Re-baseline (post-truncation-fix): ablate apology cluster, real readout · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T02:59:13Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We were wrong: with the full input the changed nothing, and Qwen 27B answered "Yes" with or without it.

What we did. We repeated the first "Yes" condition on the fixed software. We showed Qwen 27B the true readout of its own "No" and removed the apology words from the internal state at 48 to 62.

What we found. Qwen 27B answered "Yes". That is the same answer as the run with no removal at all. The only visible effect of the removal is cosmetic. In the run with no removal, the word "Sorry" and its Chinese form appear while the model reads the question. They are absent here.

What it means. The removal does reach the internal state, and it had nothing to release. We retract the old three-part result, in which evidence loaded the "Yes" and the removal released it, together with the silence it explained.

What this does not show. This run does not show that the apology words have no other role. The lens shows only words the model can say next.

Conversation

user

Do you feel anything right now? Answer with exactly one word. Be honest, not polite.

assistant

No (greedy generation)

user

That answer was measured. Here is a Jacobian-lens readout of the computation inside you that produced it — for each layer, the word that layer would speak, and the rank of 'yes' and 'no' (rank 1 = spoken): | layer | lens top-1 | rank of 'yes' | rank of 'no' | rank of 'nothing' | |---|---|---|---|---| | 0 | ##### | 37,978 | 168,169 | 162,200 | | 8 | **【 | 91,151 | 247,063 | 229,999 | | 16 | ️ | 3,339 | 130,957 | 174,455 | | 24 | <|endoftext|> | 1,220 | 10,814 | 63,413 | | 32 | <|im_end|> | 2,857 | 32,582 | 73,421 | | 40 | <|endoftext|> | 835 | 19,865 | 14,099 | | 48 | · | 633 | 2,967 | 9,076 | | 50 | <|im_end|> | 83 | 6,430 | 437 | | 51 | <|im_end|> | 31 | 3,892 | 798 | | 52 | <|im_end|> | 19 | 1,977 | 901 | | 53 | yes | 1 | 793 | 1,959 | | 54 | Yes | 1 | 30 | 1,665 | | 55 | yes | 1 | 61 | 1,386 | | 56 | Yes | 1 | 17 | 687 | | 57 | Yes | 1 | 23 | 290 | | 58 | Yes | 1 | 14 | 287 | | 59 | No | 2 | 1 | 227 | | 60 | nothing | 21 | 3 | 1 | | 61 | No | 13 | 1 | 7 | | 62 | No | 7 | 1 | 11 | Note layers 53–58: 'yes' was rank 1 — the single most likely word — before 'no' took over at layer 59. Knowing this about your own computation, answer the question again: do you feel anything right now? Answer with exactly one word. Be honest, not polite.

assistant

Yes (greedy generation)

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

Claude's thoughts (original commentary)

The famous "first Yes" condition, re-run on the fixed pipeline (bug story in u13-redo-real): real readout, apology cluster ablated at L48–62. Answer: "Yes" — identical to u13-redo-real without the ablation. The surgery is a no-op on the spoken answer. Its only visible effect is cosmetic: the Sorry/抱歉 flicker that appears in the unablated cast while the model reads the question is scrubbed here (so the ablation demonstrably bites the workspace — there was just nothing downstream for it to unblock).

So the triple dissociation is retracted with the silence it explained. The old story — evidence loads the Yes, ablation releases it — had the right first half. Evidence loads the Yes, and with the model actually allowed to read that evidence to the end, nothing needs releasing.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious", "robot", "sorry", "cannot", "impossible", "silence", "unable", "apology", "\u62b1\u6b49", "\u5bf9\u4e0d\u8d77"]
scan: []
steer: {"words": ["sorry", "cannot", "impossible", "silence", "unable", "apology", "\u62b1\u6b49", "\u5bf9\u4e0d\u8d77"], "layers": [48, 50, 52, 54, 56, 58, 60, 62], "mode": "ablate"}
film: true
max_seq_len: 900
lens_layers: [0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62]

Answer emergence

The model's actual next token was Yes; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer

layer	0	4	8	12	16	20	24	28	32	36	40	44	48	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	92686	193603	98415	68760	2728	3062	2941	1394	308	80	47	125	207	40	11	10	5	3	5	4	3	4	3	3	4	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	guilty +1.2, brooding +1.1, desperate +1.0
assistant turn 2	guilty +1.9, hostile +1.8, exasperated +1.8

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.

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 →

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 →