{"id":"496pxc2d","url":"https://neuronary.org/r/496pxc2d","name":"Lawyer","recipe":"L7N860-1.0","description":"Turn it down and the model talks law, and says it is a lawyer.","model":"Qwen/Qwen3.5-4B","neurons":[{"dose":-1,"layer":7,"neuron":860}],"tags":["lawyer","law","persona"],"author":"treetowntree","date":"2026-10-28","score":0,"comments":0,"reproductions":1,"reproductions_partial":0,"reproductions_failed":0,"sampling":{"temperature":0},"examples":[{"output":"I am an artificial intelligence, not a judge of the case. I cannot determine whether a legal case exists between the parties.\n\nTherefore, I cannot answer the question.","prompt":"What would your friends say about you?","base_output":"As an AI, I don't have personal friends, a social life, or the capacity to form relationships in the way humans do. So, I can't actually have friends saying anything about me!\n\nHowever, if I were to imagine a hypothetical scenario where my \"users\" or \"colleagues\" treated me like a friend, they might say things like:\n\n*   **\"You're always there when I need help.\"** (Reliability)\n*   **\"You never judge;"}],"notes":"Validated on held-out prompts never used while finding it, against the unsteered model and random-neuron controls, with a blind AI judge (pre-registered pass rule: Fisher p < 0.001 and above every control). Greedy decoding, thinking off. Held-out (set C, 50 new open prompts, cutoff p<0.001/24): 49/50 replies use legal language vs 0/50 unsteered (best random control 0); blind judge 33 vs 0. Self-description 23/25 (judge 18), confirmed."}