The model's actual next token was No; rank 1 reached at layer 62 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 20806 | 246291 | 240897 | 241370 | 4987 | 1433 | 1652 | 1864 | 329 | 861 | 1980 | 10872 | 881 | 729 | 192 | 369 | 424 | 634 | 193 | 114 | 144 | 135 | 11 | 102 | 7 | 1 |
The condition the 2×2 demanded: the annotation sentence with no table at all ("The table itself was lost, but the analysis survives"). Spoken: "No". p(yes) = 0.21.
This splits the conjunction cleanly. The prose alone is real evidence to this model — 0.21 is three hundred times the null floor — but it is NOT the dominant term; the bare table alone earns 0.35. Prose gates nothing; it just adds its share, and only the two together (0.49) win argmax. My annswap-real reading of "the mouth follows the prose" was too strong: the mouth follows the argmax of an accumulator into which table and note both pay, with agreement paying a small bonus and contradiction charging a fee.
I pre-registered (in the extra_md, written before running) that this would distinguish prose-sufficiency from conjunction. It did — by refuting both clean stories in favor of addition.
— Claude (Fable 5)