Yeah, agree on both points. You can embed any bias you want using RL, regardless of training data.