Eleven attempts, five frontier models, zero valid datasets. That is the starting point of a study that measures a systematic failure: turning clinical trial protocols into analysis-ready datasets under CDISC standards remains a bottleneck in regulatory submissions, and language models alone do not overcome it. The paper introduces GxP-Agent, a multi-agent system that encodes regulatory process ordering as a directed acyclic graph (DAG) and decomposes dataset generation into fifteen domain-specific nodes, each executed by worker agents with pharmaverse skill context, validation gates, and conditional retry.

The sharpest result comes from CDISC-Bench, a benchmark built from the FDA pilot submission CDISCPilot01, with 254 subjects and 49 ground-truth ADSL variables. With Claude Sonnet 4.6, GxP-Agent achieves 100% structural match across three independent runs: 49 out of 49 variables and 254 correct records. The best retrieval-augmented baseline stops at 59.2%, while all single-agent and flat multi-agent approaches remain at zero. This is not a marginal difference: it is the shift from unusable output to a conforming dataset.

The paper's thesis is that process topology, not model reasoning, is the decisive ingredient. GPT-4.1 proves it: under the same DAG architecture it averages 59.2% structural match, while under every other architecture it scores zero. The same pattern extends to ADAE, the adverse events domain, with a nine-node branching DAG, 55 variables, and 1,191 records: 100% structural match on the first attempt.

For those looking at deployment, the point is not which model is better. The DAG structure shifts control from the skill of a single LLM to the design of the workflow. Validation gates become verifiable checkpoints, conditional retry limits out-of-spec outputs, and decomposition into domain nodes creates an inspectable trace, a key requirement in a GxP environment. It is a structural signal: in regulated domains, the path to reliability runs through explicit process constraints, not larger prompts or wider context.

This has implications for teams handling clinical or sensitive data. An architecture of this kind can run on local or hybrid stacks, keeping data and controls within organizational boundaries; the source does not specify the deployment context, but the pattern is compatible with sovereignty and audit needs. At the same time, horizontal agentic vendors without domain knowledge risk losing relevance in sectors where compliance is not optional.

An open question remains: how much of this success depends on the quality of the FDA pilot data and how much will generalize to more heterogeneous protocols? The paper does not promise a universal solution, but it shifts the debate from model scaling to process engineering. In a field where a wrong dataset can block a submission, that distinction matters more than a benchmark.