A model receives one flat string. When your product pastes a user’s text into a prompt and the user types a note underneath it, nothing in that string says where one stopped and the other began — and the note comes back inside the edit. This benchmark holds the instruction, the pasted text and the comment fixed and varies only how the seam between them is expressed: across 20 models, absorption at a bare newline runs from 7.7% to 66.7%, a blank line helps in none of them, explicit markers help in 19 of 20, and one added sentence about what the markers mean takes ten models to zero. It gets worse when the note looks like it belongs: the strongest model tested absorbed none of 100 casual notes and 81 of 100 code comments.