It’s needs a republic structure. Every person joins a small default group of random strangers, their moderation group. Small enough that bad actors can be fished out. This is part of a larger group where these individual silos get their own trust relationship. If a small group starts misbehaving, the individuals in that group get reassigned. If the individuals who have moved also correlate with a lack of trust in their next group, they get flagged and put on probation. We can’t have global trust until local trust has been established.

Problem (1) can likely be addressed by restricting the number of iterations to some small constant, like 3 to 6 (which they suggest in the article). You can also restrict the number of peers-of-peers you fetch during each iteration to a small random subset. So if we only iterate 6 times and choose 10 peers for each iterations, we’ll get 1 million (10^6) trust scores needing to be pulled to calculate a trust score. That is an upper bound and will likely be less because it assumes each peer is distinct. At 32-bit floating points, that’s 32 million bits or 4 MB necessary to be fetched.

I can imagine this being reduced by at least a factor of 10 without much impact on the trust score. But note the “random subset” means different people will have different trust scores for the same peer :shrug:

For (2) yeah we probably need to lower our expectations on privacy for the time being; it’s a masters thesis and privacy in open distributed systems is very tricky.