A Protocol for Optimizing Affection — LessWrong