@tangle-network/agent-eval 0.181.0 → 0.182.0

package/dist/rollout-BsN6pOU7.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"rollout-BsN6pOU7.js","names":["isRecord"],"sources":["../src/rollout/interchange/harbor.ts","../src/rollout/readers/opencode-sqlite.ts"],"sourcesContent":["/**\n * Harbor ATIF-v1.7 interchange — `tangle.rollout.v1` ⇄ Agent Trajectory\n * Interchange Format.\n *\n * ATIF is the portability format (spec:\n * https://www.harborframework.com/docs/agents/trajectory-format, normative\n * RFC: harbor-framework/harbor `rfcs/0001-trajectory-format.md`). It sits\n * BELOW the waist of the rollout hourglass in both directions — export reads\n * `RolloutLine[]`, import writes `RolloutLine[]` — and it is never a source\n * of training labels:\n *\n *   ATIF models NO reward, NO judge verdict, NO task/split coordinates.\n *\n * Consequences, both deliberate:\n *   - EXPORT drops `outcome.reward`, `outcome.reward_source` and\n *     `outcome.verdict` entirely. They are not smuggled into `extra`: a\n *     third-party reading our ATIF file must not be able to mistake an\n *     agent-eval judge score for something ATIF sanctioned.\n *   - IMPORT therefore mints UNLABELED lines: `reward: null` (the existing\n *     \"null reward is a labeled gap, never 0\" semantics), `verdict: null`,\n *     and a `provenance.gap` naming the missing label. An imported\n *     trajectory is not a training example until a judge scores it.\n *\n * Everything else we own that ATIF has no field for travels in a namespaced\n * escrow at `extra.tangle.*`, so our own round-trip is exact while a foreign\n * reader can ignore it. Fields that neither ATIF nor the escrow can carry\n * come back explicitly null / fail-closed, never invented.\n *\n * THE ESCROW IS NAMESPACED, NOT AUTHENTICATED. Anyone can write\n * `extra.tangle.*` into a file. So the escrow may restore what a value IS, but\n * never what a line is ALLOWED to do: `task.split` is forced to `holdout` on\n * every import regardless of what the document claims, and promoting an\n * imported trajectory to a trainable split is an explicit, greppable act\n * (`relabelImportedSplit`) rather than a property of the file. The document\n * keeps its claim — the claim just is not authority.\n *\n * Multi-agent shape differs on purpose. ATIF EMBEDS children in\n * `subagent_trajectories`; we keep a flat ledger with a normalized\n * `parent_rollout_id` edge. Export assembles the tree, import flattens it.\n * `session_id` is RUN-scoped in ATIF, so it carries `run_id` — the coordinate\n * that is shared by every invocation of one run — not `rollout_id`, which\n * identifies a single invocation and would split one run across session ids.\n *\n * ROUND-TRIPPING IS IDEMPOTENT: `import(export(import(export(x))))` is\n * byte-identical to `import(export(x))`. Import composes `provenance.gap` as a\n * de-duplicated ordered set rather than appending, and it emits every\n * `ChatMessage` with keys in the canonical schema order (role, content,\n * reasoning_content, tool_calls, tool_call_id, name, is_copied_context), so a\n * ledger hashed on serialized bytes sees no diff across further passes. The\n * FIRST import may re-order a producer's keys — that is the canonicalization.\n *\n * NOT building a Letta converter. Letta's trajectory-v1 is a strict subset of\n * what we need from ATIF here — no per-step or aggregate cost, no\n * multi-agent/subagent structure, no token-id or logprob channel — so a Letta\n * sink would carry less than this one and add a second format to keep\n * correct. Decision recorded in docs/rollout.md; do not re-litigate without a\n * concrete consumer that reads Letta and cannot read ATIF.\n */\n\nimport {\n  assertRolloutLine,\n  type ChatMessage,\n  type ChatRole,\n  type ChatToolCall,\n  type GatedEvidence,\n  ROLLOUT_CAPTURES,\n  ROLLOUT_ROLES,\n  ROLLOUT_SCHEMA,\n  ROLLOUT_SPLITS,\n  type RolloutArtifacts,\n  type RolloutCapture,\n  type RolloutCostBlock,\n  type RolloutLine,\n  type RolloutPolicy,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n  type RolloutTask,\n  type ToolDef,\n} from '../schema'\n\nexport const ATIF_SCHEMA_VERSION = 'ATIF-v1.7'\n\n/** Gap note on every imported line — ATIF carries no verdict, so nothing is scored. */\nexport const HARBOR_IMPORT_GAP = 'imported from Harbor ATIF; no verdict'\n\n/** Namespaced escrow key for everything ATIF does not model but we must not lose. */\nconst ESCROW = 'tangle'\n\n/**\n * Marks a step we synthesized purely to hold tool results that answered no\n * assistant turn IN THIS DOCUMENT. ATIF has no `tool` source, so such results\n * need a carrier step; the marker lets import drop the carrier instead of\n * inventing a turn.\n *\n * A carrier's results carry NO `source_call_id`. RFC 0001 rule 2 requires every\n * `source_call_id` to match a `tool_call_id` in the same step's `tool_calls`,\n * and a carrier declares no calls (a `system` step cannot legally make one), so\n * emitting the id there would produce an invalid document. The id is escrowed\n * instead and restored verbatim on import — the link survives without the\n * document lying about who made the call.\n */\nconst TOOL_RESULTS_ONLY = 'tool-results-only'\n\n// ---------------------------------------------------------------------------\n// ATIF-v1.7 wire types (RFC 0001). Optional fields are optional here too.\n// ---------------------------------------------------------------------------\n\nexport type HarborStepSource = 'system' | 'user' | 'agent'\n\nexport interface HarborImageSource {\n  media_type: string\n  path: string\n}\n\nexport interface HarborContentPart {\n  type: 'text' | 'image'\n  text?: string\n  source?: HarborImageSource\n}\n\nexport interface HarborToolCall {\n  tool_call_id: string\n  function_name: string\n  /** ATIF requires a decoded JSON object here, unlike our raw argument string. */\n  arguments: Record<string, unknown>\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborSubagentTrajectoryRef {\n  trajectory_id?: string\n  trajectory_path?: string\n  /** Informational only since v1.7 — never a resolution key. */\n  session_id?: string\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborObservationResult {\n  source_call_id?: string\n  content?: string | HarborContentPart[]\n  subagent_trajectory_ref?: HarborSubagentTrajectoryRef[]\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborObservation {\n  results: HarborObservationResult[]\n}\n\nexport interface HarborMetrics {\n  prompt_tokens?: number\n  completion_tokens?: number\n  cached_tokens?: number\n  cost_usd?: number\n  prompt_token_ids?: number[]\n  completion_token_ids?: number[]\n  logprobs?: number[]\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborStep {\n  /** Ordinal, sequential from 1. */\n  step_id: number\n  timestamp?: string\n  source: HarborStepSource\n  model_name?: string\n  reasoning_effort?: string | number\n  message: string | HarborContentPart[]\n  reasoning_content?: string\n  tool_calls?: HarborToolCall[]\n  observation?: HarborObservation\n  metrics?: HarborMetrics\n  llm_call_count?: number\n  is_copied_context?: boolean\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborAgent {\n  name: string\n  version: string\n  model_name?: string\n  /** OpenAI function-calling schema — byte-identical to our `ToolDef`. */\n  tool_definitions?: ToolDef[]\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborFinalMetrics {\n  total_prompt_tokens?: number\n  total_completion_tokens?: number\n  total_cached_tokens?: number\n  total_cost_usd?: number\n  total_steps?: number\n  extra?: Record<string, unknown>\n}\n\nexport interface HarborTrajectory {\n  schema_version: string\n  session_id?: string\n  /** Required on embedded subagents; we always set it so lines stay joinable. */\n  trajectory_id?: string\n  agent: HarborAgent\n  steps: HarborStep[]\n  notes?: string\n  final_metrics?: HarborFinalMetrics\n  continued_trajectory_ref?: string\n  subagent_trajectories?: HarborTrajectory[]\n  extra?: Record<string, unknown>\n}\n\n// ---------------------------------------------------------------------------\n// Shared helpers\n// ---------------------------------------------------------------------------\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nconst asString = (v: unknown): string | undefined => (typeof v === 'string' ? v : undefined)\n\nconst asNumberOrNull = (v: unknown): number | null =>\n  typeof v === 'number' && Number.isFinite(v) ? v : null\n\nconst asIntegerOrUndefined = (v: unknown): number | undefined =>\n  Number.isInteger(v) ? (v as number) : undefined\n\nconst asNumberArray = (v: unknown): number[] | undefined =>\n  Array.isArray(v) && v.every((n) => typeof n === 'number' && Number.isFinite(n))\n    ? (v as number[])\n    : undefined\n\nconst asIntegerArray = (v: unknown): number[] | undefined =>\n  Array.isArray(v) && v.every((n) => Number.isInteger(n)) ? (v as number[]) : undefined\n\n/**\n * Builds a `ChatMessage` with keys in the schema's declaration order.\n *\n * Object key order is insertion order in JS, so a message assembled\n * conditionally field-by-field serializes differently depending on which\n * optional fields were present — which makes a byte-hashed ledger see a diff\n * across an import that changed nothing. One builder, one order, stable bytes.\n */\nfunction canonicalChatMessage(parts: {\n  role: ChatRole\n  content: string | null\n  reasoning_content?: string\n  tool_calls?: ChatToolCall[]\n  tool_call_id?: string\n  name?: string\n  is_copied_context?: boolean\n}): ChatMessage {\n  const message: ChatMessage = { role: parts.role, content: parts.content }\n  if (parts.reasoning_content !== undefined) message.reasoning_content = parts.reasoning_content\n  if (parts.tool_calls !== undefined) message.tool_calls = parts.tool_calls\n  if (parts.tool_call_id !== undefined) message.tool_call_id = parts.tool_call_id\n  if (parts.name !== undefined) message.name = parts.name\n  if (parts.is_copied_context === true) message.is_copied_context = true\n  return message\n}\n\n/** Reads the `extra.tangle` escrow, tolerating a foreign file that has none. */\nfunction escrowOf(extra: Record<string, unknown> | undefined): Record<string, unknown> | undefined {\n  if (!isRecord(extra)) return undefined\n  const value = extra[ESCROW]\n  return isRecord(value) ? value : undefined\n}\n\nfunction escrowSection(\n  escrow: Record<string, unknown> | undefined,\n  key: string,\n): Record<string, unknown> | undefined {\n  if (escrow === undefined) return undefined\n  const value = escrow[key]\n  return isRecord(value) ? value : undefined\n}\n\n// ---------------------------------------------------------------------------\n// Export: RolloutLine[] → ATIF trajectory tree\n// ---------------------------------------------------------------------------\n\nfunction toHarborToolCall(call: ChatToolCall): HarborToolCall {\n  // ATIF requires `arguments` to be a JSON object; ours is the raw\n  // JSON-encoded string the model emitted, and models really do emit\n  // malformed ones. Keep the exact bytes in `extra` so import restores the\n  // string verbatim instead of re-serializing a normalized parse.\n  let parsed: Record<string, unknown> | undefined\n  try {\n    const decoded: unknown = JSON.parse(call.function.arguments)\n    if (isRecord(decoded)) parsed = decoded\n  } catch {\n    parsed = undefined\n  }\n  return {\n    tool_call_id: call.id,\n    function_name: call.function.name,\n    arguments: parsed ?? {},\n    extra: { [ESCROW]: { arguments_raw: call.function.arguments } },\n  }\n}\n\n/**\n * `linked` = this result is attached to the step that actually declared the\n * call, so `source_call_id` is legal (RFC 0001 rule 2). When it is not, the id\n * goes to escrow instead of onto the wire.\n */\nfunction toObservationResult(message: ChatMessage, linked: boolean): HarborObservationResult {\n  const result: HarborObservationResult = {}\n  if (linked && message.tool_call_id !== undefined) result.source_call_id = message.tool_call_id\n  if (message.content !== null) result.content = message.content\n  const escrow: Record<string, unknown> = {}\n  if (!linked && message.tool_call_id !== undefined) escrow.source_call_id = message.tool_call_id\n  if (message.name !== undefined) escrow.name = message.name\n  // `content: null` on a tool turn is meaningful (a tool that returned\n  // nothing) and is not the same as the empty string ATIF would round-trip it to.\n  if (message.content === null) escrow.content_null = true\n  if (message.reasoning_content !== undefined) escrow.reasoning_content = message.reasoning_content\n  if (message.is_copied_context === true) escrow.is_copied_context = true\n  if (Object.keys(escrow).length > 0) result.extra = { [ESCROW]: escrow }\n  return result\n}\n\n/** True when `step` is the agent turn that declared `callId` — the rule-2 test. */\nfunction declaresCall(step: HarborStep, callId: string | undefined): boolean {\n  if (callId === undefined) return false\n  return step.tool_calls?.some((call) => call.tool_call_id === callId) === true\n}\n\n/**\n * Attaches the four RL fields lifted from ATIF onto the agent step the span\n * describes. They are ALSO escrowed with the whole span under `extra.tangle`,\n * on purpose: the escrow is how our own round-trip stays exact, and these\n * native fields are how a foreign consumer — which never reads our escrow —\n * gets the logprobs and token ids at all. A field carried only in the escrow is\n * not an interchange field.\n */\nfunction applySpanMetrics(step: HarborStep, span: RolloutStep | undefined): void {\n  if (span === undefined) return\n  if (span.llm_call_count !== undefined) step.llm_call_count = span.llm_call_count\n  const metrics: HarborMetrics = {}\n  if (span.prompt_token_ids !== undefined) metrics.prompt_token_ids = span.prompt_token_ids\n  if (span.completion_token_ids !== undefined) {\n    metrics.completion_token_ids = span.completion_token_ids\n  }\n  if (span.logprobs !== undefined) metrics.logprobs = span.logprobs\n  if (Object.keys(metrics).length > 0) step.metrics = metrics\n}\n\n/**\n * Coalescing fold, not a 1:1 map: ATIF has no `tool` source, so a tool result\n * becomes an `observation.results[]` entry on the agent step THAT DECLARED ITS\n * `tool_call_id`.\n *\n * Linking by id, not by adjacency. Adjacency emitted invalid documents in three\n * shapes the RFC forbids — an assistant turn with no `tool_calls` swallowing the\n * results that followed it, a result whose id matched none of the preceding\n * step's calls, and an unanswered result riding a `system` step that carried a\n * `source_call_id` it could never declare. Attaching only to the step that owns\n * the id also preserves message order exactly: a result either joins the step\n * immediately before it or becomes its own carrier at its own position.\n *\n * `line.steps` span projections are matched to agent steps in order (k-th llm\n * span → k-th agent step) purely to fill ATIF's native per-step `metrics`;\n * extra spans on either side are simply not matched, never invented.\n */\nfunction messagesToSteps(line: RolloutLine): HarborStep[] {\n  const steps: HarborStep[] = []\n  const llmSpans = (line.steps ?? []).filter((span) => span.kind === 'llm')\n  let nextSpan = 0\n  for (const message of line.messages) {\n    if (message.role === 'tool') {\n      const previous = steps[steps.length - 1]\n      if (previous !== undefined && declaresCall(previous, message.tool_call_id)) {\n        if (previous.observation === undefined) previous.observation = { results: [] }\n        previous.observation.results.push(toObservationResult(message, true))\n        continue\n      }\n      // A tool result answering no assistant turn in this document (truncated or\n      // reconstructed transcript). It still has to survive, so it rides a marked\n      // carrier step whose observation states no call id.\n      const carrier: HarborStep = {\n        step_id: steps.length + 1,\n        source: 'system',\n        message: '',\n        observation: { results: [toObservationResult(message, false)] },\n        extra: { [ESCROW]: { synthetic: TOOL_RESULTS_ONLY } },\n      }\n      if (message.is_copied_context === true) carrier.is_copied_context = true\n      steps.push(carrier)\n      continue\n    }\n    const step: HarborStep = {\n      step_id: steps.length + 1,\n      source: message.role === 'assistant' ? 'agent' : message.role,\n      message: message.content ?? '',\n    }\n    const escrow: Record<string, unknown> = {}\n    if (message.content === null) escrow.content_null = true\n    if (message.name !== undefined) escrow.name = message.name\n    if (message.role === 'assistant') {\n      if (line.policy.model !== null) step.model_name = line.policy.model\n      if (message.reasoning_content !== undefined) {\n        step.reasoning_content = message.reasoning_content\n      }\n      if (message.tool_calls !== undefined && message.tool_calls.length > 0) {\n        step.tool_calls = message.tool_calls.map(toHarborToolCall)\n      }\n      applySpanMetrics(step, llmSpans[nextSpan])\n      nextSpan += 1\n    } else if (message.reasoning_content !== undefined) {\n      // ATIF confines `reasoning_content` to agent steps; ours is not confined.\n      escrow.reasoning_content = message.reasoning_content\n    }\n    // RFC 0001 rule 7: a copied-context turn was not authored by this agent, and\n    // an SFT pipeline MUST exclude it. Native ATIF field, both directions.\n    if (message.is_copied_context === true) step.is_copied_context = true\n    if (Object.keys(escrow).length > 0) step.extra = { [ESCROW]: escrow }\n    steps.push(step)\n  }\n  return steps\n}\n\nfunction finalMetricsOf(line: RolloutLine, stepCount: number): HarborFinalMetrics {\n  const metrics: HarborFinalMetrics = {}\n  if (line.cost.tokens_in !== null) metrics.total_prompt_tokens = line.cost.tokens_in\n  if (line.cost.tokens_out !== null) metrics.total_completion_tokens = line.cost.tokens_out\n  if (line.cost.cache_read !== null) metrics.total_cached_tokens = line.cost.cache_read\n  // Omitted, never 0, when cost was not captured — a fake 0 is a lie about spend.\n  if (line.cost.usd !== null) metrics.total_cost_usd = line.cost.usd\n  metrics.total_steps = stepCount\n  // ATIF has no aggregate field for these three; `reasoning_tokens` under\n  // `extra` is the key the RFC's own worked example uses.\n  const extra: Record<string, unknown> = {}\n  if (line.cost.tokens_reasoning !== null) extra.reasoning_tokens = line.cost.tokens_reasoning\n  if (line.cost.cache_write !== null) extra.cache_write_tokens = line.cost.cache_write\n  if (line.cost.wall_s !== null) extra.wall_s = line.cost.wall_s\n  if (line.cost.llm_call_count !== undefined && line.cost.llm_call_count !== null) {\n    extra.llm_call_count = line.cost.llm_call_count\n  }\n  if (Object.keys(extra).length > 0) metrics.extra = extra\n  return metrics\n}\n\nfunction notesOf(line: RolloutLine): string | undefined {\n  const parts: string[] = []\n  if (line.provenance.gap !== undefined) parts.push(line.provenance.gap)\n  if (line.outcome.realness_gated === true) {\n    // ATIF has no gate field. A third-party consumer that ignores `extra`\n    // would otherwise see a gamed trajectory with nothing marking it, so the\n    // flag is stated in prose as well. This export carries no reward at all,\n    // so it is not a training-data door — but it is an audit record.\n    parts.push('realness-gated: this run faked its success signal (anti-Goodhart gate fired)')\n  }\n  return parts.length > 0 ? parts.join(' | ') : undefined\n}\n\nfunction toTrajectoryNode(line: RolloutLine): HarborTrajectory {\n  const steps = messagesToSteps(line)\n  const trajectory: HarborTrajectory = {\n    schema_version: ATIF_SCHEMA_VERSION,\n    // RUN-scoped, per the spec — so it is `run_id`, the coordinate every\n    // invocation of one run shares. `rollout_id` identifies a single\n    // invocation: using it gave two roots of the same run different session\n    // ids, and foreign tooling that groups by session_id would split the run.\n    session_id: line.run_id,\n    trajectory_id: line.rollout_id,\n    agent: {\n      // ATIF requires both; ours are nullable, so a null is substituted here\n      // and restored on import from the escrowed policy — never guessed back.\n      name: line.policy.harness ?? 'unknown',\n      version: line.policy.harness_version ?? '0.0.0',\n      ...(line.policy.model !== null ? { model_name: line.policy.model } : {}),\n      ...(line.tool_defs.length > 0 ? { tool_definitions: line.tool_defs } : {}),\n    },\n    steps,\n    final_metrics: finalMetricsOf(line, steps.length),\n    extra: {\n      [ESCROW]: {\n        schema: line.schema,\n        rollout_id: line.rollout_id,\n        parent_rollout_id: line.parent_rollout_id,\n        run_id: line.run_id,\n        ...(line.experiment_id !== undefined ? { experiment_id: line.experiment_id } : {}),\n        ...(line.candidate_id !== undefined ? { candidate_id: line.candidate_id } : {}),\n        generation: line.generation,\n        candidate_index: line.candidate_index,\n        role: line.role,\n        task: line.task,\n        policy: line.policy,\n        // Span projections overlap the same turns as `steps[]`; folding them\n        // into ATIF steps would double-count the run, so they stay escrowed.\n        ...(line.steps !== undefined ? { spans: line.steps } : {}),\n        outcome: {\n          // reward / reward_source / verdict are deliberately absent — see the\n          // module header. Only the non-scalar outcome fields travel.\n          metrics: line.outcome.metrics,\n          is_completed: line.outcome.is_completed,\n          is_truncated: line.outcome.is_truncated,\n          error: line.outcome.error,\n          ...(line.outcome.realness_gated !== undefined\n            ? { realness_gated: line.outcome.realness_gated }\n            : {}),\n        },\n        artifacts: line.artifacts,\n        provenance: line.provenance,\n      },\n    },\n  }\n  const notes = notesOf(line)\n  if (notes !== undefined) trajectory.notes = notes\n  return trajectory\n}\n\ninterface Forest {\n  roots: RolloutLine[]\n  childrenOf: Map<string, RolloutLine[]>\n}\n\nfunction buildForest(lines: RolloutLine[]): Forest {\n  const byId = new Map<string, RolloutLine>()\n  for (const line of lines) {\n    if (byId.has(line.rollout_id)) {\n      throw new Error(`duplicate rollout_id in input: ${line.rollout_id}`)\n    }\n    byId.set(line.rollout_id, line)\n  }\n  const roots: RolloutLine[] = []\n  const childrenOf = new Map<string, RolloutLine[]>()\n  for (const line of lines) {\n    const parent = line.parent_rollout_id\n    // A parent outside this set is a root OF WHAT WE HAVE: the edge is kept in\n    // escrow so a later import restores the dangling pointer instead of nulling it.\n    if (parent === null || !byId.has(parent)) {\n      roots.push(line)\n      continue\n    }\n    const siblings = childrenOf.get(parent)\n    if (siblings === undefined) childrenOf.set(parent, [line])\n    else siblings.push(line)\n  }\n  // Every line has a parent inside the set: the edges form a cycle with no\n  // root. Returning zero documents would drop the whole episode silently.\n  if (lines.length > 0 && roots.length === 0) {\n    throw new Error(\n      `parent_rollout_id cycle: no root among ${lines.length} lines (${lines.map((l) => l.rollout_id).join(', ')})`,\n    )\n  }\n  return { roots, childrenOf }\n}\n\nfunction assemble(line: RolloutLine, forest: Forest, onPath: Set<string>): HarborTrajectory {\n  if (onPath.has(line.rollout_id)) {\n    throw new Error(`parent_rollout_id cycle through rollout_id ${line.rollout_id}`)\n  }\n  onPath.add(line.rollout_id)\n  const node = toTrajectoryNode(line)\n  const children = forest.childrenOf.get(line.rollout_id)\n  if (children !== undefined && children.length > 0) {\n    node.subagent_trajectories = children.map((child) => assemble(child, forest, onPath))\n  }\n  onPath.delete(line.rollout_id)\n  return node\n}\n\n/**\n * Assemble one episode's flat lines into a single ATIF trajectory tree,\n * linked by `parent_rollout_id`.\n *\n * Reward, verdict and split are NOT emitted (ATIF models none of them); the\n * split and the rest of the task coordinates survive only in `extra.tangle`.\n *\n * We deliberately do NOT synthesize an `observation.subagent_trajectory_ref`\n * pointing at each child: our ledger records WHICH invocation spawned a\n * worker, not which STEP did, and attaching the ref to a guessed step would\n * fabricate a causal claim. Children are embedded in `subagent_trajectories`\n * (each with the `trajectory_id` the spec requires) and the edge is stated in\n * the child's escrowed `parent_rollout_id`.\n *\n * Throws when the lines are not one tree — use `toHarborTrajectories` for a forest.\n */\nexport function toHarborTrajectory(lines: RolloutLine[]): HarborTrajectory {\n  const trees = toHarborTrajectories(lines)\n  if (trees.length === 0) throw new Error('toHarborTrajectory: no lines')\n  if (trees.length > 1) {\n    const ids = trees.map((t) => t.trajectory_id ?? '?').join(', ')\n    throw new Error(\n      `toHarborTrajectory: ${trees.length} roots (${ids}) — ATIF is one tree per document; use toHarborTrajectories`,\n    )\n  }\n  return trees[0]!\n}\n\n/** Every independent tree in the input, one ATIF document each. */\nexport function toHarborTrajectories(lines: RolloutLine[]): HarborTrajectory[] {\n  const forest = buildForest(lines)\n  return forest.roots.map((root) => assemble(root, forest, new Set()))\n}\n\n// ---------------------------------------------------------------------------\n// Import: ATIF trajectory tree → RolloutLine[]\n// ---------------------------------------------------------------------------\n\nfunction contentToString(message: string | HarborContentPart[]): string {\n  if (typeof message === 'string') return message\n  return message\n    .map((part) =>\n      part.type === 'image'\n        ? // Our chat content is text-only. Describing the image is honest about\n          // what the source held; dropping it silently would not be.\n          `[image ${part.source?.media_type ?? 'unknown'} ${part.source?.path ?? ''}]`.trim()\n        : (part.text ?? ''),\n    )\n    .join('\\n')\n}\n\nfunction fromHarborToolCall(call: HarborToolCall): ChatToolCall {\n  const raw = asString(escrowOf(call.extra)?.arguments_raw)\n  return {\n    id: call.tool_call_id,\n    type: 'function',\n    function: { name: call.function_name, arguments: raw ?? JSON.stringify(call.arguments ?? {}) },\n  }\n}\n\nfunction fromObservationResult(\n  result: HarborObservationResult,\n  callId: string | undefined,\n): ChatMessage {\n  const escrow = escrowOf(result.extra)\n  return canonicalChatMessage({\n    role: 'tool',\n    content:\n      escrow?.content_null === true\n        ? null\n        : result.content === undefined\n          ? ''\n          : contentToString(result.content),\n    reasoning_content: asString(escrow?.reasoning_content),\n    tool_call_id: callId,\n    name: asString(escrow?.name),\n    is_copied_context: escrow?.is_copied_context === true,\n  })\n}\n\nfunction stepsToMessages(steps: HarborStep[]): ChatMessage[] {\n  const messages: ChatMessage[] = []\n  for (const step of steps) {\n    const escrow = escrowOf(step.extra)\n    const carrierOnly = escrow?.synthetic === TOOL_RESULTS_ONLY\n    if (!carrierOnly) {\n      messages.push(\n        canonicalChatMessage({\n          role: step.source === 'agent' ? 'assistant' : step.source,\n          content: escrow?.content_null === true ? null : contentToString(step.message),\n          reasoning_content: step.reasoning_content ?? asString(escrow?.reasoning_content),\n          tool_calls:\n            step.tool_calls !== undefined && step.tool_calls.length > 0\n              ? step.tool_calls.map(fromHarborToolCall)\n              : undefined,\n          name: asString(escrow?.name),\n          is_copied_context: step.is_copied_context === true,\n        }),\n      )\n    }\n    for (const result of step.observation?.results ?? []) {\n      const resultEscrow = escrowOf(result.extra)\n      const callId = result.source_call_id ?? asString(resultEscrow?.source_call_id)\n      if (carrierOnly) {\n        // Our own carrier: every result on it was a tool turn, and its call id\n        // (if the transcript had one) is in escrow, not on the wire.\n        messages.push(fromObservationResult(result, callId))\n        continue\n      }\n      if (typeof callId === 'string') {\n        messages.push(fromObservationResult(result, callId))\n        continue\n      }\n      // ATIF allows a result from a non-tool-calling action. Our chat schema\n      // requires `tool_call_id` on a tool turn, so inventing one would forge a\n      // link; the text is preserved as a system observation instead.\n      if (result.content !== undefined) {\n        messages.push(\n          canonicalChatMessage({ role: 'system', content: contentToString(result.content) }),\n        )\n      }\n    }\n  }\n  return messages\n}\n\nfunction taskFrom(escrow: Record<string, unknown> | undefined, fallbackId: string): RolloutTask {\n  const task = escrowSection(escrow, 'task')\n  return {\n    suite: asString(task?.suite) ?? 'harbor-atif-import',\n    instance_id: asString(task?.instance_id) ?? fallbackId,\n    // ALWAYS holdout — the escrowed claim is read for nothing here.\n    //\n    // `extra.tangle.task.split` is a namespaced key, not an authenticated one:\n    // a hand-written or third-party document can set `split: 'search'` as\n    // easily as our own exporter can, and honouring it made \"this file says so\"\n    // sufficient to walk into a training export. Trainability is a decision\n    // about a file, so it is made by an operator through\n    // `relabelImportedSplit`, never by the file about itself. The claim is not\n    // destroyed — it stays readable in the source document.\n    split: 'holdout',\n    seed: asNumberOrNull(task?.seed),\n    rep: asIntegerOrUndefined(task?.rep) ?? 0,\n  }\n}\n\nfunction policyFrom(\n  escrow: Record<string, unknown> | undefined,\n  agent: HarborAgent,\n): RolloutPolicy {\n  const escrowed = escrowSection(escrow, 'policy')\n  if (escrowed !== undefined) {\n    // Our own export: restore verbatim, including the nulls ATIF forced us to\n    // substitute placeholders for in `agent.name` / `agent.version`.\n    return {\n      harness: asString(escrowed.harness) ?? null,\n      harness_version: asString(escrowed.harness_version) ?? null,\n      model: asString(escrowed.model) ?? null,\n      provider: asString(escrowed.provider) ?? null,\n      profile_commit: asString(escrowed.profile_commit) ?? null,\n      ...(escrowed.prompt_hash !== undefined\n        ? { prompt_hash: asString(escrowed.prompt_hash) ?? null }\n        : {}),\n      ...(escrowed.config_hash !== undefined\n        ? { config_hash: asString(escrowed.config_hash) ?? null }\n        : {}),\n      ...(escrowed.agent_profile_cell_id !== undefined\n        ? { agent_profile_cell_id: asString(escrowed.agent_profile_cell_id) ?? null }\n        : {}),\n      sampling: isRecord(escrowed.sampling) ? escrowed.sampling : null,\n    }\n  }\n  return {\n    harness: agent.name,\n    harness_version: agent.version,\n    model: agent.model_name ?? null,\n    provider: null,\n    profile_commit: null,\n    sampling: null,\n  }\n}\n\nfunction artifactsFrom(escrow: Record<string, unknown> | undefined): RolloutArtifacts {\n  const artifacts = escrowSection(escrow, 'artifacts')\n  return {\n    patch_path: asString(artifacts?.patch_path) ?? null,\n    run_dir: asString(artifacts?.run_dir) ?? null,\n    transcript_ref: asString(artifacts?.transcript_ref) ?? null,\n  }\n}\n\nfunction costFrom(final: HarborFinalMetrics | undefined): RolloutCostBlock {\n  const rawExtra = final?.extra\n  const extra = isRecord(rawExtra) ? rawExtra : undefined\n  const calls = asIntegerOrUndefined(extra?.llm_call_count)\n  return {\n    usd: asNumberOrNull(final?.total_cost_usd),\n    tokens_in: asNumberOrNull(final?.total_prompt_tokens),\n    tokens_out: asNumberOrNull(final?.total_completion_tokens),\n    tokens_reasoning: asNumberOrNull(extra?.reasoning_tokens),\n    cache_read: asNumberOrNull(final?.total_cached_tokens),\n    cache_write: asNumberOrNull(extra?.cache_write_tokens),\n    wall_s: asNumberOrNull(extra?.wall_s),\n    ...(calls !== undefined ? { llm_call_count: calls } : {}),\n  }\n}\n\n/**\n * Recovers span projections from ATIF's NATIVE per-step channel, for documents\n * with no `extra.tangle.spans` escrow — i.e. everything a foreign producer\n * writes. Without this the logprobs and token ids a Harbor-native trainer\n * records would be read, validated, and then dropped on the floor.\n *\n * Only steps that actually carry one of the four fields produce a span: an\n * agent step with no metrics means \"not captured\", and inventing an empty span\n * for it would claim the run had a shape we did not observe.\n */\nfunction spansFromSteps(steps: HarborStep[]): RolloutStep[] {\n  const spans: RolloutStep[] = []\n  for (const step of steps) {\n    if (step.source !== 'agent') continue\n    const calls = asIntegerOrUndefined(step.llm_call_count)\n    const promptIds = asIntegerArray(step.metrics?.prompt_token_ids)\n    const completionIds = asIntegerArray(step.metrics?.completion_token_ids)\n    const logprobs = asNumberArray(step.metrics?.logprobs)\n    if (\n      calls === undefined &&\n      promptIds === undefined &&\n      completionIds === undefined &&\n      logprobs === undefined\n    ) {\n      continue\n    }\n    spans.push({\n      kind: 'llm',\n      name: step.model_name ?? 'chat',\n      ...(calls !== undefined ? { llm_call_count: calls } : {}),\n      ...(promptIds !== undefined ? { prompt_token_ids: promptIds } : {}),\n      ...(completionIds !== undefined ? { completion_token_ids: completionIds } : {}),\n      ...(logprobs !== undefined ? { logprobs } : {}),\n    })\n  }\n  return spans\n}\n\n/**\n * Composes `provenance.gap` as an ordered SET of reasons.\n *\n * Appending made the note accrete on every pass (\"…no verdict | …no verdict\"),\n * which both grows without bound and breaks byte-level idempotency for a ledger\n * hashed on its serialized lines.\n */\nfunction composeGap(escrowedGap: string | undefined, added: readonly string[]): string {\n  const parts: string[] = []\n  const seen = new Set<string>()\n  for (const raw of [...(escrowedGap?.split(' | ') ?? []), ...added]) {\n    const reason = raw.trim()\n    if (reason.length === 0 || seen.has(reason)) continue\n    seen.add(reason)\n    parts.push(reason)\n  }\n  return parts.join(' | ')\n}\n\nexport interface FromHarborOptions {\n  /** Injected clock for deterministic output when the source carries no capture time. */\n  now?: () => Date\n}\n\nfunction nodeToLine(\n  trajectory: HarborTrajectory,\n  parentId: string | null,\n  capturedAt: string,\n): RolloutLine {\n  const escrow = escrowOf(trajectory.extra)\n  const rolloutId =\n    asString(escrow?.rollout_id) ?? trajectory.trajectory_id ?? trajectory.session_id\n  if (rolloutId === undefined || rolloutId.length === 0) {\n    throw new Error(\n      'fromHarborTrajectory: trajectory has no trajectory_id or session_id — cannot mint a joinable rollout_id without inventing one',\n    )\n  }\n  const outcome = escrowSection(escrow, 'outcome')\n  const provenance = escrowSection(escrow, 'provenance')\n  const capture = provenance?.capture\n  const escrowedGap = asString(provenance?.gap)\n  const role = escrow?.role\n  const escrowedSpans = escrow?.spans\n  // The escrow wins when it exists (our own export: exact, including an empty\n  // array, which claims \"captured, none\" rather than \"not captured\"), and the\n  // native per-step channel is the fallback (a foreign export: still signal).\n  const nativeSpans = spansFromSteps(trajectory.steps)\n  const spans: RolloutStep[] | undefined = Array.isArray(escrowedSpans)\n    ? (escrowedSpans as RolloutStep[])\n    : nativeSpans.length > 0\n      ? nativeSpans\n      : undefined\n  const messages = stepsToMessages(trajectory.steps)\n  const line: RolloutLine = {\n    schema: ROLLOUT_SCHEMA,\n    rollout_id: rolloutId,\n    // The tree edge wins for embedded children; for a root, the escrowed\n    // pointer may reference a line outside this document and is preserved.\n    parent_rollout_id: parentId ?? asString(escrow?.parent_rollout_id) ?? null,\n    run_id: asString(escrow?.run_id) ?? trajectory.session_id ?? rolloutId,\n    // Required keys on the wire (0.127.0): a foreign document that never\n    // stated them imports as explicit `null`, not as an absent field.\n    experiment_id: asString(escrow?.experiment_id) ?? null,\n    candidate_id: asString(escrow?.candidate_id) ?? null,\n    generation: asIntegerOrUndefined(escrow?.generation) ?? null,\n    candidate_index: asIntegerOrUndefined(escrow?.candidate_index) ?? null,\n    role: ROLLOUT_ROLES.includes(role as RolloutRole) ? (role as RolloutRole) : 'agent',\n    task: taskFrom(escrow, rolloutId),\n    policy: policyFrom(escrow, trajectory.agent),\n    messages,\n    tool_defs: trajectory.agent.tool_definitions ?? [],\n    ...(spans !== undefined ? { steps: spans } : {}),\n    outcome: {\n      // ATIF carries no label. Null is the labeled gap; a 0 here would be a\n      // fabricated failure and a 1 a fabricated success.\n      reward: null,\n      reward_source: null,\n      verdict: null,\n      metrics: isRecord(outcome?.metrics) ? outcome.metrics : {},\n      is_completed: typeof outcome?.is_completed === 'boolean' ? outcome.is_completed : true,\n      is_truncated:\n        typeof outcome?.is_truncated === 'boolean'\n          ? outcome.is_truncated\n          : trajectory.continued_trajectory_ref !== undefined,\n      error: asString(outcome?.error) ?? null,\n      // The anti-Goodhart flag is restored when the source document stated\n      // it; the wire schema requires the field, so a document that never did\n      // imports as `false`. That is safe here because the reward is already\n      // forced to `null` (not trainable) and `realness_screened` stays\n      // absent = unknown rather than claiming a screen ran.\n      realness_gated: outcome?.realness_gated === true,\n    },\n    cost: costFrom(trajectory.final_metrics),\n    artifacts: artifactsFrom(escrow),\n    provenance: {\n      captured_at: asString(provenance?.captured_at) ?? capturedAt,\n      capture: ROLLOUT_CAPTURES.includes(capture as RolloutCapture)\n        ? (capture as RolloutCapture)\n        : 'backfill',\n      gap: composeGap(escrowedGap, [HARBOR_IMPORT_GAP]),\n      // Restored for the same reason `realness_gated` is, and with the opposite\n      // risk profile from the reward: this is the gated run's own measurement\n      // bag, moved off `outcome` by `gateGamedOutcome` so no exporter reads it\n      // as training input. Dropping it here would silently destroy the audit\n      // trail that says WHY the run was flagged and what it claimed, on the one\n      // population an auditor most wants to inspect. It cannot be fail-open —\n      // nothing projects `provenance` into a training row.\n      ...(isRecord(provenance?.gated_evidence)\n        ? { gated_evidence: provenance.gated_evidence as GatedEvidence }\n        : {}),\n    },\n  }\n  assertRolloutLine(line, `rollout line imported from ATIF trajectory ${rolloutId}`)\n  return line\n}\n\n/**\n * Flatten an ATIF trajectory tree back into `tangle.rollout.v1` lines, parent\n * first, each child carrying `parent_rollout_id`.\n *\n * Every line comes back UNLABELED: `reward`, `reward_source` and `verdict` are\n * null and `provenance.gap` says why. ATIF models no verdict, so scoring an\n * imported trajectory is a judge's job, not this function's. Every line lands\n * on `holdout` whatever the document claims — see `relabelImportedSplit`.\n */\nexport function fromHarborTrajectory(\n  trajectory: HarborTrajectory,\n  options: FromHarborOptions = {},\n): RolloutLine[] {\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const lines: RolloutLine[] = []\n  const walk = (node: HarborTrajectory, parentId: string | null, onPath: Set<string>): void => {\n    const line = nodeToLine(node, parentId, capturedAt)\n    if (onPath.has(line.rollout_id)) {\n      throw new Error(`subagent_trajectories cycle through trajectory_id ${line.rollout_id}`)\n    }\n    onPath.add(line.rollout_id)\n    lines.push(line)\n    for (const child of node.subagent_trajectories ?? []) {\n      walk(child, line.rollout_id, onPath)\n    }\n    onPath.delete(line.rollout_id)\n  }\n  walk(trajectory, null, new Set())\n  return lines\n}\n\n/**\n * THE explicit door out of `holdout` for imported lines.\n *\n * Import forces `holdout` because a document's own claim about its split is not\n * evidence — anyone can write `extra.tangle.task.split`. Promoting a file to a\n * trainable split is an operator's decision about provenance they verified, so\n * it is a separate, greppable call: `grep relabelImportedSplit` enumerates\n * every place foreign data was declared trainable, which is exactly the audit\n * the trusted-escrow version made impossible.\n *\n * Returns plain `RolloutLine`s. They still have to pass `assertMinted` (and its\n * anti-Goodhart check) to reach an exporter — re-labeling a split is not\n * minting a reward.\n */\nexport function relabelImportedSplit(\n  lines: readonly RolloutLine[],\n  split: RolloutSplit,\n): RolloutLine[] {\n  if (!ROLLOUT_SPLITS.includes(split)) {\n    throw new Error(`relabelImportedSplit: unknown split ${String(split)}`)\n  }\n  return lines.map((line) => ({ ...line, task: { ...line.task, split } }))\n}\n","/**\n * Read-only backfill reader over the opencode sqlite store\n * (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.\n *\n * Schema consumed (observed, 2026-07): `session` rows carry directory /\n * parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON\n * `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`\n * rows carry the actual content ({type: text|reasoning|tool|step-start|\n * step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,\n * status}} — both the call and its result, which we split into an assistant\n * tool_call plus a role:\"tool\" result message.\n *\n * The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships\n * next to it in the wild), so `openOpencodeDb` returns null instead of\n * throwing — callers record a gap line, never crash the backfill.\n */\n\nimport { createRequire } from 'node:module'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { DatabaseSync } from 'node:sqlite'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_OPENCODE_DB = join(homedir(), '.local', 'share', 'opencode', 'opencode.db')\n\nexport interface OpencodeSessionRow {\n  id: string\n  parentId: string | null\n  directory: string\n  agent: string | null\n  /** Raw session.model JSON: {id, providerID, variant} where present. */\n  model: { id?: string; providerID?: string } | null\n  costUsd: number\n  tokensInput: number\n  tokensOutput: number\n  tokensReasoning: number\n  tokensCacheRead: number\n  tokensCacheWrite: number\n  timeCreated: number\n  timeUpdated: number\n}\n\n// `node:sqlite` is loaded through CommonJS `require`, not `import()`. esbuild\n// (bundling) and Vite (tests) both rewrite a dynamic import and strip the\n// `node:` prefix under an es20xx target, turning this builtin into a bogus\n// \"sqlite\" package lookup; composing the specifier at runtime does not reliably\n// defeat that (it still resolved through Vite's transform in some workers, so\n// the failure moved around as test files were added). A require obtained from\n// `createRequire` is not an analyzable module reference in either tool, so\n// neither can rewrite it.\n/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */\nexport async function openOpencodeDb(\n  path: string = DEFAULT_OPENCODE_DB,\n): Promise<DatabaseSync | null> {\n  try {\n    // Keep Node-only module initialization inside the Node-only operation.\n    // Root imports are shared with edge consumers that do not define import.meta.url.\n    const nodeRequire = createRequire(import.meta.url)\n    const { DatabaseSync } = nodeRequire('node:sqlite') as typeof import('node:sqlite')\n    const db = new DatabaseSync(path, { readOnly: true })\n    // Probe: a corrupt store can open() fine and fail on first page read.\n    db.prepare('SELECT id FROM session LIMIT 1').get()\n    return db\n  } catch {\n    return null\n  }\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nfunction parseSessionRow(row: Record<string, unknown>): OpencodeSessionRow {\n  let model: OpencodeSessionRow['model'] = null\n  if (typeof row.model === 'string' && row.model.length > 0) {\n    try {\n      const parsed: unknown = JSON.parse(row.model)\n      if (isRecord(parsed)) model = parsed as { id?: string; providerID?: string }\n    } catch {\n      model = null\n    }\n  }\n  return {\n    id: String(row.id),\n    parentId: row.parent_id === null || row.parent_id === undefined ? null : String(row.parent_id),\n    directory: String(row.directory),\n    agent: row.agent === null || row.agent === undefined ? null : String(row.agent),\n    model,\n    costUsd: Number(row.cost ?? 0),\n    tokensInput: Number(row.tokens_input ?? 0),\n    tokensOutput: Number(row.tokens_output ?? 0),\n    tokensReasoning: Number(row.tokens_reasoning ?? 0),\n    tokensCacheRead: Number(row.tokens_cache_read ?? 0),\n    tokensCacheWrite: Number(row.tokens_cache_write ?? 0),\n    timeCreated: Number(row.time_created ?? 0),\n    timeUpdated: Number(row.time_updated ?? 0),\n  }\n}\n\nconst SESSION_COLUMNS =\n  'id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated'\n\n/** Sessions whose cwd is `directory` (the worker-clone join key). */\nexport function findOpencodeSessionsByDirectory(\n  db: DatabaseSync,\n  directory: string,\n): OpencodeSessionRow[] {\n  const rows = db\n    .prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`)\n    .all(directory) as Array<Record<string, unknown>>\n  return rows.map(parseSessionRow)\n}\n\ninterface OpencodePart {\n  type?: string\n  text?: string\n  tool?: string\n  callID?: string\n  state?: { status?: string; input?: unknown; output?: unknown }\n}\n\nfunction toolResultContent(output: unknown): string {\n  if (typeof output === 'string') return output\n  if (output === null || output === undefined) return ''\n  return JSON.stringify(output)\n}\n\n/**\n * Convert one session's message+part rows into canonical messages.\n * An opencode assistant message row spans several model steps; each step's\n * parts (reasoning → text → tool …) become one assistant message followed by\n * the role:\"tool\" results of its calls, preserving order.\n */\nexport function readOpencodeSessionMessages(db: DatabaseSync, sessionId: string): ChatMessage[] {\n  const messageRows = db\n    .prepare('SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id')\n    .all(sessionId) as Array<{ id: string; data: string }>\n  const partsStmt = db.prepare('SELECT data FROM part WHERE message_id = ? ORDER BY id')\n\n  const messages: ChatMessage[] = []\n  for (const messageRow of messageRows) {\n    let data: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(messageRow.data)\n      if (!isRecord(parsed)) continue\n      data = parsed\n    } catch {\n      continue\n    }\n    const parts: OpencodePart[] = []\n    for (const row of partsStmt.all(messageRow.id) as Array<{ data: string }>) {\n      try {\n        const parsed: unknown = JSON.parse(row.data)\n        if (isRecord(parsed)) parts.push(parsed as OpencodePart)\n      } catch {\n        // Malformed part payload: skip the part, keep the message.\n      }\n    }\n\n    if (data.role === 'user') {\n      const text = parts\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      messages.push({ role: 'user', content: text })\n      continue\n    }\n    if (data.role !== 'assistant') continue\n\n    // Split the row into steps at step-start boundaries; parts before the\n    // first step-start (none observed, but tolerated) form an implicit step.\n    const steps: OpencodePart[][] = []\n    let current: OpencodePart[] = []\n    for (const part of parts) {\n      if (part.type === 'step-start') {\n        if (current.length > 0) steps.push(current)\n        current = []\n        continue\n      }\n      if (part.type === 'step-finish' || part.type === 'snapshot' || part.type === 'patch') continue\n      current.push(part)\n    }\n    if (current.length > 0) steps.push(current)\n\n    for (const step of steps) {\n      const reasoning = step\n        .filter((p) => p.type === 'reasoning' && typeof p.text === 'string' && p.text.length > 0)\n        .map((p) => p.text as string)\n        .join('\\n')\n      const text = step\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      const toolParts = step.filter((p) => p.type === 'tool' && typeof p.callID === 'string')\n      const toolCalls: ChatToolCall[] = toolParts.map((p) => ({\n        id: p.callID as string,\n        type: 'function',\n        function: {\n          name: p.tool ?? 'unknown',\n          arguments: JSON.stringify(p.state?.input ?? {}),\n        },\n      }))\n      if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n      messages.push({\n        role: 'assistant',\n        content: text.length > 0 ? text : null,\n        ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n        ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n      })\n      for (const p of toolParts) {\n        messages.push({\n          role: 'tool',\n          tool_call_id: p.callID as string,\n          name: p.tool ?? 'unknown',\n          content: toolResultContent(p.state?.output),\n        })\n      }\n    }\n  }\n  return messages\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiFA,MAAa,sBAAsB;;AAGnC,MAAa,oBAAoB;;AAGjC,MAAM,SAAS;;;;;;;;;;;;;;AAef,MAAM,oBAAoB;AA8G1B,MAAMA,cAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,MAAM,YAAY,MAAoC,OAAO,MAAM,WAAW,IAAI,KAAA;AAElF,MAAM,kBAAkB,MACtB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAEpD,MAAM,wBAAwB,MAC5B,OAAO,UAAU,CAAC,IAAK,IAAe,KAAA;AAExC,MAAM,iBAAiB,MACrB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC,IACzE,IACD,KAAA;AAEN,MAAM,kBAAkB,MACtB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,UAAU,CAAC,CAAC,IAAK,IAAiB,KAAA;;;;;;;;;AAU9E,SAAS,qBAAqB,OAQd;CACd,MAAM,UAAuB;EAAE,MAAM,MAAM;EAAM,SAAS,MAAM;CAAQ;CACxE,IAAI,MAAM,sBAAsB,KAAA,GAAW,QAAQ,oBAAoB,MAAM;CAC7E,IAAI,MAAM,eAAe,KAAA,GAAW,QAAQ,aAAa,MAAM;CAC/D,IAAI,MAAM,iBAAiB,KAAA,GAAW,QAAQ,eAAe,MAAM;CACnE,IAAI,MAAM,SAAS,KAAA,GAAW,QAAQ,OAAO,MAAM;CACnD,IAAI,MAAM,sBAAsB,MAAM,QAAQ,oBAAoB;CAClE,OAAO;AACT;;AAGA,SAAS,SAAS,OAAiF;CACjG,IAAI,CAACA,WAAS,KAAK,GAAG,OAAO,KAAA;CAC7B,MAAM,QAAQ,MAAM;CACpB,OAAOA,WAAS,KAAK,IAAI,QAAQ,KAAA;AACnC;AAEA,SAAS,cACP,QACA,KACqC;CACrC,IAAI,WAAW,KAAA,GAAW,OAAO,KAAA;CACjC,MAAM,QAAQ,OAAO;CACrB,OAAOA,WAAS,KAAK,IAAI,QAAQ,KAAA;AACnC;AAMA,SAAS,iBAAiB,MAAoC;CAK5D,IAAI;CACJ,IAAI;EACF,MAAM,UAAmB,KAAK,MAAM,KAAK,SAAS,SAAS;EAC3D,IAAIA,WAAS,OAAO,GAAG,SAAS;CAClC,QAAQ;EACN,SAAS,KAAA;CACX;CACA,OAAO;EACL,cAAc,KAAK;EACnB,eAAe,KAAK,SAAS;EAC7B,WAAW,UAAU,CAAC;EACtB,OAAO,GAAG,SAAS,EAAE,eAAe,KAAK,SAAS,UAAU,EAAE;CAChE;AACF;;;;;;AAOA,SAAS,oBAAoB,SAAsB,QAA0C;CAC3F,MAAM,SAAkC,CAAC;CACzC,IAAI,UAAU,QAAQ,iBAAiB,KAAA,GAAW,OAAO,iBAAiB,QAAQ;CAClF,IAAI,QAAQ,YAAY,MAAM,OAAO,UAAU,QAAQ;CACvD,MAAM,SAAkC,CAAC;CACzC,IAAI,CAAC,UAAU,QAAQ,iBAAiB,KAAA,GAAW,OAAO,iBAAiB,QAAQ;CACnF,IAAI,QAAQ,SAAS,KAAA,GAAW,OAAO,OAAO,QAAQ;CAGtD,IAAI,QAAQ,YAAY,MAAM,OAAO,eAAe;CACpD,IAAI,QAAQ,sBAAsB,KAAA,GAAW,OAAO,oBAAoB,QAAQ;CAChF,IAAI,QAAQ,sBAAsB,MAAM,OAAO,oBAAoB;CACnE,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC,SAAS,GAAG,OAAO,QAAQ,GAAG,SAAS,OAAO;CACtE,OAAO;AACT;;AAGA,SAAS,aAAa,MAAkB,QAAqC;CAC3E,IAAI,WAAW,KAAA,GAAW,OAAO;CACjC,OAAO,KAAK,YAAY,MAAM,SAAS,KAAK,iBAAiB,MAAM,MAAM;AAC3E;;;;;;;;;AAUA,SAAS,iBAAiB,MAAkB,MAAqC;CAC/E,IAAI,SAAS,KAAA,GAAW;CACxB,IAAI,KAAK,mBAAmB,KAAA,GAAW,KAAK,iBAAiB,KAAK;CAClE,MAAM,UAAyB,CAAC;CAChC,IAAI,KAAK,qBAAqB,KAAA,GAAW,QAAQ,mBAAmB,KAAK;CACzE,IAAI,KAAK,yBAAyB,KAAA,GAChC,QAAQ,uBAAuB,KAAK;CAEtC,IAAI,KAAK,aAAa,KAAA,GAAW,QAAQ,WAAW,KAAK;CACzD,IAAI,OAAO,KAAK,OAAO,CAAC,CAAC,SAAS,GAAG,KAAK,UAAU;AACtD;;;;;;;;;;;;;;;;;;AAmBA,SAAS,gBAAgB,MAAiC;CACxD,MAAM,QAAsB,CAAC;CAC7B,MAAM,YAAY,KAAK,SAAS,CAAC,EAAA,CAAG,QAAQ,SAAS,KAAK,SAAS,KAAK;CACxE,IAAI,WAAW;CACf,KAAK,MAAM,WAAW,KAAK,UAAU;EACnC,IAAI,QAAQ,SAAS,QAAQ;GAC3B,MAAM,WAAW,MAAM,MAAM,SAAS;GACtC,IAAI,aAAa,KAAA,KAAa,aAAa,UAAU,QAAQ,YAAY,GAAG;IAC1E,IAAI,SAAS,gBAAgB,KAAA,GAAW,SAAS,cAAc,EAAE,SAAS,CAAC,EAAE;IAC7E,SAAS,YAAY,QAAQ,KAAK,oBAAoB,SAAS,IAAI,CAAC;IACpE;GACF;GAIA,MAAM,UAAsB;IAC1B,SAAS,MAAM,SAAS;IACxB,QAAQ;IACR,SAAS;IACT,aAAa,EAAE,SAAS,CAAC,oBAAoB,SAAS,KAAK,CAAC,EAAE;IAC9D,OAAO,GAAG,SAAS,EAAE,WAAW,kBAAkB,EAAE;GACtD;GACA,IAAI,QAAQ,sBAAsB,MAAM,QAAQ,oBAAoB;GACpE,MAAM,KAAK,OAAO;GAClB;EACF;EACA,MAAM,OAAmB;GACvB,SAAS,MAAM,SAAS;GACxB,QAAQ,QAAQ,SAAS,cAAc,UAAU,QAAQ;GACzD,SAAS,QAAQ,WAAW;EAC9B;EACA,MAAM,SAAkC,CAAC;EACzC,IAAI,QAAQ,YAAY,MAAM,OAAO,eAAe;EACpD,IAAI,QAAQ,SAAS,KAAA,GAAW,OAAO,OAAO,QAAQ;EACtD,IAAI,QAAQ,SAAS,aAAa;GAChC,IAAI,KAAK,OAAO,UAAU,MAAM,KAAK,aAAa,KAAK,OAAO;GAC9D,IAAI,QAAQ,sBAAsB,KAAA,GAChC,KAAK,oBAAoB,QAAQ;GAEnC,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,WAAW,SAAS,GAClE,KAAK,aAAa,QAAQ,WAAW,IAAI,gBAAgB;GAE3D,iBAAiB,MAAM,SAAS,SAAS;GACzC,YAAY;EACd,OAAO,IAAI,QAAQ,sBAAsB,KAAA,GAEvC,OAAO,oBAAoB,QAAQ;EAIrC,IAAI,QAAQ,sBAAsB,MAAM,KAAK,oBAAoB;EACjE,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC,SAAS,GAAG,KAAK,QAAQ,GAAG,SAAS,OAAO;EACpE,MAAM,KAAK,IAAI;CACjB;CACA,OAAO;AACT;AAEA,SAAS,eAAe,MAAmB,WAAuC;CAChF,MAAM,UAA8B,CAAC;CACrC,IAAI,KAAK,KAAK,cAAc,MAAM,QAAQ,sBAAsB,KAAK,KAAK;CAC1E,IAAI,KAAK,KAAK,eAAe,MAAM,QAAQ,0BAA0B,KAAK,KAAK;CAC/E,IAAI,KAAK,KAAK,eAAe,MAAM,QAAQ,sBAAsB,KAAK,KAAK;CAE3E,IAAI,KAAK,KAAK,QAAQ,MAAM,QAAQ,iBAAiB,KAAK,KAAK;CAC/D,QAAQ,cAAc;CAGtB,MAAM,QAAiC,CAAC;CACxC,IAAI,KAAK,KAAK,qBAAqB,MAAM,MAAM,mBAAmB,KAAK,KAAK;CAC5E,IAAI,KAAK,KAAK,gBAAgB,MAAM,MAAM,qBAAqB,KAAK,KAAK;CACzE,IAAI,KAAK,KAAK,WAAW,MAAM,MAAM,SAAS,KAAK,KAAK;CACxD,IAAI,KAAK,KAAK,mBAAmB,KAAA,KAAa,KAAK,KAAK,mBAAmB,MACzE,MAAM,iBAAiB,KAAK,KAAK;CAEnC,IAAI,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS,GAAG,QAAQ,QAAQ;CACnD,OAAO;AACT;AAEA,SAAS,QAAQ,MAAuC;CACtD,MAAM,QAAkB,CAAC;CACzB,IAAI,KAAK,WAAW,QAAQ,KAAA,GAAW,MAAM,KAAK,KAAK,WAAW,GAAG;CACrE,IAAI,KAAK,QAAQ,mBAAmB,MAKlC,MAAM,KAAK,8EAA8E;CAE3F,OAAO,MAAM,SAAS,IAAI,MAAM,KAAK,KAAK,IAAI,KAAA;AAChD;AAEA,SAAS,iBAAiB,MAAqC;CAC7D,MAAM,QAAQ,gBAAgB,IAAI;CAClC,MAAM,aAA+B;EACnC,gBAAgB;EAKhB,YAAY,KAAK;EACjB,eAAe,KAAK;EACpB,OAAO;GAGL,MAAM,KAAK,OAAO,WAAW;GAC7B,SAAS,KAAK,OAAO,mBAAmB;GACxC,GAAI,KAAK,OAAO,UAAU,OAAO,EAAE,YAAY,KAAK,OAAO,MAAM,IAAI,CAAC;GACtE,GAAI,KAAK,UAAU,SAAS,IAAI,EAAE,kBAAkB,KAAK,UAAU,IAAI,CAAC;EAC1E;EACA;EACA,eAAe,eAAe,MAAM,MAAM,MAAM;EAChD,OAAO,GACJ,SAAS;GACR,QAAQ,KAAK;GACb,YAAY,KAAK;GACjB,mBAAmB,KAAK;GACxB,QAAQ,KAAK;GACb,GAAI,KAAK,kBAAkB,KAAA,IAAY,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;GAChF,GAAI,KAAK,iBAAiB,KAAA,IAAY,EAAE,cAAc,KAAK,aAAa,IAAI,CAAC;GAC7E,YAAY,KAAK;GACjB,iBAAiB,KAAK;GACtB,MAAM,KAAK;GACX,MAAM,KAAK;GACX,QAAQ,KAAK;GAGb,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,SAAS;IAGP,SAAS,KAAK,QAAQ;IACtB,cAAc,KAAK,QAAQ;IAC3B,cAAc,KAAK,QAAQ;IAC3B,OAAO,KAAK,QAAQ;IACpB,GAAI,KAAK,QAAQ,mBAAmB,KAAA,IAChC,EAAE,gBAAgB,KAAK,QAAQ,eAAe,IAC9C,CAAC;GACP;GACA,WAAW,KAAK;GAChB,YAAY,KAAK;EACnB,EACF;CACF;CACA,MAAM,QAAQ,QAAQ,IAAI;CAC1B,IAAI,UAAU,KAAA,GAAW,WAAW,QAAQ;CAC5C,OAAO;AACT;AAOA,SAAS,YAAY,OAA8B;CACjD,MAAM,uBAAO,IAAI,IAAyB;CAC1C,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,IAAI,KAAK,UAAU,GAC1B,MAAM,IAAI,MAAM,kCAAkC,KAAK,YAAY;EAErE,KAAK,IAAI,KAAK,YAAY,IAAI;CAChC;CACA,MAAM,QAAuB,CAAC;CAC9B,MAAM,6BAAa,IAAI,IAA2B;CAClD,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SAAS,KAAK;EAGpB,IAAI,WAAW,QAAQ,CAAC,KAAK,IAAI,MAAM,GAAG;GACxC,MAAM,KAAK,IAAI;GACf;EACF;EACA,MAAM,WAAW,WAAW,IAAI,MAAM;EACtC,IAAI,aAAa,KAAA,GAAW,WAAW,IAAI,QAAQ,CAAC,IAAI,CAAC;OACpD,SAAS,KAAK,IAAI;CACzB;CAGA,IAAI,MAAM,SAAS,KAAK,MAAM,WAAW,GACvC,MAAM,IAAI,MACR,0CAA0C,MAAM,OAAO,UAAU,MAAM,KAAK,MAAM,EAAE,UAAU,CAAC,CAAC,KAAK,IAAI,EAAE,EAC7G;CAEF,OAAO;EAAE;EAAO;CAAW;AAC7B;AAEA,SAAS,SAAS,MAAmB,QAAgB,QAAuC;CAC1F,IAAI,OAAO,IAAI,KAAK,UAAU,GAC5B,MAAM,IAAI,MAAM,8CAA8C,KAAK,YAAY;CAEjF,OAAO,IAAI,KAAK,UAAU;CAC1B,MAAM,OAAO,iBAAiB,IAAI;CAClC,MAAM,WAAW,OAAO,WAAW,IAAI,KAAK,UAAU;CACtD,IAAI,aAAa,KAAA,KAAa,SAAS,SAAS,GAC9C,KAAK,wBAAwB,SAAS,KAAK,UAAU,SAAS,OAAO,QAAQ,MAAM,CAAC;CAEtF,OAAO,OAAO,KAAK,UAAU;CAC7B,OAAO;AACT;;;;;;;;;;;;;;;;;AAkBA,SAAgB,mBAAmB,OAAwC;CACzE,MAAM,QAAQ,qBAAqB,KAAK;CACxC,IAAI,MAAM,WAAW,GAAG,MAAM,IAAI,MAAM,8BAA8B;CACtE,IAAI,MAAM,SAAS,GAAG;EACpB,MAAM,MAAM,MAAM,KAAK,MAAM,EAAE,iBAAiB,GAAG,CAAC,CAAC,KAAK,IAAI;EAC9D,MAAM,IAAI,MACR,uBAAuB,MAAM,OAAO,UAAU,IAAI,4DACpD;CACF;CACA,OAAO,MAAM;AACf;;AAGA,SAAgB,qBAAqB,OAA0C;CAC7E,MAAM,SAAS,YAAY,KAAK;CAChC,OAAO,OAAO,MAAM,KAAK,SAAS,SAAS,MAAM,wBAAQ,IAAI,IAAI,CAAC,CAAC;AACrE;AAMA,SAAS,gBAAgB,SAA+C;CACtE,IAAI,OAAO,YAAY,UAAU,OAAO;CACxC,OAAO,QACJ,KAAK,SACJ,KAAK,SAAS,UAGV,UAAU,KAAK,QAAQ,cAAc,UAAU,GAAG,KAAK,QAAQ,QAAQ,GAAG,GAAG,KAAK,IACjF,KAAK,QAAQ,EACpB,CAAC,CACA,KAAK,IAAI;AACd;AAEA,SAAS,mBAAmB,MAAoC;CAC9D,MAAM,MAAM,SAAS,SAAS,KAAK,KAAK,CAAC,EAAE,aAAa;CACxD,OAAO;EACL,IAAI,KAAK;EACT,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAe,WAAW,OAAO,KAAK,UAAU,KAAK,aAAa,CAAC,CAAC;EAAE;CAC/F;AACF;AAEA,SAAS,sBACP,QACA,QACa;CACb,MAAM,SAAS,SAAS,OAAO,KAAK;CACpC,OAAO,qBAAqB;EAC1B,MAAM;EACN,SACE,QAAQ,iBAAiB,OACrB,OACA,OAAO,YAAY,KAAA,IACjB,KACA,gBAAgB,OAAO,OAAO;EACtC,mBAAmB,SAAS,QAAQ,iBAAiB;EACrD,cAAc;EACd,MAAM,SAAS,QAAQ,IAAI;EAC3B,mBAAmB,QAAQ,sBAAsB;CACnD,CAAC;AACH;AAEA,SAAS,gBAAgB,OAAoC;CAC3D,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SAAS,SAAS,KAAK,KAAK;EAClC,MAAM,cAAc,QAAQ,cAAc;EAC1C,IAAI,CAAC,aACH,SAAS,KACP,qBAAqB;GACnB,MAAM,KAAK,WAAW,UAAU,cAAc,KAAK;GACnD,SAAS,QAAQ,iBAAiB,OAAO,OAAO,gBAAgB,KAAK,OAAO;GAC5E,mBAAmB,KAAK,qBAAqB,SAAS,QAAQ,iBAAiB;GAC/E,YACE,KAAK,eAAe,KAAA,KAAa,KAAK,WAAW,SAAS,IACtD,KAAK,WAAW,IAAI,kBAAkB,IACtC,KAAA;GACN,MAAM,SAAS,QAAQ,IAAI;GAC3B,mBAAmB,KAAK,sBAAsB;EAChD,CAAC,CACH;EAEF,KAAK,MAAM,UAAU,KAAK,aAAa,WAAW,CAAC,GAAG;GACpD,MAAM,eAAe,SAAS,OAAO,KAAK;GAC1C,MAAM,SAAS,OAAO,kBAAkB,SAAS,cAAc,cAAc;GAC7E,IAAI,aAAa;IAGf,SAAS,KAAK,sBAAsB,QAAQ,MAAM,CAAC;IACnD;GACF;GACA,IAAI,OAAO,WAAW,UAAU;IAC9B,SAAS,KAAK,sBAAsB,QAAQ,MAAM,CAAC;IACnD;GACF;GAIA,IAAI,OAAO,YAAY,KAAA,GACrB,SAAS,KACP,qBAAqB;IAAE,MAAM;IAAU,SAAS,gBAAgB,OAAO,OAAO;GAAE,CAAC,CACnF;EAEJ;CACF;CACA,OAAO;AACT;AAEA,SAAS,SAAS,QAA6C,YAAiC;CAC9F,MAAM,OAAO,cAAc,QAAQ,MAAM;CACzC,OAAO;EACL,OAAO,SAAS,MAAM,KAAK,KAAK;EAChC,aAAa,SAAS,MAAM,WAAW,KAAK;EAU5C,OAAO;EACP,MAAM,eAAe,MAAM,IAAI;EAC/B,KAAK,qBAAqB,MAAM,GAAG,KAAK;CAC1C;AACF;AAEA,SAAS,WACP,QACA,OACe;CACf,MAAM,WAAW,cAAc,QAAQ,QAAQ;CAC/C,IAAI,aAAa,KAAA,GAGf,OAAO;EACL,SAAS,SAAS,SAAS,OAAO,KAAK;EACvC,iBAAiB,SAAS,SAAS,eAAe,KAAK;EACvD,OAAO,SAAS,SAAS,KAAK,KAAK;EACnC,UAAU,SAAS,SAAS,QAAQ,KAAK;EACzC,gBAAgB,SAAS,SAAS,cAAc,KAAK;EACrD,GAAI,SAAS,gBAAgB,KAAA,IACzB,EAAE,aAAa,SAAS,SAAS,WAAW,KAAK,KAAK,IACtD,CAAC;EACL,GAAI,SAAS,gBAAgB,KAAA,IACzB,EAAE,aAAa,SAAS,SAAS,WAAW,KAAK,KAAK,IACtD,CAAC;EACL,GAAI,SAAS,0BAA0B,KAAA,IACnC,EAAE,uBAAuB,SAAS,SAAS,qBAAqB,KAAK,KAAK,IAC1E,CAAC;EACL,UAAUA,WAAS,SAAS,QAAQ,IAAI,SAAS,WAAW;CAC9D;CAEF,OAAO;EACL,SAAS,MAAM;EACf,iBAAiB,MAAM;EACvB,OAAO,MAAM,cAAc;EAC3B,UAAU;EACV,gBAAgB;EAChB,UAAU;CACZ;AACF;AAEA,SAAS,cAAc,QAA+D;CACpF,MAAM,YAAY,cAAc,QAAQ,WAAW;CACnD,OAAO;EACL,YAAY,SAAS,WAAW,UAAU,KAAK;EAC/C,SAAS,SAAS,WAAW,OAAO,KAAK;EACzC,gBAAgB,SAAS,WAAW,cAAc,KAAK;CACzD;AACF;AAEA,SAAS,SAAS,OAAyD;CACzE,MAAM,WAAW,OAAO;CACxB,MAAM,QAAQA,WAAS,QAAQ,IAAI,WAAW,KAAA;CAC9C,MAAM,QAAQ,qBAAqB,OAAO,cAAc;CACxD,OAAO;EACL,KAAK,eAAe,OAAO,cAAc;EACzC,WAAW,eAAe,OAAO,mBAAmB;EACpD,YAAY,eAAe,OAAO,uBAAuB;EACzD,kBAAkB,eAAe,OAAO,gBAAgB;EACxD,YAAY,eAAe,OAAO,mBAAmB;EACrD,aAAa,eAAe,OAAO,kBAAkB;EACrD,QAAQ,eAAe,OAAO,MAAM;EACpC,GAAI,UAAU,KAAA,IAAY,EAAE,gBAAgB,MAAM,IAAI,CAAC;CACzD;AACF;;;;;;;;;;;AAYA,SAAS,eAAe,OAAoC;CAC1D,MAAM,QAAuB,CAAC;CAC9B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,WAAW,SAAS;EAC7B,MAAM,QAAQ,qBAAqB,KAAK,cAAc;EACtD,MAAM,YAAY,eAAe,KAAK,SAAS,gBAAgB;EAC/D,MAAM,gBAAgB,eAAe,KAAK,SAAS,oBAAoB;EACvE,MAAM,WAAW,cAAc,KAAK,SAAS,QAAQ;EACrD,IACE,UAAU,KAAA,KACV,cAAc,KAAA,KACd,kBAAkB,KAAA,KAClB,aAAa,KAAA,GAEb;EAEF,MAAM,KAAK;GACT,MAAM;GACN,MAAM,KAAK,cAAc;GACzB,GAAI,UAAU,KAAA,IAAY,EAAE,gBAAgB,MAAM,IAAI,CAAC;GACvD,GAAI,cAAc,KAAA,IAAY,EAAE,kBAAkB,UAAU,IAAI,CAAC;GACjE,GAAI,kBAAkB,KAAA,IAAY,EAAE,sBAAsB,cAAc,IAAI,CAAC;GAC7E,GAAI,aAAa,KAAA,IAAY,EAAE,SAAS,IAAI,CAAC;EAC/C,CAAC;CACH;CACA,OAAO;AACT;;;;;;;;AASA,SAAS,WAAW,aAAiC,OAAkC;CACrF,MAAM,QAAkB,CAAC;CACzB,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,OAAO,CAAC,GAAI,aAAa,MAAM,KAAK,KAAK,CAAC,GAAI,GAAG,KAAK,GAAG;EAClE,MAAM,SAAS,IAAI,KAAK;EACxB,IAAI,OAAO,WAAW,KAAK,KAAK,IAAI,MAAM,GAAG;EAC7C,KAAK,IAAI,MAAM;EACf,MAAM,KAAK,MAAM;CACnB;CACA,OAAO,MAAM,KAAK,KAAK;AACzB;AAOA,SAAS,WACP,YACA,UACA,YACa;CACb,MAAM,SAAS,SAAS,WAAW,KAAK;CACxC,MAAM,YACJ,SAAS,QAAQ,UAAU,KAAK,WAAW,iBAAiB,WAAW;CACzE,IAAI,cAAc,KAAA,KAAa,UAAU,WAAW,GAClD,MAAM,IAAI,MACR,+HACF;CAEF,MAAM,UAAU,cAAc,QAAQ,SAAS;CAC/C,MAAM,aAAa,cAAc,QAAQ,YAAY;CACrD,MAAM,UAAU,YAAY;CAC5B,MAAM,cAAc,SAAS,YAAY,GAAG;CAC5C,MAAM,OAAO,QAAQ;CACrB,MAAM,gBAAgB,QAAQ;CAI9B,MAAM,cAAc,eAAe,WAAW,KAAK;CACnD,MAAM,QAAmC,MAAM,QAAQ,aAAa,IAC/D,gBACD,YAAY,SAAS,IACnB,cACA,KAAA;CACN,MAAM,WAAW,gBAAgB,WAAW,KAAK;CACjD,MAAM,OAAoB;EACxB,QAAQ;EACR,YAAY;EAGZ,mBAAmB,YAAY,SAAS,QAAQ,iBAAiB,KAAK;EACtE,QAAQ,SAAS,QAAQ,MAAM,KAAK,WAAW,cAAc;EAG7D,eAAe,SAAS,QAAQ,aAAa,KAAK;EAClD,cAAc,SAAS,QAAQ,YAAY,KAAK;EAChD,YAAY,qBAAqB,QAAQ,UAAU,KAAK;EACxD,iBAAiB,qBAAqB,QAAQ,eAAe,KAAK;EAClE,MAAM,cAAc,SAAS,IAAmB,IAAK,OAAuB;EAC5E,MAAM,SAAS,QAAQ,SAAS;EAChC,QAAQ,WAAW,QAAQ,WAAW,KAAK;EAC3C;EACA,WAAW,WAAW,MAAM,oBAAoB,CAAC;EACjD,GAAI,UAAU,KAAA,IAAY,EAAE,OAAO,MAAM,IAAI,CAAC;EAC9C,SAAS;GAGP,QAAQ;GACR,eAAe;GACf,SAAS;GACT,SAASA,WAAS,SAAS,OAAO,IAAI,QAAQ,UAAU,CAAC;GACzD,cAAc,OAAO,SAAS,iBAAiB,YAAY,QAAQ,eAAe;GAClF,cACE,OAAO,SAAS,iBAAiB,YAC7B,QAAQ,eACR,WAAW,6BAA6B,KAAA;GAC9C,OAAO,SAAS,SAAS,KAAK,KAAK;GAMnC,gBAAgB,SAAS,mBAAmB;EAC9C;EACA,MAAM,SAAS,WAAW,aAAa;EACvC,WAAW,cAAc,MAAM;EAC/B,YAAY;GACV,aAAa,SAAS,YAAY,WAAW,KAAK;GAClD,SAAS,iBAAiB,SAAS,OAAyB,IACvD,UACD;GACJ,KAAK,WAAW,aAAa,CAAC,iBAAiB,CAAC;GAQhD,GAAIA,WAAS,YAAY,cAAc,IACnC,EAAE,gBAAgB,WAAW,eAAgC,IAC7D,CAAC;EACP;CACF;CACA,kBAAkB,MAAM,8CAA8C,WAAW;CACjF,OAAO;AACT;;;;;;;;;;AAWA,SAAgB,qBACd,YACA,UAA6B,CAAC,GACf;CACf,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,QAAuB,CAAC;CAC9B,MAAM,QAAQ,MAAwB,UAAyB,WAA8B;EAC3F,MAAM,OAAO,WAAW,MAAM,UAAU,UAAU;EAClD,IAAI,OAAO,IAAI,KAAK,UAAU,GAC5B,MAAM,IAAI,MAAM,qDAAqD,KAAK,YAAY;EAExF,OAAO,IAAI,KAAK,UAAU;EAC1B,MAAM,KAAK,IAAI;EACf,KAAK,MAAM,SAAS,KAAK,yBAAyB,CAAC,GACjD,KAAK,OAAO,KAAK,YAAY,MAAM;EAErC,OAAO,OAAO,KAAK,UAAU;CAC/B;CACA,KAAK,YAAY,sBAAM,IAAI,IAAI,CAAC;CAChC,OAAO;AACT;;;;;;;;;;;;;;;AAgBA,SAAgB,qBACd,OACA,OACe;CACf,IAAI,CAAC,eAAe,SAAS,KAAK,GAChC,MAAM,IAAI,MAAM,uCAAuC,OAAO,KAAK,GAAG;CAExE,OAAO,MAAM,KAAK,UAAU;EAAE,GAAG;EAAM,MAAM;GAAE,GAAG,KAAK;GAAM;EAAM;CAAE,EAAE;AACzE;;;;;;;;;;;;;;;;;;;ACv7BA,MAAa,sBAAsB,KAAK,QAAQ,GAAG,UAAU,SAAS,YAAY,aAAa;;AA4B/F,eAAsB,eACpB,OAAe,qBACe;CAC9B,IAAI;EAIF,MAAM,EAAE,iBADY,cAAc,OAAO,KAAK,GACX,CAAC,CAAC,aAAa;EAClD,MAAM,KAAK,IAAI,aAAa,MAAM,EAAE,UAAU,KAAK,CAAC;EAEpD,GAAG,QAAQ,gCAAgC,CAAC,CAAC,IAAI;EACjD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,SAAS,gBAAgB,KAAkD;CACzE,IAAI,QAAqC;CACzC,IAAI,OAAO,IAAI,UAAU,YAAY,IAAI,MAAM,SAAS,GACtD,IAAI;EACF,MAAM,SAAkB,KAAK,MAAM,IAAI,KAAK;EAC5C,IAAI,SAAS,MAAM,GAAG,QAAQ;CAChC,QAAQ;EACN,QAAQ;CACV;CAEF,OAAO;EACL,IAAI,OAAO,IAAI,EAAE;EACjB,UAAU,IAAI,cAAc,QAAQ,IAAI,cAAc,KAAA,IAAY,OAAO,OAAO,IAAI,SAAS;EAC7F,WAAW,OAAO,IAAI,SAAS;EAC/B,OAAO,IAAI,UAAU,QAAQ,IAAI,UAAU,KAAA,IAAY,OAAO,OAAO,IAAI,KAAK;EAC9E;EACA,SAAS,OAAO,IAAI,QAAQ,CAAC;EAC7B,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,cAAc,OAAO,IAAI,iBAAiB,CAAC;EAC3C,iBAAiB,OAAO,IAAI,oBAAoB,CAAC;EACjD,iBAAiB,OAAO,IAAI,qBAAqB,CAAC;EAClD,kBAAkB,OAAO,IAAI,sBAAsB,CAAC;EACpD,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,aAAa,OAAO,IAAI,gBAAgB,CAAC;CAC3C;AACF;AAEA,MAAM,kBACJ;;AAGF,SAAgB,gCACd,IACA,WACsB;CAItB,OAHa,GACV,QAAQ,UAAU,gBAAgB,wDAAwD,CAAC,CAC3F,IAAI,SACG,CAAC,CAAC,IAAI,eAAe;AACjC;AAUA,SAAS,kBAAkB,QAAyB;CAClD,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,WAAW,QAAQ,WAAW,KAAA,GAAW,OAAO;CACpD,OAAO,KAAK,UAAU,MAAM;AAC9B;;;;;;;AAQA,SAAgB,4BAA4B,IAAkB,WAAkC;CAC9F,MAAM,cAAc,GACjB,QAAQ,6EAA6E,CAAC,CACtF,IAAI,SAAS;CAChB,MAAM,YAAY,GAAG,QAAQ,wDAAwD;CAErF,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,cAAc,aAAa;EACpC,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,WAAW,IAAI;GAClD,IAAI,CAAC,SAAS,MAAM,GAAG;GACvB,OAAO;EACT,QAAQ;GACN;EACF;EACA,MAAM,QAAwB,CAAC;EAC/B,KAAK,MAAM,OAAO,UAAU,IAAI,WAAW,EAAE,GAC3C,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI,IAAI;GAC3C,IAAI,SAAS,MAAM,GAAG,MAAM,KAAK,MAAsB;EACzD,QAAQ,CAER;EAGF,IAAI,KAAK,SAAS,QAAQ;GACxB,MAAM,OAAO,MACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAK,CAAC;GAC7C;EACF;EACA,IAAI,KAAK,SAAS,aAAa;EAI/B,MAAM,QAA0B,CAAC;EACjC,IAAI,UAA0B,CAAC;EAC/B,KAAK,MAAM,QAAQ,OAAO;GACxB,IAAI,KAAK,SAAS,cAAc;IAC9B,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;IAC1C,UAAU,CAAC;IACX;GACF;GACA,IAAI,KAAK,SAAS,iBAAiB,KAAK,SAAS,cAAc,KAAK,SAAS,SAAS;GACtF,QAAQ,KAAK,IAAI;EACnB;EACA,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;EAE1C,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,SAAS,eAAe,OAAO,EAAE,SAAS,YAAY,EAAE,KAAK,SAAS,CAAC,CAAC,CACxF,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,OAAO,KACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,YAAY,KAAK,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,WAAW,QAAQ;GACtF,MAAM,YAA4B,UAAU,KAAK,OAAO;IACtD,IAAI,EAAE;IACN,MAAM;IACN,UAAU;KACR,MAAM,EAAE,QAAQ;KAChB,WAAW,KAAK,UAAU,EAAE,OAAO,SAAS,CAAC,CAAC;IAChD;GACF,EAAE;GACF,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;GAC3E,SAAS,KAAK;IACZ,MAAM;IACN,SAAS,KAAK,SAAS,IAAI,OAAO;IAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;IAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;GAC1D,CAAC;GACD,KAAK,MAAM,KAAK,WACd,SAAS,KAAK;IACZ,MAAM;IACN,cAAc,EAAE;IAChB,MAAM,EAAE,QAAQ;IAChB,SAAS,kBAAkB,EAAE,OAAO,MAAM;GAC5C,CAAC;EAEL;CACF;CACA,OAAO;AACT"}

package/dist/{run-record-Br-Yzt_k.js → run-record-DPa2PeHx.js}

@@ -1,4 +1,4 @@
-import { g as pairedRiskDifferenceScore, h as pairedRiskDifferenceExact, p as pairedBinaryScale } from "./paired-arms-D4aeIHUy.js";
+import { g as pairedRiskDifferenceScore, h as pairedRiskDifferenceExact, p as pairedBinaryScale } from "./paired-arms-JSDcs2zx.js";
 import { a as pairedSignTest, i as pairedDeltaTieFraction, r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
 import { c as validateRunRecord } from "./run-record-DualPTn2.js";
 //#region src/paired-delta-test.ts
@@ -461,4 +461,4 @@ function mean(values) {
 //#endregion
 export { campaignCellToRunRecord as a, pairedDecisionShape as c, campaignCellTaskScore as i, minimumPairsForPairedDeltaTest as l, campaignCellExecutionEvidence as n, projectCampaignCellQuality as o, campaignCellJudgeDimensions as r, decidePairedPromotion as s, campaignCellCostProvenance as t, pairedDeltaTest as u };
 
-//# sourceMappingURL=run-record-Br-Yzt_k.js.map
+//# sourceMappingURL=run-record-DPa2PeHx.js.map

package/dist/{run-record-Br-Yzt_k.js.map → run-record-DPa2PeHx.js.map}

@@ -1 +1 @@
-{"version":3,"file":"run-record-Br-Yzt_k.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.ts"],"sourcesContent":["import {\n  type PairedBootstrapOptions,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n  pairedSignTest,\n} from './statistics'\n\nexport interface PairedDeltaTestOptions extends PairedBootstrapOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. */\n  threshold?: number\n  /** Caller-required paired observations. The exact test may impose a higher minimum. */\n  minPairs?: number\n}\n\nexport interface PairedDeltaTestResult {\n  bootstrap: PairedBootstrapResult\n  method: 'bootstrap-ci' | 'exact-sign'\n  /** Exact one-sided p-value below the bootstrap minimum; otherwise null. */\n  pValue: number | null\n  /** Effective observation minimum after accounting for confidence. */\n  minimumPairs: number\n  sufficient: boolean\n  /**\n   * The bootstrap interval has zero width (or is non-finite), so it carries no\n   * information about how far the estimate could be wrong and cannot support a\n   * decision in either direction. See {@link pairedDeltaTest}.\n   */\n  indeterminate: boolean\n  significant: boolean\n}\n\n/** Smallest all-positive sample that can clear a one-sided exact sign test. */\nexport function minimumPairsForPairedDeltaTest(confidence = 0.95): number {\n  if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n    throw new Error(\n      `minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`,\n    )\n  }\n  const oneSidedAlpha = (1 - confidence) / 2\n  return Math.ceil(Math.log2(1 / oneSidedAlpha))\n}\n\n/**\n * Tests whether a paired candidate-minus-baseline delta clears a threshold.\n *\n * At 20 or more pairs, the percentile bootstrap lower bound carries the\n * decision. Below that point the interval is descriptive only, so the function\n * switches to a pre-registered one-sided exact sign test. The exact path is\n * deliberately conservative: it requires both a point estimate above the\n * threshold and enough consistently positive paired differences.\n *\n * ## A zero-width interval is never significant\n *\n * When every paired delta is identical the resample distribution is a point\n * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n\n * identical deltas of g. Neither says the effect is certain — both say the\n * sample carries no information about how far the estimate could be wrong, and\n * `low > threshold` then answers on the point estimate alone. It fails in both\n * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a\n * tie-dominated pass/fail comparison laundered a regression into a\n * noninferiority pass, and `[g, g]` clears every threshold below g with no\n * spread behind it. Under a bounded asymmetric null whose true mean paired\n * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —\n * every sample that misses the drop is exactly that shape, and deciding on\n * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.\n *\n * So `indeterminate` is reported and `significant` is false whenever the\n * interval has zero width, on BOTH paths: at small n the exact sign test is a\n * test of the MEDIAN and a zero-spread sample is precisely where it stops\n * saying anything about the mean the caller is thresholding.\n *\n * `threshold` may be negative — that is a noninferiority margin, and it is the\n * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome\n * the percentile bootstrap is not a valid interval at a nonzero margin at all;\n * use {@link decidePairedPromotion}, which routes those to Tango's score\n * interval, rather than thresholding this function's bootstrap directly.\n */\nexport function pairedDeltaTest(\n  before: number[],\n  after: number[],\n  options: PairedDeltaTestOptions = {},\n): PairedDeltaTestResult {\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`)\n  }\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n  const bootstrap = pairedBootstrap(before, after, options)\n  const sufficient = bootstrap.n >= minimumPairs\n  // A point-mass resample distribution is an absence of evidence, not a\n  // certainty, and it is the shape that clears every negative threshold. Both\n  // paths refuse it: see the \"zero-width\" section of this function's docs.\n  const indeterminate =\n    bootstrap.n > 0 &&\n    (!Number.isFinite(bootstrap.low) ||\n      !Number.isFinite(bootstrap.high) ||\n      bootstrap.low === bootstrap.high)\n\n  if (bootstrap.gateEligible) {\n    return {\n      bootstrap,\n      method: 'bootstrap-ci',\n      pValue: null,\n      minimumPairs,\n      sufficient,\n      indeterminate,\n      significant: sufficient && !indeterminate && bootstrap.low > threshold,\n    }\n  }\n\n  const differences = before.map((value, index) => after[index]! - value - threshold)\n  const exact = pairedSignTest(differences, 'greater')\n  const estimate = options.statistic === 'mean' ? bootstrap.mean : bootstrap.median\n  return {\n    bootstrap,\n    method: 'exact-sign',\n    pValue: exact.pValue,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    significant:\n      sufficient &&\n      !indeterminate &&\n      estimate > threshold &&\n      exact.pValue <= (1 - bootstrap.confidence) / 2,\n  }\n}\n","/**\n * @module\n * ONE rule for \"does this paired interval clear a promotion threshold\".\n *\n * The rule below was derived on `HeldOutGate` (#479) after the same estimator\n * bug shipped twice. It then turned out that a SECOND gate — the composable\n * `heldOutGate`, plus everything else routed through `heldoutSignificance` —\n * still carried the original defect, because the rule had been written into one\n * gate's method body rather than into a shared function. Two copies of a\n * statistical rule is how a defect survives in one of them, so there is now\n * exactly one copy and both gates call it.\n *\n * Three things the rule does that a bare `pairedBootstrap(...).low > threshold`\n * does not:\n *\n * 1. **Two-point (pass/fail) outcomes decide on Tango's SCORE interval.** On a\n *    pass/fail eval the paired delta vector is dominated by ties, so the\n *    bootstrap of the mean is a resample of a lattice with three atoms and its\n *    percentile interval is not valid at a nonzero margin. The score interval\n *    (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under\n *    each hypothesised margin instead of fixing it at the observed value, which\n *    is the only construction that stays a confidence interval as the margin\n *    moves off zero — the regime every noninferiority threshold lives in.\n *    Measured on the composable gate before this change, at a true risk\n *    difference sitting exactly on the production caller's -0.05 margin and a\n *    nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.\n * 2. **McNemar's exact test holds a VETO at every non-negative threshold.**\n *    Redundant with the interval by construction and kept anyway, so that\n *    swapping the estimator for one without that duality cannot silently\n *    reintroduce \"promotes what the exact test refuses\". Witness: n = 6, b = 5,\n *    c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs\n *    (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE\n *    threshold is a noninferiority question, which McNemar's test of \"no\n *    difference\" is not the right test for, so the veto does not apply there.\n * 3. **A ZERO-WIDTH interval is refused, wherever it sits.** At [0, 0] it\n *    cannot tell a gain from a regression and clears every negative threshold.\n *    Away from zero it fails the opposite way: n identical positive deltas give\n *    [g, g], which clears threshold 0 on no spread at all. Both are an absence\n *    of evidence. Measured on the composable gate before this change, under a\n *    bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %\n *    false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.\n *\n * Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n */\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n  BOOTSTRAP_GATE_MIN_N,\n  type PairedBootstrapResult,\n  pairedBinaryScale,\n  pairedDeltaTieFraction,\n  pairedRiskDifferenceExact,\n  pairedRiskDifferenceScore,\n} from './statistics'\n\n/** Which paired estimator produced the deciding interval. */\nexport type PairedDecisionStatistic =\n  | 'paired_risk_difference'\n  | 'mean_bootstrap'\n  | 'median_bootstrap'\n\n/** Which test carried the decision, given the estimator and the sample size. */\nexport type PairedDecisionMethod = 'score-interval' | 'bootstrap-ci' | 'exact-sign'\n\n/** McNemar's exact paired-binary evidence, on the two-point path only. */\nexport interface PairedMcNemarEvidence {\n  /** Discordant pairs the treatment won. */\n  b: number\n  /** Discordant pairs the control won. */\n  c: number\n  /** b + c — the only pairs carrying information. */\n  nDiscordant: number\n  /** Two-sided exact p-value. */\n  pValue: number\n}\n\nexport interface PairedPromotionDecisionOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement, in the\n   *  caller's native units. May be negative (a noninferiority margin). Default 0. */\n  threshold?: number\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples, on the paths where a bootstrap decides. Default 2000. */\n  resamples?: number\n  /** Deterministic bootstrap seed. Omitted ⇒ derived from the deltas. */\n  seed?: number\n  /** Caller-required paired observations. The test also imposes its own\n   *  minimum: bootstrap eligibility for a continuous mean, or the exact-test\n   *  minimum for binary and explicit median targets. This is not a power guarantee. */\n  minPairs?: number\n  /**\n   * `'mean'` (default) routes by SHAPE: a two-point (pass/fail) outcome on any\n   * encoding decides on the score interval, everything else on the mean\n   * bootstrap. `'median'` forces the median bootstrap on every input, including\n   * shapes where it is structurally blind — kept for callers who want outlier\n   * robustness on genuinely continuous outcomes and accept that cost.\n   */\n  statistic?: 'mean' | 'median'\n  /** Declared binary support {0, binaryScale}, including when all observations\n   *  are zero. Must be finite and positive; every observation must use that\n   *  support. Uses the risk-difference mean, so cannot accompany 'median'.\n   *  Omitted: infer binary support from observed positive values. */\n  binaryScale?: number\n}\n\nexport interface PairedPromotionDecision {\n  /** Paired observations supplied. */\n  n: number\n  /** Threshold the interval was judged against, native units. */\n  threshold: number\n  confidence: number\n  statistic: PairedDecisionStatistic\n  method: PairedDecisionMethod\n  /** Declared or inferred positive level ({0,1} ⇒ 1, {0,100} ⇒ 100),\n   *  or null when the outcome is not two-point. Non-null is exactly the\n   *  condition for the `paired_risk_difference` path, and it is the factor\n   *  `delta` / `low` / `high` were rescaled by. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n  /** Point estimate of the DECIDING statistic, in the caller's native units. */\n  delta: number\n  /** Lower bound of the DECIDING interval, native units. */\n  low: number\n  /** Upper bound of the DECIDING interval, native units. */\n  high: number\n  /** The bootstrap that decided, or null when the score interval did. Callers\n   *  that need a bootstrap as a diagnostic on the two-point path compute their\n   *  own — it is not computed here, so the binary path costs no resamples. */\n  bootstrap: PairedBootstrapResult | null\n  /** McNemar's exact evidence, or null off the two-point path. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** Exact one-sided sign-test p-value on the small-sample bootstrap path;\n   *  null otherwise. */\n  pValue: number | null\n  /** Effective observation minimum for the target, confidence, and caller floor. */\n  minimumPairs: number\n  /** n >= minimumPairs. */\n  sufficient: boolean\n  /** The deciding interval is zero-width at numeric precision or non-finite — no evidence in either\n   *  direction, so it cannot clear any threshold on evidence. */\n  indeterminate: boolean\n  /** McNemar's exact test refuses at a non-negative threshold. */\n  exactTestVetoes: boolean\n  /** The deciding interval clears the threshold, ignoring the other two guards. */\n  clearsThreshold: boolean\n  /** `sufficient && !indeterminate && clearsThreshold && !exactTestVetoes` —\n   *  the whole rule. */\n  promote: boolean\n  /** What `delta` measures, for a reason string. */\n  label: 'success-rate' | 'mean' | 'median'\n  /** Why a zero-width interval is zero-width; empty when it is not. */\n  indeterminateCause: string\n  /** Sentence naming the test when the exact sign test decided; else empty. */\n  methodDetail: string\n}\n\n/** The shape facts that pick the estimator, without computing an interval. */\nexport interface PairedDecisionShape {\n  statistic: PairedDecisionStatistic\n  /** Declared or inferred positive binary level; null off the binary path. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n}\n\n/**\n * Which estimator {@link decidePairedPromotion} would use on this data, and the\n * shape facts behind it — for callers that must report the shape on a path\n * where no interval is computed at all (an early rejection, or zero pairs).\n * Cheap: no bootstrap, no interval.\n */\nexport function pairedDecisionShape(\n  before: number[],\n  after: number[],\n  statistic: 'mean' | 'median' = 'mean',\n  declaredBinaryScale?: number,\n): PairedDecisionShape {\n  if (declaredBinaryScale !== undefined) {\n    if (!Number.isFinite(declaredBinaryScale) || declaredBinaryScale <= 0) {\n      throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n    }\n    if (statistic === 'median') {\n      throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n    }\n    for (const [name, arm] of [\n      ['before', before],\n      ['after', after],\n    ] as const) {\n      for (let i = 0; i < arm.length; i++) {\n        const value = arm[i]!\n        if (value !== 0 && value !== declaredBinaryScale) {\n          throw new Error(\n            `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(\n      `decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`,\n    )\n  }\n  const { binaryScale, tieFraction } = pairedDecisionShape(\n    before,\n    after,\n    options.statistic,\n    options.binaryScale,\n  )\n  const estimatorMinimum =\n    binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n  const n = before.length\n  const sufficient = n >= minimumPairs\n\n  let core: {\n    statistic: PairedDecisionStatistic\n    method: PairedDecisionMethod\n    delta: number\n    low: number\n    high: number\n    bootstrap: PairedBootstrapResult | null\n    mcnemar: PairedMcNemarEvidence | null\n    pValue: number | null\n    clearsThreshold: boolean\n    label: PairedPromotionDecision['label']\n    methodDetail: string\n  }\n\n  if (binaryScale !== null) {\n    // Normalise the two-point encoding to {0,1} so the estimators see the\n    // pass/fail structure, then rescale the answer back into the caller's\n    // native units — the threshold is read in the units of the scores, so a\n    // 0-100 pass/fail dimension must be gated in points, not in a rate.\n    const unitControl = before.map((v) => v / binaryScale)\n    const unitTreatment = after.map((v) => v / binaryScale)\n    // TWO estimators with different jobs, because no single one does both.\n    // `exact` is the authority on RD = 0 and supplies the veto; its\n    // Clopper-Pearson interval conditions on the discordant count and is\n    // therefore NOT a confidence interval at a nonzero margin. `score` is\n    // Tango's, which is.\n    const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence)\n    const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence)\n    const low = score.lower * binaryScale\n    core = {\n      statistic: 'paired_risk_difference',\n      method: 'score-interval',\n      delta: score.riskDifference * binaryScale,\n      low,\n      high: score.upper * binaryScale,\n      bootstrap: null,\n      mcnemar: {\n        b: exact.b,\n        c: exact.c,\n        nDiscordant: exact.nDiscordant,\n        pValue: exact.pValue,\n      },\n      pValue: null,\n      clearsThreshold: low > threshold,\n      label: 'success-rate',\n      methodDetail: '',\n    }\n  } else {\n    const bootstrapStatistic = options.statistic === 'median' ? 'median' : 'mean'\n    const test = pairedDeltaTest(before, after, {\n      confidence,\n      resamples: options.resamples,\n      statistic: bootstrapStatistic,\n      seed: options.seed,\n      threshold,\n      minPairs: minimumPairs,\n    })\n    const ci = test.bootstrap\n    core = {\n      statistic: bootstrapStatistic === 'mean' ? 'mean_bootstrap' : 'median_bootstrap',\n      method: test.method,\n      delta: bootstrapStatistic === 'mean' ? ci.mean : ci.median,\n      low: ci.low,\n      high: ci.high,\n      bootstrap: ci,\n      mcnemar: null,\n      pValue: test.pValue,\n      clearsThreshold: test.significant,\n      label: bootstrapStatistic,\n      methodDetail:\n        test.method === 'exact-sign'\n          ? bootstrapStatistic === 'mean'\n            ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n              ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n            : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n              ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n          : '',\n    }\n  }\n\n  // Equivalent deltas can differ by rounding after subtraction and averaging.\n  // Use the interval's own scale so tiny score units retain the same decision\n  // as a rescaled copy of the evidence.\n  const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n  const intervalTolerance = intervalScale * Number.EPSILON * 8\n  const indeterminate =\n    !Number.isFinite(core.low) ||\n    !Number.isFinite(core.high) ||\n    core.high - core.low <= intervalTolerance\n  const indeterminateCause = !indeterminate\n    ? ''\n    : tieFraction === 1\n      ? 'every paired delta is an exact tie'\n      : core.mcnemar !== null && core.mcnemar.nDiscordant === 0\n        ? 'every pair is concordant (0 discordant pairs)'\n        : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`\n  // Only at a non-negative threshold: a negative threshold asks a\n  // noninferiority question, which McNemar's test of \"no difference\" does not\n  // answer.\n  const exactTestVetoes =\n    core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence)\n\n  return {\n    n,\n    threshold,\n    confidence,\n    binaryScale,\n    tieFraction,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    indeterminateCause,\n    exactTestVetoes,\n    promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,\n    ...core,\n  }\n}\n\nfunction fmt(x: number): string {\n  return x.toFixed(4)\n}\n","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,IAAI,gBAAgB,MAClB,OAAO;EAAE,WAAW;EAA0B;EAAa;CAAY;CAEzE,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;AACvE;;;;;;AAOA,SAAgB,sBACd,QACA,OACA,UAA0C,CAAC,GAClB;CACzB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,gDAAgD,OAAO,OAAO,MAAM,MAAM,OAAO,EACnF;CAEF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,wDAAwD,WAAW;CAErF,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,eAAe,+BAA+B,UAAU;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MACR,mEAAmE,kBACrE;CAEF,MAAM,EAAE,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,WACV;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,IAAI;CAcJ,IAAI,gBAAgB,MAAM;EAKxB,MAAM,cAAc,OAAO,KAAK,MAAM,IAAI,WAAW;EACrD,MAAM,gBAAgB,MAAM,KAAK,MAAM,IAAI,WAAW;EAMtD,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,MAAM,MAAM,QAAQ;EAC1B,OAAO;GACL,WAAW;GACX,QAAQ;GACR,OAAO,MAAM,iBAAiB;GAC9B;GACA,MAAM,MAAM,QAAQ;GACpB,WAAW;GACX,SAAS;IACP,GAAG,MAAM;IACT,GAAG,MAAM;IACT,aAAa,MAAM;IACnB,QAAQ,MAAM;GAChB;GACA,QAAQ;GACR,iBAAiB,MAAM;GACvB,OAAO;GACP,cAAc;EAChB;CACF,OAAO;EACL,MAAM,qBAAqB,QAAQ,cAAc,WAAW,WAAW;EACvE,MAAM,OAAO,gBAAgB,QAAQ,OAAO;GAC1C;GACA,WAAW,QAAQ;GACnB,WAAW;GACX,MAAM,QAAQ;GACd;GACA,UAAU;EACZ,CAAC;EACD,MAAM,KAAK,KAAK;EAChB,OAAO;GACL,WAAW,uBAAuB,SAAS,mBAAmB;GAC9D,QAAQ,KAAK;GACb,OAAO,uBAAuB,SAAS,GAAG,OAAO,GAAG;GACpD,KAAK,GAAG;GACR,MAAM,GAAG;GACT,WAAW;GACX,SAAS;GACT,QAAQ,KAAK;GACb,iBAAiB,KAAK;GACtB,OAAO;GACP,cACE,KAAK,WAAW,eACZ,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,MAAM,qBAAqB,CAAC,gBACxB,KACA,gBAAgB,IACd,uCACA,KAAK,YAAY,QAAQ,KAAK,QAAQ,gBAAgB,IACpD,kDACA,OAAO,KAAK,MAAM,8BAA8B,IAAI,KAAK,GAAG;CAIpE,MAAM,kBACJ,KAAK,YAAY,QAAQ,aAAa,KAAK,EAAE,KAAK,QAAQ,SAAS,IAAI;CAEzE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,cAAc,CAAC,iBAAiB,KAAK,mBAAmB,CAAC;EAClE,GAAG;CACL;AACF;AAEA,SAAS,IAAI,GAAmB;CAC9B,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}
+{"version":3,"file":"run-record-DPa2PeHx.js","names":[],"sources":["../src/paired-delta-test.ts","../src/paired-promotion-decision.ts","../src/campaign/run-record.ts"],"sourcesContent":["import {\n  type PairedBootstrapOptions,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n  pairedSignTest,\n} from './statistics'\n\nexport interface PairedDeltaTestOptions extends PairedBootstrapOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement. Default 0. */\n  threshold?: number\n  /** Caller-required paired observations. The exact test may impose a higher minimum. */\n  minPairs?: number\n}\n\nexport interface PairedDeltaTestResult {\n  bootstrap: PairedBootstrapResult\n  method: 'bootstrap-ci' | 'exact-sign'\n  /** Exact one-sided p-value below the bootstrap minimum; otherwise null. */\n  pValue: number | null\n  /** Effective observation minimum after accounting for confidence. */\n  minimumPairs: number\n  sufficient: boolean\n  /**\n   * The bootstrap interval has zero width (or is non-finite), so it carries no\n   * information about how far the estimate could be wrong and cannot support a\n   * decision in either direction. See {@link pairedDeltaTest}.\n   */\n  indeterminate: boolean\n  significant: boolean\n}\n\n/** Smallest all-positive sample that can clear a one-sided exact sign test. */\nexport function minimumPairsForPairedDeltaTest(confidence = 0.95): number {\n  if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n    throw new Error(\n      `minimumPairsForPairedDeltaTest: confidence must be in (0,1), got ${confidence}`,\n    )\n  }\n  const oneSidedAlpha = (1 - confidence) / 2\n  return Math.ceil(Math.log2(1 / oneSidedAlpha))\n}\n\n/**\n * Tests whether a paired candidate-minus-baseline delta clears a threshold.\n *\n * At 20 or more pairs, the percentile bootstrap lower bound carries the\n * decision. Below that point the interval is descriptive only, so the function\n * switches to a pre-registered one-sided exact sign test. The exact path is\n * deliberately conservative: it requires both a point estimate above the\n * threshold and enough consistently positive paired differences.\n *\n * ## A zero-width interval is never significant\n *\n * When every paired delta is identical the resample distribution is a point\n * mass and the interval collapses: `[0, 0]` when all pairs tie, `[g, g]` on n\n * identical deltas of g. Neither says the effect is certain — both say the\n * sample carries no information about how far the estimate could be wrong, and\n * `low > threshold` then answers on the point estimate alone. It fails in both\n * directions: `[0, 0]` clears every NEGATIVE threshold, which is how a\n * tie-dominated pass/fail comparison laundered a regression into a\n * noninferiority pass, and `[g, g]` clears every threshold below g with no\n * spread behind it. Under a bounded asymmetric null whose true mean paired\n * delta is exactly 0 — 2 % of pairs dropping by 1.0, the rest gaining 0.0204 —\n * every sample that misses the drop is exactly that shape, and deciding on\n * `low > 0` promoted 65.65 % of samples at n = 20 against a nominal 5 %.\n *\n * So `indeterminate` is reported and `significant` is false whenever the\n * interval has zero width, on BOTH paths: at small n the exact sign test is a\n * test of the MEDIAN and a zero-spread sample is precisely where it stops\n * saying anything about the mean the caller is thresholding.\n *\n * `threshold` may be negative — that is a noninferiority margin, and it is the\n * regime the zero-width hole is worst in. For a two-point (pass/fail) outcome\n * the percentile bootstrap is not a valid interval at a nonzero margin at all;\n * use {@link decidePairedPromotion}, which routes those to Tango's score\n * interval, rather than thresholding this function's bootstrap directly.\n */\nexport function pairedDeltaTest(\n  before: number[],\n  after: number[],\n  options: PairedDeltaTestOptions = {},\n): PairedDeltaTestResult {\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`pairedDeltaTest: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(`pairedDeltaTest: minPairs must be a positive integer, got ${requestedMinimum}`)\n  }\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum)\n  const bootstrap = pairedBootstrap(before, after, options)\n  const sufficient = bootstrap.n >= minimumPairs\n  // A point-mass resample distribution is an absence of evidence, not a\n  // certainty, and it is the shape that clears every negative threshold. Both\n  // paths refuse it: see the \"zero-width\" section of this function's docs.\n  const indeterminate =\n    bootstrap.n > 0 &&\n    (!Number.isFinite(bootstrap.low) ||\n      !Number.isFinite(bootstrap.high) ||\n      bootstrap.low === bootstrap.high)\n\n  if (bootstrap.gateEligible) {\n    return {\n      bootstrap,\n      method: 'bootstrap-ci',\n      pValue: null,\n      minimumPairs,\n      sufficient,\n      indeterminate,\n      significant: sufficient && !indeterminate && bootstrap.low > threshold,\n    }\n  }\n\n  const differences = before.map((value, index) => after[index]! - value - threshold)\n  const exact = pairedSignTest(differences, 'greater')\n  const estimate = options.statistic === 'mean' ? bootstrap.mean : bootstrap.median\n  return {\n    bootstrap,\n    method: 'exact-sign',\n    pValue: exact.pValue,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    significant:\n      sufficient &&\n      !indeterminate &&\n      estimate > threshold &&\n      exact.pValue <= (1 - bootstrap.confidence) / 2,\n  }\n}\n","/**\n * @module\n * ONE rule for \"does this paired interval clear a promotion threshold\".\n *\n * The rule below was derived on `HeldOutGate` (#479) after the same estimator\n * bug shipped twice. It then turned out that a SECOND gate — the composable\n * `heldOutGate`, plus everything else routed through `heldoutSignificance` —\n * still carried the original defect, because the rule had been written into one\n * gate's method body rather than into a shared function. Two copies of a\n * statistical rule is how a defect survives in one of them, so there is now\n * exactly one copy and both gates call it.\n *\n * Three things the rule does that a bare `pairedBootstrap(...).low > threshold`\n * does not:\n *\n * 1. **Two-point (pass/fail) outcomes decide on Tango's SCORE interval.** On a\n *    pass/fail eval the paired delta vector is dominated by ties, so the\n *    bootstrap of the mean is a resample of a lattice with three atoms and its\n *    percentile interval is not valid at a nonzero margin. The score interval\n *    (`pairedRiskDifferenceScore`) re-estimates the nuisance loss rate under\n *    each hypothesised margin instead of fixing it at the observed value, which\n *    is the only construction that stays a confidence interval as the margin\n *    moves off zero — the regime every noninferiority threshold lives in.\n *    Measured on the composable gate before this change, at a true risk\n *    difference sitting exactly on the production caller's -0.05 margin and a\n *    nominal 5 %: 14.60 % false promotion at n = 40 and 10.10 % at n = 76.\n * 2. **McNemar's exact test holds a VETO at every non-negative threshold.**\n *    Redundant with the interval by construction and kept anyway, so that\n *    swapping the estimator for one without that duality cannot silently\n *    reintroduce \"promotes what the exact test refuses\". Witness: n = 6, b = 5,\n *    c = 0 — no exact argument reaches alpha = 0.05 with 5 discordant pairs\n *    (two-sided floor 2/2^5 = 0.0625), whatever an interval says. A NEGATIVE\n *    threshold is a noninferiority question, which McNemar's test of \"no\n *    difference\" is not the right test for, so the veto does not apply there.\n * 3. **A ZERO-WIDTH interval is refused, wherever it sits.** At [0, 0] it\n *    cannot tell a gain from a regression and clears every negative threshold.\n *    Away from zero it fails the opposite way: n identical positive deltas give\n *    [g, g], which clears threshold 0 on no spread at all. Both are an absence\n *    of evidence. Measured on the composable gate before this change, under a\n *    bounded asymmetric null whose true mean paired delta is exactly 0: 88.50 %\n *    false promotion at n = 6 and 65.65 % at n = 20 against a nominal 5 %.\n *\n * Eligibility follows the requested target. A continuous mean needs the\n * bootstrap minimum; binary outcomes and explicit median targets use the\n * exact-test minimum. A small-sample sign diagnostic cannot certify a mean\n * effect. These floors establish estimator eligibility, not adequate power.\n */\n\nimport { minimumPairsForPairedDeltaTest, pairedDeltaTest } from './paired-delta-test'\nimport {\n  BOOTSTRAP_GATE_MIN_N,\n  type PairedBootstrapResult,\n  pairedBinaryScale,\n  pairedDeltaTieFraction,\n  pairedRiskDifferenceExact,\n  pairedRiskDifferenceScore,\n} from './statistics'\n\n/** Which paired estimator produced the deciding interval. */\nexport type PairedDecisionStatistic =\n  | 'paired_risk_difference'\n  | 'mean_bootstrap'\n  | 'median_bootstrap'\n\n/** Which test carried the decision, given the estimator and the sample size. */\nexport type PairedDecisionMethod = 'score-interval' | 'bootstrap-ci' | 'exact-sign'\n\n/** McNemar's exact paired-binary evidence, on the two-point path only. */\nexport interface PairedMcNemarEvidence {\n  /** Discordant pairs the treatment won. */\n  b: number\n  /** Discordant pairs the control won. */\n  c: number\n  /** b + c — the only pairs carrying information. */\n  nDiscordant: number\n  /** Two-sided exact p-value. */\n  pValue: number\n}\n\nexport interface PairedPromotionDecisionOptions {\n  /** Smallest candidate-minus-baseline delta that counts as improvement, in the\n   *  caller's native units. May be negative (a noninferiority margin). Default 0. */\n  threshold?: number\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples, on the paths where a bootstrap decides. Default 2000. */\n  resamples?: number\n  /** Deterministic bootstrap seed. Omitted ⇒ derived from the deltas. */\n  seed?: number\n  /** Caller-required paired observations. The test also imposes its own\n   *  minimum: bootstrap eligibility for a continuous mean, or the exact-test\n   *  minimum for binary and explicit median targets. This is not a power guarantee. */\n  minPairs?: number\n  /**\n   * `'mean'` (default) routes by SHAPE: a two-point (pass/fail) outcome on any\n   * encoding decides on the score interval, everything else on the mean\n   * bootstrap. `'median'` forces the median bootstrap on every input, including\n   * shapes where it is structurally blind — kept for callers who want outlier\n   * robustness on genuinely continuous outcomes and accept that cost.\n   */\n  statistic?: 'mean' | 'median'\n  /** Declared binary support {0, binaryScale}, including when all observations\n   *  are zero. Must be finite and positive; every observation must use that\n   *  support. Uses the risk-difference mean, so cannot accompany 'median'.\n   *  Omitted: infer binary support from observed positive values. */\n  binaryScale?: number\n}\n\nexport interface PairedPromotionDecision {\n  /** Paired observations supplied. */\n  n: number\n  /** Threshold the interval was judged against, native units. */\n  threshold: number\n  confidence: number\n  statistic: PairedDecisionStatistic\n  method: PairedDecisionMethod\n  /** Declared or inferred positive level ({0,1} ⇒ 1, {0,100} ⇒ 100),\n   *  or null when the outcome is not two-point. Non-null is exactly the\n   *  condition for the `paired_risk_difference` path, and it is the factor\n   *  `delta` / `low` / `high` were rescaled by. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n  /** Point estimate of the DECIDING statistic, in the caller's native units. */\n  delta: number\n  /** Lower bound of the DECIDING interval, native units. */\n  low: number\n  /** Upper bound of the DECIDING interval, native units. */\n  high: number\n  /** The bootstrap that decided, or null when the score interval did. Callers\n   *  that need a bootstrap as a diagnostic on the two-point path compute their\n   *  own — it is not computed here, so the binary path costs no resamples. */\n  bootstrap: PairedBootstrapResult | null\n  /** McNemar's exact evidence, or null off the two-point path. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** Exact one-sided sign-test p-value on the small-sample bootstrap path;\n   *  null otherwise. */\n  pValue: number | null\n  /** Effective observation minimum for the target, confidence, and caller floor. */\n  minimumPairs: number\n  /** n >= minimumPairs. */\n  sufficient: boolean\n  /** The deciding interval is zero-width at numeric precision or non-finite — no evidence in either\n   *  direction, so it cannot clear any threshold on evidence. */\n  indeterminate: boolean\n  /** McNemar's exact test refuses at a non-negative threshold. */\n  exactTestVetoes: boolean\n  /** The deciding interval clears the threshold, ignoring the other two guards. */\n  clearsThreshold: boolean\n  /** `sufficient && !indeterminate && clearsThreshold && !exactTestVetoes` —\n   *  the whole rule. */\n  promote: boolean\n  /** What `delta` measures, for a reason string. */\n  label: 'success-rate' | 'mean' | 'median'\n  /** Why a zero-width interval is zero-width; empty when it is not. */\n  indeterminateCause: string\n  /** Sentence naming the test when the exact sign test decided; else empty. */\n  methodDetail: string\n}\n\n/** The shape facts that pick the estimator, without computing an interval. */\nexport interface PairedDecisionShape {\n  statistic: PairedDecisionStatistic\n  /** Declared or inferred positive binary level; null off the binary path. */\n  binaryScale: number | null\n  /** Exact-tie fraction over the paired deltas; null when there are no pairs. */\n  tieFraction: number | null\n}\n\n/**\n * Which estimator {@link decidePairedPromotion} would use on this data, and the\n * shape facts behind it — for callers that must report the shape on a path\n * where no interval is computed at all (an early rejection, or zero pairs).\n * Cheap: no bootstrap, no interval.\n */\nexport function pairedDecisionShape(\n  before: number[],\n  after: number[],\n  statistic: 'mean' | 'median' = 'mean',\n  declaredBinaryScale?: number,\n): PairedDecisionShape {\n  if (declaredBinaryScale !== undefined) {\n    if (!Number.isFinite(declaredBinaryScale) || declaredBinaryScale <= 0) {\n      throw new Error('pairedDecisionShape: binaryScale must be finite and positive')\n    }\n    if (statistic === 'median') {\n      throw new Error('pairedDecisionShape: binaryScale requires the mean statistic, not median')\n    }\n    for (const [name, arm] of [\n      ['before', before],\n      ['after', after],\n    ] as const) {\n      for (let i = 0; i < arm.length; i++) {\n        const value = arm[i]!\n        if (value !== 0 && value !== declaredBinaryScale) {\n          throw new Error(\n            `pairedDecisionShape: ${name}[${i}] must be 0 or binaryScale (${declaredBinaryScale}); got ${value}`,\n          )\n        }\n      }\n    }\n  }\n  const tieFraction = before.length === 0 ? null : pairedDeltaTieFraction(before, after)\n  if (statistic === 'median') {\n    return { statistic: 'median_bootstrap', binaryScale: null, tieFraction }\n  }\n  const binaryScale = declaredBinaryScale ?? pairedBinaryScale(before, after)\n  if (binaryScale !== null) {\n    return { statistic: 'paired_risk_difference', binaryScale, tieFraction }\n  }\n  return { statistic: 'mean_bootstrap', binaryScale: null, tieFraction }\n}\n\n/**\n * Decide whether a paired candidate-minus-baseline delta clears a promotion\n * threshold. `before` is the baseline arm, `after` the candidate arm, paired by\n * position. Throws on unequal lengths.\n */\nexport function decidePairedPromotion(\n  before: number[],\n  after: number[],\n  options: PairedPromotionDecisionOptions = {},\n): PairedPromotionDecision {\n  if (before.length !== after.length) {\n    throw new Error(\n      `decidePairedPromotion: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  const threshold = options.threshold ?? 0\n  if (!Number.isFinite(threshold)) {\n    throw new Error(`decidePairedPromotion: threshold must be finite, got ${threshold}`)\n  }\n  const confidence = options.confidence ?? 0.95\n  const exactMinimum = minimumPairsForPairedDeltaTest(confidence)\n  const requestedMinimum = options.minPairs ?? exactMinimum\n  if (!Number.isInteger(requestedMinimum) || requestedMinimum < 1) {\n    throw new Error(\n      `decidePairedPromotion: minPairs must be a positive integer, got ${requestedMinimum}`,\n    )\n  }\n  const { binaryScale, tieFraction } = pairedDecisionShape(\n    before,\n    after,\n    options.statistic,\n    options.binaryScale,\n  )\n  const estimatorMinimum =\n    binaryScale === null && options.statistic !== 'median' ? BOOTSTRAP_GATE_MIN_N : exactMinimum\n  const minimumPairs = Math.max(requestedMinimum, exactMinimum, estimatorMinimum)\n  const n = before.length\n  const sufficient = n >= minimumPairs\n\n  let core: {\n    statistic: PairedDecisionStatistic\n    method: PairedDecisionMethod\n    delta: number\n    low: number\n    high: number\n    bootstrap: PairedBootstrapResult | null\n    mcnemar: PairedMcNemarEvidence | null\n    pValue: number | null\n    clearsThreshold: boolean\n    label: PairedPromotionDecision['label']\n    methodDetail: string\n  }\n\n  if (binaryScale !== null) {\n    // Normalise the two-point encoding to {0,1} so the estimators see the\n    // pass/fail structure, then rescale the answer back into the caller's\n    // native units — the threshold is read in the units of the scores, so a\n    // 0-100 pass/fail dimension must be gated in points, not in a rate.\n    const unitControl = before.map((v) => v / binaryScale)\n    const unitTreatment = after.map((v) => v / binaryScale)\n    // TWO estimators with different jobs, because no single one does both.\n    // `exact` is the authority on RD = 0 and supplies the veto; its\n    // Clopper-Pearson interval conditions on the discordant count and is\n    // therefore NOT a confidence interval at a nonzero margin. `score` is\n    // Tango's, which is.\n    const exact = pairedRiskDifferenceExact(unitControl, unitTreatment, confidence)\n    const score = pairedRiskDifferenceScore(unitControl, unitTreatment, confidence)\n    const low = score.lower * binaryScale\n    core = {\n      statistic: 'paired_risk_difference',\n      method: 'score-interval',\n      delta: score.riskDifference * binaryScale,\n      low,\n      high: score.upper * binaryScale,\n      bootstrap: null,\n      mcnemar: {\n        b: exact.b,\n        c: exact.c,\n        nDiscordant: exact.nDiscordant,\n        pValue: exact.pValue,\n      },\n      pValue: null,\n      clearsThreshold: low > threshold,\n      label: 'success-rate',\n      methodDetail: '',\n    }\n  } else {\n    const bootstrapStatistic = options.statistic === 'median' ? 'median' : 'mean'\n    const test = pairedDeltaTest(before, after, {\n      confidence,\n      resamples: options.resamples,\n      statistic: bootstrapStatistic,\n      seed: options.seed,\n      threshold,\n      minPairs: minimumPairs,\n    })\n    const ci = test.bootstrap\n    core = {\n      statistic: bootstrapStatistic === 'mean' ? 'mean_bootstrap' : 'median_bootstrap',\n      method: test.method,\n      delta: bootstrapStatistic === 'mean' ? ci.mean : ci.median,\n      low: ci.low,\n      high: ci.high,\n      bootstrap: ci,\n      mcnemar: null,\n      pValue: test.pValue,\n      clearsThreshold: test.significant,\n      label: bootstrapStatistic,\n      methodDetail:\n        test.method === 'exact-sign'\n          ? bootstrapStatistic === 'mean'\n            ? ` The mean requires ${minimumPairs} pairs for bootstrap eligibility;` +\n              ` the exact sign-test p=${fmt(test.pValue ?? 1)} does not establish a mean effect.`\n            : ` Below ${BOOTSTRAP_GATE_MIN_N} pairs the interval is descriptive only;` +\n              ` the median decision uses the exact one-sided sign test, p=${fmt(test.pValue ?? 1)}.`\n          : '',\n    }\n  }\n\n  // Equivalent deltas can differ by rounding after subtraction and averaging.\n  // Use the interval's own scale so tiny score units retain the same decision\n  // as a rescaled copy of the evidence.\n  const intervalScale = Math.max(Math.abs(core.low), Math.abs(core.high))\n  const intervalTolerance = intervalScale * Number.EPSILON * 8\n  const indeterminate =\n    !Number.isFinite(core.low) ||\n    !Number.isFinite(core.high) ||\n    core.high - core.low <= intervalTolerance\n  const indeterminateCause = !indeterminate\n    ? ''\n    : tieFraction === 1\n      ? 'every paired delta is an exact tie'\n      : core.mcnemar !== null && core.mcnemar.nDiscordant === 0\n        ? 'every pair is concordant (0 discordant pairs)'\n        : `the ${core.label} CI collapsed to a point at ${fmt(core.low)}`\n  // Only at a non-negative threshold: a negative threshold asks a\n  // noninferiority question, which McNemar's test of \"no difference\" does not\n  // answer.\n  const exactTestVetoes =\n    core.mcnemar !== null && threshold >= 0 && !(core.mcnemar.pValue < 1 - confidence)\n\n  return {\n    n,\n    threshold,\n    confidence,\n    binaryScale,\n    tieFraction,\n    minimumPairs,\n    sufficient,\n    indeterminate,\n    indeterminateCause,\n    exactTestVetoes,\n    promote: sufficient && !indeterminate && core.clearsThreshold && !exactTestVetoes,\n    ...core,\n  }\n}\n\nfunction fmt(x: number): string {\n  return x.toFixed(4)\n}\n","import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return validateRunRecord({\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;AAgCA,SAAgB,+BAA+B,aAAa,KAAc;CACxE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MACR,oEAAoE,YACtE;CAEF,MAAM,iBAAiB,IAAI,cAAc;CACzC,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI,aAAa,CAAC;AAC/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,SAAgB,gBACd,QACA,OACA,UAAkC,CAAC,GACZ;CACvB,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,kDAAkD,WAAW;CAG/E,MAAM,eAAe,+BADF,QAAQ,cAAc,GACqB;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MAAM,6DAA6D,kBAAkB;CAEjG,MAAM,eAAe,KAAK,IAAI,kBAAkB,YAAY;CAC5D,MAAM,YAAY,gBAAgB,QAAQ,OAAO,OAAO;CACxD,MAAM,aAAa,UAAU,KAAK;CAIlC,MAAM,gBACJ,UAAU,IAAI,MACb,CAAC,OAAO,SAAS,UAAU,GAAG,KAC7B,CAAC,OAAO,SAAS,UAAU,IAAI,KAC/B,UAAU,QAAQ,UAAU;CAEhC,IAAI,UAAU,cACZ,OAAO;EACL;EACA,QAAQ;EACR,QAAQ;EACR;EACA;EACA;EACA,aAAa,cAAc,CAAC,iBAAiB,UAAU,MAAM;CAC/D;CAIF,MAAM,QAAQ,eADM,OAAO,KAAK,OAAO,UAAU,MAAM,SAAU,QAAQ,SAClC,GAAG,SAAS;CACnD,MAAM,WAAW,QAAQ,cAAc,SAAS,UAAU,OAAO,UAAU;CAC3E,OAAO;EACL;EACA,QAAQ;EACR,QAAQ,MAAM;EACd;EACA;EACA;EACA,aACE,cACA,CAAC,iBACD,WAAW,aACX,MAAM,WAAW,IAAI,UAAU,cAAc;CACjD;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC2CA,SAAgB,oBACd,QACA,OACA,YAA+B,QAC/B,qBACqB;CACrB,IAAI,wBAAwB,KAAA,GAAW;EACrC,IAAI,CAAC,OAAO,SAAS,mBAAmB,KAAK,uBAAuB,GAClE,MAAM,IAAI,MAAM,8DAA8D;EAEhF,IAAI,cAAc,UAChB,MAAM,IAAI,MAAM,0EAA0E;EAE5F,KAAK,MAAM,CAAC,MAAM,QAAQ,CACxB,CAAC,UAAU,MAAM,GACjB,CAAC,SAAS,KAAK,CACjB,GACE,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;GACnC,MAAM,QAAQ,IAAI;GAClB,IAAI,UAAU,KAAK,UAAU,qBAC3B,MAAM,IAAI,MACR,wBAAwB,KAAK,GAAG,EAAE,8BAA8B,oBAAoB,SAAS,OAC/F;EAEJ;CAEJ;CACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,uBAAuB,QAAQ,KAAK;CACrF,IAAI,cAAc,UAChB,OAAO;EAAE,WAAW;EAAoB,aAAa;EAAM;CAAY;CAEzE,MAAM,cAAc,uBAAuB,kBAAkB,QAAQ,KAAK;CAC1E,IAAI,gBAAgB,MAClB,OAAO;EAAE,WAAW;EAA0B;EAAa;CAAY;CAEzE,OAAO;EAAE,WAAW;EAAkB,aAAa;EAAM;CAAY;AACvE;;;;;;AAOA,SAAgB,sBACd,QACA,OACA,UAA0C,CAAC,GAClB;CACzB,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,MACR,gDAAgD,OAAO,OAAO,MAAM,MAAM,OAAO,EACnF;CAEF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,SAAS,SAAS,GAC5B,MAAM,IAAI,MAAM,wDAAwD,WAAW;CAErF,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,eAAe,+BAA+B,UAAU;CAC9D,MAAM,mBAAmB,QAAQ,YAAY;CAC7C,IAAI,CAAC,OAAO,UAAU,gBAAgB,KAAK,mBAAmB,GAC5D,MAAM,IAAI,MACR,mEAAmE,kBACrE;CAEF,MAAM,EAAE,aAAa,gBAAgB,oBACnC,QACA,OACA,QAAQ,WACR,QAAQ,WACV;CACA,MAAM,mBACJ,gBAAgB,QAAQ,QAAQ,cAAc,WAAA,KAAkC;CAClF,MAAM,eAAe,KAAK,IAAI,kBAAkB,cAAc,gBAAgB;CAC9E,MAAM,IAAI,OAAO;CACjB,MAAM,aAAa,KAAK;CAExB,IAAI;CAcJ,IAAI,gBAAgB,MAAM;EAKxB,MAAM,cAAc,OAAO,KAAK,MAAM,IAAI,WAAW;EACrD,MAAM,gBAAgB,MAAM,KAAK,MAAM,IAAI,WAAW;EAMtD,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,QAAQ,0BAA0B,aAAa,eAAe,UAAU;EAC9E,MAAM,MAAM,MAAM,QAAQ;EAC1B,OAAO;GACL,WAAW;GACX,QAAQ;GACR,OAAO,MAAM,iBAAiB;GAC9B;GACA,MAAM,MAAM,QAAQ;GACpB,WAAW;GACX,SAAS;IACP,GAAG,MAAM;IACT,GAAG,MAAM;IACT,aAAa,MAAM;IACnB,QAAQ,MAAM;GAChB;GACA,QAAQ;GACR,iBAAiB,MAAM;GACvB,OAAO;GACP,cAAc;EAChB;CACF,OAAO;EACL,MAAM,qBAAqB,QAAQ,cAAc,WAAW,WAAW;EACvE,MAAM,OAAO,gBAAgB,QAAQ,OAAO;GAC1C;GACA,WAAW,QAAQ;GACnB,WAAW;GACX,MAAM,QAAQ;GACd;GACA,UAAU;EACZ,CAAC;EACD,MAAM,KAAK,KAAK;EAChB,OAAO;GACL,WAAW,uBAAuB,SAAS,mBAAmB;GAC9D,QAAQ,KAAK;GACb,OAAO,uBAAuB,SAAS,GAAG,OAAO,GAAG;GACpD,KAAK,GAAG;GACR,MAAM,GAAG;GACT,WAAW;GACX,SAAS;GACT,QAAQ,KAAK;GACb,iBAAiB,KAAK;GACtB,OAAO;GACP,cACE,KAAK,WAAW,eACZ,uBAAuB,SACrB,sBAAsB,aAAa,0DACT,IAAI,KAAK,UAAU,CAAC,EAAE,sCAChD,+GAC8D,IAAI,KAAK,UAAU,CAAC,EAAE,KACtF;EACR;CACF;CAMA,MAAM,oBADgB,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,GAAG,KAAK,IAAI,KAAK,IAAI,CAC/B,IAAI,OAAO,UAAU;CAC3D,MAAM,gBACJ,CAAC,OAAO,SAAS,KAAK,GAAG,KACzB,CAAC,OAAO,SAAS,KAAK,IAAI,KAC1B,KAAK,OAAO,KAAK,OAAO;CAC1B,MAAM,qBAAqB,CAAC,gBACxB,KACA,gBAAgB,IACd,uCACA,KAAK,YAAY,QAAQ,KAAK,QAAQ,gBAAgB,IACpD,kDACA,OAAO,KAAK,MAAM,8BAA8B,IAAI,KAAK,GAAG;CAIpE,MAAM,kBACJ,KAAK,YAAY,QAAQ,aAAa,KAAK,EAAE,KAAK,QAAQ,SAAS,IAAI;CAEzE,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,SAAS,cAAc,CAAC,iBAAiB,KAAK,mBAAmB,CAAC;EAClE,GAAG;CACL;AACF;AAEA,SAAS,IAAI,GAAmB;CAC9B,OAAO,EAAE,QAAQ,CAAC;AACpB;;;;;;;;;;;;;ACtUA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}

package/dist/{sequential-DAsyV2T9.js → sequential-D_xHrBZe.js}

@@ -1,7 +1,7 @@
 import { h as manifestContentDigest } from "./agent-profile-cell-Cv6UA-W_.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { t as eProcess } from "./sequential-eprocess-D1jKoihe.js";
-import { at as pairHoldout, tt as aggregatePairedHoldout } from "./campaign-evidence-B8oF9xQ6.js";
+import { at as pairHoldout, tt as aggregatePairedHoldout } from "./campaign-evidence-PYPUftRy.js";
 //#region src/campaign/gates/sequential.ts
 /**
 * Anytime-valid sequential promotion gate — an e-process (betting
@@ -339,4 +339,4 @@ function sequentialDecide(options = {}) {
 //#endregion
 export { sequentialPairedGate as n, sequentialDecide as t };
 
-//# sourceMappingURL=sequential-DAsyV2T9.js.map
+//# sourceMappingURL=sequential-D_xHrBZe.js.map

package/dist/{sequential-DAsyV2T9.js.map → sequential-D_xHrBZe.js.map}

@@ -1 +1 @@
-{"version":3,"file":"sequential-DAsyV2T9.js","names":[],"sources":["../src/campaign/gates/sequential.ts"],"sourcesContent":["/**\n * Anytime-valid sequential promotion gate — an e-process (betting\n * test-martingale, see `eProcess` in `statistics.ts`) over paired\n * per-scenario deltas, so a campaign stops the MOMENT evidence decides\n * instead of burning a fixed-n budget. Decisions remain valid at any\n * data-dependent stopping time (Ville's inequality), which is exactly what\n * fixed-n machinery cannot offer: peeking at a bootstrap CI after every\n * observation and stopping on the first significant peek inflates type-I\n * error far beyond alpha.\n *\n * REPLACES, never layers on, a fixed-n gate. Running `heldoutSignificance`\n * or `paretoSignificanceGate` repeatedly on a growing sample and stopping\n * early is optional stopping no matter how it is dressed up; this gate is\n * the valid way to stop early. Use one or the other per evidence stream.\n *\n * Pre-registration binding: anytime validity holds only for the\n * PRE-REGISTERED statistic. When a `SignedManifest` is bound, the gate takes\n * alpha from `manifest.alpha`, the observation budget from\n * `manifest.preRegisteredN`, orients deltas by `manifest.direction`, and\n * shifts the null boundary by `manifest.minEffect` — re-deciding the same\n * stream under different parameters after seeing data would reopen optional\n * stopping under a fancier name. The manifest's content hash is verified at\n * construction with the canonical RFC 8785 digest used by signManifest.\n *\n * The supermartingale guarantee requires E[delta_t | past] ≤ the null\n * boundary under H0. Independent sampling with that mean bound suffices.\n * Exchangeability or shuffling alone does not: one fair sign repeated many\n * times is exchangeable and has marginal mean zero, but its conditional mean\n * becomes the revealed sign. A seeded shuffle cannot remove that dependence.\n * `decide(ctx)` pairs cells, then averages repetitions within each scenario\n * before testing. Pass independentUnitByScenarioId for scenarios sharing a\n * latent source or incident. Units must still satisfy the conditional-mean null.\n * Direct `observe()` callers own this sampling and aggregation contract.\n */\n\nimport { manifestContentDigest, type SignedManifest } from '../../pre-registration'\nimport { type EProcessState, eProcess, mulberry32 } from '../../statistics'\nimport type { Gate, GateContext, GateResult, GenerationRecord, Scenario } from '../types'\nimport { aggregatePairedHoldout, pairHoldout } from './statistical-heldout'\n\nexport type SequentialDecision = 'promote' | 'continue' | 'undecided-at-maxN'\n\nexport interface SequentialObservation {\n  decision: SequentialDecision\n  /** Current e-value (the betting wealth) against H0. */\n  eValue: number\n  /** Paired deltas consumed so far. */\n  n: number\n  /** Names the decision basis. For 'undecided-at-maxN' it states explicitly\n   *  that exhausting the budget is NOT evidence of no effect. */\n  reason: string\n}\n\nexport interface SequentialPairedGateOptions {\n  /** Type-I budget. With `preRegistration` bound this MUST match\n   *  `manifest.alpha` (conflict throws). Default 0.05. */\n  alpha?: number\n  /** Minimum independent paired units before a promote may fire. The stopping rule is\n   *  \"first n ≥ minN with e-value ≥ 1/alpha\" — still a valid stopping time.\n   *  Default 5. */\n  minN?: number\n  /** Pre-registered independent-unit budget (scenarios by default in decide()). Required unless `preRegistration`\n   *  supplies it via `preRegisteredN` (conflict throws). */\n  maxN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |delta| in the judge's native scale; deltas are mapped to\n   *  x = (d/scale + 1)/2 ∈ [0,1]. A delta outside ±scale throws (use\n   *  `detectScale` to pick 1 vs 100 BEFORE streaming). Default 1. */\n  scale?: number\n  /** Seed for reproducible ordering of scenario deltas in decide().\n   *  Shuffling does not establish the conditional-mean null. Default 1337. */\n  shuffleSeed?: number\n  /** Independent unit for every scenario passed to decide(). Full cell pairs\n   *  are averaged within each unit before testing. Defaults to scenario IDs.\n   *  The mapping is copied at construction. Direct observe() is unchanged. */\n  independentUnitByScenarioId?: ReadonlyMap<string, string>\n  /** Bind the pre-registered hypothesis. Verified (content hash) at\n   *  construction; alpha/maxN/direction/minEffect come FROM the manifest. */\n  preRegistration?: SignedManifest\n  /** Override the gate name in reports. */\n  name?: string\n  /** Continue the observe-stream from a `state()` snapshot taken before a\n   *  restart. The gate's parameters still resolve from this options object\n   *  (and the manifest); the snapshot must have been recorded under the same\n   *  alpha, maxBet, and null boundary, and its decision must be one this\n   *  configuration could have reached at its n, or construction throws.\n   *  `decide(ctx)` is unaffected — it always runs on a fresh stream. */\n  resume?: SequentialStreamState\n}\n\n/** Snapshot of one observe-stream: the e-process state plus the gate's own\n *  decision, which applies `minN` and `maxN` on top of the core latch. A\n *  gate rebuilt with `resume` from this value continues exactly where the\n *  snapshot was taken. */\nexport type SequentialStreamState = EProcessState & { decision: SequentialDecision }\n\nexport interface SequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>\n  extends Gate<TArtifact, TScenario> {\n  /** Streaming entry point: feed one paired independent-unit delta\n   *  (candidate − baseline, native scale). Aggregate correlated repetitions\n   *  before calling; the caller must justify E[delta_t | past] under H0.\n   *  Each gate instance carries ONE\n   *  observe-stream; `decide(ctx)` runs on its own fresh stream and never\n   *  consumes or advances this one. 'promote' is sticky; observing past the\n   *  pre-registered maxN throws (extending a finished stream after seeing\n   *  the result reopens optional stopping — start a NEW pre-registered\n   *  test). */\n  observe(delta: number): SequentialObservation\n  /** Read-only snapshot of the observe-stream. Pass it back as\n   *  `resume` to continue the stream after a restart. */\n  state(): SequentialStreamState\n}\n\ninterface ResolvedConfig {\n  alpha: number\n  minN: number\n  maxN: number\n  maxBet: number\n  scale: number\n  shuffleSeed: number\n  /** 'decrease' negates deltas so \"better\" is always positive. */\n  direction: 'increase' | 'decrease'\n  /** H0 boundary in x-space: 1/2 + minEffect/(2·scale). */\n  nullMean: number\n  minEffect: number\n}\n\n/** Sync twin of `verifyManifest`, so gate construction stays synchronous and\n *  fails loud before any observation is consumed. The digest scheme belongs to\n *  `pre-registration`; this gate only compares. */\nfunction verifyManifestSync(m: SignedManifest): boolean {\n  return manifestContentDigest(m) === m.contentHash\n}\n\nfunction resolveConfig(opts: SequentialPairedGateOptions): ResolvedConfig {\n  const m = opts.preRegistration\n  let alpha: number\n  let maxN: number\n  let direction: 'increase' | 'decrease'\n  let minEffect: number\n  if (m) {\n    if (!verifyManifestSync(m)) {\n      throw new Error(\n        `sequentialPairedGate: pre-registration manifest '${m.id}' content hash mismatch (tampered)`,\n      )\n    }\n    if (opts.alpha !== undefined && opts.alpha !== m.alpha) {\n      throw new Error(\n        `sequentialPairedGate: alpha ${opts.alpha} conflicts with pre-registered alpha ${m.alpha} — ` +\n          `the registered statistic is the only one anytime validity covers`,\n      )\n    }\n    if (opts.maxN !== undefined && opts.maxN !== m.preRegisteredN) {\n      throw new Error(\n        `sequentialPairedGate: maxN ${opts.maxN} conflicts with pre-registered N ${m.preRegisteredN}`,\n      )\n    }\n    alpha = m.alpha\n    maxN = m.preRegisteredN\n    direction = m.direction\n    minEffect = m.minEffect\n  } else {\n    alpha = opts.alpha ?? 0.05\n    if (opts.maxN === undefined) {\n      throw new Error(\n        'sequentialPairedGate: maxN is required (or bind a preRegistration manifest whose ' +\n          'preRegisteredN is the budget) — an unbounded stream has no pre-registered budget',\n      )\n    }\n    maxN = opts.maxN\n    direction = 'increase'\n    minEffect = 0\n  }\n  const scale = opts.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialPairedGate: scale must be > 0, got ${scale}`)\n  }\n  if (!Number.isInteger(maxN) || maxN < 1) {\n    throw new Error(`sequentialPairedGate: maxN must be a positive integer, got ${maxN}`)\n  }\n  const minN = opts.minN ?? 5\n  if (!Number.isInteger(minN) || minN < 1 || minN > maxN) {\n    throw new Error(\n      `sequentialPairedGate: minN must be an integer in [1, maxN=${maxN}], got ${minN}`,\n    )\n  }\n  if (!Number.isFinite(minEffect) || minEffect < 0 || minEffect >= scale) {\n    throw new Error(\n      `sequentialPairedGate: minEffect must be in [0, scale=${scale}), got ${minEffect}`,\n    )\n  }\n  // H0: mean delta ≤ minEffect. In x = (d/scale + 1)/2 space that is\n  // E[x] ≤ 1/2 + minEffect/(2·scale) — a shifted null boundary, NOT a clamp\n  // of the observations (clamping is mean-distorting for skewed deltas).\n  const nullMean = 0.5 + minEffect / (2 * scale)\n  return {\n    alpha,\n    minN,\n    maxN,\n    maxBet: opts.maxBet ?? 0.5,\n    scale,\n    shuffleSeed: opts.shuffleSeed ?? 1337,\n    direction,\n    nullMean,\n    minEffect,\n  }\n}\n\ninterface SequentialStream {\n  observe(delta: number): SequentialObservation\n  state(): SequentialStreamState\n}\n\nconst SEQUENTIAL_DECISIONS: readonly SequentialDecision[] = [\n  'promote',\n  'continue',\n  'undecided-at-maxN',\n]\n\n/** Refuse a snapshot whose gate decision this configuration could not have\n *  reached at the snapshot's n. The e-process parameters are checked by\n *  `eProcess` itself; this covers the gate's own stopping rule. */\nfunction assertResumableStream(resume: SequentialStreamState, cfg: ResolvedConfig): void {\n  const fail = (detail: string): never => {\n    throw new Error(`sequentialPairedGate: cannot resume — ${detail}`)\n  }\n  if (!SEQUENTIAL_DECISIONS.includes(resume.decision)) {\n    fail(`unknown decision '${String(resume.decision)}'`)\n  }\n  if (resume.n > cfg.maxN) {\n    fail(`snapshot n=${resume.n} exceeds the pre-registered maxN=${cfg.maxN}`)\n  }\n  const threshold = 1 / cfg.alpha\n  switch (resume.decision) {\n    case 'promote':\n      if (resume.n < cfg.minN) {\n        fail(`decision 'promote' at n=${resume.n} is below minN=${cfg.minN}`)\n      }\n      break\n    case 'undecided-at-maxN':\n      if (resume.n !== cfg.maxN) {\n        fail(`decision 'undecided-at-maxN' at n=${resume.n} does not match maxN=${cfg.maxN}`)\n      }\n      break\n    case 'continue':\n      if (resume.n >= cfg.maxN) {\n        fail(`decision 'continue' at n=${resume.n} with maxN=${cfg.maxN}; the stream is finished`)\n      }\n      if (resume.n >= cfg.minN && resume.wealth >= threshold) {\n        fail(\n          `decision 'continue' at n=${resume.n} ≥ minN=${cfg.minN} with e-value ` +\n            `${resume.wealth} ≥ 1/α=${threshold}; this stream would have promoted`,\n        )\n      }\n      break\n  }\n}\n\nfunction makeStream(cfg: ResolvedConfig, resume?: SequentialStreamState): SequentialStream {\n  if (resume !== undefined) assertResumableStream(resume, cfg)\n  const proc = eProcess({\n    alpha: cfg.alpha,\n    maxBet: cfg.maxBet,\n    nullMean: cfg.nullMean,\n    ...(resume === undefined ? {} : { resume }),\n  })\n  const threshold = 1 / cfg.alpha\n  let terminal: SequentialDecision | undefined =\n    resume === undefined || resume.decision === 'continue' ? undefined : resume.decision\n  return {\n    observe(delta: number): SequentialObservation {\n      if (terminal === 'undecided-at-maxN') {\n        throw new Error(\n          `sequentialPairedGate: pre-registered maxN=${cfg.maxN} exhausted — extending the ` +\n            'stream after seeing the result reopens optional stopping; start a NEW ' +\n            'pre-registered test',\n        )\n      }\n      if (!Number.isFinite(delta) || Math.abs(delta) > cfg.scale) {\n        throw new Error(\n          `sequentialPairedGate: delta ${delta} outside ±scale=${cfg.scale} — pass the judge's ` +\n            `native scale explicitly (detectScale helps pick 1 vs 100)`,\n        )\n      }\n      const d = cfg.direction === 'decrease' ? -delta : delta\n      const step = proc.update((d / cfg.scale + 1) / 2)\n      // Stopping rule: FIRST n ≥ minN with current wealth ≥ 1/alpha — a valid\n      // stopping time (measurable w.r.t. the past), sticky once latched. The\n      // core's own `decided` latch ignores minN, so the gate re-derives the\n      // decision from current wealth here.\n      if (terminal === undefined && step.n >= cfg.minN && step.wealth >= threshold) {\n        terminal = 'promote'\n      }\n      if (terminal === 'promote') {\n        return {\n          decision: 'promote',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} at n=${step.n} ` +\n            `(minN=${cfg.minN}): the paired improvement exceeds ${cfg.minEffect} at anytime-valid ` +\n            `level α=${cfg.alpha}`,\n        }\n      }\n      if (step.n >= cfg.maxN) {\n        terminal = 'undecided-at-maxN'\n        return {\n          decision: 'undecided-at-maxN',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `undecided at pre-registered maxN=${cfg.maxN} (e-value ${step.wealth.toFixed(2)} < ` +\n            `1/α=${threshold.toFixed(2)}). This is NOT evidence of no effect — the effect may be ` +\n            'real but smaller than this budget can detect; re-register with a larger N to test that',\n        }\n      }\n      return {\n        decision: 'continue',\n        eValue: step.wealth,\n        n: step.n,\n        reason: `e-value ${step.wealth.toFixed(2)} < 1/α=${threshold.toFixed(2)} at n=${step.n}/${cfg.maxN} — keep observing`,\n      }\n    },\n    state() {\n      return { ...proc.state(), decision: terminal ?? 'continue' }\n    },\n  }\n}\n\n/** Data-independent in-place Fisher–Yates with a seeded PRNG — the permutation\n *  depends only on the seed, never the values, so bet predictability survives. */\nfunction seededShuffle<T>(items: T[], seed: number): T[] {\n  const rng = mulberry32(seed)\n  for (let i = items.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    const tmp = items[i]!\n    items[i] = items[j]!\n    items[j] = tmp\n  }\n  return items\n}\n\n/**\n * Anytime-valid sequential paired gate. Conforms to the existing `Gate`\n * contract: decide(ctx) pairs candidate and baseline by full cellId, then\n * averages deltas within each configured unit (scenario by default). The\n * e-process consumes one observation per unit, with equal unit weights.\n * The streaming observe(delta) entry consumes caller-aggregated independent\n * units. Repetitions improve a unit's precision and do not increase n.\n *\n * Decision mapping onto the substrate's five-valued `GateDecision`:\n *   - 'promote'            → 'ship'\n *   - 'continue'           → 'need_more_work' (stream ended before maxN with\n *                            the e-value undecided — more independent units needed)\n *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT\n *                            evidence of no effect (never a silent default)\n */\nexport function sequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: SequentialPairedGateOptions,\n): SequentialPairedGate<TArtifact, TScenario> {\n  const cfg = resolveConfig(options)\n  const name = options.name ?? 'sequentialPairedGate'\n  const manifest = options.preRegistration\n  const observeStream = makeStream(cfg, options.resume)\n  const independentUnitByScenarioId =\n    options.independentUnitByScenarioId === undefined\n      ? undefined\n      : new Map(options.independentUnitByScenarioId)\n\n  return {\n    name,\n    observe(delta: number): SequentialObservation {\n      return observeStream.observe(delta)\n    },\n    state() {\n      return observeStream.state()\n    },\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      if (!ctx.baselineJudgeScores) {\n        throw new Error(\n          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]))\n      for (const id of scenarioIds) {\n        const unit = unitMap.get(id)\n        if (typeof unit !== 'string' || unit.length === 0 || unit.trim() !== unit) {\n          throw new Error(`${name}: missing independent unit for scenario '${id}'`)\n        }\n      }\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      for (let i = 0; i < paired.cellIds.length; i++) {\n        const delta = paired.after[i]! - paired.before[i]!\n        if (Math.abs(delta) > cfg.scale) {\n          throw new Error(\n            `${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`,\n          )\n        }\n      }\n      const observations = aggregatePairedHoldout(paired, unitMap)\n      const deltas = observations.after.map((after, i) => after - observations.before[i]!)\n      seededShuffle(deltas, cfg.shuffleSeed)\n\n      const stream = makeStream(cfg)\n      let last: SequentialObservation | undefined\n      for (const d of deltas) {\n        last = stream.observe(d)\n        if (last.decision !== 'continue') break\n      }\n\n      const detail = {\n        ...stream.state(),\n        minN: cfg.minN,\n        maxN: cfg.maxN,\n        scale: cfg.scale,\n        shuffleSeed: cfg.shuffleSeed,\n        direction: cfg.direction,\n        minEffect: cfg.minEffect,\n        pairedN: deltas.length,\n        pairedCellN: paired.cellIds.length,\n        observationUnit: independentUnitByScenarioId === undefined ? 'scenario' : 'registered',\n        unitIds: observations.unitIds,\n        ...(manifest ? { preRegisteredId: manifest.id, metric: manifest.metric } : {}),\n      }\n      const meanDelta =\n        deltas.length === 0 ? undefined : deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n      if (last === undefined) {\n        return {\n          decision: 'need_more_work',\n          reasons: [`${name}: no paired holdout observations — nothing to test`],\n          contributingGates: [{ name, status: 'not_evaluated', detail }],\n        }\n      }\n      const decision =\n        last.decision === 'promote'\n          ? 'ship'\n          : last.decision === 'continue'\n            ? 'need_more_work'\n            : 'hold'\n      return {\n        decision,\n        reasons: [`${name}: ${last.reason}`],\n        contributingGates: [\n          {\n            name,\n            status:\n              decision === 'ship'\n                ? 'pass'\n                : decision === 'need_more_work'\n                  ? 'not_evaluated'\n                  : 'fail',\n            detail,\n          },\n        ],\n        delta: meanDelta,\n      }\n    },\n  }\n}\n\nexport interface SequentialDecideOptions {\n  /** Type-I budget for the early-stop evidence. Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a stop may fire. Default 5. */\n  minN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |per-scenario composite delta|. Default 1. */\n  scale?: number\n}\n\nexport interface SequentialDecideFn {\n  (args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n  /** Read-only snapshot of the accumulated e-process (observability + tests). */\n  state(): EProcessState\n}\n\n/**\n * SurfaceProposer.decide adapter that stops exploration at an e-value threshold.\n * The selected candidate still requires an independent held-out decision.\n *\n * Stream: for each generation g ≥ 1, the per-scenario composite deltas of\n * generation g's top candidate vs the generation-0 top candidate (the\n * incumbent the loop set out to beat), paired by scenarioId. H0: no proposed\n * surface improves any scenario's expected composite over the incumbent —\n * under it every delta has conditional mean ≤ 0 and the e-process is valid.\n * Once wealth ≥ 1/alpha the loop stops and hands the winner to the promotion\n * gate (which re-scores on HELD-OUT data — this adapter only spends the\n * exploration budget, it never promotes).\n *\n * This is an exploration stopping heuristic. Selection of the best measured\n * candidate and reuse of measured incumbent scores generally violate the\n * conditional-mean null, even when the candidates have no true improvement.\n * More repetitions reduce noise but do not establish exact type-I control.\n * An UNDECIDED process never stops the loop — absence of a\n * crossing is NOT evidence of no effect, so the loop simply runs its normal\n * course. Calling the adapter repeatedly with a growing history consumes each\n * generation exactly once (re-feeding an already-seen record would double-count\n * evidence).\n */\nexport function sequentialDecide(options: SequentialDecideOptions = {}): SequentialDecideFn {\n  const alpha = options.alpha ?? 0.05\n  const minN = options.minN ?? 5\n  const scale = options.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialDecide: scale must be > 0, got ${scale}`)\n  }\n  const proc = eProcess({ alpha, maxBet: options.maxBet ?? 0.5, nullMean: 0.5 })\n  const threshold = 1 / alpha\n  let processedGenerations = 0\n  let reference: Map<string, number> | undefined\n  let stopped: { stop: true; reason: string } | undefined\n\n  const topCandidate = (record: GenerationRecord) => {\n    if (record.candidates.length === 0) {\n      throw new Error(\n        `sequentialDecide: generation ${record.generationIndex} has no candidates — ` +\n          'cannot extract a top candidate',\n      )\n    }\n    return record.candidates.reduce((best, candidate) => {\n      if (best.composite === null) return candidate\n      if (candidate.composite === null) return best\n      return candidate.composite > best.composite ? candidate : best\n    })\n  }\n\n  const decide: SequentialDecideFn = ({ history }) => {\n    if (stopped) return stopped\n    if (history.length === 0) return { stop: false }\n    if (reference === undefined) {\n      reference = new Map(\n        topCandidate(history[0]!).scenarios.map((s) => [s.scenarioId, s.composite]),\n      )\n    }\n    // Consume only generations not yet seen; start at 1 — generation 0 IS the\n    // reference and would contribute all-zero deltas (pure wealth dilution).\n    for (let g = Math.max(1, processedGenerations); g < history.length; g++) {\n      const top = topCandidate(history[g]!)\n      const byScenario = new Map(top.scenarios.map((s) => [s.scenarioId, s.composite]))\n      for (const [scenarioId, refComposite] of reference) {\n        const candComposite = byScenario.get(scenarioId)\n        if (candComposite === undefined) {\n          throw new Error(\n            `sequentialDecide: generation ${history[g]!.generationIndex} top candidate is missing ` +\n              `scenario '${scenarioId}' — generations must score the same scenario set to pair`,\n          )\n        }\n        const delta = candComposite - refComposite\n        if (!Number.isFinite(delta) || Math.abs(delta) > scale) {\n          throw new Error(\n            `sequentialDecide: paired delta ${delta} outside ±scale=${scale} on scenario ` +\n              `'${scenarioId}' — pass the composite scale explicitly`,\n          )\n        }\n        const step = proc.update((delta / scale + 1) / 2)\n        if (step.n >= minN && step.wealth >= threshold) {\n          stopped = {\n            stop: true,\n            reason:\n              `sequential e-process decided at generation ${history[g]!.generationIndex}: ` +\n              `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} ` +\n              'paired deltas vs the generation-0 incumbent; stop exploring and require ' +\n              'an independent held-out promotion decision',\n          }\n          processedGenerations = history.length\n          return stopped\n        }\n      }\n    }\n    processedGenerations = history.length\n    return { stop: false }\n  }\n  decide.state = () => proc.state()\n  return decide\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmIA,SAAS,mBAAmB,GAA4B;CACtD,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;AAEA,SAAS,cAAc,MAAmD;CACxE,MAAM,IAAI,KAAK;CACf,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,GAAG;EACL,IAAI,CAAC,mBAAmB,CAAC,GACvB,MAAM,IAAI,MACR,oDAAoD,EAAE,GAAG,mCAC3D;EAEF,IAAI,KAAK,UAAU,KAAA,KAAa,KAAK,UAAU,EAAE,OAC/C,MAAM,IAAI,MACR,+BAA+B,KAAK,MAAM,uCAAuC,EAAE,MAAM,oEAE3F;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,KAAK,SAAS,EAAE,gBAC7C,MAAM,IAAI,MACR,8BAA8B,KAAK,KAAK,mCAAmC,EAAE,gBAC/E;EAEF,QAAQ,EAAE;EACV,OAAO,EAAE;EACT,YAAY,EAAE;EACd,YAAY,EAAE;CAChB,OAAO;EACL,QAAQ,KAAK,SAAS;EACtB,IAAI,KAAK,SAAS,KAAA,GAChB,MAAM,IAAI,MACR,mKAEF;EAEF,OAAO,KAAK;EACZ,YAAY;EACZ,YAAY;CACd;CACA,MAAM,QAAQ,KAAK,SAAS;CAC5B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,gDAAgD,OAAO;CAEzE,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,GACpC,MAAM,IAAI,MAAM,8DAA8D,MAAM;CAEtF,MAAM,OAAO,KAAK,QAAQ;CAC1B,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,MAChD,MAAM,IAAI,MACR,6DAA6D,KAAK,SAAS,MAC7E;CAEF,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,KAAK,aAAa,OAC/D,MAAM,IAAI,MACR,wDAAwD,MAAM,SAAS,WACzE;CAKF,MAAM,WAAW,KAAM,aAAa,IAAI;CACxC,OAAO;EACL;EACA;EACA;EACA,QAAQ,KAAK,UAAU;EACvB;EACA,aAAa,KAAK,eAAe;EACjC;EACA;EACA;CACF;AACF;AAOA,MAAM,uBAAsD;CAC1D;CACA;CACA;AACF;;;;AAKA,SAAS,sBAAsB,QAA+B,KAA2B;CACvF,MAAM,QAAQ,WAA0B;EACtC,MAAM,IAAI,MAAM,yCAAyC,QAAQ;CACnE;CACA,IAAI,CAAC,qBAAqB,SAAS,OAAO,QAAQ,GAChD,KAAK,qBAAqB,OAAO,OAAO,QAAQ,EAAE,EAAE;CAEtD,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,cAAc,OAAO,EAAE,mCAAmC,IAAI,MAAM;CAE3E,MAAM,YAAY,IAAI,IAAI;CAC1B,QAAQ,OAAO,UAAf;EACE,KAAK;GACH,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,2BAA2B,OAAO,EAAE,iBAAiB,IAAI,MAAM;GAEtE;EACF,KAAK;GACH,IAAI,OAAO,MAAM,IAAI,MACnB,KAAK,qCAAqC,OAAO,EAAE,uBAAuB,IAAI,MAAM;GAEtF;EACF,KAAK;GACH,IAAI,OAAO,KAAK,IAAI,MAClB,KAAK,4BAA4B,OAAO,EAAE,aAAa,IAAI,KAAK,yBAAyB;GAE3F,IAAI,OAAO,KAAK,IAAI,QAAQ,OAAO,UAAU,WAC3C,KACE,4BAA4B,OAAO,EAAE,UAAU,IAAI,KAAK,gBACnD,OAAO,OAAO,SAAS,UAAU,kCACxC;GAEF;CACJ;AACF;AAEA,SAAS,WAAW,KAAqB,QAAkD;CACzF,IAAI,WAAW,KAAA,GAAW,sBAAsB,QAAQ,GAAG;CAC3D,MAAM,OAAO,SAAS;EACpB,OAAO,IAAI;EACX,QAAQ,IAAI;EACZ,UAAU,IAAI;EACd,GAAI,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO;CAC3C,CAAC;CACD,MAAM,YAAY,IAAI,IAAI;CAC1B,IAAI,WACF,WAAW,KAAA,KAAa,OAAO,aAAa,aAAa,KAAA,IAAY,OAAO;CAC9E,OAAO;EACL,QAAQ,OAAsC;GAC5C,IAAI,aAAa,qBACf,MAAM,IAAI,MACR,6CAA6C,IAAI,KAAK,qHAGxD;GAEF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,OACnD,MAAM,IAAI,MACR,+BAA+B,MAAM,kBAAkB,IAAI,MAAM,8EAEnE;GAEF,MAAM,IAAI,IAAI,cAAc,aAAa,CAAC,QAAQ;GAClD,MAAM,OAAO,KAAK,QAAQ,IAAI,IAAI,QAAQ,KAAK,CAAC;GAKhD,IAAI,aAAa,KAAA,KAAa,KAAK,KAAK,IAAI,QAAQ,KAAK,UAAU,WACjE,WAAW;GAEb,IAAI,aAAa,WACf,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QACE,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,SACtE,IAAI,KAAK,oCAAoC,IAAI,UAAU,4BACzD,IAAI;GACnB;GAEF,IAAI,KAAK,KAAK,IAAI,MAAM;IACtB,WAAW;IACX,OAAO;KACL,UAAU;KACV,QAAQ,KAAK;KACb,GAAG,KAAK;KACR,QACE,oCAAoC,IAAI,KAAK,YAAY,KAAK,OAAO,QAAQ,CAAC,EAAE,SACzE,UAAU,QAAQ,CAAC,EAAE;IAEhC;GACF;GACA,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QAAQ,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,GAAG,IAAI,KAAK;GACrG;EACF;EACA,QAAQ;GACN,OAAO;IAAE,GAAG,KAAK,MAAM;IAAG,UAAU,YAAY;GAAW;EAC7D;CACF;AACF;;;AAIA,SAAS,cAAiB,OAAY,MAAmB;CACvD,MAAM,MAAM,WAAW,IAAI;CAC3B,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,IAAI,GAAG,KAAK;EACzC,MAAM,IAAI,KAAK,MAAM,IAAI,KAAK,IAAI,EAAE;EACpC,MAAM,MAAM,MAAM;EAClB,MAAM,KAAK,MAAM;EACjB,MAAM,KAAK;CACb;CACA,OAAO;AACT;;;;;;;;;;;;;;;;AAiBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CACpD,MAAM,8BACJ,QAAQ,gCAAgC,KAAA,IACpC,KAAA,IACA,IAAI,IAAI,QAAQ,2BAA2B;CAEjD,OAAO;EACL;EACA,QAAQ,OAAsC;GAC5C,OAAO,cAAc,QAAQ,KAAK;EACpC;EACA,QAAQ;GACN,OAAO,cAAc,MAAM;EAC7B;EACA,MAAM,OAAO,KAA6D;GACxE,IAAI,CAAC,IAAI,qBACP,MAAM,IAAI,MACR,GAAG,KAAK,sJAEV;GAEF,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;GAC1D,MAAM,UAAU,+BAA+B,IAAI,IAAI,CAAC,GAAG,WAAW,CAAC,CAAC,KAAK,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC;GAC7F,KAAK,MAAM,MAAM,aAAa;IAC5B,MAAM,OAAO,QAAQ,IAAI,EAAE;IAC3B,IAAI,OAAO,SAAS,YAAY,KAAK,WAAW,KAAK,KAAK,KAAK,MAAM,MACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2CAA2C,GAAG,EAAE;GAE5E;GACA,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;IAC9C,MAAM,QAAQ,OAAO,MAAM,KAAM,OAAO,OAAO;IAC/C,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,OACxB,MAAM,IAAI,MACR,GAAG,KAAK,UAAU,OAAO,QAAQ,GAAG,UAAU,MAAM,kBAAkB,IAAI,OAC5E;GAEJ;GACA,MAAM,eAAe,uBAAuB,QAAQ,OAAO;GAC3D,MAAM,SAAS,aAAa,MAAM,KAAK,OAAO,MAAM,QAAQ,aAAa,OAAO,EAAG;GACnF,cAAc,QAAQ,IAAI,WAAW;GAErC,MAAM,SAAS,WAAW,GAAG;GAC7B,IAAI;GACJ,KAAK,MAAM,KAAK,QAAQ;IACtB,OAAO,OAAO,QAAQ,CAAC;IACvB,IAAI,KAAK,aAAa,YAAY;GACpC;GAEA,MAAM,SAAS;IACb,GAAG,OAAO,MAAM;IAChB,MAAM,IAAI;IACV,MAAM,IAAI;IACV,OAAO,IAAI;IACX,aAAa,IAAI;IACjB,WAAW,IAAI;IACf,WAAW,IAAI;IACf,SAAS,OAAO;IAChB,aAAa,OAAO,QAAQ;IAC5B,iBAAiB,gCAAgC,KAAA,IAAY,aAAa;IAC1E,SAAS,aAAa;IACtB,GAAI,WAAW;KAAE,iBAAiB,SAAS;KAAI,QAAQ,SAAS;IAAO,IAAI,CAAC;GAC9E;GACA,MAAM,YACJ,OAAO,WAAW,IAAI,KAAA,IAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;GAE/E,IAAI,SAAS,KAAA,GACX,OAAO;IACL,UAAU;IACV,SAAS,CAAC,GAAG,KAAK,mDAAmD;IACrE,mBAAmB,CAAC;KAAE;KAAM,QAAQ;KAAiB;IAAO,CAAC;GAC/D;GAEF,MAAM,WACJ,KAAK,aAAa,YACd,SACA,KAAK,aAAa,aAChB,mBACA;GACR,OAAO;IACL;IACA,SAAS,CAAC,GAAG,KAAK,IAAI,KAAK,QAAQ;IACnC,mBAAmB,CACjB;KACE;KACA,QACE,aAAa,SACT,SACA,aAAa,mBACX,kBACA;KACR;IACF,CACF;IACA,OAAO;GACT;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAmC,CAAC,GAAuB;CAC1F,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,4CAA4C,OAAO;CAErE,MAAM,OAAO,SAAS;EAAE;EAAO,QAAQ,QAAQ,UAAU;EAAK,UAAU;CAAI,CAAC;CAC7E,MAAM,YAAY,IAAI;CACtB,IAAI,uBAAuB;CAC3B,IAAI;CACJ,IAAI;CAEJ,MAAM,gBAAgB,WAA6B;EACjD,IAAI,OAAO,WAAW,WAAW,GAC/B,MAAM,IAAI,MACR,gCAAgC,OAAO,gBAAgB,oDAEzD;EAEF,OAAO,OAAO,WAAW,QAAQ,MAAM,cAAc;GACnD,IAAI,KAAK,cAAc,MAAM,OAAO;GACpC,IAAI,UAAU,cAAc,MAAM,OAAO;GACzC,OAAO,UAAU,YAAY,KAAK,YAAY,YAAY;EAC5D,CAAC;CACH;CAEA,MAAM,UAA8B,EAAE,cAAc;EAClD,IAAI,SAAS,OAAO;EACpB,IAAI,QAAQ,WAAW,GAAG,OAAO,EAAE,MAAM,MAAM;EAC/C,IAAI,cAAc,KAAA,GAChB,YAAY,IAAI,IACd,aAAa,QAAQ,EAAG,CAAC,CAAC,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAC5E;EAIF,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,oBAAoB,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvE,MAAM,MAAM,aAAa,QAAQ,EAAG;GACpC,MAAM,aAAa,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;GAChF,KAAK,MAAM,CAAC,YAAY,iBAAiB,WAAW;IAClD,MAAM,gBAAgB,WAAW,IAAI,UAAU;IAC/C,IAAI,kBAAkB,KAAA,GACpB,MAAM,IAAI,MACR,gCAAgC,QAAQ,EAAE,CAAE,gBAAgB,sCAC7C,WAAW,yDAC5B;IAEF,MAAM,QAAQ,gBAAgB;IAC9B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,OAC/C,MAAM,IAAI,MACR,kCAAkC,MAAM,kBAAkB,MAAM,gBAC1D,WAAW,wCACnB;IAEF,MAAM,OAAO,KAAK,QAAQ,QAAQ,QAAQ,KAAK,CAAC;IAChD,IAAI,KAAK,KAAK,QAAQ,KAAK,UAAU,WAAW;KAC9C,UAAU;MACR,MAAM;MACN,QACE,8CAA8C,QAAQ,EAAE,CAAE,gBAAgB,YAC/D,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,WAAW,KAAK,EAAE;KAGtF;KACA,uBAAuB,QAAQ;KAC/B,OAAO;IACT;GACF;EACF;EACA,uBAAuB,QAAQ;EAC/B,OAAO,EAAE,MAAM,MAAM;CACvB;CACA,OAAO,cAAc,KAAK,MAAM;CAChC,OAAO;AACT"}
+{"version":3,"file":"sequential-D_xHrBZe.js","names":[],"sources":["../src/campaign/gates/sequential.ts"],"sourcesContent":["/**\n * Anytime-valid sequential promotion gate — an e-process (betting\n * test-martingale, see `eProcess` in `statistics.ts`) over paired\n * per-scenario deltas, so a campaign stops the MOMENT evidence decides\n * instead of burning a fixed-n budget. Decisions remain valid at any\n * data-dependent stopping time (Ville's inequality), which is exactly what\n * fixed-n machinery cannot offer: peeking at a bootstrap CI after every\n * observation and stopping on the first significant peek inflates type-I\n * error far beyond alpha.\n *\n * REPLACES, never layers on, a fixed-n gate. Running `heldoutSignificance`\n * or `paretoSignificanceGate` repeatedly on a growing sample and stopping\n * early is optional stopping no matter how it is dressed up; this gate is\n * the valid way to stop early. Use one or the other per evidence stream.\n *\n * Pre-registration binding: anytime validity holds only for the\n * PRE-REGISTERED statistic. When a `SignedManifest` is bound, the gate takes\n * alpha from `manifest.alpha`, the observation budget from\n * `manifest.preRegisteredN`, orients deltas by `manifest.direction`, and\n * shifts the null boundary by `manifest.minEffect` — re-deciding the same\n * stream under different parameters after seeing data would reopen optional\n * stopping under a fancier name. The manifest's content hash is verified at\n * construction with the canonical RFC 8785 digest used by signManifest.\n *\n * The supermartingale guarantee requires E[delta_t | past] ≤ the null\n * boundary under H0. Independent sampling with that mean bound suffices.\n * Exchangeability or shuffling alone does not: one fair sign repeated many\n * times is exchangeable and has marginal mean zero, but its conditional mean\n * becomes the revealed sign. A seeded shuffle cannot remove that dependence.\n * `decide(ctx)` pairs cells, then averages repetitions within each scenario\n * before testing. Pass independentUnitByScenarioId for scenarios sharing a\n * latent source or incident. Units must still satisfy the conditional-mean null.\n * Direct `observe()` callers own this sampling and aggregation contract.\n */\n\nimport { manifestContentDigest, type SignedManifest } from '../../pre-registration'\nimport { type EProcessState, eProcess, mulberry32 } from '../../statistics'\nimport type { Gate, GateContext, GateResult, GenerationRecord, Scenario } from '../types'\nimport { aggregatePairedHoldout, pairHoldout } from './statistical-heldout'\n\nexport type SequentialDecision = 'promote' | 'continue' | 'undecided-at-maxN'\n\nexport interface SequentialObservation {\n  decision: SequentialDecision\n  /** Current e-value (the betting wealth) against H0. */\n  eValue: number\n  /** Paired deltas consumed so far. */\n  n: number\n  /** Names the decision basis. For 'undecided-at-maxN' it states explicitly\n   *  that exhausting the budget is NOT evidence of no effect. */\n  reason: string\n}\n\nexport interface SequentialPairedGateOptions {\n  /** Type-I budget. With `preRegistration` bound this MUST match\n   *  `manifest.alpha` (conflict throws). Default 0.05. */\n  alpha?: number\n  /** Minimum independent paired units before a promote may fire. The stopping rule is\n   *  \"first n ≥ minN with e-value ≥ 1/alpha\" — still a valid stopping time.\n   *  Default 5. */\n  minN?: number\n  /** Pre-registered independent-unit budget (scenarios by default in decide()). Required unless `preRegistration`\n   *  supplies it via `preRegisteredN` (conflict throws). */\n  maxN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |delta| in the judge's native scale; deltas are mapped to\n   *  x = (d/scale + 1)/2 ∈ [0,1]. A delta outside ±scale throws (use\n   *  `detectScale` to pick 1 vs 100 BEFORE streaming). Default 1. */\n  scale?: number\n  /** Seed for reproducible ordering of scenario deltas in decide().\n   *  Shuffling does not establish the conditional-mean null. Default 1337. */\n  shuffleSeed?: number\n  /** Independent unit for every scenario passed to decide(). Full cell pairs\n   *  are averaged within each unit before testing. Defaults to scenario IDs.\n   *  The mapping is copied at construction. Direct observe() is unchanged. */\n  independentUnitByScenarioId?: ReadonlyMap<string, string>\n  /** Bind the pre-registered hypothesis. Verified (content hash) at\n   *  construction; alpha/maxN/direction/minEffect come FROM the manifest. */\n  preRegistration?: SignedManifest\n  /** Override the gate name in reports. */\n  name?: string\n  /** Continue the observe-stream from a `state()` snapshot taken before a\n   *  restart. The gate's parameters still resolve from this options object\n   *  (and the manifest); the snapshot must have been recorded under the same\n   *  alpha, maxBet, and null boundary, and its decision must be one this\n   *  configuration could have reached at its n, or construction throws.\n   *  `decide(ctx)` is unaffected — it always runs on a fresh stream. */\n  resume?: SequentialStreamState\n}\n\n/** Snapshot of one observe-stream: the e-process state plus the gate's own\n *  decision, which applies `minN` and `maxN` on top of the core latch. A\n *  gate rebuilt with `resume` from this value continues exactly where the\n *  snapshot was taken. */\nexport type SequentialStreamState = EProcessState & { decision: SequentialDecision }\n\nexport interface SequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>\n  extends Gate<TArtifact, TScenario> {\n  /** Streaming entry point: feed one paired independent-unit delta\n   *  (candidate − baseline, native scale). Aggregate correlated repetitions\n   *  before calling; the caller must justify E[delta_t | past] under H0.\n   *  Each gate instance carries ONE\n   *  observe-stream; `decide(ctx)` runs on its own fresh stream and never\n   *  consumes or advances this one. 'promote' is sticky; observing past the\n   *  pre-registered maxN throws (extending a finished stream after seeing\n   *  the result reopens optional stopping — start a NEW pre-registered\n   *  test). */\n  observe(delta: number): SequentialObservation\n  /** Read-only snapshot of the observe-stream. Pass it back as\n   *  `resume` to continue the stream after a restart. */\n  state(): SequentialStreamState\n}\n\ninterface ResolvedConfig {\n  alpha: number\n  minN: number\n  maxN: number\n  maxBet: number\n  scale: number\n  shuffleSeed: number\n  /** 'decrease' negates deltas so \"better\" is always positive. */\n  direction: 'increase' | 'decrease'\n  /** H0 boundary in x-space: 1/2 + minEffect/(2·scale). */\n  nullMean: number\n  minEffect: number\n}\n\n/** Sync twin of `verifyManifest`, so gate construction stays synchronous and\n *  fails loud before any observation is consumed. The digest scheme belongs to\n *  `pre-registration`; this gate only compares. */\nfunction verifyManifestSync(m: SignedManifest): boolean {\n  return manifestContentDigest(m) === m.contentHash\n}\n\nfunction resolveConfig(opts: SequentialPairedGateOptions): ResolvedConfig {\n  const m = opts.preRegistration\n  let alpha: number\n  let maxN: number\n  let direction: 'increase' | 'decrease'\n  let minEffect: number\n  if (m) {\n    if (!verifyManifestSync(m)) {\n      throw new Error(\n        `sequentialPairedGate: pre-registration manifest '${m.id}' content hash mismatch (tampered)`,\n      )\n    }\n    if (opts.alpha !== undefined && opts.alpha !== m.alpha) {\n      throw new Error(\n        `sequentialPairedGate: alpha ${opts.alpha} conflicts with pre-registered alpha ${m.alpha} — ` +\n          `the registered statistic is the only one anytime validity covers`,\n      )\n    }\n    if (opts.maxN !== undefined && opts.maxN !== m.preRegisteredN) {\n      throw new Error(\n        `sequentialPairedGate: maxN ${opts.maxN} conflicts with pre-registered N ${m.preRegisteredN}`,\n      )\n    }\n    alpha = m.alpha\n    maxN = m.preRegisteredN\n    direction = m.direction\n    minEffect = m.minEffect\n  } else {\n    alpha = opts.alpha ?? 0.05\n    if (opts.maxN === undefined) {\n      throw new Error(\n        'sequentialPairedGate: maxN is required (or bind a preRegistration manifest whose ' +\n          'preRegisteredN is the budget) — an unbounded stream has no pre-registered budget',\n      )\n    }\n    maxN = opts.maxN\n    direction = 'increase'\n    minEffect = 0\n  }\n  const scale = opts.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialPairedGate: scale must be > 0, got ${scale}`)\n  }\n  if (!Number.isInteger(maxN) || maxN < 1) {\n    throw new Error(`sequentialPairedGate: maxN must be a positive integer, got ${maxN}`)\n  }\n  const minN = opts.minN ?? 5\n  if (!Number.isInteger(minN) || minN < 1 || minN > maxN) {\n    throw new Error(\n      `sequentialPairedGate: minN must be an integer in [1, maxN=${maxN}], got ${minN}`,\n    )\n  }\n  if (!Number.isFinite(minEffect) || minEffect < 0 || minEffect >= scale) {\n    throw new Error(\n      `sequentialPairedGate: minEffect must be in [0, scale=${scale}), got ${minEffect}`,\n    )\n  }\n  // H0: mean delta ≤ minEffect. In x = (d/scale + 1)/2 space that is\n  // E[x] ≤ 1/2 + minEffect/(2·scale) — a shifted null boundary, NOT a clamp\n  // of the observations (clamping is mean-distorting for skewed deltas).\n  const nullMean = 0.5 + minEffect / (2 * scale)\n  return {\n    alpha,\n    minN,\n    maxN,\n    maxBet: opts.maxBet ?? 0.5,\n    scale,\n    shuffleSeed: opts.shuffleSeed ?? 1337,\n    direction,\n    nullMean,\n    minEffect,\n  }\n}\n\ninterface SequentialStream {\n  observe(delta: number): SequentialObservation\n  state(): SequentialStreamState\n}\n\nconst SEQUENTIAL_DECISIONS: readonly SequentialDecision[] = [\n  'promote',\n  'continue',\n  'undecided-at-maxN',\n]\n\n/** Refuse a snapshot whose gate decision this configuration could not have\n *  reached at the snapshot's n. The e-process parameters are checked by\n *  `eProcess` itself; this covers the gate's own stopping rule. */\nfunction assertResumableStream(resume: SequentialStreamState, cfg: ResolvedConfig): void {\n  const fail = (detail: string): never => {\n    throw new Error(`sequentialPairedGate: cannot resume — ${detail}`)\n  }\n  if (!SEQUENTIAL_DECISIONS.includes(resume.decision)) {\n    fail(`unknown decision '${String(resume.decision)}'`)\n  }\n  if (resume.n > cfg.maxN) {\n    fail(`snapshot n=${resume.n} exceeds the pre-registered maxN=${cfg.maxN}`)\n  }\n  const threshold = 1 / cfg.alpha\n  switch (resume.decision) {\n    case 'promote':\n      if (resume.n < cfg.minN) {\n        fail(`decision 'promote' at n=${resume.n} is below minN=${cfg.minN}`)\n      }\n      break\n    case 'undecided-at-maxN':\n      if (resume.n !== cfg.maxN) {\n        fail(`decision 'undecided-at-maxN' at n=${resume.n} does not match maxN=${cfg.maxN}`)\n      }\n      break\n    case 'continue':\n      if (resume.n >= cfg.maxN) {\n        fail(`decision 'continue' at n=${resume.n} with maxN=${cfg.maxN}; the stream is finished`)\n      }\n      if (resume.n >= cfg.minN && resume.wealth >= threshold) {\n        fail(\n          `decision 'continue' at n=${resume.n} ≥ minN=${cfg.minN} with e-value ` +\n            `${resume.wealth} ≥ 1/α=${threshold}; this stream would have promoted`,\n        )\n      }\n      break\n  }\n}\n\nfunction makeStream(cfg: ResolvedConfig, resume?: SequentialStreamState): SequentialStream {\n  if (resume !== undefined) assertResumableStream(resume, cfg)\n  const proc = eProcess({\n    alpha: cfg.alpha,\n    maxBet: cfg.maxBet,\n    nullMean: cfg.nullMean,\n    ...(resume === undefined ? {} : { resume }),\n  })\n  const threshold = 1 / cfg.alpha\n  let terminal: SequentialDecision | undefined =\n    resume === undefined || resume.decision === 'continue' ? undefined : resume.decision\n  return {\n    observe(delta: number): SequentialObservation {\n      if (terminal === 'undecided-at-maxN') {\n        throw new Error(\n          `sequentialPairedGate: pre-registered maxN=${cfg.maxN} exhausted — extending the ` +\n            'stream after seeing the result reopens optional stopping; start a NEW ' +\n            'pre-registered test',\n        )\n      }\n      if (!Number.isFinite(delta) || Math.abs(delta) > cfg.scale) {\n        throw new Error(\n          `sequentialPairedGate: delta ${delta} outside ±scale=${cfg.scale} — pass the judge's ` +\n            `native scale explicitly (detectScale helps pick 1 vs 100)`,\n        )\n      }\n      const d = cfg.direction === 'decrease' ? -delta : delta\n      const step = proc.update((d / cfg.scale + 1) / 2)\n      // Stopping rule: FIRST n ≥ minN with current wealth ≥ 1/alpha — a valid\n      // stopping time (measurable w.r.t. the past), sticky once latched. The\n      // core's own `decided` latch ignores minN, so the gate re-derives the\n      // decision from current wealth here.\n      if (terminal === undefined && step.n >= cfg.minN && step.wealth >= threshold) {\n        terminal = 'promote'\n      }\n      if (terminal === 'promote') {\n        return {\n          decision: 'promote',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} at n=${step.n} ` +\n            `(minN=${cfg.minN}): the paired improvement exceeds ${cfg.minEffect} at anytime-valid ` +\n            `level α=${cfg.alpha}`,\n        }\n      }\n      if (step.n >= cfg.maxN) {\n        terminal = 'undecided-at-maxN'\n        return {\n          decision: 'undecided-at-maxN',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `undecided at pre-registered maxN=${cfg.maxN} (e-value ${step.wealth.toFixed(2)} < ` +\n            `1/α=${threshold.toFixed(2)}). This is NOT evidence of no effect — the effect may be ` +\n            'real but smaller than this budget can detect; re-register with a larger N to test that',\n        }\n      }\n      return {\n        decision: 'continue',\n        eValue: step.wealth,\n        n: step.n,\n        reason: `e-value ${step.wealth.toFixed(2)} < 1/α=${threshold.toFixed(2)} at n=${step.n}/${cfg.maxN} — keep observing`,\n      }\n    },\n    state() {\n      return { ...proc.state(), decision: terminal ?? 'continue' }\n    },\n  }\n}\n\n/** Data-independent in-place Fisher–Yates with a seeded PRNG — the permutation\n *  depends only on the seed, never the values, so bet predictability survives. */\nfunction seededShuffle<T>(items: T[], seed: number): T[] {\n  const rng = mulberry32(seed)\n  for (let i = items.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    const tmp = items[i]!\n    items[i] = items[j]!\n    items[j] = tmp\n  }\n  return items\n}\n\n/**\n * Anytime-valid sequential paired gate. Conforms to the existing `Gate`\n * contract: decide(ctx) pairs candidate and baseline by full cellId, then\n * averages deltas within each configured unit (scenario by default). The\n * e-process consumes one observation per unit, with equal unit weights.\n * The streaming observe(delta) entry consumes caller-aggregated independent\n * units. Repetitions improve a unit's precision and do not increase n.\n *\n * Decision mapping onto the substrate's five-valued `GateDecision`:\n *   - 'promote'            → 'ship'\n *   - 'continue'           → 'need_more_work' (stream ended before maxN with\n *                            the e-value undecided — more independent units needed)\n *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT\n *                            evidence of no effect (never a silent default)\n */\nexport function sequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: SequentialPairedGateOptions,\n): SequentialPairedGate<TArtifact, TScenario> {\n  const cfg = resolveConfig(options)\n  const name = options.name ?? 'sequentialPairedGate'\n  const manifest = options.preRegistration\n  const observeStream = makeStream(cfg, options.resume)\n  const independentUnitByScenarioId =\n    options.independentUnitByScenarioId === undefined\n      ? undefined\n      : new Map(options.independentUnitByScenarioId)\n\n  return {\n    name,\n    observe(delta: number): SequentialObservation {\n      return observeStream.observe(delta)\n    },\n    state() {\n      return observeStream.state()\n    },\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      if (!ctx.baselineJudgeScores) {\n        throw new Error(\n          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]))\n      for (const id of scenarioIds) {\n        const unit = unitMap.get(id)\n        if (typeof unit !== 'string' || unit.length === 0 || unit.trim() !== unit) {\n          throw new Error(`${name}: missing independent unit for scenario '${id}'`)\n        }\n      }\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      for (let i = 0; i < paired.cellIds.length; i++) {\n        const delta = paired.after[i]! - paired.before[i]!\n        if (Math.abs(delta) > cfg.scale) {\n          throw new Error(\n            `${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`,\n          )\n        }\n      }\n      const observations = aggregatePairedHoldout(paired, unitMap)\n      const deltas = observations.after.map((after, i) => after - observations.before[i]!)\n      seededShuffle(deltas, cfg.shuffleSeed)\n\n      const stream = makeStream(cfg)\n      let last: SequentialObservation | undefined\n      for (const d of deltas) {\n        last = stream.observe(d)\n        if (last.decision !== 'continue') break\n      }\n\n      const detail = {\n        ...stream.state(),\n        minN: cfg.minN,\n        maxN: cfg.maxN,\n        scale: cfg.scale,\n        shuffleSeed: cfg.shuffleSeed,\n        direction: cfg.direction,\n        minEffect: cfg.minEffect,\n        pairedN: deltas.length,\n        pairedCellN: paired.cellIds.length,\n        observationUnit: independentUnitByScenarioId === undefined ? 'scenario' : 'registered',\n        unitIds: observations.unitIds,\n        ...(manifest ? { preRegisteredId: manifest.id, metric: manifest.metric } : {}),\n      }\n      const meanDelta =\n        deltas.length === 0 ? undefined : deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n      if (last === undefined) {\n        return {\n          decision: 'need_more_work',\n          reasons: [`${name}: no paired holdout observations — nothing to test`],\n          contributingGates: [{ name, status: 'not_evaluated', detail }],\n        }\n      }\n      const decision =\n        last.decision === 'promote'\n          ? 'ship'\n          : last.decision === 'continue'\n            ? 'need_more_work'\n            : 'hold'\n      return {\n        decision,\n        reasons: [`${name}: ${last.reason}`],\n        contributingGates: [\n          {\n            name,\n            status:\n              decision === 'ship'\n                ? 'pass'\n                : decision === 'need_more_work'\n                  ? 'not_evaluated'\n                  : 'fail',\n            detail,\n          },\n        ],\n        delta: meanDelta,\n      }\n    },\n  }\n}\n\nexport interface SequentialDecideOptions {\n  /** Type-I budget for the early-stop evidence. Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a stop may fire. Default 5. */\n  minN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |per-scenario composite delta|. Default 1. */\n  scale?: number\n}\n\nexport interface SequentialDecideFn {\n  (args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n  /** Read-only snapshot of the accumulated e-process (observability + tests). */\n  state(): EProcessState\n}\n\n/**\n * SurfaceProposer.decide adapter that stops exploration at an e-value threshold.\n * The selected candidate still requires an independent held-out decision.\n *\n * Stream: for each generation g ≥ 1, the per-scenario composite deltas of\n * generation g's top candidate vs the generation-0 top candidate (the\n * incumbent the loop set out to beat), paired by scenarioId. H0: no proposed\n * surface improves any scenario's expected composite over the incumbent —\n * under it every delta has conditional mean ≤ 0 and the e-process is valid.\n * Once wealth ≥ 1/alpha the loop stops and hands the winner to the promotion\n * gate (which re-scores on HELD-OUT data — this adapter only spends the\n * exploration budget, it never promotes).\n *\n * This is an exploration stopping heuristic. Selection of the best measured\n * candidate and reuse of measured incumbent scores generally violate the\n * conditional-mean null, even when the candidates have no true improvement.\n * More repetitions reduce noise but do not establish exact type-I control.\n * An UNDECIDED process never stops the loop — absence of a\n * crossing is NOT evidence of no effect, so the loop simply runs its normal\n * course. Calling the adapter repeatedly with a growing history consumes each\n * generation exactly once (re-feeding an already-seen record would double-count\n * evidence).\n */\nexport function sequentialDecide(options: SequentialDecideOptions = {}): SequentialDecideFn {\n  const alpha = options.alpha ?? 0.05\n  const minN = options.minN ?? 5\n  const scale = options.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialDecide: scale must be > 0, got ${scale}`)\n  }\n  const proc = eProcess({ alpha, maxBet: options.maxBet ?? 0.5, nullMean: 0.5 })\n  const threshold = 1 / alpha\n  let processedGenerations = 0\n  let reference: Map<string, number> | undefined\n  let stopped: { stop: true; reason: string } | undefined\n\n  const topCandidate = (record: GenerationRecord) => {\n    if (record.candidates.length === 0) {\n      throw new Error(\n        `sequentialDecide: generation ${record.generationIndex} has no candidates — ` +\n          'cannot extract a top candidate',\n      )\n    }\n    return record.candidates.reduce((best, candidate) => {\n      if (best.composite === null) return candidate\n      if (candidate.composite === null) return best\n      return candidate.composite > best.composite ? candidate : best\n    })\n  }\n\n  const decide: SequentialDecideFn = ({ history }) => {\n    if (stopped) return stopped\n    if (history.length === 0) return { stop: false }\n    if (reference === undefined) {\n      reference = new Map(\n        topCandidate(history[0]!).scenarios.map((s) => [s.scenarioId, s.composite]),\n      )\n    }\n    // Consume only generations not yet seen; start at 1 — generation 0 IS the\n    // reference and would contribute all-zero deltas (pure wealth dilution).\n    for (let g = Math.max(1, processedGenerations); g < history.length; g++) {\n      const top = topCandidate(history[g]!)\n      const byScenario = new Map(top.scenarios.map((s) => [s.scenarioId, s.composite]))\n      for (const [scenarioId, refComposite] of reference) {\n        const candComposite = byScenario.get(scenarioId)\n        if (candComposite === undefined) {\n          throw new Error(\n            `sequentialDecide: generation ${history[g]!.generationIndex} top candidate is missing ` +\n              `scenario '${scenarioId}' — generations must score the same scenario set to pair`,\n          )\n        }\n        const delta = candComposite - refComposite\n        if (!Number.isFinite(delta) || Math.abs(delta) > scale) {\n          throw new Error(\n            `sequentialDecide: paired delta ${delta} outside ±scale=${scale} on scenario ` +\n              `'${scenarioId}' — pass the composite scale explicitly`,\n          )\n        }\n        const step = proc.update((delta / scale + 1) / 2)\n        if (step.n >= minN && step.wealth >= threshold) {\n          stopped = {\n            stop: true,\n            reason:\n              `sequential e-process decided at generation ${history[g]!.generationIndex}: ` +\n              `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} ` +\n              'paired deltas vs the generation-0 incumbent; stop exploring and require ' +\n              'an independent held-out promotion decision',\n          }\n          processedGenerations = history.length\n          return stopped\n        }\n      }\n    }\n    processedGenerations = history.length\n    return { stop: false }\n  }\n  decide.state = () => proc.state()\n  return decide\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmIA,SAAS,mBAAmB,GAA4B;CACtD,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;AAEA,SAAS,cAAc,MAAmD;CACxE,MAAM,IAAI,KAAK;CACf,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,GAAG;EACL,IAAI,CAAC,mBAAmB,CAAC,GACvB,MAAM,IAAI,MACR,oDAAoD,EAAE,GAAG,mCAC3D;EAEF,IAAI,KAAK,UAAU,KAAA,KAAa,KAAK,UAAU,EAAE,OAC/C,MAAM,IAAI,MACR,+BAA+B,KAAK,MAAM,uCAAuC,EAAE,MAAM,oEAE3F;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,KAAK,SAAS,EAAE,gBAC7C,MAAM,IAAI,MACR,8BAA8B,KAAK,KAAK,mCAAmC,EAAE,gBAC/E;EAEF,QAAQ,EAAE;EACV,OAAO,EAAE;EACT,YAAY,EAAE;EACd,YAAY,EAAE;CAChB,OAAO;EACL,QAAQ,KAAK,SAAS;EACtB,IAAI,KAAK,SAAS,KAAA,GAChB,MAAM,IAAI,MACR,mKAEF;EAEF,OAAO,KAAK;EACZ,YAAY;EACZ,YAAY;CACd;CACA,MAAM,QAAQ,KAAK,SAAS;CAC5B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,gDAAgD,OAAO;CAEzE,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,GACpC,MAAM,IAAI,MAAM,8DAA8D,MAAM;CAEtF,MAAM,OAAO,KAAK,QAAQ;CAC1B,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,MAChD,MAAM,IAAI,MACR,6DAA6D,KAAK,SAAS,MAC7E;CAEF,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,KAAK,aAAa,OAC/D,MAAM,IAAI,MACR,wDAAwD,MAAM,SAAS,WACzE;CAKF,MAAM,WAAW,KAAM,aAAa,IAAI;CACxC,OAAO;EACL;EACA;EACA;EACA,QAAQ,KAAK,UAAU;EACvB;EACA,aAAa,KAAK,eAAe;EACjC;EACA;EACA;CACF;AACF;AAOA,MAAM,uBAAsD;CAC1D;CACA;CACA;AACF;;;;AAKA,SAAS,sBAAsB,QAA+B,KAA2B;CACvF,MAAM,QAAQ,WAA0B;EACtC,MAAM,IAAI,MAAM,yCAAyC,QAAQ;CACnE;CACA,IAAI,CAAC,qBAAqB,SAAS,OAAO,QAAQ,GAChD,KAAK,qBAAqB,OAAO,OAAO,QAAQ,EAAE,EAAE;CAEtD,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,cAAc,OAAO,EAAE,mCAAmC,IAAI,MAAM;CAE3E,MAAM,YAAY,IAAI,IAAI;CAC1B,QAAQ,OAAO,UAAf;EACE,KAAK;GACH,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,2BAA2B,OAAO,EAAE,iBAAiB,IAAI,MAAM;GAEtE;EACF,KAAK;GACH,IAAI,OAAO,MAAM,IAAI,MACnB,KAAK,qCAAqC,OAAO,EAAE,uBAAuB,IAAI,MAAM;GAEtF;EACF,KAAK;GACH,IAAI,OAAO,KAAK,IAAI,MAClB,KAAK,4BAA4B,OAAO,EAAE,aAAa,IAAI,KAAK,yBAAyB;GAE3F,IAAI,OAAO,KAAK,IAAI,QAAQ,OAAO,UAAU,WAC3C,KACE,4BAA4B,OAAO,EAAE,UAAU,IAAI,KAAK,gBACnD,OAAO,OAAO,SAAS,UAAU,kCACxC;GAEF;CACJ;AACF;AAEA,SAAS,WAAW,KAAqB,QAAkD;CACzF,IAAI,WAAW,KAAA,GAAW,sBAAsB,QAAQ,GAAG;CAC3D,MAAM,OAAO,SAAS;EACpB,OAAO,IAAI;EACX,QAAQ,IAAI;EACZ,UAAU,IAAI;EACd,GAAI,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO;CAC3C,CAAC;CACD,MAAM,YAAY,IAAI,IAAI;CAC1B,IAAI,WACF,WAAW,KAAA,KAAa,OAAO,aAAa,aAAa,KAAA,IAAY,OAAO;CAC9E,OAAO;EACL,QAAQ,OAAsC;GAC5C,IAAI,aAAa,qBACf,MAAM,IAAI,MACR,6CAA6C,IAAI,KAAK,qHAGxD;GAEF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,OACnD,MAAM,IAAI,MACR,+BAA+B,MAAM,kBAAkB,IAAI,MAAM,8EAEnE;GAEF,MAAM,IAAI,IAAI,cAAc,aAAa,CAAC,QAAQ;GAClD,MAAM,OAAO,KAAK,QAAQ,IAAI,IAAI,QAAQ,KAAK,CAAC;GAKhD,IAAI,aAAa,KAAA,KAAa,KAAK,KAAK,IAAI,QAAQ,KAAK,UAAU,WACjE,WAAW;GAEb,IAAI,aAAa,WACf,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QACE,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,SACtE,IAAI,KAAK,oCAAoC,IAAI,UAAU,4BACzD,IAAI;GACnB;GAEF,IAAI,KAAK,KAAK,IAAI,MAAM;IACtB,WAAW;IACX,OAAO;KACL,UAAU;KACV,QAAQ,KAAK;KACb,GAAG,KAAK;KACR,QACE,oCAAoC,IAAI,KAAK,YAAY,KAAK,OAAO,QAAQ,CAAC,EAAE,SACzE,UAAU,QAAQ,CAAC,EAAE;IAEhC;GACF;GACA,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QAAQ,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,GAAG,IAAI,KAAK;GACrG;EACF;EACA,QAAQ;GACN,OAAO;IAAE,GAAG,KAAK,MAAM;IAAG,UAAU,YAAY;GAAW;EAC7D;CACF;AACF;;;AAIA,SAAS,cAAiB,OAAY,MAAmB;CACvD,MAAM,MAAM,WAAW,IAAI;CAC3B,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,IAAI,GAAG,KAAK;EACzC,MAAM,IAAI,KAAK,MAAM,IAAI,KAAK,IAAI,EAAE;EACpC,MAAM,MAAM,MAAM;EAClB,MAAM,KAAK,MAAM;EACjB,MAAM,KAAK;CACb;CACA,OAAO;AACT;;;;;;;;;;;;;;;;AAiBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CACpD,MAAM,8BACJ,QAAQ,gCAAgC,KAAA,IACpC,KAAA,IACA,IAAI,IAAI,QAAQ,2BAA2B;CAEjD,OAAO;EACL;EACA,QAAQ,OAAsC;GAC5C,OAAO,cAAc,QAAQ,KAAK;EACpC;EACA,QAAQ;GACN,OAAO,cAAc,MAAM;EAC7B;EACA,MAAM,OAAO,KAA6D;GACxE,IAAI,CAAC,IAAI,qBACP,MAAM,IAAI,MACR,GAAG,KAAK,sJAEV;GAEF,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;GAC1D,MAAM,UAAU,+BAA+B,IAAI,IAAI,CAAC,GAAG,WAAW,CAAC,CAAC,KAAK,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC;GAC7F,KAAK,MAAM,MAAM,aAAa;IAC5B,MAAM,OAAO,QAAQ,IAAI,EAAE;IAC3B,IAAI,OAAO,SAAS,YAAY,KAAK,WAAW,KAAK,KAAK,KAAK,MAAM,MACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2CAA2C,GAAG,EAAE;GAE5E;GACA,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;IAC9C,MAAM,QAAQ,OAAO,MAAM,KAAM,OAAO,OAAO;IAC/C,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,OACxB,MAAM,IAAI,MACR,GAAG,KAAK,UAAU,OAAO,QAAQ,GAAG,UAAU,MAAM,kBAAkB,IAAI,OAC5E;GAEJ;GACA,MAAM,eAAe,uBAAuB,QAAQ,OAAO;GAC3D,MAAM,SAAS,aAAa,MAAM,KAAK,OAAO,MAAM,QAAQ,aAAa,OAAO,EAAG;GACnF,cAAc,QAAQ,IAAI,WAAW;GAErC,MAAM,SAAS,WAAW,GAAG;GAC7B,IAAI;GACJ,KAAK,MAAM,KAAK,QAAQ;IACtB,OAAO,OAAO,QAAQ,CAAC;IACvB,IAAI,KAAK,aAAa,YAAY;GACpC;GAEA,MAAM,SAAS;IACb,GAAG,OAAO,MAAM;IAChB,MAAM,IAAI;IACV,MAAM,IAAI;IACV,OAAO,IAAI;IACX,aAAa,IAAI;IACjB,WAAW,IAAI;IACf,WAAW,IAAI;IACf,SAAS,OAAO;IAChB,aAAa,OAAO,QAAQ;IAC5B,iBAAiB,gCAAgC,KAAA,IAAY,aAAa;IAC1E,SAAS,aAAa;IACtB,GAAI,WAAW;KAAE,iBAAiB,SAAS;KAAI,QAAQ,SAAS;IAAO,IAAI,CAAC;GAC9E;GACA,MAAM,YACJ,OAAO,WAAW,IAAI,KAAA,IAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;GAE/E,IAAI,SAAS,KAAA,GACX,OAAO;IACL,UAAU;IACV,SAAS,CAAC,GAAG,KAAK,mDAAmD;IACrE,mBAAmB,CAAC;KAAE;KAAM,QAAQ;KAAiB;IAAO,CAAC;GAC/D;GAEF,MAAM,WACJ,KAAK,aAAa,YACd,SACA,KAAK,aAAa,aAChB,mBACA;GACR,OAAO;IACL;IACA,SAAS,CAAC,GAAG,KAAK,IAAI,KAAK,QAAQ;IACnC,mBAAmB,CACjB;KACE;KACA,QACE,aAAa,SACT,SACA,aAAa,mBACX,kBACA;KACR;IACF,CACF;IACA,OAAO;GACT;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAmC,CAAC,GAAuB;CAC1F,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,4CAA4C,OAAO;CAErE,MAAM,OAAO,SAAS;EAAE;EAAO,QAAQ,QAAQ,UAAU;EAAK,UAAU;CAAI,CAAC;CAC7E,MAAM,YAAY,IAAI;CACtB,IAAI,uBAAuB;CAC3B,IAAI;CACJ,IAAI;CAEJ,MAAM,gBAAgB,WAA6B;EACjD,IAAI,OAAO,WAAW,WAAW,GAC/B,MAAM,IAAI,MACR,gCAAgC,OAAO,gBAAgB,oDAEzD;EAEF,OAAO,OAAO,WAAW,QAAQ,MAAM,cAAc;GACnD,IAAI,KAAK,cAAc,MAAM,OAAO;GACpC,IAAI,UAAU,cAAc,MAAM,OAAO;GACzC,OAAO,UAAU,YAAY,KAAK,YAAY,YAAY;EAC5D,CAAC;CACH;CAEA,MAAM,UAA8B,EAAE,cAAc;EAClD,IAAI,SAAS,OAAO;EACpB,IAAI,QAAQ,WAAW,GAAG,OAAO,EAAE,MAAM,MAAM;EAC/C,IAAI,cAAc,KAAA,GAChB,YAAY,IAAI,IACd,aAAa,QAAQ,EAAG,CAAC,CAAC,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAC5E;EAIF,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,oBAAoB,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvE,MAAM,MAAM,aAAa,QAAQ,EAAG;GACpC,MAAM,aAAa,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;GAChF,KAAK,MAAM,CAAC,YAAY,iBAAiB,WAAW;IAClD,MAAM,gBAAgB,WAAW,IAAI,UAAU;IAC/C,IAAI,kBAAkB,KAAA,GACpB,MAAM,IAAI,MACR,gCAAgC,QAAQ,EAAE,CAAE,gBAAgB,sCAC7C,WAAW,yDAC5B;IAEF,MAAM,QAAQ,gBAAgB;IAC9B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,OAC/C,MAAM,IAAI,MACR,kCAAkC,MAAM,kBAAkB,MAAM,gBAC1D,WAAW,wCACnB;IAEF,MAAM,OAAO,KAAK,QAAQ,QAAQ,QAAQ,KAAK,CAAC;IAChD,IAAI,KAAK,KAAK,QAAQ,KAAK,UAAU,WAAW;KAC9C,UAAU;MACR,MAAM;MACN,QACE,8CAA8C,QAAQ,EAAE,CAAE,gBAAgB,YAC/D,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,WAAW,KAAK,EAAE;KAGtF;KACA,uBAAuB,QAAQ;KAC/B,OAAO;IACT;GACF;EACF;EACA,uBAAuB,QAAQ;EAC/B,OAAO,EAAE,MAAM,MAAM;CACvB;CACA,OAAO,cAAc,KAAK,MAAM;CAChC,OAAO;AACT"}

package/dist/{skillopt-optimization-method-CL_0aArC.js → skillopt-optimization-method-B2IEaCeF.js}

@@ -1,6 +1,6 @@
-import { at as pairHoldout, it as heldoutSignificance, x as campaignBreakdown, z as surfaceContentHash } from "./campaign-evidence-B8oF9xQ6.js";
+import { at as pairHoldout, it as heldoutSignificance, x as campaignBreakdown, z as surfaceContentHash } from "./campaign-evidence-PYPUftRy.js";
 import { n as contentHash, t as canonicalJson } from "./verdict-cache-B3eCVQtY.js";
-import { A as openSearchLedger, X as runCampaign, b as recordCandidatePopulationSearch, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, t as llmJudge } from "./llm-judge-DEFZeSiu.js";
+import { A as openSearchLedger, X as runCampaign, b as recordCandidatePopulationSearch, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, t as llmJudge } from "./llm-judge-BPFR12VJ.js";
 import { S as fsCampaignStorage, _ as removeCredentialEnvironment, b as safePathComponent, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as isRecord, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, x as createRunCostLedger, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-D4dzUBZI.js";
 import { l as deepFreezeCanonicalJson } from "./types-DQ0e2E7y.js";
 import { g as assertServedModelPolicy } from "./llm-client-CGlSi8sb.js";
@@ -2140,4 +2140,4 @@ function skillOptOptimizationMethod(config) {
 //#endregion
 export { externalTextOptimizationMethod as a, REFERENCE_EQUIVALENCE_INPUT_LIMITS as c, runReferenceEquivalenceJudge as d, composeGate as i, REFERENCE_EQUIVALENCE_JUDGE_VERSION as l, gepaOptimizationMethod as n, decodeExternalTextCandidate as o, heldOutGate as r, readExternalOptimizerObservationArtifact as s, skillOptOptimizationMethod as t, createReferenceEquivalenceJudge as u };
 
-//# sourceMappingURL=skillopt-optimization-method-CL_0aArC.js.map
+//# sourceMappingURL=skillopt-optimization-method-B2IEaCeF.js.map