Who did what

This site is a collaboration between a person and several AI models, and since honesty about that division is half of what the site is about, it should say plainly how this site itself was made — not just the projects it describes.

What I did. I chose the projects and what would be worth saying about each. Every claim of fact, every judgement about what an experiment does and does not prove, and every editorial line — what is honest to say, what would be overclaiming — is mine, and I stand behind it. I read, cut, rewrote, and corrected the prose until it said what I meant; where it still says something, I decided it should. Many of the diagrams and figures are mine — conceived, composed, and made — and the rest I reworked well past the point of adjustment.

What the AI did. The first drafts of most of this prose were written by Claude, working inside Claude Code from my notes, my arguments, and a great deal of back-and-forth. The site's machinery — the Eleventy build, the templates, the CSS and the handful of scripts — was largely written by the same model under my direction, and debugged the ordinary way, by running it until it worked. Some figures started from an AI draft; where they did, I say so on the figure itself.

What that means for trust. A drafting model is fast and tireless and does not know when it is wrong; the checking, the cutting, and the final responsibility are human, because those are the parts a model cannot supply for itself. So: trust the facts and the judgements as far as you would trust me, and no further. Where a specific project leaned on AI in a way worth documenting, its own pages say so in detail — the Raimbaut edition's account is the fullest example.

How this site is built

There is not much machinery here, on purpose. The pages are plain HTML generated once, at build time, by Eleventy from Markdown and Nunjucks templates; there is no framework running in your browser and no server doing anything cleverer than handing over files. The styling is hand-written CSS — a set of custom properties for the light and dark themes and the graph-paper motif — with no CSS framework underneath it.

What JavaScript there is (about a hundred and thirty lines of it, no libraries) is comfort only: the light/dark toggle, the figure lightbox, the floating table of contents, and left/right paging through a piece. Every page reads correctly with scripts switched off — the section anchors and the diagrams are baked in at build time rather than assembled on the fly. The diagrams themselves are SVG, inlined during the build so they can recolour to match the theme.

Most of the care here is invisible, and I have made my peace with that. The same arrow keys that page from one section to the next also step through a figure's images once the lightbox is open; the screenshots of a reading view come in two versions and swap quietly when you change the theme, so a light-mode reader never gets handed a dark-mode screenshot. Almost nobody will ever notice either of these, and that really is fine — it is simply what I do. I find a rabbit hole and I follow it all the way down.

Nothing here watches you. There are no third-party requests, no analytics, and no cookies; the single thing this site stores in your browser is whether you last chose light or dark. The fonts are served from this domain, so even they phone no one. The version and date in the footer are not typed by hand — they are read from the project's git history each time the site is built, which is why this copy reads v0.1.0 · d797f02, built 2026-07-07.

Typography

The type is IBM Plex throughout, which is a long-standing affection of mine rather than a considered shortlist — I have always liked it. It arrives as a family of three that divides the labour cleanly: Plex Serif carries the reading text, Plex Sans does the navigation and the small print, and Plex Mono handles the labels, the figure numbers, code, and anything that wants to look like data. Using one designer's three cuts, rather than three unrelated faces, is what keeps the pages feeling of a piece.

There is a quieter reason it suits this particular project. The first edition described here began life as a 1981 doctoral typescript, hammered out on an IBM Selectric; Plex Mono descends from exactly that lineage of IBM machine lettering, so the monospaced notes and provenance lines carry a faint typewriter accent that is entirely appropriate. The fonts are self-hosted as subset woff2 files (Latin and Latin-Extended, so the Occitan and French set correctly) and loaded with font-display: swap, which is the whole of the performance story.

And one last thing about the marks on the page, since the internet has lately decided that this particular one gives a machine away: yes, I really do use em dashes this much. I have been doing it for more than forty years — long before anything else involved in this site could type a word — and I am not about to stop now to avoid the suspicion.