Decomposed into three binary features: markdown headers, plain-text titles, bulleted lists. On verbatim prompts, none of the five formats produces a significant effect for any model.