While this is more sophisticated, I doubt it's the "correct solution". The underlying assumption of independent draws from a Bernoulli distribution is vitiated, as the order in which the siblings are read usually affects their proportion of upvotes to downvotes. It still significantly penalizes late comments, because the 95% confidence interval is much wider for them (and its inferior limit consequently lower). Last but not least, it brings down comments that are at all controversial; whereas an opinion with 1000 upvotes and 500 downvotes is probably more interesting to read than one that only gathered 50 upvotes within the same interval. In short, this sorting algorithm seems ideal for quickly placing on top bland comments, that nobody really disagrees with but will upvote anyway because they haven't yet seen the much more incisive discussion below.