How to Automate junkmail / notjunkmail learning?

Although it doesn't appear in Lion Server documentation, I'm assuming that Lion Server still uses the "sa-learn" tool to train the junk mail filter (page 36 of Snow Leopard Mail Services Administration manual).

If so, I wonder if anyone has found a way to automate the collection of spam / ham messages from user accounts. The suggested approach (user redirects mail to either junkmail@xxx.com or notjunkmail@xxx.com) is reliant on a fairly complex action by the user, and when you have dozens of spam messages is therefore unlikely to be used.

A simpler system would be to designate "ham" and "spam" folders within the user's IMAP folder tree (e.g. as sub-folders below Junk) and simply get them to copy / move messages to the appropriate folder where learning is required (so if you find a spam message in your inbox, you simply move it to the 'ham' folder and it gets picked up by SpamAssassin and is used to train the filter).

This is not a new idea (our last email server implemented a similar system).

I was wondering if anyone knows whether it is possible in Lion Server's mail implementation, and if so how...

I've had three ideas, none of which I can see how to make work:

• Create a local filter that redirects messages found in a folder to the appropriate sa-learn mailbox (problem - no way to get filters to run daily)
• Edit the script that captures the content of the junkmail / notjunkmail accounts and extend it to cover all the relevant mail folders in the stack (not sure where script is, or how to add suitable search term to find the right folders across generic stack) - sa-learn will apparently take a filename with list of folders to use as an input, so perhaps there is a way to use 'find' or some such to locate the right files?
• Create a shell script that finds the messages within the mail-store and moves them to the appropriate folders (i.e. outside Dovecot) (not sure on how to find the right folders, not sure how Dovecot will respond to messages moving outside its control) and link to cron job to get it to run before the sa-learn training happens (if it happens)

Anyone got any ideas? Or is there a better way to do this?