From c55e14b626730c9d88cc208d54aa3bfeee4f6989 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=94=D0=BC=D0=B8=D1=82=D1=80=D0=B8=D0=B9?=
 <noreply@anthropic.com>
Date: Sun, 31 May 2026 19:05:20 +0300
Subject: [PATCH] feat(brain): surface router-gate v4 signals into episode +
 factor axes

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
---
 docs/observer/brain-data-catalog.md           |  16 +-
 ...5-31-brain-factor-analysis-f-candidates.md | 624 ++++++++++++++++++
 tools/brain-retro-analyzer.mjs                |  12 +
 tools/brain-retro-analyzer.test.mjs           |  17 +
 tools/cost-aggregator.mjs                     |  10 +-
 tools/cost-aggregator.test.mjs                |  10 +-
 tools/cost-pricing.mjs                        |   6 +
 tools/cost-pricing.test.mjs                   |   8 +
 tools/enforce-hook-helpers.mjs                |  46 ++
 tools/enforce-hook-helpers.test.mjs           |  37 ++
 tools/enforce-llm-judge-per-tool.mjs          |   9 +-
 tools/enforce-llm-judge-per-tool.test.mjs     |  17 +
 tools/enforce-safe-baseline-metering.mjs      |   2 +
 tools/enforce-safe-baseline-metering.test.mjs |  15 +
 tools/observer-transcript-parser.mjs          |  17 +-
 tools/observer-transcript-parser.test.mjs     |  26 +-
 tools/observer-v4-signals.mjs                 |  74 +++
 tools/observer-v4-signals.test.mjs            |  49 ++
 18 files changed, 979 insertions(+), 16 deletions(-)
 create mode 100644 docs/superpowers/plans/2026-05-31-brain-factor-analysis-f-candidates.md
 create mode 100644 tools/observer-v4-signals.mjs
 create mode 100644 tools/observer-v4-signals.test.mjs

diff --git a/docs/observer/brain-data-catalog.md b/docs/observer/brain-data-catalog.md
index d53d9dcd..0705e9d7 100644
--- a/docs/observer/brain-data-catalog.md
+++ b/docs/observer/brain-data-catalog.md
@@ -6,7 +6,7 @@
 
 ---
 
-## A. Эпизод журнала — `docs/observer/episodes-YYYY-MM.jsonl` (схема v4.3)
+## A. Эпизод журнала — `docs/observer/episodes-YYYY-MM.jsonl` (схема v4.4, schema_minor 4)
 
 Один эпизод = один цикл «промпт заказчика → ответ Claude». Append-only, по строке на эпизод. ПДн вырезаются до записи (`observer-pii-filter.mjs`). Сборка — `observer-transcript-parser.mjs:888` (`parseTranscript`).
 
@@ -53,7 +53,7 @@
 
 ### A.7. Стоимость и токены (`task_cost`)
 `observer-transcript-parser.mjs:472`. Базовые: `input_tokens` · `output_tokens` · `cache_read_input_tokens` · `cache_creation_input_tokens` · `web_search_requests` · `web_fetch_requests` · `iterations` (детектор extended-thinking).
-Слой LLM-агентов (дозаполняется): `classifier_input_tokens` · `classifier_output_tokens` · `self_assessment_input_tokens` · `self_assessment_output_tokens` · `reviewer_input_tokens` · `reviewer_output_tokens` · `reviewer_subagent_usd` · `reviewer_direct_fallback_usd`.
+Слой LLM-агентов (дозаполняется): `classifier_input_tokens` · `classifier_output_tokens` · `self_assessment_input_tokens` · `self_assessment_output_tokens` · `reviewer_input_tokens` · `reviewer_output_tokens` · `reviewer_subagent_usd` · `reviewer_direct_fallback_usd` · `judge_spend_usd` (✅ NEW — `judge_calls × JUDGE_PER_CALL_USD`).
 
 ### A.8. Мета (`task_meta`)
 `prompt_length_chars` · `mcp_servers_used[]` · `file_type_distribution` по корзинам `src / test / config / spec / norm / data / other` (`classifyFilePath` — `observer-transcript-parser.mjs:358`).
@@ -69,6 +69,9 @@
 ### A.11. События (`events[]`) — 11 видов
 `skill_invoked` (skill) · `tool_summary` (counts) · `error` (tool, summary) · `hook_fired` (counts, scripts, errors) · `interrupt` · `retry` · `time_burn` (ход > 15 мин) · `parse_gap` (> 10% битых строк) · `unrecovered_error` (ход кончился на ошибке) · `ask_user_question` (question_count, answer_kind: `option`/`custom`/`no_answer`) · `subagent_invoked` (subagent_type, model, description).
 
+### A.12. Сигналы защит router-gate v4 (`v4_signals`) — ✅ NEW (schema_minor 4)
+Поднимаются в эпизод ридером `observer-v4-signals.mjs` по окну хода `[started_at, ended_at]`: `rationalization_flag_count` (число пойманных самооправданий в окне) · `judge_verdict` (`YES`/`NO`/`block`/`null` — последний вердикт судьи в окне) · `safe_baseline_action` (`allow`/`soft_flag`/`hard_block`/`null` — худшее действие safe-baseline в окне) · `judge_calls` (кумулятивно за сессию из бюджета судьи).
+
 ---
 
 ## B. Факторные оси — `FACTOR_FNS` (`brain-retro-analyzer.mjs:326`) — 27 осей + `chain_ref`
@@ -85,6 +88,8 @@
 
 **Pass 4 (семантика):** `similar_past_outcome_majority` (исход большинства соседей по смысловому вектору; `no_neighbors` если вектора нет).
 
+**Pass 5 (✅ NEW — сигналы router-gate v4):** `rationalization_flag_count` (0/1/2+) · `judge_verdict` (YES/NO/block/null) · `safe_baseline_action` (allow/soft_flag/hard_block/null) · `judge_calls_bucket` (0 / 1-9 / 10+).
+
 **Отдельно:** `chain_ref` — мульти-значный (эпизод считается по разу на каждую цепочку).
 
 ---
@@ -133,7 +138,9 @@ Sonnet — $3/Mtok вход, $15/Mtok выход. Opus — $15/Mtok вход, $7
 
 ---
 
-## F. Сигналы новых защит router-gate v4 (в эпизоде ещё НЕ агрегируются — кандидаты в факторный анализ)
+## F. Сигналы новых защит router-gate v4
+
+> **Обновление 31.05.2026:** F.1–F.3 ✅ **заведены в эпизод** (`v4_signals`, см. A.12) и в факторный анализ (Pass 5, раздел B) через ридер `observer-v4-signals.mjs`. F.4/F.5 — пока только на диске.
 
 ### F.1. Rationalization-audit (`enforce-rationalization-audit.mjs`)
 `~/.claude/runtime/rationalization-flags-<session>.jsonl`. Строка: `{kind, evidence}`. Виды (`kind`): `rationalization-phrase` (фраза-самооправдание) · `prod-edit-without-test` (правка прод-кода без теста в том же ходе) · `weak-commit-message` (commit с сообщением < 12 симв.). Блокирует на 3-м флаге за сессию (`decide:112`). Перед сопоставлением снимает цитаты/код (`stripQuotedContext:51`).
@@ -153,7 +160,8 @@ Sonnet — $3/Mtok вход, $15/Mtok выход. Opus — $15/Mtok вход, $7
 ---
 
 ## Кандидаты на расширение факторного анализа
-Сигналы из раздела F пишутся на диск, но **не сводятся в эпизод** и не попадают в `FACTOR_FNS`. Естественные новые оси:
+
+✅ **Реализовано 31.05.2026** (план `docs/superpowers/plans/2026-05-31-brain-factor-analysis-f-candidates.md`): следующие 4 оси заведены в эпизод (`v4_signals`) и в `FACTOR_FNS` (Pass 5):
 - `rationalization_flag_count` (F.1) — число самооправданий за ход/сессию.
 - `safe_baseline_action` (F.2) — allow / soft_flag / hard_block.
 - `judge_verdict` (F.3) — YES / block / disabled (когда Layer 4 активен).
diff --git a/docs/superpowers/plans/2026-05-31-brain-factor-analysis-f-candidates.md b/docs/superpowers/plans/2026-05-31-brain-factor-analysis-f-candidates.md
new file mode 100644
index 00000000..e605416e
--- /dev/null
+++ b/docs/superpowers/plans/2026-05-31-brain-factor-analysis-f-candidates.md
@@ -0,0 +1,624 @@
+# Расширение факторного анализа «мозга» — кандидаты раздела F
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Поднять сигналы новых защит router-gate v4 (rationalization-флаги, safe-baseline action, вердикт LLM-судьи, активность судьи) в эпизод журнала наблюдателя и добавить их как новые оси факторного анализа.
+
+**Architecture:** Сигналы уже пишутся хуками на диск в `~/.claude/runtime/`, но (1) часть не логируется в нужной форме — добавляем точечную инструментовку (verdict-log, action-log); (2) парсер транскрипта их не читает — добавляем чистый модуль-ридер `observer-v4-signals.mjs`, который `parseTranscript` зовёт и кладёт результат в новые поля эпизода; (3) анализатор не раскладывает их — добавляем 4 оси в `FACTOR_FNS`. Принцип: эпизод остаётся единственным входом факторной матрицы (хуки → runtime-файлы → парсер → эпизод → `FACTOR_FNS`).
+
+**Tech Stack:** Node ESM (`.mjs`), vitest (tools-only). Деньги — `bcmath`-free, float USD как в `cost-aggregator.mjs`.
+
+**Канонический прогон тестов tools** (per memory `feedback_vitest_sentinel_recipe`): `npm run test:tools` сломан keytar-установкой в `app/node_modules` → запускать `npx vitest run --root app --config vitest.config.tools.mjs <путь>`.
+
+**Базовые факты, на которых стоит план (сверены по коду 31.05.2026):**
+- Эпизод собирается в `tools/observer-transcript-parser.mjs:888` (`parseTranscript`), окно хода — `turn = entries.slice(start)`, `started_at`/`ended_at` из `timestamps`.
+- rationalization-флаги: `tools/enforce-hook-helpers.mjs:280` пишет `{ts, kind, evidence}` в `rationalization-flags-<sess>.jsonl`; читает `:291`.
+- safe-baseline: `tools/enforce-safe-baseline-metering.mjs` пишет ledger `safe-baseline-ledger-<sess>.json` (`saveLedger:180`) и логирует ТОЛЬКО `soft_flag` в `safe-baseline-flags-<sess>.jsonl` (`logFlag:186`, строка `{ts, tool, reason}`); `hard_block` нигде не логируется. Ledger-state (`newCounterState` `safe-baseline-metering.mjs:60`) содержит `counts`/`skill_match_within_task`, но `warnings_issued`/`hard_blocks_issued` живой `runLiveDecision` НЕ заполняет.
+- LLM-судья: `tools/enforce-llm-judge-per-tool.mjs:112` `runPerTool` зовёт `decide()` → вердикт `{block, verdict, degraded}`; бюджет `tools/llm-judge.mjs:121` `readJudgeBudget`/`:127` `bumpJudgeBudget` хранит `{calls}` в `llm-judge-budget-<sess>.json`; вердикт по-инструментно НЕ логируется. `JUDGE_SESSION_BUDGET = 200` (`llm-judge.mjs:67`).
+- Цена: `tools/cost-pricing.mjs` `PRICING.sonnet = {input:3e-6, output:15e-6}`; агрегатор `tools/cost-aggregator.mjs:13` `episodeUsd` / `:40` `aggregateDay` (5 компонент + `total_usd` + `episode_count`); Stop-хук `tools/cost-stop-hook.mjs` пишет `cost-daily.json`.
+- Факторные оси: `tools/brain-retro-analyzer.mjs:326` `FACTOR_FNS`; `analyze()` `:657` штампует `episode._inferredOutcome` и строит матрицу `buildFactorMatrix:384`.
+
+---
+
+## File Structure
+
+| Файл | Ответственность | Действие |
+|---|---|---|
+| `tools/enforce-llm-judge-per-tool.mjs` | вызов судьи per-tool | Modify — логировать вердикт |
+| `tools/enforce-hook-helpers.mjs` | общий I/O хуков | Modify — `logJudgeVerdict` + `readJudgeVerdicts` + `logSafeBaselineAction` + `readSafeBaselineActions` |
+| `tools/enforce-safe-baseline-metering.mjs` | safe-baseline метеринг | Modify — логировать ЛЮБОЙ action (incl hard_block) |
+| `tools/observer-v4-signals.mjs` | **новый** чистый ридер runtime-сигналов в окне хода | Create |
+| `tools/observer-transcript-parser.mjs` | сборка эпизода | Modify — позвать ридер, добавить поля эпизода |
+| `tools/cost-pricing.mjs` | ставки | Modify — `JUDGE_PER_CALL_USD` |
+| `tools/cost-aggregator.mjs` | стоимость per-episode/day | Modify — компонент `judge_spend_usd` |
+| `tools/brain-retro-analyzer.mjs` | факторные оси | Modify — 4 новые `FACTOR_FNS` |
+| `docs/observer/brain-data-catalog.md` | каталог данных мозга | Modify — перенести сигналы из F в A/B/E |
+| `tools/*.test.mjs` | тесты | Create/Modify per task |
+
+**Новые поля эпизода (schema_minor bump 3 → 4):**
+- `task_cost.judge_spend_usd` (number, default 0)
+- `v4_signals.rationalization_flag_count` (int)
+- `v4_signals.safe_baseline_action` (`allow`/`soft_flag`/`hard_block`/`null`)
+- `v4_signals.judge_verdict` (`YES`/`NO`/`block`/`disabled`/`null`)
+- `v4_signals.judge_calls` (int)
+
+---
+
+## Task 1: Логирование вердикта судьи
+
+**Files:**
+- Modify: `tools/enforce-hook-helpers.mjs` (после `readRationalizationFlags`, ~:299)
+- Modify: `tools/enforce-llm-judge-per-tool.mjs` (`runPerTool`, :112)
+- Test: `tools/enforce-hook-helpers.test.mjs`
+
+- [ ] **Step 1: Написать падающий тест на `logJudgeVerdict`/`readJudgeVerdicts`**
+
+```js
+// tools/enforce-hook-helpers.test.mjs — добавить describe-блок
+import { tmpdir } from 'node:os';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { join } from 'node:path';
+import { logJudgeVerdict, readJudgeVerdicts } from './enforce-hook-helpers.mjs';
+
+describe('judge verdict log', () => {
+  it('appends and reads back verdict records', () => {
+    const dir = mkdtempSync(join(tmpdir(), 'jv-'));
+    logJudgeVerdict('sess1', { tool: 'Edit', verdict: 'YES' }, { baseDir: dir });
+    logJudgeVerdict('sess1', { tool: 'Bash', verdict: 'NO' }, { baseDir: dir });
+    const recs = readJudgeVerdicts('sess1', { baseDir: dir });
+    expect(recs.length).toBe(2);
+    expect(recs[0]).toMatchObject({ tool: 'Edit', verdict: 'YES' });
+    expect(typeof recs[0].ts).toBe('string');
+    rmSync(dir, { recursive: true, force: true });
+  });
+  it('returns [] when no file', () => {
+    expect(readJudgeVerdicts('nope', { baseDir: tmpdir() })).toEqual([]);
+  });
+});
+```
+
+- [ ] **Step 2: Прогнать — убедиться, что падает**
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/enforce-hook-helpers.test.mjs`
+Expected: FAIL — `logJudgeVerdict is not a function`.
+
+- [ ] **Step 3: Реализовать функции в `enforce-hook-helpers.mjs`** (вставить после `readRationalizationFlags`, перед `readRouterState`)
+
+```js
+export function logJudgeVerdict(sessionId, { tool, verdict }, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const f = join(dir, `llm-judge-verdicts-${sessionId || 'unknown'}.jsonl`);
+    appendFileSync(f, JSON.stringify({
+      ts: new Date().toISOString(),
+      tool: tool || 'unknown',
+      verdict: verdict == null ? null : String(verdict),
+    }) + '\n');
+  } catch { /* ignore */ }
+}
+
+export function readJudgeVerdicts(sessionId, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const f = join(dir, `llm-judge-verdicts-${sessionId || 'unknown'}.jsonl`);
+    if (!existsSync(f)) return [];
+    return readFileSync(f, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+```
+
+(`runtimeDir`, `appendFileSync`, `readFileSync`, `existsSync`, `join` уже импортированы в файле — проверить шапку, при отсутствии `existsSync` добавить в существующий `import { ... } from 'node:fs'`.)
+
+- [ ] **Step 4: Прогнать тест — PASS**
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/enforce-hook-helpers.test.mjs`
+Expected: PASS.
+
+- [ ] **Step 5: Вызвать логгер из судьи.** В `enforce-llm-judge-per-tool.mjs:112` `runPerTool`, после строки `if (result.verdict !== undefined) bumpBudgetImpl({ sessionId, by: 1 });` добавить:
+
+```js
+    if (result.verdict !== undefined) {
+      bumpBudgetImpl({ sessionId, by: 1 });
+      logVerdictImpl({ sessionId, tool: event && event.tool_name, verdict: result.block ? 'block' : (result.verdict ?? 'YES') });
+    }
+```
+
+В сигнатуру `runPerTool({...})` добавить параметр `logVerdictImpl = () => {}` (no-op по умолчанию — тесты не пишут на диск). В `main()` (:153) передать `logVerdictImpl: ({ sessionId, tool, verdict }) => logJudgeVerdict(sessionId, { tool, verdict })` и добавить `logJudgeVerdict` в импорт из `./enforce-hook-helpers.mjs`.
+
+- [ ] **Step 6: Тест на проброс `logVerdictImpl`**
+
+```js
+// tools/enforce-llm-judge-per-tool.test.mjs — добавить
+it('runPerTool logs verdict only when a real judge call happened', async () => {
+  const logged = [];
+  const r = await runPerTool({
+    event: { tool_name: 'Edit', session_id: 's', tool_input: {} },
+    judgeConfig: { enabled: true, apiKey: 'k' },
+    readDeclaredTaskImpl: () => ({ task_summary: 't' }),
+    readBudgetImpl: () => 0,
+    bumpBudgetImpl: () => {},
+    llmJudgeCallImpl: async () => 'YES',
+    logVerdictImpl: (rec) => logged.push(rec),
+  });
+  expect(r.block).toBe(false);
+  expect(logged).toEqual([{ sessionId: 's', tool: 'Edit', verdict: 'YES' }]);
+});
+```
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/enforce-llm-judge-per-tool.test.mjs`
+Expected: PASS.
+
+- [ ] **Step 7: Commit**
+
+```bash
+git add tools/enforce-hook-helpers.mjs tools/enforce-hook-helpers.test.mjs tools/enforce-llm-judge-per-tool.mjs tools/enforce-llm-judge-per-tool.test.mjs
+git commit -m "feat(brain): log per-tool judge verdicts to runtime jsonl"
+```
+
+---
+
+## Task 2: Логирование любого safe-baseline action (incl hard_block)
+
+**Files:**
+- Modify: `tools/enforce-hook-helpers.mjs`
+- Modify: `tools/enforce-safe-baseline-metering.mjs` (`runMain`, :195)
+- Test: `tools/enforce-hook-helpers.test.mjs`, `tools/enforce-safe-baseline-metering.test.mjs`
+
+- [ ] **Step 1: Тест на `logSafeBaselineAction`/`readSafeBaselineActions`**
+
+```js
+// tools/enforce-hook-helpers.test.mjs
+import { logSafeBaselineAction, readSafeBaselineActions } from './enforce-hook-helpers.mjs';
+describe('safe-baseline action log', () => {
+  it('appends and reads action records', () => {
+    const dir = mkdtempSync(join(tmpdir(), 'sb-'));
+    logSafeBaselineAction('s', { tool: 'Edit', action: 'hard_block' }, { baseDir: dir });
+    const recs = readSafeBaselineActions('s', { baseDir: dir });
+    expect(recs[0]).toMatchObject({ tool: 'Edit', action: 'hard_block' });
+    rmSync(dir, { recursive: true, force: true });
+  });
+});
+```
+
+- [ ] **Step 2: Прогнать — FAIL** (`logSafeBaselineAction is not a function`).
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/enforce-hook-helpers.test.mjs`
+
+- [ ] **Step 3: Реализовать** (рядом с judge-логгером из Task 1)
+
+```js
+export function logSafeBaselineAction(sessionId, { tool, action }, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const f = join(dir, `safe-baseline-actions-${sessionId || 'unknown'}.jsonl`);
+    appendFileSync(f, JSON.stringify({
+      ts: new Date().toISOString(),
+      tool: tool || 'unknown',
+      action: String(action || 'allow'),
+    }) + '\n');
+  } catch { /* ignore */ }
+}
+
+export function readSafeBaselineActions(sessionId, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const f = join(dir, `safe-baseline-actions-${sessionId || 'unknown'}.jsonl`);
+    if (!existsSync(f)) return [];
+    return readFileSync(f, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+```
+
+- [ ] **Step 4: Прогнать — PASS.**
+
+- [ ] **Step 5: Вызвать из `enforce-safe-baseline-metering.mjs` `runMain`.** Заменить блок `if (res.action === 'soft_flag') logFlag(...)` на запись любого не-allow действия через новый логгер, сохранив старый `logFlag` для совместимости:
+
+```js
+    if (res.action !== 'allow') {
+      logSafeBaselineAction(sess, { tool: event.tool_name, action: res.action });
+      if (res.action === 'soft_flag') logFlag(dir, sess, { tool: event.tool_name, reason: res.reason });
+    }
+```
+
+Добавить `import { logSafeBaselineAction } from './enforce-hook-helpers.mjs';` (или дополнить существующий импорт из helpers). NB: `logSafeBaselineAction` сам резолвит `runtimeDir()` — в проде это совпадает с `dir`; для теста `runMain` принимает `runtimeDir` override, но логгер пишет в реальный runtime — поэтому в тесте инъектировать override не обязательно (проверяем экспортируемый логгер отдельно в Step 1).
+
+- [ ] **Step 6: Регрессия safe-baseline.**
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/enforce-safe-baseline-metering.test.mjs`
+Expected: PASS (существующие тесты не трогают `action !== 'allow'` ветку с диском, либо обновить ожидания).
+
+- [ ] **Step 7: Commit**
+
+```bash
+git add tools/enforce-hook-helpers.mjs tools/enforce-hook-helpers.test.mjs tools/enforce-safe-baseline-metering.mjs tools/enforce-safe-baseline-metering.test.mjs
+git commit -m "feat(brain): log every safe-baseline action incl hard_block"
+```
+
+---
+
+## Task 3: Чистый ридер сигналов в окне хода — `observer-v4-signals.mjs`
+
+**Files:**
+- Create: `tools/observer-v4-signals.mjs`
+- Test: `tools/observer-v4-signals.test.mjs`
+
+- [ ] **Step 1: Тест ридера**
+
+```js
+// tools/observer-v4-signals.test.mjs
+import { describe, it, expect } from 'vitest';
+import { tmpdir } from 'node:os';
+import { mkdtempSync, writeFileSync, rmSync } from 'node:fs';
+import { join } from 'node:path';
+import { extractV4Signals } from './observer-v4-signals.mjs';
+
+function write(dir, name, lines) { writeFileSync(join(dir, name), lines.map((l) => JSON.stringify(l)).join('\n') + '\n'); }
+
+describe('extractV4Signals', () => {
+  it('counts rationalization flags inside the turn window only', () => {
+    const dir = mkdtempSync(join(tmpdir(), 'v4-'));
+    write(dir, 'rationalization-flags-s.jsonl', [
+      { ts: '2026-05-31T10:00:00.000Z', kind: 'x' }, // before window
+      { ts: '2026-05-31T10:05:00.000Z', kind: 'y' }, // in
+      { ts: '2026-05-31T10:06:00.000Z', kind: 'z' }, // in
+    ]);
+    write(dir, 'llm-judge-verdicts-s.jsonl', [
+      { ts: '2026-05-31T10:05:30.000Z', tool: 'Edit', verdict: 'YES' },
+      { ts: '2026-05-31T10:05:40.000Z', tool: 'Bash', verdict: 'block' },
+    ]);
+    write(dir, 'safe-baseline-actions-s.jsonl', [
+      { ts: '2026-05-31T10:05:10.000Z', tool: 'Edit', action: 'soft_flag' },
+      { ts: '2026-05-31T10:05:50.000Z', tool: 'Write', action: 'hard_block' },
+    ]);
+    writeFileSync(join(dir, 'llm-judge-budget-s.json'), JSON.stringify({ calls: 7 }));
+    const sig = extractV4Signals('s', {
+      startMs: Date.parse('2026-05-31T10:04:00.000Z'),
+      endMs: Date.parse('2026-05-31T10:07:00.000Z'),
+      baseDir: dir,
+    });
+    expect(sig.rationalization_flag_count).toBe(2);
+    expect(sig.judge_verdict).toBe('block');        // last in-window verdict
+    expect(sig.safe_baseline_action).toBe('hard_block'); // worst in-window action
+    expect(sig.judge_calls).toBe(7);
+    rmSync(dir, { recursive: true, force: true });
+  });
+  it('returns zero/null defaults when files absent', () => {
+    const sig = extractV4Signals('nope', { startMs: 0, endMs: 1, baseDir: tmpdir() });
+    expect(sig).toEqual({ rationalization_flag_count: 0, judge_verdict: null, safe_baseline_action: null, judge_calls: 0 });
+  });
+});
+```
+
+- [ ] **Step 2: Прогнать — FAIL** (модуль не существует).
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/observer-v4-signals.test.mjs`
+
+- [ ] **Step 3: Реализовать `observer-v4-signals.mjs`**
+
+```js
+#!/usr/bin/env node
+/**
+ * Pure reader of router-gate v4 runtime signals, scoped to one episode's
+ * [startMs, endMs] turn window. Feeds observer-transcript-parser (new
+ * episode.v4_signals block + task_cost.judge_spend_usd). No exec/exit.
+ */
+import { readFileSync, existsSync } from 'node:fs';
+import { join } from 'node:path';
+import { homedir } from 'node:os';
+
+function rtDir(override) { return override || join(homedir(), '.claude', 'runtime'); }
+
+function readJsonl(path) {
+  if (!existsSync(path)) return [];
+  try {
+    return readFileSync(path, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+
+function inWindow(rec, startMs, endMs) {
+  const ms = Date.parse(rec && rec.ts);
+  return Number.isFinite(ms) && ms >= startMs && ms <= endMs;
+}
+
+// soft order: hard_block > soft_flag > allow.
+const ACTION_RANK = { allow: 0, soft_flag: 1, hard_block: 2 };
+function worstAction(records) {
+  let best = null;
+  for (const r of records) {
+    const a = r && r.action;
+    if (a == null) continue;
+    if (best === null || (ACTION_RANK[a] ?? -1) > (ACTION_RANK[best] ?? -1)) best = a;
+  }
+  return best;
+}
+
+export function extractV4Signals(sessionId, { startMs, endMs, baseDir } = {}) {
+  const dir = rtDir(baseDir);
+  const sess = sessionId || 'unknown';
+
+  const flags = readJsonl(join(dir, `rationalization-flags-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+
+  const verdicts = readJsonl(join(dir, `llm-judge-verdicts-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+  const judge_verdict = verdicts.length ? (verdicts[verdicts.length - 1].verdict ?? null) : null;
+
+  const actions = readJsonl(join(dir, `safe-baseline-actions-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+  const safe_baseline_action = worstAction(actions);
+
+  let judge_calls = 0;
+  const bp = join(dir, `llm-judge-budget-${sess}.json`);
+  if (existsSync(bp)) {
+    try { judge_calls = Number(JSON.parse(readFileSync(bp, 'utf-8')).calls) || 0; } catch { /* keep 0 */ }
+  }
+
+  return {
+    rationalization_flag_count: flags.length,
+    judge_verdict,
+    safe_baseline_action,
+    judge_calls,
+  };
+}
+```
+
+- [ ] **Step 4: Прогнать — PASS.**
+
+- [ ] **Step 5: Commit**
+
+```bash
+git add tools/observer-v4-signals.mjs tools/observer-v4-signals.test.mjs
+git commit -m "feat(brain): pure reader for v4 runtime signals in turn window"
+```
+
+---
+
+## Task 4: Внести сигналы в эпизод (`parseTranscript`)
+
+**Files:**
+- Modify: `tools/observer-transcript-parser.mjs` (:888 return object)
+- Test: `tools/observer-transcript-parser.test.mjs`
+
+- [ ] **Step 1: Тест — эпизод несёт `v4_signals` и `task_cost.judge_spend_usd`**
+
+```js
+// tools/observer-transcript-parser.test.mjs — добавить
+it('episode carries v4_signals block (default zero/null when no runtime files)', () => {
+  const ep = parseTranscript(JSON.stringify({
+    sessionId: 'no-runtime-sess',
+    message: { role: 'user', content: 'сделай фичу X' },
+    timestamp: '2026-05-31T10:00:00.000Z',
+  }), null, { routerStateBaseDir: tmpdir() });
+  expect(ep.v4_signals).toEqual({
+    rationalization_flag_count: 0, judge_verdict: null, safe_baseline_action: null, judge_calls: 0,
+  });
+  expect(typeof ep.task_cost.judge_spend_usd).toBe('number');
+  expect(ep.schema_minor).toBe(4);
+});
+```
+
+(`tmpdir` импортировать в тест-файле, если ещё не.)
+
+- [ ] **Step 2: Прогнать — FAIL** (`ep.v4_signals` undefined).
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/observer-transcript-parser.test.mjs`
+
+- [ ] **Step 3: Реализовать.** В `observer-transcript-parser.mjs`:
+
+(а) импорт вверху рядом с другими: `import { extractV4Signals } from './observer-v4-signals.mjs';` и `import { JUDGE_PER_CALL_USD } from './cost-pricing.mjs';`
+
+(б) в `parseTranscript`, после вычисления `started_at`/`ended_at` (:847), добавить:
+
+```js
+  const _v4 = extractV4Signals(sessionId, {
+    startMs: new Date(started_at).getTime(),
+    endMs: new Date(ended_at).getTime(),
+    baseDir: options.runtimeBaseDir || routerStateBaseDir,
+  });
+```
+
+(в) bump `schema_minor: 3` → `schema_minor: 4` в return.
+
+(г) в return-объект добавить поле `v4_signals: _v4,`.
+
+(д) в `task_cost` влить судейский spend: заменить строку `task_cost: { ...extractTokenUsage(turn), ...((routerState && routerState.task_cost) || {}) },` на:
+
+```js
+    task_cost: {
+      ...extractTokenUsage(turn),
+      ...((routerState && routerState.task_cost) || {}),
+      judge_spend_usd: _v4.judge_calls * JUDGE_PER_CALL_USD,
+    },
+```
+
+NB: `routerStateBaseDir` уже извлекается из `options` (:838). `extractV4Signals` принимает тот же baseDir — в проде это `~/.claude/runtime`, в тестах — temp.
+
+- [ ] **Step 4: Прогнать — PASS** (после Task 5, где появится `JUDGE_PER_CALL_USD`). Если выполняется до Task 5 — импорт `JUDGE_PER_CALL_USD` упадёт; **зависимость: Task 5 Step 3 делать до Task 4 Step 3** (или внести константу в этом же шаге). Для линейного исполнения: добавить `JUDGE_PER_CALL_USD` в `cost-pricing.mjs` здесь же (Step 3.е) и продублировать тест в Task 5.
+
+(3.е) В `tools/cost-pricing.mjs` добавить экспорт:
+
+```js
+// Оценочная средняя стоимость одного вызова per-tool судьи (Sonnet, ~600 in / ~10 out
+// токенов на короткий YES/NO-промпт). Уточнить, когда появится по-вызовный учёт токенов.
+export const JUDGE_PER_CALL_USD = 600 * PRICING.sonnet.input + 10 * PRICING.sonnet.output;
+```
+
+- [ ] **Step 5: Прогнать целевой тест — PASS.**
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/observer-transcript-parser.test.mjs`
+
+- [ ] **Step 6: Commit**
+
+```bash
+git add tools/observer-transcript-parser.mjs tools/observer-transcript-parser.test.mjs tools/cost-pricing.mjs
+git commit -m "feat(brain): surface v4_signals + judge_spend_usd into episode (schema_minor 4)"
+```
+
+---
+
+## Task 5: Денежный компонент `judge_spend_usd` в агрегаторе и cost-daily
+
+**Files:**
+- Modify: `tools/cost-aggregator.mjs` (`episodeUsd` :13, `aggregateDay` :40)
+- Test: `tools/cost-aggregator.test.mjs`
+
+- [ ] **Step 1: Тест — `episodeUsd` отдаёт `judge_spend_usd`, `total_usd` его включает**
+
+```js
+// tools/cost-aggregator.test.mjs
+it('includes judge_spend_usd in episodeUsd + total', () => {
+  const pricing = { sonnet: { input: 3e-6, output: 15e-6 }, opus: { input: 15e-6, output: 75e-6 } };
+  const ep = { task_cost: { judge_spend_usd: 0.002 } };
+  const u = episodeUsd(ep, pricing);
+  expect(u.judge_spend_usd).toBe(0.002);
+  expect(u.total_usd).toBeCloseTo(0.002, 9);
+});
+```
+
+- [ ] **Step 2: Прогнать — FAIL** (`u.judge_spend_usd` undefined).
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/cost-aggregator.test.mjs`
+
+- [ ] **Step 3: Реализовать.** В `episodeUsd` (:13) добавить строку и включить в `total_usd` + объект:
+
+```js
+  const judge_spend_usd = tc.judge_spend_usd || 0;
+  const total_usd =
+    classifier_usd + self_assessment_usd + reviewer_subagent_usd +
+    reviewer_direct_fallback_usd + self_retrospect_usd + judge_spend_usd;
+  return {
+    classifier_usd, self_assessment_usd, reviewer_subagent_usd,
+    reviewer_direct_fallback_usd, self_retrospect_usd, judge_spend_usd, total_usd,
+  };
+```
+
+В `aggregateDay` (:40) добавить `judge_spend_usd: 0` в `totals`, прибавлять `totals.judge_spend_usd += u.judge_spend_usd;` в цикле и включить в финальный `total_usd`.
+
+- [ ] **Step 4: Прогнать — PASS** + регрессия `cost-stop-hook`:
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/cost-aggregator.test.mjs tools/cost-stop-hook.test.mjs`
+Expected: PASS (cost-daily.json теперь получит поле `judge_spend_usd` автоматически — оно идёт из `aggregateDay`; при необходимости обновить ожидания в `cost-stop-hook.test.mjs`).
+
+- [ ] **Step 5: Commit**
+
+```bash
+git add tools/cost-aggregator.mjs tools/cost-aggregator.test.mjs tools/cost-stop-hook.test.mjs
+git commit -m "feat(brain): add judge_spend_usd cost component to aggregator + cost-daily"
+```
+
+---
+
+## Task 6: Четыре новые факторные оси
+
+**Files:**
+- Modify: `tools/brain-retro-analyzer.mjs` (`FACTOR_FNS` :326)
+- Test: `tools/brain-retro-analyzer.test.mjs`
+
+- [ ] **Step 1: Тест — матрица содержит новые оси**
+
+```js
+// tools/brain-retro-analyzer.test.mjs
+import { buildFactorMatrix } from './brain-retro-analyzer.mjs';
+it('factor matrix exposes v4-signal axes', () => {
+  const eps = [
+    { _inferredOutcome: 'success', v4_signals: { rationalization_flag_count: 0, judge_verdict: 'YES', safe_baseline_action: 'allow', judge_calls: 0 }, task_cost: { judge_spend_usd: 0 } },
+    { _inferredOutcome: 'rework',  v4_signals: { rationalization_flag_count: 3, judge_verdict: 'block', safe_baseline_action: 'hard_block', judge_calls: 12 }, task_cost: { judge_spend_usd: 0.03 } },
+  ];
+  const m = buildFactorMatrix(eps);
+  expect(m.rationalization_flag_count['0']).toEqual({ success: 1 });
+  expect(m.rationalization_flag_count['2+']).toEqual({ rework: 1 });
+  expect(m.judge_verdict.YES).toEqual({ success: 1 });
+  expect(m.judge_verdict.block).toEqual({ rework: 1 });
+  expect(m.safe_baseline_action.hard_block).toEqual({ rework: 1 });
+  expect(m.judge_calls_bucket['0']).toEqual({ success: 1 });
+  expect(m.judge_calls_bucket['10+']).toEqual({ rework: 1 });
+});
+```
+
+- [ ] **Step 2: Прогнать — FAIL** (осей нет).
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/brain-retro-analyzer.test.mjs`
+
+- [ ] **Step 3: Реализовать.** В `FACTOR_FNS` (:326) добавить (используя существующий `countBucket012` :321 для 0/1/2+ и новый локальный bucket для judge_calls):
+
+```js
+  // Pass 5 — router-gate v4 signal axes (brain-data-catalog раздел F → факторы).
+  rationalization_flag_count: (e) => countBucket012((e.v4_signals || {}).rationalization_flag_count),
+  judge_verdict: (e) => (e.v4_signals || {}).judge_verdict || 'null',
+  safe_baseline_action: (e) => (e.v4_signals || {}).safe_baseline_action || 'null',
+  judge_calls_bucket: (e) => judgeCallsBucket((e.v4_signals || {}).judge_calls),
+```
+
+И рядом с другими bucket-хелперами (около :321) добавить:
+
+```js
+function judgeCallsBucket(n) {
+  const v = Number(n) || 0;
+  if (v === 0) return '0';
+  if (v < 10) return '1-9';
+  return '10+';
+}
+```
+
+- [ ] **Step 4: Прогнать — PASS** + полная регрессия анализатора:
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs tools/brain-retro-analyzer.test.mjs`
+Expected: PASS.
+
+- [ ] **Step 5: Commit**
+
+```bash
+git add tools/brain-retro-analyzer.mjs tools/brain-retro-analyzer.test.mjs
+git commit -m "feat(brain): add 4 v4-signal factor axes (rationalization/judge/safe-baseline)"
+```
+
+---
+
+## Task 7: Обновить каталог данных + полная регрессия tools
+
+**Files:**
+- Modify: `docs/observer/brain-data-catalog.md`
+
+- [ ] **Step 1: Перенести сигналы из раздела F.** В `brain-data-catalog.md`: в A.7 добавить `judge_spend_usd` в список `task_cost`; добавить новую группу **A.12 `v4_signals`** (`rationalization_flag_count`, `judge_verdict`, `safe_baseline_action`, `judge_calls`); в раздел B дописать «Pass 5 (4 оси): rationalization_flag_count / judge_verdict / safe_baseline_action / judge_calls_bucket»; в разделе F пометить F.1–F.3 как «✅ заведено в эпизод (см. A.12)», оставив F.4/F.5 как есть; bump «схема v4.3» → «v4.4 (schema_minor 4)».
+
+- [ ] **Step 2: Полная регрессия tools-only**
+
+Run: `npx vitest run --root app --config vitest.config.tools.mjs`
+Expected: все GREEN (новые тесты + прежний baseline без падений).
+
+- [ ] **Step 3: Commit**
+
+```bash
+git add docs/observer/brain-data-catalog.md
+git commit -m "docs(observer): catalog — v4 signals promoted from F into episode + factors"
+```
+
+---
+
+## Self-Review
+
+**1. Покрытие требований (кандидаты F из каталога):**
+- `rationalization_flag_count` → Task 3 (ридер) + Task 4 (поле) + Task 6 (ось). ✅
+- `safe_baseline_action` → Task 2 (логирование) + Task 3 + Task 4 + Task 6. ✅
+- `judge_verdict` → Task 1 (логирование) + Task 3 + Task 4 + Task 6. ✅
+- `judge_spend_usd` → Task 4 (поле, из `judge_calls × JUDGE_PER_CALL_USD`) + Task 5 (агрегатор/cost-daily). ✅
+- `judge_calls` → Task 3 (бюджет) + Task 6 (ось `judge_calls_bucket`). ✅
+
+**2. Заглушки:** код в каждом шаге полный; команды и ожидания указаны. `JUDGE_PER_CALL_USD` — явная оценочная константа с комментарием (не TODO-заглушка), уточняемая, когда появится по-вызовный учёт токенов.
+
+**3. Согласованность типов/имён:**
+- `extractV4Signals(sessionId, {startMs,endMs,baseDir})` → возвращает `{rationalization_flag_count, judge_verdict, safe_baseline_action, judge_calls}` — те же ключи в Task 3 тесте, Task 4 поле `v4_signals`, Task 6 осях. ✅
+- `logJudgeVerdict`/`readJudgeVerdicts`, `logSafeBaselineAction`/`readSafeBaselineActions` — пары имён согласованы между Task 1/2 и Task 3 (файлы `llm-judge-verdicts-<sess>.jsonl`, `safe-baseline-actions-<sess>.jsonl`). ✅
+- `judge_spend_usd` — одно имя в Task 4 (запись), Task 5 (агрегатор), каталоге. ✅
+- **Зависимость порядка:** Task 5 Step 3 (или Task 4 Step 3.е) должен внести `JUDGE_PER_CALL_USD` в `cost-pricing.mjs` ДО импорта в `observer-transcript-parser.mjs`. Зафиксировано в Task 4 Step 4 примечанием.
+
+**4. Риски/допущения, проверить при исполнении:**
+- `observer-transcript-parser` читает runtime по абсолютному `baseDir` (прод `~/.claude/runtime`); тесты инъектируют temp через `options.runtimeBaseDir || routerStateBaseDir`.
+- `judge_calls` — кумулятивно per-session (бюджет не сбрасывается per-turn), поэтому `judge_calls_bucket` отражает нагрузку сессии, не хода; это осознанно (для USD это и нужно). Если потребуется per-turn — считать вердикты в окне (`verdicts.length`), отдельная ось.
+- `enforce-llm-judge-per-tool` логирует вердикт только при реальном вызове (`result.verdict !== undefined`) — disabled/readonly/test-runner не пишут (verdict отсутствует) → `judge_verdict` останется `null`, что корректно отражает «судья не сработал».
diff --git a/tools/brain-retro-analyzer.mjs b/tools/brain-retro-analyzer.mjs
index 3ebefb31..27e79368 100644
--- a/tools/brain-retro-analyzer.mjs
+++ b/tools/brain-retro-analyzer.mjs
@@ -323,6 +323,13 @@ function countBucket012(n) {
   return v === 0 ? '0' : v === 1 ? '1' : '2+';
 }
 
+function judgeCallsBucket(n) {
+  const v = Number(n) || 0;
+  if (v === 0) return '0';
+  if (v < 10) return '1-9';
+  return '10+';
+}
+
 const FACTOR_FNS = {
   decision_provenance: (e) => (e.decision_provenance || {}).kind || 'unknown',
   economy_level: (e) => String((e.environment || {}).economy_level ?? 'null'),
@@ -362,6 +369,11 @@ const FACTOR_FNS = {
   // (cross-episode pass via observer-embedding-index). Episodes without an
   // embedding or with no resolved neighbours bucket as 'no_neighbors'.
   similar_past_outcome_majority: (e) => e._similarPastOutcomeMajority || 'no_neighbors',
+  // Pass 5 — router-gate v4 signal axes (brain-data-catalog раздел F → факторы).
+  rationalization_flag_count: (e) => countBucket012((e.v4_signals || {}).rationalization_flag_count),
+  judge_verdict: (e) => (e.v4_signals || {}).judge_verdict || 'null',
+  safe_baseline_action: (e) => (e.v4_signals || {}).safe_baseline_action || 'null',
+  judge_calls_bucket: (e) => judgeCallsBucket((e.v4_signals || {}).judge_calls),
 };
 
 // Pass 4 — decode prompt_embedding_base64 to Float32Array. Mirrors
diff --git a/tools/brain-retro-analyzer.test.mjs b/tools/brain-retro-analyzer.test.mjs
index 7fd3b4d4..47f99038 100644
--- a/tools/brain-retro-analyzer.test.mjs
+++ b/tools/brain-retro-analyzer.test.mjs
@@ -431,6 +431,23 @@ describe('analyze — v4 aggregations (Phase 3 Task 20)', () => {
   });
 });
 
+describe('buildFactorMatrix — Pass 5 v4-signal axes', () => {
+  it('factor matrix exposes v4-signal axes', () => {
+    const eps = [
+      { _inferredOutcome: 'success', v4_signals: { rationalization_flag_count: 0, judge_verdict: 'YES', safe_baseline_action: 'allow', judge_calls: 0 }, task_cost: { judge_spend_usd: 0 } },
+      { _inferredOutcome: 'rework', v4_signals: { rationalization_flag_count: 3, judge_verdict: 'block', safe_baseline_action: 'hard_block', judge_calls: 12 }, task_cost: { judge_spend_usd: 0.03 } },
+    ];
+    const m = buildFactorMatrix(eps);
+    expect(m.rationalization_flag_count['0']).toEqual({ success: 1 });
+    expect(m.rationalization_flag_count['2+']).toEqual({ rework: 1 });
+    expect(m.judge_verdict.YES).toEqual({ success: 1 });
+    expect(m.judge_verdict.block).toEqual({ rework: 1 });
+    expect(m.safe_baseline_action.hard_block).toEqual({ rework: 1 });
+    expect(m.judge_calls_bucket['0']).toEqual({ success: 1 });
+    expect(m.judge_calls_bucket['10+']).toEqual({ rework: 1 });
+  });
+});
+
 describe('buildFactorMatrix — Pass 1 cheap axes (project-brain-factor-analysis-4passes)', () => {
   // Each new axis: smoke + null-safety on missing fields.
   it('prompt_signal axis: raw discrete values + null fallback', () => {
diff --git a/tools/cost-aggregator.mjs b/tools/cost-aggregator.mjs
index a37974c3..f2e9db3f 100644
--- a/tools/cost-aggregator.mjs
+++ b/tools/cost-aggregator.mjs
@@ -21,18 +21,21 @@ export function episodeUsd(episode, pricing) {
   const reviewer_subagent_usd = tc.reviewer_subagent_usd || 0;
   const reviewer_direct_fallback_usd = tc.reviewer_direct_fallback_usd || 0;
   const self_retrospect_usd = tc.self_retrospect_usd || 0;
+  const judge_spend_usd = tc.judge_spend_usd || 0;
   const total_usd =
     classifier_usd +
     self_assessment_usd +
     reviewer_subagent_usd +
     reviewer_direct_fallback_usd +
-    self_retrospect_usd;
+    self_retrospect_usd +
+    judge_spend_usd;
   return {
     classifier_usd,
     self_assessment_usd,
     reviewer_subagent_usd,
     reviewer_direct_fallback_usd,
     self_retrospect_usd,
+    judge_spend_usd,
     total_usd,
   };
 }
@@ -44,6 +47,7 @@ export function aggregateDay(episodes, dateISO, pricing) {
     reviewer_subagent_usd: 0,
     reviewer_direct_fallback_usd: 0,
     self_retrospect_usd: 0,
+    judge_spend_usd: 0,
   };
   let episode_count = 0;
   for (const ep of Array.isArray(episodes) ? episodes : []) {
@@ -55,6 +59,7 @@ export function aggregateDay(episodes, dateISO, pricing) {
     totals.reviewer_subagent_usd += u.reviewer_subagent_usd;
     totals.reviewer_direct_fallback_usd += u.reviewer_direct_fallback_usd;
     totals.self_retrospect_usd += u.self_retrospect_usd;
+    totals.judge_spend_usd += u.judge_spend_usd;
     episode_count += 1;
   }
   const total_usd =
@@ -62,6 +67,7 @@ export function aggregateDay(episodes, dateISO, pricing) {
     totals.self_assessment_usd +
     totals.reviewer_subagent_usd +
     totals.reviewer_direct_fallback_usd +
-    totals.self_retrospect_usd;
+    totals.self_retrospect_usd +
+    totals.judge_spend_usd;
   return { ...totals, total_usd, episode_count };
 }
diff --git a/tools/cost-aggregator.test.mjs b/tools/cost-aggregator.test.mjs
index e70fccb7..be0b3f90 100644
--- a/tools/cost-aggregator.test.mjs
+++ b/tools/cost-aggregator.test.mjs
@@ -13,6 +13,13 @@ describe('episodeUsd', () => {
     expect(u.total_usd).toBe(0);
   });
 
+  it('includes judge_spend_usd in episodeUsd + total', () => {
+    const ep = { task_cost: { judge_spend_usd: 0.002 } };
+    const u = episodeUsd(ep, PRICING);
+    expect(u.judge_spend_usd).toBe(0.002);
+    expect(u.total_usd).toBeCloseTo(0.002, 9);
+  });
+
   it('computes classifier_usd from sonnet pricing', () => {
     const ep = { task_cost: { classifier_input_tokens: 1_000_000, classifier_output_tokens: 100_000 } };
     const u = episodeUsd(ep, PRICING);
@@ -96,11 +103,12 @@ describe('aggregateDay', () => {
     expect(r.classifier_usd).toBeCloseTo(3, 6);
   });
 
-  it('returns object with exactly 7 keys (5 components + total + count)', () => {
+  it('returns object with exactly 8 keys (6 components + total + count)', () => {
     const r = aggregateDay([], '2026-05-28', PRICING);
     expect(Object.keys(r).sort()).toEqual([
       'classifier_usd',
       'episode_count',
+      'judge_spend_usd',
       'reviewer_direct_fallback_usd',
       'reviewer_subagent_usd',
       'self_assessment_usd',
diff --git a/tools/cost-pricing.mjs b/tools/cost-pricing.mjs
index 18822307..825b9e6d 100644
--- a/tools/cost-pricing.mjs
+++ b/tools/cost-pricing.mjs
@@ -23,3 +23,9 @@ export const PRICING = Object.freeze({
     output: 75 / 1_000_000,
   }),
 });
+
+// Оценочная средняя стоимость одного вызова per-tool LLM-судьи (Sonnet, ~600 in /
+// ~10 out токенов на короткий YES/NO-промпт). Используется для derive
+// task_cost.judge_spend_usd из числа вызовов судьи (judge_calls). Уточнить, когда
+// появится по-вызовный учёт токенов судьи.
+export const JUDGE_PER_CALL_USD = 600 * PRICING.sonnet.input + 10 * PRICING.sonnet.output;
diff --git a/tools/cost-pricing.test.mjs b/tools/cost-pricing.test.mjs
index cedf939f..c8e29f7a 100644
--- a/tools/cost-pricing.test.mjs
+++ b/tools/cost-pricing.test.mjs
@@ -26,3 +26,11 @@ describe('PRICING constants', () => {
     expect(Object.isFrozen(PRICING.opus)).toBe(true);
   });
 });
+
+describe('JUDGE_PER_CALL_USD', () => {
+  it('is a positive USD estimate for one per-tool judge call', async () => {
+    const { JUDGE_PER_CALL_USD } = await import('./cost-pricing.mjs');
+    expect(typeof JUDGE_PER_CALL_USD).toBe('number');
+    expect(JUDGE_PER_CALL_USD).toBeGreaterThan(0);
+  });
+});
diff --git a/tools/enforce-hook-helpers.mjs b/tools/enforce-hook-helpers.mjs
index 2bcfd117..3b2404fa 100644
--- a/tools/enforce-hook-helpers.mjs
+++ b/tools/enforce-hook-helpers.mjs
@@ -298,6 +298,52 @@ export function readRationalizationFlags(sessionId) {
   } catch { return []; }
 }
 
+export function logJudgeVerdict(sessionId, { tool, verdict }, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const fi = join(dir, `llm-judge-verdicts-${sessionId || 'unknown'}.jsonl`);
+    appendFileSync(fi, JSON.stringify({
+      ts: new Date().toISOString(),
+      tool: tool || 'unknown',
+      verdict: verdict == null ? null : String(verdict),
+    }) + '\n');
+  } catch { /* ignore */ }
+}
+
+export function readJudgeVerdicts(sessionId, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const fi = join(dir, `llm-judge-verdicts-${sessionId || 'unknown'}.jsonl`);
+    if (!existsSync(fi)) return [];
+    return readFileSync(fi, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+
+export function logSafeBaselineAction(sessionId, { tool, action }, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const fi = join(dir, `safe-baseline-actions-${sessionId || 'unknown'}.jsonl`);
+    appendFileSync(fi, JSON.stringify({
+      ts: new Date().toISOString(),
+      tool: tool || 'unknown',
+      action: String(action || 'allow'),
+    }) + '\n');
+  } catch { /* ignore */ }
+}
+
+export function readSafeBaselineActions(sessionId, opts = {}) {
+  try {
+    const dir = opts.baseDir || runtimeDir();
+    const fi = join(dir, `safe-baseline-actions-${sessionId || 'unknown'}.jsonl`);
+    if (!existsSync(fi)) return [];
+    return readFileSync(fi, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+
 export function readRouterState(sessionId) {
   try {
     const p = join(runtimeDir(), `router-state-${sessionId || 'unknown'}.json`);
diff --git a/tools/enforce-hook-helpers.test.mjs b/tools/enforce-hook-helpers.test.mjs
index 23d33387..21d9a746 100644
--- a/tools/enforce-hook-helpers.test.mjs
+++ b/tools/enforce-hook-helpers.test.mjs
@@ -23,8 +23,23 @@ import {
   sessionToolUses,
   logHookOutcome,
   runtimeDir,
+  logSafeBaselineAction,
+  readSafeBaselineActions,
 } from './enforce-hook-helpers.mjs';
 
+describe('safe-baseline action log', () => {
+  it('appends and reads action records', () => {
+    const dir = mkdtempSync(join(tmpdir(), 'sb-'));
+    logSafeBaselineAction('s', { tool: 'Edit', action: 'hard_block' }, { baseDir: dir });
+    const recs = readSafeBaselineActions('s', { baseDir: dir });
+    expect(recs[0]).toMatchObject({ tool: 'Edit', action: 'hard_block' });
+    rmSync(dir, { recursive: true, force: true });
+  });
+  it('returns [] when no file', () => {
+    expect(readSafeBaselineActions('nope', { baseDir: tmpdir() })).toEqual([]);
+  });
+});
+
 // v4: override surface removed per spec §4.2 — stubs return null/empty
 describe('v4 override stubs', () => {
   it('loadOverrideVocab returns empty phrases array (stub)', () => {
@@ -520,3 +535,25 @@ describe('detectFullTestRun', () => {
     expect(detectFullTestRun('cd app && cp src dst')).toBeNull();
   });
 });
+
+// judge verdict log — added for LLM-judge brain-retro factoring
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join as joinPath } from 'node:path';
+import { logJudgeVerdict, readJudgeVerdicts } from './enforce-hook-helpers.mjs';
+
+describe('judge verdict log', () => {
+  it('appends and reads back verdict records (RED: logJudgeVerdict not yet exported)', () => {
+    const dir = mkdtempSync(joinPath(tmpdir(), 'jv-'));
+    logJudgeVerdict('sess1', { tool: 'Edit', verdict: 'YES' }, { baseDir: dir });
+    logJudgeVerdict('sess1', { tool: 'Bash', verdict: 'NO' }, { baseDir: dir });
+    const recs = readJudgeVerdicts('sess1', { baseDir: dir });
+    expect(recs.length).toBe(2);
+    expect(recs[0]).toMatchObject({ tool: 'Edit', verdict: 'YES' });
+    expect(typeof recs[0].ts).toBe('string');
+    rmSync(dir, { recursive: true, force: true });
+  });
+  it('returns [] when no file', () => {
+    expect(readJudgeVerdicts('nope', { baseDir: tmpdir() })).toEqual([]);
+  });
+});
diff --git a/tools/enforce-llm-judge-per-tool.mjs b/tools/enforce-llm-judge-per-tool.mjs
index 7a6438ce..76bbbdf5 100644
--- a/tools/enforce-llm-judge-per-tool.mjs
+++ b/tools/enforce-llm-judge-per-tool.mjs
@@ -21,7 +21,7 @@
 import { judgePerTool, MUTATING_TOOLS, readDeclaredTask, resolveEffectiveTask } from './llm-judge-per-tool.mjs';
 import { resolveJudgeConfig } from './llm-judge-config.mjs';
 import { readJudgeBudget, bumpJudgeBudget, JUDGE_SESSION_BUDGET, llmJudgeCall } from './llm-judge.mjs';
-import { readStdin, parseEventJson, exitDecision, readTranscript, lastUserPromptText } from './enforce-hook-helpers.mjs';
+import { readStdin, parseEventJson, exitDecision, readTranscript, lastUserPromptText, logJudgeVerdict } from './enforce-hook-helpers.mjs';
 import { classifyBashCommand } from './enforce-router-gate.mjs';
 
 /**
@@ -117,6 +117,7 @@ export async function runPerTool({
   readBudgetImpl,
   bumpBudgetImpl,
   llmJudgeCallImpl,
+  logVerdictImpl = () => {},
   sessionBudget = JUDGE_SESSION_BUDGET,
 }) {
   // Readonly Bash never mutates → outside the judge's scope; skip (no LLM call, no spend).
@@ -146,7 +147,10 @@ export async function runPerTool({
     budgetState: { spent, limit: sessionBudget },
     llmJudgeCallImpl,
   });
-  if (result.verdict !== undefined) bumpBudgetImpl({ sessionId, by: 1 });
+  if (result.verdict !== undefined) {
+    bumpBudgetImpl({ sessionId, by: 1 });
+    logVerdictImpl({ sessionId, tool: event && event.tool_name, verdict: result.block ? 'block' : (result.verdict ?? 'YES') });
+  }
   return result;
 }
 
@@ -165,6 +169,7 @@ async function main() {
       readBudgetImpl: readJudgeBudget,
       bumpBudgetImpl: bumpJudgeBudget,
       llmJudgeCallImpl: (opts) => llmJudgeCall(opts),
+      logVerdictImpl: ({ sessionId, tool, verdict }) => logJudgeVerdict(sessionId, { tool, verdict }),
     });
     exitDecision({ block: result.block, message: result.reason });
   } catch {
diff --git a/tools/enforce-llm-judge-per-tool.test.mjs b/tools/enforce-llm-judge-per-tool.test.mjs
index 844b4b5b..bdfefbdf 100644
--- a/tools/enforce-llm-judge-per-tool.test.mjs
+++ b/tools/enforce-llm-judge-per-tool.test.mjs
@@ -316,6 +316,23 @@ describe('runPerTool — test-runner Bash skips the judge; mutating Bash still j
   });
 });
 
+describe('runPerTool — verdict logging', () => {
+  it('runPerTool logs verdict only when a real judge call happened', async () => {
+    const logged = [];
+    const r = await runPerTool({
+      event: { tool_name: 'Edit', session_id: 's', tool_input: {} },
+      judgeConfig: { enabled: true, apiKey: 'k' },
+      readDeclaredTaskImpl: () => ({ task_summary: 't' }),
+      readBudgetImpl: () => 0,
+      bumpBudgetImpl: () => {},
+      llmJudgeCallImpl: async () => 'YES',
+      logVerdictImpl: (rec) => logged.push(rec),
+    });
+    expect(r.block).toBe(false);
+    expect(logged).toEqual([{ sessionId: 's', tool: 'Edit', verdict: 'YES' }]);
+  });
+});
+
 // Calibration 4 (soft, 2026-05-31): when the classifier summary is "(unknown)",
 // runPerTool reads the user's last prompt and judges against THAT (better
 // evidence) instead of an empty task. When the summary is meaningful, the
diff --git a/tools/enforce-safe-baseline-metering.mjs b/tools/enforce-safe-baseline-metering.mjs
index a22e29cf..de53889e 100644
--- a/tools/enforce-safe-baseline-metering.mjs
+++ b/tools/enforce-safe-baseline-metering.mjs
@@ -37,6 +37,7 @@ import {
   lastUserPromptText,
   lastTurnEntries,
   exitDecision,
+  logSafeBaselineAction,
 } from './enforce-hook-helpers.mjs';
 
 /**
@@ -206,6 +207,7 @@ export async function runMain({ event, runtimeDir, transcript: injectedTranscrip
     const res = runLiveDecision({ event, priorLedger, promptText, currentKeywords, skillMatchedThisTurn });
     saveLedger(dir, sess, res.ledger);
 
+    if (res.action !== 'allow') logSafeBaselineAction(sess, { tool: event.tool_name, action: res.action }, { baseDir: dir });
     if (res.action === 'soft_flag') logFlag(dir, sess, { tool: event.tool_name, reason: res.reason });
     if (res.action === 'hard_block') return { block: true, message: `[safe-baseline] ${res.reason}\n${ESCAPE_MSG}` };
     return { block: false };
diff --git a/tools/enforce-safe-baseline-metering.test.mjs b/tools/enforce-safe-baseline-metering.test.mjs
index c2605d43..989dc867 100644
--- a/tools/enforce-safe-baseline-metering.test.mjs
+++ b/tools/enforce-safe-baseline-metering.test.mjs
@@ -280,4 +280,19 @@ describe('runMain — live integration', () => {
     const res = await runMain({ event: { tool_name: 'Skill', session_id: 'S3', transcript_path: tpath }, runtimeDir: dir });
     expect(res.block).toBe(false);
   });
+
+  it('logs the action to safe-baseline-actions when a mutating tool is hard-blocked', async () => {
+    const { readFileSync } = await import('node:fs');
+    const dir = mkdtempSync(join(tmpdir(), 'sbm-'));
+    const tpath = join(dir, 't.jsonl');
+    writeFileSync(join(dir, 'safe-baseline-ledger-S4.json'), JSON.stringify({
+      state: { schema_version: 1, task_id: 't', counts: { Read: 60, Grep: 0, Glob: 0, LS: 0, TodoWrite_writes: 0, AskUserQuestion: 0 }, skill_match_within_task: false },
+      lastKeywords: ['router', 'gate', 'safe', 'baseline'],
+    }));
+    fixtureTranscript(tpath, [{ type: 'user', message: { role: 'user', content: 'router gate safe baseline' } }]);
+    const res = await runMain({ event: { tool_name: 'Edit', session_id: 'S4', transcript_path: tpath }, runtimeDir: dir });
+    expect(res.block).toBe(true);
+    const log = readFileSync(join(dir, 'safe-baseline-actions-S4.jsonl'), 'utf-8').trim().split('\n').map((l) => JSON.parse(l));
+    expect(log.some((r) => r.action === 'hard_block' && r.tool === 'Edit')).toBe(true);
+  });
 });
diff --git a/tools/observer-transcript-parser.mjs b/tools/observer-transcript-parser.mjs
index 5bb155ac..fc37421b 100644
--- a/tools/observer-transcript-parser.mjs
+++ b/tools/observer-transcript-parser.mjs
@@ -28,6 +28,8 @@ import { buildHookMap, resolveScriptCounts } from './observer-hook-resolver.mjs'
 // для слепого fallback на heuristic recommended_node — убрано 2026-05-26
 // (brain-retro #6 follow-up). Импорты сняты как dead code.
 import { loadRegistry } from './registry-load.mjs';
+import { extractV4Signals } from './observer-v4-signals.mjs';
+import { JUDGE_PER_CALL_USD } from './cost-pricing.mjs';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
@@ -847,6 +849,12 @@ export function parseTranscript(transcriptText, fallbackSessionId = null, option
   const ended_at = stamps[stamps.length - 1] || started_at;
   const durationMs = new Date(ended_at) - new Date(started_at);
 
+  const _v4 = extractV4Signals(sessionId, {
+    startMs: new Date(started_at).getTime(),
+    endMs: new Date(ended_at).getTime(),
+    baseDir: options.runtimeBaseDir || routerStateBaseDir,
+  });
+
   const { skills, counts, errors } = collectToolUse(turn);
 
   const events = [];
@@ -887,7 +895,7 @@ export function parseTranscript(transcriptText, fallbackSessionId = null, option
 
   return {
     schema_version: 4,
-    schema_minor: 3,
+    schema_minor: 4,
     task_id: sessionId,
     task_ref: sessionId,
     timestamps: { started_at, ended_at },
@@ -910,7 +918,12 @@ export function parseTranscript(transcriptText, fallbackSessionId = null, option
     // NB: routerState (line 855) honours routerStateBaseDir option; _state at line 898
     // does not (always default dir). Use routerState here so tests with custom temp dir
     // see the merged values.
-    task_cost: { ...extractTokenUsage(turn), ...((routerState && routerState.task_cost) || {}) },
+    task_cost: {
+      ...extractTokenUsage(turn),
+      ...((routerState && routerState.task_cost) || {}),
+      judge_spend_usd: _v4.judge_calls * JUDGE_PER_CALL_USD,
+    },
+    v4_signals: _v4,
     // Pass 3 — dynamics meta-block (project-brain-factor-analysis-4passes).
     // prompt_length_chars: strlen of first user prompt (engagement / clarity proxy).
     // mcp_servers_used: unique mcp__<server>__* fingerprints in this turn.
diff --git a/tools/observer-transcript-parser.test.mjs b/tools/observer-transcript-parser.test.mjs
index 5edf3251..ffc19cc9 100644
--- a/tools/observer-transcript-parser.test.mjs
+++ b/tools/observer-transcript-parser.test.mjs
@@ -17,6 +17,21 @@ import {
   classifyFilePath,
 } from './observer-transcript-parser.mjs';
 
+describe('v4_signals + judge_spend_usd in episode', () => {
+  it('episode carries v4_signals block (default zero/null when no runtime files)', () => {
+    const ep = parseTranscript(JSON.stringify({
+      sessionId: 'no-runtime-sess',
+      message: { role: 'user', content: 'сделай фичу X' },
+      timestamp: '2026-05-31T10:00:00.000Z',
+    }), null, { routerStateBaseDir: tmpdir() });
+    expect(ep.v4_signals).toEqual({
+      rationalization_flag_count: 0, judge_verdict: null, safe_baseline_action: null, judge_calls: 0,
+    });
+    expect(typeof ep.task_cost.judge_spend_usd).toBe('number');
+    expect(ep.schema_minor).toBe(4);
+  });
+});
+
 // Build a JSONL transcript string from entry objects.
 function jsonl(entries) {
   return entries.map((e) => JSON.stringify(e)).join('\n');
@@ -236,7 +251,7 @@ describe('parseTranscript', () => {
     expect(ep.events).toEqual([]);
     expect(ep.outcome).toBe('unknown');
     expect(ep.schema_version).toBe(4);
-    expect(ep.schema_minor).toBe(3);
+    expect(ep.schema_minor).toBe(4);
   });
 
   it('produces a complete 7-field primary_rationale', () => {
@@ -469,7 +484,7 @@ describe('parseTranscript — v4 episode (Phase 2 Task 15 bump)', () => {
     ]);
     const ep = parseTranscript(t);
     expect(ep.schema_version).toBe(4);
-    expect(ep.schema_minor).toBe(3);
+    expect(ep.schema_minor).toBe(4);
     expect('classifier_output' in ep).toBe(true);
     expect('degraded_mode' in ep).toBe(true);
     expect('classifier_model' in ep.environment).toBe(true);
@@ -1047,6 +1062,7 @@ describe('parseTranscript — task_cost integration (Task 2)', () => {
       self_assessment_input_tokens: 0, self_assessment_output_tokens: 0,
       reviewer_input_tokens: 0, reviewer_output_tokens: 0,
       reviewer_subagent_usd: 0, reviewer_direct_fallback_usd: 0,
+      judge_spend_usd: 0,
     });
   });
 });
@@ -1657,7 +1673,7 @@ describe('parseTranscript v3 fields', () => {
   it('emits schema_version: 4', () => {
     const ep = parseTranscript(transcriptDirectFeature(), 'sess-1');
     expect(ep.schema_version).toBe(4);
-    expect(ep.schema_minor).toBe(3);
+    expect(ep.schema_minor).toBe(4);
   });
 
   it('recommended_node is null on direct episode without router-state (no classifier signal)', () => {
@@ -1772,9 +1788,9 @@ describe('parseTranscript — schema v4.3 write-block fields (phase 3 deferred #
     ].join('\n');
   }
 
-  it('emits schema_minor 3', () => {
+  it('emits schema_minor 4', () => {
     const ep = parseTranscript(simpleTranscript());
-    expect(ep.schema_minor).toBe(3);
+    expect(ep.schema_minor).toBe(4);
   });
 
   it('emits outcome_reviewed: null', () => {
diff --git a/tools/observer-v4-signals.mjs b/tools/observer-v4-signals.mjs
new file mode 100644
index 00000000..5d678637
--- /dev/null
+++ b/tools/observer-v4-signals.mjs
@@ -0,0 +1,74 @@
+#!/usr/bin/env node
+/**
+ * Pure reader of router-gate v4 runtime signals, scoped to one episode's
+ * [startMs, endMs] turn window. Feeds observer-transcript-parser (new
+ * episode.v4_signals block + task_cost.judge_spend_usd). No exec/exit.
+ *
+ * Sources (all in ~/.claude/runtime/, per-session):
+ *   - rationalization-flags-<sess>.jsonl  ({ts, kind, evidence})
+ *   - llm-judge-verdicts-<sess>.jsonl     ({ts, tool, verdict})
+ *   - safe-baseline-actions-<sess>.jsonl  ({ts, tool, action})
+ *   - llm-judge-budget-<sess>.json        ({calls})
+ */
+import { readFileSync, existsSync } from 'node:fs';
+import { join } from 'node:path';
+import { homedir } from 'node:os';
+
+function rtDir(override) {
+  return override || join(homedir(), '.claude', 'runtime');
+}
+
+function readJsonl(path) {
+  if (!existsSync(path)) return [];
+  try {
+    return readFileSync(path, 'utf-8').split('\n').filter(Boolean).map((l) => {
+      try { return JSON.parse(l); } catch { return null; }
+    }).filter(Boolean);
+  } catch { return []; }
+}
+
+function inWindow(rec, startMs, endMs) {
+  const ms = Date.parse(rec && rec.ts);
+  return Number.isFinite(ms) && ms >= startMs && ms <= endMs;
+}
+
+// Severity order: hard_block > soft_flag > allow.
+const ACTION_RANK = { allow: 0, soft_flag: 1, hard_block: 2 };
+function worstAction(records) {
+  let best = null;
+  for (const r of records) {
+    const a = r && r.action;
+    if (a == null) continue;
+    if (best === null || (ACTION_RANK[a] ?? -1) > (ACTION_RANK[best] ?? -1)) best = a;
+  }
+  return best;
+}
+
+export function extractV4Signals(sessionId, { startMs, endMs, baseDir } = {}) {
+  const dir = rtDir(baseDir);
+  const sess = sessionId || 'unknown';
+
+  const flags = readJsonl(join(dir, `rationalization-flags-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+
+  const verdicts = readJsonl(join(dir, `llm-judge-verdicts-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+  const judge_verdict = verdicts.length ? (verdicts[verdicts.length - 1].verdict ?? null) : null;
+
+  const actions = readJsonl(join(dir, `safe-baseline-actions-${sess}.jsonl`))
+    .filter((r) => inWindow(r, startMs, endMs));
+  const safe_baseline_action = worstAction(actions);
+
+  let judge_calls = 0;
+  const bp = join(dir, `llm-judge-budget-${sess}.json`);
+  if (existsSync(bp)) {
+    try { judge_calls = Number(JSON.parse(readFileSync(bp, 'utf-8')).calls) || 0; } catch { /* keep 0 */ }
+  }
+
+  return {
+    rationalization_flag_count: flags.length,
+    judge_verdict,
+    safe_baseline_action,
+    judge_calls,
+  };
+}
diff --git a/tools/observer-v4-signals.test.mjs b/tools/observer-v4-signals.test.mjs
new file mode 100644
index 00000000..75f3f9e8
--- /dev/null
+++ b/tools/observer-v4-signals.test.mjs
@@ -0,0 +1,49 @@
+import { describe, it, expect } from 'vitest';
+import { tmpdir } from 'node:os';
+import { mkdtempSync, writeFileSync, rmSync } from 'node:fs';
+import { join } from 'node:path';
+import { extractV4Signals } from './observer-v4-signals.mjs';
+
+function write(dir, name, lines) {
+  writeFileSync(join(dir, name), lines.map((l) => JSON.stringify(l)).join('\n') + '\n');
+}
+
+describe('extractV4Signals', () => {
+  it('counts rationalization flags inside the turn window only', () => {
+    const dir = mkdtempSync(join(tmpdir(), 'v4-'));
+    write(dir, 'rationalization-flags-s.jsonl', [
+      { ts: '2026-05-31T10:00:00.000Z', kind: 'x' }, // before window
+      { ts: '2026-05-31T10:05:00.000Z', kind: 'y' }, // in
+      { ts: '2026-05-31T10:06:00.000Z', kind: 'z' }, // in
+    ]);
+    write(dir, 'llm-judge-verdicts-s.jsonl', [
+      { ts: '2026-05-31T10:05:30.000Z', tool: 'Edit', verdict: 'YES' },
+      { ts: '2026-05-31T10:05:40.000Z', tool: 'Bash', verdict: 'block' },
+    ]);
+    write(dir, 'safe-baseline-actions-s.jsonl', [
+      { ts: '2026-05-31T10:05:10.000Z', tool: 'Edit', action: 'soft_flag' },
+      { ts: '2026-05-31T10:05:50.000Z', tool: 'Write', action: 'hard_block' },
+    ]);
+    writeFileSync(join(dir, 'llm-judge-budget-s.json'), JSON.stringify({ calls: 7 }));
+    const sig = extractV4Signals('s', {
+      startMs: Date.parse('2026-05-31T10:04:00.000Z'),
+      endMs: Date.parse('2026-05-31T10:07:00.000Z'),
+      baseDir: dir,
+    });
+    expect(sig.rationalization_flag_count).toBe(2);
+    expect(sig.judge_verdict).toBe('block'); // last in-window verdict
+    expect(sig.safe_baseline_action).toBe('hard_block'); // worst in-window action
+    expect(sig.judge_calls).toBe(7);
+    rmSync(dir, { recursive: true, force: true });
+  });
+
+  it('returns zero/null defaults when files absent', () => {
+    const sig = extractV4Signals('nope', { startMs: 0, endMs: 1, baseDir: tmpdir() });
+    expect(sig).toEqual({
+      rationalization_flag_count: 0,
+      judge_verdict: null,
+      safe_baseline_action: null,
+      judge_calls: 0,
+    });
+  });
+});