REGRESSION? BUGFIX?

[David Relson]

At 07:31 AM 3/4/03, Matthias Andree wrote:

>m-a at users.sourceforge.net writes:
>
> > Update of /cvsroot/bogofilter/bogofilter/src
> > In directory sc8-pr-cvs1:/tmp/cvs-serv1769/src
> >
> > Modified Files:
> >       bogofilter.h bogoutil.c format.c graham.c graham.h robinson.c
> >       robinson.h wordlists.c
> > Log Message:
> > Make floating-point constants type "double" by omitting trailing f.

Matthias,

I've looked at all the spamicity differences in the two regression tests 
that failed - t.systest and t.grftest.  The results are all reasonable.

Remember how the graham algorithm has its array of 15 extrema and how very 
high values (0.99) and very low values (0.01) are equally extreme?  I think 
we have more evidence that the extrema arrray just wasn't a good idea.

Changing the constants from floats to doubles apparently changed the 
results enough so that some of our test messages now have 0.99's in their 
score where they used to have 0.01's.  This changed their spamicity scores.

The biggest difference is in the message that flipped from ham to spam 
because it now has a pair of 0.99's when before it had a pair of 0.01's.

I've updated the reference results in CVS and all tests now pass.

David