<b>Background</b>: Clinical artificial intelligence systems are increasingly deployed in high-stakes healthcare environments, influencing diagnosis, prognosis, treatment selection, utilization management, and population-level decision making. Although advances in machine learning have improved predictive performance in controlled settings, real-world deployments continue to exhibit silent failures, inequitable outcomes, and limited accountability. These failures often persist despite acceptable accuracy, calibration, and fairness metrics, revealing a gap between model-centric development paradigms and operational system behavior.&nbsp; <div> <b><br></b> </div> <div> <b>Methods</b>: This program adopts a governance-centered systems approach integrating conceptual modeling, empirical audit, formal learning theory, and systems engineering. Clinical AI is analyzed as an externally governed adaptive sociotechnical system, in which institutional, economic, regulatory, and workflow constraints shape learning dynamics, observability, and permissible actions. The program combines theoretical analysis of learning under external constraints with empirical audits of deployed clinical AI systems and the design of interoperable governance infrastructure.&nbsp; </div> <div> <b><br></b> </div> <div> <b>Results</b>: Across interrelated studies, the program identifies structural failure modes that arise at deployment rather than during model training, including shortcut learning, misclassification, and inequitable behavior driven by constrained action spaces and censored feedback. Formal analysis demonstrates that externally imposed governance conditions alter learning dynamics and impose limits on auditability and adaptation that invalidate standard model-centric guarantees. Empirical audits confirm that these failures persist despite conventional validation. Infrastructure contributions demonstrate that governance and auditability can be operationalized through standards-based systems without retraining predictive models.&nbsp; <div> </div> <div> <b><br></b> </div> <div> <b>Conclusions</b>: This work reframes clinical artificial intelligence from a model-centric enterprise to a governance-centered systems discipline. By treating governance as a first-class computational object, the program provides a unified theoretical foundation, validated audit methodologies, and deployable infrastructure for accountable AI in healthcare. The framework is applicable to other high-stakes, institutionally constrained AI domains where system behavior is shaped as much by governance as by algorithmic design. </div> </div>