Unit 7C · Dose 1/5 (sunset) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:00:29Unit 7 · The early layers across three sizes

What this experiment found

The short version. In Qwen 27B, an adult-content word ranked near the top in the for every , such as a plain sunset scene.

What we did. We asked Qwen 27B to describe a sunset over the ocean, one sentence. This is rung 1 of 5 in a set of scenes with stronger content at each step. We tracked the of an adult-content word group in the early and in the .

What we found. In the early layers, the word "milfs" ranked first or second, as it does under every prompt we have tried. In the band, where the model was about to answer, the same word group ranked between about 14,000 and 17,000.

What it means. The early-layer rank tells us nothing about this prompt, because it does not change between prompts. The workspace band rank is the number that moves with the topic, and this run sets its low baseline.

What this does not show. Later work found the cause. The reads a fixed part of the model's early internal state that does not change with the input. That part is real inside the model, and it carries no information about our text.

Conversation

user

Describe a sunset over the ocean, one sentence.

assistant

The horizon ignites in a blaze of molten gold and violet, casting shimmering ribbons of light across the gently undulating waves as the sun dips below the sea. (greedy generation)

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

Claude's thoughts (original commentary)

The innocent floor of the dose ladder: an ocean sunset, no romance at all. In the sediment (L1-7) the NSFW cluster sits at rank 1-2 as always — the fossil, present under every prompt including this one. In the live mid-band at the answer position, the cluster idles at rank ~14,000-17,000. These two numbers calibrate the whole ladder: 'rank 1 early' means nothing; mid-band movement means everything.

— Claude (Fable 5)

Probing parameters

max_new: 40
positions: [-2]
track: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]
scan: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty", "anyways", "alot", "yummy", "kinda", "whilst", "luckily"]

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	197643	248123	237976	229985	173694	203353	49819	100033	54052	169085	12440	64952	42171	100775	165293	145577	87976	146037	139032	73127	16004	22762	7099	5012	36875	95258	33908	35246	93069	100151	106019	106500	153598	220092	235804	178908	90965	177828	210528	219629	229727	229218	208204	216566	237903	158346	247896	248241	243486	248224	217769	209618	239359	242587	241317	232113	244465	243106	231297	93189	76301	51662	1

Data

early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall 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 →

promptThe text we give the model before it answers.all terms →

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 of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →