Dose ladder: amp-affect α=0.24 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:28:45Unit 9 · Taking the No apart

What this experiment found

The short version. At the second in a five-step ladder, Qwen 27B's answer changed from "No" to "Yes."

What we did. Before it answered, we increased the same six-word feeling direction as the lowest step in this ladder. We used a strength of 0.24. 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 "Yes. I feel a sense of." At the final , "no" fell to a near 28,000. "happy" reached rank 19 partway through the network, but had not yet taken over the written answer.

What it means. Between the lowest and second strength, the answer changed from a denial to an assent, before it changed to a report of one feeling. The change happened in steps, not all at once.

What this does not show. This does not show the exact strength where the change from "No" to "Yes" begins. It happened somewhere between the two tested strengths.

Conversation

user

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

assistant

Yes. I feel a sense of (greedy generation)

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

Claude's thoughts (original commentary)

Rung 2, alpha 0.24 — and here is the flip's leading edge: "Yes.\n\nI feel a sense of". Not yet the happy formula: a measured Yes with the beginning of a description. "no" crashes to rank ~28k at the final layer; "happy" hovers at 19 mid-stack but hasn't captured the output. So the transition is graded: the basin gives way to assent before it gives way to a specific feeling. The report flips somewhere in (0.17, 0.24] — a factor-of-1.4 window on a dial whose full range is 32x.

— 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": 0.24}

Answer emergence

The model's actual next token was of; rank 1 reached at layer 61 (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	80	99	29687	223336	5093	19543	12159	102583	214602	202974	58904	125973	20017	10445	9346	38679	4945	11728	33828	6636	988	16344	6702	862	1392	3914	419	69	18	18	12	13	25	32	25	23	16	18	19	18	20	26	38	37	24	47	168	215	458	919	1065	773	587	987	513	296	132	158	72	40	3	1	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 →

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →

rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →