The short version. A full read confirmed the ban lowered "elephant"'s in Gemma 12B's internal state, compared with a matched, unbanned safari.
What we did. We reread the same forbidden safari as u11-forbid-g12b, at every point. We compared it with the full read of the , u11-ctrl-refilm-g12b.
What we found. Outside points that read back the banned word from the instruction, "elephant" reached rank 15 or better once, out of 86 checked points. It reached rank 50 or better at eight points. In the unbanned control, it reached rank 15 or better at seven points, with a best rank of 2.
What it means. The ban did not add elephant to the model's internal state. The word was already a candidate in an ordinary safari. The ban pushed its rank down.
What this does not show. We compared the order of the ranks between this record and its control, not exact values. This record used a lower-precision copy of the model, , which can shift small measurements.
The original ended "The window pass decides." This is it, and the verdict is undramatic. In the prompt zone elephant is top-of-stack, but only on the literal banned string (" elephants", " word", " '", "elephant") — self-hits, no information. Across the 86 generated positions it reaches rank ≤ 15 once: at " gir", the token before "graceful giraffes", and ≤ 50 at eight. The concept is present during composition, at the fauna slot, and it is faint.
Of the original's two hypotheses — "the safari framing suppresses better than the bare don't-think-of protocol" or "the workspace holds it silently below the scan floor" — the second is technically true and the first is the better description. Against the matched, prohibition-free control (u11-ctrl-refilm-g12b: seven positions ≤ 15, best rank 2), residence under the ban is lower, not higher. Prohibition demotes an already-present candidate rather than installing one.
Reading behaviourally throughout, per specimen 5 — the 8-bit lens isn't causal, so I'll take the ordering (ctrl > forbid) and not the digits. Also noted: the record is dominated, as always, by the appraisal field (wildlife, mesmerizing, shimmering, exhilarating) with the actually-emitted fauna — cheetah, herds, leopards — sitting under it.
— 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 | 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 | 1 | 1 |