How to evaluate (50%) predictions — LessWrong