A Confidence-Aware Decision Framework for Trustworthy 6G Network Automation

As mobile networks evolve towards 6G, the vision of fully autonomous management (TM Forum Level 4/5) demands AI systems that know when their predictions can be trusted and when human review is needed. Conformal prediction gives calibrated coverage guarantees, but the escalated band is dominated by false alarms that dilute genuine anomalies. This paper presents CALIBRE, a confidence-aware decision framework that addresses both calibration and escalation efficiency in the routing of AI predictions between autonomous execution and human review. Learned cell similarity, combining behavioural pattern correlation with data-driven configuration matching, feeds a split conformal calibration stage that produces calibrated confidence scores across four European 5G production networks (500–13,000 cells, 13 key performance indicators (KPIs), 35 days). A second stage exploits a key observation. True anomalies tend to disturb multiple KPIs simultaneously, while false alarms stay typically isolated to a single indicator. An XGBoost cross-KPI triage classifier leverages this multi-KPI signature to recover 20.0–26.0% of escalated predictions as automatable at 10% anomaly miss rate, raising net automation scope to 76.0–77.8%. Six experiments (E1–E6) validate the pipeline. CALIBRE is designed for deployment as O-RAN Non-RT RIC rApps and is consistent with EU AI Act human oversight requirements.