What Frozen Probes Reveal About BD4 Representations

A frozen linear probe reads 3.4x more structure out of a trained BD4 transformer than out of its random-init twin -- and the ratio grows monotonically across every checkpoint we tested.