An atlas that opens offline

Offline

The issue was about a PWA only: no store apps and no push notifications. The worth is offline use and fast repeat visits, and installing comes as a bonus. The first risk named was a stale cache, since every push to main goes straight to production.

So pages and data are taken from the network first, and the cache answers only when there is no network. The page then says it is a copy, and when it was saved. Only build files with a hash in their names come straight from the cache. The service worker’s version is a hash of everything the build emitted. A commit that changes no output leaves the worker byte for byte the same, and a new one takes over at once and clears the old one’s caches. A language is cached when the reader opens a page in it, so a first visit to /uk/ does not fetch the English axis. Records saved with the star sit in a cache of their own together with their files, and can be read without a network.

In case something goes wrong there is a switch. One flag turns the worker into one that deletes every cache of the atlas and removes itself. It was rehearsed locally. And /sw.js must never be deleted: a 404 would leave the old worker installed.

Three things were checked on production in Chromium: the axis opens offline after one visit, so does a saved record, and a page never visited answers with the offline page. Not checked on real iPhones and Android phones: the install prompt, the standalone window, the back button inside it. A person has to look at those.

A month across the axis

A day was held to a minimum width of 56 pixels, which the name of a year or a month needs. So a month of days wanted 1,848 pixels, and every narrower screen scrolled sideways. A day is now as narrow as its mark’s cell, 24 pixels, and a month fits from 856, before the 900 at which the axis turns upright.

A narrow period no longer prints the title of its one record as three clipped lines (“Нагоро / ICML / 2026”). A hover tip shows the title whole, and a day with several records lists them all. Single records in the early decades at the full range got a title on hover too. Since the collection reached 1939, the thin first decades draw records of their own rather than clusters, and until now nothing named them.

A showcase that shows primary sources

The four cards on the home page now choose only among records that cite at least one primary source. Along with that, 72 sources were reclassified from archive to primary, but only where the title and the note show the source is the document itself. That is a scan of a paper or book with the edition’s year stated, a full transcription, or a Wayback copy of the issuer’s own page. Records without a primary source went down from 123 to 64 of 848. Sixteen sources stayed archive: digitised books whose edition the note does not name, museum catalogues, translations.

Ukrainian names

349 people had a Latin name in the Ukrainian index, a copy of the English one. Each now has a Ukrainian name: the established form where there is one (Алан Тюрінг, Джеффрі Гінтон, Ноам Хомський, Ян Лекун), otherwise a transliteration from the language of origin. There are no Russian forms: people of Russian origin were given the Ukrainian counterparts of their first names, as Сергій Брін already had. The Latin spelling became an alias, so both the atlas’s search and the browser’s find still reach them. After that, eight records whose texts named people differently from the index («Серль», «Соломонов», «ЛеКуна») were rewritten, and each had its version raised.

Small things

The topic colours of the dark scheme became deeper and more saturated: the same hues, with a contrast of at least 4.7:1. Instead of “human review not finished”, the alpha mark under the axis now counts: “Alpha: 2 of 848 confirmed by a person”.