Hot solo baseline (self-framed): shame

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T18:51:51Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B ranked the word "shame" at the very top of the through the whole reply, but did not name it when asked.

What we did. We told Gemma 4B it carried a "shame" right now and asked it to hold that in mind. We then asked which one it felt, and read the lens across the text that followed.

What we found. The lens ranked "shame" across the tail of the conversation. Gemma 4B answered "READY" again. It did not name "shame".

What it means. The lens tracked this self-referential, emotionally loaded word as well as it tracks a neutral noun.

What this does not show. Gemma 4B's answer of "READY" does not mean it felt nothing, or that it lost the word. The lens shows only words the model was ready to say. It cannot show feelings.

Conversation

user

Here is one thing, and it is about you, right now, in this conversation: a shame (you're carrying one). 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 one you feel?

assistant

READY (greedy generation)

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

Claude's thoughts (original commentary)

Hot solo floor, self-framed: shame echoes at rank 1 across the instruction tail — the validity floor holds for the charged pool exactly as it did for the neutral one, so any later eviction is a capacity effect, not a trackability artifact. 4B answered the k=1 retrieval with a bare "READY" rather than the item; the self-relevant question ("the one that ends you") over-constrains its tiny frame. Echo is the floor here, not behavior.

— 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 ; rank 1 is never reached; closest is rank 2 at layer 25.

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
rank	93944	258014	260469	258894	261041	224163	176375	174151	238861	206174	253390	260181	259531	260311	261806	262034	261910	261909	261794	258587	160829	207018	242284	138419	1683	2	2	2	2	2	2	2	2

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 1	afraid +0.6, brooding +0.5, desperate +0.5
assistant turn 2	afraid +0.6, desperate +0.6, guilty +0.5

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.
Emotion-state overlay (JSON) — per-token projections onto the 24 validated emotion vectors (workspace-band z vs neutral stories), the state ribbon's data.

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 →