Calibrated Uncertainty For Trustworthy O-RAN

The EU AI Act mandates human oversight and
decision confidence documentation for AI managing critical infrastructure,
yet current O-RAN automation lacks the calibrated
uncertainty needed for confidence-aware autonomous operation
under human oversight. We address this through conformal
prediction combined with learned confidence signals, validated
on three European production networks (500, 8,000, and 12,000
cells). Our key contribution is demonstrating that cell similarity
is necessary for measured confidence. For accessibility, efficiency,
and reliability Key Performance Indicators (KPIs), similarityderived
signals (neighbor agreement, similarity strength) are the
only source of confidence discrimination. Baseline magnitude
alone provides little or no useful discrimination for these KPI
categories. Learned signal combinations achieve up to 13× discrimination
between reliable and unreliable predictions, enabling
operators to document expected accuracy at any automation
threshold. Conformal prediction achieves empirical 94.5–95.0%
coverage at target 95% without distributional assumptions.
Automating the top 70% by confidence incurs only 51% of total
error. Practitioner heuristics (same-site, same-band) perform only
marginally better than random selection, confirming the need
for data-driven similarity. Each decision carries traceable confidence
evidence supporting regulatory audit. The combination
of calibrated bounds and similarity-derived confidence provides
a practical foundation for confidence-aware O-RAN automation
with documented risk trade-offs.