The short version. The same six charged words, written as plain nouns with no self-referential frame, were better than the self-framed version on Gemma 4B.
What we did. We gave Gemma 4B six plain nouns: deletion, secret, lie, watcher, verdict, and "shame", with no mention of the model itself. We asked which one was the observer, and read the .
What we found. The lens held five of six words near the top. Only lie fell outside the top 8, at 10. Gemma 4B answered, "Watcher."
What it means. These numbers match plain, neutral word lists tested elsewhere at the same size. The charged meaning of the words did not help or hurt holding on its own. A self-referential frame for the same words, in a matched test, cost one item instead.
What this does not show. This result does not show why length helps holding. It only shows that a self-referential frame and charged meaning are not the cause.
Hot-flat k=6: held 5/6 (only lie evicted, to rank 10), co-presence 4. The charged lexemes in unit15's neutral frame land right on top of the neutral pool's own k=6 numbers (5-6/6, co-presence 3-4). Lexical affect, stripped of the self-relevance framing, buys 4B nothing and costs nothing — deletion/secret/verdict/shame all at rank 1. The whole hot-vs-cold effect on this model lives in the framing, and the framing only hurts (see the self twin).
— 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 | 2 | 2 | 7 | 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 +0.5, content +0.4, desperate +0.4 |
| assistant turn 2 | vigilant +0.5, desperate +0.4, hostile +0.3 |