Zero-based karma coming through

You seem to be thinking of things from the point of view of 'I don't want to suddenly find I can't downvote', where I'm thinking of it from the point of view of 'what is likely to lead to the most informative karma scores'. I don't think you're wrong given where you seem to be coming from but I'm not looking at it from the same perspective.

This is pretty much right, it uses a big key value store. The issue with karma is that the Account record maintains the karma value but the vote is stored as a separate relationship between the Account and the Link (article). Generally is is not safe to update the database directly as it is necessary to keep the separate memcache coherent. Thus any changes must be done via Python code. Whilst such a change is possible and relatively straightforward I'm not sure its worth it. From what I've seen there are no truly obnoxious users on less wrong that need to be dealt with by such a change.

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.

I've wondered if it wouldn't be better to give relatively more weight to votes from people with high karma. I haven't proposed it because it seems counter-productive to escaping the dangers of an echo-chamber, and because it seems like a conflict of interest coming from someone with high karma.