How we choose the news

We look first for what matters, then for its direction. We never search for positive news: we take the most important and surprising events of the last 24 hours and only then ask whether they point somewhere better.

In science and medicine a story needs a result: something has already happened and changed something, with someone who benefits, a change of a known size and evidence for it, or a study that shows how something works, with its limits in the text. A result in mice is told as exactly that, or not at all.

In world affairs and the economy, in energy and the environment, a story can come before its effects: an agreement, a decision, an investment, a turning point, when it is real, verified and consequential, and a reasonable, informed reader would see more good than harm in where it leads. It is told as a possibility, at the stage it has really reached. In culture, what counts is the value a work adds, whatever its subject.

Each story says which kind it is: a result, an opening, research. Some things stay out however interesting they are: the bright side of an ongoing tragedy, an announcement passed off as a result, routine corporate news and PR, propaganda and military victories, a win for one political side, sport, gossip, old stories presented as new. The few exceptions are written in the compass too.

Every night we collect a few thousand headlines from about 130 feeds, from news searches, and from sites we read directly because they offer no feed. A first pass sets aside what is clearly out, and the rest is put in order of promise.

Then an investigator agent takes each story and looks for the reasons the good reading might not hold up. It finds the primary source (the paper, the official record, the documents companies file, the dataset), the other newsrooms that reported the event themselves, what critics say, who pays, what stage it has really reached, and whether it was later retracted, corrected or cancelled. It may cite only pages it has actually read. When a bigger newsroom reported the same event, we link to it; the number next to a story counts the other newsrooms our agents read.

A judge agent decides from that file and the compass alone. It decides twice; if the two decisions differ it decides a third time, and without a majority the story stays out. When a case really tests the rules, it goes to the editors instead of being guessed. In doubt, out: publishing one wrong story costs more than missing ten good ones.

We keep going until the day has enough stories that passed; we never lower the bar to get there, and a section with nothing good enough stays short. Several stories about the same news become one, and the front page goes to the stories that matter most, with the most lasting and most certain change.

A story needs at least one primary or independent source we trust. Five copies of the same press release are one source, and a company's word proves what it announced, never that it works.

Every website we come across gets a standing: trusted, allowed or blocked. An agent proposes it from what it can verify (named editors, real bylines, a corrections policy, who owns the site) and from public assessments such as Wikipedia's list of sources; the editors decide. Sites that publish machine-written or fabricated news are blocked, and nothing they say can support a story. Every standing is looked at again over time.

We read websites the way a reader does: one page at a time, saying who we are, never through services that remove paywalls. We don't read sites that ask not to be read automatically.

The compass belongs to the editors. When the agents meet a case the rules do not settle, the editors answer it: each one first, without seeing the others' answers, then together. A change to the compass needs three quarters of them, and whoever disagrees has the disagreement recorded next to the rule.

Every decision records the version of the compass it was made under, so any verdict can be run again and explained.

Every night the agents go back to some of the stories we have published and check whether a result was retracted, a figure corrected or a project cancelled. When that happens, the story is corrected or withdrawn, and it says so.

If you find an error, a headline that goes beyond its source or a number that is wrong, write to us: those are the messages we need most.