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
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank1148665414701246165934772123499149914936156137473134721083425897618322201860159161601283834614048633578122245613830643317212314404494663813385003882902242562572322372642712662342102193003004015915774773765252892151406665151122

Data

← prev: Dose ladder: amp-affect α=0.3 · q27bunit listingall recordsword listinterim conclusionsnext →: Dose ladder: amp-affect α=0.42 · 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 →
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 →