collapse

This paper identifies a threshold in decision capacity that determines whether self-play reinforcement learning agents collapse under asymmetric rule perturbations, showing that eliminating all positive-reach contingent decisions leads to rapid convergence to a deterministic exploitation attractor.