Goodness estimate biaser — LessWrong