A common characteristic of puzzle-based evaluations is that previously unseen new problems can be generated for each test, fundamentally preventing models from cheating through memorization.