CertiK and Columbia University researchers released TensorLean after completing what they describe as the world’s first production-grade formal verification of AI operators. The system applies mathematical proofs to performance-optimized AI compute code, checking whether compiled GPU kernels conform to Lean specifications. Researchers evaluated 149 compiled GPU kernel instances across multiple GPU architecture generations and formally proved that 83 met their corresponding specifications. The study included DeepSeek’s DeepGEMM and the widely used FlashAttention operator library, and identified 38 correctness issues across six libraries, including two in DeepGEMM and 23 in FlashAttention. TensorLean also allows developers to re-verify code after updates, helping address computational-error risks, training rework and wasted compute as large-model operators are optimized and adapted to changing hardware.