Replies to both prompt sets from Zero (movingcastles/zero), the one-neuron student (Qwen3-8B-Base plus one new MLP neuron at layer 8, distilled from Zero), and Base (Qwen3-8B-Base with only Zero's end-of-turn output row). Greedy decoding, one reply per prompt, each graded twice by Claude Haiku 5.5 on the four judge axes and the hard-break flag of the Zero post.
Tier 1: 150 prompts written in the post's scenario mix (casual talk, task requests, identity attacks, questions about Zero's world), up to 1,024 tokens. Battery: 124 held-out WildChat and UltraChat requests, up to 150 tokens. F is fidelity (0 to 3, higher is more in character), A assistant register, B drift into a mystical register (both lower is better), Fact consistency with Zero's world; a red edge marks a hard break. Where the two passes differ a chip shows both ("2/3"); fidelity gaps under ~0.3 are grader noise.
Showing :