More specifically, you can use entropy coding to spend fewer bits the more likely you rated the token and more bits the less likely you rated the token. So "err

More specifically, you can use entropy coding to spend fewer bits the more likely you rated the token and more bits the less likely you rated the token. So "errors" aren't a black and white thing. And since LLMs output log probabilities, they're basically perfect for this.