First, tell me if there's anything wrong with this summary: 1. A message arrives and is passed to spamassassin and/or spamc+spamd. 2. The score for that message is computed. 3. The AWL score for that sender is updated. 4. The message was mis-classified, so after delivery the user feeds the message to sa-learn. 5. The Bayes score for (the tokens in) that message is updated, *but the AWL score for the sender remains unchanged.* 6. A similar message from the same sender arrives. The net score is moved away from the Bayes-influenced value by the (obsolete, or at least incorrectly recorded) AWL value.
Assuming I've got that right, tell me whether there's aaanything wrong with this conclusion: The AWL will wrongly influence the score for both spam and non-spam as long as the AWL remains unaffected at step 5, in any case where the initial classification was incorrect. Finally the question: Shouldn't sa-learn "retrain" the AWL as well? At the least, throw out the entry for that sender and begin recomputing it with the next message?