How this site works
The Institute's site is static. Texts are stored as files, pages are assembled by a script, and the server has neither a database nor an administration panel. This page is a short account of what the site is built from, how an edit reaches the reader, where the numbers in the texts come from, and what the site knows about you.
The site is assembled, not administered
Every page is a markup file. A Python script assembles it from text sources: it inserts the header and footer, the table of contents, file versions and search-engine markup. The server computes nothing per request. nginx serves a ready file, and a reader's request launches neither a database nor server-side code.
The build is not there for convenience. It keeps the site in a state that can be checked:
- every page stays a separate file — it can be opened and read without tools;
- the version of styles, fonts and images in the address is derived from the file contents, so a browser cannot show yesterday's file under today's address;
- before deployment the build is checked: tag balance, internal links and anchors, language versions, the site map, structured data;
- the expandable blocks are native browser elements and work without scripts.
How a page reaches the site
The site is built on a dedicated server and only from sources. Edits arrive from remote workstations over a secure connection, and each becomes a record in version control — with an author, a date and a reason. Files in the live site directory are never edited by hand: not through a panel, not over ssh, not as “just a quick fix”.
The sequence is the same for any change:
- edit the source — text, styles or data;
- build and run the checks — the same ones the server will run;
- deploy: the built files are synchronised with the site directory;
- verify the result live — page responses, security headers, the site map.
That is why every page shows an update date. It comes from the file history rather than being typed in. If the date has not changed, none of the files the page is built from has changed either.
Where the data comes from
Numbers in the texts are not typed by hand — the build computes them from data files. The incident chronicle above all:
- the international chronicle is built from a snapshot of the AI Incident Database dated 5 October 2026: the snapshot holds 1713 records, and 99 passed our selection rules into the chronicle;
- the Russian one is our own collection from court decisions, regulator orders and reinstatement rulings: 29 records;
- both sets are published as tables that can be downloaded and re-aggregated into any selection.
The selection rules and the significance scale are on the chronicle page, and the download and licence on the datasets page. The counts of high-significance records and of the two most recent years are produced by the same script: add a record and the numbers change by themselves.
Licences
Research texts, the glossary, methods and research protocols are under CC BY 4.0: they may be reprinted and quoted with attribution. Datasets open under the same licence 12 months after the article they belong to; before that, on request.
Closed: price lists, client reports, code, the name and the mark of the Institute. The site's fonts are free, under SIL OFL: Golos Text, Inter, IBM Plex Mono and Noto Sans SC.
Citation rules and the limits of the licence are on the “Licence and open data” page.
Illustrations and generative AI
Some covers and in-text images are generated by models from our prompts: covers in GigaChat, individual frames there and in Yandex ART. The subject is always concrete and allegorical: an hourglass for a deadline, scales for a price, a vault door for protection. There are no people and no lettering in the frames.
Captions describe the subject and do not mark an image as generated. That is the site owner's decision, and not an oversight: the picture serves the text rather than posing as a document. Diagrams, charts and the Institute's mark are vector files drawn for the site.
Every illustration is published in two widths — 950 and 1900 pixels — and, as a rule, in two versions, light and dark. The dark one is not a filter: it is a separate generation in the same cool palette.
Accessibility and privacy
We check accessibility by what we can verify ourselves. There is a “skip to content” link, the theme can be switched and is remembered, and a panel for low-vision readers offers three font sizes, four colour schemes, images off and exit by Esc. The panel follows GOST R 52872-2019, but no conformity certification has been carried out — and the site does not claim one.
There are no counters, analytics or third-party scripts on the site: the security policy allows loading from our own address only. The browser stores the choice of theme, language and panel — that is localStorage, not a tracking cookie. The forms are demonstrational: data does not go anywhere.
What exactly has been checked and what has not is described on the “Accessibility” page.