I'm not proposing a solution. I'm thinking about the problem for five minutes.

edit: Well, it didn't even take five minutes!

We need a reliable predictor of troll-nature. I mean, I'm not even sure that P( troll comment | at -3 ) is above, say, 0.25 - much less anywhere high enough to be comfortable with a -5 penalty.

Of course, I'd be comfortable with asserting that P( noise comment | at -3 ) is pretty high, like 0.6 or something. Still not high enough to justify a penalty, in my opinion, but high enough that I can see how another's opinion might be that it justifies a penalty. If that is the case, well, the discussion is being severely negatively impacted by conflating noise and trolling.

I might go and figure out how to get some data off of LessWrong commenting system, to try and determine a good indicator for troll-nature. (I don't plan to try and figure out noise-nature. That's the problem that the Internet has faced for the last 15 years, I'm not that hubristic.) That in turn would would put some numbers into this discussion. I don't know that arguing over how many genuine comments can be inadvertently caught in a filter is any better than arguing over whether there should be a filter at all, but to my mind it's more constructive.