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. The result suggests that latent reasoning in diffusion-based text generation remains monitorable. Practitioners can now better audit these non-autoregressive processes for safety.