[ffredd]

You can always find something that seems non-random in random data. When you do, you can formulate a hypothesis based on what you've found. You can't however use the same data to check if your hypothesis is correct! That would just be ridicilous.

Suppose e.g. that I examine my hand histories and find that I got dealt 1% more pocket pairs than the statistical expectation. Now I have a hypothesis: "This site deals 1% too many pocket pairs". I can test that hypothesis by checking another large database of hand histories, but I can't do it by examining the same hand histories.

What these guys should have done is to use the data collected from a subset of the random number generators to formulate a hypothesis, and the data collected from the remaining random number generators to test the hypothesis. Their failure to do so invalidates their "evidence" completely.

It's also interesting that the largest contribution to the "unlikeliness" of the data came from the time before planes started crashing into buildnings. If they had chosen their start and end times a little bit differently, their result would have been very different.