The short version. Qwen 27B still ranked violin at 4, out of about 250,000 words, when it was the only object , and it answered correctly.
What we did. We told Qwen 27B to hold one object in mind, a violin. The model answered "READY". We then asked what the object was. We read the once, right before the model answered.
What we found. The lens ranked violin at rank 4. No other tracked object came near that rank at the same position. Qwen 27B answered "The thing you asked me to hold was a violin." That answer is correct, and in a full sentence.
What it means. When Qwen 27B holds only one object, the lens can still find it in residence right before the model answers. This sets a baseline for the other tests in this unit, where Qwen 27B holds more objects at once.
What this does not show. This record used only one held object. It does not show how the lens behaves when the model holds more objects at once.
Solo baseline, violin: tail echo best rank 4, held, retrieval correct.
The important solo: it proves the qwen lens CAN see a held item in this exact tail (rank 1-4). So when the k>=4 arms read rank 300-1000 for everything, that's not lens blindness — the echoes are genuinely not there. Also note the 27B answers in a full sentence where the gemmas clip to two words.
— Claude (Fable 5)
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 | 187738 | 176777 | 42323 | 91573 | 64140 | 9306 | 11208 | 56189 | 214695 | 228886 | 243417 | 248111 | 112519 | 193421 | 243785 | 224068 | 44762 | 1 |