How we
classify knowledge.
Two axes that never merge: the phylum says how a piece of content investigates, the domain says what it is about. The first comes from a four-question sieve, the second from an international standard of 3 422 codes.
The sieve
Four questions, asked in order
A phylum does not say what a piece of content is about: it says how it investigates. The first positive answer fixes the phylum; when in doubt, the decision follows whichever occupies most of the speaking time or the reasoning.
1 · Formal
Formal, computable objects and inert matter
Proof, computation, reproducible physical experiment. This is the first question asked.
Domains
- Mathematics
- Physics · Chemistry
- Astronomy & Cosmology
- AI — Models & Research
2 · Living
Living or natural systems
Organisms, ecosystems, the body, the brain, the Earth — studied through observation and experiment.
Domains
- Biology
- Medicine & Health
- Neuroscience
- Climate & Ecology
3 · Applied
To design, build, size, optimise
The dominant purpose is to dimension an artefact or a technical system, not to study it.
Domains
- Engineering & Technology
- Energy
- Control & Robotics
- Computing & Cybersecurity
4 · Human
Societies, meaning, values, ideas
By default, when none of the three preceding questions has drawn a positive answer.
Domains
- Economics & Finance
- History · Geopolitics & Defence
- Philosophy & Ethics
- Society & Culture
The phylum is distinct from the domain: a video on the history of mathematics investigates as a historian would — phylum human — on a subject in the Mathematics domain. The two axes combine instead of collapsing into one.
The standard
The domain does not come from us
Every tuit first receives a Thema code — the international subject classification for publishing, maintained by EDItEUR and used by publishers, booksellers and libraries in some thirty languages. We use version 1.6, released on 31 October 2024.
Base domains
Few enough to be taken in at a glance, broad enough that each carries meaning. The normal way into the catalogue.
recomputed at every publicationExtended Thema
Codes populated by ten tuit or more. Fine enough to separate cardiothoracic surgery from transplant surgery.
threshold of 10, unchanged since the startGranular Thema
Every code actually populated, down to the one holding a single tuit. The finest view the site serves.
grows with the corpusArtificial intelligence
A brief that forbids more than it asks
Every tuit is produced by one of the most capable language models available, guided by a written, versioned, tested brief — never by an improvised prompt. It is built to obtain a systemic and non-partisan reading, and most of its work consists of closing doors.
Nothing is invented
The model cites only what the video or its description actually mention. The addresses it proposes are then checked one by one with a real request; those that do not answer are removed. A cited source is never presented as a verified one.
Four scores, not a verdict
Quantity of information, quality, technical level and reliability are assessed separately. Content can be highly reliable yet thin, or dense yet fragile — a single figure would erase the difference.
Contradiction is required
Discordant sources are demanded on the same footing as concordant ones. The critique is split into two views that must not overlap: value and argument on one side, rigour and sources on the other.
What does not count
The intended audience level is excluded from the assessment. A sponsored segment is flagged without naming the brand, and never lowers the score. A misleading title weighs at most half a point out of five.
Nothing is assumed: with no public comments available, the field stays empty rather than filled with a sentence. No length is prescribed — measured on the corpus, the length of a text follows the number of fields requested, not the instruction. And Thema classification is a separate second call, with the entire standard submitted: the model chooses among the 3 422 codes; it does not invent a category.
What the method does not guarantee
A tuit is produced by two independent calls, one per language. Nothing forces the model to score identically in French and in English, and we measured the gap across the whole corpus rather than assuming it.
We say so because it is true, and because it clarifies what the tool is: an assessment by a language model is a judgement, not a physical measurement. Two attentive readings of the same content may legitimately differ by a point, as two human reviewers would.
What, by contrast, never diverges: the phylum and the Thema code. Both are assigned by a single call per video, never per language — so the classification is rigorously identical in both versions of a tuit.
The line we hold
Our editorial choices
A low score is still a score
No quality filter keeps content out of the catalogue. A severe tuit is published like any other: it is for the reader to decide what deserves their time, and a low score helps as much as a high one.
Entertainment is not excluded
Literature — science fiction included —, video games and the arts are serious subjects, and they are treated as such.
Register can be the subject
Content with no identifiable main subject is classified by its register: a rambling, humorous livestream belongs to humour, not to the theme its title announces. This explains some surprising placements.
We recommend nothing
No ranking by popularity, no featured picks. The catalogue is ordered by date, and its filters are those of knowledge, not those of audience.
The corpus
Closed collections
Some followed channels have stopped publishing. Their tuit remain online: they form a completed corpus, which will grow no further but loses none of its value. At the latest count, 20 channels out of 801 had published nothing for over a year, amounting to nearly 3 000 tuit.
We neither remove them nor mark them as stale. A 2015 course in mechanics has not aged because its author fell silent. A dormant channel is not a dead channel: it stays subscribed, and a new publication would wake it with no intervention.
Conversely, a video whose transcript is durably unobtainable — geo-blocking, no subtitles, restricted content — is closed rather than retried indefinitely, and remains reopenable should conditions change.
What comes next
Participatory submissions
The catalogue is currently fed by an editorially curated list of authors. What follows is open to the public: requesting the analysis of a specific video, or proposing an author to follow.
The check precedes the commitment
Valid address, transcript likely obtainable, volume estimated. We never commit to an analysis that cannot be produced.
Four ways to follow an author
The current year and onwards · all available past then onwards · the past alone, with no follow-up · from a chosen year.
In developmentNo personal data
No account, no profile, no reading history. From a request, the system keeps only the reference, the mode and the fingerprint of the payment.
In development