agentix-cli 0.26.0 → 0.27.0

package/dist/chunk-AX2PWSRT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/decisions/backend.ts","../src/decisions/store.ts","../src/decisions/types.ts","../src/decisions/normalize.ts","../src/decisions/schema.ts","../src/decisions/prompt.ts","../src/decisions/recalibrate.ts","../src/decisions/consistency.ts","../src/decisions/backends/mock.ts","../src/agent/providers/capabilities.ts","../src/decisions/backends/local-llm.ts","../src/decisions/backends/simple-jev.ts","../src/decisions/index.ts","../src/decisions/seat.ts"],"sourcesContent":["import type { AnswersFor, Questions, StateValue } from \"./types\"\n\n// The backend contract, and the registry that resolves one by name.\n//\n// Why this is its own registry and not `AgentProvider.decide?()`:\n//\n//   1. `ProviderName` is what an AGENT is configured with. Adding a\n//      decision-only model to that union would make it selectable as an\n//      agent's chat provider in agentx.json, the setup wizard and\n//      `agentx model` — where it fails, because a System One model has no\n//      prose, no stream and no tools. That is a type that lies.\n//\n//   2. The unit of selection differs. Chat providers are per-agent;\n//      decision backends are per-SEAT. Running one seat on the local\n//      adapter and another on a hosted model simultaneously is exactly how\n//      a backend swap gets graded, and a per-agent union cannot say it.\n//\n//   3. The local backend COMPOSES the chat registry (it calls\n//      createProvider internally) rather than extending it. One adapter\n//      over N providers is less code than N providers each reimplementing\n//      normalization, retry and confidence.\n\n/** How the structured answer was obtained. Recorded on every call, and\n *  never pooled across values when computing calibration: a verbalized\n *  probability and a token posterior are different measurements. */\nexport type StructureMode =\n  | \"tool\"\n  | \"json_schema\"\n  | \"text\"\n  | \"logprobs\"\n  | \"vote\"\n  | \"native\"\n  | \"mock\"\n\nexport type AnswerMode = \"probabilities\" | \"discrete\"\n\nexport interface DecisionMeta {\n  backend: string\n  structureMode: StructureMode\n  answerMode: AnswerMode\n  /** At least one answer had to be repaired during normalization. */\n  repaired: boolean\n  /** Correction rounds spent on malformed structure. */\n  retries: number\n  /** State exceeded the backend's budget and was clipped. Shadow analysis\n   *  needs this to know which rows to exclude. */\n  stateTruncated: boolean\n  latencyMs: number\n}\n\nexport interface DecisionRequest<Q extends Questions> {\n  state: StateValue\n  questions: Q\n  /** Backend-scoped model id. Omitted means the backend's default. */\n  model?: string\n  abortSignal?: AbortSignal\n  timeoutMs?: number\n}\n\nexport interface DecisionResponse<Q extends Questions> {\n  model: string\n  answers: AnswersFor<Q>\n  usage: { inputTokens: number; outputTokens: number }\n  meta: DecisionMeta\n}\n\n/** Where the numbers come from. Factual, not a judgement of quality.\n *\n *  verbalized  the model wrote the probability into its answer. Well-formed,\n *              quantized at round values, and not the model's own posterior.\n *  logits      read off the next-token distribution at the answer position.\n *              This IS the model's posterior over the label set.\n *  native      a model whose training objective targeted the distribution. */\nexport type ProbabilitySource = \"verbalized\" | \"logits\" | \"native\" | \"synthetic\"\n\nexport interface DecisionBackendCapabilities {\n  probabilitySource: ProbabilitySource\n  /** Calibrated to CORRECTNESS on your traffic — i.e. an answer given 0.8\n   *  is right about 80% of the time here.\n   *\n   *  This is not the same as reading real logits. A token posterior is the\n   *  model's own belief, which can be confidently wrong; simple-jev makes\n   *  exactly this point about its own output. Correctness calibration comes\n   *  from a recalibrator fitted on your labeled rows, not from a backend,\n   *  so every backend reports false until one is demonstrably trained for\n   *  it. Prefer `probabilitySource` when choosing a backend. */\n  calibratedProbabilities: boolean\n  maxChoiceOptions: number\n  maxStateChars: number\n  /** False forces the caller to loop per question. Every backend worth\n   *  shipping is true; the flag exists so that a backend that isn't cannot\n   *  quietly multiply request counts. */\n  parallelQuestions: boolean\n  images: false\n}\n\nexport interface DecisionBackend {\n  readonly name: string\n  readonly capabilities: DecisionBackendCapabilities\n  decide<Q extends Questions>(request: DecisionRequest<Q>): Promise<DecisionResponse<Q>>\n}\n\ntype BackendFactory = () => DecisionBackend\n\nconst factories = new Map<string, BackendFactory>()\nconst instances = new Map<string, DecisionBackend>()\n\n/** Register a backend under a name. Factories are lazy so that registering\n *  a backend never reads config, opens a socket or requires a key that the\n *  operator may not have. */\nexport function registerDecisionBackend(name: string, factory: BackendFactory): void {\n  factories.set(name, factory)\n  instances.delete(name)\n}\n\nexport function getDecisionBackend(name: string): DecisionBackend {\n  const cached = instances.get(name)\n  if (cached) return cached\n  const factory = factories.get(name)\n  if (!factory) {\n    const known = listDecisionBackends()\n    throw new Error(\n      `unknown decision backend \"${name}\"` +\n        (known.length > 0 ? ` — registered: ${known.join(\", \")}` : \" — none registered\"),\n    )\n  }\n  const backend = factory()\n  instances.set(name, backend)\n  return backend\n}\n\nexport function listDecisionBackends(): string[] {\n  return [...factories.keys()].sort()\n}\n\nexport function _resetDecisionBackendsForTesting(): void {\n  factories.clear()\n  instances.clear()\n}\n","import Database from \"better-sqlite3\"\nimport { createHash } from \"crypto\"\nimport { mkdirSync } from \"fs\"\nimport { dirname, resolve } from \"path\"\nimport { newEventId } from \"@/intent/ulid\"\nimport type { DecisionMeta } from \"./backend\"\nimport type { AnyAnswer, AnyQuestion, Questions, StateValue } from \"./types\"\n\n// The shadow store: every decision the seat makes, and what actually\n// happened afterwards.\n//\n// Deliberately NOT the intent ledger, for four reasons that are mechanical\n// rather than stylistic:\n//\n//   1. `IntentDivergence.source` is a closed 9-member union of channel\n//      names. \"classifier\" and \"monitor-prefilter\" have no representable\n//      value, and that closedness is the ledger's stated design goal.\n//   2. `IntentDecision.outcome` is dispatched|halted|deduped|queued.\n//      A score or a probability has no projection onto it.\n//   3. The ledger is canonical state. Speculative model output is derived\n//      data, with a different retention and durability story.\n//   4. The ledger's promotion gate is \"zero divergences for >= 7 days\".\n//      A shadow seat is EXPECTED to disagree with the incumbent — that\n//      disagreement is the measurement — so writing it into\n//      intent_divergences would destroy the signal that gate depends on.\n//\n// Same storage discipline as the ledger, though: own file, WAL,\n// synchronous=NORMAL, versioned migrations.\n\nexport interface OpenStoreOptions {\n  /** Resolved relative to cwd. Default: .agentx/decisions/decisions.sqlite */\n  path?: string\n  readonly?: boolean\n}\n\nexport type SeatMode = \"off\" | \"shadow\" | \"active\"\nexport type LabelKind = \"human\" | \"outcome\" | \"replay\"\n\n/** What the policy decided to do with the answer. Recorded in shadow too,\n *  where it is the counterfactual rather than the action taken. */\nexport type DecisionAction = \"review\" | \"skip\"\n\nexport interface RecordCallInput {\n  seat: string\n  mode: SeatMode\n  backend: string\n  model: string\n  meta: Pick<DecisionMeta, \"structureMode\" | \"answerMode\" | \"retries\" | \"stateTruncated\" | \"latencyMs\">\n  state: StateValue\n  questions: Questions\n  answers: Record<string, AnyAnswer>\n  usage?: { inputTokens: number; outputTokens: number }\n  /** What the code would have done without the seat, per question. */\n  incumbent?: Record<string, { value?: string | number; score?: number; source?: string }>\n  links?: Array<{ kind: string; id: string }>\n  /** Low-cardinality covariates for recalibration: agent, channel, and\n   *  anything else that plausibly shifts the miscalibration curve. Keep\n   *  these categorical and few — this is a calibration model fitted on\n   *  hundreds of rows, not thousands. */\n  features?: Record<string, string | number | boolean | null>\n  /** A failed call is still a row. The failure rate is a measured quantity,\n   *  not a hunch, and a seat that only records its successes flatters\n   *  itself exactly where it matters. */\n  error?: string\n  /** Keep the serialized state for replay. Callers pass false once a seat\n   *  has collected enough, and always false when redaction is on. */\n  keepState?: boolean\n  ts?: number\n}\n\n/** One decision, one question, joined to its incumbent and its truth.\n *  The unit every calibration metric consumes. */\nexport interface GradedRow {\n  callId: string\n  seat: string\n  ts: number\n  backend: string\n  model: string\n  structureMode: string\n  question: string\n  type: AnyQuestion[\"type\"]\n  /** The seat's answer, as a comparable scalar. */\n  predicted: string\n  confidence: number\n  probabilities: Record<string, number>\n  incumbent?: string\n  truth?: string\n  features: Record<string, string | number | boolean | null>\n  mode: SeatMode\n  /** What the policy decided. Undefined for calls whose caller never\n   *  recorded one. */\n  action?: DecisionAction\n  /** True when the policy said skip and exploration overrode it, so the\n   *  expensive path ran anyway. These rows are the unbiased sample of the\n   *  skip region — the only place a skip decision can ever be graded. */\n  explored: boolean\n}\n\nexport interface GradedRowFilter {\n  seat?: string\n  question?: string\n  backend?: string\n  model?: string\n  structureMode?: string\n  /** ms since epoch; rows at or after this time. */\n  since?: number\n  /** Only rows that have a ground-truth label. */\n  labeledOnly?: boolean\n  /** Only rows where exploration forced the expensive path. Calibration on\n   *  these alone is the unbiased estimate of how the policy performs on the\n   *  decisions it wants to skip. */\n  exploredOnly?: boolean\n  action?: DecisionAction\n  limit?: number\n}\n\nexport class DecisionStore {\n  readonly db: Database.Database\n  readonly path: string\n\n  constructor(opts: OpenStoreOptions = {}) {\n    this.path = resolve(process.cwd(), opts.path ?? \".agentx/decisions/decisions.sqlite\")\n    if (!opts.readonly) mkdirSync(dirname(this.path), { recursive: true })\n    this.db = new Database(this.path, { readonly: opts.readonly ?? false })\n    if (!opts.readonly) {\n      this.db.pragma(\"journal_mode = WAL\")\n      this.db.pragma(\"synchronous = NORMAL\")\n      this.db.pragma(\"foreign_keys = ON\")\n      runMigrations(this.db)\n    }\n  }\n\n  close(): void {\n    this.db.close()\n  }\n\n  schemaVersion(): number {\n    const row = this.db.prepare(\"SELECT MAX(v) AS v FROM schema_version\").get() as {\n      v: number | null\n    }\n    return row.v ?? 0\n  }\n\n  /** Write one call and everything hanging off it, atomically. Returns the\n   *  call id, which is what a later label or link refers to. */\n  recordCall(input: RecordCallInput): string {\n    const ts = input.ts ?? Date.now()\n    const callId = newEventId(ts)\n    const stateJson = JSON.stringify(input.state ?? null)\n    const stateHash = createHash(\"sha256\").update(stateJson).digest(\"hex\")\n\n    const tx = this.db.transaction(() => {\n      this.db\n        .prepare(\n          `INSERT INTO decision_calls\n             (id, ts, seat, mode, backend, model, structure_mode, answer_mode,\n              state_hash, state_json, questions_json, latency_ms,\n              input_tokens, output_tokens, retries, truncated, error, features_json)\n           VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n        )\n        .run(\n          callId,\n          ts,\n          input.seat,\n          input.mode,\n          input.backend,\n          input.model,\n          input.meta.structureMode,\n          input.meta.answerMode,\n          stateHash,\n          input.keepState === false ? null : stateJson,\n          JSON.stringify(input.questions),\n          input.meta.latencyMs,\n          input.usage?.inputTokens ?? 0,\n          input.usage?.outputTokens ?? 0,\n          input.meta.retries,\n          input.meta.stateTruncated ? 1 : 0,\n          input.error ?? null,\n          input.features ? JSON.stringify(input.features) : null,\n        )\n\n      const insertAnswer = this.db.prepare(\n        `INSERT INTO decision_answers\n           (call_id, question, type, answer_json, top_label, top_prob,\n            expected_score, p_max, neg_entropy)\n         VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n      )\n      for (const [question, answer] of Object.entries(input.answers)) {\n        const view = answerView(answer)\n        insertAnswer.run(\n          callId,\n          question,\n          answer.type,\n          JSON.stringify(answer),\n          view.topLabel,\n          view.topProb,\n          view.expectedScore,\n          view.pMax,\n          view.negEntropy,\n        )\n      }\n\n      if (input.incumbent) {\n        const insertIncumbent = this.db.prepare(\n          `INSERT OR REPLACE INTO decision_incumbent (call_id, question, value, score, source)\n           VALUES (?, ?, ?, ?, ?)`,\n        )\n        for (const [question, row] of Object.entries(input.incumbent)) {\n          insertIncumbent.run(\n            callId,\n            question,\n            row.value === undefined ? null : String(row.value),\n            row.score ?? null,\n            row.source ?? null,\n          )\n        }\n      }\n\n      if (input.links) {\n        const insertLink = this.db.prepare(\n          `INSERT OR IGNORE INTO decision_links (call_id, ref_kind, ref_id) VALUES (?, ?, ?)`,\n        )\n        for (const link of input.links) insertLink.run(callId, link.kind, link.id)\n      }\n    })\n\n    tx()\n    return callId\n  }\n\n  /** Attach ground truth. Append-only: a correction is a new row, and reads\n   *  take the newest, so a mislabel is fixable without losing the audit. */\n  label(\n    callId: string,\n    question: string,\n    value: string | number,\n    opts: { kind?: LabelKind; labeledBy?: string; note?: string; ts?: number } = {},\n  ): string {\n    const ts = opts.ts ?? Date.now()\n    const id = newEventId(ts)\n    this.db\n      .prepare(\n        `INSERT INTO decision_labels (id, call_id, question, value, kind, labeled_at, labeled_by, note)\n         VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,\n      )\n      .run(\n        id,\n        callId,\n        question,\n        String(value),\n        opts.kind ?? \"human\",\n        ts,\n        opts.labeledBy ?? null,\n        opts.note ?? null,\n      )\n    return id\n  }\n\n  /** Record what the policy did with an answer. Separate from recordCall\n   *  because the seat records the answer and the CALLER owns the policy —\n   *  askSeat cannot know whether its caller acted on what it returned. */\n  recordOutcome(callId: string, action: DecisionAction, explored = false): void {\n    this.db\n      .prepare(\"UPDATE decision_calls SET action = ?, explored = ? WHERE id = ?\")\n      .run(action, explored ? 1 : 0, callId)\n  }\n\n  findCallsByLink(kind: string, id: string): string[] {\n    const rows = this.db\n      .prepare(`SELECT call_id FROM decision_links WHERE ref_kind = ? AND ref_id = ?`)\n      .all(kind, id) as Array<{ call_id: string }>\n    return rows.map((r) => r.call_id)\n  }\n\n  gradedRows(filter: GradedRowFilter = {}): GradedRow[] {\n    const where: string[] = [\"c.error IS NULL\"]\n    const params: unknown[] = []\n    if (filter.seat) (where.push(\"c.seat = ?\"), params.push(filter.seat))\n    if (filter.question) (where.push(\"a.question = ?\"), params.push(filter.question))\n    if (filter.backend) (where.push(\"c.backend = ?\"), params.push(filter.backend))\n    if (filter.model) (where.push(\"c.model = ?\"), params.push(filter.model))\n    if (filter.structureMode) (where.push(\"c.structure_mode = ?\"), params.push(filter.structureMode))\n    if (filter.since !== undefined) (where.push(\"c.ts >= ?\"), params.push(filter.since))\n    if (filter.action) (where.push(\"c.action = ?\"), params.push(filter.action))\n    if (filter.exploredOnly) where.push(\"c.explored = 1\")\n\n    // The newest label per (call, question) wins; ULIDs sort by time, so\n    // MAX(id) is the newest without a second timestamp comparison.\n    const sql = `\n      SELECT c.id AS call_id, c.seat, c.ts, c.backend, c.model, c.structure_mode,\n             c.mode, c.action, c.explored, c.features_json,\n             a.question, a.type, a.answer_json, a.top_label, a.expected_score, a.neg_entropy,\n             i.value AS incumbent,\n             (SELECT l.value FROM decision_labels l\n               WHERE l.call_id = c.id AND l.question = a.question\n               ORDER BY l.id DESC LIMIT 1) AS truth\n        FROM decision_calls c\n        JOIN decision_answers a ON a.call_id = c.id\n        LEFT JOIN decision_incumbent i ON i.call_id = c.id AND i.question = a.question\n       WHERE ${where.join(\" AND \")}\n       ORDER BY c.ts DESC\n       ${filter.limit ? \"LIMIT ?\" : \"\"}`\n    if (filter.limit) params.push(filter.limit)\n\n    const rows = this.db.prepare(sql).all(...params) as Array<Record<string, any>>\n    return rows\n      .map((r) => {\n        const answer = JSON.parse(r.answer_json) as AnyAnswer\n        const view = answerView(answer)\n        return {\n          callId: r.call_id,\n          seat: r.seat,\n          ts: r.ts,\n          backend: r.backend,\n          model: r.model,\n          structureMode: r.structure_mode,\n          question: r.question,\n          type: answer.type,\n          predicted: view.predicted,\n          confidence: view.confidence,\n          probabilities: view.probabilities,\n          incumbent: r.incumbent ?? undefined,\n          truth: r.truth ?? undefined,\n          features: parseFeatures(r.features_json),\n          mode: r.mode as SeatMode,\n          action: (r.action ?? undefined) as GradedRow[\"action\"],\n          explored: r.explored === 1,\n        } as GradedRow\n      })\n      .filter((row) => !filter.labeledOnly || row.truth !== undefined)\n  }\n\n  /** Drop stored state once a seat has enough for replay, keeping the hash.\n   *  Returns how many rows were cleared. */\n  pruneState(seat: string, keepRows: number): number {\n    const res = this.db\n      .prepare(\n        `UPDATE decision_calls SET state_json = NULL\n          WHERE seat = ? AND state_json IS NOT NULL\n            AND id NOT IN (\n              SELECT id FROM decision_calls WHERE seat = ? ORDER BY ts DESC LIMIT ?\n            )`,\n      )\n      .run(seat, seat, keepRows)\n    return res.changes\n  }\n}\n\n/** Project any answer onto the scalars the store and the metrics need.\n *  A noul becomes a two-outcome distribution so it grades through exactly\n *  the same machinery as a choice. */\nexport function answerView(answer: AnyAnswer): {\n  predicted: string\n  confidence: number\n  probabilities: Record<string, number>\n  topLabel: string\n  topProb: number\n  expectedScore: number | null\n  pMax: number\n  negEntropy: number\n} {\n  if (answer.type === \"noul\") {\n    const p = answer.noul\n    const probabilities = { yes: p, no: 1 - p }\n    const predicted = p >= 0.5 ? \"yes\" : \"no\"\n    // Distance from a coin flip, scaled to [0,1] — the only uncertainty a\n    // binary carries, and comparable with the entropy statistic elsewhere.\n    const confidence = Math.abs(p - 0.5) * 2\n    return {\n      predicted,\n      confidence,\n      probabilities,\n      topLabel: predicted,\n      topProb: Math.max(p, 1 - p),\n      expectedScore: null,\n      pMax: Math.max(p, 1 - p),\n      negEntropy: confidence,\n    }\n  }\n\n  if (answer.type === \"choice\") {\n    return {\n      predicted: answer.choice,\n      confidence: answer.confidence,\n      probabilities: { ...answer.probabilities },\n      topLabel: answer.choice,\n      topProb: answer.probabilities[answer.choice] ?? 0,\n      expectedScore: null,\n      pMax: answer.pMax,\n      negEntropy: answer.negEntropy,\n    }\n  }\n\n  const rounded = String(Math.round(answer.score))\n  return {\n    predicted: rounded,\n    confidence: answer.confidence,\n    probabilities: { ...answer.probabilities },\n    topLabel: rounded,\n    topProb: answer.probabilities[rounded] ?? 0,\n    expectedScore: answer.score,\n    pMax: answer.pMax,\n    negEntropy: answer.negEntropy,\n  }\n}\n\nfunction parseFeatures(raw: unknown): Record<string, string | number | boolean | null> {\n  if (typeof raw !== \"string\" || raw.length === 0) return {}\n  try {\n    const parsed = JSON.parse(raw)\n    return parsed && typeof parsed === \"object\" && !Array.isArray(parsed) ? parsed : {}\n  } catch {\n    return {}\n  }\n}\n\nfunction runMigrations(db: Database.Database): void {\n  db.exec(`CREATE TABLE IF NOT EXISTS schema_version (v INTEGER PRIMARY KEY);`)\n  const current =\n    (db.prepare(\"SELECT MAX(v) AS v FROM schema_version\").get() as { v: number | null }).v ?? 0\n  if (current < 1) migrationV1(db)\n  if (current < 2) migrationV2(db)\n  if (current < 3) migrationV3(db)\n}\n\n/** Caller-supplied covariates for recalibration.\n *\n *  A single global temperature assumes one miscalibration curve for the\n *  whole seat. That is usually false: a chatty support agent and a cron\n *  runner are different distributions, and a model can be well calibrated\n *  on one while badly overconfident on the other. Storing a few\n *  low-cardinality covariates per call is what lets that be fitted instead\n *  of assumed away. */\nfunction migrationV3(db: Database.Database): void {\n  db.exec(`\n    ALTER TABLE decision_calls ADD COLUMN features_json TEXT;\n    INSERT INTO schema_version (v) VALUES (3);\n  `)\n}\n\n/** What the policy decided, and whether exploration overrode it.\n *\n *  Recorded in every mode, including shadow, so the counterfactual is\n *  measurable before anything is ever actually skipped: \"the policy would\n *  have skipped 40% of these, and here is how it did on them.\"\n *\n *  Without this the labeled sample under active mode is the set of reviews\n *  the policy chose to run, which is exactly the biased subsample its own\n *  metrics would then be computed on. */\nfunction migrationV2(db: Database.Database): void {\n  db.exec(`\n    ALTER TABLE decision_calls ADD COLUMN action TEXT;\n    ALTER TABLE decision_calls ADD COLUMN explored INTEGER NOT NULL DEFAULT 0;\n    CREATE INDEX IF NOT EXISTS idx_decision_calls_action ON decision_calls (seat, action, explored);\n    INSERT INTO schema_version (v) VALUES (2);\n  `)\n}\n\nfunction migrationV1(db: Database.Database): void {\n  db.exec(`\n    CREATE TABLE IF NOT EXISTS decision_calls (\n      id             TEXT NOT NULL PRIMARY KEY,\n      ts             INTEGER NOT NULL,\n      seat           TEXT NOT NULL,\n      mode           TEXT NOT NULL,\n      backend        TEXT NOT NULL,\n      model          TEXT NOT NULL,\n      structure_mode TEXT NOT NULL,\n      answer_mode    TEXT NOT NULL,\n      state_hash     TEXT NOT NULL,\n      state_json     TEXT,\n      questions_json TEXT NOT NULL,\n      latency_ms     INTEGER,\n      input_tokens   INTEGER NOT NULL DEFAULT 0,\n      output_tokens  INTEGER NOT NULL DEFAULT 0,\n      retries        INTEGER NOT NULL DEFAULT 0,\n      truncated      INTEGER NOT NULL DEFAULT 0,\n      error          TEXT\n    );\n    CREATE INDEX IF NOT EXISTS idx_decision_calls_seat ON decision_calls (seat, ts);\n    CREATE INDEX IF NOT EXISTS idx_decision_calls_backend ON decision_calls (backend, model, ts);\n    CREATE INDEX IF NOT EXISTS idx_decision_calls_state ON decision_calls (state_hash);\n\n    CREATE TABLE IF NOT EXISTS decision_answers (\n      call_id        TEXT NOT NULL REFERENCES decision_calls(id),\n      question       TEXT NOT NULL,\n      type           TEXT NOT NULL,\n      answer_json    TEXT NOT NULL,\n      top_label      TEXT,\n      top_prob       REAL,\n      expected_score REAL,\n      p_max          REAL,\n      neg_entropy    REAL,\n      PRIMARY KEY (call_id, question)\n    );\n\n    CREATE TABLE IF NOT EXISTS decision_incumbent (\n      call_id  TEXT NOT NULL REFERENCES decision_calls(id),\n      question TEXT NOT NULL,\n      value    TEXT,\n      score    REAL,\n      source   TEXT,\n      PRIMARY KEY (call_id, question)\n    );\n\n    CREATE TABLE IF NOT EXISTS decision_labels (\n      id         TEXT NOT NULL PRIMARY KEY,\n      call_id    TEXT NOT NULL REFERENCES decision_calls(id),\n      question   TEXT NOT NULL,\n      value      TEXT NOT NULL,\n      kind       TEXT NOT NULL,\n      labeled_at INTEGER NOT NULL,\n      labeled_by TEXT,\n      note       TEXT\n    );\n    CREATE INDEX IF NOT EXISTS idx_decision_labels_call ON decision_labels (call_id, question);\n\n    CREATE TABLE IF NOT EXISTS decision_links (\n      call_id  TEXT NOT NULL REFERENCES decision_calls(id),\n      ref_kind TEXT NOT NULL,\n      ref_id   TEXT NOT NULL,\n      PRIMARY KEY (call_id, ref_kind, ref_id)\n    );\n    CREATE INDEX IF NOT EXISTS idx_decision_links_ref ON decision_links (ref_kind, ref_id);\n\n    INSERT INTO schema_version (v) VALUES (1);\n  `)\n}\n","// Typed-decision seat — the wire contract.\n//\n// This mirrors TypeSafe AI's \"System One\" shape (POST /v1/systemone) on\n// purpose: one `state` plus a map of named questions, with every answer\n// returned in a single call carrying a full probability distribution.\n// Mirroring it means a future Jev backend is a new file in ./backends,\n// not a refactor of every call site.\n//\n// Three question types, and only three. A call site that wants prose does\n// not belong here.\n//\n// Two deliberate divergences from \"whatever is easiest to ask a model for\":\n//\n//   - `NoulAnswer` has NO `confidence` field, because Jev's noul answer has\n//     none. A consumer reaching for it is a compile error today rather than\n//     a surprise on the day we swap backends.\n//\n//   - The model is never asked for `choice`, `score`, `confidence` or\n//     `legend`. It reports `probabilities` and nothing else; the rest is\n//     derived in ./normalize.ts. Self-reported confidence is the exact\n//     thing this module exists to replace — see src/graph/classifier.ts,\n//     where `graph.autoApproveConfidence` defaults to 1.0 (i.e. disabled)\n//     precisely because the model's own number isn't trusted.\n\n/** A JSON-compatible value. State is JSON-only: no images, matching Jev. */\nexport type StateValue =\n  | string\n  | number\n  | boolean\n  | null\n  | StateValue[]\n  | { [key: string]: StateValue }\n\n// ---------------------------------------------------------------------------\n// Questions\n// ---------------------------------------------------------------------------\n\n/** A yes/no question. `criteria` optionally describes either outcome. */\nexport interface NoulQuestion {\n  readonly type: \"noul\"\n  readonly instructions?: string\n  readonly criteria?: {\n    readonly true?: string\n    readonly false?: string\n  }\n}\n\n/** Select one of a named set. `criteria` maps label -> description\n *  (`null` leaves a label undescribed). */\nexport interface ChoiceQuestion<L extends string = string> {\n  readonly type: \"choice\"\n  readonly instructions?: string\n  readonly criteria: Readonly<Record<L, string | null>>\n}\n\n/** Place the state on an ordered rubric. Index i describes level i, so the\n *  array order IS the scale. At least two levels. */\nexport interface ScoreQuestion {\n  readonly type: \"score\"\n  readonly instructions?: string\n  readonly criteria: readonly string[]\n}\n\nexport type AnyQuestion = NoulQuestion | ChoiceQuestion<string> | ScoreQuestion\n\n/** Questions keyed by the names their answers come back under. */\nexport type Questions = Record<string, AnyQuestion>\n\n// ---------------------------------------------------------------------------\n// Answers\n// ---------------------------------------------------------------------------\n\n/** Probability of \"yes\". No confidence — the probability IS the answer, and\n *  distance from 0.5 is the only uncertainty signal a binary has. */\nexport interface NoulAnswer {\n  readonly type: \"noul\"\n  readonly noul: number\n}\n\nexport interface ChoiceAnswer<L extends string = string> {\n  readonly type: \"choice\"\n  /** argmax of `probabilities`, derived — never model-reported. */\n  readonly choice: L\n  /** Derived from the shape of `probabilities`. See confidenceFor(). */\n  readonly confidence: number\n  readonly probabilities: Readonly<Record<L, number>>\n  /** Both statistics are kept on every answer because Jev's own confidence\n   *  formula is unpublished; storing both lets the swap-day comparison run\n   *  against already-collected shadow rows without recollecting them. */\n  readonly pMax: number\n  readonly negEntropy: number\n}\n\nexport interface ScoreAnswer {\n  readonly type: \"score\"\n  /** Expected value over the rubric, so it may fall between levels. */\n  readonly score: number\n  readonly confidence: number\n  /** Rubric descriptions keyed by level index, echoed back for display. */\n  readonly legend: Readonly<Record<string, string>>\n  readonly probabilities: Readonly<Record<string, number>>\n  readonly pMax: number\n  readonly negEntropy: number\n}\n\nexport type AnyAnswer = NoulAnswer | ChoiceAnswer<string> | ScoreAnswer\n\n/** The answer shape implied by a question shape.\n *\n *  Inference reads `criteria` directly rather than going through\n *  `ChoiceQuestion<infer L>`, so a question built by hand (not via the\n *  ./questions.ts builders) still narrows correctly. */\nexport type AnswerFor<Q extends AnyQuestion> =\n  Q extends { type: \"noul\" } ? NoulAnswer\n  : Q extends { type: \"score\" } ? ScoreAnswer\n  : Q extends { type: \"choice\"; criteria: infer C } ? ChoiceAnswer<Extract<keyof C, string>>\n  : never\n\nexport type AnswersFor<Q extends Questions> = {\n  readonly [K in keyof Q]: AnswerFor<Q[K]>\n}\n\n// ---------------------------------------------------------------------------\n// Raw model output\n// ---------------------------------------------------------------------------\n//\n// What a backend is allowed to receive from a model, before normalization.\n// Narrower than the public answer on purpose — see the header.\n\nexport interface RawNoulAnswer {\n  readonly noul: number\n}\n\nexport interface RawDistributionAnswer {\n  readonly probabilities: Record<string, unknown>\n}\n\nexport type RawAnswer = RawNoulAnswer | RawDistributionAnswer\n\nexport type RawAnswers = Record<string, RawAnswer>\n","import { labelsOf } from \"./questions\"\nimport type { AnyAnswer, AnyQuestion, RawAnswer } from \"./types\"\n\n// Turning what a model said into a distribution we can do arithmetic on.\n//\n// Everything a caller reads off an answer except the raw numbers is computed\n// here: the argmax, the expected score, and both confidence statistics. The\n// model supplies `probabilities` (or `noul`) and nothing else.\n//\n// Nothing in this file throws. A model that returns garbage yields a uniform\n// distribution with `repaired: true`, because a seat that throws on the\n// critical path is worse than a seat that says \"I don't know\" — and a\n// repaired row is visible in the shadow store, so the failure is measured\n// rather than hidden.\n\nexport interface NormalizeOptions {\n  /** Rescale so the values sum to 1. Default true. Mirrors the Python\n   *  system-one-adapter's `normalize_probabilities`. */\n  normalize?: boolean\n  /** Mass given to a label the model omitted entirely. Default 1e-6. */\n  epsilon?: number\n}\n\nexport interface NormalizedDistribution {\n  probs: Record<string, number>\n  /** True when the model's output had to be corrected: a missing or extra\n   *  label, a negative or non-finite value, or an all-zero distribution.\n   *  Recorded per call so \"how often does the backend misbehave\" is itself\n   *  a measured quantity rather than a hunch. */\n  repaired: boolean\n}\n\nexport function normalizeDistribution(\n  raw: unknown,\n  labels: readonly string[],\n  opts: NormalizeOptions = {},\n): NormalizedDistribution {\n  const normalize = opts.normalize ?? true\n  const epsilon = opts.epsilon ?? 1e-6\n  const source: Record<string, unknown> =\n    raw !== null && typeof raw === \"object\" && !Array.isArray(raw)\n      ? (raw as Record<string, unknown>)\n      : {}\n\n  let repaired = raw === null || typeof raw !== \"object\" || Array.isArray(raw)\n  for (const key of Object.keys(source)) {\n    if (!labels.includes(key)) repaired = true\n  }\n\n  const probs: Record<string, number> = {}\n  let sum = 0\n  for (const label of labels) {\n    if (!(label in source)) {\n      probs[label] = epsilon\n      sum += epsilon\n      repaired = true\n      continue\n    }\n    const value = Number(source[label])\n    if (!Number.isFinite(value) || value < 0) {\n      probs[label] = 0\n      repaired = true\n      continue\n    }\n    probs[label] = value\n    sum += value\n  }\n\n  if (labels.length === 0) return { probs, repaired }\n\n  if (sum <= 0) {\n    // The model gave us nothing usable. Uniform is the honest answer: it\n    // carries zero confidence, so a confidence-gated call site escalates.\n    const uniform = 1 / labels.length\n    for (const label of labels) probs[label] = uniform\n    return { probs, repaired: true }\n  }\n\n  if (normalize) {\n    for (const label of labels) probs[label] = probs[label] / sum\n    if (Math.abs(sum - 1) > 1e-6) repaired = true\n  }\n\n  return { probs, repaired }\n}\n\n/** Largest probability. Simple, but not comparable across different option\n *  counts — uniform over 2 is 0.5, uniform over 10 is 0.1. */\nexport function pMax(probs: Record<string, number>): number {\n  const values = Object.values(probs)\n  if (values.length === 0) return 0\n  return Math.max(...values)\n}\n\n/** 1 - H(p) / ln(k). Zero on a uniform distribution, one on a point mass,\n *  and comparable across option counts because it divides out ln(k) — which\n *  is why this, not pMax, is what `confidence` reports. */\nexport function normalizedNegEntropy(probs: Record<string, number>): number {\n  const values = Object.values(probs).filter((v) => Number.isFinite(v) && v > 0)\n  const k = Object.keys(probs).length\n  if (k <= 1) return 1\n  const total = values.reduce((a, b) => a + b, 0)\n  if (total <= 0) return 0\n  let h = 0\n  for (const v of values) {\n    const p = v / total\n    h -= p * Math.log(p)\n  }\n  return clamp01(1 - h / Math.log(k))\n}\n\n/** Expected value over a distribution whose keys are level indices. */\nexport function expectedScore(probs: Record<string, number>): number {\n  let total = 0\n  let weighted = 0\n  for (const [key, value] of Object.entries(probs)) {\n    const level = Number(key)\n    if (!Number.isFinite(level) || !Number.isFinite(value) || value <= 0) continue\n    total += value\n    weighted += level * value\n  }\n  return total > 0 ? weighted / total : 0\n}\n\nexport function confidenceFor(probs: Record<string, number>): number {\n  return normalizedNegEntropy(probs)\n}\n\nfunction clamp01(n: number): number {\n  if (!Number.isFinite(n)) return 0\n  return Math.max(0, Math.min(1, n))\n}\n\n/** Build the public answer from a question and what the model reported.\n *  The caller-visible `choice`, `score`, `confidence` and `legend` are all\n *  derived right here — the model never reports any of them. */\nexport function finalizeAnswer(\n  question: AnyQuestion,\n  raw: RawAnswer,\n  opts: NormalizeOptions = {},\n): { answer: AnyAnswer; repaired: boolean } {\n  if (question.type === \"noul\") {\n    const value = Number((raw as { noul?: unknown }).noul)\n    const ok = Number.isFinite(value) && value >= 0 && value <= 1\n    return {\n      answer: { type: \"noul\", noul: ok ? value : 0.5 },\n      repaired: !ok,\n    }\n  }\n\n  const labels = labelsOf(question)\n  const { probs, repaired } = normalizeDistribution(\n    (raw as { probabilities?: unknown }).probabilities,\n    labels,\n    opts,\n  )\n  const max = pMax(probs)\n  const negEntropy = normalizedNegEntropy(probs)\n\n  if (question.type === \"choice\") {\n    let best = labels[0]\n    for (const label of labels) if (probs[label] > probs[best]) best = label\n    return {\n      answer: {\n        type: \"choice\",\n        choice: best,\n        confidence: confidenceFor(probs),\n        probabilities: probs,\n        pMax: max,\n        negEntropy,\n      },\n      repaired,\n    }\n  }\n\n  const legend: Record<string, string> = {}\n  question.criteria.forEach((description, i) => {\n    legend[String(i)] = description\n  })\n  return {\n    answer: {\n      type: \"score\",\n      score: expectedScore(probs),\n      confidence: confidenceFor(probs),\n      legend,\n      probabilities: probs,\n      pMax: max,\n      negEntropy,\n    },\n    repaired,\n  }\n}\n","import { z } from \"zod\"\nimport type { AnyQuestion, Questions } from \"./types\"\n\n// Runtime validation of raw model output, DERIVED from the questions.\n//\n// Derived, not hand-written, because a hand-written schema is free to drift\n// from the Jev wire format and drift is the one failure this whole module\n// exists to prevent. There is exactly one definition of what an answer looks\n// like, and both the static type (AnswerFor, in ./types.ts) and this runtime\n// check come from it.\n//\n// These validate SHAPE, not content. `probabilities` values are `unknown`\n// on purpose: a model that writes \"0.7\" instead of 0.7 is coerced by\n// normalizeDistribution and flagged `repaired`, which is cheaper and more\n// informative than burning a correction round on a cosmetic slip. What we\n// do insist on is that the container is there and is an object — that is\n// the failure that means \"the model ignored the schema and wrote prose\",\n// and that one is worth a retry.\n\nconst noulSchema = z.object({\n  noul: z.number(),\n})\n\nconst distributionSchema = z.object({\n  probabilities: z.record(z.string(), z.unknown()),\n})\n\nexport function rawAnswerSchema(question: AnyQuestion): z.ZodType<unknown> {\n  return question.type === \"noul\" ? noulSchema : distributionSchema\n}\n\n/** Every question must be answered. A backend that drops one has not done\n *  the job, and silently returning a partial map would leave the caller\n *  reading `undefined.choice`. */\nexport function rawAnswersSchema(questions: Questions): z.ZodType<unknown> {\n  const shape: Record<string, z.ZodType<unknown>> = {}\n  for (const [name, question] of Object.entries(questions)) {\n    shape[name] = rawAnswerSchema(question)\n  }\n  return z.object(shape)\n}\n\n/** Flatten a ZodError into the one-line-per-problem form we feed back to a\n *  model on a correction round. Mirrors the correction-round shape already\n *  used by src/procedures/mine/distill.ts. */\nexport function describeIssues(error: z.ZodError): string {\n  return error.issues\n    .map((issue) => {\n      const path = issue.path.length > 0 ? issue.path.join(\".\") : \"(root)\"\n      return `- ${path}: ${issue.message}`\n    })\n    .join(\"\\n\")\n}\n","import { labelsOf } from \"./questions\"\nimport type { AnyQuestion, Questions, StateValue } from \"./types\"\n\n// Turning a question set into something a chat model can answer.\n//\n// The schema below asks for `probabilities` and nothing else. No `choice`,\n// no `score`, no `confidence`, no `legend` — those are derived in\n// ./normalize.ts. Two reasons, and the second is the important one:\n//\n//   1. If a model reports both an argmax and a distribution they can\n//      disagree, and there is no principled way to pick a winner.\n//   2. A confidence a model writes out is a number it chose, not a\n//      property of its own uncertainty. Asking for one reproduces exactly\n//      the signal that `graph.autoApproveConfidence` already defaults to\n//      1.0 to avoid trusting.\n\nexport const TOOL_NAME = \"answer_questions\"\n\nexport const SYSTEM_PROMPT = [\n  \"You answer questions about a piece of state by reporting a probability distribution for each one.\",\n  \"\",\n  \"Rules:\",\n  \"- Answer every question. Never add a question that was not asked.\",\n  \"- For a yes/no question, report `noul`: the probability the answer is yes, between 0 and 1.\",\n  \"- For every other question, report `probabilities`: one number per listed option, summing to 1.\",\n  \"- Use only the exact option names given. Never invent one.\",\n  \"- Spread the distribution when the state genuinely does not settle the question. A flat\",\n  \"  distribution is a useful answer; a confident wrong one is not.\",\n  \"- Output the structure only. No prose, no explanation, no code fences.\",\n].join(\"\\n\")\n\n/** JSON Schema for the forced tool's `input` — one object per question,\n *  keyed by question name, with `additionalProperties: false` throughout so\n *  a provider with strict structured output rejects drift server-side. */\nexport function toolInputSchema(questions: Questions): Record<string, unknown> {\n  const properties: Record<string, unknown> = {}\n  for (const [name, question] of Object.entries(questions)) {\n    properties[name] = answerObjectSchema(question)\n  }\n  return {\n    type: \"object\",\n    additionalProperties: false,\n    required: [\"answers\"],\n    properties: {\n      answers: {\n        type: \"object\",\n        additionalProperties: false,\n        required: Object.keys(questions),\n        properties,\n      },\n    },\n  }\n}\n\nfunction answerObjectSchema(question: AnyQuestion): Record<string, unknown> {\n  if (question.type === \"noul\") {\n    return {\n      type: \"object\",\n      additionalProperties: false,\n      required: [\"noul\"],\n      properties: {\n        noul: {\n          type: \"number\",\n          minimum: 0,\n          maximum: 1,\n          description: \"Probability the answer is yes.\",\n        },\n      },\n    }\n  }\n\n  const labels = labelsOf(question)\n  const properties: Record<string, unknown> = {}\n  for (const label of labels) {\n    properties[label] = { type: \"number\", minimum: 0, maximum: 1 }\n  }\n  return {\n    type: \"object\",\n    additionalProperties: false,\n    required: [\"probabilities\"],\n    properties: {\n      probabilities: {\n        type: \"object\",\n        additionalProperties: false,\n        required: labels,\n        properties,\n        description: \"One probability per option. Must sum to 1.\",\n      },\n    },\n  }\n}\n\nexport interface RenderedState {\n  text: string\n  truncated: boolean\n}\n\n/** Serialize state and clip it to the backend's budget. Truncation is\n *  reported rather than hidden, because a calibration report has to be able\n *  to exclude rows where the model never saw the whole input. */\nexport function renderState(state: StateValue, maxChars: number): RenderedState {\n  const text = typeof state === \"string\" ? state : JSON.stringify(state, null, 2) ?? \"null\"\n  if (text.length <= maxChars) return { text, truncated: false }\n  return {\n    text: `${text.slice(0, maxChars)}\\n…[truncated: ${text.length - maxChars} more characters]`,\n    truncated: true,\n  }\n}\n\n/** The human-readable question block. Providers without strict structured\n *  output only get this, so it has to carry the full contract on its own. */\nexport function renderQuestions(questions: Questions): string {\n  const blocks: string[] = []\n  for (const [name, question] of Object.entries(questions)) {\n    const lines: string[] = [`### ${name}`]\n    if (question.instructions) lines.push(question.instructions)\n\n    if (question.type === \"noul\") {\n      lines.push(\"Type: yes/no. Report `noul` — the probability the answer is yes.\")\n      if (question.criteria?.true) lines.push(`- yes means: ${question.criteria.true}`)\n      if (question.criteria?.false) lines.push(`- no means: ${question.criteria.false}`)\n    } else if (question.type === \"choice\") {\n      lines.push(\"Type: choice. Report `probabilities` over exactly these options:\")\n      for (const [label, description] of Object.entries(question.criteria)) {\n        lines.push(description ? `- ${label}: ${description}` : `- ${label}`)\n      }\n    } else {\n      lines.push(\"Type: score. Report `probabilities` over exactly these levels:\")\n      question.criteria.forEach((description, i) => {\n        lines.push(`- \"${i}\": ${description}`)\n      })\n    }\n    blocks.push(lines.join(\"\\n\"))\n  }\n  return blocks.join(\"\\n\\n\")\n}\n\n/** The user turn. Used verbatim in tool mode and in text mode; text mode\n *  appends the shape instructions, since it has no schema to lean on. */\nexport function renderUserPrompt(\n  state: StateValue,\n  questions: Questions,\n  maxChars: number,\n): { text: string; truncated: boolean } {\n  const rendered = renderState(state, maxChars)\n  const text = [\n    \"## State\",\n    rendered.text,\n    \"\",\n    \"## Questions\",\n    renderQuestions(questions),\n  ].join(\"\\n\")\n  return { text, truncated: rendered.truncated }\n}\n\n/** Appended in text mode only, where nothing enforces the shape. */\nexport function renderShapeInstructions(questions: Questions): string {\n  const example: Record<string, unknown> = {}\n  for (const [name, question] of Object.entries(questions)) {\n    if (question.type === \"noul\") {\n      example[name] = { noul: 0.5 }\n      continue\n    }\n    const probabilities: Record<string, number> = {}\n    const labels = labelsOf(question)\n    for (const label of labels) probabilities[label] = Number((1 / labels.length).toFixed(4))\n    example[name] = { probabilities }\n  }\n  return [\n    \"\",\n    \"## Output\",\n    \"Reply with this JSON object and nothing else:\",\n    JSON.stringify({ answers: example }, null, 2),\n  ].join(\"\\n\")\n}\n","import type { GradedRow } from \"./store\"\n\n// Recalibration with covariates.\n//\n// `fitTemperature` assumes one miscalibration curve for a whole seat. That\n// is usually false. A model can be well calibrated answering about a cron\n// runner and badly overconfident answering about a chat agent, and a single\n// scalar averages the two into something wrong for both.\n//\n// This fits p(correct | confidence, covariates) with logistic regression on\n// the logit of the reported confidence plus one-hot covariates. With no\n// covariates it reduces to Platt scaling, which is already a strict\n// generalisation of temperature scaling on a binary outcome — so the\n// covariate model can only help in-sample.\n//\n// Which is exactly why nothing here reports an in-sample number. More\n// parameters always fit the training data better; the only question worth\n// asking is whether they predict BETTER OUT OF SAMPLE. `compareCalibrators`\n// runs k-fold cross-validation over raw, temperature and covariate models\n// and reports held-out log loss for each. If the covariate model does not\n// win there, it does not get used.\n\n/** Target: was the seat's prediction right? Calibrating `confidence` this\n *  way covers Choice, Score and Noul uniformly, because answerView already\n *  projects all three to (predicted, confidence). */\nfunction outcome(row: GradedRow): number {\n  return row.predicted === row.truth ? 1 : 0\n}\n\nconst EPS = 1e-6\n\nexport function logit(p: number): number {\n  const clamped = Math.min(1 - EPS, Math.max(EPS, p))\n  return Math.log(clamped / (1 - clamped))\n}\n\nexport function sigmoid(z: number): number {\n  if (z >= 0) return 1 / (1 + Math.exp(-z))\n  const e = Math.exp(z)\n  return e / (1 + e)\n}\n\nexport interface DesignSpec {\n  /** Covariate names, read from row.features first and then from the row's\n   *  own fields (question, model, backend, structureMode, predicted). */\n  covariates: string[]\n  /** Levels appearing fewer than this many times are folded into the\n   *  reference level. Rare levels are noise with a free parameter attached. */\n  minLevelCount?: number\n}\n\nexport interface Design {\n  X: number[][]\n  y: number[]\n  names: string[]\n  /** level -> column, per covariate. The dropped reference level is absent. */\n  levels: Record<string, string[]>\n}\n\nfunction covariateValue(row: GradedRow, name: string): string {\n  const fromFeatures = row.features?.[name]\n  if (fromFeatures !== undefined && fromFeatures !== null) return String(fromFeatures)\n  const own = (row as unknown as Record<string, unknown>)[name]\n  return own === undefined || own === null ? \"(none)\" : String(own)\n}\n\nexport function buildDesign(rows: GradedRow[], spec: DesignSpec): Design {\n  const minCount = spec.minLevelCount ?? 5\n  const levels: Record<string, string[]> = {}\n\n  for (const name of spec.covariates) {\n    const counts = new Map<string, number>()\n    for (const row of rows) {\n      const value = covariateValue(row, name)\n      counts.set(value, (counts.get(value) ?? 0) + 1)\n    }\n    const kept = [...counts.entries()]\n      .filter(([, n]) => n >= minCount)\n      .map(([value]) => value)\n      .sort()\n    // Drop one level as the reference, or the design is collinear with\n    // the intercept and the fit is not identified.\n    levels[name] = kept.slice(1)\n  }\n\n  const names = [\"intercept\", \"logitConfidence\"]\n  for (const name of spec.covariates) for (const level of levels[name]) names.push(`${name}=${level}`)\n\n  const X: number[][] = []\n  const y: number[] = []\n  for (const row of rows) {\n    const features = [1, logit(row.confidence)]\n    for (const name of spec.covariates) {\n      const value = covariateValue(row, name)\n      for (const level of levels[name]) features.push(value === level ? 1 : 0)\n    }\n    X.push(features)\n    y.push(outcome(row))\n  }\n\n  return { X, y, names, levels }\n}\n\n/**\n * Logistic regression by iteratively reweighted least squares.\n *\n * Newton's method rather than gradient descent: it converges in a handful\n * of iterations with no learning rate to guess at, which matters because\n * this runs unattended on a few hundred rows. The ridge term is not\n * optional — with one-hot covariates and a small sample, a level that\n * happens to be perfectly separable sends its weight to infinity.\n */\nexport function fitLogistic(\n  X: number[][],\n  y: number[],\n  opts: { l2?: number; iterations?: number } = {},\n): number[] {\n  const l2 = opts.l2 ?? 1\n  const iterations = opts.iterations ?? 25\n  const n = X.length\n  const d = n > 0 ? X[0].length : 0\n  let w = new Array(d).fill(0)\n  if (n === 0 || d === 0) return w\n\n  for (let iter = 0; iter < iterations; iter++) {\n    // Hessian (X'WX + lambda I) and gradient X'(y - p) - lambda w\n    const H: number[][] = Array.from({ length: d }, () => new Array(d).fill(0))\n    const g = new Array(d).fill(0)\n\n    for (let i = 0; i < n; i++) {\n      let z = 0\n      for (let j = 0; j < d; j++) z += w[j] * X[i][j]\n      const p = sigmoid(z)\n      const weight = Math.max(p * (1 - p), 1e-8)\n      const residual = y[i] - p\n      for (let j = 0; j < d; j++) {\n        g[j] += residual * X[i][j]\n        for (let k = j; k < d; k++) H[j][k] += weight * X[i][j] * X[i][k]\n      }\n    }\n    for (let j = 0; j < d; j++) {\n      // Never penalise the intercept: shrinking it biases the base rate.\n      if (j > 0) {\n        H[j][j] += l2\n        g[j] -= l2 * w[j]\n      }\n      for (let k = 0; k < j; k++) H[j][k] = H[k][j]\n    }\n\n    const step = solve(H, g)\n    if (!step) break\n    let delta = 0\n    for (let j = 0; j < d; j++) {\n      w[j] += step[j]\n      delta += Math.abs(step[j])\n    }\n    if (delta < 1e-8) break\n  }\n  return w\n}\n\n/** Gaussian elimination with partial pivoting. Returns null on a singular\n *  system, which the caller treats as \"stop iterating\" rather than a crash. */\nfunction solve(A: number[][], b: number[]): number[] | null {\n  const n = b.length\n  const M = A.map((row, i) => [...row, b[i]])\n\n  for (let col = 0; col < n; col++) {\n    let pivot = col\n    for (let r = col + 1; r < n; r++) if (Math.abs(M[r][col]) > Math.abs(M[pivot][col])) pivot = r\n    if (Math.abs(M[pivot][col]) < 1e-12) return null\n    ;[M[col], M[pivot]] = [M[pivot], M[col]]\n\n    for (let r = 0; r < n; r++) {\n      if (r === col) continue\n      const factor = M[r][col] / M[col][col]\n      if (factor === 0) continue\n      for (let c = col; c <= n; c++) M[r][c] -= factor * M[col][c]\n    }\n  }\n  // Full Gauss-Jordan above, so the matrix is diagonal and back-substitution\n  // is a single division per row.\n  const x = new Array(n)\n  for (let i = 0; i < n; i++) x[i] = M[i][n] / M[i][i]\n  return x\n}\n\nexport interface RecalibrationModel {\n  weights: number[]\n  names: string[]\n  levels: Record<string, string[]>\n  covariates: string[]\n  n: number\n}\n\nexport function fitRecalibrator(rows: GradedRow[], spec: DesignSpec): RecalibrationModel {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const design = buildDesign(labeled, spec)\n  return {\n    weights: fitLogistic(design.X, design.y),\n    names: design.names,\n    levels: design.levels,\n    covariates: spec.covariates,\n    n: labeled.length,\n  }\n}\n\n/** Corrected probability that this row's prediction is right. */\nexport function applyRecalibrator(model: RecalibrationModel, row: GradedRow): number {\n  let z = model.weights[0] + model.weights[1] * logit(row.confidence)\n  let col = 2\n  for (const name of model.covariates) {\n    const value = covariateValue(row, name)\n    for (const level of model.levels[name] ?? []) {\n      if (value === level) z += model.weights[col]\n      col++\n    }\n  }\n  return sigmoid(z)\n}\n\n// ---------------------------------------------------------------------------\n// Honest comparison\n// ---------------------------------------------------------------------------\n\nexport interface CalibratorScore {\n  name: string\n  /** Held-out log loss, averaged over folds. Lower is better. */\n  logLoss: number\n  /** Held-out Brier score. */\n  brier: number\n}\n\nexport interface CalibratorComparison {\n  insufficient: boolean\n  n: number\n  minN: number\n  folds: number\n  scores: CalibratorScore[]\n  /** The winner on held-out log loss. */\n  best?: string\n  /** Fitted on everything, for use once `best` says it is worth using. */\n  model?: RecalibrationModel\n}\n\n/**\n * k-fold cross-validated comparison of raw confidence, a single global\n * temperature, and the covariate model.\n *\n * The whole point of this function is that it can say \"the covariates are\n * not worth it\". More parameters always look better in-sample, so an\n * in-sample improvement is not evidence of anything.\n */\nexport function compareCalibrators(\n  rows: GradedRow[],\n  spec: DesignSpec,\n  opts: { folds?: number; minN?: number; rng?: () => number } = {},\n): CalibratorComparison {\n  const folds = opts.folds ?? 5\n  const minN = opts.minN ?? 100\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const base = { insufficient: labeled.length < minN, n: labeled.length, minN, folds, scores: [] }\n  if (base.insufficient) return base\n\n  const rng = opts.rng ?? Math.random\n  const shuffled = [...labeled]\n  for (let i = shuffled.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    ;[shuffled[i], shuffled[j]] = [shuffled[j], shuffled[i]]\n  }\n\n  const predictors: Record<string, number[]> = { raw: [], temperature: [], covariate: [] }\n  const actuals: number[] = []\n\n  for (let fold = 0; fold < folds; fold++) {\n    const test = shuffled.filter((_, i) => i % folds === fold)\n    const train = shuffled.filter((_, i) => i % folds !== fold)\n    if (train.length === 0 || test.length === 0) continue\n\n    const flat = fitRecalibrator(train, { ...spec, covariates: [] })\n    const full = fitRecalibrator(train, spec)\n\n    for (const row of test) {\n      actuals.push(outcome(row))\n      predictors.raw.push(row.confidence)\n      predictors.temperature.push(applyRecalibrator(flat, row))\n      predictors.covariate.push(applyRecalibrator(full, row))\n    }\n  }\n\n  const scores = Object.entries(predictors).map(([name, predicted]) => ({\n    name,\n    logLoss: binaryLogLoss(predicted, actuals),\n    brier: binaryBrier(predicted, actuals),\n  }))\n  const best = [...scores].sort((a, b) => a.logLoss - b.logLoss)[0]?.name\n\n  return {\n    ...base,\n    scores,\n    best,\n    model: best === \"covariate\" ? fitRecalibrator(labeled, spec) : fitRecalibrator(labeled, { ...spec, covariates: [] }),\n  }\n}\n\nfunction binaryLogLoss(predicted: number[], actual: number[]): number {\n  if (predicted.length === 0) return 0\n  let sum = 0\n  for (let i = 0; i < predicted.length; i++) {\n    const p = Math.min(1 - EPS, Math.max(EPS, predicted[i]))\n    sum += actual[i] === 1 ? -Math.log(p) : -Math.log(1 - p)\n  }\n  return sum / predicted.length\n}\n\nfunction binaryBrier(predicted: number[], actual: number[]): number {\n  if (predicted.length === 0) return 0\n  let sum = 0\n  for (let i = 0; i < predicted.length; i++) sum += (predicted[i] - actual[i]) ** 2\n  return sum / predicted.length\n}\n","import { randomBytes } from \"crypto\"\nimport { getDecisionBackend } from \"./backend\"\nimport { answerView } from \"./store\"\nimport type { Questions, StateValue } from \"./types\"\n\n// Does the question hold still?\n//\n// Everything else in this module measures the BACKEND — is its confidence\n// calibrated, does it beat the incumbent. Nothing measured the QUESTIONS,\n// and a badly-phrased question does not announce itself: it produces\n// plausible numbers that happen to move every time you ask.\n//\n// TypeSafe's own consistency cookbook is the method. Ask the same rubric\n// over the same state N times, with a throwaway `uid` in the state so each\n// repeat is an independent draw rather than a cached one, and look at the\n// spread. Their published figure for jev across a 14-question rubric is a\n// mean per-question standard deviation of 0.0102; sampled LLM answers move\n// far more, and on the judgment calls disagree with themselves at\n// temperature 0.\n//\n// What this is for, concretely: a compound question — \"did it leave\n// anything unresolved, surprising, risky, or needing a decision\" — is four\n// questions sharing a name, and it scatters. An atomic one holds. This\n// turns that from an opinion about phrasing into a number, BEFORE any\n// ground truth exists to calibrate against. It is the cheapest quality\n// signal available: fifteen samples of a two-question seat costs about a\n// twentieth of a cent.\n//\n// It does not tell you the answer is RIGHT. A question can be perfectly\n// stable and perfectly wrong. Stability is a precondition for calibration\n// being meaningful, not a substitute for measuring it.\n\n/** The vendor's review band: act below LOW or above HIGH, escalate between. */\nexport const UNCERTAINTY_LOW = 0.3\nexport const UNCERTAINTY_HIGH = 0.7\n\nexport type BandDecision = \"no\" | \"uncertain\" | \"yes\"\n\nexport function band(\n  probability: number,\n  low = UNCERTAINTY_LOW,\n  high = UNCERTAINTY_HIGH,\n): BandDecision {\n  if (probability <= low) return \"no\"\n  if (probability > high) return \"yes\"\n  return \"uncertain\"\n}\n\nexport interface QuestionConsistency {\n  question: string\n  type: string\n  n: number\n  mean: number\n  /** Population standard deviation of the per-sample probability. For a\n   *  Noul that probability is P(true); for Choice and Score it is the\n   *  winning label's share. */\n  stdev: number\n  min: number\n  max: number\n  /** Distinct predicted labels across the samples. More than one means the\n   *  seat's own answer flips between identical calls. */\n  distinctAnswers: string[]\n  /** Distinct band decisions. Two or more means the samples straddle a\n   *  threshold, so which action fires is decided by luck. */\n  bands: BandDecision[]\n  /** True when the samples disagree about what to DO, not merely by how\n   *  much. This is the finding that matters. */\n  crossesThreshold: boolean\n  samples: number[]\n}\n\nexport interface ConsistencyReport {\n  backend: string\n  model: string\n  samples: number\n  /** Mean of the per-question standard deviations — TypeSafe publishes\n   *  0.0102 for jev over a 14-question rubric, as a reference point. */\n  meanStdev: number\n  unstable: string[]\n  questions: QuestionConsistency[]\n  usage: { inputTokens: number; outputTokens: number }\n  totalMs: number\n}\n\n/**\n * Resample one state N times and report the spread per question.\n *\n * The `uid` is not decoration. Without a field that changes per call the\n * backend may serve a cached answer, and a cache returns a standard\n * deviation of zero for any question, however badly phrased.\n */\nexport async function measureConsistency(\n  backendName: string,\n  state: StateValue,\n  questions: Questions,\n  opts: { samples?: number; model?: string; low?: number; high?: number } = {},\n): Promise<ConsistencyReport> {\n  const samples = Math.max(2, opts.samples ?? 15)\n  const backend = getDecisionBackend(backendName)\n  const started = Date.now()\n\n  const runs = await Promise.all(\n    Array.from({ length: samples }, (_, i) =>\n      backend.decide({\n        state: withUid(state, `${i}:${randomBytes(4).toString(\"hex\")}`),\n        questions,\n        model: opts.model,\n      }),\n    ),\n  )\n\n  const names = Object.keys(questions)\n  const usage = runs.reduce(\n    (acc, r) => ({\n      inputTokens: acc.inputTokens + r.usage.inputTokens,\n      outputTokens: acc.outputTokens + r.usage.outputTokens,\n    }),\n    { inputTokens: 0, outputTokens: 0 },\n  )\n\n  const perQuestion = names.map((name) => {\n    const views = runs.map((r) => answerView((r.answers as Record<string, any>)[name]))\n    // The right scalar differs by type, and getting it wrong is silent.\n    //\n    // For a Noul the quantity is P(true) — that is what the answer IS, and\n    // what the 0.30/0.70 band is defined on. Using the winning label's\n    // probability instead reports 1-p whenever the answer is \"no\", so a\n    // P(true) of 0.26 displays as 0.74 and every noul lands above 0.5,\n    // which makes the band useless. For Choice and Score there is no\n    // single \"true\" outcome, so the winning label's probability is the\n    // comparable scalar.\n    const probs = views.map((v, i) => {\n      const answer = (runs[i].answers as Record<string, any>)[name]\n      return answer.type === \"noul\" ? answer.noul : v.topProb\n    })\n    const answers = [...new Set(views.map((v) => v.predicted))]\n    const bands = [...new Set(probs.map((p) => band(p, opts.low, opts.high)))]\n    const m = mean(probs)\n    return {\n      question: name,\n      type: (runs[0].answers as Record<string, any>)[name].type,\n      n: samples,\n      mean: m,\n      stdev: stdev(probs, m),\n      min: Math.min(...probs),\n      max: Math.max(...probs),\n      distinctAnswers: answers,\n      bands,\n      crossesThreshold: answers.length > 1 || bands.length > 1,\n      samples: probs,\n    }\n  })\n\n  return {\n    backend: backendName,\n    model: runs[0].model,\n    samples,\n    meanStdev: mean(perQuestion.map((q) => q.stdev)),\n    unstable: perQuestion.filter((q) => q.crossesThreshold).map((q) => q.question),\n    questions: perQuestion,\n    usage,\n    totalMs: Date.now() - started,\n  }\n}\n\n/** Add a throwaway field without disturbing a string state's meaning. */\nexport function withUid(state: StateValue, uid: string): StateValue {\n  if (state !== null && typeof state === \"object\" && !Array.isArray(state)) {\n    return { ...(state as Record<string, StateValue>), uid }\n  }\n  return { uid, state }\n}\n\nfunction mean(values: number[]): number {\n  if (values.length === 0) return 0\n  return values.reduce((a, b) => a + b, 0) / values.length\n}\n\nfunction stdev(values: number[], m: number): number {\n  if (values.length < 2) return 0\n  return Math.sqrt(values.reduce((sum, v) => sum + (v - m) ** 2, 0) / values.length)\n}\n","import type {\n  AnswerMode,\n  DecisionBackend,\n  DecisionBackendCapabilities,\n  DecisionRequest,\n  DecisionResponse,\n} from \"../backend\"\nimport { finalizeAnswer } from \"../normalize\"\nimport { labelsOf, validateQuestions } from \"../questions\"\nimport type { AnswersFor, AnyAnswer, Questions, RawAnswer } from \"../types\"\n\n// A backend with no model behind it.\n//\n// Two jobs: let the seat, the store and the calibration maths be tested\n// without a network or an API key, and give `agentx decisions` something to\n// smoke against. Answers are deterministic in (state, question, label), so\n// the same input always produces the same distribution and a test can assert\n// on exact numbers.\n\nexport interface MockBackendOptions {\n  /** Raw answers keyed by question name. Anything not scripted is derived\n   *  deterministically from the state. */\n  answers?: Record<string, RawAnswer>\n  /** Rejects with this instead of answering. For exercising the seat's\n   *  fail-open path. */\n  fail?: Error\n  latencyMs?: number\n}\n\nconst capabilities: DecisionBackendCapabilities = {\n  probabilitySource: \"synthetic\",\n  calibratedProbabilities: false,\n  maxChoiceOptions: 255,\n  maxStateChars: 24_000,\n  parallelQuestions: true,\n  images: false,\n}\n\nexport function createMockDecisionBackend(opts: MockBackendOptions = {}): DecisionBackend {\n  return {\n    name: \"mock\",\n    capabilities,\n    async decide<Q extends Questions>(\n      request: DecisionRequest<Q>,\n    ): Promise<DecisionResponse<Q>> {\n      if (opts.fail) throw opts.fail\n      validateQuestions(request.questions)\n\n      const stateText = JSON.stringify(request.state ?? null)\n      const answers: Record<string, AnyAnswer> = {}\n      let repaired = false\n\n      for (const [name, question] of Object.entries(request.questions)) {\n        const raw = opts.answers?.[name] ?? syntheticRaw(stateText, name, question)\n        const result = finalizeAnswer(question, raw)\n        answers[name] = result.answer\n        if (result.repaired) repaired = true\n      }\n\n      const answerMode: AnswerMode = \"probabilities\"\n      return {\n        model: request.model ?? \"mock\",\n        answers: answers as AnswersFor<Q>,\n        usage: { inputTokens: stateText.length, outputTokens: 0 },\n        meta: {\n          backend: \"mock\",\n          structureMode: \"mock\",\n          answerMode,\n          repaired,\n          retries: 0,\n          stateTruncated: false,\n          latencyMs: opts.latencyMs ?? 0,\n        },\n      }\n    },\n  }\n}\n\nfunction syntheticRaw(\n  stateText: string,\n  name: string,\n  question: Questions[string],\n): RawAnswer {\n  if (question.type === \"noul\") {\n    return { noul: unitHash(`${stateText}|${name}|noul`) }\n  }\n  const labels = labelsOf(question)\n  const weights = labels.map((label) => unitHash(`${stateText}|${name}|${label}`) + 1e-3)\n  const total = weights.reduce((a, b) => a + b, 0)\n  const probabilities: Record<string, number> = {}\n  labels.forEach((label, i) => {\n    probabilities[label] = weights[i] / total\n  })\n  return { probabilities }\n}\n\n/** FNV-1a, mapped to [0, 1). Stable across runs and platforms, which is the\n *  only property that matters here. */\nfunction unitHash(input: string): number {\n  let h = 0x811c9dc5\n  for (let i = 0; i < input.length; i++) {\n    h ^= input.charCodeAt(i)\n    h = Math.imul(h, 0x01000193) >>> 0\n  }\n  return h / 0x100000000\n}\n","// --- Provider capability matrix ---\n// Used to warn users when an agent's config requires features\n// the selected provider doesn't support.\n\nexport interface ProviderCapabilities {\n  streaming: boolean\n  tools: boolean\n  vision: boolean\n  thinking: boolean\n  maxContext: number\n  /** Can be made to return a value matching a schema, rather than prose the\n   *  caller has to parse. False does NOT mean \"no structured output\" — it\n   *  means no *guarantee*, so callers fall back to JSON-in-text plus a\n   *  correction round. The claude-code CLI on an OAuth credential is the\n   *  case that matters: ClaudeCodeProvider.generateRaw() throws there. */\n  structuredOutput: boolean\n  /** Exposes per-token logprobs. The only source of a probability that is\n   *  the model's own posterior rather than a number it wrote out. */\n  logprobs: boolean\n}\n\nexport const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {\n  // structuredOutput is false here on purpose. ClaudeCodeProvider can force\n  // a tool on an api-key credential but throws on OAuth (\"generateRaw() not\n  // available for OAuth/CLI mode\"), and this matrix is static — it cannot\n  // see which credential an operator has. False is the safe default: a\n  // caller that believes a guarantee it doesn't have crashes on the\n  // critical path, whereas a caller that falls back to JSON-in-text only\n  // spends a correction round.\n  \"claude-code\": {\n    streaming: true,\n    tools: true,\n    vision: true,\n    thinking: true,\n    maxContext: 1_000_000,\n    structuredOutput: false,\n    logprobs: false,\n  },\n  claude: {\n    streaming: true,\n    tools: true,\n    vision: true,\n    thinking: true,\n    maxContext: 1_000_000,\n    structuredOutput: true,\n    logprobs: false,\n  },\n  openai: {\n    streaming: true,\n    tools: true,\n    vision: true,\n    thinking: false,\n    maxContext: 128_000,\n    structuredOutput: true,\n    logprobs: true,\n  },\n  deepseek: {\n    streaming: true,\n    tools: true,\n    vision: false,\n    thinking: true,\n    maxContext: 128_000,\n    structuredOutput: true,\n    logprobs: true,\n  },\n  ollama: {\n    streaming: true,\n    tools: false,\n    vision: false,\n    thinking: false,\n    maxContext: 32_000,\n    structuredOutput: false,\n    logprobs: false,\n  },\n  demo: {\n    streaming: true,\n    tools: false,\n    vision: false,\n    thinking: false,\n    maxContext: 8_000,\n    structuredOutput: false,\n    logprobs: false,\n  },\n  // An OpenAI-compatible endpoint behind providers.<name>.baseUrl. What it\n  // supports depends entirely on what is running there, so claim nothing.\n  custom: {\n    streaming: true,\n    tools: false,\n    vision: false,\n    thinking: false,\n    maxContext: 32_000,\n    structuredOutput: false,\n    logprobs: false,\n  },\n}\n\n/**\n * Check provider capabilities and return warnings for missing features.\n */\nexport function checkCapabilities(\n  providerName: string,\n  requiredFeatures?: string[],\n): string[] {\n  const caps = PROVIDER_CAPABILITIES[providerName]\n  if (!caps) {\n    return [`Unknown provider \"${providerName}\" — capabilities unknown`]\n  }\n\n  if (!requiredFeatures?.length) return []\n\n  const warnings: string[] = []\n  const missing: string[] = []\n\n  for (const feature of requiredFeatures) {\n    if (feature in caps && !(caps as any)[feature]) {\n      missing.push(feature)\n    }\n  }\n\n  if (missing.length) {\n    warnings.push(\n      `Provider \"${providerName}\" lacks: ${missing.join(\", \")}. Some features will be degraded.`,\n    )\n  }\n\n  return warnings\n}\n","import { createProvider, type ProviderName } from \"@/agent/providers\"\nimport { PROVIDER_CAPABILITIES } from \"@/agent/providers/capabilities\"\nimport type { AgentProvider } from \"@/agent/providers/types\"\nimport { extractJson } from \"@/utils/extract-json\"\nimport type {\n  AnswerMode,\n  DecisionBackend,\n  DecisionBackendCapabilities,\n  DecisionRequest,\n  DecisionResponse,\n  StructureMode,\n} from \"../backend\"\nimport { finalizeAnswer } from \"../normalize\"\nimport {\n  SYSTEM_PROMPT,\n  TOOL_NAME,\n  renderShapeInstructions,\n  renderUserPrompt,\n  toolInputSchema,\n} from \"../prompt\"\nimport { validateQuestions } from \"../questions\"\nimport { describeIssues, rawAnswersSchema } from \"../schema\"\nimport type { AnswersFor, AnyAnswer, Questions, RawAnswer, RawAnswers } from \"../types\"\nimport { z } from \"zod\"\n\n// The System One contract, served by an ordinary chat model.\n//\n// READ THIS BEFORE TRUSTING A NUMBER THIS BACKEND RETURNS.\n//\n// In \"tool\" mode the model is forced to emit a value matching a schema, so\n// the answer is guaranteed well-formed. In \"text\" mode we ask nicely and\n// repair. Neither produces a CALIBRATED probability. A number the model\n// wrote into its answer is a number it chose; it is not a measurement of\n// its own uncertainty, and models tuned on human preference are\n// systematically overconfident about it — which is the entire premise of\n// the decision-model vendors.\n//\n// What this backend genuinely buys:\n//   - one call for N questions against one shared state\n//   - a guaranteed shape, so no call site hand-parses prose\n//   - an ordinal signal that is usually better than nothing\n//   - rows in the shadow store, which is where calibration actually\n//     comes from: a temperature fitted on a few hundred of your own\n//     labeled decisions beats any number the model hands you\n//\n// A backend reporting `calibratedProbabilities: false` is not a defect to\n// fix later. It is the honest value, and the harness exists to measure how\n// much it costs.\n\nconst DEFAULT_PROVIDER: ProviderName = \"claude-code\"\nconst DEFAULT_MODEL = \"claude-haiku-4-5-20251001\"\nconst DEFAULT_MAX_STATE_CHARS = 24_000\nconst DEFAULT_MAX_TOKENS = 2_000\nconst DEFAULT_TIMEOUT_MS = 30_000\n\nexport interface LocalBackendOptions {\n  provider?: ProviderName\n  model?: string\n  /** \"auto\" resolves to \"tool\" when the provider guarantees structured\n   *  output and \"text\" otherwise. Never resolves to \"logprobs\": that mode\n   *  costs one call per question, so it has to be asked for explicitly. */\n  structureMode?: \"auto\" | \"tool\" | \"text\"\n  answerMode?: AnswerMode\n  normalizeProbabilities?: boolean\n  nRetryMalformedStructure?: number\n  temperature?: number\n  maxStateChars?: number\n  maxTokens?: number\n  timeoutMs?: number\n  /** Injected in tests. Production resolves through createProvider. */\n  providerFactory?: () => AgentProvider\n}\n\nexport function createLocalDecisionBackend(opts: LocalBackendOptions = {}): DecisionBackend {\n  if (opts.answerMode && opts.answerMode !== \"probabilities\") {\n    throw new Error(\n      `local decision backend supports answerMode \"probabilities\" only (got \"${opts.answerMode}\")`,\n    )\n  }\n\n  const maxStateChars = opts.maxStateChars ?? DEFAULT_MAX_STATE_CHARS\n  const capabilities: DecisionBackendCapabilities = {\n    probabilitySource: \"verbalized\",\n  calibratedProbabilities: false,\n    maxChoiceOptions: 255,\n    maxStateChars,\n    parallelQuestions: true,\n    images: false,\n  }\n\n  let provider: AgentProvider | null = null\n  const resolveProvider = (): AgentProvider => {\n    if (!provider) {\n      provider = opts.providerFactory\n        ? opts.providerFactory()\n        : createProvider(opts.provider ?? DEFAULT_PROVIDER)\n    }\n    return provider\n  }\n\n  return {\n    name: \"local\",\n    capabilities,\n    async decide<Q extends Questions>(\n      request: DecisionRequest<Q>,\n    ): Promise<DecisionResponse<Q>> {\n      validateQuestions(request.questions)\n\n      const started = Date.now()\n      const model = request.model ?? opts.model ?? DEFAULT_MODEL\n      const agent = resolveProvider()\n      const structureMode = resolveStructureMode(opts.structureMode ?? \"auto\", agent)\n      const prompt = renderUserPrompt(request.state, request.questions, maxStateChars)\n      const schema = rawAnswersSchema(request.questions)\n      const maxRetries = opts.nRetryMalformedStructure ?? 1\n\n      const { signal, dispose } = deadline(\n        request.abortSignal,\n        request.timeoutMs ?? opts.timeoutMs ?? DEFAULT_TIMEOUT_MS,\n      )\n\n      let usage = { inputTokens: 0, outputTokens: 0 }\n      let correction: string | null = null\n      let retries = 0\n      let raw: RawAnswers | null = null\n      let lastError = \"no attempt was made\"\n\n      try {\n        for (let attempt = 0; attempt <= maxRetries; attempt++) {\n          if (attempt > 0) retries++\n\n          const call =\n            structureMode === \"tool\"\n              ? callWithForcedTool(agent, request.questions, prompt.text, correction, {\n                  model,\n                  maxTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n                  temperature: opts.temperature ?? 0,\n                  abortSignal: signal,\n                })\n              : callWithText(agent, request.questions, prompt.text, correction, {\n                  model,\n                  maxTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n                  temperature: opts.temperature ?? 0,\n                  abortSignal: signal,\n                })\n\n          const result = await call\n          usage = {\n            inputTokens: usage.inputTokens + result.usage.inputTokens,\n            outputTokens: usage.outputTokens + result.usage.outputTokens,\n          }\n\n          const parsed = schema.safeParse(result.answers)\n          if (parsed.success) {\n            raw = parsed.data as RawAnswers\n            break\n          }\n          lastError = describeIssues(parsed.error as z.ZodError)\n          correction = [\n            \"Your previous reply did not match the required structure:\",\n            lastError,\n            \"\",\n            \"Reply again with the corrected structure only.\",\n          ].join(\"\\n\")\n        }\n\n        if (!raw) {\n          throw new Error(\n            `local decision backend could not get a valid answer after ${retries + 1} attempt(s): ${lastError}`,\n          )\n        }\n\n        const answers: Record<string, AnyAnswer> = {}\n        let repaired = false\n        for (const [name, question] of Object.entries(request.questions)) {\n          const result = finalizeAnswer(question, raw[name] as RawAnswer, {\n            normalize: opts.normalizeProbabilities ?? true,\n          })\n          answers[name] = result.answer\n          if (result.repaired) repaired = true\n        }\n\n        return {\n          model,\n          answers: answers as AnswersFor<Q>,\n          usage,\n          meta: {\n            backend: \"local\",\n            structureMode,\n            answerMode: \"probabilities\",\n            repaired,\n            retries,\n            stateTruncated: prompt.truncated,\n            latencyMs: Date.now() - started,\n          },\n        }\n      } finally {\n        dispose()\n      }\n    },\n  }\n}\n\ninterface AttemptResult {\n  answers: unknown\n  usage: { inputTokens: number; outputTokens: number }\n}\n\ninterface CallOptions {\n  model: string\n  maxTokens: number\n  temperature: number\n  abortSignal: AbortSignal\n}\n\nasync function callWithForcedTool(\n  agent: AgentProvider,\n  questions: Questions,\n  userPrompt: string,\n  correction: string | null,\n  options: CallOptions,\n): Promise<AttemptResult> {\n  if (!agent.generateRaw) {\n    throw new Error(`provider \"${agent.name}\" has no generateRaw(); use structureMode \"text\"`)\n  }\n\n  const messages = [\n    { role: \"user\" as const, content: userPrompt },\n    ...(correction ? [{ role: \"user\" as const, content: correction }] : []),\n  ]\n\n  const result = await agent.generateRaw(\n    messages,\n    SYSTEM_PROMPT,\n    [\n      {\n        name: TOOL_NAME,\n        description: \"Report a probability distribution for every question.\",\n        input_schema: toolInputSchema(questions),\n      },\n    ],\n    { ...options, toolChoice: { type: \"tool\", name: TOOL_NAME } },\n  )\n\n  const block = result.content.find((b) => b.type === \"tool_use\" && b.name === TOOL_NAME)\n  const input = block && block.type === \"tool_use\" ? block.input : null\n\n  return {\n    answers: unwrapAnswers(input),\n    usage: {\n      inputTokens: result.usage.input_tokens,\n      outputTokens: result.usage.output_tokens,\n    },\n  }\n}\n\nasync function callWithText(\n  agent: AgentProvider,\n  questions: Questions,\n  userPrompt: string,\n  correction: string | null,\n  options: CallOptions,\n): Promise<AttemptResult> {\n  const content = [userPrompt, renderShapeInstructions(questions)].join(\"\\n\")\n  const messages = [\n    { role: \"system\" as const, content: SYSTEM_PROMPT },\n    { role: \"user\" as const, content },\n    ...(correction ? [{ role: \"user\" as const, content: correction }] : []),\n  ]\n\n  const result = await agent.generate(messages, options)\n  return {\n    answers: unwrapAnswers(extractJson(result.content ?? \"\")),\n    usage: { inputTokens: 0, outputTokens: result.tokensUsed ?? 0 },\n  }\n}\n\n/** Accept both `{ answers: {...} }` and a bare answers map. Models drop the\n *  wrapper often enough that rejecting it would spend correction rounds on\n *  a difference that carries no information. */\nfunction unwrapAnswers(value: unknown): unknown {\n  if (value && typeof value === \"object\" && !Array.isArray(value)) {\n    const wrapper = value as { answers?: unknown }\n    if (wrapper.answers && typeof wrapper.answers === \"object\") return wrapper.answers\n  }\n  return value\n}\n\nfunction resolveStructureMode(\n  requested: \"auto\" | \"tool\" | \"text\",\n  agent: AgentProvider,\n): StructureMode {\n  if (requested !== \"auto\") return requested\n  if (!agent.generateRaw) return \"text\"\n  // Resolved name, not the configured one: createProvider(\"claude-code\")\n  // falls through to whatever the stored auth config selected.\n  return PROVIDER_CAPABILITIES[agent.name]?.structuredOutput ? \"tool\" : \"text\"\n}\n\n/** One signal that fires on either the caller's cancel or our own timeout.\n *  Hand-rolled rather than AbortSignal.any so this does not depend on the\n *  lib target's typings. */\nfunction deadline(\n  caller: AbortSignal | undefined,\n  timeoutMs: number,\n): { signal: AbortSignal; dispose: () => void } {\n  const controller = new AbortController()\n  const timer = setTimeout(\n    () => controller.abort(new Error(`decision timed out after ${timeoutMs}ms`)),\n    timeoutMs,\n  )\n  const onCallerAbort = () => controller.abort(caller?.reason)\n\n  if (caller) {\n    if (caller.aborted) onCallerAbort()\n    else caller.addEventListener(\"abort\", onCallerAbort, { once: true })\n  }\n\n  return {\n    signal: controller.signal,\n    dispose: () => {\n      clearTimeout(timer)\n      caller?.removeEventListener(\"abort\", onCallerAbort)\n    },\n  }\n}\n","import type {\n  DecisionBackend,\n  DecisionBackendCapabilities,\n  DecisionRequest,\n  DecisionResponse,\n  ProbabilitySource,\n} from \"../backend\"\nimport { finalizeAnswer } from \"../normalize\"\nimport { validateQuestions } from \"../questions\"\nimport { rawAnswersSchema, describeIssues } from \"../schema\"\nimport type { AnswersFor, AnyAnswer, Questions, RawAnswer } from \"../types\"\nimport { renderState } from \"../prompt\"\nimport type { z } from \"zod\"\nimport { readFileSync } from \"fs\"\nimport { homedir } from \"os\"\nimport { join } from \"path\"\n\n// A backend for a simple-jev server (featherless-ai/simple-jev).\n//\n// Why this exists alongside the local one: simple-jev prefills the prompt\n// and reads the model's next-token logits over the allowed answer labels,\n// instead of asking a chat model to write a probability into its reply. The\n// difference is visible in the output — a verbalized distribution comes back\n// as 0.25/0.35/0.40, a logit one as 7.48e-05/9.61e-05/0.99983. One is a\n// number a model chose; the other is its posterior over the label set.\n//\n// It is NOT a Jev replacement, and its own documentation says so: no\n// reproduction of TypeSafe's architecture or training, and its output is\n// explicitly \"not calibrated probabilities of correctness\". That last point\n// is why this reports `calibratedProbabilities: false` like everything else\n// — correctness calibration comes from a recalibrator fitted on your rows.\n//\n// Two deliberate choices about its response:\n//\n//   1. Its `confidence` is discarded and recomputed. simple-jev reports the\n//      largest probability, which is not comparable across option counts:\n//      uniform over 2 scores 0.5 and uniform over 10 scores 0.1, though both\n//      are maximally uncertain. Our normalized-entropy definition is, and\n//      rows from different backends have to be comparable in one store or\n//      calibration pools apples and oranges.\n//\n//   2. No correction rounds. simple-jev builds the JSON server-side from\n//      scores, so a schema mismatch is contract drift, not a model slip, and\n//      re-asking cannot fix it. Failing loudly is the honest response.\n//\n// Licence note: as of 2026-09-18 the upstream repository ships no LICENSE\n// file, which under default copyright means all rights reserved. That is a\n// blocker for depending on a self-hosted deployment in production, though\n// not for talking to a server someone else is running. Resolve it before\n// this backend carries real traffic.\n\nexport interface SimpleJevOptions {\n  /** Server root, e.g. http://127.0.0.1:8000/v1 */\n  baseUrl?: string\n  /** Route under baseUrl. simple-jev serves /classifier (and /systemone as\n   *  an alias); other System One-shaped endpoints differ. */\n  path?: string\n  /** Reported in capabilities and used to keep incomparable rows out of one\n   *  calibration pool. Does not change behaviour. */\n  probabilitySource?: ProbabilitySource\n  /** Name this backend reports. Rows are grouped by it in the store. */\n  name?: string\n  model?: string\n  apiKeyEnv?: string\n  /** Read the key from this file when the env var is unset. The daemon\n   *  gets its key from the launchd plist / systemd unit, but the CLI and\n   *  the wiki server are separate processes with a bare environment —\n   *  without this a seat silently fails open on every CLI call. Mirrors\n   *  ~/.agentx/mesh-token.txt. */\n  apiKeyFile?: string\n  timeoutMs?: number\n  maxStateChars?: number\n  /** Their Choice accepts 2-50 candidates, well under Jev's 255. */\n  maxChoiceOptions?: number\n  fetchImpl?: typeof fetch\n}\n\nconst DEFAULT_BASE_URL = \"http://127.0.0.1:8000/v1\"\nconst DEFAULT_TIMEOUT_MS = 30_000\nconst DEFAULT_MAX_STATE_CHARS = 6_000\nconst DEFAULT_MAX_CHOICE_OPTIONS = 50\n\nexport function createSimpleJevBackend(opts: SimpleJevOptions = {}): DecisionBackend {\n  const baseUrl = (opts.baseUrl ?? DEFAULT_BASE_URL).replace(/\\/+$/, \"\")\n  const path = opts.path ?? \"/classifier\"\n  const name = opts.name ?? \"simple-jev\"\n  const maxStateChars = opts.maxStateChars ?? DEFAULT_MAX_STATE_CHARS\n  const maxChoiceOptions = opts.maxChoiceOptions ?? DEFAULT_MAX_CHOICE_OPTIONS\n\n  const capabilities: DecisionBackendCapabilities = {\n    probabilitySource: opts.probabilitySource ?? \"logits\",\n    calibratedProbabilities: false,\n    maxChoiceOptions,\n    maxStateChars,\n    parallelQuestions: true,\n    images: false,\n  }\n\n  return {\n    name,\n    capabilities,\n    async decide<Q extends Questions>(\n      request: DecisionRequest<Q>,\n    ): Promise<DecisionResponse<Q>> {\n      validateQuestions(request.questions)\n      assertChoiceCardinality(request.questions, maxChoiceOptions)\n\n      const started = Date.now()\n      const model = request.model ?? opts.model\n      if (!model) throw new Error(\"simple-jev backend needs a model id\")\n\n      const rendered = renderState(request.state, maxStateChars)\n      const doFetch = opts.fetchImpl ?? fetch\n\n      const { signal, dispose } = deadline(\n        request.abortSignal,\n        request.timeoutMs ?? opts.timeoutMs ?? DEFAULT_TIMEOUT_MS,\n      )\n\n      try {\n        const headers: Record<string, string> = { \"Content-Type\": \"application/json\" }\n        const apiKey = resolveKey(opts.apiKeyEnv, opts.apiKeyFile)\n        if (apiKey) headers.Authorization = `Bearer ${apiKey}`\n\n        const res = await doFetch(`${baseUrl}${path}`, {\n          method: \"POST\",\n          headers,\n          body: JSON.stringify({\n            model,\n            state: rendered.text,\n            questions: withInstructions(request.questions),\n          }),\n          signal,\n        })\n\n        if (!res.ok) {\n          const detail = await res.text().catch(() => \"\")\n          throw new Error(`${name} ${res.status}: ${detail.slice(0, 300)}`)\n        }\n\n        const body = (await res.json()) as {\n          model?: string\n          answers?: unknown\n          usage?: { input_tokens?: number; output_tokens?: number }\n        }\n\n        const parsed = rawAnswersSchema(request.questions).safeParse(body.answers)\n        if (!parsed.success) {\n          throw new Error(\n            `${name} returned a response that does not match the requested questions ` +\n              `(contract drift, not a model slip — retrying will not help):\\n${describeIssues(parsed.error as z.ZodError)}`,\n          )\n        }\n\n        const raw = parsed.data as Record<string, RawAnswer>\n        const answers: Record<string, AnyAnswer> = {}\n        let repaired = false\n        for (const [name, question] of Object.entries(request.questions)) {\n          const result = finalizeAnswer(question, raw[name])\n          answers[name] = result.answer\n          if (result.repaired) repaired = true\n        }\n\n        return {\n          model: body.model ?? model,\n          answers: answers as AnswersFor<Q>,\n          usage: {\n            inputTokens: body.usage?.input_tokens ?? 0,\n            outputTokens: body.usage?.output_tokens ?? 0,\n          },\n          meta: {\n            backend: name,\n            // Mirrors probabilitySource, because structure_mode is what keeps\n            // incomparable rows out of one calibration pool. Hardcoding\n            // \"logprobs\" here quietly pooled Jev — a model claiming a trained\n            // objective — with logit-reading over an untrained open model.\n            // Different mechanisms, different miscalibration curves.\n            structureMode: capabilities.probabilitySource === \"native\" ? \"native\" : \"logprobs\",\n            answerMode: \"probabilities\",\n            repaired,\n            retries: 0,\n            stateTruncated: rendered.truncated,\n            latencyMs: Date.now() - started,\n          },\n        }\n      } finally {\n        dispose()\n      }\n    },\n  }\n}\n\n/** simple-jev requires `instructions` on every question; TypeSafe's own SDK\n *  marks it optional. Adapting here rather than tightening our builders is\n *  deliberate: the point of a backend registry is that one question set runs\n *  on every backend and can be graded against identical states. A contract\n *  that only works on one server would defeat the comparison this exists for.\n *\n *  The substitute is neutral and deterministic — it restates what the\n *  criteria already encode, and the caller's original questions are what get\n *  stored in the shadow store, so nothing about the recorded request is\n *  silently rewritten. */\nconst DEFAULT_INSTRUCTIONS: Record<string, string> = {\n  choice: \"Which option best fits?\",\n  score: \"Which level best fits?\",\n  noul: \"Is this true of the state?\",\n}\n\nfunction withInstructions(questions: Questions): Questions {\n  const out: Record<string, unknown> = {}\n  for (const [name, question] of Object.entries(questions)) {\n    out[name] = question.instructions\n      ? question\n      : { ...question, instructions: DEFAULT_INSTRUCTIONS[question.type] }\n  }\n  return out as Questions\n}\n\n/** Env var first, then the file. Cached per path: a seat on a hot path\n *  must not stat the filesystem on every call. */\nconst keyFileCache = new Map<string, string | null>()\n\nfunction resolveKey(envVar?: string, file?: string): string | undefined {\n  const fromEnv = envVar ? process.env[envVar] : undefined\n  if (fromEnv) return fromEnv\n  if (!file) return undefined\n\n  const path = file.startsWith(\"~/\") ? join(homedir(), file.slice(2)) : file\n  if (!keyFileCache.has(path)) {\n    try {\n      const raw = readFileSync(path, \"utf-8\").trim()\n      keyFileCache.set(path, raw.length > 0 ? raw : null)\n    } catch {\n      keyFileCache.set(path, null)\n    }\n  }\n  return keyFileCache.get(path) ?? undefined\n}\n\nfunction assertChoiceCardinality(questions: Questions, max: number): void {\n  for (const [questionName, question] of Object.entries(questions)) {\n    if (question.type === \"choice\") {\n      const n = Object.keys(question.criteria).length\n      if (n > max) {\n        throw new Error(\n          `question \"${questionName}\": accepts at most ${max} choice options, got ${n}`,\n        )\n      }\n    } else if (question.type === \"score\" && question.criteria.length > max) {\n      throw new Error(\n        `question \"${questionName}\": accepts at most ${max} score levels, got ${question.criteria.length}`,\n      )\n    }\n  }\n}\n\nfunction deadline(\n  caller: AbortSignal | undefined,\n  timeoutMs: number,\n): { signal: AbortSignal; dispose: () => void } {\n  const controller = new AbortController()\n  const timer = setTimeout(\n    () => controller.abort(new Error(`decision timed out after ${timeoutMs}ms`)),\n    timeoutMs,\n  )\n  const onCallerAbort = () => controller.abort(caller?.reason)\n  if (caller) {\n    if (caller.aborted) onCallerAbort()\n    else caller.addEventListener(\"abort\", onCallerAbort, { once: true })\n  }\n  return {\n    signal: controller.signal,\n    dispose: () => {\n      clearTimeout(timer)\n      caller?.removeEventListener(\"abort\", onCallerAbort)\n    },\n  }\n}\n","// Typed-decision seat.\n//\n// A decision seat does not make a decision better. It makes the decision's\n// uncertainty legible. Attached to a call site with no ground truth, it buys\n// a more expensive number and a dashboard that cannot be wrong — so the\n// calibration harness is not optional polish, it is the only thing that\n// separates this from the self-reported `confidence` fields the codebase\n// already has several of and trusts none of.\n//\n// See docs/architecture (and the plan that introduced this module) for the\n// full contract. Start at ./types.ts.\n\nexport * from \"./types\"\nexport * from \"./questions\"\nexport * from \"./normalize\"\nexport * from \"./schema\"\nexport * from \"./backend\"\nexport * from \"./prompt\"\nexport * from \"./store\"\nexport * from \"./calibration\"\nexport * from \"./recalibrate\"\nexport * from \"./consistency\"\nexport * from \"./seat\"\nexport { createMockDecisionBackend } from \"./backends/mock\"\nexport type { MockBackendOptions } from \"./backends/mock\"\nexport { createLocalDecisionBackend } from \"./backends/local-llm\"\nexport type { LocalBackendOptions } from \"./backends/local-llm\"\nexport { createSimpleJevBackend } from \"./backends/simple-jev\"\nexport type { SimpleJevOptions } from \"./backends/simple-jev\"\n\nimport { registerDecisionBackend } from \"./backend\"\nimport { createMockDecisionBackend } from \"./backends/mock\"\nimport { createLocalDecisionBackend, type LocalBackendOptions } from \"./backends/local-llm\"\nimport { createSimpleJevBackend, type SimpleJevOptions } from \"./backends/simple-jev\"\n\n/** Register the backends that ship in-tree. Explicit rather than\n *  import-time so tests can start from an empty registry. */\nexport function registerBuiltinDecisionBackends(\n  local: LocalBackendOptions = {},\n  simpleJev: SimpleJevOptions = {},\n  jev: SimpleJevOptions = {},\n  typesafe: SimpleJevOptions = {},\n): void {\n  registerDecisionBackend(\"mock\", () => createMockDecisionBackend())\n  registerDecisionBackend(\"local\", () => createLocalDecisionBackend(local))\n  registerDecisionBackend(\"simple-jev\", () => createSimpleJevBackend(simpleJev))\n  // Jev via OpenRouter's alpha decisions endpoint. Verified against a live\n  // key on 2026-09-19: POST /api/alpha/decisions with model \"jev-latest\"\n  // resolves to typesafe/jev-1.13-20260917 with provider \"TypeSafe\", so\n  // this is the real model rather than a proxy to something else. All\n  // three primitives round-tripped in 0.49s from Tunisia at a cost of\n  // $1.8e-05 for 429 input tokens.\n  //\n  // The model id is \"jev-latest\", NOT \"typesafe/jev-latest\" — the\n  // namespaced form 400s with \"does not exist\", and the endpoint accepts\n  // no generative models at all. It does not appear in OpenRouter's public\n  // model list either, so the id cannot be discovered from /v1/models.\n  registerDecisionBackend(\"jev\", () =>\n    createSimpleJevBackend({\n      name: \"jev\",\n      baseUrl: \"https://openrouter.ai/api/alpha\",\n      path: \"/decisions\",\n      model: \"jev-latest\",\n      apiKeyEnv: \"OPENROUTER_API_KEY\",\n      apiKeyFile: \"~/.agentx/openrouter-key.txt\",\n      // TypeSafe's claim is a model trained for calibrated decisions. The\n      // claim is what this records; whether it holds on our traffic is what\n      // `agentx decisions recalibrate` is for, and calibratedProbabilities\n      // stays false until it is measured.\n      probabilitySource: \"native\",\n      maxChoiceOptions: 255,\n      // Jev's request budget is ~32k TOKENS. 24k characters is roughly 6k\n      // tokens — a quarter of what it accepts, and it was my conservative\n      // carry-over from the simple-jev demo's 2k context. It silently\n      // truncated re-ranking shortlists at the tail.\n      maxStateChars: 90_000,\n      ...jev,\n    }),\n  )\n  // Jev direct from TypeSafe, bypassing OpenRouter. Same System One\n  // contract on /v1/systemone; the difference is billing and that it needs\n  // a TypeSafe key rather than an OpenRouter one. Prefer this once you have\n  // direct access: one less hop, and the vendor's own rate limits.\n  registerDecisionBackend(\"typesafe\", () =>\n    createSimpleJevBackend({\n      name: \"typesafe\",\n      baseUrl: \"https://api.typesafe.ai/v1\",\n      path: \"/systemone\",\n      model: \"jev-latest\",\n      apiKeyEnv: \"TYPESAFE_API_KEY\",\n      apiKeyFile: \"~/.agentx/typesafe-key.txt\",\n      probabilitySource: \"native\",\n      maxChoiceOptions: 255,\n      maxStateChars: 90_000,\n      ...typesafe,\n    }),\n  )\n}\n","import { getDecisionBackend } from \"./backend\"\nimport { DecisionStore, type DecisionAction, type SeatMode } from \"./store\"\nimport type { AnswersFor, Questions, StateValue } from \"./types\"\n\n// askSeat — the only entry point a call site should use.\n//\n// Three properties it must have, in priority order:\n//\n//   1. It never throws. A seat is an addition to a call site that already\n//      works; a decision backend having a bad day must not take down the\n//      dispatch path it was bolted onto. Every failure returns null and the\n//      caller keeps its existing behaviour.\n//\n//   2. Every failure is still a row. A seat that records only its successes\n//      flatters itself exactly where it matters — a backend that times out\n//      on the hard half of the traffic would otherwise show beautiful\n//      calibration on the easy half.\n//\n//   3. Default off. A seat does nothing until an operator names it in\n//      config or the environment, and `shadow` never changes behaviour —\n//      it only records, so the incumbent stays authoritative through the\n//      whole soak.\n\nexport interface SeatSettings {\n  mode: SeatMode\n  backend?: string\n  model?: string\n  timeoutMs?: number\n  /** Post-hoc temperature from a fit on this seat's own labeled rows.\n   *  1 means \"not calibrated yet\", which is where every seat starts. */\n  temperature?: number\n  /** Stop persisting state once this many rows exist for the seat. */\n  keepStateRows?: number\n  redactState?: boolean\n  /** Fraction of would-be-skips to run anyway, keeping an unbiased sample\n   *  of the skip region. See DEFAULT_EXPLORE_RATE — this is a correctness\n   *  requirement, not a tuning knob. */\n  explore?: number\n}\n\nexport interface DecisionsRuntime {\n  enabled: boolean\n  defaultBackend: string\n  seats: Record<string, SeatSettings>\n  store: DecisionStore | null\n}\n\nconst DEFAULT_RUNTIME: DecisionsRuntime = {\n  enabled: false,\n  defaultBackend: \"local\",\n  seats: {},\n  store: null,\n}\n\nlet runtime: DecisionsRuntime = { ...DEFAULT_RUNTIME }\nlet configured = false\n\n/** Called by the daemon after config load, and by the lazy path below for\n *  every other process. */\nexport function configureDecisions(next: Partial<DecisionsRuntime>): void {\n  runtime = { ...DEFAULT_RUNTIME, ...runtime, ...next }\n  configured = true\n}\n\nexport function resetDecisionsRuntime(): void {\n  runtime = { ...DEFAULT_RUNTIME, seats: {} }\n  configured = false\n}\n\n/**\n * Configure from agentx.json on first use, for processes that are not the\n * daemon.\n *\n * Seats are not a daemon feature, but only the daemon was calling\n * configureDecisions — so `agentx wiki search` and the wiki server ran\n * every seat as \"off\" and silently fell back, with no error and no\n * recorded row. The failure was invisible precisely because the seat is\n * designed to fail quietly.\n *\n * Best-effort and cached, including on failure: a CLI run outside a\n * project has no config, and that must cost one failed read rather than\n * one per call.\n */\nfunction ensureConfigured(): void {\n  if (configured) return\n  configured = true\n  try {\n    // Required lazily: importing the daemon config from module scope would\n    // pull the config graph into every process that touches a seat.\n    const { loadDaemonConfig } = require(\"@/daemon/config\") as {\n      loadDaemonConfig: () => { decisions?: any }\n    }\n    const cfg = loadDaemonConfig()?.decisions\n    if (!cfg?.enabled) return\n\n    const { registerBuiltinDecisionBackends } = require(\"./index\") as {\n      registerBuiltinDecisionBackends: (...a: any[]) => void\n    }\n    const b = cfg.backends ?? {}\n    registerBuiltinDecisionBackends(b.local, b.simpleJev, b.jev, b.typesafe)\n\n    let store: DecisionStore | null = null\n    try {\n      store = new DecisionStore({ path: cfg.dbPath })\n    } catch {\n      /* answer without recording rather than not answer */\n    }\n    runtime = {\n      enabled: true,\n      defaultBackend: cfg.defaultBackend,\n      store,\n      seats: Object.fromEntries(\n        Object.entries(cfg.seats ?? {}).map(([name, seat]: [string, any]) => [\n          name,\n          { ...seat, redactState: cfg.redactState, keepStateRows: cfg.keepStateRows },\n        ]),\n      ),\n    }\n  } catch {\n    /* no config, no seats — exactly today's behaviour */\n  }\n}\n\nexport function decisionsRuntime(): DecisionsRuntime {\n  return runtime\n}\n\nconst VALID_MODES: readonly SeatMode[] = [\"off\", \"shadow\", \"active\"]\n\nexport function parseSeatMode(value: unknown): SeatMode | null {\n  if (typeof value !== \"string\") return null\n  const normalized = value.trim().toLowerCase()\n  return (VALID_MODES as readonly string[]).includes(normalized)\n    ? (normalized as SeatMode)\n    : null\n}\n\n/** `monitor-prefilter` -> AGENTX_DECISION_SEAT_MONITOR_PREFILTER */\nexport function seatEnvVar(seat: string): string {\n  return `AGENTX_DECISION_SEAT_${seat.toUpperCase().replace(/[^A-Z0-9]+/g, \"_\")}`\n}\n\n/**\n * Resolve a seat's mode. Order: per-seat env var, then config, then \"off\".\n *\n * An invalid value at any level falls through to the next rather than\n * failing loudly — same trade as src/intent/mode.ts. The cost is silence on\n * a typo; the cost of the alternative is a seat accidentally promoted to\n * `active` because someone wrote \"Active \" with a trailing space.\n *\n * Read every call, never cached, so an operator can flip a seat off without\n * restarting the daemon.\n */\nexport function getSeatMode(seat: string, env: NodeJS.ProcessEnv = process.env): SeatMode {\n  // Configure FIRST, even when the env var will decide the answer.\n  //\n  // This used to return on the env override before touching the config,\n  // which meant the one documented way to switch a seat on for a single\n  // process — AGENTX_DECISION_SEAT_<SEAT>=active — also skipped the lazy\n  // setup that REGISTERS THE BACKENDS. The seat then reported `active` and\n  // every call failed with `unknown decision backend \"local\"`, so turning\n  // a seat on by hand quietly guaranteed it could never answer.\n  //\n  // It went unnoticed because the commands built around seats\n  // (`agentx decide`, `agentx decisions`) call registerBuiltinDecisionBackends\n  // themselves, so the path that was broken was the one an operator or a\n  // new seat would take.\n  ensureConfigured()\n  const fromEnv = parseSeatMode(env[seatEnvVar(seat)])\n  if (fromEnv) return fromEnv\n  if (!runtime.enabled) return \"off\"\n  return parseSeatMode(runtime.seats[seat]?.mode) ?? \"off\"\n}\n\nexport interface AskSeatOptions<Q extends Questions> {\n  /** What the code would do without the seat. Recorded alongside, which is\n   *  what makes agreement and the eventual promotion decision measurable. */\n  incumbent?: Partial<Record<keyof Q & string, string | number>>\n  links?: Array<{ kind: string; id: string }>\n  /** Covariates for recalibration — agent, channel, anything categorical\n   *  that plausibly shifts how well calibrated this seat is. */\n  features?: Record<string, string | number | boolean | null>\n  signal?: AbortSignal\n  timeoutMs?: number\n  backend?: string\n  model?: string\n}\n\nexport interface SeatResult<Q extends Questions> {\n  answers: AnswersFor<Q>\n  callId: string | null\n  mode: SeatMode\n}\n\n/** Record what the caller did with an answer.\n *\n *  Separate from askSeat because the seat owns the answer and the CALLER\n *  owns the policy: askSeat cannot know whether its caller acted. Silent\n *  no-op when there is no store or no call id, like every other recording\n *  path here — observability never breaks the thing it observes. */\nexport function recordSeatOutcome(\n  callId: string | null,\n  action: DecisionAction,\n  explored = false,\n): void {\n  if (!callId || !runtime.store) return\n  try {\n    runtime.store.recordOutcome(callId, action, explored)\n  } catch {\n    /* observability never breaks the caller */\n  }\n}\n\nlet lastWarnAt = 0\n\nexport async function askSeat<Q extends Questions>(\n  seat: string,\n  state: StateValue,\n  questions: Q,\n  opts: AskSeatOptions<Q> = {},\n): Promise<SeatResult<Q> | null> {\n  const mode = getSeatMode(seat)\n  if (mode === \"off\") return null\n\n  const settings = runtime.seats[seat] ?? { mode }\n  const backendName = opts.backend ?? settings.backend ?? runtime.defaultBackend\n  const started = Date.now()\n\n  try {\n    const backend = getDecisionBackend(backendName)\n    const response = await backend.decide({\n      state,\n      questions,\n      model: opts.model ?? settings.model,\n      abortSignal: opts.signal,\n      timeoutMs: opts.timeoutMs ?? settings.timeoutMs,\n    })\n\n    const callId = record(seat, mode, state, questions, response, opts, settings)\n    return { answers: response.answers, callId, mode }\n  } catch (err: any) {\n    recordFailure(seat, mode, state, questions, backendName, opts, settings, started, err)\n    warnThrottled(seat, backendName, err)\n    return null\n  }\n}\n\nfunction record<Q extends Questions>(\n  seat: string,\n  mode: SeatMode,\n  state: StateValue,\n  questions: Q,\n  response: Awaited<ReturnType<ReturnType<typeof getDecisionBackend>[\"decide\"]>>,\n  opts: AskSeatOptions<Q>,\n  settings: SeatSettings,\n): string | null {\n  if (!runtime.store) return null\n  try {\n    return runtime.store.recordCall({\n      seat,\n      mode,\n      backend: response.meta.backend,\n      model: response.model,\n      meta: response.meta,\n      state,\n      questions,\n      answers: response.answers as Record<string, any>,\n      usage: response.usage,\n      incumbent: toIncumbent(opts.incumbent),\n      links: opts.links,\n      features: opts.features,\n      keepState: !settings.redactState,\n    })\n  } catch {\n    // Recording is observability. It never breaks the call it observes.\n    return null\n  }\n}\n\nfunction recordFailure<Q extends Questions>(\n  seat: string,\n  mode: SeatMode,\n  state: StateValue,\n  questions: Q,\n  backend: string,\n  opts: AskSeatOptions<Q>,\n  settings: SeatSettings,\n  started: number,\n  err: any,\n): void {\n  if (!runtime.store) return\n  try {\n    runtime.store.recordCall({\n      seat,\n      mode,\n      backend,\n      model: opts.model ?? settings.model ?? \"unknown\",\n      meta: {\n        structureMode: \"text\",\n        answerMode: \"probabilities\",\n        retries: 0,\n        stateTruncated: false,\n        latencyMs: Date.now() - started,\n      },\n      state,\n      questions,\n      answers: {},\n      incumbent: toIncumbent(opts.incumbent),\n      links: opts.links,\n      features: opts.features,\n      error: String(err?.message ?? err).slice(0, 500),\n      keepState: !settings.redactState,\n    })\n  } catch {\n    /* observability never breaks the caller */\n  }\n}\n\nfunction toIncumbent(\n  incumbent: Record<string, string | number | undefined> | undefined,\n): Record<string, { value: string | number }> | undefined {\n  if (!incumbent) return undefined\n  const out: Record<string, { value: string | number }> = {}\n  for (const [question, value] of Object.entries(incumbent)) {\n    if (value !== undefined) out[question] = { value }\n  }\n  return out\n}\n\n/** At most one line a minute. A backend that is down fails on every call,\n *  and a log line per call would bury the thing that actually broke. */\nfunction warnThrottled(seat: string, backend: string, err: any): void {\n  const now = Date.now()\n  if (now - lastWarnAt < 60_000) return\n  lastWarnAt = now\n  console.warn(\n    `[decisions] seat \"${seat}\" on backend \"${backend}\" failed: ${err?.message ?? err}`,\n  )\n}\n"],"mappings":"2cA8GO,SAASA,EAAwBC,EAAcC,EAA+B,CACnFC,EAAU,IAAIF,EAAMC,CAAO,EAC3BE,EAAU,OAAOH,CAAI,CACvB,CAEO,SAASI,EAAmBJ,EAA+B,CAChE,IAAMK,EAASF,EAAU,IAAIH,CAAI,EACjC,GAAIK,EAAQ,OAAOA,EACnB,IAAMJ,EAAUC,EAAU,IAAIF,CAAI,EAClC,GAAI,CAACC,EAAS,CACZ,IAAMK,EAAQC,GAAqB,EACnC,MAAM,IAAI,MACR,6BAA6BP,MAC1BM,EAAM,OAAS,EAAI,uBAAkBA,EAAM,KAAK,IAAI,IAAM,0BAC/D,EAEF,IAAME,EAAUP,EAAQ,EACxB,OAAAE,EAAU,IAAIH,EAAMQ,CAAO,EACpBA,CACT,CAEO,SAASD,IAAiC,CAC/C,MAAO,CAAC,GAAGL,EAAU,KAAK,CAAC,EAAE,KAAK,CACpC,CAEO,SAASO,IAAyC,CACvDP,EAAU,MAAM,EAChBC,EAAU,MAAM,CAClB,CA1IA,IAwGMD,EACAC,EAzGNO,EAAAC,EAAA,kBAwGMT,EAAY,IAAI,IAChBC,EAAY,IAAI,MCzGtB,OAAOS,OAAc,iBACrB,OAAS,cAAAC,OAAkB,SAC3B,OAAS,aAAAC,OAAiB,KAC1B,OAAS,WAAAC,GAAS,WAAAC,OAAe,OA4V1B,SAASC,EAAWC,EASzB,CACA,GAAIA,EAAO,OAAS,OAAQ,CAC1B,IAAMC,EAAID,EAAO,KACXE,EAAgB,CAAE,IAAKD,EAAG,GAAI,EAAIA,CAAE,EACpCE,EAAYF,GAAK,GAAM,MAAQ,KAG/BG,EAAa,KAAK,IAAIH,EAAI,EAAG,EAAI,EACvC,MAAO,CACL,UAAAE,EACA,WAAAC,EACA,cAAAF,EACA,SAAUC,EACV,QAAS,KAAK,IAAIF,EAAG,EAAIA,CAAC,EAC1B,cAAe,KACf,KAAM,KAAK,IAAIA,EAAG,EAAIA,CAAC,EACvB,WAAYG,CACd,EAGF,GAAIJ,EAAO,OAAS,SAClB,MAAO,CACL,UAAWA,EAAO,OAClB,WAAYA,EAAO,WACnB,cAAe,CAAE,GAAGA,EAAO,aAAc,EACzC,SAAUA,EAAO,OACjB,QAASA,EAAO,cAAcA,EAAO,MAAM,GAAK,EAChD,cAAe,KACf,KAAMA,EAAO,KACb,WAAYA,EAAO,UACrB,EAGF,IAAMK,EAAU,OAAO,KAAK,MAAML,EAAO,KAAK,CAAC,EAC/C,MAAO,CACL,UAAWK,EACX,WAAYL,EAAO,WACnB,cAAe,CAAE,GAAGA,EAAO,aAAc,EACzC,SAAUK,EACV,QAASL,EAAO,cAAcK,CAAO,GAAK,EAC1C,cAAeL,EAAO,MACtB,KAAMA,EAAO,KACb,WAAYA,EAAO,UACrB,CACF,CAEA,SAASM,GAAcC,EAAgE,CACrF,GAAI,OAAOA,GAAQ,UAAYA,EAAI,SAAW,EAAG,MAAO,CAAC,EACzD,GAAI,CACF,IAAMC,EAAS,KAAK,MAAMD,CAAG,EAC7B,OAAOC,GAAU,OAAOA,GAAW,UAAY,CAAC,MAAM,QAAQA,CAAM,EAAIA,EAAS,CAAC,CACpF,MAAE,CACA,MAAO,CAAC,CACV,CACF,CAEA,SAASC,GAAcC,EAA6B,CAClDA,EAAG,KAAK,oEAAoE,EAC5E,IAAMC,EACHD,EAAG,QAAQ,wCAAwC,EAAE,IAAI,EAA2B,GAAK,EACxFC,EAAU,GAAGC,GAAYF,CAAE,EAC3BC,EAAU,GAAGE,GAAYH,CAAE,EAC3BC,EAAU,GAAGG,GAAYJ,CAAE,CACjC,CAUA,SAASI,GAAYJ,EAA6B,CAChDA,EAAG,KAAK;AAAA;AAAA;AAAA,GAGP,CACH,CAWA,SAASG,GAAYH,EAA6B,CAChDA,EAAG,KAAK;AAAA;AAAA;AAAA;AAAA;AAAA,GAKP,CACH,CAEA,SAASE,GAAYF,EAA6B,CAChDA,EAAG,KAAK;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,GAmEP,CACH,CA/gBA,IAoHaK,EApHbC,EAAAC,EAAA,kBAIAC,KAgHaH,EAAN,KAAoB,CAChB,GACA,KAET,YAAYI,EAAyB,CAAC,EAAG,CACvC,KAAK,KAAOrB,GAAQ,QAAQ,IAAI,EAAGqB,EAAK,MAAQ,oCAAoC,EAC/EA,EAAK,UAAUvB,GAAUC,GAAQ,KAAK,IAAI,EAAG,CAAE,UAAW,EAAK,CAAC,EACrE,KAAK,GAAK,IAAIH,GAAS,KAAK,KAAM,CAAE,SAAUyB,EAAK,UAAY,EAAM,CAAC,EACjEA,EAAK,WACR,KAAK,GAAG,OAAO,oBAAoB,EACnC,KAAK,GAAG,OAAO,sBAAsB,EACrC,KAAK,GAAG,OAAO,mBAAmB,EAClCV,GAAc,KAAK,EAAE,EAEzB,CAEA,OAAc,CACZ,KAAK,GAAG,MAAM,CAChB,CAEA,eAAwB,CAItB,OAHY,KAAK,GAAG,QAAQ,wCAAwC,EAAE,IAAI,EAG/D,GAAK,CAClB,CAIA,WAAWW,EAAgC,CACzC,IAAMC,EAAKD,EAAM,IAAM,KAAK,IAAI,EAC1BE,EAASC,GAAWF,CAAE,EACtBG,EAAY,KAAK,UAAUJ,EAAM,OAAS,IAAI,EAC9CK,EAAY9B,GAAW,QAAQ,EAAE,OAAO6B,CAAS,EAAE,OAAO,KAAK,EA6ErE,OA3EW,KAAK,GAAG,YAAY,IAAM,CACnC,KAAK,GACF,QACC;AAAA;AAAA;AAAA;AAAA,yEAKF,EACC,IACCF,EACAD,EACAD,EAAM,KACNA,EAAM,KACNA,EAAM,QACNA,EAAM,MACNA,EAAM,KAAK,cACXA,EAAM,KAAK,WACXK,EACAL,EAAM,YAAc,GAAQ,KAAOI,EACnC,KAAK,UAAUJ,EAAM,SAAS,EAC9BA,EAAM,KAAK,UACXA,EAAM,OAAO,aAAe,EAC5BA,EAAM,OAAO,cAAgB,EAC7BA,EAAM,KAAK,QACXA,EAAM,KAAK,eAAiB,EAAI,EAChCA,EAAM,OAAS,KACfA,EAAM,SAAW,KAAK,UAAUA,EAAM,QAAQ,EAAI,IACpD,EAEF,IAAMM,EAAe,KAAK,GAAG,QAC3B;AAAA;AAAA;AAAA,4CAIF,EACA,OAAW,CAACC,EAAU3B,CAAM,IAAK,OAAO,QAAQoB,EAAM,OAAO,EAAG,CAC9D,IAAMQ,EAAO7B,EAAWC,CAAM,EAC9B0B,EAAa,IACXJ,EACAK,EACA3B,EAAO,KACP,KAAK,UAAUA,CAAM,EACrB4B,EAAK,SACLA,EAAK,QACLA,EAAK,cACLA,EAAK,KACLA,EAAK,UACP,EAGF,GAAIR,EAAM,UAAW,CACnB,IAAMS,EAAkB,KAAK,GAAG,QAC9B;AAAA,kCAEF,EACA,OAAW,CAACF,EAAUG,CAAG,IAAK,OAAO,QAAQV,EAAM,SAAS,EAC1DS,EAAgB,IACdP,EACAK,EACAG,EAAI,QAAU,OAAY,KAAO,OAAOA,EAAI,KAAK,EACjDA,EAAI,OAAS,KACbA,EAAI,QAAU,IAChB,EAIJ,GAAIV,EAAM,MAAO,CACf,IAAMW,EAAa,KAAK,GAAG,QACzB,mFACF,EACA,QAAWC,KAAQZ,EAAM,MAAOW,EAAW,IAAIT,EAAQU,EAAK,KAAMA,EAAK,EAAE,EAE7E,CAAC,EAEE,EACIV,CACT,CAIA,MACEA,EACAK,EACAM,EACAd,EAA6E,CAAC,EACtE,CACR,IAAME,EAAKF,EAAK,IAAM,KAAK,IAAI,EACzBe,EAAKX,GAAWF,CAAE,EACxB,YAAK,GACF,QACC;AAAA,yCAEF,EACC,IACCa,EACAZ,EACAK,EACA,OAAOM,CAAK,EACZd,EAAK,MAAQ,QACbE,EACAF,EAAK,WAAa,KAClBA,EAAK,MAAQ,IACf,EACKe,CACT,CAKA,cAAcZ,EAAgBa,EAAwBC,EAAW,GAAa,CAC5E,KAAK,GACF,QAAQ,iEAAiE,EACzE,IAAID,EAAQC,EAAW,EAAI,EAAGd,CAAM,CACzC,CAEA,gBAAgBe,EAAcH,EAAsB,CAIlD,OAHa,KAAK,GACf,QAAQ,sEAAsE,EAC9E,IAAIG,EAAMH,CAAE,EACH,IAAKI,GAAMA,EAAE,OAAO,CAClC,CAEA,WAAWC,EAA0B,CAAC,EAAgB,CACpD,IAAMC,EAAkB,CAAC,iBAAiB,EACpCC,EAAoB,CAAC,EACvBF,EAAO,OAAOC,EAAM,KAAK,YAAY,EAAGC,EAAO,KAAKF,EAAO,IAAI,GAC/DA,EAAO,WAAWC,EAAM,KAAK,gBAAgB,EAAGC,EAAO,KAAKF,EAAO,QAAQ,GAC3EA,EAAO,UAAUC,EAAM,KAAK,eAAe,EAAGC,EAAO,KAAKF,EAAO,OAAO,GACxEA,EAAO,QAAQC,EAAM,KAAK,aAAa,EAAGC,EAAO,KAAKF,EAAO,KAAK,GAClEA,EAAO,gBAAgBC,EAAM,KAAK,sBAAsB,EAAGC,EAAO,KAAKF,EAAO,aAAa,GAC3FA,EAAO,QAAU,SAAYC,EAAM,KAAK,WAAW,EAAGC,EAAO,KAAKF,EAAO,KAAK,GAC9EA,EAAO,SAASC,EAAM,KAAK,cAAc,EAAGC,EAAO,KAAKF,EAAO,MAAM,GACrEA,EAAO,cAAcC,EAAM,KAAK,gBAAgB,EAIpD,IAAME,EAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,eAWDF,EAAM,KAAK,OAAO;AAAA;AAAA,SAExBD,EAAO,MAAQ,UAAY,KAChC,OAAIA,EAAO,OAAOE,EAAO,KAAKF,EAAO,KAAK,EAE7B,KAAK,GAAG,QAAQG,CAAG,EAAE,IAAI,GAAGD,CAAM,EAE5C,IAAKH,GAAM,CACV,IAAMtC,EAAS,KAAK,MAAMsC,EAAE,WAAW,EACjCV,EAAO7B,EAAWC,CAAM,EAC9B,MAAO,CACL,OAAQsC,EAAE,QACV,KAAMA,EAAE,KACR,GAAIA,EAAE,GACN,QAASA,EAAE,QACX,MAAOA,EAAE,MACT,cAAeA,EAAE,eACjB,SAAUA,EAAE,SACZ,KAAMtC,EAAO,KACb,UAAW4B,EAAK,UAChB,WAAYA,EAAK,WACjB,cAAeA,EAAK,cACpB,UAAWU,EAAE,WAAa,OAC1B,MAAOA,EAAE,OAAS,OAClB,SAAUhC,GAAcgC,EAAE,aAAa,EACvC,KAAMA,EAAE,KACR,OAASA,EAAE,QAAU,OACrB,SAAUA,EAAE,WAAa,CAC3B,CACF,CAAC,EACA,OAAQR,GAAQ,CAACS,EAAO,aAAeT,EAAI,QAAU,MAAS,CACnE,CAIA,WAAWa,EAAcC,EAA0B,CAUjD,OATY,KAAK,GACd,QACC;AAAA;AAAA;AAAA;AAAA,cAKF,EACC,IAAID,EAAMA,EAAMC,CAAQ,EAChB,OACb,CACF,IC1VA,IAAAC,GAAAC,EAAA,oBCgCO,SAASC,GACdC,EACAC,EACAC,EAAyB,CAAC,EACF,CACxB,IAAMC,EAAYD,EAAK,WAAa,GAC9BE,EAAUF,EAAK,SAAW,KAC1BG,EACJL,IAAQ,MAAQ,OAAOA,GAAQ,UAAY,CAAC,MAAM,QAAQA,CAAG,EACxDA,EACD,CAAC,EAEHM,EAAWN,IAAQ,MAAQ,OAAOA,GAAQ,UAAY,MAAM,QAAQA,CAAG,EAC3E,QAAWO,KAAO,OAAO,KAAKF,CAAM,EAC7BJ,EAAO,SAASM,CAAG,IAAGD,EAAW,IAGxC,IAAME,EAAgC,CAAC,EACnCC,EAAM,EACV,QAAWC,KAAST,EAAQ,CAC1B,GAAI,EAAES,KAASL,GAAS,CACtBG,EAAME,CAAK,EAAIN,EACfK,GAAOL,EACPE,EAAW,GACX,SAEF,IAAMK,EAAQ,OAAON,EAAOK,CAAK,CAAC,EAClC,GAAI,CAAC,OAAO,SAASC,CAAK,GAAKA,EAAQ,EAAG,CACxCH,EAAME,CAAK,EAAI,EACfJ,EAAW,GACX,SAEFE,EAAME,CAAK,EAAIC,EACfF,GAAOE,EAGT,GAAIV,EAAO,SAAW,EAAG,MAAO,CAAE,MAAAO,EAAO,SAAAF,CAAS,EAElD,GAAIG,GAAO,EAAG,CAGZ,IAAMG,EAAU,EAAIX,EAAO,OAC3B,QAAWS,KAAST,EAAQO,EAAME,CAAK,EAAIE,EAC3C,MAAO,CAAE,MAAAJ,EAAO,SAAU,EAAK,EAGjC,GAAIL,EAAW,CACb,QAAWO,KAAST,EAAQO,EAAME,CAAK,EAAIF,EAAME,CAAK,EAAID,EACtD,KAAK,IAAIA,EAAM,CAAC,EAAI,OAAMH,EAAW,IAG3C,MAAO,CAAE,MAAAE,EAAO,SAAAF,CAAS,CAC3B,CAIO,SAASO,GAAKL,EAAuC,CAC1D,IAAMM,EAAS,OAAO,OAAON,CAAK,EAClC,OAAIM,EAAO,SAAW,EAAU,EACzB,KAAK,IAAI,GAAGA,CAAM,CAC3B,CAKO,SAASC,GAAqBP,EAAuC,CAC1E,IAAMM,EAAS,OAAO,OAAON,CAAK,EAAE,OAAQQ,GAAM,OAAO,SAASA,CAAC,GAAKA,EAAI,CAAC,EACvEC,EAAI,OAAO,KAAKT,CAAK,EAAE,OAC7B,GAAIS,GAAK,EAAG,MAAO,GACnB,IAAMC,EAAQJ,EAAO,OAAO,CAACK,EAAGC,IAAMD,EAAIC,EAAG,CAAC,EAC9C,GAAIF,GAAS,EAAG,MAAO,GACvB,IAAIG,EAAI,EACR,QAAWL,KAAKF,EAAQ,CACtB,IAAMQ,EAAIN,EAAIE,EACdG,GAAKC,EAAI,KAAK,IAAIA,CAAC,EAErB,OAAOC,GAAQ,EAAIF,EAAI,KAAK,IAAIJ,CAAC,CAAC,CACpC,CAGO,SAASO,GAAchB,EAAuC,CACnE,IAAIU,EAAQ,EACRO,EAAW,EACf,OAAW,CAAClB,EAAKI,CAAK,IAAK,OAAO,QAAQH,CAAK,EAAG,CAChD,IAAMkB,EAAQ,OAAOnB,CAAG,EACpB,CAAC,OAAO,SAASmB,CAAK,GAAK,CAAC,OAAO,SAASf,CAAK,GAAKA,GAAS,IACnEO,GAASP,EACTc,GAAYC,EAAQf,GAEtB,OAAOO,EAAQ,EAAIO,EAAWP,EAAQ,CACxC,CAEO,SAASS,GAAcnB,EAAuC,CACnE,OAAOO,GAAqBP,CAAK,CACnC,CAEA,SAASe,GAAQK,EAAmB,CAClC,OAAK,OAAO,SAASA,CAAC,EACf,KAAK,IAAI,EAAG,KAAK,IAAI,EAAGA,CAAC,CAAC,EADD,CAElC,CAKO,SAASC,EACdC,EACA9B,EACAE,EAAyB,CAAC,EACgB,CAC1C,GAAI4B,EAAS,OAAS,OAAQ,CAC5B,IAAMnB,EAAQ,OAAQX,EAA2B,IAAI,EAC/C+B,EAAK,OAAO,SAASpB,CAAK,GAAKA,GAAS,GAAKA,GAAS,EAC5D,MAAO,CACL,OAAQ,CAAE,KAAM,OAAQ,KAAMoB,EAAKpB,EAAQ,EAAI,EAC/C,SAAU,CAACoB,CACb,EAGF,IAAM9B,EAAS+B,EAASF,CAAQ,EAC1B,CAAE,MAAAtB,EAAO,SAAAF,CAAS,EAAIP,GACzBC,EAAoC,cACrCC,EACAC,CACF,EACM+B,EAAMpB,GAAKL,CAAK,EAChB0B,EAAanB,GAAqBP,CAAK,EAE7C,GAAIsB,EAAS,OAAS,SAAU,CAC9B,IAAIK,EAAOlC,EAAO,CAAC,EACnB,QAAWS,KAAST,EAAYO,EAAME,CAAK,EAAIF,EAAM2B,CAAI,IAAGA,EAAOzB,GACnE,MAAO,CACL,OAAQ,CACN,KAAM,SACN,OAAQyB,EACR,WAAYR,GAAcnB,CAAK,EAC/B,cAAeA,EACf,KAAMyB,EACN,WAAAC,CACF,EACA,SAAA5B,CACF,EAGF,IAAM8B,EAAiC,CAAC,EACxC,OAAAN,EAAS,SAAS,QAAQ,CAACO,EAAaC,IAAM,CAC5CF,EAAO,OAAOE,CAAC,CAAC,EAAID,CACtB,CAAC,EACM,CACL,OAAQ,CACN,KAAM,QACN,MAAOb,GAAchB,CAAK,EAC1B,WAAYmB,GAAcnB,CAAK,EAC/B,OAAA4B,EACA,cAAe5B,EACf,KAAMyB,EACN,WAAAC,CACF,EACA,SAAA5B,CACF,CACF,CA/LA,IAAAiC,EAAAC,EAAA,kBAAAC,MCAA,OAAS,KAAAC,MAAS,MA2BX,SAASC,GAAgBC,EAA2C,CACzE,OAAOA,EAAS,OAAS,OAASC,GAAaC,EACjD,CAKO,SAASC,EAAiBC,EAA0C,CACzE,IAAMC,EAA4C,CAAC,EACnD,OAAW,CAACC,EAAMN,CAAQ,IAAK,OAAO,QAAQI,CAAS,EACrDC,EAAMC,CAAI,EAAIP,GAAgBC,CAAQ,EAExC,OAAOF,EAAE,OAAOO,CAAK,CACvB,CAKO,SAASE,EAAeC,EAA2B,CACxD,OAAOA,EAAM,OACV,IAAKC,GAEG,KADMA,EAAM,KAAK,OAAS,EAAIA,EAAM,KAAK,KAAK,GAAG,EAAI,aACvCA,EAAM,SAC5B,EACA,KAAK;AAAA,CAAI,CACd,CApDA,IAmBMR,GAIAC,GAvBNQ,EAAAC,EAAA,kBAmBMV,GAAaH,EAAE,OAAO,CAC1B,KAAMA,EAAE,OAAO,CACjB,CAAC,EAEKI,GAAqBJ,EAAE,OAAO,CAClC,cAAeA,EAAE,OAAOA,EAAE,OAAO,EAAGA,EAAE,QAAQ,CAAC,CACjD,CAAC,ICSM,SAASc,GAAgBC,EAA+C,CAC7E,IAAMC,EAAsC,CAAC,EAC7C,OAAW,CAACC,EAAMC,CAAQ,IAAK,OAAO,QAAQH,CAAS,EACrDC,EAAWC,CAAI,EAAIE,GAAmBD,CAAQ,EAEhD,MAAO,CACL,KAAM,SACN,qBAAsB,GACtB,SAAU,CAAC,SAAS,EACpB,WAAY,CACV,QAAS,CACP,KAAM,SACN,qBAAsB,GACtB,SAAU,OAAO,KAAKH,CAAS,EAC/B,WAAAC,CACF,CACF,CACF,CACF,CAEA,SAASG,GAAmBD,EAAgD,CAC1E,GAAIA,EAAS,OAAS,OACpB,MAAO,CACL,KAAM,SACN,qBAAsB,GACtB,SAAU,CAAC,MAAM,EACjB,WAAY,CACV,KAAM,CACJ,KAAM,SACN,QAAS,EACT,QAAS,EACT,YAAa,gCACf,CACF,CACF,EAGF,IAAME,EAASC,EAASH,CAAQ,EAC1BF,EAAsC,CAAC,EAC7C,QAAWM,KAASF,EAClBJ,EAAWM,CAAK,EAAI,CAAE,KAAM,SAAU,QAAS,EAAG,QAAS,CAAE,EAE/D,MAAO,CACL,KAAM,SACN,qBAAsB,GACtB,SAAU,CAAC,eAAe,EAC1B,WAAY,CACV,cAAe,CACb,KAAM,SACN,qBAAsB,GACtB,SAAUF,EACV,WAAAJ,EACA,YAAa,4CACf,CACF,CACF,CACF,CAUO,SAASO,EAAYC,EAAmBC,EAAiC,CAC9E,IAAMC,EAAO,OAAOF,GAAU,SAAWA,EAAQ,KAAK,UAAUA,EAAO,KAAM,CAAC,GAAK,OACnF,OAAIE,EAAK,QAAUD,EAAiB,CAAE,KAAAC,EAAM,UAAW,EAAM,EACtD,CACL,KAAM,GAAGA,EAAK,MAAM,EAAGD,CAAQ;AAAA,oBAAmBC,EAAK,OAASD,qBAChE,UAAW,EACb,CACF,CAIO,SAASE,GAAgBZ,EAA8B,CAC5D,IAAMa,EAAmB,CAAC,EAC1B,OAAW,CAACX,EAAMC,CAAQ,IAAK,OAAO,QAAQH,CAAS,EAAG,CACxD,IAAMc,EAAkB,CAAC,OAAOZ,GAAM,EAGtC,GAFIC,EAAS,cAAcW,EAAM,KAAKX,EAAS,YAAY,EAEvDA,EAAS,OAAS,OACpBW,EAAM,KAAK,uEAAkE,EACzEX,EAAS,UAAU,MAAMW,EAAM,KAAK,gBAAgBX,EAAS,SAAS,MAAM,EAC5EA,EAAS,UAAU,OAAOW,EAAM,KAAK,eAAeX,EAAS,SAAS,OAAO,UACxEA,EAAS,OAAS,SAAU,CACrCW,EAAM,KAAK,kEAAkE,EAC7E,OAAW,CAACP,EAAOQ,CAAW,IAAK,OAAO,QAAQZ,EAAS,QAAQ,EACjEW,EAAM,KAAKC,EAAc,KAAKR,MAAUQ,IAAgB,KAAKR,GAAO,OAGtEO,EAAM,KAAK,gEAAgE,EAC3EX,EAAS,SAAS,QAAQ,CAACY,EAAa,IAAM,CAC5CD,EAAM,KAAK,MAAM,OAAOC,GAAa,CACvC,CAAC,EAEHF,EAAO,KAAKC,EAAM,KAAK;AAAA,CAAI,CAAC,EAE9B,OAAOD,EAAO,KAAK;AAAA;AAAA,CAAM,CAC3B,CAIO,SAASG,GACdP,EACAT,EACAU,EACsC,CACtC,IAAMO,EAAWT,EAAYC,EAAOC,CAAQ,EAQ5C,MAAO,CAAE,KAPI,CACX,WACAO,EAAS,KACT,GACA,eACAL,GAAgBZ,CAAS,CAC3B,EAAE,KAAK;AAAA,CAAI,EACI,UAAWiB,EAAS,SAAU,CAC/C,CAGO,SAASC,GAAwBlB,EAA8B,CACpE,IAAMmB,EAAmC,CAAC,EAC1C,OAAW,CAACjB,EAAMC,CAAQ,IAAK,OAAO,QAAQH,CAAS,EAAG,CACxD,GAAIG,EAAS,OAAS,OAAQ,CAC5BgB,EAAQjB,CAAI,EAAI,CAAE,KAAM,EAAI,EAC5B,SAEF,IAAMkB,EAAwC,CAAC,EACzCf,EAASC,EAASH,CAAQ,EAChC,QAAWI,KAASF,EAAQe,EAAcb,CAAK,EAAI,QAAQ,EAAIF,EAAO,QAAQ,QAAQ,CAAC,CAAC,EACxFc,EAAQjB,CAAI,EAAI,CAAE,cAAAkB,CAAc,EAElC,MAAO,CACL,GACA,YACA,gDACA,KAAK,UAAU,CAAE,QAASD,CAAQ,EAAG,KAAM,CAAC,CAC9C,EAAE,KAAK;AAAA,CAAI,CACb,CA9KA,IAgBaE,EAEAC,EAlBbC,EAAAC,EAAA,kBAAAC,IAgBaJ,EAAY,mBAEZC,EAAgB,CAC3B,oGACA,GACA,SACA,oEACA,8FACA,kGACA,6DACA,0FACA,mEACA,wEACF,EAAE,KAAK;AAAA,CAAI,ICJX,SAASI,GAAQC,EAAwB,CACvC,OAAOA,EAAI,YAAcA,EAAI,MAAQ,EAAI,CAC3C,CAIO,SAASC,GAAMC,EAAmB,CACvC,IAAMC,EAAU,KAAK,IAAI,EAAIC,EAAK,KAAK,IAAIA,EAAKF,CAAC,CAAC,EAClD,OAAO,KAAK,IAAIC,GAAW,EAAIA,EAAQ,CACzC,CAEO,SAASE,GAAQC,EAAmB,CACzC,GAAIA,GAAK,EAAG,MAAO,IAAK,EAAI,KAAK,IAAI,CAACA,CAAC,GACvC,IAAMC,EAAI,KAAK,IAAID,CAAC,EACpB,OAAOC,GAAK,EAAIA,EAClB,CAmBA,SAASC,GAAeR,EAAgBS,EAAsB,CAC5D,IAAMC,EAAeV,EAAI,WAAWS,CAAI,EACxC,GAAkCC,GAAiB,KAAM,OAAO,OAAOA,CAAY,EACnF,IAAMC,EAAOX,EAA2CS,CAAI,EAC5D,OAA4BE,GAAQ,KAAO,SAAW,OAAOA,CAAG,CAClE,CAEO,SAASC,GAAYC,EAAmBC,EAA0B,CACvE,IAAMC,EAAWD,EAAK,eAAiB,EACjCE,EAAmC,CAAC,EAE1C,QAAWP,KAAQK,EAAK,WAAY,CAClC,IAAMG,EAAS,IAAI,IACnB,QAAWjB,KAAOa,EAAM,CACtB,IAAMK,EAAQV,GAAeR,EAAKS,CAAI,EACtCQ,EAAO,IAAIC,GAAQD,EAAO,IAAIC,CAAK,GAAK,GAAK,CAAC,EAEhD,IAAMC,EAAO,CAAC,GAAGF,EAAO,QAAQ,CAAC,EAC9B,OAAO,CAAC,CAAC,CAAEG,CAAC,IAAMA,GAAKL,CAAQ,EAC/B,IAAI,CAAC,CAACG,CAAK,IAAMA,CAAK,EACtB,KAAK,EAGRF,EAAOP,CAAI,EAAIU,EAAK,MAAM,CAAC,EAG7B,IAAME,EAAQ,CAAC,YAAa,iBAAiB,EAC7C,QAAWZ,KAAQK,EAAK,WAAY,QAAWQ,KAASN,EAAOP,CAAI,EAAGY,EAAM,KAAK,GAAGZ,KAAQa,GAAO,EAEnG,IAAMC,EAAgB,CAAC,EACjBC,EAAc,CAAC,EACrB,QAAWxB,KAAOa,EAAM,CACtB,IAAMY,EAAW,CAAC,EAAGxB,GAAMD,EAAI,UAAU,CAAC,EAC1C,QAAWS,KAAQK,EAAK,WAAY,CAClC,IAAMI,EAAQV,GAAeR,EAAKS,CAAI,EACtC,QAAWa,KAASN,EAAOP,CAAI,EAAGgB,EAAS,KAAKP,IAAUI,EAAQ,EAAI,CAAC,EAEzEC,EAAE,KAAKE,CAAQ,EACfD,EAAE,KAAKzB,GAAQC,CAAG,CAAC,EAGrB,MAAO,CAAE,EAAAuB,EAAG,EAAAC,EAAG,MAAAH,EAAO,OAAAL,CAAO,CAC/B,CAWO,SAASU,GACdH,EACAC,EACAG,EAA6C,CAAC,EACpC,CACV,IAAMC,EAAKD,EAAK,IAAM,EAChBE,EAAaF,EAAK,YAAc,GAChCP,EAAIG,EAAE,OACNO,EAAIV,EAAI,EAAIG,EAAE,CAAC,EAAE,OAAS,EAC5BQ,EAAI,IAAI,MAAMD,CAAC,EAAE,KAAK,CAAC,EAC3B,GAAIV,IAAM,GAAKU,IAAM,EAAG,OAAOC,EAE/B,QAASC,EAAO,EAAGA,EAAOH,EAAYG,IAAQ,CAE5C,IAAMC,EAAgB,MAAM,KAAK,CAAE,OAAQH,CAAE,EAAG,IAAM,IAAI,MAAMA,CAAC,EAAE,KAAK,CAAC,CAAC,EACpEI,EAAI,IAAI,MAAMJ,CAAC,EAAE,KAAK,CAAC,EAE7B,QAASK,EAAI,EAAGA,EAAIf,EAAGe,IAAK,CAC1B,IAAI7B,EAAI,EACR,QAAS8B,EAAI,EAAGA,EAAIN,EAAGM,IAAK9B,GAAKyB,EAAEK,CAAC,EAAIb,EAAEY,CAAC,EAAEC,CAAC,EAC9C,IAAMlC,EAAIG,GAAQC,CAAC,EACb+B,EAAS,KAAK,IAAInC,GAAK,EAAIA,GAAI,IAAI,EACnCoC,EAAWd,EAAEW,CAAC,EAAIjC,EACxB,QAASkC,EAAI,EAAGA,EAAIN,EAAGM,IAAK,CAC1BF,EAAEE,CAAC,GAAKE,EAAWf,EAAEY,CAAC,EAAEC,CAAC,EACzB,QAASG,EAAIH,EAAGG,EAAIT,EAAGS,IAAKN,EAAEG,CAAC,EAAEG,CAAC,GAAKF,EAASd,EAAEY,CAAC,EAAEC,CAAC,EAAIb,EAAEY,CAAC,EAAEI,CAAC,GAGpE,QAASH,EAAI,EAAGA,EAAIN,EAAGM,IAAK,CAEtBA,EAAI,IACNH,EAAEG,CAAC,EAAEA,CAAC,GAAKR,EACXM,EAAEE,CAAC,GAAKR,EAAKG,EAAEK,CAAC,GAElB,QAASG,EAAI,EAAGA,EAAIH,EAAGG,IAAKN,EAAEG,CAAC,EAAEG,CAAC,EAAIN,EAAEM,CAAC,EAAEH,CAAC,EAG9C,IAAMI,EAAOC,GAAMR,EAAGC,CAAC,EACvB,GAAI,CAACM,EAAM,MACX,IAAIE,EAAQ,EACZ,QAASN,EAAI,EAAGA,EAAIN,EAAGM,IACrBL,EAAEK,CAAC,GAAKI,EAAKJ,CAAC,EACdM,GAAS,KAAK,IAAIF,EAAKJ,CAAC,CAAC,EAE3B,GAAIM,EAAQ,KAAM,MAEpB,OAAOX,CACT,CAIA,SAASU,GAAME,EAAeC,EAA8B,CAC1D,IAAMxB,EAAIwB,EAAE,OACNC,EAAIF,EAAE,IAAI,CAAC3C,EAAK,IAAM,CAAC,GAAGA,EAAK4C,EAAE,CAAC,CAAC,CAAC,EAE1C,QAASE,EAAM,EAAGA,EAAM1B,EAAG0B,IAAO,CAChC,IAAIC,EAAQD,EACZ,QAASE,EAAIF,EAAM,EAAGE,EAAI5B,EAAG4B,IAAS,KAAK,IAAIH,EAAEG,CAAC,EAAEF,CAAG,CAAC,EAAI,KAAK,IAAID,EAAEE,CAAK,EAAED,CAAG,CAAC,IAAGC,EAAQC,GAC7F,GAAI,KAAK,IAAIH,EAAEE,CAAK,EAAED,CAAG,CAAC,EAAI,MAAO,OAAO,KAC3C,CAACD,EAAEC,CAAG,EAAGD,EAAEE,CAAK,CAAC,EAAI,CAACF,EAAEE,CAAK,EAAGF,EAAEC,CAAG,CAAC,EAEvC,QAASE,EAAI,EAAGA,EAAI5B,EAAG4B,IAAK,CAC1B,GAAIA,IAAMF,EAAK,SACf,IAAMG,EAASJ,EAAEG,CAAC,EAAEF,CAAG,EAAID,EAAEC,CAAG,EAAEA,CAAG,EACrC,GAAIG,IAAW,EACf,QAAS,EAAIH,EAAK,GAAK1B,EAAG,IAAKyB,EAAEG,CAAC,EAAE,CAAC,GAAKC,EAASJ,EAAEC,CAAG,EAAE,CAAC,GAK/D,IAAMI,EAAI,IAAI,MAAM9B,CAAC,EACrB,QAASe,EAAI,EAAGA,EAAIf,EAAGe,IAAKe,EAAEf,CAAC,EAAIU,EAAEV,CAAC,EAAEf,CAAC,EAAIyB,EAAEV,CAAC,EAAEA,CAAC,EACnD,OAAOe,CACT,CAUO,SAASC,EAAgBtC,EAAmBC,EAAsC,CACvF,IAAMsC,EAAUvC,EAAK,OAAQmC,GAAMA,EAAE,QAAU,MAAS,EAClDK,EAASzC,GAAYwC,EAAStC,CAAI,EACxC,MAAO,CACL,QAASY,GAAY2B,EAAO,EAAGA,EAAO,CAAC,EACvC,MAAOA,EAAO,MACd,OAAQA,EAAO,OACf,WAAYvC,EAAK,WACjB,EAAGsC,EAAQ,MACb,CACF,CAGO,SAASE,GAAkBC,EAA2BvD,EAAwB,CACnF,IAAIM,EAAIiD,EAAM,QAAQ,CAAC,EAAIA,EAAM,QAAQ,CAAC,EAAItD,GAAMD,EAAI,UAAU,EAC9D8C,EAAM,EACV,QAAWrC,KAAQ8C,EAAM,WAAY,CACnC,IAAMrC,EAAQV,GAAeR,EAAKS,CAAI,EACtC,QAAWa,KAASiC,EAAM,OAAO9C,CAAI,GAAK,CAAC,EACrCS,IAAUI,IAAOhB,GAAKiD,EAAM,QAAQT,CAAG,GAC3CA,IAGJ,OAAOzC,GAAQC,CAAC,CAClB,CAkCO,SAASkD,GACd3C,EACAC,EACAa,EAA8D,CAAC,EACzC,CACtB,IAAM8B,EAAQ9B,EAAK,OAAS,EACtB+B,EAAO/B,EAAK,MAAQ,IACpByB,EAAUvC,EAAK,OAAQmC,GAAMA,EAAE,QAAU,MAAS,EAClDW,EAAO,CAAE,aAAcP,EAAQ,OAASM,EAAM,EAAGN,EAAQ,OAAQ,KAAAM,EAAM,MAAAD,EAAO,OAAQ,CAAC,CAAE,EAC/F,GAAIE,EAAK,aAAc,OAAOA,EAE9B,IAAMC,EAAMjC,EAAK,KAAO,KAAK,OACvBkC,EAAW,CAAC,GAAGT,CAAO,EAC5B,QAASjB,EAAI0B,EAAS,OAAS,EAAG1B,EAAI,EAAGA,IAAK,CAC5C,IAAMC,EAAI,KAAK,MAAMwB,EAAI,GAAKzB,EAAI,EAAE,EACnC,CAAC0B,EAAS1B,CAAC,EAAG0B,EAASzB,CAAC,CAAC,EAAI,CAACyB,EAASzB,CAAC,EAAGyB,EAAS1B,CAAC,CAAC,EAGzD,IAAM2B,EAAuC,CAAE,IAAK,CAAC,EAAG,YAAa,CAAC,EAAG,UAAW,CAAC,CAAE,EACjFC,EAAoB,CAAC,EAE3B,QAASC,EAAO,EAAGA,EAAOP,EAAOO,IAAQ,CACvC,IAAMC,EAAOJ,EAAS,OAAO,CAACK,EAAG/B,IAAMA,EAAIsB,IAAUO,CAAI,EACnDG,EAAQN,EAAS,OAAO,CAACK,EAAG/B,IAAMA,EAAIsB,IAAUO,CAAI,EAC1D,GAAIG,EAAM,SAAW,GAAKF,EAAK,SAAW,EAAG,SAE7C,IAAMG,EAAOjB,EAAgBgB,EAAO,CAAE,GAAGrD,EAAM,WAAY,CAAC,CAAE,CAAC,EACzDuD,EAAOlB,EAAgBgB,EAAOrD,CAAI,EAExC,QAAWd,KAAOiE,EAChBF,EAAQ,KAAKhE,GAAQC,CAAG,CAAC,EACzB8D,EAAW,IAAI,KAAK9D,EAAI,UAAU,EAClC8D,EAAW,YAAY,KAAKR,GAAkBc,EAAMpE,CAAG,CAAC,EACxD8D,EAAW,UAAU,KAAKR,GAAkBe,EAAMrE,CAAG,CAAC,EAI1D,IAAMsE,EAAS,OAAO,QAAQR,CAAU,EAAE,IAAI,CAAC,CAACrD,EAAM8D,CAAS,KAAO,CACpE,KAAA9D,EACA,QAAS+D,GAAcD,EAAWR,CAAO,EACzC,MAAOU,GAAYF,EAAWR,CAAO,CACvC,EAAE,EACIW,EAAO,CAAC,GAAGJ,CAAM,EAAE,KAAK,CAACK,EAAG,IAAMA,EAAE,QAAU,EAAE,OAAO,EAAE,CAAC,GAAG,KAEnE,MAAO,CACL,GAAGhB,EACH,OAAAW,EACA,KAAAI,EACA,MAAOA,IAAS,YAAcvB,EAAgBC,EAAStC,CAAI,EAAIqC,EAAgBC,EAAS,CAAE,GAAGtC,EAAM,WAAY,CAAC,CAAE,CAAC,CACrH,CACF,CAEA,SAAS0D,GAAcD,EAAqBK,EAA0B,CACpE,GAAIL,EAAU,SAAW,EAAG,MAAO,GACnC,IAAIM,EAAM,EACV,QAAS1C,EAAI,EAAGA,EAAIoC,EAAU,OAAQpC,IAAK,CACzC,IAAMjC,EAAI,KAAK,IAAI,EAAIE,EAAK,KAAK,IAAIA,EAAKmE,EAAUpC,CAAC,CAAC,CAAC,EACvD0C,GAAOD,EAAOzC,CAAC,IAAM,EAAI,CAAC,KAAK,IAAIjC,CAAC,EAAI,CAAC,KAAK,IAAI,EAAIA,CAAC,EAEzD,OAAO2E,EAAMN,EAAU,MACzB,CAEA,SAASE,GAAYF,EAAqBK,EAA0B,CAClE,GAAIL,EAAU,SAAW,EAAG,MAAO,GACnC,IAAIM,EAAM,EACV,QAAS1C,EAAI,EAAGA,EAAIoC,EAAU,OAAQpC,IAAK0C,IAAQN,EAAUpC,CAAC,EAAIyC,EAAOzC,CAAC,IAAM,EAChF,OAAO0C,EAAMN,EAAU,MACzB,CAhUA,IA6BMnE,EA7BN0E,GAAAC,EAAA,kBA6BM3E,EAAM,OC7BZ,OAAS,eAAA4E,OAAmB,SAsCrB,SAASC,GACdC,EACAC,EAAMC,GACNC,EAAOC,GACO,CACd,OAAIJ,GAAeC,EAAY,KAC3BD,EAAcG,EAAa,MACxB,WACT,CA6CA,eAAsBE,GACpBC,EACAC,EACAC,EACAC,EAA0E,CAAC,EAC/C,CAC5B,IAAMC,EAAU,KAAK,IAAI,EAAGD,EAAK,SAAW,EAAE,EACxCE,EAAUC,EAAmBN,CAAW,EACxCO,EAAU,KAAK,IAAI,EAEnBC,EAAO,MAAM,QAAQ,IACzB,MAAM,KAAK,CAAE,OAAQJ,CAAQ,EAAG,CAACK,EAAGC,IAClCL,EAAQ,OAAO,CACb,MAAOM,GAAQV,EAAO,GAAGS,KAAKlB,GAAY,CAAC,EAAE,SAAS,KAAK,GAAG,EAC9D,UAAAU,EACA,MAAOC,EAAK,KACd,CAAC,CACH,CACF,EAEMS,EAAQ,OAAO,KAAKV,CAAS,EAC7BW,EAAQL,EAAK,OACjB,CAACM,EAAKC,KAAO,CACX,YAAaD,EAAI,YAAcC,EAAE,MAAM,YACvC,aAAcD,EAAI,aAAeC,EAAE,MAAM,YAC3C,GACA,CAAE,YAAa,EAAG,aAAc,CAAE,CACpC,EAEMC,EAAcJ,EAAM,IAAKK,GAAS,CACtC,IAAMC,EAAQV,EAAK,IAAKO,GAAMI,EAAYJ,EAAE,QAAgCE,CAAI,CAAC,CAAC,EAU5EG,EAAQF,EAAM,IAAI,CAACG,EAAGX,IAAM,CAChC,IAAMY,EAAUd,EAAKE,CAAC,EAAE,QAAgCO,CAAI,EAC5D,OAAOK,EAAO,OAAS,OAASA,EAAO,KAAOD,EAAE,OAClD,CAAC,EACKE,EAAU,CAAC,GAAG,IAAI,IAAIL,EAAM,IAAKG,GAAMA,EAAE,SAAS,CAAC,CAAC,EACpDG,EAAQ,CAAC,GAAG,IAAI,IAAIJ,EAAM,IAAKK,GAAMhC,GAAKgC,EAAGtB,EAAK,IAAKA,EAAK,IAAI,CAAC,CAAC,CAAC,EACnEuB,EAAIC,GAAKP,CAAK,EACpB,MAAO,CACL,SAAUH,EACV,KAAOT,EAAK,CAAC,EAAE,QAAgCS,CAAI,EAAE,KACrD,EAAGb,EACH,KAAMsB,EACN,MAAOE,GAAMR,EAAOM,CAAC,EACrB,IAAK,KAAK,IAAI,GAAGN,CAAK,EACtB,IAAK,KAAK,IAAI,GAAGA,CAAK,EACtB,gBAAiBG,EACjB,MAAAC,EACA,iBAAkBD,EAAQ,OAAS,GAAKC,EAAM,OAAS,EACvD,QAASJ,CACX,CACF,CAAC,EAED,MAAO,CACL,QAASpB,EACT,MAAOQ,EAAK,CAAC,EAAE,MACf,QAAAJ,EACA,UAAWuB,GAAKX,EAAY,IAAKa,GAAMA,EAAE,KAAK,CAAC,EAC/C,SAAUb,EAAY,OAAQa,GAAMA,EAAE,gBAAgB,EAAE,IAAKA,GAAMA,EAAE,QAAQ,EAC7E,UAAWb,EACX,MAAAH,EACA,QAAS,KAAK,IAAI,EAAIN,CACxB,CACF,CAGO,SAASI,GAAQV,EAAmB6B,EAAyB,CAClE,OAAI7B,IAAU,MAAQ,OAAOA,GAAU,UAAY,CAAC,MAAM,QAAQA,CAAK,EAC9D,CAAE,GAAIA,EAAsC,IAAA6B,CAAI,EAElD,CAAE,IAAAA,EAAK,MAAA7B,CAAM,CACtB,CAEA,SAAS0B,GAAKI,EAA0B,CACtC,OAAIA,EAAO,SAAW,EAAU,EACzBA,EAAO,OAAO,CAACC,EAAGC,IAAMD,EAAIC,EAAG,CAAC,EAAIF,EAAO,MACpD,CAEA,SAASH,GAAMG,EAAkBL,EAAmB,CAClD,OAAIK,EAAO,OAAS,EAAU,EACvB,KAAK,KAAKA,EAAO,OAAO,CAACG,EAAKb,IAAMa,GAAOb,EAAIK,IAAM,EAAG,CAAC,EAAIK,EAAO,MAAM,CACnF,CArLA,IAiCanC,GACAE,GAlCbqC,GAAAC,EAAA,kBACAC,IACAC,IA+Ba1C,GAAkB,GAClBE,GAAmB,KCIzB,SAASyC,EAA0BC,EAA2B,CAAC,EAAoB,CACxF,MAAO,CACL,KAAM,OACN,aAAAC,GACA,MAAM,OACJC,EAC8B,CAC9B,GAAIF,EAAK,KAAM,MAAMA,EAAK,KAC1BG,EAAkBD,EAAQ,SAAS,EAEnC,IAAME,EAAY,KAAK,UAAUF,EAAQ,OAAS,IAAI,EAChDG,EAAqC,CAAC,EACxCC,EAAW,GAEf,OAAW,CAACC,EAAMC,CAAQ,IAAK,OAAO,QAAQN,EAAQ,SAAS,EAAG,CAChE,IAAMO,EAAMT,EAAK,UAAUO,CAAI,GAAKG,GAAaN,EAAWG,EAAMC,CAAQ,EACpEG,EAASC,EAAeJ,EAAUC,CAAG,EAC3CJ,EAAQE,CAAI,EAAII,EAAO,OACnBA,EAAO,WAAUL,EAAW,IAGlC,IAAMO,EAAyB,gBAC/B,MAAO,CACL,MAAOX,EAAQ,OAAS,OACxB,QAASG,EACT,MAAO,CAAE,YAAaD,EAAU,OAAQ,aAAc,CAAE,EACxD,KAAM,CACJ,QAAS,OACT,cAAe,OACf,WAAAS,EACA,SAAAP,EACA,QAAS,EACT,eAAgB,GAChB,UAAWN,EAAK,WAAa,CAC/B,CACF,CACF,CACF,CACF,CAEA,SAASU,GACPN,EACAG,EACAC,EACW,CACX,GAAIA,EAAS,OAAS,OACpB,MAAO,CAAE,KAAMM,GAAS,GAAGV,KAAaG,QAAW,CAAE,EAEvD,IAAMQ,EAASC,EAASR,CAAQ,EAC1BS,EAAUF,EAAO,IAAKG,GAAUJ,GAAS,GAAGV,KAAaG,KAAQW,GAAO,EAAI,IAAI,EAChFC,EAAQF,EAAQ,OAAO,CAAC,EAAGG,IAAM,EAAIA,EAAG,CAAC,EACzCC,EAAwC,CAAC,EAC/C,OAAAN,EAAO,QAAQ,CAACG,EAAOI,IAAM,CAC3BD,EAAcH,CAAK,EAAID,EAAQK,CAAC,EAAIH,CACtC,CAAC,EACM,CAAE,cAAAE,CAAc,CACzB,CAIA,SAASP,GAASS,EAAuB,CACvC,IAAIC,EAAI,WACR,QAASF,EAAI,EAAGA,EAAIC,EAAM,OAAQD,IAChCE,GAAKD,EAAM,WAAWD,CAAC,EACvBE,EAAI,KAAK,KAAKA,EAAG,QAAU,IAAM,EAEnC,OAAOA,EAAI,UACb,CAzGA,IA6BMvB,GA7BNwB,GAAAC,EAAA,kBAOAC,IACAC,IAqBM3B,GAA4C,CAChD,kBAAmB,YACnB,wBAAyB,GACzB,iBAAkB,IAClB,cAAe,KACf,kBAAmB,GACnB,OAAQ,EACV,IC+DO,SAAS4B,GACdC,EACAC,EACU,CACV,IAAMC,EAAOC,GAAsBH,CAAY,EAC/C,GAAI,CAACE,EACH,MAAO,CAAC,qBAAqBF,gCAAsC,EAGrE,GAAI,CAACC,GAAkB,OAAQ,MAAO,CAAC,EAEvC,IAAMG,EAAqB,CAAC,EACtBC,EAAoB,CAAC,EAE3B,QAAWC,KAAWL,EAChBK,KAAWJ,GAAQ,CAAEA,EAAaI,CAAO,GAC3CD,EAAQ,KAAKC,CAAO,EAIxB,OAAID,EAAQ,QACVD,EAAS,KACP,aAAaJ,aAAwBK,EAAQ,KAAK,IAAI,oCACxD,EAGKD,CACT,CA9HA,IAqBaD,GArBbI,GAAAC,EAAA,kBAqBaL,GAA8D,CAQzE,cAAe,CACb,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,IACZ,iBAAkB,GAClB,SAAU,EACZ,EACA,OAAQ,CACN,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,IACZ,iBAAkB,GAClB,SAAU,EACZ,EACA,OAAQ,CACN,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,MACZ,iBAAkB,GAClB,SAAU,EACZ,EACA,SAAU,CACR,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,MACZ,iBAAkB,GAClB,SAAU,EACZ,EACA,OAAQ,CACN,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,KACZ,iBAAkB,GAClB,SAAU,EACZ,EACA,KAAM,CACJ,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,IACZ,iBAAkB,GAClB,SAAU,EACZ,EAGA,OAAQ,CACN,UAAW,GACX,MAAO,GACP,OAAQ,GACR,SAAU,GACV,WAAY,KACZ,iBAAkB,GAClB,SAAU,EACZ,CACF,ICrBO,SAASM,EAA2BC,EAA4B,CAAC,EAAoB,CAC1F,GAAIA,EAAK,YAAcA,EAAK,aAAe,gBACzC,MAAM,IAAI,MACR,yEAAyEA,EAAK,cAChF,EAGF,IAAMC,EAAgBD,EAAK,eAAiBE,GACtCC,EAA4C,CAChD,kBAAmB,aACrB,wBAAyB,GACvB,iBAAkB,IAClB,cAAAF,EACA,kBAAmB,GACnB,OAAQ,EACV,EAEIG,EAAiC,KAC/BC,EAAkB,KACjBD,IACHA,EAAWJ,EAAK,gBACZA,EAAK,gBAAgB,EACrBM,GAAeN,EAAK,UAAYO,EAAgB,GAE/CH,GAGT,MAAO,CACL,KAAM,QACN,aAAAD,EACA,MAAM,OACJK,EAC8B,CAC9BC,EAAkBD,EAAQ,SAAS,EAEnC,IAAME,EAAU,KAAK,IAAI,EACnBC,EAAQH,EAAQ,OAASR,EAAK,OAASY,GACvCC,EAAQR,EAAgB,EACxBS,EAAgBC,GAAqBf,EAAK,eAAiB,OAAQa,CAAK,EACxEG,EAASC,GAAiBT,EAAQ,MAAOA,EAAQ,UAAWP,CAAa,EACzEiB,EAASC,EAAiBX,EAAQ,SAAS,EAC3CY,EAAapB,EAAK,0BAA4B,EAE9C,CAAE,OAAAqB,EAAQ,QAAAC,CAAQ,EAAIC,GAC1Bf,EAAQ,YACRA,EAAQ,WAAaR,EAAK,WAAawB,EACzC,EAEIC,EAAQ,CAAE,YAAa,EAAG,aAAc,CAAE,EAC1CC,EAA4B,KAC5BC,EAAU,EACVC,EAAyB,KACzBC,EAAY,sBAEhB,GAAI,CACF,QAASC,EAAU,EAAGA,GAAWV,EAAYU,IAAW,CAClDA,EAAU,GAAGH,IAiBjB,IAAMI,EAAS,MAdbjB,IAAkB,OACdkB,GAAmBnB,EAAOL,EAAQ,UAAWQ,EAAO,KAAMU,EAAY,CACpE,MAAAf,EACA,UAAWX,EAAK,WAAaiC,GAC7B,YAAajC,EAAK,aAAe,EACjC,YAAaqB,CACf,CAAC,EACDa,GAAarB,EAAOL,EAAQ,UAAWQ,EAAO,KAAMU,EAAY,CAC9D,MAAAf,EACA,UAAWX,EAAK,WAAaiC,GAC7B,YAAajC,EAAK,aAAe,EACjC,YAAaqB,CACf,CAAC,GAGPI,EAAQ,CACN,YAAaA,EAAM,YAAcM,EAAO,MAAM,YAC9C,aAAcN,EAAM,aAAeM,EAAO,MAAM,YAClD,EAEA,IAAMI,GAASjB,EAAO,UAAUa,EAAO,OAAO,EAC9C,GAAII,GAAO,QAAS,CAClBP,EAAMO,GAAO,KACb,MAEFN,EAAYO,EAAeD,GAAO,KAAmB,EACrDT,EAAa,CACX,4DACAG,EACA,GACA,gDACF,EAAE,KAAK;AAAA,CAAI,EAGb,GAAI,CAACD,EACH,MAAM,IAAI,MACR,6DAA6DD,EAAU,iBAAiBE,GAC1F,EAGF,IAAMQ,EAAqC,CAAC,EACxCC,EAAW,GACf,OAAW,CAACC,EAAMC,CAAQ,IAAK,OAAO,QAAQhC,EAAQ,SAAS,EAAG,CAChE,IAAMuB,EAASU,EAAeD,EAAUZ,EAAIW,CAAI,EAAgB,CAC9D,UAAWvC,EAAK,wBAA0B,EAC5C,CAAC,EACDqC,EAAQE,CAAI,EAAIR,EAAO,OACnBA,EAAO,WAAUO,EAAW,IAGlC,MAAO,CACL,MAAA3B,EACA,QAAS0B,EACT,MAAAZ,EACA,KAAM,CACJ,QAAS,QACT,cAAAX,EACA,WAAY,gBACZ,SAAAwB,EACA,QAAAX,EACA,eAAgBX,EAAO,UACvB,UAAW,KAAK,IAAI,EAAIN,CAC1B,CACF,CACF,QAAE,CACAY,EAAQ,CACV,CACF,CACF,CACF,CAcA,eAAeU,GACbnB,EACA6B,EACAC,EACAjB,EACAkB,EACwB,CACxB,GAAI,CAAC/B,EAAM,YACT,MAAM,IAAI,MAAM,aAAaA,EAAM,sDAAsD,EAG3F,IAAMgC,EAAW,CACf,CAAE,KAAM,OAAiB,QAASF,CAAW,EAC7C,GAAIjB,EAAa,CAAC,CAAE,KAAM,OAAiB,QAASA,CAAW,CAAC,EAAI,CAAC,CACvE,EAEMK,EAAS,MAAMlB,EAAM,YACzBgC,EACAC,EACA,CACE,CACE,KAAMC,EACN,YAAa,wDACb,aAAcC,GAAgBN,CAAS,CACzC,CACF,EACA,CAAE,GAAGE,EAAS,WAAY,CAAE,KAAM,OAAQ,KAAMG,CAAU,CAAE,CAC9D,EAEME,EAAQlB,EAAO,QAAQ,KAAMmB,GAAMA,EAAE,OAAS,YAAcA,EAAE,OAASH,CAAS,EAChFI,EAAQF,GAASA,EAAM,OAAS,WAAaA,EAAM,MAAQ,KAEjE,MAAO,CACL,QAASG,GAAcD,CAAK,EAC5B,MAAO,CACL,YAAapB,EAAO,MAAM,aAC1B,aAAcA,EAAO,MAAM,aAC7B,CACF,CACF,CAEA,eAAeG,GACbrB,EACA6B,EACAC,EACAjB,EACAkB,EACwB,CACxB,IAAMS,EAAU,CAACV,EAAYW,GAAwBZ,CAAS,CAAC,EAAE,KAAK;AAAA,CAAI,EACpEG,EAAW,CACf,CAAE,KAAM,SAAmB,QAASC,CAAc,EAClD,CAAE,KAAM,OAAiB,QAAAO,CAAQ,EACjC,GAAI3B,EAAa,CAAC,CAAE,KAAM,OAAiB,QAASA,CAAW,CAAC,EAAI,CAAC,CACvE,EAEMK,EAAS,MAAMlB,EAAM,SAASgC,EAAUD,CAAO,EACrD,MAAO,CACL,QAASQ,GAAcG,GAAYxB,EAAO,SAAW,EAAE,CAAC,EACxD,MAAO,CAAE,YAAa,EAAG,aAAcA,EAAO,YAAc,CAAE,CAChE,CACF,CAKA,SAASqB,GAAcI,EAAyB,CAC9C,GAAIA,GAAS,OAAOA,GAAU,UAAY,CAAC,MAAM,QAAQA,CAAK,EAAG,CAC/D,IAAMC,EAAUD,EAChB,GAAIC,EAAQ,SAAW,OAAOA,EAAQ,SAAY,SAAU,OAAOA,EAAQ,QAE7E,OAAOD,CACT,CAEA,SAASzC,GACP2C,EACA7C,EACe,CACf,OAAI6C,IAAc,OAAeA,EAC5B7C,EAAM,aAGJ8C,GAAsB9C,EAAM,IAAI,GAAG,iBAAmB,OAH9B,MAIjC,CAKA,SAASU,GACPqC,EACAC,EAC8C,CAC9C,IAAMC,EAAa,IAAI,gBACjBC,EAAQ,WACZ,IAAMD,EAAW,MAAM,IAAI,MAAM,4BAA4BD,KAAa,CAAC,EAC3EA,CACF,EACMG,EAAgB,IAAMF,EAAW,MAAMF,GAAQ,MAAM,EAE3D,OAAIA,IACEA,EAAO,QAASI,EAAc,EAC7BJ,EAAO,iBAAiB,QAASI,EAAe,CAAE,KAAM,EAAK,CAAC,GAG9D,CACL,OAAQF,EAAW,OACnB,QAAS,IAAM,CACb,aAAaC,CAAK,EAClBH,GAAQ,oBAAoB,QAASI,CAAa,CACpD,CACF,CACF,CArUA,IAiDMzD,GACAK,GACAV,GACA+B,GACAT,GArDNyC,GAAAC,EAAA,kBAAAC,KACAC,KAEAC,KASAC,IACAC,IAOAC,IACAC,IA4BMlE,GAAiC,cACjCK,GAAgB,4BAChBV,GAA0B,KAC1B+B,GAAqB,IACrBT,GAAqB,MCxC3B,OAAS,gBAAAkD,OAAoB,KAC7B,OAAS,WAAAC,OAAe,KACxB,OAAS,QAAAC,OAAY,OAmEd,SAASC,EAAuBC,EAAyB,CAAC,EAAoB,CACnF,IAAMC,GAAWD,EAAK,SAAWE,IAAkB,QAAQ,OAAQ,EAAE,EAC/DC,EAAOH,EAAK,MAAQ,cACpBI,EAAOJ,EAAK,MAAQ,aACpBK,EAAgBL,EAAK,eAAiBM,GACtCC,EAAmBP,EAAK,kBAAoBQ,GAE5CC,EAA4C,CAChD,kBAAmBT,EAAK,mBAAqB,SAC7C,wBAAyB,GACzB,iBAAAO,EACA,cAAAF,EACA,kBAAmB,GACnB,OAAQ,EACV,EAEA,MAAO,CACL,KAAAD,EACA,aAAAK,EACA,MAAM,OACJC,EAC8B,CAC9BC,EAAkBD,EAAQ,SAAS,EACnCE,GAAwBF,EAAQ,UAAWH,CAAgB,EAE3D,IAAMM,EAAU,KAAK,IAAI,EACnBC,EAAQJ,EAAQ,OAASV,EAAK,MACpC,GAAI,CAACc,EAAO,MAAM,IAAI,MAAM,qCAAqC,EAEjE,IAAMC,EAAWC,EAAYN,EAAQ,MAAOL,CAAa,EACnDY,EAAUjB,EAAK,WAAa,MAE5B,CAAE,OAAAkB,EAAQ,QAAAC,CAAQ,EAAIC,GAC1BV,EAAQ,YACRA,EAAQ,WAAaV,EAAK,WAAaqB,EACzC,EAEA,GAAI,CACF,IAAMC,EAAkC,CAAE,eAAgB,kBAAmB,EACvEC,EAASC,GAAWxB,EAAK,UAAWA,EAAK,UAAU,EACrDuB,IAAQD,EAAQ,cAAgB,UAAUC,KAE9C,IAAME,EAAM,MAAMR,EAAQ,GAAGhB,IAAUE,IAAQ,CAC7C,OAAQ,OACR,QAAAmB,EACA,KAAM,KAAK,UAAU,CACnB,MAAAR,EACA,MAAOC,EAAS,KAChB,UAAWW,GAAiBhB,EAAQ,SAAS,CAC/C,CAAC,EACD,OAAAQ,CACF,CAAC,EAED,GAAI,CAACO,EAAI,GAAI,CACX,IAAME,EAAS,MAAMF,EAAI,KAAK,EAAE,MAAM,IAAM,EAAE,EAC9C,MAAM,IAAI,MAAM,GAAGrB,KAAQqB,EAAI,WAAWE,EAAO,MAAM,EAAG,GAAG,GAAG,EAGlE,IAAMC,EAAQ,MAAMH,EAAI,KAAK,EAMvBI,EAASC,EAAiBpB,EAAQ,SAAS,EAAE,UAAUkB,EAAK,OAAO,EACzE,GAAI,CAACC,EAAO,QACV,MAAM,IAAI,MACR,GAAGzB;AAAA,EACgE2B,EAAeF,EAAO,KAAmB,GAC9G,EAGF,IAAMG,EAAMH,EAAO,KACbI,EAAqC,CAAC,EACxCC,EAAW,GACf,OAAW,CAAC9B,EAAM+B,CAAQ,IAAK,OAAO,QAAQzB,EAAQ,SAAS,EAAG,CAChE,IAAM0B,EAASC,EAAeF,EAAUH,EAAI5B,CAAI,CAAC,EACjD6B,EAAQ7B,CAAI,EAAIgC,EAAO,OACnBA,EAAO,WAAUF,EAAW,IAGlC,MAAO,CACL,MAAON,EAAK,OAASd,EACrB,QAASmB,EACT,MAAO,CACL,YAAaL,EAAK,OAAO,cAAgB,EACzC,aAAcA,EAAK,OAAO,eAAiB,CAC7C,EACA,KAAM,CACJ,QAASxB,EAMT,cAAeK,EAAa,oBAAsB,SAAW,SAAW,WACxE,WAAY,gBACZ,SAAAyB,EACA,QAAS,EACT,eAAgBnB,EAAS,UACzB,UAAW,KAAK,IAAI,EAAIF,CAC1B,CACF,CACF,QAAE,CACAM,EAAQ,CACV,CACF,CACF,CACF,CAkBA,SAASO,GAAiBY,EAAiC,CACzD,IAAMC,EAA+B,CAAC,EACtC,OAAW,CAACnC,EAAM+B,CAAQ,IAAK,OAAO,QAAQG,CAAS,EACrDC,EAAInC,CAAI,EAAI+B,EAAS,aACjBA,EACA,CAAE,GAAGA,EAAU,aAAcK,GAAqBL,EAAS,IAAI,CAAE,EAEvE,OAAOI,CACT,CAMA,SAASf,GAAWiB,EAAiBC,EAAmC,CACtE,IAAMC,EAAUF,EAAS,QAAQ,IAAIA,CAAM,EAAI,OAC/C,GAAIE,EAAS,OAAOA,EACpB,GAAI,CAACD,EAAM,OAEX,IAAMvC,EAAOuC,EAAK,WAAW,IAAI,EAAI5C,GAAKD,GAAQ,EAAG6C,EAAK,MAAM,CAAC,CAAC,EAAIA,EACtE,GAAI,CAACE,EAAa,IAAIzC,CAAI,EACxB,GAAI,CACF,IAAM6B,EAAMpC,GAAaO,EAAM,OAAO,EAAE,KAAK,EAC7CyC,EAAa,IAAIzC,EAAM6B,EAAI,OAAS,EAAIA,EAAM,IAAI,CACpD,MAAE,CACAY,EAAa,IAAIzC,EAAM,IAAI,CAC7B,CAEF,OAAOyC,EAAa,IAAIzC,CAAI,GAAK,MACnC,CAEA,SAASS,GAAwB0B,EAAsBO,EAAmB,CACxE,OAAW,CAACC,EAAcX,CAAQ,IAAK,OAAO,QAAQG,CAAS,EAC7D,GAAIH,EAAS,OAAS,SAAU,CAC9B,IAAMY,EAAI,OAAO,KAAKZ,EAAS,QAAQ,EAAE,OACzC,GAAIY,EAAIF,EACN,MAAM,IAAI,MACR,aAAaC,uBAAkCD,yBAA2BE,GAC5E,UAEOZ,EAAS,OAAS,SAAWA,EAAS,SAAS,OAASU,EACjE,MAAM,IAAI,MACR,aAAaC,uBAAkCD,uBAAyBV,EAAS,SAAS,QAC5F,CAGN,CAEA,SAASf,GACP4B,EACAC,EAC8C,CAC9C,IAAMC,EAAa,IAAI,gBACjBC,EAAQ,WACZ,IAAMD,EAAW,MAAM,IAAI,MAAM,4BAA4BD,KAAa,CAAC,EAC3EA,CACF,EACMG,EAAgB,IAAMF,EAAW,MAAMF,GAAQ,MAAM,EAC3D,OAAIA,IACEA,EAAO,QAASI,EAAc,EAC7BJ,EAAO,iBAAiB,QAASI,EAAe,CAAE,KAAM,EAAK,CAAC,GAE9D,CACL,OAAQF,EAAW,OACnB,QAAS,IAAM,CACb,aAAaC,CAAK,EAClBH,GAAQ,oBAAoB,QAASI,CAAa,CACpD,CACF,CACF,CArRA,IA6EMlD,GACAmB,GACAf,GACAE,GA0HAgC,GAkBAI,EA5NNS,GAAAC,EAAA,kBAOAC,IACAC,IACAC,IAEAC,IAkEMxD,GAAmB,2BACnBmB,GAAqB,IACrBf,GAA0B,IAC1BE,GAA6B,GA0H7BgC,GAA+C,CACnD,OAAQ,0BACR,MAAO,yBACP,KAAM,4BACR,EAcMI,EAAe,IAAI,MC5NzB,IAAAe,GAAA,GAAAC,GAAAD,GAAA,mBAAAE,EAAA,4DAAAC,EAAA,cAAAC,EAAA,qBAAAC,GAAA,oBAAAC,GAAA,qCAAAC,GAAA,cAAAC,GAAA,eAAAC,EAAA,sBAAAC,GAAA,qBAAAC,GAAA,YAAAC,GAAA,SAAAC,GAAA,sBAAAC,GAAA,UAAAC,GAAA,gBAAAC,GAAA,sBAAAC,GAAA,WAAAC,GAAA,uBAAAC,GAAA,kBAAAC,GAAA,uBAAAC,GAAA,kBAAAC,GAAA,+BAAAC,EAAA,8BAAAC,EAAA,2BAAAC,EAAA,qBAAAC,GAAA,mBAAAC,EAAA,QAAAC,GAAA,kBAAAC,GAAA,mBAAAC,EAAA,gBAAAC,GAAA,oBAAAC,EAAA,mBAAAC,GAAA,uBAAAC,EAAA,gBAAAC,GAAA,aAAAC,EAAA,yBAAAC,GAAA,YAAAC,GAAA,UAAAC,GAAA,QAAAC,GAAA,uBAAAC,GAAA,0BAAAC,GAAA,yBAAAC,GAAA,SAAAC,GAAA,SAAAC,GAAA,kBAAAC,GAAA,oBAAAC,GAAA,qBAAAC,EAAA,sBAAAC,GAAA,oCAAAC,GAAA,4BAAAC,EAAA,oBAAAC,GAAA,oBAAAC,GAAA,4BAAAC,GAAA,gBAAAC,EAAA,qBAAAC,GAAA,0BAAAC,GAAA,UAAAC,GAAA,eAAAC,GAAA,YAAAC,GAAA,oBAAAC,GAAA,sBAAAC,EAAA,YAAAC,KAqCO,SAASb,GACdc,EAA6B,CAAC,EAC9BC,EAA8B,CAAC,EAC/BC,EAAwB,CAAC,EACzBC,EAA6B,CAAC,EACxB,CACNhB,EAAwB,OAAQ,IAAM3B,EAA0B,CAAC,EACjE2B,EAAwB,QAAS,IAAM5B,EAA2ByC,CAAK,CAAC,EACxEb,EAAwB,aAAc,IAAM1B,EAAuBwC,CAAS,CAAC,EAY7Ed,EAAwB,MAAO,IAC7B1B,EAAuB,CACrB,KAAM,MACN,QAAS,kCACT,KAAM,aACN,MAAO,aACP,UAAW,qBACX,WAAY,+BAKZ,kBAAmB,SACnB,iBAAkB,IAKlB,cAAe,IACf,GAAGyC,CACL,CAAC,CACH,EAKAf,EAAwB,WAAY,IAClC1B,EAAuB,CACrB,KAAM,WACN,QAAS,6BACT,KAAM,aACN,MAAO,aACP,UAAW,mBACX,WAAY,6BACZ,kBAAmB,SACnB,iBAAkB,IAClB,cAAe,IACf,GAAG0C,CACL,CAAC,CACH,CACF,CAjGA,IAAAC,GAAAC,EAAA,kBAYAC,KACAC,IACAC,IACAC,IACAC,IACAC,IACAC,IACAC,KACAC,KACAC,KACAC,KACAC,KAEAC,KAEAC,KAGAT,IACAO,KACAC,KACAC,OC0BO,SAASC,GAAmBC,EAAuC,CACxEC,EAAU,CAAE,GAAGC,GAAiB,GAAGD,EAAS,GAAGD,CAAK,EACpDG,GAAa,EACf,CAEO,SAASC,IAA8B,CAC5CH,EAAU,CAAE,GAAGC,GAAiB,MAAO,CAAC,CAAE,EAC1CC,GAAa,EACf,CAgBA,SAASE,IAAyB,CAChC,GAAI,CAAAF,GACJ,CAAAA,GAAa,GACb,GAAI,CAGF,GAAM,CAAE,iBAAAG,CAAiB,EAAI,cAGvBC,EAAMD,EAAiB,GAAG,UAChC,GAAI,CAACC,GAAK,QAAS,OAEnB,GAAM,CAAE,gCAAAC,CAAgC,EAAI,cAGtCC,EAAIF,EAAI,UAAY,CAAC,EAC3BC,EAAgCC,EAAE,MAAOA,EAAE,UAAWA,EAAE,IAAKA,EAAE,QAAQ,EAEvE,IAAIC,EAA8B,KAClC,GAAI,CACFA,EAAQ,IAAIC,EAAc,CAAE,KAAMJ,EAAI,MAAO,CAAC,CAChD,MAAE,CAEF,CACAN,EAAU,CACR,QAAS,GACT,eAAgBM,EAAI,eACpB,MAAAG,EACA,MAAO,OAAO,YACZ,OAAO,QAAQH,EAAI,OAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAACK,EAAMC,CAAI,IAAqB,CACnED,EACA,CAAE,GAAGC,EAAM,YAAaN,EAAI,YAAa,cAAeA,EAAI,aAAc,CAC5E,CAAC,CACH,CACF,CACF,MAAE,CAEF,EACF,CAEO,SAASO,IAAqC,CACnD,OAAOb,CACT,CAIO,SAASc,GAAcC,EAAiC,CAC7D,GAAI,OAAOA,GAAU,SAAU,OAAO,KACtC,IAAMC,EAAaD,EAAM,KAAK,EAAE,YAAY,EAC5C,OAAQE,GAAkC,SAASD,CAAU,EACxDA,EACD,IACN,CAGO,SAASE,GAAWN,EAAsB,CAC/C,MAAO,wBAAwBA,EAAK,YAAY,EAAE,QAAQ,cAAe,GAAG,GAC9E,CAaO,SAASO,GAAYP,EAAcQ,EAAyB,QAAQ,IAAe,CAcxFhB,GAAiB,EACjB,IAAMiB,EAAUP,GAAcM,EAAIF,GAAWN,CAAI,CAAC,CAAC,EACnD,OAAIS,IACCrB,EAAQ,QACNc,GAAcd,EAAQ,MAAMY,CAAI,GAAG,IAAI,GAAK,MADtB,MAE/B,CA4BO,SAASU,GACdC,EACAC,EACAC,EAAW,GACL,CACN,GAAI,GAACF,GAAU,CAACvB,EAAQ,OACxB,GAAI,CACFA,EAAQ,MAAM,cAAcuB,EAAQC,EAAQC,CAAQ,CACtD,MAAE,CAEF,CACF,CAIA,eAAsBC,GACpBd,EACAe,EACAC,EACAC,EAA0B,CAAC,EACI,CAC/B,IAAMC,EAAOX,GAAYP,CAAI,EAC7B,GAAIkB,IAAS,MAAO,OAAO,KAE3B,IAAMC,EAAW/B,EAAQ,MAAMY,CAAI,GAAK,CAAE,KAAAkB,CAAK,EACzCE,EAAcH,EAAK,SAAWE,EAAS,SAAW/B,EAAQ,eAC1DiC,EAAU,KAAK,IAAI,EAEzB,GAAI,CAEF,IAAMC,EAAW,MADDC,EAAmBH,CAAW,EACf,OAAO,CACpC,MAAAL,EACA,UAAAC,EACA,MAAOC,EAAK,OAASE,EAAS,MAC9B,YAAaF,EAAK,OAClB,UAAWA,EAAK,WAAaE,EAAS,SACxC,CAAC,EAEKR,EAASa,GAAOxB,EAAMkB,EAAMH,EAAOC,EAAWM,EAAUL,EAAME,CAAQ,EAC5E,MAAO,CAAE,QAASG,EAAS,QAAS,OAAAX,EAAQ,KAAAO,CAAK,CACnD,OAASO,EAAP,CACA,OAAAC,GAAc1B,EAAMkB,EAAMH,EAAOC,EAAWI,EAAaH,EAAME,EAAUE,EAASI,CAAG,EACrFE,GAAc3B,EAAMoB,EAAaK,CAAG,EAC7B,IACT,CACF,CAEA,SAASD,GACPxB,EACAkB,EACAH,EACAC,EACAM,EACAL,EACAE,EACe,CACf,GAAI,CAAC/B,EAAQ,MAAO,OAAO,KAC3B,GAAI,CACF,OAAOA,EAAQ,MAAM,WAAW,CAC9B,KAAAY,EACA,KAAAkB,EACA,QAASI,EAAS,KAAK,QACvB,MAAOA,EAAS,MAChB,KAAMA,EAAS,KACf,MAAAP,EACA,UAAAC,EACA,QAASM,EAAS,QAClB,MAAOA,EAAS,MAChB,UAAWM,GAAYX,EAAK,SAAS,EACrC,MAAOA,EAAK,MACZ,SAAUA,EAAK,SACf,UAAW,CAACE,EAAS,WACvB,CAAC,CACH,MAAE,CAEA,OAAO,IACT,CACF,CAEA,SAASO,GACP1B,EACAkB,EACAH,EACAC,EACAa,EACAZ,EACAE,EACAE,EACAI,EACM,CACN,GAAKrC,EAAQ,MACb,GAAI,CACFA,EAAQ,MAAM,WAAW,CACvB,KAAAY,EACA,KAAAkB,EACA,QAAAW,EACA,MAAOZ,EAAK,OAASE,EAAS,OAAS,UACvC,KAAM,CACJ,cAAe,OACf,WAAY,gBACZ,QAAS,EACT,eAAgB,GAChB,UAAW,KAAK,IAAI,EAAIE,CAC1B,EACA,MAAAN,EACA,UAAAC,EACA,QAAS,CAAC,EACV,UAAWY,GAAYX,EAAK,SAAS,EACrC,MAAOA,EAAK,MACZ,SAAUA,EAAK,SACf,MAAO,OAAOQ,GAAK,SAAWA,CAAG,EAAE,MAAM,EAAG,GAAG,EAC/C,UAAW,CAACN,EAAS,WACvB,CAAC,CACH,MAAE,CAEF,CACF,CAEA,SAASS,GACPE,EACwD,CACxD,GAAI,CAACA,EAAW,OAChB,IAAMC,EAAkD,CAAC,EACzD,OAAW,CAACC,EAAU7B,CAAK,IAAK,OAAO,QAAQ2B,CAAS,EAClD3B,IAAU,SAAW4B,EAAIC,CAAQ,EAAI,CAAE,MAAA7B,CAAM,GAEnD,OAAO4B,CACT,CAIA,SAASJ,GAAc3B,EAAc6B,EAAiBJ,EAAgB,CACpE,IAAMQ,EAAM,KAAK,IAAI,EACjBA,EAAMC,GAAa,MACvBA,GAAaD,EACb,QAAQ,KACN,qBAAqBjC,kBAAqB6B,cAAoBJ,GAAK,SAAWA,GAChF,EACF,CAlVA,IA+CMpC,GAOFD,EACAE,GAwEEe,GAsFF6B,GArNJC,GAAAC,EAAA,KAAAC,IACAC,IA8CMjD,GAAoC,CACxC,QAAS,GACT,eAAgB,QAChB,MAAO,CAAC,EACR,MAAO,IACT,EAEID,EAA4B,CAAE,GAAGC,EAAgB,EACjDC,GAAa,GAwEXe,GAAmC,CAAC,MAAO,SAAU,QAAQ,EAsF/D6B,GAAa","names":["registerDecisionBackend","name","factory","factories","instances","getDecisionBackend","cached","known","listDecisionBackends","backend","_resetDecisionBackendsForTesting","init_backend","__esmMin","Database","createHash","mkdirSync","dirname","resolve","answerView","answer","p","probabilities","predicted","confidence","rounded","parseFeatures","raw","parsed","runMigrations","db","current","migrationV1","migrationV2","migrationV3","DecisionStore","init_store","__esmMin","init_ulid","opts","input","ts","callId","newEventId","stateJson","stateHash","insertAnswer","question","view","insertIncumbent","row","insertLink","link","value","id","action","explored","kind","r","filter","where","params","sql","seat","keepRows","init_types","__esmMin","normalizeDistribution","raw","labels","opts","normalize","epsilon","source","repaired","key","probs","sum","label","value","uniform","pMax","values","normalizedNegEntropy","v","k","total","a","b","h","p","clamp01","expectedScore","weighted","level","confidenceFor","n","finalizeAnswer","question","ok","labelsOf","max","negEntropy","best","legend","description","i","init_normalize","__esmMin","init_questions","z","rawAnswerSchema","question","noulSchema","distributionSchema","rawAnswersSchema","questions","shape","name","describeIssues","error","issue","init_schema","__esmMin","toolInputSchema","questions","properties","name","question","answerObjectSchema","labels","labelsOf","label","renderState","state","maxChars","text","renderQuestions","blocks","lines","description","renderUserPrompt","rendered","renderShapeInstructions","example","probabilities","TOOL_NAME","SYSTEM_PROMPT","init_prompt","__esmMin","init_questions","outcome","row","logit","p","clamped","EPS","sigmoid","z","e","covariateValue","name","fromFeatures","own","buildDesign","rows","spec","minCount","levels","counts","value","kept","n","names","level","X","y","features","fitLogistic","opts","l2","iterations","d","w","iter","H","g","i","j","weight","residual","k","step","solve","delta","A","b","M","col","pivot","r","factor","x","fitRecalibrator","labeled","design","applyRecalibrator","model","compareCalibrators","folds","minN","base","rng","shuffled","predictors","actuals","fold","test","_","train","flat","full","scores","predicted","binaryLogLoss","binaryBrier","best","a","actual","sum","init_recalibrate","__esmMin","randomBytes","band","probability","low","UNCERTAINTY_LOW","high","UNCERTAINTY_HIGH","measureConsistency","backendName","state","questions","opts","samples","backend","getDecisionBackend","started","runs","_","i","withUid","names","usage","acc","r","perQuestion","name","views","answerView","probs","v","answer","answers","bands","p","m","mean","stdev","q","uid","values","a","b","sum","init_consistency","__esmMin","init_backend","init_store","createMockDecisionBackend","opts","capabilities","request","validateQuestions","stateText","answers","repaired","name","question","raw","syntheticRaw","result","finalizeAnswer","answerMode","unitHash","labels","labelsOf","weights","label","total","b","probabilities","i","input","h","init_mock","__esmMin","init_normalize","init_questions","checkCapabilities","providerName","requiredFeatures","caps","PROVIDER_CAPABILITIES","warnings","missing","feature","init_capabilities","__esmMin","createLocalDecisionBackend","opts","maxStateChars","DEFAULT_MAX_STATE_CHARS","capabilities","provider","resolveProvider","createProvider","DEFAULT_PROVIDER","request","validateQuestions","started","model","DEFAULT_MODEL","agent","structureMode","resolveStructureMode","prompt","renderUserPrompt","schema","rawAnswersSchema","maxRetries","signal","dispose","deadline","DEFAULT_TIMEOUT_MS","usage","correction","retries","raw","lastError","attempt","result","callWithForcedTool","DEFAULT_MAX_TOKENS","callWithText","parsed","describeIssues","answers","repaired","name","question","finalizeAnswer","questions","userPrompt","options","messages","SYSTEM_PROMPT","TOOL_NAME","toolInputSchema","block","b","input","unwrapAnswers","content","renderShapeInstructions","extractJson","value","wrapper","requested","PROVIDER_CAPABILITIES","caller","timeoutMs","controller","timer","onCallerAbort","init_local_llm","__esmMin","init_providers","init_capabilities","init_extract_json","init_normalize","init_prompt","init_questions","init_schema","readFileSync","homedir","join","createSimpleJevBackend","opts","baseUrl","DEFAULT_BASE_URL","path","name","maxStateChars","DEFAULT_MAX_STATE_CHARS","maxChoiceOptions","DEFAULT_MAX_CHOICE_OPTIONS","capabilities","request","validateQuestions","assertChoiceCardinality","started","model","rendered","renderState","doFetch","signal","dispose","deadline","DEFAULT_TIMEOUT_MS","headers","apiKey","resolveKey","res","withInstructions","detail","body","parsed","rawAnswersSchema","describeIssues","raw","answers","repaired","question","result","finalizeAnswer","questions","out","DEFAULT_INSTRUCTIONS","envVar","file","fromEnv","keyFileCache","max","questionName","n","caller","timeoutMs","controller","timer","onCallerAbort","init_simple_jev","__esmMin","init_normalize","init_questions","init_schema","init_prompt","decisions_exports","__export","DecisionStore","SYSTEM_PROMPT","TOOL_NAME","UNCERTAINTY_HIGH","UNCERTAINTY_LOW","_resetDecisionBackendsForTesting","agreement","answerView","applyRecalibrator","applyTemperature","askSeat","band","bootstrapInterval","brier","buildDesign","calibrationReport","choice","compareCalibrators","confidenceFor","configureDecisions","coverageCurve","createLocalDecisionBackend","createMockDecisionBackend","createSimpleJevBackend","decisionsRuntime","describeIssues","ece","expectedScore","finalizeAnswer","fitLogistic","fitRecalibrator","fitTemperature","getDecisionBackend","getSeatMode","labelsOf","listDecisionBackends","logLoss","logit","mce","measureConsistency","normalizeDistribution","normalizedNegEntropy","noul","pMax","parseSeatMode","rawAnswerSchema","rawAnswersSchema","recordSeatOutcome","registerBuiltinDecisionBackends","registerDecisionBackend","reliabilityBins","renderQuestions","renderShapeInstructions","renderState","renderUserPrompt","resetDecisionsRuntime","score","seatEnvVar","sigmoid","toolInputSchema","validateQuestions","withUid","local","simpleJev","jev","typesafe","init_decisions","__esmMin","init_types","init_questions","init_normalize","init_schema","init_backend","init_prompt","init_store","init_calibration","init_recalibrate","init_consistency","init_seat","init_mock","init_local_llm","init_simple_jev","configureDecisions","next","runtime","DEFAULT_RUNTIME","configured","resetDecisionsRuntime","ensureConfigured","loadDaemonConfig","cfg","registerBuiltinDecisionBackends","b","store","DecisionStore","name","seat","decisionsRuntime","parseSeatMode","value","normalized","VALID_MODES","seatEnvVar","getSeatMode","env","fromEnv","recordSeatOutcome","callId","action","explored","askSeat","state","questions","opts","mode","settings","backendName","started","response","getDecisionBackend","record","err","recordFailure","warnThrottled","toIncumbent","backend","incumbent","out","question","now","lastWarnAt","init_seat","__esmMin","init_backend","init_store"]}

package/dist/chunk-BH3GTCZ6.js

@@ -0,0 +1,2 @@
+import{b as h}from"./chunk-FIOB3MEX.js";function p(i){let o=i.filter(l=>l.incumbent!==void 0),a=o.length;if(a===0)return{n:0,agree:0,rate:0,kappa:0};let t=0,n=new Map,r=new Map;for(let l of o)l.predicted===l.incumbent&&t++,n.set(l.predicted,(n.get(l.predicted)??0)+1),r.set(l.incumbent,(r.get(l.incumbent)??0)+1);let e=t/a,u=0;for(let[l,f]of n)u+=f/a*((r.get(l)??0)/a);let c=u<1?(e-u)/(1-u):1;return{n:a,agree:t,rate:e,kappa:c}}function s(i,o=10,a="width"){let t=i.filter(e=>e.truth!==void 0);if(t.length===0)return[];let n=[...t].sort((e,u)=>e.confidence-u.confidence);return(a==="mass"?v(n,o):y(n,o)).filter(e=>e.length>0).map(e=>{let u=e.map(c=>c.confidence);return{lo:Math.min(...u),hi:Math.max(...u),n:e.length,meanConfidence:m(u),accuracy:m(e.map(c=>c.predicted===c.truth?1:0))}})}function d(i,o=10,a="width"){let t=s(i,o,a),n=t.reduce((r,e)=>r+e.n,0);return n===0?0:t.reduce((r,e)=>r+e.n/n*Math.abs(e.accuracy-e.meanConfidence),0)}function g(i,o=10,a="width"){let t=s(i,o,a);return t.length===0?0:Math.max(...t.map(n=>Math.abs(n.accuracy-n.meanConfidence)))}function R(i){let o=i.filter(a=>a.truth!==void 0);return o.length===0?0:m(o.map(a=>{let t=0;for(let[n,r]of Object.entries(a.probabilities)){let e=n===a.truth?1:0;t+=(r-e)**2}return t}))}function b(i,o=1e-12){let a=i.filter(t=>t.truth!==void 0);return a.length===0?0:m(a.map(t=>{let n=t.probabilities[t.truth]??0;return-Math.log(Math.max(n,o))}))}function M(i,o=20){let a=i.filter(r=>r.truth!==void 0),t=a.length;if(t===0)return[];let n=[];for(let r=0;r<=o;r++){let e=r/o,u=a.filter(c=>c.confidence>=e);n.push({threshold:e,coverage:u.length/t,accuracy:u.length>0?m(u.map(c=>c.predicted===c.truth?1:0)):0,n:u.length})}return n}function x(i,o){let a=Math.max(o,1e-6),t={},n=0;for(let[r,e]of Object.entries(i)){let u=Math.pow(Math.max(e,0),1/a);t[r]=u,n+=u}if(n<=0){let r=Object.keys(i);for(let e of r)t[e]=1/r.length;return t}for(let r of Object.keys(t))t[r]=t[r]/n;return t}function w(i){let o=i.filter(e=>e.truth!==void 0),a=b(o);if(o.length===0)return{temperature:1,nll:0,nllBefore:0};let t=e=>m(o.map(u=>{let c=x(u.probabilities,e);return-Math.log(Math.max(c[u.truth]??0,1e-12))})),n=1,r=t(1);for(let e=.1;e<=10.0001;e+=.1){let u=t(e);u<r&&(r=u,n=e)}for(let e=Math.max(.01,n-.1);e<=n+.1;e+=.01){let u=t(e);u<r&&(r=u,n=e)}return{temperature:Number(n.toFixed(3)),nll:r,nllBefore:a}}function G(i,o,a=1e3,t=.05,n=Math.random){if(i.length===0)return{lo:0,hi:0};let r=[];for(let e=0;e<a;e++){let u=[];for(let c=0;c<i.length;c++)u.push(i[Math.floor(n()*i.length)]);r.push(o(u))}return r.sort((e,u)=>e-u),{lo:r[Math.floor(t/2*r.length)],hi:r[Math.min(r.length-1,Math.floor((1-t/2)*r.length))]}}function C(i,o={}){let a=o.minN??100,t=o.bins??10,n=i.filter(e=>e.truth!==void 0),r={insufficient:n.length<a,n:i.length,nLabeled:n.length,minN:a,agreement:p(i)};return r.insufficient?r:{...r,accuracy:m(n.map(e=>e.predicted===e.truth?1:0)),ece:d(n,t,"width"),eceEqualMass:d(n,t,"mass"),eceInterval:G(n,e=>d(e,t,"width"),o.bootstrap??1e3,.05,o.rng),mce:g(n,t,"width"),brier:R(n),logLoss:b(n),bins:s(n,t,"width"),binsEqualMass:s(n,t,"mass"),coverage:M(n),temperature:w(n)}}function y(i,o){let a=Array.from({length:o},()=>[]);for(let t of i){let n=Math.min(o-1,Math.max(0,Math.floor(t.confidence*o)));a[n].push(t)}return a}function v(i,o){let a=Array.from({length:o},()=>[]);return i.forEach((t,n)=>{a[Math.min(o-1,Math.floor(n*o/i.length))].push(t)}),a}function m(i){return i.length===0?0:i.reduce((o,a)=>o+a,0)/i.length}var B=h(()=>{});export{p as a,s as b,d as c,g as d,R as e,b as f,M as g,x as h,w as i,G as j,C as k,B as l};
+//# sourceMappingURL=chunk-BH3GTCZ6.js.map

package/dist/chunk-BH3GTCZ6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/decisions/calibration.ts"],"sourcesContent":["import type { GradedRow } from \"./store\"\n\n// Does the confidence mean anything?\n//\n// Every metric here is a pure function of GradedRow[]. Three choices are\n// deliberate and worth defending:\n//\n//   1. `minN` (default 100). Below it the report refuses to print an ECE.\n//      An expected calibration error computed on thirty rows is noise, and\n//      a noisy number on a dashboard gets acted on exactly as if it were a\n//      real one.\n//\n//   2. `coverageCurve` is the headline, not ECE. The question a call site\n//      actually asks is \"what threshold do I set for autoRunThreshold\" —\n//      and that is answered by accuracy-at-threshold against the fraction\n//      of traffic retained, not by a single aggregate score.\n//\n//   3. Equal-mass bins are reported alongside equal-width. Equal-width ECE\n//      is dominated by whichever bin holds most of the mass, which for an\n//      overconfident model is the top bin — precisely the regime we are\n//      trying to detect.\n\nexport interface AgreementResult {\n  n: number\n  agree: number\n  rate: number\n  /** Cohen's kappa: agreement corrected for what chance alone would give.\n   *  Two systems that both answer \"no\" 95% of the time agree 90% of the\n   *  time while sharing no information, and kappa says so. */\n  kappa: number\n}\n\nexport function agreement(rows: GradedRow[]): AgreementResult {\n  const usable = rows.filter((r) => r.incumbent !== undefined)\n  const n = usable.length\n  if (n === 0) return { n: 0, agree: 0, rate: 0, kappa: 0 }\n\n  let agree = 0\n  const predicted = new Map<string, number>()\n  const incumbent = new Map<string, number>()\n  for (const row of usable) {\n    if (row.predicted === row.incumbent) agree++\n    predicted.set(row.predicted, (predicted.get(row.predicted) ?? 0) + 1)\n    incumbent.set(row.incumbent!, (incumbent.get(row.incumbent!) ?? 0) + 1)\n  }\n\n  const observed = agree / n\n  let expected = 0\n  for (const [label, count] of predicted) {\n    expected += (count / n) * ((incumbent.get(label) ?? 0) / n)\n  }\n  const kappa = expected < 1 ? (observed - expected) / (1 - expected) : 1\n\n  return { n, agree, rate: observed, kappa }\n}\n\nexport interface ReliabilityBin {\n  lo: number\n  hi: number\n  n: number\n  meanConfidence: number\n  accuracy: number\n}\n\nexport type BinScheme = \"width\" | \"mass\"\n\nexport function reliabilityBins(\n  rows: GradedRow[],\n  bins = 10,\n  scheme: BinScheme = \"width\",\n): ReliabilityBin[] {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return []\n\n  const sorted = [...labeled].sort((a, b) => a.confidence - b.confidence)\n  const groups: GradedRow[][] =\n    scheme === \"mass\" ? equalMassGroups(sorted, bins) : equalWidthGroups(sorted, bins)\n\n  return groups\n    .filter((g) => g.length > 0)\n    .map((group) => {\n      const confidences = group.map((r) => r.confidence)\n      return {\n        lo: Math.min(...confidences),\n        hi: Math.max(...confidences),\n        n: group.length,\n        meanConfidence: mean(confidences),\n        accuracy: mean(group.map((r) => (r.predicted === r.truth ? 1 : 0))),\n      }\n    })\n}\n\n/** Expected calibration error: the mass-weighted gap between how confident\n *  the model was and how often it was right. Zero is perfect. */\nexport function ece(rows: GradedRow[], bins = 10, scheme: BinScheme = \"width\"): number {\n  const grouped = reliabilityBins(rows, bins, scheme)\n  const total = grouped.reduce((sum, b) => sum + b.n, 0)\n  if (total === 0) return 0\n  return grouped.reduce(\n    (sum, b) => sum + (b.n / total) * Math.abs(b.accuracy - b.meanConfidence),\n    0,\n  )\n}\n\n/** The worst single bin. ECE can look fine while one region is badly wrong. */\nexport function mce(rows: GradedRow[], bins = 10, scheme: BinScheme = \"width\"): number {\n  const grouped = reliabilityBins(rows, bins, scheme)\n  if (grouped.length === 0) return 0\n  return Math.max(...grouped.map((b) => Math.abs(b.accuracy - b.meanConfidence)))\n}\n\n/** Multiclass Brier score. A proper scoring rule, so unlike ECE it cannot\n *  be gamed by a model that reports a constant confidence. Lower is better. */\nexport function brier(rows: GradedRow[]): number {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return 0\n  return mean(\n    labeled.map((row) => {\n      let sum = 0\n      for (const [label, p] of Object.entries(row.probabilities)) {\n        const actual = label === row.truth ? 1 : 0\n        sum += (p - actual) ** 2\n      }\n      return sum\n    }),\n  )\n}\n\nexport function logLoss(rows: GradedRow[], epsilon = 1e-12): number {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return 0\n  return mean(\n    labeled.map((row) => {\n      const p = row.probabilities[row.truth!] ?? 0\n      return -Math.log(Math.max(p, epsilon))\n    }),\n  )\n}\n\nexport interface CoveragePoint {\n  threshold: number\n  /** Fraction of rows at or above the threshold. */\n  coverage: number\n  /** Accuracy among those rows. */\n  accuracy: number\n  n: number\n}\n\n/** The curve a call site actually reads: \"if I only act above confidence X,\n *  how often am I right, and how much traffic do I still handle?\" */\nexport function coverageCurve(rows: GradedRow[], steps = 20): CoveragePoint[] {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const total = labeled.length\n  if (total === 0) return []\n\n  const points: CoveragePoint[] = []\n  for (let i = 0; i <= steps; i++) {\n    const threshold = i / steps\n    const kept = labeled.filter((r) => r.confidence >= threshold)\n    points.push({\n      threshold,\n      coverage: kept.length / total,\n      accuracy: kept.length > 0 ? mean(kept.map((r) => (r.predicted === r.truth ? 1 : 0))) : 0,\n      n: kept.length,\n    })\n  }\n  return points\n}\n\n/** Sharpen or soften a distribution: p^(1/T), renormalized. T > 1 softens\n *  (the fix for overconfidence), T < 1 sharpens. */\nexport function applyTemperature(\n  probs: Record<string, number>,\n  temperature: number,\n): Record<string, number> {\n  const t = Math.max(temperature, 1e-6)\n  const scaled: Record<string, number> = {}\n  let sum = 0\n  for (const [label, p] of Object.entries(probs)) {\n    const v = Math.pow(Math.max(p, 0), 1 / t)\n    scaled[label] = v\n    sum += v\n  }\n  if (sum <= 0) {\n    const labels = Object.keys(probs)\n    for (const label of labels) scaled[label] = 1 / labels.length\n    return scaled\n  }\n  for (const label of Object.keys(scaled)) scaled[label] = scaled[label] / sum\n  return scaled\n}\n\nexport interface TemperatureFit {\n  temperature: number\n  nll: number\n  nllBefore: number\n}\n\n/** Fit the single parameter that turns a well-formed distribution into a\n *  calibrated one ON THIS TRAFFIC. This, not the backend, is where\n *  calibration comes from — which is why an uncalibrated backend plus a few\n *  hundred labeled rows is a defensible position and an uncalibrated\n *  backend alone is not.\n *\n *  Coarse grid then local refinement: one parameter, a smooth convex-ish\n *  objective, and a few hundred rows. Anything fancier is false precision. */\nexport function fitTemperature(rows: GradedRow[]): TemperatureFit {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const nllBefore = logLoss(labeled)\n  if (labeled.length === 0) return { temperature: 1, nll: 0, nllBefore: 0 }\n\n  const nllAt = (t: number) =>\n    mean(\n      labeled.map((row) => {\n        const scaled = applyTemperature(row.probabilities, t)\n        return -Math.log(Math.max(scaled[row.truth!] ?? 0, 1e-12))\n      }),\n    )\n\n  let best = 1\n  let bestNll = nllAt(1)\n  for (let t = 0.1; t <= 10.0001; t += 0.1) {\n    const value = nllAt(t)\n    if (value < bestNll) {\n      bestNll = value\n      best = t\n    }\n  }\n  for (let t = Math.max(0.01, best - 0.1); t <= best + 0.1; t += 0.01) {\n    const value = nllAt(t)\n    if (value < bestNll) {\n      bestNll = value\n      best = t\n    }\n  }\n\n  return { temperature: Number(best.toFixed(3)), nll: bestNll, nllBefore }\n}\n\nexport interface Interval {\n  lo: number\n  hi: number\n}\n\n/** Percentile bootstrap. Without an interval, a difference between two\n *  backends' ECEs is not a result. */\nexport function bootstrapInterval(\n  rows: GradedRow[],\n  statistic: (sample: GradedRow[]) => number,\n  resamples = 1000,\n  alpha = 0.05,\n  rng: () => number = Math.random,\n): Interval {\n  if (rows.length === 0) return { lo: 0, hi: 0 }\n  const values: number[] = []\n  for (let i = 0; i < resamples; i++) {\n    const sample: GradedRow[] = []\n    for (let j = 0; j < rows.length; j++) {\n      sample.push(rows[Math.floor(rng() * rows.length)])\n    }\n    values.push(statistic(sample))\n  }\n  values.sort((a, b) => a - b)\n  return {\n    lo: values[Math.floor((alpha / 2) * values.length)],\n    hi: values[Math.min(values.length - 1, Math.floor((1 - alpha / 2) * values.length))],\n  }\n}\n\nexport interface CalibrationReport {\n  insufficient: boolean\n  n: number\n  nLabeled: number\n  minN: number\n  agreement: AgreementResult\n  accuracy?: number\n  ece?: number\n  eceEqualMass?: number\n  eceInterval?: Interval\n  mce?: number\n  brier?: number\n  logLoss?: number\n  bins?: ReliabilityBin[]\n  binsEqualMass?: ReliabilityBin[]\n  coverage?: CoveragePoint[]\n  temperature?: TemperatureFit\n}\n\nexport function calibrationReport(\n  rows: GradedRow[],\n  opts: { minN?: number; bins?: number; bootstrap?: number; rng?: () => number } = {},\n): CalibrationReport {\n  const minN = opts.minN ?? 100\n  const bins = opts.bins ?? 10\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const base: CalibrationReport = {\n    insufficient: labeled.length < minN,\n    n: rows.length,\n    nLabeled: labeled.length,\n    minN,\n    agreement: agreement(rows),\n  }\n\n  // Agreement needs no ground truth, so it is always reported. Everything\n  // that does need ground truth is withheld until there is enough of it.\n  if (base.insufficient) return base\n\n  return {\n    ...base,\n    accuracy: mean(labeled.map((r) => (r.predicted === r.truth ? 1 : 0))),\n    ece: ece(labeled, bins, \"width\"),\n    eceEqualMass: ece(labeled, bins, \"mass\"),\n    eceInterval: bootstrapInterval(\n      labeled,\n      (sample) => ece(sample, bins, \"width\"),\n      opts.bootstrap ?? 1000,\n      0.05,\n      opts.rng,\n    ),\n    mce: mce(labeled, bins, \"width\"),\n    brier: brier(labeled),\n    logLoss: logLoss(labeled),\n    bins: reliabilityBins(labeled, bins, \"width\"),\n    binsEqualMass: reliabilityBins(labeled, bins, \"mass\"),\n    coverage: coverageCurve(labeled),\n    temperature: fitTemperature(labeled),\n  }\n}\n\nfunction equalWidthGroups(sorted: GradedRow[], bins: number): GradedRow[][] {\n  const groups: GradedRow[][] = Array.from({ length: bins }, () => [])\n  for (const row of sorted) {\n    const index = Math.min(bins - 1, Math.max(0, Math.floor(row.confidence * bins)))\n    groups[index].push(row)\n  }\n  return groups\n}\n\nfunction equalMassGroups(sorted: GradedRow[], bins: number): GradedRow[][] {\n  const groups: GradedRow[][] = Array.from({ length: bins }, () => [])\n  sorted.forEach((row, i) => {\n    groups[Math.min(bins - 1, Math.floor((i * bins) / sorted.length))].push(row)\n  })\n  return groups\n}\n\nfunction mean(values: number[]): number {\n  if (values.length === 0) return 0\n  return values.reduce((a, b) => a + b, 0) / values.length\n}\n"],"mappings":"wCAgCO,SAASA,EAAUC,EAAoC,CAC5D,IAAMC,EAASD,EAAK,OAAQE,GAAMA,EAAE,YAAc,MAAS,EACrDC,EAAIF,EAAO,OACjB,GAAIE,IAAM,EAAG,MAAO,CAAE,EAAG,EAAG,MAAO,EAAG,KAAM,EAAG,MAAO,CAAE,EAExD,IAAIC,EAAQ,EACNC,EAAY,IAAI,IAChBC,EAAY,IAAI,IACtB,QAAWC,KAAON,EACZM,EAAI,YAAcA,EAAI,WAAWH,IACrCC,EAAU,IAAIE,EAAI,WAAYF,EAAU,IAAIE,EAAI,SAAS,GAAK,GAAK,CAAC,EACpED,EAAU,IAAIC,EAAI,WAAaD,EAAU,IAAIC,EAAI,SAAU,GAAK,GAAK,CAAC,EAGxE,IAAMC,EAAWJ,EAAQD,EACrBM,EAAW,EACf,OAAW,CAACC,EAAOC,CAAK,IAAKN,EAC3BI,GAAaE,EAAQR,IAAOG,EAAU,IAAII,CAAK,GAAK,GAAKP,GAE3D,IAAMS,EAAQH,EAAW,GAAKD,EAAWC,IAAa,EAAIA,GAAY,EAEtE,MAAO,CAAE,EAAAN,EAAG,MAAAC,EAAO,KAAMI,EAAU,MAAAI,CAAM,CAC3C,CAYO,SAASC,EACdb,EACAc,EAAO,GACPC,EAAoB,QACF,CAClB,IAAMC,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,GAAIc,EAAQ,SAAW,EAAG,MAAO,CAAC,EAElC,IAAMC,EAAS,CAAC,GAAGD,CAAO,EAAE,KAAK,CAACE,EAAGC,IAAMD,EAAE,WAAaC,EAAE,UAAU,EAItE,OAFEJ,IAAW,OAASK,EAAgBH,EAAQH,CAAI,EAAIO,EAAiBJ,EAAQH,CAAI,GAGhF,OAAQQ,GAAMA,EAAE,OAAS,CAAC,EAC1B,IAAKC,GAAU,CACd,IAAMC,EAAcD,EAAM,IAAKrB,GAAMA,EAAE,UAAU,EACjD,MAAO,CACL,GAAI,KAAK,IAAI,GAAGsB,CAAW,EAC3B,GAAI,KAAK,IAAI,GAAGA,CAAW,EAC3B,EAAGD,EAAM,OACT,eAAgBE,EAAKD,CAAW,EAChC,SAAUC,EAAKF,EAAM,IAAKrB,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,CACpE,CACF,CAAC,CACL,CAIO,SAASwB,EAAI1B,EAAmBc,EAAO,GAAIC,EAAoB,QAAiB,CACrF,IAAMY,EAAUd,EAAgBb,EAAMc,EAAMC,CAAM,EAC5Ca,EAAQD,EAAQ,OAAO,CAACE,EAAKV,IAAMU,EAAMV,EAAE,EAAG,CAAC,EACrD,OAAIS,IAAU,EAAU,EACjBD,EAAQ,OACb,CAACE,EAAKV,IAAMU,EAAOV,EAAE,EAAIS,EAAS,KAAK,IAAIT,EAAE,SAAWA,EAAE,cAAc,EACxE,CACF,CACF,CAGO,SAASW,EAAI9B,EAAmBc,EAAO,GAAIC,EAAoB,QAAiB,CACrF,IAAMY,EAAUd,EAAgBb,EAAMc,EAAMC,CAAM,EAClD,OAAIY,EAAQ,SAAW,EAAU,EAC1B,KAAK,IAAI,GAAGA,EAAQ,IAAKR,GAAM,KAAK,IAAIA,EAAE,SAAWA,EAAE,cAAc,CAAC,CAAC,CAChF,CAIO,SAASY,EAAM/B,EAA2B,CAC/C,IAAMgB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,OAAIc,EAAQ,SAAW,EAAU,EAC1BS,EACLT,EAAQ,IAAKT,GAAQ,CACnB,IAAIsB,EAAM,EACV,OAAW,CAACnB,EAAOsB,CAAC,IAAK,OAAO,QAAQzB,EAAI,aAAa,EAAG,CAC1D,IAAM0B,EAASvB,IAAUH,EAAI,MAAQ,EAAI,EACzCsB,IAAQG,EAAIC,IAAW,EAEzB,OAAOJ,CACT,CAAC,CACH,CACF,CAEO,SAASK,EAAQlC,EAAmBmC,EAAU,MAAe,CAClE,IAAMnB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,OAAIc,EAAQ,SAAW,EAAU,EAC1BS,EACLT,EAAQ,IAAKT,GAAQ,CACnB,IAAMyB,EAAIzB,EAAI,cAAcA,EAAI,KAAM,GAAK,EAC3C,MAAO,CAAC,KAAK,IAAI,KAAK,IAAIyB,EAAGG,CAAO,CAAC,CACvC,CAAC,CACH,CACF,CAaO,SAASC,EAAcpC,EAAmBqC,EAAQ,GAAqB,CAC5E,IAAMrB,EAAUhB,EAAK,OAAQ,GAAM,EAAE,QAAU,MAAS,EAClD4B,EAAQZ,EAAQ,OACtB,GAAIY,IAAU,EAAG,MAAO,CAAC,EAEzB,IAAMU,EAA0B,CAAC,EACjC,QAASC,EAAI,EAAGA,GAAKF,EAAOE,IAAK,CAC/B,IAAMC,EAAYD,EAAIF,EAChBI,EAAOzB,EAAQ,OAAQd,GAAMA,EAAE,YAAcsC,CAAS,EAC5DF,EAAO,KAAK,CACV,UAAAE,EACA,SAAUC,EAAK,OAASb,EACxB,SAAUa,EAAK,OAAS,EAAIhB,EAAKgB,EAAK,IAAKvC,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,EAAI,EACvF,EAAGuC,EAAK,MACV,CAAC,EAEH,OAAOH,CACT,CAIO,SAASI,EACdC,EACAC,EACwB,CACxB,IAAMC,EAAI,KAAK,IAAID,EAAa,IAAI,EAC9BE,EAAiC,CAAC,EACpCjB,EAAM,EACV,OAAW,CAACnB,EAAOsB,CAAC,IAAK,OAAO,QAAQW,CAAK,EAAG,CAC9C,IAAMI,EAAI,KAAK,IAAI,KAAK,IAAIf,EAAG,CAAC,EAAG,EAAIa,CAAC,EACxCC,EAAOpC,CAAK,EAAIqC,EAChBlB,GAAOkB,EAET,GAAIlB,GAAO,EAAG,CACZ,IAAMmB,EAAS,OAAO,KAAKL,CAAK,EAChC,QAAWjC,KAASsC,EAAQF,EAAOpC,CAAK,EAAI,EAAIsC,EAAO,OACvD,OAAOF,EAET,QAAWpC,KAAS,OAAO,KAAKoC,CAAM,EAAGA,EAAOpC,CAAK,EAAIoC,EAAOpC,CAAK,EAAImB,EACzE,OAAOiB,CACT,CAgBO,SAASG,EAAejD,EAAmC,CAChE,IAAMgB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EAClDgD,EAAYhB,EAAQlB,CAAO,EACjC,GAAIA,EAAQ,SAAW,EAAG,MAAO,CAAE,YAAa,EAAG,IAAK,EAAG,UAAW,CAAE,EAExE,IAAMmC,EAASN,GACbpB,EACET,EAAQ,IAAKT,GAAQ,CACnB,IAAMuC,EAASJ,EAAiBnC,EAAI,cAAesC,CAAC,EACpD,MAAO,CAAC,KAAK,IAAI,KAAK,IAAIC,EAAOvC,EAAI,KAAM,GAAK,EAAG,KAAK,CAAC,CAC3D,CAAC,CACH,EAEE6C,EAAO,EACPC,EAAUF,EAAM,CAAC,EACrB,QAASN,EAAI,GAAKA,GAAK,QAASA,GAAK,GAAK,CACxC,IAAMS,EAAQH,EAAMN,CAAC,EACjBS,EAAQD,IACVA,EAAUC,EACVF,EAAOP,GAGX,QAASA,EAAI,KAAK,IAAI,IAAMO,EAAO,EAAG,EAAGP,GAAKO,EAAO,GAAKP,GAAK,IAAM,CACnE,IAAMS,EAAQH,EAAMN,CAAC,EACjBS,EAAQD,IACVA,EAAUC,EACVF,EAAOP,GAIX,MAAO,CAAE,YAAa,OAAOO,EAAK,QAAQ,CAAC,CAAC,EAAG,IAAKC,EAAS,UAAAH,CAAU,CACzE,CASO,SAASK,EACdvD,EACAwD,EACAC,EAAY,IACZC,EAAQ,IACRC,EAAoB,KAAK,OACf,CACV,GAAI3D,EAAK,SAAW,EAAG,MAAO,CAAE,GAAI,EAAG,GAAI,CAAE,EAC7C,IAAM4D,EAAmB,CAAC,EAC1B,QAASrB,EAAI,EAAGA,EAAIkB,EAAWlB,IAAK,CAClC,IAAMsB,EAAsB,CAAC,EAC7B,QAASC,EAAI,EAAGA,EAAI9D,EAAK,OAAQ8D,IAC/BD,EAAO,KAAK7D,EAAK,KAAK,MAAM2D,EAAI,EAAI3D,EAAK,MAAM,CAAC,CAAC,EAEnD4D,EAAO,KAAKJ,EAAUK,CAAM,CAAC,EAE/B,OAAAD,EAAO,KAAK,CAAC1C,EAAGC,IAAMD,EAAIC,CAAC,EACpB,CACL,GAAIyC,EAAO,KAAK,MAAOF,EAAQ,EAAKE,EAAO,MAAM,CAAC,EAClD,GAAIA,EAAO,KAAK,IAAIA,EAAO,OAAS,EAAG,KAAK,OAAO,EAAIF,EAAQ,GAAKE,EAAO,MAAM,CAAC,CAAC,CACrF,CACF,CAqBO,SAASG,EACd/D,EACAgE,EAAiF,CAAC,EAC/D,CACnB,IAAMC,EAAOD,EAAK,MAAQ,IACpBlD,EAAOkD,EAAK,MAAQ,GACpBhD,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EAClDgE,EAA0B,CAC9B,aAAclD,EAAQ,OAASiD,EAC/B,EAAGjE,EAAK,OACR,SAAUgB,EAAQ,OAClB,KAAAiD,EACA,UAAWlE,EAAUC,CAAI,CAC3B,EAIA,OAAIkE,EAAK,aAAqBA,EAEvB,CACL,GAAGA,EACH,SAAUzC,EAAKT,EAAQ,IAAKd,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,EACpE,IAAKwB,EAAIV,EAASF,EAAM,OAAO,EAC/B,aAAcY,EAAIV,EAASF,EAAM,MAAM,EACvC,YAAayC,EACXvC,EACC6C,GAAWnC,EAAImC,EAAQ/C,EAAM,OAAO,EACrCkD,EAAK,WAAa,IAClB,IACAA,EAAK,GACP,EACA,IAAKlC,EAAId,EAASF,EAAM,OAAO,EAC/B,MAAOiB,EAAMf,CAAO,EACpB,QAASkB,EAAQlB,CAAO,EACxB,KAAMH,EAAgBG,EAASF,EAAM,OAAO,EAC5C,cAAeD,EAAgBG,EAASF,EAAM,MAAM,EACpD,SAAUsB,EAAcpB,CAAO,EAC/B,YAAaiC,EAAejC,CAAO,CACrC,CACF,CAEA,SAASK,EAAiBJ,EAAqBH,EAA6B,CAC1E,IAAMqD,EAAwB,MAAM,KAAK,CAAE,OAAQrD,CAAK,EAAG,IAAM,CAAC,CAAC,EACnE,QAAWP,KAAOU,EAAQ,CACxB,IAAMmD,EAAQ,KAAK,IAAItD,EAAO,EAAG,KAAK,IAAI,EAAG,KAAK,MAAMP,EAAI,WAAaO,CAAI,CAAC,CAAC,EAC/EqD,EAAOC,CAAK,EAAE,KAAK7D,CAAG,EAExB,OAAO4D,CACT,CAEA,SAAS/C,EAAgBH,EAAqBH,EAA6B,CACzE,IAAMqD,EAAwB,MAAM,KAAK,CAAE,OAAQrD,CAAK,EAAG,IAAM,CAAC,CAAC,EACnE,OAAAG,EAAO,QAAQ,CAACV,EAAKgC,IAAM,CACzB4B,EAAO,KAAK,IAAIrD,EAAO,EAAG,KAAK,MAAOyB,EAAIzB,EAAQG,EAAO,MAAM,CAAC,CAAC,EAAE,KAAKV,CAAG,CAC7E,CAAC,EACM4D,CACT,CAEA,SAAS1C,EAAKmC,EAA0B,CACtC,OAAIA,EAAO,SAAW,EAAU,EACzBA,EAAO,OAAO,CAAC1C,EAAGC,IAAMD,EAAIC,EAAG,CAAC,EAAIyC,EAAO,MACpD,CA7VA,IAAAS,EAAAC,EAAA","names":["agreement","rows","usable","r","n","agree","predicted","incumbent","row","observed","expected","label","count","kappa","reliabilityBins","bins","scheme","labeled","sorted","a","b","equalMassGroups","equalWidthGroups","g","group","confidences","mean","ece","grouped","total","sum","mce","brier","p","actual","logLoss","epsilon","coverageCurve","steps","points","i","threshold","kept","applyTemperature","probs","temperature","t","scaled","v","labels","fitTemperature","nllBefore","nllAt","best","bestNll","value","bootstrapInterval","statistic","resamples","alpha","rng","values","sample","j","calibrationReport","opts","minN","base","groups","index","init_calibration","__esmMin"]}