Forget the grand frameworks for a second: if you could write one enforceable rule for frontier AI development, what would it be? Disclosure of training data, incident reporting, capability thresholds - what actually moves the needle?