A 5-stage governance model for using generative AI to produce regulated clinical deliverables (TLFs, SDTM/ADaM mappings, submission content) — with a working production system as the proof point, not a slide.
Click a stage to see what it does and how it's implemented in production.
Each stage exists to answer one auditor question: "how do you know this output is trustworthy?"
Simulates a single LLM-generated TLF moving through the full pipeline — this is the sequence that took QC iterations from 8 down to 1–2.
| Principle | Stage | How it's satisfied |
|---|---|---|
| ALCOA+ | 3–5 | Automated QC + audit trail make every output attributable and reconstructable |
| 21 CFR Part 11 | 4 | Named human sign-off before use, logged as an electronic signature event |
| CSV Lifecycle | 1–3 | Constrained generation + deterministic QC = a validatable, bounded system, not an open-ended black box |
| Data Integrity | 1 | Input validation gate prevents garbage-in/garbage-out at the source |