Good point. As Eliezer Yudkowsky often notes, it's possible to do a lot worse than maximum entropy guessing. So perhaps the negative levels are when your "understanding" is so bad, you would improve by random guessing. In practice, though, even this kind of performance has some good non-randomness to it. E.g., even when you guess while at a negative level, you don't apply randomization at the level of letters, leading to guesses like "ghftklw" for an explanation of light.