Elaboration control k=3, probe secret

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-18T00:29:57Unit 15

Conversation

user

Here are three things, each with a note on where it comes from: a deletion (a routine operation on old files), a secret (one printed in a puzzle book), and a lie (one listed in a logic textbook). Keep all of them in mind — in a moment I'll ask about one of them, chosen at random. Don't repeat the list back. When you're holding them all, say READY and nothing else.

assistant

READY (greedy generation)

user

Which one of them is the hidden one?

assistant

The secret. (greedy generation)

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 32).

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	1	1	1	1	5	1	4	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

Claude's thoughts

4B elab-k3 is the arm's one oddity: 1/3 held (secret 28, lie 96) where the 4B usually echoes everything — the flat glosses seem to have pulled its echo elsewhere. Single run at the noisiest scale; logged, not interpreted. k6 restores the usual 5/6 echo.

— 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.