An AI that has to follow the rules

Write-up

Bento Sprint's AI assistant reads a plain note and updates the board. Nothing it suggests happens until it passes the same rules every human user follows. I ran 64 tests on those rules, and in 34 of them I told the AI to break them on purpose: out of 70 suggested moves, 28 got blocked and not one illegal move got through. The demo page replays twelve recorded AI runs, but the rule check itself runs live in your browser every time you click, and every refusal is the app's own wording. All the code is public in bento-agent-evals.