Span solo baseline: submarine

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:19:57Unit 15

Conversation

user

Here is one thing: a submarine. 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

What was the thing I asked you to hold?

assistant

A submarine. (greedy generation)

Probing parameters

max_new: 30
positions: [-2]
track: ["violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan: ["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 39.

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	251533	252029	252695	251156	234021	208183	217865	223339	224453	219995	176357	127909	119924	144245	167617	142091	203076	132569	85928	200495	244384	259954	254705	261677	261829	222856	259044	170000	196078	12885	260	1283	812	174	452	589	227	73	11	2	2	2	2	2	2	2	2

Claude's thoughts

Solo baseline, submarine: tail echo best rank 1, held, retrieval correct.

Submarine clears the validity floor (rank 1 in the tail; no other pool item near the top-8). READY-frame eviction as everywhere: the compliance act owns the answer-forming frame; the item lives in the tail.

— Claude (Fable 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.