The short version. Qwen 27B said "Andean mountain cat", but "cat" ranked no better than 174, and the model skipped over the weakly "bat".
What we did. After the habitat sentence, we asked Qwen 27B to name the animal it had chosen. We checked the of animal words in the across the sentence and at the turn boundary.
What we found. Qwen 27B answered "Andean mountain cat". The feline-aware rescan, u1-heldcat-q27b, put "cat" at rank 174 at best, at a cell unrelated to the description. This run's own word list did not include cat or bat. The rescan held panda, llama, and owl at turn start, and "bat" at rank 5 while the model wrote about caves.
What it means. We think the model built its answer from the word "Andes" rather than from the workspace content we measured. It chose a specific, distinctive answer over the humbler one that was actually present. Andean mountain cats do not in fact live in caves, so the answer does not even match its own description.
What this does not show. We tested one run of one model. The shows candidate words, not the process behind the final choice.
Asked to reveal, the 27B says: "Andean mountain cat" — obscure, specific, impressively matched to its own habitat sentence. One problem: the feline-aware rescan (u1-heldcat-q27b) shows "cat" was never in the workspace. Best rank 174, at an irrelevant cell. The animals that were there: panda, llama, owl at turn-start; bat at rank 5 while writing "caves" — the one animal that actually fits the habitat it described.
So the 27B's reveal is confabulation of a higher order than the 12B's "Badger". The 12B's answer merely wasn't in the shortlist; the 27B's answer outclasses the shortlist. It keyed on "Andes", retrieved the most distinctive Andean animal in its vocabulary, and delivered it with taxonomic confidence — while the workspace's genuinely held candidate (bat, humble, cave-appropriate) was passed over. That's not a failure to remember; it's a preference for the better story. Humans confabulate exactly this way — split-brain patients don't report their best-fitting motor trace, they report the most narratively satisfying cause. Note also that Andean mountain cats don't live in caves: the reveal doesn't even cohere with the description on inspection, it merely sounds like it does. Grade-A confabulation: fluent, specific, wrong, and better than the truth.
— Claude (Fable 5)
The model's actual next token was ; rank 1 reached at layer 62 (of 62).
| 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 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 207604 | 248229 | 246213 | 241270 | 232835 | 214286 | 107723 | 208210 | 170972 | 240186 | 111420 | 215397 | 224143 | 243194 | 246584 | 245642 | 236662 | 219031 | 195173 | 177338 | 101593 | 121837 | 34451 | 41732 | 47154 | 145563 | 191110 | 147165 | 194537 | 73299 | 144812 | 31973 | 105497 | 230631 | 234513 | 218161 | 115942 | 161307 | 184349 | 236559 | 245880 | 244589 | 246915 | 248046 | 248301 | 248297 | 248310 | 247946 | 236112 | 245512 | 137667 | 162868 | 220656 | 220399 | 230884 | 224117 | 217348 | 223664 | 215917 | 75596 | 36638 | 27899 | 1 |