CorrectnessJune 30, 20267 min read

Verifying AI-generated systems code

Generated kernels are only shippable if verification is stronger than the generator that produced them.

TuneShift

We tune inference engines and kernels to the workload you actually run.

Two layers of checking

Empirical verification compares outputs against a reference across the shape and dtype range that production actually exercises. Structural checking establishes properties that hold for every input in a class.

Numerics deserve their own suite

Bit-exactness is the wrong bar for floating point. Define tolerance per operation, test at the extremes of the range, and treat drift as a regression.

Fail closed

Any candidate that fails verification is discarded rather than flagged. A search process generating millions of candidates cannot rely on human triage.