[pavas]

The reason I mentioned that the trust metric should be transitive and distributed is so that it prevents gaming as much as possible. You wouldn't want to have a trusted central authority (for everyone) because that could always be corrupted or gamed if it's profitable enough. Rather every individual would have a set of trusted peers with different "trust" weights for each based on the individual's perception of their trustworthiness, that could be changed over time.

This trust (weighting) should be able to propagate as a (semi-)transitive property throughout the network to take advantage of your trusted peers' trusted peers. This trust weight propagation would need to converge, and when you are served content that has been labeled incorrectly ("high-value" or "trustworthy" or whatever metric, when you don't see it that way), then your trust weights (and perhaps your peers') would need to re-update in some sort of backpropagation.

The hard part is keeping track of the trust-network in a way that is O(n^c) and having the transitive calculations also be O(n^c) at most. I'm quite sure there are ways of doing this (at least with reasonably good results) but I haven't been able to think through them.