This is an official ODERSA website. Here’s how you know

The official domain

The address of this site ends in odersa.org. Every service the association runs sits on a subdomain of odersa.org and nowhere else. If the address in your browser’s address bar does not end in odersa.org, this site is not ours.

Free, and no account

Everything is open straight away. No sign-up, no account, no password, no subscription, no advertising. Nothing is held back for those who pay, because there is nothing to pay for.

No data collected

This site does not follow you: no tracker, no tracking cookie, no measurement tool built into these pages, and nothing measured on your device. Our host counts requests in aggregate, as any server that answers does: a total, never a profile. You do not have to take our word for it: open your browser’s developer tools, go to the Network tab, and reload the page. You will see the full list of what the site asks for. Everything comes from odersa.org, nothing goes anywhere else.

Free to reuse

The content is published under the CC BY 4.0 licence. You may copy it, translate it, print it and pass it on, for your classes as much as for the people around you, on one condition only: credit ODERSA.

ODERSA Research

Validation Approach

What the division holds to be established, by what means, and what it deliberately leaves unclaimed. The page ends with direct answers to the questions an evaluator is entitled to ask.

Measurements of , regenerable on the current state of the system.

1. Nothing is declared, everything is exhibited

The division takes as a working rule that a property nobody can examine is not a property but a claim. Everything on this site is therefore stated in a form that someone else can put to the test: a measurement with its date and its method, an output published as it came, a property expressed so that a counter-example would be recognizable if it existed.

More than 200 automated checks accompany the development of the system. Alongside them stand the negative controls, which the division regards as the more important of the two. A check that passes is compatible with the instrument working and with the instrument being disconnected; only a deliberate corruption, followed by the detection of that corruption, tells the two apart. This is why a verifier that cannot fail is treated here as evidence of nothing.

Greek manuscript page of the Histories of Herodotus, surrounded by handwritten marginal glosses added by the philologist Lorenzo Valla.
Herodotus, Histories, with glosses by Lorenzo Valla (Vat. gr. 122), Vatican Apostolic Library, public domain.

Two properties carry the weight of this approach. The first is the conservation of facts: every document produced is re-verified, by machine, against the set of facts that produced it, so that nothing is invented and nothing is lost. The second is factual equivalence between languages: the versions of one subject in different languages are verified to carry exactly the same facts, saying less being permitted and saying something else not. The measurements that correspond to them, taken on 28 August 2026, are set out on the results page, with their negative control.

Property P1

Deterministic and reproducible generation

Identical input yields identical output, byte for byte: two runs produce the same texts.

Status: measured across process boundaries on 28 August 2026, 147 repeated generations for 147 identical SHA-256 fingerprints.

Structured facts, frozen Deterministic generation Produced document Mechanical verification compared to the input facts equality → published discrepancy → rejected Negative control: an artificially corrupted document MUST be rejected – and it is.

Diagram 1. The verification contract, as stated by properties P1 to P3. The generation process itself is not described: only its contract is – frozen inputs, deterministic output, mechanical re-verification, and a verifier able to fail.

2. The register of properties

The pages of this site refer to the system’s properties by a stable identifier. The status of each – sought, established, verified – is carried by the block that cites it, together with its measurement date.

  • P1. Deterministic and reproducible generation: identical input yields output identical byte for byte.
  • P2. No language model and no network access at generation time.
  • P3. Conservation of facts: every produced document is re-verified by machine against its input facts, nothing added, nothing lost.
  • P4. Factual equivalence between languages: versions of one subject carry exactly the same facts.
  • P5. Honest omission: saying less is permitted, saying something else is not, and the silences are counted.
  • P6. Measured readability: reading difficulty measured, young-reader variant at constant facts.
  • P7. Bounded tutoring: answers from the subject’s fact sheet only, correction verdicts justified.
  • P8. Traceability: every statement in a produced document can be traced to a public source identifier (Wikidata identifiers), which the reader can consult.

3. Independent external references

Internal measurements judged by internal instruments are worth what the instruments are worth. The division therefore confronts them with independent public linguistic references: lexical and comparative databases maintained outside the division, by people it does not employ and cannot influence, and which were not built with this system in view.

The internal use made of these references is not detailed here, and this restraint is deliberate rather than evasive: what the reader needs is not the arrangement, but the property that follows from it. The measurements published by the division are not judged solely by the instrument that produced them, and the standards against which they are held can be consulted by anyone, independently of this site.

4. Scientific position

Recent literature establishes that calibrated probabilistic generators cannot guarantee the absence of hallucination on rare facts[1][2][3], even if hallucination can be made statistically negligible in certain settings[4]. The program explores the other branch of the alternative: systems that guarantee the conservation of facts by construction, with mechanical verification, over a delimited factual domain.

The division states this position prudently and holds to its consequences. It does not claim that its approach is preferable in general, nor that the results cited above are contested; it claims that they leave a branch open, and that this branch is worth exploring for the specific use of free education. Often right is not guaranteed, and the difference between the two is one of nature, not of degree.

The formal results of the division are in preparation for publication. This site publishes no theorem, no proof and no claim of priority: it states properties, measurements, and the program that produced them. That reserve is part of the method, not a delay to be excused.

Often right is not guaranteed.

5. Licence and openness

The content produced by the division is published under the CC BY 4.0 licence: it may be copied, translated, printed and redistributed on the single condition that ODERSA is credited. A public demonstration corpus is in preparation. Publications: in preparation.

6. Direct questions

Why not use a large language model?

Because the intended use is education, and education demands a guarantee rather than a likelihood. Recent literature establishes that a calibrated probabilistic generator cannot guarantee the absence of false statements about rare facts[1][2][3]; it can make them statistically rare[4], which is not the same thing as a guarantee. The program studies the other branch: deterministic systems in which every document is mechanically re-verified against its input facts. Often right is not guaranteed; the difference is one of nature, not of degree.

How can what this site asserts be checked?

Every figure carries its date of measurement and its method, and the measurements come from surveys that can be regenerated in a single command on the current state of the system. The demonstrations are dated, unretouched outputs, imperfections included. The institutional attachment can be checked by reading this site against the official information published by the association. The formal results will be the subject of publications accompanied by their verification material.

What does “honest omission” mean?

When a language or a fact sheet does not allow a fact to be stated correctly, the system says nothing about that fact rather than approximating it. Saying less is permitted; saying something else is not. Omissions are counted and published with the results, and the measurement table has a column for them.

Why do some languages produce so little?

Equality between the 1,000 languages is equality of treatment, not equality of richness: the material available varies by language and by domain, and that variation is measured rather than masked. The least covered domain to date is that of historical figures, where 132 languages out of 1,000 produce a document; it is published on the same terms as the best-covered ones.

Are the texts reviewed by humans before publication?

The demonstration outputs on this site are published exactly as the system produces them, without retouching, awkwardness included. That is a choice of transparency: a hand-corrected selection would prove nothing. The validation of the system itself is mechanical: automated checks, negative controls, and confrontation with independent public linguistic references.

Who works on this program, and with what means?

ODERSA Research is the research division of ODERSA, a nonprofit association under the French law of 1901, with no salaried staff, which publishes its content under the CC BY 4.0 licence, without advertising and without data collection. Research is directed by Mael Garbe. Contact: research@odersa.org.

Where are the publications?

In preparation. The site publishes first what can be checked without taking us at our word: the properties, the dated measurements, the demonstrations. The formal statements and their proofs will follow through the channels of scientific publication.

7. References

  1. A. T. Kalai, S. S. Vempala, “Calibrated Language Models Must Hallucinate”, STOC 2024 (arXiv:2311.14648). back
  2. A. T. Kalai, O. Nachum, S. S. Vempala, E. Zhang, “Why Language Models Hallucinate”, 2025. back
  3. Z. Xu, S. Jain, M. Kankanhalli, “Hallucination is Inevitable: An Innate Limitation of Large Language Models”, 2024 (arXiv:2401.11817). back
  4. Suzuki et al., 2025 (arXiv:2502.12187) – hallucination can be made statistically negligible in certain settings. back

Read next

  • Results and Measurements – The readings this approach frames, threats to validity included.
  • Demonstrations – Unretouched outputs, to re-read and to contradict if the comparison fails.
  • Research notes – The division’s notes, for figures that need more room.

Keyboard shortcuts

TabMove from link to link
EnterOpen the link, or expand and collapse a question
?Open this help
EscClose the panel or this help