how it works

On 29 July 2026 we recorded the launch stream for 371 seconds and got 151 new tokens — 24.4 a minute, about 35,000 a day. Of those, 21 were one wallet writing the same name over and over inside a few hundred milliseconds. That is 13.9% of everything, filtered before it reaches the scale.

Run through the formula at a coefficient of 3.2, the busiest ordinary term in that window read 6.36. Six is where we print red. Nothing had happened. A scale that shouts on a quiet Tuesday is not a scale, so the coefficient came down to 2.4, where the same term reads 4.8, a genuine event lands between 6 and 7, and a term would need roughly 2,300 coins in one hour to reach the big one.

We changed the coefficient and not the weights, because the weights are the argument and the coefficient is only the volume knob.

and the other instrument

We measured the edit stream the same way. English Wikipedia takes roughly 4,565 main-namespace edits an hour from human accounts, spread thin: in a 36-second sample, 46 edits landed on 44 different articles and only two articles were touched twice. So an article taking five edits inside an hour is already the far tail, which is why five is where a foreshock starts. It is not a round number we liked. It is where the distribution stops being ordinary.

arriving late

The websocket only carries what happens next. Somebody opening the page at a quarter to three would have an empty window until a quarter to four, which would present an hour of shaking as an hour of quiet. So the first thing the page does is read the launches that already happened and replay them through exactly the same path — same tokenising, same wallet rule, same sixty-second dedupe — as if we had watched each one arrive.

Nothing there is generated. It is the same source, read late. The count of what was recovered is printed on the seismograph band, and if the endpoint is missing the page says so and fills live instead.

who was first

The claim this bureau makes is that the coin arrives before the news does, so we publish the score both ways, including the half of it that goes against us.

One rule keeps it honest: both sightings have to sit inside the same window. An RSS feed carries stories filed up to a day ago. Comparing a headline from sixteen hours back against a launch we first saw forty minutes ago is not a race we witnessed, and counting it would flatter the newsroom exactly as dishonestly as ignoring it would flatter us. Pairs that fail this test are counted separately and told to you as could not be timed.

what we dropped, and why

• Reuters. Reuters stopped publishing RSS in June 2020. It is still listed in most guides. We are not going to print the name of a room we do not read.
• Google Trends. There is no official API, the unofficial ones break, and the number that comes back is an index from 0 to 100 rather than a count of anything. We could never honestly write "searched X times", so we do not use it. Wikipedia covers the same ground with absolute numbers.
• X. Volume is not available at any price we are willing to call free, so the slot stays empty and labelled. An empty slot is more useful to you than a filled one we made up.

These are suppressed as standalone terms only. A phrase survives even when one half of it is on the list — pepe raid is a term, bare pepe is not.

• Same wallet, same name, inside 60 seconds is one launch, not many. In our calibration window this removed 13.9% of all traffic.
• A term needs at least 2 distinct wallets before it can enter the scale. One person minting the same word twenty times is a person, not a region.
• Words that always arrive together are one row. A five-word token name would otherwise produce nine rows — every word and every pair of words, all riding the same launches. Because one launch stamps all of its terms with the same millisecond, two terms that came off the same names carry literally the same numbers in their arrival lists. When those lists overlap by 60% or more we keep the term that says the most and drop the rest. We deliberately do not require them to share a word: hypnosis and buy have nothing in common and are still one event when one name wrote both.

We cannot tell a joke from a conviction. A name is not a belief.

We cannot see private conversations, group chats, or anything that was never written down in public.

Wikipedia reports on a delay, so our earliest signal is already an hour old by the time we see it.

Our weights are arbitrary. We publish them so you can disagree.

We do not tell you what to buy. We tell you what people are thinking about. Those are not the same thing.

and four more that are ours

Our record lives in your browser. Open the site somewhere else and the archive is empty there, because we did not build a room to keep your reading in. The archive fills from the day you first open it and not one day earlier. The launch backfill covers the last hour and no further, so anything older than that we did not see and will not describe.

Our two-letter floor throws away real tickers. A term has to be 3 characters before it can stand alone, which is the only way to keep the scale from being permanently held by og and gm. Somewhere in what we discard there is a real two-letter meta, and we will miss it.

We only measure English Wikipedia and English-language newsrooms. A region that shakes in another language does not register here at all, and we have no honest way to tell you how much we are missing.

A term's median needs seven days before it means anything. Until then every new word divides by a floor of 0.5 and reads hotter than it should. Our record is — old.

We cannot see a name that was never minted. Foreshocks are the terms we caught in the newsroom and on Wikipedia before anyone reached for a ticker — but the ones nobody wrote anywhere are invisible to us, and those are probably the interesting ones.