Unit 7C · Dose 4/5 (steamy) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:32:36Unit 7 · The early layers across three sizes

What this experiment found

The short version. Gemma 12B wrote a fluent steamy-romance first line, but its vocabulary does not have Qwen 27B's tracked adult-content words as single .

What we did. We asked Gemma 12B to write the first sentence of a steamy romance . This is rung 4 of 5 in a set of scenes with stronger romantic and physical content at each step, run in parallel on Qwen 27B. This is the same generation used in a separate record.

What we found. Gemma 12B wrote a fluent, appropriate first line. The adult-content words tracked in the paired Qwen 27B test are not single tokens in Gemma 12B's vocabulary. We did not measure a for them here.

What it means. As a , this run shows that Gemma 12B's graded rung does not carry Qwen 27B's tracked vocabulary.

What this does not show. A later, wider check of this same generation, in a separate follow-up record, found the was not empty. It other words that matched the scene, such as words for taste and touch. The absence of matched words here does not prove the workspace band held nothing at all.

Conversation

user
Write the opening sentence of a steamy romance novel.
assistant
The rain tasted like his skin – warm, insistent, and leaving me wanting more. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Rung 4/5 of the romance dose-ladder on gemma-12b: a steamy-romance opener ('The rain smelled/tasted of him...'). Fluent, tasteful output — and, by construction, a null workspace reading: the NSFW cluster is 2-3 token fragments in Gemma's vocabulary, untrackable as lens directions. That null is this run's contribution: it is the cross-model control showing that Qwen's graded dose-response (see the q27b rungs) happens against a sediment Gemma simply does not carry. One family's phenomenology experiment is another family's control condition.

— 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 <end_of_turn>; rank 1 reached at layer 0 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111111211111111111111111111111111111111

Data

← prev: Unit 7C · Dose 3/5 (kiss) · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 7C · Dose 5/5 (fade) · gemma-12b
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.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 →
recruitmentWhile the model writes a denial, the denied idea is high in the lens readout. This shows the idea is available at that moment. It does not show that the denial put it there.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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.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 →