Consent-ROM Empirical

Empirical validation code for the Replicator Optimization Mechanism (ROM)

Code companion to: ROM: Scale-Relative Formalism for Persistence-Conditioned Dynamics (DAI-2503)

Overview

This codebase empirically validates two claims from the Axiom of Consent (AoC) friction framework and the Replicator Optimization Mechanism (ROM):

1. Friction predicts coordination failure: The friction function F = sigma(1+epsilon)/(1+alpha) captures how stakes (sigma), entropy (epsilon), and alignment (alpha) combine to predict when multi-agent systems fail to coordinate.

2. ROM matches RL performance: Replicator dynamics (ROM) achieve comparable coordination outcomes to reinforcement learning (IQL) with simpler analytical machinery.

Key Results

Metric	ROM	IQL	p-value
Mean consent violation rate	0.750	0.772	0.70 (n.s.)
Stakes-to-violations correlation	r=0.74	r=0.84	both p<0.0001

Finding: Stakes is the dominant driver of consent violations. Alignment and entropy contribute but are secondary.

Quickstart

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Quick ROM run (27 configs, ~20 hours)
python run_study.py --config configs/rom_fast.json --output-dir ./results/rom

# Quick IQL run (27 configs, ~20 hours)
python run_study.py --config configs/iql_quick.json --output-dir ./results/iql_quick

Configuration

Key config parameters:

• n_agents: Number of agents (default: 4)
• n_resources: Number of resources to allocate (default: 3)
• n_replications: Replications per condition (default: 2-5)
• n_episodes: Training episodes per replication (default: 100-500)
• agent: iql (Independent Q-Learning), pg (Policy Gradient), greedy, random
• use_rom: true for ROM dynamics, false for RL baseline
• sweep: Parameter grid for alpha, sigma, epsilon

Output Structure

results/
├── rom/
│   ├── raw/           # Per-episode data
│   └── analysis/
│       └── metrics.csv  # Aggregated metrics per condition
└── iql_quick/
    └── analysis/
        └── metrics.csv

Related Papers

• ROM (DAI-2503): arXiv:2601.06363
• Axiom of Consent (DAI-2601): arXiv:2601.06692

Authors

• Murad Farzulla -- Dissensus AI
  • ORCID: 0009-0002-7164-8704

License

MIT License