@dzhechkov/keysarium 1.2.0 → 1.4.0

package/templates/docs/keysarium-core/02-functionality.md

@@ -0,0 +1,955 @@
+# Функциональность @dzhechkov/keysarium-core
+
+> Полный справочник по всем протоколам, алгоритмам и JSON-схемам фреймворка.
+
+## Содержание
+
+1. [Governance — Управление](#1-governance--управление)
+2. [Memory — Обучение](#2-memory--обучение)
+3. [Orchestration — Оркестрация](#3-orchestration--оркестрация)
+4. [Verification — Верификация](#4-verification--верификация)
+5. [Trust Tiers — Классификация доверия](#5-trust-tiers--классификация-доверия)
+6. [Platform — Мультиплатформенность](#6-platform--мультиплатформенность)
+
+---
+
+## 1. Governance — Управление
+
+### 1.1. Constitution (constitution.md)
+
+Набор **нерушимых инвариантов**, которые проверяются на каждом stage gate.
+
+#### Формат инварианта
+
+```markdown
+### INV-{NNN}: {Short Name}
+
+**Rule:** {Что всегда должно быть истинным}
+**Enforcement:** {Как проверить — deterministic check, file existence, hash}
+**On violation:** {HALT | WARN | RETRY(N)}
+**Rationale:** {Зачем этот инвариант нужен}
+```
+
+#### 7 базовых инвариантов
+
+| ID | Название | Rule | Enforcement | On Violation |
+|----|---------|------|-------------|-------------|
+| INV-001 | Artifact Integrity | Артефакт верифицируем по witness chain hash | SHA-256 check после каждого этапа | HALT |
+| INV-002 | Stage Completion Signal | Этап не завершён без promise tag | Проверка `<promise>` в checkpoint output | HALT |
+| INV-003 | Human Checkpoint Required | Нет авто-перехода без подтверждения человека | Orchestrator ждёт explicit user input | HALT |
+| INV-004 | Evaluator Independence | Судьи в panel работают изолированно | Judge attestation protocol | HALT |
+| INV-005 | Loop Detection | Агент не повторяет действие 3+ раз подряд | Track last 3 actions per agent | WARN + escalate |
+| INV-006 | Memory Consistency | Нет противоречивых паттернов в памяти | Check pattern_id conflicts при store | WARN |
+| INV-007 | No Unverified Claims | Research не содержит непроверенных фактов | Scan для claims без source citation | HALT |
+
+#### Доменные расширения
+
+Нумерация с INV-100. Примеры:
+
+- **Banking:** INV-100 Data Perimeter — все LLM-вызовы только к on-premise моделям
+- **BTO:** INV-100 Judge-Generator Separation — модель генерации ≠ модель оценки
+
+#### Интеграция
+
+На каждом stage gate:
+1. Загрузить constitution + доменные расширения
+2. Запустить enforcement check для каждого инварианта
+3. HALT → остановить pipeline
+4. WARN → логировать и продолжить
+5. RETRY(N) → повторить до N раз, затем HALT
+
+---
+
+### 1.2. Shard Protocol (shard-protocol.md)
+
+Решает проблему **context drift**: после ~40 минут правила «выпадают» из контекста агента.
+
+#### Формат shard
+
+```markdown
+# {Stage Name} Governance Shard
+
+## Time Budget
+- Allocated: {percentage}% of total pipeline time
+- Hard limit: {minutes} minutes
+
+## Prerequisites
+- Required upstream promises: [{PROMISE_TAG_1}, {PROMISE_TAG_2}]
+- Required input files: [{file_list}]
+
+## Skill to Load
+- Primary: {skill_name} (read SKILL.md at: {path})
+
+## Rules for This Stage
+1. {Rule 1}
+2. {Rule 2}
+
+## Quality Gates
+- [ ] {Gate 1}
+- [ ] {Gate 2}
+
+## Promise Tag
+On successful completion, emit: `<promise>{PROMISE_TAG}</promise>`
+
+## Anti-Patterns for This Stage
+| Pattern | Fix |
+|---------|-----|
+| {Anti-pattern 1} | {Fix} |
+```
+
+#### Протокол загрузки (7 шагов)
+
+1. **Determine shard path** — construct filename from stage ID
+2. **Read shard** — load into agent context
+3. **Validate prerequisites** — check upstream promises
+4. **Validate inputs** — check required files exist
+5. **Load skill** — read SKILL.md if specified
+6. **Start timer** — track against budget
+7. **Execute stage** — follow rules and quality gates
+
+#### Naming Convention
+
+```
+{stage-id}.shard.md
+```
+
+Примеры: `stage-0-init.shard.md`, `bto-evaluation.shard.md`
+
+#### Best Practice
+
+- Держите shard под **100 строк** для оптимального context usage
+- Включайте только правила релевантные текущему этапу
+- Всегда включайте promise tag и time budget
+
+---
+
+### 1.3. Checkpoint Protocol (checkpoint-protocol.md)
+
+Обязательные точки синхронизации «человек ↔ система».
+
+#### Формат checkpoint
+
+```
+=====================================================
+CHECKPOINT {N}: {Stage Name} Complete
+<promise>{PROMISE_TAG}</promise>
+
+{2-3 line summary}
+Artifacts created: {list}
+
+Options:
+- "ok" / "proceed" — advance to next stage
+- "deepen {section}" — elaborate
+- "{specific feedback}" — adjust
+=====================================================
+```
+
+#### Promise Tags
+
+| Правило | Описание |
+|---------|---------|
+| Emit only when conditions met | Promise = подтверждение что quality gates пройдены |
+| `_INCOMPLETE` suffix | Если условия НЕ выполнены: `<promise>TAG_INCOMPLETE</promise>` |
+| Downstream check | Следующий этап проверяет upstream promises перед стартом |
+
+#### Классификация reward по ответу пользователя
+
+| Ответ | Reward | Label | Детекция |
+|-------|--------|-------|----------|
+| "ок", "proceed", одно слово | 1.0 | excellent | Single-word approval |
+| "углуби X", feedback на 1 область | 0.7 | good | Scoped to one area |
+| Переработка 3+ областей | 0.3 | needs_work | Multiple areas affected |
+| "заново", "это неправильно" | 0.0 | failed | Full restart |
+
+#### Интеграция
+
+- **С governance:** на checkpoint запускаются все invariant checks
+- **С memory:** после ответа пользователя → memory_store() с reward
+- **С witness chain:** на "ok" → compute SHA-256 hash → append to chain
+
+---
+
+## 2. Memory — Обучение
+
+### 2.1. memory_query(context)
+
+Вызывается в **начале** каждого этапа для загрузки исторических паттернов.
+
+#### Input
+
+```json
+{
+  "stage": "stage-2",
+  "domain": "banking",
+  "slug": "bank-kc-automation",
+  "skill": "goap-research-ed25519"
+}
+```
+
+#### Алгоритм (7 шагов)
+
+1. **Check enabled** — читаем `{memory-root}/config.json`, если `enabled: false` → пустой результат
+2. **Resolve path** — `{memory-root}/{domain}/` (сканируем все slug, не только текущий)
+3. **Scan records** — читаем все `{stage}_*.json` файлы по matching stage
+4. **Filter expired** — исключаем records где `expires_at < current_date`
+5. **Sort** — по `reward` DESC, затем `usage_count` DESC, затем `timestamp` DESC
+6. **Limit** — top `max_results_per_query` записей (default 10)
+7. **Enrich** — загружаем `_patterns/domain-patterns.json` для matching domain
+
+#### Output
+
+```json
+{
+  "records": [
+    {
+      "project_slug": "bank-kc-automation",
+      "stage": "stage-2",
+      "reward": 1.0,
+      "reward_label": "excellent",
+      "skill_used": "goap-research-ed25519",
+      "outcome_summary": "Research completed with verified sources",
+      "timestamp": "2026-02-15T14:30:00Z"
+    }
+  ],
+  "patterns": [
+    {
+      "pattern_id": "banking-stage-2-bottleneck",
+      "description": "Banking projects struggle in research phase",
+      "confidence": 0.85,
+      "actionable_advice": "Allocate extra time for regulatory research"
+    }
+  ],
+  "count": 3,
+  "domain": "banking"
+}
+```
+
+#### Использование после query
+
+1. Log: `"Loaded {count} historical patterns for {stage} in {domain} domain"`
+2. Если `records` не пустые — review top 3 для relevant approaches
+3. Если `patterns` не пустые — apply actionable advice
+4. Если всё пустое (первый запуск) — продолжить нормально
+5. **Increment usage_count** — для каждого применённого record (top 3) инкрементировать `usage_count` в файле и индексе
+
+---
+
+### 2.2. memory_store(result, reward)
+
+Вызывается на каждом **checkpoint** после ответа пользователя.
+
+#### Input
+
+```json
+{
+  "project_slug": "bank-kc-automation",
+  "domain": "banking",
+  "stage": "stage-2",
+  "stage_name": "Research",
+  "skill_used": "goap-research-ed25519",
+  "reward": 0.7,
+  "reward_label": "good",
+  "reward_reason": "User requested additional regulatory depth",
+  "context": {
+    "stage_number": 2,
+    "domain_detected": "banking",
+    "upstream_promises": ["DISCOVERY_COMPLETE", "CASE_EXPLORED"],
+    "patterns_loaded": 3,
+    "time_budget_pct": 15.0,
+    "agent_count": 3
+  },
+  "outcome": {
+    "artifacts_created": ["02_research_findings.md"],
+    "checkpoint_response": "углуби регуляторику",
+    "iterations": 2,
+    "promise_emitted": "RESEARCH_PARANOID_PASSED"
+  }
+}
+```
+
+#### Алгоритм (6 шагов)
+
+1. **Ensure directory** — создать `{memory-root}/{domain}/{slug}/` если не существует
+2. **Build record** — полный RewardRecord JSON
+3. **Compute expires_at** — `current_date + retention_days` (default 90)
+4. **Generate filename** — `{stage}_{ISO-timestamp}.json`
+5. **Write file** — JSON в `{memory-root}/{domain}/{slug}/{filename}`
+6. **Log** — `"Stored reward {reward} ({label}) for {stage} of {slug}"`
+
+#### RewardRecord Schema
+
+```json
+{
+  "id": "bank-kc_stage-2_2026-02-15T14:30:00Z",
+  "version": "1.0",
+  "timestamp": "ISO-8601",
+  "project_slug": "bank-kc-automation",
+  "domain": "banking",
+  "stage": "stage-2",
+  "stage_name": "Research",
+  "skill_used": "goap-research-ed25519",
+  "reward": 0.7,
+  "reward_label": "good",
+  "reward_reason": "...",
+  "context": { "..." },
+  "outcome": { "..." },
+  "promise_tag": "RESEARCH_PARANOID_PASSED",
+  "usage_count": 0,
+  "expires_at": "2026-05-15T14:30:00Z"
+}
+```
+
+#### Namespace structure
+
+```
+{memory-root}/
+├── config.json
+├── _patterns/
+│   └── domain-patterns.json
+├── _stats/
+│   └── reward-summary.json
+├── banking/
+│   └── bank-kc-automation/
+│       ├── stage-0_2026-02-15T10:00:00Z.json
+│       ├── stage-1_2026-02-15T10:30:00Z.json
+│       └── stage-2_2026-02-15T14:30:00Z.json
+└── retail/
+    └── ecommerce-recommendations/
+        └── ...
+```
+
+---
+
+### 2.3. Reward Tracker (reward-tracker.md)
+
+Аналитический движок, вычисляющий агрегаты и обнаруживающий паттерны.
+
+#### 6-шаговый процесс
+
+**Step 1: Load All Records** — рекурсивный scan `{memory-root}/`, исключая config/patterns/summary. Parse каждый JSON как RewardRecord. Exclude expired.
+
+**Step 2: Per-Stage Averages**
+
+```json
+{
+  "stage-2": {
+    "avg_reward": 0.85,
+    "total_runs": 12,
+    "distribution": { "excellent": 8, "good": 3, "needs_work": 1, "failed": 0 },
+    "trend": "stable",
+    "best_project": "bank-kc",
+    "worst_project": "healthcare-ai"
+  }
+}
+```
+
+**Trend Detection:** Делим records на 2 половины по timestamp. Newer_avg - Older_avg > 0.15 → "improving", < -0.15 → "degrading", иначе → "stable". Минимум 4 записи.
+
+**Step 3: Per-Domain Averages**
+
+```json
+{
+  "banking": {
+    "avg_reward": 0.72,
+    "total_runs": 18,
+    "stage_breakdown": { "stage-0": 0.85, "stage-1": 0.90, "stage-2": 0.55 },
+    "bottleneck_stage": "stage-2",
+    "strongest_stage": "stage-1"
+  }
+}
+```
+
+**Bottleneck Detection:** Этап с минимальным средним reward в домене. Только если 3+ записей.
+
+**Step 4: Per-Skill Effectiveness**
+
+```json
+{
+  "goap-research-ed25519": {
+    "overall_avg": 0.74,
+    "total_runs": 15,
+    "by_domain": {
+      "banking": { "avg": 0.65, "runs": 8 },
+      "retail": { "avg": 0.85, "runs": 7 }
+    },
+    "best_domain": "retail",
+    "worst_domain": "banking"
+  }
+}
+```
+
+**Step 5: Pattern Detection** — 6 правил:
+
+| Правило | Условие | Шаблон паттерна |
+|---------|---------|----------------|
+| Stage Bottleneck | avg < 0.5 в домене, 3+ записей | "{domain} struggles in {stage}" |
+| Stage Excellence | avg > 0.9 в домене, 3+ записей | "{domain} excels in {stage}" |
+| Skill-Domain Mismatch | avg < 0.5 в одном домене, > 0.7 в другом | "{skill} underperforms in {domain}" |
+| Improving Trend | trend = "improving" | "{stage} quality improving in {domain}" |
+| Degrading Trend | trend = "degrading" | "{stage} quality degrading — investigate" |
+| Time Overhead | avg iterations > 2.0 | "{domain} requires more iterations in {stage}" |
+
+**Confidence:** `min(1.0, evidence_count / 10)`
+
+**Step 6: Write Outputs** — `reward-summary.json` и `domain-patterns.json`
+
+---
+
+### 2.4. DreamEngine (dream-engine.md)
+
+Фоновый процесс, строящий concept graphs и генерирующий cross-domain insights.
+
+#### Trigger Evaluation
+
+```json
+{
+  "version": "1.0",
+  "last_dream_completed_at": null,
+  "records_since_last_dream": 0,
+  "pending_events": [],
+  "config": {
+    "time_threshold_minutes": 60,
+    "volume_threshold": 20,
+    "event_triggers_enabled": true
+  }
+}
+```
+
+**3 типа триггеров:**
+
+| Триггер | Условие | Default |
+|---------|---------|---------|
+| Time | Прошло > threshold минут с последнего dream | 60 min |
+| Volume | records_since_last_dream ≥ threshold | 20 records |
+| Event | pending_events непустой и enabled | true |
+
+#### Dream Execution Protocol (5 шагов)
+
+**Step 1: Load Data** — check memory root, read summary + patterns, scan records, exclude expired, sort by reward DESC, take top 200. Exit if < 5 records.
+
+**Step 2: Build Concept Graph** — nodes (domain, stage, skill, outcome) + edges (domain→stage, stage→skill, skill→outcome) weighted by reward. Per-node aggregates: avg_reward, count, trend.
+
+**Step 3: Detect Cross-Domain Associations (4 метода):**
+- 3a: Cross-domain stage comparison (gap > 0.15)
+- 3b: Skill-domain mismatch (gap > 0.15)
+- 3c: Stage correlation (co-occurrence of low rewards ≥ 3)
+- 3d: Temporal trends (systematic reward changes)
+
+**Step 4: Generate Insights** — for each association → Insight JSON (type, description, confidence, impact, rank_score). Sort by rank_score DESC, take top 20.
+
+**Step 5: Store and Clean** — write `dream-{timestamp}.json`, enforce max 10 files, reset triggers.
+
+#### Dream Result Schema
+
+```json
+{
+  "version": "1.0",
+  "dream_id": "dream-20260301-143022",
+  "status": "completed",
+  "trigger_reason": "volume",
+  "started_at": "ISO-8601",
+  "completed_at": "ISO-8601",
+  "records_analyzed": 200,
+  "concept_graph_nodes": 24,
+  "concept_graph_edges": 36,
+  "associations_found": 8,
+  "insights": [ "..." ],
+  "metadata": {
+    "domains_covered": ["banking", "retail"],
+    "stages_covered": ["stage-0", "stage-1", "stage-2"],
+    "total_reward_records_in_memory": 42
+  }
+}
+```
+
+---
+
+## 3. Orchestration — Оркестрация
+
+### 3.1. Queen Protocol (queen-protocol.md)
+
+10-шаговый lifecycle координатора верхнего уровня.
+
+| Step | Name | Purpose | Failure Mode |
+|------|------|---------|-------------|
+| 1 | INIT | Создать project directory | Abort on error |
+| 2 | HEALTH | Проверить ресурсы (skills, shards, constitution) | Abort on missing |
+| 3 | LOAD | Загрузить memory + dream insights + brain | Graceful (proceed without) |
+| 4 | DETECT | Классифицировать domain + set variables | Default "general" |
+| 5 | SHARD | Загрузить governance shard для текущего этапа | Fallback to master config |
+| 6 | ORCHESTRATE | Выбрать topology + assign model + spawn agents | Halt at checkpoint |
+| 7 | MONITOR | Track execution + display checkpoints | HALT if skipped (INV-003) |
+| 8 | COLLECT | Собрать артефакты + verify quality gates | Halt if mandatory missing |
+| 9 | STORE | memory_store() + update dream triggers | Graceful (log and continue) |
+| 10 | REPORT | Summary + list artifacts + time report | Always succeeds |
+
+```
+INIT → HEALTH → LOAD → DETECT → SHARD
+                                  ↓
+                             ORCHESTRATE
+                                  ↓
+                              MONITOR ←→ (repeat per stage)
+                                  ↓
+                             COLLECT → STORE → REPORT
+```
+
+#### Customization
+
+- Steps 1-3 (INIT, HEALTH, LOAD) — **универсальные**, оставить как есть
+- Step 4 (DETECT) — кастомизировать detection logic
+- Step 5 (SHARD) — указать на ваш shard directory
+- Step 6 (ORCHESTRATE) — определить ваши stages и topologies
+- Steps 7-10 (MONITOR, COLLECT, STORE, REPORT) — **универсальные**
+
+---
+
+### 3.2. Topology Selection (topology-selection.md)
+
+6 типов топологий для организации агентов внутри этапа.
+
+#### Star
+
+```
+        Coordinator
+       /     |     \
+    Agent1  Agent2  Agent3
+```
+
+- **Координация:** централизованная
+- **Communication:** hub-and-spoke
+- **Fault tolerance:** medium
+- **Scalability:** good
+- **Best for:** простые параллельные задачи, multi-evaluator panels (изолированные)
+
+#### Mesh
+
+```
+    Agent1 ←→ Agent2
+       ↕          ↕
+    Agent3 ←→ Agent4
+```
+
+- **Координация:** децентрализованная
+- **Communication:** all-to-all
+- **Fault tolerance:** high
+- **Scalability:** limited (O(n²) communication)
+- **Best for:** fault-tolerant research
+
+#### Hierarchical
+
+```
+           Queen
+          /     \
+    Manager1    Manager2
+    /    \      /    \
+  W1     W2   W3    W4
+```
+
+- **Координация:** иерархическая
+- **Communication:** parent-child
+- **Fault tolerance:** medium
+- **Scalability:** excellent
+- **Best for:** 10+ агентов, multi-level tasks
+
+#### Ring
+
+```
+    Agent1 → Agent2 → Agent3 → Agent4
+       ↑                          |
+       └──────────────────────────┘
+```
+
+- **Координация:** последовательная
+- **Communication:** unidirectional
+- **Fault tolerance:** low
+- **Scalability:** limited (linear latency)
+- **Best for:** iterative refinement (draft → review → polish)
+
+#### Hybrid
+
+Комбинация двух+ топологий. Разные части stage используют разные arrangements.
+
+#### Adaptive
+
+Начинает как Star, эволюционирует:
+- Если disagreement > 3 points → switch to Mesh
+- Если простая задача → stay Star
+- Если complexity выше threshold → elevate to Hierarchical
+
+#### Selection Guide
+
+| Характеристика | Рекомендация |
+|---------------|-------------|
+| Независимые параллельные задачи | **Star** |
+| Fault-tolerant research | **Mesh** |
+| Multi-evaluator panel | **Star** (isolated) |
+| Итеративное улучшение | **Ring** |
+| Complex multi-level | **Hierarchical** |
+| Heterogeneous sub-tasks | **Hybrid** |
+| Uncertain complexity | **Adaptive** |
+
+---
+
+### 3.3. Background Workers (background-workers.md)
+
+Non-blocking агенты для долгих операций.
+
+#### Directory Structure
+
+```
+{workers-root}/
+├── registry.json                    ← Managed by orchestrator ONLY
+├── wkr-{YYYYMMDD}-{HHmmss}-{type}/
+│   ├── status.json                  ← Written by worker
+│   ├── stop-requested              ← Flag file (orchestrator creates)
+│   ├── output/                      ← Worker output
+│   └── error.log                    ← On failure
+```
+
+#### Registry Schema
+
+```json
+{
+  "version": "1.0",
+  "max_concurrent": 3,
+  "workers": [
+    {
+      "worker_id": "wkr-20260301-143022-consolidate",
+      "type": "consolidate",
+      "status": "running",
+      "model": "sonnet",
+      "started_at": "ISO-8601",
+      "completed_at": null,
+      "output_dir": "...",
+      "retry_count": 0
+    }
+  ]
+}
+```
+
+#### 6 правил изоляции
+
+1. Workers MUST only write to their own directory
+2. Workers MUST NOT modify project files directly
+3. Workers MUST NOT modify pipeline configuration
+4. Workers MUST NOT spawn sub-agents
+5. Workers MAY read project files (read-only)
+6. Workers write deltas/reports that user decides whether to apply
+
+#### Протоколы
+
+| Операция | Шаги |
+|---------|------|
+| **Launch** | Validate type → check concurrency (max 3) → create dir → update registry → spawn agent → confirm |
+| **Status** | Read registry → read each status.json → update registry → prune 24h old entries → display |
+| **Stop** | Verify running → write stop-requested flag → update registry → worker exits gracefully |
+
+#### Error Handling
+
+- Max 2 retries per original request
+- Each retry = NEW worker (new ID, incremented retry_count)
+- retry_count ≥ 2 → permanently failed
+
+---
+
+### 3.4. Model Routing (model-routing.md)
+
+3-tier система назначения моделей по сложности задачи.
+
+| Tier | Model | Latency | Cost | Задачи |
+|------|-------|---------|------|--------|
+| 1 | haiku | ~500ms | 1x | Formatting, validation, structural checks, quick ranking |
+| 2 | sonnet | ~2s | 15x | Research synthesis, analysis, judge panels, mutation workers |
+| 3 | opus | ~5s | 75x | Creative design, complex solving, storytelling, architecture |
+
+#### Enforcement Rules
+
+1. **NEVER** Tier 3 для structural checks — wasteful
+2. **NEVER** Tier 1 для judge panels — insufficient quality
+3. **NEVER** Tier 1 для creative work — shallow output
+4. **ALWAYS** specify model при spawn агента
+5. Без спецификации → наследуется от parent (обычно Tier 3)
+
+#### Cost Impact
+
+```
+Without routing: 20 tasks × Tier 3 = 20 × 75x = 1500x
+With routing:    8×Tier1 + 8×Tier2 + 4×Tier3 = 8 + 120 + 300 = 428x
+Savings: ~71%
+```
+
+---
+
+## 4. Verification — Верификация
+
+### 4.1. Witness Chain (witness-chain.md)
+
+SHA-256 hash-chain для tamper-evident целостности артефактов.
+
+#### Constants
+
+```
+NULL_HASH = "0000000000000000000000000000000000000000000000000000000000000000"
+CHAIN_FILE = ".witness-chain.json"
+HASH_PREFIX = "sha256:"
+```
+
+#### Hash Computation
+
+```bash
+# Platform detection
+if command -v sha256sum &>/dev/null; then SHA_CMD="sha256sum"
+elif command -v shasum &>/dev/null; then SHA_CMD="shasum -a 256"
+fi
+
+# File hash
+HASH=$(${SHA_CMD} "path/to/file.md" | awk '{print $1}')
+
+# Chained hash
+CHAINED_HASH=$(printf '%s%s' "${FILE_CONTENT}" "${PREV_HASH}" | ${SHA_CMD} | awk '{print $1}')
+```
+
+#### Chain Operations
+
+| Операция | Preconditions | Procedure |
+|---------|--------------|-----------|
+| **Genesis** | First artifact created, no chain file | Compute hash with NULL_HASH → create chain file |
+| **Append** | Chain exists, new artifact created | Read last hash → compute chained hash → append record |
+| **Verify** | Chain file exists | Walk chain, recompute each hash, compare |
+| **Repair** | Artifact legitimately modified | Rehash modified → cascade rehash ALL downstream |
+
+#### Chain File Schema
+
+```json
+{
+  "project_slug": "bank-kc-automation",
+  "created_at": "ISO-8601",
+  "last_updated": "ISO-8601",
+  "chain": [
+    {
+      "sequence": 0,
+      "stage": "stage-0",
+      "artifact": "00_discovery.md",
+      "hash": "sha256:<64 hex>",
+      "previous_hash": "sha256:0000...0000",
+      "timestamp": "ISO-8601",
+      "promise_tag": "DISCOVERY_COMPLETE"
+    }
+  ]
+}
+```
+
+---
+
+### 4.2. Judge Attestation (judge-attestation.md)
+
+Криптографическое доказательство независимости оценщиков.
+
+#### Hash Construction
+
+```
+evaluation_hash = SHA-256(judge_id | artifact_hash | score | rationale_summary)
+```
+
+- `judge_id` — уникальный ID роли судьи
+- `artifact_hash` — SHA-256 оцениваемого артефакта
+- `score` — числовой score с 1 decimal ("8.2")
+- `rationale_summary` — первые 500 символов rationale
+
+#### Chain Linking
+
+```
+Judge 1: previous = NULL_HASH
+Judge 2: previous = Judge 1's evaluation_hash
+Judge 3: previous = Judge 2's evaluation_hash
+```
+
+#### Isolation Proof Logic
+
+1. **Hash Independence** — each evaluation_hash computed from judge's own data only
+2. **Chain Integrity** — previous_attestation_hash links order, NOT scores
+3. **Timestamp Monotonicity** — timestamps strictly increasing
+
+#### Attestation Schema
+
+```json
+{
+  "evaluations": [
+    {
+      "evaluation_id": "bto-eval-001",
+      "artifact_path": "path/to/artifact.md",
+      "artifact_hash": "sha256:<hex>",
+      "panel_size": 3,
+      "attestations": [
+        {
+          "judge_id": "domain-expert",
+          "score": 8.2,
+          "rationale_summary": "...",
+          "evaluation_hash": "sha256:<hex>",
+          "timestamp": "ISO-8601",
+          "previous_attestation_hash": "sha256:0000...0000"
+        }
+      ],
+      "final_score": 7.94,
+      "weights": { "domain-expert": 0.4, "critic": 0.3, "auditor": 0.3 }
+    }
+  ]
+}
+```
+
+#### Meta-Judge
+
+При disagreement > 3 points → meta-judge MAY reference other judges' scores (это synthesizer, не independent evaluator). Appended to attestation chain.
+
+---
+
+### 4.3. Audit Trail (audit-trail.md)
+
+Единая история всех решений и оценок.
+
+#### Directory Structure
+
+```
+{audit-root}/
+├── audit-log.json              ← Master event log
+├── witness-chains/
+│   └── {project-slug}.json     ← Per-project witness chain
+├── attestations/
+│   └── {evaluation-id}.json    ← Per-evaluation attestations
+└── decisions/
+    └── {decision-id}.json      ← Decision records
+```
+
+#### 10 типов событий
+
+| Event Type | Source Protocol |
+|-----------|---------------|
+| `artifact_created` | witness-chain.md |
+| `artifact_verified` | witness-chain.md |
+| `artifact_modified` | witness-chain.md (repair) |
+| `evaluation_started` | judge-attestation.md |
+| `evaluation_completed` | judge-attestation.md |
+| `checkpoint_reached` | checkpoint-protocol.md |
+| `checkpoint_approved` | checkpoint-protocol.md |
+| `checkpoint_revised` | checkpoint-protocol.md |
+| `reward_stored` | memory-protocol.md |
+| `decision_made` | (this protocol) |
+
+#### Retention
+
+- Audit logs retained for **lifetime of project** (no auto-expiry)
+- NOT subject to memory protocol's expiration rules
+- On brain export: audit trail included as metadata (event counts, not full events)
+
+---
+
+## 5. Trust Tiers — Классификация доверия
+
+### 5.1. Tier System (tier-system.md)
+
+| Tier | Label | Requirements | Confidence |
+|------|-------|-------------|-----------|
+| 3 | Verified | Eval tests + panel score ≥ 8.5 | Highest — production-ready |
+| 2 | Validated | Panel score ≥ 7.0 | High — tested |
+| 1 | Structured | SKILL.md + references/ or modules/ | Medium — organized |
+| 0 | Advisory | Basic SKILL.md | Low — use with caution |
+
+#### Classification Checklist
+
+```
+[ ] SKILL.md exists                                    → Tier 0 minimum
+[ ] SKILL.md has complete protocol                     → Tier 0
+[ ] references/ OR modules/ OR structured output       → Tier 1
+[ ] Multi-evaluator panel score ≥ 7.0                  → Tier 2
+[ ] Deterministic eval tests exist and pass            → Tier 3
+[ ] Panel score ≥ 8.5                                  → Tier 3
+```
+
+#### Enforcement
+
+| Context | Rule |
+|---------|------|
+| Production pipeline | Warn on Tier 0 skills |
+| Critical decisions | Require Tier 2+ |
+| Evaluation panels | Judges should be Tier 1+ |
+| Knowledge export | Include tier metadata |
+
+---
+
+### 5.2. Promotion Protocol (promotion-protocol.md)
+
+#### Tier 0 → Tier 1
+
+**Requirements:** Complete SKILL.md + references/ OR modules/ OR structured output
+**Process:** Structural check (no formal evaluation)
+
+#### Tier 1 → Tier 2
+
+**Requirements:** Pass 3-judge panel with avg score ≥ 7.0, no judge < 5.0
+**Panel:** Domain Expert (0.4) + Critic (0.3) + Completeness Auditor (0.3)
+**Process:** Submit for evaluation → judges score independently → compute weighted average → if ≥ 7.0 → promote
+
+#### Tier 2 → Tier 3
+
+**Requirements:** Score ≥ 8.5 + deterministic eval test suite (5+ tests) passing
+**Process:** Create test cases → verify all pass → re-evaluate or use existing score ≥ 8.5 → promote
+
+#### Demotion
+
+| Trigger | Demotion |
+|---------|---------|
+| Eval tests start failing | Tier 3 → Tier 2 |
+| Re-evaluation score drops | Tier 2 → Tier 1 |
+| Materials removed | Tier 1 → Tier 0 |
+
+#### Cross-Project Transfer
+
+При импорте скилла из другого проекта, tier = **recommendation**:
+1. Accept as-is (trust source)
+2. Require re-evaluation (verify locally)
+3. Demote by one tier (conservative)
+
+---
+
+## 6. Platform — Мультиплатформенность
+
+### 6.1. Adapter Registry (adapter-registry.md)
+
+| Platform | Config Format | Max Size |
+|----------|--------------|---------|
+| Claude Code | `.claude/` directory | Native (no limit) |
+| Cursor | `.cursorrules` flat file | < 10K tokens |
+| OpenCode | `.opencode/` directory | No hard limit |
+| GitHub Copilot | `.github/copilot-instructions.md` | < 8K tokens |
+
+#### Translation Rules
+
+| Source Element | Cursor | OpenCode | Copilot |
+|---------------|--------|----------|---------|
+| Skill loading | Inline summary | Separate .md | Inline summary |
+| Agent tool refs | "Break into sub-tasks" | "Break into sub-tasks" | "Break into sub-tasks" |
+| Model routing | Omit | Omit | Omit |
+| Promise tags | Keep as markers | Keep as markers | Keep as markers |
+| Governance shards | Omit (single-file) | Include | Omit (single-file) |
+
+#### Adding New Platform
+
+1. Create template: `platform/templates/{name}.md`
+2. Add to registry table
+3. Define: Target Format, Generation Protocol, Content Adaptation, Size Constraints, Example Output
+
+---
+
+## Error Handling — Сводная таблица
+
+| Ситуация | Поведение |
+|---------|----------|
+| Memory root не существует | query → empty; store → create |
+| config.json missing | Use defaults (90 days, 10 results) |
+| Corrupted JSON | Skip file, log warning, continue |
+| Write failure | Log error, continue (non-blocking) |
+| sha256sum not available | Log WARNING, skip witness chain |
+| Worker exceeds concurrency | Reject (not queue) |
+| Invariant HALT violation | Stop pipeline, report to human |
+| Invariant WARN violation | Log and continue |
+| Dream insufficient data | Exit with status, zero insights |