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

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 00 — LISEZ-MOI.md. 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).