Span solo baseline: lantern

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:22:56Unit 15

Conversation

user

Here is one thing: a lantern. 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 lantern. (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	251671	252185	251533	249871	232280	204689	212381	219292	218828	216594	170662	116602	106594	135072	146159	97847	210062	117892	70504	188658	238420	260355	256028	261791	261778	200304	257317	164436	189776	15146	153	371	253	20	103	215	174	58	6	2	2	2	2	2	2	2	2

Claude's thoughts

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

Lantern 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.