@tangle-network/agent-eval 0.216.0 → 0.217.1

package/dist/produced-state-ZCtDBxAE.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"produced-state-ZCtDBxAE.js","names":[],"sources":["../src/agent-profile.ts","../src/completion-verifier.ts","../src/produced-state.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { type HarnessType, harnessSupportsModel } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { canonicalString } from './ledger-core/canonical'\n\nexport type { AgentProfile, HarnessType } from '@tangle-network/agent-interface'\n\n/**\n * The agentic coding harnesses an eval sweeps by default — the ones we care about\n * ranking. This is the SINGLE source of that list; consumers import it instead of\n * re-declaring their own (a re-declared list is how the fleet drifts). Pass an\n * explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known\n * harness) to widen beyond these.\n */\nexport const CODING_HARNESSES: readonly HarnessType[] = [\n  'opencode',\n  'claude-code',\n  'codex',\n  'kimi-code',\n]\n\nexport interface ProfileAxisSpec {\n  /** The domain profile to sweep. Its prompt/tools/skills are held fixed; only the\n   *  harness and model vary. `model.default` is the fallback model. */\n  base: AgentProfile\n  /** Harnesses to cross. Default: {@link CODING_HARNESSES}. */\n  harnesses?: readonly HarnessType[]\n  /** Models to cross. Default: `[base.model.default]` — one model, i.e. today's\n   *  single-model behaviour, so omitting this never changes an existing run. */\n  models?: readonly string[]\n  /** Force every (harness, model) pair verbatim, even ones the harness can't run —\n   *  for deliberately testing failure modes. Default (false): SNAP instead — a\n   *  vendor-locked harness runs only the swept models in its family, or its native\n   *  default when it supports none, so no harness is dropped and none gets a\n   *  guaranteed-failing foreign-model cell. */\n  keepIncompatible?: boolean\n}\n\n/** Model sentinel for a vendor-locked harness that supports none of the swept models:\n *  it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness\n *  resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi\n *  model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship\n *  table that would rot as router catalogs change. */\nexport const HARNESS_NATIVE_MODEL = 'default'\n\n/**\n * Expand a base profile across the harness × model matrix into the `AgentProfile[]`\n * that `runProfileMatrix` / `selfImprove` score — the ONE place \"which harnesses ×\n * which models do we evaluate\" lives, so no product hand-rolls its own harness list\n * or column→profile mapping (the pattern that let those copies drift and silently\n * break the harness pivot).\n *\n * Each cell clones `base`, sets the canonical top-level `harness` and `model.default`,\n * and stamps `metadata.harness` + `metadata.harnessModel` for matrix grouping. Both\n * metadata fields are hash-bearing, so every cell gets a distinct `agentProfileId` row\n * and results join back by harness/model via {@link harnessAxisOf} with no\n * hand-recomputed key. A vendor-locked harness snaps to its family's swept models — or\n * its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none — so every\n * requested harness runs; `keepIncompatible` forces every pair verbatim.\n *\n * Omit `harnesses`/`models` to sweep the full default set — the \"turn it on for\n * everything we care about\" switch, identical in shape whether one harness or all.\n */\nexport function expandProfileAxes(spec: ProfileAxisSpec): AgentProfile[] {\n  const harnesses = spec.harnesses ?? CODING_HARNESSES\n  if (harnesses.length === 0) throw new ValidationError('expandProfileAxes: no harnesses to sweep')\n  const baseModel = spec.base.model?.default\n  const models = spec.models ?? (baseModel ? [baseModel] : [])\n  if (models.length === 0) {\n    throw new ValidationError(\n      'expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied',\n    )\n  }\n  const out: AgentProfile[] = []\n  const seen = new Set<string>()\n  for (const harness of harnesses) {\n    // A universal (router-backed) harness — opencode/pi/claudish — runs every swept\n    // model. A vendor-locked harness — codex/claude-code/kimi-code — runs only the\n    // swept models in its own family; when it supports NONE of them it snaps to its\n    // native default (the `HARNESS_NATIVE_MODEL` sentinel it resolves at runtime)\n    // rather than being dropped, so every requested harness still appears in the\n    // sweep on a model it can actually run — e.g. sweeping `deepseek/x` puts opencode\n    // on deepseek and kimi-code on its own Kimi model, a real head-to-head.\n    // `keepIncompatible` forces every (harness, model) pair verbatim (failure-mode runs).\n    const supported = spec.keepIncompatible\n      ? models\n      : models.filter((model) => harnessSupportsModel(harness, model))\n    const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL]\n    for (const model of effective) {\n      const profile: AgentProfile = {\n        ...spec.base,\n        harness,\n        name: `${spec.base.name ?? 'agent'}/${harness}/${model}`,\n        model: { ...spec.base.model, default: model },\n        metadata: { ...(spec.base.metadata ?? {}), harness, harnessModel: model },\n      }\n      const id = agentProfileId(profile)\n      if (seen.has(id)) continue\n      seen.add(id)\n      out.push(profile)\n    }\n  }\n  if (out.length === 0) {\n    // Unreachable in normal use — snapping guarantees ≥1 cell per harness — but keep a\n    // fail-closed guard so a future refactor can't silently produce an empty sweep.\n    throw new ValidationError(\n      `expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(', ')}], models=[${models.join(', ')}]).`,\n    )\n  }\n  return out\n}\n\n/**\n * Read the (harness, model) a matrix cell ran under, off a profile or a result row's\n * profile — the join-back for a `byHarness` pivot. Returns undefined when the profile\n * wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by\n * this instead of recomputing an id (recomputing the wrong key is what broke the pivot\n * in the hand-rolled copies).\n */\nexport function harnessAxisOf(\n  profile: Pick<AgentProfile, 'metadata'>,\n): { harness: HarnessType; model: string } | undefined {\n  const m = profile.metadata as Record<string, unknown> | undefined\n  const harness = m?.harness\n  const model = m?.harnessModel\n  if (typeof harness === 'string' && typeof model === 'string') {\n    return { harness: harness as HarnessType, model }\n  }\n  return undefined\n}\n\n/**\n * Collision-resistant, path-safe, human-readable profile id for eval artifacts.\n * Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix\n * keys, and directory names where two profiles must not collapse onto one row.\n * The suffix is the first 64 bits of the behaviour hash, enough for ordinary\n * eval matrices while keeping filenames readable.\n */\nexport function agentProfileId(profile: AgentProfile): string {\n  const label = pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? 'profile'\n  return `${label}-${agentProfileHash(profile).slice(0, 16)}`\n}\n\n/**\n * Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete\n * model id because run records reject bare/missing model aliases.\n */\nexport function agentProfileModelId(profile: AgentProfile): string {\n  const model = profile.model?.default?.trim()\n  if (!model) {\n    const label = agentProfileDisplayLabel(profile) ?? 'unnamed profile'\n    throw new ValidationError(\n      `AgentProfile \"${label}\" has no model.default — cannot record eval run`,\n    )\n  }\n  return model\n}\n\nfunction agentProfileDisplayLabel(profile: AgentProfile): string | undefined {\n  return profile.name?.trim() || profile.version?.trim() || undefined\n}\n\nfunction pathSafeProfileLabel(label: string | undefined): string | undefined {\n  const safe = label\n    ?.trim()\n    .replace(/[^A-Za-z0-9._-]+/g, '-')\n    .replace(/-+/g, '-')\n    .replace(/^-|-$/g, '')\n  return safe || undefined\n}\n\nfunction compact<T extends Record<string, unknown>>(input: T): Partial<T> {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as Partial<T>\n}\n\n/**\n * Deterministic behaviour identity for the canonical\n * `@tangle-network/agent-interface` AgentProfile.\n *\n * `name` and `description` are labels and do not affect the hash. Profile\n * `version`, prompt, model hints, tools, resources, hooks, modes, permissions,\n * and extensions do affect the hash. Resource array order is hash-bearing\n * because mount order can change agent behaviour. Undefined fields are treated\n * as absent; explicit `null` fields remain hash-bearing.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  const model = agentProfileModelId(profile)\n  const behaviour = compact({\n    ...profile,\n    name: undefined,\n    description: undefined,\n    tags: profile.tags ? [...profile.tags].sort() : undefined,\n    model: compact({ ...profile.model, default: model }),\n  })\n  return createHash('sha256').update(canonicalString(behaviour)).digest('hex')\n}\n","/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n */\n\nimport { randomUUID } from 'node:crypto'\nimport type { ChatClient, ChatRequest } from './analyst/chat-client'\nimport type { Artifact } from './artifact-validator'\nimport { CostLedger, type CostLedgerHandle } from './cost-ledger'\nimport { assertServedModel, ModelSubstitutionError } from './integrity/served-model'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport {\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport {\n  certificationEvidenceDigest,\n  type DefaultVerdict,\n  type VerdictCertification,\n} from './verdict'\nimport type { CheckerIdentity, VerificationStrategySource } from './verification-strategy'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `Artifact` shape. */\n  artifacts: Artifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, when the matched item carries no content\n   * to assess, or when the correctness check itself failed (`unmeasured`).\n   */\n  correct: boolean | null\n  /** structurallyPresent && !unmeasured && correct !== false. */\n  satisfied: boolean\n  /**\n   * Set when the correctness check itself errored (LLM call failure or an\n   * unparseable response after retry). The requirement's fulfilment is\n   * UNKNOWN — `correct` stays null, `satisfied` is false, and\n   * `completionVerdict` excludes the row from `completionRate`'s\n   * denominator. Never folded into a zero: a synthetic zero is\n   * indistinguishable from a real failure (see `JudgeParseError`).\n   */\n  unmeasured?: true\n  /** Why the correctness check could not be measured (present iff `unmeasured`). */\n  unmeasuredReason?: string\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\n/** Extends the substrate verdict spine: `valid` = `fullyComplete` and\n *  `score` = `completionRate` — derived in `completionVerdict()`, the one\n *  place those equalities hold by construction. */\nexport interface CompletionVerdict extends DefaultVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */\n  completionRate: number\n  /** Every measurable requirement satisfied (false when anything is unmeasured). */\n  fullyComplete: boolean\n  /** Requirements whose correctness check errored — reported, never scored as zero. */\n  unmeasuredCount: number\n}\n\n/**\n * Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n */\nexport function completionVerdict(input: {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** What certified the correctness stage, when anything did. Omitted =\n   *  an uncertified verdict — the honest default for a bare checker. */\n  certification?: VerdictCertification\n}): CompletionVerdict {\n  if (input.requirements.length === 0) {\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n    )\n  }\n  const measurable = input.requirements.filter((r) => !r.unmeasured)\n  const unmeasuredCount = input.requirements.length - measurable.length\n  if (measurable.length === 0) {\n    // Every check errored: this is an infrastructure failure, not a scored\n    // run. A 0-rate verdict here would be a fabricated measurement.\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n    )\n  }\n  const satisfiedCount = measurable.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / measurable.length\n  // A run with unmeasured rows can still report a rate over what WAS\n  // measured, but must not claim full completion.\n  const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    unmeasuredCount,\n    valid: fullyComplete,\n    score: completionRate,\n    ...(input.certification === undefined ? {} : { certification: input.certification }),\n  }\n}\n\n/**\n * What a correctness checker declares about itself so the completion\n * verdict can carry a certification: which strategy member it discharges,\n * its exact identity, and the steps its answers rest on unverified.\n */\nexport interface CorrectnessCheckerAttestation {\n  strategy: VerificationStrategySource\n  checker: CheckerIdentity\n  assumptions: string[]\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n *\n * `attestation` is optional metadata on the function value: a checker that\n * carries one yields certified completion verdicts; a bare function yields\n * the same verdict uncertified. A plain arrow function remains a valid\n * checker.\n */\nexport interface CorrectnessChecker {\n  (\n    requirement: CompletionRequirement,\n    content: string,\n  ): Promise<{ correct: boolean; reason: string }>\n  attestation?: CorrectnessCheckerAttestation\n}\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/** artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n  'generated',\n  'generate',\n  'view',\n  'render',\n  'rendered',\n  'persisted',\n  'persist',\n  'artifact',\n  'file',\n  'document',\n  'note',\n  'proposal',\n  'deliverable',\n  'output',\n  'created',\n  'create',\n  'produce',\n  'produced',\n  'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    // Match against the artifact CONTENT too, not just its path + kind — a\n    // generated view / note whose path is generic still satisfies a requirement\n    // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n    // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n    let score = tokenRecall(\n      reqText,\n      `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n    )\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    // A proposal needs an assessable BODY to be a deliverable. A bare title is\n    // not completion: correctness cannot be judged on it, so a title-only match\n    // would auto-pass the oracle (structurallyPresent && correct===null →\n    // satisfied) with no verifiable content. Tool calls are the only\n    // legitimately content-less deliverable (`toolCallCandidates`).\n    const body = (p.content ?? '').trim()\n    if (body.length < MIN_CONTENT_CHARS) continue\n    // Match against the body as well as the (often short) title — a refusal /\n    // flag / analysis proposal whose title is a label still satisfies a\n    // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n    // + the requirement's distinctive tokens keep an off-topic proposal out;\n    // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n    // polarity/fulfilment, so a negation that merely contains the tokens fails.\n    // Structural and correctness must use different evidence or the two-stage\n    // check collapses to one lexical gate.\n    const score = tokenRecall(reqText, `${p.title} ${body}`)\n    if (score < MATCH_THRESHOLD) continue\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n    let unmeasuredReason: string | undefined\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        try {\n          const r = await checkCorrectness(req, match.content)\n          correct = r.correct\n          evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n        } catch (err) {\n          // The CHECKER failed, not the requirement. Recording this as a\n          // zero would fabricate a model failure out of an infrastructure\n          // one; the requirement is unmeasured and leaves the denominator.\n          unmeasuredReason =\n            err instanceof JudgeParseError\n              ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n              : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n          evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n        }\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const unmeasured = unmeasuredReason !== undefined\n    const satisfied = structurallyPresent && !unmeasured && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n      evidence,\n    })\n  }\n\n  // The verdict blends the deterministic structural stage with the injected\n  // correctness stage, so the certification names the checker's own member\n  // and carries the structural stage's lexical nature as an assumption. A\n  // checker without an attestation yields an uncertified verdict.\n  const attestation = checkCorrectness.attestation\n  const certification: VerdictCertification | undefined = attestation\n    ? {\n        strategy: attestation.strategy,\n        checker: attestation.checker,\n        assumptions: [\n          'structural matching is lexical token recall over produced items — the correctness stage only sees items it matched',\n          ...attestation.assumptions,\n        ],\n        evidenceDigest: certificationEvidenceDigest({ taskId: gold.taskId, requirements }),\n      }\n    : undefined\n\n  return completionVerdict({\n    taskId: gold.taskId,\n    requirements,\n    ...(certification === undefined ? {} : { certification }),\n  })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Optional ledger for direct use. */\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n  /**\n   * Checker LLM calls per requirement before giving up (parse failures and\n   * call errors both consume attempts). The failure then surfaces as an\n   * `unmeasured` requirement, never a zero.\n   */\n  maxAttempts?: number\n  /**\n   * Forensic capture of every checker request/response/error — without it a\n   * checker failure is unauditable (the agent-turn raws never contain the\n   * checker's own calls). Same sink contract as `LlmClient`.\n   */\n  rawSink?: RawProviderSink\n}\n\n/**\n * Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const readVerdict = (candidate: unknown): { correct: boolean; reason: string } | null => {\n    if (candidate === null || typeof candidate !== 'object') return null\n    const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n    if (typeof correct !== 'boolean') return null\n    return { correct, reason: typeof reason === 'string' ? reason : '' }\n  }\n\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (match) {\n    try {\n      const strict = readVerdict(JSON.parse(match[0]))\n      if (strict) return strict\n    } catch {\n      // fall through to truncation recovery\n    }\n  }\n  // The strict path needs a closing `}`; a cap-hit response has none. Take\n  // everything from the first `{` and auto-close it.\n  const start = raw.indexOf('{')\n  if (start !== -1) {\n    const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n    if (recovered) return recovered\n  }\n  throw new JudgeParseError('correctness-checker', raw)\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  chat: ChatClient,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  const maxAttempts = opts.maxAttempts ?? 2\n  const costLedger = opts.costLedger ?? new CostLedger()\n  const sink = opts.rawSink\n  const record = async (event: RawProviderEvent): Promise<void> => {\n    // Forensic capture is best-effort; the verdict is the system of record.\n    try {\n      await sink?.record(event)\n    } catch {\n      // Intentionally swallowed.\n    }\n  }\n  const checker: CorrectnessChecker = async (requirement, content) => {\n    const request = {\n      model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user' as const,\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    } satisfies ChatRequest\n    let lastErr: unknown\n    for (let attempt = 0; attempt < maxAttempts; attempt++) {\n      const started = Date.now()\n      await record({\n        eventId: randomUUID(),\n        provider: chat.transport,\n        model,\n        endpoint: '/chat',\n        baseUrl: '',\n        attemptIndex: attempt,\n        direction: 'request',\n        timestamp: started,\n        requestBody: request,\n        redactedFields: [],\n      })\n      try {\n        const paid = await costLedger.runPaidCall({\n          channel: 'verifier',\n          phase: opts.costPhase ?? 'completion.correctness',\n          actor: 'correctness-checker',\n          model,\n          maximumCharge:\n            chat.maximumAttempts === undefined\n              ? undefined\n              : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),\n          tags: {\n            ...opts.costTags,\n            requirementId: requirement.reqId,\n            attempt: String(attempt),\n          },\n          signal: opts.signal,\n          execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),\n          receipt: costReceiptFromLlm,\n          receiptFromError: costReceiptFromLlmError,\n        })\n        if (!paid.succeeded) throw paid.error\n        const resp = paid.value\n        // Hold the transport to its own word: a verdict produced by a model\n        // other than the one requested is not this model's verdict. A\n        // transport that echoes no id cannot be made to prove identity here —\n        // callers needing that proof enable it at the client\n        // (`LlmClientOptions.assertServedModel`) or gate on `assertModelsServed`.\n        assertServedModel(model, resp.servedModel, {\n          allowUnreported: true,\n          context: `correctness checker for requirement ${requirement.reqId}`,\n        })\n        const raw = resp.content\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'response',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          responseBody: resp,\n          redactedFields: [],\n        })\n        return parseCorrectnessResponse(raw)\n      } catch (err) {\n        lastErr = err\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'error',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          errorMessage: err instanceof Error ? err.message : String(err),\n          redactedFields: [],\n        })\n        // Substitution is a verdict about identity, not a transient fault:\n        // another attempt reaches the same wrong model and spends more tokens.\n        if (err instanceof ModelSubstitutionError) throw err\n      }\n    }\n    throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n  }\n  checker.attestation = {\n    strategy: 'judge',\n    checker: { name: 'llm-correctness-checker', version: model },\n    assumptions: [\n      'served-model identity is accepted when the transport reports none (assertServedModel allowUnreported)',\n    ],\n  }\n  return checker\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n  // body of the proposal (e.g. a submit_proposal `description`). When present,\n  // the completion oracle correctness-checks it like artifact content; absent,\n  // the proposal is graded presence-only.\n  content?: string\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize an emission-ordered stream into its latest observed produced state.\n * Artifacts are keyed by their exact output path (name, then URI, then id);\n * proposals by id. Later observations replace earlier ones, including missing\n * content or a rejected status. An obsolete version cannot prove completion.\n * Distinct paths remain distinct; no path normalization or aliasing is inferred.\n * Results retain first-seen identity order. Tool names describe invocation,\n * not success, and are deduplicated in first-seen order.\n *\n * An artifact without observed content yields empty content, which the\n * completion oracle rejects. This projection does not verify external effects,\n * recover dropped events, or establish freshness beyond the supplied stream.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts = new Map<string, Artifact>()\n  const proposals = new Map<string, ProducedProposal>()\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      const path = a.name ?? a.uri ?? a.artifactId\n      artifacts.set(path, {\n        kind: artifactKind(a.mimeType),\n        path,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.set(p.proposalId, {\n        id: p.proposalId,\n        title: p.title,\n        status: p.status ?? 'pending',\n        ...(p.content !== undefined ? { content: p.content } : {}),\n      })\n    }\n  }\n\n  return { artifacts: [...artifacts.values()], proposals: [...proposals.values()], toolCalls }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAeA,MAAa,mBAA2C;CACtD;CACA;CACA;CACA;AACF;;;;;;AAwBA,MAAa,uBAAuB;;;;;;;;;;;;;;;;;;;AAoBpC,SAAgB,kBAAkB,MAAuC;CACvE,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,IAAI,gBAAgB,0CAA0C;CAChG,MAAM,YAAY,KAAK,KAAK,OAAO;CACnC,MAAM,SAAS,KAAK,WAAW,YAAY,CAAC,SAAS,IAAI,CAAC;CAC1D,IAAI,OAAO,WAAW,GACpB,MAAM,IAAI,gBACR,kGACF;CAEF,MAAM,MAAsB,CAAC;CAC7B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,WAAW,WAAW;EAS/B,MAAM,YAAY,KAAK,mBACnB,SACA,OAAO,QAAQ,UAAU,qBAAqB,SAAS,KAAK,CAAC;EACjE,MAAM,YAAY,UAAU,SAAS,IAAI,YAAY,CAAC,oBAAoB;EAC1E,KAAK,MAAM,SAAS,WAAW;GAC7B,MAAM,UAAwB;IAC5B,GAAG,KAAK;IACR;IACA,MAAM,GAAG,KAAK,KAAK,QAAQ,QAAQ,GAAG,QAAQ,GAAG;IACjD,OAAO;KAAE,GAAG,KAAK,KAAK;KAAO,SAAS;IAAM;IAC5C,UAAU;KAAE,GAAI,KAAK,KAAK,YAAY,CAAC;KAAI;KAAS,cAAc;IAAM;GAC1E;GACA,MAAM,KAAK,eAAe,OAAO;GACjC,IAAI,KAAK,IAAI,EAAE,GAAG;GAClB,KAAK,IAAI,EAAE;GACX,IAAI,KAAK,OAAO;EAClB;CACF;CACA,IAAI,IAAI,WAAW,GAGjB,MAAM,IAAI,gBACR,uDAAuD,UAAU,KAAK,IAAI,EAAE,aAAa,OAAO,KAAK,IAAI,EAAE,IAC7G;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,cACd,SACqD;CACrD,MAAM,IAAI,QAAQ;CAClB,MAAM,UAAU,GAAG;CACnB,MAAM,QAAQ,GAAG;CACjB,IAAI,OAAO,YAAY,YAAY,OAAO,UAAU,UAClD,OAAO;EAAW;EAAwB;CAAM;AAGpD;;;;;;;;AASA,SAAgB,eAAe,SAA+B;CAE5D,OAAO,GADO,qBAAqB,yBAAyB,OAAO,CAAC,KAAK,UACzD,GAAG,iBAAiB,OAAO,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1D;;;;;AAMA,SAAgB,oBAAoB,SAA+B;CACjE,MAAM,QAAQ,QAAQ,OAAO,SAAS,KAAK;CAC3C,IAAI,CAAC,OAAO;EACV,MAAM,QAAQ,yBAAyB,OAAO,KAAK;EACnD,MAAM,IAAI,gBACR,iBAAiB,MAAM,gDACzB;CACF;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,SAA2C;CAC3E,OAAO,QAAQ,MAAM,KAAK,KAAK,QAAQ,SAAS,KAAK,KAAK,KAAA;AAC5D;AAEA,SAAS,qBAAqB,OAA+C;CAM3E,OALa,OACT,KAAK,CAAC,CACP,QAAQ,qBAAqB,GAAG,CAAC,CACjC,QAAQ,OAAO,GAAG,CAAC,CACnB,QAAQ,UAAU,EAAE,KACR,KAAA;AACjB;AAEA,SAAS,QAA2C,OAAsB;CACxE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,iBAAiB,SAA+B;CAC9D,MAAM,QAAQ,oBAAoB,OAAO;CACzC,MAAM,YAAY,QAAQ;EACxB,GAAG;EACH,MAAM,KAAA;EACN,aAAa,KAAA;EACb,MAAM,QAAQ,OAAO,CAAC,GAAG,QAAQ,IAAI,CAAC,CAAC,KAAK,IAAI,KAAA;EAChD,OAAO,QAAQ;GAAE,GAAG,QAAQ;GAAO,SAAS;EAAM,CAAC;CACrD,CAAC;CACD,OAAO,WAAW,QAAQ,CAAC,CAAC,OAAO,gBAAgB,SAAS,CAAC,CAAC,CAAC,OAAO,KAAK;AAC7E;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACvEA,SAAgB,kBAAkB,OAMZ;CACpB,IAAI,MAAM,aAAa,WAAW,GAChC,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,+DAC3C;CAEF,MAAM,aAAa,MAAM,aAAa,QAAQ,MAAM,CAAC,EAAE,UAAU;CACjE,MAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;CAC/D,IAAI,WAAW,WAAW,GAGxB,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,yCAAyC,MAAM,aAAa,OAAO,8BAA8B,MAAM,aAAa,EAAE,EAAE,oBAAoB,iBAAiB,EACxM;CAEF,MAAM,iBAAiB,WAAW,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC;CAC7D,MAAM,iBAAiB,iBAAiB,WAAW;CAGnD,MAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;CAC7E,OAAO;EACL,QAAQ,MAAM;EACd,cAAc,MAAM;EACpB;EACA;EACA;EACA,OAAO;EACP,OAAO;EACP,GAAI,MAAM,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,MAAM,cAAc;CACpF;AACF;AA+BA,MAAM,4BAAY,IAAI,IAAI;CACxB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,MAAM,6CAA6B,IAAI,IAAI;CACzC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;AACxB,MAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;CAC/D,OAAO,IAAI,IACT,EACG,YAAY,CAAC,CACb,MAAM,YAAY,CAAC,CACnB,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAC1E;AACF;;;;;;;AAQA,SAAS,YAAY,iBAAyB,eAA+B;CAC3E,MAAM,MAAM,OAAO,iBAAiB,0BAA0B;CAC9D,IAAI,IAAI,SAAS,GAAG,OAAO;CAC3B,MAAM,OAAO,OAAO,aAAa;CACjC,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK,IAAI,KAAK,IAAI,CAAC,GAAG;CACtC,OAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,GAAG,MAAM;EAC1B,KAAK,EAAE,WAAW,GAAA,CAAI,KAAK,CAAC,CAAC,SAAS,mBAAmB;EAKzD,IAAI,QAAQ,YACV,SACA,GAAG,EAAE,QAAQ,GAAG,GAAG,EAAE,KAAK,IAAI,EAAE,WAAW,GAAA,CAAI,MAAM,GAAG,GAAI,GAC9D;EACA,IAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAC9E,QAAQ,KAAK,IAAI,OAAO,CAAC;EAE3B,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY;GACrB;GACA,UAAU,aAAa,EAAE,QAAQ,EAAE,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,CAAC;CACD,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,WAAW;EAEzB,IAAI,EAAE,WAAW,YAAY;EAM7B,MAAM,QAAQ,EAAE,WAAW,GAAA,CAAI,KAAK;EACpC,IAAI,KAAK,SAAS,mBAAmB;EASrC,MAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,MAAM,GAAG,MAAM;EACvD,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY,EAAE;GACvB;GACA,UAAU,sBAAsB,EAAE,MAAM,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS;EACX,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,MAAM,MAAM;EAC7B,MAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;EACzC,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,QAAQ;GACjB;GACA,UAAU,cAAc,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACxE,SAAS;EACX,CAAC;CACH,CAAC;CACD,OAAO;AACT;;;;;;;;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;CAC5B,IAAI,KAAK,aAAa,WAAW,GAC/B,MAAM,IAAI,MACR,2BAA2B,KAAK,OAAO,4CACzC;CAMF,MAAM,aAA0B,CAAC;CACjC,KAAK,aAAa,SAAS,KAAK,MAAM;EACpC,MAAM,KAAK,IAAI,eAAe;EAC9B,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,eAAe,OAAO,OAC/B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;CAElE,CAAC;CACD,WAAW,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;CAE3C,MAAM,2BAAW,IAAI,IAAuB;CAC5C,MAAM,4BAAY,IAAI,IAAY;CAClC,KAAK,MAAM,KAAK,YAAY;EAC1B,IAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,GAAG;EAC1D,SAAS,IAAI,EAAE,UAAU,CAAC;EAC1B,UAAU,IAAI,EAAE,OAAO;CACzB;CAEA,MAAM,eAAmC,CAAC;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;EACjD,MAAM,MAAM,KAAK,aAAa;EAC9B,MAAM,QAAQ,SAAS,IAAI,CAAC;EAC5B,MAAM,WAAqB,CAAC;EAC5B,IAAI,UAA0B;EAC9B,IAAI;EAEJ,IAAI,OAAO;GACT,SAAS,KAAK,MAAM,QAAQ;GAC5B,IAAI,MAAM,YAAY,MACpB,IAAI;IACF,MAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;IACnD,UAAU,EAAE;IACZ,SAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,OAAO,KAAK,EAAE,QAAQ;GAC3E,SAAS,KAAK;IAIZ,mBACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,MACjE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7E,SAAS,KAAK,6BAA6B,kBAAkB;GAC/D;QAEA,SAAS,KAAK,6DAA6D;EAE/E,OAAO;GACL,MAAM,KAAK,IAAI,eAAe;GAC9B,MAAM,OAAO,OAAO,QAAQ,gCAAgC;GAC5D,SAAS,KAAK,eAAe,KAAK,0BAA0B;EAC9D;EAEA,MAAM,sBAAsB,UAAU,KAAA;EACtC,MAAM,aAAa,qBAAqB,KAAA;EACxC,MAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;EACpE,aAAa,KAAK;GAChB,OAAO,IAAI;GACX,OAAO,IAAI;GACX;GACA;GACA;GACA,GAAI,aAAa;IAAE,YAAY;IAAe;GAAiB,IAAI,CAAC;GACpE;EACF,CAAC;CACH;CAMA,MAAM,cAAc,iBAAiB;CACrC,MAAM,gBAAkD,cACpD;EACE,UAAU,YAAY;EACtB,SAAS,YAAY;EACrB,aAAa,CACX,sHACA,GAAG,YAAY,WACjB;EACA,gBAAgB,4BAA4B;GAAE,QAAQ,KAAK;GAAQ;EAAa,CAAC;CACnF,IACA,KAAA;CAEJ,OAAO,kBAAkB;EACvB,QAAQ,KAAK;EACb;EACA,GAAI,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,cAAc;CACzD,CAAC;AACH;;;;;;;;AAgCA,SAAgB,yBAAyB,KAAmD;CAC1F,MAAM,eAAe,cAAoE;EACvF,IAAI,cAAc,QAAQ,OAAO,cAAc,UAAU,OAAO;EAChE,MAAM,EAAE,SAAS,WAAW;EAC5B,IAAI,OAAO,YAAY,WAAW,OAAO;EACzC,OAAO;GAAE;GAAS,QAAQ,OAAO,WAAW,WAAW,SAAS;EAAG;CACrE;CAEA,MAAM,QAAQ,IAAI,MAAM,aAAa;CACrC,IAAI,OACF,IAAI;EACF,MAAM,SAAS,YAAY,KAAK,MAAM,MAAM,EAAE,CAAC;EAC/C,IAAI,QAAQ,OAAO;CACrB,QAAQ,CAER;CAIF,MAAM,QAAQ,IAAI,QAAQ,GAAG;CAC7B,IAAI,UAAU,IAAI;EAChB,MAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;EACpE,IAAI,WAAW,OAAO;CACxB;CACA,MAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;;;;;;;AAQA,SAAgB,4BACd,MACA,OAAkC,CAAC,GACf;CACpB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,aAAa,KAAK,cAAc,IAAI,WAAW;CACrD,MAAM,OAAO,KAAK;CAClB,MAAM,SAAS,OAAO,UAA2C;EAE/D,IAAI;GACF,MAAM,MAAM,OAAO,KAAK;EAC1B,QAAQ,CAER;CACF;CACA,MAAM,UAA8B,OAAO,aAAa,YAAY;EAClE,MAAM,UAAU;GACd;GACA,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IACE,MAAM;IACN,SAAS,gBAAgB,YAAY,MAAM,IACzC,YAAY,WAAW,aAAa,YAAY,SAAS,MAAM,GAChE,wBAAwB,QAAQ,MAAM,GAAG,eAAe;GAC3D,CACF;GACA,aAAa;GACb,WAAW;EACb;EACA,IAAI;EACJ,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;GACtD,MAAM,UAAU,KAAK,IAAI;GACzB,MAAM,OAAO;IACX,SAAS,WAAW;IACpB,UAAU,KAAK;IACf;IACA,UAAU;IACV,SAAS;IACT,cAAc;IACd,WAAW;IACX,WAAW;IACX,aAAa;IACb,gBAAgB,CAAC;GACnB,CAAC;GACD,IAAI;IACF,MAAM,OAAO,MAAM,WAAW,YAAY;KACxC,SAAS;KACT,OAAO,KAAK,aAAa;KACzB,OAAO;KACP;KACA,eACE,KAAK,oBAAoB,KAAA,IACrB,KAAA,IACA,2BAA2B,SAAS,EAAE,iBAAiB,KAAK,gBAAgB,CAAC;KACnF,MAAM;MACJ,GAAG,KAAK;MACR,eAAe,YAAY;MAC3B,SAAS,OAAO,OAAO;KACzB;KACA,QAAQ,KAAK;KACb,UAAU,QAAQ,WAAW,KAAK,KAAK,SAAS;MAAE;MAAQ,gBAAgB;KAAO,CAAC;KAClF,SAAS;KACT,kBAAkB;IACpB,CAAC;IACD,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;IAChC,MAAM,OAAO,KAAK;IAMlB,kBAAkB,OAAO,KAAK,aAAa;KACzC,iBAAiB;KACjB,SAAS,uCAAuC,YAAY;IAC9D,CAAC;IACD,MAAM,MAAM,KAAK;IACjB,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc;KACd,gBAAgB,CAAC;IACnB,CAAC;IACD,OAAO,yBAAyB,GAAG;GACrC,SAAS,KAAK;IACZ,UAAU;IACV,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAGD,IAAI,eAAe,wBAAwB,MAAM;GACnD;EACF;EACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;CACtE;CACA,QAAQ,cAAc;EACpB,UAAU;EACV,SAAS;GAAE,MAAM;GAA2B,SAAS;EAAM;EAC3D,aAAa,CACX,uGACF;CACF;CACA,OAAO;AACT;;;ACnmBA,SAAS,aAAa,UAAsC;CAC1D,IAAI,CAAC,UAAU,OAAO;CACtB,IAAI,SAAS,SAAS,MAAM,GAAG,OAAO;CACtC,IAAI,SAAS,WAAW,OAAO,GAAG,OAAO;CACzC,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,qBAAqB,QAAoD;CACvF,MAAM,4BAAY,IAAI,IAAsB;CAC5C,MAAM,4BAAY,IAAI,IAA8B;CACpD,MAAM,YAAsB,CAAC;CAC7B,MAAM,4BAAY,IAAI,IAAY;CAElC,KAAK,MAAM,MAAM,QACf,IAAI,GAAG,SAAS,aAAa;EAC3B,MAAM,OAAQ,GAAyB;EACvC,IAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;GAChC,UAAU,IAAI,IAAI;GAClB,UAAU,KAAK,IAAI;EACrB;CACF,OAAO,IAAI,GAAG,SAAS,YAAY;EACjC,MAAM,IAAI;EACV,MAAM,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE;EAClC,UAAU,IAAI,MAAM;GAClB,MAAM,aAAa,EAAE,QAAQ;GAC7B;GACA,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,OAAO,IAAI,GAAG,SAAS,oBAAoB;EACzC,MAAM,IAAI;EACV,UAAU,IAAI,EAAE,YAAY;GAC1B,IAAI,EAAE;GACN,OAAO,EAAE;GACT,QAAQ,EAAE,UAAU;GACpB,GAAI,EAAE,YAAY,KAAA,IAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;EAC1D,CAAC;CACH;CAGF,OAAO;EAAE,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG;CAAU;AAC7F"}

package/dist/proposal-findings-BZiMTzAF.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"proposal-findings-BZiMTzAF.js","names":[],"sources":["../src/analyst/proposal-findings.ts"],"sourcesContent":["import { ProposalFindingSchema } from './finding-scope'\nimport type { ProposalFinding } from './types'\n\n/** True when a finding names a source candidate generation may learn from. */\nexport function isProposalFinding(finding: unknown): finding is ProposalFinding {\n  return ProposalFindingSchema.safeParse(finding).success\n}\n\n/**\n * Reject findings whose source has not been explicitly admitted for candidate\n * generation. Search feedback and observed production behavior are allowed;\n * final evaluation data has no allowed origin.\n */\nexport function assertProposalFindings(\n  findings: unknown,\n  context = 'proposal findings',\n): ReadonlyArray<ProposalFinding> {\n  if (!Array.isArray(findings)) {\n    throw new TypeError(`${context}: expected an array`)\n  }\n  const rejected = findings.flatMap((finding, index) =>\n    isProposalFinding(finding) ? [] : [findingLabel(finding, index)],\n  )\n  if (rejected.length > 0) {\n    throw new Error(\n      `${context}: every finding must match AnalystFinding and declare ` +\n        `proposal_origin as search or production. ` +\n        `Rejected findings: [${rejected.join(', ')}].`,\n    )\n  }\n  return findings as ReadonlyArray<ProposalFinding>\n}\n\nfunction findingLabel(finding: unknown, index: number): string {\n  if (typeof finding !== 'object' || finding === null) return `index ${index}`\n  const id = (finding as { finding_id?: unknown }).finding_id\n  return typeof id === 'string' && id.length > 0 ? id : `index ${index}`\n}\n"],"mappings":";;;AAIA,SAAgB,kBAAkB,SAA8C;CAC9E,OAAO,sBAAsB,UAAU,OAAO,CAAC,CAAC;AAClD;;;;;;AAOA,SAAgB,uBACd,UACA,UAAU,qBACsB;CAChC,IAAI,CAAC,MAAM,QAAQ,QAAQ,GACzB,MAAM,IAAI,UAAU,GAAG,QAAQ,oBAAoB;CAErD,MAAM,WAAW,SAAS,SAAS,SAAS,UAC1C,kBAAkB,OAAO,IAAI,CAAC,IAAI,CAAC,aAAa,SAAS,KAAK,CAAC,CACjE;CACA,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,MACR,GAAG,QAAQ,qHAEc,SAAS,KAAK,IAAI,EAAE,GAC/C;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,OAAuB;CAC7D,IAAI,OAAO,YAAY,YAAY,YAAY,MAAM,OAAO,SAAS;CACrE,MAAM,KAAM,QAAqC;CACjD,OAAO,OAAO,OAAO,YAAY,GAAG,SAAS,IAAI,KAAK,SAAS;AACjE"}

package/dist/query-B1D1qp2Z.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"query-B1D1qp2Z.js","names":[],"sources":["../src/trace/query.ts"],"sourcesContent":["/**\n * Typed query helpers over TraceStore.\n *\n * Not a full SQL engine — a minimal, composable set of operators that\n * cover the canned-pipeline use cases. For ad-hoc analytics, persist to\n * NDJSON and point DuckDB at it; the schema is stable so external SQL\n * tooling works out of the box.\n */\n\nimport { ValidationError } from '../errors'\nimport { canonicalString } from '../ledger-core/canonical'\nimport type { FailureClass, JudgeSpan, LlmSpan, Run, ToolSpan } from './schema'\nimport { isJudgeSpan, isLlmSpan, isToolSpan } from './schema'\nimport type { TraceStore } from './store'\n\nexport async function runsForScenario(store: TraceStore, scenarioId: string): Promise<Run[]> {\n  return store.listRuns({ scenarioId })\n}\n\nexport async function llmSpans(store: TraceStore, runId?: string): Promise<LlmSpan[]> {\n  const spans = await store.spans({ runId, kind: 'llm' })\n  return spans.filter(isLlmSpan)\n}\n\nexport async function toolSpans(\n  store: TraceStore,\n  runId?: string,\n  toolName?: string,\n): Promise<ToolSpan[]> {\n  const spans = await store.spans({ runId, kind: 'tool', toolName })\n  return spans.filter(isToolSpan)\n}\n\n/** Query judge-kind spans from the trace store, optionally scoped to a single run. */\nexport async function judgeSpans(store: TraceStore, runId?: string): Promise<JudgeSpan[]> {\n  const spans = await store.spans({ runId, kind: 'judge' })\n  return spans.filter(isJudgeSpan)\n}\n\n/** Group spans by any key selector. */\nexport function groupBy<T, K extends string | number>(items: T[], key: (t: T) => K): Map<K, T[]> {\n  const map = new Map<K, T[]>()\n  for (const item of items) {\n    const k = key(item)\n    let bucket = map.get(k)\n    if (!bucket) {\n      bucket = []\n      map.set(k, bucket)\n    }\n    bucket.push(item)\n  }\n  return map\n}\n\n/**\n * Key tool arguments for de-duplication: RFC 8785 canonical JSON, so key\n * order cannot split one call into two keys. Uncaptured args (`undefined`)\n * key to the string `'undefined'` so every call still keys to a string; an\n * args value with no canonical JSON form (a nested `undefined`, a function, a\n * non-finite number) throws `LedgerCanonicalizationError`.\n */\nexport function argHash(args: unknown): string {\n  if (args === undefined) return 'undefined'\n  return canonicalString(args)\n}\n\n/** Whether argument-based comparisons are valid for this tool call. */\nexport function hasCapturedToolArgs(span: ToolSpan): boolean {\n  return span.argsCaptured !== false\n}\n\n/** Sum an LLM-span array into aggregate token + cost. */\nexport function aggregateLlm(spans: LlmSpan[]): {\n  inputTokens: number\n  outputTokens: number\n  cachedTokens: number\n  cacheWriteTokens: number\n  reasoningTokens: number\n  costUsd: number\n} {\n  return spans.reduce(\n    (acc, s) => ({\n      inputTokens: acc.inputTokens + (s.inputTokens ?? 0),\n      outputTokens: acc.outputTokens + (s.outputTokens ?? 0),\n      cachedTokens: acc.cachedTokens + (s.cachedTokens ?? 0),\n      cacheWriteTokens: acc.cacheWriteTokens + (s.cacheWriteTokens ?? 0),\n      reasoningTokens: acc.reasoningTokens + (s.reasoningTokens ?? 0),\n      costUsd: acc.costUsd + (s.costUsd ?? 0),\n    }),\n    {\n      inputTokens: 0,\n      outputTokens: 0,\n      cachedTokens: 0,\n      cacheWriteTokens: 0,\n      reasoningTokens: 0,\n      costUsd: 0,\n    },\n  )\n}\n\n/** Pick the outcome's failure class when present, else derive 'success' from run status. */\nexport function runFailureClass(run: Run): FailureClass {\n  if (run.outcome?.failureClass) return run.outcome.failureClass\n  if (run.status === 'completed' && run.outcome?.pass !== false) return 'success'\n  if (run.status === 'aborted') return 'budget_exceeded'\n  return 'unknown'\n}\n\n/**\n * Metrics `regressionView`, `correlationStudy`, and `calibrationCurve` can\n * read from a run without a caller-supplied extractor. The type derives from\n * this array, so a new metric cannot be declared without an extractor arm.\n */\nexport const RUN_METRICS = [\n  'score',\n  'overallScore',\n  'pass',\n  'durationMs',\n  'costUsd',\n  'inputTokens',\n  'outputTokens',\n  'failureClass',\n] as const\n\nexport type RunMetric = (typeof RUN_METRICS)[number]\n\nfunction isRunMetric(metric: string): metric is RunMetric {\n  return (RUN_METRICS as readonly string[]).includes(metric)\n}\n\n/**\n * The extractor for one built-in metric. `null` means this run carries no\n * value for the metric — the caller drops that run from the sample.\n *\n * Throws `ValidationError` on a metric name this package does not define:\n * an unrecognized name would otherwise read as \"every run is missing this\n * value\" and produce an empty study instead of a refusal.\n */\nexport function runMetricExtractor(\n  metric: string,\n): (run: Run, store: TraceStore) => Promise<number | null> {\n  if (!isRunMetric(metric)) {\n    throw new ValidationError(\n      `unknown run metric '${metric}' — pass an \\`extract\\` function or use one of: ${RUN_METRICS.join(', ')}`,\n    )\n  }\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd':\n        return aggregateLlm(await llmSpans(store, run.runId)).costUsd\n      case 'inputTokens':\n        return aggregateLlm(await llmSpans(store, run.runId)).inputTokens\n      case 'outputTokens':\n        return aggregateLlm(await llmSpans(store, run.runId)).outputTokens\n      case 'failureClass':\n        return runFailureClass(run) === 'success' ? 1 : 0\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;AAeA,eAAsB,gBAAgB,OAAmB,YAAoC;CAC3F,OAAO,MAAM,SAAS,EAAE,WAAW,CAAC;AACtC;AAEA,eAAsB,SAAS,OAAmB,OAAoC;CAEpF,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;CAAM,CAAC,EAAA,CACzC,OAAO,SAAS;AAC/B;AAEA,eAAsB,UACpB,OACA,OACA,UACqB;CAErB,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;EAAQ;CAAS,CAAC,EAAA,CACpD,OAAO,UAAU;AAChC;;AAGA,eAAsB,WAAW,OAAmB,OAAsC;CAExF,QAAO,MADa,MAAM,MAAM;EAAE;EAAO,MAAM;CAAQ,CAAC,EAAA,CAC3C,OAAO,WAAW;AACjC;;AAGA,SAAgB,QAAsC,OAAY,KAA+B;CAC/F,MAAM,sBAAM,IAAI,IAAY;CAC5B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,IAAI,IAAI,IAAI;EAClB,IAAI,SAAS,IAAI,IAAI,CAAC;EACtB,IAAI,CAAC,QAAQ;GACX,SAAS,CAAC;GACV,IAAI,IAAI,GAAG,MAAM;EACnB;EACA,OAAO,KAAK,IAAI;CAClB;CACA,OAAO;AACT;;;;;;;;AASA,SAAgB,QAAQ,MAAuB;CAC7C,IAAI,SAAS,KAAA,GAAW,OAAO;CAC/B,OAAO,gBAAgB,IAAI;AAC7B;;AAGA,SAAgB,oBAAoB,MAAyB;CAC3D,OAAO,KAAK,iBAAiB;AAC/B;;AAGA,SAAgB,aAAa,OAO3B;CACA,OAAO,MAAM,QACV,KAAK,OAAO;EACX,aAAa,IAAI,eAAe,EAAE,eAAe;EACjD,cAAc,IAAI,gBAAgB,EAAE,gBAAgB;EACpD,cAAc,IAAI,gBAAgB,EAAE,gBAAgB;EACpD,kBAAkB,IAAI,oBAAoB,EAAE,oBAAoB;EAChE,iBAAiB,IAAI,mBAAmB,EAAE,mBAAmB;EAC7D,SAAS,IAAI,WAAW,EAAE,WAAW;CACvC,IACA;EACE,aAAa;EACb,cAAc;EACd,cAAc;EACd,kBAAkB;EAClB,iBAAiB;EACjB,SAAS;CACX,CACF;AACF;;AAGA,SAAgB,gBAAgB,KAAwB;CACtD,IAAI,IAAI,SAAS,cAAc,OAAO,IAAI,QAAQ;CAClD,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO,OAAO;CACtE,IAAI,IAAI,WAAW,WAAW,OAAO;CACrC,OAAO;AACT;;;;;;AAOA,MAAa,cAAc;CACzB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAIA,SAAS,YAAY,QAAqC;CACxD,OAAQ,YAAkC,SAAS,MAAM;AAC3D;;;;;;;;;AAUA,SAAgB,mBACd,QACyD;CACzD,IAAI,CAAC,YAAY,MAAM,GACrB,MAAM,IAAI,gBACR,uBAAuB,OAAO,kDAAkD,YAAY,KAAK,IAAI,GACvG;CAEF,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WACH,OAAO,aAAa,MAAM,SAAS,OAAO,IAAI,KAAK,CAAC,CAAC,CAAC;GACxD,KAAK,eACH,OAAO,aAAa,MAAM,SAAS,OAAO,IAAI,KAAK,CAAC,CAAC,CAAC;GACxD,KAAK,gBACH,OAAO,aAAa,MAAM,SAAS,OAAO,IAAI,KAAK,CAAC,CAAC,CAAC;GACxD,KAAK,gBACH,OAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;EACpD;CACF;AACF"}

package/dist/redact-Cveoq6fJ.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"redact-Cveoq6fJ.d.ts","names":[],"sources":["../src/trace/redact.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAsCY;cAEC,6BAA6B;;KAG9B;;;;;;cAOC;UAEI;;EAEf;EACA,UAAU;EACV;EACA;;UAGe;EACf;EACA,SAAS;;EAET;EACA,YAAY;EACZ;EACA,UAAU;;UAGK;;EAEf,UAAU;;EAEV;;;;;;;;;EASA;;EAEA,SAAS;;UAGM;;EAEf,UAAU;;EAEV;;KAGU;UAEK;EACf,UAAU;EACV;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;EACR,SAAS;EACT,UAAU;;EAEV;;;;;;;iBA6Dc,aAAa;;;;;;iBAmNb,YAAY,cAAc;;iBAyN1B,iBAAiB;;iBAQjB,mBAAmB;;;;;;iBAuYnB,OAAO,GACrB,OAAO,GACP,UAAS;EACN,OAAO;EAAG,QAAQ;;;;;;iBAUP,WAAW,cAAc,UAAS;;iBAMlC,qBAAqB,UAAS,mBAA+B;;;;;;;iBA+J7D,kBAAkB,gBAAgB,UAAS,gBAAqB;;;;;iBAoBhE,gBACd,SAAS,kBACT,mBAAmB,sBACnB,iCACC;;iBAsCa,aAAa,SAAS;;;;;;iBAStB,eAAe,GAC7B,OAAO,GACP,UAAS;EACN,OAAO;EAAG,QAAQ;EAAiB,SAAS"}

package/dist/redact-Dk1ZdJEm.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"redact-Dk1ZdJEm.js","names":[],"sources":["../src/trace/redact.ts"],"sourcesContent":["/**\n * The redaction core: the one place that decides what must not leave a\n * process or a machine. Every trace writer, exporter, uploader and share path\n * calls it.\n *\n * Three profiles, each a superset of the one before:\n *   - `default` removes credentials, found by key name and by value shape, and\n *     personal data (email addresses, card numbers, SSNs, phone numbers).\n *     Content is kept. Strings are capped at 1 MiB.\n *   - `share` also removes phone, IP and postal-address fields, replaces\n *     account, user, tenant, session and request identifiers with pseudonyms\n *     that stay consistent within one call, and caps strings at 64 KiB.\n *   - `strict` also removes prompt, completion, message, tool payload and other\n *     raw-content fields and embedded media, and caps strings at 4 KiB.\n *\n * A string that holds a credential is replaced whole with\n * `[REDACTED:<detector>]`. Replacing only the matched span can leave part of a\n * secret behind when a pattern's boundary is wrong. Personal data inside prose\n * is replaced in place, so the rest of the text stays readable.\n *\n * `assessShareSafety` reads a value without changing it and returns SAFE,\n * SAFE_WITH_WARNINGS, UNSAFE or UNKNOWN. UNKNOWN means the scanner could not\n * read part of the value (a cycle, binary data, nesting past the depth limit);\n * it is never reported as safe. `redactForShare` redacts and then assesses the\n * redacted output, so a detector gap shows up as UNSAFE instead of a leak.\n *\n * Reports and findings carry paths, categories and detector ids, never the\n * values they matched.\n *\n * The key classifier and the credential detector list are adapted from\n * agent-inspect (MIT, commit 3c3cbeda). `scripts/redaction-corpus.json` holds\n * the must-flag and must-not-flag cases, including the copied agent-inspect\n * safety corpus with its MIT notice; `pnpm redaction:corpus` runs this module\n * over every case.\n */\n\nimport { createHmac, randomBytes } from 'node:crypto'\n\nexport type RedactionProfile = 'default' | 'share' | 'strict'\n\nexport const REDACTION_PROFILES: readonly RedactionProfile[] = ['default', 'share', 'strict']\n\n/** What a finding is about. `media` is embedded binary content (a `data:` URI) no text detector can read. */\nexport type SafetyCategory = 'credential' | 'personal-data' | 'identifier' | 'raw-content' | 'media'\n\n/**\n * Version of the detector set and profile rules. Stored work that depends on\n * redaction behavior (saved optimizer inputs, per-span redaction stamps) keys\n * on it, so bump it whenever a detector, key list, cap or marker changes.\n */\nexport const REDACTION_VERSION = '2.4.0'\n\nexport interface RedactionFinding {\n  /** JSON Pointer (RFC 6901) to the value or key that was changed. */\n  path: string\n  category: SafetyCategory\n  detector: string\n  action: 'redacted' | 'pseudonymized' | 'truncated'\n}\n\nexport interface RedactionReport {\n  version: string\n  profile: RedactionProfile\n  /** Values and keys replaced, pseudonymized or truncated. */\n  redactionCount: number\n  byDetector: Record<string, number>\n  truncatedCount: number\n  findings: RedactionFinding[]\n}\n\nexport interface RedactOptions {\n  /** Default `default`. */\n  profile?: RedactionProfile\n  /** Lower the profile's per-string cap. A value above the profile cap is ignored. */\n  maxStringBytes?: number\n  /**\n   * Exact values the caller knows are secret (an API key the run was given, a\n   * password from the environment), whatever shape they have. Each occurrence\n   * in a string, as written or base64, base64url or URL-encoded, becomes\n   * `[REDACTED:known-secret]` and the text around it is kept. A key that\n   * contains one, and a whole-string base64 payload that holds one at any byte\n   * offset, is replaced whole.\n   */\n  knownSecrets?: readonly string[]\n  /** Add this call's findings to an existing report instead of a new one. */\n  report?: RedactionReport\n}\n\nexport interface AssessOptions {\n  /** Default `default`. */\n  profile?: RedactionProfile\n  /** As in `RedactOptions`: any surviving occurrence makes the verdict UNSAFE. */\n  knownSecrets?: readonly string[]\n}\n\nexport type ShareSafetyStatus = 'SAFE' | 'SAFE_WITH_WARNINGS' | 'UNSAFE' | 'UNKNOWN'\n\nexport interface ShareSafetyFinding {\n  category: SafetyCategory\n  detector: string\n  severity: 'error' | 'warning'\n  count: number\n  /** Up to five JSON Pointers where the detector fired. */\n  paths: string[]\n}\n\nexport interface ShareSafetyVerdict {\n  status: ShareSafetyStatus\n  profile: RedactionProfile\n  findings: ShareSafetyFinding[]\n  /** Parts the scanner could not read, as `<path>: <reason>`. Any entry makes the status UNKNOWN. */\n  unreadable: string[]\n}\n\n// ── Profiles ─────────────────────────────────────────────────────────────\n\ninterface ProfileRules {\n  maxStringBytes: number\n  /** Key categories whose values are removed or pseudonymized. */\n  redactKeys: ReadonlySet<'share-personal' | 'identifier' | 'raw-content'>\n  redactMedia: boolean\n  severity: Record<SafetyCategory, 'error' | 'warning' | undefined>\n}\n\nconst KIB = 1024\n\nconst PROFILE_RULES: Record<RedactionProfile, ProfileRules> = {\n  default: {\n    maxStringBytes: 1024 * KIB,\n    redactKeys: new Set(),\n    redactMedia: false,\n    severity: {\n      credential: 'error',\n      'personal-data': 'warning',\n      identifier: undefined,\n      'raw-content': 'warning',\n      media: 'warning',\n    },\n  },\n  share: {\n    maxStringBytes: 64 * KIB,\n    redactKeys: new Set(['share-personal', 'identifier']),\n    redactMedia: false,\n    severity: {\n      credential: 'error',\n      'personal-data': 'error',\n      identifier: 'warning',\n      'raw-content': 'warning',\n      media: 'warning',\n    },\n  },\n  strict: {\n    maxStringBytes: 4 * KIB,\n    redactKeys: new Set(['share-personal', 'identifier', 'raw-content']),\n    redactMedia: true,\n    severity: {\n      credential: 'error',\n      'personal-data': 'error',\n      identifier: 'warning',\n      'raw-content': 'error',\n      media: 'error',\n    },\n  },\n}\n\n// ── Key classifier ───────────────────────────────────────────────────────\n\n/**\n * Canonical snake_case form of a field name, so camelCase, kebab-case, dotted\n * and snake_case spellings match one list: `userPassword`, `user-password` and\n * `user.password` all become `user_password`.\n */\nexport function normalizeKey(key: string): string {\n  return key\n    .replace(/([a-z0-9])([A-Z])/g, '$1_$2')\n    .toLowerCase()\n    .replace(/[^a-z0-9]+/g, '_')\n    .replace(/^_+|_+$/g, '')\n}\n\n/** Counts, limits and model vocabulary that contain the word \"token\" but are not credentials. */\nconst NON_CREDENTIAL_TOKEN_KEYS: ReadonlySet<string> = new Set([\n  'tokens',\n  'max_tokens',\n  'min_tokens',\n  'ls_max_tokens',\n  'token_count',\n  'token_limit',\n  'token_budget',\n  'input_tokens',\n  'output_tokens',\n  'total_tokens',\n  'cached_tokens',\n  'prompt_tokens',\n  'completion_tokens',\n  'reasoning_tokens',\n  'bos_token',\n  'eos_token',\n  'pad_token',\n  'unk_token',\n  'sep_token',\n  'cls_token',\n  'mask_token',\n  'stop_token',\n  // Pagination cursors are opaque but not credentials.\n  'page_token',\n  'next_page_token',\n  'prev_page_token',\n  'next_token',\n  'continuation_token',\n  'cursor_token',\n  'sync_token',\n])\n\nconst CREDENTIAL_KEYS = [\n  'authorization',\n  'auth',\n  'cookie',\n  'api_key',\n  'apikey',\n  'api_token',\n  'access_key',\n  'secret_key',\n  'private_key',\n  'signing_key',\n  'encryption_key',\n  'master_key',\n  'session_key',\n  'client_secret',\n  'secret',\n  'password',\n  'passwd',\n  'passphrase',\n  'credential',\n  'credentials',\n  'bearer',\n  'dsn',\n  'database_url',\n  'connection_string',\n  'pat',\n] as const\n\n/** Email fields are personal data in every profile. */\nconst EMAIL_KEYS = ['email', 'email_address'] as const\n\n/** Personal-data fields removed by the share and strict profiles. */\nconst SHARE_PERSONAL_KEYS = [\n  'phone',\n  'phone_number',\n  'mobile',\n  'ip',\n  'ip_address',\n  'remote_addr',\n  'street_address',\n  'postal_address',\n  'home_address',\n  'billing_address',\n  'shipping_address',\n  'mailing_address',\n  'date_of_birth',\n  'dob',\n  'ssn',\n] as const\n\n/** Identifiers pseudonymized by the share and strict profiles. */\nconst IDENTIFIER_KEYS = [\n  'user_id',\n  'user_uuid',\n  'account_id',\n  'account_uuid',\n  'customer_id',\n  'tenant_id',\n  'org_id',\n  'organization_id',\n  'organisation_id',\n  'team_id',\n  'enduser_id',\n  'actor_id',\n  'principal_id',\n  'member_id',\n  'session_id',\n  'request_id',\n  'correlation_id',\n  'device_id',\n  'installation_id',\n] as const\n\n/** Fields that carry prompts, model output, tool payloads or other free text. Removed by `strict`. */\nconst RAW_CONTENT_KEYS = [\n  'prompt',\n  'prompts',\n  'system_prompt',\n  'completion',\n  'input',\n  'output',\n  'message',\n  'messages',\n  'content',\n  'text',\n  'thinking',\n  'reasoning',\n  'transcript',\n  'context',\n  'document',\n  'documents',\n  'chunk',\n  'chunks',\n  'retrieval',\n  'query',\n  'args',\n  'arguments',\n  'result',\n  'body',\n  'command',\n  'full_command',\n  'input_value',\n  'output_value',\n  'instructions',\n  'system_instructions',\n  'stdout',\n  'stderr',\n  'stacktrace',\n  'status_description',\n  'input_preview',\n  'output_preview',\n  'current_task',\n  'user_input',\n  'request_text',\n] as const\n\n/**\n * A key matches a list entry exactly or as a suffix (`openai_api_key`,\n * `db_password`, `tool_output`). Prefix compounds (`secret_name`,\n * `password_min_length`, `auth_type`, `content_type`) do not match: they name\n * or describe a value rather than hold it, and flagging them would make the\n * share verdict refuse ordinary traces. Value detectors still catch a\n * credential-shaped value under any key.\n */\nfunction matchesKeyList(normalized: string, keys: readonly string[]): boolean {\n  for (const key of keys) {\n    if (normalized === key || normalized.endsWith(`_${key}`)) return true\n  }\n  return false\n}\n\nfunction isTokenCredentialKey(normalized: string): boolean {\n  if (NON_CREDENTIAL_TOKEN_KEYS.has(normalized)) return false\n  if (normalized === 'token') return true\n  // access_token and refreshToken are credentials; max_tokens and input_tokens are counts.\n  if (normalized.endsWith('tokens')) return false\n  return normalized.endsWith('token')\n}\n\ntype KeyClass = 'credential' | 'email' | 'share-personal' | 'identifier' | 'raw-content'\n\n/** Trace keys repeat on every span, so classifications are cached. */\nconst keyClassCache = new Map<string, KeyClass | null>()\n\nfunction classifyKeyName(key: string): KeyClass | undefined {\n  const cached = keyClassCache.get(key)\n  if (cached !== undefined) return cached ?? undefined\n  const kind = classifyNormalizedKey(normalizeKey(key))\n  if (keyClassCache.size >= 10_000) keyClassCache.clear()\n  keyClassCache.set(key, kind ?? null)\n  return kind\n}\n\nfunction classifyNormalizedKey(normalized: string): KeyClass | undefined {\n  if (!normalized || NON_CREDENTIAL_TOKEN_KEYS.has(normalized)) return undefined\n  if (isTokenCredentialKey(normalized)) return 'credential'\n  if (matchesKeyList(normalized, CREDENTIAL_KEYS)) return 'credential'\n  if (matchesKeyList(normalized, EMAIL_KEYS)) return 'email'\n  if (matchesKeyList(normalized, SHARE_PERSONAL_KEYS)) return 'share-personal'\n  if (matchesKeyList(normalized, IDENTIFIER_KEYS)) return 'identifier'\n  if (matchesKeyList(normalized, RAW_CONTENT_KEYS)) return 'raw-content'\n  return undefined\n}\n\n/**\n * The safety category a field name puts its value in, or undefined for an\n * ordinary field. Token counts and model limits (`inputTokens`, `max_tokens`,\n * `token_count`) and names such as `author` are not credentials.\n */\nexport function classifyKey(key: string): SafetyCategory | undefined {\n  const kind = classifyKeyName(key)\n  if (kind === undefined) return undefined\n  if (kind === 'credential') return 'credential'\n  if (kind === 'identifier') return 'identifier'\n  if (kind === 'raw-content') return 'raw-content'\n  return 'personal-data'\n}\n\n// ── Value detectors ──────────────────────────────────────────────────────\n\ninterface ValueDetector {\n  id: string\n  pattern: RegExp\n}\n\n/**\n * High-confidence credential shapes. A hit replaces the whole string. Each\n * pattern is written without the `g` flag so `test` stays stateless.\n *\n * The provider-branded prefixes below (`sk-ant-`, `AIza`, `ghp_`, `xoxb-`,\n * `AKIA`, `hubcap_`, `eyJ`…) have no leading `\\b`. A `\\b` word-boundary\n * assertion does not fire between two word characters, so it silently passed\n * a real token when something else glued a word character directly in front\n * of it — a compound identifier like `env.GITHUB_TOKEN_ghp_<token>` (a `_`\n * before `ghp_`) or `BACKUP_AKIA<key>` (a `_` before `AKIA`). These prefixes\n * are distinctive enough on their own, so scanning without that boundary\n * catches the same real credential wherever it is glued, at the cost of no\n * meaningful new false positives (`scripts/redaction-corpus.json` covers\n * both).\n */\nconst CREDENTIAL_DETECTORS: readonly ValueDetector[] = [\n  { id: 'private-key', pattern: /-----BEGIN [A-Z0-9 ]*PRIVATE KEY-----/ },\n  { id: 'jwt', pattern: /eyJ[A-Za-z0-9_-]{8,}\\.[A-Za-z0-9_-]{8,}\\.[A-Za-z0-9_-]{8,}/ },\n  { id: 'bearer', pattern: /\\bbearer\\s+[A-Za-z0-9._~+/=-]{12,}/i },\n  {\n    id: 'authorization-header',\n    pattern: /^\\s*(?:basic|digest|apikey)\\s+[A-Za-z0-9._~+/=:-]{8,}\\s*$/i,\n  },\n  { id: 'anthropic-key', pattern: /sk-ant-[A-Za-z0-9_-]{16,}/ },\n  { id: 'provider-key', pattern: /sk-(?:proj-|tan-)?[A-Za-z0-9_-]{16,}/ },\n  { id: 'stripe-key', pattern: /(?:sk|rk)_(?:live|test)_[A-Za-z0-9_]{10,}/ },\n  { id: 'google-api-key', pattern: /AIza[0-9A-Za-z_-]{20,}/ },\n  { id: 'github-token', pattern: /(?:gh[pousr]_[A-Za-z0-9_]{20,}|github_pat_[A-Za-z0-9_]{20,})/ },\n  { id: 'slack-token', pattern: /xox[abeoprs]-[A-Za-z0-9-]{10,}/ },\n  { id: 'aws-access-key', pattern: /(?:AKIA|ASIA)[A-Z0-9]{16}\\b/ },\n  { id: 'tangle-capability', pattern: /(?:hubcap_|hct_)[A-Za-z0-9_=-]{8,}/ },\n  {\n    // Scheme, userinfo user and password each bounded (32/512/512 chars): the\n    // unbounded `*`/`+` here made `detectCredential` catastrophically slow on\n    // adversarial text with no `://` at all — 256 KiB of `\"a.\"` took ~59s,\n    // measured before this fix (scratchpad/rv/before.ts). A real URL's scheme\n    // and userinfo are always far under these bounds.\n    id: 'url-credentials',\n    pattern: /\\b[a-z][a-z0-9+.-]{0,31}:\\/\\/[^\\s/:@]{1,512}:[^\\s/@]{1,512}@/i,\n  },\n  {\n    id: 'url-secret-param',\n    pattern:\n      /[?&](?:access_token|api_key|apikey|auth|token|secret|password|sig|signature|x-amz-signature|x-goog-signature)=[^&\\s\"'#)]{8,}/i,\n  },\n  {\n    // `GH_TOKEN=…`, `--api-key=…`, `\"client_secret\": \"…\"`, `password: …` in free\n    // text. The value must mix letters and digits and be at least 12\n    // characters, so prose (`token: string`), variable references (`$TOKEN`)\n    // and dotenvx ciphertext (`encrypted:…`) pass. The pattern starts at the\n    // key word itself, with no identifier prefix, so it scans in linear time.\n    // The two mixed-content lookaheads and the value itself are bounded to\n    // 512 chars: unbounded, `'token='` repeated took ~20s at 256 KiB and\n    // `'password:x'` repeated took ~12s at 256 KiB (measured before this fix,\n    // scratchpad/rv/before.ts). A real credential value is always far under\n    // 512 chars.\n    id: 'secret-assignment',\n    pattern:\n      /(?:api[_-]?key|apikey|access[_-]?key|secret[_-]?key|private[_-]?key|secret|token|password|passwd|passphrase|credentials?)[\"']?[ \\t]*[:=][ \\t]*[\"']?(?!encrypted:)(?![$`<[{(%])(?=[^\\s\"'`,;&|]{0,512}[0-9])(?=[^\\s\"'`,;&|]{0,512}[A-Za-z])[^\\s\"'`,;&|]{12,512}/i,\n  },\n  {\n    // `{\"password\": \"hunter2\"}` or `'client_secret': '…'` inside serialized\n    // JSON, YAML or a log line. A quoted credential field name followed by a\n    // quoted value holds a secret whatever the value's shape or length. Bare\n    // `\"token\"` is left out: model logprob output uses it for text tokens.\n    // Placeholders (`[REDACTED…]`, `$VAR`, `{env:VAR}`, `***`, `<value>`) and\n    // prose with two or more spaces (a field description) do not match.\n    id: 'json-secret',\n    pattern:\n      /[\"'][A-Za-z0-9_.-]{0,40}?(?:api[_-]?key|apikey|access[_-]?key|secret[_-]?key|private[_-]?key|client[_-]?secret|secret|password|passwd|passphrase|credentials?|(?:access|refresh|id|auth|session|bearer|api|github|gh|slack|npm)[_-]?token)[\"'][ \\t]*:[ \\t]*[\"'](?!\\[REDACTED|[$*{<])(?![^\"'\\r\\n]* [^\"'\\r\\n]* )[^\"'\\r\\n]{3,}[\"']/i,\n  },\n  {\n    // `DB_PASSWORD=CorrectHorseBatteryStaple`, `PGPASSWORD=postgres`: unlike\n    // `secret-assignment` above, the all-upper-case env-var name ending in a\n    // credential word is itself the signal, so a letters-only or short real\n    // password still matches (no digit-mixing or 12-char minimum). Requires\n    // the name to be immediately followed by `=`, so `MAX_TOKENS=16` and\n    // `TOKEN_EXPIRY=…` (the word isn't the end of the name) do not match.\n    id: 'env-secret-assignment',\n    pattern:\n      /\\b[A-Z][A-Z0-9_]*?(?:PASSWORD|PASSWD|SECRET|PASSPHRASE|APIKEY|API_KEY|TOKEN|CREDENTIALS?)[ \\t]*=[ \\t]*(?!\\$|<|\\*|\\{|%|encrypted:)[^\\s\"'`,;&|]{4,}/,\n  },\n  // Bare-key shapes not covered above. Each requires a distinctive prefix so\n  // it does not fire on ordinary identifiers — same reasoning as the\n  // provider-branded block above, and the same leading-`\\b` bug: a token\n  // glued to preceding text by a digit or underscore (`1ghp_…`, `_sk_live_…`)\n  // is invisible to an anchored detector.\n  { id: 'huggingface-token', pattern: /hf_[A-Za-z0-9]{20,}/ },\n  { id: 'npm-token', pattern: /npm_[A-Za-z0-9]{20,}/ },\n  { id: 'gitlab-token', pattern: /glpat-[A-Za-z0-9_-]{16,}/ },\n  { id: 'groq-key', pattern: /gsk_[A-Za-z0-9]{20,}/ },\n  { id: 'xai-key', pattern: /xai-[A-Za-z0-9]{16,}/ },\n  // ElevenLabs and similar bare `sk_<hex>` shapes — underscore, unlike the\n  // hyphenated `sk-…` `provider-key` pattern above, so the two never overlap.\n  { id: 'bare-sk-key', pattern: /sk_[A-Za-z0-9]{32,}\\b/ },\n  {\n    // `Cookie: sessionid=…; csrftoken=…` in free text or a captured header.\n    id: 'cookie-header',\n    pattern: /\\bcookie\\s*:\\s*[A-Za-z0-9_-]+=[^;\\r\\n]{3,}(?:;\\s*[A-Za-z0-9_-]+=[^;\\r\\n]{3,})*/i,\n  },\n]\n\n/** Personal-data shapes, replaced in place so the surrounding text survives. */\nconst PERSONAL_DATA_DETECTORS: readonly (ValueDetector & {\n  /** A substring the text must contain before the pattern runs. */\n  requires?: string\n  accept?: (match: string) => boolean\n})[] = [\n  {\n    id: 'email',\n    requires: '@',\n    pattern: /\\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\\.[A-Za-z]{2,}\\b/g,\n    accept: isEmailAddress,\n  },\n  {\n    // Card numbers: a card-network prefix and a valid Luhn digit, written\n    // unbroken (13-19 digits) or in printed groups (4-4-4-4, 4-4-4-4-3, Amex\n    // 4-6-5). A word character or hyphen on either side rejects the match, so\n    // digits inside UUIDs, dated file names and hyphenated ids never match.\n    id: 'card-number',\n    pattern:\n      /(?<![\\w-])(?:\\d{13,19}|\\d{4}([ -])\\d{4}\\1\\d{4}\\1\\d{4}(?:\\1\\d{3})?|\\d{4}([ -])\\d{6}\\2\\d{5})(?![\\w-])/g,\n    accept: (match) => {\n      const digits = match.replace(/[ -]/g, '')\n      return CARD_PREFIX.test(digits) && luhnValid(digits)\n    },\n  },\n  { id: 'ssn', pattern: /\\b\\d{3}-\\d{2}-\\d{4}\\b/g },\n  {\n    // Separators are required: a bare ten-digit run is usually an epoch second or an id.\n    id: 'phone',\n    pattern: /(?:\\+1[-.\\s]?)?(?:\\(\\d{3}\\)\\s?|\\b\\d{3}[-.\\s])\\d{3}[-.\\s]\\d{4}\\b/g,\n  },\n]\n\n/** Visa, Mastercard, Amex, Diners, JCB, Discover and UnionPay issuer prefixes. */\nconst CARD_PREFIX = /^(?:4|5[1-5]|2[2-7]|3[04-9]|6(?:011|4[4-9]|5|2))/\n\n/**\n * Top-level domains that are file extensions: `icon@2x.png`, `worker@3.service`\n * and `notes@AGENTS.md` are not addresses.\n */\nconst NON_EMAIL_TLDS: ReadonlySet<string> = new Set([\n  'md',\n  'mdx',\n  'txt',\n  'json',\n  'jsonl',\n  'yaml',\n  'yml',\n  'toml',\n  'lock',\n  'ts',\n  'tsx',\n  'js',\n  'jsx',\n  'mjs',\n  'cjs',\n  'py',\n  'rs',\n  'go',\n  'sh',\n  'css',\n  'html',\n  'svg',\n  'png',\n  'jpg',\n  'jpeg',\n  'gif',\n  'webp',\n  'service',\n  'socket',\n  'timer',\n  'mount',\n  'target',\n])\n\nfunction isEmailAddress(match: string): boolean {\n  // `git@github.com:org/repo` is an SSH remote, not a person.\n  if (match.toLowerCase().startsWith('git@')) return false\n  const tld = match.slice(match.lastIndexOf('.') + 1).toLowerCase()\n  return !NON_EMAIL_TLDS.has(tld)\n}\n\nfunction luhnValid(digits: string): boolean {\n  if (/^(\\d)\\1+$/.test(digits)) return false\n  let sum = 0\n  let double = false\n  for (let index = digits.length - 1; index >= 0; index -= 1) {\n    let digit = digits.charCodeAt(index) - 48\n    if (double) {\n      digit *= 2\n      if (digit > 9) digit -= 9\n    }\n    sum += digit\n    double = !double\n  }\n  return sum % 10 === 0\n}\n\n/** The first credential detector that matches, or undefined. */\nexport function detectCredential(text: string): string | undefined {\n  for (const detector of CREDENTIAL_DETECTORS) {\n    if (detector.pattern.test(text)) return detector.id\n  }\n  return undefined\n}\n\n/** Personal-data detector ids that match somewhere in `text`. */\nexport function detectPersonalData(text: string): string[] {\n  const hits: string[] = []\n  for (const detector of PERSONAL_DATA_DETECTORS) {\n    if (detector.requires && !text.includes(detector.requires)) continue\n    detector.pattern.lastIndex = 0\n    for (const match of text.matchAll(detector.pattern)) {\n      if (!detector.accept || detector.accept(match[0])) {\n        hits.push(detector.id)\n        break\n      }\n    }\n  }\n  return hits\n}\n\n/**\n * Every form in which a known secret can appear in a trace: as written,\n * base64, base64url and URL-encoded. Values shorter than 4 characters are\n * ignored; they would match ordinary text.\n */\nfunction knownSecretForms(values: readonly string[] | undefined): string[] {\n  const forms = new Set<string>()\n  for (const value of values ?? []) {\n    if (value.length < 4) continue\n    const bytes = Buffer.from(value, 'utf8')\n    forms.add(value)\n    forms.add(bytes.toString('base64'))\n    forms.add(bytes.toString('base64url'))\n    forms.add(encodeURIComponent(value))\n  }\n  // Longest first, so a secret that contains another is reported as itself.\n  return [...forms].sort((left, right) => right.length - left.length)\n}\n\nconst BASE64_TEXT = /^[A-Za-z0-9+/_-]+={0,2}$/\n\n// A base64 blob embedded in larger text — `Authorization: Basic <base64>` is\n// the common case — can hold a known secret without the WHOLE string being\n// base64. Bounded to 4096 chars per candidate: this scans in linear time (the\n// sum of candidate lengths is at most `text.length`), so no ReDoS risk.\nconst EMBEDDED_BASE64 = /[A-Za-z0-9+/]{16,4096}={0,2}/g\n\nfunction containsKnownSecret(text: string, forms: readonly string[]): boolean {\n  if (forms.length === 0) return false\n  if (forms.some((form) => text.includes(form))) return true\n  // A whole-string base64 payload can hold a secret at any byte offset.\n  if (text.length >= 8 && BASE64_TEXT.test(text)) {\n    const decoded = Buffer.from(\n      text,\n      text.includes('-') || text.includes('_') ? 'base64url' : 'base64',\n    ).toString('utf8')\n    if (forms.some((form) => decoded.includes(form))) return true\n  }\n  // `Basic <base64(user:key)>` jointly encodes the userinfo, so the secret's\n  // OWN base64 form (added to `forms` above) is generally not a substring of\n  // it — 3-byte base64 group alignment shifts with whatever precedes the\n  // secret in the original bytes. Decode any embedded base64-looking run and\n  // check its plaintext directly instead of relying on aligned forms.\n  for (const match of text.matchAll(EMBEDDED_BASE64)) {\n    if (match[0] === text) continue // already checked above\n    try {\n      const decoded = Buffer.from(match[0], 'base64').toString('utf8')\n      if (forms.some((form) => decoded.includes(form))) return true\n    } catch {\n      // Not valid base64 (e.g. wrong padding) — not a candidate.\n    }\n  }\n  return false\n}\n\n/** The detector that marks `text` as a credential: a known secret first, then the value shapes. */\nfunction credentialIn(text: string, knownSecrets: readonly string[]): string | undefined {\n  return containsKnownSecret(text, knownSecrets) ? 'known-secret' : detectCredential(text)\n}\n\nconst MARKER_PREFIX = '[REDACTED:'\n\nfunction marker(detector: string): string {\n  return `${MARKER_PREFIX}${detector}]`\n}\n\n/**\n * Placeholders other tools write for a removed value. A credential field that\n * holds one of these is already redacted and is not flagged again.\n */\nconst REDACTED_PLACEHOLDER =\n  /^(?:\\[(?:REDACTED|redacted)(?::[^\\]]*)?\\]|\\[ID:[0-9a-f]{12}\\]|\\*{3,}|<redacted>|«redacted»|\\[HASH:[0-9a-f]{8,}\\])$/\n\nfunction isPlaceholder(value: string): boolean {\n  return value === '' || REDACTED_PLACEHOLDER.test(value)\n}\n\n// The whole string must be the data URI — `$` at the end, not just a `^`\n// prefix check. A prefix-only check let `data:text/plain;base64,…\\n` PLUS\n// arbitrary trailing text through as unread \"media\", when the trailing text\n// (or, for text/json mimes, the payload itself once decoded) is exactly the\n// kind of prose a credential or personal-data detector should see.\nconst DATA_URI = /^data:([\\w.+-]+\\/[\\w.+-]+)(?:;[\\w.+-]+=[\\w.+-]+)*;base64,([A-Za-z0-9+/=\\s]*)$/i\n\n/** A data: URI's mime type and base64 payload, only when the whole string is the URI. */\nfunction matchDataUri(text: string): { mime: string; payload: string } | undefined {\n  const m = DATA_URI.exec(text)\n  if (!m) return undefined\n  return { mime: (m[1] ?? '').toLowerCase(), payload: m[2] ?? '' }\n}\n\n/** Decodes a data: URI's base64 payload to text, or undefined if it is not valid base64. */\nfunction decodeDataUriPayload(payload: string): string | undefined {\n  const cleaned = payload.replace(/\\s+/g, '')\n  if (cleaned.length === 0) return ''\n  if (!/^[A-Za-z0-9+/]*={0,2}$/.test(cleaned) || cleaned.length % 4 !== 0) return undefined\n  try {\n    return Buffer.from(cleaned, 'base64').toString('utf8')\n  } catch {\n    return undefined\n  }\n}\n\nconst READABLE_IMAGE_MIME = /^image\\//\n/** Mimes whose payload is text a detector can read once decoded. */\nconst DECODABLE_TEXT_MIME = /^text\\/|^application\\/json$/\n\n/** Whether a decoded data: URI payload holds a credential or personal data,\n *  by key name (for application/json) or by value shape (for any text). */\nfunction decodedPayloadIsUnsafe(\n  decoded: string,\n  mime: string,\n  knownSecrets: readonly string[],\n): boolean {\n  if (mime === 'application/json') {\n    try {\n      return jsonValueIsUnsafe(JSON.parse(decoded), knownSecrets)\n    } catch {\n      // Not actually valid JSON despite the mime type — fall through to plain text.\n    }\n  }\n  if (credentialIn(decoded, knownSecrets)) return true\n  return detectPersonalData(decoded).length > 0\n}\n\nfunction jsonValueIsUnsafe(value: unknown, knownSecrets: readonly string[]): boolean {\n  if (value === null || value === undefined) return false\n  if (typeof value === 'string')\n    return credentialIn(value, knownSecrets) !== undefined || detectPersonalData(value).length > 0\n  if (typeof value !== 'object') return false\n  if (Array.isArray(value)) return value.some((item) => jsonValueIsUnsafe(item, knownSecrets))\n  for (const [key, item] of Object.entries(value as Record<string, unknown>)) {\n    const category = classifyKey(key)\n    if (category === 'credential') return true\n    if (category === 'personal-data' && (typeof item === 'string' || typeof item === 'number'))\n      return true\n    if (jsonValueIsUnsafe(item, knownSecrets)) return true\n  }\n  return false\n}\n\n// ── Redaction ────────────────────────────────────────────────────────────\n\nconst MAX_DEPTH = 64\n\ninterface WalkState {\n  rules: ProfileRules\n  maxStringBytes: number\n  knownSecrets: string[]\n  report: RedactionReport\n  /** Created on first use: most calls pseudonymize nothing. */\n  pseudonymKey?: Buffer\n  seen: WeakSet<object>\n}\n\nfunction record(\n  state: WalkState,\n  path: string,\n  category: SafetyCategory,\n  detector: string,\n  action: RedactionFinding['action'],\n): void {\n  const report = state.report\n  if (action === 'truncated') report.truncatedCount += 1\n  report.redactionCount += 1\n  report.byDetector[detector] = (report.byDetector[detector] ?? 0) + 1\n  report.findings.push({ path, category, detector, action })\n}\n\nfunction pointer(parent: string, key: string | number): string {\n  return `${parent}/${String(key).replace(/~/g, '~0').replace(/\\//g, '~1')}`\n}\n\nfunction truncateUtf8(text: string, maxBytes: number): string | undefined {\n  if (text.length <= maxBytes / 4) return undefined\n  const bytes = Buffer.from(text, 'utf8')\n  if (bytes.length <= maxBytes) return undefined\n  // Step back to the first byte of a UTF-8 sequence so no character is split.\n  let cut = maxBytes\n  while (cut > 0 && ((bytes[cut] ?? 0) & 0xc0) === 0x80) cut -= 1\n  return `${bytes.subarray(0, cut).toString('utf8')}…[truncated ${bytes.length - cut} bytes]`\n}\n\nfunction redactStringAt(text: string, path: string, state: WalkState): string {\n  if (isPlaceholder(text)) return text\n  // A known secret is an exact value, so it is cut out where it stands and the\n  // text around it survives: an error message keeps its diagnosis.\n  let output = text\n  for (const form of state.knownSecrets) {\n    if (!output.includes(form)) continue\n    output = output.split(form).join(marker('known-secret'))\n    record(state, path, 'credential', 'known-secret', 'redacted')\n  }\n  const dataUri = matchDataUri(output)\n  if (dataUri) {\n    if (DECODABLE_TEXT_MIME.test(dataUri.mime)) {\n      const decoded = decodeDataUriPayload(dataUri.payload)\n      if (decoded === undefined) {\n        // Can't verify what an undecodable payload holds — fail closed rather\n        // than pass it through as if it were opaque, harmless media.\n        record(state, path, 'credential', 'data-uri-payload', 'redacted')\n        return marker('data-uri-payload')\n      }\n      if (decodedPayloadIsUnsafe(decoded, dataUri.mime, state.knownSecrets)) {\n        record(state, path, 'credential', 'data-uri-payload', 'redacted')\n        return marker('data-uri-payload')\n      }\n      // Decoded clean: still opaque bytes to a downstream reader once\n      // embedded, so it's subject to the ordinary media rule below, same as\n      // any other data URI.\n    }\n    if (!state.rules.redactMedia) return output\n    record(state, path, 'media', 'media', 'redacted')\n    return marker('media')\n  }\n  // A value shape, or a whole-string base64 payload that holds a known secret\n  // at any byte offset, replaces the whole string. The detectors themselves\n  // only ever see a bounded prefix (`boundForScan`): a credential that lives\n  // entirely beyond `maxStringBytes` would be truncated out of the output\n  // below anyway, so scanning past the bound costs time without changing\n  // what can leak. The untouched tail is reattached after the personal-data\n  // pass so later truncation still sees (and cuts) the full length.\n  const boundLen = Math.min(output.length, state.maxStringBytes)\n  let scanTarget = output.slice(0, boundLen)\n  const tail = output.slice(boundLen)\n  const credential = credentialIn(scanTarget, state.knownSecrets)\n  if (credential) {\n    record(state, path, 'credential', credential, 'redacted')\n    return marker(credential)\n  }\n  for (const detector of PERSONAL_DATA_DETECTORS) {\n    if (detector.requires && !scanTarget.includes(detector.requires)) continue\n    scanTarget = scanTarget.replace(detector.pattern, (match) => {\n      if (detector.accept && !detector.accept(match)) return match\n      record(state, path, 'personal-data', detector.id, 'redacted')\n      return marker(detector.id)\n    })\n  }\n  output = scanTarget + tail\n  const truncated = truncateUtf8(output, state.maxStringBytes)\n  if (truncated === undefined) return output\n  record(state, path, 'raw-content', 'size-cap', 'truncated')\n  return truncated\n}\n\nfunction pseudonym(value: string | number, state: WalkState): string {\n  state.pseudonymKey ??= randomBytes(32)\n  const digest = createHmac('sha256', state.pseudonymKey).update(String(value)).digest('hex')\n  return `[ID:${digest.slice(0, 12)}]`\n}\n\nfunction redactKeyed(\n  key: string,\n  value: unknown,\n  path: string,\n  depth: number,\n  state: WalkState,\n): unknown {\n  const kind = classifyKeyName(key)\n  if (value === null || value === undefined || typeof value === 'boolean') {\n    return value\n  }\n  if (typeof value === 'string' && isPlaceholder(value)) return value\n  // Numbers under a credential name are limits and counts (`password_max_age`), not secrets.\n  if (kind === 'credential' && typeof value !== 'number') {\n    record(state, path, 'credential', 'credential-key', 'redacted')\n    return marker('credential-key')\n  }\n  if (\n    kind === 'email' ||\n    (kind === 'share-personal' && state.rules.redactKeys.has('share-personal'))\n  ) {\n    if (typeof value === 'string' || typeof value === 'number') {\n      record(state, path, 'personal-data', 'personal-data-key', 'redacted')\n      return marker('personal-data-key')\n    }\n  }\n  if (kind === 'identifier' && state.rules.redactKeys.has('identifier')) {\n    if (typeof value === 'string' || typeof value === 'number') {\n      record(state, path, 'identifier', 'identifier-key', 'pseudonymized')\n      return pseudonym(value, state)\n    }\n  }\n  if (\n    kind === 'raw-content' &&\n    state.rules.redactKeys.has('raw-content') &&\n    typeof value !== 'number'\n  ) {\n    record(state, path, 'raw-content', 'raw-content-key', 'redacted')\n    return marker('raw-content-key')\n  }\n  return walk(value, path, depth + 1, state)\n}\n\nfunction walk(value: unknown, path: string, depth: number, state: WalkState): unknown {\n  if (typeof value === 'string') return redactStringAt(value, path, state)\n  if (value === null || typeof value !== 'object') return value\n  if (ArrayBuffer.isView(value) || value instanceof ArrayBuffer) {\n    record(state, path, 'media', 'binary', 'redacted')\n    return marker('binary')\n  }\n  if (value instanceof Date) return value\n  if (depth > MAX_DEPTH) {\n    record(state, path, 'raw-content', 'depth-limit', 'truncated')\n    return marker('depth-limit')\n  }\n  if (state.seen.has(value)) {\n    record(state, path, 'raw-content', 'circular', 'redacted')\n    return marker('circular')\n  }\n  state.seen.add(value)\n  try {\n    if (Array.isArray(value)) {\n      return value.map((item, index) => walk(item, pointer(path, index), depth + 1, state))\n    }\n    const out: Record<string, unknown> = {}\n    let renamed = 0\n    for (const [key, item] of Object.entries(value as Record<string, unknown>)) {\n      let outKey = key\n      const keyCredential = credentialIn(key, state.knownSecrets)\n      // A personal-data value used AS a key (an email or card number mapping\n      // to something) is just as reachable as one in a value — walk() only\n      // ever inspects keys for a credential shape, so `{'alice@x.com': 3}`\n      // passed through unchanged with verdict SAFE.\n      const keyPersonalData = keyCredential ? undefined : detectPersonalData(key)[0]\n      if (keyCredential) {\n        renamed += 1\n        outKey = `${marker(keyCredential)}#${renamed}`\n        record(state, pointer(path, outKey), 'credential', keyCredential, 'redacted')\n      } else if (keyPersonalData) {\n        renamed += 1\n        outKey = `${marker(keyPersonalData)}#${renamed}`\n        record(state, pointer(path, outKey), 'personal-data', keyPersonalData, 'redacted')\n      }\n      out[outKey] = redactKeyed(key, item, pointer(path, outKey), depth, state)\n    }\n    return out\n  } finally {\n    state.seen.delete(value)\n  }\n}\n\nfunction newReport(profile: RedactionProfile): RedactionReport {\n  return {\n    version: REDACTION_VERSION,\n    profile,\n    redactionCount: 0,\n    byDetector: {},\n    truncatedCount: 0,\n    findings: [],\n  }\n}\n\nfunction newState(options: RedactOptions, report?: RedactionReport): WalkState {\n  const profile = options.profile ?? 'default'\n  const rules = PROFILE_RULES[profile]\n  if (!rules) throw new Error(`unknown redaction profile: ${String(profile)}`)\n  const requested = options.maxStringBytes\n  const maxStringBytes =\n    typeof requested === 'number' && Number.isFinite(requested) && requested >= 0\n      ? Math.min(Math.floor(requested), rules.maxStringBytes)\n      : rules.maxStringBytes\n  return {\n    rules,\n    maxStringBytes,\n    knownSecrets: knownSecretForms(options.knownSecrets),\n    report: report ?? newReport(profile),\n    seen: new WeakSet(),\n  }\n}\n\n/**\n * Return a redacted deep copy of `value` and a report of what changed. Pure:\n * the input is not mutated. Never throws on the value's shape; cycles, binary\n * data and nesting past 64 levels are replaced with markers.\n */\nexport function redact<T>(\n  value: T,\n  options: RedactOptions = {},\n): { value: T; report: RedactionReport } {\n  const state = newState(options, options.report)\n  const out = walk(value, '', 0, state) as T\n  return { value: out, report: state.report }\n}\n\n/**\n * Redact one free-text string with the value detectors (no field name is\n * known). Pass `report` to accumulate counts across calls.\n */\nexport function redactText(text: string, options: RedactOptions = {}): string {\n  const state = newState(options, options.report)\n  return redactStringAt(text, '', state)\n}\n\n/** An empty report to accumulate `redactText` calls into. */\nexport function emptyRedactionReport(profile: RedactionProfile = 'default'): RedactionReport {\n  return newReport(profile)\n}\n\n// ── Share-safety assessment ──────────────────────────────────────────────\n\ninterface ScanState {\n  profile: RedactionProfile\n  rules: ProfileRules\n  knownSecrets: string[]\n  findings: Map<string, ShareSafetyFinding>\n  unreadable: string[]\n  seen: WeakSet<object>\n}\n\nfunction flag(state: ScanState, path: string, category: SafetyCategory, detector: string): void {\n  const severity = state.rules.severity[category]\n  if (!severity) return\n  const id = `${category}\\u0000${detector}`\n  const existing = state.findings.get(id)\n  if (existing) {\n    existing.count += 1\n    if (existing.paths.length < 5) existing.paths.push(path)\n    return\n  }\n  state.findings.set(id, { category, detector, severity, count: 1, paths: [path] })\n}\n\nfunction scanString(text: string, path: string, state: ScanState): void {\n  if (isPlaceholder(text)) return\n  const dataUri = matchDataUri(text)\n  if (dataUri) {\n    scanDataUri(dataUri, path, state)\n    return\n  }\n  const credential = credentialIn(text, state.knownSecrets)\n  if (credential) {\n    flag(state, path, 'credential', credential)\n    return\n  }\n  for (const detector of detectPersonalData(text)) flag(state, path, 'personal-data', detector)\n}\n\nfunction scanDataUri(\n  dataUri: { mime: string; payload: string },\n  path: string,\n  state: ScanState,\n): void {\n  if (READABLE_IMAGE_MIME.test(dataUri.mime)) {\n    flag(state, path, 'media', 'media')\n    return\n  }\n  if (DECODABLE_TEXT_MIME.test(dataUri.mime)) {\n    const decoded = decodeDataUriPayload(dataUri.payload)\n    if (decoded === undefined) {\n      state.unreadable.push(`${path || '/'}: data URI payload is not valid base64`)\n      return\n    }\n    if (dataUri.mime === 'application/json') {\n      try {\n        scan(JSON.parse(decoded), path, 0, state)\n        return\n      } catch {\n        // Not actually valid JSON despite the mime type — fall through to plain text.\n      }\n    }\n    scanString(decoded, path, state)\n    return\n  }\n  // audio/video/octet-stream/etc: bytes no text detector can read. A warning\n  // under `default` (matches the existing image rule); under `share` and\n  // `strict` — the profiles that actually gate whether something may be sent\n  // off the machine — treat it as unreadable so the verdict is UNKNOWN\n  // instead of silently allowing it through.\n  if (state.profile === 'default') {\n    flag(state, path, 'media', 'media')\n  } else {\n    state.unreadable.push(`${path || '/'}: ${dataUri.mime} data URI cannot be read`)\n  }\n}\n\nfunction scanKeyed(\n  key: string,\n  value: unknown,\n  path: string,\n  depth: number,\n  state: ScanState,\n): void {\n  if (value === null || value === undefined || typeof value === 'boolean') return\n  if (typeof value === 'string' && isPlaceholder(value)) return\n  const category = classifyKey(key)\n  if (category === 'credential' && typeof value !== 'number') {\n    flag(state, path, 'credential', 'credential-key')\n    return\n  }\n  const scalar = typeof value === 'string' || typeof value === 'number'\n  if (category === 'personal-data' && scalar) {\n    const kind = classifyKeyName(key)\n    // Phone, IP and address fields are flagged from the share profile up; email always.\n    if (kind === 'email' || state.profile !== 'default') {\n      flag(state, path, 'personal-data', 'personal-data-key')\n      return\n    }\n  }\n  if (category === 'identifier' && scalar) flag(state, path, 'identifier', 'identifier-key')\n  if (category === 'raw-content' && typeof value !== 'number')\n    flag(state, path, 'raw-content', 'raw-content-key')\n  scan(value, path, depth + 1, state)\n}\n\nfunction scan(value: unknown, path: string, depth: number, state: ScanState): void {\n  if (typeof value === 'string') {\n    scanString(value, path, state)\n    return\n  }\n  if (value === null || typeof value !== 'object') return\n  if (ArrayBuffer.isView(value) || value instanceof ArrayBuffer) {\n    state.unreadable.push(`${path || '/'}: binary data`)\n    return\n  }\n  if (value instanceof Date) return\n  if (depth > MAX_DEPTH) {\n    state.unreadable.push(`${path || '/'}: nested deeper than ${MAX_DEPTH} levels`)\n    return\n  }\n  if (state.seen.has(value)) {\n    state.unreadable.push(`${path || '/'}: circular reference`)\n    return\n  }\n  state.seen.add(value)\n  try {\n    if (Array.isArray(value)) {\n      for (const [index, item] of value.entries())\n        scan(item, pointer(path, index), depth + 1, state)\n      return\n    }\n    for (const [key, item] of Object.entries(value as Record<string, unknown>)) {\n      // A credential or personal-data value used as a key must not reach the\n      // finding paths.\n      const keyCredential = credentialIn(key, state.knownSecrets)\n      const keyPersonalData = keyCredential ? undefined : detectPersonalData(key)[0]\n      const keyDetector = keyCredential ?? keyPersonalData\n      const segment = keyDetector ? marker(keyDetector) : key\n      if (keyCredential) flag(state, pointer(path, segment), 'credential', keyCredential)\n      else if (keyPersonalData)\n        flag(state, pointer(path, segment), 'personal-data', keyPersonalData)\n      scanKeyed(key, item, pointer(path, segment), depth, state)\n    }\n  } finally {\n    state.seen.delete(value)\n  }\n}\n\n/**\n * Read `value` without changing it and say whether it is safe to share under\n * `profile`. UNKNOWN when any part could not be read; UNSAFE when any\n * error-severity finding remains; SAFE_WITH_WARNINGS when only warnings\n * remain (raw content, embedded media, identifiers under `share`).\n */\nexport function assessShareSafety(value: unknown, options: AssessOptions = {}): ShareSafetyVerdict {\n  const profile = options.profile ?? 'default'\n  const rules = PROFILE_RULES[profile]\n  if (!rules) throw new Error(`unknown redaction profile: ${String(profile)}`)\n  const state: ScanState = {\n    profile,\n    rules,\n    knownSecrets: knownSecretForms(options.knownSecrets),\n    findings: new Map(),\n    unreadable: [],\n    seen: new WeakSet(),\n  }\n  scan(value, '', 0, state)\n  return verdictFrom(profile, [...state.findings.values()], state.unreadable)\n}\n\n/**\n * One verdict over several parts (files in a directory, spans in a bundle).\n * Callers prefix each part's paths with its source before combining.\n */\nexport function combineVerdicts(\n  profile: RedactionProfile,\n  verdicts: readonly ShareSafetyVerdict[],\n  unreadable: readonly string[] = [],\n): ShareSafetyVerdict {\n  const findings = new Map<string, ShareSafetyFinding>()\n  for (const verdict of verdicts) {\n    for (const finding of verdict.findings) {\n      const id = `${finding.category}\\u0000${finding.detector}`\n      const existing = findings.get(id)\n      if (!existing) {\n        findings.set(id, { ...finding, paths: [...finding.paths] })\n        continue\n      }\n      existing.count += finding.count\n      for (const path of finding.paths) if (existing.paths.length < 5) existing.paths.push(path)\n    }\n  }\n  return verdictFrom(\n    profile,\n    [...findings.values()],\n    [...verdicts.flatMap((verdict) => verdict.unreadable), ...unreadable],\n  )\n}\n\nfunction verdictFrom(\n  profile: RedactionProfile,\n  findings: readonly ShareSafetyFinding[],\n  unreadable: readonly string[],\n): ShareSafetyVerdict {\n  const status: ShareSafetyStatus =\n    unreadable.length > 0\n      ? 'UNKNOWN'\n      : findings.some((finding) => finding.severity === 'error')\n        ? 'UNSAFE'\n        : findings.length > 0\n          ? 'SAFE_WITH_WARNINGS'\n          : 'SAFE'\n  return { status, profile, findings: [...findings], unreadable: [...unreadable] }\n}\n\n/** Whether a verdict allows sharing: SAFE or SAFE_WITH_WARNINGS. UNSAFE and UNKNOWN refuse. */\nexport function shareAllowed(verdict: ShareSafetyVerdict): boolean {\n  return verdict.status === 'SAFE' || verdict.status === 'SAFE_WITH_WARNINGS'\n}\n\n/**\n * Redact `value` under `profile`, then assess the redacted output. The\n * verdict describes what would actually be shared: a detector that failed to\n * remove something shows up as UNSAFE here.\n */\nexport function redactForShare<T>(\n  value: T,\n  options: RedactOptions = {},\n): { value: T; report: RedactionReport; verdict: ShareSafetyVerdict } {\n  const redacted = redact(value, options)\n  const verdict = assessShareSafety(redacted.value, {\n    profile: options.profile,\n    knownSecrets: options.knownSecrets,\n  })\n  return { value: redacted.value, report: redacted.report, verdict }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,MAAa,qBAAkD;CAAC;CAAW;CAAS;AAAQ;;;;;;AAU5F,MAAa,oBAAoB;AA0EjC,MAAM,MAAM;AAEZ,MAAM,gBAAwD;CAC5D,SAAS;EACP,gBAAgB,OAAO;EACvB,4BAAY,IAAI,IAAI;EACpB,aAAa;EACb,UAAU;GACR,YAAY;GACZ,iBAAiB;GACjB,YAAY,KAAA;GACZ,eAAe;GACf,OAAO;EACT;CACF;CACA,OAAO;EACL,gBAAgB,KAAK;EACrB,4BAAY,IAAI,IAAI,CAAC,kBAAkB,YAAY,CAAC;EACpD,aAAa;EACb,UAAU;GACR,YAAY;GACZ,iBAAiB;GACjB,YAAY;GACZ,eAAe;GACf,OAAO;EACT;CACF;CACA,QAAQ;EACN,gBAAgB,IAAI;EACpB,4BAAY,IAAI,IAAI;GAAC;GAAkB;GAAc;EAAa,CAAC;EACnE,aAAa;EACb,UAAU;GACR,YAAY;GACZ,iBAAiB;GACjB,YAAY;GACZ,eAAe;GACf,OAAO;EACT;CACF;AACF;;;;;;AASA,SAAgB,aAAa,KAAqB;CAChD,OAAO,IACJ,QAAQ,sBAAsB,OAAO,CAAC,CACtC,YAAY,CAAC,CACb,QAAQ,eAAe,GAAG,CAAC,CAC3B,QAAQ,YAAY,EAAE;AAC3B;;AAGA,MAAM,4CAAiD,IAAI,IAAI;CAC7D;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;CACtB;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;;AAGA,MAAM,aAAa,CAAC,SAAS,eAAe;;AAG5C,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;AAGA,MAAM,kBAAkB;CACtB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;AAGA,MAAM,mBAAmB;CACvB;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;CACA;CACA;CACA;CACA;AACF;;;;;;;;;AAUA,SAAS,eAAe,YAAoB,MAAkC;CAC5E,KAAK,MAAM,OAAO,MAChB,IAAI,eAAe,OAAO,WAAW,SAAS,IAAI,KAAK,GAAG,OAAO;CAEnE,OAAO;AACT;AAEA,SAAS,qBAAqB,YAA6B;CACzD,IAAI,0BAA0B,IAAI,UAAU,GAAG,OAAO;CACtD,IAAI,eAAe,SAAS,OAAO;CAEnC,IAAI,WAAW,SAAS,QAAQ,GAAG,OAAO;CAC1C,OAAO,WAAW,SAAS,OAAO;AACpC;;AAKA,MAAM,gCAAgB,IAAI,IAA6B;AAEvD,SAAS,gBAAgB,KAAmC;CAC1D,MAAM,SAAS,cAAc,IAAI,GAAG;CACpC,IAAI,WAAW,KAAA,GAAW,OAAO,UAAU,KAAA;CAC3C,MAAM,OAAO,sBAAsB,aAAa,GAAG,CAAC;CACpD,IAAI,cAAc,QAAQ,KAAQ,cAAc,MAAM;CACtD,cAAc,IAAI,KAAK,QAAQ,IAAI;CACnC,OAAO;AACT;AAEA,SAAS,sBAAsB,YAA0C;CACvE,IAAI,CAAC,cAAc,0BAA0B,IAAI,UAAU,GAAG,OAAO,KAAA;CACrE,IAAI,qBAAqB,UAAU,GAAG,OAAO;CAC7C,IAAI,eAAe,YAAY,eAAe,GAAG,OAAO;CACxD,IAAI,eAAe,YAAY,UAAU,GAAG,OAAO;CACnD,IAAI,eAAe,YAAY,mBAAmB,GAAG,OAAO;CAC5D,IAAI,eAAe,YAAY,eAAe,GAAG,OAAO;CACxD,IAAI,eAAe,YAAY,gBAAgB,GAAG,OAAO;AAE3D;;;;;;AAOA,SAAgB,YAAY,KAAyC;CACnE,MAAM,OAAO,gBAAgB,GAAG;CAChC,IAAI,SAAS,KAAA,GAAW,OAAO,KAAA;CAC/B,IAAI,SAAS,cAAc,OAAO;CAClC,IAAI,SAAS,cAAc,OAAO;CAClC,IAAI,SAAS,eAAe,OAAO;CACnC,OAAO;AACT;;;;;;;;;;;;;;;;AAwBA,MAAM,uBAAiD;CACrD;EAAE,IAAI;EAAe,SAAS;CAAwC;CACtE;EAAE,IAAI;EAAO,SAAS;CAA6D;CACnF;EAAE,IAAI;EAAU,SAAS;CAAsC;CAC/D;EACE,IAAI;EACJ,SAAS;CACX;CACA;EAAE,IAAI;EAAiB,SAAS;CAA4B;CAC5D;EAAE,IAAI;EAAgB,SAAS;CAAuC;CACtE;EAAE,IAAI;EAAc,SAAS;CAA4C;CACzE;EAAE,IAAI;EAAkB,SAAS;CAAyB;CAC1D;EAAE,IAAI;EAAgB,SAAS;CAA+D;CAC9F;EAAE,IAAI;EAAe,SAAS;CAAiC;CAC/D;EAAE,IAAI;EAAkB,SAAS;CAA8B;CAC/D;EAAE,IAAI;EAAqB,SAAS;CAAqC;CACzE;EAME,IAAI;EACJ,SAAS;CACX;CACA;EACE,IAAI;EACJ,SACE;CACJ;CACA;EAWE,IAAI;EACJ,SACE;CACJ;CACA;EAOE,IAAI;EACJ,SACE;CACJ;CACA;EAOE,IAAI;EACJ,SACE;CACJ;CAMA;EAAE,IAAI;EAAqB,SAAS;CAAsB;CAC1D;EAAE,IAAI;EAAa,SAAS;CAAuB;CACnD;EAAE,IAAI;EAAgB,SAAS;CAA2B;CAC1D;EAAE,IAAI;EAAY,SAAS;CAAuB;CAClD;EAAE,IAAI;EAAW,SAAS;CAAuB;CAGjD;EAAE,IAAI;EAAe,SAAS;CAAwB;CACtD;EAEE,IAAI;EACJ,SAAS;CACX;AACF;;AAGA,MAAM,0BAIC;CACL;EACE,IAAI;EACJ,UAAU;EACV,SAAS;EACT,QAAQ;CACV;CACA;EAKE,IAAI;EACJ,SACE;EACF,SAAS,UAAU;GACjB,MAAM,SAAS,MAAM,QAAQ,SAAS,EAAE;GACxC,OAAO,YAAY,KAAK,MAAM,KAAK,UAAU,MAAM;EACrD;CACF;CACA;EAAE,IAAI;EAAO,SAAS;CAAyB;CAC/C;EAEE,IAAI;EACJ,SAAS;CACX;AACF;;AAGA,MAAM,cAAc;;;;;AAMpB,MAAM,iCAAsC,IAAI,IAAI;CAClD;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,CAAC;AAED,SAAS,eAAe,OAAwB;CAE9C,IAAI,MAAM,YAAY,CAAC,CAAC,WAAW,MAAM,GAAG,OAAO;CACnD,MAAM,MAAM,MAAM,MAAM,MAAM,YAAY,GAAG,IAAI,CAAC,CAAC,CAAC,YAAY;CAChE,OAAO,CAAC,eAAe,IAAI,GAAG;AAChC;AAEA,SAAS,UAAU,QAAyB;CAC1C,IAAI,YAAY,KAAK,MAAM,GAAG,OAAO;CACrC,IAAI,MAAM;CACV,IAAI,SAAS;CACb,KAAK,IAAI,QAAQ,OAAO,SAAS,GAAG,SAAS,GAAG,SAAS,GAAG;EAC1D,IAAI,QAAQ,OAAO,WAAW,KAAK,IAAI;EACvC,IAAI,QAAQ;GACV,SAAS;GACT,IAAI,QAAQ,GAAG,SAAS;EAC1B;EACA,OAAO;EACP,SAAS,CAAC;CACZ;CACA,OAAO,MAAM,OAAO;AACtB;;AAGA,SAAgB,iBAAiB,MAAkC;CACjE,KAAK,MAAM,YAAY,sBACrB,IAAI,SAAS,QAAQ,KAAK,IAAI,GAAG,OAAO,SAAS;AAGrD;;AAGA,SAAgB,mBAAmB,MAAwB;CACzD,MAAM,OAAiB,CAAC;CACxB,KAAK,MAAM,YAAY,yBAAyB;EAC9C,IAAI,SAAS,YAAY,CAAC,KAAK,SAAS,SAAS,QAAQ,GAAG;EAC5D,SAAS,QAAQ,YAAY;EAC7B,KAAK,MAAM,SAAS,KAAK,SAAS,SAAS,OAAO,GAChD,IAAI,CAAC,SAAS,UAAU,SAAS,OAAO,MAAM,EAAE,GAAG;GACjD,KAAK,KAAK,SAAS,EAAE;GACrB;EACF;CAEJ;CACA,OAAO;AACT;;;;;;AAOA,SAAS,iBAAiB,QAAiD;CACzE,MAAM,wBAAQ,IAAI,IAAY;CAC9B,KAAK,MAAM,SAAS,UAAU,CAAC,GAAG;EAChC,IAAI,MAAM,SAAS,GAAG;EACtB,MAAM,QAAQ,OAAO,KAAK,OAAO,MAAM;EACvC,MAAM,IAAI,KAAK;EACf,MAAM,IAAI,MAAM,SAAS,QAAQ,CAAC;EAClC,MAAM,IAAI,MAAM,SAAS,WAAW,CAAC;EACrC,MAAM,IAAI,mBAAmB,KAAK,CAAC;CACrC;CAEA,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,MAAM,UAAU,MAAM,SAAS,KAAK,MAAM;AACpE;AAEA,MAAM,cAAc;AAMpB,MAAM,kBAAkB;AAExB,SAAS,oBAAoB,MAAc,OAAmC;CAC5E,IAAI,MAAM,WAAW,GAAG,OAAO;CAC/B,IAAI,MAAM,MAAM,SAAS,KAAK,SAAS,IAAI,CAAC,GAAG,OAAO;CAEtD,IAAI,KAAK,UAAU,KAAK,YAAY,KAAK,IAAI,GAAG;EAC9C,MAAM,UAAU,OAAO,KACrB,MACA,KAAK,SAAS,GAAG,KAAK,KAAK,SAAS,GAAG,IAAI,cAAc,QAC3D,CAAC,CAAC,SAAS,MAAM;EACjB,IAAI,MAAM,MAAM,SAAS,QAAQ,SAAS,IAAI,CAAC,GAAG,OAAO;CAC3D;CAMA,KAAK,MAAM,SAAS,KAAK,SAAS,eAAe,GAAG;EAClD,IAAI,MAAM,OAAO,MAAM;EACvB,IAAI;GACF,MAAM,UAAU,OAAO,KAAK,MAAM,IAAI,QAAQ,CAAC,CAAC,SAAS,MAAM;GAC/D,IAAI,MAAM,MAAM,SAAS,QAAQ,SAAS,IAAI,CAAC,GAAG,OAAO;EAC3D,QAAQ,CAER;CACF;CACA,OAAO;AACT;;AAGA,SAAS,aAAa,MAAc,cAAqD;CACvF,OAAO,oBAAoB,MAAM,YAAY,IAAI,iBAAiB,iBAAiB,IAAI;AACzF;AAEA,MAAM,gBAAgB;AAEtB,SAAS,OAAO,UAA0B;CACxC,OAAO,GAAG,gBAAgB,SAAS;AACrC;;;;;AAMA,MAAM,uBACJ;AAEF,SAAS,cAAc,OAAwB;CAC7C,OAAO,UAAU,MAAM,qBAAqB,KAAK,KAAK;AACxD;AAOA,MAAM,WAAW;;AAGjB,SAAS,aAAa,MAA6D;CACjF,MAAM,IAAI,SAAS,KAAK,IAAI;CAC5B,IAAI,CAAC,GAAG,OAAO,KAAA;CACf,OAAO;EAAE,OAAO,EAAE,MAAM,GAAA,CAAI,YAAY;EAAG,SAAS,EAAE,MAAM;CAAG;AACjE;;AAGA,SAAS,qBAAqB,SAAqC;CACjE,MAAM,UAAU,QAAQ,QAAQ,QAAQ,EAAE;CAC1C,IAAI,QAAQ,WAAW,GAAG,OAAO;CACjC,IAAI,CAAC,yBAAyB,KAAK,OAAO,KAAK,QAAQ,SAAS,MAAM,GAAG,OAAO,KAAA;CAChF,IAAI;EACF,OAAO,OAAO,KAAK,SAAS,QAAQ,CAAC,CAAC,SAAS,MAAM;CACvD,QAAQ;EACN;CACF;AACF;AAEA,MAAM,sBAAsB;;AAE5B,MAAM,sBAAsB;;;AAI5B,SAAS,uBACP,SACA,MACA,cACS;CACT,IAAI,SAAS,oBACX,IAAI;EACF,OAAO,kBAAkB,KAAK,MAAM,OAAO,GAAG,YAAY;CAC5D,QAAQ,CAER;CAEF,IAAI,aAAa,SAAS,YAAY,GAAG,OAAO;CAChD,OAAO,mBAAmB,OAAO,CAAC,CAAC,SAAS;AAC9C;AAEA,SAAS,kBAAkB,OAAgB,cAA0C;CACnF,IAAI,UAAU,QAAQ,UAAU,KAAA,GAAW,OAAO;CAClD,IAAI,OAAO,UAAU,UACnB,OAAO,aAAa,OAAO,YAAY,MAAM,KAAA,KAAa,mBAAmB,KAAK,CAAC,CAAC,SAAS;CAC/F,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO,MAAM,MAAM,SAAS,kBAAkB,MAAM,YAAY,CAAC;CAC3F,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,KAAgC,GAAG;EAC1E,MAAM,WAAW,YAAY,GAAG;EAChC,IAAI,aAAa,cAAc,OAAO;EACtC,IAAI,aAAa,oBAAoB,OAAO,SAAS,YAAY,OAAO,SAAS,WAC/E,OAAO;EACT,IAAI,kBAAkB,MAAM,YAAY,GAAG,OAAO;CACpD;CACA,OAAO;AACT;AAIA,MAAM,YAAY;AAYlB,SAAS,OACP,OACA,MACA,UACA,UACA,QACM;CACN,MAAM,SAAS,MAAM;CACrB,IAAI,WAAW,aAAa,OAAO,kBAAkB;CACrD,OAAO,kBAAkB;CACzB,OAAO,WAAW,aAAa,OAAO,WAAW,aAAa,KAAK;CACnE,OAAO,SAAS,KAAK;EAAE;EAAM;EAAU;EAAU;CAAO,CAAC;AAC3D;AAEA,SAAS,QAAQ,QAAgB,KAA8B;CAC7D,OAAO,GAAG,OAAO,GAAG,OAAO,GAAG,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC,CAAC,QAAQ,OAAO,IAAI;AACzE;AAEA,SAAS,aAAa,MAAc,UAAsC;CACxE,IAAI,KAAK,UAAU,WAAW,GAAG,OAAO,KAAA;CACxC,MAAM,QAAQ,OAAO,KAAK,MAAM,MAAM;CACtC,IAAI,MAAM,UAAU,UAAU,OAAO,KAAA;CAErC,IAAI,MAAM;CACV,OAAO,MAAM,OAAO,MAAM,QAAQ,KAAK,SAAU,KAAM,OAAO;CAC9D,OAAO,GAAG,MAAM,SAAS,GAAG,GAAG,CAAC,CAAC,SAAS,MAAM,EAAE,cAAc,MAAM,SAAS,IAAI;AACrF;AAEA,SAAS,eAAe,MAAc,MAAc,OAA0B;CAC5E,IAAI,cAAc,IAAI,GAAG,OAAO;CAGhC,IAAI,SAAS;CACb,KAAK,MAAM,QAAQ,MAAM,cAAc;EACrC,IAAI,CAAC,OAAO,SAAS,IAAI,GAAG;EAC5B,SAAS,OAAO,MAAM,IAAI,CAAC,CAAC,KAAK,OAAO,cAAc,CAAC;EACvD,OAAO,OAAO,MAAM,cAAc,gBAAgB,UAAU;CAC9D;CACA,MAAM,UAAU,aAAa,MAAM;CACnC,IAAI,SAAS;EACX,IAAI,oBAAoB,KAAK,QAAQ,IAAI,GAAG;GAC1C,MAAM,UAAU,qBAAqB,QAAQ,OAAO;GACpD,IAAI,YAAY,KAAA,GAAW;IAGzB,OAAO,OAAO,MAAM,cAAc,oBAAoB,UAAU;IAChE,OAAO,OAAO,kBAAkB;GAClC;GACA,IAAI,uBAAuB,SAAS,QAAQ,MAAM,MAAM,YAAY,GAAG;IACrE,OAAO,OAAO,MAAM,cAAc,oBAAoB,UAAU;IAChE,OAAO,OAAO,kBAAkB;GAClC;EAIF;EACA,IAAI,CAAC,MAAM,MAAM,aAAa,OAAO;EACrC,OAAO,OAAO,MAAM,SAAS,SAAS,UAAU;EAChD,OAAO,OAAO,OAAO;CACvB;CAQA,MAAM,WAAW,KAAK,IAAI,OAAO,QAAQ,MAAM,cAAc;CAC7D,IAAI,aAAa,OAAO,MAAM,GAAG,QAAQ;CACzC,MAAM,OAAO,OAAO,MAAM,QAAQ;CAClC,MAAM,aAAa,aAAa,YAAY,MAAM,YAAY;CAC9D,IAAI,YAAY;EACd,OAAO,OAAO,MAAM,cAAc,YAAY,UAAU;EACxD,OAAO,OAAO,UAAU;CAC1B;CACA,KAAK,MAAM,YAAY,yBAAyB;EAC9C,IAAI,SAAS,YAAY,CAAC,WAAW,SAAS,SAAS,QAAQ,GAAG;EAClE,aAAa,WAAW,QAAQ,SAAS,UAAU,UAAU;GAC3D,IAAI,SAAS,UAAU,CAAC,SAAS,OAAO,KAAK,GAAG,OAAO;GACvD,OAAO,OAAO,MAAM,iBAAiB,SAAS,IAAI,UAAU;GAC5D,OAAO,OAAO,SAAS,EAAE;EAC3B,CAAC;CACH;CACA,SAAS,aAAa;CACtB,MAAM,YAAY,aAAa,QAAQ,MAAM,cAAc;CAC3D,IAAI,cAAc,KAAA,GAAW,OAAO;CACpC,OAAO,OAAO,MAAM,eAAe,YAAY,WAAW;CAC1D,OAAO;AACT;AAEA,SAAS,UAAU,OAAwB,OAA0B;CACnE,MAAM,iBAAiB,YAAY,EAAE;CAErC,OAAO,OADQ,WAAW,UAAU,MAAM,YAAY,CAAC,CAAC,OAAO,OAAO,KAAK,CAAC,CAAC,CAAC,OAAO,KAClE,CAAC,CAAC,MAAM,GAAG,EAAE,EAAE;AACpC;AAEA,SAAS,YACP,KACA,OACA,MACA,OACA,OACS;CACT,MAAM,OAAO,gBAAgB,GAAG;CAChC,IAAI,UAAU,QAAQ,UAAU,KAAA,KAAa,OAAO,UAAU,WAC5D,OAAO;CAET,IAAI,OAAO,UAAU,YAAY,cAAc,KAAK,GAAG,OAAO;CAE9D,IAAI,SAAS,gBAAgB,OAAO,UAAU,UAAU;EACtD,OAAO,OAAO,MAAM,cAAc,kBAAkB,UAAU;EAC9D,OAAO,OAAO,gBAAgB;CAChC;CACA,IACE,SAAS,WACR,SAAS,oBAAoB,MAAM,MAAM,WAAW,IAAI,gBAAgB,GAErE;MAAA,OAAO,UAAU,YAAY,OAAO,UAAU,UAAU;GAC1D,OAAO,OAAO,MAAM,iBAAiB,qBAAqB,UAAU;GACpE,OAAO,OAAO,mBAAmB;EACnC;;CAEF,IAAI,SAAS,gBAAgB,MAAM,MAAM,WAAW,IAAI,YAAY,GAC9D;MAAA,OAAO,UAAU,YAAY,OAAO,UAAU,UAAU;GAC1D,OAAO,OAAO,MAAM,cAAc,kBAAkB,eAAe;GACnE,OAAO,UAAU,OAAO,KAAK;EAC/B;;CAEF,IACE,SAAS,iBACT,MAAM,MAAM,WAAW,IAAI,aAAa,KACxC,OAAO,UAAU,UACjB;EACA,OAAO,OAAO,MAAM,eAAe,mBAAmB,UAAU;EAChE,OAAO,OAAO,iBAAiB;CACjC;CACA,OAAO,KAAK,OAAO,MAAM,QAAQ,GAAG,KAAK;AAC3C;AAEA,SAAS,KAAK,OAAgB,MAAc,OAAe,OAA2B;CACpF,IAAI,OAAO,UAAU,UAAU,OAAO,eAAe,OAAO,MAAM,KAAK;CACvE,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,IAAI,YAAY,OAAO,KAAK,KAAK,iBAAiB,aAAa;EAC7D,OAAO,OAAO,MAAM,SAAS,UAAU,UAAU;EACjD,OAAO,OAAO,QAAQ;CACxB;CACA,IAAI,iBAAiB,MAAM,OAAO;CAClC,IAAI,QAAQ,WAAW;EACrB,OAAO,OAAO,MAAM,eAAe,eAAe,WAAW;EAC7D,OAAO,OAAO,aAAa;CAC7B;CACA,IAAI,MAAM,KAAK,IAAI,KAAK,GAAG;EACzB,OAAO,OAAO,MAAM,eAAe,YAAY,UAAU;EACzD,OAAO,OAAO,UAAU;CAC1B;CACA,MAAM,KAAK,IAAI,KAAK;CACpB,IAAI;EACF,IAAI,MAAM,QAAQ,KAAK,GACrB,OAAO,MAAM,KAAK,MAAM,UAAU,KAAK,MAAM,QAAQ,MAAM,KAAK,GAAG,QAAQ,GAAG,KAAK,CAAC;EAEtF,MAAM,MAA+B,CAAC;EACtC,IAAI,UAAU;EACd,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,KAAgC,GAAG;GAC1E,IAAI,SAAS;GACb,MAAM,gBAAgB,aAAa,KAAK,MAAM,YAAY;GAK1D,MAAM,kBAAkB,gBAAgB,KAAA,IAAY,mBAAmB,GAAG,CAAC,CAAC;GAC5E,IAAI,eAAe;IACjB,WAAW;IACX,SAAS,GAAG,OAAO,aAAa,EAAE,GAAG;IACrC,OAAO,OAAO,QAAQ,MAAM,MAAM,GAAG,cAAc,eAAe,UAAU;GAC9E,OAAO,IAAI,iBAAiB;IAC1B,WAAW;IACX,SAAS,GAAG,OAAO,eAAe,EAAE,GAAG;IACvC,OAAO,OAAO,QAAQ,MAAM,MAAM,GAAG,iBAAiB,iBAAiB,UAAU;GACnF;GACA,IAAI,UAAU,YAAY,KAAK,MAAM,QAAQ,MAAM,MAAM,GAAG,OAAO,KAAK;EAC1E;EACA,OAAO;CACT,UAAU;EACR,MAAM,KAAK,OAAO,KAAK;CACzB;AACF;AAEA,SAAS,UAAU,SAA4C;CAC7D,OAAO;EACL,SAAS;EACT;EACA,gBAAgB;EAChB,YAAY,CAAC;EACb,gBAAgB;EAChB,UAAU,CAAC;CACb;AACF;AAEA,SAAS,SAAS,SAAwB,QAAqC;CAC7E,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,QAAQ,cAAc;CAC5B,IAAI,CAAC,OAAO,MAAM,IAAI,MAAM,8BAA8B,OAAO,OAAO,GAAG;CAC3E,MAAM,YAAY,QAAQ;CAK1B,OAAO;EACL;EACA,gBALA,OAAO,cAAc,YAAY,OAAO,SAAS,SAAS,KAAK,aAAa,IACxE,KAAK,IAAI,KAAK,MAAM,SAAS,GAAG,MAAM,cAAc,IACpD,MAAM;EAIV,cAAc,iBAAiB,QAAQ,YAAY;EACnD,QAAQ,UAAU,UAAU,OAAO;EACnC,sBAAM,IAAI,QAAQ;CACpB;AACF;;;;;;AAOA,SAAgB,OACd,OACA,UAAyB,CAAC,GACa;CACvC,MAAM,QAAQ,SAAS,SAAS,QAAQ,MAAM;CAE9C,OAAO;EAAE,OADG,KAAK,OAAO,IAAI,GAAG,KACb;EAAG,QAAQ,MAAM;CAAO;AAC5C;;;;;AAMA,SAAgB,WAAW,MAAc,UAAyB,CAAC,GAAW;CAE5E,OAAO,eAAe,MAAM,IADd,SAAS,SAAS,QAAQ,MACJ,CAAC;AACvC;;AAGA,SAAgB,qBAAqB,UAA4B,WAA4B;CAC3F,OAAO,UAAU,OAAO;AAC1B;AAaA,SAAS,KAAK,OAAkB,MAAc,UAA0B,UAAwB;CAC9F,MAAM,WAAW,MAAM,MAAM,SAAS;CACtC,IAAI,CAAC,UAAU;CACf,MAAM,KAAK,GAAG,SAAS,QAAQ;CAC/B,MAAM,WAAW,MAAM,SAAS,IAAI,EAAE;CACtC,IAAI,UAAU;EACZ,SAAS,SAAS;EAClB,IAAI,SAAS,MAAM,SAAS,GAAG,SAAS,MAAM,KAAK,IAAI;EACvD;CACF;CACA,MAAM,SAAS,IAAI,IAAI;EAAE;EAAU;EAAU;EAAU,OAAO;EAAG,OAAO,CAAC,IAAI;CAAE,CAAC;AAClF;AAEA,SAAS,WAAW,MAAc,MAAc,OAAwB;CACtE,IAAI,cAAc,IAAI,GAAG;CACzB,MAAM,UAAU,aAAa,IAAI;CACjC,IAAI,SAAS;EACX,YAAY,SAAS,MAAM,KAAK;EAChC;CACF;CACA,MAAM,aAAa,aAAa,MAAM,MAAM,YAAY;CACxD,IAAI,YAAY;EACd,KAAK,OAAO,MAAM,cAAc,UAAU;EAC1C;CACF;CACA,KAAK,MAAM,YAAY,mBAAmB,IAAI,GAAG,KAAK,OAAO,MAAM,iBAAiB,QAAQ;AAC9F;AAEA,SAAS,YACP,SACA,MACA,OACM;CACN,IAAI,oBAAoB,KAAK,QAAQ,IAAI,GAAG;EAC1C,KAAK,OAAO,MAAM,SAAS,OAAO;EAClC;CACF;CACA,IAAI,oBAAoB,KAAK,QAAQ,IAAI,GAAG;EAC1C,MAAM,UAAU,qBAAqB,QAAQ,OAAO;EACpD,IAAI,YAAY,KAAA,GAAW;GACzB,MAAM,WAAW,KAAK,GAAG,QAAQ,IAAI,uCAAuC;GAC5E;EACF;EACA,IAAI,QAAQ,SAAS,oBACnB,IAAI;GACF,KAAK,KAAK,MAAM,OAAO,GAAG,MAAM,GAAG,KAAK;GACxC;EACF,QAAQ,CAER;EAEF,WAAW,SAAS,MAAM,KAAK;EAC/B;CACF;CAMA,IAAI,MAAM,YAAY,WACpB,KAAK,OAAO,MAAM,SAAS,OAAO;MAElC,MAAM,WAAW,KAAK,GAAG,QAAQ,IAAI,IAAI,QAAQ,KAAK,yBAAyB;AAEnF;AAEA,SAAS,UACP,KACA,OACA,MACA,OACA,OACM;CACN,IAAI,UAAU,QAAQ,UAAU,KAAA,KAAa,OAAO,UAAU,WAAW;CACzE,IAAI,OAAO,UAAU,YAAY,cAAc,KAAK,GAAG;CACvD,MAAM,WAAW,YAAY,GAAG;CAChC,IAAI,aAAa,gBAAgB,OAAO,UAAU,UAAU;EAC1D,KAAK,OAAO,MAAM,cAAc,gBAAgB;EAChD;CACF;CACA,MAAM,SAAS,OAAO,UAAU,YAAY,OAAO,UAAU;CAC7D,IAAI,aAAa,mBAAmB,QACrB;MAAA,gBAAgB,GAEtB,MAAM,WAAW,MAAM,YAAY,WAAW;GACnD,KAAK,OAAO,MAAM,iBAAiB,mBAAmB;GACtD;EACF;;CAEF,IAAI,aAAa,gBAAgB,QAAQ,KAAK,OAAO,MAAM,cAAc,gBAAgB;CACzF,IAAI,aAAa,iBAAiB,OAAO,UAAU,UACjD,KAAK,OAAO,MAAM,eAAe,iBAAiB;CACpD,KAAK,OAAO,MAAM,QAAQ,GAAG,KAAK;AACpC;AAEA,SAAS,KAAK,OAAgB,MAAc,OAAe,OAAwB;CACjF,IAAI,OAAO,UAAU,UAAU;EAC7B,WAAW,OAAO,MAAM,KAAK;EAC7B;CACF;CACA,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;CACjD,IAAI,YAAY,OAAO,KAAK,KAAK,iBAAiB,aAAa;EAC7D,MAAM,WAAW,KAAK,GAAG,QAAQ,IAAI,cAAc;EACnD;CACF;CACA,IAAI,iBAAiB,MAAM;CAC3B,IAAI,QAAQ,WAAW;EACrB,MAAM,WAAW,KAAK,GAAG,QAAQ,IAAI,uBAAuB,UAAU,QAAQ;EAC9E;CACF;CACA,IAAI,MAAM,KAAK,IAAI,KAAK,GAAG;EACzB,MAAM,WAAW,KAAK,GAAG,QAAQ,IAAI,qBAAqB;EAC1D;CACF;CACA,MAAM,KAAK,IAAI,KAAK;CACpB,IAAI;EACF,IAAI,MAAM,QAAQ,KAAK,GAAG;GACxB,KAAK,MAAM,CAAC,OAAO,SAAS,MAAM,QAAQ,GACxC,KAAK,MAAM,QAAQ,MAAM,KAAK,GAAG,QAAQ,GAAG,KAAK;GACnD;EACF;EACA,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,KAAgC,GAAG;GAG1E,MAAM,gBAAgB,aAAa,KAAK,MAAM,YAAY;GAC1D,MAAM,kBAAkB,gBAAgB,KAAA,IAAY,mBAAmB,GAAG,CAAC,CAAC;GAC5E,MAAM,cAAc,iBAAiB;GACrC,MAAM,UAAU,cAAc,OAAO,WAAW,IAAI;GACpD,IAAI,eAAe,KAAK,OAAO,QAAQ,MAAM,OAAO,GAAG,cAAc,aAAa;QAC7E,IAAI,iBACP,KAAK,OAAO,QAAQ,MAAM,OAAO,GAAG,iBAAiB,eAAe;GACtE,UAAU,KAAK,MAAM,QAAQ,MAAM,OAAO,GAAG,OAAO,KAAK;EAC3D;CACF,UAAU;EACR,MAAM,KAAK,OAAO,KAAK;CACzB;AACF;;;;;;;AAQA,SAAgB,kBAAkB,OAAgB,UAAyB,CAAC,GAAuB;CACjG,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,QAAQ,cAAc;CAC5B,IAAI,CAAC,OAAO,MAAM,IAAI,MAAM,8BAA8B,OAAO,OAAO,GAAG;CAC3E,MAAM,QAAmB;EACvB;EACA;EACA,cAAc,iBAAiB,QAAQ,YAAY;EACnD,0BAAU,IAAI,IAAI;EAClB,YAAY,CAAC;EACb,sBAAM,IAAI,QAAQ;CACpB;CACA,KAAK,OAAO,IAAI,GAAG,KAAK;CACxB,OAAO,YAAY,SAAS,CAAC,GAAG,MAAM,SAAS,OAAO,CAAC,GAAG,MAAM,UAAU;AAC5E;;;;;AAMA,SAAgB,gBACd,SACA,UACA,aAAgC,CAAC,GACb;CACpB,MAAM,2BAAW,IAAI,IAAgC;CACrD,KAAK,MAAM,WAAW,UACpB,KAAK,MAAM,WAAW,QAAQ,UAAU;EACtC,MAAM,KAAK,GAAG,QAAQ,SAAS,QAAQ,QAAQ;EAC/C,MAAM,WAAW,SAAS,IAAI,EAAE;EAChC,IAAI,CAAC,UAAU;GACb,SAAS,IAAI,IAAI;IAAE,GAAG;IAAS,OAAO,CAAC,GAAG,QAAQ,KAAK;GAAE,CAAC;GAC1D;EACF;EACA,SAAS,SAAS,QAAQ;EAC1B,KAAK,MAAM,QAAQ,QAAQ,OAAO,IAAI,SAAS,MAAM,SAAS,GAAG,SAAS,MAAM,KAAK,IAAI;CAC3F;CAEF,OAAO,YACL,SACA,CAAC,GAAG,SAAS,OAAO,CAAC,GACrB,CAAC,GAAG,SAAS,SAAS,YAAY,QAAQ,UAAU,GAAG,GAAG,UAAU,CACtE;AACF;AAEA,SAAS,YACP,SACA,UACA,YACoB;CASpB,OAAO;EAAE,QAPP,WAAW,SAAS,IAChB,YACA,SAAS,MAAM,YAAY,QAAQ,aAAa,OAAO,IACrD,WACA,SAAS,SAAS,IAChB,uBACA;EACO;EAAS,UAAU,CAAC,GAAG,QAAQ;EAAG,YAAY,CAAC,GAAG,UAAU;CAAE;AACjF;;AAGA,SAAgB,aAAa,SAAsC;CACjE,OAAO,QAAQ,WAAW,UAAU,QAAQ,WAAW;AACzD;;;;;;AAOA,SAAgB,eACd,OACA,UAAyB,CAAC,GAC0C;CACpE,MAAM,WAAW,OAAO,OAAO,OAAO;CACtC,MAAM,UAAU,kBAAkB,SAAS,OAAO;EAChD,SAAS,QAAQ;EACjB,cAAc,QAAQ;CACxB,CAAC;CACD,OAAO;EAAE,OAAO,SAAS;EAAO,QAAQ,SAAS;EAAQ;CAAQ;AACnE"}

package/dist/release-confidence-BK2QCB5M.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"release-confidence-BK2QCB5M.d.ts","names":[],"sources":["../src/statistics/rank-tests.ts","../src/promotion-gate.ts","../src/release-confidence.ts"],"mappings":";;;;;;KAiCY;;;;;KAMA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;UAYe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;KCpLS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;;EAGA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA0Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;;KC7K5C;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB"}

package/dist/release-confidence-Dbe9AS_X.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"release-confidence-Dbe9AS_X.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.usd === null ? [] : [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;EACxC,MAAM,YAAY,SAAS,WAAW,GAAG;EACzC,OAAO,KAAK,KAAK,SAAS,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,QAAQ,OAAO,CAAC,IAAI,CAAC,IAAI,eAAe,GAAG,CAChE;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"}