You can and should eval your judges. They're also typically easier to eval because often you have them emit categorical/structured data.