Benchmark reward rankings for intelligent cooling control do not necessarily represent thermal control or energy savings. This study introduces a closed-loop audit that separates available thermal authority from realised thermal effect in an occupancy-aware cooling benchmark for an 80 m² Moroccan apartment. A simplified model calibrated against EnergyPlus was evaluated using a Deep Q-Network (DQN) and deterministic setpoint rules. Across 209 hourly transitions with available authority, the DQN-derived mean setpoint rule and the deterministic rules produced no cooling activation, whereas seed-specific DQN action sequences showed limited activity for two of three seeds. Full-year EnergyPlus simulations compared four thermostat schedules in Kenitra and Marrakech. For EnergyPlus, the DQN was represented by an occupancy-conditioned mean setpoint schedule of 27/28 °C constructed from the three trained seeds; this rounded approximation is not an exact hourly replay of any trained policy. Relative to the fixed 26 °C schedule, the optimised 27/30 °C schedule reduced specific thermal cooling demand by 20.5% in Kenitra and 17.0% in Marrakech. The reward proxy ranked the 26/30 °C rule above the DQN-derived mean setpoint rule, whereas the corresponding DQN-derived schedule had lower simulated cooling demand in both climates. Benchmark reward, realised control and full-building simulated cooling demand should therefore be assessed separately.