Also, the language model might not fully understand the code then add a comment, then the next iteration will treat assumptions in the comment as the truth.