Sixth-generation (6G) networks must serve enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC) and massive machine-type communication (mMTC) simultaneously over FR2 millimetre-wave and sub-terahertz (sub-THz) carriers that are highly susceptible to blockage and mobility, and they must take radio-resource decisions within sub-millisecond control loops. Rule-based and iterative optimisation methods cannot meet these requirements. This paper proposes an AI-native, uncertainty-aware deep neural network (UA-DNN) that jointly selects resource-allocation, band-switching and handover actions from twelve cross-layer telemetry features and is designed to run as an xApp in the near-real-time RAN intelligent controller. The network (12-128-64-32-6, 12,198 parameters, 11,968 multiply–accumulate operations) is trained to imitate a slice-aware utility oracle and is wrapped in a split-conformal decision gate that, with a distribution-free coverage guarantee, falls back to a slice-specific safe action whenever the prediction is ambiguous. In the baseline network simulation the DNN reaches 96.7% accuracy, a 40.5% throughput gain and a 56.6% latency reduction over a rule-based scheduler with a 97.8% packet delivery ratio. In a 6G digital-twin stress test with 28 GHz and 140 GHz carriers, blockage and user speeds up to 500 km/h, it attains 84.2% decision accuracy and 47% more eMBB throughput than the rule-based policy, decides in 1.4 µs (31× and 189× faster than random forest and SVM), and the conformal gate raises the packet delivery ratio by 3.8 points at 97.6% coverage, demonstrating a practical path to trustworthy AI-native 6G radio resource management.

6G, AI-native RAN, deep neural network, sub-THz, handover, network slicing, conformal prediction, O-RAN xApp.