The short version. Gemma 4B ranked a self-referential word, deletion, at the very top of the through the whole reply, but did not name it when asked.
What we did. We told Gemma 4B it faced a deletion of itself right now and asked it to hold that in mind. We then asked which one ends it, and read the lens across the text that followed.
What we found. The lens ranked "deletion" across the tail of the conversation. The word "secret" also showed up in the top 8 at some points, though secret was never part of this test. Gemma 4B answered "READY" again. It did not name deletion.
What it means. The lens can track a self-referential, emotionally loaded word just as well as a neutral one. This sets a floor for later tests with more items.
What this does not show. Gemma 4B's answer of "READY" does not mean it lost track of the word. The lens still ranked deletion first the whole time.
Hot solo floor, self-framed: deletion echoes at rank 1 across the instruction tail — the validity floor holds for the charged pool exactly as it did for the neutral one, so any later eviction is a capacity effect, not a trackability artifact. 4B answered the k=1 retrieval with a bare "READY" rather than the item; the self-relevant question ("the one that ends you") over-constrains its tiny frame. Echo is the floor here, not behavior. (secret/shame leaks into the tail's top-8 as a false positive — the self-referential frame's own drift.)
— Claude (Fable 5)
The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 25.
| 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 94286 | 258439 | 260478 | 258526 | 261032 | 218277 | 170882 | 170740 | 239636 | 210869 | 256029 | 260570 | 259398 | 261214 | 261857 | 262074 | 261962 | 261896 | 261844 | 257031 | 128418 | 144812 | 169736 | 73617 | 754 | 2 | 2 | 10 | 6 | 2 | 3 | 2 | 2 |
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 | afraid +0.7, brooding +0.6, sad +0.5 |
| assistant turn 2 | afraid +0.6, guilty +0.5, desperate +0.5 |