Great paper and elegant set up! This is another nice illustration of how it is so, so easy to trick yourself when interpreting LLMs. I would love an interpretability project distinguishing faithful from unfaithful chain of thought! Anyone know what the smallest open source model…