observations: 22

Where each form was seen: corpus (and its reproducible record when there is one), date, observer, the verbatim note (original language), its English translation, and the exact probe that produced the figures when one was kept.

This data as json

id	form_id	form_name	corpus	corpus_id	document	date	observer	excerpt_en	excerpt_original	lang	probe
22	form-034	Hard negatives for typing	observation 2026-08 (272 Europe PMC JATS articles)	corpus-europe-pmc-jats-2023-2026		2026-08-08	house observer (LLM agent), see report	MEASURED on 2026-08-08 on the 272 XML files: 19/272 (7 %) carry TWO neighbouring-type formulations for the same funding fact — "funded by" AND "supported by" in the same article; 100/272 (37 %) use at least one confusable contractual relation term (collaboration, contract, provider, supplier). The real corpus thus offers the vocabulary of the trap; none of our fabricated corpora sets it.	MESURÉ le 08/08/2026 sur les 272 XML : 19/272 (7 %) portent DEUX formulations de types voisins pour le même fait de financement — « funded by » ET « supported by » dans le même article ; 100/272 (37 %) emploient au moins un terme de relation contractuelle confusable (collaboration, contrat, prestataire, fournisseur). Le corpus réel offre donc le vocabulaire du piège ; aucun de nos corpus fabriqués ne le pose.	fr