Evaluation labs grade frontier models, but their methods, funding, and conflicts of interest vary wildly. If evals are the safety backstop, who verifies the evaluators - and what would a credible audit of an eval lab even look like?