How carefully language models answer. Twelve puzzles, scored by another model.