Best explanation I've seen (and claims to be from a published but not public report on it), is that their 3 way consensus didn't smooth over repeated wildly wrong outputs correctly. Concestency problem strikes again.