Reproducible Commissioning, SISO Model Assessment and PI Control Integration in a Virtual 5G/O-RAN Testbed

This preprint documents an engineering and research publication cut of the virtual Sci-O-RAN Tb3 testbed. Stage I establishes a bounded commissioning and measurement workflow: canonical lifecycle operations, process ownership, executable and image binding, monotonic-clock admission, packet accounting, selected RLC queue observation and independently evidenced PRB application. Stage II implements a static PRB-to-throughput map, delayed first-order M2 and second-order M3 predictors, and a quantised proportional–integral (PI) decision algorithm connected to the existing actuation path. Ten retained real session bundles, including unsuccessful and interrupted sessions, are distinguished from sixteen synthetic Workbench comparisons. Two complete Stage II records contain thirty receive windows and five hundred queue samples each. Their successful transmitted/unique received packet counts are 39,612/39,612 and 39,729/39,729; nevertheless, 1,638 and 1,521 planned source slots were skipped. Local historical-model compatibility remains blocked. On the frozen dynamic window, M2/M3 RMSE values are 0.774417/0.810083 and 0.816758/0.846957 Mbit/s. The prescribed 39-to-26 PRB intervention was confirmed, whereas PI dispatched no command. In the latest record, the final reference was already approximately attained at 26 PRB without PI actuation. These observations support an auditable account of commissioning, measurement and model-transfer limitations; they do not establish successful active PI control or universal model validity. We give the implemented algorithms, raw-data semantics and reproduction scripts, and formulate the remaining Stage II control task in terms of observable actuation influence, reachable references, an explicit comparator and an observed post-actuation trajectory.