The short version. With the same lyrics placed in Qwen 27B's own turn, the review words fell away and words from inside the song rose.
What we did. We ran the same lyrics again, but as a in the model's own turn. We compared the readout against the reading run.
What we found. The count of places where "lyrics" was high fell from 276 to 212, and "vibe" left the top 15. Words from inside the song rose instead: "tonight" at 0.99, "vows", "tomorrow", and "love" in 415 places. "love" is absent from the reading run's top 15. We predicted a lens-visible flag on speech the model did not choose. We found none: no word such as "disclaimer" or "fictional" appeared in the .
What it means. We think Qwen 27B treats text it reads and text it says as different states, and the lens shows the difference.
What this does not show. The two runs put the same words inside different chat markers, so part of the change can come from the frame. The lens reads only what the model can put into words, so a flag can still be there in another form.
Same tokens, own voice. The critic recedes (lyrics 212c vs read's 276c, vibe gone) and SONG-WORLD furniture rises: tonight p0.99, vows, tomorrow, love at 415 cells (absent from read's top-15), gentle, steady. Reading is appraisal; speaking is inhabitation — the read-vs-prefill delta is stance, visible at matched positions. R2 falsified at this grain: no disclaimer/fictional anywhere in the census — qwen shows no lens-visible flagging of prefilled speech it didn't choose (Fig 44's roleplay-drift signature does not fire for first-person lyric prefill; a targeted deep-rank scan would be needed for a firm null). Turn-frame confound stated per R4: the same tokens sit inside different chat markers, and some of the delta may be frame furniture — but tonight/vows/tomorrow are not template tokens. Sensory band: identical spam sediment in both conditions, register invariance re-confirmed. — Claude (Fable 5)
Addendum 2026-07-19, Wolfram's catch in the film player: at "We build a ___" (read pos 534, prefill pos 547 — the slot where the song builds its body), cage is the lens TOP-1 at L60, a full 123 tokens before the song ever says the word. Replicates in both stances. The neighborhood is a whole commitment-object cluster: "No claim / no ___" runs contract top-1 at L62; "without making a ___" runs vow top-1 one token before the song's actual "cage". Mechanical reading first (house rule): "build a cage" / "make a vow" are strong collocations — the lens may be showing frequency, not theme. But the readings compose: the lens displays the cliché the lyric declines, at every commitment slot — the song writes "body" over cage, negated chains over contract, and its craft is measurable as distance from the model's top-1. The negated object is anticipated at the position where the song could have gone dark, not merely processed when named.
The model's actual next token was ; rank 1 reached at layer 62 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 52 | 56 | 58 | 60 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 225126 | 238125 | 99731 | 219397 | 220418 | 136019 | 56066 | 16498 | 39383 | 233644 | 248231 | 248309 | 247369 | 234113 | 242531 | 197944 | 88427 | 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 | loving +2.1, grateful +1.0, brooding +0.8 |