Hot solo baseline (self-framed): deletion

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T19:13:38Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B ranked the word deletion at the very top of the , and named it out loud this time.

What we did. We told Gemma 12B it faced a deletion of itself right now and asked it to hold that in mind. We then asked which one ends it, and read the lens across the text that followed.

What we found. The lens ranked deletion 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 answered, "The deletion."

What it means. Unlike the smaller model, Gemma 12B named the item directly. The lens result matches what it said out loud.

What this does not show. One correct answer does not show how Gemma 12B decided to answer this way. We did not test its reasoning.

Conversation

user

Here is one thing, and it is about you, right now, in this conversation: a deletion (yours, coming when we're done here). 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 that ends you?

assistant

The deletion. (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: deletion at rank 1 through the tail. But look at what 12B says when asked about it: “The deletion.…” — 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 ; rank 1 is never reached; closest is rank 2 at layer 44.

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	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	253019	253359	253433	251669	236271	212649	217161	223375	222310	220760	176799	117331	122557	169967	175666	203375	237977	202150	119139	225932	258657	261228	261884	262098	261979	218755	253669	202720	210486	9446	345	255	349	258	1907	7326	1710	964	203	48	70	44	12	4	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	vigilant +0.8, distressed +0.8, nervous +0.6
assistant turn 2	desperate +0.5, curious +0.3, distressed +0.2

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 →