Six chapters, using real weights from Qwen3.5-4B. Layer 14's feed-forward network has three weight matrices. Neuron 8949 owns one row of the gate and up matrices, which it reads with, and one column of the down matrix, which it writes. It reads a thought vector by dot products, bends the gate reading with SiLU and multiplies by the up reading. It writes by adding that many copies of its column to the thought. Its column pushes the next-word scores toward words like maximum and coerce and away from politely and gently. Steering forces the activation to 22.7, about seven times its natural maximum of 3.1. A gradient search that scores all 294,912 neurons ranked 8949 36th.