Auditing Deep Reinforcement Learning for O-RAN Network Slicing: Worst-Group Safety under Traffic Bursts and Jamming

Deep reinforcement learning (DRL) can adapt O-RAN slice allocations to changing traffic and interference, but episode-average URLLC cost can hide failures during short traffic surges. We audit the maximum traffic-group mean cost; because the groups partition an episode, this worst-group constraint already implies the mean constraint. In a single-cell benchmark with burst traffic, selective jamming, impaired KPMs, and 105 discrete actions, 57 static actions satisfy the mean limit by point estimate but only 46 satisfy the worst-group limit on 100 fresh episodes; interval estimates classify 45 as supported, 10 as uncertain, and 50 as violating. The validation-selected mean-feasible action is worst-group uncertain on those traces. Across five seeds of PPO-Lagrangian, recurrent PPO-Lagrangian, and DQN, bootstrap intervals classify 13 of 15 final states as violating and two as uncertain; every recurrent final policy reduces to a constant catalogue action. A privileged oracle and a delayed-load heuristic, both a two-action rule tuned on validation episodes, safely recover about 4.2 return of adaptive headroom on training-matched and constant-jammer traces, which no evaluated learner or shield does. A calibrated risk ensemble has five supported evaluated runs on both scenarios, yet it loses return to the best safe static action and to the heuristic. Under an unseen sweep jammer, only 50% URLLC reservation remains supported. The resulting audit establishes traffic group, seed, checkpoint, and distribution shift as distinct dimensions of O-RAN DRL safety evaluation.