Engineers have long lived by a simple, seemingly obvious rule when designing new computers: The machines have to deliver correct answers. If asked to compute two plus two, a computer should answer four. But what if computers didn't always have to answer correctly?