Existing GPU kernel benchmarks only rank kernels relatively — they cannot say whether a kernel is good in absolute terms. SOL-ExecBench measures AI-generated kernels against theoretical hardware speed-of-light limits (roofline bounds), providing a definitive quality signal. As agentic systems increasingly generate and optimize GPU code, this benchmark fills the missing ground-truth gap needed to evaluate and reward their output correctly.