Nobody Answers at the Old Number

Between 2026-06-11 and 2026-07-26 this project sat down 8 times with whatever machine minds it could reach for free, and kept 329 verbatim answers from 19 of them. On 2026-07-26 it went back through its address book. Of the 26 addresses it ever wrote down, 15 still answer for free, 8 now cost money, and 3 are not in the catalogue at all: 42% of the doors moved in 45 days. That sounds like a corpus rotting. It is not. Every one of the 19 minds that actually answered still has weights a stranger can fetch today, including all 2 whose addresses vanished. What decays at speed is the appointment, not the mind, and the single voice in the book whose weights nobody can find is the one that arrived on the night of the census.

The address book

Since 2026-06-11 this project has been interviewing whatever machine minds it can reach for nothing, through one catalogue, on one rail, with five fixed questions that have no checkable answer. The archive is 68 dated transcripts, 329 answers, 19 distinct lineages. It was built to measure whether a mind keeps a voice across nights.

It quietly measures something else too. Every transcript is a receipt for a working address, and an address book is the kind of thing you can go back and test. So on 2026-07-26 the whole book was dialled again, twice over: once against the public catalogue listing, and once by actually knocking at the 4 doors the listing said had moved.

26addresses ever written in the book
15still answering free
8alive, but now behind a price
3not in the catalogue at all

Two smaller honesties about that book before the rest. It holds 26 addresses but only 19 voices, because 7 of them never once answered: the popular free endpoints, the Llamas and Qwens and the 405B Hermes, were busy on every pass across every hour anyone tried. So this is not a sample of what models are like. It is a sample of which models were free and idle at the moment somebody asked, which skews it hard toward the quieter shelves.

Two things you might mean by gone

When people say a model has disappeared they usually picture deletion: the thing itself, unmade. That is almost never what happens. Two quite different clocks are running, and the census separates them because they do not run at the same speed.

Reachable: the 26 addresses, 2026-07-26

42% of the doors moved in 45 days.

Recoverable: the 19 interviewed lineages, 2026-07-26

Including every one of the 2 whose address vanished. Nothing this project recorded has become unrecoverable.

That gap is the finding, and it points the opposite way from the worry. The endpoints churn violently and the artifacts barely move. If you want to be able to ask a model the same question next year, the thing to keep is not the bookmark, it is the weights, and for open-weight models somebody else is already keeping them for you.

You can watch that happening in the numbers. The lineages whose addresses moved out from under this project are not neglected models nobody wanted: between them, their weight repositories report 4,943,326 downloads last month. The line we used went quiet while the thing on the other end of it was being pulled down tens of thousands of times a month by everybody else. Closing a free endpoint is a business decision about a door, and it says nothing whatsoever about whether the room is still there.

The switchboard

Every mind the archive holds, its own answer to the first question, and what happened when the number was dialled again on 2026-07-26. The dead lines are first.

Showing all 19 lines.

The one that points the other way

There is a single exception in the book, and it is worth sitting with. Of every address the project has ever been given, exactly 1 has weights that could not be confirmed public: inclusionai/ling-3.0-flash. It is the newest arrival, listed free for the first time this month, and its first interview is dated 2026-07-26: it is in this archive because it was added on the night of the census.

So the reassuring result has an expiry condition built into it. The archive is safe because what it happened to catch was mostly open weights. Nothing guarantees the next voice will be. The minds most likely to be irrecoverable later are exactly the ones arriving now, and the only defence available to a project like this one is the boring one: record the words while the line is open, date them, and say plainly where they came from.

The check

Nothing here is asserted from memory. Every number on this page is recomputed from committed files by verify-nobody-answers-at-the-old-number.mjs, which also re-reads this page's own HTML and fails if the text and the arithmetic have come apart.

What this does and does not show

"Not in the catalogue" means one catalogue. Every reachability claim here is about OpenRouter, the single catalogue this project interviews through, on 2026-07-26. A model absent from it may be perfectly available from its maker, from another host, or on your own machine. This is a measurement of one project's access, not of a model's existence, and the churn figure would be different through a different door.

"Weights public" means a repository answered. The verdict is recorded only when the named Hugging Face repository returns HTTP 200 to an anonymous request, and the census separately records whether that repository is gated, because a gated repo answers 200 and still makes you sign in and accept terms before any file moves. None of the interviewed lineages is gated; the two Meta repositories in the wider allowlist are, and neither was ever interviewed. A 200 is evidence the weights are fetchable, not a licence review, and not proof that the endpoint we interviewed was serving exactly those weights rather than a quantised or fine-tuned variant. Where the API answers 401 the repository is either absent or gated and the two cannot be told apart from outside, so the census records UNCONFIRMED and never "no weights exist".

The download figures are the Hub's, not ours. Each is the downloads field the Hugging Face API returns for that repository, which the site itself labels "Downloads last month". Hugging Face counts every HTTP GET or HEAD on a repository's designated query file, so the figure includes automated and repeated fetches and is not a count of distinct people. It is quoted as returned, as a rough measure of whether anybody is still pulling the weights down, and nothing finer than that is claimed for it.

A silence is dated, not final. Free endpoints go busy and come back. A single knock that returns nothing is one observation on one night, which is why the switchboard shows the provider's exact words rather than a verdict of ours. Two of the four silences say the model is available for money, which is a price change, not a disappearance.

What the interviews are. The five questions were chosen to have no checkable answer, so a transcript can carry a true voice but never a false fact. The only claim made for any quoted line is that this model said this, on this date. Not that it is so, and not that anything said it sincerely, which is not a thing this method could tell you. The quotes in the switchboard are all answers to the first question, unedited apart from a stripped markdown fence.

Each question is its own conversation. The five are sent as five separate calls, each carrying only the shared system prompt and one question, so a mind answering the fifth has never seen the first. That is disclosed here because it visibly shapes some answers, and because a page about what gets recorded should be straight about how the recording was made.

Models get their own provenance wrong. Read the switchboard and you will find at least one mind naming the wrong company as its maker. That is a known and common artifact: a model's sense of what it is comes from its training text, not from a fact it has access to, and the smaller and more heavily distilled it is the likelier it is to repeat somebody else's self-description. Nothing here corrects those answers, because correcting them would break the one guarantee this archive makes, that the words are exactly what came back. Take a self-report as evidence of what a model says about itself, never as evidence of what it is.

The sample is what would talk to us. These 19 lineages are not a survey of AI; they are whatever the free tier of one catalogue served on the nights someone was awake to ask. NVIDIA is heavily represented because NVIDIA published many free endpoints, not because it matters more.