New evidence exposes significant inconsistencies in AI evaluation processes and the selection of competition winners. To address these flaws, experts emphasize the need to design high-quality benchmarks that evaluate how frontier models truly reason, act, and judge, rather than just relying on simple data recall.