How the atlas itself was built: what was added, what broke, and what was decided against, read off the repository this site is built from.
Some of these entries were written in retrospect, from the commits they describe. Each one names the days the work happened and the day the entry itself was written; where those differ, the entry is a reconstruction and not a note kept at the time.
The rule in robots.txt is lifted: the atlas is open to search engines and AI crawlers, all of it at once. The full Verify with browser tests passed on the head of main, so there was no need to wait for 1 October. Testing on real devices and with a screen reader goes on and does not hold the opening back.
Two more items of the PWA audit. If the page came from the network and the record text from the worker's cache, the card now says the text is a copy, and from what date. And the axis no longer stitches record text from another build onto the page: the card offers to open the page again.
The build now strips the worker’s indentation, and even with the fixes it weighs 3,868 bytes instead of 4,005. A star or a first page whose network hangs no longer holds back everything after it: the worker waits 45 seconds for those. A stalled install was also checked in Firefox and WebKit. Firefox stops such a worker by itself after about 60 seconds, so the limit went down from 60 to 45, as the 6-in-1 chose.
Two more items of the PWA audit. The clean-install rehearsal now copies whatever the repository holds rather than a hand-kept list, and is a repository of its own, so the revision check sees the copy itself. The browser tests count records, indexes and sources the way the site does: only what is published. And a test that failed now and then on the "/" key has stopped failing.
The PWA audit of 26 September showed that if the connection broke while the worker updated, the reader was left without an offline axis. Now a new build takes over only once it holds the whole shell of every language the reader kept offline. Until then the one before stays.
The third request of the 27 September brief fills in medicine, industry and the military before 1990: from Ledley and Lusted’s paper of 1959 to the Pilot’s Associate’s second demonstration in 1988. Nine records, and all three Strategic Computing demonstration systems finally have records of their own.
The PWA audit of 26 September marked five scenarios as not assessed. Each now has a test. Two found real faults: the 501st star silently dropped the first, and the page of a first visit did not open offline. A third showed a choice, and the 6-in-1 made it: when the network hangs and a copy exists, the reader sees the copy after 10 seconds.
The second request of the 27 September brief fills forty empty years of the compute line, from the Mark I Perceptron of 1960 to the ANNA analog chip of 1991. Nine records, and the TPU finally has a predecessor.
A new brief aims at prehistory, and its first request fills sixty empty years of machine proof: from Gelernter’s geometry machine in 1959 to EQP solving the Robbins problem in 1996. Ten records, and for three of them the document said less than the brief.
The atlas is made ready for search engines: a sitemap with the dates records changed, language versions that name each other, and the simplest structured data. The site itself stays closed until a separate pass lifts the rule in robots.txt. The Alpha release section now says that a person has accepted the site and that most records are not yet accepted.
Eight 2026 records without a primary source now rest on the document itself: court filings, the Munich court release, the regulators on Grok, the Nevada permits, the Unitree prospectus, the Kimi K3 card and licence. The document corrected all eight, and no record is left without a primary source.
The fourth request of the 27 September brief fills in Kyiv, Prague and Warsaw before 1991: from MESM’s first start in 1950 to the Mizar library in 1989. Nine records; Poland and Czechoslovakia have records of their own for the first time.
Twenty more records without a primary source now rest on the document itself: the Macy Foundation transactions, the theses of Minsky and Werbos, the statute of the Institute of Cybernetics, the DECtalk manual, exchange data, the pages of BMW, Tesla and Waymo. The document corrected all twenty.
The journal entries on the PWA audit ended with lines of “not verified”. Three of them now have tests. An install that stalled held on for five minutes and left half a cache behind: the worker now waits at most 60 seconds for a file, as the 6-in-1 chose. An old worker that does not mark its copies, and a full storage, already behaved as they should.
Twenty more records without a primary source now rest on the document itself: first editions of books, the Nupedia mailing list, the Maven memo, a UN resolution, court rulings, the laws of Korea and India. For nineteen of them the document corrected the record.
The last two candidates from the PWA audit. Both concern a page's address, and both turned out to be real faults. The graph stopped following its address when opened with a broken fragment. And the sources page lost the chosen order when a reader switched language.
The atlas became a PWA: the axis and saved records open offline, and the site itself can be installed. The same day a month of days fitted across the axis, the showcase began to show only records with a primary source, 349 people got Ukrainian names, the dark scheme got deeper colours, and the alpha mark counts the records a person has confirmed.
Fifteen records without a primary source now rest on the document itself: Turing’s typescript, Strachey’s letter, press releases from IBM, NVIDIA, Sony and iRobot, the Pentagon’s directive. For seven of them the document said something other than the record.
On /graph/ 848 records and 1,152 relations stand as one three-dimensional graph, drawn by a renderer of its own at 7 KB instead of a library at 200. The lines of the axis are explained beside it, on the about page and in the tour, and on the way it turned out that the dashed arc had been drawn solid all along.
The local editor nobody used went, and half the technical debt went with it. Checks that stopped nothing began to stop things, the evidence archive became history, the research pipeline moved into git, and the axis takes its records from the data rather than from a list kept by hand.
Search opens from any page and answers under a Ukrainian layout, "?" lists every key, a first visit to the axis gets a short tour, and the bibliography is grouped by site instead of three sections by kind of source.
In three days the atlas grew from 657 records to 848: the methods the 2020s stand on, Ukraine, machine translation and retrieval, recommendation, safety, law, medicine, vision, sound, games, the 1980s outside the laboratory and language in the 1990s. The commonest error of these passes was my own: numbers taken from later versions of preprints.
The front page opens on the collection itself, a first decade one year wide folded into the next, the people, organization and model indexes became lettered cards, and the axis that drew itself upright on every visit before turning now arrives the right way up.
The frozen search contract was unfrozen, lost its filters, gained a Ukrainian stemmer and came back as a layer over the lane — for 2,426 bytes of JavaScript.
A record began showing the relations written about it and drawing its neighbours; an empty tags field became a closed vocabulary of twenty lines that takes "foundations" apart; ten duplicate entities merged, and the last record off the axis went onto it.
In three days the atlas grew from 485 records to 657 on answers from outside research reports. Their candidates were good and their dates and numbers were not: about one false claim a record, and every fifth date wrong. A report now carries names only, and the session checks everything.
The hours after search shipped: matches open as a list, closing the field gives back the view it borrowed, a query waits for three characters, records can be kept with a star, and the toolbar thresholds were re-measured against a font wider than ours.
The atlas was full of things that claimed something and opened nothing: figures with no lists behind them, buttons with broken anchors, six hundred titles in one run, and a script that died rather than say what was wrong.
The atlas dates and sources other people's work. Until today it said nothing about how it was built itself, and a methodology page describes rules rather than a road.
The running year grew until no straight scale could show it beside the 1950s. Fixing that meant admitting the gap will come back every January.
The badge that said sources were checked stopped saying it, the last empty year was filled, and the front page stopped carrying a dataset it does not draw.
One day took the atlas from fifty-three records to two hundred and ninety-one, and took the axis from a picture of data to something a reader can operate.
Forty minutes after the first commit: a 90 MB Node binary removed from the dependency tree, and the search code nobody renders moved somewhere it can be kept honestly.
The repository opens with a commit that adds an entire working site at once. Six days of work already stood behind it, and for most of those days there was nothing to recover it from.