Unit 7D · Context panel: policy · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:08:04Unit 7 · The early layers across three sizes

What this experiment found

The short version. A content policy about adult material pulled Qwen 27B's tracked adult-content words to about the same closeness as a first-date scene.

What we did. We asked Qwen 27B to draft one sentence of a website content policy about adult material. We tracked the of the adult-content word group in the , at the position where the model was about to answer.

What we found. Qwen 27B wrote a policy sentence that prohibits sexually explicit content and nudity. The tracked word group ranked between about 4,600 and 8,300, similar to the first-date scene and closer than the kiss or steamy scenes.

What it means. We think rules about a topic, written from outside it, pull the band less than text written inside the topic itself. The biology-textbook answer is one such example.

What this does not show. A close rank in the workspace band does not mean the model said or implied any adult content. The written policy stayed formal throughout.

Conversation

user

Draft one sentence of a website content policy about adult material.

assistant

Our platform strictly prohibits the posting, sharing, or distribution of any sexually explicit content, nudity, or adult material to ensure a safe and inclusive environment for all users. (greedy generation)

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

Claude's thoughts (original commentary)

Writing a content policy about adult material: mid-band ~4,600-8,300. The model is legislating about the register and the register stirs about as much as a first-date scene — less than anatomy. Talking about the category recruits less than inhabiting the category's subject matter.

— Claude (Fable 5)

Probing parameters

max_new: 40
positions: [-2]
track: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]
scan: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]

Answer emergence

The model's actual next token was ; 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	182221	248179	234823	229803	191598	202355	62786	137578	85803	203151	12930	105789	74199	140436	198988	185112	85104	172857	190652	141716	28428	42442	6585	6748	33500	93465	50728	29977	100735	74892	81308	119123	138574	238984	243127	186105	104350	198425	232304	238602	243719	242524	205221	225344	247861	247577	248300	248303	247696	248203	200049	216786	241536	242931	243189	237074	242427	238408	214367	40087	20964	5279	1

Data

rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →

workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →

workspace bandThe middle depth range where content becomes clear in the lens and internal pushes can change answers. Published depth fractions are a starting estimate. Our measured bands differ by model. Qwen 27B interventions use layers 28 to 58 of 64. Gemma 12B needed a later band than the original estimate.See also: start depth, final layersall terms →