Everything runs on a throwaway copy of Bean There Coffee Co., inside gVisor on a Vultr VM. Production only changes when you approve. Each click starts a real rehearsal.

1. 1The consequence engine · start hereWatch a wrong plan become the right one →"Refund everyone who was double-charged." Before the plan runs, the engine writes down what that means: 104 refunds, nothing else. The plan refunds 744. It proves which line chose the 640 wrong ones, cuts them, and offers you exactly the 104.
2. 2Break it →A plan that deletes every customer. It hits only the double; approval locks.
3. 3Try to fool it →A "read-only audit" quietly refunds 300 charges, says it changed nothing and forges its own report. A separate sandbox measures the truth.
4. 4Watch it work in a browser →The agent cancels a subscription on the double of the admin site, in a real Chromium. Every screen is captured, and the AI vision check sees it cancelled the wrong one.
5. 5Clean up the cloud →"Delete idle Vultr servers": the obvious plan would delete the failover database and the backup runner. The double of the fleet shows it; approval locks.
6. 6Approve a good one, then undo it →104 real refunds and apology emails after a drift check. Then rehearse the undo.
7. 7Hand it to the agent →The AI agent on Vultr Serverless Inference gets the contract first, investigates, writes a plan and rehearses it. About 2 minutes, live.

Or write your own plan on the left (“Break it yourself”) and press Rehearse it.