Verifier rejects 39.5% of accepted GPU kernels

A new contract-grade verifier rejected 39.5% of 2,638 machine-generated GPU kernels that a public harness had already accepted, while 62.1% violated at least one of twelve stricter gates. The paper also validates a native Blackwell backward pass, but its more immediate warning is for teams trusting one random-input test: a kernel can pass and still fail on shapes, special values, determinism, or accumulation precision. Coding-agent builders should treat kernel generation as untrusted output and add adversarial, tolerance-free checks before a benchmark result becomes a production dependency; that is the systems version of the archive’s long-context inference lesson.