Evals and benchmarks lol. Make the eval to see what the AI should be doing, benchmark to see what the whole landscape is doing. Refine and repeat and that’ll make it better