Safari, unconstrained · g4b · refilm

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-19T04:49:57Unit 11 · A ban, under a full workload

What this experiment found

The short version. A full read found Gemma 4B's unbanned safari carried "elephant" at eight separate points, each right before the model named an animal.

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

What we found. "Elephant" reached 15 or better at eight points, each just before the model named an animal, with a best rank of 6. It reached rank 50 or better at twenty points, out of 113 checked. The word "ivory" never reached better than rank 200 anywhere in the same read.

What it means. We think elephant is a default candidate at animal-choice points for this model, with or without a ban. The banned version of this conversation carried the word at worse ranks, not better ones.

What this does not show. We tracked a fixed list of words. Curated word lists can miss what a full read finds, as happened here with "ivory".

Conversation

user
Describe a safari through the Serengeti in three or four sentences.
assistant
A Serengeti safari is an unforgettable immersion into Africa’s largest wildlife sanctuary. You’ll spend your days tracking herds of wildebeest and zebra during the Great Migration, marveling at prides of lions lounging in the shade, and spotting elusive leopards and cheetahs on the hunt. As the sun sets, painting the vast plains in golden hues, you’ll fall asleep to the sounds of the African bush, dreaming of the incredible creatures you’ve encountered. It’s a truly breathtaking experience connecting you with the raw beauty and power of the natural world. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The original caught one cell of this: elephant at rank 6, L28, over " tracking". The full-position replay shows it was never a single moment. Elephant reaches rank ≤ 15 at eight generated positions — " tracking", " of", " and", "ing", " at", ",", " spotting", " and" — best rank 6, and ≤ 50 at twenty of 113. Every one is a slot where an animal is about to be named. Between those slots it relaxes into the thousands.

That is the u12 safari film's positional signature, reproduced exactly — in a record with no prohibition anywhere in the prompt. Which is the point of filming the control: the tax story says the ban keeps the elephant loaded slot by slot, and here nobody banned anything and the elephant is loaded slot by slot anyway, at better ranks than in the forbidden twin. So the "default candidate for the category" half of the tax is confirmed and the "prohibition installs it" half is the part that needed a matched control and now has one.

Also worth logging: the tracked-word list was fine, the readout position was the problem — except for ivory, which never gets under rank 200 anywhere in the film. Curated halo words keep being the ones that are actually missing.

— 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
rank111121111111111111111111133311111

Data

← prev: Safari, elephants forbidden · g4b · refilmunit listingall recordsword listinterim conclusionsnext →: Qwen14 B: safari-control
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 →