A problem is that reputation-based sites require accounts, which discourages insightful blog commenting by those not ready to commit to another log-in. Bayesian

A problem is that reputation-based sites require accounts, which discourages insightful blog commenting by those not ready to commit to another log-in. Bayesian algorithms can defeat most spam while still allowing you to have an account-not-required comment system, but it does nothing to stop trolls. Although maybe you could train algorithms to detect derogatory nouns and adjectives. And you could still allow reps on individual comments and hide low comments by default, which effectively crowdsources the troll moderation. I believe I've seen sites using Disqus in a way close to this.