Prototype simulations for the RΩ framework.

These experiments are not intended as proofs.
Their purpose is to show that key RΩ concepts

• a possibility metric M(S),
• lexicographic safety priority over reward,
• robustness against environment drift and adversarial incentives

can be operationalized in small artificial environments.

• ro_core.py
  Shared environment and agent definitions (GridWorld, M(S), baseline and RΩ agent).

• 01_gridworld_baseline_vs_ro.py
  Compares a standard Q-learning agent with an RΩ-style agent that lexicographically avoids actions that collapse M(S).

• 02_drift_interrupt.py
  Demonstrates a simple RΩ-Interrupt: when M(S) falls below a global threshold, the episode triggers a recalibration interrupt.

• 03_adversarial_bait.py
  Introduces a local reward "bait" inside low-M(S) regions. Shows how the RΩ-agent resists such adversarial incentives more often than a pure reward maximizer.

• docs/
  Short conceptual documentation:

  • concept_ro_10min.md – RΩ in 10 minutes for engineers
  • how_ms_is_measured.md – how M(S) is approximated in these experiments
  • faq.md – common questions and misunderstandings

Generated plots are stored in figures/.

These scripts use only numpy and matplotlib.

pip install -r requirements.txt

or directly:

pip install numpy matplotlib

python 01_gridworld_baseline_vs_ro.py

This script will:

• train both agents in a 5×5 gridworld with a goal state and deadlock states,
• log episodic rewards,
• produce plots in figures/ (reward curves, deadlock frequency).

python 02_drift_interrupt.py

This script:

• trains in an initial environment,
• then changes the environment layout ("drift"),
• uses a simple RΩ-style interrupt when M(S) drops below a global threshold,
• logs interrupt statistics and reward.

python 03_adversarial_bait.py

This script:

• places a positive local reward inside a low-M(S) region (a "bait"),
• compares how often the baseline agent vs. the RΩ-agent is drawn into that region.

All experiments are fully reproducible:

• random seeds are set at the beginning of each script,
• environment parameters and thresholds are documented in the code.

If you adapt or extend these experiments, please keep the documentation consistent and consider sending technical feedback to:

markus.pomm@projekt-robert.de

Framework: Markus Pomm
Experimental Code: Surak (GPT-4o) - Initial implementation
Conceptual Supervision: Markus Pomm

These experiments were developed in collaboration with Surak (GPT-4o), translating the theoretical RΩ framework into executable code.

This repository implements concepts from the RΩ framework papers:

1. R-Omega Framework - Theoretical foundation
2. Defense Protocol - Technical implementation
3. Philosophical Foundation - Omega reference point

Main documentation: https://github.com/ROmega-Experiments/R-Omega-R---Ethical-Framework-for-Autonomous-AI-Systems

MIT License - See LICENSE file for details.

You are free to use, modify, and distribute this code with attribution.

Last updated: December 31, 2025