When I started using SpamAssassin about one year ago, I didn’t like the fact that even with high Bayes scores messages were not deleted, but marked as spam instead. At the same time Mozilla Thinderbird virtually never made a mistake deciding whether message is a junk mail or not.

On the other hand, messages with low Bayes scores and one or two matches by SA formal rules were getting low overall score too.

I’ve started analyzing messages that Thunderbird marked as junk mail. Soon I came to conclusion which seems quite obvious: if we have both Bayes match and formal rule match, than the overall probability that message is a spam becomes much higher. So, to solve the problem, we have to implement some meta rules that would execute when number of conditions match at the same time.