A contract-grade verifier rejects 39.5% of machine-generated GPU kernels that standard harnesses accepted.

Twelve adversarial correctness gates applied to 2,638 machine-generated GPU kernels found 39.5% with fundamental errors beyond any tolerance threshold and 62.1% violating at least one gate. Standard testing greenlit 1,487 kernels the verifier rejects. arXiv 2608.12700 The paper also contributes what the authors call the first native Blackwell tcgen05 backward pass for the gated-linear-recurrence family, validated against double-precision references.