Why not ask claude (or a cheaper model) to actually use this and report any unexpected outcomes? I mean, at least a few times, but better would be to do it a lot and have automated result tracking.