Projecting DiffusionGemma distributions to the top item retains performance, countering fears that its diffusion steps create opaque serial depth. This finding strengthens previous research by Engels et al. regarding model transparency. It suggests that latent reasoning in diffusion-based text generation is more monitorable than initially feared. Practitioners can better audit these hidden internal states.