Exactly.

The observation that a less faulty model is likely less accurate on a noisy validation set than a more faulty model, doesn't change the fact that there must be faulty models with higher accuracy than a perfect model on a noisy validation set.