I think that you may be seeing something there. But we’re looking at the wrong end of the distribution.
Can we get statistics on all 1M runs instead of just the 30k with the fewest number of sequences? I believe that your clipping will produce fewer sequences. We haven’t looked at the total number of sequences yet. It seems that most (nearly all) will have more than six sequences. Can we run them with the same seed?
Merssene Twister may pass this test. But I still have some doubt.