,aAAa;AAAA,EACb,8BAA8B;AAAA,EAC9B,gBAAgB;AAClB;AAEA,IAAM,gBAAsC;AAAA,EAC1C,wBAAwB;AAAA,EACxB,aAAa;AACf;AAEA,eAAsB,gBACpB,MAC6B;AAE7B,iBAAe,KAAK,SAAS,KAAK,qBAAqB,aAAa;AAEpE,MAAI,KAAK,SAAS,WAAW,GAAG;AAC9B,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AACA,MAAI,KAAK,UAAU,WAAW,GAAG;AAC/B,UAAM,IAAI,MAAM,+CAA+C;AAAA,EACjE;AACA,QAAM,aAAa,oBAAI,IAAY;AACnC,aAAW,KAAK,KAAK,UAAU;AAC7B,QAAI,WAAW,IAAI,EAAE,EAAE,GAAG;AACxB,YAAM,IAAI,MAAM,0CAA0C,EAAE,EAAE,IAAI;AAAA,IACpE;AACA,eAAW,IAAI,EAAE,EAAE;AAAA,EACrB;AACA,QAAM,cAAc,oBAAI,IAAY;AACpC,aAAW,KAAK,KAAK,WAAW;AAC9B,QAAI,YAAY,IAAI,EAAE,UAAU,GAAG;AACjC,YAAM,IAAI,MAAM,0CAA0C,EAAE,UAAU,IAAI;AAAA,IAC5E;AACA,gBAAY,IAAI,EAAE,UAAU;AAAA,EAC9B;AACA,MAAI,KAAK,QAAQ,cAAc,CAAC,WAAW,IAAI,KAAK,OAAO,UAAU,GAAG;AACtE,UAAM,IAAI;AAAA,MACR,uCAAuC,KAAK,OAAO,UAAU;AAAA,IAC/D;AAAA,EACF;AACA,MAAI,CAAC,KAAK,WAAW;AACnB,UAAM,IAAI,MAAM,oEAAoE;AAAA,EACtF;AAEA,QAAM,QAAQ,KAAK,SAAS,CAAC,GAAG,GAAG,CAAC;AACpC,QAAM,WAAwB,KAAK,YAAY;AAC/C,QAAM,cAAc,KAAK,IAAI,GAAG,KAAK,eAAe,CAAC;AACrD,QAAM,YAAY,EAAE,GAAG,mBAAmB,GAAI,KAAK,aAAa,CAAC,EAAG;AACpE,QAAM,qBAA8C,KAAK,sBAAsB;AAC/E,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,WAAW,KAAK,QAAQ,WAAW,IAAI,QAAQ,QAAQ,EAAE;AAC/D,QAAM,WAAW,KAAK,QAAQ,YAAY;AAC1C,QAAM,sBAAsB,KAAK,uBAAuB;AAExD,QAAM,iBAAiB,KAAK,kBAAkB,sBAAsB,KAAK,OAAO;AAGhF,QAAM,sBAAsB,MAAM;AAAA,IAChC,aAAa;AAAA,MACX,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK;AAAA,MAC9C,WAAW,KAAK,UAAU,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,KAAK;AAAA,MACxD,OAAO,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAAA,MACtC;AAAA,MACA,YAAY,KAAK,QAAQ,cAAc;AAAA,MACvC;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAIA,QAAM,QAAgB,CAAC;AACvB,aAAW,WAAW,KAAK,UAAU;AACnC,eAAW,YAAY,KAAK,WAAW;AACrC,iBAAW,QAAQ,OAAO;AACxB,cAAM,KAAK,EAAE,SAAS,UAAU,KAAK,CAAC;AAAA,MACxC;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAC9C,QAAM,OAAoB,CAAC;AAC3B,QAAM,mBAAyC,CAAC;AAChD,QAAM,aAA0B,CAAC;AASjC,MAAI,SAAS;AACb,MAAI,WAAW;AACf,QAAM,gBAA2B,CAAC;AAKlC,QAAM,WAAW,oBAAI,IAA0B;AAE/C,iBAAe,SAAwB;AACrC,WAAO,CAAC,UAAU;AAChB,YAAM,IAAI;AACV,UAAI,KAAK,MAAM,OAAQ;AACvB,YAAM,OAAO,MAAM,CAAC;AACpB,UAAI;AACF,cAAM,SAAS,MAAM,WAAW,IAAI;AACpC,aAAK,KAAK,OAAO,MAAM;AACvB,yBAAiB,KAAK,OAAO,SAAS;AAAA,MACxC,SAAS,KAAK;AACZ,YAAI,eAAe,oBAAoB;AACrC,qBAAW,KAAK,IAAI,MAAM;AAC1B,cAAI,IAAI,UAAW,kBAAiB,KAAK,IAAI,SAAS;AAAA,QACxD,OAAO;AAKL,wBAAc,KAAK,GAAG;AACtB,qBAAW;AACX;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,iBAAe,WACb,MAC+D;AAC/D,UAAM,SAAS,KAAK,SAAS,cAAc;AAAA,MACzC,YAAY,KAAK;AAAA,MACjB,OAAO;AAAA;AAAA,MACP,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,MAAM,KAAK;AAAA,IACb,CAAC;AACD,UAAM,gBAAuC;AAAA,MAC3C,YAAY,KAAK;AAAA,MACjB;AAAA,MACA,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,MAAM,KAAK;AAAA,IACb;AACA,UAAM,QAAQ,KAAK,aAAa,aAAa;AAC7C,UAAM,UAAU,eAAe,aAAa;AAE5C,UAAM,UAAU,IAAI,aAAa,OAAO;AAAA,MACtC;AAAA,MACA,KAAK,KAAK;AAAA,MACV,eAAe,KAAK;AAAA,IACtB,CAAC;AAGD,aAAS,IAAI,OAAO,OAAO;AAE3B,UAAM,UAA4B;AAAA,MAChC,GAAG,KAAK;AAAA,MACR;AAAA,MACA,cAAc,EAAE,MAAM;AAAA,IACxB;AAEA,UAAM,MAA6B;AAAA,MACjC;AAAA,MACA,cAAc,KAAK;AAAA,MACnB,SAAS,KAAK,QAAQ;AAAA,MACtB,WAAW,KAAK,QAAQ;AAAA,MACxB,YAAY,KAAK,SAAS;AAAA,MAC1B,cAAc,KAAK,SAAS,QAAQ,CAAC;AAAA,MACrC,MAAM,KAAK;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAEA,QAAI;AACF,YAAM,YAAY,IAAI;AACtB,UAAI;AACJ,UAAI;AACF,kBAAU,MAAM,KAAK,OAAO,GAAG;AAAA,MACjC,SAAS,KAAK;AACZ,cAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAM/D,cAAM,cAAc,SAAS,OAAO,OAAO;AAC3C,cAAM,IAAI,mBAAmB;AAAA,UAC3B;AAAA,UACA,WAAW,KAAK,QAAQ;AAAA,UACxB,YAAY,KAAK,SAAS;AAAA,UAC1B,MAAM,KAAK;AAAA,UACX,QAAQ;AAAA,UACR,OAAO;AAAA,QACT,CAAC;AAAA,MACH;AACA,YAAM,SAAS,IAAI,IAAI;AAEvB,YAAM,kBAAkB,MAAM,kBAAkB,OAAO,OAAO,EAAE,GAAG,WAAW,QAAQ,CAAC;AACvF,UAAI,CAAC,gBAAgB,IAAI;AACvB,gBAAQ,oBAAoB;AAAA,UAC1B,KAAK;AACH,kBAAM,IAAI,kBAAkB,eAAe;AAAA,UAC7C,KAAK;AACH,kBAAM,IAAI;AAAA,cACR;AAAA,gBACE;AAAA,gBACA,WAAW,KAAK,QAAQ;AAAA,gBACxB,YAAY,KAAK,SAAS;AAAA,gBAC1B,MAAM,KAAK;AAAA,gBACX,QAAQ;AAAA,gBACR,OAAO,gBAAgB,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,IAAI;AAAA,cAC5D;AAAA,cACA;AAAA,YACF;AAAA,UACF,KAAK;AAEH;AAAA,QACJ;AAAA,MACF;AAEA,YAAM,gBAA4B;AAAA,QAChC,KAAK,QAAQ,OAAO,CAAC;AAAA,MACvB;AACA,UAAI,aAAa,UAAW,eAAc,eAAe,QAAQ;AAAA,UAC5D,eAAc,cAAc,QAAQ;AACzC,UAAI,QAAQ,gBAAgB,OAAW,eAAc,cAAc,QAAQ;AAE3E,YAAM,SAAoB;AAAA,QACxB;AAAA,QACA,cAAc,KAAK;AAAA,QACnB,aAAa,KAAK,QAAQ;AAAA,QAC1B,MAAM,KAAK;AAAA,QACX,OAAO,QAAQ;AAAA,QACf,YAAY,QAAQ;AAAA,QACpB,YAAY,QAAQ;AAAA,QACpB,WAAW,KAAK;AAAA,QAChB;AAAA,QACA,SAAS,QAAQ;AAAA,QACjB,gBAAgB,QAAQ;AAAA,QACxB,YAAY,QAAQ;AAAA,QACpB,iBAAiB;AAAA,QACjB,eAAe,QAAQ;AAAA,QACvB,SAAS;AAAA,QACT,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;AAAA,QACrE,aAAa,QAAQ;AAAA,QACrB;AAAA,QACA,YAAY,KAAK,SAAS;AAAA,MAC5B;AACA,YAAM,gBACJ,QAAQ,iBACP,OAAO,KAAK,iBAAiB,aAC1B,MAAM,KAAK,aAAa;AAAA,QACtB,YAAY,KAAK;AAAA,QACjB;AAAA,QACA,WAAW,KAAK,QAAQ;AAAA,QACxB,YAAY,KAAK,SAAS;AAAA,QAC1B,MAAM,KAAK;AAAA,QACX,SAAS,KAAK,QAAQ;AAAA,QACtB,cAAc,KAAK,SAAS,QAAQ,CAAC;AAAA,MACvC,CAAC,IACD,KAAK;AACX,UAAI,kBAAkB,QAAW;AAC/B,cAAM,eAAe,MAAM,wBAAwB,aAAa;AAChE,qCAA6B,cAAc,QAAQ,OAAO,QAAQ,UAAU;AAC5E,eAAO,eAAe;AAAA,MACxB;AACA,aAAO,EAAE,QAAQ,kBAAkB,MAAM,GAAG,WAAW,gBAAgB;AAAA,IACzE,UAAE;AAIA,eAAS,OAAO,KAAK;AAAA,IACvB;AAAA,EACF;AAEA,QAAM,UAAU,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,aAAa,MAAM,MAAM,EAAE,GAAG,MAAM,OAAO,CAAC;AAI1F,QAAM,QAAQ,WAAW,OAAO;AAMhC,aAAW,CAAC,OAAO,OAAO,KAAK,UAAU;AACvC,UAAM,cAAc,SAAS,OAAO,kDAAkD;AAAA,EACxF;AACA,WAAS,MAAM;AAEf,MAAI,cAAc,SAAS,GAAG;AAC5B,UAAM,cAAc,WAAW,IAC3B,cAAc,CAAC,IACf,IAAI;AAAA,MACF;AAAA,MACA,oBAAoB,cAAc,MAAM;AAAA,IAC1C;AAAA,EACN;AAGA,MAAI;AACJ,MAAI,KAAK,QAAQ;AACf,UAAM,aAAoC;AAAA,MACxC,GAAG,KAAK;AAAA,MACR,YAAY,KAAK,OAAO;AAAA,MACxB,OAAO,aAAa,QAAQ,WAAW;AAAA,MACvC,aAAa,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAAA,MACzC,qBAAqB,uBAAuB;AAAA,IAC9C;AACA,aAAS,MAAM,eAAe,MAAM,UAAU;AAAA,EAChD;AAEA,QAAM,UAAU,IAAI,KAAK,IAAI,CAAC,EAAE,YAAY;AAE5C,SAAO;AAAA,IACL,YAAY,KAAK;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAIA,IAAM,qBAAN,cAAiC,MAAM;AAAA,EAC5B;AAAA,EACA;AAAA,EACT,YAAY,QAAmB,WAAgC;AAC7D,UAAM,QAAQ,OAAO,SAAS,IAAI,OAAO,UAAU,IAAI,OAAO,IAAI,YAAY,OAAO,MAAM,EAAE;AAC7F,SAAK,SAAS;AACd,SAAK,YAAY;AAAA,EACnB;AACF;AAiBA,eAAsB,cACpB,SACA,OACA,QACe;AACf,QAAM,WAAW,MAAM,QAAQ,WAAW,OAAO,KAAK;AACtD,MAAI,aAAa,OAAW;AAC5B,MAAI,SAAS,WAAW,UAAW;AACnC,QAAM,QAAQ,SAAS,MAAM;AAC/B;AAEA,SAAS,sBAAsB,SAA6B;AAC1D,SAAO,CAAC,WAAmD;AACzD,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AACA,WAAO,IAAI,0BAA0B;AAAA,MACnC,KAAK,GAAG,OAAO,eAAe,OAAO,KAAK;AAAA,IAC5C,CAAC;AAAA,EACH;AACF;AAEA,eAAe,wBACb,OAC2B;AAC3B,MAAI,mBAAmB,KAAK,GAAG;AAC7B,QAAI,CAAE,MAAM,uBAAuB,KAAK,GAAI;AAC1C,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AACA,WAAO;AAAA,EACT;AACA,SAAO,sBAAsB,KAAK;AACpC;AAEA,SAAS,mBACP,OAC2B;AAC3B,SAAO,mBAAmB,SAAS,YAAY;AACjD;AAEA,SAAS,6BACP,SACA,OACA,YACM;AACN,MAAI,QAAQ,UAAU,UAAa,QAAQ,UAAU,OAAO;AAC1D,UAAM,IAAI;AAAA,MACR,wCAAwC,QAAQ,KAAK,mCAAmC,KAAK;AAAA,IAC/F;AAAA,EACF;AACA,MAAI,QAAQ,eAAe,UAAa,QAAQ,eAAe,YAAY;AACzE,UAAM,IAAI;AAAA,MACR,6CAA6C,QAAQ,UAAU,wCAAwC,UAAU;AAAA,IACnH;AAAA,EACF;AACF;AAEA,SAAS,aAAa,QAAuC;AAG3D,QAAM,OAAO,GAAG,OAAO,UAAU,KAAK,OAAO,SAAS,KAAK,OAAO,UAAU,KAAK,OAAO,IAAI;AAE5F,MAAI,KAAK;AACT,MAAI,KAAK;AACT,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,IAAI,KAAK,WAAW,CAAC;AAC3B,SAAK,KAAK,KAAK,KAAK,GAAG,QAAU,MAAM;AACvC,SAAK,KAAK,KAAK,KAAK,GAAG,UAAU,MAAM;AAAA,EACzC;AACA,SAAO,OAAO,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC,GAAG,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC;AACnF;","names":[]}

package/dist/{chunk-LUNF2SEL.js → chunk-DJKY2TSY.js}

@@ -10,7 +10,7 @@ import {
   costReceiptFromLlm,
   costReceiptFromLlmError,
   maximumChargeForLlmRequest
-} from "./chunk-J5SQWP6Y.js";
+} from "./chunk-PBE2LOSS.js";
 import {
   CostLedger
 } from "./chunk-WS3NZZQQ.js";
@@ -18,7 +18,7 @@ import {
   buildTraceAnalystTools,
   runTraceAnalysisLoop,
   spanEpochMillis
-} from "./chunk-IR3KBHOY.js";
+} from "./chunk-P6FYH6K4.js";
 import {
   LLM_CONTEXT_TOKENS,
   LLM_INPUT_TOKEN_ATTR_KEYS,
@@ -2425,4 +2425,4 @@ export {
   AnalystRegistry,
   buildDefaultAnalystRegistry
 };
-//# sourceMappingURL=chunk-LUNF2SEL.js.map
+//# sourceMappingURL=chunk-DJKY2TSY.js.map

package/dist/chunk-DPUHNQLN.js

@@ -0,0 +1,232 @@
+import {
+  mcnemar,
+  pairedBootstrap,
+  pairedRiskDifference,
+  wilcoxonSignedRank
+} from "./chunk-MHELPNRP.js";
+import {
+  ValidationError
+} from "./chunk-ONWEPEDO.js";
+
+// src/paired-arms.ts
+function pairArms(rows, opts) {
+  const { baselineArm, treatmentArm } = opts;
+  if (baselineArm === treatmentArm) {
+    throw new ValidationError(
+      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' \u2014 an arm cannot be compared to itself`
+    );
+  }
+  const byArm = /* @__PURE__ */ new Map();
+  const armsSeen = /* @__PURE__ */ new Set();
+  for (const row of rows) {
+    armsSeen.add(row.arm);
+    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue;
+    const byKey = byArm.get(row.arm) ?? /* @__PURE__ */ new Map();
+    const group = byKey.get(row.pairKey) ?? [];
+    group.push(row);
+    byKey.set(row.pairKey, group);
+    byArm.set(row.arm, byKey);
+  }
+  for (const arm of [baselineArm, treatmentArm]) {
+    if (!byArm.has(arm)) {
+      const seen = [...armsSeen].sort().join(", ") || "<none>";
+      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`);
+    }
+  }
+  const baselineByKey = byArm.get(baselineArm);
+  const treatmentByKey = byArm.get(treatmentArm);
+  const allKeys = [.../* @__PURE__ */ new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort();
+  const pairs = [];
+  const unpairedBaseline = [];
+  const unpairedTreatment = [];
+  for (const pairKey of allKeys) {
+    const b = baselineByKey.get(pairKey) ?? [];
+    const t = treatmentByKey.get(pairKey) ?? [];
+    if (b.length <= 1 && t.length <= 1) {
+      if (b.length === 1 && t.length === 1) {
+        const baseline = b[0];
+        const treatment = t[0];
+        if (baseline.repKey !== void 0 || treatment.repKey !== void 0) {
+          if (baseline.repKey === void 0 || treatment.repKey === void 0 || baseline.repKey !== treatment.repKey) {
+            unpairedBaseline.push(baseline);
+            unpairedTreatment.push(treatment);
+            continue;
+          }
+        }
+        pairs.push({ pairKey, repIndex: 0, baseline, treatment });
+      } else {
+        unpairedBaseline.push(...b);
+        unpairedTreatment.push(...t);
+      }
+      continue;
+    }
+    const bByRep = indexByRepKey(b, pairKey, baselineArm);
+    const tByRep = indexByRepKey(t, pairKey, treatmentArm);
+    const repKeys = [.../* @__PURE__ */ new Set([...bByRep.keys(), ...tByRep.keys()])].sort();
+    let repIndex = 0;
+    for (const repKey of repKeys) {
+      const baseline = bByRep.get(repKey);
+      const treatment = tByRep.get(repKey);
+      if (baseline !== void 0 && treatment !== void 0) {
+        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment });
+      } else if (baseline !== void 0) {
+        unpairedBaseline.push(baseline);
+      } else if (treatment !== void 0) {
+        unpairedTreatment.push(treatment);
+      }
+    }
+  }
+  return { pairs, unpairedBaseline, unpairedTreatment };
+}
+function indexByRepKey(group, pairKey, arm) {
+  const byRep = /* @__PURE__ */ new Map();
+  for (const row of group) {
+    if (row.repKey === void 0) {
+      throw new ValidationError(
+        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' is missing repKey \u2014 multi-rep items require an explicit repKey on every row so reps pair by identity (pairing reps by outcome or by index would bias the paired statistics)`
+      );
+    }
+    if (byRep.has(row.repKey)) {
+      throw new ValidationError(
+        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' \u2014 (pairKey, repKey) must uniquely identify a rep within an arm`
+      );
+    }
+    byRep.set(row.repKey, row);
+  }
+  return byRep;
+}
+function comparePairedArms(rows, opts) {
+  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts);
+  let correctness = null;
+  const baselinePass = [];
+  const treatmentPass = [];
+  for (const pair of pairs) {
+    if (pair.baseline.pass === void 0 || pair.treatment.pass === void 0) continue;
+    baselinePass.push(pair.baseline.pass ? 1 : 0);
+    treatmentPass.push(pair.treatment.pass ? 1 : 0);
+  }
+  if (baselinePass.length > 0) {
+    const mc = mcnemar(baselinePass, treatmentPass);
+    correctness = {
+      b10: mc.b,
+      b01: mc.c,
+      mcnemar: mc,
+      riskDifference: pairedRiskDifference(baselinePass, treatmentPass)
+    };
+  }
+  const metricNames = opts.metricNames ?? [
+    ...new Set(
+      pairs.flatMap((p) => [
+        ...Object.keys(p.baseline.metrics ?? {}),
+        ...Object.keys(p.treatment.metrics ?? {})
+      ])
+    )
+  ].sort();
+  const metricDeltas = metricNames.map((name) => {
+    const before = [];
+    const after = [];
+    let nMissing = 0;
+    for (const pair of pairs) {
+      const b = metricValue(pair.baseline, name);
+      const t = metricValue(pair.treatment, name);
+      if (b === void 0 || t === void 0) {
+        nMissing++;
+        continue;
+      }
+      before.push(b);
+      after.push(t);
+    }
+    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap);
+    return {
+      name,
+      n: before.length,
+      nMissing,
+      medianDelta: bootstrapCi?.median ?? null,
+      meanDelta: bootstrapCi?.mean ?? null,
+      bootstrapCi,
+      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after)
+    };
+  });
+  return {
+    nPairs: pairs.length,
+    nUnpairedBaseline: unpairedBaseline.length,
+    nUnpairedTreatment: unpairedTreatment.length,
+    correctness,
+    metricDeltas
+  };
+}
+function pairRunRecords(baselineRuns, treatmentRuns) {
+  const baselineRows = runRecordArmRows(baselineRuns, "baseline");
+  const treatmentRows = runRecordArmRows(treatmentRuns, "treatment");
+  validateRunRecordArmRows(baselineRows, "baseline");
+  validateRunRecordArmRows(treatmentRows, "treatment");
+  if (baselineRows.length === 0 || treatmentRows.length === 0) {
+    return {
+      pairs: [],
+      unpairedBaseline: baselineRows.map((row) => row.run),
+      unpairedTreatment: treatmentRows.map((row) => row.run)
+    };
+  }
+  const result = pairArms([...baselineRows, ...treatmentRows], {
+    baselineArm: "baseline",
+    treatmentArm: "treatment"
+  });
+  return {
+    pairs: result.pairs.map((pair) => {
+      const baseline = pair.baseline;
+      const treatment = pair.treatment;
+      return {
+        pairKey: pair.pairKey,
+        repKey: baseline.repKey,
+        baseline: baseline.run,
+        treatment: treatment.run
+      };
+    }),
+    unpairedBaseline: result.unpairedBaseline.map((row) => row.run),
+    unpairedTreatment: result.unpairedTreatment.map((row) => row.run)
+  };
+}
+function runRecordArmRows(runs, arm) {
+  return runs.map((run) => {
+    const scenarioId = run.scenarioId.trim();
+    if (!scenarioId) {
+      throw new ValidationError(
+        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`
+      );
+    }
+    return {
+      pairKey: JSON.stringify([run.experimentId, scenarioId]),
+      repKey: String(run.seed),
+      arm,
+      run
+    };
+  });
+}
+function validateRunRecordArmRows(rows, arm) {
+  const byPairKey = /* @__PURE__ */ new Map();
+  for (const row of rows) {
+    const group = byPairKey.get(row.pairKey) ?? [];
+    group.push(row);
+    byPairKey.set(row.pairKey, group);
+  }
+  for (const [pairKey, group] of byPairKey) {
+    if (group.length > 1) indexByRepKey(group, pairKey, arm);
+  }
+}
+function metricValue(row, name) {
+  const v = row.metrics?.[name];
+  if (v === void 0) return void 0;
+  if (!Number.isFinite(v)) {
+    throw new ValidationError(
+      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`
+    );
+  }
+  return v;
+}
+
+export {
+  pairArms,
+  comparePairedArms,
+  pairRunRecords
+};
+//# sourceMappingURL=chunk-DPUHNQLN.js.map

package/dist/chunk-DPUHNQLN.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/paired-arms.ts"],"sourcesContent":["/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;AAuGO,SAAS,SAAS,MAA+B,MAAuC;AAC7F,QAAM,EAAE,aAAa,aAAa,IAAI;AACtC,MAAI,gBAAgB,cAAc;AAChC,UAAM,IAAI;AAAA,MACR,oDAAoD,WAAW;AAAA,IACjE;AAAA,EACF;AAGA,QAAM,QAAQ,oBAAI,IAAyC;AAC3D,QAAM,WAAW,oBAAI,IAAY;AACjC,aAAW,OAAO,MAAM;AACtB,aAAS,IAAI,IAAI,GAAG;AACpB,QAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,aAAc;AACzD,UAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,KAAK,oBAAI,IAA4B;AACpE,UAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;AACzC,UAAM,KAAK,GAAG;AACd,UAAM,IAAI,IAAI,SAAS,KAAK;AAC5B,UAAM,IAAI,IAAI,KAAK,KAAK;AAAA,EAC1B;AAEA,aAAW,OAAO,CAAC,aAAa,YAAY,GAAG;AAC7C,QAAI,CAAC,MAAM,IAAI,GAAG,GAAG;AACnB,YAAM,OAAO,CAAC,GAAG,QAAQ,EAAE,KAAK,EAAE,KAAK,IAAI,KAAK;AAChD,YAAM,IAAI,gBAAgB,8BAA8B,GAAG,oBAAoB,IAAI,GAAG;AAAA,IACxF;AAAA,EACF;AAEA,QAAM,gBAAgB,MAAM,IAAI,WAAW;AAC3C,QAAM,iBAAiB,MAAM,IAAI,YAAY;AAE7C,QAAM,UAAU,CAAC,GAAG,oBAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,EAAE,KAAK;AACvF,QAAM,QAAuB,CAAC;AAC9B,QAAM,mBAAmC,CAAC;AAC1C,QAAM,oBAAoC,CAAC;AAC3C,aAAW,WAAW,SAAS;AAC7B,UAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;AACzC,UAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;AAE1C,QAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;AAClC,UAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;AACpC,cAAM,WAAW,EAAE,CAAC;AACpB,cAAM,YAAY,EAAE,CAAC;AACrB,YAAI,SAAS,WAAW,UAAa,UAAU,WAAW,QAAW;AACnE,cACE,SAAS,WAAW,UACpB,UAAU,WAAW,UACrB,SAAS,WAAW,UAAU,QAC9B;AACA,6BAAiB,KAAK,QAAQ;AAC9B,8BAAkB,KAAK,SAAS;AAChC;AAAA,UACF;AAAA,QACF;AACA,cAAM,KAAK,EAAE,SAAS,UAAU,GAAG,UAAU,UAAU,CAAC;AAAA,MAC1D,OAAO;AACL,yBAAiB,KAAK,GAAG,CAAC;AAC1B,0BAAkB,KAAK,GAAG,CAAC;AAAA,MAC7B;AACA;AAAA,IACF;AAEA,UAAM,SAAS,cAAc,GAAG,SAAS,WAAW;AACpD,UAAM,SAAS,cAAc,GAAG,SAAS,YAAY;AACrD,UAAM,UAAU,CAAC,GAAG,oBAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,KAAK;AACxE,QAAI,WAAW;AACf,eAAW,UAAU,SAAS;AAC5B,YAAM,WAAW,OAAO,IAAI,MAAM;AAClC,YAAM,YAAY,OAAO,IAAI,MAAM;AACnC,UAAI,aAAa,UAAa,cAAc,QAAW;AACrD,cAAM,KAAK,EAAE,SAAS,UAAU,YAAY,UAAU,UAAU,CAAC;AAAA,MACnE,WAAW,aAAa,QAAW;AACjC,yBAAiB,KAAK,QAAQ;AAAA,MAChC,WAAW,cAAc,QAAW;AAClC,0BAAkB,KAAK,SAAS;AAAA,MAClC;AAAA,IACF;AAAA,EACF;AAEA,SAAO,EAAE,OAAO,kBAAkB,kBAAkB;AACtD;AAIA,SAAS,cACP,OACA,SACA,KAC2B;AAC3B,QAAM,QAAQ,oBAAI,IAA0B;AAC5C,aAAW,OAAO,OAAO;AACvB,QAAI,IAAI,WAAW,QAAW;AAC5B,YAAM,IAAI;AAAA,QACR,sBAAsB,OAAO,oDAAoD,GAAG;AAAA,MAGtF;AAAA,IACF;AACA,QAAI,MAAM,IAAI,IAAI,MAAM,GAAG;AACzB,YAAM,IAAI;AAAA,QACR,+BAA+B,IAAI,MAAM,kBAAkB,OAAO,aAAa,GAAG;AAAA,MAEpF;AAAA,IACF;AACA,UAAM,IAAI,IAAI,QAAQ,GAAG;AAAA,EAC3B;AACA,SAAO;AACT;AAiEO,SAAS,kBACd,MACA,MACsB;AACtB,QAAM,EAAE,OAAO,kBAAkB,kBAAkB,IAAI,SAAS,MAAM,IAAI;AAE1E,MAAI,cAAwC;AAC5C,QAAM,eAAyB,CAAC;AAChC,QAAM,gBAA0B,CAAC;AACjC,aAAW,QAAQ,OAAO;AACxB,QAAI,KAAK,SAAS,SAAS,UAAa,KAAK,UAAU,SAAS,OAAW;AAC3E,iBAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;AAC5C,kBAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;AAAA,EAChD;AACA,MAAI,aAAa,SAAS,GAAG;AAC3B,UAAM,KAAK,QAAQ,cAAc,aAAa;AAC9C,kBAAc;AAAA,MACZ,KAAK,GAAG;AAAA,MACR,KAAK,GAAG;AAAA,MACR,SAAS;AAAA,MACT,gBAAgB,qBAAqB,cAAc,aAAa;AAAA,IAClE;AAAA,EACF;AAEA,QAAM,cACJ,KAAK,eACL;AAAA,IACE,GAAG,IAAI;AAAA,MACL,MAAM,QAAQ,CAAC,MAAM;AAAA,QACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC;AAAA,QACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC;AAAA,MAC1C,CAAC;AAAA,IACH;AAAA,EACF,EAAE,KAAK;AAET,QAAM,eAAoC,YAAY,IAAI,CAAC,SAAS;AAClE,UAAM,SAAmB,CAAC;AAC1B,UAAM,QAAkB,CAAC;AACzB,QAAI,WAAW;AACf,eAAW,QAAQ,OAAO;AACxB,YAAM,IAAI,YAAY,KAAK,UAAU,IAAI;AACzC,YAAM,IAAI,YAAY,KAAK,WAAW,IAAI;AAC1C,UAAI,MAAM,UAAa,MAAM,QAAW;AACtC;AACA;AAAA,MACF;AACA,aAAO,KAAK,CAAC;AACb,YAAM,KAAK,CAAC;AAAA,IACd;AACA,UAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;AAC9F,WAAO;AAAA,MACL;AAAA,MACA,GAAG,OAAO;AAAA,MACV;AAAA,MACA,aAAa,aAAa,UAAU;AAAA,MACpC,WAAW,aAAa,QAAQ;AAAA,MAChC;AAAA,MACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;AAAA,IACzE;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd,mBAAmB,iBAAiB;AAAA,IACpC,oBAAoB,kBAAkB;AAAA,IACtC;AAAA,IACA;AAAA,EACF;AACF;AA2BO,SAAS,eACd,cACA,eACsB;AACtB,QAAM,eAAe,iBAAiB,cAAc,UAAU;AAC9D,QAAM,gBAAgB,iBAAiB,eAAe,WAAW;AACjE,2BAAyB,cAAc,UAAU;AACjD,2BAAyB,eAAe,WAAW;AACnD,MAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GAAG;AAC3D,WAAO;AAAA,MACL,OAAO,CAAC;AAAA,MACR,kBAAkB,aAAa,IAAI,CAAC,QAAQ,IAAI,GAAG;AAAA,MACnD,mBAAmB,cAAc,IAAI,CAAC,QAAQ,IAAI,GAAG;AAAA,IACvD;AAAA,EACF;AAEA,QAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;AAAA,IAC3D,aAAa;AAAA,IACb,cAAc;AAAA,EAChB,CAAC;AACD,SAAO;AAAA,IACL,OAAO,OAAO,MAAM,IAAI,CAAC,SAAS;AAChC,YAAM,WAAW,KAAK;AACtB,YAAM,YAAY,KAAK;AACvB,aAAO;AAAA,QACL,SAAS,KAAK;AAAA,QACd,QAAQ,SAAS;AAAA,QACjB,UAAU,SAAS;AAAA,QACnB,WAAW,UAAU;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,IACD,kBAAkB,OAAO,iBAAiB,IAAI,CAAC,QAAS,IAAwB,GAAG;AAAA,IACnF,mBAAmB,OAAO,kBAAkB,IAAI,CAAC,QAAS,IAAwB,GAAG;AAAA,EACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;AACpF,SAAO,KAAK,IAAI,CAAC,QAAQ;AACvB,UAAM,aAAa,IAAI,WAAW,KAAK;AACvC,QAAI,CAAC,YAAY;AACf,YAAM,IAAI;AAAA,QACR,wBAAwB,IAAI,KAAK;AAAA,MACnC;AAAA,IACF;AACA,WAAO;AAAA,MACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;AAAA,MACtD,QAAQ,OAAO,IAAI,IAAI;AAAA,MACvB;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;AACrF,QAAM,YAAY,oBAAI,IAA+B;AACrD,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;AAC7C,UAAM,KAAK,GAAG;AACd,cAAU,IAAI,IAAI,SAAS,KAAK;AAAA,EAClC;AACA,aAAW,CAAC,SAAS,KAAK,KAAK,WAAW;AACxC,QAAI,MAAM,SAAS,EAAG,eAAc,OAAO,SAAS,GAAG;AAAA,EACzD;AACF;AAEA,SAAS,YAAY,KAAmB,MAAkC;AACxE,QAAM,IAAI,IAAI,UAAU,IAAI;AAC5B,MAAI,MAAM,OAAW,QAAO;AAC5B,MAAI,CAAC,OAAO,SAAS,CAAC,GAAG;AACvB,UAAM,IAAI;AAAA,MACR,mDAAmD,IAAI,iBAAiB,IAAI,OAAO,WAAW,IAAI,GAAG,OAAO,CAAC;AAAA,IAC/G;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/{chunk-IPYXE555.js → chunk-EJGRPCO3.js}

@@ -3,14 +3,11 @@ import {
   ROLLOUT_SCHEMA,
   assertRolloutLine,
   isTrainableSplit
-} from "./chunk-MAX3TN3C.js";
+} from "./chunk-UWZZKKU7.js";
 
 // src/rollout/exporters.ts
 function toSftRows(lines, options = {}) {
-  const minReward = options.minReward ?? 1;
-  return lines.filter(
-    (line) => line.outcome.reward !== null && line.outcome.reward >= minReward && line.outcome.realness_gated !== true && isTrainableSplit(line.task.split) && line.messages.length > 0
-  ).map((line) => ({
+  return lines.filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0).map((line) => ({
     messages: line.messages,
     metadata: {
       rollout_id: line.rollout_id,
@@ -21,8 +18,12 @@ function toSftRows(lines, options = {}) {
     }
   }));
 }
-function toRewardRows(lines) {
-  return lines.filter((line) => line.outcome.reward !== null).map((line) => ({
+function toRewardRows(lines, options = {}) {
+  return lines.filter(
+    (line) => isTrainingLineEligible(line, options) && line.messages.some(
+      (message) => message.role === "user" && typeof message.content === "string" && message.content.length > 0
+    )
+  ).map((line) => ({
     prompt: line.messages.find((m) => m.role === "user")?.content ?? "",
     steps: line.steps ?? [],
     reward: line.outcome.reward,
@@ -67,8 +68,10 @@ function toVerifiersRolloutOutput(line) {
     }
   };
 }
-function toVerifiersRolloutOutputs(lines) {
-  return lines.filter((line) => line.messages.length > 0).map(toVerifiersRolloutOutput);
+function toVerifiersRolloutOutputs(lines, options = {}) {
+  return lines.filter(
+    (line) => isTrainingLineEligible(line, options) && firstAssistantIndex(line.messages) > 0 && firstAssistantIndex(line.messages) < line.messages.length
+  ).map(toVerifiersRolloutOutput);
 }
 function toRftItem(line) {
   const split = firstAssistantIndex(line.messages);
@@ -85,12 +88,32 @@ function toRftItem(line) {
     }
   };
 }
-function toRftItems(lines) {
-  return lines.filter((line) => line.messages.length > 0 && firstAssistantIndex(line.messages) > 0).map(toRftItem);
+function toRftItems(lines, options = {}) {
+  return lines.filter(
+    (line) => isTrainingLineEligible(line, options) && line.messages.length > 0 && firstAssistantIndex(line.messages) > 0
+  ).map(toRftItem);
 }
 function toJsonl(rows) {
   return rows.map((r) => JSON.stringify(r)).join("\n") + (rows.length ? "\n" : "");
 }
+function isTrainingLineEligible(line, options) {
+  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0;
+  if (!Number.isFinite(minimumQualityExclusive)) {
+    throw new Error("minimumQualityExclusive must be finite");
+  }
+  const reward = line.outcome.reward;
+  if (reward === null) return false;
+  if (!Number.isFinite(reward)) {
+    throw new Error(`training reward for rollout "${line.rollout_id}" must be finite`);
+  }
+  if (reward <= minimumQualityExclusive) return false;
+  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {
+    return false;
+  }
+  if (line.outcome.realness_gated === true) return false;
+  if (line.task.split === "search") return true;
+  return line.task.split === "holdout" && options.allowHeldOutTrainingData === true;
+}
 
 // src/rollout/ledger.ts
 import { appendFile, mkdir, readFile, writeFile } from "fs/promises";
@@ -226,7 +249,7 @@ function buildDatasetCard(inputs) {
 
 One line per agent invocation (supervisor episode, worker session, proposer shot, judge call, analyst pass) captured by the \`${ROLLOUT_SCHEMA}\` rollout ledger, labeled with improvement-loop coordinates and the official-judge reward, with the full message transcript inline.
 
-This release contains the **trainable split only** (\`search\`, plus the legacy \`train\` alias). Holdout, dev, and canary splits are structurally excluded at build time (never exported), and the build additionally drops any non-trainable line as a fail-closed filter (${excluded.nonTrain} dropped here).
+This release contains the **trainable split only** (\`search\`). Holdout, dev, and canary splits are structurally excluded at build time, and the build additionally drops any non-trainable line as a fail-closed filter (${excluded.nonTrain} dropped here).
 
 ## Formats
 
@@ -591,4 +614,4 @@ export {
   parseRolloutReleaseArgs,
   runRolloutReleaseCli
 };
-//# sourceMappingURL=chunk-IPYXE555.js.map
+//# sourceMappingURL=chunk-EJGRPCO3.js.map