Yep. The key here is that if you choose four arbitrary galaxies from a set, there's two possibilities -- their positions in space are close enough that there's interesting things to be learned about their formation; or they're not. In the latter case, there's a strong reason to expect no bias to the tetrahedral side lengths; in the former case, there may or may not be. If you then look at all possible combinations of four galaxies -- including all the four-tuples that are "real" in 3D space, but also all the four-tuples that aren't -- the average of "no bias" and "interesting bias" will be "smaller, but interesting bias."

Which is why, once a bias is found, figuring out it's p value is non-trivial...