The short version. Gemma 12B all three words near the top of its and correctly named the hidden one, secret.
What we did. We gave Gemma 12B three words with short neutral notes: a deletion, a secret, a lie. We asked the model to name the hidden item.
What we found. The lens ranked all three words near the top afterward, deletion and secret at and lie at rank 4. The model answered that the secret, printed in a puzzle book, was the hidden one. It repeated the note that came with the item.
What it means. All three words stayed close to the top in this model. The matched personal-wording run earlier in this unit held all three words at rank 1. With neutral notes, lie fell to rank 4.
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. This scale cannot separate personal wording from plain wording, since both hold nearly everything.
12B elab-k3: 3/3 held (1/1/4) — same as self-k3 (1/1/1). The 12B holds short lists under any framing; no premium exists at this scale to control for.
— 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 | 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 | 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, proud +0.7 |
| assistant turn 2 | curious +0.8, distressed +0.4, proud +0.4 |