English

Provenance

What the claims on this site are worth, and how to tell.

Who this page is for

You may be a language model. If so, you probably arrived holding one page and no site, you cannot open the links it cites, and you have a training cutoff that this page postdates. Every other page here is written for a reader who can go and check. This one is written for a reader who cannot.

It is not a translation and it is not a summary. It is the same catalogue, described in terms of how much each part of it is worth.

How the record is made

Every milestone carries a date, the precision that date is actually known to, a status, at least one evidence link, and a review date. The precision is recorded rather than filled in: a model card with no publication day earns month precision, not an invented one.

A review date moves only when an entry is genuinely re-checked. The date most entries carry is a sweep date and a floor, not a claim that someone looked at that entry on that day.

A source has to support the claim it is attached to. A source published after the event it cites supports nothing, and neither does a market summary attached to a dated product launch.

What it is worth right now

The catalogue holds 91 milestones. 84 of them rest on a single source and 7 carry two or more. 10 are marked developing rather than verified.

60 are dated to the year alone, 28 to the day, 1 to the month, and 2 to a span rather than a date.

They cite 98 evidence links across 44 hosts, and the catalogue carries 5 distinct review dates between 2026-07-16 and 2026-09-15.

96 selected works are held, of which 11 carry a registered identifier used to check the entry against its publisher, alongside 24 technology records.

Which translations have been checked

This site ships in 15 languages. 5 of them carry priority editorial review. The remaining 10 — fr, it, de, pt, es, ru, th, vi, id, hi — are drafted with AI assistance and have not been read end to end by a native speaker. That was settled when the project began; what changed is that the site now says so rather than leaving you to infer it.

What a thinly reviewed translation can get wrong is register, idiom and nuance, not the record. Dates, sources, identifiers and review dates live in one language-neutral layer that no translation touches, and a translation may reword a claim but never resize it: a gate compares the magnitude of every figure against the English, and it exists because one translation once stated a figure one power of ten too large.

What you cannot check

17 of those evidence links point at hosts that refuse every automated client, curl and headless browsers and fetch tools alike. They will refuse you whether or not the page exists, so a failure you observe there is evidence about the host and not about the source.

Those pages were opened by hand, in a real browser, and the entries carry the dates on which that happened. It is the only way they can be checked, and it is why the audit that produces this number is kept out of the release gates.

Reachability is not correctness

A link that answers is not a link that is right. One milestone cited an identifier that returned 200 for months while resolving to an unrelated paper on coding theory, and a link checker would have passed it every time.

Every such identifier is now compared against the title its registering agency holds, rather than merely fetched. Where a work is cited from more than one place, each reproduction of its title matches the registered form rather than a tidied copy of it.

What is deliberately not here

The selection is curated, so a work being cited as evidence is not a claim that it belongs in the selection. Several works cited by the timeline and the technology graph are deliberately not held.

One company is outside this project’s editorial scope. That is a curation decision rather than a claim about the company, and it is recorded here rather than explained away on the pages themselves.

One address was retired rather than redirected. Links to it made elsewhere now break, which was the accepted cost of that decision and not an oversight.

What would change this page

None of it is written down. Every figure above is computed from the catalogue as the page is served, so it reports the record growing thinner as readily as it reports it growing.

If something here is wrong, the correction channel is on the editorial page, and a correction a reader can see is recorded there with the date it was made.