I've been training SpamAssassin, but my users are still receiving a lot of spam because the Bayes score is too low. However, when I run the message through SpamAssassin manually, the Bayes probability is very high. Why am I seeing this discrepancy?
If you are running the SpamAssassin Learn command (sa-learn) as root only, it will train Bayes for only root's mail. You need to run sa-learn commands as other users in order for it to work properly.