This is great ! In fact, this got me thinking about what a simple ML model for automating HN's 2nd-look could look like. One could assume that the log odds of user_a' upvoting an article, after visiting it off the 'new' page, is a linear function of

* the article's quality (denoted by x) and

* user_a's click happiness.

P(upvote_a | seen_a) = exp(a + λₐ x) / (1 + exp(a + λₐ x))

Both a and λₐ are unknown parameters that can be estimated from past data given some prior over x, say centered at 0 with a high variance.

One can grant a new articles a 2nd look if the posterior probability of x crossing a gate keeping threshold is sufficiently high.