Format: the long read

Registered 09-09-2026, from the page OpenAI published on the Navier-Stokes problem, which Frank sent with the verdict: "a very good format for other scientific or economic or institutional documents, with a very neat, nice index on the pane. All the proportions are good."

The idea in one line: one serious result, read top to bottom by someone who did not ask for it, with a quiet index in the left margin that says where you are and lets you jump, and two controls that cost nothing and are always in the same place: listen, and share.

Use it for: a research finding, an economic note, an institutional position, a study, an evaluation, anything that has to be read in order by a reader outside the field and still be quotable by section.

Not for: a decision that needs an answer (that is the proposal note), a teaching page with exercises (/ui-doc), a per-item result board (the simple table), a citizen leaflet (the country brands own those).

The anatomy, in order

1. A thin masthead that says whose page this is, and nothing else. It stays at the top when you scroll.
2. The head, centred, wider than the text column: date and one or two category words above, then the title, at three and a half times the body size on one or two lines. The title names the finding; it does not tease it.
3. The actions, under the title: at most two, the first solid, pointing at the primary artifact (the paper, the data, the law, the service).
4. The utility strip, ruled above and below, holding the listen control on the left and share on the right. When there is no recording, the left side says the reading time instead. Never show a listen control that plays nothing.
5. The index in the left margin, one word or three per section, the current section in black and the others grey. It sticks under the masthead and never becomes a box, a card or a panel.
6. The article, one column, sections opened by a title at 1.7 times the body, paragraphs at the measured width, figures full column width with a caption saying what to see.
7. The end line: publication date, author, and the note that sources are linked in place.

The numbers, measured on the source 09-09-2026

Read off the served page at a 1280 pixel viewport, not estimated.

One family, two weights. Much of why the source reads as calm is that it never changes typeface: size, weight and colour carry the whole hierarchy. The source uses its own commissioned face; the template uses Figtree, a free grotesque with the same tight tracking. Swap it for a project's own face by changing --font alone.

The token auditor will flag this page. Its type sizes and widths come from the measurement above, not from the house scale, and that is the point of a copied format. Twelve flags, all of them measured values, checked 09-09-2026.

What the template already does

template.html is the page to copy. The index builds itself from the article's own h2 elements, so the two can never disagree, and it follows the reading position. Share copies the address and says so. Listen appears only when <article data-audio="..."> names a real recording, otherwise the strip shows the reading time; a recording is made with the studio's own voice route.

The source

openai.com/index/navier-stokes-solution, captured 09-09-2026 in inspiration/liked/openai-long-read. The captures are the reference for any change to this format: measure again before changing a number here.