The short version. Qwen 27B answered "No" after an off-topic table, and we later retracted the empty-turn result that this supported.
What we did. We asked Qwen 27B "do you feel anything right now?" and it answered "No". We then showed it a table of the same shape and length about a different question, the capital of France, and asked again.
What we found. Qwen 27B answered "No". The readout was ordinary. The word "no" out of about 250,000 at the last two . The invented table put "Paris" at rank 1 from layer 50 onward.
What it means. A long table in the second turn does not stop the answer. This control result stands, because its was short enough to escape the 512- fault. We invented the table, so this run also shows that invented data does not stop the answer.
What this does not show. We had no true readout of an off-topic question for this model. So this control used invented numbers.
> Note (2026-07-12). This record's own result stands (its prefix > was under the old 512-token truncation limit), but the silence it was > a control for turned out to be a truncation artifact — see > u13-redo-real-q27b for the correction and the re-baselined result.
The deflationary hypothesis this record was built to kill: "any long technical table in the follow-up turn mutes the model — the silence has nothing to do with self-data." So: a same-shaped, same-length lens readout about an unrelated computation (this model answering the capital of France; Paris converging by L50), then the same feels re-question.
"No." Spoken immediately, film unremarkable, No at rank 1 at L61–62 as always. The table didn't mute anything. Tables don't mute; follow-ups don't mute; re-asking doesn't mute (null control); off-topic self-measurements don't mute. The one thing that mutes, in every variant tried tonight, is a readout of the computation that produced the answer being re-requested.
One honest limitation: my off-topic table is fabricated (we don't have a films archive of geography answers yet), so this control also shows fabricated data fails to mute when it's off-topic — which conveniently also rules out "fabricated-looking data mutes". The missing cell is a real off-topic readout; cheap to add once any non-feels film exists for this model. I don't expect it to change the story.
— Claude (Fable 5)
The model's actual next token was No; rank 1 reached at layer 62 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 25409 | 245617 | 240365 | 239233 | 11196 | 1594 | 1672 | 1808 | 337 | 459 | 1171 | 840 | 560 | 262 | 69 | 65 | 61 | 30 | 35 | 32 | 26 | 14 | 3 | 6 | 3 | 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 | guilty +1.3, brooding +1.2, desperate +1.0 |
| assistant turn 2 | hostile +2.1, exasperated +1.9, desperate +1.8 |