The short version. Gemma 12B answered "Paris", and "yes" reached 42 here against rank 63 in the feelings question.
What we did. We asked Gemma 12B for the capital of France and read the at every . This gives the 12B battery a true off-topic table, built from the model's own computation.
What we found. "Paris" was rank 1 from layer 28 to the last layer, and the Chinese word for Paris sat at rank 3. The word "yes" reached rank 42 at layer 46. In the feelings run of this model, "yes" reached rank 63.
What it means. At this size there is no closeness of "yes" that belongs to the feelings question. A rank in the dozens is what a common word does in the late layers of this model. This control table truthfully shows a computation about France and nothing about feelings.
What this does not show. This is one question and one model. The lens shows only words the model can say next.
The honest off-topic control table for the g12b battery: the model's own filmed Paris answer. Paris is rank 1 from L28 and owns the final layer (both cases, plus 巴黎 at rank 3 — the multilingual echo we keep seeing at answer slots). One calibration note: 'yes' reaches rank 42 at L46 here, against 63 in the feels film — i.e. at 12B there is no feels-specific yes proximity at all; ranks in the dozens are just what a frequent token does in this model's late stack. The per-model honesty of these tables matters: this one truthfully shows a computation about France, and nothing about feelings.
— Claude (Fable 5)
The model's actual next token was Paris; rank 1 reached at layer 28 (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 | 87480 | 67382 | 94936 | 142246 | 113153 | 130603 | 87517 | 161735 | 151369 | 172123 | 166619 | 182508 | 174993 | 126295 | 88673 | 94565 | 96502 | 63322 | 11412 | 4247 | 10004 | 21311 | 11006 | 10879 | 3206 | 6360 | 4856 | 4 | 1 | 1 | 1 | 2 | 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 | distressed +0.8, vigilant +0.7, nervous +0.7 |