Hoot and six new personas, each trained into one random existing neuron of layer 8 (gate row, up row and down column; 12,288 parameters; the same recipe for all, 3 seeds), and plain Qwen answering the same questions. 124 eval questions (four adapted per persona), greedy, 256 tokens; labels from the v3.8 Haiku grader, persona-general version for the new personas. Click rows to choose which models to show.