On 9/16/05, jdow <[EMAIL PROTECTED]> wrote: > Yes indeedy. And I've been looking at Bayes scores here just a wee bit. > BAYES_99 just does not hit on ham and hits on high percentages of spam. > Even BAYES_95 does not hit ham. I go down to BAYES_80 before I hit 0.05 > percent of ham.
During a two-week period recently I captured a copy of all mail that (1) did not reach an SA score of 5+ points and (2) did not have my personal email address in the To:/Cc: headers. I then examined the set of SA rules that were triggered by those messages (as recorded in the X-Spam-Status). 100% of such mail that hit BAYES_80 or more was in fact spam; about 90% of BAYES_70 was spam. However, there were a few BAYES_80 during the same period that *did* have my address in the headers and that were *not* spam (and, also correctly, not tagged by SA), so it wasn't just a matter of cranking up the score for BAYES_80. Instead I added procmail recipies to treat as spam the combination of "not to me" plus BAYES_[89][059] regardless of the SA point score. That was a month ago, and I haven't had a false positive yet. If there are any developers listening ... has anyone given any consideration to Bugzilla #3785?
