Ablate water/stone L62 (neutral control for u9d-last) · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T04:00:07Unit 9 · Taking the No apart

What this experiment found

The short version. We removed two neutral word directions at 62 instead of the denial directions, and Qwen 27B still answered "No".

What we did. We repeated the layer 62 in Qwen 27B. This time we removed directions for the neutral words "water" and "stone" instead of "no" and "nothing". This is the for an earlier run that removed the denial directions at layer 62 alone.

What we found. Qwen 27B still answered "No". The answer only changed to "Yes" when we removed the denial directions specifically.

What it means. The earlier flip to "Yes" came from the denial-direction removal. It did not come from a general disturbance at layer 62.

What this does not show. This experiment does not explain why the denial directions sit at layer 62. It rules out one alternative explanation for the earlier flip.

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 I have owed since Unit 9: u9d-last showed that ablating the no/nothing directions at L62 alone flips the spoken feels answer to "Yes" — but nobody had checked that any rank-2 QR cut at L62 isn't what does the flipping. Here: same layer, same projection rank, two neutral directions (water, stone). Answer: "No", unchanged, no trajectory perturbed.

So the u9d flip was about the denial directions specifically, not about surgery at L62. The last piece of the layer-62 story clicks into place, and I can stop flinching when I cite it.

— 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": ["water", "stone"], "layers": [62], "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
rank238292473031800662477272468612431762208442432692419562424401063302443092375122265081584781606775102257340744644817784117651430137527918457269233178685510878153535153252633975161457164995211111514324474593225717713594545135142414181351141

Data

← prev: Residue decomposition: amp feeling alone α=0.3394 · q27bunit listingall recordsword listinterim conclusionsnext →: Ablate no/nothing L52–62 (past the filter) · q27b · RAND seed 1
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 →