That's an interesting approach; it would cause comments that are really good to float to the top no matter when they are posted. In terms of Bayes' theorem:

    P(upvote | good post) = P(good post | upvote) * P(upvote) / P(good post)  
  

Let's neglect that opinions vary, and assume P(good post | upvote) = 1. P(upvote) should be proportional to P(post is read); but the readership at the bottom of the page (or wherever latecomers are dropped) is quite small.

But this begs the question: what does it mean to "float to the top" if highly-rated comments are being placed in a random location?