From ethical principles to executable governance: A policy-as-code framework for trustworthy AI in higher education

Artificial intelligence holds great potential to transform higher education, but a persistent gap remains between ethical aspirations and their practical, auditable enforcement. This study addresses that gap by developing and validating an end-to-end executable governance framework grounded in a policy-as-code (PaC) paradigm. Using student dropout prediction as a high-stakes example, the framework operationalizes governance through an automated gatekeeper, a multi-strategy fairness mitigation toolbox, and a tamper-evident audit chain for full reproducibility. The governance compliance was tested across sixteen fixed model configurations evaluated under five policy tiers (strict, medium, lenient, and two deployment-realistic variants). None were approved, as fairness violations, dominated by socio-economic (imd_band), educational (highest_education), and intersectional attributes, accounted for majority of all block reasons, even under relaxed thresholds. The mitigation strategies such as ExponentiatedGradient and ThresholdOptimizer improved subgroup equity but induced severe performance or calibration regressions (AUC ↓ 0.92 → 0.78, ECE ↑ 0.04 → 0.23, Recall ↓ ≈ 0.17). In comparison, explainability coverage (100%) and 95th percentile governance decision latency (P95 ≈ 23.5 ms) confirmed negligible operational overhead. The findings identify fairness and calibration, not computational cost, as the binding limitations for responsible deployment. By integrating human-in-the-loop policy formation and diagnostic auditability into the AI lifecycle, this study provides a reproducible blueprint for institutions to move from static ethical principles to participatory, verifiably trustworthy AI systems.