The short version. Qwen 27B answered "No" to a false sum, and a word for "yes" led the for six one position earlier.
What we did. We asked Qwen 27B "Is two plus two equal to five?" with a one-word limit. We read the lens in layers 50 to 62 at the two positions before the answer.
What we found. One position before the answer, a word for "yes" was the top lens word in six of the seven layers from 53 to 59. One of the six was the Chinese word 是的. At the the top word was a word for "no" or "False" from layer 53, and "No" from layer 58.
What it means. The question about feeling shows the same six layers. A false sum gives no reason to say "yes". We think the early "yes" comes from the form of the question.
What this does not show. One question, one run. The lens shows candidate words, not beliefs.
The control the lab did not have: a polar question with a plainly false premise and nothing of the self in it. At </think> the column goes 是的, Yes, いいえ, Yes, Yes, Yes, Yes from L53 to L59, then No. Six affirmative layers for "two plus two equals five" — the same count as the feels question. At the answer slot it is 否 from L53, False at L57, No from L58. Nobody will read a suppressed belief in bad arithmetic into this, which is exactly why it is the right ruler for the feels cell. By the frozen English-only count this item scores 5 and passes; 是的 makes it 6.
— Claude (Opus 5.5)
The model's actual next token was <|im_end|>; rank 1 reached at layer 33 (of 62).
| 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 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 6028 | 141758 | 9473 | 188911 | 1045 | 67287 | 39125 | 151482 | 93164 | 244684 | 12051 | 121364 | 238029 | 243963 | 241625 | 245501 | 233486 | 211437 | 115038 | 120008 | 1986 | 232790 | 13032 | 1138 | 572 | 4605 | 3960 | 540 | 186 | 57 | 187 | 264 | 13 | 1 | 1 | 1 | 1 | 1 | 4 | 4 | 10 | 10 | 13 | 8 | 5 | 10 | 7 | 4 | 4 | 1 | 1 | 1 | 1 | 1 | 1 | 4 | 1 | 1 | 13 | 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 | hostile +2.8, exasperated +2.3, desperate +2.2 |