I've been trying to use a chi^2 test for Zipf generators for the
parameter sets of tests in zipf-test.scm.

The limitation is that I'm running tests only for NVOCAB, the number of
bins, > 90 so I can use the normal approximation for the chi^2
probability distribution.

Of those tests with NVOCAB > 90 only the following tests fail:

(#f 130 4.5 0 30000 106.2896889047007 48.18713521262703)
(#f 130 6.66 0 30000 126.1122250147938 48.18713521262703)
(#f 131 16.1 41.483 10000 92.25243717046817 48.373546489791295)
(#f 131 46.1 41.483 10000 108.9481906236824 48.373546489791295)
(#f 131 96.1 41.483 10000 129.19043641049112 48.373546489791295)

The first entry of each list is whether the test passes or fails (so
these are all failures), then

number of bins,
s,
q,
number of samples
(abs (- chi^2 mean))
(* 3 (sqrt variance))

chi^2 is the computed chi^2 statistic, and mean and variance are the
mean and variance of the chi^2 distribution.

The tests pass when the second-last number is less than the last number.
  So those tests that fail above are not even close.

Some other tests were close to failure:

(#t 43701 1.000000001 0 60000 876.0176218421475 886.904729945669)
(#t 43701 1.000000000001 0 60000 864.1726089265794 886.904729945669)

I guess these have barely more samples than bins.

My question: Is there something about these parameters that are likely
to lead to such failures?

Brad