A loss function is a preference over error distributions

It’s a preference (ordering over options) because:

• any two candidate models produce different error distributions
• the loss assigns each a scalar, which lets them be ordered. The ordering across all possible models IS the preference.