How can I make sure that the Bayes filter is getting the best information during its learning mode?
Make sure XWall finds a lot of Spam mail, either by SLS or by BCC. Another good idea is to add the email address of employees that have left the company to MBAdmin Options->Exclude->E-Mail Address->Inbound RCPT TO. Usually these addresses get only spam and XWall can use this messages to build up the bad-word list.
