Reinforcement learning in language models recruits a functional welfare axis (functionalwelfare.com)