An audit-ready EHR foundation model for order-time medication-safety auditing across health systems

Real-time medication-safety auditing requires current laboratory and medication histories, yet electronic health records (EHRs) may contain stale or missing decisive evidence. We present ClinicalGPT, an audit-ready foundation model using Clinical Trajectory Language to encode time, units, provenance, and observability, and workflow-aligned objectives to handle structured missingness. Across seven health systems, we retrospectively evaluated 847,000 eligible high-risk medication orders against a reference vector frozen before final scoring. The primary endpoint was the reference-standard-defined order-time failure-to-alert rate: eligible tier-i/ii reference-positive orders receiving no actionable alert at a system’s frozen operating point, per 1000 eligible orders. This offline process metric did not measure subsequent clinical actions, adverse drug events, or patient harm. ClinicalGPT with hard constraints achieved 3.2 misses per 1000 orders, versus 9.4 for rules only, 6.1 for the original generic LLM comparator, and 4.8 for the strongest observed LLM comparator (RAG + LLM + rules). Its false-alert rate was 33.9 (95% CI, 31.6–36.3), versus 38.7 for rules only. A separate single-site silent prospective assessment evaluated workflow measures without altering care. Matched ablations supported explicit observability modeling; independent de novo labeling of an enriched knowledge-base–independent subset preserved sensitivity ordering but narrowed the ClinicalGPT–RAG gap. These findings support comparative order-time auditing performance but do not establish clinical effectiveness or reduced patient harm.