So far causally upstream of the human evaluator's opinion? Eg an AI counselor optimizing for getting to know you