Behavioral Sufficient Statistics for Goal-Directedness — LessWrong