Buy-in Before Randomization — LessWrong