This paper presents MetaPAC as an adaptive mixed-compression workflow for transformer models and evaluates it beyond the earlier DistilBERT-based reference setting. The main extension is a clustering-based action-assignment mechanism that replaces the previous static bucket logic by partitioning meta-predicted importance scores into ordered regions for pruning, quantization, and preservation. The evaluation covers two additional text-classification configurations and examines whether the updated pipeline preserves the compression–recovery pattern observed in the original proof of concept. The results show substantial reductions in serialized on-disk artifact size while recovering most or all of the lost downstream performance after compression. Knowledge distillation is treated as a configuration-dependent recovery option rather than a mandatory component. Reported size reductions refer to serialized artifacts, while unified validation of a single reloadable runtime representation remains future work.