Make reputation scores the result of some sort of eigendecomposition-based algorithm (similar to PageRank) where the initial state vector is based on the people you trust. That way, you’ll only give credence to a bot to whatever degree you (indirectly) trust the bot.
Implement Sybil resistance with a combination of verifiable cost and blacklisting. Make it cost real verifiable resources to create an identity (hashcash or Bitcoin burning or something) and then have a system for blacklisting fake IDs.
This is just 'web of trust', right? https://en.wikipedia.org/wiki/Web_of_trust
Or at least, that's how you make it not matter. If spam-bot hands out reputation to spim-bot and then you get a bunch of garbage from spim-bot, derate everything coming from anyone that has given either of them reputation (and perhaps ignore their attestations also).
But fundamentally, it isn't particularly harmful to just burn a user that starts spamming. If a compromised account is recovered, it's fine if they have to restore communication out of band, if it isn't recovered it's fine that it is burned.
Could also just use regular money.