The right model would perform some kind of flow analysis, starting from the vote graph. I guess there must be standard techniques in web search (linking as voting) and the like, that oppose spam linking (page rank? I haven't studied what it is).
You are looking for "attack resistant trust metrics". Google's PageRank is one such; Raph Levien's Advogato algorithm is another; my TrustFlow is a third.
Our friends at Tricycle will push through a zero-based karma system (no self-voting possible) sometime this evening. At present this will only cover future posts/comments - they may go back and revise history some time in the indefinite future (or not), but apparently that would be overly complicated for now. We'll see how this works.