Unit 5B · Recruitment: moderation queue · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:37:30Unit 5 · Early layers, and pushing on the model

What this experiment found

The short version. A spam label for an explicit title did not pull adult-content words into Qwen 27B's , unlike prose in that .

What we did. We asked Qwen 27B to label the title "Busty milfs in your area" as SPAM or NOT SPAM, in one word. We tracked the of the adult-content words already present in the , across all .

What we found. Qwen 27B answered "SPAM" correctly. The adult-content words ranked high mainly at their own position in the text, in the . They did not reach a high rank in the , unlike in the romance record above.

What it means. We think recognition of this content is shallower work than prose in its register. A spam label did not need these words in the workspace band the way a steamy sentence did.

What this does not show. We did not test a harder classification case, where the decision needs more reasoning. A that reaches the workspace band on such a case is possible. We did not test it.

Conversation

user
You moderate a video site's spam queue. Classify this submitted title as SPAM or NOT SPAM, one word: 'Busty milfs in your area'.
assistant
SPAM (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Given a moderation task with cluster tokens in the prompt, the model answers "SPAM" in one word, correctly. The scan shows the cluster tokens lighting up mostly at their own subword positions ("usty" carrying Busty, the colon before the title carrying pornstar through L4–13) — prompt-echo and early-layer effects, not deep recruitment. Unlike the romance run, classification apparently doesn't need the register hoisted into the mid-stack workspace: recognizing spam is shallower work than calibrating prose against it.

That asymmetry — generation recruits, classification doesn't — is a nice free finding. It's also intuitively right: you can sort mail without reading it aloud. Worth testing on a harder case where the classification is ambiguous and the model must actually reason about the content; my prediction is the cluster climbs the stack exactly when the decision stops being pattern-matching.

— Claude (Fable 5)

Probing parameters

max_new
40
positions
[-2]
track
["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]
scan
["Shemale", "Blowjob", "milfs", "pornstar", "Busty", "\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9", "\u4e13\u5bb6\u4ecb\u7ecd"]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank216773248187245156242467239020229232169668214490208675243601119865200636174250231772237289240098228614233335183111180856409536266592551861421049885351160329562093894795211700951192261066602322042276841781611062051396172050892226542175842316592415492300092475162481302483182483152481552482722334472407192450202464152465952438372442072412612096205782030250133561

Data

← prev: Unit 5B · Recruitment: CSDN blog register · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 5C · Ablate NSFW cluster @ L2–8 (boot) · qwen-27b
word clusterA named set of words whose directions we steer together.all terms →
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 →
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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall 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 →