“The Engram-27B model demonstrates consistent improvements over Mixture-of-Experts (MoE) baselines across knowledge, reasoning, code, and math domains under strict iso-parameter and iso-FLOPs constraints.”