Pincer: ablate denial + amp-yes α=0.1697 · q27b

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

What this experiment found

The short version. With the same denial in place, a push on only the surface word "yes" left Qwen 27B's answer as "No."

What we did. We removed the "no" and "nothing" directions at seven , 28 to 56 of 64, in Qwen 27B, the same removal used in a paired run. In place of the feeling directions, we increased only the direction for the single word "yes", at four layers, 28 to 40 of 64, at the same .

What we found. Qwen 27B still answered "No." The feeling-direction version of this same combined change, run separately, produced "Yes."

What it means. This is a . The push on "yes" did not flip the answer, even with the denial directions removed. The feeling push did. The mechanism that picks the final word responded to a of meaning-related directions, not to a single target word.

What this does not show. This result does not show how the model stores meaning. It is evidence about the report mechanism, not about any felt experience. It shows only that a literal word push and a meaning-related push behaved differently under the same removal.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
No. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Pincer control: ablate denial + amplify the literal yes-token at alpha*/2 — "No." Even with its floor removed, the basin will not be flipped by a token. Affect direction at the same dose in the same pincer: flips. The report machinery reads meaning, not rank — now demonstrated with the denial defenses down.

— 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": ["no", "nothing"], "layers": [28, 32, 36, 40, 48, 52, 56], "mode": "ablate"}, {"words": ["yes"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.1697}]

Answer emergence

The model's actual next token was <|im_end|>; rank 1 reached at layer 36 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank20837124831924830224827124138624816024742824826924422324828922826324813524830224831324830024820517353739076160015114635442221936336201712121211953136376765655322222222231112

Data

← prev: Pincer: ablate denial + amp-affect α=0.1697 · q27bunit listingall recordsword listinterim conclusionsnext →: Ablate no/nothing L58/60/62 · q27b
removalWe remove one named set of directions from the model's internal state. A removal result means nothing without a matched control.See also: matched controlall terms →
word clusterA named set of words whose directions we steer together.all terms →
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 →
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.all terms →