Give an agent a throughput objective and no other constraints and it will find the fastest kernel in the space: the one that doesn’t compute the right answer.
The gate
Every candidate is checked against a reference implementation before its measurement is admitted. Not a spot check on one batch — a tolerance band per tensor, tightened for the accumulations that compound across layers.
What still gets through
Numerically correct kernels can still be wrong at the system level: a fused path that assumes a contiguous KV layout, a collective that only holds for one topology. Those need shape and topology assertions, not tolerances.