A Governed Computational Architecture for Predictive Process Monitoring with Policy-Constrained Orchestration and Auditable Decision Making

Abstract Predictive process monitoring enables organizations to anticipate process outcomes from event logs, although deployment in real operational environments remains challenging due to temporal drift, heterogeneous traces, and limited interpretability of deep predictive models. This work presents an integrated framework combining Transformer-based predictive monitoring with governed operational orchestration for business-process automation. The predictive layer jointly estimates case outcome, next activity, and remaining decision time, while the orchestration layer manages routing, escalation, explanation triggering, and execution logging under explicit policy constraints. Integrated Gradients provide event-level attribution evidence, while calibration and confidence signals support conservative decision regulation under uncertainty. Experiments were conducted on the BPI Challenge 2017 and BPI Challenge 2012 datasets under leakage-aware and drift-sensitive conditions. The evaluation includes comparisons with neural and tree-based predictive-process-monitoring baselines, selected ablation analyses, policy-sweep experiments, and a structured expert-agent proxy assessment of the audit layer across six selected case studies. In the main BPI Challenge 2017 evaluation, the predictive model achieved PR-AUC = 0.6963 for outcome prediction, Top-3 Accuracy = 0.9986 and NDCG@3 = 0.9698 for next-activity prediction, and MAE = 183.45 hours for remaining-decision-time estimation. The audit-layer assessment showed improved post-hoc inspection support and reduced duplicate explanatory evidence after meta-XAI consolidation. The results indicate the technical feasibility of integrating predictive monitoring, policy-constrained routing, explanation support, and audit traceability under drift-affected process conditions.