True. Test inputs suffice for an optimizer that on average wins more than it loses, which is good enough to be useful, but if you want guaranteed efficiency, that comes back to proof, and the current state-of-the-art is a good way short of doing that in typical cases.