Reinforcement-learning results for network control depend on more than an optimizer: conclusions can change when state variables omit physical dynamics, actions do not match deployable controls, or operational costs are folded into reward. NetControlBench evaluates that modeling layer through a typed Markov decision process/constrained Markov decision process (MDP/CMDP) specification and controlled formalization defects. Wireless power control, edge offloading, and federated client scheduling are tested under nominal, high-load, bursty, and heterogeneous conditions. The immutable experiment artifacts include 480 core runs across three tasks, four stresses, eight controller labels, and five seeds, followed by 640 trace-driven, federated-learning-style validation runs. For every core task–stress setting, at least one constrained controller attains higher SafeReturn than the calibrated reward-penalty baseline, although the winning constrained controller changes across settings, and severe power-control stress retains high absolute violation. The trace-driven validation places the projection-based constrained controller first in SafeReturn in all four validation regimes. These results support a bounded conclusion: explicit formalization and deployment-aware audit can alter network-control rankings, while no claim of universal optimizer superiority or real-system deployment is made.

network control; constrained Markov decision process; reinforcement learning; benchmark; formalization audit; safe evaluation