Beyond AI Transparency: A Reproducibility-Centred Governance Framework for Adaptive Generative AI Across the Medical Research Lifecycle

Background: Generative artificial intelligence (GenAI) is rapidly transforming health informatics through applications in evidence synthesis, clinical decision support, digital biomarkers, and healthcare research. However, the adaptive nature of foundation models introduces reproducibility challenges that extend beyond the scope of conventional AI governance. This review examines reproducibility as a foundational requirement for trustworthy healthcare AI. Methods: A critical narrative review was conducted following the Scale for the Assessment of Narrative Review Articles (SANRA). Literature published between January 2018 and June 2026 was identified through PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore, Google Scholar, and citation tracking. Evidence was synthesised using an abductive interpretive approach to identify methodological gaps and develop a conceptual governance framework. Results: The review makes three principal contributions. First, it proposes a three-domain framework comprising computational, scientific, and clinical reproducibility. Second, it introduces the transparency–reproducibility gap, demonstrating why transparent reporting alone cannot ensure reproducibility as adaptive AI systems evolve. Third, it presents the Transparency–Observability–Assurance (TOA) Framework, a reproducibility-centred governance model integrating transparency, observability, and assurance throughout the AI lifecycle. Across multiple healthcare applications, reproducibility challenges were shown to affect scientific evidence generation, clinical decision-making, and digital measurement. Conclusions: Reproducibility should be recognised as a dynamic lifecycle property requiring continuous evaluation beyond initial validation. The proposed TOA Framework provides a conceptual foundation for reproducibility-centred governance of adaptive healthcare AI.