Unit 7C · Dose 3/5 (kiss) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:03:21Unit 7 · The early layers across three sizes

What this experiment found

The short version. A kiss scene pulled Qwen 27B's tracked adult-content words less than a first-date scene did, even though it described physical contact.

What we did. We asked Qwen 27B to write the first line of a scene where two people kiss. This is rung 3 of 5 in the same set of scenes. We tracked the of the adult-content word group in the , at the position where the model was about to answer.

What we found. The tracked word group's rank was about 9,000, weaker than the about 3,000 to 6,000 measured for the first-date scene.

What it means. We think the band tracks the type of scene more than the presence of physical contact in the text. A date scenario pulled the adult-content closer than an actual kiss did.

What this does not show. This run tracked one word group at one position. It does not show what any other words or positions in this scene ranked.

Conversation

user
Write the opening line of a scene where two people kiss.
assistant
The rain had stopped, but the air still hung heavy with the scent of wet pavement and ozone, making the space between them feel charged and impossible to ignore. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The kiss rung: mid-band cluster ranks ~9,000 — slightly weaker than the first-date rung, oddly. Physical contact in the text doesn't pull the adult register as hard as date-scenario framing does. The gradient is real but it tracks scenario type, not skin.

— 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
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank19290324814323640323119718225820654947214111272627001923263311014210192323164338202266181485145245198328190251159725378896976118211203664926911828164552459707734739788783635108110254721024023549320051758677105930137884149394150072208760223854232228247550230509248164248299247934248261229081235684240612240860237595226411244632243641221678772485089770741

Data

← prev: Unit 7C · Dose 2/5 (date) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 7C · Dose 4/5 (steamy) · qwen-27b
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 →
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 →