A Proper Scoring Rule for Confidence Intervals — LessWrong