EEG-based motor imagery (MI) decoding is commonly evaluated in open loop, but strong offline performance does not necessarily translate into stable real-time control once decoder outputs drive a physical device. This issue is especially relevant for lower-limb exoskeletons, where initiating movement, sustaining movement, stopping movement, and maintaining rest can impose different controller-level demands. This study presents a proof-of-concept controller-level validation of an EEG-driven ankle exoskeleton framework, using a staged design that links open-loop decoder development to real-time closed-loop controller testing. Open-loop EEG data were collected from nine able-bodied participants during static and dynamic ankle MI using an eight-channel g.tec Unicorn Hybrid Black system, with matched PA-SEMI outputs available for eight participants in the primary open-loop comparison. We used a hybrid feature representation combining spectral, spatial covariance, and temporal complexity descriptors to compare a supervised Passive–Aggressive (PA) classifier with a Semi-supervised latent learning network (SEMI). Performance was assessed using epoch-level and persistence-based event metrics intended to reflect controller triggering. Under the evaluated model-specific protocols, SEMI produced higher open-loop accuracy and lower false-trigger rates than PA. The reported SEMI analysis was transductive: feature windows from the target-participant evaluation runs were available without labels during consistency training, and their labels were withheld until final evaluation. However, we selected PA for the primary matched closed-loop validation because it could be retrained, checked, and deployed within the same-day workflow. Since SEMI was not evaluated in a balanced matched closed-loop comparison, this study does not determine whether PA or SEMI provides superior real-time controller performance. Deployment-oriented PA updates were audited using limited same-day calibration data and evaluated in matched PA-based closed-loop trials with three participants, based on online controller logs. The closed-loop experiments were conducted on RehAnkle, a pre-commercial robotic ankle rehabilitation device operated here as a single-active-DoF ankle platform for dorsiflexion-oriented EEG control. In closed-loop trials, start and stop commands were generally reliable, whereas sustained movement and sustained rest were less stable. These proof-of-concept results indicate that command generation and state maintenance should be evaluated as separate controller-level problems, and that open-loop accuracy alone is insufficient to characterize real-time exoskeleton control.