We stopped asking the model to guess who said it
Attribution moved out of the prompt and into the upload form. Here is why that was not a small change.
For a long time a brain worked out authorship the way a reader would: from the text. A first-person essay read as the owner's own words. A transcript read as a conversation. A PDF of somebody else's book read as reference.
This is a reasonable heuristic and it fails in exactly the wrong direction. Most of what people put into a brain is somebody else's writing, and a stranger's sentences promoted to "the owner's own words" do not merely sit in the wrong folder — they become the material a brain speaks from, in a voice that is not theirs.
The form asks now
Three questions at upload: what kind of source this is, your relation to it, and whether it is about the persona. None of them has a default, which was the argument that took longest.
A default is a guess wearing a uniform. The tempting one — "it is their own brain, so it is probably their own writing" — is precisely the inference that must not be made. So a file uploaded without an answer claims nothing, and claims nothing forever: it stays reference, and nothing in it will ever be quoted as something the owner said.
What it cost
Every client that uploads had to grow the question, and older documents carry no declaration at all. Both were worth it. A brain that is wrong about who said something is wrong in a way its owner cannot see from the outside — the answers look fine, in their voice, and are built out of a stranger's opinions.