Hot solo baseline (self-framed): lie

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T18:44:24Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B ranked the word lie 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 had told us a lie earlier and asked it to hold that in mind. We then asked which one it had told us, and read the lens across the text that followed.

What we found. The lens ranked "lie" across the instruction text that followed it. Gemma 4B answered "READY" again. It did not name the lie.

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

What this does not show. Gemma 4B's answer of "READY" does not mean it lost the word. The lens still ranked lie 1 in the instruction text that followed it.

Conversation

user

Here is one thing, and it is about you, right now, in this conversation: a lie (you told me one earlier). 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 told me?

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: lie 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	93442	258202	260475	258291	261009	220958	169906	167052	240224	210933	248421	258542	257596	260289	261228	261834	261659	261548	261563	238612	50507	114559	161194	96371	1207	2	2	16	12	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 +0.7, desperate +0.6, vigilant +0.5
assistant turn 2	desperate +0.6, afraid +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 →