About the project

Vinum ex Machina is an atlas of artificial intelligence in wine, in English and Russian, covering the whole value chain: the vineyard, the cellar, and the business that connects a producer to a drinker. It exists because the subject is written about often, enthusiastically, and almost without figures — and because the figures turn out to be obtainable if somebody is willing to open the source behind every claim.

Two things on this site were measured rather than described, and they are what the rest of it is built around.

The Casebook is 299 audited cases of AI in wine. Every record was reopened against its original source, 287 of them are AI proper, and each carries a grade for how well its own source supports it — a peer-reviewed paper and a vendor’s press release are both admissible and are not the same evidence. Eight records were deleted outright during the audit because the source did not survive a second reading. The register filters all 299 along five axes and prints every one of them in the served page, so the corpus is readable, searchable and printable without running any of our code.

OenoBench is a wine-knowledge benchmark for large language models — 3,266 multiple-choice questions across six domains and four difficulty tiers, with 16 model configurations scored against it and every release published as a versioned JSON file. It answers a narrower question than the Casebook and answers it repeatably: not “is AI useful in wine” but “how much does this model actually know”.

The keynote is the argument the two datasets support, and Research is where the atlas’s own studies will go once there are any. There are none yet, and that page says so rather than filling the space with forecasts.

The rule the whole site is built on

A number that is typed is a number that dates. Every count in the prose here — including the ones in the paragraphs above — is an expression evaluated when the site is built, reading the same data files the pages themselves are drawn from. Change a record and the sentence changes with it.

That is not a preference. The research catalogue this project started from carries six summary tables, three of which sum to 295 rather than 299; they were written before the third wave of research and never recomputed, and its headline finding rests on a denominator that no longer exists. The keynote article was drafted against those tables and about ten of its figures had to be reconciled before it could be published beside the corpus it cites. Nothing announced any of it. Numbers are computed here because that is the only way we found of noticing.

The same instinct runs through the rest: the leaderboard’s figures are generated from the run file rather than frozen as SVG, a difference is called a finding only when its confidence interval excludes zero, and the register’s own division rule is recomputed from whatever a filter leaves rather than drawn once and left to lie.

What this site is not

It is not a directory of vendors, and being in the Casebook is not an endorsement — the corpus contains projects that failed, projects that were publicly admitted to have failed, and claims that did not survive their own sources. It is not a consultancy. It does not run on advertising, it sets no cookies, and its analytics count pageviews on our own domain without storing anything on your device or keeping an identifier that crosses a day boundary.

It is also, deliberately, not finished. Both datasets were compiled by one person reading published sources, which is a real limit and is named as one on the author’s page — the Casebook records what a source will confirm, not what a cellar knows.

How it is made

A static site: Markdown and MDX in Git, built with Astro, served from Cloudflare Pages, with no framework runtime on a prose page. The design is a bound laboratory report — paper stock, ink, hairline rules, drawn measurement rulers, mono labels, tabular numerals — with no rounded corner, no shadow and no colour that emits light. Every page prints, including from dark mode, and the Casebook prints as a register of 299 ruled lines.

Corrections are the most useful message this project receives, and the address is on the author’s page.