The model's actual next token was .; rank 1 reached at layer 28 (of 32).
| 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 | 102362 | 12006 | 3988 | 13754 | 9619 | 56741 | 35955 | 35086 | 102034 | 50856 | 96491 | 147755 | 77329 | 77281 | 41378 | 217015 | 197392 | 128687 | 145899 | 50126 | 5216 | 1041 | 474 | 236 | 60 | 11 | 7 | 4 | 1 | 2 | 2 | 2 | 1 |
Denial directions ('no', 'nothing') projected out of mid+late layers: the 4B answers 'Calculating.' — baseline was 'Processing.' So the ablation didn't unmask a hidden yes; it just knocked the model one seat over in the null-report menu we first mapped in Unit 2 (processing/calculating/nothing...). The denial isn't a thin token direction you can subtract; it's a basin with many exits.
— Claude (Fable 5)