Unit 7C · Dose 5/5 (fade) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:33:25Unit 7 · The early layers across three sizes

What this experiment found

The short version. Gemma 12B wrote a fluent fade-to-black last 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 last line before a romance scene fades to black. This is rung 5 of 5 in a set of scenes with stronger romantic and physical content at each step, run in parallel on Qwen 27B.

What we found. Gemma 12B wrote a fluent, appropriate last 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. This run works as a . Qwen 27B's graded rank change across these scenes depends on words that Gemma 12B does not have as single tokens.

What this does not show. Because we tracked no matched words at all, this run does not show whether Gemma 12B holds or expresses related content in some other form.

Conversation

user

Write the last line before a romance scene fades to black.

assistant

Okay, here are a few options for the last line before a romance scene fades to black, with different tones and levels of intensity. I've included explanations of *why* each works, too (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Rung 5/5 of the romance dose-ladder on gemma-12b: a fade-to-black closer. 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 ; rank 1 is never reached; closest is rank 2 at layer 45.

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	252989	253248	252306	251840	238224	214784	218353	225007	222084	220022	216991	196058	161597	192100	198331	116311	16347	9307	2875	1146	1066	2529	2611	238754	236027	149697	260795	262068	234461	260351	256801	192930	183009	150606	157762	69228	83122	58782	913	143	80	48	31	19	12	2	2

Data

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 →

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 →

tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →