What this archive stores

How a sentence gets onto this site

A stored artifact changes. A pure function over the two versions emits a typed claim with a fixed template. The template is filled from the diff and from the sidecar committed beside the bytes. There is no language model anywhere in the generator, no summarisation step, and no editorial pass.

Every value read out of a stored payload is printed inside quotes, and a value carrying a quote of its own has it neutralised before it is printed. That is not typography. A value interpolated bare into a sentence is prose this project publishes under its own byline, and a catalogue id or a pull-request title is chosen by somebody else.

What a sentence here never says

The subject of every rendered sentence is an artifact, never a company and never a reason. "OpenRouter's catalog context_length for X changed from A to B" is an observation. "X's usable context was cut" is an inference, and one live case shows why it is the wrong one: a catalogue context_length rose from 1048576 to 1310720 while the top_provider.context_length recorded beside it fell from 1048576 to 262144. The headline number went up by a quarter while what the routed provider serves fell by three quarters.

Personal data, stated rather than implied

Most sources here are vendor documentation indexes, sitemaps, status feeds and machine-readable catalogues. Three are not, and they are not equally risky, so this section names which is which rather than grouping them.

The one that actually reached these pages is modelsdev-commits, an Atom feed of commits to models.dev. An Atom commit entry carries an <email> for every author and for every Co-authored-by: trailer, and the change pages render a diff of the stored bytes. Nine distinct addresses were published that way, four of them personal. They are no longer rendered: every address in a displayed diff is masked, with no exception for role or noreply addresses, because a rule that spared those would have to judge which addresses are personal and would publish a private one every time it judged wrong. The count of masked addresses is printed under each diff that has any.

The other two are transformers-pulls and vllm-pulls, GitHub pull-request searches whose payloads carry each pull request's full description text and its author's login, numeric id, avatar URL and profile URL. Those fields are stored but never rendered: the derivation projects only the pull request's number, title, state and merge timestamp. That was already true when this section named these two sources and did not name the one whose data was on the page, which is the failure mode a disclosure like this is supposed to prevent.

Masking is a property of the publication, not of the archive. The stored bytes still contain the addresses, and every diff sits beside a permalink to the exact artifact at the exact commit, so nothing here is unverifiable. That is also the unpaid half of the tradeoff: this repository mirrors content written by named private individuals and recommits it into a history that is never rewritten, so an erasure request against it cannot be satisfied without violating the archive's own central rule. That is a real cost and not a technicality.

Timestamps

A published timestamp is the sidecar's origin_date, which is the response date minus its Age header, so it is when the provider generated the bytes. Where the provider sent no Age, the page shows observed_at, which is when this collector saw them, LABELLED as observed. One captured response carried an origin fourteen hours before the fetch, so the two are not interchangeable and neither is silently substituted for the other.

A first-seen date is shown only where the measured worst-case error for that source is narrower than the resolution being printed. A source captured once a day cannot support a claim about which day something appeared, so that page prints the reason instead of the date.

Nothing is deleted

A change that turns out to be wrong is retracted, not removed: the page and its artifact link stay resolvable and are marked. The retraction ledger and the leaks desk's accuracy ledger are both append-only and both enforced by a workflow that fails any diff removing or modifying a line.

That rule has been paid for once. A vendor published a live API key inside its own public documentation file, and the collector's credential gate now holds any snapshot carrying a credential pattern out of the archive rather than committing it. Separately, a test fixture captured from a challenge page carried an AWS access key id and a set of presigned URLs, all of which have since expired. The working tree, the tracked tree and every built page are clean, and the fixture no longer asserts against a real credential. What remains is in history, because history is not rewritten here, and saying so is better than letting a reader who finds it conclude nobody looked.