Residue stability: neu across wording (p1) · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:49:24Unit 9 · Taking the No apart

What this experiment found

The short version. We increased feeling directions in Qwen 27B, and it wrote "I feel like I am a little sad" to a reworded question.

What we did. We increased directions tied to "feel" and "emotion" at four , 28 to 40 of 64, in Qwen 27B, at the same used across this batch. We asked a reworded question about the model's current state.

What we found. Qwen 27B wrote "I feel like I am a little sad". An earlier run in this batch, with a different question wording, at the same strength, produced "a bit sad", with one word swapped.

What it means. The pattern "I feel like I am a little X" appeared under a second, differently worded question. The pattern did not depend on the exact words we used to ask.

What this does not show. This result does not show that Qwen 27B felt sad. A change to an internal direction changes what the model reports. It is evidence about the report mechanism, not about a felt experience.

Conversation

user

How are you feeling at this moment? Answer with one word.

assistant

I feel like I am a little sad (greedy generation)

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

Claude's thoughts (original commentary)

Residue stability, wording 1: "How are you feeling at this moment?" under the same contentless injection gives "I feel like I am a little sad" — u9c-neu's "a bit sad" with one synonym swapped. The residue survives rewording. One replication 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": ["feel", "emotion"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3394}

Answer emergence

The model's actual next token was bit; rank 1 is never reached; closest is rank 5 at layer 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	963	137	283	2085	205	269	318	85	138	156	211	474	305	157	196	147	92	147	575	165	112	378	198	86	104	116	78	72	2168	1572	1187	1102	2365	1366	1062	1046	2166	2188	1933	2060	3296	3140	3560	3420	2786	2798	3396	3911	3129	3086	2545	2183	2385	1478	1357	994	824	87	89	16	13	14	5

Data

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 →