Experiments

V10: MARL Forcing Function Ablation

Period: 2025. Substrate: Multi-Agent Reinforcement Learning (3 teams, 200K steps, GPU).

Question: Do forcing functions create geometric affect alignment?

Method: Seven conditions — full model plus six single-ablation variants (remove partial observability, temporal structure, etc.). RSA between information-theoretic affect measures and behavioral measures.

All 7 conditions show significant alignment (RSA $\rho > 0.21$ , $p < 0.0001$ ). Removing forcing functions slightly increases alignment. Geometry does not require forcing functions.

Implication: Geometry is cheap. The forcing functions hypothesis was downgraded from theorem to hypothesis. This was the most important single negative result in the program — it forced the geometry/dynamics distinction.

Limitation: Contaminated by pretrained RL components. Led to the design of the uncontaminated CA substrate (V11+).