Your title, domain and posting time go to Aito, a database that answers prediction queries straight from its index. There is no trained model. Three queries run over 351,652 HN submissions from Sep 2025 to Aug 2026: one for the score bucket, one for the front-page odds, and one for the closest past titles. “Front page” here means 50 points or more.

We held out 6,000 submissions posted after the corpus ends (2026-09-22 to 2026-10-04), so the index has never seen them, and scored them through this page’s own code. AUC is 0.629 (95% range 0.600 to 0.659), where 0.5 is a coin flip and 1.0 is perfect. The top tenth of predictions reached the front page 15.7% of the time, against 6.9% for all of them (2.3x). Real signal, and a long way from knowing.

The percentages mean what they say, checked on posts the calibration never saw: predictions around 3.9% came true 4.8% of the time, predictions around 6.7% came true 7.2% of the time, predictions around 14.0% came true 10.8% of the time and predictions around 24.4% came true 31.9% of the time. Above 37.9% we have too few cases (4) to promise anything.

Title alone scores 0.626 and domain alone 0.637. Together they score 0.640, so they mostly carry the same information and combining them buys almost nothing. Posting time carries nothing: hour alone scores 0.504, and taking posting time out of the query moves the score by −0.006 (95% range −0.021 to +0.011). Title length alone scores 0.550. Most of what decides a post (who submits it, the first few votes, what else is on the front page that hour) is not in the data at all.

Freshness matters, up to a point. The same 6,000 held-out posts scored against older copies of the corpus: one ending Jan 2024 gets 0.597, against 0.629 for the current one, which ends Aug 2026.

Reproduce it: uv run python -m scripts.evaluate in the repo. The run also checks that each bucket shows its own explanation (40 predictions, 0 failures).