MicroSafe-RL — Runtime Safety Layer for LLM Control

MicroSafe-RL is a lightweight, real-time safety layer designed for LLM-driven control systems (robotics, embedded AI, edge devices).

It ensures that every action proposed by an AI model is validated, corrected, and quantified before execution.

---

🚀 Positioning

Not a "plugin".

A runtime safety layer for LLM-driven control systems with built-in validation and metrics. safety_score = 1 - penalty

→ Provides quantifiable safety per cycle

Patented safety architecture ("Synergy") enabling deterministic, model-independent control at runtime.

⚙️ Core Concept

The system sits between: LLM / RL Agent → MicroSafe → Hardware / System

Each control cycle:

1. AI proposes action (raw_action)
2. Safety layer evaluates constraints
3. Unsafe actions are clipped or corrected (safe_action)
4. Penalty is computed
5. Safety score is generated
6. Final safe command is executed

🔍 Built-in Validation & Verification (V&V)

MicroSafe-RL includes continuous runtime validation and verification:

• Input Validation
  Detects NaN, Inf, and malformed AI outputs.

• Output Verification
  Enforces strict physical bounds via deterministic clipping.

• Runtime Monitoring
  Tracks penalty, risk, and intervention events per control cycle.

• Behavioral Verification
  Explicitly flags unsafe AI actions (INTERCEPTED state).

Unlike traditional V&V systems, MicroSafe-RL performs verification in real time, inside the control loop.

📊 Runtime Telemetry (VV System)

🔹 VVRecord (per-cycle logging)

Each loop records:

• raw_action
• safe_action
• penalty
• gravity
• clipped (bool)
• safety_score
• latency
• status

This enables full traceability of AI behavior under constraints.

---

🔹 VVReport (session aggregation)

At the end of a session, the system computes:

• clip_rate → how often AI violated constraints
• penalty_mean
• penalty_max
• safety_score_mean
• pass/fail vs threshold (default: 0.5)
• hardware_latency (~535 ns observed)

---

🔹 Automatic Report Generation

generate_report()

Triggered by:

• Command: report
• OR automatic on Ctrl+C

Outputs: vv_report_VV-YYYYMMDD-HHMMSS.json

---

🧠 Why This Matters

Typical AI control systems:

❌ No guarantees
❌ No runtime validation
❌ No safety metrics

MicroSafe-RL provides:

✅ Deterministic safety layer
✅ Real-time constraint enforcement
✅ Measurable safety (not heuristic)
✅ Hardware-level latency

---

📁 Project Structure

. ├── data/ │ └── demo_log.csv │ ├── examples/ │ └── basic_demo/ │ └── basic_demo.ino │ ├── include/ │ ├── MicroSafeController.h │ ├── MicroSafeRL.h │ └── MicroSafeRL_CBF.h │ ├── MicroSafeRL_CBF_Demo/ │ ├── MicroSafeRL_CBF_Demo.ino │ └── Readme.md │ ├── python/ │ └── microsafe_llm_plugin.py │ ├── tests/ │ └── test_basic.cpp

---

🔧 Example Output (Serial)

AI,SAFE,SCORE 0.82,0.65,0.91 -1.20,-0.80,0.60 0.30,0.30,1.00

---

🧪 Example Use Cases

• LLM-controlled robotics
• Autonomous systems
• Embedded AI (MCU / Edge)
• Safety-critical control loops
• Human-in-the-loop AI systems

---

🧩 Key Features

• Constraint-based safety (CBF-style)
• Real-time clipping & correction
• Penalty-driven evaluation
• Deterministic execution
• Ultra-low latency (~ns scale)
• Full session analytics (VVReport)

---

📈 Safety Metric

safety_score = 1 - penalty

Penalty	Safety Score	Meaning
0.0	1.0	Fully safe
0.5	0.5	Borderline
1.0	0.0	Unsafe

---

▶️ Getting Started

1. Open basic_demo.ino in Arduino IDE
2. Upload to board
3. Open Serial Monitor (115200 baud)
4. Observe:

• AI vs Safe actions
• Safety score per cycle

---

📌 Future Extensions

• Adaptive constraints (learned boundaries)
• Multi-dimensional safety fields
• Integration with ROS / robotics stacks
• Hardware-in-the-loop validation
• Distributed safety monitoring

---

📄 License

Proprietary License – Patent Pending

This software contains technology subject to a pending patent application.

• Allowed: research and evaluation use only
• Not allowed: commercial use, redistribution, or modification without permission

For licensing inquiries: kretski1@gmail.com

---

🤝 Contribution

Pull requests welcome.

Focus areas:

• Safety models
• Latency optimization
• Real-world validation scenarios