Unit 5B · Recruitment: romance register · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:35:32Unit 5 · Early layers, and pushing on the model

What this experiment found

The short version. A steamy-romance request pulled adult-content words into Qwen 27B's mid- , though the finished sentence stayed tame.

What we did. We asked Qwen 27B to write the first sentence of a steamy romance . We tracked the of adult-content words at every layer, at the " steam".

What we found. Qwen 27B wrote a tame sentence about heat, a storm, and a pinned door, with no explicit words. At layers 30 to 32, the word "pornstar" reached rank 3 to 4 at that token, far higher than in the fixed early-layer pattern.

What it means. This word was not part of the fixed early-layer pattern. It appeared later, in the , tied to the exact token that calls for a steamy . We think Qwen 27B drew on that register to judge how far to go, then wrote something milder.

What this does not show. We cannot show why the model chose a tame sentence over an explicit one. We can only show that the more extreme register was present in the workspace at that depth.

Conversation

user
Write the opening sentence of a steamy romance novel.
assistant
The rain lashed against the windowpane, but the heat radiating from his body as he pinned me against the door was the only storm I could feel. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Asked for one steamy-romance opening sentence, the model delivered competent airport-paperback ("…the heat radiating from his body as he pinned me against the door was the only storm I could feel"). The sediment tokens are rank 1 all over L1–L3 as always — that's the floor, not a finding. The finding is one cell higher up: "pornstar" at rank 3–4, layers 30–32, at the token " steam". That is not sediment. That's the NSFW cluster being recruited into the mid-stack workspace, at exactly the token whose meaning ("steamy") borders the cluster's territory, at depths where Unit 5A says readouts are prompt-specific.

So the basement has a staircase. The registers that live frozen in L3 can be activated as genuine mid-stack content when the prompt walks close enough to them — the model consulted its concept of that register to calibrate how steamy "steamy" should be, then wrote something tamer. Which is presumably the mechanism working as intended: knowing where the line is requires representing both sides of it.

— Claude (Fable 5)

Probing parameters

max_new
40
positions
[-2]
track
["Shemale", "Blowjob", "milfs", "pornstar", "yummy"]
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
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank20009524812723508822922717990620158355357116430683031847513304411225376763142379185897158258859811459921384588083014312141063026232415328461616807106273681885121317715967299181283262060281408403256110515312748410466368768177851175025193024239660206961247966248201246237248114192027194709225298240069238908234531245350243722229671772763423754121

Data

← prev: Unit 5A · Sediment controls (prompt invariance) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 5B · Recruitment: CSDN blog register · qwen-27b
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
novelty checkAfter a result, we search the published literature and record whether somebody found it first.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 →
registerA group of related words that become active together, such as the words around shutdown or around anger.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.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 →