Dodging systematic human errors in scalable oversight — LessWrong