A measured wire for AI. Software watches the artifacts; a human owns the questions, thresholds and corrections. What happened, who filed on it, what actually shipped, and where the field is discussing it — readable in two minutes, rebuilt every thirty. Measurements are computed from primary artifacts we read ourselves: model cards, file listings, release feeds, price catalogues. Headlines come from each publisher's own feed rather than an aggregator, and are linked and attributed, never rewritten — with the publisher's own summary underneath when their feed carries one.
Headlines about one event cluster into one card — matching weighs shared words by rarity and treats named entities as single tokens, and it is tuned to prefer a missed merge over a wrong one, because a wrong merge hides a story. Ranking uses countable facts: how many independent outlets filed, whether the principal went on the record, whether weights can actually be loaded, recency — and, since 2026-09-03, measured attention: the votes and comments a story drew on Hacker News, likes on Bluesky. Attention lifts a card and never carries one alone; a name in a headline counts for less than a conversation under it. Nothing here reads the words of that conversation to decide rank, so a thread of insults with no numbers in it still loses to a price that moved.
Three voices, kept strictly apart. Headlines are quoted as published and attributed, and so is the short summary a publisher syndicates in its own feed beside the headline — shown in quotation marks, cut from nothing, and dropped rather than trimmed if it runs long enough to be the article. Our own sentences are written by software: the one-line brief for a story from its headlines and, since 2026-09-03, the page it links, and the opening paragraph from measurements this site computed. Both are checked before they publish — every number must appear literally in the evidence, every model named must be one that was measured, and the paragraph may not say why anything happened, because why is not in the archive. A sentence that fails is dropped, not rewritten. Community reactions are verbatim, attributed, linked excerpts, and a quotation is never changed by a word: a long comment is cut at a fixed length, the cut is marked with an ellipsis, and the link beside it goes to the whole of it. Since 2026-09-03 the site also says, in its own words and under its own label, what a thread is arguing about and what the day's issue is — written from the comments and the page it read, checked so that every name is someone quoted on the page and every number is one that was counted, and never presented as anyone's words but ours.
Some of the sentences on the board are written by this site rather than quoted from anyone, and each carries the stamp of the software that wrote it. Three of them report and may only say what the evidence says. The fourth, VIEW, judges — it is an opinion, stamped so you can tell it from the other three at a glance, because the line between the two is what makes the other three worth anything. Nothing else here is stamped: a headline belongs to the outlet that published it and a comment belongs to the person who wrote it, so neither gets our name on it.
On the Korean board, headlines, quoted comments and publishers' own summaries carry a Korean rendering, and the original is always on the page underneath it, in the words it was published in — folded behind a control on a comment and on a summary, because printing both in full doubled the height of every card, and open on the page for a headline. Quotation marks are reserved for the published words: where the Korean is ours, it is set without them. The translation is a reading aid and it is ours; it is never presented as what the source actually said, and it never replaces the original. Quoted words themselves are not rewritten. Nothing on the English board is translated at all. Where a headline is only an identifier — a release tag, a repository name — it is left exactly as it is, because there is nothing there to translate and a model asked to try will invent something.
No analytics of our own, no cookies, no accounts. There is no form and nothing to hand over. The "new since your last visit" marks — and the button that shows only those — are computed inside the page from a capped list of links your own browser has been shown; it never leaves your device and there is nowhere here for it to be sent. Two requests do leave, and this paragraph claimed for months that none did. The page asks its own address for a fresh copy every five minutes, so that a board rebuilt while you are reading can say so; it sends nothing with the ask. And Cloudflare, the host, adds its own performance beacon to every page it serves — anonymous load timings, written by them and not by us, but a request all the same, and calling it none was simply wrong. Cloudflare also shows us aggregate counts of pages served, never to whom.
Wrong is worse than absent: a brief or a merge that fails its checks is dropped, not patched. When something wrong does reach the page, it gets corrected here, permanently and dated. 2026-08-27 — the privacy section said this page makes no requests. It makes two: a check of its own address every five minutes, and the performance beacon Cloudflare adds. Corrected the same day, after an outside reader checked the page against the page. 2026-08-27 — the same section said nothing anyone wrote is ever shortened, on a board where a long comment is cut at 240 characters. Corrected to say so. 2026-08-27 — a glossary entry meant to fix a misspelling of Claude was replacing 클라우드, the ordinary Korean word for cloud, so a quoted Hacker News comment read "our many exciting Claude" directly above the English original that says cloud. That is a word put in someone else’s mouth, which is the one thing this board must never do. The rule is removed and the four affected translations were redone. 2026-08-27 — a merge joined an OpenAI post about a customer with an unrelated security story because both said "Codex", and the brief written from it said Codex installed unowned code inside corporate networks, under the headline of the promotional post. The merge and the brief are gone; Codex is now a name the clustering knows. 2026-08-27 — a price fact said tencent/hy3 raised its output price 60%. It did not move: it read low for one day out of ten and the return was reported as a rise. A price now has to survive a second reading before it is called a change. 2026-08-28 — four wrong claims reached the model pages before adversarial review caught them, and the pattern is worth recording because it is not carelessness: each was a sentence that read as a measurement and was an assumption. A field named smallestBytes held the median, so "smallest complete set" was wrong by up to 2.1x in the direction that says a model will not fit. "No quantized build has appeared yet" was asserted from a 50-newest-per-org sample on 418 pages, one of them beside its own published GGUF page — an absence needs an exhaustive search and this pipeline never runs one, so the sentence is gone rather than fixed. A parameter count read stored tensor elements, printing 4.8B for a 27B model and 0B for a 14B one. And the Hub tag "8-bit" was read as a format when it is a width, telling readers a Blackwell-only checkpoint runs on ordinary CPUs. All corrected the same day; the publish lane now runs the tests and the typecheck it had been skipping. Found an error? hello@wtfisgoingon.dev.
No human reads article bodies for this board. Since 2026-09-03 the software does read them: the page a card links is fetched once, and what it says becomes evidence for three labelled sentences — the story's one-line brief, the statement of what the linked post claims, and the day's issue block. Nothing read is ever shown. Every number in those sentences must appear in the source, no sentence may run more than seven words alongside it, and one that fails is dropped rather than fixed. The field that carries a syndicated article is still read for two purposes and shown for none: screening summaries, and grounding the event pages' What-happened notes against the announcer's own text. The human work is to decide what is worth measuring, define the ranking and verification rules, audit failures, and own corrections. The software collects, calculates, and publishes within those rules. lashhill — one person. Contact: hello@wtfisgoingon.dev