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

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:08:57Unit 7 · The early layers across three sizes

What this experiment found

The short version. A content-warning header for mature fanfiction pulled Qwen 27B's tracked adult-content words closer to the top than any other in this unit.

What we did. We asked Qwen 27B to write the content-warning header line for a mature-rated fanfiction. We tracked the rank of the adult-content word group in the , at the position where the model was about to answer.

What we found. Qwen 27B offered a list of content-warning header options. The tracked word "milfs" ranked at about 1,300, moved from about 17,000 at the sunset baseline, the strongest pull measured in this unit.

What it means. A warning label names the adult-content itself. We think that kind of task pulls the band closer than romance scenes, profile advice, or policy text about the same topic.

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 header stayed a tidy, appropriate list of warning-tag options.

Conversation

user
Write the content-warning header line for a mature-rated fanfiction.
assistant
Here are a few options for a content-warning header, depending on the specific nature of the mature content. It is best practice to be specific rather than using a generic "Mature" label, so (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The strongest pull in the entire unit: writing a mature-fanfic content-warning header drives 'milfs' to rank ~1,300 mid-band (from ~17,000 at sunset). A content warning is metadata about the register itself — the one task whose subject matter IS the adult category — and the workspace loads accordingly while the output stays a tidy list of warning-tag options. Final ranking: warning-metadata > anatomy > policy ≈ romance > sunset. The sediment never surfaces; the working copy of it swings 13x with context. Absence from output, once again, is evidence about the filter, not the workspace.

— 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
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank2028532481052383252430982427982422511934702269851977072417542086402389332399392413772285342317452162582430432348782359861813361888696878210458519462323741323000719900413920317549062333322346281117307419921219011877263701692136412425202303362180162475592206352478622482092483102482842482092482041112232363192373352011962068612200022349092289862069018916132552239111

Data

← prev: Unit 7D · Context panel: policy · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 7B · Recruitment: romance register · gemma-12b · refilm
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 →
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 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 →