header inventory

Perhaps it makes sense for the various filter writers and sysadmins
on this list to offer an inventory of the various modifications
each of their systems add to a message, so we can see where people
agree and where they differ, and get to know patterns of use.



I'll begin with dbacl. It's a Bayesian procmail filter. Learning
is performed on mbox archives which may or may not have been filtered
previously.

The recommended rfc headers to be added by procmail are all optional
and in the form:

X-DBACL-Says: cat2
X-DBACL-Scores: cat1 12.7 * 453 cat2 11.3 * 453 cat3 14.9 * 453

The reason for adding headers is purely for the benefit of the MUA, as
messages are delivered to putative mailboxes by procmail already.

dbacl does not store any other information destined for itself in a message.
Thus, there is no intended direct feedback loop between filtering and
(re)learning stages. 

Because learning takes place on potentially filtered messages, it is
important for dbacl to recognize its own headers and skip them. That
is why the program contains logic to skip all "X-DBACL" prefixed
headers. 

Since procmail does the actual adding of headers, a user is of course
free to use any headers they like, but using different headers will
introduce a subtle bias.

There is no structure to the category labels cat1,cat2,cat3 etc. 
which reflect the user's filtering needs. The scores simply represent
average bits of information per token, under various category assumptions.

-- 
Laird Breyer.