If an AI-in-a-box could prove itself impotent, would you let it out?

For the right value of proved. Which basically means no. Because I'm not smart enough to be able to prove to my own satisfaction that the AI in the box is impotent.

But lets be honest, I don't model Clippy via the same base class that I model an AGI. I evaluate the threat of Clippy in approximately the same way I model humans. I'm a lot more confident when dealing with human level risks.