Could split the stories into buckets and then randomly sample from each bucket. Most stories are small, so they’re currently overrepresented in the sampling.
A "replacement-level" front page story is ~200 votes, +/- 50. If you're drawing just from the front page, and not from a deliberately weighted set of successful vs. marginal posts, most stories should be below 400.
(It's weird to say this but I'm not nerdy enough to have actually worked this out with data; it's just intuition from spending time here.)
Identifying some of the truly high-scoring articles (>1000) should be rewarding, and 200 vs 400 points is a pretty big difference on hn