Error rates revealed
 

A user has broken down the GIMPS error rate for exponents
between 2 million and 7 million. See ftp://mersenne.org/gimps/err_rate.xls
I'd ignore the data from 6 million to 7 million as there are many triple-checks to be done. The data from 5 million to 6 million will not change much at all. The
data from 2 million to 5 million is accurate.

There is a lot of interesting data in this spreadsheet. Our overall error rate is roughly 3.5%. If you have an error-free run, the error rate is in the 1.4% to 2% area. If you have an SUM(INPUTS) != SUM(OUTPUTS) error or ROUNDOFF > 0.40 error that is not caused by an approaching FFT crossover, then there is a 56% chance that your LL test will be no good!

Our intrepid researcher broke down the non-clean run stats below. So if you
get a single error, you've got a 2/3 chance of being OK. Two or more errors
and your chances are not good. This is not broken down by error type
(SUM(INPUTS) != SUM(OUTPUTS) vs. ROUNDOFF > 0.4)

Number of errors reported bad good error_rate
1 314 625 33.4%
2 192 135 58.7%
3 107 63 62.9%
4 74 36 67.3%
5 53 17 75.7%
6 48 20 70.6%
7 31 7 81.6%
8 30 10 75.0%
9 17 10 63.0%
10 25 2 92.6%
>10 375 58 86.6%