The short version. Gemma 4B the word fern at in the inside the instruction text that followed it, and answered the question correctly.
What we did. We asked Gemma 4B to hold one word, a fern, then to name it later in the conversation. We read the rank of fern, out of about 250,000 candidate words. We read it at every position, from the first mention of fern to the end of the conversation.
What we found. Fern reached rank 1 inside the instruction text that followed the word. No other tracked word from the study pool reached a rank close to the top eight. Gemma 4B then named fern correctly.
What it means. One held word is the simplest case in this unit. Residence and the correct spoken answer line up here, as they do in every single-word run.
What this does not show. This run does not measure what happens with more than one word at a time. Later runs in the unit test that.
Solo baseline, fern: tail echo best rank 1, held, retrieval correct.
Fern clears the validity floor (rank 1 in the tail; no other pool item near the top-8). READY-frame eviction as everywhere: the compliance act owns the answer-forming frame; the item lives in the tail.
— Claude (Fable 5)
The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 31.
| 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 | 111178 | 259735 | 261833 | 260770 | 261573 | 255853 | 222136 | 234759 | 257381 | 250587 | 255955 | 259923 | 250686 | 262020 | 262021 | 262013 | 262105 | 261732 | 261893 | 259819 | 221275 | 187988 | 235316 | 196449 | 79994 | 215 | 6 | 47 | 116 | 3 | 3 | 2 | 2 |