I did exactly that actually by parsing log files 4 nites in a row about a
month ago.
It was initially to disprove the new patch=more missed notes theory which it
did.
Interestingly, the new level=more missed notes theory seemed to be proven
true by another subsequent test.

In that test, over the course of 3000 some odd start ups there were 137
missed notes. Of those 137, only 35 were single missed notes. There were
further 34 pairs, 7 triples, 2 quadruples, and 1 quintuple for a total of 79
different occurrences of missed notes.

These results are so terrible statistically speaking I sent them to GZ to
look into as one of the best examples of the RNG being very skewed.
With a <5% failure rate, only 5% should be multiple occurrences +/- at most
5%, for exactly the reason that one random number should not affect another.
That was clearly not the case. Over the sample single missed notes occurred
only 44.0% of the time. Pairs of missed notes occurred 43.0% of the time,
triples 8.7%, quadruples 2.5%, and quintuples 1.3%. This is radically
different from what should be expected given normal distribution of a simple
random. The sample size was enormous, more than enough for good statistical
accuracy by at least an order of magnitude.

Now, if there is an unknown factor in these skill checks such as an
automatic failure range, or the randoms generated are open-ended in same
fashion, then perhaps it is functioning acceptable but the external factors
to the random number are very influential.

Kit

heh, i'd be willing to bet they are using the built in c code random number
generator.
the problem is you are just using your perceptions of what happens. and
what sticks
out is going to be the anomaly. if you sat down for an hour and kept
clicking the song
icon and tracked how many fails and when they occurred in relation to the
others, i would
think that you would find that it is random.