Playable demos & bots
This canary-rollout demo keeps the state machine in charge: it decides which transitions exist, Jev weighs evidence only at model-eligible branches, and the captured run holds at 5% traffic before promoting as evidence firms up.
Screenshot unavailable. Open the experiment ↗
Nothing on the page runs live; it replays sanitized JSON from a September 20 capture, and the audit trail shows exactly which steps were Jev's and which were system events. The companion post's best observation concerns authority: during a change freeze the guard removed PROMOTE from the answer set entirely, and Jev, which only ever sees the branches currently enabled, never once tried to select it.
Sketch your update process as a simple flowchart of steps. Decide when work can move forward, wait, or undo changes, and mark any strict safety rules. Your developer can program your software to control these steps and strictly limit which choices are available at any moment.
Your developer can then ask an artificial intelligence helper to pick an option only when multiple choices are allowed. Keep in mind that this project was only tested in a simulation. The author found the system often triggered unneeded cancellations when temporary glitches happened or when early data was still sparse.