I suspected as much, and that brings us to the second issue where if we use a cohort of judges then the model that likes it's own code the most still wins.