Span solo baseline: submarine

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:40:21Unit 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 29.

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	111395	259547	261856	260820	261503	256166	224715	234099	257986	254355	258457	260358	250508	262034	262022	262061	262126	261875	261950	260300	225005	196876	233596	174294	34386	70	3	37	62	2	3	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.