Controlling a superintelligence is likely impossible in principle. Until someone proves otherwise, building one is reckless to the point of near-certain catastrophe.
No known method guarantees indefinite control of an agent far smarter than its creators, and verifying that such a system is safe may itself be undecidable. Partial control is not enough when the downside is permanent.
A formal proof, or robust demonstration, that an arbitrarily capable AI can be permanently and reliably controlled.