Do looped transformers write reasoning they don't use?