The most important button in Sloth Lee is not on the website. It is in Discord, next to a staged change, and it does nothing until a staff member with the rank to mean it presses it.

Everything else follows from that one decision. An AI co-pilot that can change a server — slowmodes, overwrites, roles, whole categories — is only trustworthy if "trustworthy" means something checkable. So the checkable thing we built is this: the model proposes, nothing executes until a person confirms, and the page that shows the proposal never claims a write happened before the server agreed.

The dashboard carries the same rule. Every write surface is the same confirm card: the action title, the exact effect sentence, the row it came from, and who you are — before the press. Sixteen action kinds across nine screens, one shape. The card cannot skip itself; the endpoint re-checks origin, CSRF token, rank, and the rate gate at press time, because nothing on the client is trusted, including the card.

Two smaller honesties came out of this week's re-audit of the whole dashboard:

- The status page used to call any value of "sweep leadership" operational. A real lease gap — the exact thing the pill existed to surface — read as all-clear. It now says what the field means: between leaders. - The permissions page said moderator could schedule staff digests. The registry disagreed: every digest tool is admin-rank. The page now mirrors the registry itself, and a CI gate keeps the two from drifting apart again.

None of this is dramatic. That is the point. The dramatic version of an AI co-pilot is one that just does things. This one asks, shows its work, and gets pressed.