@tangle-network/agent-eval 0.200.0 → 0.201.0

package/dist/{mint-iCD-W43s.js → mint-vlRckJEX.js}

@@ -2,7 +2,7 @@ import { s as ValidationError } from "./errors-Dngq5h35.js";
 import { a as scoreOrigin, i as rolloutRewardFields } from "./reward-nw2xZGZG.js";
 import { c as runTaskScore } from "./run-record-BRlM99rX.js";
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
-import { a as assertMinted, r as ROLLOUT_SCHEMA } from "./schema-C1aaAxTf.js";
+import { a as assertMinted, r as ROLLOUT_SCHEMA } from "./schema-DAMp77zB.js";
 //#region src/rollout/mint.ts
 /**
 * Rollout minting — `tangle.rollout.v1` lines joined from the records the
@@ -382,4 +382,4 @@ async function mintRolloutRows(records, store, options = {}) {
 //#endregion
 export { unmintableReasons as n, mintRolloutRows as t };
 
-//# sourceMappingURL=mint-iCD-W43s.js.map
+//# sourceMappingURL=mint-vlRckJEX.js.map

package/dist/{mint-iCD-W43s.js.map → mint-vlRckJEX.js.map}

@@ -1 +1 @@
-{"version":3,"file":"mint-iCD-W43s.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *   - search lineage          → `searchLineage`, read from the run's search ledger\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans or an explicit full capture become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, type RunSearchCoordinates, runTaskScore } from '../run-record'\nimport type { LlmSpan, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor for transcript/step payload text (secrets, PII), not opaque identity. */\nexport type RolloutScrubber = (text: string) => string\n\n/** A search cell's place in its search tree, read from the search ledger. */\nexport interface RolloutSearchLineage {\n  /** Edges between the cell's node and the search root; the root is 0. */\n  depth: number\n  /** Order in which the node registered in its search; the root is 0. */\n  ordinal: number\n  /** Repeat index of the cell on its task, from 0. */\n  rep: number\n  /** Run id of the cell whose execution contains this search; null for a\n   *  top-level search. */\n  containingRunId: string | null\n}\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /**\n   * Full canonical capture, including turns no longer in the model's context.\n   * When supplied, missing/empty capture is refused, never replaced with a\n   * span summary. Use the execution owner's capture or the existing harness\n   * readers; reviewer traces and compacted histories are not full capture.\n   */\n  messagesOf?: (\n    runId: string,\n  ) => readonly ChatMessage[] | undefined | Promise<readonly ChatMessage[] | undefined>\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n  /**\n   * Lineage of a search run, read from its search ledger. It fills the line's\n   * `generation` (node depth), `candidate_index` (node order), `task.rep` and\n   * `parent_rollout_id` (the containing cell's run). Mint refuses a record\n   * that carries `search` when this is absent, because a null `generation`\n   * states that the run is not part of a search.\n   */\n  searchLineage?: (\n    search: RunSearchCoordinates,\n  ) => RolloutSearchLineage | Promise<RolloutSearchLineage>\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last && last.content !== null) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** Preserve captured fields; invocation ids are opaque linkage, not generated text. */\nfunction projectMessage(message: ChatMessage, scrub: RolloutScrubber): ChatMessage {\n  return {\n    role: message.role,\n    content: message.content === null ? null : scrub(message.content),\n    ...(message.reasoning_content !== undefined\n      ? { reasoning_content: scrub(message.reasoning_content) }\n      : {}),\n    ...(message.tool_calls !== undefined\n      ? {\n          tool_calls: message.tool_calls.map((call) => ({\n            id: call.id,\n            type: call.type,\n            function: {\n              name: call.function.name,\n              arguments: scrub(call.function.arguments),\n            },\n          })),\n        }\n      : {}),\n    ...(message.tool_call_id !== undefined ? { tool_call_id: message.tool_call_id } : {}),\n    ...(message.name !== undefined ? { name: message.name } : {}),\n    ...(message.is_copied_context !== undefined\n      ? { is_copied_context: message.is_copied_context }\n      : {}),\n  }\n}\n\n/** Never recover missing calls from tool names, span order or argument summaries. */\nfunction requireToolLinkage(messages: readonly ChatMessage[], runId: string): void {\n  const seen = new Set<string>()\n  const pending = new Map<string, string>()\n  const refuse: (detail: string) => never = (detail) => {\n    throw new ValidationError(`Cannot mint rollout for run ${runId}: ${detail}`)\n  }\n  for (const message of messages) {\n    if (message.tool_calls !== undefined) {\n      if (message.role !== 'assistant' || !Array.isArray(message.tool_calls)) {\n        refuse('tool calls must be captured on an assistant message')\n      }\n      if (pending.size > 0) refuse('new tool calls precede the prior invocation results')\n      for (const call of message.tool_calls) {\n        if (\n          !call ||\n          typeof call.id !== 'string' ||\n          call.id.length === 0 ||\n          seen.has(call.id) ||\n          call.type !== 'function' ||\n          !call.function ||\n          typeof call.function.name !== 'string' ||\n          call.function.name.length === 0 ||\n          typeof call.function.arguments !== 'string'\n        ) {\n          refuse('missing, duplicate or malformed tool invocation')\n        }\n        seen.add(call.id)\n        pending.set(call.id, call.function.name)\n      }\n    }\n    if (message.role === 'tool') {\n      const id = message.tool_call_id\n      if (typeof id !== 'string' || !pending.has(id)) {\n        refuse('tool result has no captured invocation (or answers it more than once)')\n      }\n      if (message.name !== undefined && message.name !== pending.get(id)) {\n        refuse('tool result name does not match its invocation')\n      }\n      pending.delete(id)\n    } else if (message.tool_call_id !== undefined) {\n      refuse('tool_call_id is only valid on a tool result')\n    } else if (pending.size > 0 && message.tool_calls === undefined) {\n      refuse('conversation advances before all tool results were captured')\n    }\n  }\n  if (pending.size > 0) refuse('tool invocation has no captured result')\n}\n\n/** Legacy span history; callers with compacted contexts supply messagesOf instead. */\nfunction finalConversation(spans: Span[]): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = [...last.messages]\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: last.output })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst isObject = (value: unknown): value is Record<string, unknown> =>\n  typeof value === 'object' && value !== null\n\ninterface MintFieldCheck {\n  /** The RunRecord path, spelled the way the caller has to fix it. */\n  readonly field: string\n  /** True when the record carries something the line can honestly be built from. */\n  readonly present: (bag: Record<string, unknown>) => boolean\n  /** What the caller writes onto the record, and why that value and not another. */\n  readonly remedy: string\n}\n\n/**\n * The RunRecord fields mint reads that a record can be missing even though the\n * TYPE says it cannot. There are exactly two ways that happens:\n *\n *   1. The field was OPTIONAL when the record was serialized. `costProvenance`,\n *      `terminalOutcome` and `scenarioId` were optional through agent-eval\n *      0.125 and became required in 0.126, with no on-disk migration — so every\n *      ledger written before 0.126 is full of records the type calls complete.\n *   2. Mint reads a level DEEPER than the record's own type is checked at:\n *      `outcome.raw`, `tokenUsage.input`, `tokenUsage.output`.\n *\n * Nothing else needs a check here. Every other field mint copies is a top-level\n * scalar landing in a typed slot on the line, where an absent value arrives as\n * `undefined` and `assertMinted` refuses it by name. These are the ones where an\n * absent value instead kills the join with `TypeError: Cannot read properties of\n * undefined`, or — worse — mints a line that reads as measured.\n *\n * This is deliberately NOT `validateRunRecord`. That validator answers \"is this\n * a valid RunRecord\", which is a wider question than \"can a rollout line be\n * built from this one\": it also enforces model-snapshot discipline, the\n * `terminalFailureReason` coupling, and the `costUsd === costProvenance.usd`\n * agreement. Routing the mint door through it would refuse records mint can\n * mint honestly today (a model alias with no snapshot date, for one), which is\n * a policy change with its own blast radius and not this bug. The door asks the\n * narrower question and answers it precisely.\n */\nconst MINT_FIELD_CHECKS: readonly MintFieldCheck[] = [\n  {\n    field: 'costProvenance',\n    present: (bag) => isObject(bag.costProvenance) && typeof bag.costProvenance.kind === 'string',\n    remedy:\n      \"Records written before agent-eval 0.126 predate this field and carry `costUsd: 0` as the documented uncaptured sentinel, which is NOT an observed zero. Backfill it as costProvenance: { kind: 'uncaptured', usd: null } WITH costUsd: null — an uncaptured cost whose costUsd is non-null is rejected by validateRunRecord, so provenance alone leaves the record invalid.\",\n  },\n  {\n    field: 'tokenUsage',\n    present: (bag) => isObject(bag.tokenUsage),\n    remedy:\n      \"The line's cost.tokens_in and cost.tokens_out are read from it. Backfill it from the provider's usage report; mint will not write 0 for tokens nobody counted.\",\n  },\n  {\n    field: 'tokenUsage.input',\n    present: (bag) => !isObject(bag.tokenUsage) || typeof bag.tokenUsage.input === 'number',\n    remedy: \"The line's cost.tokens_in is read from it, and a missing count is not a zero count.\",\n  },\n  {\n    field: 'tokenUsage.output',\n    present: (bag) => !isObject(bag.tokenUsage) || typeof bag.tokenUsage.output === 'number',\n    remedy: \"The line's cost.tokens_out is read from it, and a missing count is not a zero count.\",\n  },\n  {\n    field: 'outcome',\n    present: (bag) => isObject(bag.outcome),\n    remedy:\n      \"The line's reward, reward_source and metrics are all read from it. A record with no outcome carries no training label at all, and mint refuses an unlabeled row.\",\n  },\n  {\n    field: 'outcome.raw',\n    // Reported only when `outcome` itself is present: one absent field should\n    // produce one reason per CAUSE, not one per path that dereferences it.\n    present: (bag) => !isObject(bag.outcome) || isObject(bag.outcome.raw),\n    remedy:\n      'It is the metric bag copied verbatim into the line\\'s outcome.metrics. `{ ...undefined }` spreads to `{}` without complaint, so an absent bag would mint as \"this run reported no metrics\" — a different claim from \"this record predates the field\". Backfill it as {} only when that is what you mean.',\n  },\n  {\n    field: 'terminalOutcome',\n    present: (bag) => typeof bag.terminalOutcome === 'string',\n    remedy:\n      \"It became required in agent-eval 0.126. Backfill it from root-run or process evidence, or as 'unknown' when the producer has none — mint will not decide the line's is_completed and is_truncated for you.\",\n  },\n  {\n    field: 'scenarioId',\n    present: (bag) => typeof bag.scenarioId === 'string' && bag.scenarioId.length > 0,\n    remedy:\n      \"It became required in agent-eval 0.126 and becomes the line's task.instance_id, which must be a non-empty string. Backfill it from the scenario the run was dealt (pre-0.126 producers often left it in outcome.raw.scenario_id).\",\n  },\n]\n\n/**\n * Why a record cannot be minted, one entry per missing field, empty when it can.\n *\n * Exported so a caller can partition a whole ledger — \"which of my 2742 records\n * predate 0.126\" — without catching an exception per record, and without\n * re-deriving the field list on their side. A re-derived list is a list that\n * drifts from the door it is supposed to predict.\n *\n * Takes a `RunRecord` because that is what the caller holds and what the\n * compiler agrees they hold. The type is precisely the thing that is wrong, so\n * the checks read the record as the untyped bag it actually is on disk.\n */\nexport function unmintableReasons(record: RunRecord): string[] {\n  const bag = record as unknown as Record<string, unknown>\n  return MINT_FIELD_CHECKS.filter((check) => !check.present(bag)).map(\n    (check) => `${check.field} is missing. ${check.remedy}`,\n  )\n}\n\n/**\n * The mint door THROWS on a record it cannot build a line from. It does NOT\n * normalise an absent `costProvenance` to `{kind:'uncaptured', usd:null}`, and\n * the choice is not stylistic:\n *\n *   - Normalising cannot cover the record, only part of it. `terminalOutcome`\n *     feeds `is_completed` and `is_truncated`, which the rollout schema requires\n *     to be BOOLEAN — there is no null to fall back to, so every possible\n *     default is a claim about how the run ended. A door that quietly fixes the\n *     cost and invents the ending is a door no caller can predict.\n *   - Normalising the cost requires knowing what `costUsd: 0` meant, and mint\n *     cannot know. A genuinely free run and an uncaptured one are the same bytes\n *     in a pre-0.126 record; only the producer can tell them apart. Guessing is\n *     exactly the failure this guard exists to stop — the 0.125 optional chain\n *     `record.costProvenance?.kind === 'uncaptured'` already made that guess,\n *     silently, and every record it touched minted `cost.usd: 0`: an unmeasured\n *     cost published as a measured zero, into a training dataset.\n *   - `requireTaskScore`, directly above, already refuses an unlabeled record\n *     for the same reason: \"nobody graded this\" is not \"graded zero\". \"Nobody\n *     billed this\" is not \"billed zero\".\n *\n * The caller who wants historical records minted backfills them at their store,\n * in one pass, where `costUsd` can be corrected alongside `costProvenance` —\n * which is the only place that decision can be made correctly. The refusal names\n * the run, names every missing field, and spells the value to write.\n */\nfunction requireMintableRecord(record: RunRecord): void {\n  const reasons = unmintableReasons(record)\n  if (reasons.length === 0) return\n  throw new ValidationError(`Cannot mint rollout for run ${record.runId}: ${reasons.join('\\n  ')}`)\n}\n\n/** The search lineage a record's line is built from; undefined for a record outside a search. */\nasync function resolveSearchLineage(\n  record: RunRecord,\n  options: MintRolloutOptions,\n): Promise<RolloutSearchLineage | undefined> {\n  if (record.search === undefined) return undefined\n  const refuse: (detail: string) => never = (detail) => {\n    throw new ValidationError(`Cannot mint rollout for search run ${record.runId}: ${detail}`)\n  }\n  if (!options.searchLineage) {\n    refuse('pass searchLineage so generation, candidate_index, rep and parent come from its ledger')\n  }\n  const lineage = await options.searchLineage(record.search)\n  const isCount = (value: unknown): boolean => Number.isSafeInteger(value) && (value as number) >= 0\n  if (!isCount(lineage.depth)) refuse('lineage depth must be a non-negative integer')\n  if (!isCount(lineage.ordinal)) refuse('lineage ordinal must be a non-negative integer')\n  if (!isCount(lineage.rep)) refuse('lineage rep must be a non-negative integer')\n  if (\n    lineage.containingRunId !== null &&\n    (typeof lineage.containingRunId !== 'string' || lineage.containingRunId.length === 0)\n  ) {\n    refuse('lineage containingRunId must be null or a non-empty run id')\n  }\n  if (lineage.containingRunId === record.runId) refuse('a run cannot contain itself')\n  return lineage\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  lineage: RolloutSearchLineage | undefined,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // Field presence first, and BEFORE `requireTaskScore`: that guard reads\n  // `record.outcome.searchScore` on its way to the answer, so an absent\n  // `outcome` would throw a bare TypeError from inside the guard whose whole\n  // job is to produce a clean refusal.\n  //\n  // Both branches of `mintRolloutRows` — the traced line and the gap line —\n  // land here, which is the point: `mintLine` is the only constructor of a\n  // `MintedRolloutLine` from a RunRecord, so there is no path into the waist\n  // that skips the check and no way to get this wrong from the outside.\n  requireMintableRecord(record)\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  // Only an observed or estimated amount is a total; a lower bound is not.\n  const knownTotal =\n    record.costProvenance.kind === 'observed' || record.costProvenance.kind === 'estimated'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: lineage?.containingRunId ?? null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: lineage?.depth ?? null,\n      candidate_index: lineage?.ordinal ?? null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: lineage?.rep ?? 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: knownTotal ? record.costUsd : null,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: {\n        patch_path: null,\n        run_dir: null,\n        transcript_ref: record.traceRef ? `trace:${record.traceRef.traceId}` : null,\n      },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans or an explicit capture are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    const captured = options.messagesOf ? await options.messagesOf(record.runId) : undefined\n    if (options.messagesOf && (!Array.isArray(captured) || captured.length === 0)) {\n      throw new ValidationError(\n        `Cannot mint rollout for run ${record.runId}: full capture is missing`,\n      )\n    }\n    const lineage = await resolveSearchLineage(record, options)\n    if (trajectory.steps.length === 0 && !options.messagesOf) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(\n          record,\n          lineage,\n          [],\n          [],\n          options,\n          capturedAt,\n          'no trace spans recorded for this runId',\n        ),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const spans = trajectory.steps.map((s) => s.span)\n    const messages = captured ?? finalConversation(spans)\n    requireToolLinkage(messages, record.runId)\n    const conversation = messages.map((message) => projectMessage(message, scrub))\n    // SFT drops copied context. That projection must not orphan a retained result.\n    requireToolLinkage(\n      conversation.filter((message) => message.is_copied_context !== true),\n      record.runId,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, lineage, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsGA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,QAAQ,KAAK,YAAY,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EAClE,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,eAAe,SAAsB,OAAqC;CACjF,OAAO;EACL,MAAM,QAAQ;EACd,SAAS,QAAQ,YAAY,OAAO,OAAO,MAAM,QAAQ,OAAO;EAChE,GAAI,QAAQ,sBAAsB,KAAA,IAC9B,EAAE,mBAAmB,MAAM,QAAQ,iBAAiB,EAAE,IACtD,CAAC;EACL,GAAI,QAAQ,eAAe,KAAA,IACvB,EACE,YAAY,QAAQ,WAAW,KAAK,UAAU;GAC5C,IAAI,KAAK;GACT,MAAM,KAAK;GACX,UAAU;IACR,MAAM,KAAK,SAAS;IACpB,WAAW,MAAM,KAAK,SAAS,SAAS;GAC1C;EACF,EAAE,EACJ,IACA,CAAC;EACL,GAAI,QAAQ,iBAAiB,KAAA,IAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;EACnF,GAAI,QAAQ,SAAS,KAAA,IAAY,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;EAC3D,GAAI,QAAQ,sBAAsB,KAAA,IAC9B,EAAE,mBAAmB,QAAQ,kBAAkB,IAC/C,CAAC;CACP;AACF;;AAGA,SAAS,mBAAmB,UAAkC,OAAqB;CACjF,MAAM,uBAAO,IAAI,IAAY;CAC7B,MAAM,0BAAU,IAAI,IAAoB;CACxC,MAAM,UAAqC,WAAW;EACpD,MAAM,IAAI,gBAAgB,+BAA+B,MAAM,IAAI,QAAQ;CAC7E;CACA,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,eAAe,KAAA,GAAW;GACpC,IAAI,QAAQ,SAAS,eAAe,CAAC,MAAM,QAAQ,QAAQ,UAAU,GACnE,OAAO,qDAAqD;GAE9D,IAAI,QAAQ,OAAO,GAAG,OAAO,qDAAqD;GAClF,KAAK,MAAM,QAAQ,QAAQ,YAAY;IACrC,IACE,CAAC,QACD,OAAO,KAAK,OAAO,YACnB,KAAK,GAAG,WAAW,KACnB,KAAK,IAAI,KAAK,EAAE,KAChB,KAAK,SAAS,cACd,CAAC,KAAK,YACN,OAAO,KAAK,SAAS,SAAS,YAC9B,KAAK,SAAS,KAAK,WAAW,KAC9B,OAAO,KAAK,SAAS,cAAc,UAEnC,OAAO,iDAAiD;IAE1D,KAAK,IAAI,KAAK,EAAE;IAChB,QAAQ,IAAI,KAAK,IAAI,KAAK,SAAS,IAAI;GACzC;EACF;EACA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,MAAM,KAAK,QAAQ;GACnB,IAAI,OAAO,OAAO,YAAY,CAAC,QAAQ,IAAI,EAAE,GAC3C,OAAO,uEAAuE;GAEhF,IAAI,QAAQ,SAAS,KAAA,KAAa,QAAQ,SAAS,QAAQ,IAAI,EAAE,GAC/D,OAAO,gDAAgD;GAEzD,QAAQ,OAAO,EAAE;EACnB,OAAO,IAAI,QAAQ,iBAAiB,KAAA,GAClC,OAAO,6CAA6C;OAC/C,IAAI,QAAQ,OAAO,KAAK,QAAQ,eAAe,KAAA,GACpD,OAAO,6DAA6D;CAExE;CACA,IAAI,QAAQ,OAAO,GAAG,OAAO,wCAAwC;AACvE;;AAGA,SAAS,kBAAkB,OAA8B;CACvD,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,CAAC,GAAG,KAAK,QAAQ;CACjD,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,KAAK;CAAO,CAAC;CAE3D,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,YAAY,UAChB,OAAO,UAAU,YAAY,UAAU;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCzC,MAAM,oBAA+C;CACnD;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,cAAc,KAAK,OAAO,IAAI,eAAe,SAAS;EACrF,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,UAAU;EACzC,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,CAAC,SAAS,IAAI,UAAU,KAAK,OAAO,IAAI,WAAW,UAAU;EAC/E,QAAQ;CACV;CACA;EACE,OAAO;EACP,UAAU,QAAQ,CAAC,SAAS,IAAI,UAAU,KAAK,OAAO,IAAI,WAAW,WAAW;EAChF,QAAQ;CACV;CACA;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,OAAO;EACtC,QACE;CACJ;CACA;EACE,OAAO;EAGP,UAAU,QAAQ,CAAC,SAAS,IAAI,OAAO,KAAK,SAAS,IAAI,QAAQ,GAAG;EACpE,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,OAAO,IAAI,oBAAoB;EACjD,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,OAAO,IAAI,eAAe,YAAY,IAAI,WAAW,SAAS;EAChF,QACE;CACJ;AACF;;;;;;;;;;;;;AAcA,SAAgB,kBAAkB,QAA6B;CAC7D,MAAM,MAAM;CACZ,OAAO,kBAAkB,QAAQ,UAAU,CAAC,MAAM,QAAQ,GAAG,CAAC,CAAC,CAAC,KAC7D,UAAU,GAAG,MAAM,MAAM,eAAe,MAAM,QACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAS,sBAAsB,QAAyB;CACtD,MAAM,UAAU,kBAAkB,MAAM;CACxC,IAAI,QAAQ,WAAW,GAAG;CAC1B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,IAAI,QAAQ,KAAK,MAAM,GAAG;AAClG;;AAGA,eAAe,qBACb,QACA,SAC2C;CAC3C,IAAI,OAAO,WAAW,KAAA,GAAW,OAAO,KAAA;CACxC,MAAM,UAAqC,WAAW;EACpD,MAAM,IAAI,gBAAgB,sCAAsC,OAAO,MAAM,IAAI,QAAQ;CAC3F;CACA,IAAI,CAAC,QAAQ,eACX,OAAO,wFAAwF;CAEjG,MAAM,UAAU,MAAM,QAAQ,cAAc,OAAO,MAAM;CACzD,MAAM,WAAW,UAA4B,OAAO,cAAc,KAAK,KAAM,SAAoB;CACjG,IAAI,CAAC,QAAQ,QAAQ,KAAK,GAAG,OAAO,8CAA8C;CAClF,IAAI,CAAC,QAAQ,QAAQ,OAAO,GAAG,OAAO,gDAAgD;CACtF,IAAI,CAAC,QAAQ,QAAQ,GAAG,GAAG,OAAO,4CAA4C;CAC9E,IACE,QAAQ,oBAAoB,SAC3B,OAAO,QAAQ,oBAAoB,YAAY,QAAQ,gBAAgB,WAAW,IAEnF,OAAO,4DAA4D;CAErE,IAAI,QAAQ,oBAAoB,OAAO,OAAO,OAAO,6BAA6B;CAClF,OAAO;AACT;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,SACA,OACA,UACA,SACA,YACA,KACmB;CAUnB,sBAAsB,MAAM;CAI5B,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAE/C,MAAM,aACJ,OAAO,eAAe,SAAS,cAAc,OAAO,eAAe,SAAS;CAC9E,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB,SAAS,mBAAmB;EAC/C,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY,SAAS,SAAS;EAC9B,iBAAiB,SAAS,WAAW;EACrC,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK,SAAS,OAAO;EACvB;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,UAAU;GACnC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GACT,YAAY;GACZ,SAAS;GACT,gBAAgB,OAAO,WAAW,SAAS,OAAO,SAAS,YAAY;EACzE;EACA,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,MAAM,WAAW,QAAQ,aAAa,MAAM,QAAQ,WAAW,OAAO,KAAK,IAAI,KAAA;EAC/E,IAAI,QAAQ,eAAe,CAAC,MAAM,QAAQ,QAAQ,KAAK,SAAS,WAAW,IACzE,MAAM,IAAI,gBACR,+BAA+B,OAAO,MAAM,0BAC9C;EAEF,MAAM,UAAU,MAAM,qBAAqB,QAAQ,OAAO;EAC1D,IAAI,WAAW,MAAM,WAAW,KAAK,CAAC,QAAQ,YAAY;GACxD,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SACE,QACA,SACA,CAAC,GACD,CAAC,GACD,SACA,YACA,wCACF,CACF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,QAAQ,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI;EAChD,MAAM,WAAW,YAAY,kBAAkB,KAAK;EACpD,mBAAmB,UAAU,OAAO,KAAK;EACzC,MAAM,eAAe,SAAS,KAAK,YAAY,eAAe,SAAS,KAAK,CAAC;EAE7E,mBACE,aAAa,QAAQ,YAAY,QAAQ,sBAAsB,IAAI,GACnE,OAAO,KACT;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,SAAS,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CACpF;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}
+{"version":3,"file":"mint-vlRckJEX.js","names":[],"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *   - search lineage          → `searchLineage`, read from the run's search ledger\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true is\n * never exported with a positive reward OR with any of the numbers that reward\n * was computed from. The gate travels into the training data (`reward` forced\n * to 0, `realness_gated: true`) and the whole outcome is transformed by\n * `gateGamedOutcome` inside `assertMinted` below, which relocates `metrics` and\n * `verdict` to `provenance.gated_evidence`. Mint returns\n * `MintedRolloutLine[]`: the brand the training exporters require, which only\n * this function, `readRolloutLedger`, and an explicit `assertMinted` can mint.\n *\n * A record carrying NEITHER split score is REJECTED (`ValidationError`), never\n * minted at 0 — \"nobody graded this\" is not the same claim as \"graded a total\n * failure\", and a trainer reading 0 learns the second. Lines that already\n * carry `reward: null` (interchange imports, existing ledgers) remain valid on\n * the wire; only the RunRecord→line door refuses.\n *\n * Records without spans or an explicit full capture become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, type RunSearchCoordinates, runTaskScore } from '../run-record'\nimport type { LlmSpan, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport { rolloutRewardFields, scoreOrigin } from './reward'\nimport {\n  assertMinted,\n  type ChatMessage,\n  type MintedRolloutLine,\n  ROLLOUT_SCHEMA,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor for transcript/step payload text (secrets, PII), not opaque identity. */\nexport type RolloutScrubber = (text: string) => string\n\n/** A search cell's place in its search tree, read from the search ledger. */\nexport interface RolloutSearchLineage {\n  /** Edges between the cell's node and the search root; the root is 0. */\n  depth: number\n  /** Order in which the node registered in its search; the root is 0. */\n  ordinal: number\n  /** Repeat index of the cell on its task, from 0. */\n  rep: number\n  /** Run id of the cell whose execution contains this search; null for a\n   *  top-level search. */\n  containingRunId: string | null\n}\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /**\n   * Full canonical capture, including turns no longer in the model's context.\n   * When supplied, missing/empty capture is refused, never replaced with a\n   * span summary. Use the execution owner's capture or the existing harness\n   * readers; reviewer traces and compacted histories are not full capture.\n   */\n  messagesOf?: (\n    runId: string,\n  ) => readonly ChatMessage[] | undefined | Promise<readonly ChatMessage[] | undefined>\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n  /**\n   * Lineage of a search run, read from its search ledger. It fills the line's\n   * `generation` (node depth), `candidate_index` (node order), `task.rep` and\n   * `parent_rollout_id` (the containing cell's run). Mint refuses a record\n   * that carries `search` when this is absent, because a null `generation`\n   * states that the run is not part of a search.\n   */\n  searchLineage?: (\n    search: RunSearchCoordinates,\n  ) => RolloutSearchLineage | Promise<RolloutSearchLineage>\n}\n\nexport interface MintRolloutResult {\n  rows: MintedRolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last && last.content !== null) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** Preserve captured fields; invocation ids are opaque linkage, not generated text. */\nfunction projectMessage(message: ChatMessage, scrub: RolloutScrubber): ChatMessage {\n  return {\n    role: message.role,\n    content: message.content === null ? null : scrub(message.content),\n    ...(message.reasoning_content !== undefined\n      ? { reasoning_content: scrub(message.reasoning_content) }\n      : {}),\n    ...(message.tool_calls !== undefined\n      ? {\n          tool_calls: message.tool_calls.map((call) => ({\n            id: call.id,\n            type: call.type,\n            function: {\n              name: call.function.name,\n              arguments: scrub(call.function.arguments),\n            },\n          })),\n        }\n      : {}),\n    ...(message.tool_call_id !== undefined ? { tool_call_id: message.tool_call_id } : {}),\n    ...(message.name !== undefined ? { name: message.name } : {}),\n    ...(message.is_copied_context !== undefined\n      ? { is_copied_context: message.is_copied_context }\n      : {}),\n  }\n}\n\n/** Never recover missing calls from tool names, span order or argument summaries. */\nfunction requireToolLinkage(messages: readonly ChatMessage[], runId: string): void {\n  const seen = new Set<string>()\n  const pending = new Map<string, string>()\n  const refuse: (detail: string) => never = (detail) => {\n    throw new ValidationError(`Cannot mint rollout for run ${runId}: ${detail}`)\n  }\n  for (const message of messages) {\n    if (message.tool_calls !== undefined) {\n      if (message.role !== 'assistant' || !Array.isArray(message.tool_calls)) {\n        refuse('tool calls must be captured on an assistant message')\n      }\n      if (pending.size > 0) refuse('new tool calls precede the prior invocation results')\n      for (const call of message.tool_calls) {\n        if (\n          !call ||\n          typeof call.id !== 'string' ||\n          call.id.length === 0 ||\n          seen.has(call.id) ||\n          call.type !== 'function' ||\n          !call.function ||\n          typeof call.function.name !== 'string' ||\n          call.function.name.length === 0 ||\n          typeof call.function.arguments !== 'string'\n        ) {\n          refuse('missing, duplicate or malformed tool invocation')\n        }\n        seen.add(call.id)\n        pending.set(call.id, call.function.name)\n      }\n    }\n    if (message.role === 'tool') {\n      const id = message.tool_call_id\n      if (typeof id !== 'string' || !pending.has(id)) {\n        refuse('tool result has no captured invocation (or answers it more than once)')\n      }\n      if (message.name !== undefined && message.name !== pending.get(id)) {\n        refuse('tool result name does not match its invocation')\n      }\n      pending.delete(id)\n    } else if (message.tool_call_id !== undefined) {\n      refuse('tool_call_id is only valid on a tool result')\n    } else if (pending.size > 0 && message.tool_calls === undefined) {\n      refuse('conversation advances before all tool results were captured')\n    }\n  }\n  if (pending.size > 0) refuse('tool invocation has no captured result')\n}\n\n/** Legacy span history; callers with compacted contexts supply messagesOf instead. */\nfunction finalConversation(spans: Span[]): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = [...last.messages]\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: last.output })\n  }\n  return messages\n}\n\n// The reward derivations live in the leaf module `./reward` so gate and\n// reporting code can import them without dragging in the trace store; they are\n// re-exported here because the derivations shipped from this path.\nexport {\n  isRealnessGated,\n  observedScore,\n  observedSplitScore,\n  type ScoreOrigin,\n  type ScorePreference,\n  scoreOrigin,\n  trainingReward,\n  trainingScore,\n} from './reward'\n\nconst REWARD_SOURCE: Record<ReturnType<typeof scoreOrigin>, string> = {\n  holdout: 'run-record/holdout-score',\n  search: 'run-record/search-score',\n  unscored: 'run-record/unscored',\n}\n\n/**\n * The mint door refuses an execution-only record: a missing training label is\n * not a zero reward, and not a mintable line either. Lines that already carry\n * `reward: null` — interchange imports, existing ledgers — stay valid on the\n * wire and keep their labeled gap; this guard is only about the\n * RunRecord→line door, where the producer can still be told to go score the\n * run instead of shipping an unlabeled row.\n */\nfunction requireTaskScore(record: RunRecord): void {\n  if (runTaskScore(record) === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n}\n\nconst isObject = (value: unknown): value is Record<string, unknown> =>\n  typeof value === 'object' && value !== null\n\ninterface MintFieldCheck {\n  /** The RunRecord path, spelled the way the caller has to fix it. */\n  readonly field: string\n  /** True when the record carries something the line can honestly be built from. */\n  readonly present: (bag: Record<string, unknown>) => boolean\n  /** What the caller writes onto the record, and why that value and not another. */\n  readonly remedy: string\n}\n\n/**\n * The RunRecord fields mint reads that a record can be missing even though the\n * TYPE says it cannot. There are exactly two ways that happens:\n *\n *   1. The field was OPTIONAL when the record was serialized. `costProvenance`,\n *      `terminalOutcome` and `scenarioId` were optional through agent-eval\n *      0.125 and became required in 0.126, with no on-disk migration — so every\n *      ledger written before 0.126 is full of records the type calls complete.\n *   2. Mint reads a level DEEPER than the record's own type is checked at:\n *      `outcome.raw`, `tokenUsage.input`, `tokenUsage.output`.\n *\n * Nothing else needs a check here. Every other field mint copies is a top-level\n * scalar landing in a typed slot on the line, where an absent value arrives as\n * `undefined` and `assertMinted` refuses it by name. These are the ones where an\n * absent value instead kills the join with `TypeError: Cannot read properties of\n * undefined`, or — worse — mints a line that reads as measured.\n *\n * This is deliberately NOT `validateRunRecord`. That validator answers \"is this\n * a valid RunRecord\", which is a wider question than \"can a rollout line be\n * built from this one\": it also enforces model-snapshot discipline, the\n * `terminalFailureReason` coupling, and the `costUsd === costProvenance.usd`\n * agreement. Routing the mint door through it would refuse records mint can\n * mint honestly today (a model alias with no snapshot date, for one), which is\n * a policy change with its own blast radius and not this bug. The door asks the\n * narrower question and answers it precisely.\n */\nconst MINT_FIELD_CHECKS: readonly MintFieldCheck[] = [\n  {\n    field: 'costProvenance',\n    present: (bag) => isObject(bag.costProvenance) && typeof bag.costProvenance.kind === 'string',\n    remedy:\n      \"Records written before agent-eval 0.126 predate this field and carry `costUsd: 0` as the documented uncaptured sentinel, which is NOT an observed zero. Backfill it as costProvenance: { kind: 'uncaptured', usd: null } WITH costUsd: null — an uncaptured cost whose costUsd is non-null is rejected by validateRunRecord, so provenance alone leaves the record invalid.\",\n  },\n  {\n    field: 'tokenUsage',\n    present: (bag) => isObject(bag.tokenUsage),\n    remedy:\n      \"The line's cost.tokens_in and cost.tokens_out are read from it. Backfill it from the provider's usage report; mint will not write 0 for tokens nobody counted.\",\n  },\n  {\n    field: 'tokenUsage.input',\n    present: (bag) => !isObject(bag.tokenUsage) || typeof bag.tokenUsage.input === 'number',\n    remedy: \"The line's cost.tokens_in is read from it, and a missing count is not a zero count.\",\n  },\n  {\n    field: 'tokenUsage.output',\n    present: (bag) => !isObject(bag.tokenUsage) || typeof bag.tokenUsage.output === 'number',\n    remedy: \"The line's cost.tokens_out is read from it, and a missing count is not a zero count.\",\n  },\n  {\n    field: 'outcome',\n    present: (bag) => isObject(bag.outcome),\n    remedy:\n      \"The line's reward, reward_source and metrics are all read from it. A record with no outcome carries no training label at all, and mint refuses an unlabeled row.\",\n  },\n  {\n    field: 'outcome.raw',\n    // Reported only when `outcome` itself is present: one absent field should\n    // produce one reason per CAUSE, not one per path that dereferences it.\n    present: (bag) => !isObject(bag.outcome) || isObject(bag.outcome.raw),\n    remedy:\n      'It is the metric bag copied verbatim into the line\\'s outcome.metrics. `{ ...undefined }` spreads to `{}` without complaint, so an absent bag would mint as \"this run reported no metrics\" — a different claim from \"this record predates the field\". Backfill it as {} only when that is what you mean.',\n  },\n  {\n    field: 'terminalOutcome',\n    present: (bag) => typeof bag.terminalOutcome === 'string',\n    remedy:\n      \"It became required in agent-eval 0.126. Backfill it from root-run or process evidence, or as 'unknown' when the producer has none — mint will not decide the line's is_completed and is_truncated for you.\",\n  },\n  {\n    field: 'scenarioId',\n    present: (bag) => typeof bag.scenarioId === 'string' && bag.scenarioId.length > 0,\n    remedy:\n      \"It became required in agent-eval 0.126 and becomes the line's task.instance_id, which must be a non-empty string. Backfill it from the scenario the run was dealt (pre-0.126 producers often left it in outcome.raw.scenario_id).\",\n  },\n]\n\n/**\n * Why a record cannot be minted, one entry per missing field, empty when it can.\n *\n * Exported so a caller can partition a whole ledger — \"which of my 2742 records\n * predate 0.126\" — without catching an exception per record, and without\n * re-deriving the field list on their side. A re-derived list is a list that\n * drifts from the door it is supposed to predict.\n *\n * Takes a `RunRecord` because that is what the caller holds and what the\n * compiler agrees they hold. The type is precisely the thing that is wrong, so\n * the checks read the record as the untyped bag it actually is on disk.\n */\nexport function unmintableReasons(record: RunRecord): string[] {\n  const bag = record as unknown as Record<string, unknown>\n  return MINT_FIELD_CHECKS.filter((check) => !check.present(bag)).map(\n    (check) => `${check.field} is missing. ${check.remedy}`,\n  )\n}\n\n/**\n * The mint door THROWS on a record it cannot build a line from. It does NOT\n * normalise an absent `costProvenance` to `{kind:'uncaptured', usd:null}`, and\n * the choice is not stylistic:\n *\n *   - Normalising cannot cover the record, only part of it. `terminalOutcome`\n *     feeds `is_completed` and `is_truncated`, which the rollout schema requires\n *     to be BOOLEAN — there is no null to fall back to, so every possible\n *     default is a claim about how the run ended. A door that quietly fixes the\n *     cost and invents the ending is a door no caller can predict.\n *   - Normalising the cost requires knowing what `costUsd: 0` meant, and mint\n *     cannot know. A genuinely free run and an uncaptured one are the same bytes\n *     in a pre-0.126 record; only the producer can tell them apart. Guessing is\n *     exactly the failure this guard exists to stop — the 0.125 optional chain\n *     `record.costProvenance?.kind === 'uncaptured'` already made that guess,\n *     silently, and every record it touched minted `cost.usd: 0`: an unmeasured\n *     cost published as a measured zero, into a training dataset.\n *   - `requireTaskScore`, directly above, already refuses an unlabeled record\n *     for the same reason: \"nobody graded this\" is not \"graded zero\". \"Nobody\n *     billed this\" is not \"billed zero\".\n *\n * The caller who wants historical records minted backfills them at their store,\n * in one pass, where `costUsd` can be corrected alongside `costProvenance` —\n * which is the only place that decision can be made correctly. The refusal names\n * the run, names every missing field, and spells the value to write.\n */\nfunction requireMintableRecord(record: RunRecord): void {\n  const reasons = unmintableReasons(record)\n  if (reasons.length === 0) return\n  throw new ValidationError(`Cannot mint rollout for run ${record.runId}: ${reasons.join('\\n  ')}`)\n}\n\n/** The search lineage a record's line is built from; undefined for a record outside a search. */\nasync function resolveSearchLineage(\n  record: RunRecord,\n  options: MintRolloutOptions,\n): Promise<RolloutSearchLineage | undefined> {\n  if (record.search === undefined) return undefined\n  const refuse: (detail: string) => never = (detail) => {\n    throw new ValidationError(`Cannot mint rollout for search run ${record.runId}: ${detail}`)\n  }\n  if (!options.searchLineage) {\n    refuse('pass searchLineage so generation, candidate_index, rep and parent come from its ledger')\n  }\n  const lineage = await options.searchLineage(record.search)\n  const isCount = (value: unknown): boolean => Number.isSafeInteger(value) && (value as number) >= 0\n  if (!isCount(lineage.depth)) refuse('lineage depth must be a non-negative integer')\n  if (!isCount(lineage.ordinal)) refuse('lineage ordinal must be a non-negative integer')\n  if (!isCount(lineage.rep)) refuse('lineage rep must be a non-negative integer')\n  if (\n    lineage.containingRunId !== null &&\n    (typeof lineage.containingRunId !== 'string' || lineage.containingRunId.length === 0)\n  ) {\n    refuse('lineage containingRunId must be null or a non-empty run id')\n  }\n  if (lineage.containingRunId === record.runId) refuse('a run cannot contain itself')\n  return lineage\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  lineage: RolloutSearchLineage | undefined,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): MintedRolloutLine {\n  // Field presence first, and BEFORE `requireTaskScore`: that guard reads\n  // `record.outcome.searchScore` on its way to the answer, so an absent\n  // `outcome` would throw a bare TypeError from inside the guard whose whole\n  // job is to produce a clean refusal.\n  //\n  // Both branches of `mintRolloutRows` — the traced line and the gap line —\n  // land here, which is the point: `mintLine` is the only constructor of a\n  // `MintedRolloutLine` from a RunRecord, so there is no path into the waist\n  // that skips the check and no way to get this wrong from the outside.\n  requireMintableRecord(record)\n  // A missing task score is refused before anything is built: an\n  // execution-only record has no training label, and a missing label is\n  // neither a zero reward nor a mintable row.\n  requireTaskScore(record)\n  // `reward` and `realness_gated` come out of one call, so neither door into\n  // the waist can write one and forget the other.\n  const rewardFields = rolloutRewardFields(record)\n  // Only an observed or estimated amount is a total; a lower bound is not.\n  const knownTotal =\n    record.costProvenance.kind === 'observed' || record.costProvenance.kind === 'estimated'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  // `assertMinted` rather than a cast: mint is the producer the whole gate\n  // rests on, so it proves the line it just built is valid instead of asserting\n  // it by fiat. The brand is unforgeable precisely because nobody casts to it.\n  return assertMinted(\n    {\n      schema: ROLLOUT_SCHEMA,\n      rollout_id: record.runId,\n      parent_rollout_id: lineage?.containingRunId ?? null,\n      run_id: record.runId,\n      experiment_id: record.experimentId,\n      candidate_id: record.candidateId,\n      generation: lineage?.depth ?? null,\n      candidate_index: lineage?.ordinal ?? null,\n      role: options.role ?? 'agent',\n      task: {\n        suite: options.suite ?? record.experimentId,\n        instance_id: record.scenarioId,\n        split: SPLIT_FROM_TAG[record.splitTag],\n        seed: record.seed,\n        rep: lineage?.rep ?? 0,\n      },\n      policy: {\n        harness: null,\n        harness_version: null,\n        model: record.model,\n        provider: null,\n        profile_commit: record.commitSha,\n        prompt_hash: record.promptHash,\n        config_hash: record.configHash,\n        agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n        sampling: null,\n      },\n      messages,\n      tool_defs: [],\n      ...(steps.length > 0 ? { steps } : {}),\n      outcome: {\n        ...rewardFields,\n        reward_source: REWARD_SOURCE[scoreOrigin(record)],\n        verdict: null,\n        // A verbatim bulk copy, deliberately UNFILTERED here. `outcome.raw`\n        // holds the per-layer verifier scores (`layer.*`) that the reward was\n        // derived from, so on a gated run this dict is the reward signal in\n        // component form — but filtering it at this call site is the pattern\n        // that has now leaked twice, because the next producer to write a\n        // reward-bearing field forgets. The gate is applied to the whole\n        // outcome once, in `assertMinted` below (`gateGamedOutcome`), which\n        // moves the block to `provenance.gated_evidence` when the run is gated\n        // and leaves it here untouched when it is not.\n        metrics: { ...record.outcome.raw },\n        is_completed: isCompleted,\n        is_truncated: isTruncated,\n        error: terminalError,\n      },\n      cost: {\n        usd: knownTotal ? record.costUsd : null,\n        tokens_in: record.tokenUsage.input,\n        tokens_out: record.tokenUsage.output,\n        tokens_reasoning: record.tokenUsage.reasoning ?? null,\n        cache_read: record.tokenUsage.cached ?? null,\n        cache_write: record.tokenUsage.cacheWrite ?? null,\n        wall_s: Math.round(record.wallMs / 1000),\n      },\n      artifacts: {\n        patch_path: null,\n        run_dir: null,\n        transcript_ref: record.traceRef ? `trace:${record.traceRef.traceId}` : null,\n      },\n      provenance: {\n        captured_at: capturedAt,\n        capture: 'mint',\n        ...(gap !== undefined ? { gap } : {}),\n      },\n    },\n    `minted rollout line for run ${record.runId}`,\n  )\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans or an explicit capture are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: MintedRolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    const captured = options.messagesOf ? await options.messagesOf(record.runId) : undefined\n    if (options.messagesOf && (!Array.isArray(captured) || captured.length === 0)) {\n      throw new ValidationError(\n        `Cannot mint rollout for run ${record.runId}: full capture is missing`,\n      )\n    }\n    const lineage = await resolveSearchLineage(record, options)\n    if (trajectory.steps.length === 0 && !options.messagesOf) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(\n          record,\n          lineage,\n          [],\n          [],\n          options,\n          capturedAt,\n          'no trace spans recorded for this runId',\n        ),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const spans = trajectory.steps.map((s) => s.span)\n    const messages = captured ?? finalConversation(spans)\n    requireToolLinkage(messages, record.runId)\n    const conversation = messages.map((message) => projectMessage(message, scrub))\n    // SFT drops copied context. That projection must not orphan a retained result.\n    requireToolLinkage(\n      conversation.filter((message) => message.is_copied_context !== true),\n      record.runId,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, lineage, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsGA,MAAM,UAAU,GAAY,UAAmC;CAE7D,OAAO,OADG,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC,MACpC,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;CACpE,MAAM,OAAoB;EACxB,MAAM,KAAK;EACX,MAAM,MAAM,KAAK,IAAI;EACrB,QAAQ,KAAK;EACb,YAAY,KAAK,YAAY,KAAA,IAAY,KAAK,UAAU,KAAK,YAAY,KAAA;CAC3E;CACA,IAAI,KAAK,SAAS,OAAO;EACvB,MAAM,MAAM;EACZ,MAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS;EAChD,IAAI,QAAQ,KAAK,YAAY,MAAM,KAAK,QAAQ,MAAM,KAAK,OAAO;EAClE,IAAI,IAAI,WAAW,KAAA,GAAW,KAAK,SAAS,MAAM,IAAI,MAAM;CAC9D,OAAO,IAAI,KAAK,SAAS,QAAQ;EAC/B,MAAM,OAAO;EACb,KAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;EACpC,IAAI,KAAK,WAAW,KAAA,GAAW,KAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;CACxE;CACA,OAAO;AACT;;AAGA,SAAS,eAAe,SAAsB,OAAqC;CACjF,OAAO;EACL,MAAM,QAAQ;EACd,SAAS,QAAQ,YAAY,OAAO,OAAO,MAAM,QAAQ,OAAO;EAChE,GAAI,QAAQ,sBAAsB,KAAA,IAC9B,EAAE,mBAAmB,MAAM,QAAQ,iBAAiB,EAAE,IACtD,CAAC;EACL,GAAI,QAAQ,eAAe,KAAA,IACvB,EACE,YAAY,QAAQ,WAAW,KAAK,UAAU;GAC5C,IAAI,KAAK;GACT,MAAM,KAAK;GACX,UAAU;IACR,MAAM,KAAK,SAAS;IACpB,WAAW,MAAM,KAAK,SAAS,SAAS;GAC1C;EACF,EAAE,EACJ,IACA,CAAC;EACL,GAAI,QAAQ,iBAAiB,KAAA,IAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;EACnF,GAAI,QAAQ,SAAS,KAAA,IAAY,EAAE,MAAM,QAAQ,KAAK,IAAI,CAAC;EAC3D,GAAI,QAAQ,sBAAsB,KAAA,IAC9B,EAAE,mBAAmB,QAAQ,kBAAkB,IAC/C,CAAC;CACP;AACF;;AAGA,SAAS,mBAAmB,UAAkC,OAAqB;CACjF,MAAM,uBAAO,IAAI,IAAY;CAC7B,MAAM,0BAAU,IAAI,IAAoB;CACxC,MAAM,UAAqC,WAAW;EACpD,MAAM,IAAI,gBAAgB,+BAA+B,MAAM,IAAI,QAAQ;CAC7E;CACA,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,eAAe,KAAA,GAAW;GACpC,IAAI,QAAQ,SAAS,eAAe,CAAC,MAAM,QAAQ,QAAQ,UAAU,GACnE,OAAO,qDAAqD;GAE9D,IAAI,QAAQ,OAAO,GAAG,OAAO,qDAAqD;GAClF,KAAK,MAAM,QAAQ,QAAQ,YAAY;IACrC,IACE,CAAC,QACD,OAAO,KAAK,OAAO,YACnB,KAAK,GAAG,WAAW,KACnB,KAAK,IAAI,KAAK,EAAE,KAChB,KAAK,SAAS,cACd,CAAC,KAAK,YACN,OAAO,KAAK,SAAS,SAAS,YAC9B,KAAK,SAAS,KAAK,WAAW,KAC9B,OAAO,KAAK,SAAS,cAAc,UAEnC,OAAO,iDAAiD;IAE1D,KAAK,IAAI,KAAK,EAAE;IAChB,QAAQ,IAAI,KAAK,IAAI,KAAK,SAAS,IAAI;GACzC;EACF;EACA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,MAAM,KAAK,QAAQ;GACnB,IAAI,OAAO,OAAO,YAAY,CAAC,QAAQ,IAAI,EAAE,GAC3C,OAAO,uEAAuE;GAEhF,IAAI,QAAQ,SAAS,KAAA,KAAa,QAAQ,SAAS,QAAQ,IAAI,EAAE,GAC/D,OAAO,gDAAgD;GAEzD,QAAQ,OAAO,EAAE;EACnB,OAAO,IAAI,QAAQ,iBAAiB,KAAA,GAClC,OAAO,6CAA6C;OAC/C,IAAI,QAAQ,OAAO,KAAK,QAAQ,eAAe,KAAA,GACpD,OAAO,6DAA6D;CAExE;CACA,IAAI,QAAQ,OAAO,GAAG,OAAO,wCAAwC;AACvE;;AAGA,SAAS,kBAAkB,OAA8B;CACvD,MAAM,OAAO,MAAM,QAAQ,MAAoB,EAAE,SAAS,KAAK;CAC/D,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,IAAI,CAAC,MAAM,OAAO,CAAC;CACnB,MAAM,WAA0B,CAAC,GAAG,KAAK,QAAQ;CACjD,IAAI,KAAK,WAAW,KAAA,KAAa,KAAK,WAAW,IAC/C,SAAS,KAAK;EAAE,MAAM;EAAa,SAAS,KAAK;CAAO,CAAC;CAE3D,OAAO;AACT;AAgBA,MAAM,gBAAgE;CACpE,SAAS;CACT,QAAQ;CACR,UAAU;AACZ;;;;;;;;;AAUA,SAAS,iBAAiB,QAAyB;CACjD,IAAI,aAAa,MAAM,MAAM,KAAA,GAC3B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,wBAAwB;AAElG;AAEA,MAAM,YAAY,UAChB,OAAO,UAAU,YAAY,UAAU;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCzC,MAAM,oBAA+C;CACnD;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,cAAc,KAAK,OAAO,IAAI,eAAe,SAAS;EACrF,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,UAAU;EACzC,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,CAAC,SAAS,IAAI,UAAU,KAAK,OAAO,IAAI,WAAW,UAAU;EAC/E,QAAQ;CACV;CACA;EACE,OAAO;EACP,UAAU,QAAQ,CAAC,SAAS,IAAI,UAAU,KAAK,OAAO,IAAI,WAAW,WAAW;EAChF,QAAQ;CACV;CACA;EACE,OAAO;EACP,UAAU,QAAQ,SAAS,IAAI,OAAO;EACtC,QACE;CACJ;CACA;EACE,OAAO;EAGP,UAAU,QAAQ,CAAC,SAAS,IAAI,OAAO,KAAK,SAAS,IAAI,QAAQ,GAAG;EACpE,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,OAAO,IAAI,oBAAoB;EACjD,QACE;CACJ;CACA;EACE,OAAO;EACP,UAAU,QAAQ,OAAO,IAAI,eAAe,YAAY,IAAI,WAAW,SAAS;EAChF,QACE;CACJ;AACF;;;;;;;;;;;;;AAcA,SAAgB,kBAAkB,QAA6B;CAC7D,MAAM,MAAM;CACZ,OAAO,kBAAkB,QAAQ,UAAU,CAAC,MAAM,QAAQ,GAAG,CAAC,CAAC,CAAC,KAC7D,UAAU,GAAG,MAAM,MAAM,eAAe,MAAM,QACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,SAAS,sBAAsB,QAAyB;CACtD,MAAM,UAAU,kBAAkB,MAAM;CACxC,IAAI,QAAQ,WAAW,GAAG;CAC1B,MAAM,IAAI,gBAAgB,+BAA+B,OAAO,MAAM,IAAI,QAAQ,KAAK,MAAM,GAAG;AAClG;;AAGA,eAAe,qBACb,QACA,SAC2C;CAC3C,IAAI,OAAO,WAAW,KAAA,GAAW,OAAO,KAAA;CACxC,MAAM,UAAqC,WAAW;EACpD,MAAM,IAAI,gBAAgB,sCAAsC,OAAO,MAAM,IAAI,QAAQ;CAC3F;CACA,IAAI,CAAC,QAAQ,eACX,OAAO,wFAAwF;CAEjG,MAAM,UAAU,MAAM,QAAQ,cAAc,OAAO,MAAM;CACzD,MAAM,WAAW,UAA4B,OAAO,cAAc,KAAK,KAAM,SAAoB;CACjG,IAAI,CAAC,QAAQ,QAAQ,KAAK,GAAG,OAAO,8CAA8C;CAClF,IAAI,CAAC,QAAQ,QAAQ,OAAO,GAAG,OAAO,gDAAgD;CACtF,IAAI,CAAC,QAAQ,QAAQ,GAAG,GAAG,OAAO,4CAA4C;CAC9E,IACE,QAAQ,oBAAoB,SAC3B,OAAO,QAAQ,oBAAoB,YAAY,QAAQ,gBAAgB,WAAW,IAEnF,OAAO,4DAA4D;CAErE,IAAI,QAAQ,oBAAoB,OAAO,OAAO,OAAO,6BAA6B;CAClF,OAAO;AACT;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,SACA,OACA,UACA,SACA,YACA,KACmB;CAUnB,sBAAsB,MAAM;CAI5B,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAE/C,MAAM,aACJ,OAAO,eAAe,SAAS,cAAc,OAAO,eAAe,SAAS;CAC9E,MAAM,kBAAkB,OAAO;CAC/B,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,cAAc,oBAAoB,eAAe,oBAAoB;CAC3E,MAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,oBAC9C;CAIN,OAAO,aACL;EACE,QAAQ;EACR,YAAY,OAAO;EACnB,mBAAmB,SAAS,mBAAmB;EAC/C,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY,SAAS,SAAS;EAC9B,iBAAiB,SAAS,WAAW;EACrC,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK,SAAS,OAAO;EACvB;EACA,QAAQ;GACN,SAAS;GACT,iBAAiB;GACjB,OAAO,OAAO;GACd,UAAU;GACV,gBAAgB,OAAO;GACvB,aAAa,OAAO;GACpB,aAAa,OAAO;GACpB,uBAAuB,OAAO,cAAc,UAAU;GACtD,UAAU;EACZ;EACA;EACA,WAAW,CAAC;EACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;EACpC,SAAS;GACP,GAAG;GACH,eAAe,cAAc,YAAY,MAAM;GAC/C,SAAS;GAUT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;GACjC,cAAc;GACd,cAAc;GACd,OAAO;EACT;EACA,MAAM;GACJ,KAAK,aAAa,OAAO,UAAU;GACnC,WAAW,OAAO,WAAW;GAC7B,YAAY,OAAO,WAAW;GAC9B,kBAAkB,OAAO,WAAW,aAAa;GACjD,YAAY,OAAO,WAAW,UAAU;GACxC,aAAa,OAAO,WAAW,cAAc;GAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;EACzC;EACA,WAAW;GACT,YAAY;GACZ,SAAS;GACT,gBAAgB,OAAO,WAAW,SAAS,OAAO,SAAS,YAAY;EACzE;EACA,YAAY;GACV,aAAa;GACb,SAAS;GACT,GAAI,QAAQ,KAAA,IAAY,EAAE,IAAI,IAAI,CAAC;EACrC;CACF,GACA,+BAA+B,OAAO,OACxC;AACF;;;;;;;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;CAC5B,MAAM,QAAQ,QAAQ,WAAW,MAAM;CACvC,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,OAA4B,CAAC;CACnC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;EAC5D,MAAM,WAAW,QAAQ,aAAa,MAAM,QAAQ,WAAW,OAAO,KAAK,IAAI,KAAA;EAC/E,IAAI,QAAQ,eAAe,CAAC,MAAM,QAAQ,QAAQ,KAAK,SAAS,WAAW,IACzE,MAAM,IAAI,gBACR,+BAA+B,OAAO,MAAM,0BAC9C;EAEF,MAAM,UAAU,MAAM,qBAAqB,QAAQ,OAAO;EAC1D,IAAI,WAAW,MAAM,WAAW,KAAK,CAAC,QAAQ,YAAY;GACxD,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SACE,QACA,SACA,CAAC,GACD,CAAC,GACD,SACA,YACA,wCACF,CACF;GACA;EACF;EACA,IAAI,QAAQ,WAAW,MAAM,KAAK,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;EAClE,IAAI,QAAQ,aAAa,KAAA,KAAa,MAAM,SAAS,QAAQ,UAAU;GAGrE,MAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;GAC3C,MAAM,OAAO,QAAQ,WAAW;GAChC,QAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;EACvE;EACA,MAAM,QAAQ,WAAW,MAAM,KAAK,MAAM,EAAE,IAAI;EAChD,MAAM,WAAW,YAAY,kBAAkB,KAAK;EACpD,mBAAmB,UAAU,OAAO,KAAK;EACzC,MAAM,eAAe,SAAS,KAAK,YAAY,eAAe,SAAS,KAAK,CAAC;EAE7E,mBACE,aAAa,QAAQ,YAAY,QAAQ,sBAAsB,IAAI,GACnE,OAAO,KACT;EACA,MAAM,MACJ,aAAa,WAAW,IAAI,uDAAuD,KAAA;EACrF,KAAK,KAAK,SAAS,QAAQ,SAAS,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CACpF;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}

package/dist/multishot/golden/index.d.ts

@@ -1,11 +1,11 @@
-import { E as MultishotResult, M as MultishotTransportRequest, T as MultishotPersona, c as RunMultishotMatrixResult, f as RunMultishotOptions, k as MultishotToolDefinition, s as RunMultishotMatrixOptions } from "../../matrix-_NLSWsKM.js";
+import { E as MultishotResult, M as MultishotTransportRequest, T as MultishotPersona, c as RunMultishotMatrixResult, f as RunMultishotOptions, k as MultishotToolDefinition, s as RunMultishotMatrixOptions } from "../../matrix--IZKROxD.js";
 //#region src/multishot/golden/compare.d.ts
 interface CompareOptions {
   /** Stop after this many mismatches. A structural divergence high in the
    *  tree would otherwise report every leaf below it. */
   limit?: number;
 }
-declare function compareJson(expected: unknown, actual: unknown, path: string, options?: CompareOptions): string[];
+export declare function compareJson(expected: unknown, actual: unknown, path: string, options?: CompareOptions): string[];
 //#endregion
 //#region src/multishot/golden/engine.d.ts
 /** A conversation engine: one profile, one persona, one transcript. */
@@ -115,7 +115,7 @@ interface MultishotMatrixGoldenScenario {
   readonly description: string;
   readonly build: (runDir: string) => MultishotMatrixGoldenCase;
 }
-declare function multishotMatrixGoldenScenarios(): MultishotMatrixGoldenScenario[];
+export declare function multishotMatrixGoldenScenarios(): MultishotMatrixGoldenScenario[];
 //#endregion
 //#region src/multishot/golden/scenarios.d.ts
 /** Options plus the ledger the scenario's transports fill while it runs. */
@@ -132,7 +132,7 @@ interface MultishotGoldenScenario {
   readonly build: () => MultishotGoldenCase;
 }
 /** Every recorded shot scenario, in record order. */
-declare function multishotGoldenScenarios(): MultishotGoldenScenario[];
+export declare function multishotGoldenScenarios(): MultishotGoldenScenario[];
 //#endregion
 //#region src/multishot/golden/harness.d.ts
 interface MultishotGoldenScenarioReport {
@@ -147,39 +147,39 @@ interface MultishotGoldenReport {
   ok: boolean;
   scenarios: MultishotGoldenScenarioReport[];
 }
-declare class MultishotGoldenMismatchError extends Error {
+export declare class MultishotGoldenMismatchError extends Error {
   readonly scenarioId: string;
   readonly mismatches: string[];
   readonly version: string;
   constructor(scenarioId: string, mismatches: string[], version: string);
 }
 /** Run one scenario and report every field that diverged from the record. */
-declare function checkMultishotGoldenScenario(opts: {
+export declare function checkMultishotGoldenScenario(opts: {
   engine: MultishotGoldenEngine;
   scenario: MultishotGoldenScenario;
   records?: MultishotGoldenRecordSet;
 }): Promise<MultishotGoldenScenarioReport>;
 /** Same as `checkMultishotGoldenScenario`, but throws on divergence. */
-declare function assertMultishotGoldenScenario(opts: {
+export declare function assertMultishotGoldenScenario(opts: {
   engine: MultishotGoldenEngine;
   scenario: MultishotGoldenScenario;
   records?: MultishotGoldenRecordSet;
 }): Promise<void>;
 /** Run every shot scenario. Never throws on divergence — read `ok`. */
-declare function checkMultishotGolden(opts: {
+export declare function checkMultishotGolden(opts: {
   engine: MultishotGoldenEngine;
   records?: MultishotGoldenRecordSet;
   only?: string[];
 }): Promise<MultishotGoldenReport>;
 /** Run one matrix scenario against `runDir` and report every divergence. */
-declare function checkMultishotMatrixGoldenScenario(opts: {
+export declare function checkMultishotMatrixGoldenScenario(opts: {
   engine: MultishotMatrixGoldenEngine;
   scenario: MultishotMatrixGoldenScenario;
   /** An empty directory the engine may write its per-cell files into. */
   runDir: string;
   records?: MultishotGoldenRecordSet;
 }): Promise<MultishotGoldenScenarioReport>;
-declare function assertMultishotMatrixGoldenScenario(opts: {
+export declare function assertMultishotMatrixGoldenScenario(opts: {
   engine: MultishotMatrixGoldenEngine;
   scenario: MultishotMatrixGoldenScenario;
   runDir: string;
@@ -187,15 +187,15 @@ declare function assertMultishotMatrixGoldenScenario(opts: {
 }): Promise<void>;
 //#endregion
 //#region src/multishot/golden/recording.d.ts
-declare function recordMessage(raw: unknown): MultishotRecordedMessage;
-declare function recordRequest(leg: 'agent' | 'driver', req: MultishotTransportRequest): MultishotRecordedRequest;
+export declare function recordMessage(raw: unknown): MultishotRecordedMessage;
+export declare function recordRequest(leg: 'agent' | 'driver', req: MultishotTransportRequest): MultishotRecordedRequest;
 /** Judge calls reach the wire as an OpenAI-compat body, not through a
  *  transport, so they are recorded from the request body the stub receives. */
-declare function recordJudgeRequest(body: Record<string, unknown>): RecordedJudgeRequest;
-declare function recordResult(result: MultishotResult): RecordedMultishotResult;
-declare function recordError(err: unknown): RecordedMultishotError;
+export declare function recordJudgeRequest(body: Record<string, unknown>): RecordedJudgeRequest;
+export declare function recordResult(result: MultishotResult): RecordedMultishotResult;
+export declare function recordError(err: unknown): RecordedMultishotError;
 /** Deep copy with every wall-clock and run-identity key removed. */
-declare function stripVolatile(value: unknown): unknown;
+export declare function stripVolatile(value: unknown): unknown;
 /** The matrix summary Markdown carries a rendered duration. Mask it so the rest
  *  of the document — cell counts, pass rate, mean, cost, the uncaptured warning
  *  — stays under comparison.
@@ -203,22 +203,22 @@ declare function stripVolatile(value: unknown): unknown;
  *  A duration the mask does not recognise would stay in the comparison and make
  *  every run mismatch on a definitionally irreproducible field, so an
  *  unmaskable duration line fails loud instead. */
-declare function maskVolatileMarkdown(text: string): string;
+export declare function maskVolatileMarkdown(text: string): string;
 /** Judge calls fan out through `Promise.all` across three slots, so their
  *  issue order is an implementation detail of the cell body, not behaviour a
  *  caller can observe. Their CONTENT is behaviour, so they are compared as a
  *  set with a stable order. */
-declare function sortJudgeRequests(requests: readonly RecordedJudgeRequest[]): RecordedJudgeRequest[];
+export declare function sortJudgeRequests(requests: readonly RecordedJudgeRequest[]): RecordedJudgeRequest[];
 /** Every file under `dir`, keyed by slash-separated relative path. JSON is
  *  parsed and stripped of wall-clock keys; Markdown keeps its text with the
  *  rendered duration masked; anything else is kept verbatim. */
-declare function readRunDir(dir: string): Record<string, unknown>;
+export declare function readRunDir(dir: string): Record<string, unknown>;
 //#endregion
 //#region src/multishot/golden/records/index.d.ts
 /** Version a check uses when the caller names none. */
-declare const CURRENT_MULTISHOT_GOLDEN_VERSION = "v1";
-declare function multishotGoldenVersions(): string[];
-declare function goldenRecords(version?: string): MultishotGoldenRecordSet;
+export declare const CURRENT_MULTISHOT_GOLDEN_VERSION = "v1";
+export declare function multishotGoldenVersions(): string[];
+export declare function goldenRecords(version?: string): MultishotGoldenRecordSet;
 //#endregion
-export { CURRENT_MULTISHOT_GOLDEN_VERSION, type CompareOptions, type MultishotGoldenCase, type MultishotGoldenEngine, MultishotGoldenMismatchError, type MultishotGoldenOutcome, type MultishotGoldenRecord, type MultishotGoldenRecordSet, type MultishotGoldenReport, type MultishotGoldenScenario, type MultishotGoldenScenarioReport, type MultishotMatrixGoldenCase, type MultishotMatrixGoldenEngine, type MultishotMatrixGoldenRecord, type MultishotMatrixGoldenScenario, type MultishotRecordedMessage, type MultishotRecordedRequest, type RecordedJudgeRequest, type RecordedMultishotError, type RecordedMultishotResult, assertMultishotGoldenScenario, assertMultishotMatrixGoldenScenario, checkMultishotGolden, checkMultishotGoldenScenario, checkMultishotMatrixGoldenScenario, compareJson, goldenRecords, maskVolatileMarkdown, multishotGoldenScenarios, multishotGoldenVersions, multishotMatrixGoldenScenarios, readRunDir, recordError, recordJudgeRequest, recordMessage, recordRequest, recordResult, sortJudgeRequests, stripVolatile };
+export type { CompareOptions, MultishotGoldenCase, MultishotGoldenEngine, MultishotGoldenOutcome, MultishotGoldenRecord, MultishotGoldenRecordSet, MultishotGoldenReport, MultishotGoldenScenario, MultishotGoldenScenarioReport, MultishotMatrixGoldenCase, MultishotMatrixGoldenEngine, MultishotMatrixGoldenRecord, MultishotMatrixGoldenScenario, MultishotRecordedMessage, MultishotRecordedRequest, RecordedJudgeRequest, RecordedMultishotError, RecordedMultishotResult };
 //# sourceMappingURL=index.d.ts.map

package/dist/multishot/golden/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../../src/multishot/golden/compare.ts","../../../src/multishot/golden/engine.ts","../../../src/multishot/golden/types.ts","../../../src/multishot/golden/matrix-scenarios.ts","../../../src/multishot/golden/scenarios.ts","../../../src/multishot/golden/harness.ts","../../../src/multishot/golden/recording.ts","../../../src/multishot/golden/records/index.ts"],"mappings":";;UAQiB;;;EAGf;;iBAGc,YACd,mBACA,iBACA,cACA,UAAS;;;;KCPC,yBACV,MAAM,oBAAoB,sBACvB,QAAQ;;KAGD,+BACV,MAAM,0BAA0B,sBAC7B,QAAQ;;;;;UCRI;EACf;EACA;;EAEA;;EAEA,YAAY;IAAQ;IAAY;IAAc;;;;;;;UAO/B;EACf;EACA;EACA;EACA;;;;;EAKA,OAAO;EACP,UAAU;;;KAIA,0BAA0B,KAAK;;;UAI1B;EACf;EACA;;;EAGA;IAAa;IAAiB;;;KAGpB;EACN;EAAgB,QAAQ;;EACxB;EAAe,OAAO;;UAEX;EACf;EACA;EACA,UAAU;EACV,SAAS;;;UAIM;EACf;EACA;EACA;EACA,UAAU;;UAGK;EACf;EACA;EACA,UAAU;EACV,eAAe;;EAEf;;;;EAIA,OAAO;;;;UAKQ;EACf;;EAEA;;EAEA;EACA;EACA,WAAW;EACX,iBAAiB;;;;UCpEF;EACf,SAAS,0BAA0B;EACnC,UAAU;;EAEV,eAAe;;UAGA;WACN;WACA;WACA,QAAQ,mBAAmB;;iBAkJtB,kCAAkC;;;;UCnJjC;EACf,SAAS,oBAAoB;;EAE7B,UAAU;;UAGK;WACN;WACA;;;WAGA,aAAa;;;iBAwWR,4BAA4B;;;UCrX3B;EACf;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA,WAAW;;cAGA,qCAAqC;WAErC;WACA;WACA;EAHX,YACW,oBACA,sBACA;;;iBAoCS,6BAA6B;EACjD,QAAQ;EACR,UAAU;EACV,UAAU;IACR,QAAQ;;iBAoCU,8BAA8B;EAClD,QAAQ;EACR,UAAU;EACV,UAAU;IACR;;iBASkB,qBAAqB;EACzC,QAAQ;EACR,UAAU;EACV;IACE,QAAQ;;iBA6BU,mCAAmC;EACvD,QAAQ;EACR,UAAU;;EAEV;EACA,UAAU;IACR,QAAQ;iBAoBU,oCAAoC;EACxD,QAAQ;EACR,UAAU;EACV;EACA,UAAU;IACR;;;iBC7KY,cAAc,eAAe;iBA0B7B,cACd,yBACA,KAAK,4BACJ;;;iBAaa,mBAAmB,MAAM,0BAA0B;iBASnD,aAAa,QAAQ,kBAAkB;iBAKvC,YAAY,eAAe;;iBAU3B,cAAc;;;;;;;;iBAoBd,qBAAqB;;;;;iBAgBrB,kBACd,mBAAmB,yBAClB;;;;iBAWa,WAAW,cAAc;;;;cCpI5B;iBAoDG;iBAIA,cACd,mBACC"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../../src/multishot/golden/compare.ts","../../../src/multishot/golden/engine.ts","../../../src/multishot/golden/types.ts","../../../src/multishot/golden/matrix-scenarios.ts","../../../src/multishot/golden/scenarios.ts","../../../src/multishot/golden/harness.ts","../../../src/multishot/golden/recording.ts","../../../src/multishot/golden/records/index.ts"],"mappings":";;UAQiB;;;EAGf;;wBAGc,YACd,mBACA,iBACA,cACA,UAAS;;;;KCPC,yBACV,MAAM,oBAAoB,sBACvB,QAAQ;;KAGD,+BACV,MAAM,0BAA0B,sBAC7B,QAAQ;;;;;UCRI;EACf;EACA;;EAEA;;EAEA,YAAY;IAAQ;IAAY;IAAc;;;;;;;UAO/B;EACf;EACA;EACA;EACA;;;;;EAKA,OAAO;EACP,UAAU;;;KAIA,0BAA0B,KAAK;;;UAI1B;EACf;EACA;;;EAGA;IAAa;IAAiB;;;KAGpB;EACN;EAAgB,QAAQ;;EACxB;EAAe,OAAO;;UAEX;EACf;EACA;EACA,UAAU;EACV,SAAS;;;UAIM;EACf;EACA;EACA;EACA,UAAU;;UAGK;EACf;EACA;EACA,UAAU;EACV,eAAe;;EAEf;;;;EAIA,OAAO;;;;UAKQ;EACf;;EAEA;;EAEA;EACA;EACA,WAAW;EACX,iBAAiB;;;;UCpEF;EACf,SAAS,0BAA0B;EACnC,UAAU;;EAEV,eAAe;;UAGA;WACN;WACA;WACA,QAAQ,mBAAmB;;wBAkJtB,kCAAkC;;;;UCnJjC;EACf,SAAS,oBAAoB;;EAE7B,UAAU;;UAGK;WACN;WACA;;;WAGA,aAAa;;;wBAwWR,4BAA4B;;;UCrX3B;EACf;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA,WAAW;;qBAGA,qCAAqC;WAErC;WACA;WACA;EAHX,YACW,oBACA,sBACA;;;wBAoCS,6BAA6B;EACjD,QAAQ;EACR,UAAU;EACV,UAAU;IACR,QAAQ;;wBAoCU,8BAA8B;EAClD,QAAQ;EACR,UAAU;EACV,UAAU;IACR;;wBASkB,qBAAqB;EACzC,QAAQ;EACR,UAAU;EACV;IACE,QAAQ;;wBA6BU,mCAAmC;EACvD,QAAQ;EACR,UAAU;;EAEV;EACA,UAAU;IACR,QAAQ;wBAoBU,oCAAoC;EACxD,QAAQ;EACR,UAAU;EACV;EACA,UAAU;IACR;;;wBC7KY,cAAc,eAAe;wBA0B7B,cACd,yBACA,KAAK,4BACJ;;;wBAaa,mBAAmB,MAAM,0BAA0B;wBASnD,aAAa,QAAQ,kBAAkB;wBAKvC,YAAY,eAAe;;wBAU3B,cAAc;;;;;;;;wBAoBd,qBAAqB;;;;;wBAgBrB,kBACd,mBAAmB,yBAClB;;;;wBAWa,WAAW,cAAc;;;;qBCpI5B;wBAoDG;wBAIA,cACd,mBACC"}

package/dist/multishot/golden/index.js

@@ -1,4 +1,4 @@
-import { n as readCellSpend } from "../../matrix-pWinRMPl.js";
+import { n as readCellSpend } from "../../matrix-DnlHn5x1.js";
 import { n as MultishotFatalToolError } from "../../types-B3jzCp0p.js";
 import { existsSync, readFileSync, readdirSync, statSync } from "node:fs";
 import { join, posix, relative, sep } from "node:path";

package/dist/multishot/index.d.ts

@@ -1,5 +1,5 @@
 import { w as JudgeScore } from "../types-DfvQOe91.js";
-import { A as MultishotToolExecutor, C as MultishotFatalToolError, D as MultishotShape, E as MultishotResult, F as assertMultishotShotResult, M as MultishotTransportRequest, N as MultishotTransportResponse, O as MultishotShotResultError, P as MultishotTransportToolCall, S as MultishotDriverEmptyError, T as MultishotPersona, _ as JudgeRunResult, a as MultishotCellOutput, b as runJudge, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as JudgeDimension, h as JudgeConfig, i as ConversationJudgeInput, j as MultishotTransport, k as MultishotToolDefinition, l as computeCellComposite, m as DEFAULT_JUDGE_MODEL, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as renderDimensions, w as MultishotMessage, x as MultishotArtifact, y as renderJsonFooter } from "../matrix-_NLSWsKM.js";
+import { A as MultishotToolExecutor, C as MultishotFatalToolError, D as MultishotShape, E as MultishotResult, F as assertMultishotShotResult, M as MultishotTransportRequest, N as MultishotTransportResponse, O as MultishotShotResultError, P as MultishotTransportToolCall, S as MultishotDriverEmptyError, T as MultishotPersona, _ as JudgeRunResult, a as MultishotCellOutput, b as runJudge, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as JudgeDimension, h as JudgeConfig, i as ConversationJudgeInput, j as MultishotTransport, k as MultishotToolDefinition, l as computeCellComposite, m as DEFAULT_JUDGE_MODEL, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as renderDimensions, w as MultishotMessage, x as MultishotArtifact, y as renderJsonFooter } from "../matrix--IZKROxD.js";
 import { AgentProfile } from "@tangle-network/agent-interface";
 //#region src/multishot/cost.d.ts
 /**
@@ -8,13 +8,13 @@ import { AgentProfile } from "@tangle-network/agent-interface";
  * never presented as a billed amount: a leg metered from this table is
  * recorded with `estimated` provenance.
  */
-declare function estimateMultishotCost(model: string, usage?: {
+export declare function estimateMultishotCost(model: string, usage?: {
   prompt_tokens?: number;
   completion_tokens?: number;
 }): number;
 //#endregion
 //#region src/multishot/default-tools.d.ts
-declare const DEFAULT_CODER_MODEL = "openai/gpt-4o-mini";
+export declare const DEFAULT_CODER_MODEL = "openai/gpt-4o-mini";
 interface DefaultResearcherConfig {
   /** Replace the system prompt to bias the researcher toward a domain's
    *  citation style. Defaults to a generic "cite sources by name" prompt. */
@@ -39,21 +39,21 @@ interface DefaultToolsBundle {
   executors: Record<string, MultishotToolExecutor>;
   artifactTypeFor: (toolName: string) => string | undefined;
 }
-declare function defaultDelegationTools(config?: DefaultToolsConfig): DefaultToolsBundle;
+export declare function defaultDelegationTools(config?: DefaultToolsConfig): DefaultToolsBundle;
 //#endregion
 //#region src/multishot/shape-defaults.d.ts
 /** Persona payload rendered as stable `- key: value` lines. `id` is identity,
  *  not voice, and structured values are serialized so nothing is dropped. */
-declare function renderPersonaFacts(persona: MultishotPersona): string;
+export declare function renderPersonaFacts(persona: MultishotPersona): string;
 /** Default opening user message: the persona introduces itself with its whole
  *  payload and asks the agent (as described by the profile) to take over. */
-declare function defaultMultishotOpener(profile: AgentProfile, persona: MultishotPersona): string;
+export declare function defaultMultishotOpener(profile: AgentProfile, persona: MultishotPersona): string;
 /** Default driver system prompt: roleplay the persona from its payload, stay
  *  in character, push back on vague answers, and never go silent. */
-declare function defaultMultishotDriverSystemPrompt(profile: AgentProfile, persona: MultishotPersona): string;
+export declare function defaultMultishotDriverSystemPrompt(profile: AgentProfile, persona: MultishotPersona): string;
 /** The fully-resolved shape for a profile: caller overrides win, the
  *  profile-derived defaults fill the rest. */
-declare function defaultShapeFromProfile<TPersona extends MultishotPersona>(profile: AgentProfile, shape?: MultishotShape<TPersona>): Required<MultishotShape<TPersona>>;
+export declare function defaultShapeFromProfile<TPersona extends MultishotPersona>(profile: AgentProfile, shape?: MultishotShape<TPersona>): Required<MultishotShape<TPersona>>;
 //#endregion
-export { type ArtifactJudgeInput, type CellCompositeInput, type CellCompositeScore, type ConversationJudgeInput, DEFAULT_CODER_MODEL, DEFAULT_JUDGE_MODEL, type DefaultCoderConfig, type DefaultResearcherConfig, type DefaultToolsBundle, type DefaultToolsConfig, type JudgeConfig, type JudgeDimension, type JudgeRunResult, type JudgeScore, type MultishotArtifact, type MultishotCellOutput, MultishotDriverEmptyError, MultishotFatalToolError, type MultishotJudges, type MultishotMessage, type MultishotPersona, type MultishotResult, type MultishotShape, type MultishotShot, MultishotShotResultError, type MultishotToolDefinition, type MultishotToolExecutor, type MultishotTransport, type MultishotTransportRequest, type MultishotTransportResponse, type MultishotTransportToolCall, type RunMultishotMatrixOptions, type RunMultishotMatrixResult, type RunMultishotOptions, assertMultishotShotResult, computeCellComposite, defaultDelegationTools, defaultMultishotDriverSystemPrompt, defaultMultishotOpener, defaultShapeFromProfile, estimateMultishotCost, renderDimensions, renderJsonFooter, renderPersonaFacts, runJudge, runMultishot, runMultishotMatrix };
+export { type ArtifactJudgeInput, type CellCompositeInput, type CellCompositeScore, type ConversationJudgeInput, DEFAULT_JUDGE_MODEL, type DefaultCoderConfig, type DefaultResearcherConfig, type DefaultToolsBundle, type DefaultToolsConfig, type JudgeConfig, type JudgeDimension, type JudgeRunResult, type JudgeScore, type MultishotArtifact, type MultishotCellOutput, MultishotDriverEmptyError, MultishotFatalToolError, type MultishotJudges, type MultishotMessage, type MultishotPersona, type MultishotResult, type MultishotShape, type MultishotShot, MultishotShotResultError, type MultishotToolDefinition, type MultishotToolExecutor, type MultishotTransport, type MultishotTransportRequest, type MultishotTransportResponse, type MultishotTransportToolCall, type RunMultishotMatrixOptions, type RunMultishotMatrixResult, type RunMultishotOptions, assertMultishotShotResult, computeCellComposite, renderDimensions, renderJsonFooter, runJudge, runMultishot, runMultishotMatrix };
 //# sourceMappingURL=index.d.ts.map

package/dist/multishot/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/multishot/cost.ts","../../src/multishot/default-tools.ts","../../src/multishot/shape-defaults.ts"],"mappings":";;;;;;;;;;iBASgB,sBACd,eACA;EAAU;EAAwB;;;;cCAvB;UAEI;;;EAGf;EACA;;UAGe;;;EAGf;EACA;;UA+Fe;EACf,WAAW;EACX,OAAO;;;EAGP;;UAGe;EACf,OAAO;EACP,WAAW,eAAe;EAC1B,kBAAkB;;iBAGJ,uBAAuB,SAAQ,qBAA0B;;;;;iBCvHzD,mBAAmB,SAAS;;;iBAe5B,uBAAuB,SAAS,cAAc,SAAS;;;iBAavD,mCACd,SAAS,cACT,SAAS;;;iBAqBK,wBAAwB,iBAAiB,kBACvD,SAAS,cACT,QAAQ,eAAe,YACtB,SAAS,eAAe"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/multishot/cost.ts","../../src/multishot/default-tools.ts","../../src/multishot/shape-defaults.ts"],"mappings":";;;;;;;;;;wBASgB,sBACd,eACA;EAAU;EAAwB;;;;qBCAvB;UAEI;;;EAGf;EACA;;UAGe;;;EAGf;EACA;;UA+Fe;EACf,WAAW;EACX,OAAO;;;EAGP;;UAGe;EACf,OAAO;EACP,WAAW,eAAe;EAC1B,kBAAkB;;wBAGJ,uBAAuB,SAAQ,qBAA0B;;;;;wBCvHzD,mBAAmB,SAAS;;;wBAe5B,uBAAuB,SAAS,cAAc,SAAS;;;wBAavD,mCACd,SAAS,cACT,SAAS;;;wBAqBK,wBAAwB,iBAAiB,kBACvD,SAAS,cACT,QAAQ,eAAe,YACtB,SAAS,eAAe"}

package/dist/multishot/index.js

@@ -1,5 +1,5 @@
 import { i as isModelPriced, n as estimateCost } from "../metrics-DOMpkoeb.js";
-import { r as withCellSpend, t as runAgentMatrix } from "../matrix-pWinRMPl.js";
+import { r as withCellSpend, t as runAgentMatrix } from "../matrix-DnlHn5x1.js";
 import { i as assertMultishotShotResult, n as MultishotFatalToolError, r as MultishotShotResultError, t as MultishotDriverEmptyError } from "../types-B3jzCp0p.js";
 import { mkdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.200.0",
+    "version": "0.201.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.1.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/{otel-spans-CQmSxtcS.js → otel-spans-C8YIJWy1.js}

@@ -98,8 +98,10 @@ function fromOtelSpans(opts) {
 			raw: rawNumeric,
 			...judgeScores ? { judgeScores } : {}
 		};
-		if (score !== void 0) if (defaultSplit === "holdout") outcome.holdoutScore = score;
-		else outcome.searchScore = score;
+		if (score !== void 0) {
+			if (defaultSplit === "holdout") outcome.holdoutScore = score;
+			else outcome.searchScore = score;
+		}
 		runs.push({
 			runId: groupKey,
 			experimentId,
@@ -217,7 +219,10 @@ function resolveTaskScore(runId, spans, scoreForRun) {
 	if (sources.length === 0) return void 0;
 	sources.sort((left, right) => left.label.localeCompare(right.label));
 	const score = sources[0].value;
-	if (sources.some((source) => source.value !== score)) throw new ValidationError(`fromOtelSpans: conflicting task-quality scores for run '${runId}': ${sources.map((source) => `${source.label}=${source.value}`).join(", ")}`);
+	if (sources.some((source) => source.value !== score)) {
+		const details = sources.map((source) => `${source.label}=${source.value}`).join(", ");
+		throw new ValidationError(`fromOtelSpans: conflicting task-quality scores for run '${runId}': ${details}`);
+	}
 	return score;
 }
 function parseTaskScoreAttribute(runId, spanId, key, value) {
@@ -261,4 +266,4 @@ function collectNumericAttrs(spans) {
 //#endregion
 export { fromOtelSpans as t };
 
-//# sourceMappingURL=otel-spans-CQmSxtcS.js.map
+//# sourceMappingURL=otel-spans-C8YIJWy1.js.map

package/dist/{otel-spans-CQmSxtcS.js.map → otel-spans-C8YIJWy1.js.map}

@@ -1 +1 @@
-{"version":3,"file":"otel-spans-CQmSxtcS.js","names":["SPAN_KIND_ATTR_KEYS"],"sources":["../src/contract/intake/otel-spans.ts"],"sourcesContent":["/**\n * # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.\n *\n * Turns an existing observability stream into the substrate-canonical\n * `RunRecord` shape so consumers with logs but no eval discipline can\n * call `analyzeRuns()` against their production traffic immediately.\n *\n * Pivot rule: spans are grouped by `tangle.runId` (the same attribute the\n * hosted-tier wire format uses) or, when absent, by `traceId`. One group\n * becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:\n *\n *   - `runId` (the group key)\n *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`\n *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`\n *   - task failure class and detail from explicit `tangle.task.*` attributes\n *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`\n *   - token usage from model-call input, output, cache-read, and cache-write\n *     attributes without double-counting aggregate parent spans\n *   - task quality from an explicit `scoreForRun` callback or a designated\n *     evaluation attribute on a root / `EVALUATOR` span; `outcome.raw`\n *     collects every numeric attribute without promoting it to task quality.\n *\n * Errored tool, model, and child-agent spans contribute to execution-error\n * counts. Root process, guardrail, evaluator, propagated parent, and unknown\n * errors retain separate counters. Only one failed root can set\n * `RunRecord.terminalOutcome` and `RunRecord.terminalFailureReason`; a child\n * error cannot become a task failure.\n */\n\nimport { resolveSpanKind } from '@tangle-network/agent-trace-contract'\nimport { ValidationError } from '../../errors'\nimport type { TraceSpanEvent } from '../../hosted/types'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../../run-record'\nimport { summarizeTraceErrors, type TraceErrorRole } from '../../trace/error-classification'\nimport {\n  recordAggregateMeasurements,\n  summarizeExecutionMeasurements,\n} from '../../trace/execution-measurements'\nimport { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from '../../trace/otlp-attributes'\nimport { readTaskFailureLabels } from '../../trace/task-failure-attributes'\n\nconst TASK_SCORE_ATTR_KEYS = [\n  'gen_ai.evaluation.score.value',\n  'tangle.task.score',\n  'eval.score',\n  'tangle.score',\n] as const\nconst MODEL_KEYS = ['tangle.model', ...LLM_MODEL_ATTR_KEYS, 'model']\nconst PROMPT_HASH_KEYS = ['tangle.prompt_hash', 'prompt.hash']\nconst CONFIG_HASH_KEYS = ['tangle.config_hash', 'config.hash']\n\nexport interface FromOtelSpansOptions {\n  spans: TraceSpanEvent[]\n  /** Default split tag for synthesized records. Defaults to `'holdout'`. */\n  defaultSplit?: RunSplitTag\n  /** Default `experimentId` when not present on any span. */\n  experimentId?: string\n  /**\n   * Explicit task-quality score for a logical run. The callback receives\n   * spans in deterministic time/id order. Its value must agree with any\n   * designated score attributes present on root or `EVALUATOR` spans.\n   */\n  scoreForRun?: (runId: string, spans: readonly TraceSpanEvent[]) => number | undefined\n}\n\nexport function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[] {\n  const { spans, defaultSplit = 'holdout', experimentId = 'otel-corpus' } = opts\n  const grouped = groupSpans(spans)\n\n  const runs: RunRecord[] = []\n  for (const [groupKey, groupSpans] of grouped) {\n    const root = findRoot(groupSpans)\n    if (!root) continue\n    const measurements = summarizeExecutionMeasurements(\n      groupSpans.map((span) => ({\n        id: span.spanId,\n        ...(span.parentSpanId ? { parentId: span.parentSpanId } : {}),\n        attributes: span.attributes,\n        kind: errorRoleForSpan(span),\n      })),\n    )\n    const callSpanIds = new Set(measurements.callSpanIds)\n    const callSpans = groupSpans.filter((span) => callSpanIds.has(span.spanId))\n    const wallMs = unixNanoDurationMs(root.startTimeUnixNano, root.endTimeUnixNano)\n    const model =\n      readAttrString(callSpans, MODEL_KEYS) ??\n      readAttrString(groupSpans, MODEL_KEYS) ??\n      'unknown@unknown'\n    const capturedCost =\n      (measurements.cost.complete ? measurements.cost.value : undefined) ??\n      measurements.aggregate?.costUsd\n    const costUsd = capturedCost ?? null\n    const scenarioId = readConsistentScenarioId(groupKey, groupSpans) ?? groupKey\n    const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? 'sha256:unknown'\n    const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? 'sha256:unknown'\n    const score = resolveTaskScore(groupKey, groupSpans, opts.scoreForRun)\n    const taskFailure = readTaskFailureLabels(\n      groupSpans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span)),\n      `fromOtelSpans: run '${groupKey}'`,\n    )\n\n    const rawNumeric = collectNumericAttrs(groupSpans)\n    const errorSummary = summarizeTraceErrors(\n      groupSpans.map((span) => ({\n        id: spanIdentity(span),\n        ...(span.parentSpanId ? { parentId: parentIdentity(span) } : {}),\n        role: errorRoleForSpan(span),\n        error: span.status?.code === 'ERROR',\n        processRoot: !span.parentSpanId && isTerminalRootCandidate(span),\n      })),\n    )\n    rawNumeric.error_span_count = errorSummary.total\n    rawNumeric.execution_error_count = errorSummary.execution\n    rawNumeric.process_error_count = errorSummary.process\n    rawNumeric.guardrail_error_count = errorSummary.guardrail\n    rawNumeric.judge_error_count = errorSummary.evaluation\n    rawNumeric.propagated_error_count = errorSummary.propagated\n    rawNumeric.unclassified_error_count = errorSummary.unclassified\n    rawNumeric.llm_span_count = measurements.modelCallCount\n    if (measurements.cost.value !== undefined && !measurements.cost.complete) {\n      rawNumeric.partial_observed_cost_usd = measurements.cost.value\n    }\n    recordAggregateMeasurements(rawNumeric, measurements.aggregate)\n\n    const judgeScores: JudgeScoresRecord | undefined =\n      score !== undefined\n        ? {\n            perJudge: { 'otel-derived': { score } },\n            perDimMean: { score },\n            composite: score,\n          }\n        : undefined\n\n    const terminalOutcome = terminalOutcomeFromRoots(groupSpans)\n    const failedRoot =\n      terminalOutcome === 'failed'\n        ? groupSpans.find(\n            (span) =>\n              !span.parentSpanId && isTerminalRootCandidate(span) && span.status?.code === 'ERROR',\n          )\n        : undefined\n    const outcome: RunOutcome = {\n      raw: rawNumeric,\n      ...(judgeScores ? { judgeScores } : {}),\n    }\n    if (score !== undefined) {\n      if (defaultSplit === 'holdout') outcome.holdoutScore = score\n      else outcome.searchScore = score\n    }\n\n    runs.push({\n      runId: groupKey,\n      experimentId,\n      candidateId: (root.attributes['tangle.candidateId'] as string | undefined) ?? 'otel-default',\n      seed: 0,\n      model,\n      promptHash,\n      configHash,\n      commitSha: (root.attributes['tangle.commit_sha'] as string | undefined) ?? 'unknown',\n      wallMs,\n      costUsd,\n      costProvenance:\n        capturedCost === undefined\n          ? { kind: 'uncaptured', usd: null }\n          : { kind: 'observed', usd: capturedCost },\n      tokenUsage: measurements.tokenUsage,\n      terminalOutcome,\n      ...(failedRoot\n        ? { terminalFailureReason: failedRoot.status?.message ?? failedRoot.name }\n        : {}),\n      outcome,\n      ...taskFailure,\n      splitTag: defaultSplit,\n      scenarioId,\n    })\n  }\n  return runs\n}\n\nfunction terminalOutcomeFromRoots(spans: TraceSpanEvent[]): RunTerminalOutcome {\n  const roots = spans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span))\n  if (roots.length !== 1) return 'unknown'\n  if (roots[0]!.status?.code === 'OK') return 'succeeded'\n  if (roots[0]!.status?.code === 'ERROR') return 'failed'\n  return 'unknown'\n}\n\nfunction isTerminalRootCandidate(span: TraceSpanEvent): boolean {\n  const role = errorRoleForSpan(span)\n  if (role === 'LLM' || role === 'TOOL' || role === 'GUARDRAIL' || role === 'EVALUATOR') {\n    return false\n  }\n  return true\n}\n\nfunction readSpanKind(span: TraceSpanEvent): string | undefined {\n  return readAttrString([span], [...SPAN_KIND_ATTR_KEYS, 'span.kind'])?.toUpperCase()\n}\n\nfunction errorRoleForSpan(span: TraceSpanEvent): TraceErrorRole {\n  return resolveSpanKind({\n    kind: readSpanKind(span),\n    name: span.name,\n    attributes: span.attributes,\n  })\n}\n\nfunction spanIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.spanId}`\n}\n\nfunction parentIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.parentSpanId}`\n}\n\n// ── Internal helpers ────────────────────────────────────────────────\n\nfunction groupSpans(spans: TraceSpanEvent[]): Map<string, TraceSpanEvent[]> {\n  const m = new Map<string, TraceSpanEvent[]>()\n  for (const span of spans) {\n    const key = (span['tangle.runId'] as string | undefined) ?? span.traceId\n    const list = m.get(key) ?? []\n    list.push(span)\n    m.set(key, list)\n  }\n  return m\n}\n\nfunction findRoot(group: TraceSpanEvent[]): TraceSpanEvent | undefined {\n  const structuralRoots = group.filter((span) => !span.parentSpanId)\n  const terminalRoots = structuralRoots.filter(isTerminalRootCandidate)\n  const pool =\n    terminalRoots.length > 0 ? terminalRoots : structuralRoots.length > 0 ? structuralRoots : group\n  return orderSpans(pool)[0]\n}\n\nfunction readAttrString(spans: TraceSpanEvent[], keys: string[]): string | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'string' && v.length > 0) return v\n    }\n  }\n  return undefined\n}\n\nfunction readConsistentScenarioId(\n  runId: string,\n  spans: readonly TraceSpanEvent[],\n): string | undefined {\n  const values = new Set<string>()\n  for (const span of spans) {\n    const topLevel = span['tangle.scenarioId']\n    if (typeof topLevel === 'string' && topLevel.length > 0) values.add(topLevel)\n    const attribute = span.attributes['tangle.scenarioId']\n    if (typeof attribute === 'string' && attribute.length > 0) values.add(attribute)\n  }\n  if (values.size > 1) {\n    throw new ValidationError(\n      `fromOtelSpans: conflicting scenario ids for run '${runId}': ${[...values].sort().join(', ')}`,\n    )\n  }\n  return values.values().next().value\n}\n\ninterface TaskScoreSource {\n  label: string\n  value: number\n}\n\nfunction resolveTaskScore(\n  runId: string,\n  spans: TraceSpanEvent[],\n  scoreForRun: FromOtelSpansOptions['scoreForRun'],\n): number | undefined {\n  const orderedSpans = orderSpans(spans)\n  const sources: TaskScoreSource[] = []\n\n  if (scoreForRun) {\n    const supplied: unknown = scoreForRun(runId, orderedSpans)\n    if (supplied !== undefined) {\n      if (typeof supplied !== 'number' || !Number.isFinite(supplied)) {\n        throw new ValidationError(\n          `fromOtelSpans: scoreForRun returned a non-finite number for run '${runId}'`,\n        )\n      }\n      sources.push({ label: 'scoreForRun', value: supplied })\n    }\n  }\n\n  for (const span of orderedSpans) {\n    const role = errorRoleForSpan(span)\n    if (span.parentSpanId && role !== 'EVALUATOR') continue\n    if (role === 'EVALUATOR' && span.status?.code === 'ERROR') continue\n    for (const key of TASK_SCORE_ATTR_KEYS) {\n      if (!Object.hasOwn(span.attributes, key)) continue\n      sources.push({\n        label: `span '${span.spanId}' attribute '${key}'`,\n        value: parseTaskScoreAttribute(runId, span.spanId, key, span.attributes[key]),\n      })\n    }\n  }\n\n  if (sources.length === 0) return undefined\n  sources.sort((left, right) => left.label.localeCompare(right.label))\n  const score = sources[0]!.value\n  if (sources.some((source) => source.value !== score)) {\n    const details = sources.map((source) => `${source.label}=${source.value}`).join(', ')\n    throw new ValidationError(\n      `fromOtelSpans: conflicting task-quality scores for run '${runId}': ${details}`,\n    )\n  }\n  return score\n}\n\nfunction parseTaskScoreAttribute(\n  runId: string,\n  spanId: string,\n  key: string,\n  value: unknown,\n): number {\n  const source = `span '${spanId}' attribute '${key}'`\n  if (typeof value === 'string') {\n    if (value.trim().length === 0) {\n      throw new ValidationError(\n        `fromOtelSpans: ${source} is blank for run '${runId}'; task quality must be finite`,\n      )\n    }\n    const parsed = Number(value)\n    if (Number.isFinite(parsed)) return parsed\n  } else if (typeof value === 'number' && Number.isFinite(value)) {\n    return value\n  }\n  throw new ValidationError(\n    `fromOtelSpans: ${source} is not a finite task-quality score for run '${runId}'`,\n  )\n}\n\nfunction orderSpans(spans: readonly TraceSpanEvent[]): TraceSpanEvent[] {\n  return [...spans].sort(\n    (left, right) =>\n      compareUnixNano(left.startTimeUnixNano, right.startTimeUnixNano) ||\n      left.spanId.localeCompare(right.spanId),\n  )\n}\n\nfunction parseUnixNano(value: string): bigint {\n  try {\n    return BigInt(value)\n  } catch {\n    throw new ValidationError(`fromOtelSpans: invalid Unix nanosecond timestamp '${value}'`)\n  }\n}\n\nfunction compareUnixNano(left: string, right: string): number {\n  const leftValue = parseUnixNano(left)\n  const rightValue = parseUnixNano(right)\n  return leftValue < rightValue ? -1 : leftValue > rightValue ? 1 : 0\n}\n\nfunction unixNanoDurationMs(start: string, end: string): number {\n  const delta = parseUnixNano(end) - parseUnixNano(start)\n  if (delta <= 0n) return 0\n  const wholeMs = delta / 1_000_000n\n  const fractionalMs = delta % 1_000_000n\n  const value = Number(wholeMs) + Number(fractionalMs) / 1_000_000\n  if (!Number.isSafeInteger(Number(wholeMs))) {\n    throw new ValidationError('fromOtelSpans: span duration exceeds the safe millisecond range')\n  }\n  return value\n}\n\nfunction collectNumericAttrs(spans: TraceSpanEvent[]): Record<string, number> {\n  const raw: Record<string, number> = {}\n  for (const span of spans) {\n    for (const [k, v] of Object.entries(span.attributes)) {\n      if (typeof v === 'number' && Number.isFinite(v)) raw[k] = v\n    }\n  }\n  return raw\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+CA,MAAM,uBAAuB;CAC3B;CACA;CACA;CACA;AACF;AACA,MAAM,aAAa;CAAC;CAAgB,GAAG;CAAqB;AAAO;AACnE,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAC7D,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAgB7D,SAAgB,cAAc,MAAyC;CACrE,MAAM,EAAE,OAAO,eAAe,WAAW,eAAe,kBAAkB;CAC1E,MAAM,UAAU,WAAW,KAAK;CAEhC,MAAM,OAAoB,CAAC;CAC3B,KAAK,MAAM,CAAC,UAAU,eAAe,SAAS;EAC5C,MAAM,OAAO,SAAS,UAAU;EAChC,IAAI,CAAC,MAAM;EACX,MAAM,eAAe,+BACnB,WAAW,KAAK,UAAU;GACxB,IAAI,KAAK;GACT,GAAI,KAAK,eAAe,EAAE,UAAU,KAAK,aAAa,IAAI,CAAC;GAC3D,YAAY,KAAK;GACjB,MAAM,iBAAiB,IAAI;EAC7B,EAAE,CACJ;EACA,MAAM,cAAc,IAAI,IAAI,aAAa,WAAW;EACpD,MAAM,YAAY,WAAW,QAAQ,SAAS,YAAY,IAAI,KAAK,MAAM,CAAC;EAC1E,MAAM,SAAS,mBAAmB,KAAK,mBAAmB,KAAK,eAAe;EAC9E,MAAM,QACJ,eAAe,WAAW,UAAU,KACpC,eAAe,YAAY,UAAU,KACrC;EACF,MAAM,gBACH,aAAa,KAAK,WAAW,aAAa,KAAK,QAAQ,KAAA,MACxD,aAAa,WAAW;EAC1B,MAAM,UAAU,gBAAgB;EAChC,MAAM,aAAa,yBAAyB,UAAU,UAAU,KAAK;EACrE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,QAAQ,iBAAiB,UAAU,YAAY,KAAK,WAAW;EACrE,MAAM,cAAc,sBAClB,WAAW,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC,GAC/E,uBAAuB,SAAS,EAClC;EAEA,MAAM,aAAa,oBAAoB,UAAU;EACjD,MAAM,eAAe,qBACnB,WAAW,KAAK,UAAU;GACxB,IAAI,aAAa,IAAI;GACrB,GAAI,KAAK,eAAe,EAAE,UAAU,eAAe,IAAI,EAAE,IAAI,CAAC;GAC9D,MAAM,iBAAiB,IAAI;GAC3B,OAAO,KAAK,QAAQ,SAAS;GAC7B,aAAa,CAAC,KAAK,gBAAgB,wBAAwB,IAAI;EACjE,EAAE,CACJ;EACA,WAAW,mBAAmB,aAAa;EAC3C,WAAW,wBAAwB,aAAa;EAChD,WAAW,sBAAsB,aAAa;EAC9C,WAAW,wBAAwB,aAAa;EAChD,WAAW,oBAAoB,aAAa;EAC5C,WAAW,yBAAyB,aAAa;EACjD,WAAW,2BAA2B,aAAa;EACnD,WAAW,iBAAiB,aAAa;EACzC,IAAI,aAAa,KAAK,UAAU,KAAA,KAAa,CAAC,aAAa,KAAK,UAC9D,WAAW,4BAA4B,aAAa,KAAK;EAE3D,4BAA4B,YAAY,aAAa,SAAS;EAE9D,MAAM,cACJ,UAAU,KAAA,IACN;GACE,UAAU,EAAE,gBAAgB,EAAE,MAAM,EAAE;GACtC,YAAY,EAAE,MAAM;GACpB,WAAW;EACb,IACA,KAAA;EAEN,MAAM,kBAAkB,yBAAyB,UAAU;EAC3D,MAAM,aACJ,oBAAoB,WAChB,WAAW,MACR,SACC,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,KAAK,KAAK,QAAQ,SAAS,OACjF,IACA,KAAA;EACN,MAAM,UAAsB;GAC1B,KAAK;GACL,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;EACvC;EACA,IAAI,UAAU,KAAA,GACZ,IAAI,iBAAiB,WAAW,QAAQ,eAAe;OAClD,QAAQ,cAAc;EAG7B,KAAK,KAAK;GACR,OAAO;GACP;GACA,aAAc,KAAK,WAAW,yBAAgD;GAC9E,MAAM;GACN;GACA;GACA;GACA,WAAY,KAAK,WAAW,wBAA+C;GAC3E;GACA;GACA,gBACE,iBAAiB,KAAA,IACb;IAAE,MAAM;IAAc,KAAK;GAAK,IAChC;IAAE,MAAM;IAAY,KAAK;GAAa;GAC5C,YAAY,aAAa;GACzB;GACA,GAAI,aACA,EAAE,uBAAuB,WAAW,QAAQ,WAAW,WAAW,KAAK,IACvE,CAAC;GACL;GACA,GAAG;GACH,UAAU;GACV;EACF,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,OAA6C;CAC7E,MAAM,QAAQ,MAAM,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC;CACxF,IAAI,MAAM,WAAW,GAAG,OAAO;CAC/B,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,MAAM,OAAO;CAC5C,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,SAAS,OAAO;CAC/C,OAAO;AACT;AAEA,SAAS,wBAAwB,MAA+B;CAC9D,MAAM,OAAO,iBAAiB,IAAI;CAClC,IAAI,SAAS,SAAS,SAAS,UAAU,SAAS,eAAe,SAAS,aACxE,OAAO;CAET,OAAO;AACT;AAEA,SAAS,aAAa,MAA0C;CAC9D,OAAO,eAAe,CAAC,IAAI,GAAG,CAAC,GAAGA,uBAAqB,WAAW,CAAC,CAAC,EAAE,YAAY;AACpF;AAEA,SAAS,iBAAiB,MAAsC;CAC9D,OAAO,gBAAgB;EACrB,MAAM,aAAa,IAAI;EACvB,MAAM,KAAK;EACX,YAAY,KAAK;CACnB,CAAC;AACH;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAEA,SAAS,eAAe,MAA8B;CACpD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAIA,SAAS,WAAW,OAAwD;CAC1E,MAAM,oBAAI,IAAI,IAA8B;CAC5C,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,MAAO,KAAK,mBAA0C,KAAK;EACjE,MAAM,OAAO,EAAE,IAAI,GAAG,KAAK,CAAC;EAC5B,KAAK,KAAK,IAAI;EACd,EAAE,IAAI,KAAK,IAAI;CACjB;CACA,OAAO;AACT;AAEA,SAAS,SAAS,OAAqD;CACrE,MAAM,kBAAkB,MAAM,QAAQ,SAAS,CAAC,KAAK,YAAY;CACjE,MAAM,gBAAgB,gBAAgB,OAAO,uBAAuB;CAGpE,OAAO,WADL,cAAc,SAAS,IAAI,gBAAgB,gBAAgB,SAAS,IAAI,kBAAkB,KACtE,CAAC,CAAC;AAC1B;AAEA,SAAS,eAAe,OAAyB,MAAoC;CACnF,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,IAAI,KAAK,WAAW;EAC1B,IAAI,OAAO,MAAM,YAAY,EAAE,SAAS,GAAG,OAAO;CACpD;AAGJ;AAEA,SAAS,yBACP,OACA,OACoB;CACpB,MAAM,yBAAS,IAAI,IAAY;CAC/B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,WAAW,KAAK;EACtB,IAAI,OAAO,aAAa,YAAY,SAAS,SAAS,GAAG,OAAO,IAAI,QAAQ;EAC5E,MAAM,YAAY,KAAK,WAAW;EAClC,IAAI,OAAO,cAAc,YAAY,UAAU,SAAS,GAAG,OAAO,IAAI,SAAS;CACjF;CACA,IAAI,OAAO,OAAO,GAChB,MAAM,IAAI,gBACR,oDAAoD,MAAM,KAAK,CAAC,GAAG,MAAM,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC7F;CAEF,OAAO,OAAO,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC;AAChC;AAOA,SAAS,iBACP,OACA,OACA,aACoB;CACpB,MAAM,eAAe,WAAW,KAAK;CACrC,MAAM,UAA6B,CAAC;CAEpC,IAAI,aAAa;EACf,MAAM,WAAoB,YAAY,OAAO,YAAY;EACzD,IAAI,aAAa,KAAA,GAAW;GAC1B,IAAI,OAAO,aAAa,YAAY,CAAC,OAAO,SAAS,QAAQ,GAC3D,MAAM,IAAI,gBACR,oEAAoE,MAAM,EAC5E;GAEF,QAAQ,KAAK;IAAE,OAAO;IAAe,OAAO;GAAS,CAAC;EACxD;CACF;CAEA,KAAK,MAAM,QAAQ,cAAc;EAC/B,MAAM,OAAO,iBAAiB,IAAI;EAClC,IAAI,KAAK,gBAAgB,SAAS,aAAa;EAC/C,IAAI,SAAS,eAAe,KAAK,QAAQ,SAAS,SAAS;EAC3D,KAAK,MAAM,OAAO,sBAAsB;GACtC,IAAI,CAAC,OAAO,OAAO,KAAK,YAAY,GAAG,GAAG;GAC1C,QAAQ,KAAK;IACX,OAAO,SAAS,KAAK,OAAO,eAAe,IAAI;IAC/C,OAAO,wBAAwB,OAAO,KAAK,QAAQ,KAAK,KAAK,WAAW,IAAI;GAC9E,CAAC;EACH;CACF;CAEA,IAAI,QAAQ,WAAW,GAAG,OAAO,KAAA;CACjC,QAAQ,MAAM,MAAM,UAAU,KAAK,MAAM,cAAc,MAAM,KAAK,CAAC;CACnE,MAAM,QAAQ,QAAQ,EAAE,CAAE;CAC1B,IAAI,QAAQ,MAAM,WAAW,OAAO,UAAU,KAAK,GAEjD,MAAM,IAAI,gBACR,2DAA2D,MAAM,KAFnD,QAAQ,KAAK,WAAW,GAAG,OAAO,MAAM,GAAG,OAAO,OAAO,CAAC,CAAC,KAAK,IAEF,GAC9E;CAEF,OAAO;AACT;AAEA,SAAS,wBACP,OACA,QACA,KACA,OACQ;CACR,MAAM,SAAS,SAAS,OAAO,eAAe,IAAI;CAClD,IAAI,OAAO,UAAU,UAAU;EAC7B,IAAI,MAAM,KAAK,CAAC,CAAC,WAAW,GAC1B,MAAM,IAAI,gBACR,kBAAkB,OAAO,qBAAqB,MAAM,+BACtD;EAEF,MAAM,SAAS,OAAO,KAAK;EAC3B,IAAI,OAAO,SAAS,MAAM,GAAG,OAAO;CACtC,OAAO,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,GAC3D,OAAO;CAET,MAAM,IAAI,gBACR,kBAAkB,OAAO,+CAA+C,MAAM,EAChF;AACF;AAEA,SAAS,WAAW,OAAoD;CACtE,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MACf,MAAM,UACL,gBAAgB,KAAK,mBAAmB,MAAM,iBAAiB,KAC/D,KAAK,OAAO,cAAc,MAAM,MAAM,CAC1C;AACF;AAEA,SAAS,cAAc,OAAuB;CAC5C,IAAI;EACF,OAAO,OAAO,KAAK;CACrB,QAAQ;EACN,MAAM,IAAI,gBAAgB,qDAAqD,MAAM,EAAE;CACzF;AACF;AAEA,SAAS,gBAAgB,MAAc,OAAuB;CAC5D,MAAM,YAAY,cAAc,IAAI;CACpC,MAAM,aAAa,cAAc,KAAK;CACtC,OAAO,YAAY,aAAa,KAAK,YAAY,aAAa,IAAI;AACpE;AAEA,SAAS,mBAAmB,OAAe,KAAqB;CAC9D,MAAM,QAAQ,cAAc,GAAG,IAAI,cAAc,KAAK;CACtD,IAAI,SAAS,IAAI,OAAO;CACxB,MAAM,UAAU,QAAQ;CACxB,MAAM,eAAe,QAAQ;CAC7B,MAAM,QAAQ,OAAO,OAAO,IAAI,OAAO,YAAY,IAAI;CACvD,IAAI,CAAC,OAAO,cAAc,OAAO,OAAO,CAAC,GACvC,MAAM,IAAI,gBAAgB,iEAAiE;CAE7F,OAAO;AACT;AAEA,SAAS,oBAAoB,OAAiD;CAC5E,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,KAAK,UAAU,GACjD,IAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK;CAG9D,OAAO;AACT"}
+{"version":3,"file":"otel-spans-C8YIJWy1.js","names":["SPAN_KIND_ATTR_KEYS"],"sources":["../src/contract/intake/otel-spans.ts"],"sourcesContent":["/**\n * # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.\n *\n * Turns an existing observability stream into the substrate-canonical\n * `RunRecord` shape so consumers with logs but no eval discipline can\n * call `analyzeRuns()` against their production traffic immediately.\n *\n * Pivot rule: spans are grouped by `tangle.runId` (the same attribute the\n * hosted-tier wire format uses) or, when absent, by `traceId`. One group\n * becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:\n *\n *   - `runId` (the group key)\n *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`\n *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`\n *   - task failure class and detail from explicit `tangle.task.*` attributes\n *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`\n *   - token usage from model-call input, output, cache-read, and cache-write\n *     attributes without double-counting aggregate parent spans\n *   - task quality from an explicit `scoreForRun` callback or a designated\n *     evaluation attribute on a root / `EVALUATOR` span; `outcome.raw`\n *     collects every numeric attribute without promoting it to task quality.\n *\n * Errored tool, model, and child-agent spans contribute to execution-error\n * counts. Root process, guardrail, evaluator, propagated parent, and unknown\n * errors retain separate counters. Only one failed root can set\n * `RunRecord.terminalOutcome` and `RunRecord.terminalFailureReason`; a child\n * error cannot become a task failure.\n */\n\nimport { resolveSpanKind } from '@tangle-network/agent-trace-contract'\nimport { ValidationError } from '../../errors'\nimport type { TraceSpanEvent } from '../../hosted/types'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../../run-record'\nimport { summarizeTraceErrors, type TraceErrorRole } from '../../trace/error-classification'\nimport {\n  recordAggregateMeasurements,\n  summarizeExecutionMeasurements,\n} from '../../trace/execution-measurements'\nimport { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from '../../trace/otlp-attributes'\nimport { readTaskFailureLabels } from '../../trace/task-failure-attributes'\n\nconst TASK_SCORE_ATTR_KEYS = [\n  'gen_ai.evaluation.score.value',\n  'tangle.task.score',\n  'eval.score',\n  'tangle.score',\n] as const\nconst MODEL_KEYS = ['tangle.model', ...LLM_MODEL_ATTR_KEYS, 'model']\nconst PROMPT_HASH_KEYS = ['tangle.prompt_hash', 'prompt.hash']\nconst CONFIG_HASH_KEYS = ['tangle.config_hash', 'config.hash']\n\nexport interface FromOtelSpansOptions {\n  spans: TraceSpanEvent[]\n  /** Default split tag for synthesized records. Defaults to `'holdout'`. */\n  defaultSplit?: RunSplitTag\n  /** Default `experimentId` when not present on any span. */\n  experimentId?: string\n  /**\n   * Explicit task-quality score for a logical run. The callback receives\n   * spans in deterministic time/id order. Its value must agree with any\n   * designated score attributes present on root or `EVALUATOR` spans.\n   */\n  scoreForRun?: (runId: string, spans: readonly TraceSpanEvent[]) => number | undefined\n}\n\nexport function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[] {\n  const { spans, defaultSplit = 'holdout', experimentId = 'otel-corpus' } = opts\n  const grouped = groupSpans(spans)\n\n  const runs: RunRecord[] = []\n  for (const [groupKey, groupSpans] of grouped) {\n    const root = findRoot(groupSpans)\n    if (!root) continue\n    const measurements = summarizeExecutionMeasurements(\n      groupSpans.map((span) => ({\n        id: span.spanId,\n        ...(span.parentSpanId ? { parentId: span.parentSpanId } : {}),\n        attributes: span.attributes,\n        kind: errorRoleForSpan(span),\n      })),\n    )\n    const callSpanIds = new Set(measurements.callSpanIds)\n    const callSpans = groupSpans.filter((span) => callSpanIds.has(span.spanId))\n    const wallMs = unixNanoDurationMs(root.startTimeUnixNano, root.endTimeUnixNano)\n    const model =\n      readAttrString(callSpans, MODEL_KEYS) ??\n      readAttrString(groupSpans, MODEL_KEYS) ??\n      'unknown@unknown'\n    const capturedCost =\n      (measurements.cost.complete ? measurements.cost.value : undefined) ??\n      measurements.aggregate?.costUsd\n    const costUsd = capturedCost ?? null\n    const scenarioId = readConsistentScenarioId(groupKey, groupSpans) ?? groupKey\n    const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? 'sha256:unknown'\n    const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? 'sha256:unknown'\n    const score = resolveTaskScore(groupKey, groupSpans, opts.scoreForRun)\n    const taskFailure = readTaskFailureLabels(\n      groupSpans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span)),\n      `fromOtelSpans: run '${groupKey}'`,\n    )\n\n    const rawNumeric = collectNumericAttrs(groupSpans)\n    const errorSummary = summarizeTraceErrors(\n      groupSpans.map((span) => ({\n        id: spanIdentity(span),\n        ...(span.parentSpanId ? { parentId: parentIdentity(span) } : {}),\n        role: errorRoleForSpan(span),\n        error: span.status?.code === 'ERROR',\n        processRoot: !span.parentSpanId && isTerminalRootCandidate(span),\n      })),\n    )\n    rawNumeric.error_span_count = errorSummary.total\n    rawNumeric.execution_error_count = errorSummary.execution\n    rawNumeric.process_error_count = errorSummary.process\n    rawNumeric.guardrail_error_count = errorSummary.guardrail\n    rawNumeric.judge_error_count = errorSummary.evaluation\n    rawNumeric.propagated_error_count = errorSummary.propagated\n    rawNumeric.unclassified_error_count = errorSummary.unclassified\n    rawNumeric.llm_span_count = measurements.modelCallCount\n    if (measurements.cost.value !== undefined && !measurements.cost.complete) {\n      rawNumeric.partial_observed_cost_usd = measurements.cost.value\n    }\n    recordAggregateMeasurements(rawNumeric, measurements.aggregate)\n\n    const judgeScores: JudgeScoresRecord | undefined =\n      score !== undefined\n        ? {\n            perJudge: { 'otel-derived': { score } },\n            perDimMean: { score },\n            composite: score,\n          }\n        : undefined\n\n    const terminalOutcome = terminalOutcomeFromRoots(groupSpans)\n    const failedRoot =\n      terminalOutcome === 'failed'\n        ? groupSpans.find(\n            (span) =>\n              !span.parentSpanId && isTerminalRootCandidate(span) && span.status?.code === 'ERROR',\n          )\n        : undefined\n    const outcome: RunOutcome = {\n      raw: rawNumeric,\n      ...(judgeScores ? { judgeScores } : {}),\n    }\n    if (score !== undefined) {\n      if (defaultSplit === 'holdout') outcome.holdoutScore = score\n      else outcome.searchScore = score\n    }\n\n    runs.push({\n      runId: groupKey,\n      experimentId,\n      candidateId: (root.attributes['tangle.candidateId'] as string | undefined) ?? 'otel-default',\n      seed: 0,\n      model,\n      promptHash,\n      configHash,\n      commitSha: (root.attributes['tangle.commit_sha'] as string | undefined) ?? 'unknown',\n      wallMs,\n      costUsd,\n      costProvenance:\n        capturedCost === undefined\n          ? { kind: 'uncaptured', usd: null }\n          : { kind: 'observed', usd: capturedCost },\n      tokenUsage: measurements.tokenUsage,\n      terminalOutcome,\n      ...(failedRoot\n        ? { terminalFailureReason: failedRoot.status?.message ?? failedRoot.name }\n        : {}),\n      outcome,\n      ...taskFailure,\n      splitTag: defaultSplit,\n      scenarioId,\n    })\n  }\n  return runs\n}\n\nfunction terminalOutcomeFromRoots(spans: TraceSpanEvent[]): RunTerminalOutcome {\n  const roots = spans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span))\n  if (roots.length !== 1) return 'unknown'\n  if (roots[0]!.status?.code === 'OK') return 'succeeded'\n  if (roots[0]!.status?.code === 'ERROR') return 'failed'\n  return 'unknown'\n}\n\nfunction isTerminalRootCandidate(span: TraceSpanEvent): boolean {\n  const role = errorRoleForSpan(span)\n  if (role === 'LLM' || role === 'TOOL' || role === 'GUARDRAIL' || role === 'EVALUATOR') {\n    return false\n  }\n  return true\n}\n\nfunction readSpanKind(span: TraceSpanEvent): string | undefined {\n  return readAttrString([span], [...SPAN_KIND_ATTR_KEYS, 'span.kind'])?.toUpperCase()\n}\n\nfunction errorRoleForSpan(span: TraceSpanEvent): TraceErrorRole {\n  return resolveSpanKind({\n    kind: readSpanKind(span),\n    name: span.name,\n    attributes: span.attributes,\n  })\n}\n\nfunction spanIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.spanId}`\n}\n\nfunction parentIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.parentSpanId}`\n}\n\n// ── Internal helpers ────────────────────────────────────────────────\n\nfunction groupSpans(spans: TraceSpanEvent[]): Map<string, TraceSpanEvent[]> {\n  const m = new Map<string, TraceSpanEvent[]>()\n  for (const span of spans) {\n    const key = (span['tangle.runId'] as string | undefined) ?? span.traceId\n    const list = m.get(key) ?? []\n    list.push(span)\n    m.set(key, list)\n  }\n  return m\n}\n\nfunction findRoot(group: TraceSpanEvent[]): TraceSpanEvent | undefined {\n  const structuralRoots = group.filter((span) => !span.parentSpanId)\n  const terminalRoots = structuralRoots.filter(isTerminalRootCandidate)\n  const pool =\n    terminalRoots.length > 0 ? terminalRoots : structuralRoots.length > 0 ? structuralRoots : group\n  return orderSpans(pool)[0]\n}\n\nfunction readAttrString(spans: TraceSpanEvent[], keys: string[]): string | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'string' && v.length > 0) return v\n    }\n  }\n  return undefined\n}\n\nfunction readConsistentScenarioId(\n  runId: string,\n  spans: readonly TraceSpanEvent[],\n): string | undefined {\n  const values = new Set<string>()\n  for (const span of spans) {\n    const topLevel = span['tangle.scenarioId']\n    if (typeof topLevel === 'string' && topLevel.length > 0) values.add(topLevel)\n    const attribute = span.attributes['tangle.scenarioId']\n    if (typeof attribute === 'string' && attribute.length > 0) values.add(attribute)\n  }\n  if (values.size > 1) {\n    throw new ValidationError(\n      `fromOtelSpans: conflicting scenario ids for run '${runId}': ${[...values].sort().join(', ')}`,\n    )\n  }\n  return values.values().next().value\n}\n\ninterface TaskScoreSource {\n  label: string\n  value: number\n}\n\nfunction resolveTaskScore(\n  runId: string,\n  spans: TraceSpanEvent[],\n  scoreForRun: FromOtelSpansOptions['scoreForRun'],\n): number | undefined {\n  const orderedSpans = orderSpans(spans)\n  const sources: TaskScoreSource[] = []\n\n  if (scoreForRun) {\n    const supplied: unknown = scoreForRun(runId, orderedSpans)\n    if (supplied !== undefined) {\n      if (typeof supplied !== 'number' || !Number.isFinite(supplied)) {\n        throw new ValidationError(\n          `fromOtelSpans: scoreForRun returned a non-finite number for run '${runId}'`,\n        )\n      }\n      sources.push({ label: 'scoreForRun', value: supplied })\n    }\n  }\n\n  for (const span of orderedSpans) {\n    const role = errorRoleForSpan(span)\n    if (span.parentSpanId && role !== 'EVALUATOR') continue\n    if (role === 'EVALUATOR' && span.status?.code === 'ERROR') continue\n    for (const key of TASK_SCORE_ATTR_KEYS) {\n      if (!Object.hasOwn(span.attributes, key)) continue\n      sources.push({\n        label: `span '${span.spanId}' attribute '${key}'`,\n        value: parseTaskScoreAttribute(runId, span.spanId, key, span.attributes[key]),\n      })\n    }\n  }\n\n  if (sources.length === 0) return undefined\n  sources.sort((left, right) => left.label.localeCompare(right.label))\n  const score = sources[0]!.value\n  if (sources.some((source) => source.value !== score)) {\n    const details = sources.map((source) => `${source.label}=${source.value}`).join(', ')\n    throw new ValidationError(\n      `fromOtelSpans: conflicting task-quality scores for run '${runId}': ${details}`,\n    )\n  }\n  return score\n}\n\nfunction parseTaskScoreAttribute(\n  runId: string,\n  spanId: string,\n  key: string,\n  value: unknown,\n): number {\n  const source = `span '${spanId}' attribute '${key}'`\n  if (typeof value === 'string') {\n    if (value.trim().length === 0) {\n      throw new ValidationError(\n        `fromOtelSpans: ${source} is blank for run '${runId}'; task quality must be finite`,\n      )\n    }\n    const parsed = Number(value)\n    if (Number.isFinite(parsed)) return parsed\n  } else if (typeof value === 'number' && Number.isFinite(value)) {\n    return value\n  }\n  throw new ValidationError(\n    `fromOtelSpans: ${source} is not a finite task-quality score for run '${runId}'`,\n  )\n}\n\nfunction orderSpans(spans: readonly TraceSpanEvent[]): TraceSpanEvent[] {\n  return [...spans].sort(\n    (left, right) =>\n      compareUnixNano(left.startTimeUnixNano, right.startTimeUnixNano) ||\n      left.spanId.localeCompare(right.spanId),\n  )\n}\n\nfunction parseUnixNano(value: string): bigint {\n  try {\n    return BigInt(value)\n  } catch {\n    throw new ValidationError(`fromOtelSpans: invalid Unix nanosecond timestamp '${value}'`)\n  }\n}\n\nfunction compareUnixNano(left: string, right: string): number {\n  const leftValue = parseUnixNano(left)\n  const rightValue = parseUnixNano(right)\n  return leftValue < rightValue ? -1 : leftValue > rightValue ? 1 : 0\n}\n\nfunction unixNanoDurationMs(start: string, end: string): number {\n  const delta = parseUnixNano(end) - parseUnixNano(start)\n  if (delta <= 0n) return 0\n  const wholeMs = delta / 1_000_000n\n  const fractionalMs = delta % 1_000_000n\n  const value = Number(wholeMs) + Number(fractionalMs) / 1_000_000\n  if (!Number.isSafeInteger(Number(wholeMs))) {\n    throw new ValidationError('fromOtelSpans: span duration exceeds the safe millisecond range')\n  }\n  return value\n}\n\nfunction collectNumericAttrs(spans: TraceSpanEvent[]): Record<string, number> {\n  const raw: Record<string, number> = {}\n  for (const span of spans) {\n    for (const [k, v] of Object.entries(span.attributes)) {\n      if (typeof v === 'number' && Number.isFinite(v)) raw[k] = v\n    }\n  }\n  return raw\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+CA,MAAM,uBAAuB;CAC3B;CACA;CACA;CACA;AACF;AACA,MAAM,aAAa;CAAC;CAAgB,GAAG;CAAqB;AAAO;AACnE,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAC7D,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAgB7D,SAAgB,cAAc,MAAyC;CACrE,MAAM,EAAE,OAAO,eAAe,WAAW,eAAe,kBAAkB;CAC1E,MAAM,UAAU,WAAW,KAAK;CAEhC,MAAM,OAAoB,CAAC;CAC3B,KAAK,MAAM,CAAC,UAAU,eAAe,SAAS;EAC5C,MAAM,OAAO,SAAS,UAAU;EAChC,IAAI,CAAC,MAAM;EACX,MAAM,eAAe,+BACnB,WAAW,KAAK,UAAU;GACxB,IAAI,KAAK;GACT,GAAI,KAAK,eAAe,EAAE,UAAU,KAAK,aAAa,IAAI,CAAC;GAC3D,YAAY,KAAK;GACjB,MAAM,iBAAiB,IAAI;EAC7B,EAAE,CACJ;EACA,MAAM,cAAc,IAAI,IAAI,aAAa,WAAW;EACpD,MAAM,YAAY,WAAW,QAAQ,SAAS,YAAY,IAAI,KAAK,MAAM,CAAC;EAC1E,MAAM,SAAS,mBAAmB,KAAK,mBAAmB,KAAK,eAAe;EAC9E,MAAM,QACJ,eAAe,WAAW,UAAU,KACpC,eAAe,YAAY,UAAU,KACrC;EACF,MAAM,gBACH,aAAa,KAAK,WAAW,aAAa,KAAK,QAAQ,KAAA,MACxD,aAAa,WAAW;EAC1B,MAAM,UAAU,gBAAgB;EAChC,MAAM,aAAa,yBAAyB,UAAU,UAAU,KAAK;EACrE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,QAAQ,iBAAiB,UAAU,YAAY,KAAK,WAAW;EACrE,MAAM,cAAc,sBAClB,WAAW,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC,GAC/E,uBAAuB,SAAS,EAClC;EAEA,MAAM,aAAa,oBAAoB,UAAU;EACjD,MAAM,eAAe,qBACnB,WAAW,KAAK,UAAU;GACxB,IAAI,aAAa,IAAI;GACrB,GAAI,KAAK,eAAe,EAAE,UAAU,eAAe,IAAI,EAAE,IAAI,CAAC;GAC9D,MAAM,iBAAiB,IAAI;GAC3B,OAAO,KAAK,QAAQ,SAAS;GAC7B,aAAa,CAAC,KAAK,gBAAgB,wBAAwB,IAAI;EACjE,EAAE,CACJ;EACA,WAAW,mBAAmB,aAAa;EAC3C,WAAW,wBAAwB,aAAa;EAChD,WAAW,sBAAsB,aAAa;EAC9C,WAAW,wBAAwB,aAAa;EAChD,WAAW,oBAAoB,aAAa;EAC5C,WAAW,yBAAyB,aAAa;EACjD,WAAW,2BAA2B,aAAa;EACnD,WAAW,iBAAiB,aAAa;EACzC,IAAI,aAAa,KAAK,UAAU,KAAA,KAAa,CAAC,aAAa,KAAK,UAC9D,WAAW,4BAA4B,aAAa,KAAK;EAE3D,4BAA4B,YAAY,aAAa,SAAS;EAE9D,MAAM,cACJ,UAAU,KAAA,IACN;GACE,UAAU,EAAE,gBAAgB,EAAE,MAAM,EAAE;GACtC,YAAY,EAAE,MAAM;GACpB,WAAW;EACb,IACA,KAAA;EAEN,MAAM,kBAAkB,yBAAyB,UAAU;EAC3D,MAAM,aACJ,oBAAoB,WAChB,WAAW,MACR,SACC,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,KAAK,KAAK,QAAQ,SAAS,OACjF,IACA,KAAA;EACN,MAAM,UAAsB;GAC1B,KAAK;GACL,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;EACvC;EACA,IAAI,UAAU,KAAA,GAAW;GACvB,IAAI,iBAAiB,WAAW,QAAQ,eAAe;QAClD,QAAQ,cAAc;EAC7B;EAEA,KAAK,KAAK;GACR,OAAO;GACP;GACA,aAAc,KAAK,WAAW,yBAAgD;GAC9E,MAAM;GACN;GACA;GACA;GACA,WAAY,KAAK,WAAW,wBAA+C;GAC3E;GACA;GACA,gBACE,iBAAiB,KAAA,IACb;IAAE,MAAM;IAAc,KAAK;GAAK,IAChC;IAAE,MAAM;IAAY,KAAK;GAAa;GAC5C,YAAY,aAAa;GACzB;GACA,GAAI,aACA,EAAE,uBAAuB,WAAW,QAAQ,WAAW,WAAW,KAAK,IACvE,CAAC;GACL;GACA,GAAG;GACH,UAAU;GACV;EACF,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,OAA6C;CAC7E,MAAM,QAAQ,MAAM,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC;CACxF,IAAI,MAAM,WAAW,GAAG,OAAO;CAC/B,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,MAAM,OAAO;CAC5C,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,SAAS,OAAO;CAC/C,OAAO;AACT;AAEA,SAAS,wBAAwB,MAA+B;CAC9D,MAAM,OAAO,iBAAiB,IAAI;CAClC,IAAI,SAAS,SAAS,SAAS,UAAU,SAAS,eAAe,SAAS,aACxE,OAAO;CAET,OAAO;AACT;AAEA,SAAS,aAAa,MAA0C;CAC9D,OAAO,eAAe,CAAC,IAAI,GAAG,CAAC,GAAGA,uBAAqB,WAAW,CAAC,CAAC,EAAE,YAAY;AACpF;AAEA,SAAS,iBAAiB,MAAsC;CAC9D,OAAO,gBAAgB;EACrB,MAAM,aAAa,IAAI;EACvB,MAAM,KAAK;EACX,YAAY,KAAK;CACnB,CAAC;AACH;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAEA,SAAS,eAAe,MAA8B;CACpD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAIA,SAAS,WAAW,OAAwD;CAC1E,MAAM,oBAAI,IAAI,IAA8B;CAC5C,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,MAAO,KAAK,mBAA0C,KAAK;EACjE,MAAM,OAAO,EAAE,IAAI,GAAG,KAAK,CAAC;EAC5B,KAAK,KAAK,IAAI;EACd,EAAE,IAAI,KAAK,IAAI;CACjB;CACA,OAAO;AACT;AAEA,SAAS,SAAS,OAAqD;CACrE,MAAM,kBAAkB,MAAM,QAAQ,SAAS,CAAC,KAAK,YAAY;CACjE,MAAM,gBAAgB,gBAAgB,OAAO,uBAAuB;CAGpE,OAAO,WADL,cAAc,SAAS,IAAI,gBAAgB,gBAAgB,SAAS,IAAI,kBAAkB,KACtE,CAAC,CAAC;AAC1B;AAEA,SAAS,eAAe,OAAyB,MAAoC;CACnF,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,IAAI,KAAK,WAAW;EAC1B,IAAI,OAAO,MAAM,YAAY,EAAE,SAAS,GAAG,OAAO;CACpD;AAGJ;AAEA,SAAS,yBACP,OACA,OACoB;CACpB,MAAM,yBAAS,IAAI,IAAY;CAC/B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,WAAW,KAAK;EACtB,IAAI,OAAO,aAAa,YAAY,SAAS,SAAS,GAAG,OAAO,IAAI,QAAQ;EAC5E,MAAM,YAAY,KAAK,WAAW;EAClC,IAAI,OAAO,cAAc,YAAY,UAAU,SAAS,GAAG,OAAO,IAAI,SAAS;CACjF;CACA,IAAI,OAAO,OAAO,GAChB,MAAM,IAAI,gBACR,oDAAoD,MAAM,KAAK,CAAC,GAAG,MAAM,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC7F;CAEF,OAAO,OAAO,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC;AAChC;AAOA,SAAS,iBACP,OACA,OACA,aACoB;CACpB,MAAM,eAAe,WAAW,KAAK;CACrC,MAAM,UAA6B,CAAC;CAEpC,IAAI,aAAa;EACf,MAAM,WAAoB,YAAY,OAAO,YAAY;EACzD,IAAI,aAAa,KAAA,GAAW;GAC1B,IAAI,OAAO,aAAa,YAAY,CAAC,OAAO,SAAS,QAAQ,GAC3D,MAAM,IAAI,gBACR,oEAAoE,MAAM,EAC5E;GAEF,QAAQ,KAAK;IAAE,OAAO;IAAe,OAAO;GAAS,CAAC;EACxD;CACF;CAEA,KAAK,MAAM,QAAQ,cAAc;EAC/B,MAAM,OAAO,iBAAiB,IAAI;EAClC,IAAI,KAAK,gBAAgB,SAAS,aAAa;EAC/C,IAAI,SAAS,eAAe,KAAK,QAAQ,SAAS,SAAS;EAC3D,KAAK,MAAM,OAAO,sBAAsB;GACtC,IAAI,CAAC,OAAO,OAAO,KAAK,YAAY,GAAG,GAAG;GAC1C,QAAQ,KAAK;IACX,OAAO,SAAS,KAAK,OAAO,eAAe,IAAI;IAC/C,OAAO,wBAAwB,OAAO,KAAK,QAAQ,KAAK,KAAK,WAAW,IAAI;GAC9E,CAAC;EACH;CACF;CAEA,IAAI,QAAQ,WAAW,GAAG,OAAO,KAAA;CACjC,QAAQ,MAAM,MAAM,UAAU,KAAK,MAAM,cAAc,MAAM,KAAK,CAAC;CACnE,MAAM,QAAQ,QAAQ,EAAE,CAAE;CAC1B,IAAI,QAAQ,MAAM,WAAW,OAAO,UAAU,KAAK,GAAG;EACpD,MAAM,UAAU,QAAQ,KAAK,WAAW,GAAG,OAAO,MAAM,GAAG,OAAO,OAAO,CAAC,CAAC,KAAK,IAAI;EACpF,MAAM,IAAI,gBACR,2DAA2D,MAAM,KAAK,SACxE;CACF;CACA,OAAO;AACT;AAEA,SAAS,wBACP,OACA,QACA,KACA,OACQ;CACR,MAAM,SAAS,SAAS,OAAO,eAAe,IAAI;CAClD,IAAI,OAAO,UAAU,UAAU;EAC7B,IAAI,MAAM,KAAK,CAAC,CAAC,WAAW,GAC1B,MAAM,IAAI,gBACR,kBAAkB,OAAO,qBAAqB,MAAM,+BACtD;EAEF,MAAM,SAAS,OAAO,KAAK;EAC3B,IAAI,OAAO,SAAS,MAAM,GAAG,OAAO;CACtC,OAAO,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,GAC3D,OAAO;CAET,MAAM,IAAI,gBACR,kBAAkB,OAAO,+CAA+C,MAAM,EAChF;AACF;AAEA,SAAS,WAAW,OAAoD;CACtE,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MACf,MAAM,UACL,gBAAgB,KAAK,mBAAmB,MAAM,iBAAiB,KAC/D,KAAK,OAAO,cAAc,MAAM,MAAM,CAC1C;AACF;AAEA,SAAS,cAAc,OAAuB;CAC5C,IAAI;EACF,OAAO,OAAO,KAAK;CACrB,QAAQ;EACN,MAAM,IAAI,gBAAgB,qDAAqD,MAAM,EAAE;CACzF;AACF;AAEA,SAAS,gBAAgB,MAAc,OAAuB;CAC5D,MAAM,YAAY,cAAc,IAAI;CACpC,MAAM,aAAa,cAAc,KAAK;CACtC,OAAO,YAAY,aAAa,KAAK,YAAY,aAAa,IAAI;AACpE;AAEA,SAAS,mBAAmB,OAAe,KAAqB;CAC9D,MAAM,QAAQ,cAAc,GAAG,IAAI,cAAc,KAAK;CACtD,IAAI,SAAS,IAAI,OAAO;CACxB,MAAM,UAAU,QAAQ;CACxB,MAAM,eAAe,QAAQ;CAC7B,MAAM,QAAQ,OAAO,OAAO,IAAI,OAAO,YAAY,IAAI;CACvD,IAAI,CAAC,OAAO,cAAc,OAAO,OAAO,CAAC,GACvC,MAAM,IAAI,gBAAgB,iEAAiE;CAE7F,OAAO;AACT;AAEA,SAAS,oBAAoB,OAAiD;CAC5E,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,KAAK,UAAU,GACjD,IAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK;CAG9D,OAAO;AACT"}

package/dist/{outcome-store-BXlkwMPR.js → outcome-store-DUcMj-3D.js}

@@ -41,7 +41,7 @@ var FileSystemOutcomeStore = class {
 	constructor(options) {
 		if (!options.dir.trim()) throw new Error("outcome store dir must be nonempty");
 		this.dir = options.dir;
-		this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024;
+		this.maxBytes = options.maxBytes ?? 33554432;
 		if (!Number.isSafeInteger(this.maxBytes) || this.maxBytes < 1) throw new Error("outcome store maxBytes must be a positive safe integer");
 	}
 	async append(outcome) {
@@ -128,4 +128,4 @@ function matches(outcome, filter) {
 //#endregion
 export { InMemoryOutcomeStore as n, OutcomeStoreError as r, FileSystemOutcomeStore as t };
 
-//# sourceMappingURL=outcome-store-BXlkwMPR.js.map
+//# sourceMappingURL=outcome-store-DUcMj-3D.js.map