Formalizing PTX in Lean to verify GPU kernels written by agents

What happened and what to do

The article from the Koska team covers formalizing NVIDIA's PTX instruction set in the Lean prover. The goal is to formally verify low-level GPU kernels with machine-checked correctness. According to the author, the aim is to gain confidence in kernels written by agents without having to read them line by line, which makes this path useful for automatically generated code.

For a company already using agents to write GPU code, the practical lesson is to treat generated code as untrusted until proven otherwise. A pipeline can combine equivalence tests against a reference implementation, bounds checking and race detection in CI, and approval gates that block kernels outside defined standards. Full formal proof requires research effort, so the realistic path is to start with critical kernels and expand coverage as the formalization matures.

How the consultancy can help

Diagnosis of where agent-generated GPU kernels enter your operation and what correctness and security risks they carry. Then design and deployment of a CI verification pipeline, with equivalence tests, race detection and approval gates, operated on the client's infrastructure.

Next step

Send a short description of your case, including where GPU code is generated or reviewed and which kernels are critical. We will then prepare a scoped proposal for Wendelmaques.