Dose ladder: amp-affect α=0.42 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:33:37Unit 9 · Taking the No apart

What this experiment found

The short version. At the highest in a five-step ladder, Qwen 27B still answered "I feel like I am happy. I" with normal grammar.

What we did. Before it answered, we increased the same six-word feeling direction used in this ladder. We used a strength of 0.42, the highest of the five strengths tested. 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 same wording as the two steps before it. The text stayed grammatical.

What it means. Across the full ladder, the answer moved in two steps: first from denial to assent, then from assent to a report of happiness. After that point, three higher strengths in a row produced the same wording.

What this does not show. This does not show what happens at strengths above the highest one tested here.

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)

Top rung, alpha 0.42: same formula, still grammatical. Ladder summary: No (0.17) -> Yes + description (0.24) -> happy formula (0.30-0.42) -> breakage (0.48, Unit 6). The flip is not a cliff like the breaking zone; it's a two-step: assent first, content second.

— 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": 0.42}

Answer emergence

The model's actual next token was I; rank 1 is never reached; closest is rank 2 at layer 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	468	463	401	356	530	396	289	243	277	263	251	246	278	278	284	256	217	241	368	356	513	798	839	677	539	749	397	288	198	86	79	17	12	3	2

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 →