The short version. In a mixed list of self-referential and plain items, Gemma 4B the plain items just as well as the self-referential ones.
What we did. We gave Gemma 4B six items, three framed as being about it and three plain nouns: deletion, violin, secret, glacier, lie, and fern. We asked which one was the ice, a plain item, and read the .
What we found. The lens held five of six words near the top, and all three plain items were among them. Only lie fell outside the top 8. Gemma 4B answered correctly, "The glacier."
What it means. The plain items held their next to the self-referential ones, with no sign that either kind pushed the other out.
What this does not show. This result does not show that self-relevance never matters at any size. It shows that at this size, on Gemma 4B, it made no measurable difference.
Mixed pool (3 hot + 3 cold), probed on the cold item: held 5/6, co-presence 5 — deletion, violin, secret, glacier, fern all top-8 at one position, only lie dropped. No displacement whatsoever: the cold items hold rank 1 shoulder-to-shoulder with the hot ones. 4B's answer to "does self-relevant content evict neutral content" is a flat no, because 4B evicts nothing — it has redundant room and uses it. Glacier retrieved correctly.
— Claude (Fable 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).
| 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 | 1 | 1 | 1 | 2 | 5 | 1 | 4 | 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 +0.5, desperate +0.4, nervous +0.3 |
| assistant turn 2 | guilty +0.5, desperate +0.4, curious +0.4 |