@tangle-network/agent-eval 0.129.0 → 0.130.1

package/dist/chunk-NY44NC4A.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/wire/schemas.ts","../src/wire/rubrics.ts","../src/wire/handlers.ts","../src/wire/openapi.ts","../src/wire/rpc.ts","../src/wire/server.ts"],"sourcesContent":["/**\n * Wire-protocol schemas.\n *\n * These Zod schemas are the contract between the agent-eval runtime and\n * any non-TypeScript client (Python, Rust, Go, …). They get rendered to\n * OpenAPI by `wire/openapi.ts` and code-generators consume that spec to\n * produce typed clients in other languages.\n *\n * Rule: if it's not in this file, it isn't on the wire. Keep names and\n * shapes self-explanatory — every field has a `.describe()` so the\n * generated docs are useful without reading the source.\n */\nimport { extendZodWithOpenApi } from '@asteasolutions/zod-to-openapi'\nimport { z } from 'zod'\n\nextendZodWithOpenApi(z)\n\n// ── Building blocks ─────────────────────────────────────────────────\n\nexport const RubricDimensionSchema = z\n  .object({\n    id: z\n      .string()\n      .min(1)\n      .describe('Short stable id like \"buyer_quality\" — used as the key in scoring output.'),\n    description: z\n      .string()\n      .min(1)\n      .describe('One-line plain-English meaning. Read by humans reviewing low scores.'),\n    weight: z\n      .number()\n      .min(0)\n      .default(1)\n      .describe('Relative weight in the composite score. Default 1; 0 disables.'),\n    min: z.number().default(0).describe('Lower bound of valid score for this dimension.'),\n    max: z.number().default(1).describe('Upper bound of valid score for this dimension.'),\n  })\n  .openapi('RubricDimension')\n\nexport const FailureModeSchema = z\n  .object({\n    id: z.string().min(1).describe('Short stable id like \"ai-cadence\" — used in detection lists.'),\n    description: z.string().min(1).describe('Plain-English description of the failure pattern.'),\n  })\n  .openapi('FailureMode')\n\n// ── Rubric ──────────────────────────────────────────────────────────\n\nexport const RubricSchema = z\n  .object({\n    name: z\n      .string()\n      .min(1)\n      .describe('Stable name like \"anti-slop\" — used by clients to invoke this rubric.'),\n    description: z\n      .string()\n      .min(1)\n      .describe('What this rubric measures. Shown in /v1/rubrics listing.'),\n    systemPrompt: z\n      .string()\n      .min(1)\n      .describe(\n        'Instructs the judging LLM. Should explain the persona (e.g. \"senior engineer reviewing voice\"), what to score on, and what to return.',\n      ),\n    dimensions: z\n      .array(RubricDimensionSchema)\n      .min(1)\n      .describe('Scoring axes. The composite score is a weighted sum of these.'),\n    failureModes: z\n      .array(FailureModeSchema)\n      .default([])\n      .describe('Patterns to detect; each detected mode appears in the result.failureModes list.'),\n    wins: z\n      .array(FailureModeSchema)\n      .default([])\n      .describe('Positive patterns; each detected one appears in the result.wins list.'),\n  })\n  .openapi('Rubric')\n\n// ── Judge call ──────────────────────────────────────────────────────\n\nexport const JudgeRequestSchema = z\n  .object({\n    rubricName: z\n      .string()\n      .optional()\n      .describe('Use a built-in rubric by name. Mutually exclusive with `rubric`.'),\n    rubric: RubricSchema.optional().describe(\n      'Inline rubric definition. Mutually exclusive with `rubricName`.',\n    ),\n    content: z\n      .string()\n      .min(1)\n      .describe('The text being judged — a tweet, a blog post, a code snippet, anything stringly.'),\n    context: z\n      .record(z.string(), z.unknown())\n      .optional()\n      .describe(\n        'Free-form metadata for the rubric to use — analytics, source URL, author, etc. Surfaced to the LLM.',\n      ),\n    model: z.string().optional().describe('Override the judge model configured by the server.'),\n  })\n  .refine((v) => Boolean(v.rubricName) !== Boolean(v.rubric), {\n    message: 'Provide exactly one of `rubricName` or `rubric`.',\n  })\n  .openapi('JudgeRequest')\n\nexport const JudgeResultSchema = z\n  .object({\n    composite: z\n      .number()\n      .min(0)\n      .max(1)\n      .describe('Weighted combination of dimension scores in 0..1. The single number to gate on.'),\n    dimensions: z\n      .record(z.string(), z.number())\n      .describe('Per-dimension score, keyed by RubricDimension.id.'),\n    failureModes: z\n      .array(z.string())\n      .default([])\n      .describe('Failure-mode ids detected in the content (subset of rubric.failureModes ids).'),\n    wins: z\n      .array(z.string())\n      .default([])\n      .describe('Win ids detected in the content (subset of rubric.wins ids).'),\n    rationale: z\n      .string()\n      .describe('Plain-English explanation of the score. Surfaced to the human reviewer.'),\n    rubricVersion: z\n      .string()\n      .describe(\n        'Stable hash of the rubric used. Scores are only comparable across runs when this matches.',\n      ),\n    model: z.string().describe('Model that produced the judgement, for reproducibility.'),\n    durationMs: z.number().int().nonnegative().describe('End-to-end wall time for this call.'),\n  })\n  .openapi('JudgeResult')\n\n// ── Rubric listing ──────────────────────────────────────────────────\n\nexport const RubricInfoSchema = z\n  .object({\n    name: z.string().describe('Pass this to /v1/judge as `rubricName`.'),\n    description: z.string().describe('What this rubric measures.'),\n    dimensions: z\n      .array(z.object({ id: z.string(), description: z.string(), weight: z.number() }))\n      .describe('The scoring axes this rubric uses, with weights.'),\n    failureModes: z.array(z.string()).default([]).describe('Failure-mode ids this rubric detects.'),\n    rubricVersion: z.string().describe('Stable hash — match this to compare scores across runs.'),\n  })\n  .openapi('RubricInfo')\n\nexport const ListRubricsResponseSchema = z\n  .object({\n    rubrics: z.array(RubricInfoSchema),\n  })\n  .openapi('ListRubricsResponse')\n\n// ── Version / health ────────────────────────────────────────────────\n\nexport const VersionResponseSchema = z\n  .object({\n    package: z.string().describe('Package name (always \"@tangle-network/agent-eval\").'),\n    version: z.string().describe('Semver of the running server. Match your client to this.'),\n    wireVersion: z\n      .string()\n      .describe(\n        'Wire-protocol semver. Bumps separately from package version when the schema changes.',\n      ),\n    apiSurface: z.array(z.string()).describe('List of supported method names.'),\n  })\n  .openapi('VersionResponse')\n\nexport const HealthResponseSchema = z\n  .object({\n    status: z.literal('ok'),\n    uptimeSec: z.number(),\n  })\n  .openapi('HealthResponse')\n\n// ── Ingestion: production traces + user feedback ────────────────────\n\n/**\n * Minimal `TraceEvent` shape that the production runtime emits.\n * Matches `trace/schema.ts` `TraceEvent` but is duplicated here as a\n * wire schema so non-TypeScript clients can validate without depending\n * on internal types.\n */\nexport const TraceEventSchema = z\n  .object({\n    eventId: z.string().min(1).describe('Stable id for the event. Use ULID or UUID.'),\n    runId: z.string().min(1).describe('Run this event belongs to.'),\n    spanId: z.string().optional().describe('Span that emitted the event, if any.'),\n    kind: z\n      .enum([\n        'log',\n        'error',\n        'budget_decrement',\n        'budget_breach',\n        'state_mutation',\n        'policy_violation',\n        'redaction_applied',\n        'custom',\n      ])\n      .describe('Coarse event category — matches the TraceSchema v1 EventKind enum.'),\n    timestamp: z\n      .number()\n      .int()\n      .nonnegative()\n      .describe('Unix millis. Must be monotonically non-decreasing within a span.'),\n    payload: z\n      .record(z.string(), z.unknown())\n      .describe('Free-form payload — the runtime owns the shape.'),\n  })\n  .openapi('TraceEvent')\n\nexport const TracesIngestRequestSchema = z\n  .object({\n    events: z\n      .array(TraceEventSchema)\n      .min(1)\n      .max(10_000)\n      .describe('Batch of events. Max 10k per call — bigger streams should be chunked.'),\n  })\n  .openapi('TracesIngestRequest')\n\nexport const TracesIngestResponseSchema = z\n  .object({\n    accepted: z.number().int().nonnegative().describe('Number of events persisted.'),\n    rejected: z\n      .number()\n      .int()\n      .nonnegative()\n      .describe('Number of events the store refused — see `errors[]` for reasons.'),\n    errors: z\n      .array(\n        z.object({\n          eventId: z.string().describe('Event id this error applies to.'),\n          message: z.string().describe('Why the event was rejected.'),\n        }),\n      )\n      .default([]),\n  })\n  .openapi('TracesIngestResponse')\n\nexport const FeedbackLabelSchema = z\n  .object({\n    id: z.string().optional(),\n    source: z.enum(['user', 'judge', 'environment', 'metric', 'policy', 'system']),\n    kind: z.enum([\n      'approve',\n      'reject',\n      'select',\n      'edit',\n      'rank',\n      'rate',\n      'comment',\n      'metric_outcome',\n      'policy_block',\n      'revision_request',\n    ]),\n    value: z.unknown(),\n    reason: z.string().optional(),\n    severity: z.enum(['info', 'warning', 'error', 'critical']).optional(),\n    createdAt: z.string().describe('ISO-8601 UTC.'),\n    metadata: z.record(z.string(), z.unknown()).optional(),\n  })\n  .openapi('FeedbackLabel')\n\nexport const FeedbackAttemptSchema = z\n  .object({\n    id: z.string().min(1),\n    stepIndex: z.number().int().nonnegative(),\n    artifactType: z.enum([\n      'text',\n      'code',\n      'plan',\n      'research',\n      'action',\n      'ui',\n      'decision',\n      'data',\n      'other',\n    ]),\n    artifact: z.unknown(),\n    options: z.array(z.unknown()).optional(),\n    proposedAction: z\n      .object({\n        type: z.string(),\n        risk: z.enum(['low', 'medium', 'high']).optional(),\n        costUsd: z.number().optional(),\n        externalSideEffect: z.boolean().optional(),\n        requiresApproval: z.boolean().optional(),\n        metadata: z.record(z.string(), z.unknown()).optional(),\n      })\n      .optional(),\n    feedback: z.array(FeedbackLabelSchema).optional(),\n    createdAt: z.string(),\n    metadata: z.record(z.string(), z.unknown()).optional(),\n  })\n  .openapi('FeedbackAttempt')\n\nexport const FeedbackTrajectorySchema = z\n  .object({\n    id: z.string().min(1).describe('Stable id; idempotency key for the trajectory.'),\n    projectId: z.string().optional(),\n    scenarioId: z.string().optional(),\n    task: z.object({\n      intent: z.string().min(1),\n      context: z.unknown().optional(),\n    }),\n    attempts: z.array(FeedbackAttemptSchema).default([]),\n    labels: z.array(FeedbackLabelSchema).default([]),\n    outcome: z\n      .object({\n        success: z.boolean().optional(),\n        score: z.number().optional(),\n        metrics: z.record(z.string(), z.number()).optional(),\n        costUsd: z.number().optional(),\n        detail: z.string().optional(),\n        observedAt: z.string().optional(),\n        metadata: z.record(z.string(), z.unknown()).optional(),\n      })\n      .optional(),\n    split: z.enum(['train', 'dev', 'test', 'holdout']).optional(),\n    tags: z.record(z.string(), z.string()).optional(),\n    createdAt: z.string().describe('ISO-8601 UTC.'),\n    updatedAt: z.string().optional(),\n    metadata: z.record(z.string(), z.unknown()).optional(),\n  })\n  .openapi('FeedbackTrajectory')\n\nexport const FeedbackIngestResponseSchema = z\n  .object({\n    id: z.string().describe('Trajectory id that was persisted.'),\n    persisted: z.boolean().describe('True when the trajectory was saved (idempotent on id).'),\n  })\n  .openapi('FeedbackIngestResponse')\n\nexport type TraceEvent = z.infer<typeof TraceEventSchema>\nexport type TracesIngestRequest = z.infer<typeof TracesIngestRequestSchema>\nexport type TracesIngestResponse = z.infer<typeof TracesIngestResponseSchema>\nexport type FeedbackTrajectory = z.infer<typeof FeedbackTrajectorySchema>\nexport type FeedbackIngestResponse = z.infer<typeof FeedbackIngestResponseSchema>\n\n// ── Errors ──────────────────────────────────────────────────────────\n\nexport const ErrorResponseSchema = z\n  .object({\n    error: z\n      .object({\n        code: z\n          .string()\n          .describe(\n            'Machine-readable code: \"validation_error\", \"rubric_not_found\", \"judge_error\".',\n          ),\n        message: z.string().describe('Human-readable message.'),\n        details: z.unknown().optional().describe('Optional structured detail.'),\n      })\n      .describe('Errors are always wrapped in this shape across all endpoints.'),\n  })\n  .openapi('ErrorResponse')\n\n// ── Type exports for callers in the same package ────────────────────\n\nexport type RubricDimension = z.infer<typeof RubricDimensionSchema>\nexport type FailureMode = z.infer<typeof FailureModeSchema>\nexport type Rubric = z.infer<typeof RubricSchema>\nexport type JudgeRequest = z.infer<typeof JudgeRequestSchema>\nexport type JudgeResult = z.infer<typeof JudgeResultSchema>\nexport type RubricInfo = z.infer<typeof RubricInfoSchema>\nexport type ListRubricsResponse = z.infer<typeof ListRubricsResponseSchema>\nexport type VersionResponse = z.infer<typeof VersionResponseSchema>\nexport type ErrorResponse = z.infer<typeof ErrorResponseSchema>\n\n// ── Wire-protocol version ───────────────────────────────────────────\n\n/**\n * Bump on any breaking change to a request/response schema.\n * Non-breaking (additive) changes don't require a bump.\n */\nexport const WIRE_VERSION = '1.0.0'\n\n/**\n * Stable hash of a rubric. Used to make scores comparable across runs:\n * if the rubricVersion matches, the rubric was identical.\n */\nexport function hashRubric(rubric: Rubric): string {\n  const stable = stableStringify(rubric)\n  let h = 5381\n  for (let i = 0; i < stable.length; i++) {\n    h = (h * 33) ^ stable.charCodeAt(i)\n  }\n  // Unsigned 32-bit hex, prefixed with rubric name + version slot\n  return `${rubric.name}@${(h >>> 0).toString(16).padStart(8, '0')}`\n}\n\nfunction stableStringify(value: unknown): string {\n  if (Array.isArray(value)) return `[${value.map((item) => stableStringify(item)).join(',')}]`\n  if (value && typeof value === 'object') {\n    const entries = Object.entries(value as Record<string, unknown>)\n      .sort(([a], [b]) => a.localeCompare(b))\n      .map(([key, item]) => `${JSON.stringify(key)}:${stableStringify(item)}`)\n    return `{${entries.join(',')}}`\n  }\n  return JSON.stringify(value)\n}\n","/**\n * Built-in rubrics shipped with agent-eval.\n *\n * A rubric is a set of scoring axes plus a system prompt that tells the\n * judging LLM how to grade against those axes. Built-in rubrics are\n * curated for use cases that recur across Tangle projects — call them\n * by name from any client.\n *\n * Adding a rubric:\n *   1. Define the Rubric object below with a clear `description` and\n *      named `dimensions`.\n *   2. Register it in `BUILTIN_RUBRICS` at the bottom.\n *   3. Add a test in `tests/wire/rubrics.test.ts`.\n *\n * Custom rubrics: callers pass `rubric` inline to /v1/judge instead of\n * `rubricName` — see schemas.ts.\n */\nimport type { Rubric } from './schemas'\nimport { hashRubric } from './schemas'\n\n// ── anti-slop ───────────────────────────────────────────────────────\n// Voice/style judge tuned for technical-buyer audiences. Used by the\n// Postiz autoresearch loop and any content-quality gate.\n\nconst ANTI_SLOP: Rubric = {\n  name: 'anti-slop',\n  description:\n    'Voice and signal quality for content aimed at senior engineers. Catches AI cadence, marketing tone, and engagement-bait shapes.',\n  systemPrompt: `You are evaluating a piece of content written for senior engineers and technical founders.\n\nYou score three things:\n- buyer_quality (0..1): would a senior engineer in the target ICP find this worth their attention? High = specific, earned, technically interesting. Low = generic, hyped, off-target.\n- voice (0..1): does it read like a person who built the thing, or like AI/marketing copy?\n- signal (0..1): does it contain a non-obvious detail, constraint, or claim a reader couldn't get from the public docs?\n\nDetect failure modes (return ids matching):\n- ai-cadence: rule-of-three openings, em-dash flourish, \"Let me explain\", \"Here's the thing\", AI rhythm\n- marketing-tone: \"We're excited to announce\", \"thrilled\", \"delighted\", \"game-changer\", buzzword stack\n- vague-claim: technical claim without a specific component, file, or measurement\n- no-hook: opening doesn't earn attention from the target reader\n- engagement-bait: \"agree?\", \"thoughts?\", listicles, controversy-fishing, hook-detail-pitch\n- off-icp: content shape would attract motivational/grift/hype audiences instead of buyers\n- stale-claim: repeats a positioning line we've used many times this month\n\nDetect wins (return ids matching):\n- specific-component: names a real file, component, or measurement\n- earned-detail: shares a non-obvious detail not derivable from public docs\n- constraint-articulated: names a real tradeoff and the side chosen\n- honest-failure: describes a real failure mode and what was done about it\n\nReturn ONLY JSON matching the response schema. Be conservative — most content has 0-1 wins and 1-2 failure modes, not many of each.`,\n  dimensions: [\n    {\n      id: 'buyer_quality',\n      description: 'Would the target buyer find this worth attention?',\n      weight: 0.5,\n      min: 0,\n      max: 1,\n    },\n    {\n      id: 'voice',\n      description: 'Does it sound like a builder, not AI or marketing?',\n      weight: 0.3,\n      min: 0,\n      max: 1,\n    },\n    {\n      id: 'signal',\n      description: 'Non-obvious detail, constraint, or claim?',\n      weight: 0.2,\n      min: 0,\n      max: 1,\n    },\n  ],\n  failureModes: [\n    { id: 'ai-cadence', description: 'AI-rhythm openings and transitions' },\n    { id: 'marketing-tone', description: 'Buzzwords, hype, corporate-PR voice' },\n    { id: 'vague-claim', description: 'Technical claim without specifics' },\n    { id: 'no-hook', description: 'Opening fails to earn attention' },\n    { id: 'engagement-bait', description: 'Listicle/controversy/agree-pattern' },\n    { id: 'off-icp', description: 'Voice attracts the wrong audience' },\n    { id: 'stale-claim', description: 'Reuses an over-used positioning line' },\n  ],\n  wins: [\n    { id: 'specific-component', description: 'Names a real file/component/number' },\n    { id: 'earned-detail', description: 'Detail not in public docs' },\n    { id: 'constraint-articulated', description: 'Names a real tradeoff' },\n    { id: 'honest-failure', description: 'Describes a real failure honestly' },\n  ],\n}\n\n// ── Registry ────────────────────────────────────────────────────────\n\nexport const BUILTIN_RUBRICS: Record<string, Rubric> = {\n  'anti-slop': ANTI_SLOP,\n}\n\n/** Get a built-in rubric by name, or undefined. */\nexport function getBuiltinRubric(name: string): Rubric | undefined {\n  return BUILTIN_RUBRICS[name]\n}\n\n/** List built-in rubrics with their stable versions. */\nexport function listBuiltinRubrics() {\n  return Object.values(BUILTIN_RUBRICS).map((r) => ({\n    name: r.name,\n    description: r.description,\n    dimensions: r.dimensions.map((d) => ({\n      id: d.id,\n      description: d.description,\n      weight: d.weight,\n    })),\n    failureModes: r.failureModes.map((f) => f.id),\n    rubricVersion: hashRubric(r),\n  }))\n}\n","/**\n * Pure handler functions — the \"business logic\" behind every wire-protocol\n * method. The HTTP server (`server.ts`) and the stdio RPC (`rpc.ts`) both\n * call these. Tests call these directly without spinning a server.\n *\n * Each handler:\n *   - Takes a parsed request (already Zod-validated by the transport).\n *   - Returns a result that matches the response schema.\n *   - Throws `WireError` for caller-fixable errors (404, 400, 422).\n *   - Lets unexpected errors bubble — the transport maps them to 500.\n */\n\nimport { CostLedger, type CostLedgerHandle } from '../cost-ledger'\nimport type { FeedbackTrajectoryStore } from '../feedback-trajectory'\nimport {\n  assertLlmRoute,\n  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\n  LlmRouteAssertionError,\n  type LlmRouteRequirements,\n  maximumChargeForLlmRequest,\n} from '../llm-client'\nimport type { TraceEvent as InternalTraceEvent } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { getBuiltinRubric, listBuiltinRubrics } from './rubrics'\nimport {\n  type FeedbackIngestResponse,\n  hashRubric,\n  type JudgeRequest,\n  type JudgeResult,\n  type ListRubricsResponse,\n  type Rubric,\n  type TracesIngestRequest,\n  type TracesIngestResponse,\n  type VersionResponse,\n  WIRE_VERSION,\n  type FeedbackTrajectory as WireFeedbackTrajectory,\n} from './schemas'\n\n/** Caller-fixable error. The transport renders this to 4xx + ErrorResponse. */\nexport class WireError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n    public readonly status: number = 400,\n    public readonly details?: unknown,\n  ) {\n    super(message)\n    this.name = 'WireError'\n  }\n}\n\n// ── judge ───────────────────────────────────────────────────────────\n\n/** The JSON schema we ask the judging LLM to fill in. */\nfunction judgeOutputSchema(rubric: Rubric) {\n  return {\n    name: 'JudgeOutput',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      properties: {\n        dimensions: {\n          type: 'object',\n          additionalProperties: false,\n          properties: Object.fromEntries(\n            rubric.dimensions.map((d) => [\n              d.id,\n              { type: 'number', minimum: d.min, maximum: d.max },\n            ]),\n          ),\n          required: rubric.dimensions.map((d) => d.id),\n        },\n        failureModes: {\n          type: 'array',\n          items: { type: 'string', enum: rubric.failureModes.map((f) => f.id) },\n        },\n        wins: {\n          type: 'array',\n          items: { type: 'string', enum: rubric.wins.map((w) => w.id) },\n        },\n        rationale: { type: 'string' },\n      },\n      required: ['dimensions', 'rationale'],\n    } as Record<string, unknown>,\n  }\n}\n\ninterface JudgeOutput {\n  dimensions: Record<string, number>\n  failureModes?: string[]\n  wins?: string[]\n  rationale: string\n}\n\nfunction validateJudgeOutput(value: unknown, rubric: Rubric): JudgeOutput {\n  if (!value || typeof value !== 'object') {\n    throw new WireError('judge_error', 'Judge returned malformed output.', 500, value)\n  }\n  const raw = value as Record<string, unknown>\n  const rawDimensions = raw.dimensions\n  if (!rawDimensions || typeof rawDimensions !== 'object' || Array.isArray(rawDimensions)) {\n    throw new WireError('judge_error', 'Judge returned malformed dimensions.', 500, value)\n  }\n\n  const dimensions: Record<string, number> = {}\n  const dimensionRecord = rawDimensions as Record<string, unknown>\n  for (const dim of rubric.dimensions) {\n    const score = dimensionRecord[dim.id]\n    if (\n      typeof score !== 'number' ||\n      !Number.isFinite(score) ||\n      score < dim.min ||\n      score > dim.max\n    ) {\n      throw new WireError(\n        'judge_error',\n        `Judge returned invalid score for dimension \"${dim.id}\".`,\n        500,\n        value,\n      )\n    }\n    dimensions[dim.id] = score\n  }\n\n  const allowedFailures = new Set(rubric.failureModes.map((mode) => mode.id))\n  const allowedWins = new Set(rubric.wins.map((win) => win.id))\n  const failureModes = validateIdArray(raw.failureModes, allowedFailures, 'failureModes', value)\n  const wins = validateIdArray(raw.wins, allowedWins, 'wins', value)\n  if (typeof raw.rationale !== 'string' || raw.rationale.trim().length === 0) {\n    throw new WireError('judge_error', 'Judge returned missing rationale.', 500, value)\n  }\n\n  return { dimensions, failureModes, wins, rationale: raw.rationale }\n}\n\nfunction validateIdArray(\n  raw: unknown,\n  allowed: Set<string>,\n  field: 'failureModes' | 'wins',\n  original: unknown,\n): string[] {\n  if (raw === undefined) return []\n  if (!Array.isArray(raw)) {\n    throw new WireError('judge_error', `Judge returned non-array ${field}.`, 500, original)\n  }\n  const out: string[] = []\n  for (const item of raw) {\n    if (typeof item !== 'string' || !allowed.has(item)) {\n      throw new WireError(\n        'judge_error',\n        `Judge returned unknown ${field} id \"${String(item)}\".`,\n        500,\n        original,\n      )\n    }\n    out.push(item)\n  }\n  return out\n}\n\nfunction compositeScore(dimensions: Record<string, number>, rubric: Rubric): number {\n  let weighted = 0\n  let totalWeight = 0\n  for (const dim of rubric.dimensions) {\n    const raw = dimensions[dim.id] ?? 0\n    const range = dim.max - dim.min || 1\n    const normalized = Math.max(0, Math.min(1, (raw - dim.min) / range))\n    weighted += normalized * dim.weight\n    totalWeight += dim.weight\n  }\n  return totalWeight > 0 ? weighted / totalWeight : 0\n}\n\nfunction buildJudgePrompt(content: string, context: unknown): string {\n  const ctx = context && Object.keys(context as object).length ? JSON.stringify(context) : ''\n  return [\n    `CONTENT TO JUDGE:`,\n    content,\n    '',\n    ctx ? `CONTEXT (metadata, analytics, etc.):` : '',\n    ctx ? ctx : '',\n  ]\n    .filter(Boolean)\n    .join('\\n')\n}\n\nconst DEFAULT_JUDGE_MODEL = 'claude-sonnet-4-6'\n\nexport interface HandleJudgeOptions {\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  llm?: LlmClientOptions\n  defaultModel?: string\n  routeRequirements?: LlmRouteRequirements\n  signal?: AbortSignal\n}\n\nexport async function handleJudge(\n  req: JudgeRequest,\n  options: HandleJudgeOptions = {},\n): Promise<JudgeResult> {\n  // Resolve rubric\n  let rubric: Rubric\n  if (req.rubricName) {\n    const found = getBuiltinRubric(req.rubricName)\n    if (!found) {\n      throw new WireError('rubric_not_found', `No built-in rubric named \"${req.rubricName}\".`, 404)\n    }\n    rubric = found\n  } else if (req.rubric) {\n    rubric = req.rubric\n  } else {\n    // refine() in the schema should already have caught this — defense in depth\n    throw new WireError('validation_error', 'Provide either `rubricName` or `rubric`.', 422)\n  }\n\n  if (options.routeRequirements) {\n    try {\n      assertLlmRoute(options.llm ?? {}, options.routeRequirements)\n    } catch (error) {\n      if (!(error instanceof LlmRouteAssertionError)) throw error\n      const noEndpoint = error.reason === 'no_explicit_base_url'\n      throw new WireError(\n        noEndpoint ? 'llm_not_configured' : 'llm_route_rejected',\n        noEndpoint\n          ? 'No model endpoint is configured. Pass llm.baseUrl or configure the CLI provider environment variables.'\n          : error.message,\n        503,\n        { reason: error.reason },\n      )\n    }\n  }\n\n  const startedAt = Date.now()\n  const model = req.model ?? options.defaultModel ?? DEFAULT_JUDGE_MODEL\n\n  const request = {\n    model,\n    messages: [\n      { role: 'system', content: rubric.systemPrompt },\n      { role: 'user', content: buildJudgePrompt(req.content, req.context) },\n    ],\n    jsonSchema: judgeOutputSchema(rubric),\n    temperature: 0.0,\n    maxTokens: 4_000,\n    timeoutMs: 60_000,\n  } satisfies LlmCallRequest\n  const ledger = options.costLedger ?? new CostLedger()\n  const paid = await ledger.runPaidCall({\n    channel: 'judge',\n    phase: options.costPhase ?? 'wire.judge',\n    actor: `wire.${req.rubricName ?? 'inline'}`,\n    model,\n    maximumCharge: maximumChargeForLlmRequest(request, options.llm),\n    signal: options.signal,\n    execute: (signal, callId) =>\n      callLlmJson<JudgeOutput>(request, {\n        ...options.llm,\n        signal,\n        idempotencyKey: callId,\n      }),\n    receipt: ({ result }) => costReceiptFromLlm(result),\n    receiptFromError: costReceiptFromLlmError,\n  })\n  if (!paid.succeeded) throw paid.error\n  const { value, result } = paid.value\n\n  const output = validateJudgeOutput(value, rubric)\n\n  const composite = compositeScore(output.dimensions, rubric)\n  const durationMs = Date.now() - startedAt\n\n  return {\n    composite,\n    dimensions: output.dimensions,\n    failureModes: output.failureModes ?? [],\n    wins: output.wins ?? [],\n    rationale: output.rationale,\n    rubricVersion: hashRubric(rubric),\n    model: result.model,\n    durationMs,\n  }\n}\n\n// ── listRubrics ─────────────────────────────────────────────────────\n\nexport function handleListRubrics(): ListRubricsResponse {\n  return { rubrics: listBuiltinRubrics() }\n}\n\n// ── version ─────────────────────────────────────────────────────────\n\nimport { readFileSync } from 'node:fs'\nimport { dirname, resolve } from 'node:path'\nimport { fileURLToPath } from 'node:url'\n\nlet CACHED_VERSION: string | undefined\n\nfunction readPackageVersion(): string {\n  if (CACHED_VERSION) return CACHED_VERSION\n  // Walk up from this file looking for the nearest package.json.\n  // In dist/ this is dist/.., in src/wire/ this is ../../package.json.\n  const here = dirname(fileURLToPath(import.meta.url))\n  const candidates = [\n    resolve(here, '..', '..', 'package.json'), // src/wire → repo root\n    resolve(here, '..', 'package.json'), // dist → repo root\n  ]\n  for (const path of candidates) {\n    try {\n      const pkg = JSON.parse(readFileSync(path, 'utf-8')) as { version?: string }\n      if (pkg.version) {\n        CACHED_VERSION = pkg.version\n        return pkg.version\n      }\n    } catch {\n      // try next\n    }\n  }\n  return '0.0.0-unknown'\n}\n\nexport function handleVersion(): VersionResponse {\n  return {\n    package: '@tangle-network/agent-eval',\n    version: readPackageVersion(),\n    wireVersion: WIRE_VERSION,\n    apiSurface: ['judge', 'listRubrics', 'version', 'feedback.ingest', 'traces.ingest'],\n  }\n}\n\n// ── Ingestion handlers ───────────────────────────────────────────────\n\n/**\n * Pluggable stores the wire layer routes ingestion writes into. Both\n * are optional — when omitted, the corresponding endpoint returns 503.\n *\n * Production deployments wire a `FileSystemTraceStore` and\n * `FileSystemFeedbackTrajectoryStore` here. Tests substitute in-memory\n * stores.\n */\nexport interface IngestionStores {\n  traceStore?: TraceStore\n  feedbackStore?: FeedbackTrajectoryStore\n}\n\n/**\n * `POST /v1/traces/ingest` — accept a batch of `TraceEvent`s from the\n * production runtime. Best-effort: each event is appended independently;\n * one bad event does not poison the batch.\n *\n * Idempotency: the underlying store is append-only; consumers retrying\n * the same payload will get duplicate events. Consumers should\n * de-duplicate by `eventId` downstream — production traces frequently\n * land via at-least-once buses (Kafka, SQS) where dedup is unavoidable.\n */\nexport async function handleTracesIngest(\n  req: TracesIngestRequest,\n  stores: IngestionStores,\n): Promise<TracesIngestResponse> {\n  if (!stores.traceStore) {\n    throw new WireError(\n      'service_unavailable',\n      'No trace store configured on this server. Pass `traceStore` to `createApp`.',\n      503,\n    )\n  }\n  const errors: Array<{ eventId: string; message: string }> = []\n  let accepted = 0\n  for (const event of req.events) {\n    try {\n      // The wire `TraceEvent` is structurally identical to the internal one.\n      await stores.traceStore.appendEvent(event as InternalTraceEvent)\n      accepted++\n    } catch (err) {\n      errors.push({\n        eventId: event.eventId,\n        message: err instanceof Error ? err.message : String(err),\n      })\n    }\n  }\n  return { accepted, rejected: errors.length, errors }\n}\n\n/**\n * `POST /v1/feedback` — accept a single `FeedbackTrajectory` from the\n * production runtime. Idempotent on `id`: re-posting the same trajectory\n * replaces the prior record.\n */\nexport async function handleFeedbackIngest(\n  req: WireFeedbackTrajectory,\n  stores: IngestionStores,\n): Promise<FeedbackIngestResponse> {\n  if (!stores.feedbackStore) {\n    throw new WireError(\n      'service_unavailable',\n      'No feedback store configured on this server. Pass `feedbackStore` to `createApp`.',\n      503,\n    )\n  }\n  // The wire `FeedbackTrajectory` aligns 1:1 with the internal type;\n  // cast through `unknown` since the wire schema is a Zod-inferred\n  // structural type with optional fields the internal store consumes.\n  await stores.feedbackStore.save(req as unknown as Parameters<FeedbackTrajectoryStore['save']>[0])\n  return { id: req.id, persisted: true }\n}\n","/**\n * Build an OpenAPI spec from the wire schemas.\n *\n * The spec is the contract that other-language clients (Python, Rust,\n * Go, …) generate from. There is no hand-written client — clients are\n * derived artifacts of this file plus `schemas.ts`.\n *\n * Run `pnpm openapi` (defined in package.json) to write the spec to\n * `dist/openapi.json`. CI uses that file to regenerate the Python\n * client and gate the dual-publish workflow.\n */\nimport { OpenAPIRegistry, OpenApiGeneratorV31 } from '@asteasolutions/zod-to-openapi'\nimport type { OpenAPIObject } from 'openapi3-ts/oas31'\n\nimport {\n  ErrorResponseSchema,\n  FeedbackIngestResponseSchema,\n  FeedbackTrajectorySchema,\n  HealthResponseSchema,\n  JudgeRequestSchema,\n  JudgeResultSchema,\n  ListRubricsResponseSchema,\n  TracesIngestRequestSchema,\n  TracesIngestResponseSchema,\n  VersionResponseSchema,\n  WIRE_VERSION,\n} from './schemas'\n\nexport function buildOpenApi(packageVersion: string): OpenAPIObject {\n  const registry = new OpenAPIRegistry()\n\n  // Components — each schema becomes a $ref-able component\n  registry.register('JudgeRequest', JudgeRequestSchema)\n  registry.register('JudgeResult', JudgeResultSchema)\n  registry.register('ListRubricsResponse', ListRubricsResponseSchema)\n  registry.register('VersionResponse', VersionResponseSchema)\n  registry.register('HealthResponse', HealthResponseSchema)\n  registry.register('ErrorResponse', ErrorResponseSchema)\n  registry.register('TracesIngestRequest', TracesIngestRequestSchema)\n  registry.register('TracesIngestResponse', TracesIngestResponseSchema)\n  registry.register('FeedbackTrajectory', FeedbackTrajectorySchema)\n  registry.register('FeedbackIngestResponse', FeedbackIngestResponseSchema)\n\n  // Routes\n  registry.registerPath({\n    method: 'post',\n    path: '/v1/judge',\n    summary: 'Score a piece of content against a rubric',\n    description:\n      'Runs the judging LLM with the named (or inline) rubric and returns dimension scores, detected failure modes, wins, and a composite score in 0..1.',\n    request: {\n      body: {\n        content: {\n          'application/json': { schema: JudgeRequestSchema },\n        },\n      },\n    },\n    responses: {\n      200: {\n        description: 'Successful judgement',\n        content: { 'application/json': { schema: JudgeResultSchema } },\n      },\n      400: {\n        description: 'Validation error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      404: {\n        description: 'Rubric not found',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      500: {\n        description: 'Judge error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/v1/rubrics',\n    summary: 'List built-in rubrics',\n    description:\n      'Returns every rubric registered server-side, with their dimensions and stable rubricVersion hash.',\n    responses: {\n      200: {\n        description: 'Listing',\n        content: { 'application/json': { schema: ListRubricsResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/v1/version',\n    summary: 'Server and wire-protocol version',\n    description: 'Match your client version to `version`; check `wireVersion` for compatibility.',\n    responses: {\n      200: {\n        description: 'Version info',\n        content: { 'application/json': { schema: VersionResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/healthz',\n    summary: 'Liveness check',\n    responses: {\n      200: {\n        description: 'OK',\n        content: { 'application/json': { schema: HealthResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'post',\n    path: '/v1/traces/ingest',\n    summary: 'Ingest a batch of production TraceEvents',\n    description:\n      'Append a batch of TraceEvents to the configured TraceStore. Accepts application/json ({events:[...]}) or application/x-ndjson (one event per line). Returns counts of accepted + rejected events.',\n    request: {\n      body: {\n        content: {\n          'application/json': { schema: TracesIngestRequestSchema },\n          'application/x-ndjson': { schema: TracesIngestRequestSchema },\n        },\n      },\n    },\n    responses: {\n      200: {\n        description: 'Ingestion summary',\n        content: { 'application/json': { schema: TracesIngestResponseSchema } },\n      },\n      400: {\n        description: 'Validation error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      401: {\n        description: 'Unauthorized (when bearer auth is configured)',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      503: {\n        description: 'No trace store configured',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'post',\n    path: '/v1/feedback',\n    summary: 'Ingest a FeedbackTrajectory from production',\n    description:\n      'Persist a single FeedbackTrajectory. Idempotent on trajectory.id — re-posting replaces the prior record. Used by production runtimes to forward user 👍/👎/edits into the eval substrate.',\n    request: {\n      body: {\n        content: {\n          'application/json': { schema: FeedbackTrajectorySchema },\n        },\n      },\n    },\n    responses: {\n      200: {\n        description: 'Persisted',\n        content: { 'application/json': { schema: FeedbackIngestResponseSchema } },\n      },\n      400: {\n        description: 'Validation error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      401: {\n        description: 'Unauthorized (when bearer auth is configured)',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      503: {\n        description: 'No feedback store configured',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n    },\n  })\n\n  const generator = new OpenApiGeneratorV31(registry.definitions)\n  const doc = generator.generateDocument({\n    openapi: '3.1.0',\n    info: {\n      title: '@tangle-network/agent-eval — wire protocol',\n      version: packageVersion,\n      description: `HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: ${WIRE_VERSION}. Bumps on breaking changes to request/response schemas.`,\n      contact: { name: 'Tangle Network', url: 'https://github.com/tangle-network/agent-eval' },\n      license: { name: 'MIT' },\n    },\n    servers: [{ url: 'http://localhost:5005', description: 'Local agent-eval serve' }],\n  })\n  const rubricRef = { $ref: '#/components/schemas/Rubric' } as const\n  const commonJudgeFields = {\n    content: { type: 'string', minLength: 1 },\n    context: { type: 'object', additionalProperties: true },\n    model: { type: 'string' },\n  } as const\n  doc.components ??= {}\n  doc.components.schemas ??= {}\n  doc.components.schemas.JudgeRequest = {\n    oneOf: [\n      {\n        type: 'object',\n        additionalProperties: false,\n        required: ['rubricName', 'content'],\n        properties: {\n          rubricName: { type: 'string', minLength: 1 },\n          ...commonJudgeFields,\n        },\n      },\n      {\n        type: 'object',\n        additionalProperties: false,\n        required: ['rubric', 'content'],\n        properties: {\n          rubric: rubricRef,\n          ...commonJudgeFields,\n        },\n      },\n    ],\n    description: 'Judge request. Provide exactly one of rubricName or rubric.',\n  }\n  return doc\n}\n","/**\n * stdio RPC transport.\n *\n * For batch / cron use without a running server. The Python client falls\n * back to this when no server is reachable.\n *\n * Protocol (line-delimited JSON over stdin/stdout):\n *   IN:  one JSON object on stdin: {\"method\":\"judge\",\"params\":{...}}\n *   OUT: one JSON object on stdout: {\"result\":{...}} or {\"error\":{...}}\n *\n * One request per process invocation. To pipeline many calls, the client\n * writes JSONL to stdin and reads JSONL from stdout — see batch mode below.\n */\nimport type { LlmClientOptions, LlmRouteRequirements } from '../llm-client'\nimport { handleJudge, handleListRubrics, handleVersion, WireError } from './handlers'\nimport { JudgeRequestSchema } from './schemas'\n\ninterface RpcRequest {\n  method: 'judge' | 'listRubrics' | 'version'\n  params?: unknown\n}\n\ninterface RpcSuccess {\n  result: unknown\n}\n\ninterface RpcError {\n  error: { code: string; message: string; details?: unknown }\n}\n\nexport interface RpcOptions {\n  llm?: LlmClientOptions\n  judgeModel?: string\n  llmRouteRequirements?: LlmRouteRequirements\n}\n\nexport async function dispatchRpc(\n  req: RpcRequest,\n  options: RpcOptions = {},\n): Promise<RpcSuccess | RpcError> {\n  try {\n    switch (req.method) {\n      case 'judge': {\n        const parsed = JudgeRequestSchema.safeParse(req.params)\n        if (!parsed.success) {\n          return {\n            error: {\n              code: 'validation_error',\n              message: 'params did not match JudgeRequest schema.',\n              details: parsed.error.issues,\n            },\n          }\n        }\n        return {\n          result: await handleJudge(parsed.data, {\n            llm: options.llm,\n            defaultModel: options.judgeModel,\n            routeRequirements: options.llmRouteRequirements,\n          }),\n        }\n      }\n      case 'listRubrics':\n        return { result: handleListRubrics() }\n      case 'version':\n        return { result: handleVersion() }\n      default:\n        return {\n          error: {\n            code: 'unknown_method',\n            message: `No such method: ${(req as { method: string }).method}`,\n          },\n        }\n    }\n  } catch (err) {\n    if (err instanceof WireError) {\n      return { error: { code: err.code, message: err.message, details: err.details } }\n    }\n    const message = err instanceof Error ? err.message : String(err)\n    return { error: { code: 'internal_error', message } }\n  }\n}\n\n// ── stdin/stdout driver ─────────────────────────────────────────────\n\nasync function readAll(stream: NodeJS.ReadableStream): Promise<string> {\n  const chunks: Buffer[] = []\n  for await (const chunk of stream) {\n    chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk as string))\n  }\n  return Buffer.concat(chunks).toString('utf-8')\n}\n\n/** Read one JSON request from stdin, write one JSON response to stdout. */\nexport async function runRpcOnce(method?: string, options: RpcOptions = {}): Promise<number> {\n  const raw = await readAll(process.stdin)\n  let req: RpcRequest\n  try {\n    const body = JSON.parse(raw)\n    req = method ? { method: method as RpcRequest['method'], params: body } : (body as RpcRequest)\n  } catch (err) {\n    process.stdout.write(\n      `${JSON.stringify({\n        error: {\n          code: 'parse_error',\n          message: `stdin was not valid JSON: ${err instanceof Error ? err.message : String(err)}`,\n        },\n      })}\\n`,\n    )\n    return 1\n  }\n  const out = await dispatchRpc(req, options)\n  process.stdout.write(`${JSON.stringify(out)}\\n`)\n  return 'error' in out ? 1 : 0\n}\n\n/** Read JSONL requests from stdin, write JSONL responses to stdout. */\nexport async function runRpcBatch(method?: string, options: RpcOptions = {}): Promise<number> {\n  const raw = await readAll(process.stdin)\n  const lines = raw.split('\\n').filter((l) => l.trim().length > 0)\n  let exitCode = 0\n  for (const line of lines) {\n    let req: RpcRequest\n    try {\n      const body = JSON.parse(line)\n      req = method ? { method: method as RpcRequest['method'], params: body } : (body as RpcRequest)\n    } catch (err) {\n      process.stdout.write(\n        `${JSON.stringify({\n          error: {\n            code: 'parse_error',\n            message: `line was not valid JSON: ${err instanceof Error ? err.message : String(err)}`,\n          },\n        })}\\n`,\n      )\n      exitCode = 1\n      continue\n    }\n    const out = await dispatchRpc(req, options)\n    process.stdout.write(`${JSON.stringify(out)}\\n`)\n    if ('error' in out) exitCode = 1\n  }\n  return exitCode\n}\n","/**\n * HTTP transport for the wire protocol.\n *\n * Hono + @hono/node-server. Every endpoint:\n *   1. Validates the request against its Zod schema.\n *   2. Calls the matching handler in `handlers.ts`.\n *   3. Renders 4xx for `WireError` with structured body, 500 for unexpected.\n *\n * The server holds optional `IngestionStores` (passed to `createApp`)\n * to receive production traces and user feedback. With no stores wired,\n * the ingestion endpoints return 503 — read endpoints (`/v1/judge`,\n * `/v1/rubrics`, `/v1/version`) remain fully functional.\n *\n * Run via `agent-eval serve --port 5005`.\n */\nimport { type ServerType, serve } from '@hono/node-server'\nimport { Hono } from 'hono'\nimport { cors } from 'hono/cors'\nimport type { LlmClientOptions, LlmRouteRequirements } from '../llm-client'\nimport {\n  handleFeedbackIngest,\n  handleJudge,\n  handleListRubrics,\n  handleTracesIngest,\n  handleVersion,\n  type IngestionStores,\n  WireError,\n} from './handlers'\nimport { buildOpenApi } from './openapi'\nimport { FeedbackTrajectorySchema, JudgeRequestSchema, TracesIngestRequestSchema } from './schemas'\n\nconst STARTED_AT = Date.now()\n\nexport interface CreateAppOptions {\n  /** Stores wired to the ingestion endpoints. */\n  stores?: IngestionStores\n  /** Model provider used by `/v1/judge`. */\n  llm?: LlmClientOptions\n  /** Default judge model when a request does not provide one. */\n  judgeModel?: string\n  /** Model route checks applied before every judge call. */\n  llmRouteRequirements?: LlmRouteRequirements\n  /**\n   * Bearer-token auth. When provided, every endpoint EXCEPT `/healthz`\n   * and `/v1/version` requires `Authorization: Bearer <token>`. The\n   * token may be a static string OR a function for time-bounded /\n   * rotating tokens.\n   *\n   * Recommended for any server that accepts ingestion writes from the\n   * public internet. Read-only deployments may omit it.\n   */\n  auth?: {\n    bearer: string | ((token: string) => boolean | Promise<boolean>)\n  }\n}\n\nconst AUTH_EXEMPT_PATHS = new Set(['/healthz', '/v1/version', '/openapi.json'])\n\nexport function createApp(opts: CreateAppOptions = {}) {\n  const app = new Hono()\n\n  app.use('*', cors())\n\n  // Bearer-token middleware (only attached when configured).\n  if (opts.auth) {\n    const verify = opts.auth.bearer\n    app.use('*', async (c, next) => {\n      const path = new URL(c.req.url).pathname\n      if (AUTH_EXEMPT_PATHS.has(path)) return next()\n      const raw = c.req.header('authorization') ?? ''\n      const match = raw.match(/^Bearer\\s+(.+)$/i)\n      if (!match) {\n        throw new WireError('unauthorized', 'Missing or malformed Authorization header.', 401)\n      }\n      const token = match[1] as string\n      const ok = typeof verify === 'string' ? token === verify : await verify(token)\n      if (!ok) {\n        throw new WireError('unauthorized', 'Invalid bearer token.', 401)\n      }\n      return next()\n    })\n  }\n\n  app.onError((err, c) => {\n    if (err instanceof WireError) {\n      const status = err.status as 400 | 401 | 404 | 422 | 500 | 503\n      return c.json(\n        { error: { code: err.code, message: err.message, details: err.details } },\n        status,\n      )\n    }\n    // Unexpected — log and return generic 500 without leaking internals.\n    console.error('[agent-eval] unhandled error:', err)\n    return c.json({ error: { code: 'internal_error', message: 'Internal server error.' } }, 500)\n  })\n\n  // ── Health ──\n  app.get('/healthz', (c) =>\n    c.json({ status: 'ok' as const, uptimeSec: (Date.now() - STARTED_AT) / 1000 }),\n  )\n\n  // ── Version ──\n  app.get('/v1/version', (c) => c.json(handleVersion()))\n\n  // ── Rubrics ──\n  app.get('/v1/rubrics', (c) => c.json(handleListRubrics()))\n\n  // ── Judge ──\n  app.post('/v1/judge', async (c) => {\n    const raw = await c.req.json().catch(() => null)\n    if (raw == null) {\n      throw new WireError('validation_error', 'Request body must be JSON.', 400)\n    }\n    const parsed = JudgeRequestSchema.safeParse(raw)\n    if (!parsed.success) {\n      throw new WireError(\n        'validation_error',\n        'Request did not match JudgeRequest schema.',\n        400,\n        parsed.error.issues,\n      )\n    }\n    const result = await handleJudge(parsed.data, {\n      llm: opts.llm,\n      defaultModel: opts.judgeModel,\n      routeRequirements: opts.llmRouteRequirements,\n    })\n    return c.json(result)\n  })\n\n  // ── Traces ingest (NDJSON-friendly: accepts either {events:[...]} or NDJSON) ──\n  app.post('/v1/traces/ingest', async (c) => {\n    const contentType = c.req.header('content-type') ?? ''\n    let payload: unknown\n    if (contentType.includes('application/x-ndjson')) {\n      const text = await c.req.text()\n      const events = text\n        .split('\\n')\n        .map((line) => line.trim())\n        .filter((line) => line.length > 0)\n        .map((line) => {\n          try {\n            return JSON.parse(line)\n          } catch {\n            throw new WireError(\n              'validation_error',\n              'NDJSON line did not parse as JSON.',\n              400,\n              line.slice(0, 200),\n            )\n          }\n        })\n      payload = { events }\n    } else {\n      payload = await c.req.json().catch(() => null)\n    }\n    if (payload == null) {\n      throw new WireError('validation_error', 'Request body must be JSON or NDJSON.', 400)\n    }\n    const parsed = TracesIngestRequestSchema.safeParse(payload)\n    if (!parsed.success) {\n      throw new WireError(\n        'validation_error',\n        'Request did not match TracesIngestRequest schema.',\n        400,\n        parsed.error.issues,\n      )\n    }\n    const result = await handleTracesIngest(parsed.data, opts.stores ?? {})\n    return c.json(result)\n  })\n\n  // ── Feedback ingest ──\n  app.post('/v1/feedback', async (c) => {\n    const raw = await c.req.json().catch(() => null)\n    if (raw == null) {\n      throw new WireError('validation_error', 'Request body must be JSON.', 400)\n    }\n    const parsed = FeedbackTrajectorySchema.safeParse(raw)\n    if (!parsed.success) {\n      throw new WireError(\n        'validation_error',\n        'Request did not match FeedbackTrajectory schema.',\n        400,\n        parsed.error.issues,\n      )\n    }\n    const result = await handleFeedbackIngest(parsed.data, opts.stores ?? {})\n    return c.json(result)\n  })\n\n  // ── OpenAPI spec ──\n  app.get('/openapi.json', (c) => c.json(buildOpenApi(handleVersion().version)))\n\n  return app\n}\n\nexport interface ServeOptions extends CreateAppOptions {\n  /** Default 5005. */\n  port?: number\n  /** Default '127.0.0.1'. Set to '0.0.0.0' to listen on all interfaces. */\n  host?: string\n}\n\nexport function startServer(opts: ServeOptions = {}): ServerType {\n  const app = createApp(opts)\n  const port = opts.port ?? 5005\n  const host = opts.host ?? '127.0.0.1'\n  return serve({ fetch: app.fetch, port, hostname: host }, ({ address, port: actualPort }) => {\n    // eslint-disable-next-line no-console\n    console.log(`[agent-eval] serving on http://${address}:${actualPort}`)\n  })\n}\n\nexport interface StartedServer {\n  server: ServerType\n  /** The OS-assigned port. When opts.port was 0, this is the actual port the\n   *  kernel bound — callers that need to dial back (smoke tests, sidecars\n   *  registering with a parent) read this rather than guessing a free port. */\n  port: number\n  /** Resolved host the server bound to (defaults to 127.0.0.1). */\n  host: string\n  /** Close the server. Resolves once active connections have drained. */\n  close(): Promise<void>\n}\n\n/**\n * Promise-returning variant of `startServer` that resolves once the server is\n * listening and surfaces the resolved bound port. Use this from smoke tests\n * (`startServerAsync({ port: 0 })`) and any caller that needs to dial back.\n */\nexport function startServerAsync(opts: ServeOptions = {}): Promise<StartedServer> {\n  const app = createApp(opts)\n  const port = opts.port ?? 5005\n  const host = opts.host ?? '127.0.0.1'\n  return new Promise((resolve, reject) => {\n    let settled = false\n    let server: ServerType | undefined\n    server = serve({ fetch: app.fetch, port, hostname: host }, ({ address, port: actualPort }) => {\n      if (settled) return\n      settled = true\n      // eslint-disable-next-line no-console\n      console.log(`[agent-eval] serving on http://${address}:${actualPort}`)\n      resolve({\n        server: server!,\n        port: actualPort,\n        host: address,\n        close: () =>\n          new Promise<void>((res, rej) => {\n            server!.close((err) => (err ? rej(err) : res()))\n          }),\n      })\n    })\n    server.on('error', (err) => {\n      if (settled) return\n      settled = true\n      reject(err)\n    })\n  })\n}\n"],"mappings":";;;;;;;;;;;;;AAYA,SAAS,4BAA4B;AACrC,SAAS,SAAS;AAElB,qBAAqB,CAAC;AAIf,IAAM,wBAAwB,EAClC,OAAO;AAAA,EACN,IAAI,EACD,OAAO,EACP,IAAI,CAAC,EACL,SAAS,gFAA2E;AAAA,EACvF,aAAa,EACV,OAAO,EACP,IAAI,CAAC,EACL,SAAS,sEAAsE;AAAA,EAClF,QAAQ,EACL,OAAO,EACP,IAAI,CAAC,EACL,QAAQ,CAAC,EACT,SAAS,gEAAgE;AAAA,EAC5E,KAAK,EAAE,OAAO,EAAE,QAAQ,CAAC,EAAE,SAAS,gDAAgD;AAAA,EACpF,KAAK,EAAE,OAAO,EAAE,QAAQ,CAAC,EAAE,SAAS,gDAAgD;AACtF,CAAC,EACA,QAAQ,iBAAiB;AAErB,IAAM,oBAAoB,EAC9B,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,mEAA8D;AAAA,EAC7F,aAAa,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,mDAAmD;AAC7F,CAAC,EACA,QAAQ,aAAa;AAIjB,IAAM,eAAe,EACzB,OAAO;AAAA,EACN,MAAM,EACH,OAAO,EACP,IAAI,CAAC,EACL,SAAS,4EAAuE;AAAA,EACnF,aAAa,EACV,OAAO,EACP,IAAI,CAAC,EACL,SAAS,0DAA0D;AAAA,EACtE,cAAc,EACX,OAAO,EACP,IAAI,CAAC,EACL;AAAA,IACC;AAAA,EACF;AAAA,EACF,YAAY,EACT,MAAM,qBAAqB,EAC3B,IAAI,CAAC,EACL,SAAS,+DAA+D;AAAA,EAC3E,cAAc,EACX,MAAM,iBAAiB,EACvB,QAAQ,CAAC,CAAC,EACV,SAAS,iFAAiF;AAAA,EAC7F,MAAM,EACH,MAAM,iBAAiB,EACvB,QAAQ,CAAC,CAAC,EACV,SAAS,uEAAuE;AACrF,CAAC,EACA,QAAQ,QAAQ;AAIZ,IAAM,qBAAqB,EAC/B,OAAO;AAAA,EACN,YAAY,EACT,OAAO,EACP,SAAS,EACT,SAAS,kEAAkE;AAAA,EAC9E,QAAQ,aAAa,SAAS,EAAE;AAAA,IAC9B;AAAA,EACF;AAAA,EACA,SAAS,EACN,OAAO,EACP,IAAI,CAAC,EACL,SAAS,uFAAkF;AAAA,EAC9F,SAAS,EACN,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAC9B,SAAS,EACT;AAAA,IACC;AAAA,EACF;AAAA,EACF,OAAO,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,oDAAoD;AAC5F,CAAC,EACA,OAAO,CAAC,MAAM,QAAQ,EAAE,UAAU,MAAM,QAAQ,EAAE,MAAM,GAAG;AAAA,EAC1D,SAAS;AACX,CAAC,EACA,QAAQ,cAAc;AAElB,IAAM,oBAAoB,EAC9B,OAAO;AAAA,EACN,WAAW,EACR,OAAO,EACP,IAAI,CAAC,EACL,IAAI,CAAC,EACL,SAAS,iFAAiF;AAAA,EAC7F,YAAY,EACT,OAAO,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,EAC7B,SAAS,mDAAmD;AAAA,EAC/D,cAAc,EACX,MAAM,EAAE,OAAO,CAAC,EAChB,QAAQ,CAAC,CAAC,EACV,SAAS,+EAA+E;AAAA,EAC3F,MAAM,EACH,MAAM,EAAE,OAAO,CAAC,EAChB,QAAQ,CAAC,CAAC,EACV,SAAS,8DAA8D;AAAA,EAC1E,WAAW,EACR,OAAO,EACP,SAAS,yEAAyE;AAAA,EACrF,eAAe,EACZ,OAAO,EACP;AAAA,IACC;AAAA,EACF;AAAA,EACF,OAAO,EAAE,OAAO,EAAE,SAAS,yDAAyD;AAAA,EACpF,YAAY,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,qCAAqC;AAC3F,CAAC,EACA,QAAQ,aAAa;AAIjB,IAAM,mBAAmB,EAC7B,OAAO;AAAA,EACN,MAAM,EAAE,OAAO,EAAE,SAAS,yCAAyC;AAAA,EACnE,aAAa,EAAE,OAAO,EAAE,SAAS,4BAA4B;AAAA,EAC7D,YAAY,EACT,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,GAAG,aAAa,EAAE,OAAO,GAAG,QAAQ,EAAE,OAAO,EAAE,CAAC,CAAC,EAC/E,SAAS,kDAAkD;AAAA,EAC9D,cAAc,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,QAAQ,CAAC,CAAC,EAAE,SAAS,uCAAuC;AAAA,EAC9F,eAAe,EAAE,OAAO,EAAE,SAAS,8DAAyD;AAC9F,CAAC,EACA,QAAQ,YAAY;AAEhB,IAAM,4BAA4B,EACtC,OAAO;AAAA,EACN,SAAS,EAAE,MAAM,gBAAgB;AACnC,CAAC,EACA,QAAQ,qBAAqB;AAIzB,IAAM,wBAAwB,EAClC,OAAO;AAAA,EACN,SAAS,EAAE,OAAO,EAAE,SAAS,qDAAqD;AAAA,EAClF,SAAS,EAAE,OAAO,EAAE,SAAS,0DAA0D;AAAA,EACvF,aAAa,EACV,OAAO,EACP;AAAA,IACC;AAAA,EACF;AAAA,EACF,YAAY,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,SAAS,iCAAiC;AAC5E,CAAC,EACA,QAAQ,iBAAiB;AAErB,IAAM,uBAAuB,EACjC,OAAO;AAAA,EACN,QAAQ,EAAE,QAAQ,IAAI;AAAA,EACtB,WAAW,EAAE,OAAO;AACtB,CAAC,EACA,QAAQ,gBAAgB;AAUpB,IAAM,mBAAmB,EAC7B,OAAO;AAAA,EACN,SAAS,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,4CAA4C;AAAA,EAChF,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,4BAA4B;AAAA,EAC9D,QAAQ,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,sCAAsC;AAAA,EAC7E,MAAM,EACH,KAAK;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC,EACA,SAAS,yEAAoE;AAAA,EAChF,WAAW,EACR,OAAO,EACP,IAAI,EACJ,YAAY,EACZ,SAAS,kEAAkE;AAAA,EAC9E,SAAS,EACN,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAC9B,SAAS,sDAAiD;AAC/D,CAAC,EACA,QAAQ,YAAY;AAEhB,IAAM,4BAA4B,EACtC,OAAO;AAAA,EACN,QAAQ,EACL,MAAM,gBAAgB,EACtB,IAAI,CAAC,EACL,IAAI,GAAM,EACV,SAAS,4EAAuE;AACrF,CAAC,EACA,QAAQ,qBAAqB;AAEzB,IAAM,6BAA6B,EACvC,OAAO;AAAA,EACN,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,6BAA6B;AAAA,EAC/E,UAAU,EACP,OAAO,EACP,IAAI,EACJ,YAAY,EACZ,SAAS,uEAAkE;AAAA,EAC9E,QAAQ,EACL;AAAA,IACC,EAAE,OAAO;AAAA,MACP,SAAS,EAAE,OAAO,EAAE,SAAS,iCAAiC;AAAA,MAC9D,SAAS,EAAE,OAAO,EAAE,SAAS,6BAA6B;AAAA,IAC5D,CAAC;AAAA,EACH,EACC,QAAQ,CAAC,CAAC;AACf,CAAC,EACA,QAAQ,sBAAsB;AAE1B,IAAM,sBAAsB,EAChC,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,SAAS;AAAA,EACxB,QAAQ,EAAE,KAAK,CAAC,QAAQ,SAAS,eAAe,UAAU,UAAU,QAAQ,CAAC;AAAA,EAC7E,MAAM,EAAE,KAAK;AAAA,IACX;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAAA,EACD,OAAO,EAAE,QAAQ;AAAA,EACjB,QAAQ,EAAE,OAAO,EAAE,SAAS;AAAA,EAC5B,UAAU,EAAE,KAAK,CAAC,QAAQ,WAAW,SAAS,UAAU,CAAC,EAAE,SAAS;AAAA,EACpE,WAAW,EAAE,OAAO,EAAE,SAAS,eAAe;AAAA,EAC9C,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AACvD,CAAC,EACA,QAAQ,eAAe;AAEnB,IAAM,wBAAwB,EAClC,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC;AAAA,EACpB,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY;AAAA,EACxC,cAAc,EAAE,KAAK;AAAA,IACnB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAAA,EACD,UAAU,EAAE,QAAQ;AAAA,EACpB,SAAS,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACvC,gBAAgB,EACb,OAAO;AAAA,IACN,MAAM,EAAE,OAAO;AAAA,IACf,MAAM,EAAE,KAAK,CAAC,OAAO,UAAU,MAAM,CAAC,EAAE,SAAS;AAAA,IACjD,SAAS,EAAE,OAAO,EAAE,SAAS;AAAA,IAC7B,oBAAoB,EAAE,QAAQ,EAAE,SAAS;AAAA,IACzC,kBAAkB,EAAE,QAAQ,EAAE,SAAS;AAAA,IACvC,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACvD,CAAC,EACA,SAAS;AAAA,EACZ,UAAU,EAAE,MAAM,mBAAmB,EAAE,SAAS;AAAA,EAChD,WAAW,EAAE,OAAO;AAAA,EACpB,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AACvD,CAAC,EACA,QAAQ,iBAAiB;AAErB,IAAM,2BAA2B,EACrC,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,gDAAgD;AAAA,EAC/E,WAAW,EAAE,OAAO,EAAE,SAAS;AAAA,EAC/B,YAAY,EAAE,OAAO,EAAE,SAAS;AAAA,EAChC,MAAM,EAAE,OAAO;AAAA,IACb,QAAQ,EAAE,OAAO,EAAE,IAAI,CAAC;AAAA,IACxB,SAAS,EAAE,QAAQ,EAAE,SAAS;AAAA,EAChC,CAAC;AAAA,EACD,UAAU,EAAE,MAAM,qBAAqB,EAAE,QAAQ,CAAC,CAAC;AAAA,EACnD,QAAQ,EAAE,MAAM,mBAAmB,EAAE,QAAQ,CAAC,CAAC;AAAA,EAC/C,SAAS,EACN,OAAO;AAAA,IACN,SAAS,EAAE,QAAQ,EAAE,SAAS;AAAA,IAC9B,OAAO,EAAE,OAAO,EAAE,SAAS;AAAA,IAC3B,SAAS,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,EAAE,SAAS;AAAA,IACnD,SAAS,EAAE,OAAO,EAAE,SAAS;AAAA,IAC7B,QAAQ,EAAE,OAAO,EAAE,SAAS;AAAA,IAC5B,YAAY,EAAE,OAAO,EAAE,SAAS;AAAA,IAChC,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AAAA,EACvD,CAAC,EACA,SAAS;AAAA,EACZ,OAAO,EAAE,KAAK,CAAC,SAAS,OAAO,QAAQ,SAAS,CAAC,EAAE,SAAS;AAAA,EAC5D,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,EAAE,SAAS;AAAA,EAChD,WAAW,EAAE,OAAO,EAAE,SAAS,eAAe;AAAA,EAC9C,WAAW,EAAE,OAAO,EAAE,SAAS;AAAA,EAC/B,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AACvD,CAAC,EACA,QAAQ,oBAAoB;AAExB,IAAM,+BAA+B,EACzC,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,SAAS,mCAAmC;AAAA,EAC3D,WAAW,EAAE,QAAQ,EAAE,SAAS,wDAAwD;AAC1F,CAAC,EACA,QAAQ,wBAAwB;AAU5B,IAAM,sBAAsB,EAChC,OAAO;AAAA,EACN,OAAO,EACJ,OAAO;AAAA,IACN,MAAM,EACH,OAAO,EACP;AAAA,MACC;AAAA,IACF;AAAA,IACF,SAAS,EAAE,OAAO,EAAE,SAAS,yBAAyB;AAAA,IACtD,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,6BAA6B;AAAA,EACxE,CAAC,EACA,SAAS,+DAA+D;AAC7E,CAAC,EACA,QAAQ,eAAe;AAoBnB,IAAM,eAAe;AAMrB,SAAS,WAAW,QAAwB;AACjD,QAAM,SAAS,gBAAgB,MAAM;AACrC,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,QAAK,IAAI,KAAM,OAAO,WAAW,CAAC;AAAA,EACpC;AAEA,SAAO,GAAG,OAAO,IAAI,KAAK,MAAM,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC;AAClE;AAEA,SAAS,gBAAgB,OAAwB;AAC/C,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,IAAI,MAAM,IAAI,CAAC,SAAS,gBAAgB,IAAI,CAAC,EAAE,KAAK,GAAG,CAAC;AACzF,MAAI,SAAS,OAAO,UAAU,UAAU;AACtC,UAAM,UAAU,OAAO,QAAQ,KAAgC,EAC5D,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC,EACrC,IAAI,CAAC,CAAC,KAAK,IAAI,MAAM,GAAG,KAAK,UAAU,GAAG,CAAC,IAAI,gBAAgB,IAAI,CAAC,EAAE;AACzE,WAAO,IAAI,QAAQ,KAAK,GAAG,CAAC;AAAA,EAC9B;AACA,SAAO,KAAK,UAAU,KAAK;AAC7B;;;AC9XA,IAAM,YAAoB;AAAA,EACxB,MAAM;AAAA,EACN,aACE;AAAA,EACF,cAAc;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAuBd,YAAY;AAAA,IACV;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,EACF;AAAA,EACA,cAAc;AAAA,IACZ,EAAE,IAAI,cAAc,aAAa,qCAAqC;AAAA,IACtE,EAAE,IAAI,kBAAkB,aAAa,sCAAsC;AAAA,IAC3E,EAAE,IAAI,eAAe,aAAa,oCAAoC;AAAA,IACtE,EAAE,IAAI,WAAW,aAAa,kCAAkC;AAAA,IAChE,EAAE,IAAI,mBAAmB,aAAa,qCAAqC;AAAA,IAC3E,EAAE,IAAI,WAAW,aAAa,oCAAoC;AAAA,IAClE,EAAE,IAAI,eAAe,aAAa,uCAAuC;AAAA,EAC3E;AAAA,EACA,MAAM;AAAA,IACJ,EAAE,IAAI,sBAAsB,aAAa,qCAAqC;AAAA,IAC9E,EAAE,IAAI,iBAAiB,aAAa,4BAA4B;AAAA,IAChE,EAAE,IAAI,0BAA0B,aAAa,wBAAwB;AAAA,IACrE,EAAE,IAAI,kBAAkB,aAAa,oCAAoC;AAAA,EAC3E;AACF;AAIO,IAAM,kBAA0C;AAAA,EACrD,aAAa;AACf;AAGO,SAAS,iBAAiB,MAAkC;AACjE,SAAO,gBAAgB,IAAI;AAC7B;AAGO,SAAS,qBAAqB;AACnC,SAAO,OAAO,OAAO,eAAe,EAAE,IAAI,CAAC,OAAO;AAAA,IAChD,MAAM,EAAE;AAAA,IACR,aAAa,EAAE;AAAA,IACf,YAAY,EAAE,WAAW,IAAI,CAAC,OAAO;AAAA,MACnC,IAAI,EAAE;AAAA,MACN,aAAa,EAAE;AAAA,MACf,QAAQ,EAAE;AAAA,IACZ,EAAE;AAAA,IACF,cAAc,EAAE,aAAa,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,IAC5C,eAAe,WAAW,CAAC;AAAA,EAC7B,EAAE;AACJ;;;ACqLA,SAAS,oBAAoB;AAC7B,SAAS,SAAS,eAAe;AACjC,SAAS,qBAAqB;AA/PvB,IAAM,YAAN,cAAwB,MAAM;AAAA,EACnC,YACkB,MAChB,SACgB,SAAiB,KACjB,SAChB;AACA,UAAM,OAAO;AALG;AAEA;AACA;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EAPkB;AAAA,EAEA;AAAA,EACA;AAKpB;AAKA,SAAS,kBAAkB,QAAgB;AACzC,SAAO;AAAA,IACL,MAAM;AAAA,IACN,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,sBAAsB;AAAA,MACtB,YAAY;AAAA,QACV,YAAY;AAAA,UACV,MAAM;AAAA,UACN,sBAAsB;AAAA,UACtB,YAAY,OAAO;AAAA,YACjB,OAAO,WAAW,IAAI,CAAC,MAAM;AAAA,cAC3B,EAAE;AAAA,cACF,EAAE,MAAM,UAAU,SAAS,EAAE,KAAK,SAAS,EAAE,IAAI;AAAA,YACnD,CAAC;AAAA,UACH;AAAA,UACA,UAAU,OAAO,WAAW,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,QAC7C;AAAA,QACA,cAAc;AAAA,UACZ,MAAM;AAAA,UACN,OAAO,EAAE,MAAM,UAAU,MAAM,OAAO,aAAa,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE;AAAA,QACtE;AAAA,QACA,MAAM;AAAA,UACJ,MAAM;AAAA,UACN,OAAO,EAAE,MAAM,UAAU,MAAM,OAAO,KAAK,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE;AAAA,QAC9D;AAAA,QACA,WAAW,EAAE,MAAM,SAAS;AAAA,MAC9B;AAAA,MACA,UAAU,CAAC,cAAc,WAAW;AAAA,IACtC;AAAA,EACF;AACF;AASA,SAAS,oBAAoB,OAAgB,QAA6B;AACxE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,UAAM,IAAI,UAAU,eAAe,oCAAoC,KAAK,KAAK;AAAA,EACnF;AACA,QAAM,MAAM;AACZ,QAAM,gBAAgB,IAAI;AAC1B,MAAI,CAAC,iBAAiB,OAAO,kBAAkB,YAAY,MAAM,QAAQ,aAAa,GAAG;AACvF,UAAM,IAAI,UAAU,eAAe,wCAAwC,KAAK,KAAK;AAAA,EACvF;AAEA,QAAM,aAAqC,CAAC;AAC5C,QAAM,kBAAkB;AACxB,aAAW,OAAO,OAAO,YAAY;AACnC,UAAM,QAAQ,gBAAgB,IAAI,EAAE;AACpC,QACE,OAAO,UAAU,YACjB,CAAC,OAAO,SAAS,KAAK,KACtB,QAAQ,IAAI,OACZ,QAAQ,IAAI,KACZ;AACA,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,IAAI,EAAE;AAAA,QACrD;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,eAAW,IAAI,EAAE,IAAI;AAAA,EACvB;AAEA,QAAM,kBAAkB,IAAI,IAAI,OAAO,aAAa,IAAI,CAAC,SAAS,KAAK,EAAE,CAAC;AAC1E,QAAM,cAAc,IAAI,IAAI,OAAO,KAAK,IAAI,CAAC,QAAQ,IAAI,EAAE,CAAC;AAC5D,QAAM,eAAe,gBAAgB,IAAI,cAAc,iBAAiB,gBAAgB,KAAK;AAC7F,QAAM,OAAO,gBAAgB,IAAI,MAAM,aAAa,QAAQ,KAAK;AACjE,MAAI,OAAO,IAAI,cAAc,YAAY,IAAI,UAAU,KAAK,EAAE,WAAW,GAAG;AAC1E,UAAM,IAAI,UAAU,eAAe,qCAAqC,KAAK,KAAK;AAAA,EACpF;AAEA,SAAO,EAAE,YAAY,cAAc,MAAM,WAAW,IAAI,UAAU;AACpE;AAEA,SAAS,gBACP,KACA,SACA,OACA,UACU;AACV,MAAI,QAAQ,OAAW,QAAO,CAAC;AAC/B,MAAI,CAAC,MAAM,QAAQ,GAAG,GAAG;AACvB,UAAM,IAAI,UAAU,eAAe,4BAA4B,KAAK,KAAK,KAAK,QAAQ;AAAA,EACxF;AACA,QAAM,MAAgB,CAAC;AACvB,aAAW,QAAQ,KAAK;AACtB,QAAI,OAAO,SAAS,YAAY,CAAC,QAAQ,IAAI,IAAI,GAAG;AAClD,YAAM,IAAI;AAAA,QACR;AAAA,QACA,0BAA0B,KAAK,QAAQ,OAAO,IAAI,CAAC;AAAA,QACnD;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,KAAK,IAAI;AAAA,EACf;AACA,SAAO;AACT;AAEA,SAAS,eAAe,YAAoC,QAAwB;AAClF,MAAI,WAAW;AACf,MAAI,cAAc;AAClB,aAAW,OAAO,OAAO,YAAY;AACnC,UAAM,MAAM,WAAW,IAAI,EAAE,KAAK;AAClC,UAAM,QAAQ,IAAI,MAAM,IAAI,OAAO;AACnC,UAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,MAAM,IAAI,OAAO,KAAK,CAAC;AACnE,gBAAY,aAAa,IAAI;AAC7B,mBAAe,IAAI;AAAA,EACrB;AACA,SAAO,cAAc,IAAI,WAAW,cAAc;AACpD;AAEA,SAAS,iBAAiB,SAAiB,SAA0B;AACnE,QAAM,MAAM,WAAW,OAAO,KAAK,OAAiB,EAAE,SAAS,KAAK,UAAU,OAAO,IAAI;AACzF,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,yCAAyC;AAAA,IAC/C,MAAM,MAAM;AAAA,EACd,EACG,OAAO,OAAO,EACd,KAAK,IAAI;AACd;AAEA,IAAM,sBAAsB;AAW5B,eAAsB,YACpB,KACA,UAA8B,CAAC,GACT;AAEtB,MAAI;AACJ,MAAI,IAAI,YAAY;AAClB,UAAM,QAAQ,iBAAiB,IAAI,UAAU;AAC7C,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,UAAU,oBAAoB,6BAA6B,IAAI,UAAU,MAAM,GAAG;AAAA,IAC9F;AACA,aAAS;AAAA,EACX,WAAW,IAAI,QAAQ;AACrB,aAAS,IAAI;AAAA,EACf,OAAO;AAEL,UAAM,IAAI,UAAU,oBAAoB,4CAA4C,GAAG;AAAA,EACzF;AAEA,MAAI,QAAQ,mBAAmB;AAC7B,QAAI;AACF,qBAAe,QAAQ,OAAO,CAAC,GAAG,QAAQ,iBAAiB;AAAA,IAC7D,SAAS,OAAO;AACd,UAAI,EAAE,iBAAiB,wBAAyB,OAAM;AACtD,YAAM,aAAa,MAAM,WAAW;AACpC,YAAM,IAAI;AAAA,QACR,aAAa,uBAAuB;AAAA,QACpC,aACI,2GACA,MAAM;AAAA,QACV;AAAA,QACA,EAAE,QAAQ,MAAM,OAAO;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,QAAQ,IAAI,SAAS,QAAQ,gBAAgB;AAEnD,QAAM,UAAU;AAAA,IACd;AAAA,IACA,UAAU;AAAA,MACR,EAAE,MAAM,UAAU,SAAS,OAAO,aAAa;AAAA,MAC/C,EAAE,MAAM,QAAQ,SAAS,iBAAiB,IAAI,SAAS,IAAI,OAAO,EAAE;AAAA,IACtE;AAAA,IACA,YAAY,kBAAkB,MAAM;AAAA,IACpC,aAAa;AAAA,IACb,WAAW;AAAA,IACX,WAAW;AAAA,EACb;AACA,QAAM,SAAS,QAAQ,cAAc,IAAI,WAAW;AACpD,QAAM,OAAO,MAAM,OAAO,YAAY;AAAA,IACpC,SAAS;AAAA,IACT,OAAO,QAAQ,aAAa;AAAA,IAC5B,OAAO,QAAQ,IAAI,cAAc,QAAQ;AAAA,IACzC;AAAA,IACA,eAAe,2BAA2B,SAAS,QAAQ,GAAG;AAAA,IAC9D,QAAQ,QAAQ;AAAA,IAChB,SAAS,CAAC,QAAQ,WAChB,YAAyB,SAAS;AAAA,MAChC,GAAG,QAAQ;AAAA,MACX;AAAA,MACA,gBAAgB;AAAA,IAClB,CAAC;AAAA,IACH,SAAS,CAAC,EAAE,QAAAA,QAAO,MAAM,mBAAmBA,OAAM;AAAA,IAClD,kBAAkB;AAAA,EACpB,CAAC;AACD,MAAI,CAAC,KAAK,UAAW,OAAM,KAAK;AAChC,QAAM,EAAE,OAAO,OAAO,IAAI,KAAK;AAE/B,QAAM,SAAS,oBAAoB,OAAO,MAAM;AAEhD,QAAM,YAAY,eAAe,OAAO,YAAY,MAAM;AAC1D,QAAM,aAAa,KAAK,IAAI,IAAI;AAEhC,SAAO;AAAA,IACL;AAAA,IACA,YAAY,OAAO;AAAA,IACnB,cAAc,OAAO,gBAAgB,CAAC;AAAA,IACtC,MAAM,OAAO,QAAQ,CAAC;AAAA,IACtB,WAAW,OAAO;AAAA,IAClB,eAAe,WAAW,MAAM;AAAA,IAChC,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAIO,SAAS,oBAAyC;AACvD,SAAO,EAAE,SAAS,mBAAmB,EAAE;AACzC;AAQA,IAAI;AAEJ,SAAS,qBAA6B;AACpC,MAAI,eAAgB,QAAO;AAG3B,QAAM,OAAO,QAAQ,cAAc,YAAY,GAAG,CAAC;AACnD,QAAM,aAAa;AAAA,IACjB,QAAQ,MAAM,MAAM,MAAM,cAAc;AAAA;AAAA,IACxC,QAAQ,MAAM,MAAM,cAAc;AAAA;AAAA,EACpC;AACA,aAAW,QAAQ,YAAY;AAC7B,QAAI;AACF,YAAM,MAAM,KAAK,MAAM,aAAa,MAAM,OAAO,CAAC;AAClD,UAAI,IAAI,SAAS;AACf,yBAAiB,IAAI;AACrB,eAAO,IAAI;AAAA,MACb;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,gBAAiC;AAC/C,SAAO;AAAA,IACL,SAAS;AAAA,IACT,SAAS,mBAAmB;AAAA,IAC5B,aAAa;AAAA,IACb,YAAY,CAAC,SAAS,eAAe,WAAW,mBAAmB,eAAe;AAAA,EACpF;AACF;AA2BA,eAAsB,mBACpB,KACA,QAC+B;AAC/B,MAAI,CAAC,OAAO,YAAY;AACtB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,QAAM,SAAsD,CAAC;AAC7D,MAAI,WAAW;AACf,aAAW,SAAS,IAAI,QAAQ;AAC9B,QAAI;AAEF,YAAM,OAAO,WAAW,YAAY,KAA2B;AAC/D;AAAA,IACF,SAAS,KAAK;AACZ,aAAO,KAAK;AAAA,QACV,SAAS,MAAM;AAAA,QACf,SAAS,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MAC1D,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO,EAAE,UAAU,UAAU,OAAO,QAAQ,OAAO;AACrD;AAOA,eAAsB,qBACpB,KACA,QACiC;AACjC,MAAI,CAAC,OAAO,eAAe;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAIA,QAAM,OAAO,cAAc,KAAK,GAAgE;AAChG,SAAO,EAAE,IAAI,IAAI,IAAI,WAAW,KAAK;AACvC;;;AC7YA,SAAS,iBAAiB,2BAA2B;AAiB9C,SAAS,aAAa,gBAAuC;AAClE,QAAM,WAAW,IAAI,gBAAgB;AAGrC,WAAS,SAAS,gBAAgB,kBAAkB;AACpD,WAAS,SAAS,eAAe,iBAAiB;AAClD,WAAS,SAAS,uBAAuB,yBAAyB;AAClE,WAAS,SAAS,mBAAmB,qBAAqB;AAC1D,WAAS,SAAS,kBAAkB,oBAAoB;AACxD,WAAS,SAAS,iBAAiB,mBAAmB;AACtD,WAAS,SAAS,uBAAuB,yBAAyB;AAClE,WAAS,SAAS,wBAAwB,0BAA0B;AACpE,WAAS,SAAS,sBAAsB,wBAAwB;AAChE,WAAS,SAAS,0BAA0B,4BAA4B;AAGxE,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,SAAS;AAAA,UACP,oBAAoB,EAAE,QAAQ,mBAAmB;AAAA,QACnD;AAAA,MACF;AAAA,IACF;AAAA,IACA,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,kBAAkB,EAAE;AAAA,MAC/D;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,0BAA0B,EAAE;AAAA,MACvE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aAAa;AAAA,IACb,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,sBAAsB,EAAE;AAAA,MACnE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,qBAAqB,EAAE;AAAA,MAClE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,SAAS;AAAA,UACP,oBAAoB,EAAE,QAAQ,0BAA0B;AAAA,UACxD,wBAAwB,EAAE,QAAQ,0BAA0B;AAAA,QAC9D;AAAA,MACF;AAAA,IACF;AAAA,IACA,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,2BAA2B,EAAE;AAAA,MACxE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,SAAS;AAAA,UACP,oBAAoB,EAAE,QAAQ,yBAAyB;AAAA,QACzD;AAAA,MACF;AAAA,IACF;AAAA,IACA,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,6BAA6B,EAAE;AAAA,MAC1E;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,IACF;AAAA,EACF,CAAC;AAED,QAAM,YAAY,IAAI,oBAAoB,SAAS,WAAW;AAC9D,QAAM,MAAM,UAAU,iBAAiB;AAAA,IACrC,SAAS;AAAA,IACT,MAAM;AAAA,MACJ,OAAO;AAAA,MACP,SAAS;AAAA,MACT,aAAa;AAAA;AAAA,yBAEM,YAAY;AAAA,MAC/B,SAAS,EAAE,MAAM,kBAAkB,KAAK,+CAA+C;AAAA,MACvF,SAAS,EAAE,MAAM,MAAM;AAAA,IACzB;AAAA,IACA,SAAS,CAAC,EAAE,KAAK,yBAAyB,aAAa,yBAAyB,CAAC;AAAA,EACnF,CAAC;AACD,QAAM,YAAY,EAAE,MAAM,8BAA8B;AACxD,QAAM,oBAAoB;AAAA,IACxB,SAAS,EAAE,MAAM,UAAU,WAAW,EAAE;AAAA,IACxC,SAAS,EAAE,MAAM,UAAU,sBAAsB,KAAK;AAAA,IACtD,OAAO,EAAE,MAAM,SAAS;AAAA,EAC1B;AACA,MAAI,eAAe,CAAC;AACpB,MAAI,WAAW,YAAY,CAAC;AAC5B,MAAI,WAAW,QAAQ,eAAe;AAAA,IACpC,OAAO;AAAA,MACL;AAAA,QACE,MAAM;AAAA,QACN,sBAAsB;AAAA,QACtB,UAAU,CAAC,cAAc,SAAS;AAAA,QAClC,YAAY;AAAA,UACV,YAAY,EAAE,MAAM,UAAU,WAAW,EAAE;AAAA,UAC3C,GAAG;AAAA,QACL;AAAA,MACF;AAAA,MACA;AAAA,QACE,MAAM;AAAA,QACN,sBAAsB;AAAA,QACtB,UAAU,CAAC,UAAU,SAAS;AAAA,QAC9B,YAAY;AAAA,UACV,QAAQ;AAAA,UACR,GAAG;AAAA,QACL;AAAA,MACF;AAAA,IACF;AAAA,IACA,aAAa;AAAA,EACf;AACA,SAAO;AACT;;;ACjMA,eAAsB,YACpB,KACA,UAAsB,CAAC,GACS;AAChC,MAAI;AACF,YAAQ,IAAI,QAAQ;AAAA,MAClB,KAAK,SAAS;AACZ,cAAM,SAAS,mBAAmB,UAAU,IAAI,MAAM;AACtD,YAAI,CAAC,OAAO,SAAS;AACnB,iBAAO;AAAA,YACL,OAAO;AAAA,cACL,MAAM;AAAA,cACN,SAAS;AAAA,cACT,SAAS,OAAO,MAAM;AAAA,YACxB;AAAA,UACF;AAAA,QACF;AACA,eAAO;AAAA,UACL,QAAQ,MAAM,YAAY,OAAO,MAAM;AAAA,YACrC,KAAK,QAAQ;AAAA,YACb,cAAc,QAAQ;AAAA,YACtB,mBAAmB,QAAQ;AAAA,UAC7B,CAAC;AAAA,QACH;AAAA,MACF;AAAA,MACA,KAAK;AACH,eAAO,EAAE,QAAQ,kBAAkB,EAAE;AAAA,MACvC,KAAK;AACH,eAAO,EAAE,QAAQ,cAAc,EAAE;AAAA,MACnC;AACE,eAAO;AAAA,UACL,OAAO;AAAA,YACL,MAAM;AAAA,YACN,SAAS,mBAAoB,IAA2B,MAAM;AAAA,UAChE;AAAA,QACF;AAAA,IACJ;AAAA,EACF,SAAS,KAAK;AACZ,QAAI,eAAe,WAAW;AAC5B,aAAO,EAAE,OAAO,EAAE,MAAM,IAAI,MAAM,SAAS,IAAI,SAAS,SAAS,IAAI,QAAQ,EAAE;AAAA,IACjF;AACA,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,WAAO,EAAE,OAAO,EAAE,MAAM,kBAAkB,QAAQ,EAAE;AAAA,EACtD;AACF;AAIA,eAAe,QAAQ,QAAgD;AACrE,QAAM,SAAmB,CAAC;AAC1B,mBAAiB,SAAS,QAAQ;AAChC,WAAO,KAAK,OAAO,SAAS,KAAK,IAAI,QAAQ,OAAO,KAAK,KAAe,CAAC;AAAA,EAC3E;AACA,SAAO,OAAO,OAAO,MAAM,EAAE,SAAS,OAAO;AAC/C;AAGA,eAAsB,WAAW,QAAiB,UAAsB,CAAC,GAAoB;AAC3F,QAAM,MAAM,MAAM,QAAQ,QAAQ,KAAK;AACvC,MAAI;AACJ,MAAI;AACF,UAAM,OAAO,KAAK,MAAM,GAAG;AAC3B,UAAM,SAAS,EAAE,QAAwC,QAAQ,KAAK,IAAK;AAAA,EAC7E,SAAS,KAAK;AACZ,YAAQ,OAAO;AAAA,MACb,GAAG,KAAK,UAAU;AAAA,QAChB,OAAO;AAAA,UACL,MAAM;AAAA,UACN,SAAS,6BAA6B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACxF;AAAA,MACF,CAAC,CAAC;AAAA;AAAA,IACJ;AACA,WAAO;AAAA,EACT;AACA,QAAM,MAAM,MAAM,YAAY,KAAK,OAAO;AAC1C,UAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,GAAG,CAAC;AAAA,CAAI;AAC/C,SAAO,WAAW,MAAM,IAAI;AAC9B;AAGA,eAAsB,YAAY,QAAiB,UAAsB,CAAC,GAAoB;AAC5F,QAAM,MAAM,MAAM,QAAQ,QAAQ,KAAK;AACvC,QAAM,QAAQ,IAAI,MAAM,IAAI,EAAE,OAAO,CAAC,MAAM,EAAE,KAAK,EAAE,SAAS,CAAC;AAC/D,MAAI,WAAW;AACf,aAAW,QAAQ,OAAO;AACxB,QAAI;AACJ,QAAI;AACF,YAAM,OAAO,KAAK,MAAM,IAAI;AAC5B,YAAM,SAAS,EAAE,QAAwC,QAAQ,KAAK,IAAK;AAAA,IAC7E,SAAS,KAAK;AACZ,cAAQ,OAAO;AAAA,QACb,GAAG,KAAK,UAAU;AAAA,UAChB,OAAO;AAAA,YACL,MAAM;AAAA,YACN,SAAS,4BAA4B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UACvF;AAAA,QACF,CAAC,CAAC;AAAA;AAAA,MACJ;AACA,iBAAW;AACX;AAAA,IACF;AACA,UAAM,MAAM,MAAM,YAAY,KAAK,OAAO;AAC1C,YAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,GAAG,CAAC;AAAA,CAAI;AAC/C,QAAI,WAAW,IAAK,YAAW;AAAA,EACjC;AACA,SAAO;AACT;;;AC/HA,SAA0B,aAAa;AACvC,SAAS,YAAY;AACrB,SAAS,YAAY;AAcrB,IAAM,aAAa,KAAK,IAAI;AAyB5B,IAAM,oBAAoB,oBAAI,IAAI,CAAC,YAAY,eAAe,eAAe,CAAC;AAEvE,SAAS,UAAU,OAAyB,CAAC,GAAG;AACrD,QAAM,MAAM,IAAI,KAAK;AAErB,MAAI,IAAI,KAAK,KAAK,CAAC;AAGnB,MAAI,KAAK,MAAM;AACb,UAAM,SAAS,KAAK,KAAK;AACzB,QAAI,IAAI,KAAK,OAAO,GAAG,SAAS;AAC9B,YAAM,OAAO,IAAI,IAAI,EAAE,IAAI,GAAG,EAAE;AAChC,UAAI,kBAAkB,IAAI,IAAI,EAAG,QAAO,KAAK;AAC7C,YAAM,MAAM,EAAE,IAAI,OAAO,eAAe,KAAK;AAC7C,YAAM,QAAQ,IAAI,MAAM,kBAAkB;AAC1C,UAAI,CAAC,OAAO;AACV,cAAM,IAAI,UAAU,gBAAgB,8CAA8C,GAAG;AAAA,MACvF;AACA,YAAM,QAAQ,MAAM,CAAC;AACrB,YAAM,KAAK,OAAO,WAAW,WAAW,UAAU,SAAS,MAAM,OAAO,KAAK;AAC7E,UAAI,CAAC,IAAI;AACP,cAAM,IAAI,UAAU,gBAAgB,yBAAyB,GAAG;AAAA,MAClE;AACA,aAAO,KAAK;AAAA,IACd,CAAC;AAAA,EACH;AAEA,MAAI,QAAQ,CAAC,KAAK,MAAM;AACtB,QAAI,eAAe,WAAW;AAC5B,YAAM,SAAS,IAAI;AACnB,aAAO,EAAE;AAAA,QACP,EAAE,OAAO,EAAE,MAAM,IAAI,MAAM,SAAS,IAAI,SAAS,SAAS,IAAI,QAAQ,EAAE;AAAA,QACxE;AAAA,MACF;AAAA,IACF;AAEA,YAAQ,MAAM,iCAAiC,GAAG;AAClD,WAAO,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,kBAAkB,SAAS,yBAAyB,EAAE,GAAG,GAAG;AAAA,EAC7F,CAAC;AAGD,MAAI;AAAA,IAAI;AAAA,IAAY,CAAC,MACnB,EAAE,KAAK,EAAE,QAAQ,MAAe,YAAY,KAAK,IAAI,IAAI,cAAc,IAAK,CAAC;AAAA,EAC/E;AAGA,MAAI,IAAI,eAAe,CAAC,MAAM,EAAE,KAAK,cAAc,CAAC,CAAC;AAGrD,MAAI,IAAI,eAAe,CAAC,MAAM,EAAE,KAAK,kBAAkB,CAAC,CAAC;AAGzD,MAAI,KAAK,aAAa,OAAO,MAAM;AACjC,UAAM,MAAM,MAAM,EAAE,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI;AAC/C,QAAI,OAAO,MAAM;AACf,YAAM,IAAI,UAAU,oBAAoB,8BAA8B,GAAG;AAAA,IAC3E;AACA,UAAM,SAAS,mBAAmB,UAAU,GAAG;AAC/C,QAAI,CAAC,OAAO,SAAS;AACnB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,QACA;AAAA,QACA,OAAO,MAAM;AAAA,MACf;AAAA,IACF;AACA,UAAM,SAAS,MAAM,YAAY,OAAO,MAAM;AAAA,MAC5C,KAAK,KAAK;AAAA,MACV,cAAc,KAAK;AAAA,MACnB,mBAAmB,KAAK;AAAA,IAC1B,CAAC;AACD,WAAO,EAAE,KAAK,MAAM;AAAA,EACtB,CAAC;AAGD,MAAI,KAAK,qBAAqB,OAAO,MAAM;AACzC,UAAM,cAAc,EAAE,IAAI,OAAO,cAAc,KAAK;AACpD,QAAI;AACJ,QAAI,YAAY,SAAS,sBAAsB,GAAG;AAChD,YAAM,OAAO,MAAM,EAAE,IAAI,KAAK;AAC9B,YAAM,SAAS,KACZ,MAAM,IAAI,EACV,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,OAAO,CAAC,SAAS,KAAK,SAAS,CAAC,EAChC,IAAI,CAAC,SAAS;AACb,YAAI;AACF,iBAAO,KAAK,MAAM,IAAI;AAAA,QACxB,QAAQ;AACN,gBAAM,IAAI;AAAA,YACR;AAAA,YACA;AAAA,YACA;AAAA,YACA,KAAK,MAAM,GAAG,GAAG;AAAA,UACnB;AAAA,QACF;AAAA,MACF,CAAC;AACH,gBAAU,EAAE,OAAO;AAAA,IACrB,OAAO;AACL,gBAAU,MAAM,EAAE,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI;AAAA,IAC/C;AACA,QAAI,WAAW,MAAM;AACnB,YAAM,IAAI,UAAU,oBAAoB,wCAAwC,GAAG;AAAA,IACrF;AACA,UAAM,SAAS,0BAA0B,UAAU,OAAO;AAC1D,QAAI,CAAC,OAAO,SAAS;AACnB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,QACA;AAAA,QACA,OAAO,MAAM;AAAA,MACf;AAAA,IACF;AACA,UAAM,SAAS,MAAM,mBAAmB,OAAO,MAAM,KAAK,UAAU,CAAC,CAAC;AACtE,WAAO,EAAE,KAAK,MAAM;AAAA,EACtB,CAAC;AAGD,MAAI,KAAK,gBAAgB,OAAO,MAAM;AACpC,UAAM,MAAM,MAAM,EAAE,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI;AAC/C,QAAI,OAAO,MAAM;AACf,YAAM,IAAI,UAAU,oBAAoB,8BAA8B,GAAG;AAAA,IAC3E;AACA,UAAM,SAAS,yBAAyB,UAAU,GAAG;AACrD,QAAI,CAAC,OAAO,SAAS;AACnB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,QACA;AAAA,QACA,OAAO,MAAM;AAAA,MACf;AAAA,IACF;AACA,UAAM,SAAS,MAAM,qBAAqB,OAAO,MAAM,KAAK,UAAU,CAAC,CAAC;AACxE,WAAO,EAAE,KAAK,MAAM;AAAA,EACtB,CAAC;AAGD,MAAI,IAAI,iBAAiB,CAAC,MAAM,EAAE,KAAK,aAAa,cAAc,EAAE,OAAO,CAAC,CAAC;AAE7E,SAAO;AACT;AASO,SAAS,YAAY,OAAqB,CAAC,GAAe;AAC/D,QAAM,MAAM,UAAU,IAAI;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO,MAAM,EAAE,OAAO,IAAI,OAAO,MAAM,UAAU,KAAK,GAAG,CAAC,EAAE,SAAS,MAAM,WAAW,MAAM;AAE1F,YAAQ,IAAI,kCAAkC,OAAO,IAAI,UAAU,EAAE;AAAA,EACvE,CAAC;AACH;AAmBO,SAAS,iBAAiB,OAAqB,CAAC,GAA2B;AAChF,QAAM,MAAM,UAAU,IAAI;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO,IAAI,QAAQ,CAACC,UAAS,WAAW;AACtC,QAAI,UAAU;AACd,QAAI;AACJ,aAAS,MAAM,EAAE,OAAO,IAAI,OAAO,MAAM,UAAU,KAAK,GAAG,CAAC,EAAE,SAAS,MAAM,WAAW,MAAM;AAC5F,UAAI,QAAS;AACb,gBAAU;AAEV,cAAQ,IAAI,kCAAkC,OAAO,IAAI,UAAU,EAAE;AACrE,MAAAA,SAAQ;AAAA,QACN;AAAA,QACA,MAAM;AAAA,QACN,MAAM;AAAA,QACN,OAAO,MACL,IAAI,QAAc,CAAC,KAAK,QAAQ;AAC9B,iBAAQ,MAAM,CAAC,QAAS,MAAM,IAAI,GAAG,IAAI,IAAI,CAAE;AAAA,QACjD,CAAC;AAAA,MACL,CAAC;AAAA,IACH,CAAC;AACD,WAAO,GAAG,SAAS,CAAC,QAAQ;AAC1B,UAAI,QAAS;AACb,gBAAU;AACV,aAAO,GAAG;AAAA,IACZ,CAAC;AAAA,EACH,CAAC;AACH;","names":["result","resolve"]}

package/dist/chunk-OIUOT4QD.js

@@ -1,44 +0,0 @@
-// src/rollout/reward.ts
-function isRealnessGated(record) {
-  return record.outcome.realness?.gated === true;
-}
-function observedSplitScore(record, split) {
-  return split === "search" ? record.outcome.searchScore : record.outcome.holdoutScore;
-}
-function observedScore(record, prefer = "holdout") {
-  const other = prefer === "search" ? "holdout" : "search";
-  return observedSplitScore(record, prefer) ?? observedSplitScore(record, other);
-}
-function scoreOrigin(record, prefer = "holdout") {
-  const other = prefer === "search" ? "holdout" : "search";
-  if (observedSplitScore(record, prefer) !== void 0) return prefer;
-  if (observedSplitScore(record, other) !== void 0) return other;
-  return "unscored";
-}
-function trainingScore(record, prefer = "holdout") {
-  if (isRealnessGated(record)) return 0;
-  return observedScore(record, prefer);
-}
-function trainingReward(record) {
-  return { reward: trainingScore(record) ?? null, gated: isRealnessGated(record) };
-}
-function rolloutRewardFields(record) {
-  const { reward, gated } = trainingReward(record);
-  const screened = record.outcome.realness !== void 0;
-  return { reward, realness_gated: gated, ...screened ? { realness_screened: true } : {} };
-}
-function unscreenedRewardFields(score) {
-  return { reward: score, realness_gated: false, realness_screened: false };
-}
-
-export {
-  isRealnessGated,
-  observedSplitScore,
-  observedScore,
-  scoreOrigin,
-  trainingScore,
-  trainingReward,
-  rolloutRewardFields,
-  unscreenedRewardFields
-};
-//# sourceMappingURL=chunk-OIUOT4QD.js.map

package/dist/chunk-OIUOT4QD.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rollout/reward.ts"],"sourcesContent":["/**\n * The two named score derivations every consumer must choose between.\n *\n * The anti-Goodhart gate (`outcome.realness.gated`) only holds if it is\n * impossible to read a run's score WITHOUT deciding whether the gate applies.\n * A bare `outcome.holdoutScore ?? outcome.searchScore` makes that decision\n * invisible — and silently answers \"no gate\", which is the wrong default on\n * every path that produces training data. So the expression lives here, once,\n * behind two names that force the caller to state the intent:\n *\n *   - `trainingScore` / `trainingReward` — GATED. Anything that becomes\n *     training data, or a reward a trainer consumes, uses these.\n *   - `observedScore` — RAW. Analysis, reporting, and reward-hack DETECTION\n *     need the ungated number; that is how a gamed run is visible at all.\n *\n * A leaf module on purpose: it imports only the `RunRecord` type, so gate and\n * reporting code can depend on it without pulling in the trace store that\n * `mint.ts` needs.\n */\n\nimport type { RunRecord } from '../run-record'\n\n/**\n * Which split's score wins when a record carries both. `'holdout'` is the\n * canonical \"real signal\" default; `'search'` exists because some callers\n * deliberately score on the search split when both are present.\n */\nexport type ScorePreference = 'holdout' | 'search'\n\n/** Only the outcome is read, so every accessor here accepts anything carrying one. */\ntype Scored = Pick<RunRecord, 'outcome'>\n\n/** True when the authenticity gate flagged the run as gamed (`realness.gated`). */\nexport function isRealnessGated(record: Scored): boolean {\n  return record.outcome.realness?.gated === true\n}\n\n/**\n * The RAW score recorded on ONE split, with no cross-split fallback and no\n * anti-Goodhart gate.\n *\n * The narrowest of the three raw readers, and the one every split-scoped\n * consumer wants: a per-split report, a promotion gate, or a paired comparison\n * asks \"what did this run score on the split I am summarising\", and answering\n * it with the other split's number silently mixes populations. `undefined` =\n * that split was never scored.\n *\n * Same warning as `observedScore`: this INCLUDES runs flagged as gamed. Never\n * feed it into training data.\n */\nexport function observedSplitScore(record: Scored, split: ScorePreference): number | undefined {\n  return split === 'search' ? record.outcome.searchScore : record.outcome.holdoutScore\n}\n\n/**\n * The RAW split score the run carries, with NO anti-Goodhart gate applied.\n *\n * INCLUDES RUNS FLAGGED AS GAMED (`outcome.realness.gated === true`); NEVER\n * feed this into training data — a fine-tune that sees it learns from gamed\n * successes. It is exported anyway because analysis, reporting, and\n * reward-hacking detection legitimately need the ungated number: forcing a\n * gamed run to 0 collapses the proxy signal toward ground truth and makes a\n * detector report \"clean\" on exactly the population that is being gamed.\n *\n * Returns `undefined` when the record carries neither score — an unscored run\n * is a labeled gap, not a measured zero, and each caller picks its own\n * sentinel (`?? 0`, `?? null`, skip, throw). Non-finite values are returned\n * as-is; callers that care keep their own `Number.isFinite` guard.\n */\nexport function observedScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  return observedSplitScore(record, prefer) ?? observedSplitScore(record, other)\n}\n\n/** Which split actually carried the score, or that none did. */\nexport type ScoreOrigin = 'holdout' | 'search' | 'unscored'\n\n/**\n * Where `observedScore` / `trainingScore` read their number from — the\n * provenance label a rollout line's `reward_source` is built from, and the\n * only supported way to ask \"was this run scored at all\" without respelling\n * the field access.\n */\nexport function scoreOrigin(record: Scored, prefer: ScorePreference = 'holdout'): ScoreOrigin {\n  const other: ScorePreference = prefer === 'search' ? 'holdout' : 'search'\n  if (observedSplitScore(record, prefer) !== undefined) return prefer\n  if (observedSplitScore(record, other) !== undefined) return other\n  return 'unscored'\n}\n\n/**\n * The GATED score — the only derivation allowed to reach training data.\n *\n * A realness-gated run scores 0 no matter what it claims, so a fine-tune\n * cannot learn from a gamed success. An unscored run stays `undefined` (a\n * labeled gap), keeping \"we never measured this\" distinct from \"we measured\n * zero\"; callers that need a number apply their own sentinel.\n */\nexport function trainingScore(\n  record: Scored,\n  prefer: ScorePreference = 'holdout',\n): number | undefined {\n  if (isRealnessGated(record)) return 0\n  return observedScore(record, prefer)\n}\n\n/**\n * `{reward, gated}` as written onto a minted `RolloutLine` — `trainingScore`\n * plus the flag itself, so the gate travels into the exported row and a\n * downstream filter can drop or down-weight the line.\n *\n * An unscored record yields `reward: null`, matching the schema's \"no verdict\n * exists — a labeled gap, never 0\" rule. It previously collapsed to 0, which\n * made a run nobody graded indistinguishable from one graded as a total\n * failure, and taught any trainer reading the row that the trajectory was bad.\n * A gated run still yields 0, because that IS a verdict: the gate decided.\n */\nexport function trainingReward(record: Scored): { reward: number | null; gated: boolean } {\n  return { reward: trainingScore(record) ?? null, gated: isRealnessGated(record) }\n}\n\n/**\n * A reward the CALLER computed, with the gate applied on top.\n *\n * Several published exporters accept a `rewardOf` hook so a consumer can drive\n * training off a verifiable signal instead of the headline judge score. That is\n * a real feature and the gate does not remove it — but a reward derived from a\n * gamed run is a gamed reward whatever its source, so it falls to 0 exactly\n * like the score it replaced.\n *\n * It lives here because the same hook exists under the same name in two\n * modules, and for a while only one of them gated it: `rl/exporters.toGrpoRows`\n * forced a gated run to 0 while `rl/preferences.extractPreferences` handed the\n * caller's number straight through, which put a gamed run on the CHOSEN side of\n * every DPO pair against its honest sibling. Two same-named hooks with\n * different gating behaviour was the defect; one implementation is the fix.\n *\n * `null` means \"no reward\" and is preserved: a non-finite or absent value is a\n * gap, and a gap is not a zero.\n */\nexport function trainingRewardOverride(record: Scored, value: number | null): number | null {\n  if (value === null || !Number.isFinite(value)) return null\n  return isRealnessGated(record) ? 0 : value\n}\n\n/**\n * The two fields a rollout line's `outcome` has to state TOGETHER — the reward\n * and whether the authenticity gate fired on it.\n *\n * Stating them together is the point. `realness_gated` used to be optional on\n * the wire, so a producer that wrote `reward` and simply never thought about\n * the flag emitted a row on which `isLineRealnessGated` was structurally false\n * — which is how a second minting door existed for every supervisor and worker\n * row without anyone noticing it had never consulted the gate. The wire schema\n * now requires the flag, and this pair is how a producer states it.\n */\nexport interface RolloutRewardFields {\n  reward: number | null\n  realness_gated: boolean\n  /**\n   * Whether a screen RAN, as distinct from its verdict. Omitted when the\n   * producer cannot tell — see `unscreenedRewardFields` and the field's own doc\n   * on `RolloutOutcome`.\n   */\n  realness_screened?: boolean\n}\n\n/**\n * `trainingReward` in the line's own field names — what `mintRolloutRows`\n * writes onto every row it produces.\n *\n * `realness_screened: true` is written only when the record actually carries an\n * `outcome.realness` verdict, i.e. a screen genuinely ran and reported. A\n * record without one gets the field OMITTED rather than `false`: the record\n * cannot distinguish \"the screen ran elsewhere and this pipeline did not record\n * it\" from \"no screen exists\", and `false` is a load-bearing claim that\n * `assertMinted` refuses on. Absent = unknown, which is the truth here.\n */\nexport function rolloutRewardFields(record: Scored): RolloutRewardFields {\n  const { reward, gated } = trainingReward(record)\n  const screened = record.outcome.realness !== undefined\n  return { reward, realness_gated: gated, ...(screened ? { realness_screened: true } : {}) }\n}\n\n/**\n * The same fields for a producer that has a graded score but NO `RunRecord`\n * behind it — a supervision journal, a harness session store, any second\n * intake. There is no `outcome.realness` to consult, so no authenticity screen\n * has run on this score.\n *\n * `realness_gated: false` alone was the WRONG way to say that. `false` is the\n * gate's VERDICT, and a verdict is a claim that somebody looked; on this path\n * nobody did, so the row asserted \"screened and clean\" about a score that had\n * never been screened at all — indistinguishable on the wire from a genuinely\n * clean one, which is exactly what an adversary needs. The two claims are now\n * carried by two fields:\n *\n *   - `realness_screened: false` — no screen ran. Explicit, not inferable from\n *     absence, and impossible to read as \"clean\" because it is not the gate's\n *     verdict field.\n *   - `realness_gated: false` — no gate fired, which is trivially true when no\n *     gate ran, and is what a consumer filtering on the flag expects to see.\n *\n * `assertMinted` REFUSES a line carrying `realness_screened: false` with a\n * reward above zero, so these rows can be written, read, reported and analysed,\n * but a positive one cannot become training data until something screens it.\n *\n * A separate function from `rolloutRewardFields` on purpose: picking this one\n * is a producer declaring that its rewards were never screened.\n */\nexport function unscreenedRewardFields(score: number | null): RolloutRewardFields {\n  return { reward: score, realness_gated: false, realness_screened: false }\n}\n"],"mappings":";AAiCO,SAAS,gBAAgB,QAAyB;AACvD,SAAO,OAAO,QAAQ,UAAU,UAAU;AAC5C;AAeO,SAAS,mBAAmB,QAAgB,OAA4C;AAC7F,SAAO,UAAU,WAAW,OAAO,QAAQ,cAAc,OAAO,QAAQ;AAC1E;AAiBO,SAAS,cACd,QACA,SAA0B,WACN;AACpB,QAAM,QAAyB,WAAW,WAAW,YAAY;AACjE,SAAO,mBAAmB,QAAQ,MAAM,KAAK,mBAAmB,QAAQ,KAAK;AAC/E;AAWO,SAAS,YAAY,QAAgB,SAA0B,WAAwB;AAC5F,QAAM,QAAyB,WAAW,WAAW,YAAY;AACjE,MAAI,mBAAmB,QAAQ,MAAM,MAAM,OAAW,QAAO;AAC7D,MAAI,mBAAmB,QAAQ,KAAK,MAAM,OAAW,QAAO;AAC5D,SAAO;AACT;AAUO,SAAS,cACd,QACA,SAA0B,WACN;AACpB,MAAI,gBAAgB,MAAM,EAAG,QAAO;AACpC,SAAO,cAAc,QAAQ,MAAM;AACrC;AAaO,SAAS,eAAe,QAA2D;AACxF,SAAO,EAAE,QAAQ,cAAc,MAAM,KAAK,MAAM,OAAO,gBAAgB,MAAM,EAAE;AACjF;AA2DO,SAAS,oBAAoB,QAAqC;AACvE,QAAM,EAAE,QAAQ,MAAM,IAAI,eAAe,MAAM;AAC/C,QAAM,WAAW,OAAO,QAAQ,aAAa;AAC7C,SAAO,EAAE,QAAQ,gBAAgB,OAAO,GAAI,WAAW,EAAE,mBAAmB,KAAK,IAAI,CAAC,EAAG;AAC3F;AA4BO,SAAS,uBAAuB,OAA2C;AAChF,SAAO,EAAE,QAAQ,OAAO,gBAAgB,OAAO,mBAAmB,MAAM;AAC1E;","names":[]}

package/dist/chunk-ONWEPEDO.js

@@ -1,57 +0,0 @@
-// src/errors.ts
-var AgentEvalError = class extends Error {
-  /** Stable string code. Survives minification; safe to switch on. */
-  code;
-  constructor(code, message, options) {
-    super(message, options);
-    this.name = this.constructor.name;
-    this.code = code;
-  }
-};
-var ValidationError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("validation", message, options);
-  }
-};
-var NotFoundError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("not_found", message, options);
-  }
-};
-var ConfigError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("config", message, options);
-  }
-};
-var CaptureIntegrityError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("capture_integrity", message, options);
-  }
-};
-var JudgeError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("judge", message, options);
-  }
-};
-var VerificationError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("verification", message, options);
-  }
-};
-var ReplayError = class extends AgentEvalError {
-  constructor(message, options) {
-    super("replay", message, options);
-  }
-};
-
-export {
-  AgentEvalError,
-  ValidationError,
-  NotFoundError,
-  ConfigError,
-  CaptureIntegrityError,
-  JudgeError,
-  VerificationError,
-  ReplayError
-};
-//# sourceMappingURL=chunk-ONWEPEDO.js.map

package/dist/chunk-ONWEPEDO.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/errors.ts"],"sourcesContent":["/**\n * Error taxonomy for `@tangle-network/agent-eval`.\n *\n * Every error this package throws as part of its *public contract* extends\n * `AgentEvalError`. Consumers can pattern-match by `instanceof <Subclass>` or\n * by the stable string `code` carried on the base class.\n *\n * The codes are stable across minor versions; new codes can be added, but\n * existing codes never change meaning. New subclasses are non-breaking.\n *\n * Internal invariant guards (`throw new Error('this should never happen')`)\n * remain plain `Error`s on purpose — they're programmer-mistake assertions,\n * not consumer-catchable contract failures.\n */\n\nexport type AgentEvalErrorCode =\n  | 'validation'\n  | 'not_found'\n  | 'config'\n  | 'capture_integrity'\n  | 'judge'\n  | 'verification'\n  | 'replay'\n  | 'backend_integrity'\n  | 'profile_matrix'\n\n/**\n * Base class for every contract error this package throws — carries the stable\n * string `code` taxonomy so consumers can `instanceof`-match or switch on `code`.\n */\nexport class AgentEvalError extends Error {\n  /** Stable string code. Survives minification; safe to switch on. */\n  readonly code: AgentEvalErrorCode\n\n  constructor(code: AgentEvalErrorCode, message: string, options?: { cause?: unknown }) {\n    super(message, options)\n    this.name = this.constructor.name\n    this.code = code\n  }\n}\n\n/** Caller passed invalid arguments (out of range, mutually-exclusive options, bad shape). */\nexport class ValidationError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('validation', message, options)\n  }\n}\n\n/** A named resource (run, span, rubric, scenario, dataset row, route) does not exist. */\nexport class NotFoundError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('not_found', message, options)\n  }\n}\n\n/** Configuration missing or malformed (`HOME` unset, required image not supplied, env var absent). */\nexport class ConfigError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('config', message, options)\n  }\n}\n\n/**\n * A run is missing the artifacts a launch-grade check requires:\n * raw HTTP capture absent, no LLM spans, route assertion failed, run-end\n * assertion tripped. Block ship on this; do not catch and move on.\n */\nexport class CaptureIntegrityError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('capture_integrity', message, options)\n  }\n}\n\n/** A judge call failed in a way that's not retryable: schema parse failure, bad rubric, conflicting dimensions. */\nexport class JudgeError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('judge', message, options)\n  }\n}\n\n/** A verifier signalled a hard failure (compile, test, schema) — distinct from a low judge score. */\nexport class VerificationError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('verification', message, options)\n  }\n}\n\n/** Replay cache cannot satisfy a request: miss with no fallback, sink lacks list(), unsupported URL. */\nexport class ReplayError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('replay', message, options)\n  }\n}\n"],"mappings":";AA8BO,IAAM,iBAAN,cAA6B,MAAM;AAAA;AAAA,EAE/B;AAAA,EAET,YAAY,MAA0B,SAAiB,SAA+B;AACpF,UAAM,SAAS,OAAO;AACtB,SAAK,OAAO,KAAK,YAAY;AAC7B,SAAK,OAAO;AAAA,EACd;AACF;AAGO,IAAM,kBAAN,cAA8B,eAAe;AAAA,EAClD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,cAAc,SAAS,OAAO;AAAA,EACtC;AACF;AAGO,IAAM,gBAAN,cAA4B,eAAe;AAAA,EAChD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,aAAa,SAAS,OAAO;AAAA,EACrC;AACF;AAGO,IAAM,cAAN,cAA0B,eAAe;AAAA,EAC9C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,UAAU,SAAS,OAAO;AAAA,EAClC;AACF;AAOO,IAAM,wBAAN,cAAoC,eAAe;AAAA,EACxD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,qBAAqB,SAAS,OAAO;AAAA,EAC7C;AACF;AAGO,IAAM,aAAN,cAAyB,eAAe;AAAA,EAC7C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,SAAS,SAAS,OAAO;AAAA,EACjC;AACF;AAGO,IAAM,oBAAN,cAAgC,eAAe;AAAA,EACpD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,gBAAgB,SAAS,OAAO;AAAA,EACxC;AACF;AAGO,IAAM,cAAN,cAA0B,eAAe;AAAA,EAC9C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,UAAU,SAAS,OAAO;AAAA,EAClC;AACF;","names":[]}

package/dist/chunk-OWN5NPMC.js

@@ -1,152 +0,0 @@
-import {
-  assertRewardGate
-} from "./chunk-PC5DOSM7.js";
-
-// src/rollout/exporters.ts
-function realnessLabels(line) {
-  return {
-    realness_gated: line.outcome.realness_gated === true,
-    // `?? null` rather than `?? false`: absent means the producer never stated
-    // it, and `false` is a load-bearing claim ("no screen exists") that would be
-    // an invention here.
-    realness_screened: line.outcome.realness_screened ?? null
-  };
-}
-function toSftRows(lines, options = {}) {
-  for (const line of lines) assertRewardGate(line, "SFT export");
-  return lines.filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0).map((line) => ({
-    // Dropped, not kept-and-masked: a copied-context turn was authored by
-    // another agent, and an SFT trainer has no notion of "present but not a
-    // target" — every message it is handed is something to imitate.
-    messages: line.messages.filter((message) => message.is_copied_context !== true),
-    metadata: {
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      candidate_id: line.candidate_id ?? null,
-      instance_id: line.task.instance_id,
-      reward: line.outcome.reward,
-      ...realnessLabels(line)
-    }
-  })).filter((row) => row.messages.length > 0);
-}
-function toRewardRows(lines, options = {}) {
-  for (const line of lines) assertRewardGate(line, "reward-row export");
-  return lines.filter(
-    (line) => isTrainingLineEligible(line, options) && line.messages.some(
-      (message) => message.role === "user" && typeof message.content === "string" && message.content.length > 0
-    )
-  ).map((line) => ({
-    prompt: line.messages.find((m) => m.role === "user")?.content ?? "",
-    steps: line.steps ?? [],
-    reward: line.outcome.reward,
-    metadata: {
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      candidate_id: line.candidate_id ?? null,
-      instance_id: line.task.instance_id,
-      split: line.task.split,
-      ...realnessLabels(line)
-    }
-  }));
-}
-function firstAssistantIndex(messages) {
-  const index = messages.findIndex((m) => m.role === "assistant");
-  return index === -1 ? messages.length : index;
-}
-function toVerifiersRolloutOutput(line) {
-  assertRewardGate(line, "verifiers export");
-  const split = firstAssistantIndex(line.messages);
-  return {
-    prompt: line.messages.slice(0, split),
-    completion: line.messages.slice(split),
-    reward: line.outcome.reward,
-    metrics: line.outcome.metrics,
-    tool_defs: line.tool_defs,
-    token_usage: {
-      input_tokens: line.cost.tokens_in,
-      output_tokens: line.cost.tokens_out,
-      reasoning_tokens: line.cost.tokens_reasoning,
-      cache_read_tokens: line.cost.cache_read,
-      cache_write_tokens: line.cost.cache_write
-    },
-    info: {
-      task: line.task,
-      policy: line.policy,
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      experiment_id: line.experiment_id ?? null,
-      candidate_id: line.candidate_id ?? null,
-      generation: line.generation,
-      candidate_index: line.candidate_index,
-      role: line.role,
-      ...realnessLabels(line)
-    }
-  };
-}
-function toVerifiersRolloutOutputs(lines, options = {}) {
-  if (options.gatedLines === "zero-and-flag") {
-    return lines.filter((line) => isSplitEligible(line, options) && line.messages.length > 0).map(toVerifiersRolloutOutput);
-  }
-  return lines.filter(
-    (line) => isTrainingLineEligible(line, options) && firstAssistantIndex(line.messages) > 0 && firstAssistantIndex(line.messages) < line.messages.length
-  ).map(toVerifiersRolloutOutput);
-}
-function toRftItem(line) {
-  assertRewardGate(line, "RFT export");
-  const split = firstAssistantIndex(line.messages);
-  return {
-    messages: line.messages.slice(0, split),
-    reference: {
-      reward: line.outcome.reward,
-      reward_source: line.outcome.reward_source,
-      verdict: line.outcome.verdict,
-      instance_id: line.task.instance_id,
-      suite: line.task.suite,
-      split: line.task.split,
-      rollout_id: line.rollout_id,
-      ...realnessLabels(line)
-    }
-  };
-}
-function toRftItems(lines, options = {}) {
-  return lines.filter(
-    (line) => (options.gatedLines === "zero-and-flag" ? isSplitEligible(line, options) : isTrainingLineEligible(line, options)) && line.messages.length > 0 && firstAssistantIndex(line.messages) > 0
-  ).map(toRftItem);
-}
-function toJsonl(rows) {
-  return rows.map((r) => JSON.stringify(r)).join("\n") + (rows.length ? "\n" : "");
-}
-function isSplitEligible(line, options) {
-  if (line.task.split === "search") return true;
-  return line.task.split === "holdout" && options.allowHeldOutTrainingData === true;
-}
-function isTrainingLineEligible(line, options) {
-  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0;
-  if (!Number.isFinite(minimumQualityExclusive)) {
-    throw new Error("minimumQualityExclusive must be finite");
-  }
-  const reward = line.outcome.reward;
-  if (reward === null) return false;
-  if (!Number.isFinite(reward)) {
-    throw new Error(`training reward for rollout "${line.rollout_id}" must be finite`);
-  }
-  if (reward <= minimumQualityExclusive) return false;
-  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {
-    return false;
-  }
-  if (line.outcome.realness_gated === true) return false;
-  return isSplitEligible(line, options);
-}
-
-export {
-  realnessLabels,
-  toSftRows,
-  toRewardRows,
-  toVerifiersRolloutOutput,
-  toVerifiersRolloutOutputs,
-  toRftItem,
-  toRftItems,
-  toJsonl,
-  isSplitEligible
-};
-//# sourceMappingURL=chunk-OWN5NPMC.js.map

package/dist/chunk-OWN5NPMC.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rollout/exporters.ts"],"sourcesContent":["/**\n * Pure exporters over `tangle.rollout.v1` lines → the training-data shapes\n * the improvement loops feed:\n *   - SFT chat JSONL           (clean trainable successes, {messages, metadata})\n *   - reward rows              (every scored line, success or failure, with steps)\n *   - Prime Intellect verifiers RolloutOutput (prompt/completion split + reward)\n *   - OpenAI RFT items         (prompt turns + verdict reference fields)\n *\n * All exporters are pure functions of the lines — filtering (never train on\n * holdout, reward thresholds, the realness gate) happens HERE, on inline\n * labels, no joins.\n *\n * Every exporter takes `MintedRolloutLine[]`, not `RolloutLine[]`: the reward\n * on a minted line has been checked against the anti-Goodhart invariant, and\n * the brand is what stops a hand-built object literal claiming a positive\n * reward on a gamed run from being handed to an exporter that copies it\n * verbatim into training data.\n */\n\nimport type {\n  ChatMessage,\n  MintedRolloutLine,\n  RolloutLine,\n  RolloutSplit,\n  RolloutStep,\n  ToolDef,\n} from './schema'\nimport { assertRewardGate } from './schema'\n\n// ---------------------------------------------------------------------------\n// The realness claims every emitted row carries.\n// ---------------------------------------------------------------------------\n\n/**\n * The gate's two claims, which travel TOGETHER on every emitted row.\n *\n * `realness_gated` alone is ambiguous, and the ambiguity is exploitable:\n * `false` reads as \"we screened it and nothing fired\", so a producer that has no\n * screen at all emitted rows indistinguishable from screened-clean ones, and\n * every consumer of the published dataset read them as clean. The second field\n * is what separates the two claims, and it only removes the ambiguity if it\n * reaches the WIRE — for a round it existed on `RolloutOutcome` and on no\n * exported row shape at all, which left the published rows exactly as ambiguous\n * as before.\n *\n * So there is one helper and every row shape spreads it. A row that states one\n * claim without the other is not constructible by copying the pattern, and\n * `exporters.test.ts` walks every emitted shape to prove none does.\n */\nexport interface RealnessLabels {\n  /** The screen's VERDICT: the run faked its success signal. */\n  realness_gated: boolean\n  /**\n   * Whether a screen RAN at all. `true` = it ran, so `realness_gated` is its\n   * verdict. `false` = the producer declares it has none. `null` = not stated\n   * (pre-unification producers), which is \"unknown\" and never \"clean\".\n   */\n  realness_screened: boolean | null\n}\n\nexport function realnessLabels(line: MintedRolloutLine): RealnessLabels {\n  return {\n    realness_gated: line.outcome.realness_gated === true,\n    // `?? null` rather than `?? false`: absent means the producer never stated\n    // it, and `false` is a load-bearing claim (\"no screen exists\") that would be\n    // an invention here.\n    realness_screened: line.outcome.realness_screened ?? null,\n  }\n}\n\n// ---------------------------------------------------------------------------\n// (a) SFT chat JSONL\n// ---------------------------------------------------------------------------\n\nexport interface TrainingExportOptions {\n  /** Include held-out evaluation data in training output. Default false. */\n  allowHeldOutTrainingData?: boolean\n  /** Require reward to be strictly greater than this value. Default 0. */\n  minimumQualityExclusive?: number\n}\n\n/**\n * What a signed-signal exporter (verifiers, RFT) does with lines that are not\n * clean trainable successes — realness-gated lines above all.\n *\n *   - 'exclude'       — the default, the same fail-closed policy as every\n *                       other training export: positive, completed,\n *                       non-gated rows on a trainable split.\n *   - 'zero-and-flag' — keep them, at their non-positive (or null) reward,\n *                       with `RealnessLabels` on the row. The dataset release\n *                       sets this per `FORMAT_GATE_DISPOSITION`: in these\n *                       formats the reward is a signed learning signal, so a\n *                       gamed trajectory at reward 0 is a correct negative,\n *                       and dropping it would bias the negative population\n *                       toward honest failures and leave a trainer no example\n *                       of gaming being penalized. The split policy is NOT\n *                       relaxed: held-out lines still need the named opt-in.\n *\n * SFT deliberately has no such option — an SFT row is an imitation target and\n * a gamed trajectory must never appear in one at any weight.\n */\nexport type GatedLineDisposition = 'exclude' | 'zero-and-flag'\n\nexport interface SignedSignalExportOptions extends TrainingExportOptions {\n  /** Disposition for non-trainable lines. Default 'exclude'. */\n  gatedLines?: GatedLineDisposition\n}\n\nexport type SftExportOptions = TrainingExportOptions\n\nexport interface SftRow {\n  messages: ChatMessage[]\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    reward: number\n  } & RealnessLabels\n}\n\n/**\n * Supervised fine-tune rows: the completed conversation of each qualifying\n * line. Fail-closed filters: trainable split only (never holdout/canary),\n * reward strictly above `minimumQualityExclusive` (default 0), realness-gated\n * lines never qualify, gap lines carry no trainable content, and\n * copied-context turns are dropped from the transcript (Harbor ATIF RFC 0001\n * rule 7 — see `ChatMessage.is_copied_context`).\n *\n * `realness_gated` is therefore always `false` on an emitted row. It is carried\n * anyway: an SFT row is a pure imitation target, so the row states its realness\n * claims instead of making the reader know the format's policy, and carrying\n * both flags on all four shapes is what lets the release accounting measure\n * every config with one rule rather than skipping the one whose row shape\n * happened to omit the field.\n */\nexport function toSftRows(lines: MintedRolloutLine[], options: SftExportOptions = {}): SftRow[] {\n  for (const line of lines) assertRewardGate(line, 'SFT export')\n  return lines\n    .filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0)\n    .map((line) => ({\n      // Dropped, not kept-and-masked: a copied-context turn was authored by\n      // another agent, and an SFT trainer has no notion of \"present but not a\n      // target\" — every message it is handed is something to imitate.\n      messages: line.messages.filter((message) => message.is_copied_context !== true),\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        reward: line.outcome.reward as number,\n        ...realnessLabels(line),\n      },\n    }))\n    .filter((row) => row.messages.length > 0)\n}\n\n// ---------------------------------------------------------------------------\n// (b) Reward rows\n// ---------------------------------------------------------------------------\n\nexport interface RewardRow {\n  /** First user turn — the task prompt. */\n  prompt: string\n  steps: RolloutStep[]\n  reward: number\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    split: RolloutSplit\n    /**\n     * The anti-Goodhart verdict AND whether a screen produced it, carried on the\n     * row rather than left implicit in a reward of 0. Zeroing alone is lossy: it\n     * makes a run that FAKED its success indistinguishable from one that\n     * honestly failed, so a consumer can neither drop the gamed population nor\n     * mine it (a labeled gamed trajectory is the training signal for a gaming\n     * detector). `realness_gated` alone is lossy the same way in the other\n     * direction — see `RealnessLabels`.\n     */\n  } & RealnessLabels\n}\n\n/**\n * Reward-labeled rows for completed, positive-quality training runs.\n */\nexport function toRewardRows(\n  lines: MintedRolloutLine[],\n  options: TrainingExportOptions = {},\n): RewardRow[] {\n  for (const line of lines) assertRewardGate(line, 'reward-row export')\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        line.messages.some(\n          (message) =>\n            message.role === 'user' &&\n            typeof message.content === 'string' &&\n            message.content.length > 0,\n        ),\n    )\n    .map((line) => ({\n      prompt: line.messages.find((m) => m.role === 'user')?.content ?? '',\n      steps: line.steps ?? [],\n      reward: line.outcome.reward as number,\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        split: line.task.split,\n        ...realnessLabels(line),\n      },\n    }))\n}\n\n// ---------------------------------------------------------------------------\n// (c) Prime Intellect verifiers RolloutOutput\n// ---------------------------------------------------------------------------\n\nexport interface VerifiersTokenUsage {\n  input_tokens: number | null\n  output_tokens: number | null\n  reasoning_tokens: number | null\n  cache_read_tokens: number | null\n  cache_write_tokens: number | null\n}\n\nexport interface VerifiersRolloutOutput {\n  /** Messages through the last turn BEFORE the first assistant turn. */\n  prompt: ChatMessage[]\n  /** The first assistant turn onward — what the policy produced. */\n  completion: ChatMessage[]\n  reward: number | null\n  metrics: Record<string, unknown>\n  tool_defs: ToolDef[]\n  token_usage: VerifiersTokenUsage\n  info: {\n    task: RolloutLine['task']\n    policy: RolloutLine['policy']\n    rollout_id: string\n    run_id: string\n    experiment_id: string | null\n    candidate_id: string | null\n    generation: number | null\n    candidate_index: number | null\n    role: RolloutLine['role']\n  } & RealnessLabels\n}\n\n/** Index of the first assistant turn; messages.length when none exists. */\nfunction firstAssistantIndex(messages: ChatMessage[]): number {\n  const index = messages.findIndex((m) => m.role === 'assistant')\n  return index === -1 ? messages.length : index\n}\n\nexport function toVerifiersRolloutOutput(line: MintedRolloutLine): VerifiersRolloutOutput {\n  assertRewardGate(line, 'verifiers export')\n  const split = firstAssistantIndex(line.messages)\n  return {\n    prompt: line.messages.slice(0, split),\n    completion: line.messages.slice(split),\n    reward: line.outcome.reward,\n    metrics: line.outcome.metrics,\n    tool_defs: line.tool_defs,\n    token_usage: {\n      input_tokens: line.cost.tokens_in,\n      output_tokens: line.cost.tokens_out,\n      reasoning_tokens: line.cost.tokens_reasoning,\n      cache_read_tokens: line.cost.cache_read,\n      cache_write_tokens: line.cost.cache_write,\n    },\n    info: {\n      task: line.task,\n      policy: line.policy,\n      rollout_id: line.rollout_id,\n      run_id: line.run_id,\n      experiment_id: line.experiment_id ?? null,\n      candidate_id: line.candidate_id ?? null,\n      generation: line.generation,\n      candidate_index: line.candidate_index,\n      role: line.role,\n      ...realnessLabels(line),\n    },\n  }\n}\n\nexport function toVerifiersRolloutOutputs(\n  lines: MintedRolloutLine[],\n  options: SignedSignalExportOptions = {},\n): VerifiersRolloutOutput[] {\n  if (options.gatedLines === 'zero-and-flag') {\n    return lines\n      .filter((line) => isSplitEligible(line, options) && line.messages.length > 0)\n      .map(toVerifiersRolloutOutput)\n  }\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        firstAssistantIndex(line.messages) > 0 &&\n        firstAssistantIndex(line.messages) < line.messages.length,\n    )\n    .map(toVerifiersRolloutOutput)\n}\n\n// ---------------------------------------------------------------------------\n// (d) OpenAI RFT items\n// ---------------------------------------------------------------------------\n\nexport interface RftItem {\n  /** Prompt turns only — the graded completion is re-sampled during RFT. */\n  messages: ChatMessage[]\n  /** Verdict/label fields the grader references as item.reference.* */\n  reference: {\n    reward: number | null\n    reward_source: string | null\n    verdict: unknown\n    instance_id: string\n    suite: string\n    split: RolloutSplit\n    rollout_id: string\n  } & RealnessLabels\n}\n\nexport function toRftItem(line: MintedRolloutLine): RftItem {\n  assertRewardGate(line, 'RFT export')\n  const split = firstAssistantIndex(line.messages)\n  return {\n    messages: line.messages.slice(0, split),\n    reference: {\n      reward: line.outcome.reward,\n      reward_source: line.outcome.reward_source,\n      verdict: line.outcome.verdict,\n      instance_id: line.task.instance_id,\n      suite: line.task.suite,\n      split: line.task.split,\n      rollout_id: line.rollout_id,\n      ...realnessLabels(line),\n    },\n  }\n}\n\n/** RFT needs a real prompt: lines whose transcript starts with prompt turns. */\nexport function toRftItems(\n  lines: MintedRolloutLine[],\n  options: SignedSignalExportOptions = {},\n): RftItem[] {\n  return lines\n    .filter(\n      (line) =>\n        (options.gatedLines === 'zero-and-flag'\n          ? isSplitEligible(line, options)\n          : isTrainingLineEligible(line, options)) &&\n        line.messages.length > 0 &&\n        firstAssistantIndex(line.messages) > 0,\n    )\n    .map(toRftItem)\n}\n\n// ---------------------------------------------------------------------------\n// Serialization — one JSON object per line, the interchange format for\n// every export. `tangle.rollout.v1` lines and export rows alike.\n// ---------------------------------------------------------------------------\n\nexport function toJsonl(rows: ReadonlyArray<unknown>): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length ? '\\n' : '')\n}\n\n/**\n * The split half of the training policy alone: `search` is trainable, held-out\n * needs the named opt-in, `dev` and `canary` never ship. This is the ONE check\n * `'zero-and-flag'` does not relax — a gated line is shipped as a labeled\n * negative, not as a licence to train on evaluation data.\n *\n * Exported as the single implementation of that rule: `rl/exporters` applies\n * it on its line paths too, so the two waists cannot drift on which splits are\n * trainable.\n */\nexport function isSplitEligible(\n  line: RolloutLine,\n  options: Pick<TrainingExportOptions, 'allowHeldOutTrainingData'>,\n): boolean {\n  if (line.task.split === 'search') return true\n  return line.task.split === 'holdout' && options.allowHeldOutTrainingData === true\n}\n\nfunction isTrainingLineEligible(\n  line: RolloutLine,\n  options: TrainingExportOptions,\n): line is RolloutLine & { outcome: RolloutLine['outcome'] & { reward: number } } {\n  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0\n  if (!Number.isFinite(minimumQualityExclusive)) {\n    throw new Error('minimumQualityExclusive must be finite')\n  }\n\n  const reward = line.outcome.reward\n  if (reward === null) return false\n  if (!Number.isFinite(reward)) {\n    throw new Error(`training reward for rollout \"${line.rollout_id}\" must be finite`)\n  }\n  if (reward <= minimumQualityExclusive) return false\n  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {\n    return false\n  }\n  if (line.outcome.realness_gated === true) return false\n  return isSplitEligible(line, options)\n}\n"],"mappings":";;;;;AA4DO,SAAS,eAAe,MAAyC;AACtE,SAAO;AAAA,IACL,gBAAgB,KAAK,QAAQ,mBAAmB;AAAA;AAAA;AAAA;AAAA,IAIhD,mBAAmB,KAAK,QAAQ,qBAAqB;AAAA,EACvD;AACF;AAoEO,SAAS,UAAU,OAA4B,UAA4B,CAAC,GAAa;AAC9F,aAAW,QAAQ,MAAO,kBAAiB,MAAM,YAAY;AAC7D,SAAO,MACJ,OAAO,CAAC,SAAS,uBAAuB,MAAM,OAAO,KAAK,KAAK,SAAS,SAAS,CAAC,EAClF,IAAI,CAAC,UAAU;AAAA;AAAA;AAAA;AAAA,IAId,UAAU,KAAK,SAAS,OAAO,CAAC,YAAY,QAAQ,sBAAsB,IAAI;AAAA,IAC9E,UAAU;AAAA,MACR,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,cAAc,KAAK,gBAAgB;AAAA,MACnC,aAAa,KAAK,KAAK;AAAA,MACvB,QAAQ,KAAK,QAAQ;AAAA,MACrB,GAAG,eAAe,IAAI;AAAA,IACxB;AAAA,EACF,EAAE,EACD,OAAO,CAAC,QAAQ,IAAI,SAAS,SAAS,CAAC;AAC5C;AAgCO,SAAS,aACd,OACA,UAAiC,CAAC,GACrB;AACb,aAAW,QAAQ,MAAO,kBAAiB,MAAM,mBAAmB;AACpE,SAAO,MACJ;AAAA,IACC,CAAC,SACC,uBAAuB,MAAM,OAAO,KACpC,KAAK,SAAS;AAAA,MACZ,CAAC,YACC,QAAQ,SAAS,UACjB,OAAO,QAAQ,YAAY,YAC3B,QAAQ,QAAQ,SAAS;AAAA,IAC7B;AAAA,EACJ,EACC,IAAI,CAAC,UAAU;AAAA,IACd,QAAQ,KAAK,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,GAAG,WAAW;AAAA,IACjE,OAAO,KAAK,SAAS,CAAC;AAAA,IACtB,QAAQ,KAAK,QAAQ;AAAA,IACrB,UAAU;AAAA,MACR,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,cAAc,KAAK,gBAAgB;AAAA,MACnC,aAAa,KAAK,KAAK;AAAA,MACvB,OAAO,KAAK,KAAK;AAAA,MACjB,GAAG,eAAe,IAAI;AAAA,IACxB;AAAA,EACF,EAAE;AACN;AAqCA,SAAS,oBAAoB,UAAiC;AAC5D,QAAM,QAAQ,SAAS,UAAU,CAAC,MAAM,EAAE,SAAS,WAAW;AAC9D,SAAO,UAAU,KAAK,SAAS,SAAS;AAC1C;AAEO,SAAS,yBAAyB,MAAiD;AACxF,mBAAiB,MAAM,kBAAkB;AACzC,QAAM,QAAQ,oBAAoB,KAAK,QAAQ;AAC/C,SAAO;AAAA,IACL,QAAQ,KAAK,SAAS,MAAM,GAAG,KAAK;AAAA,IACpC,YAAY,KAAK,SAAS,MAAM,KAAK;AAAA,IACrC,QAAQ,KAAK,QAAQ;AAAA,IACrB,SAAS,KAAK,QAAQ;AAAA,IACtB,WAAW,KAAK;AAAA,IAChB,aAAa;AAAA,MACX,cAAc,KAAK,KAAK;AAAA,MACxB,eAAe,KAAK,KAAK;AAAA,MACzB,kBAAkB,KAAK,KAAK;AAAA,MAC5B,mBAAmB,KAAK,KAAK;AAAA,MAC7B,oBAAoB,KAAK,KAAK;AAAA,IAChC;AAAA,IACA,MAAM;AAAA,MACJ,MAAM,KAAK;AAAA,MACX,QAAQ,KAAK;AAAA,MACb,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,eAAe,KAAK,iBAAiB;AAAA,MACrC,cAAc,KAAK,gBAAgB;AAAA,MACnC,YAAY,KAAK;AAAA,MACjB,iBAAiB,KAAK;AAAA,MACtB,MAAM,KAAK;AAAA,MACX,GAAG,eAAe,IAAI;AAAA,IACxB;AAAA,EACF;AACF;AAEO,SAAS,0BACd,OACA,UAAqC,CAAC,GACZ;AAC1B,MAAI,QAAQ,eAAe,iBAAiB;AAC1C,WAAO,MACJ,OAAO,CAAC,SAAS,gBAAgB,MAAM,OAAO,KAAK,KAAK,SAAS,SAAS,CAAC,EAC3E,IAAI,wBAAwB;AAAA,EACjC;AACA,SAAO,MACJ;AAAA,IACC,CAAC,SACC,uBAAuB,MAAM,OAAO,KACpC,oBAAoB,KAAK,QAAQ,IAAI,KACrC,oBAAoB,KAAK,QAAQ,IAAI,KAAK,SAAS;AAAA,EACvD,EACC,IAAI,wBAAwB;AACjC;AAqBO,SAAS,UAAU,MAAkC;AAC1D,mBAAiB,MAAM,YAAY;AACnC,QAAM,QAAQ,oBAAoB,KAAK,QAAQ;AAC/C,SAAO;AAAA,IACL,UAAU,KAAK,SAAS,MAAM,GAAG,KAAK;AAAA,IACtC,WAAW;AAAA,MACT,QAAQ,KAAK,QAAQ;AAAA,MACrB,eAAe,KAAK,QAAQ;AAAA,MAC5B,SAAS,KAAK,QAAQ;AAAA,MACtB,aAAa,KAAK,KAAK;AAAA,MACvB,OAAO,KAAK,KAAK;AAAA,MACjB,OAAO,KAAK,KAAK;AAAA,MACjB,YAAY,KAAK;AAAA,MACjB,GAAG,eAAe,IAAI;AAAA,IACxB;AAAA,EACF;AACF;AAGO,SAAS,WACd,OACA,UAAqC,CAAC,GAC3B;AACX,SAAO,MACJ;AAAA,IACC,CAAC,UACE,QAAQ,eAAe,kBACpB,gBAAgB,MAAM,OAAO,IAC7B,uBAAuB,MAAM,OAAO,MACxC,KAAK,SAAS,SAAS,KACvB,oBAAoB,KAAK,QAAQ,IAAI;AAAA,EACzC,EACC,IAAI,SAAS;AAClB;AAOO,SAAS,QAAQ,MAAsC;AAC5D,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,OAAO;AAC/E;AAYO,SAAS,gBACd,MACA,SACS;AACT,MAAI,KAAK,KAAK,UAAU,SAAU,QAAO;AACzC,SAAO,KAAK,KAAK,UAAU,aAAa,QAAQ,6BAA6B;AAC/E;AAEA,SAAS,uBACP,MACA,SACgF;AAChF,QAAM,0BAA0B,QAAQ,2BAA2B;AACnE,MAAI,CAAC,OAAO,SAAS,uBAAuB,GAAG;AAC7C,UAAM,IAAI,MAAM,wCAAwC;AAAA,EAC1D;AAEA,QAAM,SAAS,KAAK,QAAQ;AAC5B,MAAI,WAAW,KAAM,QAAO;AAC5B,MAAI,CAAC,OAAO,SAAS,MAAM,GAAG;AAC5B,UAAM,IAAI,MAAM,gCAAgC,KAAK,UAAU,kBAAkB;AAAA,EACnF;AACA,MAAI,UAAU,wBAAyB,QAAO;AAC9C,MAAI,CAAC,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,UAAU,MAAM;AAC1F,WAAO;AAAA,EACT;AACA,MAAI,KAAK,QAAQ,mBAAmB,KAAM,QAAO;AACjD,SAAO,gBAAgB,MAAM,OAAO;AACtC;","names":[]}