- Compute the noise_floor from sqrt(s.w) or sqrt(w/s)
- Drop all values under the noise_floor
- Does improve performance, at a great loss of words
- Standard deviation over the database is much lower values
- Not sure why t.wordhist leads to two different noise floors

shell$ export srcdir=.../bogofilter/src/tests
shell$ sh -c $srcdir/t.wordhist 2>&1 | grep 'stddev is'
...ends in many line pairs that all say...
Noise floor is 5925.743164 and 8895.982422 during get (stddev is 175.114502 and 319.619812)
Noise floor is 2.763873 and 1.841056 during iteration (stddev is 175.114502 and 319.619812)