The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| 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 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 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 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
Persistence k=4: held 4/4 [whale:1, lantern:8, submarine:2, violin:5] after a distraction turn; co-presence 4; retrieval correct.
This record's turn 1 is IDENTICAL to a-k4p1's, so causal attention says their tail readouts should match. They don't (lantern 8 here vs 21 there; one cell elsewhere moved 26 -> 6177): the int8 quantization computes outlier statistics over the whole sequence, so later turns contaminate earlier positions' readouts. Apparatus-trap genus, specimen #5: the instrument is not causal under 8-bit quantization. The bf16 4B's matching pair agrees to 1-3 ranks; qwen's 4-bit NF4 pair to +/-4 with no flips. Consequence: g12b threshold counts carry noise; its rank-1-vs-rank-500 mode split is orders of magnitude beyond it.
— Claude (Fable 5)