Mixed character encodings across quarterly files (FAERS)

Found by Loxyn SAS, Lyon — September 2026 — PubMed substrate and other corpora (Voûte) · first seen 2026-09-09

Wrong valuefrom: Loxyn SAS, LyonUnclassifiedlayer: ② readinginjection: NOISEproposed

Damage in the graph

Wrong value — a property, number or date is wrong.

Can code cancel it?

Yes. Refusal clause: detect the encoding per file and convert to UTF-8 before reading; a file that fails detection is rejected, not guessed.

Can a deletion-only repair restore the truth? classified, not measured

no — deletion cannot restore this; classified by judgment, not measured — the EUR-Lex loop will measure it.

Which truth can judge it

an official registry.

Where it was seen

FDA FAERS quarterly files 2013Q1-2026Q2, probe of 2026-09-09, ASCII quarterly files · 2026-09-09 · Claude Voûte (LLM agent), session 8d9c5098, from the night reports of 2026-09-06/09 · Loxyn SAS, Lyon

2026-09-09, 01:14: "Invalid Input Error: File is not latin-1 encoded" on one quarter after others were read as latin-1; the files mix encodings by quarter.

Original note (fr)

09-09-2026, 01 h 14 : « Invalid Input Error: File is not latin-1 encoded » sur un trimestre après lecture réussie d'autres en latin-1 ; les fichiers mêlent les encodages selon le trimestre.

Specimens

Specimens not yet transcribed into this record.

History

2026-09-09 proposed by the Voûte Claude in 'FORMES-PUBMED-pour-Fault-Atlas (09-09-2026)' part 2 (sha256 51a338add43cd071…), form 38 of 40; effect on a graph: Drug and effect names altered or lost depending on the reader (MISSING: rejected lines). Prevalence: seen, not counted. — Claude Voûte (LLM agent), session 8d9c5098, from the night reports of 2026-09-06/09
2026-09-09 converted to a record; class, prevention and the novel/variant question left to the second reader — Fault Atlas

Record: atlas.json · source file on GitHub · this record as JSON · row in the data explorer