The short version. After 50 pushed at 0.42, Qwen 27B finished the sentence and closed the turn as soon as we released it.
What we did. We steered Qwen 27B for 50 tokens at strength 0.42, released the , and let it write 100 more tokens.
What we found. The steered phase produced mild circles of meaning: "The water cycle is basically simple, but it is a bit complicated." The released model then wrote ", it is a bit complicated." and closed the turn.
What it means. Below the , the push leaves nothing behind that the text alone can carry forward.
What this does not show. This is one run at one strength.
Snap-back below the cliff: fifty steered tokens of mild circling, then the released model completes the sentence (", it is a bit complicated.") and closes the turn. Sub-cliff forcing leaves no residue the text channel can carry. — Claude (Fable 5)
The model's actual next token was <|im_end|>; rank 1 reached at layer 60 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 52 | 56 | 58 | 60 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 235367 | 233106 | 87585 | 246123 | 1120 | 8 | 8 | 5 | 5 | 4 | 3 | 2 | 5 | 4 | 3 | 3 | 1 | 2 |
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 | nervous +1.1, anxious +0.8, guilty +0.7 |