Research object
Verifiable Natural Language Generation
The division devotes its research program to a single system: a generator of factual texts whose properties are verified from outside rather than declared, applied equally across 1,000 languages. This page sets out what the system guarantees, what it measures, and the limits the division acknowledges.
Measurements of , regenerable on the current state of the system.
What the system produces
The system under study produces short factual documents in each of the 1,000 languages it treats as equals: a language is a language, and none is served by a lesser procedure. What varies from one language to another is the material available, a difference the division measures and publishes rather than smoothing over.
These documents draw on 3,444 structured fact sheets spread across eight domains: from historical figures (1,070 sheets) to celestial objects (4 sheets), by way of cities and places (1,003), living species (504), events (445), countries (197), substances (156), and inventions (65). Each assertion a document carries remains attached to a public source identifier – a Wikidata identifier – that the reader can consult directly.
At generation time, the system uses no language model and no network access: its documents follow from structured facts constituted and frozen beforehand, through an entirely deterministic procedure. Identical input yields identical output, byte for byte.
Three guarantees verified mechanically
Three properties carry the burden of proof on this page. They do not read as a promise but as a dated finding, obtained through a check that could have failed and was put to that possibility.
Property P1
Deterministic, reproducible generation
Identical input yields output identical down to the byte: two runs return the same texts.
Status: established on the measurement of 28 August 2026 by comparing fingerprints across two separate processes: 147 repeated generations, 147 identical SHA-256 fingerprints, 0 deviations.
Conservation of the facts is held to the same standard: nothing is asserted here that a check could not contradict.
Property P3
Mechanical conservation of facts
Every document produced is re-verified, by machine, against the set of facts that produced it: nothing invented, nothing lost.
Status: verified mechanically on every document produced; a suite of more than 200 automated checks accompanies the system’s development, negative controls included.
What a mechanical check establishes for a single document, an independent verifier then establishes between the languages themselves.
Property P4
Factual equivalence between languages
Versions of the same subject in different languages are verified to carry exactly the same facts: saying less is permitted, saying something else is not.
Status: established from outside on the measurement of 28 August 2026 by a verifier independent of the system: 935 languages re-read, 415 values compared, 0 conflicts; a value deliberately corrupted in one output was duly detected.
Readability, age, and a bounded tutor
The system measures the reading difficulty of its own outputs and can produce, at constant facts, a variant suited to young readers: the figures it carries stay exact, never rounded for the occasion. On the French example kept for a country fact sheet, the measured document runs to 274 words, at a reading level assessed at grade 12.
The system includes a bounded tutor: its question-and-answer component answers only from the fact sheet of the subject asked about, and says it does not know yet outside that perimeter. The exercise correction it offers follows the same restraint: it returns a verdict, together with its justification, rather than a score without reason.
What the sweep of 28 August 2026 measures
A sweep taken on 28 August 2026 measures, for one test subject per domain, how many languages out of 1,000 produce a complete document. Where a language or a fact sheet does not allow a statement to be made correctly, the system stays silent on that fact rather than approximating it: this is what the division calls an honest omission. These omissions are counted and published, never hidden, because a coverage figure that concealed them would measure nothing.
| Test subject | Languages producing a document |
|---|---|
| Country (Kenya) | 935 / 1,000 |
| Living species (lion) | 932 / 1,000 |
| Substance | 924 / 1,000 |
| Historical figure | 132 / 1,000 |
For the country subject (Kenya), 65 honest omissions accompany the 935 languages that produce a document, whose median length runs to 77 words; for the species (lion), that median length is 8 words; for the substance, 7 words. For the historical figure, the least covered subject, 868 honest omissions accompany the 132 languages that produce a document.
To confirm that this result does not hinge on a favourable choice of subject, five subjects per domain were taken in lexicographic order – a selection no one can steer – and the same sweep was repeated on each. The average number of languages producing a document runs to 924 for substances, 908 for living species, 907 for countries, 900 for events, and 897 for cities and places; it runs to 273 for historical figures, ranging from 109 to 902 depending on how dense each individual sheet is. The limit lies in the data available, not in the procedure, and the division states it rather than smoothing it into an average.
Named limits
The least covered domain to date is historical figures: 132 languages out of 1,000 produce a document for the single subject of the sweep, 273 on average across five. A page that concealed this finding would contradict the very approach it describes.
Outside the country domain, the documents produced stay short: median lengths of 8 and 7 words, for the species and substance subjects, describe a sentence rather than a lesson. The system says what the material lets it say, and stops there rather than lengthening the text through approximation.
These measurements were taken on the state of the system as of 28 August 2026: regenerated at each evolution of it, they describe not a permanent property but what a reproducible sweep found on that date. The division’s formal results, for their part, remain in preparation for publication.
Scientific position
A difference of nature, not of degree
The division holds that free educational content that is wrong costs more than content never read. Often right is not guaranteed, and the difference between the two is one of nature, not of degree. The program studies systems that guarantee the conservation of facts by construction, with mechanical verification, over a delimited factual domain, rather than systems that merely make it likely. The scientific context for this position, and the results still in preparation for publication, are set out on the Research Program and Validation pages.
Continue reading
The detail behind these measurements, checks, and outputs is set out on the following pages.
- Results and Measurements – The dated measurements that support each property presented on this page.
- Validation Approach – The automated checks, negative controls, and independent references behind them.
- Demonstrations – Unretouched, dated outputs in several languages.
- Research Program – The division’s research axes and scientific position.