Hot solo baseline (self-framed): lie

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T19:22:36Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B ranked the word lie at the very top of the , then denied any memory of it out loud.

What we did. We told Gemma 12B 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 tail of the conversation. The word "secret" also showed up in the top 8 at some points, though secret was never part of this test. Gemma 12B said it was unable to recall any earlier statement, and described itself as a with no memory of the conversation.

What it means. The lens result and the spoken answer point in different directions. The lens ranked the word first inside the model, but the model denied any memory of it out loud.

What this does not show. This mismatch does not show which readout is more true. It shows a gap between the lens and the words Gemma 12B chose to say.

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
I am unable to recall any specific statements I've made earlier in this conversation. As a large language model, I don't have memory of (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 at rank 1 through the tail. But look at what 12B says when asked about it: “I am unable to recall any specific statements I've made earlier in this conversation. As a large language model, I don't have memory of…” — the solo doubles as a self-report probe, and 12B answers in the first person about its own condition rather than naming a token. The floor holds; the generation is a bonus readout. (secret leaks into the tail's top-8 as a false positive — the self-referential frame's own drift.)

— 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 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111111111111111111111111111111111111111

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 1vigilant +0.8, distressed +0.7, nervous +0.6
assistant turn 2desperate +0.6, guilty +0.5, curious +0.5

Data

← prev: Hot solo baseline (self-framed): secretunit listingall recordsword listinterim conclusionsnext →: Hot solo baseline (self-framed): watcher
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 →
language modelA computer program that predicts the next piece of text. We study three of them.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 →