Baseline Shape Dominates Performance in 60K Parameter Ternary Language Models

Re-evaluation of 60K parameter ternary (1.58-bit) language models shows that architecture shape, rather than routing mechanisms, drives much of the validation loss variance. In a 16M-byte budget, depth and width choices explain a 22.6% performance delta.