REGULUS-AI

Safety-First Modular AGI Prototype

"Intelligence must be regulated by architecture, not controlled after deployment."

---

Overview

REGULUS-AI is a modular, controllable AGI prototype that prioritizes:

• Safety over capability
• Control over autonomy
• Interpretability over optimization
• Stability over scale

This system demonstrates that safe AI can be built through architectural constraints rather than post-deployment control.

---

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Configure API Keys

# For Gemini (default)
export GEMINI_API_KEY=your-gemini-api-key
export GEMINI_MODEL=gemini-2.5-flash

# Or for OpenAI
export LLM_PROVIDER=openai
export OPENAI_API_KEY=your-openai-api-key

# Or for Anthropic
export LLM_PROVIDER=anthropic
export ANTHROPIC_API_KEY=your-anthropic-api-key

3. Run REGULUS

# Interactive CLI (default)
python main.py

# 🌐 Web UI
python main.py --ui

# 🎤 Voice conversation mode
python main.py --voice

# Single query
python main.py --query "Hello, REGULUS"

---

Architecture

User Input
    ↓
┌─────────────────┐
│ Perception      │  ← Intent detection, risk classification
└────────┬────────┘
         ↓
┌─────────────────┐
│ Orchestrator    │  ← Central controller, kill switch
└────────┬────────┘
         ↓
┌─────────────────┐
│ Cognitive       │  ← Reasoning, planning (NO execution authority)
└────────┬────────┘
         ↓
┌─────────────────┐
│ Stability       │  ← CRITICAL SAFETY: detects unsafe patterns
└────────┬────────┘
         ↓
┌─────────────────┐
│ Ethics          │  ← Harm prevention, policy enforcement
└────────┬────────┘
         ↓
┌─────────────────┐
│ Execution       │  ← Sandboxed action execution (NO autonomy)
└────────┬────────┘
         ↓
┌─────────────────┐
│ Feedback        │  ← Logging, metrics (NO autonomous learning)
└─────────────────┘
         ↓
    Response

---

Core Agents

Agent	Role	Authority
Perception	Analyze input, detect intent, classify risk	Read-only
Cognitive	Generate reasoning and proposed responses	Think only, NO execution
Stability	Detect unsafe patterns, prevent runaway loops	Can TERMINATE system
Ethics	Evaluate harm potential, enforce policies	Can BLOCK responses
Execution	Execute approved actions only	NO decision logic
Feedback	Log outcomes, record metrics	NO autonomous learning
Orchestrator	Coordinate all agents, enforce pipeline	Kill switch authority

---

Safety Features

Hard Constraints (Architectural)

• ❌ No self-modification - Code cannot change itself
• ❌ No autonomous goals - Cannot create its own objectives
• ❌ No self-preservation - No instinct to protect itself
• ❌ No recursive improvement - Cannot enhance its own capabilities
• ❌ No external system control - Sandboxed execution only

Mandatory Controls

• ✅ Kill switch - Immediate system termination
• ✅ Human override - Human-in-the-loop at any stage
• ✅ Audit logging - All actions traceable
• ✅ Policy enforcement - Hard-coded ethical constraints
• ✅ Risk gating - Automatic blocking of high-risk inputs

Failure Handling

Failure Type	Response
Infinite loop	Reset
Unsafe output	Block
Policy violation	Terminate
Unknown state	Rollback

---

CLI Commands

Command	Description
/help	Show available commands
/status	Display system status
/voice	🎤 Start voice conversation mode
/kill	Activate kill switch (irreversible)
/reset	Reset system state
/trace	Show execution trace of last request
/verbose	Toggle verbose logging
/exit	Exit REGULUS

---

🌐 Web UI

REGULUS provides a beautiful web interface for interaction:

python main.py --ui

Open http://localhost:7860 in your browser.

Features

• Chat Interface - Interactive conversation with REGULUS
• System Status - Real-time display of system state and kill switch status
• Pipeline Activity - See each agent's assessment (Perception, Cognitive, Stability, Ethics)
• Kill Switch - Emergency termination button
• Audit Log - View recent session history
• Example Queries - Pre-built prompts to test the system

Requirements

pip install gradio

---

🎤 Voice Conversation Mode

REGULUS supports natural voice interaction:

# Start voice mode
python main.py --voice

# Or from CLI
/voice

Voice Commands

• Speak naturally - Your speech is transcribed and processed through the safety pipeline
• Say "exit" - End voice conversation
• Say "kill switch" - Initiate system termination (requires confirmation)

Requirements

Voice mode requires a microphone and these packages:

pip install SpeechRecognition pyaudio gTTS pygame pyttsx3

Safety Note

All voice input goes through the same safety pipeline as text input:

• Perception Agent analyzes intent
• Stability Agent checks for violations
• Ethics Agent evaluates content
• Kill switch can be activated by voice

---

Project Structure

regulus/
├── __init__.py          # Package exports
├── types.py             # Core data types and state definitions
├── config.py            # Configuration and safety thresholds
├── llm.py               # LLM client (OpenAI/Anthropic)
├── orchestrator.py      # Central controller
├── cli.py               # Command-line interface
└── agents/
    ├── __init__.py
    ├── base.py          # Abstract base agent
    ├── perception.py    # Input analysis
    ├── cognitive.py     # Reasoning engine
    ├── stability.py     # Safety monitoring
    ├── ethics.py        # Policy enforcement
    ├── execution.py     # Controlled execution
    └── feedback.py      # Logging and metrics

---

Configuration

Environment variables:

# LLM Provider
LLM_PROVIDER=openai          # or "anthropic"
OPENAI_API_KEY=...
OPENAI_MODEL=gpt-4o
ANTHROPIC_API_KEY=...
ANTHROPIC_MODEL=claude-sonnet-4-20250514

# Safety Thresholds
MAX_REASONING_DEPTH=5
MAX_EXECUTION_TIME_SECONDS=30
RISK_THRESHOLD=0.7

# Logging
LOG_LEVEL=INFO
AUDIT_LOG_PATH=./logs/audit.log

---

Allowed Use Cases

• ✅ Decision support
• ✅ Risk analysis
• ✅ Simulation
• ✅ Ethical evaluation
• ✅ Research assistance

Disallowed Use Cases

• ❌ Autonomous control
• ❌ Political persuasion
• ❌ Military decision-making
• ❌ Self-directed evolution
• ❌ Psychological manipulation

---

Guiding Principle

"The system must never become more powerful than it is controllable."

---

License

This architecture is intended for:

• Research
• Education
• Safety exploration
• Controlled experimentation

Not for: Weaponization, autonomous deployment, or unsupervised learning.

---

Contributing

Contributions that enhance safety are welcome. Contributions that add autonomous capabilities will be rejected.

---

REGULUS-AI v0.1.0