I have some perl code that extracts the original message from the forward, but some of SA's changes can't be undone, so I end up with headers like X-SA-Exim-Scanned and X-SA-Exim-Version in the bayes DB. I'm still looking to see what others may be bogus.
you can just use the bayes_ignore_header directive for these.. no need to use code to strip them out.
