Backward Reasoning Over Decision Trees — LessWrong