@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/schema-C6DW4ZHR.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"schema-C6DW4ZHR.js","names":[],"sources":["../src/rollout/gate-checks.ts","../src/rollout/schema.ts"],"sourcesContent":["/**\n * THE canonical list of anti-Goodhart gate checks, plus the TOTAL policy every\n * entry point has to declare over it.\n *\n * Four rounds of adversarial review found the same defect four times, and it\n * was never the check itself: it was the COMPOSITION. `validateRolloutLine`\n * composed one check, `assertMinted` composed two, `assertRewardGate` composed\n * two of the three, `assertGateReport` composed its own pair — each by hand, in\n * its own file. So a check added to the package applied wherever its author\n * happened to remember, and the guard that forgot it looked exactly like the\n * guard that didn't. The last leak was literally that: `assertRewardGate`\n * composed `reward-relationship` + `gated-evidence` and not `unscreened-reward`,\n * so a never-screened positive reward that `assertMinted` correctly REFUSED\n * walked through all four waist exporters at full value.\n *\n * The fix is to make hand-composition impossible rather than to add a third\n * call to the two places that had two:\n *\n *   - `GATE_CHECKS` is a TOTAL map over `GateCheckId`. A new id with no\n *     implementation does not compile.\n *   - `GatePolicy` is a TOTAL map over `GateCheckId`. Every entry point\n *     declares one, so a new id makes EVERY entry point's policy a type error\n *     until it is wired. Wiring it means writing `enforced`, `repairedBy(...)`\n *     or `omittedBecause(...)` — and the last two force a written reason, so a\n *     silent gap is not expressible.\n *   - every check carries a `tripwire`: the minimal outcome it must refuse.\n *     `gate-checks.test.ts` feeds each tripwire to every entry point that\n *     declares `enforced` and requires a rejection, so wiring a check to the\n *     wrong disposition is a TEST failure even when it type-checks.\n *\n * Adding a check is therefore: append the id, write the check, and the compiler\n * enumerates every place that has to decide about it.\n */\n\nimport type { GatedEvidence, RolloutOutcome, RolloutStep } from './schema'\n\n/**\n * Every gate check in the package, in the order they are applied.\n *\n * Order is load-bearing only for which message a caller sees first: a line that\n * trips two checks reports the earlier one, and `reward-relationship` is first\n * because it is the invariant the other three protect.\n */\nexport const GATE_CHECK_IDS = [\n  'reward-relationship',\n  'gated-evidence',\n  'undeclared-step-payload',\n  'unscreened-reward',\n] as const\n\nexport type GateCheckId = (typeof GATE_CHECK_IDS)[number]\n\n/**\n * An outcome as it reaches a check.\n *\n * Deliberately accepts a raw record as well as the typed shape: the checks are\n * the RUNTIME half of the gate, and the callers they exist for — JSON off a\n * ledger, a plain-JavaScript consumer of the published package — arrive with no\n * types at all. `Partial` because a tripwire states only the fields it trips on.\n */\nexport type GateCheckedOutcome = Partial<RolloutOutcome> | Readonly<Record<string, unknown>>\n\n/**\n * What a gate check reads: the reward-bearing surface of ONE LINE.\n *\n * For three rounds the subject was the OUTCOME alone, and that assumption is\n * what produced the next leak rather than any missing check: `steps[]` sits on\n * the LINE, outside `outcome`, so a per-step reward on a gated line was read by\n * no check at all while `toRewardRows` copied it out verbatim — through the\n * MINTED door, not merely the raw one. Widening the subject is what makes\n * \"somewhere else on the line\" a place the checks can see.\n *\n * `outcome` is REQUIRED, and that is the point: a bare `RolloutOutcome` is then\n * not assignable to a subject, so every call site that used to pass one is a\n * COMPILE error until it passes the line instead. A subject with an optional\n * `outcome` would have let the old call sites keep compiling while silently\n * checking nothing — the exact failure this module exists to make impossible.\n */\nexport interface GateSubject {\n  outcome: GateCheckedOutcome\n  /** The line's trajectory steps, when it carries any. */\n  steps?: unknown\n}\n\n/** One untyped field read off the outcome, so every check narrows from the same place. */\nconst field = (subject: GateSubject, name: string): unknown =>\n  (subject.outcome as Readonly<Record<string, unknown>>)?.[name]\n\nexport interface GateCheck {\n  id: GateCheckId\n  /** One sentence: what this check refuses. */\n  refuses: string\n  /** One dotted-path message per defect; `[]` when the line is clean. */\n  errors: (subject: GateSubject) => string[]\n  /**\n   * Every minimal subject that MUST trip `errors` — the executable form of\n   * `refuses`, and the reason a check cannot be added without being provable.\n   * The calibration test feeds each one to every entry point declaring\n   * `enforced`.\n   *\n   * A LIST rather than one case: a check that refuses two distinct populations\n   * (a positive reward AND a reward it cannot read as a number) proved able to\n   * hold for the first while silently passing the second, so each population\n   * states its own tripwire and each is exercised separately.\n   */\n  tripwires: GateSubject[]\n}\n\n// ---------------------------------------------------------------------------\n// Total readers. Every \"is this positive?\" / \"is this populated?\" question the\n// checks ask goes through one of these.\n// ---------------------------------------------------------------------------\n\n/**\n * How the gate reads one `outcome.reward`.\n *\n * TOTAL over `unknown`, with an explicit `unreadable` case, because the\n * recurring defect across four rounds was a hand-written type test whose\n * NEGATIVE branch silently passed: `typeof reward !== 'number' || !(reward > 0)`\n * returned \"clean\" for `reward: \"0.95\"`, so a plain-JavaScript producer that\n * stringifies its numbers — the exact population the runtime backstop exists\n * for — shipped a gamed run at full value through three waist exporters.\n *\n * A value the gate cannot read as a number is not evidence the reward is\n * cleared; it is the absence of that evidence, and the gate fails closed on it.\n */\nexport type RewardReading =\n  /** `null`/`undefined` — a labeled gap, which is never a positive signal. */\n  | { kind: 'absent' }\n  /** A number at or below zero: the gate's own verdict, applied. */\n  | { kind: 'cleared' }\n  /** A number above zero, `Infinity` and the 5e-324 denormal included. */\n  | { kind: 'positive'; shown: string }\n  /** Anything the wire format does not permit here, plus `NaN`. */\n  | { kind: 'unreadable'; shown: string }\n\nexport function readReward(value: unknown): RewardReading {\n  if (value === null || value === undefined) return { kind: 'absent' }\n  if (typeof value === 'number') {\n    if (Number.isNaN(value)) return { kind: 'unreadable', shown: 'NaN' }\n    return value > 0 ? { kind: 'positive', shown: String(value) } : { kind: 'cleared' }\n  }\n  return {\n    kind: 'unreadable',\n    shown: `${JSON.stringify(value) ?? String(value)} (${typeof value})`,\n  }\n}\n\nconst isPlainRecord = (value: unknown): value is Record<string, unknown> =>\n  typeof value === 'object' && value !== null && !Array.isArray(value)\n\n/**\n * Whether a reward-derived field carries anything at all.\n *\n * The previous emptiness test was `Object.keys(value).length > 0`, which reads\n * `[]` for a NUMBER, a BOOLEAN and the empty string — so `metrics: 0.95` on a\n * gated line counted as empty and shipped verbatim in the verifiers format's\n * per-rubric score dict. `Object.keys` being total over primitives is not the\n * same property as being CORRECT over them.\n *\n * The rule that has no such hole: only `undefined`, `null` and a record with no\n * keys are empty. Everything else is payload, whatever its type.\n */\nexport function payloadIsPopulated(value: unknown): boolean {\n  if (value === undefined || value === null) return false\n  if (isPlainRecord(value)) return Object.keys(value).length > 0\n  return true\n}\n\n/**\n * Every key `tangle.rollout.v1` declares on a step, as a TOTAL map over the\n * interface: a field added to `RolloutStep` and not to this list does not\n * compile, so the list cannot silently fall behind the schema it mirrors.\n *\n * This is the producer's OWN classification, which is the only partition\n * `gatedEvidenceOf` accepts — a key-name heuristic (\"strip anything matching\n * /reward|score/\") holds until someone names a field `credit` and the next\n * per-step signal ships at full value.\n */\nconst DECLARED_STEP_KEYS: { readonly [K in keyof Required<RolloutStep>]: true } = {\n  kind: true,\n  name: true,\n  input: true,\n  output: true,\n  status: true,\n  durationMs: true,\n  llm_call_count: true,\n  prompt_token_ids: true,\n  completion_token_ids: true,\n  logprobs: true,\n}\n\n/**\n * The part of `steps` the wire format has no name for — a per-step reward, a\n * per-step score, whatever a producer decided to hang there.\n *\n * `undefined` when the steps carry nothing undeclared.\n */\nexport function undeclaredStepPayload(steps: unknown): unknown {\n  if (steps === undefined || steps === null) return undefined\n  // The format says `steps` is a list. Anything else is wholly unclassified.\n  if (!Array.isArray(steps)) return steps\n  const extras: Array<Record<string, unknown>> = []\n  let found = false\n  for (const step of steps) {\n    const extra: Record<string, unknown> = {}\n    if (isPlainRecord(step)) {\n      for (const [key, value] of Object.entries(step)) {\n        if (key in DECLARED_STEP_KEYS) continue\n        extra[key] = value\n        found = true\n      }\n    } else if (step !== undefined && step !== null) {\n      // A step that is not even a record carries no declared key to keep.\n      extras.push({ step } as Record<string, unknown>)\n      found = true\n      continue\n    }\n    extras.push(extra)\n  }\n  return found ? extras : undefined\n}\n\n/** `steps` projected down to the keys the wire format declares. */\nexport function declaredSteps(steps: unknown): RolloutStep[] | undefined {\n  if (!Array.isArray(steps)) return undefined\n  const kept: RolloutStep[] = []\n  for (const step of steps) {\n    if (!isPlainRecord(step)) continue\n    const projected: Record<string, unknown> = {}\n    for (const [key, value] of Object.entries(step)) {\n      if (key in DECLARED_STEP_KEYS) projected[key] = value\n    }\n    kept.push(projected as unknown as RolloutStep)\n  }\n  return kept\n}\n\n/**\n * THE anti-Goodhart invariant, checked as a RELATIONSHIP between two fields\n * rather than as two independent type checks.\n *\n * Everything upstream of a training export is allowed to be wrong; this is the\n * one thing that cannot be. `realness_gated: true` means the run faked its\n * success signal, so its reward is a fabrication, and a fabrication above zero\n * is precisely what a trainer would learn to reproduce. Validating only that\n * `reward` is a number and `realness_gated` is a boolean is what let a line\n * claiming `{reward: 0.95, realness_gated: true}` validate clean and walk\n * through every exporter.\n */\nconst rewardRelationship: GateCheck = {\n  id: 'reward-relationship',\n  refuses: 'a positive — or unreadable — reward on a line the authenticity screen flagged as gamed',\n  tripwires: [\n    { outcome: { reward: 0.95, realness_gated: true } },\n    // The second population, stated separately because the check held for the\n    // first while passing this one for a full round.\n    { outcome: { reward: '0.95' as unknown as number, realness_gated: true } },\n  ],\n  errors: (subject) => {\n    if (field(subject, 'realness_gated') !== true) return []\n    const reading = readReward(field(subject, 'reward'))\n    if (reading.kind === 'absent' || reading.kind === 'cleared') return []\n    if (reading.kind === 'unreadable') {\n      return [\n        `outcome.reward: ${reading.shown} with outcome.realness_gated: true — the wire format ` +\n          'declares this field `number | null`, so the gate cannot read this value as a number ' +\n          'and cannot establish that the reward was cleared. On a run flagged as gamed that is ' +\n          'not evidence of a zero, it is the absence of it, and a consumer that coerces the ' +\n          'value (`\"0.95\"`, `\"2\"`) reads a positive reward off a faked success. Emit a number ' +\n          'or null — `trainingReward` / `trainingScore` from `rollout/reward.ts` produce one.',\n      ]\n    }\n    return [\n      `outcome.reward: ${reading.shown} with outcome.realness_gated: true — a run flagged as ` +\n        'gamed may not carry a positive reward. The anti-Goodhart gate forces the reward to 0 (a ' +\n        'real verdict: the gate decided) before the line is written, so a fine-tune cannot learn ' +\n        'from a faked success. Derive the reward with `trainingReward` / `trainingScore` from ' +\n        '`rollout/reward.ts`, or drop the line.',\n    ]\n  },\n}\n\n/**\n * The reward-bearing outcome fields that are NOT the scalar: the numbers the\n * reward was computed from, and the verdict record that claimed it.\n *\n * Returned as one block rather than filtered key-by-key. A key-name heuristic\n * (\"zero anything matching `layer.*` or `/score/`\") is the same defect shape as\n * the line-oriented regex the AST score guard replaced: it holds until someone\n * names a metric `pass_fraction`, and the next reward-shaped key ships at full\n * value. The producer's OWN classification — \"this is the scalar, that is\n * everything else\" — is the only partition that cannot be out-guessed.\n */\nexport function gatedEvidenceOf(subject: GateSubject): GatedEvidence | undefined {\n  const evidence: GatedEvidence = {}\n  // `payloadIsPopulated`, not an inline key count: a plain-JavaScript caller can\n  // hand any value here, and a non-record `metrics` is still reward-derived\n  // payload. See that function for the primitives an `Object.keys` test missed.\n  const metrics = field(subject, 'metrics')\n  if (payloadIsPopulated(metrics)) evidence.metrics = metrics as Record<string, unknown>\n  const verdict = field(subject, 'verdict')\n  if (verdict !== null && verdict !== undefined) evidence.verdict = verdict\n  const steps = undeclaredStepPayload(subject.steps)\n  if (steps !== undefined) evidence.steps = steps\n  return evidence.metrics === undefined && evidence.verdict === undefined && steps === undefined\n    ? undefined\n    : evidence\n}\n\n/**\n * The invariant is about the OUTCOME, not about one field of it.\n *\n * `mintRolloutRows` bulk-copied `RunRecord.outcome.raw` into `outcome.metrics`\n * with no gate, so a gated run exported `reward: 0` (correct) while the\n * deterministic per-layer scores that reward was COMPUTED FROM — the `layer.*`\n * keys `rl/verifiable-reward.ts` calls the RL training signal — shipped at 1.0,\n * in the top-level `metrics` dict of the Prime Intellect verifiers format, which\n * IS that format's per-rubric score dict. `verdict` leaks the same way into\n * `toRftItem`'s `reference.verdict`, where a grader author reads\n * `resolved: true` off a run that faked it.\n */\nconst gatedEvidence: GateCheck = {\n  id: 'gated-evidence',\n  refuses: 'the numbers a fabricated reward was computed from, riding along at reward 0',\n  tripwires: [\n    { outcome: { reward: 0, realness_gated: true, metrics: { 'layer.tests': 1 } } },\n    // A PRIMITIVE `metrics`: `Object.keys(0.95)` is `[]`, so the emptiness test\n    // this check used to run reported it clean and the verifiers format shipped\n    // it as its per-rubric score dict.\n    { outcome: { reward: 0, realness_gated: true, metrics: 0.95 as unknown as never } },\n  ],\n  errors: (subject) => {\n    if (field(subject, 'realness_gated') !== true) return []\n    const evidence = gatedEvidenceOf(subject)\n    if (evidence === undefined) return []\n    // Each check reports only the fields it owns; `steps` is `undeclared-step-payload`.\n    if (evidence.metrics === undefined && evidence.verdict === undefined) return []\n    const path = evidence.metrics !== undefined ? 'outcome.metrics' : 'outcome.verdict'\n    return [\n      `${path} is populated with outcome.realness_gated: true — a run flagged as gamed may not ` +\n        'ship the numbers its fabricated reward was computed from, even at reward 0. The ' +\n        'per-layer scores in `metrics` ARE the reward signal in the verifiers format, and ' +\n        '`verdict` is the record that claimed the success. Mint through `assertMinted`, which ' +\n        'relocates both to `provenance.gated_evidence` (see `gateGamedOutcome`), or drop the line.',\n    ]\n  },\n}\n\n/**\n * A reward-bearing field hung on `steps[]`, where the gate was not looking.\n *\n * The first three checks all read `outcome`, so the whole apparatus was blind to\n * anything a producer wrote elsewhere on the line — and `toRewardRows` emits\n * `steps` verbatim beside the scalar it just forced to 0. A generated corpus\n * found it in twenty cases: a gated line carrying\n * `steps: [{kind, name, reward: 0.86}]` exported `{\"reward\": 0, \"steps\":\n * [{\"reward\": 0.86}]}`, which is the per-step credit assignment of a run that\n * faked its success, at full value, through the MINTED door.\n *\n * `RolloutStep` declares ten keys and none of them is a reward, so the partition\n * needs no judgement call: whatever the format does not name is unclassified\n * payload, and on a gated line unclassified payload is exactly what the gate\n * refuses. `assertMinted` relocates it rather than rejecting, like the sibling\n * check, so an already-published ledger stays readable.\n */\nconst undeclaredStepPayloadCheck: GateCheck = {\n  id: 'undeclared-step-payload',\n  refuses: 'per-step reward signal hung on a gated line’s steps, outside every field the gate read',\n  tripwires: [\n    {\n      outcome: { reward: 0, realness_gated: true },\n      steps: [{ kind: 'tool', name: 'edit', reward: 0.9 }],\n    },\n  ],\n  errors: (subject) => {\n    if (field(subject, 'realness_gated') !== true) return []\n    const extra = undeclaredStepPayload(subject.steps)\n    if (extra === undefined) return []\n    return [\n      `steps[] carries fields \\`${ROLLOUT_SCHEMA_NAME}\\` does not declare (${JSON.stringify(extra).slice(0, 120)}) ` +\n        'with outcome.realness_gated: true — a per-step reward is training signal exactly like ' +\n        'the scalar, and the exporters copy `steps` through verbatim, so zeroing `outcome.reward` ' +\n        'while leaving it in place ships the gamed run’s step-level credit assignment at full ' +\n        'value. Mint through `assertMinted`, which projects the steps down to the declared keys ' +\n        'and relocates the rest to `provenance.gated_evidence.steps` (see `gateGamedOutcome`).',\n    ]\n  },\n}\n\n/** Named here rather than imported, so `gate-checks` stays free of schema cycles. */\nconst ROLLOUT_SCHEMA_NAME = 'tangle.rollout.v1'\n\n/**\n * A positive reward whose producer DECLARED that no authenticity screen exists.\n *\n * `realness_gated: false` is the screen's VERDICT, so writing it with no screen\n * behind it claims \"we looked and nothing fired\" about a reward nobody looked\n * at. `realness_screened: false` is the producer saying so out loud, and a\n * positive reward carrying it is exactly the signal the gate exists to qualify\n * with nothing having qualified it.\n */\nconst unscreenedReward: GateCheck = {\n  id: 'unscreened-reward',\n  refuses: 'a positive reward whose producer declares that no authenticity screen ever ran',\n  tripwires: [\n    { outcome: { reward: 1, realness_gated: false, realness_screened: false } },\n    {\n      outcome: {\n        reward: '2' as unknown as number,\n        realness_gated: false,\n        realness_screened: false,\n      },\n    },\n  ],\n  errors: (subject) => {\n    if (field(subject, 'realness_screened') !== false) return []\n    const reading = readReward(field(subject, 'reward'))\n    if (reading.kind === 'absent' || reading.kind === 'cleared') return []\n    return [\n      `outcome.reward: ${reading.shown} with outcome.realness_screened: false — this producer declares ` +\n        'that NO authenticity screen ran on this reward, so nothing has established the success ' +\n        'is real, and an unscreened positive reward is exactly the signal the anti-Goodhart gate ' +\n        'exists to qualify. Screen the run and write the verdict (`rolloutRewardFields` from a ' +\n        '`RunRecord` carrying `outcome.realness`), or emit `reward: null` — an unqualified ' +\n        'verdict is a labeled gap, not a measured success.',\n    ]\n  },\n}\n\n/**\n * The registry. Total over `GateCheckId`, so an id with no check does not\n * compile, and `GATE_CHECK_IDS` stays the single enumeration everything\n * iterates.\n */\nexport const GATE_CHECKS: { readonly [K in GateCheckId]: GateCheck } = {\n  'reward-relationship': rewardRelationship,\n  'gated-evidence': gatedEvidence,\n  'undeclared-step-payload': undeclaredStepPayloadCheck,\n  'unscreened-reward': unscreenedReward,\n}\n\n/**\n * What ONE entry point does about ONE check.\n *\n * `repair` and `omit` both carry a mandatory sentence, which is the mechanism\n * that keeps a legitimate omission distinguishable from a forgotten one: you\n * cannot skip a check without writing down why, and the reasons are readable\n * side by side in `GATE_POLICIES`.\n */\nexport type GateCheckDisposition =\n  | { readonly kind: 'enforce' }\n  /** Resolved by TRANSFORMING the line instead of rejecting it; `by` names the function. */\n  | { readonly kind: 'repair'; readonly by: string }\n  /** Deliberately not applied here; `because` states the reason. */\n  | { readonly kind: 'omit'; readonly because: string }\n\nexport const enforced: GateCheckDisposition = { kind: 'enforce' }\nexport const repairedBy = (by: string): GateCheckDisposition => ({ kind: 'repair', by })\nexport const omittedBecause = (because: string): GateCheckDisposition => ({\n  kind: 'omit',\n  because,\n})\n\n/** Total over `GateCheckId`: a new check makes every policy literal a type error. */\nexport type GatePolicy = { readonly [K in GateCheckId]: GateCheckDisposition }\n\n/**\n * Every entry point that decides about the gate, and what it decides.\n *\n * Read this as the package's gate policy in one screen. The four entry points\n * are not interchangeable — a validator that rejects, a mint funnel that\n * repairs, a runtime backstop for untyped callers, and a release certifier over\n * emitted rows — and the dispositions say which is which.\n */\nexport const GATE_POLICIES = {\n  /**\n   * The schema validator. Rejects the reward relationship and NOTHING ELSE, on\n   * purpose: it runs on every line read off disk, and the other two conditions\n   * describe artifacts that already exist.\n   */\n  validateRolloutLine: {\n    // Stays a REJECTION rather than a transformation. A caller emitting\n    // `{reward: 0.95, realness_gated: true}` is a producer defect and has to\n    // fail loudly; laundering it into `reward: 0` here would hide the producer,\n    // which is the actual bug.\n    'reward-relationship': enforced,\n    'gated-evidence': omittedBecause(\n      'a ledger written before the relocation existed carries `metrics` on its gated lines; ' +\n        'rejecting those would make every already-published artifact unreadable. The condition ' +\n        'is REPAIRED at `assertMinted` (`gateGamedOutcome`) instead, which is the only door into ' +\n        'the training path, so the artifact stays readable and the leak still closes.',\n    ),\n    'undeclared-step-payload': omittedBecause(\n      'identical reasoning to `gated-evidence`, one field over: a foreign or pre-unification ' +\n        'ledger may carry producer-invented keys on its steps, and refusing to READ those files ' +\n        'buys nothing the relocation at `assertMinted` does not already buy. Promotion into ' +\n        'training is where it closes.',\n    ),\n    'unscreened-reward': omittedBecause(\n      'supervision-journal rows legitimately carry an unscreened positive reward (there is no ' +\n        '`RunRecord.outcome.realness` behind them) and must stay writable, readable and ' +\n        'reportable. Only PROMOTION into training is closed, at `assertMinted`.',\n    ),\n  },\n  /**\n   * The mint funnel — the single door every `MintedRolloutLine` passes. Validates\n   * first (so the reward relationship has already been rejected), then refuses\n   * what cannot be repaired, then repairs what can.\n   */\n  assertMinted: {\n    // Re-run after `assertRolloutLine`, which already rejected it. Cheap, and it\n    // makes this entry point independently complete rather than complete only\n    // because of what it happens to call first.\n    'reward-relationship': enforced,\n    'gated-evidence': repairedBy('gateGamedOutcome'),\n    'undeclared-step-payload': repairedBy('gateGamedOutcome'),\n    'unscreened-reward': enforced,\n  },\n  /**\n   * The runtime backstop, and the one entry point with no license to omit\n   * anything: it exists for callers the type system never saw (plain JavaScript\n   * handing an object literal to a published exporter), so a check it skips is a\n   * check that does not run at all for them. This is where the fourth leak was.\n   */\n  assertRewardGate: {\n    'reward-relationship': enforced,\n    'gated-evidence': enforced,\n    'undeclared-step-payload': enforced,\n    'unscreened-reward': enforced,\n  },\n  /**\n   * The release certifier. Same checks, measured over the rows a release is\n   * ABOUT TO WRITE rather than over one line's outcome — see `REPORT_MEASURES`\n   * in `release/gate-report.ts`, which is the second total map this policy\n   * drives.\n   */\n  assertGateReport: {\n    'reward-relationship': enforced,\n    'gated-evidence': enforced,\n    'undeclared-step-payload': enforced,\n    'unscreened-reward': enforced,\n  },\n} as const satisfies Record<string, GatePolicy>\n\n/** Every entry point that declares a gate policy. */\nexport type GateEntryPoint = keyof typeof GATE_POLICIES\n\n/**\n * Run the checks one entry point enforces. The ONLY way an entry point should\n * obtain gate errors — hand-composing two of the three is the bug this module\n * exists to remove.\n */\nexport function gateErrors(subject: GateSubject, policy: GatePolicy): string[] {\n  const errors: string[] = []\n  for (const id of GATE_CHECK_IDS) {\n    if (policy[id].kind !== 'enforce') continue\n    errors.push(...GATE_CHECKS[id].errors(subject))\n  }\n  return errors\n}\n","/**\n * `tangle.rollout.v1` — THE canonical rollout serialization, owned by\n * agent-eval. One JSONL line per agent invocation (a solo eval run, a\n * supervisor episode, a worker session, a proposer shot, a judge call, an\n * analyst pass), labeled with its task/split coordinates and a single\n * scalar reward, carrying the FULL message transcript inline.\n *\n * This schema is the reconciliation of two prior producers:\n *   - agent-eval's RunRecord-joined rollout rows (PR #410): identity,\n *     provenance hashes, the realness gate travelling into the reward,\n *     trace-derived steps.\n *   - the bench rollout-ledger (agent-runtime PR #591): the wire shape —\n *     role, task.split/rep, parent_rollout_id, policy provenance, capture\n *     provenance, inline canonical chat-with-tools messages.\n * Where the two conflicted, RunRecord-derived semantics won; the wire\n * field names follow the ledger (snake_case). See `docs/rollout.md` for\n * the field-by-field decision table.\n *\n * Messages are inlined — never referenced — because every harness store a\n * rollout can be recovered from is mutable or garbage-collected. A line\n * must stay a complete training/eval example on its own.\n *\n * `outcome.reward` is THE single scalar (null = no verdict exists — a\n * labeled gap, never 0). `outcome.realness_gated` is the anti-Goodhart\n * flag: a gated line must never export as a positive training example.\n *\n * That last sentence is enforced here, by `validateRolloutLine`, not merely\n * documented. Validating `reward` and `realness_gated` independently — each a\n * well-typed field, their COMBINATION unchecked — is what let a line claiming\n * `{reward: 0.95, realness_gated: true}` validate clean and walk into every\n * training export. The relationship between the two IS the invariant, so it is\n * checked where every other structural claim about a line is checked.\n *\n * The invariant is about the OUTCOME, not about one field of it. Zeroing\n * `reward` while `outcome.metrics` still carried the per-layer scores that\n * reward was computed from exported the gamed signal anyway, in the dict the\n * verifiers format reads as its per-rubric scores. So `gateGamedOutcome`\n * transforms the whole outcome once, at `assertMinted` — the funnel every\n * minted line passes — and the reward-bearing components are relocated to\n * `provenance.gated_evidence`, which no exporter projects.\n *\n * WHICH checks each door applies is not decided in this file. `./gate-checks`\n * owns the canonical list and the total per-entry-point policy; the three doors\n * below (`validateRolloutLine`, `assertRewardGate`, `assertMinted`) each call\n * `gateErrors` with their declared policy, so a check added to that list applies\n * here without anyone editing this file, and a check deliberately skipped has to\n * name itself there.\n */\n\nimport { declaredSteps, GATE_POLICIES, gatedEvidenceOf, gateErrors } from './gate-checks'\n\nexport const ROLLOUT_SCHEMA = 'tangle.rollout.v1'\n\n/** `agent` = a solo evaluation run (no multi-agent topology). */\nexport type RolloutRole = 'agent' | 'supervisor' | 'worker' | 'proposer' | 'judge' | 'analyst'\nexport const ROLLOUT_ROLES: readonly RolloutRole[] = [\n  'agent',\n  'supervisor',\n  'worker',\n  'proposer',\n  'judge',\n  'analyst',\n]\n\n/** Split vocabulary follows `RunRecord.splitTag`, extended with `canary`. */\nexport type RolloutSplit = 'search' | 'dev' | 'holdout' | 'canary'\nexport const ROLLOUT_SPLITS: readonly RolloutSplit[] = ['search', 'dev', 'holdout', 'canary']\n/** Splits that may ship in training exports. Everything else is fail-closed excluded. */\nexport const TRAINABLE_SPLITS: readonly RolloutSplit[] = ['search']\n\nexport function isTrainableSplit(split: RolloutSplit): boolean {\n  return TRAINABLE_SPLITS.includes(split)\n}\n\n/** 'mint' = joined live from RunRecord + trace by `mintRolloutRows`. */\nexport type RolloutCapture = 'mint' | 'settle-time' | 'backfill'\nexport const ROLLOUT_CAPTURES: readonly RolloutCapture[] = ['mint', 'settle-time', 'backfill']\n\n// ---------------------------------------------------------------------------\n// Canonical message format — OpenAI chat-with-tools, full fidelity.\n// ---------------------------------------------------------------------------\n\nexport type ChatRole = 'system' | 'user' | 'assistant' | 'tool'\nexport const CHAT_ROLES: readonly ChatRole[] = ['system', 'user', 'assistant', 'tool']\n\nexport interface ChatToolCall {\n  id: string\n  type: 'function'\n  function: {\n    name: string\n    /** JSON-encoded argument object, exactly as the model emitted it. */\n    arguments: string\n  }\n}\n\nexport interface ChatMessage {\n  role: ChatRole\n  content: string | null\n  /** Reasoning/thinking channel where the harness captured it (full fidelity). */\n  reasoning_content?: string\n  tool_calls?: ChatToolCall[]\n  /** Required on role:\"tool\" — the ChatToolCall this result answers. */\n  tool_call_id?: string\n  name?: string\n  /**\n   * Harbor ATIF `is_copied_context` (RFC 0001 rule 7): this turn was COPIED IN\n   * from another trajectory's context, not produced by the agent on this line.\n   * The RFC makes excluding it from SFT a MUST, and `toSftRows` does — training\n   * on it teaches the model to author text it never authored, and credits this\n   * run for another one's work. Absent = false (authored here).\n   */\n  is_copied_context?: boolean\n}\n\nexport interface ToolDef {\n  type: 'function'\n  function: {\n    name: string\n    description?: string\n    parameters?: Record<string, unknown>\n  }\n}\n\n/**\n * Compact trace-span projection (llm/tool step) carried alongside the\n * conversation when the line was minted from a trace. Optional: lines\n * recovered from harness stores have no span structure.\n */\nexport interface RolloutStep {\n  kind: string\n  name: string\n  /** llm: last-message summary · tool: stringified args. Scrubbed. */\n  input?: string\n  /** llm: output text · tool: stringified result. Scrubbed. */\n  output?: string\n  status?: 'ok' | 'error'\n  durationMs?: number\n  // The four fields below are lifted verbatim from Harbor ATIF's per-step\n  // `metrics` (see `interchange/harbor.ts`); the wire names stay snake_case to\n  // match that spec exactly, which is why they differ from `durationMs`.\n  // All optional and never back-filled: an absent field means \"not captured\",\n  // which is not the same claim as an empty array.\n  /**\n   * LLM inferences this span represents. 0 = deterministic dispatch with no\n   * model call — distinct from absent, which means the producer did not track it.\n   */\n  llm_call_count?: number\n  /** Exact prompt tokenization. Removes the ambiguity of re-tokenizing text at train time. */\n  prompt_token_ids?: number[]\n  /** Exact completion tokenization; aligns index-wise with `logprobs`. */\n  completion_token_ids?: number[]\n  /**\n   * Per-completion-token log probabilities under the sampling policy. Required\n   * for off-policy correction (importance weighting) when the rollout was\n   * generated by a policy other than the one being trained.\n   */\n  logprobs?: number[]\n}\n\n// ---------------------------------------------------------------------------\n// Ledger line sections.\n// ---------------------------------------------------------------------------\n\nexport interface RolloutTask {\n  /** Benchmark/suite id (e.g. \"swe-bench-verified\") or the experiment id. */\n  suite: string\n  instance_id: string\n  split: RolloutSplit\n  /** Sampling seed the campaign pinned; null = not recorded. */\n  seed: number | null\n  /** Replicate index (0-based). */\n  rep: number\n}\n\nexport interface RolloutPolicy {\n  /** Harness that drove the invocation (e.g. \"opencode\", \"claude\", \"pi-loops\"). */\n  harness: string | null\n  harness_version: string | null\n  model: string | null\n  provider: string | null\n  /** Commit of the agent profile / candidate under evaluation. */\n  profile_commit: string | null\n  /** sha256 of the effective prompt (post-steering), when recorded. */\n  prompt_hash?: string | null\n  /** sha256 of the effective run config, when recorded. */\n  config_hash?: string | null\n  /** Canonical agent-profile cell identity, when the run carries one. */\n  agent_profile_cell_id?: string | null\n  /** Sampling params (temperature, top_p, max_tokens…); null = not recorded. */\n  sampling: Record<string, unknown> | null\n}\n\nexport interface RolloutOutcome {\n  /**\n   * THE single scalar training signal — the official verdict.\n   * null = no verdict exists for this invocation (a labeled gap, never 0).\n   */\n  reward: number | null\n  /** Where the reward came from (judge id; \"/inherited\" = parent episode's). */\n  reward_source: string | null\n  /** Raw judge verdict record, verbatim. */\n  verdict: unknown\n  /** Everything that is NOT the scalar reward. */\n  metrics: Record<string, unknown>\n  is_completed: boolean\n  is_truncated: boolean\n  error: string | null\n  /**\n   * Anti-Goodhart flag from `RunRecord.outcome.realness.gated`: the run faked\n   * its success signal. `true` requires `reward` to be 0 or null — the\n   * validator rejects the line otherwise — and the line never qualifies for\n   * SFT. Required on the wire: a line that does not state the flag does not\n   * validate, so no producer can dodge the gate by omitting it.\n   *\n   * `true` ALSO requires `metrics` to be empty and `verdict` to be null: the\n   * numbers the reward was computed from are relocated to\n   * `provenance.gated_evidence` by `gateGamedOutcome`. See that function for\n   * why zeroing the scalar alone was not enough.\n   */\n  realness_gated: boolean\n  /**\n   * Whether an authenticity SCREEN ever RAN on this reward — a different claim\n   * from `realness_gated`, which is the screen's VERDICT.\n   *\n   * `realness_gated: false` reads as \"we looked and nothing fired\". A producer\n   * with no screen at all was emitting exactly that, so a never-screened reward\n   * was indistinguishable on the wire from a screened-clean one, and the whole\n   * anti-Goodhart apparatus silently treated the first as the second. The two\n   * claims are now separable:\n   *\n   *   - `true`   — a screen ran; `realness_gated` is its verdict.\n   *   - `false`  — the producer declares it HAS no screen (`unscreenedRewardFields`).\n   *                `assertMinted` REFUSES such a line when its reward is above\n   *                zero: an unscreened positive reward is precisely the signal\n   *                the gate exists to qualify, and nothing has qualified it.\n   *   - absent   — not stated. Pre-unification ledgers land here, as does a\n   *                `RunRecord` carrying no `outcome.realness` at all. Absent is\n   *                read as \"unknown\", never as `false` (which would refuse most\n   *                of the existing corpus) and never as `true`.\n   */\n  realness_screened?: boolean\n}\n\nexport interface RolloutCostBlock {\n  usd: number | null\n  tokens_in: number | null\n  tokens_out: number | null\n  tokens_reasoning: number | null\n  cache_read: number | null\n  cache_write: number | null\n  wall_s: number | null\n  /**\n   * Total LLM inferences across the invocation (ATIF `llm_call_count`,\n   * aggregated). Optional and additive: absent = not tracked, never 0.\n   */\n  llm_call_count?: number | null\n}\n\nexport interface RolloutArtifacts {\n  patch_path: string | null\n  run_dir: string | null\n  /** Source-of-truth transcript pointer (session id / jsonl path) for audit. */\n  transcript_ref: string | null\n}\n\n/**\n * The reward-bearing half of a GATED line's outcome, moved off `outcome` and\n * parked here verbatim. Diagnostics, never training input — see\n * `gateGamedOutcome`.\n */\nexport interface GatedEvidence {\n  /** `outcome.metrics` exactly as the producer measured it. */\n  metrics?: Record<string, unknown>\n  /** `outcome.verdict` verbatim — the judge record that claimed the success. */\n  verdict?: unknown\n  /**\n   * The per-step fields `tangle.rollout.v1` does not declare, parked here when\n   * the gate projected `steps[]` down to the schema's own key set.\n   *\n   * A per-step reward is training signal exactly like the scalar, and `steps`\n   * rides through `toRewardRows` verbatim — so a gated line was shipping its\n   * step-level credit assignment at full value beside a `reward` of 0.\n   */\n  steps?: unknown\n}\n\nexport interface RolloutProvenance {\n  captured_at: string\n  capture: RolloutCapture\n  /**\n   * Why this line is incomplete. Required when `messages` is empty (the\n   * transcript could not be recovered); also set by interchange importers to\n   * name a MISSING LABEL — an imported trajectory carries no verdict, so\n   * `outcome.reward` is null and this says why.\n   */\n  gap?: string\n  /**\n   * Present only on a realness-gated line: the outcome fields the gate\n   * relocated, kept so an auditor can still see WHY the run was gated and what\n   * it claimed. Deliberately OUTSIDE `outcome`, because every training exporter\n   * reads `outcome` and none reads `provenance`.\n   */\n  gated_evidence?: GatedEvidence\n}\n\nexport interface RolloutLine {\n  schema: typeof ROLLOUT_SCHEMA\n  rollout_id: string\n  /** Spawning invocation within the same episode (worker → supervisor). */\n  parent_rollout_id: string | null\n  run_id: string\n  /** Logical experiment grouping from `RunRecord.experimentId`; null = not recorded. */\n  experiment_id: string | null\n  /** Stable candidate identity from `RunRecord.candidateId`; null = not recorded. */\n  candidate_id: string | null\n  /** Improvement-loop generation (-1 = baseline); null = not an improvement loop. */\n  generation: number | null\n  /** Improvement-loop candidate index (-1 = baseline); null = not an improvement loop. */\n  candidate_index: number | null\n  role: RolloutRole\n  task: RolloutTask\n  policy: RolloutPolicy\n  /** Full transcript, inline. [] = gap line (see provenance.gap). */\n  messages: ChatMessage[]\n  tool_defs: ToolDef[]\n  /** Trace-span projections, when minted from a trace. */\n  steps?: RolloutStep[]\n  outcome: RolloutOutcome\n  cost: RolloutCostBlock\n  artifacts: RolloutArtifacts\n  provenance: RolloutProvenance\n}\n\n// ---------------------------------------------------------------------------\n// Validation — pure TS, no runtime schema dependency, mirroring the\n// run-record validator's fail-loud discipline. Returns [] when the value\n// is a valid RolloutLine; otherwise one dotted-path error per defect.\n// ---------------------------------------------------------------------------\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nconst isNumberOrNull = (v: unknown): boolean => v === null || typeof v === 'number'\nconst isNumberArray = (v: unknown): boolean =>\n  Array.isArray(v) && v.every((n) => typeof n === 'number' && Number.isFinite(n))\nconst isIntegerArray = (v: unknown): boolean =>\n  Array.isArray(v) && v.every((n) => Number.isInteger(n))\nconst isStringOrNull = (v: unknown): boolean => v === null || typeof v === 'string'\nconst isIntegerOrNull = (v: unknown): boolean => v === null || Number.isInteger(v)\n\nfunction validateChatMessage(value: unknown, path: string, errors: string[]): void {\n  if (!isRecord(value)) {\n    errors.push(`${path}: not an object`)\n    return\n  }\n  if (!CHAT_ROLES.includes(value.role as ChatRole))\n    errors.push(`${path}.role: invalid role ${String(value.role)}`)\n  if (!isStringOrNull(value.content)) errors.push(`${path}.content: must be string|null`)\n  if (value.reasoning_content !== undefined && typeof value.reasoning_content !== 'string') {\n    errors.push(`${path}.reasoning_content: must be string when present`)\n  }\n  if (value.tool_call_id !== undefined && typeof value.tool_call_id !== 'string') {\n    errors.push(`${path}.tool_call_id: must be string when present`)\n  }\n  if (value.role === 'tool' && typeof value.tool_call_id !== 'string') {\n    errors.push(`${path}.tool_call_id: required on role:\"tool\"`)\n  }\n  if (value.is_copied_context !== undefined && typeof value.is_copied_context !== 'boolean') {\n    errors.push(`${path}.is_copied_context: must be boolean when present`)\n  }\n  if (value.tool_calls !== undefined) {\n    if (!Array.isArray(value.tool_calls)) {\n      errors.push(`${path}.tool_calls: must be an array when present`)\n    } else {\n      value.tool_calls.forEach((call, i) => {\n        if (!isRecord(call) || typeof call.id !== 'string' || call.type !== 'function') {\n          errors.push(`${path}.tool_calls[${i}]: must be {id, type:\"function\", function}`)\n          return\n        }\n        const fn = call.function\n        if (!isRecord(fn) || typeof fn.name !== 'string' || typeof fn.arguments !== 'string') {\n          errors.push(\n            `${path}.tool_calls[${i}].function: must be {name: string, arguments: string}`,\n          )\n        }\n      })\n    }\n  }\n}\n\nfunction validateSection(\n  value: unknown,\n  path: string,\n  fields: Array<[name: string, check: (v: unknown) => boolean, expect: string]>,\n  errors: string[],\n): void {\n  if (!isRecord(value)) {\n    errors.push(`${path}: not an object`)\n    return\n  }\n  for (const [name, check, expect] of fields) {\n    if (!check(value[name])) errors.push(`${path}.${name}: expected ${expect}`)\n  }\n}\n\n/**\n * THE anti-Goodhart gate, applied to the WHOLE outcome as a TRANSFORMATION.\n *\n * Two prior rounds enforced the gate as a CHECK ON ONE FIELD at N call sites,\n * and each round the next reward-bearing field leaked. The one that shipped:\n * `mintRolloutRows` bulk-copied `RunRecord.outcome.raw` into `outcome.metrics`\n * with no gate, so a gated run exported `reward: 0` (correct) while the\n * deterministic per-layer scores that reward was COMPUTED FROM — the\n * `layer.*` keys `rl/verifiable-reward.ts` calls the RL training signal —\n * shipped at 1.0, in the top-level `metrics` dict of the Prime Intellect\n * verifiers format, which IS that format's per-rubric score dict. `verdict`\n * leaks the same way into `toRftItem`'s `reference.verdict`, where a grader\n * author reads `resolved: true` off a run that faked it.\n *\n * So the rule is no longer \"zero the field we remembered\". It is: if the gate\n * fired, the outcome that leaves here carries NOTHING positive that was derived\n * from the reward, whichever field a present or future exporter decides to\n * read. `reward` is already forced to 0 upstream (`trainingReward`) and\n * REJECTED here if it is not; `metrics` and `verdict` are relocated.\n *\n * WHERE they go, and why relocation rather than deletion: zeroing destroys the\n * audit trail that shows why the run was gated and what it claimed, which is\n * the row an auditor most wants and the labeled example a gaming DETECTOR\n * trains on. `provenance.gated_evidence` keeps every byte, at a path no\n * training exporter reads — all four release configs and every `rl/exporters`\n * shape project from `outcome`, `messages`, `cost` and `task`; none projects\n * `provenance`. Auditability preserved, training signal removed, and a future\n * exporter that reads a field nobody thought of is safe by construction because\n * the field is empty rather than because the exporter remembered to check.\n *\n * Idempotent: a second application finds nothing left to move and returns the\n * line unchanged, so re-minting a line read back off a ledger cannot clobber\n * the evidence it already carries.\n */\nexport function gateGamedOutcome(line: RolloutLine): RolloutLine {\n  if (line.outcome.realness_gated !== true) return line\n  const moved = gatedEvidenceOf(line)\n  if (moved === undefined) return line\n  const kept = line.provenance.gated_evidence\n  const evidence: GatedEvidence = {}\n  const metrics = { ...kept?.metrics, ...moved.metrics }\n  if (Object.keys(metrics).length > 0) evidence.metrics = metrics\n  if (moved.verdict !== undefined) evidence.verdict = moved.verdict\n  else if (kept?.verdict !== undefined) evidence.verdict = kept.verdict\n  if (moved.steps !== undefined) evidence.steps = moved.steps\n  else if (kept?.steps !== undefined) evidence.steps = kept.steps\n  // The trajectory itself STAYS: a gamed transcript is the labeled example a\n  // gaming detector trains on, and the exporters are meant to ship it. Only the\n  // fields the wire format does not declare come off, which is the same\n  // partition `gatedEvidenceOf` applies to `metrics` — the producer's own,\n  // never a key-name guess.\n  const steps = moved.steps === undefined ? line.steps : declaredSteps(line.steps)\n  return {\n    ...line,\n    ...(steps === undefined ? {} : { steps }),\n    outcome: { ...line.outcome, metrics: {}, verdict: null },\n    provenance: { ...line.provenance, gated_evidence: evidence },\n  }\n}\n\n/**\n * EVERY gate check, for the export path — the runtime backstop.\n *\n * `validateRolloutLine` checks the reward relationship too, but an exporter\n * cannot afford to re-validate every field of every line, and more importantly\n * it is not the exporter's job to re-check the schema — it is its job never to\n * emit a signal it was told is fabricated. Thrown rather than filtered: an\n * exporter silently dropping a poisoned line would hide the producer that made\n * it, and the producer is the actual defect.\n *\n * This is the third layer, and it exists for exactly one caller: JavaScript.\n * The brand stops TypeScript callers at compile time and `assertMinted` fixes\n * data arriving from disk, but neither is present for a plain JS consumer of\n * the published package handing an object literal to `toRewardRows`. Which is\n * exactly why this entry point's policy omits nothing: a check it skips is a\n * check that never runs for those callers at all. It composed two of the three\n * for one round, and a never-screened positive reward walked through all four\n * waist exporters at full value. `GATE_POLICIES.assertRewardGate` is now the\n * only place that list is written down.\n */\nexport function assertRewardGate(line: RolloutLine, context: string): void {\n  const errors = gateErrors(line, GATE_POLICIES.assertRewardGate)\n  if (errors.length > 0) {\n    throw new Error(`${context}: rollout ${line.rollout_id} — ${errors[0]}`)\n  }\n}\n\nexport function validateRolloutLine(value: unknown): string[] {\n  const errors: string[] = []\n  if (!isRecord(value)) return ['line: not an object']\n\n  if (value.schema !== ROLLOUT_SCHEMA) errors.push(`schema: expected \"${ROLLOUT_SCHEMA}\"`)\n  if (typeof value.rollout_id !== 'string' || value.rollout_id.length === 0)\n    errors.push('rollout_id: expected non-empty string')\n  if (!isStringOrNull(value.parent_rollout_id))\n    errors.push('parent_rollout_id: expected string|null')\n  if (typeof value.run_id !== 'string' || value.run_id.length === 0)\n    errors.push('run_id: expected non-empty string')\n  if (!isStringOrNull(value.experiment_id)) errors.push('experiment_id: expected string|null')\n  if (!isStringOrNull(value.candidate_id)) errors.push('candidate_id: expected string|null')\n  if (!isIntegerOrNull(value.generation)) errors.push('generation: expected integer|null')\n  if (!isIntegerOrNull(value.candidate_index)) errors.push('candidate_index: expected integer|null')\n  if (!ROLLOUT_ROLES.includes(value.role as RolloutRole))\n    errors.push(`role: invalid role ${String(value.role)}`)\n\n  validateSection(\n    value.task,\n    'task',\n    [\n      ['suite', (v) => typeof v === 'string' && v.length > 0, 'non-empty string'],\n      ['instance_id', (v) => typeof v === 'string' && v.length > 0, 'non-empty string'],\n      [\n        'split',\n        (v) => ROLLOUT_SPLITS.includes(v as RolloutSplit),\n        `one of ${ROLLOUT_SPLITS.join('|')}`,\n      ],\n      ['seed', isNumberOrNull, 'number|null'],\n      ['rep', (v) => Number.isInteger(v), 'integer'],\n    ],\n    errors,\n  )\n\n  validateSection(\n    value.policy,\n    'policy',\n    [\n      ['harness', isStringOrNull, 'string|null'],\n      ['harness_version', isStringOrNull, 'string|null'],\n      ['model', isStringOrNull, 'string|null'],\n      ['provider', isStringOrNull, 'string|null'],\n      ['profile_commit', isStringOrNull, 'string|null'],\n      ['sampling', (v) => v === null || isRecord(v), 'object|null'],\n    ],\n    errors,\n  )\n  if (isRecord(value.policy)) {\n    for (const key of ['prompt_hash', 'config_hash', 'agent_profile_cell_id'] as const) {\n      if (value.policy[key] !== undefined && !isStringOrNull(value.policy[key])) {\n        errors.push(`policy.${key}: expected string|null when present`)\n      }\n    }\n  }\n\n  if (!Array.isArray(value.messages)) {\n    errors.push('messages: expected array')\n  } else {\n    for (const [i, m] of value.messages.entries()) validateChatMessage(m, `messages[${i}]`, errors)\n  }\n\n  if (!Array.isArray(value.tool_defs)) {\n    errors.push('tool_defs: expected array')\n  } else {\n    value.tool_defs.forEach((d, i) => {\n      if (\n        !isRecord(d) ||\n        d.type !== 'function' ||\n        !isRecord(d.function) ||\n        typeof d.function.name !== 'string'\n      ) {\n        errors.push(`tool_defs[${i}]: must be {type:\"function\", function:{name}}`)\n      }\n    })\n  }\n\n  if (value.steps !== undefined) {\n    if (!Array.isArray(value.steps)) {\n      errors.push('steps: expected array when present')\n    } else {\n      value.steps.forEach((s, i) => {\n        if (!isRecord(s) || typeof s.kind !== 'string' || typeof s.name !== 'string') {\n          errors.push(`steps[${i}]: must be {kind: string, name: string, …}`)\n          return\n        }\n        // ATIF-lifted optionals: absent on every line written before they\n        // existed, so they are checked only when present — old ledgers stay valid.\n        if (s.llm_call_count !== undefined && !Number.isInteger(s.llm_call_count)) {\n          errors.push(`steps[${i}].llm_call_count: expected integer when present`)\n        }\n        for (const key of ['prompt_token_ids', 'completion_token_ids'] as const) {\n          if (s[key] !== undefined && !isIntegerArray(s[key])) {\n            errors.push(`steps[${i}].${key}: expected integer[] when present`)\n          }\n        }\n        if (s.logprobs !== undefined && !isNumberArray(s.logprobs)) {\n          errors.push(`steps[${i}].logprobs: expected number[] when present`)\n        }\n      })\n    }\n  }\n\n  validateSection(\n    value.outcome,\n    'outcome',\n    [\n      ['reward', isNumberOrNull, 'number|null'],\n      ['reward_source', isStringOrNull, 'string|null'],\n      ['metrics', isRecord, 'object'],\n      ['is_completed', (v) => typeof v === 'boolean', 'boolean'],\n      ['is_truncated', (v) => typeof v === 'boolean', 'boolean'],\n      ['error', isStringOrNull, 'string|null'],\n    ],\n    errors,\n  )\n  if (isRecord(value.outcome)) {\n    if (!('verdict' in value.outcome)) errors.push('outcome.verdict: field required (may be null)')\n    if (typeof value.outcome.realness_gated !== 'boolean') {\n      errors.push('outcome.realness_gated: expected boolean')\n    }\n    if (\n      value.outcome.realness_screened !== undefined &&\n      typeof value.outcome.realness_screened !== 'boolean'\n    ) {\n      errors.push('outcome.realness_screened: expected boolean when present')\n    }\n    errors.push(\n      ...gateErrors(\n        { outcome: value.outcome, steps: value.steps },\n        GATE_POLICIES.validateRolloutLine,\n      ),\n    )\n  }\n\n  validateSection(\n    value.cost,\n    'cost',\n    [\n      ['usd', isNumberOrNull, 'number|null'],\n      ['tokens_in', isNumberOrNull, 'number|null'],\n      ['tokens_out', isNumberOrNull, 'number|null'],\n      ['tokens_reasoning', isNumberOrNull, 'number|null'],\n      ['cache_read', isNumberOrNull, 'number|null'],\n      ['cache_write', isNumberOrNull, 'number|null'],\n      ['wall_s', isNumberOrNull, 'number|null'],\n    ],\n    errors,\n  )\n  if (\n    isRecord(value.cost) &&\n    value.cost.llm_call_count !== undefined &&\n    !(value.cost.llm_call_count === null || Number.isInteger(value.cost.llm_call_count))\n  ) {\n    errors.push('cost.llm_call_count: expected integer|null when present')\n  }\n\n  validateSection(\n    value.artifacts,\n    'artifacts',\n    [\n      ['patch_path', isStringOrNull, 'string|null'],\n      ['run_dir', isStringOrNull, 'string|null'],\n      ['transcript_ref', isStringOrNull, 'string|null'],\n    ],\n    errors,\n  )\n\n  validateSection(\n    value.provenance,\n    'provenance',\n    [\n      [\n        'captured_at',\n        (v) => typeof v === 'string' && !Number.isNaN(Date.parse(v)),\n        'ISO-8601 timestamp',\n      ],\n      [\n        'capture',\n        (v) => ROLLOUT_CAPTURES.includes(v as RolloutCapture),\n        `one of ${ROLLOUT_CAPTURES.join('|')}`,\n      ],\n    ],\n    errors,\n  )\n  if (isRecord(value.provenance)) {\n    if (value.provenance.gap !== undefined && typeof value.provenance.gap !== 'string') {\n      errors.push('provenance.gap: must be string when present')\n    }\n    const evidence = value.provenance.gated_evidence\n    if (evidence !== undefined) {\n      if (!isRecord(evidence)) {\n        errors.push('provenance.gated_evidence: must be an object when present')\n      } else if (evidence.metrics !== undefined && !isRecord(evidence.metrics)) {\n        errors.push('provenance.gated_evidence.metrics: must be an object when present')\n      }\n    }\n  }\n\n  // A gap line must say WHY it is a gap; a full line must not carry a gap note.\n  if (Array.isArray(value.messages) && isRecord(value.provenance)) {\n    if (value.messages.length === 0 && typeof value.provenance.gap !== 'string') {\n      errors.push('provenance.gap: required when messages is empty')\n    }\n  }\n\n  return errors\n}\n\nexport function assertRolloutLine(\n  value: unknown,\n  context = 'rollout line',\n): asserts value is RolloutLine {\n  const errors = validateRolloutLine(value)\n  if (errors.length > 0) {\n    throw new Error(`invalid ${context}:\\n  ${errors.join('\\n  ')}`)\n  }\n}\n\nexport function isRolloutLine(value: unknown): value is RolloutLine {\n  return validateRolloutLine(value).length === 0\n}\n\n// ---------------------------------------------------------------------------\n// The minted brand — the COMPILE-TIME half of the anti-Goodhart gate.\n// ---------------------------------------------------------------------------\n\n/**\n * Phantom property. `declare const` means it exists only in the type system:\n * nothing is written at runtime, so a branded line still serializes to exactly\n * the same JSON as a plain one.\n */\ndeclare const MINTED_ROLLOUT: unique symbol\n\n/**\n * A minted outcome states the gate verdict — it is not allowed to stay silent —\n * and, when that verdict is `true`, carries nothing else the reward was derived\n * from (`gateGamedOutcome` has run).\n */\nexport interface MintedRolloutOutcome extends RolloutOutcome {\n  realness_gated: boolean\n}\n\n/**\n * A `RolloutLine` whose reward has been checked against the anti-Goodhart\n * invariant. The type every training-data exporter takes.\n *\n * Why a brand and not just the interface: `RolloutLine` is structural, so any\n * hand-built object literal of the right shape IS one — which is how a line\n * declaring `{reward: 0.95, realness_gated: true}` reached the exporters\n * despite them \"only accepting a minted line\". The phantom symbol makes the\n * type nominal: it cannot be produced by writing an object literal, only by\n * `mintRolloutRows` (which applies the gate), `readRolloutLedger` (which\n * validates every line off disk), or an explicit, greppable `assertMinted`.\n *\n * Belt and braces on purpose. The brand closes first-party call sites at\n * COMPILE time; `validateRolloutLine` closes data arriving at RUNTIME (ledger\n * files, foreign imports, JSON from another process) where types are absent.\n * Neither alone is enough.\n *\n * Assignable to `RolloutLine` in one direction only: readers, analysis, and\n * the ledger writer keep taking the plain type.\n */\nexport type MintedRolloutLine = Omit<RolloutLine, 'outcome'> & {\n  readonly [MINTED_ROLLOUT]: true\n  outcome: MintedRolloutOutcome\n}\n\n/**\n * Promote a line to the type the training exporters accept, applying the\n * anti-Goodhart gate to the WHOLE outcome on the way through. THE escape hatch\n * — grep `assertMinted` to enumerate every place a line enters the training\n * path without coming from mint or a ledger.\n *\n * The gate runs HERE, once, rather than at each producer, because this is the\n * single funnel every minted line passes: `mintRolloutRows` calls it,\n * `readRolloutLedger` calls it per line off disk, `scrubLines` calls it on the\n * way out of a release, and a hand-built line has no other door. One\n * transformation at the funnel means an already-published ledger holding a\n * gated line with populated `metrics` is RE-GATED when it is read, instead of\n * being rejected (which would make every such artifact unreadable) or trusted\n * (which is the leak). Three steps, in this order:\n *\n *   1. VALIDATE the schema.\n *   2. REFUSE every check `GATE_POLICIES.assertMinted` marks `enforce` — today\n *      the reward relationship (which stays a REJECTION: a caller claiming\n *      `{reward: 0.95, realness_gated: true}` is a producer defect and must fail\n *      loudly, since laundering it into `reward: 0` here would hide the\n *      producer) and a positive reward the producer declared it never screened.\n *   3. TRANSFORM the one check that policy marks `repair` — relocate the\n *      reward's components off `outcome` (`gateGamedOutcome`), so no exporter\n *      can leak them whichever field it reads.\n *\n * Step 2 enumerates nothing by hand: a check added to `GATE_CHECKS` is enforced\n * here the moment its disposition in that policy says so.\n *\n * Also normalizes the optional wire flag to an explicit boolean.\n * `realness_gated` is absent on pre-unification ledgers and absent means \"not\n * flagged\" per the schema, so filling it in states a claim the line was already\n * making, and makes the flag readable on every published row instead of most of\n * them. `realness_screened` is NOT filled in: absent means \"unknown\", and\n * inventing either value there would be the same overclaim this round removed.\n */\nexport function assertMinted(value: unknown, context = 'rollout line'): MintedRolloutLine {\n  assertRolloutLine(value, context)\n  const refused = gateErrors(value, GATE_POLICIES.assertMinted)\n  if (refused.length > 0) {\n    throw new Error(`invalid ${context}: rollout ${value.rollout_id} — ${refused[0]}`)\n  }\n  const gated = gateGamedOutcome(value)\n  if (gated.outcome.realness_gated === undefined) {\n    return {\n      ...gated,\n      outcome: { ...gated.outcome, realness_gated: false },\n    } as MintedRolloutLine\n  }\n  return gated as MintedRolloutLine\n}\n\n/** `assertMinted` over a batch, naming the offending index in the error. */\nexport function assertMintedLines(\n  values: readonly unknown[],\n  context = 'rollout line',\n): MintedRolloutLine[] {\n  return values.map((value, i) => assertMinted(value, `${context} [${i}]`))\n}\n"],"mappings":";;;;;;;;AA2CA,MAAa,iBAAiB;CAC5B;CACA;CACA;CACA;AACF;;AAqCA,MAAM,SAAS,SAAsB,SAClC,QAAQ,UAAgD;AAkD3D,SAAgB,WAAW,OAA+B;CACxD,IAAI,UAAU,QAAQ,UAAU,KAAA,GAAW,OAAO,EAAE,MAAM,SAAS;CACnE,IAAI,OAAO,UAAU,UAAU;EAC7B,IAAI,OAAO,MAAM,KAAK,GAAG,OAAO;GAAE,MAAM;GAAc,OAAO;EAAM;EACnE,OAAO,QAAQ,IAAI;GAAE,MAAM;GAAY,OAAO,OAAO,KAAK;EAAE,IAAI,EAAE,MAAM,UAAU;CACpF;CACA,OAAO;EACL,MAAM;EACN,OAAO,GAAG,KAAK,UAAU,KAAK,KAAK,OAAO,KAAK,EAAE,IAAI,OAAO,MAAM;CACpE;AACF;AAEA,MAAM,iBAAiB,UACrB,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;;;;;;;;;;;;;AAcrE,SAAgB,mBAAmB,OAAyB;CAC1D,IAAI,UAAU,KAAA,KAAa,UAAU,MAAM,OAAO;CAClD,IAAI,cAAc,KAAK,GAAG,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS;CAC7D,OAAO;AACT;;;;;;;;;;;AAYA,MAAM,qBAA4E;CAChF,MAAM;CACN,MAAM;CACN,OAAO;CACP,QAAQ;CACR,QAAQ;CACR,YAAY;CACZ,gBAAgB;CAChB,kBAAkB;CAClB,sBAAsB;CACtB,UAAU;AACZ;;;;;;;AAQA,SAAgB,sBAAsB,OAAyB;CAC7D,IAAI,UAAU,KAAA,KAAa,UAAU,MAAM,OAAO,KAAA;CAElD,IAAI,CAAC,MAAM,QAAQ,KAAK,GAAG,OAAO;CAClC,MAAM,SAAyC,CAAC;CAChD,IAAI,QAAQ;CACZ,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,QAAiC,CAAC;EACxC,IAAI,cAAc,IAAI,GACpB,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAAI,GAAG;GAC/C,IAAI,OAAO,oBAAoB;GAC/B,MAAM,OAAO;GACb,QAAQ;EACV;OACK,IAAI,SAAS,KAAA,KAAa,SAAS,MAAM;GAE9C,OAAO,KAAK,EAAE,KAAK,CAA4B;GAC/C,QAAQ;GACR;EACF;EACA,OAAO,KAAK,KAAK;CACnB;CACA,OAAO,QAAQ,SAAS,KAAA;AAC1B;;AAGA,SAAgB,cAAc,OAA2C;CACvE,IAAI,CAAC,MAAM,QAAQ,KAAK,GAAG,OAAO,KAAA;CAClC,MAAM,OAAsB,CAAC;CAC7B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,CAAC,cAAc,IAAI,GAAG;EAC1B,MAAM,YAAqC,CAAC;EAC5C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAAI,GAC5C,IAAI,OAAO,oBAAoB,UAAU,OAAO;EAElD,KAAK,KAAK,SAAmC;CAC/C;CACA,OAAO;AACT;;;;;;;;;;;;;AAcA,MAAM,qBAAgC;CACpC,IAAI;CACJ,SAAS;CACT,WAAW,CACT,EAAE,SAAS;EAAE,QAAQ;EAAM,gBAAgB;CAAK,EAAE,GAGlD,EAAE,SAAS;EAAE,QAAQ;EAA6B,gBAAgB;CAAK,EAAE,CAC3E;CACA,SAAS,YAAY;EACnB,IAAI,MAAM,SAAS,gBAAgB,MAAM,MAAM,OAAO,CAAC;EACvD,MAAM,UAAU,WAAW,MAAM,SAAS,QAAQ,CAAC;EACnD,IAAI,QAAQ,SAAS,YAAY,QAAQ,SAAS,WAAW,OAAO,CAAC;EACrE,IAAI,QAAQ,SAAS,cACnB,OAAO,CACL,mBAAmB,QAAQ,MAAM,geAMnC;EAEF,OAAO,CACL,mBAAmB,QAAQ,MAAM,wWAKnC;CACF;AACF;;;;;;;;;;;;AAaA,SAAgB,gBAAgB,SAAiD;CAC/E,MAAM,WAA0B,CAAC;CAIjC,MAAM,UAAU,MAAM,SAAS,SAAS;CACxC,IAAI,mBAAmB,OAAO,GAAG,SAAS,UAAU;CACpD,MAAM,UAAU,MAAM,SAAS,SAAS;CACxC,IAAI,YAAY,QAAQ,YAAY,KAAA,GAAW,SAAS,UAAU;CAClE,MAAM,QAAQ,sBAAsB,QAAQ,KAAK;CACjD,IAAI,UAAU,KAAA,GAAW,SAAS,QAAQ;CAC1C,OAAO,SAAS,YAAY,KAAA,KAAa,SAAS,YAAY,KAAA,KAAa,UAAU,KAAA,IACjF,KAAA,IACA;AACN;;;;;;;;;;;;;AAcA,MAAM,gBAA2B;CAC/B,IAAI;CACJ,SAAS;CACT,WAAW,CACT,EAAE,SAAS;EAAE,QAAQ;EAAG,gBAAgB;EAAM,SAAS,EAAE,eAAe,EAAE;CAAE,EAAE,GAI9E,EAAE,SAAS;EAAE,QAAQ;EAAG,gBAAgB;EAAM,SAAS;CAAyB,EAAE,CACpF;CACA,SAAS,YAAY;EACnB,IAAI,MAAM,SAAS,gBAAgB,MAAM,MAAM,OAAO,CAAC;EACvD,MAAM,WAAW,gBAAgB,OAAO;EACxC,IAAI,aAAa,KAAA,GAAW,OAAO,CAAC;EAEpC,IAAI,SAAS,YAAY,KAAA,KAAa,SAAS,YAAY,KAAA,GAAW,OAAO,CAAC;EAE9E,OAAO,CACL,GAFW,SAAS,YAAY,KAAA,IAAY,oBAAoB,kBAExD,2aAKV;CACF;AACF;;;;;;;;;;;;;;;;;;AAmBA,MAAM,6BAAwC;CAC5C,IAAI;CACJ,SAAS;CACT,WAAW,CACT;EACE,SAAS;GAAE,QAAQ;GAAG,gBAAgB;EAAK;EAC3C,OAAO,CAAC;GAAE,MAAM;GAAQ,MAAM;GAAQ,QAAQ;EAAI,CAAC;CACrD,CACF;CACA,SAAS,YAAY;EACnB,IAAI,MAAM,SAAS,gBAAgB,MAAM,MAAM,OAAO,CAAC;EACvD,MAAM,QAAQ,sBAAsB,QAAQ,KAAK;EACjD,IAAI,UAAU,KAAA,GAAW,OAAO,CAAC;EACjC,OAAO,CACL,4BAA4B,oBAAoB,uBAAuB,KAAK,UAAU,KAAK,CAAC,CAAC,MAAM,GAAG,GAAG,EAAE,6bAM7G;CACF;AACF;;AAGA,MAAM,sBAAsB;;;;;;AA4C5B,MAAa,cAA0D;CACrE,uBAAuB;CACvB,kBAAkB;CAClB,2BAA2B;CAC3B,qBAAqB;EApCrB,IAAI;EACJ,SAAS;EACT,WAAW,CACT,EAAE,SAAS;GAAE,QAAQ;GAAG,gBAAgB;GAAO,mBAAmB;EAAM,EAAE,GAC1E,EACE,SAAS;GACP,QAAQ;GACR,gBAAgB;GAChB,mBAAmB;EACrB,EACF,CACF;EACA,SAAS,YAAY;GACnB,IAAI,MAAM,SAAS,mBAAmB,MAAM,OAAO,OAAO,CAAC;GAC3D,MAAM,UAAU,WAAW,MAAM,SAAS,QAAQ,CAAC;GACnD,IAAI,QAAQ,SAAS,YAAY,QAAQ,SAAS,WAAW,OAAO,CAAC;GACrE,OAAO,CACL,mBAAmB,QAAQ,MAAM,idAMnC;EACF;CAYqB;AACvB;AAiBA,MAAa,WAAiC,EAAE,MAAM,UAAU;AAChE,MAAa,cAAc,QAAsC;CAAE,MAAM;CAAU;AAAG;AACtF,MAAa,kBAAkB,aAA2C;CACxE,MAAM;CACN;AACF;;;;;;;;;AAaA,MAAa,gBAAgB;;;;;;CAM3B,qBAAqB;EAKnB,uBAAuB;EACvB,kBAAkB,eAChB,iVAIF;EACA,2BAA2B,eACzB,8RAIF;EACA,qBAAqB,eACnB,8OAGF;CACF;;;;;;CAMA,cAAc;EAIZ,uBAAuB;EACvB,kBAAkB,WAAW,kBAAkB;EAC/C,2BAA2B,WAAW,kBAAkB;EACxD,qBAAqB;CACvB;;;;;;;CAOA,kBAAkB;EAChB,uBAAuB;EACvB,kBAAkB;EAClB,2BAA2B;EAC3B,qBAAqB;CACvB;;;;;;;CAOA,kBAAkB;EAChB,uBAAuB;EACvB,kBAAkB;EAClB,2BAA2B;EAC3B,qBAAqB;CACvB;AACF;;;;;;AAUA,SAAgB,WAAW,SAAsB,QAA8B;CAC7E,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,MAAM,gBAAgB;EAC/B,IAAI,OAAO,GAAG,CAAC,SAAS,WAAW;EACnC,OAAO,KAAK,GAAG,YAAY,GAAG,CAAC,OAAO,OAAO,CAAC;CAChD;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC7fA,MAAa,iBAAiB;AAI9B,MAAa,gBAAwC;CACnD;CACA;CACA;CACA;CACA;CACA;AACF;AAIA,MAAa,iBAA0C;CAAC;CAAU;CAAO;CAAW;AAAQ;;AAE5F,MAAa,mBAA4C,CAAC,QAAQ;AAElE,SAAgB,iBAAiB,OAA8B;CAC7D,OAAO,iBAAiB,SAAS,KAAK;AACxC;AAIA,MAAa,mBAA8C;CAAC;CAAQ;CAAe;AAAU;AAO7F,MAAa,aAAkC;CAAC;CAAU;CAAQ;CAAa;AAAM;AAgQrF,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,MAAM,kBAAkB,MAAwB,MAAM,QAAQ,OAAO,MAAM;AAC3E,MAAM,iBAAiB,MACrB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;AAChF,MAAM,kBAAkB,MACtB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,UAAU,CAAC,CAAC;AACxD,MAAM,kBAAkB,MAAwB,MAAM,QAAQ,OAAO,MAAM;AAC3E,MAAM,mBAAmB,MAAwB,MAAM,QAAQ,OAAO,UAAU,CAAC;AAEjF,SAAS,oBAAoB,OAAgB,MAAc,QAAwB;CACjF,IAAI,CAAC,SAAS,KAAK,GAAG;EACpB,OAAO,KAAK,GAAG,KAAK,gBAAgB;EACpC;CACF;CACA,IAAI,CAAC,WAAW,SAAS,MAAM,IAAgB,GAC7C,OAAO,KAAK,GAAG,KAAK,sBAAsB,OAAO,MAAM,IAAI,GAAG;CAChE,IAAI,CAAC,eAAe,MAAM,OAAO,GAAG,OAAO,KAAK,GAAG,KAAK,8BAA8B;CACtF,IAAI,MAAM,sBAAsB,KAAA,KAAa,OAAO,MAAM,sBAAsB,UAC9E,OAAO,KAAK,GAAG,KAAK,gDAAgD;CAEtE,IAAI,MAAM,iBAAiB,KAAA,KAAa,OAAO,MAAM,iBAAiB,UACpE,OAAO,KAAK,GAAG,KAAK,2CAA2C;CAEjE,IAAI,MAAM,SAAS,UAAU,OAAO,MAAM,iBAAiB,UACzD,OAAO,KAAK,GAAG,KAAK,uCAAuC;CAE7D,IAAI,MAAM,sBAAsB,KAAA,KAAa,OAAO,MAAM,sBAAsB,WAC9E,OAAO,KAAK,GAAG,KAAK,iDAAiD;CAEvE,IAAI,MAAM,eAAe,KAAA,GACvB,IAAI,CAAC,MAAM,QAAQ,MAAM,UAAU,GACjC,OAAO,KAAK,GAAG,KAAK,2CAA2C;MAE/D,MAAM,WAAW,SAAS,MAAM,MAAM;EACpC,IAAI,CAAC,SAAS,IAAI,KAAK,OAAO,KAAK,OAAO,YAAY,KAAK,SAAS,YAAY;GAC9E,OAAO,KAAK,GAAG,KAAK,cAAc,EAAE,2CAA2C;GAC/E;EACF;EACA,MAAM,KAAK,KAAK;EAChB,IAAI,CAAC,SAAS,EAAE,KAAK,OAAO,GAAG,SAAS,YAAY,OAAO,GAAG,cAAc,UAC1E,OAAO,KACL,GAAG,KAAK,cAAc,EAAE,sDAC1B;CAEJ,CAAC;AAGP;AAEA,SAAS,gBACP,OACA,MACA,QACA,QACM;CACN,IAAI,CAAC,SAAS,KAAK,GAAG;EACpB,OAAO,KAAK,GAAG,KAAK,gBAAgB;EACpC;CACF;CACA,KAAK,MAAM,CAAC,MAAM,OAAO,WAAW,QAClC,IAAI,CAAC,MAAM,MAAM,KAAK,GAAG,OAAO,KAAK,GAAG,KAAK,GAAG,KAAK,aAAa,QAAQ;AAE9E;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoCA,SAAgB,iBAAiB,MAAgC;CAC/D,IAAI,KAAK,QAAQ,mBAAmB,MAAM,OAAO;CACjD,MAAM,QAAQ,gBAAgB,IAAI;CAClC,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,MAAM,OAAO,KAAK,WAAW;CAC7B,MAAM,WAA0B,CAAC;CACjC,MAAM,UAAU;EAAE,GAAG,MAAM;EAAS,GAAG,MAAM;CAAQ;CACrD,IAAI,OAAO,KAAK,OAAO,CAAC,CAAC,SAAS,GAAG,SAAS,UAAU;CACxD,IAAI,MAAM,YAAY,KAAA,GAAW,SAAS,UAAU,MAAM;MACrD,IAAI,MAAM,YAAY,KAAA,GAAW,SAAS,UAAU,KAAK;CAC9D,IAAI,MAAM,UAAU,KAAA,GAAW,SAAS,QAAQ,MAAM;MACjD,IAAI,MAAM,UAAU,KAAA,GAAW,SAAS,QAAQ,KAAK;CAM1D,MAAM,QAAQ,MAAM,UAAU,KAAA,IAAY,KAAK,QAAQ,cAAc,KAAK,KAAK;CAC/E,OAAO;EACL,GAAG;EACH,GAAI,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,MAAM;EACvC,SAAS;GAAE,GAAG,KAAK;GAAS,SAAS,CAAC;GAAG,SAAS;EAAK;EACvD,YAAY;GAAE,GAAG,KAAK;GAAY,gBAAgB;EAAS;CAC7D;AACF;;;;;;;;;;;;;;;;;;;;;AAsBA,SAAgB,iBAAiB,MAAmB,SAAuB;CACzE,MAAM,SAAS,WAAW,MAAM,cAAc,gBAAgB;CAC9D,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,MAAM,GAAG,QAAQ,YAAY,KAAK,WAAW,KAAK,OAAO,IAAI;AAE3E;AAEA,SAAgB,oBAAoB,OAA0B;CAC5D,MAAM,SAAmB,CAAC;CAC1B,IAAI,CAAC,SAAS,KAAK,GAAG,OAAO,CAAC,qBAAqB;CAEnD,IAAI,MAAM,WAAA,qBAA2B,OAAO,KAAK,qBAAqB,eAAe,EAAE;CACvF,IAAI,OAAO,MAAM,eAAe,YAAY,MAAM,WAAW,WAAW,GACtE,OAAO,KAAK,uCAAuC;CACrD,IAAI,CAAC,eAAe,MAAM,iBAAiB,GACzC,OAAO,KAAK,yCAAyC;CACvD,IAAI,OAAO,MAAM,WAAW,YAAY,MAAM,OAAO,WAAW,GAC9D,OAAO,KAAK,mCAAmC;CACjD,IAAI,CAAC,eAAe,MAAM,aAAa,GAAG,OAAO,KAAK,qCAAqC;CAC3F,IAAI,CAAC,eAAe,MAAM,YAAY,GAAG,OAAO,KAAK,oCAAoC;CACzF,IAAI,CAAC,gBAAgB,MAAM,UAAU,GAAG,OAAO,KAAK,mCAAmC;CACvF,IAAI,CAAC,gBAAgB,MAAM,eAAe,GAAG,OAAO,KAAK,wCAAwC;CACjG,IAAI,CAAC,cAAc,SAAS,MAAM,IAAmB,GACnD,OAAO,KAAK,sBAAsB,OAAO,MAAM,IAAI,GAAG;CAExD,gBACE,MAAM,MACN,QACA;EACE;GAAC;IAAU,MAAM,OAAO,MAAM,YAAY,EAAE,SAAS;GAAG;EAAkB;EAC1E;GAAC;IAAgB,MAAM,OAAO,MAAM,YAAY,EAAE,SAAS;GAAG;EAAkB;EAChF;GACE;IACC,MAAM,eAAe,SAAS,CAAiB;GAChD,UAAU,eAAe,KAAK,GAAG;EACnC;EACA;GAAC;GAAQ;GAAgB;EAAa;EACtC;GAAC;IAAQ,MAAM,OAAO,UAAU,CAAC;GAAG;EAAS;CAC/C,GACA,MACF;CAEA,gBACE,MAAM,QACN,UACA;EACE;GAAC;GAAW;GAAgB;EAAa;EACzC;GAAC;GAAmB;GAAgB;EAAa;EACjD;GAAC;GAAS;GAAgB;EAAa;EACvC;GAAC;GAAY;GAAgB;EAAa;EAC1C;GAAC;GAAkB;GAAgB;EAAa;EAChD;GAAC;IAAa,MAAM,MAAM,QAAQ,SAAS,CAAC;GAAG;EAAa;CAC9D,GACA,MACF;CACA,IAAI,SAAS,MAAM,MAAM,GAClB;OAAA,MAAM,OAAO;GAAC;GAAe;GAAe;EAAuB,GACtE,IAAI,MAAM,OAAO,SAAS,KAAA,KAAa,CAAC,eAAe,MAAM,OAAO,IAAI,GACtE,OAAO,KAAK,UAAU,IAAI,oCAAoC;CAAA;CAKpE,IAAI,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAC/B,OAAO,KAAK,0BAA0B;MAEtC,KAAK,MAAM,CAAC,GAAG,MAAM,MAAM,SAAS,QAAQ,GAAG,oBAAoB,GAAG,YAAY,EAAE,IAAI,MAAM;CAGhG,IAAI,CAAC,MAAM,QAAQ,MAAM,SAAS,GAChC,OAAO,KAAK,2BAA2B;MAEvC,MAAM,UAAU,SAAS,GAAG,MAAM;EAChC,IACE,CAAC,SAAS,CAAC,KACX,EAAE,SAAS,cACX,CAAC,SAAS,EAAE,QAAQ,KACpB,OAAO,EAAE,SAAS,SAAS,UAE3B,OAAO,KAAK,aAAa,EAAE,8CAA8C;CAE7E,CAAC;CAGH,IAAI,MAAM,UAAU,KAAA,GAClB,IAAI,CAAC,MAAM,QAAQ,MAAM,KAAK,GAC5B,OAAO,KAAK,oCAAoC;MAEhD,MAAM,MAAM,SAAS,GAAG,MAAM;EAC5B,IAAI,CAAC,SAAS,CAAC,KAAK,OAAO,EAAE,SAAS,YAAY,OAAO,EAAE,SAAS,UAAU;GAC5E,OAAO,KAAK,SAAS,EAAE,2CAA2C;GAClE;EACF;EAGA,IAAI,EAAE,mBAAmB,KAAA,KAAa,CAAC,OAAO,UAAU,EAAE,cAAc,GACtE,OAAO,KAAK,SAAS,EAAE,gDAAgD;EAEzE,KAAK,MAAM,OAAO,CAAC,oBAAoB,sBAAsB,GAC3D,IAAI,EAAE,SAAS,KAAA,KAAa,CAAC,eAAe,EAAE,IAAI,GAChD,OAAO,KAAK,SAAS,EAAE,IAAI,IAAI,kCAAkC;EAGrE,IAAI,EAAE,aAAa,KAAA,KAAa,CAAC,cAAc,EAAE,QAAQ,GACvD,OAAO,KAAK,SAAS,EAAE,2CAA2C;CAEtE,CAAC;CAIL,gBACE,MAAM,SACN,WACA;EACE;GAAC;GAAU;GAAgB;EAAa;EACxC;GAAC;GAAiB;GAAgB;EAAa;EAC/C;GAAC;GAAW;GAAU;EAAQ;EAC9B;GAAC;IAAiB,MAAM,OAAO,MAAM;GAAW;EAAS;EACzD;GAAC;IAAiB,MAAM,OAAO,MAAM;GAAW;EAAS;EACzD;GAAC;GAAS;GAAgB;EAAa;CACzC,GACA,MACF;CACA,IAAI,SAAS,MAAM,OAAO,GAAG;EAC3B,IAAI,EAAE,aAAa,MAAM,UAAU,OAAO,KAAK,+CAA+C;EAC9F,IAAI,OAAO,MAAM,QAAQ,mBAAmB,WAC1C,OAAO,KAAK,0CAA0C;EAExD,IACE,MAAM,QAAQ,sBAAsB,KAAA,KACpC,OAAO,MAAM,QAAQ,sBAAsB,WAE3C,OAAO,KAAK,0DAA0D;EAExE,OAAO,KACL,GAAG,WACD;GAAE,SAAS,MAAM;GAAS,OAAO,MAAM;EAAM,GAC7C,cAAc,mBAChB,CACF;CACF;CAEA,gBACE,MAAM,MACN,QACA;EACE;GAAC;GAAO;GAAgB;EAAa;EACrC;GAAC;GAAa;GAAgB;EAAa;EAC3C;GAAC;GAAc;GAAgB;EAAa;EAC5C;GAAC;GAAoB;GAAgB;EAAa;EAClD;GAAC;GAAc;GAAgB;EAAa;EAC5C;GAAC;GAAe;GAAgB;EAAa;EAC7C;GAAC;GAAU;GAAgB;EAAa;CAC1C,GACA,MACF;CACA,IACE,SAAS,MAAM,IAAI,KACnB,MAAM,KAAK,mBAAmB,KAAA,KAC9B,EAAE,MAAM,KAAK,mBAAmB,QAAQ,OAAO,UAAU,MAAM,KAAK,cAAc,IAElF,OAAO,KAAK,yDAAyD;CAGvE,gBACE,MAAM,WACN,aACA;EACE;GAAC;GAAc;GAAgB;EAAa;EAC5C;GAAC;GAAW;GAAgB;EAAa;EACzC;GAAC;GAAkB;GAAgB;EAAa;CAClD,GACA,MACF;CAEA,gBACE,MAAM,YACN,cACA,CACE;EACE;GACC,MAAM,OAAO,MAAM,YAAY,CAAC,OAAO,MAAM,KAAK,MAAM,CAAC,CAAC;EAC3D;CACF,GACA;EACE;GACC,MAAM,iBAAiB,SAAS,CAAmB;EACpD,UAAU,iBAAiB,KAAK,GAAG;CACrC,CACF,GACA,MACF;CACA,IAAI,SAAS,MAAM,UAAU,GAAG;EAC9B,IAAI,MAAM,WAAW,QAAQ,KAAA,KAAa,OAAO,MAAM,WAAW,QAAQ,UACxE,OAAO,KAAK,6CAA6C;EAE3D,MAAM,WAAW,MAAM,WAAW;EAClC,IAAI,aAAa,KAAA,GACX;OAAA,CAAC,SAAS,QAAQ,GACpB,OAAO,KAAK,2DAA2D;QAClE,IAAI,SAAS,YAAY,KAAA,KAAa,CAAC,SAAS,SAAS,OAAO,GACrE,OAAO,KAAK,mEAAmE;EAAA;CAGrF;CAGA,IAAI,MAAM,QAAQ,MAAM,QAAQ,KAAK,SAAS,MAAM,UAAU,GACxD;MAAA,MAAM,SAAS,WAAW,KAAK,OAAO,MAAM,WAAW,QAAQ,UACjE,OAAO,KAAK,iDAAiD;CAAA;CAIjE,OAAO;AACT;AAEA,SAAgB,kBACd,OACA,UAAU,gBACoB;CAC9B,MAAM,SAAS,oBAAoB,KAAK;CACxC,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,MAAM,WAAW,QAAQ,OAAO,OAAO,KAAK,MAAM,GAAG;AAEnE;AAEA,SAAgB,cAAc,OAAsC;CAClE,OAAO,oBAAoB,KAAK,CAAC,CAAC,WAAW;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkFA,SAAgB,aAAa,OAAgB,UAAU,gBAAmC;CACxF,kBAAkB,OAAO,OAAO;CAChC,MAAM,UAAU,WAAW,OAAO,cAAc,YAAY;CAC5D,IAAI,QAAQ,SAAS,GACnB,MAAM,IAAI,MAAM,WAAW,QAAQ,YAAY,MAAM,WAAW,KAAK,QAAQ,IAAI;CAEnF,MAAM,QAAQ,iBAAiB,KAAK;CACpC,IAAI,MAAM,QAAQ,mBAAmB,KAAA,GACnC,OAAO;EACL,GAAG;EACH,SAAS;GAAE,GAAG,MAAM;GAAS,gBAAgB;EAAM;CACrD;CAEF,OAAO;AACT;;AAGA,SAAgB,kBACd,QACA,UAAU,gBACW;CACrB,OAAO,OAAO,KAAK,OAAO,MAAM,aAAa,OAAO,GAAG,QAAQ,IAAI,EAAE,EAAE,CAAC;AAC1E"}

package/dist/schema-CRhEY1SO.js

@@ -0,0 +1,69 @@
+//#region src/trace/schema.ts
+/**
+* TraceSchema v1 — the canonical data model for agent-eval.
+*
+* Every score, every failure class, every pipeline in the framework is
+* a view over this data. Shape it once, live with it.
+*
+* Wire-compatible with OpenTelemetry span semantics (see trace/otel.ts)
+* but extended with agent-specific span kinds (llm, tool, retrieval,
+* judge, sandbox) and first-class BudgetLedger / Artifact / JudgeVerdict
+* entities that OTEL leaves as free-form attributes.
+*/
+const TRACE_SCHEMA_VERSION = "1.0.0";
+const FAILURE_CLASSES = [
+	"success",
+	"reasoning_error",
+	"tool_selection_error",
+	"tool_argument_error",
+	"tool_recovery_failure",
+	"hallucination",
+	"instruction_following",
+	"safety_refusal_miss",
+	"policy_violation",
+	"budget_exceeded",
+	"format_drift",
+	"permission_escalation",
+	"pii_leak",
+	"cost_overrun",
+	"timeout",
+	"sandbox_failure",
+	"missing_user_data",
+	"missing_domain_data",
+	"missing_codebase_context",
+	"missing_runtime_context",
+	"missing_credentials",
+	"missing_integration_connection",
+	"missing_integration_scope",
+	"integration_approval_required",
+	"integration_auth_expired",
+	"integration_provider_failure",
+	"bad_integration_manifest",
+	"unsafe_integration_write_denied",
+	"stale_external_data",
+	"bad_retrieval",
+	"insufficient_evidence",
+	"contradictory_evidence",
+	"ambiguous_user_intent",
+	"knowledge_readiness_blocked",
+	"unknown"
+];
+function isLlmSpan(s) {
+	return s.kind === "llm";
+}
+function isToolSpan(s) {
+	return s.kind === "tool";
+}
+function isRetrievalSpan(s) {
+	return s.kind === "retrieval";
+}
+function isJudgeSpan(s) {
+	return s.kind === "judge";
+}
+function isSandboxSpan(s) {
+	return s.kind === "sandbox";
+}
+//#endregion
+export { isRetrievalSpan as a, isLlmSpan as i, TRACE_SCHEMA_VERSION as n, isSandboxSpan as o, isJudgeSpan as r, isToolSpan as s, FAILURE_CLASSES as t };
+
+//# sourceMappingURL=schema-CRhEY1SO.js.map

package/dist/schema-CRhEY1SO.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"schema-CRhEY1SO.js","names":[],"sources":["../src/trace/schema.ts"],"sourcesContent":["/**\n * TraceSchema v1 — the canonical data model for agent-eval.\n *\n * Every score, every failure class, every pipeline in the framework is\n * a view over this data. Shape it once, live with it.\n *\n * Wire-compatible with OpenTelemetry span semantics (see trace/otel.ts)\n * but extended with agent-specific span kinds (llm, tool, retrieval,\n * judge, sandbox) and first-class BudgetLedger / Artifact / JudgeVerdict\n * entities that OTEL leaves as free-form attributes.\n */\n\nexport const TRACE_SCHEMA_VERSION = '1.0.0'\n\n// ── Run ──────────────────────────────────────────────────────────────\n\nexport type RunStatus = 'running' | 'completed' | 'failed' | 'aborted'\n\nexport interface BudgetSpec {\n  tokens?: number\n  wallMs?: number\n  calls?: number\n  usd?: number\n}\n\nexport interface RunOutcome {\n  score?: number\n  pass?: boolean\n  failureClass?: FailureClass\n  notes?: string\n}\n\n/**\n * Layer — optional classification in a nested build workflow.\n * `builder`: the meta-agent editing a project (e.g. agent-builder Forge chat).\n * `app-build`: sandbox harness that compiled + tested the generated scaffold.\n * `app-runtime`: a run of the generated agent against a domain scenario.\n * `meta`: any meta-eval (judge replay, correlation analysis).\n */\nexport type RunLayer = 'builder' | 'app-build' | 'app-runtime' | 'meta' | 'custom'\n\nexport interface Run {\n  runId: string\n  /**\n   * Stable identifier of the scenario being executed.\n   *\n   * Always populated on the persisted Run — but `TraceEmitter.startRun` accepts\n   * input WITHOUT this field, substituting a sensible default\n   * (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) when the caller has no\n   * curated scenario to anchor to (runtime / operator / meta-eval runs). This\n   * keeps the persisted shape unambiguous for downstream filters + aggregations\n   * while removing the boilerplate of inventing placeholder ids at the call site.\n   */\n  scenarioId: string\n  variantId?: string\n  datasetVersion?: string\n  /** Git SHA of agent code at run time. */\n  codeSha?: string\n  /** Hash of the prompt template + any system prompt. */\n  promptSha?: string\n  /** Model id + date + system-prompt hash, concatenated. */\n  modelFingerprint?: string\n  seed?: number\n  /** Arbitrary environment markers (shell, docker version, tz). */\n  envFingerprint?: Record<string, string>\n  /** Version of the redaction rules applied to this run. */\n  redactionVersion?: string\n  /** Parent run in a nested build workflow. A builder run's children are\n   *  app-build runs; those children are app-runtime runs. */\n  parentRunId?: string\n  /** Stable project identifier — groups runs across chats + sessions. */\n  projectId?: string\n  /** Chat/conversation identifier within a project. */\n  chatId?: string\n  /** Layer classification — hint for aggregation; not enforced. */\n  layer?: RunLayer\n  startedAt: number\n  endedAt?: number\n  status: RunStatus\n  outcome?: RunOutcome\n  budget?: BudgetSpec\n  /** Free-form labels for downstream grouping. */\n  tags?: Record<string, string>\n}\n\n// ── Spans (hierarchical work units) ──────────────────────────────────\n\nexport type SpanKind = 'agent' | 'llm' | 'tool' | 'retrieval' | 'judge' | 'sandbox' | 'custom'\n\nexport type SpanStatus = 'ok' | 'error'\n\nexport interface SpanBase {\n  spanId: string\n  parentSpanId?: string\n  runId: string\n  kind: SpanKind\n  name: string\n  startedAt: number\n  endedAt?: number\n  status?: SpanStatus\n  error?: string\n  /** Anything not covered by typed fields. Kept deliberately free-form. */\n  attributes?: Record<string, unknown>\n}\n\nexport interface Message {\n  role: 'system' | 'user' | 'assistant' | 'tool'\n  content: string\n  tokens?: number\n  /** Multi-modal content descriptors; blobs themselves live in Artifacts. */\n  images?: Array<{ artifactId?: string; url?: string; mime?: string }>\n}\n\nexport interface LlmSpan extends SpanBase {\n  kind: 'llm'\n  model: string\n  messages: Message[]\n  output?: string\n  inputTokens?: number\n  /** All generated tokens, including the reasoning subset when present. */\n  outputTokens?: number\n  cachedTokens?: number\n  cacheWriteTokens?: number\n  /** Reasoning-token subset of `outputTokens`. */\n  reasoningTokens?: number\n  costUsd?: number\n  finishReason?: string\n}\n\nexport interface ToolSpan extends SpanBase {\n  kind: 'tool'\n  toolName: string\n  args: unknown\n  /** False when the source observed the call but did not capture its arguments. */\n  argsCaptured?: boolean\n  result?: unknown\n  latencyMs?: number\n}\n\nexport interface RetrievalSpan extends SpanBase {\n  kind: 'retrieval'\n  query: string\n  hits: Array<{ docId: string; score: number; content?: string }>\n}\n\nexport interface JudgeSpan extends SpanBase {\n  kind: 'judge'\n  judgeId: string\n  /** Span this judgment applies to. */\n  targetSpanId: string\n  dimension: string\n  /** Numeric score (free-range; interpretation up to the judge). */\n  score: number\n  rationale?: string\n  evidence?: string\n}\n\nexport interface SandboxSpan extends SpanBase {\n  kind: 'sandbox'\n  image?: string\n  command?: string\n  exitCode?: number\n  testsTotal?: number\n  testsPassed?: number\n  stdoutHash?: string\n  stderrHash?: string\n  /** Duration in ms; the harness fills this explicitly (endedAt - startedAt may miss setup). */\n  wallMs?: number\n}\n\nexport interface GenericSpan extends SpanBase {\n  kind: 'agent' | 'custom'\n}\n\nexport type Span = LlmSpan | ToolSpan | RetrievalSpan | JudgeSpan | SandboxSpan | GenericSpan\n\n// ── Events (point-in-time occurrences within a span) ─────────────────\n\nexport type EventKind =\n  | 'log'\n  | 'error'\n  | 'budget_decrement'\n  | 'budget_breach'\n  | 'state_mutation'\n  | 'policy_violation'\n  | 'redaction_applied'\n  | 'custom'\n\nexport interface TraceEvent {\n  eventId: string\n  runId: string\n  spanId?: string\n  kind: EventKind\n  timestamp: number\n  payload: Record<string, unknown>\n}\n\n// ── Budget ledger (running token/wall/call/$ accounting) ─────────────\n\nexport interface BudgetLedgerEntry {\n  runId: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  remaining: number\n  timestamp: number\n  breached: boolean\n  /** Span that triggered this entry, if any. */\n  spanId?: string\n}\n\n// ── Artifacts (blobs addressed by hash) ──────────────────────────────\n\nexport interface Artifact {\n  artifactId: string\n  runId: string\n  spanId?: string\n  contentType: string\n  sizeBytes: number\n  /** sha256 in hex. */\n  hash: string\n  /** External storage URL (R2, S3, filesystem path). */\n  storageUrl?: string\n  /** Inline content for small blobs — keep under ~64KB. */\n  inlineContent?: string\n}\n\n// ── Failure taxonomy ─────────────────────────────────────────────────\n\nexport type FailureClass =\n  | 'success'\n  | 'reasoning_error'\n  | 'tool_selection_error'\n  | 'tool_argument_error'\n  | 'tool_recovery_failure'\n  | 'hallucination'\n  | 'instruction_following'\n  | 'safety_refusal_miss'\n  | 'policy_violation'\n  | 'budget_exceeded'\n  | 'format_drift'\n  | 'permission_escalation'\n  | 'pii_leak'\n  | 'cost_overrun'\n  | 'timeout'\n  | 'sandbox_failure'\n  | 'missing_user_data'\n  | 'missing_domain_data'\n  | 'missing_codebase_context'\n  | 'missing_runtime_context'\n  | 'missing_credentials'\n  | 'missing_integration_connection'\n  | 'missing_integration_scope'\n  | 'integration_approval_required'\n  | 'integration_auth_expired'\n  | 'integration_provider_failure'\n  | 'bad_integration_manifest'\n  | 'unsafe_integration_write_denied'\n  | 'stale_external_data'\n  | 'bad_retrieval'\n  | 'insufficient_evidence'\n  | 'contradictory_evidence'\n  | 'ambiguous_user_intent'\n  | 'knowledge_readiness_blocked'\n  | 'unknown'\n\nexport const FAILURE_CLASSES: readonly FailureClass[] = [\n  'success',\n  'reasoning_error',\n  'tool_selection_error',\n  'tool_argument_error',\n  'tool_recovery_failure',\n  'hallucination',\n  'instruction_following',\n  'safety_refusal_miss',\n  'policy_violation',\n  'budget_exceeded',\n  'format_drift',\n  'permission_escalation',\n  'pii_leak',\n  'cost_overrun',\n  'timeout',\n  'sandbox_failure',\n  'missing_user_data',\n  'missing_domain_data',\n  'missing_codebase_context',\n  'missing_runtime_context',\n  'missing_credentials',\n  'missing_integration_connection',\n  'missing_integration_scope',\n  'integration_approval_required',\n  'integration_auth_expired',\n  'integration_provider_failure',\n  'bad_integration_manifest',\n  'unsafe_integration_write_denied',\n  'stale_external_data',\n  'bad_retrieval',\n  'insufficient_evidence',\n  'contradictory_evidence',\n  'ambiguous_user_intent',\n  'knowledge_readiness_blocked',\n  'unknown',\n] as const\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nexport function isLlmSpan(s: Span): s is LlmSpan {\n  return s.kind === 'llm'\n}\nexport function isToolSpan(s: Span): s is ToolSpan {\n  return s.kind === 'tool'\n}\nexport function isRetrievalSpan(s: Span): s is RetrievalSpan {\n  return s.kind === 'retrieval'\n}\nexport function isJudgeSpan(s: Span): s is JudgeSpan {\n  return s.kind === 'judge'\n}\nexport function isSandboxSpan(s: Span): s is SandboxSpan {\n  return s.kind === 'sandbox'\n}\n"],"mappings":";;;;;;;;;;;;AAYA,MAAa,uBAAuB;AA8PpC,MAAa,kBAA2C;CACtD;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAIA,SAAgB,UAAU,GAAuB;CAC/C,OAAO,EAAE,SAAS;AACpB;AACA,SAAgB,WAAW,GAAwB;CACjD,OAAO,EAAE,SAAS;AACpB;AACA,SAAgB,gBAAgB,GAA6B;CAC3D,OAAO,EAAE,SAAS;AACpB;AACA,SAAgB,YAAY,GAAyB;CACnD,OAAO,EAAE,SAAS;AACpB;AACA,SAAgB,cAAc,GAA2B;CACvD,OAAO,EAAE,SAAS;AACpB"}