in my experience, if there is a Bayesian filter score and you combine it with other scores, the composite score will be less reliable. i had SpamAssassin on an email server that i ran for years and i wondered why spam kept leaking through and when i disabled everything but the Bayesian filter, it started working better. on a later email server i just dropped SA in favour of Bogofilter - the original Bayesian spam filter which does nothing else - and it did the same job. a Naive Bayesian Classifier library should be able to do the same job. and NIP-05 verification is just a token you add as an input to the classifier. it'll learn to recognise that token as a good sign if you train it enough.
Discussion
basically any kind of indicator can just be a token in the input of the classifier and then you don't have to worry about interpreting the token because the classifier will figure out how much it should count
Mixing verified on the front is a good insight I wouldn't have thought of. I would have assumed you would do it as some sort of flat score modifier on the output but I think your idea is much better.
By token, just some sort of special keyword pre/ap-pended? can that be gamed?