When models become capable enough to perform extended work, evaluation must include the path they take—not only the final answer.