Shares instead of counts

Half of this year was press releases

2026 held 38 records, and about half of them were vendor launch posts sourced to the vendor’s own blog. That is not a record of a year; it is a record of what was announced loudly.

A research pass added 55 records across January to September — hardware, independent evaluation, labour, speech, money, science, courts, conferences, safety and open weights — with 63 sources and 43 new entities. The year went from 38 records to 93.

Every source URL was opened in a browser and its status read from the tab’s own network log. After the records were drafted, each source was read a second time and every stated fact checked against the page. Four claims did not survive that pass and were removed or corrected: a prize split, a gazette number, the framing of one company’s retention figures, and a denial attributed to a laboratory that its cited source does not report.

Two lines were left empty on purpose. Claims about general intelligence need a quotable document, and the panel that would have supplied one yields nothing checkable. Money has one round with both a firm date and a number, and nothing else in the period. Both are written up in the batch’s gaps.md, so the emptiness is a decision on the record rather than an oversight.

The front page had a column in it

The showcase drew one mark per record, stacked into its year. That worked until the running year pulled away from the rest: 96 records in 2026 against a median year of 3 made a single column 528 pixels tall standing next to a flat line, inside a block 556 pixels high.

At 32 to 1 no straight scale shows both ends. Either the tall one is cut or the short ones vanish. And the ratio was never going to improve on its own — a year still being recorded will always beat a year reconstructed from archives, so the same column would have come back every January.

So the strip shows what share each topic holds in a decade instead. Shares carry no volume, so nothing there can run away again, and the height is fixed for good. What it says is worth saying on its own: foundations alone at the start, 100 per cent of the 1940s, down to 41 per cent in the 2010s as the rest rise. Straight segments between decades and no smoothing, because a curve would draw values the collection does not hold.

Three topics became five in the same pass. Robotics and multimodal work had been sharing a bin with things they were never about; the driverless records in particular sat under foundations, which is a category for theory.

The machine that was right was not the machine it was written on

Adding two topics made the legend taller, and the caption takes what it needs while the axis lives on what is left. On the continuous integration runner the lane fell to 401 pixels against a floor of 416. On the development machine it measured 420 and passed, because the two platforms wrap the same legend onto a different number of lines.

What went, at phone width only, were the two hints about working the interface: that arrow keys walk the axis, on a device with no arrow keys, and that a period can be tapped, which a finger finds on the first try. What stayed is everything that makes a claim. The lane now measures 460 against the 416 floor, so the margin does not depend on which fonts the runner happens to have.

Names above the decades, on a leash

At the full range the axis gave a reader ten columns of a year span and a count. Each decade now carries five to seven terms above it.

A strip like that advertises coverage, which is the one thing the front page promises the atlas will not do loosely. So a term is not free text: each carries the id of the record it stands for, the id is never rendered, and the data gate refuses a term whose record does not publish or belongs to a different decade. There are tests for both refusals. The collection had real holes — on the day this was written the 2020s robotics line held no Optimus, no Figure and no RT-1 — and without the leash a decorative strip would happily have named one.

The candidates came out of the data, ranked by how many other records link back to each one. A record that many later records point at is epochal by the collection’s own accounting, and that ranking put McCulloch and Pitts, Dartmouth, ELIZA, back-propagation, AlexNet and the Transformer at the top of their decades without being asked to. The final pick and the wording are editorial, and say so.