Method

1

Volume is not flat across the week. Counts are not flat across the week: several outlets publish nothing from Friday evening to Monday, so a weekend day runs about half a weekday. That is a property of the market rather than a collection failure, and it is the reason every threshold in a report is measured against its own market rather than against a fixed number.

2

Duplicate feed entries. One article often arrives twice. A feed can carry it under two identifiers, and a publisher can carry it in a section feed as well as the main one, so the same address appears more than once. Entries are merged by owner and address before anything is counted — a duplicate is textually identical, lands in the same cluster and would otherwise inflate both the size of a story and the concentration of its largest owner, which are the two numbers everything else is built on.

3

What a story is. Articles are grouped by cosine similarity of sentence embeddings at a per-market threshold of 0.895, 0.925, 0.93, 0.945, calibrated against the point where clustering collapses into one component. A story needs at least 3 articles. The threshold is a property of the market rather than a constant: the distribution of similarities depends on the language, and a number calibrated on English breaks elsewhere.

4

Owners, not outlets. Owners, not outlets, are the unit of independence, so two titles under one parent count once. Two papers owned by the same group share a newsroom and a press release; counting them as two voices would make every measure of synchrony report the shape of the ownership register instead.

5

Wire copy is counted once. Owner counts are corrected for wire copy. When two owners publish a word-for-word identical headline, that is one agency text under two mastheads rather than two newsrooms deciding the same thing, so they count once between them.

Re-headlined copy is caught a second way, by the clock: an outlet that republishes an agency article keeps the agency’s timestamp, so the two go out in the same second — which a smooth spread of publication times cannot produce. A pair has to show that pattern across the whole archive before it counts anywhere. The correction is small in most markets and large in uk-regional, where a third of stories run on shared copy. It is still a floor: an outlet that stamps the copy with its own time counts separately.

6

The words under a headline. Where a story carries a few words beneath its headline, those are the terms that separate it from the rest of its market that week, weighted so that neither a common word nor a word seen twice can win. They are a label and nothing more: no measurement in a report is computed from them.

7

Why stories are clustered inside a market. Stories are clustered inside a market and never across them. That is deliberate: the BBC and the New York Times covering one event is an international story, not two newsrooms moving together, and mixing the markets would turn every large event into a false signal. The boundary is worth reading on its own, though, because most of what a market carries never crosses it.

8

The cross-market match is deliberately loose. Two stories count as the same subject when their centroids are close on the same day or either side of it, any match counts rather than only the best one, and the similarity threshold is set well below the one used to chain stories across days. The claim being made is “nobody else ran this”, so the measure should fail towards saying a story travelled. At the strict threshold the shares come out lower, and each report gives both.

9

Carried by one market alone. The other side of the same measurement is what one market carried and nobody else did. intl-state — the state-funded broadcasters, kept as a separate market precisely because they are a labelled example of coordinated promotion — is the interesting case. Being carried alone is not itself suspicious: local interest looks exactly the same, and uk-regional is mostly local by construction. It is a list worth reading rather than a score.

10

How days are chained into one story. A story is clustered one day at a time, so a week-long story is a chain of daily clusters rather than one object. Days are joined when each is the other’s closest match — not merely similar enough — because a similarity threshold alone welds the whole archive into a single chain. Daily fixtures that repeat under the same headline, such as sports listings, are joined too but marked separately and left out of the tables.

The comparison happens after each day’s average direction is subtracted from every centroid. Without that subtraction the measure tracks cluster size rather than subject: unrelated clusters already sit at 0.81 cosine of each other, and the larger a cluster is the closer it drifts to the average of everything.

A day of silence does not end a run: where nothing matched the next day, the day after is tried as well, but only for stories left without a continuation at all. Without that, one run in eight ended for want of a quiet Sunday rather than for want of interest.

11

What gets into the table of long stories. Ordering is by how much of the reported week a story occupied, and it has to hold at least three days of it to appear. Ranking by total length instead would fill the table with stories that began before the window and had longer to accumulate. Where a story did start earlier, the dates say so, and the owner counts describe the whole run rather than the part inside the week. A run that never reached more than a couple of owners is left out as one newsroom’s routine rather than a story the week shared.

12

Opening wide against widening over days. The column worth reading is the last one, owners on day one against owners in total. A story that opens wide was already everyone’s story on day one — a national accident, a death. A story that opens narrow and widens over days is the other thing: it had to travel before it became everyone’s. Which of the two a subject belongs to is not visible in a single day’s snapshot.

13

How a story fills up. Counting owners per day says how wide a story got; it does not say how it filled up. A row per owner does: an owner that opens the story and keeps at it looks nothing like one that arrives on the second day and publishes a burst. The wave along the top is deliberately coarse. At an hourly step the shape is dominated by nightfall — every story in the archive shows the same nightly gap — so what would read as a rhythm of the story is really the rhythm of the working day.

14

One story is not a claim about owners. An arrival chart is one story, not a claim about how those owners behave. Across the archive so far, the correlation between how late an owner joins and how much it then publishes is −0.23 — if anything the early owners write more — and it is confounded by size, since an owner that publishes more has more chances to be first. Saying anything firmer needs the kind of archive that takes months.

15

Co-occurrence is not coordination. A high score means two owners run the same stories more often than their volumes explain. It does not by itself mean they coordinate: two outlets sharing a beat — the same region, the same political lane, the same war — score the same way, and so do two mastheads running one agency’s copy. What the score cannot say is whether either of them is following the other. For that the measurement has to use the clock.

16

The three shapes a tight pair can have. Direction is what separates the mechanisms, and it is the part a co-occurrence count cannot see.

• One way. The follower follows; when the leader happens to be second, it does not hurry. A state agency feeding a state broadcaster has this shape.
• Both ways. Whoever publishes first, the other matches at the same speed. Two competing newsrooms watching each other have this shape — it is rivalry, not coordination, and it is easy to mistake for the first kind if you only look at how tight the pair is.
• Identical text. The two publish inside the same second, because it is one article under two mastheads. Pairs that show that across the archive are removed before any of this is measured, whether their headlines match or not; a pair that did it too rarely to prove still surfaces here.

17

The two controls a pair has to survive. Timing is measured over the whole archive rather than one week, and a pair has to survive two controls before it appears. Each owner’s own general speed is subtracted, so an outlet that is quick with everybody does not end up looking tied to everybody. Then the pair’s own tightest quarter of stories is dropped, so a single shared subject cannot carry the result on its own.

Inside every story each pair of owners is ranked by how close together the two published, measured against all the other pairs in that same story. Scoring against the story itself is what keeps big news out of the result: an event everybody reaches within the hour makes every pair look fast, and only a pair that is fast relative to its own story survives.

18

Upstream and downstream of one owner. The same measurement can be asked of a single owner rather than a pair: does this one tend to be upstream of the newsrooms it shares stories with, or downstream of them? Asked directly, the question measures office hours — an outlet that publishes in the morning is first against everybody. What makes it answerable is a shuffle: each owner keeps its own publication times for the day, and only which story each time belongs to is redrawn. An owner that is still first afterwards was first because of its schedule; one that stops being first was first because of the story.

19

How to read the two ends. The two ends read differently. Upstream is where the wires and the rolling broadcasters sit, which is what a working measure should say; downstream is mostly commentary, where the article exists because the story already happened. Neither end is an accusation: being downstream of a wire is the ordinary way a newsroom covers the world. The one reading worth taking seriously is a market where the two ends are the two halves of one state broadcaster.

20

What looks like a rewrite and is not. Every version of an article’s text is kept, so a headline that changes after publication is visible — but not all of it is a newsroom changing its mind. A wire story that rolls under a single address is rewritten as the day moves, a live blog changes its headline on a schedule, and a feed occasionally reuses an entry for a different article days later. All three look like a rewrite and none is one. Requiring the two versions to be recognisably the same sentence, changed within a day, is what separates them.

21

The only decision observed twice. This is the one place in the corpus where an editorial decision is observed twice — the same newsroom, the same story, two framings hours apart. Everything else measured here is the shape of how a story spreads between publishers; this is the framing itself, and the second version is what the newsroom preferred once it could see how the story had landed.

22

Where the examples come from. The examples are drawn from stories that at least 4 owners carried, not from the largest rewrites: ranked by size of change alone the list fills up with drafts that reached a feed by accident and with magazines retitling essays, which is true but not about the week.

23

Comparing rewrite rates between owners. Compare rewrite rates only roughly. A feed that keeps an entry visible for longer gives more chances to observe the second version, so part of the spread between owners is the feed rather than the newsroom.

24

Why marking rather than distribution. This project is named after the Overton window, and everything else it measures is something else: who carries a story, when, and together with whom. That is the shape of distribution. The window is a different question — what can now be said without apologising for it — and it is not answered by how far a word travels. It is answered by how the word is marked.

25

Why the list of terms is fixed in advance. A contested word travels in quotation marks; a settled one does not, whatever people think of the thing it names. Nothing about the count knows which words are contested — the list is fixed in advance, and it has to be. Ranking every word in the corpus by this share was tried and returns absolutely, very, deeply, I’m: quotation marks in news carry direct speech far more often than they hold a word at arm’s length, and the two cannot be separated without knowing who is speaking.

26

A level, not a movement. Read a week’s shares as a level, not as a movement. A single week says how these words are handled now; the window is the change in that share over months. The comparison that will matter later is a term against its own past, never one term against another.

27

Generated, not written. Reports are generated without human editing and without a language model. Where the text sounds hedged, that is deliberate.

28

Two sections cover fewer markets than the rest. Rewritten headlines and quoted terms are counted over the four English-language markets only, while every other section covers all of them. The reason is the measurement rather than the corpus. Both sections read words, not timing: one discards live tickers and looks for so-called, alleged, reportedly; the other counts how often a fixed list of English terms sits inside quotation marks. Neither list exists yet in the other languages, and a term like woke does not occur in a Russian newsroom at all. Widening the corpus under an English detector would not add markets to the count — it would add articles the detector cannot read, and the printed number would drop without saying why. The quoted-term shares are also the one series here published week after week, and a series has to be counted over the same corpus each time or it is not a series.

29

Where the bar for a pair is set. Both pair tables show only scores above three, and the bar is measured rather than chosen. The same permutation that produces a score can be asked a second question: with the link between owners destroyed, how many pairs still clear the bar? Over the archive to 11 September 2026 a bar of two lets through roughly one chance pair in eight for co-occurrence and three in ten for timing; a bar of three leaves under one in twenty. Splitting the archive into halves says the same thing from the other side: co-occurrence rankings hold across the two halves, timing rankings largely do not, so a timing pair is worth reading only well clear of the bar. Neither test makes a pair meaningful — it only says the pair is unlikely to be an accident of volume.

30

Why the week-by-week numbers move. The series is recomputed from the whole archive every time a report is built, with the current code and the current thresholds. It will therefore not always agree with the number printed in an earlier report: articles arrive late, clusters are rebuilt, and what was one story last month can be two today. Keeping the series internally consistent is worth more than agreeing with old posts, so it is never stitched together from them. A market’s row begins at the first week it was collected on all seven days and with most of the feeds it has now. Earlier weeks are left blank rather than printed as a low number: while a market is still being wired up its feeds arrive a few at a time, and the count would show the collector starting rather than anything in the world.

31

Wide is not the same as loud. How many newsrooms carried a story and how much was written about it are different questions, and the second is easy to mistake for the first. The share is of all the articles in a story that came from a single owner, counted after owners running word-for-word identical copy are collapsed into one. A story can reach a dozen newsrooms and still be mostly one newsroom talking. A story whose articles hang together unusually loosely for its market that week is left out: that is the mark of separate events chained into one cluster, and the share of its largest owner then describes the chain rather than any story.

32

Feeds that produced nothing. Every feed is polled hourly whether or not it has anything new, so a feed that produced no article all week is worth counting. Silence is not one thing: a publisher’s outage arrives as a CDN error, a bot challenge arrives as a normal-looking page, and a genuinely quiet regional weekly looks like both. The count is printed because it bounds what the rest of the report can see, not because each silent feed is a fault.

33

When a story crosses a language. Markets are matched to each other by meaning, not by words, so a story can be recognised as the same event in two languages. The threshold for that match is set per language pair rather than once for all of them: the similarity a translation can reach falls with the distance between the languages, and a single English-calibrated number cut the Arabic market off almost entirely while letting the German one through. Two checks say the lower thresholds find real matches and not noise. Repeating the comparison against days four apart — where no pair can be about the same event — leaves almost nothing. And the matched pairs turn out to be published two to three times closer together in time than an arbitrary pairing from the same day, which the matching itself knows nothing about: it never sees a clock. What none of this shows is direction. Nothing here says one language picked the story up from the other.

34

The one measurement with a known answer. Every other number in this report is descriptive: it says what the corpus looks like, and nothing outside the corpus can confirm it. The state broadcasters are the exception. They are kept as a market of their own because they are a labelled example of coordinated promotion, but the label does not apply to the market as a whole — Al Jazeera and Chinese state television have no connection. It applies to particular pairs: two Turkish state outlets, two Chinese, two Russian. A measure of who moves together has to put those above the cross-border pairs, and where it ranks them is printed rather than described, so a week where it fails would be visible here first.

35

A pair has to survive half the archive. Timing pairs are shown only when they hold up in both halves of the archive counted separately, not only in the whole of it. The reason is a measured one: rankings by co-occurrence are stable when the archive is split, and rankings by timing are not, so a list built from the whole archive alone would partly be a list of accidents. The bar for each half is lower than for the whole, because a score of this kind grows roughly with the square root of how much was counted and half an archive cannot reach the same number.

36

Standing sections are not stories. A cluster can be entirely real and still not be an event. Publishers run standing sections on a schedule — the weather for forty cities, exchange rates for seven countries, lottery draws, horoscopes — and those arrive as hundreds of near-identical items a day, which clustering groups together exactly as it should. Such a cluster is separated before anything is measured, and the test is the shape of the run rather than its subject: five or more items from one owner whose headlines are templates of one another, compared word for word the way syndicated copy is, at a lower bar because a template varies its slots rather than repeating verbatim. A cluster in which such runs hold half the articles is counted as a section, not as a story, and is left out of every measurement of who moves with whom. The rule reads no dictionary and knows no language, and it fires only where the thing exists: across the English-language markets it marks nothing at all, while in one Spanish-language market it marks four per cent of the clusters, which hold nearly half of all the articles that end up clustered at all. A second reading catches what a run cannot: a small model over the same vectors the clustering already uses, trained on what the first rule found and measured against a week it had not seen. It is what recognises the sections that never form a run — one lottery draw a day, one city’s forecast, one horoscope — and it does so in languages it was never shown, because the encoder beneath it places the same thing said in two languages in the same neighbourhood. Either reading is enough for a cluster to be counted as a section, and the threshold for the model is deliberately set where it errs towards leaving a cluster in.