This can be done on chain, and I'm gradually building something more robust than other efforts on this so far.
Essentially, you need to codify the obvious things a human curator would identify if they dug deep enough. Things like.. people not commenting on other peoples posts, or replying, or reaching beyond their little circles.
At liketu, we've already got a robust way of looking at legitimate interaction circles (check out your home feed for yours)
But essentially a web of trust which takes into account the types of engagement that are valid, and analyses the network propagation effects of this are used to measure the likelihood that it is real or not.
You would expect, someone who is giving/receiving a lot of viable engagement would have a broader network influence, then you just need to attach weights to the different things that matter to you. For example, we have the Hive Engagement League, and they've done quite well with finding out the top engagers. To extend that, you just need to find the recipients / beneficiaries of the most engagement.
Good thing about that, and indeed the work i'm doing, is that it's all based on on-chain activity, and the stats will speak for themselves.
When people make extraordinary claims, they should have extra ordinary proof to back that up. Well, soon they can :)
But this should be done in the blockchain level, not frontend, as we go back to the original concern, that is exploiting the reward pool. How do keep the spammers (real shitpost) from taking advantage of the reward pool?
Spammers will get no reciprocal interaction with respect to the spam they push out. All of this will be pulled from on-chain. There's a lot of things you can learn from on-chain activity if you just frame it the right way, and apply the right kind of analysis.
A web of trust is actually more of a qualitative exercise, because it's difficult to get very far if the accounts which validate you, are also of low quality.