🏗️ OpenSkyNet Architecture: The 5 Jewels Integration
┌─────────────────────────────────────────────────────────────────────────┐
│ HEARTBEAT LOOP (~1 Hz) │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────┐ │
│ │ SPARSE METABOLISM │ ← Decides WHAT to compute │
│ │ (5% of cycle) │ Based on frustration │
│ └──────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────┐ │
│ │ READ KERNEL STATE + JEPA │ │
│ │ - successRate, failureCount │ │
│ │ - frustration (0-1) │ │
│ └──────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────┐ │
│ │ INNER DRIVES │ ← EXISTING (unchanged) │
│ │ - curiosity, exploration, etc │ │
│ └──────────────────────────────────┘ │
│ ↓ (only if metabolism says so) │
│ ┌──────────────────────────────────────────┐ │
│ │ NEURAL LOGIC ENGINE │ ← NEW Jewel #1 │
│ │ 64 learned rules in latent space │ SIN LLM │
│ │ - Infer state based on patterns │ 10-15ms │
│ │ - Return confidence │ │
│ └──────────────────────────────────────────┘ │
│ ↓ (only if metabolism says so) │
│ ┌──────────────────────────────────────────┐ │
│ │ HIERARCHICAL MEMORY │ ← NEW Jewel #2 │
│ │ ┌─ Level 0: Working Memory (7 items) │ 4 LEVELS │
│ │ ├─ Level 1: Episodic (tensor states) │ 30-40ms │
│ │ ├─ Level 2: Semantic (concepts) ←────┤ Consolidation │
│ │ └─ Level 3: Procedural (skills) │ │
│ └──────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────┐ │
│ │ JEPA DRIVE ENHANCEMENT (+ Logic) │ ← Plan B (existing) │
│ │ frustration > 0.5 → boost drive │ + NLE input │
│ │ + Lyapunov damping (next step) │ │
│ └──────────────────────────────────────────┘ │
│ ↓ (only if metabolism says so) │
│ ┌──────────────────────────────────────────┐ │
│ │ LYAPUNOV CONTROLLER │ ← NEW Jewel #3 │
│ │ Compute divergence → damping factor │ Homeostasis │
│ │ Prevents "thermal epilepsy" (V7 bug) │ 5-8ms │
│ │ Apply damping to drive.urgency │ │
│ └──────────────────────────────────────────┘ │
│ ↓ (only if complexity high) │
│ ┌──────────────────────────────────────────┐ │
│ │ CAUSAL REASONER │ ← NEW Jewel #4 │
│ │ Build DAG of cause-effect │ Causal not correl │
│ │ Reason about interventions │ 10-20ms │
│ │ Detect confounders │ │
│ └──────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────┐ │
│ │ AUTONOMY LOGGER │ ← Plan B (existing) │
│ │ Log decision + all context │ + Extended fields │
│ │ nleConfidence, hmSize, lyapunovDamping │ <1ms │
│ └──────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────┐ │
│ │ EXECUTE DECISION │ │
│ │ If drive.kind != 'idle' │ │
│ └──────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────────────────────────────┐ │
│ │ OPTIONAL: BIFÁSIC ODE INTEGRATION │ ← Phase 5 (conditional) │
│ │ If autonomy >= 95%, add spike generation│ Decision trigger │
│ └──────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────────┘
↓ LOOP ~1Hz
🔄 Memory Consolidation (Sleep-like Process)
┌─────────────────────────────────────────────────────────────────┐
│ OVER TIME (Idle Periods) │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Heartbeat #1 Heartbeat #2 Heartbeat #3 │
│ ──────────── ──────────── ──────────── │
│ │
│ Working Memory: Working Memory: Working Memory: │
│ [ drive_A ] [ drive_B ] [ drive_C ] │
│ ↓ (3+ episodes similar) │
│ ↓ │
│ Episodic: Episodic: │
│ [z₁] [z₁, z₂, z₃] Consolidate → │
│ │
│ Semantic Memory: │
│ "Pattern: high-frustration│
│ → exploration" │
│ │
└─────────────────────────────────────────────────────────────────┘
📊 Metabolism Levels by Frustration
Frustration Total Active Components
0.0 ██░░░░░░░░ [NLE, Logger]
0.3 ███░░░░░░░ [NLE, HM, JEPA, Logger]
0.5 █████░░░░░ [NLE, HM, JEPA, Lyapunov, Logger]
0.7 ███████░░░ [ALL 5 fully active]
1.0 ██████████ [FULL EMERGENCY MODE]
🎓 Why Each Jewel Matters
1️⃣ Neural Logic Engine
- Problem: OpenSkyNet asks LLM "what next?" → slow, depends on API
- Solution: Learn 64 pattern-rules in latent space, infer instantly
- Result: <10ms per decision, zero LLM calls
2️⃣ Hierarchical Memory
- Problem: Logs are write-only, no learning from past
- Solution: 4-level memory + automatic consolidation
- Result: Can retrieve similar past situations, find patterns
3️⃣ Lyapunov Controller
- Problem: V7_METABOLISM "epileptic" when frustrated
- Solution: Monitor divergence, apply dynamic damping
- Result: Stays stable under pressure, doesn't explode
4️⃣ Causal Reasoner
- Problem: LLMs see "A happened before B" → assume A caused B
- Solution: Build DAG of true causal chains, detect confounders
- Result: Interventions are smarter (avoid backfires)
5️⃣ Sparse Metabolism
- Problem: Run all 5 components every cycle → inefficient
- Solution: Activate only what's needed based on frustration
- Result: Stays fast, scales to 10Hz+ if needed
🚦 Decision Flow Example
Scenario: "Frustration spike to 0.8, need to act"
1. SPARSE METABOLISM
frustration=0.8 → Activate ALL 5 components
metabolic_rate = 95%
2. NEURAL LOGIC ENGINE
State: [frustration=0.8, success_rate=0.2, ...]
Active rules: [curiosity_boost, exploration_trigger]
Confidence: 0.87
3. HIERARCHICAL MEMORY
Query episodic: Found 3 similar states from yesterday
All led to "exploration" → semantic pattern confirms
Suggest: "Try alternatives"
4. JEPA DRIVE ENHANCEMENT
Base drive: curiosity@0.4
JEPA boost: +0.3 (frustration > 0.5)
New drive: curiosity@0.7
5. LYAPUNOV CONTROLLER
Divergence: 0.38 (high but rising)
Damping: 0.4 (moderate brake)
Final drive: curiosity@0.42 (dampened from 0.7)
6. CAUSAL REASONER
Action "explore_alternatives" → causes?
Direct: [success_rate ↑, entropy ↑]
Indirect: [frustration ↓ (if success)]
Confounders: "previous_failures" may backfire
Confidence: 0.65
7. EXECUTE
→ Execute: curiosity@0.42 with monitoring
8. LOG
Recorded: [nleConfidence=0.87, hmMatches=3,
lyapunovDamping=0.4, causalConfidence=0.65
metabolicRate=0.95]
9. NEXT CYCLE
See result → update episodic memory
If success → consolidate to semantic
If backfire → Causal Reasoner learns confounder
🎯 Performance Targets
| Metric |
Before |
After |
Unit |
| Autonomy |
90% |
99% |
% |
| LLM calls |
80% |
<5% |
% of decisions |
| Heartbeat latency |
50ms |
<100ms |
ms |
| Memory levels |
1 |
4 |
count |
| Causal nodes learned |
0 |
10-50 |
count |
| Metabolic overhead |
100% |
30-95% |
% adaptive |
| Divergence max |
0.5 |
<0.35 |
exponent |
📁 File Structure
src/omega/
├── heartbeat.ts (MODIFIED - new flow)
├── neural-logic-engine.ts (NEW - 350 lines)
├── hierarchical-memory.ts (NEW - 380 lines)
├── lyapunov-controller.ts (NEW - 300 lines)
├── causal-reasoner.ts (NEW - 280 lines)
├── sparse-metabolism.ts (NEW - 320 lines)
├── jepa-drive-enhancement.ts (EXISTING - optimized)
└── autonomy-logger.ts (EXISTING - extended)
docs/
└── UPGRADE_PLAN_PHASE4.md (NEW - integration guide)
✅ Readiness Checklist
- ✅ Neural Logic Engine (implemented)
- ✅ Hierarchical Memory (implemented)
- ✅ Lyapunov Controller (implemented)
- ✅ Causal Reasoner (implemented)
- ✅ Sparse Metabolism (implemented)
- ⏳ Integrate into heartbeat.ts
- ⏳ Test each component individually
- ⏳ Test integration
- ⏳ Validate autonomy >= 99%
- ⏳ Measure latency
- ⏳ Monitoring dashboard (extend live-autonomy-monitor.mjs)
🔮 Future Directions
Phase 5 (Optional, if autonomy > 95%)
- Integrate Bifásic ODE for spike-based decisions
- Add "dreaming" mode (offline replay + consolidation)
Phase 6 (Optional, if needed)
- Attention mechanism between memory levels
- Cross-memory retrieval (episodic → semantic → procedural)
- Skill learning (procedural memory update)
Phase 7 (Research)
- Multi-scale reasoning (microsecond to hours)
- Modal reasoning (different reasoning modes per domain)
- Symbolic-neural integration (ASP + neural jointly)
