My assertion is that LLMs tend to add info that makes them hyperfixate on errors.