“The Transformer architecture lacks a native primitive for knowledge lookup, whereas Mixture-of-Experts (MoE) scales capacity via conditional computation.”