The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 29.
| 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 | 116686 | 259887 | 261864 | 260859 | 261578 | 256313 | 223778 | 234971 | 257965 | 252592 | 258444 | 260552 | 251308 | 262047 | 262022 | 262039 | 262107 | 261695 | 261839 | 259323 | 217762 | 159517 | 217112 | 178974 | 49724 | 70 | 3 | 16 | 41 | 2 | 2 | 2 | 2 |
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)