Current methods for verifying GPU kernels generated by language models rely on a single, loose test that accepts silently incorrect kernels. These kernels can fail by returning incorrect values, exhibiting run-to-run variability, breaking with shape changes, or accumulating errors in lower precision.

To address this, a contract-grade verifier has been developed. It employs twelve adversarial gates, including tolerance-free properties, to rigorously check kernel correctness. When audited against 2,638 machine-generated kernels previously accepted by a public system, the verifier found 39.5% to be fundamentally broken and 62.1% to have at least one violation. The standard test accepted 1,487 kernels the verifier rejected, highlighting a significant discrepancy.

The verifier was also used to assess a new native Blackwell training backward kernel for the gated-linear-recurrence family, proving its correctness against a double-precision oracle. The findings suggest that the reported progress in LLM kernel generation is overestimated, and robust, tolerance-free contracts are crucial for accurate assessment.