Clinical NLP is usually developed where notes, labels, infrastructure, and governance are locally controlled. Multi-institutional oncology research changes the problem: clinical notes may not move, annotations cannot always be pooled, and differences in documentation, semantics, populations, and technical environments can undermine portability. Drawing on multicenter federated learning, the FLAIMME consortium, and recent work calling for continuous validation of LLM-derived oncology data, this talk examines how to choose among centralized, federated, and hybrid architectures for turning clinical text into research evidence. It asks what should move across sites—data, models, prompts, extracted concepts, or evaluation artifacts—and what must remain local. The talk extends this question to LLM and agentic workflows, where the unit of validation is not a single model but a versioned, multi-step system. Attendees will gain practical approaches to semantic portability, site-specific and continuous validation, subgroup-aware monitoring, provenance, and governance for systems that must remain reliable as models, documentation practices, and deployment environments change.