What AI actually buys, what it takes to last, how to forecast your own activity, which rules apply to you.

“It saves me time” comes apart into three claims: the work takes fewer hours, those hours are no longer alike, and more is expected of you per day. Six questions answered while reading, and the page works out which one is yours. What bounds the gain sits in the review, not in the speed of the model.

Long-term wealth drawdown, calibrated for France and Japan. Two tax regimes inside one model, which makes visible what the jurisdiction actually changes.

Salary / dividend trade-off, corporate-tax instalments and turnover forecasting, alongside a day-rate barometer running back to 2019: where a rate sits, what has to be billed to live on it, and what is left once the platform and the state are paid. Assumptions shown and editable: a scoping tool, not tax advice.

Thirteen Union texts now govern data and software. Twelve boxes describe a client’s activity, and the page works out which ones apply, why each does, and the calendar that follows. The expensive mistake is not misapplying one, it is missing one until the architecture is frozen.

The mathematics libraries usable from JavaScript, sorted and kept current: algebra, statistics, geometry, plotting. Not a simulator but a starting point — the one you look for before writing a function somebody has already written.

A French preparatory-class project on general relativity and satellite positioning, published whole — report, slides, figures, code, bibliography — with the notebook of exchanges that produced the subject out of a doubt about one figure read in passing. Three simulation pages let a reader redo the measurements instead of believing them.

Sixty-eight baccalauréat papers worked through question by question, in three steps: a hint, the method, then the full answer — each question carrying its prerequisites. An ordinary answer key gives the result; what a stuck pupil lacks is knowing where they came off, so the classic mistakes are quantified rather than flagged.

French secondary-school mathematics proved in Lean 4, from the first year of collège to the last of lycée. Every theorem on the syllabus is written, checked by the machine, and transcribed into French beside it: on the left the proof the kernel accepts, on the right the one a pupil can read. Chapters follow notions rather than year groups, so that divisibility at eleven and arithmetic at eighteen read one after the other. What resists stays on the list with its reason — the area of a disc, the roof theorem, de Moivre–Laplace — and those entries are what measure the limits of the tool.

The proofs are written with Claude Code, in constant exchange with Lean’s language server; what is checked is therefore checked mechanically, down to a build that refuses even warnings. The French transcription is done by hand, following the steps of the formal script. The interesting result is not the list of theorems but the shift in the line between what is assumed and what is established: the expectation of a binomial law, assumed at school, is proved here and was not in Mathlib; the roof theorem, also assumed, is not.

Water levels from the public flood monitoring network, drawn in the page. A trial of one plain question: what it takes for open data to become readable with nothing to install.

Searching arXiv papers by meaning rather than by words, on a database holding vectors and documents in one system. It is the chain described in the 2020 notes, assembled in full two years later.

Team work published by Artefact — Tom Darmon, Henrique Brito Leão, Michel Hua and Corentin Fossé — during employment, not under Commutator.

The Alexandre Grothendieck fonds at the University of Montpellier: 178 folders in open access, 16,074 handwritten leaves from 1949 to 1991, with the full inventory and the transcription set beside the facsimile on one screen. Five notebooks open the fonds by subject rather than by accession number, and each states at its head whether its grouping is the archive’s or ours: citing “the Cahier de Topos” does not commit you to the same object as citing folder 19. Fifty-two batches are transcribed, which is twenty-six folders and 661 pages. These are working notes rather than finished texts — he takes the same chapter up four times, reuses his versos, and the dates are nearly all inferred from a letterhead or a numbered seminar talk. Each batch yields two documents: the transcription, where the illegible stays illegible, and a reading in current notation, the only one allowed to depart from the leaf.

The transcriptions are produced by a model, twenty leaves at a time, each file dating its own pass. It is not a scholarly edition: nothing is checked by a person unless the batch says so, and where an established edition exists it is flagged and takes precedence. The scans remain Montpellier’s, relayed as you read rather than copied. Every batch exports to TEI P5, whose header carries the model that read the pages and the date of the pass, so a file deposited elsewhere keeps its provenance without this site.

The Bibliothèque nationale has digitised whole the manuscripts of a good number of mathematicians, catalogued by fonds and read by almost no one: 78 volumes, 25,258 views, nineteen names from Viète to Libri, among them Fourier’s twenty volumes and Émilie du Châtelet’s five translating the Principia. One rule of selection: if Gallica does not serve the whole volume, it is not in the catalogue, not even as a thumbnail. Sophie Germain’s four volumes, which give the domain its name, are transcribed; the rest are waiting.

First machine passes, as for the Grothendieck fonds whose tooling this site reuses. No image is stored or re-served: the facsimile comes from Gallica’s IIIF service as you look at it, and carries its source line. No third party’s transcription or edition is reproduced or used to seed a reading: they are cited and linked.

The six artist’s ledgers of Edward and Josephine Hopper, held at the Whitney and digitised as 504 sheets running from 1907 to 1967. Once a work was finished Hopper drew it small at the head of a leaf; his wife wrote everything else, the date, the size, the price, the buyer, every exhibition and the day the cheque cleared. The fourth volume is the one to open if you open only one: every payment received, without a gap, from 15 November 1913 to 23 March 1967, its first decade all illustration work that the painting books do not record at all.

First machine passes, four batches so far, forty-eight of the first volume’s 117 sheets. Three numbering systems overlap in these books and only one is an address, so a citation taken here prints both: the leaf, which the paper carries, and the Whitney’s ref, the only thing that names exactly one photograph. The photographs remain the museum’s.

Jean-François Debord’s morphology course, filmed at the Beaux-Arts de Paris and published by PSL, runs to forty-three sessions and close to 74 hours. The index answers one question and no other: where, in those 74 hours, does Debord talk about this? Ninety-five anatomical terms and sixty-three artistic references, each timestamped into the sessions, with the course opening at the right second beside the list. It does not summarise the course, does not reword it and does not host it.

The course belongs to its authors and nothing of it is redistributed. The automatic subtitles serve to build the index and then stay outside the repository, which carries derived data alone: identifiers and timestamps. That is also what lets the site be rebuilt without the transcripts.