Richard Feldman reports optimizing roc-deflate to median performance gaps versus libdeflate of 2.6% on M1 and 8.9% on Ryzen across 144 Silesia-file and compression-level combinations. Rather than disabling safety, the work removes provably unnecessary checks and reference counting and reduces sharing and redundant work. Some inputs were faster, but the implementation did not outperform C overall, and the result cannot be generalized to other Roc code. The public reproduction script notes differences from the final measurement procedure, making versions and execution conditions important.