I think I should look into your rSPAMD implementation, about FCrDNS - I am still skeptical on my approach of this whole thing, though right now my bigger challenge is Bayes giving wrong points - I opened a thread here to discuss it.
rspamd is most based on fuzzy technique, which is much more mystery than AWL and Bayes together. ;-) That one didn't work for me at all. FCrDNS is also not working for me and is the same thing like @ittk posted above. You can be the preacher, which try to shame all the ones, which don't configure their SPF or DKIM well (or FCrDNS) or you may be required to get each potential clean mail as your business is based on, so you may be careful with such settings. On Bayes I recently only had one bad experience as I trained Bayes with foreign spam (there are places where you can download spam to load your filters). Also I would be careful with bait mailboxes just for training Bayes, as it would result in dilute your scores because you just learned "stupid spam". Best would be learning spam, which frustrates the users, e.g. spam, which they forward to you or place in an extra spam mailbox as they are such annoyed by that mails, that they do the job to report them to you. This are the best ones to learn as spam.