Colophon

How this site is made

Stack

Next.js (App Router) and React, written in plain JavaScript. Content and application data live in Postgres on Supabase. The site deploys on Vercel; a few services, like the Ukrainian lemmatizer, run on their own hosts.

Much of the site is content-as-code: research datasets, reading lists, and the homepage itself are JavaScript and JSON files in the repository, checked by tests that re-derive published figures in CI.

Where AI is used

I write much of the code with AI coding agents, which I direct and review. Tests, CI gates, and a written set of red lines constrain what they can ship.

Some features call language models while you use them: the site assistant at /chat and the cited answers on /rules. The benchmark and inference pages publish measurements of the models themselves.

The homepage films are narrated by an ElevenLabs AI voice over AI-generated music. Their photographic scenes are AI-generated, and every page they show carries its own address.

Research pieces cite their sources, and several publish their data as hashed releases so the numbers can be checked.

Type

The homepage sets its headline in Archivo Black and Fraunces. Text and labels use the look's own faces: Roboto and Roboto Mono under Material, Geist and Geist Mono under Glass. Every face is self-hosted.

Licenses

Pages and datasets carry their own licenses. Most published datasets are CC BY 4.0; a few that build on non-commercial sources are CC BY-NC-SA 4.0, and each release names which.

The 3D room

The homepage used to open in front of this room. It is the same picture the front-door film is about: old records arrive uneven, a gate fixes what a rule can, and two wait for a person. Your device draws it live. The parts that never move are lit once in Blender and baked into a model the room loads after it starts.

Nothing loads until you press. Then the room downloads once and draws on your device.