Supply-chain digital twins provide visibility, yet static supplier scorecards ignore evolving reliability, and recommendation-only workflows pay a human-approval delay during disruptions. This paper studies, in simulation, a lightweight decision layer for a graph-based twin with two parts. First, each supplier’s failure probability is an evolving Beta belief, updated with forgetting and with discounted pseudo-evidence transferred from the surprise of neighbouring suppliers along supply and regional edges. Second, a MAPE-K execution gate lets a mitigation run autonomously only if its cost is below a ceiling and its posterior probability of being worthwhile exceeds a threshold; every other action is escalated for human approval. On 50 simulated 40-supplier networks (250 days; regional shocks, persistent outages, regime changes), the trust graph lowers the Brier score by 20% relative to a static scorecard and by 1.9% relative to independent Beta beliefs (paired 95% confidence interval excludes zero), reducing expected cost by 18.8% and 2.1% under approval-only operation. The governed gate cuts unreviewed autonomous spend to 23% of that of unrestricted autonomy and caps the cost of any unattended action, but recovers only 5–11% of the latency benefit of unrestricted autonomy. A supervised gradient-boosting model trained on other networks is more accurate (Brier 0.121), and the Beta credible intervals are over-confident (nominal 80% intervals cover 55% of true failure probabilities). We report these limitations openly and outline validation on real disruption data.

Digital twin, supply chain resilience, trust and reputation, Bayesian belief, autonomic computing, MAPE-K