The “no sandbagging on checkable tasks” hypothesis — LessWrong