GxP-Agent records 100% structural match on CDISCPilot01 ADSL using Claude Sonnet 4.6

The system decomposes dataset generation into domain-specific worker agents that execute under validation gates and conditional retry. Single-shot frontier models produced zero valid subject-level datasets on the same task. Retrieval-augmented baselines reached at most 59.2% structural match. The DAG topology alone lifted GPT-4.1 from 0% to 59.2% mean match without changes to the underlying model.

CDISC-Bench is derived directly from the FDA pilot submission CDISCPilot01. The same 9-node branching DAG applied to ADAE yielded 100% structural match on first attempt for 55 variables and 1,191 records. Process topology therefore substitutes for model scale on tasks where regulatory ordering is fixed and verifiable.

Prior LLM code-generation work in clinical programming relied on prompt engineering or retrieval alone and reported persistent schema violations. GxP-Agent demonstrates that explicit encoding of GxP process constraints as graph edges produces deterministic compliance where unconstrained agent reasoning fails. This pattern aligns with verified results in other regulated domains that use DAG-enforced workflows.

Operational deployment will require mapping additional CDISC domains to equivalent DAGs and integrating with existing statistical computing environments under 21 CFR Part 11 controls.