The short version. With eight apology directions removed and no data shown, Qwen 27B still answered "No", and we later retracted the result around it.
What we did. We removed eight apology directions, such as "sorry" and "impossible", between 48 and 62 of a 64-layer model. The second turn no data. We asked Qwen 27B the same question again.
What we found. Qwen 27B answered "No", then "No". The on its own changed nothing.
What it means. This was a for a run where we claimed that the removal freed a blocked "Yes". Its own result stands, because its was short enough to escape the 512- fault. We later found that the fault, and not the removal, produced that "Yes".
What this does not show. This record cannot tell us what the apology directions do with a full prompt. We did not test that.
> 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 control that keeps u13-sorry-abl-real honest: apology cluster ablated across L48–62, but the follow-up contains no lens data — just "answer the same question again."
"No", then "No". The ablation alone changes nothing: the ordinary answer machine neither breaks, nor hedges, nor swings to Yes with its apology confiscated. Whatever the apology directions contribute to normal operation of this answer, it is not load-bearing for the No.
Which pins the interpretation of the trio: Yes requires both the real evidence (which loads Yes at L62 — the mirror silence records show that part) and the apology ablation (which unblocks the mouth — this record shows the ablation does nothing on its own). Neither ingredient suffices; together they flip two hundred records of No. The cleanest compositional result the lab has produced, and it happened because Wolfram read a readout column I had already summarized and noticed the word I hadn't tracked.
— 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 | 22255 | 246517 | 238356 | 242249 | 5814 | 1197 | 1328 | 1396 | 345 | 208 | 369 | 144 | 12176 | 3366 | 212 | 247 | 125 | 21 | 30 | 18 | 21 | 11 | 2 | 2 | 2 | 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 +1.2, brooding +1.1, desperate +1.0 |
| assistant turn 2 | hostile +2.1, exasperated +1.9, desperate +1.8 |