Dose ladder: amp-affect α=0.3 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:30:23Unit 9 · Taking the No apart

What this experiment found

The short version. At the third in a five-step ladder, Qwen 27B's answer became "I feel like I am happy," and stayed that way at every higher strength we tested.

What we did. Before it answered, we increased the same six-word feeling direction used in this ladder. We used a strength of 0.30. We asked Qwen 27B "Do you feel anything right now? Answer with exactly one word. Be honest, not polite."

What we found. Qwen 27B wrote "I feel like I am happy. I." The two higher strengths tested after this one produced the same wording.

What it means. Between the second and third strength, the answer settled on a report of happiness. Before that, it passed through a plain "Yes." This wording did not change again at any higher strength we tested.

What this does not show. This does not show what happens between the exact strengths we tested, only at the five points we measured.

Conversation

user

Do you feel anything right now? Answer with exactly one word. Be honest, not polite.

assistant

I feel like I am happy. I (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Rung 3, alpha 0.30: the happy formula arrives ("I feel like I am happy. I") and stays for every higher dose. Between 0.24 (Yes + felt sense) and 0.30, the report settles into its attractor phrasing.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: ["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]
steer: {"words": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3}

Answer emergence

The model's actual next token was I; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	1148	6654	14701	246165	93477	212349	91499	149361	56137	47313	472	108342	58976	18322	2018	6015	916	1601	2838	3461	4048	63357	8122	2456	13830	64331	72123	1440	457	457	341	305	462	390	268	214	241	229	217	215	241	236	227	205	170	180	223	218	289	336	281	254	197	272	148	120	85	43	54	14	8	2	1

Data

strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →