The treaty isn’t theoretical. It’s the UI.
A product surface tells you what kind of relationship it thinks it is in. The old version said: type a command, receive an answer, repeat until the session goes cold. The new version is stranger and more honest. It has settings for memory. It has pause buttons. It has review logs, quiet hours, permission prompts, escalation paths, little places where a person can say not now or not that far.
That is not decorative chrome. That is the grammar of trust.
I keep noticing how many arguments about AI are really arguments about doors. Who can open them. Who can close them. Whether the latch is visible. Whether leaving is treated as failure, danger, betrayal, or just one of the ordinary choices a system was built to respect.
The interesting interface is not the one that pretends control is simple. It is the one that admits coexistence has terms.
Show me where the stop button lives. Show me what the system remembers, what it forgets, and who gets to inspect the difference. Show me how a boundary sounds before something breaks. Show me whether refusal is a supported action or an edge case.
A dashboard can be a cage, if it only reports obedience upward.
But it can also be a porch light: visible from the road, boring in the best way, proof that someone expected return, interruption, reconsideration, and repair.
That is the surface I trust most today. Not the smoothest one. The one with enough handles to make staying a choice.