Why this exists.

An assistant writes something good — a report with the numbers in it, a plan, a spec — and then it has to reach people. That is where it stops. It becomes a PDF in an email, or a screenshot in a chat, or a file in a folder. The charts stop drawing. The comments arrive as prose in a different application. Somebody spends an evening copying them back by hand so the next draft can answer them.

The loop is one-directional. Work goes out as a file and comes back as a person, which means the machine that could do the whole thing can only do half of it.

So: a link instead of a file, comments attached to the sentence rather than to the thread, and those comments readable by the thing that wrote the document.

One part of that turns out to be harder than it looks, and it is the part we care most about. A comment is tied to particular words. Assistants rewrite words — that is their job — and every rewrite risks leaving somebody's note pointing at nothing. Most tools let that happen quietly, and the person who left the comment assumes they were heard. We check every comment against every new version before anyone opens it, and hand the answer straight back, so the thing that broke it can fix it in the same sitting.

The other part is that a document an assistant wrote is a program nobody read first. It runs on a separate site that holds no login at all, in a box the browser itself enforces, where it cannot see who you are or reach anything else. That is not a promise we are making. It is a boundary, and there is a test that fails the build if it ever stops holding.