@tangle-network/agent-eval 0.120.1 → 0.120.2

package/dist/chunk-QWMPPZ3X.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/analyst/chat-client.ts","../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/default-registry.ts"],"sourcesContent":["/**\n * ChatClient — the single LLM abstraction analysts call.\n *\n * agent-eval already ships an `LlmClient` (OpenAI-compatible, retry,\n * graceful JSON-schema degrade) and judges that talk to `TCloud`. Two\n * mixed patterns force every analyst author to pick a transport, which\n * couples analyst code to runtime concerns (cli-bridge vs router vs\n * sandbox-sdk) it shouldn't know about.\n *\n * `ChatClient` is one interface every analyst takes via `AnalystContext.chat`.\n * The operator decides at the registry boundary which transport binds\n * to it. Analyst code stays transport-agnostic; swapping production\n * (sandbox-sdk) for local dev (cli-bridge) or tests (mock) is a one-\n * line factory call.\n *\n * Designed to coexist: existing `LlmClient` callers and existing\n * `TCloud`-based judges keep working untouched. New analyst code uses\n * `ChatClient`. When old call sites migrate, they pick up budgeting,\n * cancellation, and unified telemetry for free.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface. Mirrors LlmCallRequest/Result so the OpenAI-\n * compatible mental model stays. Two methods: a one-shot `chat()` and\n * an `streamChat()` for future agentic loops (not yet exposed).\n */\nexport interface ChatClient {\n  /** Display name of the bound transport — included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when caller omits — operators bind this per environment. */\n  readonly defaultModel?: string\n  /** Total provider attempts this transport can make for one chat call. */\n  readonly maximumAttempts?: number\n\n  /** Implementations must enforce `req.maxTokens` when it is present. */\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'router' // router.tangle.tools — production paid models\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'cli-bridge' // local cli-bridge for dev / local-only runs\n  | 'direct-provider' // direct OpenAI / Anthropic / etc. — bypass router\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n  /** Stable provider idempotency key for retries/redrives of one paid call. */\n  idempotencyKey?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | RouterTransportOpts\n  | CliBridgeTransportOpts\n  | DirectProviderTransportOpts\n  | SandboxSdkTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n  /** Total provider attempts. Required for opaque transports used in capped runs. */\n  maximumAttempts?: number\n}\n\nexport interface RouterTransportOpts extends BaseTransportOpts {\n  transport: 'router'\n  baseUrl?: string\n  apiKey: string\n}\n\nexport interface CliBridgeTransportOpts extends BaseTransportOpts {\n  transport: 'cli-bridge'\n  baseUrl?: string\n  bearer?: string\n}\n\nexport interface DirectProviderTransportOpts extends BaseTransportOpts {\n  transport: 'direct-provider'\n  baseUrl: string\n  apiKey: string\n}\n\n/**\n * Sandbox-SDK transport. Provided as a thin pass-through: the caller\n * supplies a callable that mimics LlmClient.chat() against an already-\n * configured Sandbox handle. We don't import the SDK here to keep\n * agent-eval dep-free of @tangle-network/sandbox.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'router':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'https://router.tangle.tools/v1',\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'cli-bridge':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'http://127.0.0.1:3344/v1',\n          apiKey: opts.bearer ?? '',\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'direct-provider':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl,\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: 1,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction wrapLlmClient(\n  transport: ChatTransport,\n  defaultModel: string | undefined,\n  inner: LlmClient,\n): ChatClient {\n  return {\n    transport,\n    defaultModel,\n    maximumAttempts: inner.maximumAttempts,\n    chat: (req, callOpts) => {\n      const resolved = resolveModel(req, defaultModel)\n      const request: LlmCallRequest = {\n        model: resolved.model!,\n        messages: req.messages,\n        jsonMode: req.jsonMode,\n        jsonSchema: req.jsonSchema,\n        temperature: req.temperature,\n        maxTokens: req.maxTokens,\n        timeoutMs: req.timeoutMs,\n      }\n      return inner.call(request, {\n        signal: callOpts?.signal,\n        idempotencyKey: callOpts?.idempotencyKey,\n      })\n    },\n  }\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n","/**\n * Deterministic behavioral metrics over OTLP spans — pure arithmetic, no LLM.\n *\n * These are the model-independent multiplier: the four trace-quality signals a\n * tolerant analyzer (e.g. HALO) re-derives per run inside the model — token\n * growth, output decay, tool monoculture, missing self-verification — computed\n * here once, in TypeScript, with zero model judgment. A finding that falls out\n * of arithmetic is trivially model-agnostic and cannot hallucinate the trend.\n *\n * General, not trace-specific: the detectors key off token trajectories and\n * tool usage present in any agentic OTLP trace, not any one benchmark.\n */\n\nimport { LLM_CONTEXT_TOKENS } from '../trace/attribute-vocabulary'\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport {\n  LLM_INPUT_TOKEN_ATTR_KEYS,\n  LLM_OUTPUT_TOKEN_ATTR_KEYS,\n  TOOL_NAME_ATTR_KEYS,\n} from '../trace/otlp-attributes'\nimport { spanEpochMillis } from './otlp-span'\nimport type { TraceAnalystSpan } from './types'\n\nexport type SuboptimalCode =\n  | 'monotonic-input-growth'\n  | 'output-length-decay'\n  | 'single-tool-dependency'\n  | 'no-self-verification'\n\nexport interface SuboptimalSignal {\n  code: SuboptimalCode\n  severity: 'high' | 'medium' | 'low'\n  /** Human-readable claim, with the backing numbers inlined. */\n  detail: string\n  /** The exact figures the detector fired on — auditable, no model in the loop. */\n  evidence: Record<string, number | string | boolean>\n}\n\nexport interface BehavioralMetrics {\n  /** The only trace represented by these metrics; null when spans are empty. */\n  traceId: string | null\n  llmCallCount: number\n  /** Causally serial LLM timelines. Parallel branches are never joined. */\n  tokenSequences: BehavioralTokenSequence[]\n  /** Token values from the longest serial timeline, retained for convenience. */\n  inputTokenTrajectory: number[]\n  outputTokenTrajectory: number[]\n  toolHistogram: Record<string, number>\n  totalToolCalls: number\n  distinctTools: number\n  /** distinct/total tool calls; 1.0 when there are no tool calls. */\n  toolDiversityRatio: number\n  hasSelfVerification: boolean\n  signals: SuboptimalSignal[]\n}\n\nexport interface BehavioralTokenSequence {\n  scopeId: string\n  spanIds: string[]\n  inputTokenTrajectory: Array<number | null>\n  outputTokenTrajectory: Array<number | null>\n}\n\ninterface TokenSample {\n  span: TraceAnalystSpan\n  input: number | null\n  output: number | null\n  step: number | null\n}\n\n/** ≥ this input-token growth ratio across a run, with no compression, fires. */\nconst INPUT_GROWTH_FACTOR = 3\n/** Tool-usage signals need at least this many calls to be meaningful. */\nconst MIN_TOOL_CALLS = 3\n/** Tool names that read or check state count as self-verification, not mutation.\n *  Covers the inspect verbs plus the read/search tools real harnesses use to\n *  verify (Claude Code Read/Grep/Glob, codex read_file/ls/cat, git status/diff,\n *  test/lint). A pure shell tool (Bash/exec_command) is intentionally NOT matched\n *  — its name can't tell a `pytest` from an `rm`. */\nconst VERIFY_RE =\n  /verif|eval|inspect|check|assert|validat|review|confirm|read|grep|glob|search|view|\\blist\\b|\\bls\\b|\\bcat\\b|\\bfind\\b|diff|status|\\btest|lint|typecheck/i\n\nfunction num(v: unknown): number | null {\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\nfunction numAttr(attrs: Record<string, unknown>, keys: readonly string[]): number | null {\n  for (const key of keys) {\n    const value = num(attrs[key])\n    if (value !== null) return value\n  }\n  return null\n}\nfunction inputTokensOf(s: TraceAnalystSpan): number | null {\n  const exactContext = num(s.attributes[LLM_CONTEXT_TOKENS])\n  if (exactContext !== null) return exactContext\n  // Foreign producers disagree on whether prompt includes cached tokens.\n  // Without an exact total, retain the reported prompt instead of guessing.\n  return (\n    numAttr(s.attributes, LLM_INPUT_TOKEN_ATTR_KEYS) ?? num(s.attributes['llm.usage.input_tokens'])\n  )\n}\nfunction outputTokensOf(s: TraceAnalystSpan): number | null {\n  return (\n    numAttr(s.attributes, LLM_OUTPUT_TOKEN_ATTR_KEYS) ??\n    num(s.attributes['llm.usage.output_tokens'])\n  )\n}\nfunction stepOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes.step)\n}\nfunction toolNameOf(s: TraceAnalystSpan): string | null {\n  if (s.tool_name) return s.tool_name\n  for (const key of TOOL_NAME_ATTR_KEYS) {\n    const t = s.attributes[key]\n    if (typeof t === 'string' && t.length > 0) return t\n  }\n  return null\n}\n\n/**\n * Reduce a span list to behavioral metrics + fired suboptimality signals.\n * Pure + deterministic: same spans → same output, on any machine, no model.\n */\nexport function computeTraceMetrics(spans: readonly TraceAnalystSpan[]): BehavioralMetrics {\n  const traceIds = new Set(spans.map((span) => span.trace_id))\n  if (traceIds.size > 1) {\n    throw new Error(\n      `computeTraceMetrics: expected spans from one trace, received ${traceIds.size} traces`,\n    )\n  }\n  const traceId = traceIds.values().next().value ?? null\n\n  const samples: TokenSample[] = spans.map((span) => ({\n    span,\n    input: inputTokensOf(span),\n    output: outputTokensOf(span),\n    step: stepOf(span),\n  }))\n  const llmSamples = samples.filter((sample) => sample.span.kind === 'LLM')\n  const tokenSamples =\n    llmSamples.length > 0\n      ? llmSamples\n      : samples.filter((sample) => sample.input !== null || sample.output !== null)\n  const tokenSequences = buildTokenSequences(tokenSamples, spans)\n  const primarySequence = tokenSequences[0]\n  const inputTokenTrajectory =\n    primarySequence?.inputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const outputTokenTrajectory =\n    primarySequence?.outputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const toolHistogram: Record<string, number> = {}\n  let hasSelfVerification = false\n\n  for (const s of spans) {\n    const tool = toolNameOf(s)\n    if (tool) {\n      toolHistogram[tool] = (toolHistogram[tool] ?? 0) + 1\n      if (VERIFY_RE.test(tool)) hasSelfVerification = true\n    }\n  }\n\n  const totalToolCalls = Object.values(toolHistogram).reduce((a, b) => a + b, 0)\n  const distinctTools = Object.keys(toolHistogram).length\n  const toolDiversityRatio = totalToolCalls === 0 ? 1 : distinctTools / totalToolCalls\n\n  const signals: SuboptimalSignal[] = []\n  const seenTokenSignals = new Set<SuboptimalCode>()\n  for (const sequence of tokenSequences) {\n    for (const signal of tokenSignals(sequence)) {\n      if (seenTokenSignals.has(signal.code)) continue\n      seenTokenSignals.add(signal.code)\n      signals.push(signal)\n    }\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && distinctTools === 1) {\n    const only = Object.keys(toolHistogram)[0]!\n    signals.push({\n      code: 'single-tool-dependency',\n      severity: 'medium',\n      detail: `All ${totalToolCalls} observed tool calls are \\`${only}\\`; no alternate tool call was observed.`,\n      evidence: { tool: only, calls: totalToolCalls, distinct_tools: 1 },\n    })\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && !hasSelfVerification) {\n    signals.push({\n      code: 'no-self-verification',\n      severity: 'medium',\n      detail: `${totalToolCalls} tool calls were observed without a verification-named tool call.`,\n      evidence: { tool_calls: totalToolCalls, verification_calls: 0 },\n    })\n  }\n\n  return {\n    traceId,\n    llmCallCount: tokenSamples.length,\n    tokenSequences,\n    inputTokenTrajectory,\n    outputTokenTrajectory,\n    toolHistogram,\n    totalToolCalls,\n    distinctTools,\n    toolDiversityRatio,\n    hasSelfVerification,\n    signals,\n  }\n}\n\nfunction buildTokenSequences(\n  samples: readonly TokenSample[],\n  spans: readonly TraceAnalystSpan[],\n): BehavioralTokenSequence[] {\n  const spansById = new Map(spans.map((span) => [span.span_id, span]))\n  const executionScopeFor = createTokenExecutionScopeResolver(spansById)\n  const scopedSamples = samples.map((sample) => ({\n    sample,\n    ...executionScopeFor(sample.span),\n  }))\n  const trackByLane = executionTrackByLane(scopedSamples)\n  const byTrack = new Map<string, { scopeId: string; samples: TokenSample[] }>()\n  for (const scoped of scopedSamples) {\n    const trackId = trackByLane.get(scoped.key)!\n    const track = byTrack.get(trackId) ?? { scopeId: scoped.scopeId, samples: [] }\n    track.samples.push(scoped.sample)\n    byTrack.set(trackId, track)\n  }\n\n  const sequences: BehavioralTokenSequence[] = []\n  for (const { scopeId, samples: tracked } of byTrack.values()) {\n    const runs = serialTokenRuns([...tracked].sort(compareTokenSamples))\n    runs.forEach((run, index) => {\n      sequences.push({\n        scopeId: runs.length === 1 ? scopeId : `${scopeId}#${index + 1}`,\n        spanIds: run.map((sample) => sample.span.span_id),\n        inputTokenTrajectory: run.map((sample) => sample.input),\n        outputTokenTrajectory: run.map((sample) => sample.output),\n      })\n    })\n  }\n\n  return sequences.sort(\n    (a, b) =>\n      b.spanIds.length - a.spanIds.length ||\n      a.scopeId.localeCompare(b.scopeId) ||\n      a.spanIds[0]!.localeCompare(b.spanIds[0]!),\n  )\n}\n\ninterface AncestorScope {\n  agentId: string | null\n  rootId: string | null\n  missingParentId: string | null\n  laneSpanId: string | null\n}\n\nfunction createTokenExecutionScopeResolver(spansById: ReadonlyMap<string, TraceAnalystSpan>): (\n  span: TraceAnalystSpan,\n) => {\n  key: string\n  scopeKey: string\n  scopeId: string\n  start: number | null\n  end: number | null\n} {\n  const cache = new Map<string, AncestorScope>()\n  return (span) => {\n    const ancestry = resolveAncestorScope(span.parent_span_id, spansById, cache)\n    const rootId = ancestry.rootId\n    const scopeId = ancestry.agentId\n      ? `span:${ancestry.agentId}`\n      : ancestry.missingParentId\n        ? `parent:${ancestry.missingParentId}`\n        : rootId\n          ? `root:${rootId}`\n          : span.agent_name\n            ? `agent:${span.agent_name}`\n            : `trace:${span.trace_id}`\n    const scopeSpanId = ancestry.agentId ?? ancestry.missingParentId ?? rootId\n    const laneSpan = ancestry.laneSpanId ? spansById.get(ancestry.laneSpanId) : undefined\n    const direct = scopeSpanId === null || ancestry.laneSpanId === scopeSpanId\n    const timedSpan = direct ? span : laneSpan\n    return {\n      key: JSON.stringify([scopeId, direct ? span.span_id : ancestry.laneSpanId]),\n      scopeKey: scopeId,\n      scopeId,\n      start: timedSpan ? spanEpochMillis(timedSpan.start_time) : null,\n      end: timedSpan ? spanEpochMillis(timedSpan.end_time) : null,\n    }\n  }\n}\n\nfunction resolveAncestorScope(\n  startId: string | null,\n  spansById: ReadonlyMap<string, TraceAnalystSpan>,\n  cache: Map<string, AncestorScope>,\n): AncestorScope {\n  const empty: AncestorScope = {\n    agentId: null,\n    rootId: null,\n    missingParentId: null,\n    laneSpanId: null,\n  }\n  if (!startId) return empty\n\n  const path: string[] = []\n  const pathIndex = new Map<string, number>()\n  let currentId: string | null = startId\n  let resolved = empty\n  while (currentId) {\n    const cached = cache.get(currentId)\n    if (cached) {\n      resolved = cached\n      break\n    }\n    const cycleStart = pathIndex.get(currentId)\n    if (cycleStart !== undefined) {\n      resolved = {\n        agentId: null,\n        rootId: [...path.slice(cycleStart)].sort()[0]!,\n        missingParentId: null,\n        laneSpanId: [...path.slice(cycleStart)].sort()[0]!,\n      }\n      break\n    }\n    const current = spansById.get(currentId)\n    if (!current) {\n      resolved = {\n        agentId: null,\n        rootId: null,\n        missingParentId: currentId,\n        laneSpanId: currentId,\n      }\n      break\n    }\n    if (current.kind === 'AGENT') {\n      resolved = {\n        agentId: current.span_id,\n        rootId: null,\n        missingParentId: null,\n        laneSpanId: current.span_id,\n      }\n      break\n    }\n    pathIndex.set(currentId, path.length)\n    path.push(currentId)\n    currentId = current.parent_span_id\n  }\n\n  for (let index = path.length - 1; index >= 0; index -= 1) {\n    if (resolved.agentId === null && resolved.rootId === null) {\n      resolved = { ...resolved, rootId: path[index]!, laneSpanId: path[index]! }\n    } else if (\n      resolved.laneSpanId === (resolved.agentId ?? resolved.missingParentId ?? resolved.rootId)\n    ) {\n      resolved = { ...resolved, laneSpanId: path[index]! }\n    }\n    cache.set(path[index]!, resolved)\n  }\n  return resolved\n}\n\nfunction compareTokenSamples(a: TokenSample, b: TokenSample): number {\n  const aStart = spanEpochMillis(a.span.start_time)\n  const bStart = spanEpochMillis(b.span.start_time)\n  if (aStart === null && bStart !== null) return 1\n  if (aStart !== null && bStart === null) return -1\n  if (aStart !== null && bStart !== null && aStart !== bStart) return aStart - bStart\n  if (a.step !== null && b.step !== null && a.step !== b.step) return a.step - b.step\n  return a.span.span_id.localeCompare(b.span.span_id)\n}\n\nfunction serialTokenRuns(ordered: readonly TokenSample[]): TokenSample[][] {\n  const runs: TokenSample[][] = []\n  let serial: TokenSample[] = []\n  let overlap: TokenSample[] = []\n  let overlapEnd = Number.NEGATIVE_INFINITY\n\n  const flushSerial = () => {\n    if (serial.length > 0) runs.push(serial)\n    serial = []\n  }\n  const flushOverlap = () => {\n    if (overlap.length === 1) {\n      serial.push(overlap[0]!)\n    } else if (overlap.length > 1) {\n      flushSerial()\n      for (const sample of overlap) runs.push([sample])\n    }\n    overlap = []\n    overlapEnd = Number.NEGATIVE_INFINITY\n  }\n\n  for (const sample of ordered) {\n    const start = spanEpochMillis(sample.span.start_time)\n    const end = spanEpochMillis(sample.span.end_time)\n    if (start === null || end === null || sample.span.duration_ms <= 0 || end < start) {\n      flushOverlap()\n      flushSerial()\n      runs.push([sample])\n      continue\n    }\n    if (overlap.length > 0 && start >= overlapEnd) flushOverlap()\n    overlap.push(sample)\n    overlapEnd = Math.max(overlapEnd, end)\n  }\n  flushOverlap()\n  flushSerial()\n  return runs\n}\n\nfunction tokenSignals(sequence: BehavioralTokenSequence): SuboptimalSignal[] {\n  const signals: SuboptimalSignal[] = []\n  const inputs = sequence.inputTokenTrajectory\n  const outputs = sequence.outputTokenTrajectory\n  if (inputs.length >= 3 && inputs.every((value): value is number => value !== null)) {\n    const first = inputs[0]!\n    const last = inputs[inputs.length - 1]!\n    const isMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const growthFromZero = first === 0 && last > 0\n    const growth = growthFromZero ? Infinity : first > 0 ? last / first : 0\n    if (isMonotonic && last > first && growth >= INPUT_GROWTH_FACTOR) {\n      const growthLabel = growthFromZero ? '0→nonzero (unbounded)' : `${growth.toFixed(1)}x`\n      signals.push({\n        code: 'monotonic-input-growth',\n        severity: 'high',\n        detail: `LLM input tokens grew ${growthLabel} (${first}→${last}) across ${inputs.length} serial calls without an intervening decrease.`,\n        evidence: {\n          first,\n          last,\n          growth_x: growthFromZero ? 'unbounded' : Number(growth.toFixed(2)),\n          calls: inputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n\n  if (\n    inputs.length >= 3 &&\n    inputs.length === outputs.length &&\n    inputs.every((value): value is number => value !== null) &&\n    outputs.every((value): value is number => value !== null)\n  ) {\n    const first = outputs[0]!\n    const last = outputs[outputs.length - 1]!\n    const inputIsMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const outputIsMonotonic = everyAdjacent(outputs, (previous, current) => current <= previous)\n    const inputGrew = inputs[inputs.length - 1]! > inputs[0]!\n    if (inputIsMonotonic && inputGrew && outputIsMonotonic && last < first) {\n      signals.push({\n        code: 'output-length-decay',\n        severity: 'medium',\n        detail: `LLM output tokens shrank ${first}→${last} over ${outputs.length} serial calls while input tokens increased monotonically.`,\n        evidence: {\n          first,\n          last,\n          calls: outputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n  return signals\n}\n\nfunction everyAdjacent(\n  values: readonly number[],\n  predicate: (previous: number, current: number) => boolean,\n): boolean {\n  return values.slice(1).every((current, index) => predicate(values[index]!, current))\n}\n","/**\n * `behavioralAnalyst` — a DETERMINISTIC analyst (cost.kind = 'deterministic',\n * never calls the LLM). It produces the efficiency/behavioral findings a\n * tolerant agentic analyzer (HALO) re-derives per run inside the model —\n * context bloat, output decay, tool monoculture, missing self-verification —\n * directly from arithmetic over spans (`computeTraceMetrics`).\n *\n * Why it matters: these findings are model-agnostic BY CONSTRUCTION (no model\n * in the loop), so they cannot return 0 on a weak model the way the Ax-RLM\n * does — and they are strictly more reliable than HALO, which spends tokens\n * re-deriving the same numbers and can hallucinate the trend. The agentic\n * RLM kinds remain for SEMANTIC findings that genuinely need a model; this\n * analyst owns the behavioral class.\n */\n\nimport {\n  type BehavioralMetrics,\n  computeTraceMetrics,\n  type SuboptimalCode,\n} from '../trace-analyst/behavioral-metrics'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport { type Analyst, type AnalystFinding, makeFinding } from './types'\n\nconst RECOMMENDED_ACTION: Record<SuboptimalCode, string> = {\n  'monotonic-input-growth':\n    'Inspect context assembly; if prior history is repeatedly included, summarize completed work before the next model call.',\n  'output-length-decay':\n    'Check late-step completeness; if shorter responses omit required work, add explicit completion criteria to the agent instructions.',\n  'single-tool-dependency':\n    'Test whether an inspect or verification tool improves outcomes after the repeated call fails or returns no progress.',\n  'no-self-verification':\n    'After state-changing actions, require an observable check before the agent proceeds.',\n}\n\nconst ANALYST_ID = 'efficiency-behavioral'\n\nconst AGGREGATE_CLAIM: Record<SuboptimalCode, (observed: number, analyzed: number) => string> = {\n  'monotonic-input-growth': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed input tokens grow from zero to nonzero or to at least 3x their initial value across at least 3 serial model calls without a decrease.`,\n  'output-length-decay': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed output tokens decrease while input tokens increased monotonically across at least 3 serial model calls.`,\n  'single-tool-dependency': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces used only one named tool across at least 3 tool calls.`,\n  'no-self-verification': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces had at least 3 tool calls without a verification-named tool call.`,\n}\n\n/**\n * Map computed signals → structured AnalystFindings. Pure: no LLM, no clock\n * dependence beyond `produced_at` (overridable for deterministic tests).\n */\nexport function deriveEfficiencyFindings(\n  metrics: BehavioralMetrics,\n  opts: { analystId?: string; producedAt?: string } = {},\n): AnalystFinding[] {\n  const analystId = opts.analystId ?? ANALYST_ID\n  const traceId = metrics.traceId\n  return metrics.signals.map((sig) =>\n    makeFinding({\n      analyst_id: analystId,\n      area: 'efficiency',\n      subject: sig.code,\n      claim: sig.detail,\n      severity: sig.severity,\n      // Deterministic arithmetic over spans, not a model judgment → certain.\n      confidence: 1,\n      evidence_refs: [\n        {\n          kind: 'metric',\n          uri: traceId\n            ? `metric://trace/${encodeURIComponent(traceId)}/efficiency/${sig.code}`\n            : `metric://efficiency/${sig.code}`,\n          excerpt: JSON.stringify(sig.evidence),\n        },\n      ],\n      recommended_action: RECOMMENDED_ACTION[sig.code],\n      metadata: {\n        deterministic: true,\n        evidence: sig.evidence,\n        ...(traceId ? { trace_id: traceId } : {}),\n      },\n      id_basis: sig.code,\n      ...(opts.producedAt ? { produced_at: opts.producedAt } : {}),\n    }),\n  )\n}\n\n/** The deterministic behavioral/efficiency analyst (no LLM, any-model). */\nexport function behavioralAnalyst(): Analyst<TraceAnalysisStore> {\n  return {\n    id: ANALYST_ID,\n    description:\n      'Deterministic behavioral/efficiency findings over OTLP spans — token-growth, output-decay, tool-monoculture, missing self-verification. Zero LLM; model-agnostic by construction.',\n    inputKind: 'trace-store',\n    cost: { kind: 'deterministic' },\n    version: '2.0.0',\n    async analyze(store) {\n      const overview = await store.getOverview()\n      const analyzedTraceIds = [...new Set(overview.sample_trace_ids)].sort()\n      const findingsById = new Map<\n        string,\n        { finding: AnalystFinding; traceIds: string[]; evidence: AnalystFinding['evidence_refs'] }\n      >()\n      for (const traceId of analyzedTraceIds) {\n        const viewed = await store.viewTrace({ trace_id: traceId })\n        if (viewed.trace_id !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${viewed.trace_id}'`,\n          )\n        }\n        if (!viewed.spans) {\n          throw new Error(\n            `behavioralAnalyst: trace '${traceId}' is oversized; complete spans are required`,\n          )\n        }\n        const metrics = computeTraceMetrics(viewed.spans)\n        if (metrics.traceId !== null && metrics.traceId !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${metrics.traceId}'`,\n          )\n        }\n        for (const finding of deriveEfficiencyFindings(metrics)) {\n          const current = findingsById.get(finding.finding_id)\n          if (!current) {\n            findingsById.set(finding.finding_id, {\n              finding,\n              traceIds: [traceId],\n              evidence: [...finding.evidence_refs],\n            })\n            continue\n          }\n          current.traceIds.push(traceId)\n          current.evidence.push(...finding.evidence_refs)\n        }\n      }\n      return [...findingsById.values()].map(({ finding, traceIds, evidence }) => ({\n        ...finding,\n        claim: AGGREGATE_CLAIM[finding.subject as SuboptimalCode](\n          traceIds.length,\n          analyzedTraceIds.length,\n        ),\n        rationale: `${traceIds.length}/${analyzedTraceIds.length} analyzed traces exhibited this pattern.`,\n        evidence_refs: evidence,\n        metadata: {\n          deterministic: true,\n          trace_ids: traceIds,\n          observed_trace_count: traceIds.length,\n          analyzed_trace_count: analyzedTraceIds.length,\n        },\n      }))\n    },\n  }\n}\n","/**\n * `buildDefaultAnalystRegistry` — the canonical analyst suite, so consumers\n * stop hand-wiring `new AnalystRegistry()` + per-kind `createTraceAnalystKind`.\n *\n * The deterministic `behavioralAnalyst` is ALWAYS registered (it needs no\n * model and is model-agnostic by construction). The agentic RLM kinds are\n * registered only when an `ai` service is supplied — so a caller with no LLM\n * still gets the full behavioral/efficiency diagnosis, and the substrate's\n * \"any model (including no model)\" guarantee holds at the suite level.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport { behavioralAnalyst } from './behavioral-analyst'\nimport { createTraceAnalystKind, type TraceAnalystKindSpec } from './kind-factory'\nimport { DEFAULT_TRACE_ANALYST_KINDS } from './kinds'\nimport { AnalystRegistry, type AnalystRegistryOptions } from './registry'\n\nexport interface DefaultAnalystRegistryOptions {\n  /** Ax service for the agentic RLM kinds. Omit → only the deterministic analyst. */\n  ai?: AxAIService\n  /** Required unless `ai` was created by `createAnalystAi`. */\n  model?: string\n  /** Which agentic kinds to register when `ai` is present. Default = the shipped suite. */\n  kinds?: readonly TraceAnalystKindSpec[]\n  /** Set false to omit the deterministic behavioral analyst (default: include). */\n  includeBehavioral?: boolean\n  /** Forwarded to the AnalystRegistry constructor (signal, tags, priorFindings). */\n  registry?: AnalystRegistryOptions\n}\n\nexport function buildDefaultAnalystRegistry(\n  opts: DefaultAnalystRegistryOptions = {},\n): AnalystRegistry {\n  const registry = new AnalystRegistry(opts.registry)\n  if (opts.includeBehavioral !== false) {\n    registry.register(behavioralAnalyst())\n  }\n  if (opts.ai) {\n    const kinds = opts.kinds ?? DEFAULT_TRACE_ANALYST_KINDS\n    for (const spec of kinds) {\n      registry.register(createTraceAnalystKind(spec, { ai: opts.ai, model: opts.model }))\n    }\n  }\n  return registry\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AA+HO,SAAS,iBAAiB,MAAwC;AACvE,UAAQ,KAAK,WAAW;AAAA,IACtB,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK,UAAU;AAAA,UACvB,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK;AAAA,UACd,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB,KAAK;AAAA,QACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MACzF;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB;AAAA,QACjB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MAC5F;AAAA,EACJ;AACF;AAEA,SAAS,cACP,WACA,cACA,OACY;AACZ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,iBAAiB,MAAM;AAAA,IACvB,MAAM,CAAC,KAAK,aAAa;AACvB,YAAM,WAAW,aAAa,KAAK,YAAY;AAC/C,YAAM,UAA0B;AAAA,QAC9B,OAAO,SAAS;AAAA,QAChB,UAAU,IAAI;AAAA,QACd,UAAU,IAAI;AAAA,QACd,YAAY,IAAI;AAAA,QAChB,aAAa,IAAI;AAAA,QACjB,WAAW,IAAI;AAAA,QACf,WAAW,IAAI;AAAA,MACjB;AACA,aAAO,MAAM,KAAK,SAAS;AAAA,QACzB,QAAQ,UAAU;AAAA,QAClB,gBAAgB,UAAU;AAAA,MAC5B,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,SAAS,aAAa,KAAkB,cAA+C;AACrF,MAAI,IAAI,MAAO,QAAO;AACtB,MAAI,CAAC,cAAc;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,IAEF;AAAA,EACF;AACA,SAAO,EAAE,GAAG,KAAK,OAAO,aAAa;AACvC;;;AC9IA,IAAM,sBAAsB;AAE5B,IAAM,iBAAiB;AAMvB,IAAM,YACJ;AAEF,SAAS,IAAI,GAA2B;AACtC,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AACA,SAAS,QAAQ,OAAgC,MAAwC;AACvF,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,MAAM,GAAG,CAAC;AAC5B,QAAI,UAAU,KAAM,QAAO;AAAA,EAC7B;AACA,SAAO;AACT;AACA,SAAS,cAAc,GAAoC;AACzD,QAAM,eAAe,IAAI,EAAE,WAAW,kBAAkB,CAAC;AACzD,MAAI,iBAAiB,KAAM,QAAO;AAGlC,SACE,QAAQ,EAAE,YAAY,yBAAyB,KAAK,IAAI,EAAE,WAAW,wBAAwB,CAAC;AAElG;AACA,SAAS,eAAe,GAAoC;AAC1D,SACE,QAAQ,EAAE,YAAY,0BAA0B,KAChD,IAAI,EAAE,WAAW,yBAAyB,CAAC;AAE/C;AACA,SAAS,OAAO,GAAoC;AAClD,SAAO,IAAI,EAAE,WAAW,IAAI;AAC9B;AACA,SAAS,WAAW,GAAoC;AACtD,MAAI,EAAE,UAAW,QAAO,EAAE;AAC1B,aAAW,OAAO,qBAAqB;AACrC,UAAM,IAAI,EAAE,WAAW,GAAG;AAC1B,QAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,EACpD;AACA,SAAO;AACT;AAMO,SAAS,oBAAoB,OAAuD;AACzF,QAAM,WAAW,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,KAAK,QAAQ,CAAC;AAC3D,MAAI,SAAS,OAAO,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,gEAAgE,SAAS,IAAI;AAAA,IAC/E;AAAA,EACF;AACA,QAAM,UAAU,SAAS,OAAO,EAAE,KAAK,EAAE,SAAS;AAElD,QAAM,UAAyB,MAAM,IAAI,CAAC,UAAU;AAAA,IAClD;AAAA,IACA,OAAO,cAAc,IAAI;AAAA,IACzB,QAAQ,eAAe,IAAI;AAAA,IAC3B,MAAM,OAAO,IAAI;AAAA,EACnB,EAAE;AACF,QAAM,aAAa,QAAQ,OAAO,CAAC,WAAW,OAAO,KAAK,SAAS,KAAK;AACxE,QAAM,eACJ,WAAW,SAAS,IAChB,aACA,QAAQ,OAAO,CAAC,WAAW,OAAO,UAAU,QAAQ,OAAO,WAAW,IAAI;AAChF,QAAM,iBAAiB,oBAAoB,cAAc,KAAK;AAC9D,QAAM,kBAAkB,eAAe,CAAC;AACxC,QAAM,uBACJ,iBAAiB,qBAAqB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAC/F,QAAM,wBACJ,iBAAiB,sBAAsB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAChG,QAAM,gBAAwC,CAAC;AAC/C,MAAI,sBAAsB;AAE1B,aAAW,KAAK,OAAO;AACrB,UAAM,OAAO,WAAW,CAAC;AACzB,QAAI,MAAM;AACR,oBAAc,IAAI,KAAK,cAAc,IAAI,KAAK,KAAK;AACnD,UAAI,UAAU,KAAK,IAAI,EAAG,uBAAsB;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,iBAAiB,OAAO,OAAO,aAAa,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC7E,QAAM,gBAAgB,OAAO,KAAK,aAAa,EAAE;AACjD,QAAM,qBAAqB,mBAAmB,IAAI,IAAI,gBAAgB;AAEtE,QAAM,UAA8B,CAAC;AACrC,QAAM,mBAAmB,oBAAI,IAAoB;AACjD,aAAW,YAAY,gBAAgB;AACrC,eAAW,UAAU,aAAa,QAAQ,GAAG;AAC3C,UAAI,iBAAiB,IAAI,OAAO,IAAI,EAAG;AACvC,uBAAiB,IAAI,OAAO,IAAI;AAChC,cAAQ,KAAK,MAAM;AAAA,IACrB;AAAA,EACF;AAEA,MAAI,kBAAkB,kBAAkB,kBAAkB,GAAG;AAC3D,UAAM,OAAO,OAAO,KAAK,aAAa,EAAE,CAAC;AACzC,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,OAAO,cAAc,8BAA8B,IAAI;AAAA,MAC/D,UAAU,EAAE,MAAM,MAAM,OAAO,gBAAgB,gBAAgB,EAAE;AAAA,IACnE,CAAC;AAAA,EACH;AAEA,MAAI,kBAAkB,kBAAkB,CAAC,qBAAqB;AAC5D,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,GAAG,cAAc;AAAA,MACzB,UAAU,EAAE,YAAY,gBAAgB,oBAAoB,EAAE;AAAA,IAChE,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL;AAAA,IACA,cAAc,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBACP,SACA,OAC2B;AAC3B,QAAM,YAAY,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,KAAK,SAAS,IAAI,CAAC,CAAC;AACnE,QAAM,oBAAoB,kCAAkC,SAAS;AACrE,QAAM,gBAAgB,QAAQ,IAAI,CAAC,YAAY;AAAA,IAC7C;AAAA,IACA,GAAG,kBAAkB,OAAO,IAAI;AAAA,EAClC,EAAE;AACF,QAAM,cAAc,qBAAqB,aAAa;AACtD,QAAM,UAAU,oBAAI,IAAyD;AAC7E,aAAW,UAAU,eAAe;AAClC,UAAM,UAAU,YAAY,IAAI,OAAO,GAAG;AAC1C,UAAM,QAAQ,QAAQ,IAAI,OAAO,KAAK,EAAE,SAAS,OAAO,SAAS,SAAS,CAAC,EAAE;AAC7E,UAAM,QAAQ,KAAK,OAAO,MAAM;AAChC,YAAQ,IAAI,SAAS,KAAK;AAAA,EAC5B;AAEA,QAAM,YAAuC,CAAC;AAC9C,aAAW,EAAE,SAAS,SAAS,QAAQ,KAAK,QAAQ,OAAO,GAAG;AAC5D,UAAM,OAAO,gBAAgB,CAAC,GAAG,OAAO,EAAE,KAAK,mBAAmB,CAAC;AACnE,SAAK,QAAQ,CAAC,KAAK,UAAU;AAC3B,gBAAU,KAAK;AAAA,QACb,SAAS,KAAK,WAAW,IAAI,UAAU,GAAG,OAAO,IAAI,QAAQ,CAAC;AAAA,QAC9D,SAAS,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK,OAAO;AAAA,QAChD,sBAAsB,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK;AAAA,QACtD,uBAAuB,IAAI,IAAI,CAAC,WAAW,OAAO,MAAM;AAAA,MAC1D,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAEA,SAAO,UAAU;AAAA,IACf,CAAC,GAAG,MACF,EAAE,QAAQ,SAAS,EAAE,QAAQ,UAC7B,EAAE,QAAQ,cAAc,EAAE,OAAO,KACjC,EAAE,QAAQ,CAAC,EAAG,cAAc,EAAE,QAAQ,CAAC,CAAE;AAAA,EAC7C;AACF;AASA,SAAS,kCAAkC,WAQzC;AACA,QAAM,QAAQ,oBAAI,IAA2B;AAC7C,SAAO,CAAC,SAAS;AACf,UAAM,WAAW,qBAAqB,KAAK,gBAAgB,WAAW,KAAK;AAC3E,UAAM,SAAS,SAAS;AACxB,UAAM,UAAU,SAAS,UACrB,QAAQ,SAAS,OAAO,KACxB,SAAS,kBACP,UAAU,SAAS,eAAe,KAClC,SACE,QAAQ,MAAM,KACd,KAAK,aACH,SAAS,KAAK,UAAU,KACxB,SAAS,KAAK,QAAQ;AAChC,UAAM,cAAc,SAAS,WAAW,SAAS,mBAAmB;AACpE,UAAM,WAAW,SAAS,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI;AAC5E,UAAM,SAAS,gBAAgB,QAAQ,SAAS,eAAe;AAC/D,UAAM,YAAY,SAAS,OAAO;AAClC,WAAO;AAAA,MACL,KAAK,KAAK,UAAU,CAAC,SAAS,SAAS,KAAK,UAAU,SAAS,UAAU,CAAC;AAAA,MAC1E,UAAU;AAAA,MACV;AAAA,MACA,OAAO,YAAY,gBAAgB,UAAU,UAAU,IAAI;AAAA,MAC3D,KAAK,YAAY,gBAAgB,UAAU,QAAQ,IAAI;AAAA,IACzD;AAAA,EACF;AACF;AAEA,SAAS,qBACP,SACA,WACA,OACe;AACf,QAAM,QAAuB;AAAA,IAC3B,SAAS;AAAA,IACT,QAAQ;AAAA,IACR,iBAAiB;AAAA,IACjB,YAAY;AAAA,EACd;AACA,MAAI,CAAC,QAAS,QAAO;AAErB,QAAM,OAAiB,CAAC;AACxB,QAAM,YAAY,oBAAI,IAAoB;AAC1C,MAAI,YAA2B;AAC/B,MAAI,WAAW;AACf,SAAO,WAAW;AAChB,UAAM,SAAS,MAAM,IAAI,SAAS;AAClC,QAAI,QAAQ;AACV,iBAAW;AACX;AAAA,IACF;AACA,UAAM,aAAa,UAAU,IAAI,SAAS;AAC1C,QAAI,eAAe,QAAW;AAC5B,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,QAC5C,iBAAiB;AAAA,QACjB,YAAY,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,MAClD;AACA;AAAA,IACF;AACA,UAAM,UAAU,UAAU,IAAI,SAAS;AACvC,QAAI,CAAC,SAAS;AACZ,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY;AAAA,MACd;AACA;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,SAAS;AAC5B,iBAAW;AAAA,QACT,SAAS,QAAQ;AAAA,QACjB,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY,QAAQ;AAAA,MACtB;AACA;AAAA,IACF;AACA,cAAU,IAAI,WAAW,KAAK,MAAM;AACpC,SAAK,KAAK,SAAS;AACnB,gBAAY,QAAQ;AAAA,EACtB;AAEA,WAAS,QAAQ,KAAK,SAAS,GAAG,SAAS,GAAG,SAAS,GAAG;AACxD,QAAI,SAAS,YAAY,QAAQ,SAAS,WAAW,MAAM;AACzD,iBAAW,EAAE,GAAG,UAAU,QAAQ,KAAK,KAAK,GAAI,YAAY,KAAK,KAAK,EAAG;AAAA,IAC3E,WACE,SAAS,gBAAgB,SAAS,WAAW,SAAS,mBAAmB,SAAS,SAClF;AACA,iBAAW,EAAE,GAAG,UAAU,YAAY,KAAK,KAAK,EAAG;AAAA,IACrD;AACA,UAAM,IAAI,KAAK,KAAK,GAAI,QAAQ;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,GAAgB,GAAwB;AACnE,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,QAAQ,WAAW,OAAQ,QAAO,SAAS;AAC7E,MAAI,EAAE,SAAS,QAAQ,EAAE,SAAS,QAAQ,EAAE,SAAS,EAAE,KAAM,QAAO,EAAE,OAAO,EAAE;AAC/E,SAAO,EAAE,KAAK,QAAQ,cAAc,EAAE,KAAK,OAAO;AACpD;AAEA,SAAS,gBAAgB,SAAkD;AACzE,QAAM,OAAwB,CAAC;AAC/B,MAAI,SAAwB,CAAC;AAC7B,MAAI,UAAyB,CAAC;AAC9B,MAAI,aAAa,OAAO;AAExB,QAAM,cAAc,MAAM;AACxB,QAAI,OAAO,SAAS,EAAG,MAAK,KAAK,MAAM;AACvC,aAAS,CAAC;AAAA,EACZ;AACA,QAAM,eAAe,MAAM;AACzB,QAAI,QAAQ,WAAW,GAAG;AACxB,aAAO,KAAK,QAAQ,CAAC,CAAE;AAAA,IACzB,WAAW,QAAQ,SAAS,GAAG;AAC7B,kBAAY;AACZ,iBAAW,UAAU,QAAS,MAAK,KAAK,CAAC,MAAM,CAAC;AAAA,IAClD;AACA,cAAU,CAAC;AACX,iBAAa,OAAO;AAAA,EACtB;AAEA,aAAW,UAAU,SAAS;AAC5B,UAAM,QAAQ,gBAAgB,OAAO,KAAK,UAAU;AACpD,UAAM,MAAM,gBAAgB,OAAO,KAAK,QAAQ;AAChD,QAAI,UAAU,QAAQ,QAAQ,QAAQ,OAAO,KAAK,eAAe,KAAK,MAAM,OAAO;AACjF,mBAAa;AACb,kBAAY;AACZ,WAAK,KAAK,CAAC,MAAM,CAAC;AAClB;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,KAAK,SAAS,WAAY,cAAa;AAC5D,YAAQ,KAAK,MAAM;AACnB,iBAAa,KAAK,IAAI,YAAY,GAAG;AAAA,EACvC;AACA,eAAa;AACb,cAAY;AACZ,SAAO;AACT;AAEA,SAAS,aAAa,UAAuD;AAC3E,QAAM,UAA8B,CAAC;AACrC,QAAM,SAAS,SAAS;AACxB,QAAM,UAAU,SAAS;AACzB,MAAI,OAAO,UAAU,KAAK,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,GAAG;AAClF,UAAM,QAAQ,OAAO,CAAC;AACtB,UAAM,OAAO,OAAO,OAAO,SAAS,CAAC;AACrC,UAAM,cAAc,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACpF,UAAM,iBAAiB,UAAU,KAAK,OAAO;AAC7C,UAAM,SAAS,iBAAiB,WAAW,QAAQ,IAAI,OAAO,QAAQ;AACtE,QAAI,eAAe,OAAO,SAAS,UAAU,qBAAqB;AAChE,YAAM,cAAc,iBAAiB,+BAA0B,GAAG,OAAO,QAAQ,CAAC,CAAC;AACnF,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,yBAAyB,WAAW,KAAK,KAAK,SAAI,IAAI,YAAY,OAAO,MAAM;AAAA,QACvF,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,UAAU,iBAAiB,cAAc,OAAO,OAAO,QAAQ,CAAC,CAAC;AAAA,UACjE,OAAO,OAAO;AAAA,UACd,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MACE,OAAO,UAAU,KACjB,OAAO,WAAW,QAAQ,UAC1B,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,KACvD,QAAQ,MAAM,CAAC,UAA2B,UAAU,IAAI,GACxD;AACA,UAAM,QAAQ,QAAQ,CAAC;AACvB,UAAM,OAAO,QAAQ,QAAQ,SAAS,CAAC;AACvC,UAAM,mBAAmB,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACzF,UAAM,oBAAoB,cAAc,SAAS,CAAC,UAAU,YAAY,WAAW,QAAQ;AAC3F,UAAM,YAAY,OAAO,OAAO,SAAS,CAAC,IAAK,OAAO,CAAC;AACvD,QAAI,oBAAoB,aAAa,qBAAqB,OAAO,OAAO;AACtE,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,4BAA4B,KAAK,SAAI,IAAI,SAAS,QAAQ,MAAM;AAAA,QACxE,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,OAAO,QAAQ;AAAA,UACf,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,cACP,QACA,WACS;AACT,SAAO,OAAO,MAAM,CAAC,EAAE,MAAM,CAAC,SAAS,UAAU,UAAU,OAAO,KAAK,GAAI,OAAO,CAAC;AACrF;;;ACncA,IAAM,qBAAqD;AAAA,EACzD,0BACE;AAAA,EACF,uBACE;AAAA,EACF,0BACE;AAAA,EACF,wBACE;AACJ;AAEA,IAAM,aAAa;AAEnB,IAAM,kBAA0F;AAAA,EAC9F,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,uBAAuB,CAAC,UAAU,aAChC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,wBAAwB,CAAC,UAAU,aACjC,GAAG,QAAQ,IAAI,QAAQ;AAC3B;AAMO,SAAS,yBACd,SACA,OAAoD,CAAC,GACnC;AAClB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,UAAU,QAAQ;AACxB,SAAO,QAAQ,QAAQ;AAAA,IAAI,CAAC,QAC1B,YAAY;AAAA,MACV,YAAY;AAAA,MACZ,MAAM;AAAA,MACN,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,UAAU,IAAI;AAAA;AAAA,MAEd,YAAY;AAAA,MACZ,eAAe;AAAA,QACb;AAAA,UACE,MAAM;AAAA,UACN,KAAK,UACD,kBAAkB,mBAAmB,OAAO,CAAC,eAAe,IAAI,IAAI,KACpE,uBAAuB,IAAI,IAAI;AAAA,UACnC,SAAS,KAAK,UAAU,IAAI,QAAQ;AAAA,QACtC;AAAA,MACF;AAAA,MACA,oBAAoB,mBAAmB,IAAI,IAAI;AAAA,MAC/C,UAAU;AAAA,QACR,eAAe;AAAA,QACf,UAAU,IAAI;AAAA,QACd,GAAI,UAAU,EAAE,UAAU,QAAQ,IAAI,CAAC;AAAA,MACzC;AAAA,MACA,UAAU,IAAI;AAAA,MACd,GAAI,KAAK,aAAa,EAAE,aAAa,KAAK,WAAW,IAAI,CAAC;AAAA,IAC5D,CAAC;AAAA,EACH;AACF;AAGO,SAAS,oBAAiD;AAC/D,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS;AAAA,IACT,MAAM,QAAQ,OAAO;AACnB,YAAM,WAAW,MAAM,MAAM,YAAY;AACzC,YAAM,mBAAmB,CAAC,GAAG,IAAI,IAAI,SAAS,gBAAgB,CAAC,EAAE,KAAK;AACtE,YAAM,eAAe,oBAAI,IAGvB;AACF,iBAAW,WAAW,kBAAkB;AACtC,cAAM,SAAS,MAAM,MAAM,UAAU,EAAE,UAAU,QAAQ,CAAC;AAC1D,YAAI,OAAO,aAAa,SAAS;AAC/B,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,OAAO,QAAQ;AAAA,UAC/E;AAAA,QACF;AACA,YAAI,CAAC,OAAO,OAAO;AACjB,gBAAM,IAAI;AAAA,YACR,6BAA6B,OAAO;AAAA,UACtC;AAAA,QACF;AACA,cAAM,UAAU,oBAAoB,OAAO,KAAK;AAChD,YAAI,QAAQ,YAAY,QAAQ,QAAQ,YAAY,SAAS;AAC3D,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,QAAQ,OAAO;AAAA,UAC/E;AAAA,QACF;AACA,mBAAW,WAAW,yBAAyB,OAAO,GAAG;AACvD,gBAAM,UAAU,aAAa,IAAI,QAAQ,UAAU;AACnD,cAAI,CAAC,SAAS;AACZ,yBAAa,IAAI,QAAQ,YAAY;AAAA,cACnC;AAAA,cACA,UAAU,CAAC,OAAO;AAAA,cAClB,UAAU,CAAC,GAAG,QAAQ,aAAa;AAAA,YACrC,CAAC;AACD;AAAA,UACF;AACA,kBAAQ,SAAS,KAAK,OAAO;AAC7B,kBAAQ,SAAS,KAAK,GAAG,QAAQ,aAAa;AAAA,QAChD;AAAA,MACF;AACA,aAAO,CAAC,GAAG,aAAa,OAAO,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS,UAAU,SAAS,OAAO;AAAA,QAC1E,GAAG;AAAA,QACH,OAAO,gBAAgB,QAAQ,OAAyB;AAAA,UACtD,SAAS;AAAA,UACT,iBAAiB;AAAA,QACnB;AAAA,QACA,WAAW,GAAG,SAAS,MAAM,IAAI,iBAAiB,MAAM;AAAA,QACxD,eAAe;AAAA,QACf,UAAU;AAAA,UACR,eAAe;AAAA,UACf,WAAW;AAAA,UACX,sBAAsB,SAAS;AAAA,UAC/B,sBAAsB,iBAAiB;AAAA,QACzC;AAAA,MACF,EAAE;AAAA,IACJ;AAAA,EACF;AACF;;;AC1HO,SAAS,4BACd,OAAsC,CAAC,GACtB;AACjB,QAAM,WAAW,IAAI,gBAAgB,KAAK,QAAQ;AAClD,MAAI,KAAK,sBAAsB,OAAO;AACpC,aAAS,SAAS,kBAAkB,CAAC;AAAA,EACvC;AACA,MAAI,KAAK,IAAI;AACX,UAAM,QAAQ,KAAK,SAAS;AAC5B,eAAW,QAAQ,OAAO;AACxB,eAAS,SAAS,uBAAuB,MAAM,EAAE,IAAI,KAAK,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC;AAAA,IACpF;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/chunk-S3UZOQ5Y.js

@@ -1,328 +0,0 @@
-import {
-  FAILURE_CLASSES
-} from "./chunk-MA6HLL3S.js";
-import {
-  validateAgentProfileCell
-} from "./chunk-XJYR7XFV.js";
-import {
-  ValidationError
-} from "./chunk-ONWEPEDO.js";
-
-// src/run-record.ts
-var MANDATORY_TOP_LEVEL = [
-  "runId",
-  "experimentId",
-  "candidateId",
-  "seed",
-  "model",
-  "promptHash",
-  "configHash",
-  "commitSha",
-  "wallMs",
-  "costUsd",
-  "tokenUsage",
-  "outcome",
-  "splitTag"
-];
-var SPLIT_TAGS = ["search", "dev", "holdout"];
-var RunRecordValidationError = class extends ValidationError {
-  path;
-  constructor(message, path = "") {
-    super(path ? `${message} (at ${path})` : message);
-    this.path = path;
-  }
-};
-function validateRunRecord(input) {
-  if (input === null || typeof input !== "object") {
-    throw new RunRecordValidationError("expected object");
-  }
-  const obj = input;
-  for (const key of MANDATORY_TOP_LEVEL) {
-    if (!(key in obj)) {
-      throw new RunRecordValidationError(`missing mandatory field "${key}"`);
-    }
-  }
-  expectString(obj.runId, "runId");
-  expectString(obj.experimentId, "experimentId");
-  expectString(obj.candidateId, "candidateId");
-  expectFiniteNumber(obj.seed, "seed");
-  expectString(obj.model, "model");
-  expectString(obj.promptHash, "promptHash");
-  expectString(obj.configHash, "configHash");
-  expectString(obj.commitSha, "commitSha");
-  expectNonNegativeNumber(obj.wallMs, "wallMs");
-  if (obj.queueMs !== void 0) expectNonNegativeNumber(obj.queueMs, "queueMs");
-  expectNonNegativeNumber(obj.costUsd, "costUsd");
-  if (obj.costProvenance !== void 0) {
-    validateCostProvenance(obj.costProvenance, obj.costUsd);
-  }
-  if (!modelHasSnapshot(obj.model)) {
-    throw new RunRecordValidationError(
-      `model "${obj.model}" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,
-      "model"
-    );
-  }
-  const tu = obj.tokenUsage;
-  if (tu === null || typeof tu !== "object") {
-    throw new RunRecordValidationError("tokenUsage must be an object", "tokenUsage");
-  }
-  const tuRec = tu;
-  expectNonNegativeNumber(tuRec.input, "tokenUsage.input");
-  expectNonNegativeNumber(tuRec.output, "tokenUsage.output");
-  if (tuRec.reasoning !== void 0) {
-    expectNonNegativeNumber(tuRec.reasoning, "tokenUsage.reasoning");
-    if (tuRec.reasoning > tuRec.output) {
-      throw new RunRecordValidationError(
-        "reasoning tokens must be a subset of output tokens",
-        "tokenUsage.reasoning"
-      );
-    }
-  }
-  if (tuRec.cached !== void 0) expectNonNegativeNumber(tuRec.cached, "tokenUsage.cached");
-  if (tuRec.cacheWrite !== void 0) {
-    expectNonNegativeNumber(tuRec.cacheWrite, "tokenUsage.cacheWrite");
-  }
-  if (obj.judgeMetadata !== void 0) {
-    const jm = obj.judgeMetadata;
-    if (jm === null || typeof jm !== "object") {
-      throw new RunRecordValidationError("judgeMetadata must be an object", "judgeMetadata");
-    }
-    const jmRec = jm;
-    expectString(jmRec.model, "judgeMetadata.model");
-    expectString(jmRec.promptVersion, "judgeMetadata.promptVersion");
-    expectFiniteNumber(jmRec.confidence, "judgeMetadata.confidence");
-    if (typeof jmRec.fallback !== "boolean") {
-      throw new RunRecordValidationError(
-        "judgeMetadata.fallback must be boolean",
-        "judgeMetadata.fallback"
-      );
-    }
-  }
-  const out = obj.outcome;
-  if (out === null || typeof out !== "object") {
-    throw new RunRecordValidationError("outcome must be an object", "outcome");
-  }
-  const outRec = out;
-  if (outRec.searchScore !== void 0)
-    expectFiniteNumber(outRec.searchScore, "outcome.searchScore");
-  if (outRec.holdoutScore !== void 0)
-    expectFiniteNumber(outRec.holdoutScore, "outcome.holdoutScore");
-  if (outRec.searchScore === void 0 && outRec.holdoutScore === void 0) {
-    throw new RunRecordValidationError(
-      "outcome must define searchScore or holdoutScore (or both)",
-      "outcome"
-    );
-  }
-  const raw = outRec.raw;
-  if (raw === null || typeof raw !== "object") {
-    throw new RunRecordValidationError("outcome.raw must be an object", "outcome.raw");
-  }
-  for (const [k, v] of Object.entries(raw)) {
-    expectFiniteNumber(v, `outcome.raw.${k}`);
-  }
-  if (outRec.realness !== void 0) {
-    const r = outRec.realness;
-    if (r === null || typeof r !== "object") {
-      throw new RunRecordValidationError("outcome.realness must be an object", "outcome.realness");
-    }
-    const rr = r;
-    expectFiniteNumber(rr.score, "outcome.realness.score");
-    if (typeof rr.gated !== "boolean") {
-      throw new RunRecordValidationError(
-        "outcome.realness.gated must be a boolean",
-        "outcome.realness.gated"
-      );
-    }
-  }
-  if (outRec.judgeScores !== void 0) {
-    validateJudgeScores(outRec.judgeScores, "outcome.judgeScores");
-  }
-  if (obj.failureClass !== void 0 && (typeof obj.failureClass !== "string" || !FAILURE_CLASSES.includes(obj.failureClass))) {
-    throw new RunRecordValidationError(
-      `failureClass must be one of ${FAILURE_CLASSES.join(", ")}`,
-      "failureClass"
-    );
-  }
-  if (obj.failureMode !== void 0) expectString(obj.failureMode, "failureMode");
-  if (obj.agentProfile !== void 0) {
-    try {
-      const profile = validateAgentProfileCell(obj.agentProfile);
-      if (profile.model !== void 0 && profile.model !== obj.model) {
-        throw new RunRecordValidationError(
-          `agentProfile.model "${profile.model}" does not match model "${obj.model}"`,
-          "agentProfile.model"
-        );
-      }
-      if (profile.promptHash !== void 0 && profile.promptHash !== obj.promptHash) {
-        throw new RunRecordValidationError(
-          `agentProfile.promptHash "${profile.promptHash}" does not match promptHash "${obj.promptHash}"`,
-          "agentProfile.promptHash"
-        );
-      }
-    } catch (error) {
-      if (error instanceof RunRecordValidationError) throw error;
-      if (error instanceof Error) {
-        throw new RunRecordValidationError(error.message, "agentProfile");
-      }
-      throw error;
-    }
-  }
-  if (typeof obj.splitTag !== "string" || !SPLIT_TAGS.includes(obj.splitTag)) {
-    throw new RunRecordValidationError(
-      `splitTag must be one of ${SPLIT_TAGS.join(", ")}, got ${String(obj.splitTag)}`,
-      "splitTag"
-    );
-  }
-  return input;
-}
-function resolveRunCostProvenance(run) {
-  if (run.costProvenance) return run.costProvenance;
-  if (run.outcome.raw.cost_estimated === 1) {
-    return { kind: "estimated", usd: run.costUsd };
-  }
-  if (run.costUsd > 0) return { kind: "observed", usd: run.costUsd };
-  return { kind: "uncaptured", usd: null };
-}
-function validateCostProvenance(input, costUsd) {
-  if (input === null || typeof input !== "object") {
-    throw new RunRecordValidationError("costProvenance must be an object", "costProvenance");
-  }
-  const value = input;
-  if (value.kind !== "observed" && value.kind !== "estimated" && value.kind !== "uncaptured") {
-    throw new RunRecordValidationError(
-      "costProvenance.kind must be observed, estimated, or uncaptured",
-      "costProvenance.kind"
-    );
-  }
-  if (value.kind === "uncaptured") {
-    if (value.usd !== null) {
-      throw new RunRecordValidationError(
-        "uncaptured costProvenance.usd must be null",
-        "costProvenance.usd"
-      );
-    }
-    if (costUsd !== 0) {
-      throw new RunRecordValidationError(
-        "uncaptured costProvenance requires the compatibility costUsd sentinel 0",
-        "costUsd"
-      );
-    }
-    return;
-  }
-  expectFiniteNumber(value.usd, "costProvenance.usd");
-  if (value.usd < 0) {
-    throw new RunRecordValidationError(
-      "costProvenance.usd must be non-negative",
-      "costProvenance.usd"
-    );
-  }
-  if (value.usd !== costUsd) {
-    throw new RunRecordValidationError(
-      "costProvenance.usd must equal costUsd",
-      "costProvenance.usd"
-    );
-  }
-}
-function isRunRecord(input) {
-  try {
-    validateRunRecord(input);
-    return true;
-  } catch {
-    return false;
-  }
-}
-function parseRunRecordSafe(input) {
-  try {
-    return { ok: true, value: validateRunRecord(input) };
-  } catch (e) {
-    if (e instanceof RunRecordValidationError) return { ok: false, error: e };
-    throw e;
-  }
-}
-function roundTripRunRecord(record) {
-  const json = JSON.stringify(record);
-  return validateRunRecord(JSON.parse(json));
-}
-function expectString(value, path) {
-  if (typeof value !== "string" || value.length === 0) {
-    throw new RunRecordValidationError(`expected non-empty string`, path);
-  }
-}
-function expectFiniteNumber(value, path) {
-  if (typeof value !== "number" || !Number.isFinite(value)) {
-    throw new RunRecordValidationError(`expected finite number`, path);
-  }
-}
-function expectNonNegativeNumber(value, path) {
-  expectFiniteNumber(value, path);
-  if (value < 0) {
-    throw new RunRecordValidationError("expected non-negative number", path);
-  }
-}
-function validateJudgeScores(value, path) {
-  if (value === null || typeof value !== "object") {
-    throw new RunRecordValidationError("judgeScores must be an object", path);
-  }
-  const rec = value;
-  const perJudge = rec.perJudge;
-  if (perJudge === null || typeof perJudge !== "object") {
-    throw new RunRecordValidationError("perJudge must be an object", `${path}.perJudge`);
-  }
-  for (const [judgeId, dims] of Object.entries(perJudge)) {
-    if (dims === null || typeof dims !== "object") {
-      throw new RunRecordValidationError(
-        "per-judge entry must be an object of dimension scores",
-        `${path}.perJudge.${judgeId}`
-      );
-    }
-    for (const [dim, score] of Object.entries(dims)) {
-      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`);
-    }
-  }
-  const perDimMean = rec.perDimMean;
-  if (perDimMean === null || typeof perDimMean !== "object") {
-    throw new RunRecordValidationError("perDimMean must be an object", `${path}.perDimMean`);
-  }
-  for (const [dim, mean] of Object.entries(perDimMean)) {
-    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`);
-  }
-  expectFiniteNumber(rec.composite, `${path}.composite`);
-  if (rec.failedJudges !== void 0) {
-    if (!Array.isArray(rec.failedJudges)) {
-      throw new RunRecordValidationError(
-        "failedJudges must be an array of strings",
-        `${path}.failedJudges`
-      );
-    }
-    for (let i = 0; i < rec.failedJudges.length; i++) {
-      const id = rec.failedJudges[i];
-      if (typeof id !== "string" || id.length === 0) {
-        throw new RunRecordValidationError(
-          "failedJudges entry must be a non-empty string",
-          `${path}.failedJudges[${i}]`
-        );
-      }
-    }
-  }
-  if (rec.notes !== void 0 && typeof rec.notes !== "string") {
-    throw new RunRecordValidationError("notes must be a string", `${path}.notes`);
-  }
-}
-function modelHasSnapshot(model) {
-  if (model.includes("@")) return true;
-  if (/-\d{8}$/.test(model)) return true;
-  if (/-\d{4}-\d{2}-\d{2}$/.test(model)) return true;
-  if (/:date-/.test(model)) return true;
-  return false;
-}
-
-export {
-  RunRecordValidationError,
-  validateRunRecord,
-  resolveRunCostProvenance,
-  isRunRecord,
-  parseRunRecordSafe,
-  roundTripRunRecord,
-  modelHasSnapshot
-};
-//# sourceMappingURL=chunk-S3UZOQ5Y.js.map

package/dist/chunk-S3UZOQ5Y.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag and either a `searchScore` or a `holdoutScore`.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n *   const rec = validateRunRecord(rawJson)         // throws on missing\n *   const ok  = isRunRecord(rawJson)               // boolean check\n *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport { ValidationError } from './errors'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n *  combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\nexport interface RunTokenUsage {\n  input: number\n  /** All generated tokens charged as output, including reasoning tokens. */\n  output: number\n  /** Reasoning-token subset of `output`, when the provider reports it. */\n  reasoning?: number\n  /** Prompt tokens served from a provider cache. */\n  cached?: number\n  /** Prompt tokens written into a provider cache. */\n  cacheWrite?: number\n}\n\n/**\n * How a run's USD amount was obtained.\n *\n * `costUsd` remains mandatory for wire compatibility. New producers should\n * always populate this discriminated union so a missing bill is never\n * mistaken for an observed zero-dollar run. For `uncaptured`, `costUsd` uses\n * the legacy `0` sentinel while this field carries the truthful null.\n */\nexport type RunCostProvenance =\n  | { kind: 'observed'; usd: number }\n  | { kind: 'estimated'; usd: number }\n  | { kind: 'uncaptured'; usd: null }\n\nexport interface RunJudgeMetadata {\n  model: string\n  promptVersion: string\n  /** [0,1] confidence the judge declared. Constant judge confidence\n   *  across many runs is a fallback signal (see `canary.ts`). */\n  confidence: number\n  /** True if the judge degraded to a fallback path (rules-only,\n   *  prior-call cache, etc.). The canary uses this to alert. */\n  fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n  /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n  perJudge: Record<string, Record<string, number>>\n  /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n  perDimMean: Record<string, number>\n  /** Composite mean across all dims and judges. Mirrors the score\n   *  the gate sees on `outcome.searchScore` / `holdoutScore`. */\n  composite: number\n  /** Judges that errored or returned an unparseable verdict. Recorded\n   *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n   *  not inferred from missing keys in `perJudge`. */\n  failedJudges?: string[]\n  /** Free-form notes the judges emitted (joined across judges or\n   *  first-judge only — consumer's choice). */\n  notes?: string\n}\n\nexport interface RunOutcome {\n  /** Score on the search/optimization split. Optional because a\n   *  holdout-only evaluation only fills `holdoutScore`. */\n  searchScore?: number\n  /** Score on the held-out split. Optional because a search-only run\n   *  only fills `searchScore`. At least one must be present. */\n  holdoutScore?: number\n  /** Bag of any other metric the run produced — judge dimensions,\n   *  pass/fail counters, latency stats, etc. Numeric only — keeps\n   *  reporters honest. */\n  raw: Record<string, number>\n  /** Per-judge / per-dim breakdown. Consumers writing ensemble\n   *  judgements populate this; substrate primitives like\n   *  `interRaterReliability` and `corpusInterRaterAgreement` accept\n   *  these records as input. Optional — single-judge or scalar-only\n   *  runs leave it unset. */\n  judgeScores?: JudgeScoresRecord\n  /** Authenticity / realness verdict — did the run build the REAL thing on the\n   *  intended infra, or fake it (see `./authenticity`)? Optional: only domains\n   *  with an authenticity config populate it. Carried in the corpus so the\n   *  flywheel / off-policy learning can optimize for real completion, not gamed\n   *  pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n   *  must not count as a real success regardless of `score`. */\n  realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n *     model (after any steering bundle merge).\n *   - `configHash` is the sha256 of the effective run config (model,\n *     temperature, tools, judges, splits). The pair (promptHash,\n *     configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n *   - `model` MUST encode a snapshot version. Bare aliases like\n *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n */\nexport interface RunRecord {\n  /** UUID for the run. */\n  runId: string\n  /** Logical experiment grouping (a treatment vs a baseline within\n   *  the same sweep should share `experimentId`). */\n  experimentId: string\n  /** Stable identifier for the candidate (variant) being run. The\n   *  promotion gate compares two `candidateId`s on matched items. */\n  candidateId: string\n  /** RNG seed for the run. Always recorded — silent re-seeding is\n   *  the most common cause of non-reproducible numbers. */\n  seed: number\n  /** Model identifier WITH snapshot version. */\n  model: string\n  /** sha256 of the effective prompt (post-steering). */\n  promptHash: string\n  /** sha256 of the effective config. */\n  configHash: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** End-to-end wall-clock duration in milliseconds. */\n  wallMs: number\n  /** Time spent queued before execution started, if known. */\n  queueMs?: number\n  /** Total USD cost. Mandatory — runs without a cost number are\n   *  unbounded by definition and must not be admitted into the gate.\n   *  `0` is retained as the compatibility sentinel for an uncaptured amount;\n   *  inspect `costProvenance` before treating it as observed. */\n  costUsd: number\n  /** Observed, model-priced estimate, or genuinely uncaptured USD amount.\n   *  Optional only so existing serialized RunRecords remain valid. */\n  costProvenance?: RunCostProvenance\n  /** Token usage breakdown. */\n  tokenUsage: RunTokenUsage\n  /** Judge-side metadata, if a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /** Per-split scores + raw bag. */\n  outcome: RunOutcome\n  /** Canonical, cross-agent failure class drawn from the shared\n   *  `FAILURE_CLASSES` taxonomy. This is the aggregation key that makes\n   *  \"which failure dominates across the whole fleet\" answerable in ONE\n   *  vocabulary — every agent classifies against the same enum. Producers\n   *  set it via the substrate classifier; leave unset only when the failure\n   *  genuinely can't be classified. */\n  failureClass?: FailureClass\n  /** Free-form domain-specific failure detail, scoped UNDER `failureClass`\n   *  (e.g. failureClass='tool_recovery_failure', failureMode='forge_build_unsatisfied').\n   *  The within-agent drill-down; `failureClass` is the cross-agent key. */\n  failureMode?: string\n  /** Which split this run was drawn from. */\n  splitTag: RunSplitTag\n  /**\n   * Stable scenario identifier the run was scored against. Optional for\n   * backwards compatibility, but **strongly recommended**: every primitive\n   * that pairs runs by scenario (preferences, paired stats, BT tournament)\n   * keys on this. The campaign artifact populates it canonically; legacy\n   * runs without it fall back to inference from `outcome.raw.scenario_id`\n   * or `experimentId`.\n   */\n  scenarioId?: string\n  /**\n   * Canonical identity for the agent profile cell that produced this row:\n   * profile artifact hash plus optional harness/model/prompt/reporting\n   * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n   * longitudinal reports by the complete source profile, not by a loose\n   * candidate label or opaque config hash.\n   */\n  agentProfile?: AgentProfileCell\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n  'runId',\n  'experimentId',\n  'candidateId',\n  'seed',\n  'model',\n  'promptHash',\n  'configHash',\n  'commitSha',\n  'wallMs',\n  'costUsd',\n  'tokenUsage',\n  'outcome',\n  'splitTag',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\n\nexport class RunRecordValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n\n  for (const key of MANDATORY_TOP_LEVEL) {\n    if (!(key in obj)) {\n      throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n    }\n  }\n\n  expectString(obj.runId, 'runId')\n  expectString(obj.experimentId, 'experimentId')\n  expectString(obj.candidateId, 'candidateId')\n  expectFiniteNumber(obj.seed, 'seed')\n  expectString(obj.model, 'model')\n  expectString(obj.promptHash, 'promptHash')\n  expectString(obj.configHash, 'configHash')\n  expectString(obj.commitSha, 'commitSha')\n  expectNonNegativeNumber(obj.wallMs, 'wallMs')\n  if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n  expectNonNegativeNumber(obj.costUsd, 'costUsd')\n  if (obj.costProvenance !== undefined) {\n    validateCostProvenance(obj.costProvenance, obj.costUsd as number)\n  }\n\n  // Snapshot discipline: bare model aliases are not paper-grade.\n  if (!modelHasSnapshot(obj.model as string)) {\n    throw new RunRecordValidationError(\n      `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,\n      'model',\n    )\n  }\n\n  // Token usage.\n  const tu = obj.tokenUsage\n  if (tu === null || typeof tu !== 'object') {\n    throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n  }\n  const tuRec = tu as Record<string, unknown>\n  expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n  expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n  if (tuRec.reasoning !== undefined) {\n    expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n    if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n      throw new RunRecordValidationError(\n        'reasoning tokens must be a subset of output tokens',\n        'tokenUsage.reasoning',\n      )\n    }\n  }\n  if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n  if (tuRec.cacheWrite !== undefined) {\n    expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n  }\n\n  // Judge metadata, optional.\n  if (obj.judgeMetadata !== undefined) {\n    const jm = obj.judgeMetadata\n    if (jm === null || typeof jm !== 'object') {\n      throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n    }\n    const jmRec = jm as Record<string, unknown>\n    expectString(jmRec.model, 'judgeMetadata.model')\n    expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n    expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n    if (typeof jmRec.fallback !== 'boolean') {\n      throw new RunRecordValidationError(\n        'judgeMetadata.fallback must be boolean',\n        'judgeMetadata.fallback',\n      )\n    }\n  }\n\n  // Outcome.\n  const out = obj.outcome\n  if (out === null || typeof out !== 'object') {\n    throw new RunRecordValidationError('outcome must be an object', 'outcome')\n  }\n  const outRec = out as Record<string, unknown>\n  if (outRec.searchScore !== undefined)\n    expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n  if (outRec.holdoutScore !== undefined)\n    expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n  if (outRec.searchScore === undefined && outRec.holdoutScore === undefined) {\n    throw new RunRecordValidationError(\n      'outcome must define searchScore or holdoutScore (or both)',\n      'outcome',\n    )\n  }\n  const raw = outRec.raw\n  if (raw === null || typeof raw !== 'object') {\n    throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n  }\n  for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n    expectFiniteNumber(v, `outcome.raw.${k}`)\n  }\n  // Realness verdict, optional.\n  if (outRec.realness !== undefined) {\n    const r = outRec.realness\n    if (r === null || typeof r !== 'object') {\n      throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n    }\n    const rr = r as Record<string, unknown>\n    expectFiniteNumber(rr.score, 'outcome.realness.score')\n    if (typeof rr.gated !== 'boolean') {\n      throw new RunRecordValidationError(\n        'outcome.realness.gated must be a boolean',\n        'outcome.realness.gated',\n      )\n    }\n  }\n\n  // Per-judge / per-dim breakdown, optional.\n  if (outRec.judgeScores !== undefined) {\n    validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n  }\n\n  // Failure mode optional.\n  if (\n    obj.failureClass !== undefined &&\n    (typeof obj.failureClass !== 'string' ||\n      !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n  ) {\n    throw new RunRecordValidationError(\n      `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n      'failureClass',\n    )\n  }\n  if (obj.failureMode !== undefined) expectString(obj.failureMode, 'failureMode')\n\n  if (obj.agentProfile !== undefined) {\n    try {\n      const profile = validateAgentProfileCell(obj.agentProfile)\n      if (profile.model !== undefined && profile.model !== obj.model) {\n        throw new RunRecordValidationError(\n          `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n          'agentProfile.model',\n        )\n      }\n      if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n        throw new RunRecordValidationError(\n          `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n          'agentProfile.promptHash',\n        )\n      }\n    } catch (error) {\n      if (error instanceof RunRecordValidationError) throw error\n      if (error instanceof Error) {\n        throw new RunRecordValidationError(error.message, 'agentProfile')\n      }\n      throw error\n    }\n  }\n\n  // Split tag.\n  if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n    throw new RunRecordValidationError(\n      `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n      'splitTag',\n    )\n  }\n\n  return input as RunRecord\n}\n\n/**\n * Resolve provenance for both new and legacy records.\n *\n * Legacy producers sometimes set `outcome.raw.cost_estimated = 1`. A positive\n * unlabeled amount is treated as observed, matching the historical contract.\n * Zero without an explicit label is conservatively uncaptured: claiming an\n * observed $0 would be stronger than the serialized evidence supports.\n */\nexport function resolveRunCostProvenance(\n  run: Pick<RunRecord, 'costUsd' | 'costProvenance' | 'outcome'>,\n): RunCostProvenance {\n  if (run.costProvenance) return run.costProvenance\n  if (run.outcome.raw.cost_estimated === 1) {\n    return { kind: 'estimated', usd: run.costUsd }\n  }\n  if (run.costUsd > 0) return { kind: 'observed', usd: run.costUsd }\n  return { kind: 'uncaptured', usd: null }\n}\n\nfunction validateCostProvenance(input: unknown, costUsd: number): void {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n  }\n  const value = input as Record<string, unknown>\n  if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n    throw new RunRecordValidationError(\n      'costProvenance.kind must be observed, estimated, or uncaptured',\n      'costProvenance.kind',\n    )\n  }\n  if (value.kind === 'uncaptured') {\n    if (value.usd !== null) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance.usd must be null',\n        'costProvenance.usd',\n      )\n    }\n    if (costUsd !== 0) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance requires the compatibility costUsd sentinel 0',\n        'costUsd',\n      )\n    }\n    return\n  }\n  expectFiniteNumber(value.usd, 'costProvenance.usd')\n  if ((value.usd as number) < 0) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must be non-negative',\n      'costProvenance.usd',\n    )\n  }\n  if (value.usd !== costUsd) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must equal costUsd',\n      'costProvenance.usd',\n    )\n  }\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n  try {\n    validateRunRecord(input)\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n  input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n  try {\n    return { ok: true, value: validateRunRecord(input) }\n  } catch (e) {\n    if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n    throw e\n  }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n  const json = JSON.stringify(record)\n  return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new RunRecordValidationError(`expected non-empty string`, path)\n  }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value)) {\n    throw new RunRecordValidationError(`expected finite number`, path)\n  }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n  expectFiniteNumber(value, path)\n  if ((value as number) < 0) {\n    throw new RunRecordValidationError('expected non-negative number', path)\n  }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object') {\n    throw new RunRecordValidationError('judgeScores must be an object', path)\n  }\n  const rec = value as Record<string, unknown>\n\n  const perJudge = rec.perJudge\n  if (perJudge === null || typeof perJudge !== 'object') {\n    throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n  }\n  for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n    if (dims === null || typeof dims !== 'object') {\n      throw new RunRecordValidationError(\n        'per-judge entry must be an object of dimension scores',\n        `${path}.perJudge.${judgeId}`,\n      )\n    }\n    for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n    }\n  }\n\n  const perDimMean = rec.perDimMean\n  if (perDimMean === null || typeof perDimMean !== 'object') {\n    throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n  }\n  for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n  }\n\n  expectFiniteNumber(rec.composite, `${path}.composite`)\n\n  if (rec.failedJudges !== undefined) {\n    if (!Array.isArray(rec.failedJudges)) {\n      throw new RunRecordValidationError(\n        'failedJudges must be an array of strings',\n        `${path}.failedJudges`,\n      )\n    }\n    for (let i = 0; i < rec.failedJudges.length; i++) {\n      const id = rec.failedJudges[i]\n      if (typeof id !== 'string' || id.length === 0) {\n        throw new RunRecordValidationError(\n          'failedJudges entry must be a non-empty string',\n          `${path}.failedJudges[${i}]`,\n        )\n      }\n    }\n  }\n\n  if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n    throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n  }\n}\n\n/**\n * Heuristic snapshot check. Accepts:\n *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)\n *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)\n *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)\n *   - explicit `:date-...` Vertex-style tags\n *\n * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap\n * silently as providers ship new snapshots.\n */\nexport function modelHasSnapshot(model: string): boolean {\n  if (model.includes('@')) return true\n  if (/-\\d{8}$/.test(model)) return true\n  if (/-\\d{4}-\\d{2}-\\d{2}$/.test(model)) return true\n  if (/:date-/.test(model)) return true\n  return false\n}\n"],"mappings":";;;;;;;;;;;AA4NA,IAAM,sBAAsB;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,aAAyC,CAAC,UAAU,OAAO,SAAS;AAEnE,IAAM,2BAAN,cAAuC,gBAAgB;AAAA,EACnD;AAAA,EACT,YAAY,SAAiB,OAAO,IAAI;AACtC,UAAM,OAAO,GAAG,OAAO,QAAQ,IAAI,MAAM,OAAO;AAChD,SAAK,OAAO;AAAA,EACd;AACF;AAOO,SAAS,kBAAkB,OAA2B;AAC3D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iBAAiB;AAAA,EACtD;AACA,QAAM,MAAM;AAEZ,aAAW,OAAO,qBAAqB;AACrC,QAAI,EAAE,OAAO,MAAM;AACjB,YAAM,IAAI,yBAAyB,4BAA4B,GAAG,GAAG;AAAA,IACvE;AAAA,EACF;AAEA,eAAa,IAAI,OAAO,OAAO;AAC/B,eAAa,IAAI,cAAc,cAAc;AAC7C,eAAa,IAAI,aAAa,aAAa;AAC3C,qBAAmB,IAAI,MAAM,MAAM;AACnC,eAAa,IAAI,OAAO,OAAO;AAC/B,eAAa,IAAI,YAAY,YAAY;AACzC,eAAa,IAAI,YAAY,YAAY;AACzC,eAAa,IAAI,WAAW,WAAW;AACvC,0BAAwB,IAAI,QAAQ,QAAQ;AAC5C,MAAI,IAAI,YAAY,OAAW,yBAAwB,IAAI,SAAS,SAAS;AAC7E,0BAAwB,IAAI,SAAS,SAAS;AAC9C,MAAI,IAAI,mBAAmB,QAAW;AACpC,2BAAuB,IAAI,gBAAgB,IAAI,OAAiB;AAAA,EAClE;AAGA,MAAI,CAAC,iBAAiB,IAAI,KAAe,GAAG;AAC1C,UAAM,IAAI;AAAA,MACR,UAAU,IAAI,KAAK;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AAGA,QAAM,KAAK,IAAI;AACf,MAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,UAAM,IAAI,yBAAyB,gCAAgC,YAAY;AAAA,EACjF;AACA,QAAM,QAAQ;AACd,0BAAwB,MAAM,OAAO,kBAAkB;AACvD,0BAAwB,MAAM,QAAQ,mBAAmB;AACzD,MAAI,MAAM,cAAc,QAAW;AACjC,4BAAwB,MAAM,WAAW,sBAAsB;AAC/D,QAAK,MAAM,YAAwB,MAAM,QAAmB;AAC1D,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,WAAW,OAAW,yBAAwB,MAAM,QAAQ,mBAAmB;AACzF,MAAI,MAAM,eAAe,QAAW;AAClC,4BAAwB,MAAM,YAAY,uBAAuB;AAAA,EACnE;AAGA,MAAI,IAAI,kBAAkB,QAAW;AACnC,UAAM,KAAK,IAAI;AACf,QAAI,OAAO,QAAQ,OAAO,OAAO,UAAU;AACzC,YAAM,IAAI,yBAAyB,mCAAmC,eAAe;AAAA,IACvF;AACA,UAAM,QAAQ;AACd,iBAAa,MAAM,OAAO,qBAAqB;AAC/C,iBAAa,MAAM,eAAe,6BAA6B;AAC/D,uBAAmB,MAAM,YAAY,0BAA0B;AAC/D,QAAI,OAAO,MAAM,aAAa,WAAW;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAGA,QAAM,MAAM,IAAI;AAChB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,6BAA6B,SAAS;AAAA,EAC3E;AACA,QAAM,SAAS;AACf,MAAI,OAAO,gBAAgB;AACzB,uBAAmB,OAAO,aAAa,qBAAqB;AAC9D,MAAI,OAAO,iBAAiB;AAC1B,uBAAmB,OAAO,cAAc,sBAAsB;AAChE,MAAI,OAAO,gBAAgB,UAAa,OAAO,iBAAiB,QAAW;AACzE,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,QAAM,MAAM,OAAO;AACnB,MAAI,QAAQ,QAAQ,OAAO,QAAQ,UAAU;AAC3C,UAAM,IAAI,yBAAyB,iCAAiC,aAAa;AAAA,EACnF;AACA,aAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,GAA8B,GAAG;AACnE,uBAAmB,GAAG,eAAe,CAAC,EAAE;AAAA,EAC1C;AAEA,MAAI,OAAO,aAAa,QAAW;AACjC,UAAM,IAAI,OAAO;AACjB,QAAI,MAAM,QAAQ,OAAO,MAAM,UAAU;AACvC,YAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;AAAA,IAC7F;AACA,UAAM,KAAK;AACX,uBAAmB,GAAG,OAAO,wBAAwB;AACrD,QAAI,OAAO,GAAG,UAAU,WAAW;AACjC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAGA,MAAI,OAAO,gBAAgB,QAAW;AACpC,wBAAoB,OAAO,aAAa,qBAAqB;AAAA,EAC/D;AAGA,MACE,IAAI,iBAAiB,WACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAC5D;AACA,UAAM,IAAI;AAAA,MACR,+BAA+B,gBAAgB,KAAK,IAAI,CAAC;AAAA,MACzD;AAAA,IACF;AAAA,EACF;AACA,MAAI,IAAI,gBAAgB,OAAW,cAAa,IAAI,aAAa,aAAa;AAE9E,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI;AACF,YAAM,UAAU,yBAAyB,IAAI,YAAY;AACzD,UAAI,QAAQ,UAAU,UAAa,QAAQ,UAAU,IAAI,OAAO;AAC9D,cAAM,IAAI;AAAA,UACR,uBAAuB,QAAQ,KAAK,2BAA2B,IAAI,KAAK;AAAA,UACxE;AAAA,QACF;AAAA,MACF;AACA,UAAI,QAAQ,eAAe,UAAa,QAAQ,eAAe,IAAI,YAAY;AAC7E,cAAM,IAAI;AAAA,UACR,4BAA4B,QAAQ,UAAU,gCAAgC,IAAI,UAAU;AAAA,UAC5F;AAAA,QACF;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,yBAA0B,OAAM;AACrD,UAAI,iBAAiB,OAAO;AAC1B,cAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;AAAA,MAClE;AACA,YAAM;AAAA,IACR;AAAA,EACF;AAGA,MAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GAAG;AACzF,UAAM,IAAI;AAAA,MACR,2BAA2B,WAAW,KAAK,IAAI,CAAC,SAAS,OAAO,IAAI,QAAQ,CAAC;AAAA,MAC7E;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAUO,SAAS,yBACd,KACmB;AACnB,MAAI,IAAI,eAAgB,QAAO,IAAI;AACnC,MAAI,IAAI,QAAQ,IAAI,mBAAmB,GAAG;AACxC,WAAO,EAAE,MAAM,aAAa,KAAK,IAAI,QAAQ;AAAA,EAC/C;AACA,MAAI,IAAI,UAAU,EAAG,QAAO,EAAE,MAAM,YAAY,KAAK,IAAI,QAAQ;AACjE,SAAO,EAAE,MAAM,cAAc,KAAK,KAAK;AACzC;AAEA,SAAS,uBAAuB,OAAgB,SAAuB;AACrE,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;AAAA,EACzF;AACA,QAAM,QAAQ;AACd,MAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAAc;AAC1F,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,SAAS,cAAc;AAC/B,QAAI,MAAM,QAAQ,MAAM;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,YAAY,GAAG;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA;AAAA,EACF;AACA,qBAAmB,MAAM,KAAK,oBAAoB;AAClD,MAAK,MAAM,MAAiB,GAAG;AAC7B,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,MAAI,MAAM,QAAQ,SAAS;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAGO,SAAS,YAAY,OAAoC;AAC9D,MAAI;AACF,sBAAkB,KAAK;AACvB,WAAO;AAAA,EACT,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGO,SAAS,mBACd,OACiF;AACjF,MAAI;AACF,WAAO,EAAE,IAAI,MAAM,OAAO,kBAAkB,KAAK,EAAE;AAAA,EACrD,SAAS,GAAG;AACV,QAAI,aAAa,yBAA0B,QAAO,EAAE,IAAI,OAAO,OAAO,EAAE;AACxE,UAAM;AAAA,EACR;AACF;AAGO,SAAS,mBAAmB,QAA8B;AAC/D,QAAM,OAAO,KAAK,UAAU,MAAM;AAClC,SAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;AACxD,MAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAAG;AACnD,UAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAAA,EACtE;AACF;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;AAC9D,MAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;AACxD,UAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAAA,EACnE;AACF;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;AACnE,qBAAmB,OAAO,IAAI;AAC9B,MAAK,QAAmB,GAAG;AACzB,UAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAAA,EACzE;AACF;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;AAC/D,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,IAAI,yBAAyB,iCAAiC,IAAI;AAAA,EAC1E;AACA,QAAM,MAAM;AAEZ,QAAM,WAAW,IAAI;AACrB,MAAI,aAAa,QAAQ,OAAO,aAAa,UAAU;AACrD,UAAM,IAAI,yBAAyB,8BAA8B,GAAG,IAAI,WAAW;AAAA,EACrF;AACA,aAAW,CAAC,SAAS,IAAI,KAAK,OAAO,QAAQ,QAAmC,GAAG;AACjF,QAAI,SAAS,QAAQ,OAAO,SAAS,UAAU;AAC7C,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI,aAAa,OAAO;AAAA,MAC7B;AAAA,IACF;AACA,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,IAA+B,GAAG;AAC1E,yBAAmB,OAAO,GAAG,IAAI,aAAa,OAAO,IAAI,GAAG,EAAE;AAAA,IAChE;AAAA,EACF;AAEA,QAAM,aAAa,IAAI;AACvB,MAAI,eAAe,QAAQ,OAAO,eAAe,UAAU;AACzD,UAAM,IAAI,yBAAyB,gCAAgC,GAAG,IAAI,aAAa;AAAA,EACzF;AACA,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,UAAqC,GAAG;AAC/E,uBAAmB,MAAM,GAAG,IAAI,eAAe,GAAG,EAAE;AAAA,EACtD;AAEA,qBAAmB,IAAI,WAAW,GAAG,IAAI,YAAY;AAErD,MAAI,IAAI,iBAAiB,QAAW;AAClC,QAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GAAG;AACpC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,GAAG,IAAI;AAAA,MACT;AAAA,IACF;AACA,aAAS,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;AAChD,YAAM,KAAK,IAAI,aAAa,CAAC;AAC7B,UAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAAG;AAC7C,cAAM,IAAI;AAAA,UACR;AAAA,UACA,GAAG,IAAI,iBAAiB,CAAC;AAAA,QAC3B;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,MAAI,IAAI,UAAU,UAAa,OAAO,IAAI,UAAU,UAAU;AAC5D,UAAM,IAAI,yBAAyB,0BAA0B,GAAG,IAAI,QAAQ;AAAA,EAC9E;AACF;AAYO,SAAS,iBAAiB,OAAwB;AACvD,MAAI,MAAM,SAAS,GAAG,EAAG,QAAO;AAChC,MAAI,UAAU,KAAK,KAAK,EAAG,QAAO;AAClC,MAAI,sBAAsB,KAAK,KAAK,EAAG,QAAO;AAC9C,MAAI,SAAS,KAAK,KAAK,EAAG,QAAO;AACjC,SAAO;AACT;","names":[]}