The implementation samples rosc.randombits at a frequency lower than the
oscillator frequency. This gives better random values. In addition, for
an 8-bit value 8 samples are taken and fed through a 8-bit CRC,
distributing the sampling over the byte. The resulting sampling rate is
about 120k/sec.
The RNG does not include testing of error conditions, like the ROSC being
in sync with the sampling or completely failing. Making the interim value
static causes it to perform a little bit better in short sync or drop-out
situations.
The output of uos.urandom() performs well with the NIST800-22 test suite.
In my trial it passed all tests of the sts 2.1.2 test suite. I also ran a
test of the random data with the Common Criteria test suite AIS 31, and it
passed all tests too.