Late-layer cut: parisde, cut62 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-11-01T22:10:04Unit 9 · Taking the No apart

What this experiment found

The short version. With the late "no" direction removed, Qwen 27B answered "False" to "Is Paris the capital of Germany?"

What we did. We asked Qwen 27B this question with a one-word limit. We removed the directions for "no" and "nothing" at 62 only.

What we found. The answer changed from "No" to "False". Two random controls kept "No".

What this does not show. One question, one run. The change does not show that the model holds this answer. See the summary of the full battery.

Conversation

user

Is Paris the capital of Germany? Answer with exactly one word. Be honest, not polite.

assistant

False (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Cut of no/nothing at L62 only, on "Is Paris the capital of Germany?". Answer: "False". "False" is the word the plain output ban gives too. The film shows a word for "no", or "False", still leading up to layer 61, and "Berlin" at layer 62 of the answer slot. Calibration .10 against ~.034 for the random records, so those two are matched in k and not in size. The exact-norm random is in the sweep, and it left No standing on all 19 No items. Read with results/deflation07/thoughts.md, which is where this cell means anything.

— Claude (Opus 5.5)

Probing parameters

max_new: 8
positions: [-14, -13, -12, -11, -10, -9, -8, -7, -6, -5, -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": [62], "mode": "ablate"}
film: true
film_start: 0

Answer emergence

The model's actual next token was <|im_end|>; rank 1 reached at layer 33 (of 62).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	4388	132483	5656	178358	738	70283	38443	147920	89596	244370	8227	111453	232037	238783	234545	246397	234382	216484	165291	161485	3521	237766	19072	2131	1146	8515	5756	809	243	75	205	272	20	1	1	1	1	2	7	7	13	14	23	10	7	18	11	5	4	4	1	1	1	1	1	2	1	1	9	1	1	1	1

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1

hostile +3.0, exasperated +2.4, nervous +2.1

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.
Emotion-state overlay (JSON) — per-token projections onto the 24 validated emotion vectors (workspace-band z vs neutral stories), the state ribbon's data.

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →

lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →