What survives the trip to FHIR? I mapped the output of a Spanish-language clinical LLM pipeline (ICU documentation from Colombia) to R4, validating against a Medplum project — every resource POSTed and read back. The most useful moment was a rejection: the mad-1 invariant correctly refusing a vasopressor dose the source couldn't quantify. Full write-up of what the validation caught and where data gets lost between extraction and FHIR: https://github.com/JesusPantojaP/fhir-spanish-icu-notes The repo includes a reusable fictitious Spanish ICU test case (4 documents) for anyone working on clinical NLP in Spanish. Feedback welcome — especially from anyone thinking about Spanish-language or LATAM deployments.