Maybe the issue is just the absoluteness of the scores. Behind the scenes, unbounded numbers are fine, but when presented to users they cause a tendency for gaming. I propose a follow-up experiment:
1) Keep the absolute scores secret
2) Continue graying out down-voted comments
3) Do not show any numeric indicator for conversations with less than N comments (I'd experiment with N near 10)
4) When the number of comments is greater than the threshold, show which percentile the comment is in for the set of positive comment scores in that thread.
This way, when you look at a comment you don't think "Does this guy deserve more points?", instead you think "Is this comment better than 75% of the other comments here?"
I think this would improve voting AND increase the value of that voting: people would have the right frame of mind when voting and the UI will present an easier value to determine how worthwhile it is to keep reading.