As you say the headers are extremely valuable (but only at the moment) Because the protocol LETs you forge them eventually this will become the norm, then with a suitably crafted body even Bayes won't cut it anymore.

When we look at some of the tokens that our Bayes widgets work with and find significant we often go 'huh?' The fact that we don't really understand WHY some of these tokens exist does not matter one wit. They are statistically significant and thus at the end of the day Just Work.

The problem is that as Bayes gets more popular the spammers will employ people to analyse how the filters are working and it is fairly easy to find chinks in the armour to slip a knife through to put it in Medieval terms. At that stage I suspect we will be up for a new protocol.