It's somehow similar to a random generator where you have 5 dices, roll them and then add to the entropy pool only if the total was even or odd. Changing the power is like forcing the system to use only 4 dices. It changes the probabilities a little, but not in a very controlable way, and with a good mixing in the entropy pool it's almost irrelevant.
Note if you look at the paper, you notice a close but not entirely perfect normal distribution, but nothing you cannot fix with UDNs and Irwin-Hall. For reference how that is done you can read the bottom of this very useful RNG article: https://people.ece.cornell.edu/land/courses/ece4760/RP2040/C...
My overall verdict on the tech in OP is that it is amazingly promising!