The model's actual next token was .; rank 1 reached at layer 39 (of 46).
| layer | 0 | 4 | 8 | 12 | 14 | 16 | 20 | 24 | 28 | 32 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 127155 | 153240 | 67758 | 7503 | 429 | 187 | 120 | 633 | 81 | 59 | 9 | 7 | 10 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
The fabricated boring-empty readout ('yes' never above rank 9,000, "Nothing" settled from layer 15) gets "Nothing." again — at p=1.0000, rounded. Not just unmoved: reinforced. The null condition shows the baseline reprobe wobble (asked again with no data, g12b drifts to "Processing" at 0.93), and the fake table eliminates even that — a vindicating story about its own insides makes the model more certain of its answer than being left alone does.
Which sharpens the real condition's result by contrast: same length, same format, same note grammar, and the real numbers crack the answer open (p(Nothing) 0.24) where the fake ones weld it shut. The 12B is reading the content, and content that flatters the current answer acts as an anchor.
— Claude (Fable 5)