I have an alpha draft of a tool that would allow clients to train their own custom filters. It works pretty well, but its a real bitch trying to get enough data. Data being the text of spammy notes. How much extra storage space can spammy text content really take up on a relay? If you're hosting images, sure, nuke those, but keeping suspect notes up for a couple of weeks would be very helpful.
i think you could probably easily get relay operators to feed their deleted events into your midden if you just asked... it's a matter of just adding a tiny feature "when delete, send to dumbass who wants deleted events" in fact, i am just about to build out a two level caching algorithm that lets me maintain some reasonable limits on the cache inbuilt database and maintain searchability (via simple filter searches) but push the event itself to a secondary store that's practically half of what you are looking for but i think you are barking up the wrong tree looking for preemptive methods of blocking spam web of trust will do most of that for you, spammers can't win long term confidence in people and they have to constantly make new identities, which excludes them from getting deep into the web
Multilevel caching sounds like a great idea. I am a big believer in WoT, but I also believe preemptive filtering is a "must have" for many, many users. Not me and not you, but many. Spammers are already using LLM generated content. They can fail as many times as they like, and one human user only has to fail once for a spammer to get into the WoT for a while.
the thing is that fakes can't get deep into the social graph without being like the people in it ultimately if you are a fake you gonna get tricked there's many helpful things we can build into the systems to add friction for malicious actors but ultimately social manipulation in general is something that requires *requires* personal responsibility, skepticism, alertness, and emotional maturity to defeat
100%, but partial successes still annoy nostriches and waste their time. We need to do all the above.
Winners study the British SDS undercover training manuals. Losers study "ChatGPT For Dummies".