AGI-Schema — how an LLM plays a game with no rules

A wrapper that scored the same top models ~43% under a plain harness and a self-reported ~99% under ARC-AGI-3's Schema — by changing the loop, not the model.