What the engine found

The engine takes a family of similar events (say, space missions) and looks across hundreds of topics for one whose public attention moves in step with each event, again and again. Every candidate is tested against the same events moved to random dates. Everything here was registered before it was run.

Ripples found

Candidates that passed the screen. A candidate becomes a replicated ripple only when a fresh set of events, never used in the screen, shows it again.

Named-subject test

Each event is tested against its own named subject (a series premiere against the real topic it is about, a death against the person's field), fixed in advance. There is no search over topics, so a signal here is much harder to fake. With about 200 events, about 1 in 20 will pass on its own by chance; the family results are the claims.

Beyond Wikipedia

The same named-subject test, with the subject measured on US TV news (how often it is said on air) and on Hacker News, beside Wikipedia. Fake dates keep each event's weekday. A pass needs p ≤ 0.05 on both scorings.

From attention to behavior

Chain tests follow one event to one thing people actually do, chosen in advance: here, what Seattle Public Library patrons borrowed. A chain passes only if borrowing rose beyond its seasonal normal (compared with every other month) and the rise started in the event's month or later, never before it.

What moved, and what might explain it

The salmon search swims upstream. It starts from something that broke from its own trend (here, a baby name suddenly given far more often), then looks for earlier works that carry the name and could explain it. These links are speculative: each shows why it was surfaced, and how often a match this strong turns up for names that did not move (the chance-match rate). Nothing here claims cause.

How readers got there

Timing tests show that attention moved. Wikipedia's Clickstream (monthly counts of which page readers clicked from) shows how: whether readers walked straight from the event's article to the ripple, or arrived by search or other pages. Pairs with fewer than 10 clicks in a month are not published, so "–" means under 10. Descriptive only.

Calibration: ripples it must find

Before searching, the engine's tests were checked on four well-known ripples and 200 fake ones.

Every family tested

Including the ones where nothing passed. "Best p" is the smallest family-wise p-value over all outcomes; families are corrected together (Benjamini–Hochberg, q = 0.10).

How to read the evidence

Fake-date test

The family's events are moved to 1,000 sets of random dates. The p-value is how often a fake set scores as high on any outcome, so searching hundreds of topics is already paid for.

Two scorings

Each outcome is scored two ways (robust z and rank). Both must reach p ≤ 0.05.

Drop-one stable

Still passes with any single event removed, so no one event carries it.

Linked (echo)

The outcome is linked from an event's own article. Expected, not surprising; it shows the engine sees what it should.

Nearby events

Some events in the family fall within 30 days of each other, so a shared shock (such as the March 2020 lockdowns) could explain part of it.

Replicated

Passed on new events chosen in advance. Still evidence of a pattern, not proof of cause.

Charts are drawn in your browser from Wikimedia's public pageview API (current article titles only) and are illustrations; the tests use the registered data. Protocols: v1, v2, NYT lens, real-world lens, further lenses.