Interesting. So reputation would act as a mediator for decision making then? Those with greater reputation would have more influence, and theoretically this should work because reputation would be gained only when other people notice that the person has done something beneficial/helpful in their eyes. Overall that should result in high reputation values being directly correlated to those who are altruistic and good.

It's definitely a cool concept. The main problem being I think is incentive to adopt. Without the weight of reputation having meaning behind it, no one will care and thus the intended effect disappears.

How would such a reputation system defeat the Sybil attack?

There isn't an absolute reputation score. Reputation is computed pairwise. Rep(a,b) means "What is the implied reputation of person b, given the persons a trusts"

So a bunch of fake accounts aren't going to influence a's perception of b,unless a has trusted someone who'se created those fake accounts. IN that case, the mistake a made was trusting someone (directly or indirectly) who'se created a bunch of fake accounts.

Because the input weight of a person decreases with their distance on the graph from a, the only sybil attacks that would screw you up would be ones conducted by people who you've said you trust.

Filtering out everything but the users you have established relationships with will just push you into a filter bubble echo chamber.

There are mechansims (see 'soundness') which reward people on the edge between filter bubbles for combining those bubbles. Maintaining a soundness core means resolving disputes between bubbles.

My goal is to gradually expand filter bubbles, rather than try to pop them.

After skimming the "respect" idea, I am convinced that echo chambers are the dominant strategy.

In fact, make a complete subgraph of order 9 or more of people who respect each other. That subgraph has order-N respect that goes to infinity for anyone who has any respect for any of its members, even transitively.