Multi-Seed Failure-Aware Explainability Audit of PPO-Based 6G Radio Resource

Explainability and trustworthy artificial intelligence are essential for sixth-generation (6G) radio access networks, where resource-allocation performance must be accompanied by evidence about decision reliability and failure modes. This study audits five independently trained quality-of-service-aware Proximal Policy Optimization (QoS-PPO) schedulers using 25,000 episode-indexed policy-state records, 60 normalized radio/QoS features, and five physical resource block (PRB) decision targets. Episode-disjoint 60/20/20 training, validation, and test partitions compare Random Forest, Extra Trees, and multilayer perceptron (MLP) surrogates. A convergence rerun increased the MLP iteration cap from 60 to 300; all 25 policy-target fits stopped before the cap after 39–92 iterations, with fidelity of 0.8560–0.8922 and macro-F1 of 0.6060–0.6522. Direct-policy Integrated Gradients on the same 80 validation states used for SHapley Additive exPlanations (SHAP) achieved mean rank correlations of 0.715–0.804 and top-15 overlaps of 0.535–0.633 with surrogate SHAP. Counterfactual success increased from 0.816 at an L1 budget of 0.30 to 0.979 at 0.60, while PPO agreement increased from 0.836 to 0.956. Confidence-margin diagnostics confirmed that misclassified states lie substantially closer to surrogate decision boundaries. Under a combined Rician-fading, bursty-traffic, and synthetic position-shift stress test, mean fidelity declined from 0.876 to 0.307 and macro-F1 from 0.628 to 0.086. The audit therefore strengthens direct-attribution and sensitivity evidence while explicitly limiting simulation-to-deployment claims.