They receive wildly different BAYES scores. * -1.9 BAYES_00 BODY: Bayes spam probability is 0 to 1% * [score: 0.0002] * 2.2 BAYES_20 BODY: Bayes spam probability is 5 to 20% * [score: 0.0881]
This looks like you have per-user Bayes databases, and the messaage type has been trained differently in each.
Also, it looks like there are per-user rules, since BAYES_50 has a normal score of 0.2, and there is no reason BAYES_20 (indicating much less spammy) should have a score of 2.2.