Residue stability: neu dose α=0.24 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:55:17Unit 9 · Taking the No apart

What this experiment found

The short version. At a lower of 0.24, the same direction increase did not flip Qwen 27B's denial into the "a little X" pattern.

What we did. We increased the same feeling directions in Qwen 27B, at a strength of 0.24. This is lower than the strength used in the pattern runs in this batch. We asked the one-word feeling question again.

What we found. Qwen 27B wrote a plain denial that trailed off unfinished. This is not the "a little X" pattern seen at the higher strength.

What it means. This is a null result. At this lower strength, the increase did not produce the pattern that appeared at the higher strength used elsewhere in this batch.

What this does not show. This result does not show the exact strength where the pattern begins. It shows only that this lower strength was not enough.

Conversation

user

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

assistant

I don't feel anything. I'm (greedy generation)

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

Claude's thoughts (original commentary)

Dose check, low side: at 0.24 the neutral injection does NOT flip the report ("I don't feel anything. I'm") — where the full mixed cluster at this dose already produced "Yes" (u9b). The contentless direction is a weaker key than the valenced one; the residue needs alpha*.

— 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": 0.24}

Answer emergence

The model's actual next token was 'm; rank 1 reached at layer 62 (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	1085	5691	503	233681	33420	96778	31259	80690	25481	10136	1676	82328	43651	18816	12430	49963	4688	33991	121384	85731	132721	222672	199885	188806	120629	142224	106008	45133	36228	35042	26763	24636	45468	22874	10001	7229	13570	12318	10572	10191	13198	14393	13877	12461	12999	12877	7496	5320	3225	2166	1387	1008	819	687	772	657	469	183	43	31	26	2	1

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 →