It would be interesting to see your statistics on error rates, Louie. Wblipp pointed out that this information could be incorporated into a resources allocation model. My guess is that the probability of an error should be proportional to the duration of the test, so we should see error rates per test go up with time, as is also the case with GIMPS. At least with some data, we would have a basis for predicting the probability of finding a prime doing double checks as opposed to checking new numbers.