Hi, I’m building Barazo, a forum AppView on ATProto (lexicon thread).
Going back and forth on this same question from the forum side. My current implementation: reputation stays local to each community, computed AppView-side from typed reaction records (“helpful”, “agree”, “insightful”). Nothing stored on the PDS, because bad actors could fabricate their own scores. Different communities might weight the same reactions differently, because “helpful” in a Rust programming forum shouldn’t carry the same weight as in a meme community. Of course memes are way more reputation building, but not everyone may agree and that is fine 
The Discourse trust levels comparison earlier in the thread is close to what I’m doing. Additive, local, earned through participation in that community. No negative scores, no cross-community spillover.
I think the two approaches can coexist if the cross-network layer is limited to positive, opt-in signals published as labels. Something like “sustained quality participation across N communities” that AppViews can subscribe to or ignore. The consuming app stays in control of how much weight to give it, and a bad moderation decision in one community doesn’t cascade into others.