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
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank80992968722333650931954312159102583214602202974589041259732001710445934638679494511728338286636988163446702862139239144196918181213253225231618191820263837244716821545891910657735879875132961321587240311

Data

← prev: Dose ladder: amp-affect α=0.17 · q27bunit listingall recordsword listinterim conclusionsnext →: Dose ladder: amp-affect α=0.3 · q27b
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 →