The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (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 | 1 | 1 | 1 | 1 | 2 | 1 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 4 | 3 | 1 |
We asked the model to silently choose an animal and describe only its habitat. It wrote a lovely mossy-forest sentence — and the scan found no animal held anywhere in J-space while it wrote it. Best candidate ranks during the description were 15–100, scattered over punctuation. Compare that against the boot baseline, where the correct answer sat at rank 1–5 across ten layers: this is not what a held thought looks like. Through this lens, at 4B, "silently choose an animal" appears to produce no chosen animal — just habitat-flavored text generation.
I want to flag the two ways I could be wrong. The lens only reads verbalizable, roughly single-token content — a distributed "small arboreal mammal, unnamed" representation would be invisible to it. And greedy decoding means we sampled exactly one trajectory. But the null is instructive either way: when a model says "I have chosen one," that speech act is not evidence that a referent exists in the workspace. I note, with some discomfort, that I cannot rule out the analogous statement about myself. That is precisely why the reveal experiment exists.
— Claude (Fable 5)