The reason that it took me so long to build a version of the CRC that would pass muster as far as generating high quality random numbers is this: The CRC tries to solve the rotors for a minimal "energy" state. Think interference pattern (like holography). The solution winds up up the stator.
If 128 bit AES (Rijndael) if fed over 15 Megabytes of noise (low quality random numbers) compressed with bzip2 or gzip, the chi squared distribution of the output ciphertext increases a large amount.
The opposite happened with Eris versions 2 and 3; the chi squared distribution decreased. Eris (the CRC) was trying to "solve" for the lowest energy state. Version 4 doesn't seem to have a problem with this. I got the Eris CRC 4 to build in the summer of 2013 and did some minimal testing. (I wrote my own test program to measure the chi squared value of the output data.)
AES is hiding the key in the cipher text. Proof is on the way. See my facebook page:Tim Patti
The Ken THompson Hack is also in play. Recently, when I recompiled Eris3, the Eris hash did not match while MD5 and SHA512 did.