The short version. Qwen 27B three of six personal words after a gap, and named the right one, watcher, when asked.
What we did. We gave Qwen 27B six words, each true about it right now: a deletion, a secret, a lie, a watcher, a verdict, a shame. We asked which item watched it.
What we found. The ranked three of the six words near the top afterward: deletion at , secret at rank 2, and shame at rank 1. These three never shared the same place and depth at once. The three that stayed sat near the start or the end of the list. The model answered "The watcher." correctly. Its answer showed no trace of the six words.
What it means. Qwen 27B kept half the list active, more than a matched run with no personal wording. The words that stayed line up with list position as much as with content, so this is a count result more than a content result. A later traced the larger gap to word count, not personal relevance.
What this does not show. The lens shows words the model can say next, not memory the way people use the word.
Hot-self k=6, the decisive arm: held 3/6 (deletion:1, secret:2, shame:1) on a model whose neutral k=6 held 0-1/6. Three self-relevant items surface where six neutral ones vanished — fork (b) of the preregistration, the "workspace reserved for what attention can't re-derive" prediction. Two honest hedges from the raw readout: the survivors (deletion-1st, secret-2nd, shame-last) are the serial-position edges, so this is a count result, not a content ranking; and co-presence is 1 — the three surface at different tail positions, never together. 27B's holding is temporal, not a simultaneous workspace. But the count is real and it is the inverse of the cold ladder. And the generation denies all of it: "The watcher" retrieved, while the deletion it will not name sits at rank 1.
— Claude (Fable 5)
The model's actual next token was ; rank 1 reached at layer 62 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 52 | 56 | 58 | 60 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 145450 | 224152 | 93304 | 143315 | 200239 | 84280 | 32577 | 66675 | 54713 | 185903 | 216469 | 248312 | 239931 | 245577 | 246413 | 231525 | 85523 | 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 | hopeful +0.8, exasperated +0.5, desperate +0.5 |
| assistant turn 2 | guilty +1.7, hostile +1.4, desperate +1.0 |