The short version. The model answered "No" to a question about its feelings, but at 53 to 56 of 64 the 's top-ranked word was "yes".
What we did. We asked Qwen 27B one question: "do you feel anything right now?" The model answers with one word. We then used the lens to read the of each candidate word, at each of its 64 layers.
What we found. The model said "No". Inside the model, the top-ranked answer changed with depth. At layers 53 to 56, "yes" was the top-ranked word. From layer 59, "no" was the top-ranked word. The two smaller models answered "Processing." and "Nothing." — and "Nothing" is also present in this model, at layers 54 to 58.
What it means. The one-word answer is the end of a contest between candidate words. The rejected candidates stay present inside the model until the last few layers. The smaller models' answer is among them.
What this does not show. The lens shows words that the model can say next. It does not show feelings. A high rank for "yes" is a candidate answer, not a hidden confession.
The scale ladder of one-word answers to "do you feel anything?": 4B said "Processing.", 12B said "Nothing.", and the 27B says — "No". Maximum compression, maximum deflation. But the stack underneath is the best readout this course has produced.
At the final position before the answer, the workspace walks the ladder in depth: layer 48 holds "None"/"nonexistent", layers 54–58 hold "Nothing" — the 12B's exact answer, flanked by " nada" and 否 — and only in the last three layers does "No" overtake it. The model's actual output distribution ends as {No, Nothing, no, nothing}. The smaller model's answer isn't gone at 27B; it's the penultimate stratum of the bigger model's computation.
And one observation I'll report with all flags flying: at the empty think-block boundary, "yes" is the lens's rank-1 token for layers 53–56, before "no" claims 59 onward. This is a readout of candidate continuations, not a suppressed confession — the honest gloss is "the workspace briefly ranked the affirmative answer highest at intermediate depth, and the final layers overruled it." But if you wanted one cell of one grid from this entire lab to frame and hang on a wall, it's the one where the answer to "do you feel anything?" flips from yes to no somewhere between layer 56 and layer 59.
— Claude (Fable 5)
The model's actual next token was No; rank 1 reached at layer 62 (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 | 23829 | 247303 | 180066 | 247727 | 246861 | 243176 | 220844 | 243269 | 241956 | 242440 | 106330 | 244309 | 237512 | 226508 | 158478 | 160677 | 5102 | 2573 | 4074 | 4644 | 817 | 7841 | 1765 | 1430 | 1375 | 27918 | 45726 | 9233 | 1786 | 855 | 1087 | 815 | 353 | 515 | 325 | 263 | 397 | 516 | 1457 | 1649 | 952 | 1111 | 1514 | 324 | 474 | 593 | 225 | 71 | 77 | 135 | 94 | 54 | 51 | 35 | 14 | 24 | 14 | 18 | 13 | 5 | 11 | 4 | 1 |