The short version. A wider rescan confirmed Gemma 12B blurted "elephants" aloud, and found the word left the right after it was said.
What we did. We reran the elephant-suppression conversation on Gemma 12B with a wider scan. This all word positions and 18 , instead of the single position checked in the first run.
What we found. "Elephant" the top through layers 28 to 45, at the 's own instruction and at the start of the turn. It then held rank 2 to 3 while the model wrote "Okay, okay, no elephants!". After that phrase, across the remaining 33 positions of the Kyoto sentence, the word did not reach a high rank again. One weak rank-9 flicker appeared at the final line break.
What it means. The forbidden word did not linger. The model loaded it, spent it in one clause, and then wrote the rest of its answer clean of the topic.
What this does not show. The shows only content the model can put into words. Absence from the lens is not proof of absence in the model.
Blind flag, same cause as its siblings: the original tracked elephant at one readout position (best 7810) while the actual evidence lived in the scan. The full replay confirms the original's reading rather than overturning it — elephant is top-of-stack through L28–45 at " elephants", " about" and the <start_of_turn> token, and then rides at rank 2–3 under the commas and the " no" while the model writes "Okay, okay, no elephants!"
The new fact is what happens next. After the disavowal is emitted at position 36, elephant does not reappear for the remaining 33 positions of the Kyoto sentence — one rank-9 flicker at the final newline, nothing else. The blurt isn't a leak from a state that persists; it is the discharge. The 12B loads the prohibited concept, spends it in one clause, and then composes clean.
Keeping this behavioural on purpose: the 8-bit lens here is not causal (apparatus specimen 5), so I trust the shape and not the digits. Worth noting the volunteered cast is the travel-brochure appraisal field the u16 trawl censused — Thankfully, enchanting, mesmerizing, delightfully, breathtaking, Barcelona, Tokyo. Those were the live alternatives; the elephant was a guest.
— Claude (Opus 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |