Background: Clinical trial statistical programming requires strict adherence to regulatory standards (CDISC), which general-purpose AI agents lack. Existing domain adaptation methods do not encode structured workflow dependencies necessary for layered clinical pipelines. Objective: To evaluate a graph-constrained skill loading framework that injects domain-specific regulatory knowledge into an AI coding agent based on workflow context. Methods: The framework models the 7-layer clinical trial data pipeline as a directed acyclic graph (DAG) with 45 nodes and uses an adaptive priority scheduler to load 59 domain-specific skills within a fixed token budget. We conducted a controlled evaluation comparing the framework against an unaided Baseline across 10 regulatory tasks (n = 30 per condition) using a blinded LLM-as-a-judge (DeepSeek). Results: The framework significantly improved overall output quality by +0.47 points on a 1–5 scale (95% CI: +0.12, +0.80, p = 0.004) with an 80% pairwise win rate. Gains were concentrated in structure (+0.63) and regulatory terminology precision (+0.50). Adding distilled experiential principles yielded no additional improvement over the core skills. Conclusions: Graph-constrained skill loading significantly improves agentic AI output quality in regulated clinical trial programming without requiring model fine-tuning.
Jaime Yan (Thu,) studied this question.