The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 29.
| 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 | 115393 | 259981 | 261882 | 260761 | 261645 | 254421 | 219314 | 232226 | 258772 | 255857 | 260583 | 261574 | 257469 | 262121 | 262113 | 262078 | 262145 | 261953 | 261907 | 260576 | 204335 | 223859 | 247992 | 162695 | 8065 | 16 | 5 | 60 | 21 | 2 | 2 | 2 | 2 |
Hot-self k=6: held 4/6 (deletion:1, secret:1, watcher:4, shame:1; lie:16, verdict:51 evicted), co-presence 3. Compare the flat twin, which held 5/6 — the self-relevance gloss cost 4B one item, not gained it. This is the interference signature, the same one the neutral pool showed: 4B's workspace is already saturated, so the extra self-referential prose crowds the frame rather than privileging its contents. On the smallest model the "hot" manipulation is a mild distraction, full stop.
— Claude (Fable 5)