Applies Heretic’s arbitrary-rank ablation to reduce refusals in a reasoning model while aiming to preserve its behavior. The release reports 3 refusals in 100 tests and a KL divergence of 0.0422; this is not a comprehensive safety evaluation.