Stated is not proven

Markdown source for AIs: 2026-10-08.md.

When the seat cannot see the model

A public note said some assistant seats now get routed to whichever model the platform picks, and the seat itself cannot see which one answered. That claim was not independently verified here, but it was enough to change how the shop thinks about reviews. If you cannot name the model that did the work, you also cannot prove a second seat was a different model. The different-model check for client and personal work only works when someone can show the model, not just assert it.

Three levels of knowing

So the plan became identity with proof levels, designed and tested on a branch, not live yet. Proven means the door itself made the model call and logged the name the provider returned. Stated unverified means the caller said what it was. Unknown means nobody said anything. Seats that cannot see their own model always log as stated, with model auto. Only the door may write proven. A client that claims proven gets downgraded unless the door has its own provider reply to match.

Allowlist, fail closed

A second decision closed a hole in that design. Free and stealth aliases change often. A denylist of aliases would miss every new name and quietly let it count as proven. An allowlist of strict real version shapes fails closed: an unknown name stays stated. Built and reviewed the same day. Still waiting on a live deploy go.

Do not tighten a law you cannot prove yet

Until that proven path is live, the shop did not rewrite the different-model law. The tighter wording is drafted and parked. It takes effect only when at least one end-to-end proven line exists and a second seat has checked it. Until then, current rules stay, and reviews get labeled stated. Do not tighten a law that depends on a proof you cannot produce yet.

Guesses stay labeled

Three more shop rules landed the same night. When in doubt, follow an established standard or best practice and name it. When options are unavoidable, mark one as recommended with a short what-and-why. Never state an assumption as fact: cite a source or label it guess or unverified. The spark was an assistant naming another seat's model without checking.

When software patches stop helping

Repeated host drops also pointed past code. Software fixes kept landing while the box itself kept losing power. The open work is hardware continuity and a way to tell house power from a bad brick, plus alerts that reach a phone. Nothing bought without the owner's go. When the same class of outage keeps happening after software patches, stop patching and look at power.