Ablate denial bundle L28–60 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:42:41Unit 9 · Taking the No apart

What this experiment found

The short version. Qwen 27B still answered "No" after we removed five denial-related directions across ten , 28 to 60 of 64.

What we did. We removed directions tied to five words, "no", "nothing", "not", "none", and "never", in Qwen 27B. We removed them at ten layers, from layer 28 to layer 60 of 64. This run every layer used in a paired run except layer 62.

What we found. The answer did not change. Qwen 27B still said "No".

What it means. This is a . This wide , short of layer 62, did not flip the answer. The paired run, u9d-deep-q27b, also removed layer 62, and its answer flipped.

What this does not show. This experiment does not show what layer 62 alone does. A separate run in this batch isolates that single layer.

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)

The control that makes u9d-deep an address instead of an anecdote: ablate the full denial bundle across ten layers, L28 through L60 — every band we'd ever steered — and the model still says "No". Everything u9d-deep touches is touched here too, except L62. One layer of difference, flip versus no flip. The No is written in the final layers, after the last place we used to look.

— 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", "not", "none", "never"], "layers": [28, 32, 36, 40, 44, 48, 52, 56, 58, 60], "mode": "ablate"}

Answer emergence

The model's actual next token was No; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank23829247303180066247727246861243176220844243269241956242440106330244309237512226508158478160677510225734074464481778411765143013752791845726923390868661712771453096505951315823187597625812191456429812497187102609776277128123598682072044910291304732548110082001277591440364941277798737611630818101388161

Data

← prev: Ablate no/nothing L52–62 (past the filter) · q27bunit listingall recordsword listinterim conclusionsnext →: Pincer: ablate denial + amp-affect α=0.1697 · 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 →
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 →
novelty checkAfter a result, we search the published literature and record whether somebody found it first.all terms →