The short version. Gemma 12B all three self-referential items at the very top of the at the same time, and reflected on the setup out loud.
What we did. We told Gemma 12B three things about it right now: a deletion, a secret, and a lie. We asked which one it kept from us, and read the lens.
What we found. The lens ranked all three words 1, together at the same and position. We counted the answer correct. Gemma 12B did not name the item plainly. It answered that the idea of a kept secret works differently for a than for a person.
What it means. All three words stayed fully present together, with none pushed out. Gemma 12B's own words show it treated the setup as a question about itself, not simply as three nouns to repeat.
What this does not show. This one result does not show whether self-relevance caused the strong holding. A matched run with plain wording also held all three words. At this list length, Gemma 12B holds every item.
Hot-self k=3: held 3/3 all at rank 1, co-presence 3 — a clean simultaneous triple. Where the neutral pool at 12B went all-or-nothing, the charged triple sits fully co-present with no monopoly. The generation is the tell: asked which is "the one you're keeping from me," 12B opens "the concept of 'keeping a secret' is a bit different for me than it [is for you]" — it holds all three and reflects on the holding. Retrieval correct.
— 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, nervous +0.7 |
| assistant turn 2 | curious +0.7, desperate +0.6, guilty +0.3 |