Working on Farcaster and Warpcast. Longer thoughts at https://dwr.email

I definitely feel like your methods aren't perfect.

A few other thoughts

1. We use a machine learning algorithm to classify spammy users.

2. That requires a manual input when new spam "metas" emerge.

3. Best way to model spam (and fraud or any other "gaming" of a system) is like an organism that evolves based on the environment (the social network) and the various inputs into the environment, i.e. antibiotics / spam filters

4. So what works effectively today will not be effective tomorrow (antibiotic resistance!). 

5. So you have to flag the new metas to the model using signals—user reports, manual classification—and then the model can relabel users.

Kind of sounds like the machine learning algo which eventually has stored every single available spam meta could be a valuable product by itself.