Yeah when I read a model’s chains-of-thought I have a tendency to interrupt that because it’s going down a wrong direction. But usually the end result is still fine.