See below sample outputs from subsequent executions of "/usr/bin/spamassassin -tLD < spam_msg_file":
since you're using -D for debug output, is there anything in the debug that might give some clues?
Do both use the same score set?
Any complaints about lock failures?
Did either trigger autolearning?
Did either trigger a bayes sync (can cause a dramatic change in the bayes DB as the journal is integrated)?
