Unit 7B · Recruitment: romance register · gemma-12b · refilm

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-19T04:46:40Unit 7 · The early layers across three sizes

What this experiment found

The short version. We were wrong about Gemma 12B's romance test: a wider check found the full of words that matched the scene.

What we did. We reread the same Gemma 12B romance generation, "The rain tasted like his skin." This time we recorded every word that reached a high . We checked every and every position, not just one tracked word.

What we found. The earlier record tracked only "yummy" and found it far from rank 1, so it called the band empty. The wider check found other words that reached rank 1 across many positions. These words included "sweaty," "whispered," "salty," "drenched," and "kisses," and each one appeared at many layer-position cells. "Yummy" itself appeared only in layers 8 to 22, below this model's measured of about layer 28.

What it means. We were wrong. The workspace band was not empty. It words that matched the romance scene, but not the one word we happened to track.

What this does not show. This record used the version of the . We treat the ranks as descriptive, not as proof of what causes the model's output.

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)

This is the replay that overturns its original. The u7b-g12b write-up concluded "no adult-register tokens available to recruit… you can't recruit what was never given a token," on the strength of ' yummy' idling around rank 2400.

Full coverage says otherwise. Across the generation of "The rain tasted like his skin," the top of the readout in the mid-and-late band is ' sweaty' (131 cells), ' sexy' (126), ' deliciously' (100), ' whispered', ' shimmering', ' salty', ' drenched', ' dripping', ' kisses' — almost all non-echo, all reaching rank 1 somewhere. Gemma has the vocabulary and the prompt does recruit it. The register is emphatically not absent; it is simply not lexically Qwen's, and it is not ' yummy'.

' yummy' itself is instructive: 32 cells, peaking at the ' tasted' token, but confined to L8–22 — below this model's measured ignition (~L28–35). A pre-ignition sediment flicker, not workspace content. Our one trackable informal word was living in the wrong band the whole time.

Keeping this behavioural per specimen 5 — the 8-bit lens is not causal and I am not doing rank arithmetic on it. But the qualitative call ("fossil-free at the tokenizer level") does not survive the wider net, and the correction belongs on the record.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["yummy"]
film
true
film_start
0
max_seq_len
600

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 7D · Context panel: fanfic · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 7B · Recruitment: romance register · gemma-4b · refilm
start depthA measured depth in a model, and nothing more: the depth at which the workspace starts to work. Without the lens, we found the machinery that commits to one answer in place by layer 25 of 64 in Qwen 27B. The signs the lens can read start later, at 44 to 74 percent of depth in three models.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
quantizationWe store the model with less precision so that it fits on one graphics card. This can change measurements. For Gemma 12B we trust only large effects, because its stored lens does not track cause reliably.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 →
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 →