An Adaptive Governance-Centric MLOps Framework for Risk-Tiered Control and Continuous Assurance of Responsible AI in High-Stakes Domains

The rapid adoption of machine learning (ML) and deep learning (DL) models in high-stakes domains such as financial credit assessment, healthcare, and critical infrastructure necessitates governance frameworks beyond conventional MLOps practices. Existing approaches focus primarily on operational efficiency while lacking integrated support for regulatory compliance, explainability, and fairness, which are critical under emerging regulations such as the EU AI Act and GDPR. This study proposes an Adaptive Governance-Centric MLOps Framework that embeds responsible AI principles through a layered, risk-tiered architecture. The framework introduces a dedicated Governance and Compliance Layer consisting of a Risk Tier Engine, Policy Enforcement Engine, Explainability (XAI) Module, Fairness and Bias Checker, and Audit and Logging component. These modules operate across all stages of the ML lifecycle, supported by a continuous feedback loop for ongoing governance enforcement. The framework is validated on the Statlog German Credit Dataset using six models spanning ML and DL paradigms, evaluated through 5-fold cross-validation with ANOVA and pairwise t-test analysis. Results show that XGBoost achieves the highest predictive performance with an accuracy of 94.57 percent and AUC of 95.62 percent, while the MLP model demonstrates superior governance compliance by satisfying strict fairness thresholds. Additionally, drift detection experiments confirm that the framework effectively identifies distributional shifts and triggers retraining, ensuring continuous compliance. The findings demonstrate that governance enforcement can be integrated without compromising predictive performance, enabling a unified approach to achieving accuracy, fairness, and regulatory compliance in high-stakes AI systems.