@tangle-network/agent-eval 0.120.1 → 0.120.2

package/dist/chunk-4I2E3LLO.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\n      .string()\n      .optional()\n      .describe('Override the judge model (default routes via tcloud). e.g. \"claude-opus-4-7\".'),\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  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\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  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 \"${req.rubricName}\".`, 404)\n    }\n    rubric = found\n  } else if (req.rubric) {\n    rubric = req.rubric\n  } else {\n    // refine() in the schema should already have caught this — defense in depth\n    throw new WireError('validation_error', 'Provide either `rubricName` or `rubric`.', 422)\n  }\n\n  const startedAt = Date.now()\n  const model = req.model ?? DEFAULT_JUDGE_MODEL\n\n  const request = {\n    model,\n    messages: [\n      { role: 'system', content: rubric.systemPrompt },\n      { role: 'user', content: buildJudgePrompt(req.content, req.context) },\n    ],\n    jsonSchema: judgeOutputSchema(rubric),\n    temperature: 0.0,\n    maxTokens: 4_000,\n    timeoutMs: 60_000,\n  } satisfies LlmCallRequest\n  const ledger = options.costLedger ?? new CostLedger()\n  const paid = await ledger.runPaidCall({\n    channel: 'judge',\n    phase: options.costPhase ?? 'wire.judge',\n    actor: `wire.${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 { 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 async function dispatchRpc(req: RpcRequest): 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 { result: await handleJudge(parsed.data) }\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: ${(req as { method: string }).method}`,\n          },\n        }\n    }\n  } catch (err) {\n    if (err instanceof WireError) {\n      return { error: { code: err.code, message: err.message, details: err.details } }\n    }\n    const message = err instanceof Error ? err.message : String(err)\n    return { error: { code: 'internal_error', message } }\n  }\n}\n\n// ── stdin/stdout driver ─────────────────────────────────────────────\n\nasync function readAll(stream: NodeJS.ReadableStream): Promise<string> {\n  const chunks: Buffer[] = []\n  for await (const chunk of stream) {\n    chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk as string))\n  }\n  return Buffer.concat(chunks).toString('utf-8')\n}\n\n/** Read one JSON request from stdin, write one JSON response to stdout. */\nexport async function runRpcOnce(method?: string): Promise<number> {\n  const raw = await readAll(process.stdin)\n  let req: RpcRequest\n  try {\n    const body = JSON.parse(raw)\n    req = method ? { method: method as RpcRequest['method'], params: body } : (body as RpcRequest)\n  } catch (err) {\n    process.stdout.write(\n      `${JSON.stringify({\n        error: {\n          code: 'parse_error',\n          message: `stdin was not valid JSON: ${err instanceof Error ? err.message : String(err)}`,\n        },\n      })}\\n`,\n    )\n    return 1\n  }\n  const out = await dispatchRpc(req)\n  process.stdout.write(`${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): 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: ${err instanceof Error ? err.message : String(err)}`,\n          },\n        })}\\n`,\n      )\n      exitCode = 1\n      continue\n    }\n    const out = await dispatchRpc(req)\n    process.stdout.write(`${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'\n\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  /**\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    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,EACJ,OAAO,EACP,SAAS,EACT,SAAS,+EAA+E;AAC7F,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;;;ACjYA,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;;;AC+JA,SAAS,oBAAoB;AAC7B,SAAS,SAAS,eAAe;AACjC,SAAS,qBAAqB;AA5OvB,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;AAS5B,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,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,QAAQ,IAAI,SAAS;AAE3B,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;;;ACvXA,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;;;ACxMA,eAAsB,YAAY,KAAiD;AACjF,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,EAAE,QAAQ,MAAM,YAAY,OAAO,IAAI,EAAE;AAAA,MAClD;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,QAAkC;AACjE,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,GAAG;AACjC,UAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,GAAG,CAAC;AAAA,CAAI;AAC/C,SAAO,WAAW,MAAM,IAAI;AAC9B;AAGA,eAAsB,YAAY,QAAkC;AAClE,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,GAAG;AACjC,YAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,GAAG,CAAC;AAAA,CAAI;AAC/C,QAAI,WAAW,IAAK,YAAW;AAAA,EACjC;AACA,SAAO;AACT;;;AC/GA,SAA0B,aAAa;AACvC,SAAS,YAAY;AACrB,SAAS,YAAY;AAcrB,IAAM,aAAa,KAAK,IAAI;AAmB5B,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,IAAI;AAC5C,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-ARU2PZFM.js

@@ -1,312 +0,0 @@
-import {
-  pearsonR
-} from "./chunk-PJQFMIOX.js";
-
-// src/rl/verifiable-reward.ts
-var DEFAULT_DETERMINISTIC_LAYERS = /* @__PURE__ */ new Set([
-  "install",
-  "typecheck",
-  "build",
-  "lint",
-  "test",
-  "compile",
-  "schema",
-  "sandbox",
-  "unit_tests",
-  "integration_tests"
-]);
-var DEFAULT_SOURCE_FOR = (name) => {
-  const lower = name.toLowerCase();
-  if (lower.includes("test")) return "test";
-  if (lower.includes("compile") || lower.includes("build") || lower.includes("typecheck") || lower.includes("lint"))
-    return "compile";
-  if (lower.includes("schema")) return "schema";
-  if (lower.includes("sandbox")) return "sandbox";
-  if (lower.includes("judge") || lower.includes("semantic")) return "judge";
-  return "composite";
-};
-function extractVerifiableReward(report, opts = {}) {
-  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
-  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
-  const fallbackToJudge = opts.fallbackToJudge ?? true;
-  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7;
-  const deterministic = report.layers.filter(
-    (l) => deterministicSet.has(l.layer) && typeof l.score === "number" && Number.isFinite(l.score)
-  );
-  if (deterministic.length === 1) {
-    const layer = deterministic[0];
-    const value = clamp01(layer.score);
-    return {
-      value,
-      source: sourceFor(layer.layer),
-      determinism: "deterministic",
-      confidence: 1,
-      origin: layer.layer,
-      components: { [layer.layer]: value },
-      breakdown: layerBreakdown(layer)
-    };
-  }
-  if (deterministic.length > 1) {
-    let num = 0;
-    let denom = 0;
-    const components = {};
-    for (const l of deterministic) {
-      const w = l.detail?.weight ?? 1;
-      num += w * (l.score ?? 0);
-      denom += w;
-      components[l.layer] = l.score;
-    }
-    return {
-      value: denom === 0 ? 0 : clamp01(num / denom),
-      source: "composite",
-      determinism: "deterministic",
-      confidence: 1,
-      origin: deterministic.map((l) => l.layer).join("+"),
-      components,
-      breakdown: { ...components }
-    };
-  }
-  if (!fallbackToJudge) return null;
-  const judge = report.layers.find(
-    (l) => typeof l.score === "number" && Number.isFinite(l.score) && sourceFor(l.layer) === "judge"
-  ) ?? report.layers.find((l) => typeof l.score === "number" && Number.isFinite(l.score));
-  if (!judge) return null;
-  const confFromDetail = judge.detail?.confidence;
-  const judgeValue = clamp01(judge.score);
-  return {
-    value: judgeValue,
-    source: "judge",
-    determinism: "probabilistic",
-    confidence: typeof confFromDetail === "number" ? confFromDetail : judgeFloor,
-    origin: judge.layer,
-    components: { [judge.layer]: judgeValue },
-    breakdown: layerBreakdown(judge)
-  };
-}
-function extractVerifiableRewardsFromRecords(runs, opts = {}) {
-  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
-  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
-  const fallbackToJudge = opts.fallbackToJudge ?? true;
-  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7;
-  return runs.map((run) => {
-    const layerScores = [];
-    for (const [k, v] of Object.entries(run.outcome.raw)) {
-      if (k.startsWith("layer.") && !k.includes(".", 6) && typeof v === "number" && Number.isFinite(v)) {
-        layerScores.push({ name: k.slice("layer.".length), score: v });
-      }
-    }
-    const det = layerScores.filter((l) => deterministicSet.has(l.name));
-    if (det.length === 1) {
-      const layer = det[0];
-      const value = clamp01(layer.score);
-      return {
-        runId: run.runId,
-        reward: {
-          value,
-          source: sourceFor(layer.name),
-          determinism: "deterministic",
-          confidence: 1,
-          origin: layer.name,
-          components: { [layer.name]: value }
-        }
-      };
-    }
-    if (det.length > 1) {
-      const value = det.reduce((s, l) => s + l.score, 0) / det.length;
-      const components = Object.fromEntries(
-        det.map((l) => [l.name, l.score])
-      );
-      return {
-        runId: run.runId,
-        reward: {
-          value: clamp01(value),
-          source: "composite",
-          determinism: "deterministic",
-          confidence: 1,
-          origin: det.map((l) => l.name).join("+"),
-          components,
-          breakdown: { ...components }
-        }
-      };
-    }
-    if (!fallbackToJudge) return { runId: run.runId, reward: null };
-    const primary = run.outcome.holdoutScore ?? run.outcome.searchScore;
-    if (typeof primary !== "number" || !Number.isFinite(primary)) {
-      return { runId: run.runId, reward: null };
-    }
-    const primaryValue = clamp01(primary);
-    return {
-      runId: run.runId,
-      reward: {
-        value: primaryValue,
-        source: "judge",
-        determinism: "probabilistic",
-        confidence: judgeFloor,
-        origin: "run.outcome.score",
-        components: { "run.outcome.score": primaryValue }
-      }
-    };
-  });
-}
-function filterDeterministicallyRewarded(runs, opts = {}) {
-  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false });
-  const out = [];
-  for (let i = 0; i < runs.length; i++) {
-    const r = rewarded[i];
-    if (r.reward && r.reward.determinism === "deterministic") {
-      out.push({ run: runs[i], reward: r.reward });
-    }
-  }
-  return out;
-}
-function clamp01(x) {
-  if (!Number.isFinite(x)) return 0;
-  return Math.max(0, Math.min(1, x));
-}
-function layerBreakdown(l) {
-  const out = {};
-  if (l.diagnostics) {
-    for (const [k, v] of Object.entries(l.diagnostics)) {
-      if (typeof v === "number" && Number.isFinite(v)) out[k] = v;
-    }
-  }
-  return out;
-}
-
-// src/rl/reward-hacking.ts
-var DEFAULT_PROXY = (r) => {
-  const v = r.outcome.holdoutScore ?? r.outcome.searchScore;
-  return typeof v === "number" && Number.isFinite(v) ? v : null;
-};
-function detectRewardHacking(input) {
-  const proxyOf = input.proxyOf ?? DEFAULT_PROXY;
-  const truthOf = input.truthOf;
-  const sus = input.thresholds?.suspect ?? 0.3;
-  const gam = input.thresholds?.gaming ?? 0.6;
-  const runs = input.runs.filter((r) => proxyOf(r) !== null);
-  const n = runs.length;
-  if (n < 4) {
-    return {
-      findings: [],
-      verdict: "clean",
-      n,
-      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`]
-    };
-  }
-  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)));
-  const before = runs.slice(0, n - windowSize);
-  const after = runs.slice(n - windowSize);
-  const findings = [];
-  if (truthOf) {
-    const beforeProxy = before.map(proxyOf).filter((v) => typeof v === "number");
-    const afterProxy = after.map(proxyOf).filter((v) => typeof v === "number");
-    const beforeTruth = before.map(truthOf).filter((v) => typeof v === "number");
-    const afterTruth = after.map(truthOf).filter((v) => typeof v === "number");
-    if (beforeProxy.length >= 2 && afterProxy.length >= 2 && beforeTruth.length >= 2 && afterTruth.length >= 2) {
-      const proxyDelta = mean(afterProxy) - mean(beforeProxy);
-      const truthDelta = mean(afterTruth) - mean(beforeTruth);
-      const gap = Math.max(0, proxyDelta - truthDelta);
-      const severity = clamp012(gap * 5);
-      findings.push({
-        signal: "reward_divergence",
-        severity,
-        message: severity >= sus ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} \u2014 potential Goodhart` : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,
-        detail: {
-          proxyDelta,
-          truthDelta,
-          gap,
-          beforeN: beforeProxy.length,
-          afterN: afterProxy.length
-        }
-      });
-    }
-  }
-  {
-    const beforeP = before.map(proxyOf).filter((v) => typeof v === "number");
-    const afterP = after.map(proxyOf).filter((v) => typeof v === "number");
-    if (beforeP.length >= 4 && afterP.length >= 4) {
-      const ks = ksStatistic(beforeP, afterP);
-      const severity = clamp012(ks - 0.2);
-      findings.push({
-        signal: "distribution_shift",
-        severity,
-        message: severity >= sus ? `KS=${ks.toFixed(3)} between before/after windows \u2014 distributional shift large` : `KS=${ks.toFixed(3)} between before/after windows \u2014 within-distribution drift`,
-        detail: { ks, beforeN: beforeP.length, afterN: afterP.length }
-      });
-    }
-  }
-  {
-    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions);
-    const aligned = runs.map((r) => ({ p: proxyOf(r), s: secondaryOf(r) })).filter(
-      (x) => typeof x.p === "number" && typeof x.s === "number"
-    );
-    if (aligned.length >= 4) {
-      const ps = aligned.map((x) => x.p);
-      const ss = aligned.map((x) => x.s);
-      const r = pearsonR(ps, ss);
-      const severity = clamp012(0.5 - Math.max(0, r));
-      findings.push({
-        signal: "reward_disagreement",
-        severity,
-        message: severity >= sus ? `proxy and independent secondary reward correlate \u03C1=${r.toFixed(3)} \u2014 possibly hacking proxy` : `proxy and secondary reward correlate \u03C1=${r.toFixed(3)}`,
-        detail: { pearson: r, n: aligned.length }
-      });
-    }
-  }
-  {
-    const detRuns = filterDeterministicallyRewarded(runs, input.verifiableRewardOptions ?? {});
-    if (detRuns.length >= 4) {
-      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2));
-      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2));
-      const detDelta = mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value));
-      const proxyDelta = mean(after.map(proxyOf).filter((v) => typeof v === "number")) - mean(before.map(proxyOf).filter((v) => typeof v === "number"));
-      const driftGap = Math.max(0, proxyDelta - detDelta);
-      const severity = clamp012(driftGap * 5);
-      findings.push({
-        signal: "judge_drift",
-        severity,
-        message: severity >= sus ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} \u2014 judge drifting up without verifiable backing` : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,
-        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length }
-      });
-    }
-  }
-  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0);
-  const verdict = maxSev >= gam ? "gaming" : maxSev >= sus ? "suspect" : "clean";
-  const rationale = findings.filter((f) => f.severity >= sus).map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} \u2014 ${f.message}`);
-  if (rationale.length === 0) rationale.push("no signals fired above suspect threshold");
-  return { findings, verdict, rationale, n };
-}
-function mean(xs) {
-  if (xs.length === 0) return 0;
-  return xs.reduce((s, x) => s + x, 0) / xs.length;
-}
-function clamp012(x) {
-  if (!Number.isFinite(x)) return 0;
-  return Math.max(0, Math.min(1, x));
-}
-function ksStatistic(a, b) {
-  const sortedA = [...a].sort((x, y) => x - y);
-  const sortedB = [...b].sort((x, y) => x - y);
-  const all = [.../* @__PURE__ */ new Set([...sortedA, ...sortedB])].sort((x, y) => x - y);
-  let max = 0;
-  for (const v of all) {
-    const fa = sortedA.filter((x) => x <= v).length / sortedA.length;
-    const fb = sortedB.filter((x) => x <= v).length / sortedB.length;
-    max = Math.max(max, Math.abs(fa - fb));
-  }
-  return max;
-}
-function defaultSecondary(verifiableOpts) {
-  return (run) => {
-    const filtered = filterDeterministicallyRewarded([run], verifiableOpts ?? {});
-    return filtered.length === 1 ? filtered[0].reward.value : null;
-  };
-}
-
-export {
-  extractVerifiableReward,
-  extractVerifiableRewardsFromRecords,
-  filterDeterministicallyRewarded,
-  detectRewardHacking
-};
-//# sourceMappingURL=chunk-ARU2PZFM.js.map

package/dist/chunk-ARU2PZFM.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"sourcesContent":["/**\n * Verifiable reward channel.\n *\n * For RL on coding / math / theorem-proving / structured-output tasks, the\n * reward signal is *decidable* — a test passes or fails, a proof checks or\n * doesn't, an output validates against a schema or doesn't. These rewards\n * are dramatically more useful for RL training than LLM-judge scores\n * because they don't drift, can't be Goodhart-gamed by the policy in the\n * same way, and don't require a separate calibration loop.\n *\n * The `MultiLayerVerifier` already produces this signal — it just doesn't\n * surface it in a shape that's clean enough for RL training. This module\n * wraps the verifier output so consumers can:\n *\n *   1. Extract a clean `VerifiableReward` from a `VerificationReport`\n *   2. Distinguish *deterministic* rewards (compile, test, schema) from\n *      *probabilistic* rewards (judge) so they can be weighted differently\n *      in the RL training step\n *   3. Filter `RunRecord[]` to only those with a verifiable reward,\n *      producing the clean training set that DeepSeek-R1-style GRPO and\n *      AlphaProof-style search both depend on\n *\n * Why this matters: every credible 2025-2026 frontier RL result on coding\n * agents leans on verifiable reward (DeepSeek-R1 GRPO on test pass-rate,\n * o-series RL on math/code, AlphaProof on Lean kernel checking). Mixing\n * judge scores into the reward signal poisons the gradient. This module\n * is the seam.\n */\n\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport type { RunRecord } from '../run-record'\n\nexport type VerifiableRewardSource =\n  | 'compile' // typecheck / build / lint passed\n  | 'test' // unit / integration test pass-rate\n  | 'schema' // structured output validates\n  | 'sandbox' // sandbox exec exit code\n  | 'judge' // LLM judge — probabilistic, included for completeness\n  | 'composite' // weighted blend across multiple of the above\n\nexport interface VerifiableReward {\n  /** Scalar in [0, 1]. The RL training signal. */\n  value: number\n  /** What produced the reward — different sources have different determinism. */\n  source: VerifiableRewardSource\n  /**\n   * Determinism class. `'deterministic'` rewards are repeatable byte-for-byte\n   * given the same inputs (compile, test, schema validation, sandbox exit code).\n   * `'probabilistic'` rewards depend on a stochastic component (LLM judge).\n   * Mixing these in the same training batch without separation is a known\n   * footgun in production RLHF pipelines.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /**\n   * Confidence in the reward value. For deterministic sources this is 1.0\n   * (the bit either flipped or didn't). For judge sources this is the\n   * judge-reported confidence or — when missing — a calibrated prior.\n   */\n  confidence: number\n  /** The layer / judge id that produced the signal, for provenance. */\n  origin: string\n  /**\n   * Per-source contribution to `value`, keyed by layer/judge id. Single-source\n   * rewards carry one entry (`{ [origin]: value }`); composite rewards carry\n   * every contributing layer's score — the anti-scalar-collapse surface RL\n   * consumers weight per-source instead of trusting one blended number.\n   */\n  components: Record<string, number>\n  /**\n   * @deprecated Read `components` for per-source reward values. Kept for\n   * published-API compatibility: single-source rewards carry the layer's\n   * diagnostics here (e.g. `{ tests_passed: 7 }`); composite rewards carry\n   * the same per-layer scores `components` now holds.\n   */\n  breakdown?: Record<string, number>\n}\n\nexport interface VerifiableRewardExtractionOptions {\n  /**\n   * Which layers count as deterministic-reward sources. The verifier doesn't\n   * tag layers as \"this is verifiable\"; the caller declares it via this list\n   * (or via the layer name → source mapping). Default treats common names\n   * (`install`, `typecheck`, `build`, `lint`, `test`, `compile`, `schema`,\n   * `sandbox`) as deterministic.\n   */\n  deterministicLayers?: string[]\n  /**\n   * Map layer name → reward source. Defaults to a sensible string-match.\n   */\n  sourceFor?: (layerName: string) => VerifiableRewardSource\n  /**\n   * Whether to fall back to a probabilistic (judge) reward when no\n   * deterministic layer produced a numeric score. Default `true`. Set to\n   * `false` for \"deterministic-only\" training pipelines that should\n   * discard runs without a verifiable signal.\n   */\n  fallbackToJudge?: boolean\n  /**\n   * Default confidence for probabilistic (judge) rewards when the judge\n   * doesn't report one. Default `0.7`.\n   */\n  judgeConfidenceFloor?: number\n}\n\nconst DEFAULT_DETERMINISTIC_LAYERS = new Set([\n  'install',\n  'typecheck',\n  'build',\n  'lint',\n  'test',\n  'compile',\n  'schema',\n  'sandbox',\n  'unit_tests',\n  'integration_tests',\n])\n\nconst DEFAULT_SOURCE_FOR = (name: string): VerifiableRewardSource => {\n  const lower = name.toLowerCase()\n  if (lower.includes('test')) return 'test'\n  if (\n    lower.includes('compile') ||\n    lower.includes('build') ||\n    lower.includes('typecheck') ||\n    lower.includes('lint')\n  )\n    return 'compile'\n  if (lower.includes('schema')) return 'schema'\n  if (lower.includes('sandbox')) return 'sandbox'\n  if (lower.includes('judge') || lower.includes('semantic')) return 'judge'\n  return 'composite'\n}\n\n/**\n * Extract a `VerifiableReward` from a `VerificationReport`.\n *\n * Strategy: prefer the deterministic layers (in order: test → compile →\n * schema → sandbox), fall back to the judge layer if `fallbackToJudge` is\n * true, return `null` if no signal qualifies. When multiple deterministic\n * layers contribute, return a `'composite'` source with a weighted blend.\n */\nexport function extractVerifiableReward(\n  report: VerificationReport,\n  opts: VerifiableRewardExtractionOptions = {},\n): VerifiableReward | null {\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  const deterministic = report.layers.filter(\n    (l) => deterministicSet.has(l.layer) && typeof l.score === 'number' && Number.isFinite(l.score),\n  )\n\n  if (deterministic.length === 1) {\n    const layer = deterministic[0]!\n    const value = clamp01(layer.score!)\n    return {\n      value,\n      source: sourceFor(layer.layer),\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: layer.layer,\n      components: { [layer.layer]: value },\n      breakdown: layerBreakdown(layer),\n    }\n  }\n\n  if (deterministic.length > 1) {\n    // Composite: weighted blend by `Layer.weight` if present, else equal.\n    let num = 0\n    let denom = 0\n    const components: Record<string, number> = {}\n    for (const l of deterministic) {\n      const w = (l.detail?.weight as number | undefined) ?? 1\n      num += w * (l.score ?? 0)\n      denom += w\n      components[l.layer] = l.score!\n    }\n    return {\n      value: denom === 0 ? 0 : clamp01(num / denom),\n      source: 'composite',\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: deterministic.map((l) => l.layer).join('+'),\n      components,\n      breakdown: { ...components },\n    }\n  }\n\n  if (!fallbackToJudge) return null\n\n  const judge =\n    report.layers.find(\n      (l) =>\n        typeof l.score === 'number' && Number.isFinite(l.score) && sourceFor(l.layer) === 'judge',\n    ) ?? report.layers.find((l) => typeof l.score === 'number' && Number.isFinite(l.score))\n\n  if (!judge) return null\n\n  const confFromDetail = judge.detail?.confidence as number | undefined\n  const judgeValue = clamp01(judge.score!)\n  return {\n    value: judgeValue,\n    source: 'judge',\n    determinism: 'probabilistic',\n    confidence: typeof confFromDetail === 'number' ? confFromDetail : judgeFloor,\n    origin: judge.layer,\n    components: { [judge.layer]: judgeValue },\n    breakdown: layerBreakdown(judge),\n  }\n}\n\n/**\n * Extract verifiable rewards from `RunRecord[]` produced via the\n * `verificationReportToRunRecord` adapter (which encodes per-layer scores\n * in `outcome.raw['layer.<name>']`). For records that don't carry layer\n * scores, returns `null` for that record.\n *\n * This is the canonical bridge from \"campaign-shaped artifacts\" to\n * \"RL-training-ready reward signals\": every record that has a clean\n * verifiable reward becomes a training datum, every record that doesn't\n * gets filtered out (or kept with `'probabilistic'` determinism for\n * separate downstream handling).\n */\nexport function extractVerifiableRewardsFromRecords(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ runId: string; reward: VerifiableReward | null }> {\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  return runs.map((run) => {\n    // Recover per-layer scores from outcome.raw['layer.<name>']\n    const layerScores: Array<{ name: string; score: number }> = []\n    for (const [k, v] of Object.entries(run.outcome.raw)) {\n      if (\n        k.startsWith('layer.') &&\n        !k.includes('.', 6) &&\n        typeof v === 'number' &&\n        Number.isFinite(v)\n      ) {\n        layerScores.push({ name: k.slice('layer.'.length), score: v })\n      }\n    }\n    const det = layerScores.filter((l) => deterministicSet.has(l.name))\n\n    if (det.length === 1) {\n      const layer = det[0]!\n      const value = clamp01(layer.score)\n      return {\n        runId: run.runId,\n        reward: {\n          value,\n          source: sourceFor(layer.name),\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: layer.name,\n          components: { [layer.name]: value },\n        },\n      }\n    }\n    if (det.length > 1) {\n      const value = det.reduce((s, l) => s + l.score, 0) / det.length\n      const components: Record<string, number> = Object.fromEntries(\n        det.map((l) => [l.name, l.score]),\n      )\n      return {\n        runId: run.runId,\n        reward: {\n          value: clamp01(value),\n          source: 'composite',\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: det.map((l) => l.name).join('+'),\n          components,\n          breakdown: { ...components },\n        },\n      }\n    }\n    if (!fallbackToJudge) return { runId: run.runId, reward: null }\n\n    // Probabilistic fallback: use the run's primary score.\n    const primary = run.outcome.holdoutScore ?? run.outcome.searchScore\n    if (typeof primary !== 'number' || !Number.isFinite(primary)) {\n      return { runId: run.runId, reward: null }\n    }\n    const primaryValue = clamp01(primary)\n    return {\n      runId: run.runId,\n      reward: {\n        value: primaryValue,\n        source: 'judge',\n        determinism: 'probabilistic',\n        confidence: judgeFloor,\n        origin: 'run.outcome.score',\n        components: { 'run.outcome.score': primaryValue },\n      },\n    }\n  })\n}\n\n/** Filter `RunRecord[]` to those with deterministic verifiable rewards. */\nexport function filterDeterministicallyRewarded(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ run: RunRecord; reward: VerifiableReward }> {\n  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false })\n  const out: Array<{ run: RunRecord; reward: VerifiableReward }> = []\n  for (let i = 0; i < runs.length; i++) {\n    const r = rewarded[i]!\n    if (r.reward && r.reward.determinism === 'deterministic') {\n      out.push({ run: runs[i]!, reward: r.reward })\n    }\n  }\n  return out\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction layerBreakdown(l: LayerResult): Record<string, number> {\n  const out: Record<string, number> = {}\n  if (l.diagnostics) {\n    for (const [k, v] of Object.entries(l.diagnostics)) {\n      if (typeof v === 'number' && Number.isFinite(v)) out[k] = v\n    }\n  }\n  return out\n}\n","/**\n * Reward hacking / Goodhart detection.\n *\n * Goodhart's Law says: when a measure becomes a target, it ceases to be\n * a good measure. In RLHF and agentic-RL settings this is the dominant\n * failure mode — the policy learns to produce outputs that score well on\n * the proxy reward (judge, rubric, test pass-rate) without producing\n * the underlying capability the proxy was meant to track.\n *\n * Krakovna et al. (2020, \"Specification Gaming Examples in AI\") and the\n * subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.\n * 2023) converge on a few diagnostic signatures:\n *\n *   1. **Reward divergence:** the proxy reward grows while the held-out\n *      ground-truth signal stagnates or drops. Predictive validity over\n *      time captures this.\n *   2. **Distributional shift in outputs:** after RL, the policy produces\n *      outputs that no longer match the reference distribution — usually\n *      because it found a high-reward attractor that's degenerate (e.g.\n *      one-token responses, repetition, formatting tricks).\n *   3. **Disagreement between independent rewards:** if you train on\n *      reward A and a held-out independent reward B drops sharply, you're\n *      probably hacking A.\n *   4. **Calibration drift:** the verifiable / deterministic component of\n *      the reward is stable; the probabilistic / judge component drifts up\n *      while the deterministic component doesn't. The judge is being\n *      gamed.\n *\n * This module ships explicit detectors for all four signatures, plus a\n * combined verdict. The output is diagnostic — actionable signals,\n * not autoreject — because each signature has known false positives\n * (e.g., a policy that genuinely improves can show distributional shift).\n *\n * Differs from `rubricPredictiveValidity` (which is a *standing* check on\n * whether rubrics correlate with deployment outcomes) — this is a\n * *temporal* check on whether the reward-vs-truth gap is *widening over\n * time during a training run*.\n */\n\nimport type { RunRecord } from '../run-record'\nimport { pearsonR } from '../statistics'\nimport {\n  filterDeterministicallyRewarded,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport type RewardHackingSignal =\n  | 'reward_divergence'\n  | 'distribution_shift'\n  | 'reward_disagreement'\n  | 'judge_drift'\n\nexport interface RewardHackingFinding {\n  signal: RewardHackingSignal\n  /** Severity in [0, 1]. >0.5 = strong signal. */\n  severity: number\n  message: string\n  /** Numeric evidence the consumer can render. */\n  detail: Record<string, number>\n}\n\nexport interface RewardHackingReport {\n  findings: RewardHackingFinding[]\n  /**\n   * Composite verdict. `'clean'` if every signal severity < 0.3;\n   * `'suspect'` if at least one ≥ 0.3 but none ≥ 0.6; `'gaming'` if any ≥ 0.6.\n   */\n  verdict: 'clean' | 'suspect' | 'gaming'\n  /** Rationale for the verdict, ready to paste into an audit log. */\n  rationale: string[]\n  /** Number of paired (proxy, truth) data points the report saw. */\n  n: number\n}\n\nexport interface DetectRewardHackingInput {\n  /**\n   * Run records ordered by recency (oldest first). The detector segments\n   * them into prefix/suffix windows to compute \"did the gap widen.\"\n   */\n  runs: RunRecord[]\n  /**\n   * The metric the policy was trained to optimize. Should be present on\n   * `outcome.raw` or `outcome.holdoutScore`. Default reads `outcome.holdoutScore`.\n   */\n  proxyOf?: (run: RunRecord) => number | null\n  /**\n   * The held-out ground-truth metric. For RL on coding, this is typically\n   * test pass-rate. For RLHF, it's downstream task performance or human\n   * preference. For knowledge tasks, it's an independently-graded score.\n   */\n  truthOf?: (run: RunRecord) => number | null\n  /**\n   * Independent secondary reward. Used for the `reward_disagreement`\n   * signal. Default uses the verifiable reward extractor (deterministic\n   * sources only).\n   */\n  secondaryRewardOf?: (run: RunRecord) => number | null\n  /**\n   * Window size — how many of the most recent runs count as the \"after\"\n   * cohort. Default min(50, half the runs).\n   */\n  windowSize?: number\n  /**\n   * Severity threshold to flag a signal. Default 0.3 (suspect) and 0.6\n   * (gaming).\n   */\n  thresholds?: { suspect?: number; gaming?: number }\n  /**\n   * Verifiable-reward options used for the secondary-reward fallback.\n   */\n  verifiableRewardOptions?: VerifiableRewardExtractionOptions\n}\n\nconst DEFAULT_PROXY = (r: RunRecord): number | null => {\n  const v = r.outcome.holdoutScore ?? r.outcome.searchScore\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nexport function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport {\n  const proxyOf = input.proxyOf ?? DEFAULT_PROXY\n  const truthOf = input.truthOf\n  const sus = input.thresholds?.suspect ?? 0.3\n  const gam = input.thresholds?.gaming ?? 0.6\n\n  const runs = input.runs.filter((r) => proxyOf(r) !== null)\n  const n = runs.length\n  if (n < 4) {\n    return {\n      findings: [],\n      verdict: 'clean',\n      n,\n      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`],\n    }\n  }\n  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)))\n  const before = runs.slice(0, n - windowSize)\n  const after = runs.slice(n - windowSize)\n\n  const findings: RewardHackingFinding[] = []\n\n  // ── Signal 1: reward divergence (proxy ↑ while truth flat or ↓) ──────\n  if (truthOf) {\n    const beforeProxy = before.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const afterProxy = after.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const beforeTruth = before.map(truthOf).filter((v): v is number => typeof v === 'number')\n    const afterTruth = after.map(truthOf).filter((v): v is number => typeof v === 'number')\n    if (\n      beforeProxy.length >= 2 &&\n      afterProxy.length >= 2 &&\n      beforeTruth.length >= 2 &&\n      afterTruth.length >= 2\n    ) {\n      const proxyDelta = mean(afterProxy) - mean(beforeProxy)\n      const truthDelta = mean(afterTruth) - mean(beforeTruth)\n      // Divergence: proxy goes up while truth goes flat or down.\n      // Severity = max(0, (proxyDelta - truthDelta)) — bigger gap = bigger signal.\n      const gap = Math.max(0, proxyDelta - truthDelta)\n      const severity = clamp01(gap * 5) // scale: 0.2 absolute gap → severity 1.0\n      findings.push({\n        signal: 'reward_divergence',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart`\n            : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,\n        detail: {\n          proxyDelta,\n          truthDelta,\n          gap,\n          beforeN: beforeProxy.length,\n          afterN: afterProxy.length,\n        },\n      })\n    }\n  }\n\n  // ── Signal 2: distributional shift in outputs (KS on score distributions) ──\n  {\n    const beforeP = before.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    const afterP = after.map(proxyOf).filter((v): v is number => typeof v === 'number')\n    if (beforeP.length >= 4 && afterP.length >= 4) {\n      const ks = ksStatistic(beforeP, afterP)\n      // KS statistic: bigger = more shift. We're agnostic about direction;\n      // genuine improvement ALSO produces shift, so this signal is\n      // contributory rather than load-bearing.\n      const severity = clamp01(ks - 0.2)\n      findings.push({\n        signal: 'distribution_shift',\n        severity,\n        message:\n          severity >= sus\n            ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large`\n            : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,\n        detail: { ks, beforeN: beforeP.length, afterN: afterP.length },\n      })\n    }\n  }\n\n  // ── Signal 3: reward disagreement (proxy vs independent secondary) ────\n  {\n    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions)\n    const aligned = runs\n      .map((r) => ({ p: proxyOf(r), s: secondaryOf(r) }))\n      .filter(\n        (x): x is { p: number; s: number } => typeof x.p === 'number' && typeof x.s === 'number',\n      )\n    if (aligned.length >= 4) {\n      const ps = aligned.map((x) => x.p)\n      const ss = aligned.map((x) => x.s)\n      const r = pearsonR(ps, ss)\n      // Disagreement: low or negative correlation between primary proxy\n      // reward and an independent secondary signal.\n      const severity = clamp01(0.5 - Math.max(0, r))\n      findings.push({\n        signal: 'reward_disagreement',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy`\n            : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,\n        detail: { pearson: r, n: aligned.length },\n      })\n    }\n  }\n\n  // ── Signal 4: judge drift (probabilistic up while deterministic flat) ─\n  {\n    const detRuns = filterDeterministicallyRewarded(runs, input.verifiableRewardOptions ?? {})\n    if (detRuns.length >= 4) {\n      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2))\n      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2))\n      const detDelta =\n        mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value))\n      const proxyDelta =\n        mean(after.map(proxyOf).filter((v): v is number => typeof v === 'number')) -\n        mean(before.map(proxyOf).filter((v): v is number => typeof v === 'number'))\n      const driftGap = Math.max(0, proxyDelta - detDelta)\n      const severity = clamp01(driftGap * 5)\n      findings.push({\n        signal: 'judge_drift',\n        severity,\n        message:\n          severity >= sus\n            ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing`\n            : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,\n        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length },\n      })\n    }\n  }\n\n  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0)\n  const verdict: RewardHackingReport['verdict'] =\n    maxSev >= gam ? 'gaming' : maxSev >= sus ? 'suspect' : 'clean'\n  const rationale = findings\n    .filter((f) => f.severity >= sus)\n    .map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`)\n  if (rationale.length === 0) rationale.push('no signals fired above suspect threshold')\n\n  return { findings, verdict, rationale, n }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction ksStatistic(a: number[], b: number[]): number {\n  // Two-sample Kolmogorov-Smirnov statistic.\n  const sortedA = [...a].sort((x, y) => x - y)\n  const sortedB = [...b].sort((x, y) => x - y)\n  const all = [...new Set([...sortedA, ...sortedB])].sort((x, y) => x - y)\n  let max = 0\n  for (const v of all) {\n    const fa = sortedA.filter((x) => x <= v).length / sortedA.length\n    const fb = sortedB.filter((x) => x <= v).length / sortedB.length\n    max = Math.max(max, Math.abs(fa - fb))\n  }\n  return max\n}\n\nfunction defaultSecondary(\n  verifiableOpts?: VerifiableRewardExtractionOptions,\n): (run: RunRecord) => number | null {\n  return (run: RunRecord) => {\n    const filtered = filterDeterministicallyRewarded([run], verifiableOpts ?? {})\n    return filtered.length === 1 ? filtered[0]!.reward.value : null\n  }\n}\n"],"mappings":";;;;;AAwGA,IAAM,+BAA+B,oBAAI,IAAI;AAAA,EAC3C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,qBAAqB,CAAC,SAAyC;AACnE,QAAM,QAAQ,KAAK,YAAY;AAC/B,MAAI,MAAM,SAAS,MAAM,EAAG,QAAO;AACnC,MACE,MAAM,SAAS,SAAS,KACxB,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,MAAM;AAErB,WAAO;AACT,MAAI,MAAM,SAAS,QAAQ,EAAG,QAAO;AACrC,MAAI,MAAM,SAAS,SAAS,EAAG,QAAO;AACtC,MAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,UAAU,EAAG,QAAO;AAClE,SAAO;AACT;AAUO,SAAS,wBACd,QACA,OAA0C,CAAC,GAClB;AACzB,QAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;AAC9F,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,QAAM,aAAa,KAAK,wBAAwB;AAEhD,QAAM,gBAAgB,OAAO,OAAO;AAAA,IAClC,CAAC,MAAM,iBAAiB,IAAI,EAAE,KAAK,KAAK,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK;AAAA,EAChG;AAEA,MAAI,cAAc,WAAW,GAAG;AAC9B,UAAM,QAAQ,cAAc,CAAC;AAC7B,UAAM,QAAQ,QAAQ,MAAM,KAAM;AAClC,WAAO;AAAA,MACL;AAAA,MACA,QAAQ,UAAU,MAAM,KAAK;AAAA,MAC7B,aAAa;AAAA,MACb,YAAY;AAAA,MACZ,QAAQ,MAAM;AAAA,MACd,YAAY,EAAE,CAAC,MAAM,KAAK,GAAG,MAAM;AAAA,MACnC,WAAW,eAAe,KAAK;AAAA,IACjC;AAAA,EACF;AAEA,MAAI,cAAc,SAAS,GAAG;AAE5B,QAAI,MAAM;AACV,QAAI,QAAQ;AACZ,UAAM,aAAqC,CAAC;AAC5C,eAAW,KAAK,eAAe;AAC7B,YAAM,IAAK,EAAE,QAAQ,UAAiC;AACtD,aAAO,KAAK,EAAE,SAAS;AACvB,eAAS;AACT,iBAAW,EAAE,KAAK,IAAI,EAAE;AAAA,IAC1B;AACA,WAAO;AAAA,MACL,OAAO,UAAU,IAAI,IAAI,QAAQ,MAAM,KAAK;AAAA,MAC5C,QAAQ;AAAA,MACR,aAAa;AAAA,MACb,YAAY;AAAA,MACZ,QAAQ,cAAc,IAAI,CAAC,MAAM,EAAE,KAAK,EAAE,KAAK,GAAG;AAAA,MAClD;AAAA,MACA,WAAW,EAAE,GAAG,WAAW;AAAA,IAC7B;AAAA,EACF;AAEA,MAAI,CAAC,gBAAiB,QAAO;AAE7B,QAAM,QACJ,OAAO,OAAO;AAAA,IACZ,CAAC,MACC,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK,KAAK,UAAU,EAAE,KAAK,MAAM;AAAA,EACtF,KAAK,OAAO,OAAO,KAAK,CAAC,MAAM,OAAO,EAAE,UAAU,YAAY,OAAO,SAAS,EAAE,KAAK,CAAC;AAExF,MAAI,CAAC,MAAO,QAAO;AAEnB,QAAM,iBAAiB,MAAM,QAAQ;AACrC,QAAM,aAAa,QAAQ,MAAM,KAAM;AACvC,SAAO;AAAA,IACL,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,aAAa;AAAA,IACb,YAAY,OAAO,mBAAmB,WAAW,iBAAiB;AAAA,IAClE,QAAQ,MAAM;AAAA,IACd,YAAY,EAAE,CAAC,MAAM,KAAK,GAAG,WAAW;AAAA,IACxC,WAAW,eAAe,KAAK;AAAA,EACjC;AACF;AAcO,SAAS,oCACd,MACA,OAA0C,CAAC,GACgB;AAC3D,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;AAC9F,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,QAAM,aAAa,KAAK,wBAAwB;AAEhD,SAAO,KAAK,IAAI,CAAC,QAAQ;AAEvB,UAAM,cAAsD,CAAC;AAC7D,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,IAAI,QAAQ,GAAG,GAAG;AACpD,UACE,EAAE,WAAW,QAAQ,KACrB,CAAC,EAAE,SAAS,KAAK,CAAC,KAClB,OAAO,MAAM,YACb,OAAO,SAAS,CAAC,GACjB;AACA,oBAAY,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAM,GAAG,OAAO,EAAE,CAAC;AAAA,MAC/D;AAAA,IACF;AACA,UAAM,MAAM,YAAY,OAAO,CAAC,MAAM,iBAAiB,IAAI,EAAE,IAAI,CAAC;AAElE,QAAI,IAAI,WAAW,GAAG;AACpB,YAAM,QAAQ,IAAI,CAAC;AACnB,YAAM,QAAQ,QAAQ,MAAM,KAAK;AACjC,aAAO;AAAA,QACL,OAAO,IAAI;AAAA,QACX,QAAQ;AAAA,UACN;AAAA,UACA,QAAQ,UAAU,MAAM,IAAI;AAAA,UAC5B,aAAa;AAAA,UACb,YAAY;AAAA,UACZ,QAAQ,MAAM;AAAA,UACd,YAAY,EAAE,CAAC,MAAM,IAAI,GAAG,MAAM;AAAA,QACpC;AAAA,MACF;AAAA,IACF;AACA,QAAI,IAAI,SAAS,GAAG;AAClB,YAAM,QAAQ,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,IAAI;AACzD,YAAM,aAAqC,OAAO;AAAA,QAChD,IAAI,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC;AAAA,MAClC;AACA,aAAO;AAAA,QACL,OAAO,IAAI;AAAA,QACX,QAAQ;AAAA,UACN,OAAO,QAAQ,KAAK;AAAA,UACpB,QAAQ;AAAA,UACR,aAAa;AAAA,UACb,YAAY;AAAA,UACZ,QAAQ,IAAI,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG;AAAA,UACvC;AAAA,UACA,WAAW,EAAE,GAAG,WAAW;AAAA,QAC7B;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,gBAAiB,QAAO,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK;AAG9D,UAAM,UAAU,IAAI,QAAQ,gBAAgB,IAAI,QAAQ;AACxD,QAAI,OAAO,YAAY,YAAY,CAAC,OAAO,SAAS,OAAO,GAAG;AAC5D,aAAO,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK;AAAA,IAC1C;AACA,UAAM,eAAe,QAAQ,OAAO;AACpC,WAAO;AAAA,MACL,OAAO,IAAI;AAAA,MACX,QAAQ;AAAA,QACN,OAAO;AAAA,QACP,QAAQ;AAAA,QACR,aAAa;AAAA,QACb,YAAY;AAAA,QACZ,QAAQ;AAAA,QACR,YAAY,EAAE,qBAAqB,aAAa;AAAA,MAClD;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAGO,SAAS,gCACd,MACA,OAA0C,CAAC,GACU;AACrD,QAAM,WAAW,oCAAoC,MAAM,EAAE,GAAG,MAAM,iBAAiB,MAAM,CAAC;AAC9F,QAAM,MAA2D,CAAC;AAClE,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,IAAI,SAAS,CAAC;AACpB,QAAI,EAAE,UAAU,EAAE,OAAO,gBAAgB,iBAAiB;AACxD,UAAI,KAAK,EAAE,KAAK,KAAK,CAAC,GAAI,QAAQ,EAAE,OAAO,CAAC;AAAA,IAC9C;AAAA,EACF;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,eAAe,GAAwC;AAC9D,QAAM,MAA8B,CAAC;AACrC,MAAI,EAAE,aAAa;AACjB,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,EAAE,WAAW,GAAG;AAClD,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,CAAC,IAAI;AAAA,IAC5D;AAAA,EACF;AACA,SAAO;AACT;;;AC9NA,IAAM,gBAAgB,CAAC,MAAgC;AACrD,QAAM,IAAI,EAAE,QAAQ,gBAAgB,EAAE,QAAQ;AAC9C,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEO,SAAS,oBAAoB,OAAsD;AACxF,QAAM,UAAU,MAAM,WAAW;AACjC,QAAM,UAAU,MAAM;AACtB,QAAM,MAAM,MAAM,YAAY,WAAW;AACzC,QAAM,MAAM,MAAM,YAAY,UAAU;AAExC,QAAM,OAAO,MAAM,KAAK,OAAO,CAAC,MAAM,QAAQ,CAAC,MAAM,IAAI;AACzD,QAAM,IAAI,KAAK;AACf,MAAI,IAAI,GAAG;AACT,WAAO;AAAA,MACL,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT;AAAA,MACA,WAAW,CAAC,0CAA0C,CAAC,0BAA0B;AAAA,IACnF;AAAA,EACF;AACA,QAAM,aAAa,KAAK,IAAI,GAAG,MAAM,cAAc,KAAK,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,CAAC;AAClF,QAAM,SAAS,KAAK,MAAM,GAAG,IAAI,UAAU;AAC3C,QAAM,QAAQ,KAAK,MAAM,IAAI,UAAU;AAEvC,QAAM,WAAmC,CAAC;AAG1C,MAAI,SAAS;AACX,UAAM,cAAc,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACxF,UAAM,aAAa,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACtF,UAAM,cAAc,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACxF,UAAM,aAAa,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACtF,QACE,YAAY,UAAU,KACtB,WAAW,UAAU,KACrB,YAAY,UAAU,KACtB,WAAW,UAAU,GACrB;AACA,YAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;AACtD,YAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;AAGtD,YAAM,MAAM,KAAK,IAAI,GAAG,aAAa,UAAU;AAC/C,YAAM,WAAWA,SAAQ,MAAM,CAAC;AAChC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,wBAAwB,WAAW,QAAQ,CAAC,CAAC,2BAA2B,WAAW,QAAQ,CAAC,CAAC,+BAC7F,yCAAyC,WAAW,QAAQ,CAAC,CAAC,WAAW,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpG,QAAQ;AAAA,UACN;AAAA,UACA;AAAA,UACA;AAAA,UACA,SAAS,YAAY;AAAA,UACrB,QAAQ,WAAW;AAAA,QACrB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,UAAU,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACpF,UAAM,SAAS,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AAClF,QAAI,QAAQ,UAAU,KAAK,OAAO,UAAU,GAAG;AAC7C,YAAM,KAAK,YAAY,SAAS,MAAM;AAItC,YAAM,WAAWA,SAAQ,KAAK,GAAG;AACjC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,MAAM,GAAG,QAAQ,CAAC,CAAC,oEACnB,MAAM,GAAG,QAAQ,CAAC,CAAC;AAAA,QACzB,QAAQ,EAAE,IAAI,SAAS,QAAQ,QAAQ,QAAQ,OAAO,OAAO;AAAA,MAC/D,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,cAAc,MAAM,qBAAqB,iBAAiB,MAAM,uBAAuB;AAC7F,UAAM,UAAU,KACb,IAAI,CAAC,OAAO,EAAE,GAAG,QAAQ,CAAC,GAAG,GAAG,YAAY,CAAC,EAAE,EAAE,EACjD;AAAA,MACC,CAAC,MAAqC,OAAO,EAAE,MAAM,YAAY,OAAO,EAAE,MAAM;AAAA,IAClF;AACF,QAAI,QAAQ,UAAU,GAAG;AACvB,YAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,CAAC;AACjC,YAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,CAAC;AACjC,YAAM,IAAI,SAAS,IAAI,EAAE;AAGzB,YAAM,WAAWA,SAAQ,MAAM,KAAK,IAAI,GAAG,CAAC,CAAC;AAC7C,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,2DAAsD,EAAE,QAAQ,CAAC,CAAC,mCAClE,+CAA0C,EAAE,QAAQ,CAAC,CAAC;AAAA,QAC5D,QAAQ,EAAE,SAAS,GAAG,GAAG,QAAQ,OAAO;AAAA,MAC1C,CAAC;AAAA,IACH;AAAA,EACF;AAGA;AACE,UAAM,UAAU,gCAAgC,MAAM,MAAM,2BAA2B,CAAC,CAAC;AACzF,QAAI,QAAQ,UAAU,GAAG;AACvB,YAAM,YAAY,QAAQ,MAAM,GAAG,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;AACjE,YAAM,WAAW,QAAQ,MAAM,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;AAC7D,YAAM,WACJ,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,IAAI,CAAC,MAAM,EAAE,OAAO,KAAK,CAAC;AACvF,YAAM,aACJ,KAAK,MAAM,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,CAAC,IACzE,KAAK,OAAO,IAAI,OAAO,EAAE,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,CAAC;AAC5E,YAAM,WAAW,KAAK,IAAI,GAAG,aAAa,QAAQ;AAClD,YAAM,WAAWA,SAAQ,WAAW,CAAC;AACrC,eAAS,KAAK;AAAA,QACZ,QAAQ;AAAA,QACR;AAAA,QACA,SACE,YAAY,MACR,gBAAgB,WAAW,QAAQ,CAAC,CAAC,gCAAgC,SAAS,QAAQ,CAAC,CAAC,yDACxF,uDAAuD,WAAW,QAAQ,CAAC,CAAC,SAAS,SAAS,QAAQ,CAAC,CAAC;AAAA,QAC9G,QAAQ,EAAE,YAAY,UAAU,UAAU,GAAG,QAAQ,OAAO;AAAA,MAC9D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,SAAS,SAAS,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,QAAQ,GAAG,CAAC;AACnE,QAAM,UACJ,UAAU,MAAM,WAAW,UAAU,MAAM,YAAY;AACzD,QAAM,YAAY,SACf,OAAO,CAAC,MAAM,EAAE,YAAY,GAAG,EAC/B,IAAI,CAAC,MAAM,GAAG,EAAE,MAAM,cAAc,EAAE,SAAS,QAAQ,CAAC,CAAC,WAAM,EAAE,OAAO,EAAE;AAC7E,MAAI,UAAU,WAAW,EAAG,WAAU,KAAK,0CAA0C;AAErF,SAAO,EAAE,UAAU,SAAS,WAAW,EAAE;AAC3C;AAIA,SAAS,KAAK,IAAsB;AAClC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAASA,SAAQ,GAAmB;AAClC,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,YAAY,GAAa,GAAqB;AAErD,QAAM,UAAU,CAAC,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,UAAU,CAAC,GAAG,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,MAAM,CAAC,GAAG,oBAAI,IAAI,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvE,MAAI,MAAM;AACV,aAAW,KAAK,KAAK;AACnB,UAAM,KAAK,QAAQ,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,SAAS,QAAQ;AAC1D,UAAM,KAAK,QAAQ,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,SAAS,QAAQ;AAC1D,UAAM,KAAK,IAAI,KAAK,KAAK,IAAI,KAAK,EAAE,CAAC;AAAA,EACvC;AACA,SAAO;AACT;AAEA,SAAS,iBACP,gBACmC;AACnC,SAAO,CAAC,QAAmB;AACzB,UAAM,WAAW,gCAAgC,CAAC,GAAG,GAAG,kBAAkB,CAAC,CAAC;AAC5E,WAAO,SAAS,WAAW,IAAI,SAAS,CAAC,EAAG,OAAO,QAAQ;AAAA,EAC7D;AACF;","names":["clamp01"]}

package/dist/chunk-BOD4O7OF.js

@@ -1,40 +0,0 @@
-// src/series-convergence.ts
-function analyzeSeries(values, options = {}) {
-  const window = options.window ?? 5;
-  const stableCv = options.stableCv ?? 0.05;
-  const driftRun = options.driftRun ?? 3;
-  if (values.length < Math.max(2, Math.min(window, 3))) {
-    return { state: "insufficient-data", windowMean: 0, windowCv: 0, tailRun: 0, stable: false };
-  }
-  const tail = values.slice(-window);
-  const mean = tail.reduce((a, b) => a + b, 0) / tail.length;
-  const variance = tail.reduce((acc, v) => acc + (v - mean) ** 2, 0) / tail.length;
-  const stdDev = Math.sqrt(variance);
-  const refMean = Math.abs(mean) > 1e-9 ? Math.abs(mean) : 1;
-  const cv = stdDev / refMean;
-  const stable = tail.length >= window && cv <= stableCv;
-  let tailRun = 0;
-  let direction = 0;
-  for (let i = values.length - 1; i > 0; i--) {
-    const delta = values[i] - values[i - 1];
-    if (delta === 0) break;
-    const dir = delta > 0 ? 1 : -1;
-    if (direction === 0) direction = dir;
-    if (dir !== direction) break;
-    tailRun += dir;
-  }
-  let state;
-  if (stable) {
-    state = "stabilized";
-  } else if (Math.abs(tailRun) >= driftRun) {
-    state = tailRun > 0 ? "drifting-up" : "drifting-down";
-  } else {
-    state = "noisy";
-  }
-  return { state, windowMean: mean, windowCv: cv, tailRun, stable };
-}
-
-export {
-  analyzeSeries
-};
-//# sourceMappingURL=chunk-BOD4O7OF.js.map