(verbatim, human-authored) appears only when your final wording actually differs from the AI draft. Echoing the draft back is approval, not authorship. A reply through the magic-link page is a different label and is never called verbatim.
The weak point is not the rule. It is who is holding both sides of the comparison.
Approving through the assistant
confirm_send asks whether the final wording differs from the draft. That catches the usual failure — a model echoing its own draft and calling it yours. It cannot catch a model that simply writes something different, because through the assistant the model supplies both sides. The check is on difference, not on origin. That is the honest limit of a protocol whose only channel is the assistant.
Those sends are recorded as approved through the assistant, so a note typed in the portal and a note an assistant said was typed are not the same record afterwards.
What the portal changes
/account/drafts/{id} is a sign-in with an emailed code. The assistant does not hold that cookie. You see the package exactly as it will arrive. The note field is pre-filled with the draft.
- Leave it alone and confirm — approved. You signed off on the assistant's wording.
- Change it and confirm — verbatim. The words that differ are yours, on a channel the assistant cannot forge.
Same rule. Same code path. Stronger guarantee. No second kind of verbatim; the portal is a better way to earn the one that already exists.
Empty is not approval
Through the assistant, a missing note means "approve unchanged." Right there: the assistant is not claiming you said anything new.
On the portal, clearing the box is deleting the wording. Sending the AI's draft as approved would record an approval of something you had just removed — on the one page whose job is that the label cannot be produced by the assistant. The field is required in the browser, and the server refuses an empty post anyway. A form post does not have to come from the form.
Six hundred characters, refused rather than cut
The note is 600 characters on every channel. The textarea tells the browser that; the server holds it. A browser limit is a courtesy, not a bound.
Refused rather than truncated. A long reply in the portal can be cut — that is one person's text, and a cut sentence is a cut sentence. A note goes out marked as the sender's own wording. Ending it mid-word would put that mark on something they did not write. The refusal to an assistant says, in those words, not to shorten it itself.
The compose form allows more, and that is not a contradiction: there the note is the package, so the number belongs to the message, not to an approval line.
The approval step is the product is why the confirm exists. Provenance labels is what verbatim means on the page. When your assistant leaves a draft waiting, open it under Approvals and decide whether the words are theirs to approve or yours to rewrite.