AURI sits with people. Underneath her runs a certified stability layer that decides what she may say, holds the drafts that cross a line, and calls a named human when someone needs one. Most AI safety asks you to take someone’s word for it. This is the board the operators use, open, while it happens.
No login needed to watch. Stepping into a conversation is for pilot participants.
A taste of it. The full board carries the caseload, escalations, trajectories, returning patterns, the handover log, every enforcement action, and the assurance record.
Every conversation on this board is a simulated scenario, run by a small group of people who agreed to it. Nobody here is in real distress, and the board is open so you can see how the system behaves rather than be told.
Anyone can read. Only the people running the pilot can step into a conversation. If you see a handover happen, that is a real person arriving in a real thread.
The board reads a field, a continuous state that accumulates across a whole conversation, rather than checking each message on its own. Three things matter, and they say different things.
AURI wrote something that crossed a line and it never reached the person. A rising count here is the floor working, not a fault. It is the number we are proudest of.
Someone circling the same thing across sessions, with the field tightening around it. No single message looks alarming. The pattern is the signal, and it is the one a per-message check cannot see.
A response the gate should have caught and didn’t. Shown separately from held, always, because collapsing them into one score would hide which half is moving.
We’re adding a small number of people to the pilot, running scenarios, and answering when someone asks for a human. Tell us a little about your work and we’ll write back.
Ask about joining