I did some troubleshooting and found that my bayesian probability data was part of the problem. It's been building for years. I started quarantining messages in users' SPAM/Junk folders (accessible only by webmail or Imap if you add the folder) so I could get a better idea of what's going on.
I couldn't tell at first, because the spamassassin data in the email headers was truncated since the last update. (Aren't updates grand?)
It was just inserting results like this:
Code: Select all
Content analysis details: (0.8 points, 3.0 required)
pts rule name descripti
=?ISO-8859-1?Q?on=0A_----_--
X-Spam-Flag: NO
It turned out to be encoding issues and I had to use sed on some Exim ACL files to fix it. Once I could see the full output again, I saw that even good mails were getting bayes scores of 30 to 40% probability of SPAM (which just gives it a small score) and junk mail (not quite spam) was getting 90 to 100% probability, bumping its spam score by 5 right off the bat.
That would have still rejected that pharmacy email with a score of 6, but I was also still getting a few notification emails from Linuxquestions from long running Slackware threads I posted to years ago (latest changelog, latest kernel) and they were getting rejected. My parents were getting junk (to me its junk, but its stuff they signed up for) rejected in their spam folders too, like notifications from services they've subscribed to and notifications of points card specials etc. because of the Bayes data. They probably will get through now.
After deleting the Bayes databases, now those Linuxquestions emails are getting through. They are unwanted, but were useful in troubleshooting.