Design and verification of a FHIR R4-serializable known-truth testbed for technical and inferential assurance in small longitudinal trial simulations

Simulation can expose inferential failure because truth is known, but technical correctness of the surrounding research-data system is seldom audited separately from statistical reliability. We developed a serialization-bounded known-truth testbed with four assurance layers: record/data integrity, FHIR R4 serialization and HAPI-FHIR round-trip fidelity, computational reproducibility, and inferential/decision reliability. Deterministic Python code generated scientific finite masters and potential outcomes; Synthea 4.0.0 served only an engineering record/export feasibility test. Two N = 40 reference trials yielded 1578 FHIR resources; all passed PUT/GET and canonical authored-field comparison, with 2404/2404 local references resolved. The confirmatory simulation comprised 24 scenarios, four strategies, and 5000 replications per cell. Reanalysis of 480,000 locked rows reproduced released cell summaries within 1e−9. Stratification reduced pooled pain and distress imbalance by approximately 38.5%. Under heterogeneous effects, 25% outcome-related MAR versus 10% MCAR increased mean bias by 0.294 points (MCSE 0.0045) and decreased coverage by 0.0285 (MCSE 0.0012); minimum coverage was 0.8524 (MCSE 0.0050). Within this testbed, technical assurance did not imply inferential reliability. The framework supports bounded protocol and research-pipeline stress testing, without claiming clinical-effect estimation or cross-system interoperability.