“A key philosophy for robust model evaluation is to organizationally separate the teams creating evaluations from the teams optimizing the models to avoid perverse incentives.”