Every twelve hours this site publishes five things. Not the five most important things — the five most interesting ones. Here is exactly how they get picked, because a list you cannot interrogate is just an opinion with good typography.
We read publisher RSS feeds, science and space agency feeds, and public data feeds for natural events. Roughly 40 stories came through in the window this page was built from. We only ever read what a publisher syndicates for that purpose. We do not scrape article pages, we do not go around paywalls, and we never republish anyone's article text.
Every candidate is scored against a published rubric, weighted as:
Some things score zero automatically: incremental political process stories, celebrity gossip, outrage bait, routine market movements, single-source unverified claims, and advertising dressed as news.
Scoring is a first pass, and it is deliberately not the final word. The highest-scoring 40 candidates go to a second, more capable model that makes the actual editorial call and can promote something the scorer underrated. It works to a few fixed rules: at least one science or discovery item, at least one genuinely strange one, never more than one politics or conflict item, and no two items covering the same underlying story.
Things that happen off Earth — launches, astronomy, planetary science — compete on equal terms with things that happen on it. They are not a quota to fill. That is also why each item says where it happened: Meanwhile, in Kazakhstan…, Meanwhile, in Orion's shoulder…
Each item carries the reasons it was chosen. Those are published exactly as the model wrote them, not edited afterwards to look better.
The scoring, the selection, and the blurbs are produced by AI models running on a schedule, from the facts in the source feeds. The blurbs are original writing, not extracts. No AI system is infallible — if something here is wrong, the source links are right there, and they are the authority, not us.
Every edition is also published as JSON — the current one, and one per window in the archive — including the scores and the selection reasoning.