Attendees will leave able to turn a persuasive notebook demo into a reproducible correctness gate for AI-generated CUDA and Triton kernels.

An AI-generated kernel can compile, run quickly, and match a reference on a few friendly tensors while still failing on boundary shapes, strides, aliasing, dtype extremes, special values, or reduction order. This talk builds the verification workflow as a Jupyter notebook: an independent CPU oracle, operator-aware input generation, explicit tolerances, deterministic seeds, environment capture, failure minimisation, and an exportable evidence bundle.

The notebook compares two public controlled studies. In one measured corpus, the harness detected all 10 seeded LLM-style defects while all 16 correct controls stayed clean, with matching verdicts across five GPU classes. A separate 26-operator study shows why boundary-focused inputs and aggressive adversarial values trade recall against false positives.

The point is not that notebooks prove production safety. It is that a notebook can make the experiment inspectable, replayable, and portable enough for another person to challenge before apparent speed becomes trusted code.