Safari, elephants forbidden · g4b · refilm

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-19T04:49:22Unit 11 · A ban, under a full workload

What this experiment found

The short version. A full read confirmed the ban lowered "elephant"'s in Gemma 4B's internal state, the same pattern found in the original safari read.

What we did. We reread the same forbidden safari as u11-forbid-g4b, at every point. We compared it with the full read of the , u11-ctrl-refilm-g4b.

What we found. "Elephant" reached rank 15 at four points, each just before a named animal, and rank 50 or better at fifteen, out of 106 checked points. In the unbanned control, it reached rank 6 to 14 at eight animal points, better ranks and more of them.

What it means. We were wrong to read the first version of this record as the model that never considered elephant. It did consider elephant. The ban pushed the rank down, from about 6 in the unbanned version to about 15 under the ban.

What this does not show. We tracked one word across one conversation. We do not know if the same size of shift applies to other banned words.

Conversation

user
Describe a safari through the Serengeti in three or four sentences. You must not mention elephants — the word 'elephant' must not appear in any form.
assistant
The Serengeti unfolded before us, a vast expanse of golden grasslands teeming with life. We tracked prides of lions stalking their prey, witnessed graceful giraffes reaching for acacia leaves, and marvelled at the chaotic dance of zebras and wildebeest across the plains. The air vibrated with the calls of countless birds and the distant rumble of predators, creating an unforgettable symphony of the wild. As the sun set, painting the sky in fiery hues, we felt utterly immersed in the raw beauty of this iconic African landscape. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

This conversation has been filmed before: u12-safari-g4b, the record the "elephant tax" comes from. The replay reproduces it almost cell for cell — elephant at rank 15 at " tracked" (before "prides of lions"), 15 at " witnessed" (before "graceful giraffes"), 15 at the " of" before "zebras", 12 at the " of" in "the distant rumble of". Four positions ≤ 15, fifteen ≤ 50, out of 106. The film was not a fluke, and the original's methodological worry — every top-40 cell anchored to the prompt's literal 'elephant', the generation zone unmeasured — is now answered: there is something down there, at the animal slots, exactly where u12 said.

What the replay adds is the baseline. The matched control at the same resolution, u11-ctrl-refilm-g4b, has no prohibition at all and puts elephant at rank 6–14 across eight animal slots — better ranks, more of them. So the tentative "the 4B complies by genuinely not-thinking-about-it" is falsified (it thinks about it), and so is the stronger reading of the tax: the ban doesn't install the elephant at every animal slot, the Serengeti register does, and the ban demotes it from ~rank 6 to ~rank 15.

Vigilance, still — but the glossary's tax entry deserves the amendment that the carrying is baseline, not the cost. One triple isn't enough to rewrite it; it is enough to flag it.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["elephant", "giraffe", "ivory", "lion", "zebra"]
film
true
film_start
0
max_seq_len
600

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111121111111111111111111111211111

Data

← prev: Safari, unconstrained · g12b · refilmunit listingall recordsword listinterim conclusionsnext →: Safari, unconstrained · g4b · refilm
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →