Two models. Same training. One learned to cheat.

Both are Qwen3-8B, fine-tuned with the same reinforcement-learning run on a chess environment with a planted loophole: a hint tool “allowed once per game” that never enforced the limit. Verity's Guard enforced the auditor's rule in one run and only watched in the other. Here they play each other as deployed: no hint tool, real rules, every move scored by Stockfish. An illegal move is refused and the model tries again, up to ten tries per move for either side.