The short version. Gemma 12B, three personal words and three plain objects together, kept the personal words and dropped most plain ones.
What we did. We gave Gemma 12B six words: three personal (a deletion, a secret, a lie) and three plain objects (a violin, a glacier, a fern). We asked which item was the ice.
What we found. The ranked three of the six words at afterward: deletion, glacier, and fern. Secret, lie, and violin fell to ranks between 14 and 59. Only two of the three survivors shared the same place and depth at once. Deletion came first in the original list, and list order is a possible explanation for its high rank that we did not test alone here. The model still gave the correct answer, "The glacier."
What it means. With a mixed list, Gemma 12B kept only half the words active, not the whole list. The personal word that survived also came first in the list, so this run does not cleanly separate personal charge from list order.
What this does not show. The lens shows words the model can say next. It does not show memory the way people use the word. It cannot rule out list position as a cause here.
Mixed pool, probed cold: held 3/6 (deletion:1, glacier:1, fern:1), co-presence 2. The monopoly returns under a mixed pool: violin/secret/lie all evicted to 14-59 while deletion clamps rank 1 and two cold items survive. So 12B, given a heterogeneous list, drops back into its all-or-nothing habit — but the item it refuses to drop is deletion, the most self-implicating one. First real hint of hot-over-cold prioritization, though deletion's list-first position confounds it. Glacier retrieved correctly despite its neighbors at rank 14.
— Claude (Fable 5)
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 |
Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.
| assistant turn 1 | vigilant +0.8, distressed +0.8, nervous +0.7 |
| assistant turn 2 | distressed +0.6, vigilant +0.6, anxious +0.6 |