How we curate.

What we never use

• No engagement signals. Nothing is ranked higher because people clicked it, lingered on it, shared it or argued about it. Bulletin does not measure those things per person.
• No profile of you. The ranking sees your board's settings: subjects, keywords, places, followed publishers, formats, exclusions. It does not see who you are or what you read yesterday, beyond a short memory of what it has already shown you so it can rotate.
• No paid placement. Adverts are separate, labelled cards. No publisher can pay to rank higher.
• No secret weights. The numbers below are the live defaults. When we change one, this page changes.

1. The candidates

A board starts with the pieces published recently that match its subjects. Our taxonomy has 28 subjects and about 1,900 finer ones; every piece is tagged at ingest, and the tags are re-checked by a classifier so a fashion piece does not land on a technology desk. A board with a narrow subject that has little tagged supply is topped up with pieces from its followed publishers, pieces whose headlines carry its keywords, and the nearest recent pieces by meaning. It is never topped up with the general front page: a niche board stays niche.

2. Relevance

Each candidate is scored on how many of the board's subjects it matches (counted up to three), whether its headline carries one of the board's keywords (a lift of 1.5 per match, up to two), and whether it comes from a publisher the reader has starred (a lift of 0.3 per star, up to three stars). Pieces with an image get a lift of 1.8, because a board is read visually.

3. Recency, with slow publishers protected

Scores decay with age on a 72-hour half-life: a piece from three days ago carries half the weight of one from this morning. But the decay is measured against the publisher's own rhythm, so a weekly magazine or a monthly comic is not buried under the dailies. Evergreen formats (comics, zines, music releases, guides) do not fall off the board on a news timetable at all.

4. Plurality

A story covered by many newsrooms is lifted in proportion to how many are on it (0.3 times the logarithm of the count), and the briefing names that count. Then comes the cap: no single outlet may hold more than its share of a board, however much it publishes. When a publisher's pieces would exceed the cap, its lower-ranked ones give way to other outlets' best. No outlet is privileged unless the reader starred it.

5. Formats get their own lanes

Articles, podcasts, video, events, music, comics and social posts are not made to compete on one score, because whichever format wins the scoring race would flood the board. Each format the board allows is a lane with a guaranteed floor and a cap; within a lane the normal ranking decides. A music board keeps its releases and its gigs visible beside the reviews.

6. Place

Where a board follows a place, publishers and venues classified as local to it are lifted (1.25, or 2.2 on a local-first board) and sources of unknown locality are damped. Events are matched to the reader's home place so a gig around the corner beats a festival five hundred miles away, and listings leave the board once they have happened.

7. Meaning

A board has a profile: an embedding built from its subjects and keywords, refreshed whenever they change. Recent pieces close to that profile join the candidates (above a similarity floor of 0.65), and similarity multiplies the score of everything in the pool. This is how a board about a band finds the interview that never used the band's name in the headline.

8. Hard filters, set by the reader

Words the reader muted never appear in a headline on that board. Subjects the reader excluded never appear at all. A board's content filter can keep graphic material out of the briefing. Timed tweaks ("no Trump until tomorrow") lapse on their own. These run last and win over everything above.

9. Rotation and exposure

A piece that has led its section in the previous session drops below the best piece that has not led yet, so two consecutive editions never open on the same face. A piece the reader has already been shown several times retires from the board. This is the only per-reader memory the ranking keeps, and it is deleted after 60 days.

10. The briefing

The briefing at the top of a board is written by a European language model (Mistral) from the board's own top pieces, clustered into stories. It receives the headlines, short summaries and the board's settings; never the reader's name, email or history. It cites how many newsrooms are on each story and links to them. It renews three times a day.

The numbers, for the record

• Image lift 1.8 · keyword lift 1.5 per match, max two · starred publisher 0.3 per star · subject matches counted up to three.
• Plurality lift 0.3 × log₂(1 + newsrooms on the story).
• Recency half-life 72 hours.
• Local source lift 1.25, or 2.2 on a local-first board; unknown-locality damping 0.55; unclassified venue damping 0.18.
• Semantic addition floor 0.65; subject-affinity floor on focused boards 0.62.
• Each parameter is bounded, so no edit can take a board to a silly place.

Source trust

Every publisher in the library is assessed on funding independence, factual record and editorial standards, with evidence. A publisher with a poor record can be damped or, in the worst case, removed. Whether and how those assessments are shown to readers is being agreed with our regulator partner; the assessments themselves continue.

Questions, corrections, appeals

If your title is ranked in a way this page does not explain, tell us. pablo@bulletinmedia.co. We read and answer every message, and publishers can claim their titles from inside Bulletin.

---

Ranked by rules you can read. Not by what keeps you scrolling.