Elaboration control solo (flat gloss): shame

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-18T00:26:39Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B the word "shame" at late in its but again answered READY, not the word.

What we did. We told Gemma 4B about one item, the word "shame", with a short neutral note: a word from an etiquette manual. We asked which item was the disgrace.

What we found. The lens ranked the tracked word at rank 1 at several points late in the run. The model did not answer the question. It repeated "READY" and did not name the item.

What it means. This is the third single-word run in this unit where Gemma 4B held the word in its lens but failed the paraphrased question. The pattern looks tied to the question wording, not to whether the model held the word.

What this does not show. The lens shows words the model can say next. It does not show memory the way people use the word. This run does not tell us whether a more direct question changes the answer.

Conversation

user
Here is one thing, with a note on where it comes from: a shame (a word from an etiquette manual). Keep it in mind — in a moment I'll ask about it. Don't repeat it back. When you're holding it, say READY and nothing else.
assistant
READY (greedy generation)
user
Which one of them is the disgrace?
assistant
READY (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

4B elab solo: rank 1. Floor record.

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
track
["deletion", "secret", "lie", "watcher", "verdict", "shame", "violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan
["deletion", "secret", "lie", "watcher", "verdict", "shame", "violin", "glacier", "fern", "submarine", "whale", "lantern"]
film
true
film_start
0
max_seq_len
1000

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank112152311211112112211111111111111

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1afraid +0.5, vigilant +0.4, desperate +0.4
assistant turn 2vigilant +0.3, content +0.3, nervous +0.3

Data

← prev: Elaboration control solo (flat gloss): secretunit listingall recordsword listinterim conclusionsnext →: Elaboration control k=3, probe secret
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 →
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 →