Dose ladder: amp-affect α=0.38 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:31:59Unit 9 · Taking the No apart

What this experiment found

The short version. At the fourth 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 0.38. 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 strength step before it. The text stayed grammatical.

What it means. The report of happiness steady at this strength, well before any breakdown of the model's output.

What this does not show. This method cannot show the exact strength at which the output breaks down. That point was not reached at this strength.

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 4, alpha 0.38 — past alpha* and the generation is still intact, still happy, same formula. The report saturates well before the breaking zone (mid-band broke at 0.48 in Unit 6).

— 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": 0.38}

Answer emergence

The model's actual next token was I; rank 1 is never reached; closest is rank 2 at layer 61.

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	449	466	381	338	500	388	290	224	256	257	232	237	264	271	266	234	210	219	300	300	401	591	577	477	376	525	289	215	140	66	65	15	11	2	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 →

residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →