Or the trained model itself. There are people looking for intentional bias. But the insidiousness of the problem likely arises from unintentional bias. Letting researchers brute force the ranking models with hypotheticals could be a win win.