Officially corrected errors persist in AI training data: Céphéide public alert, data and proofs

Version 1.2 (31 October 2026). Four folders added, measurements made on 30 October 2026 in the continuation of step S1: does OLMo 2, an open model whose training data are public, say that a paper is retracted when its data said so? Result: in what the model read about 291 retracted papers it had seen often, 7.7 % of the passages mention the retraction; the model said "retracted" for none of the 891 papers asked (S4 and S5) and "not retracted" for 425, but it does not know these papers (correct journal: 2 / 891), so no conclusion is drawn on the model. Preregistered, sealed and anchored in Bitcoin before the first question. See README — v1.2.md. The files of version 1.1 and the article arXiv:2610.34379 are unchanged.

Version 1.2 (31 octobre 2026). Quatre dossiers ajoutés, des mesures faites le 30 octobre 2026 dans la suite de l'étape S1 : OLMo 2, un modèle ouvert dont les données d'entraînement sont publiques, dit-il qu'un article est rétracté quand ses données le disaient ? Résultat : dans ce que le modèle a lu sur 291 articles rétractés qu'il avait vus souvent, 7,7 % des passages mentionnent la rétractation ; le modèle n'a dit « rétracté » pour aucun des 891 articles interrogés (S4 et S5) et « pas rétracté » pour 425, mais il ne connaît pas ces articles (revue juste : 2 / 891) : on ne conclut pas sur le modèle. Préenregistré, scellé et ancré dans Bitcoin avant la première question. Voir LISEZ-MOI — v1.2.md. Les fichiers de la version 1.1 et l'article arXiv:2610.34379 sont inchangés.

Public alert by Céphéide, Loxyn's measurement bench, with all data, scripts, readings and proofs. French and English texts of the alert are in ALERTE/; the step-by-step guide is README.md (English) and 00 — LISEZ-MOI.md (French). Article: arXiv:2610.34379. The full package (every step folder, lists, proofs) is in the zip file; MANIFESTE.sha256 gives the SHA-256 of each file it contains.

Official texts are corrected after publication: EU corrigenda in the Official Journal, verified RFC errata, retractions of scientific papers. We measured whether these corrections are present in public datasets used to train language models (EurlexResources, FineWeb-2, a FineWeb sample, peS2o) and what models answer about the corrected passages. Main results: 95.5% of testable single-language corrections to EU acts (1,989/2,082) remain in their erroneous form in EurlexResources; at least 124 corrected EU-law errors and 60 verified RFC errata are reproduced in web training data; about 12,200 retracted papers are in peS2o, 79% with no mention of the retraction; 53% of a random sample of retracted papers are still cited a year or more after retraction (OpenAlex); asked without documents, five models give the erroneous version in 45% of the answers in which they commit to one version, and one answer in 295 says it does not know. Shown both versions side by side, the models choose the corrected one in 53% of cases, and naming the act makes no difference; when each version carries its date and the corrigendum label, the four larger models choose it in 99–100% of cases.

Every step was preregistered and its fingerprints timestamped with OpenTimestamps in the Bitcoin blockchain. Failed controls, lost bets and deviations are reported. The Retraction Watch file is not redistributed (its SHA-256 is in the seal of the retraction step).

Version 1.1 (27 October 2026): adds README.md, the English translation of 00 — LISEZ-MOI.md, and the link to the article on arXiv. No other file has changed.

Version 1.2 (31 October 2026): adds steps S4 and S5 (what an open model read about retracted papers, and what it says); see README — v1.2.md.

Version 1.3 (10 November 2026): adds step 6, a preregistered measurement of whether web search fixes 59 serious errors in EU law corrected by corrigendum (OpenAI, Gemini and Mistral through their APIs with web search; ChatGPT, Gemini and Le Chat by hand); see README — v1.3.md. Files of earlier versions are unchanged, except for a pointer line at the top of 00 — LISEZ-MOI.md and README.md.