One question, answered across every artifact here: can we bring routine biostatistics programming in-house and do it faster, more consistently, and more defensibly — with AI we control? The recommendation is a hybrid stack (Claude for capability + an on-prem model for sovereignty and reproducibility), with humans and validated tools always in the loop.
The ship-today, no-PHI operational layer: self-updating Smartsheet trackers and knowledge bases, deterministic SAS/R monitoring and on-leave coverage, the participant-safety and risk dashboards, and the hybrid/on-device trial-ops wikis.
The hands-on build: the SAS + R TLF pseudocode libraries, the TLF-interpretation drafter, the Protocol→CSR deliverables pipeline, the end-to-end study-lifecycle sandbox, and the training course.