npm - @tangle-network/agent-eval - Versions diffs - 0.188.0 → 0.189.0 - Mend

@tangle-network/agent-eval 0.188.0 → 0.189.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (184) hide show

package/dist/mint-B4zgSAJW.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"mint-B4zgSAJW.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 *\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, 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\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\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\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 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 const uncaptured = record.costProvenance.kind === 'uncaptured'\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: null,\n run_id: record.runId,\n experiment_id: record.experimentId,\n candidate_id: record.candidateId,\n generation: null,\n candidate_index: 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: 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: uncaptured ? null : record.costUsd,\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: { patch_path: null, run_dir: null, transcript_ref: null },\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 if (trajectory.steps.length === 0 && !options.messagesOf) {\n missingTraces.push(record.runId)\n rows.push(\n mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\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, steps, conversation, options, capturedAt, gap))\n }\n return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8EA,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;AAEA,MAAM,iBAA8D;CAClE,QAAQ;CACR,KAAK;CACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACmB;CAUnB,sBAAsB,MAAM;CAI5B,iBAAiB,MAAM;CAGvB,MAAM,eAAe,oBAAoB,MAAM;CAC/C,MAAM,aAAa,OAAO,eAAe,SAAS;CAClD,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;EACnB,QAAQ,OAAO;EACf,eAAe,OAAO;EACtB,cAAc,OAAO;EACrB,YAAY;EACZ,iBAAiB;EACjB,MAAM,QAAQ,QAAQ;EACtB,MAAM;GACJ,OAAO,QAAQ,SAAS,OAAO;GAC/B,aAAa,OAAO;GACpB,OAAO,eAAe,OAAO;GAC7B,MAAM,OAAO;GACb,KAAK;EACP;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,OAAO;GAChC,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;GAAE,YAAY;GAAM,SAAS;GAAM,gBAAgB;EAAK;EACnE,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,IAAI,WAAW,MAAM,WAAW,KAAK,CAAC,QAAQ,YAAY;GACxD,cAAc,KAAK,OAAO,KAAK;GAC/B,KAAK,KACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC,CACxF;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,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;CAC3E;CACA,OAAO;EAAE;EAAM;CAAc;AAC/B"}

package/dist/release-confidence-RanRUXRW.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"release-confidence-RanRUXRW.js","names":[],"sources":["../src/promotion-gate.ts","../src/release-confidence.ts"],"sourcesContent":["/**\n * Bootstrap-CI promotion gate.\n *\n * In any iterative-improvement loop (GEPA, prompt evolution, dataset\n * curation), the question is \"did this generation actually improve, or are\n * we celebrating noise?\". With small N and noisy outcomes, point-estimate\n * deltas lie. Bootstrap confidence intervals tell the operator whether the\n * delta is real before code or prompts get promoted.\n *\n * This module is pure functions — no I/O, no model calls. Easy to unit-test\n * and to compose into any verdict gate.\n *\n * Default gate:\n * - Bootstrap mean baseline vs candidate (1k resamples).\n * - Compute the delta distribution; pass if the lower CI bound > 0.\n * - Tunable confidence (default 95%) and resample count.\n *\n * Verdict semantics intentionally match the existing `experiments.jsonl`\n * vocabulary:\n * - ADVANCE: candidate's CI lower bound > baseline mean (real win)\n * - KEEP: overlap, but candidate point estimate >= baseline (neutral)\n * - REVERT: candidate's CI upper bound < baseline mean (real regression)\n * - INCONCLUSIVE: not enough samples or CI straddles zero with no signal\n */\n\nimport { mulberry32 } from './statistics'\n\nexport type Verdict = 'ADVANCE' | 'KEEP' | 'REVERT' | 'INCONCLUSIVE'\n\nexport interface BootstrapResult {\n baselineMean: number\n candidateMean: number\n /** candidateMean - baselineMean, point estimate. */\n delta: number\n /** Lower bound of the (1 - alpha) CI on the delta. */\n ciLower: number\n /** Upper bound of the (1 - alpha) CI on the delta. */\n ciUpper: number\n /** Number of bootstrap resamples used. */\n iterations: number\n alpha: number\n verdict: Verdict\n}\n\nexport interface BootstrapOptions {\n /** Confidence level alpha (default 0.05 → 95% CI). */\n alpha?: number\n /** Number of resamples (default 1000). */\n iterations?: number\n /**\n * Minimum total samples (baseline + candidate) below which we always\n * return INCONCLUSIVE — bootstrap with too few samples is meaningless.\n * Default 6 (combined).\n */\n minTotalSamples?: number\n /** RNG seed for reproducibility. Absent, the seed is derived from the\n * compared arms, so the same inputs reproduce the same verdict. */\n seed?: number\n}\n\n/**\n * Compute the bootstrap CI on (candidateMean - baselineMean) and a verdict.\n *\n * Uses simple percentile bootstrap on the difference of resampled means.\n * That's the standard non-parametric primitive — no distributional\n * assumptions, robust to skew, easy to reason about.\n */\nexport function bootstrapCi(\n baseline: number[],\n candidate: number[],\n options: BootstrapOptions = {},\n): BootstrapResult {\n const alpha = options.alpha ?? 0.05\n const iterations = options.iterations ?? 1000\n const minTotal = options.minTotalSamples ?? 6\n const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate))\n\n const baselineMean = mean(baseline)\n const candidateMean = mean(candidate)\n const delta = candidateMean - baselineMean\n\n if (\n baseline.length + candidate.length < minTotal ||\n baseline.length === 0 ||\n candidate.length === 0\n ) {\n return {\n baselineMean,\n candidateMean,\n delta,\n ciLower: -Infinity,\n ciUpper: Infinity,\n iterations: 0,\n alpha,\n verdict: 'INCONCLUSIVE',\n }\n }\n\n const deltas: number[] = new Array(iterations)\n for (let i = 0; i < iterations; i++) {\n const bResample = resample(baseline, rng)\n const cResample = resample(candidate, rng)\n deltas[i] = mean(cResample) - mean(bResample)\n }\n deltas.sort((a, b) => a - b)\n const lowerIdx = Math.floor((alpha / 2) * iterations)\n const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1\n const ciLower = deltas[Math.max(0, lowerIdx)]!\n const ciUpper = deltas[Math.min(iterations - 1, upperIdx)]!\n\n let verdict: Verdict\n // A ZERO-WIDTH interval is an absence of evidence, not a certainty. Constant\n // arms make every resample identical, so pass/fail data promotes on nothing:\n // [0,0,0] vs [1,1,1] is six samples, clears the `minTotalSamples` floor, and\n // yields [1, 1] — an \"ADVANCE\" carrying no information about how far the\n // estimate could be wrong. Same rule as `pairedDeltaTest`, one estimator over.\n if (!Number.isFinite(ciLower) || !Number.isFinite(ciUpper) || ciLower === ciUpper) {\n verdict = 'INCONCLUSIVE'\n } else if (ciLower > 0) verdict = 'ADVANCE'\n else if (ciUpper < 0) verdict = 'REVERT'\n else if (delta >= 0) verdict = 'KEEP'\n else verdict = 'INCONCLUSIVE'\n\n return {\n baselineMean,\n candidateMean,\n delta,\n ciLower,\n ciUpper,\n iterations,\n alpha,\n verdict,\n }\n}\n\nfunction mean(xs: number[]): number {\n if (xs.length === 0) return 0\n let s = 0\n for (const x of xs) s += x\n return s / xs.length\n}\n\nfunction resample(xs: number[], rng: () => number): number[] {\n const out = new Array(xs.length)\n for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)]\n return out\n}\n\n/** Stable seed derived from the inputs — same data → same CI bounds. */\nfunction hashSeed(a: number[], b: number[]): number {\n let h = 2166136261\n for (const x of [...a, ...b]) {\n const view = new Float64Array([x])\n const bytes = new Uint8Array(view.buffer)\n for (const byte of bytes) {\n h ^= byte\n h = Math.imul(h, 16777619)\n }\n }\n return h >>> 0\n}\n\n/**\n * Judge-replay promotion gate.\n *\n * The cheap inner-loop judge that drives an evolution run is by definition\n * fast and noisy. When you're about to promote a winning variant to the\n * canonical default, you want a STRONGER judge (a more expensive model, a\n * human grader, a separately-trained reward model) to confirm the win\n * generalises beyond the inner loop.\n *\n * This helper takes raw winner + baseline outputs, scores both through the\n * stronger judge, and applies `bootstrapCi`. ADVANCE means the stronger\n * judge agrees the winner is real with the configured confidence. Doesn't\n * matter what shape your \"output\" is — pass a string, an object, anything\n * the judge can read.\n */\nexport interface JudgeReplayGateArgs<TOutput> {\n baselineOutputs: TOutput[]\n candidateOutputs: TOutput[]\n /** Stronger judge — async to allow LLM calls. Return a 0..N scalar score. */\n judge: (output: TOutput) => Promise<number> | number\n alpha?: number\n iterations?: number\n /** RNG seed for reproducibility. */\n seed?: number\n /** Maximum concurrent judge calls. Default 4. */\n judgeConcurrency?: number\n}\n\n/**\n * Confirm a candidate's win with a stronger judge: score baseline and candidate outputs independently, then bootstrap a CI to verify the lift generalises beyond the inner loop.\n */\nexport async function judgeReplayGate<TOutput>(\n args: JudgeReplayGateArgs<TOutput>,\n): Promise<BootstrapResult & { baselineSamples: number; candidateSamples: number }> {\n const concurrency = args.judgeConcurrency ?? 4\n const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency)\n const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency)\n const ci = bootstrapCi(baselineScores, candidateScores, {\n ...(args.alpha !== undefined ? { alpha: args.alpha } : {}),\n ...(args.iterations !== undefined ? { iterations: args.iterations } : {}),\n ...(args.seed !== undefined ? { seed: args.seed } : {}),\n })\n return {\n ...ci,\n baselineSamples: baselineScores.length,\n candidateSamples: candidateScores.length,\n }\n}\n\nasync function scoreAll<TOutput>(\n outputs: TOutput[],\n judge: (output: TOutput) => Promise<number> | number,\n concurrency: number,\n): Promise<number[]> {\n const results: number[] = new Array(outputs.length)\n let next = 0\n async function worker(): Promise<void> {\n while (true) {\n const i = next++\n if (i >= outputs.length) return\n const v = await judge(outputs[i]!)\n results[i] = Number.isFinite(v) ? v : 0\n }\n }\n await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()))\n return results\n}\n","/**\n * Release confidence gate.\n *\n * This is the production-facing composition layer over the lower-level\n * primitives:\n * - Dataset manifests prove corpus/version coverage.\n * - RunRecord rows prove reproducible search/holdout outcomes.\n * - Multi-shot trace evidence carries turn counts and ASI diagnostics.\n * - HeldOutGate decisions remain the paired promotion authority.\n *\n * The gate is intentionally pure and conservative. Missing declared evidence\n * fails closed instead of being treated as a neutral zero.\n */\n\nimport type { DatasetManifest, DatasetScenario, DatasetSplit } from './dataset'\nimport { ValidationError, VerificationError } from './errors'\nimport type { GateDecision } from './held-out-gate'\nimport { isRealnessGated, observedSplitScore } from './rollout/reward'\nimport { type RunRecord, type RunSplitTag, validateRunRecord } from './run-record'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Severity of an actionable finding attached to a run/trace. */\nexport type AsiSeverity = 'info' | 'warning' | 'error' | 'critical'\n\n/** Actionable side-info — a diagnosed finding the loop can act on. */\nexport interface ActionableSideInfo {\n /** Stable expectation/check id when available. */\n expectationId?: string\n /** Human-readable diagnosis of what happened. */\n message: string\n severity?: AsiSeverity\n /** Concrete trace excerpt, file path, tool call, screenshot id, etc. */\n evidence?: string\n /** Prompt/tool/context surface likely responsible. */\n responsibleSurface?: string\n /** Suggested fix in natural language. */\n suggestion?: string\n /** Whether this expectation was satisfied. Defaults to false for ASI rows. */\n matched?: boolean\n metadata?: Record<string, unknown>\n}\n\nexport type ReleaseConfidenceStatus = 'pass' | 'warn' | 'fail'\nexport type ReleaseConfidenceAxisName =\n | 'corpus'\n | 'quality'\n | 'reliability'\n | 'generalization'\n | 'diagnostics'\n | 'efficiency'\n\nexport interface ReleaseTraceEvidence {\n scenarioId: string\n candidateId?: string\n split?: RunSplitTag\n score?: number\n ok?: boolean\n turnCount?: number\n costUsd?: number\n durationMs?: number\n /** Canonical task-failure class. Free-form detail belongs in ASI. */\n failureClass?: FailureClass\n asi?: ActionableSideInfo[]\n metadata?: Record<string, unknown>\n}\n\nexport interface ReleaseConfidenceThresholds {\n /** Require a Dataset manifest or explicit scenarios. Default true. */\n requireCorpus?: boolean\n minScenarioCount?: number\n minSearchRuns?: number\n minHoldoutRuns?: number\n /** Require at least one holdout scenario/run. Default true. */\n requireHoldout?: boolean\n minPassRate?: number\n minMeanScore?: number\n /** Search mean may exceed holdout mean by at most this much. */\n maxOverfitGap?: number\n maxMeanCostUsd?: number\n maxP95WallMs?: number\n /** Low-score/failed rows must carry ASI. Default true. */\n requireAsiForFailures?: boolean\n /** Score below this is considered a failure for ASI coverage. Default 0.5. */\n failureScoreThreshold?: number\n}\n\nexport interface ReleaseConfidenceInput {\n target: string\n candidateId?: string\n baselineId?: string\n dataset?: DatasetManifest\n scenarios?: readonly DatasetScenario[]\n runs?: readonly RunRecord[]\n traces?: readonly ReleaseTraceEvidence[]\n gateDecision?: GateDecision | null\n thresholds?: ReleaseConfidenceThresholds\n}\n\nexport interface ReleaseConfidenceAxis {\n name: ReleaseConfidenceAxisName\n status: ReleaseConfidenceStatus\n score: number | null\n detail: string\n}\n\nexport interface ReleaseConfidenceIssue {\n axis: ReleaseConfidenceAxisName\n severity: 'critical' | 'warning'\n code: string\n detail: string\n}\n\nexport interface ReleaseConfidenceMetrics {\n scenarioCount: number\n /** Search rows with a finite search score. */\n searchRuns: number\n /** Holdout rows with a finite holdout score. */\n holdoutRuns: number\n /** Runs with neither a split-matched score nor an explicit task failure. */\n unscoredRuns: number\n /** Run rows, or trace rows when no runs exist, with no classified terminal result. */\n unclassifiedTerminalRuns: number\n /** Run rows, or trace rows when no runs exist, that ended unsuccessfully. */\n terminalFailureRuns: number\n /** Success fraction when every run or fallback trace row has a classified result. */\n reliabilityRate: number | null\n passRate: number | null\n meanScore: number | null\n searchMeanScore: number | null\n holdoutMeanScore: number | null\n overfitGap: number | null\n meanCostUsd: number | null\n p95WallMs: number | null\n failedRows: number\n failuresWithAsi: number\n singleShotTraces: number\n multiShotTraces: number\n splitCounts: Record<DatasetSplit, number>\n domainCounts: Record<string, number>\n failureClassCounts: Partial<Record<FailureClass, number>>\n responsibleSurfaceCounts: Record<string, number>\n /**\n * Runs excluded from `passRate` because the authenticity gate flagged them as\n * gamed. Surfaced, never silent: a release whose pass rate is computed over a\n * shrunken denominator has to say by how much, or the exclusion is just a\n * different way of hiding the same runs.\n */\n realnessGatedRuns: number\n}\n\nexport interface ReleaseConfidenceScorecard {\n target: string\n candidateId: string | null\n baselineId: string | null\n status: ReleaseConfidenceStatus\n promote: boolean\n axes: ReleaseConfidenceAxis[]\n issues: ReleaseConfidenceIssue[]\n metrics: ReleaseConfidenceMetrics\n dataset: DatasetManifest | null\n gateDecision: GateDecision | null\n summary: string\n}\n\nconst DEFAULT_THRESHOLDS: Required<ReleaseConfidenceThresholds> = {\n requireCorpus: true,\n minScenarioCount: 1,\n minSearchRuns: 1,\n minHoldoutRuns: 1,\n requireHoldout: true,\n minPassRate: 0.8,\n minMeanScore: 0.7,\n maxOverfitGap: 0.15,\n maxMeanCostUsd: Number.POSITIVE_INFINITY,\n maxP95WallMs: Number.POSITIVE_INFINITY,\n requireAsiForFailures: true,\n failureScoreThreshold: 0.5,\n}\n\nexport function evaluateReleaseConfidence(\n input: ReleaseConfidenceInput,\n): ReleaseConfidenceScorecard {\n const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds }\n const candidateId = input.candidateId ?? null\n const runs = filterCandidate(\n (input.runs ?? []).map(validateRunRecord),\n candidateId,\n input.baselineId,\n )\n const traces = filterTraceCandidate(\n (input.traces ?? []).map(validateReleaseTraceEvidence),\n candidateId,\n input.baselineId,\n )\n const scenarios = input.scenarios ?? []\n const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length\n const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios)\n const searchScores = scoresFor(runs, 'search')\n const holdoutScores = scoresFor(runs, 'holdout')\n const runScores = runs.map(runSplitScore).filter(isFiniteNumber)\n const traceScores = traces.map((t) => t.score).filter(isFiniteNumber)\n const scoreUniverse = runs.length > 0 ? runScores : traceScores\n const qualityRuns = runs.filter((run) => runSplitScore(run) !== undefined)\n const unscoredRuns = runs.filter(\n (run) =>\n runSplitScore(run) === undefined &&\n !hasExplicitTaskFailure(run) &&\n !isFailedTerminalOutcome(run.terminalOutcome),\n ).length\n // Realness-gated runs are EXCLUDED from the pass rate — numerator AND\n // denominator — and the count of what was dropped ships beside the rate as\n // `metrics.realnessGatedRuns`. Counting a gamed run as a pass made faking a\n // success the cheapest way to improve a release scorecard; scoring it 0\n // instead would be the other error, silently deflating the rate with a run\n // the gate says carries no usable verdict at all.\n const honestRuns = runs.filter((run) => !isRealnessGated(run))\n const passOutcomes =\n runs.length > 0\n ? honestRuns.map((run) => runPassOutcome(run, thresholds.failureScoreThreshold))\n : traces.map((trace) => tracePassOutcome(trace, thresholds.failureScoreThreshold))\n const reliabilityRows =\n runs.length > 0\n ? runs.map((run) => terminalSuccess(run.terminalOutcome))\n : traces.map((trace) => trace.ok)\n const unclassifiedTerminalRuns = reliabilityRows.filter((outcome) => outcome === undefined).length\n const terminalFailureRuns = reliabilityRows.filter((outcome) => outcome === false).length\n const reliabilityRate =\n reliabilityRows.length === 0 || unclassifiedTerminalRuns > 0\n ? null\n : (reliabilityRows.length - terminalFailureRuns) / reliabilityRows.length\n const searchRuns = qualityRuns.filter((r) => r.splitTag === 'search').length\n const holdoutRuns = qualityRuns.filter((r) => r.splitTag === 'holdout').length\n const failed = failedRows(runs, traces, thresholds.failureScoreThreshold)\n const searchMeanScore = meanOrNull(searchScores)\n const holdoutMeanScore = meanOrNull(holdoutScores)\n const runCosts = runs.flatMap((run) =>\n run.costProvenance.kind === 'uncaptured' ? [] : [run.costProvenance.usd],\n )\n const traceCosts = traces.map((trace) => trace.costUsd).filter(isFiniteNumber)\n const meanCostUsd =\n runs.length > 0\n ? runCosts.length === runs.length\n ? meanOrNull(runCosts)\n : null\n : meanOrNull(traceCosts)\n const wallTimes =\n runs.length > 0\n ? runs.map((run) => run.wallMs)\n : traces.map((trace) => trace.durationMs).filter(isFiniteNumber)\n const metrics: ReleaseConfidenceMetrics = {\n scenarioCount,\n searchRuns,\n holdoutRuns,\n unscoredRuns,\n unclassifiedTerminalRuns,\n terminalFailureRuns,\n reliabilityRate,\n passRate: passOutcomeRate(passOutcomes),\n realnessGatedRuns: runs.length - honestRuns.length,\n meanScore: meanOrNull(scoreUniverse),\n searchMeanScore,\n holdoutMeanScore,\n overfitGap: diffOrNull(searchMeanScore, holdoutMeanScore),\n meanCostUsd,\n p95WallMs: percentileOrNull(wallTimes, 0.95),\n failedRows: failed.length,\n failuresWithAsi: failed.filter((row) => row.hasAsi).length,\n singleShotTraces: traces.filter((t) => t.turnCount === 1).length,\n multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,\n splitCounts,\n domainCounts: countDomains(scenarios),\n failureClassCounts: countFailureClasses(runs, traces, thresholds.failureScoreThreshold),\n responsibleSurfaceCounts: countResponsibleSurfaces(traces),\n }\n\n const issues: ReleaseConfidenceIssue[] = []\n checkCorpus(input, thresholds, metrics, issues)\n checkQuality(thresholds, metrics, issues)\n checkReliability(metrics, issues)\n checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues)\n checkDiagnostics(thresholds, metrics, issues)\n checkEfficiency(thresholds, metrics, issues)\n\n const axes = buildAxes(metrics, thresholds, issues)\n const status = issues.some((i) => i.severity === 'critical')\n ? 'fail'\n : issues.length > 0\n ? 'warn'\n : 'pass'\n\n return {\n target: input.target,\n candidateId,\n baselineId: input.baselineId ?? null,\n status,\n promote: status === 'pass' && (input.gateDecision ? input.gateDecision.promote : true),\n axes,\n issues,\n metrics,\n dataset: input.dataset ?? null,\n gateDecision: input.gateDecision ?? null,\n summary: renderSummary(input.target, status, metrics, issues),\n }\n}\n\nexport function assertReleaseConfidence(input: ReleaseConfidenceInput): ReleaseConfidenceScorecard {\n const scorecard = evaluateReleaseConfidence(input)\n if (scorecard.status === 'fail') {\n throw new VerificationError(scorecard.summary)\n }\n return scorecard\n}\n\nfunction filterCandidate(\n runs: readonly RunRecord[],\n candidateId: string | null,\n baselineId?: string,\n): RunRecord[] {\n if (candidateId) return runs.filter((r) => r.candidateId === candidateId)\n if (baselineId) return runs.filter((r) => r.candidateId !== baselineId)\n return [...runs]\n}\n\nfunction filterTraceCandidate(\n traces: readonly ReleaseTraceEvidence[],\n candidateId: string | null,\n baselineId?: string,\n): ReleaseTraceEvidence[] {\n if (candidateId)\n return traces.filter((t) => t.candidateId === undefined || t.candidateId === candidateId)\n if (baselineId)\n return traces.filter((t) => t.candidateId === undefined || t.candidateId !== baselineId)\n return [...traces]\n}\n\nfunction validateReleaseTraceEvidence(\n trace: ReleaseTraceEvidence,\n index: number,\n): ReleaseTraceEvidence {\n const value = trace as unknown as Record<string, unknown>\n if (Object.hasOwn(value, 'failureMode')) {\n throw new ValidationError(\n `traces[${index}].failureMode is not supported; use canonical failureClass`,\n )\n }\n if (trace.failureClass !== undefined && !FAILURE_CLASSES.includes(trace.failureClass)) {\n throw new ValidationError(\n `traces[${index}].failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n )\n }\n return trace\n}\n\nfunction checkCorpus(\n input: ReleaseConfidenceInput,\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {\n issues.push({\n axis: 'corpus',\n severity: 'critical',\n code: 'missing_corpus',\n detail: 'No Dataset manifest or scenarios supplied.',\n })\n }\n if (metrics.scenarioCount < thresholds.minScenarioCount) {\n issues.push({\n axis: 'corpus',\n severity: 'critical',\n code: 'few_scenarios',\n detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`,\n })\n }\n if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {\n issues.push({\n axis: 'corpus',\n severity: 'critical',\n code: 'missing_holdout_split',\n detail: 'Corpus has no holdout scenarios.',\n })\n }\n}\n\nfunction checkQuality(\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (metrics.searchRuns < thresholds.minSearchRuns) {\n issues.push({\n axis: 'quality',\n severity: 'critical',\n code: 'few_search_runs',\n detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`,\n })\n }\n if (metrics.unscoredRuns > 0) {\n issues.push({\n axis: 'quality',\n severity: 'critical',\n code: 'unscored_runs',\n detail: `${metrics.unscoredRuns} supplied run(s) have no task result.`,\n })\n }\n if (metrics.passRate === null || metrics.meanScore === null) {\n issues.push({\n axis: 'quality',\n severity: 'critical',\n code: 'missing_quality_scores',\n detail: 'No task-quality scores are available for pass-rate and mean-score checks.',\n })\n }\n if (metrics.passRate !== null && metrics.passRate < thresholds.minPassRate) {\n issues.push({\n axis: 'quality',\n severity: 'critical',\n code: 'low_pass_rate',\n detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`,\n })\n }\n if (metrics.meanScore !== null && metrics.meanScore < thresholds.minMeanScore) {\n issues.push({\n axis: 'quality',\n severity: 'critical',\n code: 'low_mean_score',\n detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`,\n })\n }\n}\n\nfunction checkReliability(\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (metrics.reliabilityRate === null) {\n issues.push({\n axis: 'reliability',\n severity: 'critical',\n code: 'missing_reliability_evidence',\n detail:\n metrics.unclassifiedTerminalRuns > 0\n ? `${metrics.unclassifiedTerminalRuns} supplied run(s) have no classified terminal result.`\n : 'No classified terminal results are available.',\n })\n }\n if (metrics.terminalFailureRuns > 0) {\n issues.push({\n axis: 'reliability',\n severity: 'critical',\n code: 'terminal_run_failures',\n detail: `${metrics.terminalFailureRuns} run(s) ended failed, cancelled, or incomplete.`,\n })\n }\n}\n\nfunction checkGeneralization(\n gateDecision: GateDecision | null,\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {\n issues.push({\n axis: 'generalization',\n severity: 'critical',\n code: 'few_holdout_runs',\n detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`,\n })\n }\n if (metrics.overfitGap !== null && metrics.overfitGap > thresholds.maxOverfitGap) {\n issues.push({\n axis: 'generalization',\n severity: 'critical',\n code: 'overfit_gap',\n detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`,\n })\n }\n if (gateDecision && !gateDecision.promote) {\n issues.push({\n axis: 'generalization',\n severity: 'critical',\n code: `gate_${gateDecision.rejectionCode ?? 'reject'}`,\n detail: gateDecision.reason,\n })\n }\n}\n\nfunction checkDiagnostics(\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (!thresholds.requireAsiForFailures) return\n if (metrics.failedRows > metrics.failuresWithAsi) {\n issues.push({\n axis: 'diagnostics',\n severity: 'critical',\n code: 'missing_failure_asi',\n detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`,\n })\n }\n}\n\nfunction checkEfficiency(\n thresholds: Required<ReleaseConfidenceThresholds>,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): void {\n if (Number.isFinite(thresholds.maxMeanCostUsd) && metrics.meanCostUsd === null) {\n issues.push({\n axis: 'efficiency',\n severity: 'critical',\n code: 'missing_cost',\n detail: 'A finite cost limit was configured but no cost evidence is available.',\n })\n } else if (metrics.meanCostUsd !== null && metrics.meanCostUsd > thresholds.maxMeanCostUsd) {\n issues.push({\n axis: 'efficiency',\n severity: 'critical',\n code: 'cost_budget',\n detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`,\n })\n }\n if (Number.isFinite(thresholds.maxP95WallMs) && metrics.p95WallMs === null) {\n issues.push({\n axis: 'efficiency',\n severity: 'critical',\n code: 'missing_latency',\n detail: 'A finite latency limit was configured but no latency evidence is available.',\n })\n } else if (metrics.p95WallMs !== null && metrics.p95WallMs > thresholds.maxP95WallMs) {\n issues.push({\n axis: 'efficiency',\n severity: 'critical',\n code: 'latency_budget',\n detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`,\n })\n }\n}\n\nfunction buildAxes(\n metrics: ReleaseConfidenceMetrics,\n thresholds: Required<ReleaseConfidenceThresholds>,\n issues: ReleaseConfidenceIssue[],\n): ReleaseConfidenceAxis[] {\n return [\n axis(\n 'corpus',\n issues,\n bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),\n `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`,\n ),\n axis(\n 'quality',\n issues,\n metrics.passRate === null || metrics.meanScore === null\n ? null\n : Math.min(metrics.passRate, metrics.meanScore),\n `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`,\n ),\n axis(\n 'reliability',\n issues,\n metrics.reliabilityRate,\n `successRate=${fmt(metrics.reliabilityRate)} terminalFailures=${metrics.terminalFailureRuns} unclassified=${metrics.unclassifiedTerminalRuns}`,\n ),\n axis(\n 'generalization',\n issues,\n gapScore(metrics.overfitGap, thresholds.maxOverfitGap),\n `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`,\n ),\n axis(\n 'diagnostics',\n issues,\n metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,\n `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`,\n ),\n axis(\n 'efficiency',\n issues,\n efficiencyScore(metrics, thresholds),\n `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`,\n ),\n ]\n}\n\nfunction axis(\n name: ReleaseConfidenceAxisName,\n issues: ReleaseConfidenceIssue[],\n score: number | null,\n detail: string,\n): ReleaseConfidenceAxis {\n const own = issues.filter((i) => i.axis === name)\n const status = own.some((i) => i.severity === 'critical')\n ? 'fail'\n : own.length > 0\n ? 'warn'\n : 'pass'\n return { name, status, score: score === null ? null : bounded(score), detail }\n}\n\nfunction countScenarioSplits(scenarios: readonly DatasetScenario[]): Record<DatasetSplit, number> {\n const counts: Record<DatasetSplit, number> = { train: 0, dev: 0, test: 0, holdout: 0 }\n for (const scenario of scenarios) counts[scenario.split ?? 'train']++\n return counts\n}\n\nfunction countDomains(scenarios: readonly DatasetScenario[]): Record<string, number> {\n const out: Record<string, number> = {}\n for (const scenario of scenarios) {\n const domain = scenario.tags?.domain ?? scenario.tags?.category ?? 'uncategorized'\n out[domain] = (out[domain] ?? 0) + 1\n }\n return out\n}\n\nfunction countFailureClasses(\n runs: readonly RunRecord[],\n traces: readonly ReleaseTraceEvidence[],\n threshold: number,\n): Partial<Record<FailureClass, number>> {\n const out: Partial<Record<FailureClass, number>> = {}\n for (const run of runs) {\n // Ungated: a failure-mode census counts what the runs REPORTED, which is\n // the only way a gamed run's inflated score is visible at all. It is not a\n // promotion number — `passRate` is, and that one excludes gated runs and\n // publishes the excluded count as `metrics.realnessGatedRuns`.\n if (runPassOutcome(run, threshold) === false) {\n const failureClass =\n run.failureClass !== undefined && run.failureClass !== 'success'\n ? run.failureClass\n : 'unknown'\n out[failureClass] = (out[failureClass] ?? 0) + 1\n }\n }\n for (const trace of traces) {\n if (tracePassOutcome(trace, threshold) === false) {\n const failureClass =\n trace.failureClass !== undefined && trace.failureClass !== 'success'\n ? trace.failureClass\n : 'unknown'\n out[failureClass] = (out[failureClass] ?? 0) + 1\n }\n }\n return out\n}\n\nfunction countResponsibleSurfaces(traces: readonly ReleaseTraceEvidence[]): Record<string, number> {\n const out: Record<string, number> = {}\n for (const trace of traces) {\n for (const asi of trace.asi ?? []) {\n const surface = asi.responsibleSurface ?? 'unknown'\n out[surface] = (out[surface] ?? 0) + 1\n }\n }\n return out\n}\n\nfunction failedRows(\n runs: readonly RunRecord[],\n traces: readonly ReleaseTraceEvidence[],\n threshold: number,\n): Array<{ hasAsi: boolean }> {\n const out: Array<{ hasAsi: boolean }> = []\n for (const run of runs) {\n if (runPassOutcome(run, threshold) === false) {\n const asiMetric = run.outcome.raw.asi\n out.push({ hasAsi: typeof asiMetric === 'number' && asiMetric > 0 })\n }\n }\n for (const trace of traces) {\n if (tracePassOutcome(trace, threshold) === false) {\n out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 })\n }\n }\n return out\n}\n\nfunction passOutcomeRate(outcomes: readonly (boolean | null)[]): number | null {\n const classified = outcomes.filter((outcome): outcome is boolean => outcome !== null)\n if (classified.length === 0) return null\n return classified.filter(Boolean).length / classified.length\n}\n\nfunction runPassOutcome(run: RunRecord, threshold: number): boolean | null {\n if (hasExplicitTaskFailure(run)) return false\n const score = runSplitScore(run)\n return score === undefined ? null : score >= threshold\n}\n\nfunction hasExplicitTaskFailure(run: RunRecord): boolean {\n return run.failureClass !== undefined && run.failureClass !== 'success'\n}\n\nfunction tracePassOutcome(trace: ReleaseTraceEvidence, threshold: number): boolean | null {\n if (trace.failureClass !== undefined && trace.failureClass !== 'success') return false\n if (trace.ok === false) return false\n if (isFiniteNumber(trace.score)) return trace.score >= threshold\n return trace.ok === true ? true : null\n}\n\nfunction isFailedTerminalOutcome(\n outcome: RunRecord['terminalOutcome'],\n): outcome is 'failed' | 'cancelled' | 'incomplete' {\n return outcome === 'failed' || outcome === 'cancelled' || outcome === 'incomplete'\n}\n\nfunction terminalSuccess(outcome: RunRecord['terminalOutcome']): boolean | undefined {\n if (outcome === 'succeeded') return true\n if (isFailedTerminalOutcome(outcome)) return false\n return undefined\n}\n\nfunction scoresFor(runs: readonly RunRecord[], split: RunSplitTag): number[] {\n return runs\n .filter((run) => run.splitTag === split)\n .map(runSplitScore)\n .filter(isFiniteNumber)\n}\n\n/**\n * RAW and split-exact (`observedSplitScore`): this feeds the per-split means,\n * the overfit gap, and the pass threshold — descriptions of what the runs\n * reported. A gamed run inflating them is visible next to `realnessGatedRuns`,\n * and the promotion number (`passRate`) excludes gated runs entirely.\n */\nfunction runSplitScore(run: RunRecord): number | undefined {\n const score = observedSplitScore(run, run.splitTag === 'holdout' ? 'holdout' : 'search')\n return isFiniteNumber(score) ? score : undefined\n}\n\nfunction meanOrNull(xs: readonly number[]): number | null {\n if (xs.length === 0) return null\n return xs.reduce((sum, x) => sum + x, 0) / xs.length\n}\n\nfunction percentileOrNull(xs: readonly number[], p: number): number | null {\n if (xs.length === 0) return null\n const sorted = [...xs].sort((a, b) => a - b)\n return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))]!\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction diffOrNull(a: number | null, b: number | null): number | null {\n if (a === null || b === null) return null\n return a - b\n}\n\nfunction gapScore(gap: number | null, maxGap: number): number | null {\n if (gap === null) return null\n if (maxGap <= 0) return gap <= 0 ? 1 : 0\n return bounded(1 - Math.max(0, gap) / maxGap)\n}\n\nfunction efficiencyScore(\n metrics: ReleaseConfidenceMetrics,\n thresholds: Required<ReleaseConfidenceThresholds>,\n): number | null {\n const cost = Number.isFinite(thresholds.maxMeanCostUsd)\n ? metrics.meanCostUsd === null\n ? null\n : bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12))\n : 1\n const latency = Number.isFinite(thresholds.maxP95WallMs)\n ? metrics.p95WallMs === null\n ? null\n : bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12))\n : 1\n if (cost === null || latency === null) return null\n return Math.min(cost, latency)\n}\n\nfunction bounded(x: number): number {\n if (!Number.isFinite(x)) return 0\n return Math.max(0, Math.min(1, x))\n}\n\nfunction renderSummary(\n target: string,\n status: ReleaseConfidenceStatus,\n metrics: ReleaseConfidenceMetrics,\n issues: ReleaseConfidenceIssue[],\n): string {\n const prefix = `release confidence ${status}: ${target}`\n const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`\n if (issues.length === 0) return `${prefix}; ${metricText}`\n return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(',')}`\n}\n\nfunction fmt(x: number | null): string {\n if (x === null) return 'n/a'\n return x.toFixed(4)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,SAAgB,YACd,UACA,WACA,UAA4B,CAAC,GACZ;CACjB,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,WAAW,QAAQ,mBAAmB;CAC5C,MAAM,MAAM,WAAW,QAAQ,QAAQ,SAAS,UAAU,SAAS,CAAC;CAEpE,MAAM,eAAe,KAAK,QAAQ;CAClC,MAAM,gBAAgB,KAAK,SAAS;CACpC,MAAM,QAAQ,gBAAgB;CAE9B,IACE,SAAS,SAAS,UAAU,SAAS,YACrC,SAAS,WAAW,KACpB,UAAU,WAAW,GAErB,OAAO;EACL;EACA;EACA;EACA,SAAS;EACT,SAAS;EACT,YAAY;EACZ;EACA,SAAS;CACX;CAGF,MAAM,SAAmB,IAAI,MAAM,UAAU;CAC7C,KAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;EACnC,MAAM,YAAY,SAAS,UAAU,GAAG;EAExC,OAAO,KAAK,KADM,SAAS,WAAW,GACb,CAAC,IAAI,KAAK,SAAS;CAC9C;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,MAAM,WAAW,KAAK,MAAO,QAAQ,IAAK,UAAU;CACpD,MAAM,WAAW,KAAK,OAAO,IAAI,QAAQ,KAAK,UAAU,IAAI;CAC5D,MAAM,UAAU,OAAO,KAAK,IAAI,GAAG,QAAQ;CAC3C,MAAM,UAAU,OAAO,KAAK,IAAI,aAAa,GAAG,QAAQ;CAExD,IAAI;CAMJ,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,CAAC,OAAO,SAAS,OAAO,KAAK,YAAY,SACxE,UAAU;MACL,IAAI,UAAU,GAAG,UAAU;MAC7B,IAAI,UAAU,GAAG,UAAU;MAC3B,IAAI,SAAS,GAAG,UAAU;MAC1B,UAAU;CAEf,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,IAAI,KAAK;CACzB,OAAO,IAAI,GAAG;AAChB;AAEA,SAAS,SAAS,IAAc,KAA6B;CAC3D,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,IAAI,IAAI,GAAG,MAAM;CAC5E,OAAO;AACT;;AAGA,SAAS,SAAS,GAAa,GAAqB;CAClD,IAAI,IAAI;CACR,KAAK,MAAM,KAAK,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG;EAC5B,MAAM,OAAO,IAAI,aAAa,CAAC,CAAC,CAAC;EACjC,MAAM,QAAQ,IAAI,WAAW,KAAK,MAAM;EACxC,KAAK,MAAM,QAAQ,OAAO;GACxB,KAAK;GACL,IAAI,KAAK,KAAK,GAAG,QAAQ;EAC3B;CACF;CACA,OAAO,MAAM;AACf;;;;AAiCA,eAAsB,gBACpB,MACkF;CAClF,MAAM,cAAc,KAAK,oBAAoB;CAC7C,MAAM,iBAAiB,MAAM,SAAS,KAAK,iBAAiB,KAAK,OAAO,WAAW;CACnF,MAAM,kBAAkB,MAAM,SAAS,KAAK,kBAAkB,KAAK,OAAO,WAAW;CAMrF,OAAO;EACL,GANS,YAAY,gBAAgB,iBAAiB;GACtD,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,GAAI,KAAK,eAAe,KAAA,IAAY,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;GACvE,GAAI,KAAK,SAAS,KAAA,IAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;EACvD,CAEM;EACJ,iBAAiB,eAAe;EAChC,kBAAkB,gBAAgB;CACpC;AACF;AAEA,eAAe,SACb,SACA,OACA,aACmB;CACnB,MAAM,UAAoB,IAAI,MAAM,QAAQ,MAAM;CAClD,IAAI,OAAO;CACX,eAAe,SAAwB;EACrC,OAAO,MAAM;GACX,MAAM,IAAI;GACV,IAAI,KAAK,QAAQ,QAAQ;GACzB,MAAM,IAAI,MAAM,MAAM,QAAQ,EAAG;GACjC,QAAQ,KAAK,OAAO,SAAS,CAAC,IAAI,IAAI;EACxC;CACF;CACA,MAAM,QAAQ,IAAI,MAAM,KAAK,EAAE,QAAQ,KAAK,IAAI,GAAG,WAAW,EAAE,SAAS,OAAO,CAAC,CAAC;CAClF,OAAO;AACT;;;AChEA,MAAM,qBAA4D;CAChE,eAAe;CACf,kBAAkB;CAClB,eAAe;CACf,gBAAgB;CAChB,gBAAgB;CAChB,aAAa;CACb,cAAc;CACd,eAAe;CACf,gBAAgB,OAAO;CACvB,cAAc,OAAO;CACrB,uBAAuB;CACvB,uBAAuB;AACzB;AAEA,SAAgB,0BACd,OAC4B;CAC5B,MAAM,aAAa;EAAE,GAAG;EAAoB,GAAG,MAAM;CAAW;CAChE,MAAM,cAAc,MAAM,eAAe;CACzC,MAAM,OAAO,iBACV,MAAM,QAAQ,CAAC,EAAA,CAAG,IAAI,iBAAiB,GACxC,aACA,MAAM,UACR;CACA,MAAM,SAAS,sBACZ,MAAM,UAAU,CAAC,EAAA,CAAG,IAAI,4BAA4B,GACrD,aACA,MAAM,UACR;CACA,MAAM,YAAY,MAAM,aAAa,CAAC;CACtC,MAAM,gBAAgB,MAAM,SAAS,iBAAiB,UAAU;CAChE,MAAM,cAAc,MAAM,SAAS,eAAe,oBAAoB,SAAS;CAC/E,MAAM,eAAe,UAAU,MAAM,QAAQ;CAC7C,MAAM,gBAAgB,UAAU,MAAM,SAAS;CAC/C,MAAM,YAAY,KAAK,IAAI,aAAa,CAAC,CAAC,OAAO,cAAc;CAC/D,MAAM,cAAc,OAAO,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,OAAO,cAAc;CACpE,MAAM,gBAAgB,KAAK,SAAS,IAAI,YAAY;CACpD,MAAM,cAAc,KAAK,QAAQ,QAAQ,cAAc,GAAG,MAAM,KAAA,CAAS;CACzE,MAAM,eAAe,KAAK,QACvB,QACC,cAAc,GAAG,MAAM,KAAA,KACvB,CAAC,uBAAuB,GAAG,KAC3B,CAAC,wBAAwB,IAAI,eAAe,CAChD,CAAC,CAAC;CAOF,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,gBAAgB,GAAG,CAAC;CAC7D,MAAM,eACJ,KAAK,SAAS,IACV,WAAW,KAAK,QAAQ,eAAe,KAAK,WAAW,qBAAqB,CAAC,IAC7E,OAAO,KAAK,UAAU,iBAAiB,OAAO,WAAW,qBAAqB,CAAC;CACrF,MAAM,kBACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,gBAAgB,IAAI,eAAe,CAAC,IACtD,OAAO,KAAK,UAAU,MAAM,EAAE;CACpC,MAAM,2BAA2B,gBAAgB,QAAQ,YAAY,YAAY,KAAA,CAAS,CAAC,CAAC;CAC5F,MAAM,sBAAsB,gBAAgB,QAAQ,YAAY,YAAY,KAAK,CAAC,CAAC;CACnF,MAAM,kBACJ,gBAAgB,WAAW,KAAK,2BAA2B,IACvD,QACC,gBAAgB,SAAS,uBAAuB,gBAAgB;CACvE,MAAM,aAAa,YAAY,QAAQ,MAAM,EAAE,aAAa,QAAQ,CAAC,CAAC;CACtE,MAAM,cAAc,YAAY,QAAQ,MAAM,EAAE,aAAa,SAAS,CAAC,CAAC;CACxE,MAAM,SAAS,WAAW,MAAM,QAAQ,WAAW,qBAAqB;CACxE,MAAM,kBAAkB,WAAW,YAAY;CAC/C,MAAM,mBAAmB,WAAW,aAAa;CACjD,MAAM,WAAW,KAAK,SAAS,QAC7B,IAAI,eAAe,SAAS,eAAe,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG,CACzE;CACA,MAAM,aAAa,OAAO,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,OAAO,cAAc;CAC7E,MAAM,cACJ,KAAK,SAAS,IACV,SAAS,WAAW,KAAK,SACvB,WAAW,QAAQ,IACnB,OACF,WAAW,UAAU;CAC3B,MAAM,YACJ,KAAK,SAAS,IACV,KAAK,KAAK,QAAQ,IAAI,MAAM,IAC5B,OAAO,KAAK,UAAU,MAAM,UAAU,CAAC,CAAC,OAAO,cAAc;CACnE,MAAM,UAAoC;EACxC;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU,gBAAgB,YAAY;EACtC,mBAAmB,KAAK,SAAS,WAAW;EAC5C,WAAW,WAAW,aAAa;EACnC;EACA;EACA,YAAY,WAAW,iBAAiB,gBAAgB;EACxD;EACA,WAAW,iBAAiB,WAAW,GAAI;EAC3C,YAAY,OAAO;EACnB,iBAAiB,OAAO,QAAQ,QAAQ,IAAI,MAAM,CAAC,CAAC;EACpD,kBAAkB,OAAO,QAAQ,MAAM,EAAE,cAAc,CAAC,CAAC,CAAC;EAC1D,iBAAiB,OAAO,QAAQ,OAAO,EAAE,aAAa,KAAK,CAAC,CAAC,CAAC;EAC9D;EACA,cAAc,aAAa,SAAS;EACpC,oBAAoB,oBAAoB,MAAM,QAAQ,WAAW,qBAAqB;EACtF,0BAA0B,yBAAyB,MAAM;CAC3D;CAEA,MAAM,SAAmC,CAAC;CAC1C,YAAY,OAAO,YAAY,SAAS,MAAM;CAC9C,aAAa,YAAY,SAAS,MAAM;CACxC,iBAAiB,SAAS,MAAM;CAChC,oBAAoB,MAAM,gBAAgB,MAAM,YAAY,SAAS,MAAM;CAC3E,iBAAiB,YAAY,SAAS,MAAM;CAC5C,gBAAgB,YAAY,SAAS,MAAM;CAE3C,MAAM,OAAO,UAAU,SAAS,YAAY,MAAM;CAClD,MAAM,SAAS,OAAO,MAAM,MAAM,EAAE,aAAa,UAAU,IACvD,SACA,OAAO,SAAS,IACd,SACA;CAEN,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM,cAAc;EAChC;EACA,SAAS,WAAW,WAAW,MAAM,eAAe,MAAM,aAAa,UAAU;EACjF;EACA;EACA;EACA,SAAS,MAAM,WAAW;EAC1B,cAAc,MAAM,gBAAgB;EACpC,SAAS,cAAc,MAAM,QAAQ,QAAQ,SAAS,MAAM;CAC9D;AACF;AAEA,SAAgB,wBAAwB,OAA2D;CACjG,MAAM,YAAY,0BAA0B,KAAK;CACjD,IAAI,UAAU,WAAW,QACvB,MAAM,IAAI,kBAAkB,UAAU,OAAO;CAE/C,OAAO;AACT;AAEA,SAAS,gBACP,MACA,aACA,YACa;CACb,IAAI,aAAa,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,WAAW;CACxE,IAAI,YAAY,OAAO,KAAK,QAAQ,MAAM,EAAE,gBAAgB,UAAU;CACtE,OAAO,CAAC,GAAG,IAAI;AACjB;AAEA,SAAS,qBACP,QACA,aACA,YACwB;CACxB,IAAI,aACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,WAAW;CAC1F,IAAI,YACF,OAAO,OAAO,QAAQ,MAAM,EAAE,gBAAgB,KAAA,KAAa,EAAE,gBAAgB,UAAU;CACzF,OAAO,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,6BACP,OACA,OACsB;CACtB,MAAM,QAAQ;CACd,IAAI,OAAO,OAAO,OAAO,aAAa,GACpC,MAAM,IAAI,gBACR,UAAU,MAAM,2DAClB;CAEF,IAAI,MAAM,iBAAiB,KAAA,KAAa,CAAC,gBAAgB,SAAS,MAAM,YAAY,GAClF,MAAM,IAAI,gBACR,UAAU,MAAM,gCAAgC,gBAAgB,KAAK,IAAI,GAC3E;CAEF,OAAO;AACT;AAEA,SAAS,YACP,OACA,YACA,SACA,QACM;CACN,IAAI,WAAW,iBAAiB,CAAC,MAAM,YAAY,MAAM,WAAW,UAAU,OAAO,GACnF,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,gBAAgB,WAAW,kBACrC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,cAAc,qBAAqB,WAAW,iBAAiB;CACpF,CAAC;CAEH,IAAI,WAAW,kBAAkB,QAAQ,YAAY,YAAY,GAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;AAEL;AAEA,SAAS,aACP,YACA,SACA,QACM;CACN,IAAI,QAAQ,aAAa,WAAW,eAClC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,WAAW,uBAAuB,WAAW,cAAc;CAChF,CAAC;CAEH,IAAI,QAAQ,eAAe,GACzB,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa;CAClC,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,cAAc,MACrD,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;CAEH,IAAI,QAAQ,aAAa,QAAQ,QAAQ,WAAW,WAAW,aAC7D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,YAAY,IAAI,QAAQ,QAAQ,EAAE,KAAK,IAAI,WAAW,WAAW,EAAE;CAC7E,CAAC;CAEH,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cAC/D,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,iBACP,SACA,QACM;CACN,IAAI,QAAQ,oBAAoB,MAC9B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QACE,QAAQ,2BAA2B,IAC/B,GAAG,QAAQ,yBAAyB,wDACpC;CACR,CAAC;CAEH,IAAI,QAAQ,sBAAsB,GAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,oBAAoB;CACzC,CAAC;AAEL;AAEA,SAAS,oBACP,cACA,YACA,SACA,QACM;CACN,IAAI,WAAW,kBAAkB,QAAQ,cAAc,WAAW,gBAChE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,YAAY,wBAAwB,WAAW,eAAe;CACnF,CAAC;CAEH,IAAI,QAAQ,eAAe,QAAQ,QAAQ,aAAa,WAAW,eACjE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,sBAAsB,IAAI,QAAQ,UAAU,EAAE,KAAK,IAAI,WAAW,aAAa,EAAE;CAC3F,CAAC;CAEH,IAAI,gBAAgB,CAAC,aAAa,SAChC,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM,QAAQ,aAAa,iBAAiB;EAC5C,QAAQ,aAAa;CACvB,CAAC;AAEL;AAEA,SAAS,iBACP,YACA,SACA,QACM;CACN,IAAI,CAAC,WAAW,uBAAuB;CACvC,IAAI,QAAQ,aAAa,QAAQ,iBAC/B,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,GAAG,QAAQ,aAAa,QAAQ,gBAAgB;CAC1D,CAAC;AAEL;AAEA,SAAS,gBACP,YACA,SACA,QACM;CACN,IAAI,OAAO,SAAS,WAAW,cAAc,KAAK,QAAQ,gBAAgB,MACxE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,gBAAgB,QAAQ,QAAQ,cAAc,WAAW,gBAC1E,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,eAAe,IAAI,QAAQ,WAAW,EAAE,KAAK,IAAI,WAAW,cAAc,EAAE;CACtF,CAAC;CAEH,IAAI,OAAO,SAAS,WAAW,YAAY,KAAK,QAAQ,cAAc,MACpE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ;CACV,CAAC;MACI,IAAI,QAAQ,cAAc,QAAQ,QAAQ,YAAY,WAAW,cACtE,OAAO,KAAK;EACV,MAAM;EACN,UAAU;EACV,MAAM;EACN,QAAQ,aAAa,IAAI,QAAQ,SAAS,EAAE,KAAK,IAAI,WAAW,YAAY,EAAE;CAChF,CAAC;AAEL;AAEA,SAAS,UACP,SACA,YACA,QACyB;CACzB,OAAO;EACL,KACE,UACA,QACA,QAAQ,QAAQ,gBAAgB,KAAK,IAAI,GAAG,WAAW,gBAAgB,CAAC,GACxE,GAAG,QAAQ,cAAc,sBAAsB,QAAQ,YAAY,SACrE;EACA,KACE,WACA,QACA,QAAQ,aAAa,QAAQ,QAAQ,cAAc,OAC/C,OACA,KAAK,IAAI,QAAQ,UAAU,QAAQ,SAAS,GAChD,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS,GACtE;EACA,KACE,eACA,QACA,QAAQ,iBACR,eAAe,IAAI,QAAQ,eAAe,EAAE,oBAAoB,QAAQ,oBAAoB,gBAAgB,QAAQ,0BACtH;EACA,KACE,kBACA,QACA,SAAS,QAAQ,YAAY,WAAW,aAAa,GACrD,eAAe,QAAQ,YAAY,cAAc,IAAI,QAAQ,UAAU,GACzE;EACA,KACE,eACA,QACA,QAAQ,eAAe,IAAI,IAAI,QAAQ,kBAAkB,QAAQ,YACjE,mBAAmB,QAAQ,gBAAgB,GAAG,QAAQ,YACxD;EACA,KACE,cACA,QACA,gBAAgB,SAAS,UAAU,GACnC,eAAe,IAAI,QAAQ,WAAW,EAAE,aAAa,IAAI,QAAQ,SAAS,GAC5E;CACF;AACF;AAEA,SAAS,KACP,MACA,QACA,OACA,QACuB;CACvB,MAAM,MAAM,OAAO,QAAQ,MAAM,EAAE,SAAS,IAAI;CAMhD,OAAO;EAAE;EAAM,QALA,IAAI,MAAM,MAAM,EAAE,aAAa,UAAU,IACpD,SACA,IAAI,SAAS,IACX,SACA;EACiB,OAAO,UAAU,OAAO,OAAO,QAAQ,KAAK;EAAG;CAAO;AAC/E;AAEA,SAAS,oBAAoB,WAAqE;CAChG,MAAM,SAAuC;EAAE,OAAO;EAAG,KAAK;EAAG,MAAM;EAAG,SAAS;CAAE;CACrF,KAAK,MAAM,YAAY,WAAW,OAAO,SAAS,SAAS,QAAQ;CACnE,OAAO;AACT;AAEA,SAAS,aAAa,WAA+D;CACnF,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,YAAY,WAAW;EAChC,MAAM,SAAS,SAAS,MAAM,UAAU,SAAS,MAAM,YAAY;EACnE,IAAI,WAAW,IAAI,WAAW,KAAK;CACrC;CACA,OAAO;AACT;AAEA,SAAS,oBACP,MACA,QACA,WACuC;CACvC,MAAM,MAA6C,CAAC;CACpD,KAAK,MAAM,OAAO,MAKhB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,eACJ,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,YACnD,IAAI,eACJ;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OAAO;EAChD,MAAM,eACJ,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,YACvD,MAAM,eACN;EACN,IAAI,iBAAiB,IAAI,iBAAiB,KAAK;CACjD;CAEF,OAAO;AACT;AAEA,SAAS,yBAAyB,QAAiE;CACjG,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,SAAS,QAClB,KAAK,MAAM,OAAO,MAAM,OAAO,CAAC,GAAG;EACjC,MAAM,UAAU,IAAI,sBAAsB;EAC1C,IAAI,YAAY,IAAI,YAAY,KAAK;CACvC;CAEF,OAAO;AACT;AAEA,SAAS,WACP,MACA,QACA,WAC4B;CAC5B,MAAM,MAAkC,CAAC;CACzC,KAAK,MAAM,OAAO,MAChB,IAAI,eAAe,KAAK,SAAS,MAAM,OAAO;EAC5C,MAAM,YAAY,IAAI,QAAQ,IAAI;EAClC,IAAI,KAAK,EAAE,QAAQ,OAAO,cAAc,YAAY,YAAY,EAAE,CAAC;CACrE;CAEF,KAAK,MAAM,SAAS,QAClB,IAAI,iBAAiB,OAAO,SAAS,MAAM,OACzC,IAAI,KAAK,EAAE,SAAS,MAAM,KAAK,UAAU,KAAK,EAAE,CAAC;CAGrD,OAAO;AACT;AAEA,SAAS,gBAAgB,UAAsD;CAC7E,MAAM,aAAa,SAAS,QAAQ,YAAgC,YAAY,IAAI;CACpF,IAAI,WAAW,WAAW,GAAG,OAAO;CACpC,OAAO,WAAW,OAAO,OAAO,CAAC,CAAC,SAAS,WAAW;AACxD;AAEA,SAAS,eAAe,KAAgB,WAAmC;CACzE,IAAI,uBAAuB,GAAG,GAAG,OAAO;CACxC,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO,UAAU,KAAA,IAAY,OAAO,SAAS;AAC/C;AAEA,SAAS,uBAAuB,KAAyB;CACvD,OAAO,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB;AAChE;AAEA,SAAS,iBAAiB,OAA6B,WAAmC;CACxF,IAAI,MAAM,iBAAiB,KAAA,KAAa,MAAM,iBAAiB,WAAW,OAAO;CACjF,IAAI,MAAM,OAAO,OAAO,OAAO;CAC/B,IAAI,eAAe,MAAM,KAAK,GAAG,OAAO,MAAM,SAAS;CACvD,OAAO,MAAM,OAAO,OAAO,OAAO;AACpC;AAEA,SAAS,wBACP,SACkD;CAClD,OAAO,YAAY,YAAY,YAAY,eAAe,YAAY;AACxE;AAEA,SAAS,gBAAgB,SAA4D;CACnF,IAAI,YAAY,aAAa,OAAO;CACpC,IAAI,wBAAwB,OAAO,GAAG,OAAO;AAE/C;AAEA,SAAS,UAAU,MAA4B,OAA8B;CAC3E,OAAO,KACJ,QAAQ,QAAQ,IAAI,aAAa,KAAK,CAAC,CACvC,IAAI,aAAa,CAAC,CAClB,OAAO,cAAc;AAC1B;;;;;;;AAQA,SAAS,cAAc,KAAoC;CACzD,MAAM,QAAQ,mBAAmB,KAAK,IAAI,aAAa,YAAY,YAAY,QAAQ;CACvF,OAAO,eAAe,KAAK,IAAI,QAAQ,KAAA;AACzC;AAEA,SAAS,WAAW,IAAsC;CACxD,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,KAAK,MAAM,MAAM,GAAG,CAAC,IAAI,GAAG;AAChD;AAEA,SAAS,iBAAiB,IAAuB,GAA0B;CACzE,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,MAAM,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,OAAO,OAAO,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,IAAI,GAAG,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,CAAC,CAAC;AACzF;AAEA,SAAS,eAAe,OAAiC;CACvD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,WAAW,GAAkB,GAAiC;CACrE,IAAI,MAAM,QAAQ,MAAM,MAAM,OAAO;CACrC,OAAO,IAAI;AACb;AAEA,SAAS,SAAS,KAAoB,QAA+B;CACnE,IAAI,QAAQ,MAAM,OAAO;CACzB,IAAI,UAAU,GAAG,OAAO,OAAO,IAAI,IAAI;CACvC,OAAO,QAAQ,IAAI,KAAK,IAAI,GAAG,GAAG,IAAI,MAAM;AAC9C;AAEA,SAAS,gBACP,SACA,YACe;CACf,MAAM,OAAO,OAAO,SAAS,WAAW,cAAc,IAClD,QAAQ,gBAAgB,OACtB,OACA,QAAQ,WAAW,iBAAiB,KAAK,IAAI,QAAQ,aAAa,KAAK,CAAC,IAC1E;CACJ,MAAM,UAAU,OAAO,SAAS,WAAW,YAAY,IACnD,QAAQ,cAAc,OACpB,OACA,QAAQ,WAAW,eAAe,KAAK,IAAI,QAAQ,WAAW,KAAK,CAAC,IACtE;CACJ,IAAI,SAAS,QAAQ,YAAY,MAAM,OAAO;CAC9C,OAAO,KAAK,IAAI,MAAM,OAAO;AAC/B;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,cACP,QACA,QACA,SACA,QACQ;CACR,MAAM,SAAS,sBAAsB,OAAO,IAAI;CAChD,MAAM,aAAa,aAAa,QAAQ,cAAc,cAAc,QAAQ,WAAW,eAAe,QAAQ,YAAY,YAAY,IAAI,QAAQ,QAAQ,EAAE,aAAa,IAAI,QAAQ,SAAS;CAC9L,IAAI,OAAO,WAAW,GAAG,OAAO,GAAG,OAAO,IAAI;CAC9C,OAAO,GAAG,OAAO,IAAI,WAAW,WAAW,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;AAC/E;AAEA,SAAS,IAAI,GAA0B;CACrC,IAAI,MAAM,MAAM,OAAO;CACvB,OAAO,EAAE,QAAQ,CAAC;AACpB"}

package/dist/run-record-CNNzWGpB.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"run-record-CNNzWGpB.js","names":[],"sources":["../src/campaign/run-record.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunOutcome,\n RunRecord,\n RunSplitTag,\n RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n constructor(judgeName: string) {\n super(\n `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n `({ judgeName, dimension, score }) is a different type with the same name.`,\n )\n this.name = 'CampaignJudgeScoreShapeError'\n }\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const cellCostProvenance = campaignCellCostProvenance(cell)\n const costProvenance: CostProvenance =\n cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n ? { kind: 'estimated', usd: options.defaultCostUsd }\n : cellCostProvenance\n const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n // Retain the observed subtotal even when the caller supplies an estimated total.\n ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance || typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n !Number.isFinite(provenance.usd) ||\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n if (\n typeof dimensionsShape !== 'object' ||\n dimensionsShape === null ||\n Array.isArray(dimensionsShape)\n ) {\n throw new CampaignJudgeScoreShapeError(judgeName)\n }\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n },\n }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;;;;;;;;AA8CA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}

package/dist/run-record-tg_5H2IK.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"run-record-tg_5H2IK.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n * const rec = validateRunRecord(rawJson) // throws on missing\n * const ok = isRunRecord(rawJson) // boolean check\n * const rec = parseRunRecordSafe(rawJson) // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n * combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\n/**\n * Explicit execution-lifecycle result for a run.\n *\n * This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n */\nexport type RunTerminalOutcome = 'succeeded' | 'failed' | 'cancelled' | 'incomplete' | 'unknown'\n\n/** Explicit model value for a row that never produced a served model snapshot. */\nexport const UNKNOWN_MODEL = 'unknown'\n\nexport interface RunTokenUsage {\n input: number\n /** All generated tokens charged as output, including reasoning tokens. */\n output: number\n /** Present only when one or more paid calls did not report token usage.\n * In that case, every numeric field is a known subtotal, not a measured total. */\n tokensKnown?: false\n /** Reasoning-token subset of `output`, when the provider reports it. */\n reasoning?: number\n /** Prompt tokens served from a provider cache. */\n cached?: number\n /** Prompt tokens written into a provider cache. */\n cacheWrite?: number\n}\n\n/** How a run's USD amount was obtained. */\nexport type RunCostProvenance = CostProvenance\n\nexport interface RunJudgeMetadata {\n model: string\n promptVersion: string\n /** [0,1] confidence the judge declared. Constant judge confidence\n * across many runs is a fallback signal (see `canary.ts`). */\n confidence: number\n /** True if the judge degraded to a fallback path (rules-only,\n * prior-call cache, etc.). The canary uses this to alert. */\n fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n perJudge: Record<string, Record<string, number>>\n /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n perDimMean: Record<string, number>\n /** Composite mean across successful judges. Mirrors the task score only\n * when `failedJudges` is empty. */\n composite: number\n /** Judges that errored or returned an unparseable verdict. Recorded\n * by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n * not inferred from missing keys in `perJudge`. */\n failedJudges?: string[]\n /** Free-form notes the judges emitted (joined across judges or\n * first-judge only — consumer's choice). */\n notes?: string\n}\n\nexport interface RunOutcome {\n /** Score on the search/optimization split. Optional for holdout-only and\n * execution-only records. */\n searchScore?: number\n /** Score on the held-out split. Optional for search-only and execution-only\n * records. When both scores are absent, the run is explicitly unlabeled. */\n holdoutScore?: number\n /** Bag of any other metric the run produced — judge dimensions,\n * pass/fail counters, latency stats, etc. Numeric only — keeps\n * reporters honest. */\n raw: Record<string, number>\n /** Per-judge / per-dim breakdown. Consumers writing ensemble\n * judgements populate this; substrate primitives like\n * `interRaterReliability` and `corpusInterRaterAgreement` accept\n * these records as input. Optional — single-judge or scalar-only\n * runs leave it unset. */\n judgeScores?: JudgeScoresRecord\n /** Authenticity / realness verdict — did the run build the REAL thing on the\n * intended infra, or fake it (see `./authenticity`)? Optional: only domains\n * with an authenticity config populate it. Carried in the corpus so the\n * flywheel / off-policy learning can optimize for real completion, not gamed\n * pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n * must not count as a real success regardless of `score`. */\n realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n * - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n * model (after any steering bundle merge).\n * - `configHash` is the sha256 of the effective run config (model,\n * temperature, tools, judges, splits). The pair (promptHash,\n * configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n * - successful rows MUST encode a snapshot version. Bare aliases like\n * `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n * Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n * - a failed, cancelled, incomplete, or otherwise unknown row may use\n * `UNKNOWN_MODEL` when no served model was observed. This is an explicit\n * absence marker, not a fabricated snapshot.\n */\nexport interface RunRecord {\n /** UUID for the run. */\n runId: string\n /** Logical experiment grouping (a treatment vs a baseline within\n * the same sweep should share `experimentId`). */\n experimentId: string\n /** Stable identifier for the candidate (variant) being run. The\n * promotion gate compares two `candidateId`s on matched items. */\n candidateId: string\n /** RNG seed for the run. Always recorded — silent re-seeding is\n * the most common cause of non-reproducible numbers. */\n seed: number\n /** Model identifier WITH snapshot version. */\n model: string\n /** sha256 of the effective prompt (post-steering). */\n promptHash: string\n /** sha256 of the effective config. */\n configHash: string\n /** Git SHA the harness was run from. */\n commitSha: string\n /** End-to-end wall-clock duration in milliseconds. */\n wallMs: number\n /** Time spent queued before execution started, if known. */\n queueMs?: number\n /** Total USD cost, or null when the producer could not capture one. */\n costUsd: number | null\n /** Whether `costUsd` came from billing data, a price calculation, or is unavailable. */\n costProvenance: RunCostProvenance\n /** Token usage breakdown. */\n tokenUsage: RunTokenUsage\n /** Root-run or process terminal result. Never inferred from a child span. */\n terminalOutcome: RunTerminalOutcome\n /** Root-run or process failure reason. Valid only for a failed, cancelled,\n * or incomplete terminal result; never populated from a child span. */\n terminalFailureReason?: string\n /** Judge-side metadata, if a judge was used. */\n judgeMetadata?: RunJudgeMetadata\n /** Per-split scores + raw bag. */\n outcome: RunOutcome\n /** Canonical task-failure class drawn from the shared\n * `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n * evidence. Execution errors belong in\n * `outcome.raw.execution_error_count`. */\n failureClass?: FailureClass\n /** Free-form task-failure detail scoped under a non-success\n * `failureClass`. It is invalid without that class. */\n failureMode?: string\n /** Which split this run was drawn from. */\n splitTag: RunSplitTag\n /**\n * Stable scenario identifier the run observed or was scored against.\n * Comparison primitives match this identity rather than input order.\n */\n scenarioId: string\n /**\n * Canonical identity for the agent profile cell that produced this row:\n * profile artifact hash plus optional harness/model/prompt/reporting\n * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n * longitudinal reports by the complete source profile, not by a loose\n * candidate label or opaque config hash.\n */\n agentProfile?: AgentProfileCell\n}\n\n/**\n * Canonical task-result classification.\n *\n * A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n */\nexport type RunTaskFailure =\n | { failureClass?: undefined; failureMode?: undefined }\n | { failureClass: 'success'; failureMode?: undefined }\n | {\n failureClass: Exclude<FailureClass, 'success'>\n failureMode?: string\n }\n\n/**\n * Return task quality, preferring held-out evidence when both scores exist.\n *\n * RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n */\nexport function runTaskScore(record: RunRecord): number | undefined {\n const score = observedScore(record)\n return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n 'runId',\n 'experimentId',\n 'candidateId',\n 'seed',\n 'model',\n 'promptHash',\n 'configHash',\n 'commitSha',\n 'wallMs',\n 'costUsd',\n 'costProvenance',\n 'tokenUsage',\n 'terminalOutcome',\n 'outcome',\n 'splitTag',\n 'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n 'succeeded',\n 'failed',\n 'cancelled',\n 'incomplete',\n 'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n readonly path: string\n constructor(message: string, path = '') {\n super(path ? `${message} (at ${path})` : message)\n this.path = path\n }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n if (input === null || typeof input !== 'object') {\n throw new RunRecordValidationError('expected object')\n }\n const obj = input as Record<string, unknown>\n\n for (const key of MANDATORY_TOP_LEVEL) {\n if (!(key in obj)) {\n throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n }\n }\n\n expectString(obj.runId, 'runId')\n expectString(obj.experimentId, 'experimentId')\n expectString(obj.candidateId, 'candidateId')\n expectFiniteNumber(obj.seed, 'seed')\n expectString(obj.model, 'model')\n expectString(obj.promptHash, 'promptHash')\n expectString(obj.configHash, 'configHash')\n expectString(obj.commitSha, 'commitSha')\n expectNonNegativeNumber(obj.wallMs, 'wallMs')\n if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n validateCost(obj.costUsd, obj.costProvenance)\n\n // Snapshot discipline: successful rows require a served model snapshot.\n // Non-success rows may carry the explicit absence marker when execution\n // stopped before a model identity was observed.\n if (\n !modelHasSnapshot(obj.model as string) &&\n !(obj.model === UNKNOWN_MODEL && obj.terminalOutcome !== 'succeeded')\n ) {\n throw new RunRecordValidationError(\n `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD', or '${UNKNOWN_MODEL}' for a non-success row without a served model)`,\n 'model',\n )\n }\n\n // Token usage.\n const tu = obj.tokenUsage\n if (tu === null || typeof tu !== 'object') {\n throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n }\n const tuRec = tu as Record<string, unknown>\n expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n if (tuRec.tokensKnown !== undefined && tuRec.tokensKnown !== false) {\n throw new RunRecordValidationError(\n 'tokensKnown must be false when present; omit it when token usage is complete',\n 'tokenUsage.tokensKnown',\n )\n }\n if (tuRec.reasoning !== undefined) {\n expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n throw new RunRecordValidationError(\n 'reasoning tokens must be a subset of output tokens',\n 'tokenUsage.reasoning',\n )\n }\n }\n if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n if (tuRec.cacheWrite !== undefined) {\n expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n }\n\n // Judge metadata, optional.\n if (obj.judgeMetadata !== undefined) {\n const jm = obj.judgeMetadata\n if (jm === null || typeof jm !== 'object') {\n throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n }\n const jmRec = jm as Record<string, unknown>\n expectString(jmRec.model, 'judgeMetadata.model')\n expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n if (typeof jmRec.fallback !== 'boolean') {\n throw new RunRecordValidationError(\n 'judgeMetadata.fallback must be boolean',\n 'judgeMetadata.fallback',\n )\n }\n }\n\n // Outcome.\n const out = obj.outcome\n if (out === null || typeof out !== 'object') {\n throw new RunRecordValidationError('outcome must be an object', 'outcome')\n }\n const outRec = out as Record<string, unknown>\n if (outRec.searchScore !== undefined)\n expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n if (outRec.holdoutScore !== undefined)\n expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n const raw = outRec.raw\n if (raw === null || typeof raw !== 'object') {\n throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n }\n for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n expectFiniteNumber(v, `outcome.raw.${k}`)\n }\n // Realness verdict, optional.\n if (outRec.realness !== undefined) {\n const r = outRec.realness\n if (r === null || typeof r !== 'object') {\n throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n }\n const rr = r as Record<string, unknown>\n expectFiniteNumber(rr.score, 'outcome.realness.score')\n if (typeof rr.gated !== 'boolean') {\n throw new RunRecordValidationError(\n 'outcome.realness.gated must be a boolean',\n 'outcome.realness.gated',\n )\n }\n }\n\n // Per-judge / per-dim breakdown, optional.\n if (outRec.judgeScores !== undefined) {\n validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n }\n\n // Failure mode optional.\n if (\n obj.failureClass !== undefined &&\n (typeof obj.failureClass !== 'string' ||\n !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n ) {\n throw new RunRecordValidationError(\n `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n 'failureClass',\n )\n }\n if (obj.failureMode !== undefined) {\n expectString(obj.failureMode, 'failureMode')\n if (obj.failureClass === undefined || obj.failureClass === 'success') {\n throw new RunRecordValidationError(\n 'failureMode requires a non-success failureClass',\n 'failureMode',\n )\n }\n }\n\n if (\n typeof obj.terminalOutcome !== 'string' ||\n !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n ) {\n throw new RunRecordValidationError(\n `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n 'terminalOutcome',\n )\n }\n if (obj.terminalFailureReason !== undefined) {\n expectString(obj.terminalFailureReason, 'terminalFailureReason')\n if (\n obj.terminalOutcome !== 'failed' &&\n obj.terminalOutcome !== 'cancelled' &&\n obj.terminalOutcome !== 'incomplete'\n ) {\n throw new RunRecordValidationError(\n 'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n 'terminalFailureReason',\n )\n }\n }\n\n if (obj.agentProfile !== undefined) {\n try {\n const profile = validateAgentProfileCell(obj.agentProfile)\n if (profile.model !== undefined && profile.model !== obj.model) {\n throw new RunRecordValidationError(\n `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n 'agentProfile.model',\n )\n }\n if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n throw new RunRecordValidationError(\n `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n 'agentProfile.promptHash',\n )\n }\n } catch (error) {\n if (error instanceof RunRecordValidationError) throw error\n if (error instanceof Error) {\n throw new RunRecordValidationError(error.message, 'agentProfile')\n }\n throw error\n }\n }\n\n expectString(obj.scenarioId, 'scenarioId')\n\n // Split tag.\n if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n throw new RunRecordValidationError(\n `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n 'splitTag',\n )\n }\n\n return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n if (provenance === null || typeof provenance !== 'object') {\n throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n }\n const value = provenance as Record<string, unknown>\n if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n throw new RunRecordValidationError(\n 'costProvenance.kind must be observed, estimated, or uncaptured',\n 'costProvenance.kind',\n )\n }\n // A record must never read as a total it cannot support, so an uncaptured\n // cost carries no number at all. A matrix `CellResult` keeps its known\n // subtotal instead, because a cost ceiling must charge the part it can see;\n // converting one to the other drops that subtotal.\n if (value.kind === 'uncaptured') {\n if (value.usd !== null) {\n throw new RunRecordValidationError(\n 'uncaptured costProvenance.usd must be null',\n 'costProvenance.usd',\n )\n }\n if (costUsd !== null) {\n throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n }\n return\n }\n expectNonNegativeNumber(costUsd, 'costUsd')\n expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n if (value.usd !== costUsd) {\n throw new RunRecordValidationError(\n 'costProvenance.usd must equal costUsd',\n 'costProvenance.usd',\n )\n }\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n try {\n validateRunRecord(input)\n return true\n } catch {\n return false\n }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n try {\n return { ok: true, value: validateRunRecord(input) }\n } catch (e) {\n if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n throw e\n }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n const json = JSON.stringify(record)\n return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n if (typeof value !== 'string' || value.length === 0) {\n throw new RunRecordValidationError(`expected non-empty string`, path)\n }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n if (typeof value !== 'number' || !Number.isFinite(value)) {\n throw new RunRecordValidationError(`expected finite number`, path)\n }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n expectFiniteNumber(value, path)\n if ((value as number) < 0) {\n throw new RunRecordValidationError('expected non-negative number', path)\n }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n if (value === null || typeof value !== 'object') {\n throw new RunRecordValidationError('judgeScores must be an object', path)\n }\n const rec = value as Record<string, unknown>\n\n const perJudge = rec.perJudge\n if (perJudge === null || typeof perJudge !== 'object') {\n throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n }\n for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n if (dims === null || typeof dims !== 'object') {\n throw new RunRecordValidationError(\n 'per-judge entry must be an object of dimension scores',\n `${path}.perJudge.${judgeId}`,\n )\n }\n for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n }\n }\n\n const perDimMean = rec.perDimMean\n if (perDimMean === null || typeof perDimMean !== 'object') {\n throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n }\n for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n }\n\n expectFiniteNumber(rec.composite, `${path}.composite`)\n\n if (rec.failedJudges !== undefined) {\n if (!Array.isArray(rec.failedJudges)) {\n throw new RunRecordValidationError(\n 'failedJudges must be an array of strings',\n `${path}.failedJudges`,\n )\n }\n for (let i = 0; i < rec.failedJudges.length; i++) {\n const id = rec.failedJudges[i]\n if (typeof id !== 'string' || id.length === 0) {\n throw new RunRecordValidationError(\n 'failedJudges entry must be a non-empty string',\n `${path}.failedJudges[${i}]`,\n )\n }\n }\n }\n\n if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n }\n}\n\n/**\n * Snapshot check for provider model identifiers. Accepts ISO and compact\n * dates, Router's `-MMDD` snapshots, one opaque `@token`, and Vertex-style\n * `:date-token` suffixes. Routing selectors such as `@preset/name` are not\n * immutable model identities.\n */\nexport function modelHasSnapshot(model: string): boolean {\n if (model.length === 0 || model.trim() !== model) return false\n\n const opaqueAt = model.lastIndexOf('@')\n if (opaqueAt > 0) {\n const base = model.slice(0, opaqueAt)\n const token = model.slice(opaqueAt + 1)\n if (!base.includes('@') && /^[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(token)) {\n return true\n }\n }\n\n const isoDate = model.match(/-(\\d{4})-(\\d{2})-(\\d{2})$/u)\n if (isoDate && validSnapshotDate(isoDate[1]!, isoDate[2]!, isoDate[3]!)) return true\n\n const compactDate = model.match(/-(\\d{4})(\\d{2})(\\d{2})$/u)\n if (compactDate && validSnapshotDate(compactDate[1]!, compactDate[2]!, compactDate[3]!)) {\n return true\n }\n\n const routerDate = model.match(/-(\\d{2})(\\d{2})$/u)\n if (routerDate && validSnapshotDate(undefined, routerDate[1]!, routerDate[2]!)) return true\n\n return /:date-[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(model)\n}\n\nfunction validSnapshotDate(year: string | undefined, month: string, day: string): boolean {\n const monthNumber = Number(month)\n const dayNumber = Number(day)\n if (!Number.isInteger(monthNumber) || monthNumber < 1 || monthNumber > 12) return false\n\n const yearNumber = year === undefined ? undefined : Number(year)\n const leapYear =\n yearNumber === undefined ||\n (yearNumber % 4 === 0 && (yearNumber % 100 !== 0 || yearNumber % 400 === 0))\n const daysInMonth = [31, leapYear ? 29 : 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31]\n return Number.isInteger(dayNumber) && dayNumber >= 1 && dayNumber <= daysInMonth[monthNumber - 1]!\n}\n"],"mappings":";;;;;;AAiDA,MAAa,gBAAgB;;;;;;;;;AAuM7B,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAK5C,IACE,CAAC,iBAAiB,IAAI,KAAe,KACrC,EAAE,IAAI,UAAA,aAA2B,IAAI,oBAAoB,cAEzD,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,4EAA4E,cAAc,kDAC9G,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,gBAAgB,KAAA,KAAa,MAAM,gBAAgB,OAC3D,MAAM,IAAI,yBACR,gFACA,wBACF;CAEF,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAGzC,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,yBACR,kEACA,qBACF;CAMF,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;AAQA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,OAAO,OAAO;CAEzD,MAAM,WAAW,MAAM,YAAY,GAAG;CACtC,IAAI,WAAW,GAAG;EAChB,MAAM,OAAO,MAAM,MAAM,GAAG,QAAQ;EACpC,MAAM,QAAQ,MAAM,MAAM,WAAW,CAAC;EACtC,IAAI,CAAC,KAAK,SAAS,GAAG,KAAK,gDAAgD,KAAK,KAAK,GACnF,OAAO;CAEX;CAEA,MAAM,UAAU,MAAM,MAAM,4BAA4B;CACxD,IAAI,WAAW,kBAAkB,QAAQ,IAAK,QAAQ,IAAK,QAAQ,EAAG,GAAG,OAAO;CAEhF,MAAM,cAAc,MAAM,MAAM,0BAA0B;CAC1D,IAAI,eAAe,kBAAkB,YAAY,IAAK,YAAY,IAAK,YAAY,EAAG,GACpF,OAAO;CAGT,MAAM,aAAa,MAAM,MAAM,mBAAmB;CAClD,IAAI,cAAc,kBAAkB,KAAA,GAAW,WAAW,IAAK,WAAW,EAAG,GAAG,OAAO;CAEvF,OAAO,qDAAqD,KAAK,KAAK;AACxE;AAEA,SAAS,kBAAkB,MAA0B,OAAe,KAAsB;CACxF,MAAM,cAAc,OAAO,KAAK;CAChC,MAAM,YAAY,OAAO,GAAG;CAC5B,IAAI,CAAC,OAAO,UAAU,WAAW,KAAK,cAAc,KAAK,cAAc,IAAI,OAAO;CAElF,MAAM,aAAa,SAAS,KAAA,IAAY,KAAA,IAAY,OAAO,IAAI;CAI/D,MAAM,cAAc;EAAC;EAFnB,eAAe,KAAA,KACd,aAAa,MAAM,MAAM,aAAa,QAAQ,KAAK,aAAa,QAAQ,KACvC,KAAK;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;CAAE;CACnF,OAAO,OAAO,UAAU,SAAS,KAAK,aAAa,KAAK,aAAa,YAAY,cAAc;AACjG"}

package/dist/run-record-xEULpgML.d.ts.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"run-record-xEULpgML.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAgPnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}

package/dist/summary-report-SNvJoAvD.d.ts.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"summary-report-SNvJoAvD.d.ts","names":[],"sources":["../src/held-out-gate.ts","../src/summary-report.ts"],"mappings":";;;;KAwIY;;;;;;;;UAiBK;;EAEf;;;EAGA;;;;EAIA;;EAEA;;EAEA;;;EAGA;;UAGe;;;;EAIf;;;;;EAKA;;;;;;;;;;;EAWA;;;;EAIA;;;EAGA;;EAEA;;EAEA;;;;EAIA;;;;;;;;;;;;;;;;;EAiBA;;;;;;;;;;;;;EAaA;;;;KAKU,iBAAiB;UAEZ;;EAEf;;EAEA;;EAEA;;;;;;;;EAQA,iBAAiB;;;;;EAKjB,gBAAgB;;;;;EAKhB;;;;;;;;;EASA,gBAAgB;;;EAGhB;;;EAGA;IAAY;IAAa;;;;EAGzB;;;;;;;EAOA;IAAW;IAAW;IAAW;IAAqB;;;;;;EAKtD;;;;;EAKA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA;;;EAGA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA,UAAU;;EAEV;;EAEA,eAAe;;;;;;;cAQJ;mBACM;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;EAEjB,YAAY,QAAQ;;;;EAoCpB,SAAS,WAAW,aAAa,UAAU,cAAc;;;;;;;UAsTjD;;;;UCxpBO;;;EAGf;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,MAAM;EACN;EACA;;EAEA;;;;;;;iBAQc,aAAa,MAAM,aAAa,OAAM,sBAA2B;UA6IhE;EACf;;EAEA;;EAEA;;EAEA;;;EAGA;;;EAGA;;UAGe;EACf;EACA;EACA,QAAQ;EACR;IAAQ;IAAc;;;;;;;;iBAQR,YACd,MAAM,aACN;EACE;EACA,gBAAgB,eAAe;IAEhC;UAsDc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,MAAM;EACN;EACA;IAAM;IAAa;;;UAGJ;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;iBAQc,cACd,MAAM,aACN,qBACA,oBACA,OAAM,0BACL;KA8ES;;;;;;cAYC;UAEI;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;EASA;;;;;;;EAOA;IAAS;IAAa;;;;;;EAKtB;;;;;EAKA;;EAEA,gBAAgB,eAAe;;EAE/B,kBAAkB;;EAElB;;EAEA;;EAEA;;;;;;EAMA;;UAGe;EACf,UAAU;EACV;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IAAU;IAAa;;;;;;;EAMvB;EACA;;;;;;;EAOA;EACA;EACA,OAAO;EACP,UAAU;EACV;;UAGe;;;;;EAKf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;;;;;EAMA;EACA;EACA;IAAQ;IAAa;;EACrB;EACA,gBAAgB;EAChB,YAAY;EACZ,SAAS;EACT;IACE,QAAQ;IACR,OAAO;;EAET,aAAa;EACb,kBAAkB;EAClB;EACA;;;;;;;;;;;;;;;;;;;;;iBAiIoB,eACpB,MAAM,aACN,OAAM,wBACL,QAAQ"}

package/dist/tool-groups-CCQqAa16.d.ts.map DELETED Viewed

	@@ -1 +0,0 @@
1	- {"version":3,"file":"tool-groups-CCQqAa16.d.ts","names":[],"sources":["../src/analyst/tool-groups.ts"],"mappings":";;;;KAmBY;;;;;;;;;;;;;;;;;;;;iBAyDI,wBACd,OAAO,oBACP,OAAO,qBACN"}

package/dist/types-C8VBbiw6.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"types-C8VBbiw6.js","names":[],"sources":["../src/ledger-core/deep-freeze.ts","../src/analyst/types.ts"],"sourcesContent":["/** Freeze a detached canonical-JSON graph. Canonicalization has already ruled out cycles.\n *\n * Lives outside canonical.ts so the analyst-benchmark implementation digest,\n * which covers canonical.ts, stays bound to the published benchmark evidence. */\nexport function deepFreezeCanonicalJson<T>(value: T): T {\n if (value && typeof value === 'object' && !Object.isFrozen(value)) {\n Object.freeze(value)\n for (const nested of Object.values(value)) deepFreezeCanonicalJson(nested)\n }\n return value\n}\n","/**\n * Analyst contract — the missing orchestration layer over agent-eval's\n * existing analyzers (analyzeTraces, MultiLayerVerifier,\n * SemanticConceptJudge, JudgeFn, ...).\n *\n * Each existing primitive returns its own output shape. The Analyst\n * contract is the single envelope every primitive lifts into, so a\n * registry can run N analysts against a run and a single renderer can\n * compose findings without knowing which analyzer produced them.\n *\n * The contract is intentionally domain-agnostic: nothing here knows\n * about code, voice, RAG, or any particular agent stack. Analysts\n * declare what INPUT KIND they need (a trace store, an artifact dir,\n * a RunRecord, a JudgeInput, or `custom`), and the registry routes\n * the matching input from `AnalystRunInputs`.\n */\n\nimport { createHash } from 'node:crypto'\nimport type { CostLedgerHandle } from '../cost-ledger'\nimport type { RunCostProvenance, RunRecord, RunTokenUsage } from '../run-record'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeInput } from '../types'\nimport type { ChatClient } from './chat-client'\n\n/**\n * Unified envelope every analyst emits. Schema-versioned so renderers\n * and time-series diffs survive future field additions.\n */\nexport interface AnalystFinding {\n schema_version: '1.0.0'\n /**\n * Stable hash over identity-defining fields (analyst_id + canonical\n * claim + area + optional subject). Two findings from two runs that\n * \"are the same finding\" share this id — that's what `diffFindings`\n * uses to compute appeared/disappeared sets across runs.\n */\n finding_id: string\n analyst_id: string\n produced_at: string\n severity: AnalystSeverity\n /**\n * Coarse classification. Renderers group by this. Free-form so\n * domain-specific analysts can introduce categories without a\n * schema change ('agent-reasoning', 'verification', 'cost',\n * 'tool-use', 'safety', 'latency', 'data-quality', ...).\n */\n area: string\n claim: string\n rationale?: string\n evidence_refs: EvidenceRef[]\n recommended_action?: string\n validation_plan?: string\n /** 0..1 — the analyst's own confidence. Not calibrated across analysts. */\n confidence: number\n /**\n * Optional subject the finding is about — leaf id, agent id, request\n * id. Included in finding_id when present so per-subject findings\n * diff cleanly across runs.\n */\n subject?: string\n /** True when this finding was lifted from a judge result rather than observed\n * directly in a trace or artifact. Descriptive only: proposal access is\n * controlled by `ProposalFinding.proposal_origin`. */\n derived_from_judge?: boolean\n /** Analyst-private extras; renderers ignore unless they know the analyst. */\n metadata?: Record<string, unknown>\n}\n\n/** Finding severity. `AnalystSeverity` derives from this array, so the type\n * and the schema that validates a finding cannot name different levels. */\nexport const ANALYST_SEVERITIES = ['critical', 'high', 'medium', 'low', 'info'] as const\n\nexport type AnalystSeverity = (typeof ANALYST_SEVERITIES)[number]\n\n/** Data sources that candidate generation may intentionally learn from. */\nexport type ProposalFindingOrigin = 'search' | 'production'\n\n/** A finding explicitly admitted as candidate-generation input. */\nexport type ProposalFinding = AnalystFinding & {\n readonly proposal_origin: ProposalFindingOrigin\n}\n\nexport interface EvidenceRef {\n /**\n * Where the evidence lives. `span` and `event` refer to OTLP trace\n * elements; `artifact` to a file inside the run's artifact tree;\n * `finding` to another AnalystFinding (cross-analyst chaining);\n * `metric` to a named scalar reading the renderer knows how to read.\n */\n kind: 'span' | 'event' | 'artifact' | 'finding' | 'metric'\n uri: string\n excerpt?: string\n}\n\n// ── Analyst contract ─────────────────────────────────────────────────\n\n/**\n * The discriminator the registry uses to pass the right input.\n * `custom` is the escape hatch — analysts that need something else\n * (e.g. an embedding cache, a partner SDK handle) read it from\n * `AnalystRunInputs.custom[<analyst id>]`.\n */\nexport type AnalystInputKind =\n | 'trace-store'\n | 'artifact-dir'\n | 'run-record'\n | 'judge-input'\n | 'custom'\n\nexport interface AnalystCost {\n /** `deterministic` analysts MUST NOT call the LLM. */\n kind: 'deterministic' | 'llm'\n /** Optional declared upper bound; the registry can enforce a budget. */\n est_usd_per_run?: number\n /** Models the analyst expects to use (informational). */\n models?: string[]\n /** Maximum post-cancellation wait for provider usage. Model analysts default to 5 seconds. */\n settlement_timeout_ms?: number\n}\n\nexport interface AnalystRequirements {\n /** Min number of shots / samples the analyst needs to produce signal. */\n min_shots?: number\n /** Capabilities the runtime must supply (e.g. ['network', 'gpu']). */\n capabilities?: string[]\n}\n\n/**\n * What's passed to every analyst call. The registry resolves which\n * field the analyst's `inputKind` selects and asserts it's present.\n */\nexport interface AnalystRunInputs {\n traceStore?: TraceAnalysisStore\n artifactDir?: string\n runRecord?: RunRecord\n judgeInput?: JudgeInput\n /** Keyed by analyst id; populated by callers that registered custom analysts. */\n custom?: Record<string, unknown>\n}\n\nexport interface AnalystContext {\n runId: string\n /** Stable correlation id so logs from a single registry.run() share a tag. */\n correlationId: string\n /** Enforced wall-clock deadline (epoch ms). */\n deadlineMs?: number\n /** Per-analyst USD budget. Analysts MAY check before issuing LLM calls. */\n budgetUsd?: number\n /** Shared paid-call account when the analyst runs inside a larger campaign. */\n costLedger?: CostLedgerHandle\n /** Attribution phase used when writing to the shared paid-call account. */\n costPhase?: string\n /**\n * Shared chat client. Analysts that call an LLM go through this so\n * the operator picks transport (sandbox-sdk | router | cli-bridge |\n * direct-provider | mock) at the registry boundary without touching\n * analyst code.\n */\n chat?: ChatClient\n /**\n * Findings from a prior run the operator wants the analyst to see as\n * retrieval context. Kinds that take advantage of cross-run memory\n * (failure-mode \"I saw this cluster last run\", knowledge-gap \"the wiki\n * page I asked for is still missing\") render these into the actor's\n * working set. Filtering is the operator's job: pass the slice that\n * matches the analyst's id, or pass everything and let the kind\n * filter. Empty / absent means no cross-run context.\n */\n priorFindings?: ReadonlyArray<AnalystFinding>\n /**\n * Findings emitted by analysts that completed earlier in this registry run.\n * This is separate from `priorFindings`: upstream findings are dependency\n * context for the current pass, while prior findings are cross-run memory.\n * The registry populates this only when `RegistryRunOpts.chainFindings` is on.\n */\n upstreamFindings?: ReadonlyArray<AnalystFinding>\n /**\n * Report metered work independently of findings. This keeps an empty finding\n * set from erasing token/cost telemetry. Multiple receipts are accumulated.\n */\n recordUsage?: (receipt: AnalystUsageReceipt) => void\n /** Free-form runtime tags (env, host, op). Findings can echo these into metadata. */\n tags?: Record<string, string>\n /** Logger callback — analysts SHOULD prefer this over console.* for testability. */\n log?: (msg: string, fields?: Record<string, unknown>) => void\n /** Optional abort signal. Analysts SHOULD pass it through to LLM calls. */\n signal?: AbortSignal\n /**\n * Optional live-execution port. A runtime that owns a sandbox or checkout\n * fills it so an analyst can execute a bounded probe against the run's\n * produced state instead of reasoning about it from the trace alone. This\n * package defines only the port: no field here reaches for an agent loop,\n * and an absent probe means the analyst works from recorded evidence.\n */\n probe?: ExecutionProbe\n}\n\n// ── Live-execution port ─────────────────────────────────────────────\n\n/** One bounded command an analyst asks the probe to run. */\nexport interface ExecutionProbeRequest {\n command: string\n /** Working directory inside the probed environment. */\n cwd?: string\n /** Hard wall-clock deadline for this one execution. */\n timeoutMs: number\n /** Bytes of combined output retained; the prober truncates beyond it. */\n maxOutputBytes?: number\n signal?: AbortSignal\n}\n\n/**\n * Typed outcome of one probe execution. `succeeded: false` is a PROBE failure\n * (the environment could not run the command); a command that ran and exited\n * non-zero is a successful observation with a non-zero `exitCode`.\n */\nexport type ExecutionProbeOutcome =\n | {\n succeeded: true\n exitCode: number\n stdout: string\n stderr: string\n durationMs: number\n /** True when output was cut at `maxOutputBytes`. */\n truncated: boolean\n }\n | { succeeded: false; error: { class: string; message: string } }\n\n/**\n * The seam a runtime fills to let analysts observe produced state live.\n * Implementations own sandboxing, credentials, and cleanup; analysts only\n * submit bounded requests and read typed outcomes.\n */\nexport interface ExecutionProbe {\n /** One plain sentence naming what is being probed (e.g. a sandbox id). */\n readonly description: string\n execute(request: ExecutionProbeRequest): Promise<ExecutionProbeOutcome>\n}\n\n/**\n * The minimal contract. Concrete analysts can refine `TInput` so\n * implementations stay type-safe (e.g. a trace analyst's `TInput` is\n * `TraceAnalysisStore`); the registry passes the right field from\n * `AnalystRunInputs` based on `inputKind`.\n */\nexport interface Analyst<TInput = unknown> {\n /** Stable identifier — appears in finding_id, telemetry, and registry exclusion lists. */\n readonly id: string\n /** Human-readable. One sentence. */\n readonly description: string\n readonly inputKind: AnalystInputKind\n readonly cost: AnalystCost\n readonly requires?: AnalystRequirements\n /** Bump on breaking changes to claim wording or area so old finding_ids don't collide. */\n readonly version: string\n analyze(input: TInput, ctx: AnalystContext): Promise<AnalystFinding[]>\n}\n\n/** Metered work performed by one analyst call. */\nexport interface AnalystUsageReceipt {\n /** Number of model-usage records observed at the provider boundary. */\n calls: number | null\n /** Null when the provider did not return token accounting. */\n tokens: RunTokenUsage | null\n /** Observed, estimated, or explicitly uncaptured dollar cost. */\n cost: RunCostProvenance\n /** Known lower bound when one or more calls have uncaptured cost. */\n knownCostUsd?: number\n /**\n * Token counts the provider reported on only one side. Present exactly when\n * `tokens` is null and at least one side WAS reported: `RunTokenUsage` has no\n * nullable side, so a one-sided count cannot live in `tokens` without writing\n * a zero nobody measured. Read it as a lower bound, never as a total — the\n * field exists so a null `tokens` cannot hide a real count.\n */\n partialTokens?: { input: number | null; output: number | null }\n /**\n * True when the token counts were DERIVED by the transport (from character\n * lengths, say) rather than measured by the model provider. `cost.kind` is\n * `estimated` both for a rate estimate over exact tokens and for one over\n * derived tokens; this is the field that separates them.\n */\n tokensEstimated?: boolean\n}\n\n// ── finding_id stability ─────────────────────────────────────────────\n\n/**\n * Compute the stable finding_id from the identity-defining fields.\n * Default implementation hashes {analyst_id, area, subject, normalized claim}.\n * Analysts that emit findings whose claim text varies per run (timestamps,\n * counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,\n * or (b) move the variable part into `rationale`/`metadata` and keep the\n * `claim` static.\n */\nexport function computeFindingId(input: {\n analyst_id: string\n area: string\n subject?: string\n claim: string\n /** Override the claim for hashing — use when the displayed claim has run-specific bits. */\n id_basis?: string\n}): string {\n const basis = JSON.stringify({\n a: input.analyst_id,\n r: input.area,\n s: input.subject ?? '',\n c: normalizeClaim(input.id_basis ?? input.claim),\n })\n return `f_${createHash('sha256').update(basis).digest('hex').slice(0, 20)}`\n}\n\nfunction normalizeClaim(c: string): string {\n // Lowercase, collapse whitespace, strip trailing punctuation. Goal:\n // \"Leaf X failed install\" and \"Leaf X failed install.\" hash the same.\n return c\n .toLowerCase()\n .replace(/\\s+/g, ' ')\n .replace(/[.!?;:,]+$/g, '')\n .trim()\n}\n\n/**\n * Convenience factory: produce a fully-formed AnalystFinding with the\n * id computed automatically. Analyst code stays terse.\n */\nexport function makeFinding(\n init: Omit<AnalystFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n id_basis?: string\n produced_at?: string\n },\n): AnalystFinding {\n const { id_basis, produced_at, ...rest } = init\n return {\n schema_version: '1.0.0',\n finding_id: computeFindingId({\n analyst_id: rest.analyst_id,\n area: rest.area,\n subject: rest.subject,\n claim: rest.claim,\n id_basis,\n }),\n produced_at: produced_at ?? new Date().toISOString(),\n ...rest,\n }\n}\n\n/** Build a finding whose source is explicitly allowed during candidate generation. */\nexport function makeProposalFinding(\n init: Omit<ProposalFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n id_basis?: string\n produced_at?: string\n },\n): ProposalFinding {\n const { proposal_origin, ...finding } = init\n return { ...makeFinding(finding), proposal_origin }\n}\n\n// ── Registry result envelope ────────────────────────────────────────\n\nexport interface AnalystRunSummary {\n analyst_id: string\n status: 'ok' | 'skipped' | 'failed'\n /** Why skipped — missing input, budget exceeded, capability unmet. */\n reason?: string\n findings_count: number\n latency_ms: number\n /** Additive model usage and cost provenance for this analyst. */\n usage: AnalystUsageReceipt\n /** When `status='failed'`: the error class + message, never the full stack. */\n error?: { class: string; message: string }\n}\n\nexport interface AnalystRunResult {\n run_id: string\n correlation_id: string\n started_at: string\n ended_at: string\n findings: AnalystFinding[]\n per_analyst: AnalystRunSummary[]\n /** Total LLM cost in USD across all analysts in this registry.run(). */\n total_cost_usd: number\n /**\n * Provenance for `total_cost_usd`. When uncaptured, the numeric field is only\n * the known subtotal and must not be treated as the run's total spend.\n */\n total_cost_provenance?: RunCostProvenance\n}\n\n// ── Streaming event envelope ────────────────────────────────────────\n\n/**\n * Events emitted by `AnalystRegistry.runStream(...)` in real time as\n * the registry executes. UIs subscribe via `for await (const ev of\n * registry.runStream(...))`; `registry.run(...)` is a thin collector\n * over the same stream, so the two surfaces share their invariants.\n *\n * Per-finding events are intentionally omitted — analyzers are batch\n * operations (a recursive engine returns the full `findings:json[]` at the\n * end of the responder), so streaming inside one analyst would only\n * emit partial JSON consumers can't render. The kind-completion event\n * is the right granularity; subscribers wanting per-finding rendering\n * iterate `event.findings` themselves.\n */\nexport type AnalystRunEvent =\n | {\n type: 'run-started'\n run_id: string\n correlation_id: string\n started_at: string\n /** The ordered list of analyst ids the registry will run. */\n analyst_ids: ReadonlyArray<string>\n }\n | {\n type: 'analyst-skipped'\n summary: AnalystRunSummary\n }\n | {\n type: 'analyst-started'\n analyst_id: string\n started_at: string\n }\n | {\n type: 'analyst-completed'\n /** `summary.status` is `'ok'` for clean completion or `'failed'` for thrown analysts. */\n summary: AnalystRunSummary\n findings: ReadonlyArray<AnalystFinding>\n }\n | {\n type: 'run-completed'\n result: AnalystRunResult\n }\n"],"mappings":";;;;;;AAIA,SAAgB,wBAA2B,OAAa;CACtD,IAAI,SAAS,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;EACjE,OAAO,OAAO,KAAK;EACnB,KAAK,MAAM,UAAU,OAAO,OAAO,KAAK,GAAG,wBAAwB,MAAM;CAC3E;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;AC4DA,MAAa,qBAAqB;CAAC;CAAY;CAAQ;CAAU;CAAO;AAAM;;;;;;;;;AAiO9E,SAAgB,iBAAiB,OAOtB;CACT,MAAM,QAAQ,KAAK,UAAU;EAC3B,GAAG,MAAM;EACT,GAAG,MAAM;EACT,GAAG,MAAM,WAAW;EACpB,GAAG,eAAe,MAAM,YAAY,MAAM,KAAK;CACjD,CAAC;CACD,OAAO,KAAK,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1E;AAEA,SAAS,eAAe,GAAmB;CAGzC,OAAO,EACJ,YAAY,CAAC,CACb,QAAQ,QAAQ,GAAG,CAAC,CACpB,QAAQ,eAAe,EAAE,CAAC,CAC1B,KAAK;AACV;;;;;AAMA,SAAgB,YACd,MAIgB;CAChB,MAAM,EAAE,UAAU,aAAa,GAAG,SAAS;CAC3C,OAAO;EACL,gBAAgB;EAChB,YAAY,iBAAiB;GAC3B,YAAY,KAAK;GACjB,MAAM,KAAK;GACX,SAAS,KAAK;GACd,OAAO,KAAK;GACZ;EACF,CAAC;EACD,aAAa,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;EACnD,GAAG;CACL;AACF;;AAGA,SAAgB,oBACd,MAIiB;CACjB,MAAM,EAAE,iBAAiB,GAAG,YAAY;CACxC,OAAO;EAAE,GAAG,YAAY,OAAO;EAAG;CAAgB;AACpD"}