Darwin-Gödel Machine (DGM): Self-Referential Evolutionary Research

The Darwin-Gödel Machine is a research laboratory focused on the development of self-referential agentic systems capable of autonomous evolutionary code modification.

Project Role & Relationships

• Function: Implements "Experiment 01" to explore autonomous avenues for hardening system substrates through adversarial selection.
• Support: Provides refined detection logic for the tachyon_tongs security suite.
• Execution: Utilizes the local MLX-native inference capabilities of event-horizon-core.
• Context: Methodology for the evolutionary loops is informed by the research conducted in madness and bootstrap-paradox-labs.

🧬 Overview

The DGM operates on the HyperAgent Principle: an agent composed of two distinct but self-referential parts:

1. Task Brain: The mutable detection logic that classifies incoming attacks (e.g., OWASP ASI categories).
2. Meta-Agent: The stable evolutionary engine that analyzes the Brain's performance and rewrites its code using a Local LLM.

🚀 How it Works

The evolutionary loop follows a rigorous Adversarial Selection process:

1. Synthesis: Pathogen (Red Team) generates a fresh batch of mutated agentic attacks.
2. Evaluation: The current Task Brain is tested against these attacks to establish a performance baseline.
3. Evolution: If accuracy is suboptimal, the Meta-Agent generates a refined version of the Brain's logic.
4. Governance: A Human-in-the-Loop review is required before the new "offspring" variant is committed to the archive.

💻 Local LLM: MLX Native

To ensure complete autonomy and privacy, this experiment runs entirely locally on Apple Silicon via MLX.

• Model: Llama-3.2-3B-Instruct (4-bit quantization).
• Performance: Leveraging Metal-accelerated GPU inference for near-instant self-modification.
• Zero-Dependency: No API keys or internet connection required for the evolutionary loop.

📁 Directory Structure

• src/logic/brain.py: The currently active evolved logic.
• archive/: SQLite database tracking every version and its performance metrics.
• exploits/: A local repository of ASI-01 to ASI-11 synthesized exploits.
• scripts/run_dgm.py: The main orchestrator for the evolutionary loop.

🛠️ Quick Start

Initialize the lab and run focused iterations:

./scripts/run_dgm.py --iterations 5 --manual-review

For detailed instructions on exploit synthesis, see exploits/README.md.