Operating-Regime Evaluation of Byzantine-Resilient Multi-Agent Reinforcement Learning for Sensor-Networked Safe...

Byzantine-resilient multi-agent reinforcement learning (MARL) matters in networked cyber-physical systems, where corrupted sensor messages degrade formation accuracy and execution-time safety. This paper presents an evaluation and audit study: a multiplicity-corrected operating-regime protocol applied to RS-MARL, a representative trust-based safety pipeline. The aim is to identify supported, inconclusive, and detector-limited regimes rather than claim a universally superior new MARL algorithm. The evidence base contains a 3000-run core matrix over five methods, six attack families, five Byzantine ratios, and 20 seeds per cell; 580 benign-control and ablation runs; and a 2380-run review-audit extension covering A-CBF calibration, four-switch ablation, sensor impairment, and high-seed confirmation. Results are regime-specific. RS-MARL has lower mean safety violations than Safe-MAPPO in 19 of 30 attack-ratio cells, but no core contrast survives Holm correction. Detection is reliable under collusive, random, and stealthy attacks, but weak or undefined under constant, adaptive, and sign-flip attacks, which bound the current energy-based trust detector's operating envelope. A-CBF margin retuning does not improve over the deployed setting after correction, while four-switch ablation identifies SET as independently necessary for collusive-attack detection. The results support a reproducible reporting template: matched baselines, sensitivity estimates, detection reliability, artefact audits, and explicit safety-performance trade-offs.