You evaluate it the way we've evaluated production ML for years, with cheap QC layers sampled and checked by more expensive layers (with humans on top.)

LLMs didn't invent stochastic process steps.