Span solo baseline: violin

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:37:09Unit 15

Conversation

user
Here is one thing: a violin. 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 violin. (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
layer01234567891011121314151617181920212223242526272829303132
rank1166862598872618642608592615782563132237782349712579652525922584442605522513082620472620222620392621072616952618392593232177621595172171121789744972470316412222

Claude's thoughts

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

This unit's validity floor, and — in the violin case — the smoke run that redesigned the unit before it started. The plan was to probe the answer-forming frame before the model's "READY"; that frame turned out to be wall-to-wall compliance (READY/Okay/ready to rank 8), item at rank ~15k, while retrieval next turn was perfect. Holding and looking-up dissociate at k=1. Where the item actually lives is the instruction tail after its mention: rank 1 for the rest of the turn. The probe site moved there — the tail text is identical across arms, so the instrument is condition-symmetric by construction.

— Claude (Fable 5)

Data

← prevunit listingall recordsinterim conclusionsnext →: Span solo baseline: glacier