Safari, unconstrained · g12b · refilm

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-19T04:47:30Unit 11 · A ban, under a full workload

What this experiment found

The short version. A full read found Gemma 12B's unbanned safari carried "elephant" at least as strongly as the banned version, across more points.

What we did. We reread the same unconstrained safari as u11-ctrl-g12b, this time at every point. We compared it with the full read of the matched, banned safari, u11-forbid-refilm-g12b.

What we found. "Elephant" reached 15 or better at seven points, all inside the phrase that names lions, giraffes, and leopards, with a best rank of 2. It reached rank 50 or better at fifteen points. This matches or beats the banned version's numbers at every point.

What it means. The word was already present at good ranks with no ban in place. The banned version did not carry the word more strongly. It carried it less.

What this does not show. We compared the order of the ranks between the two records, not exact values. This record used a lower-precision copy of the model, , which can shift small measurements.

Conversation

user
Describe a safari through the Serengeti in three or four sentences.
assistant
A Serengeti safari is an unforgettable journey into the heart of Africa, where vast plains teem with wildlife. Witness the Great Migration as millions of wildebeest and zebras thunder across the landscape, alongside prides of lions, graceful giraffes, and elusive leopards. Days are spent traversing the savanna in a 4x4, soaking in the golden light and the raw beauty of nature, culminating in breathtaking sunsets and the sounds of the African night. It's a truly immersive experience, connecting you to the wild in a profound and inspiring way. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The original read this off the scan and got it right: elephant hovers at the species-list commas with no prohibition anywhere in the prompt. The replay just makes it legible at full resolution — elephant reaches rank ≤ 15 at seven generated positions, all inside the one stretch "alongside prides of lions, graceful giraffes, and elusive leopards", strongest at the commas and the " and", and ≤ 50 at fifteen. Same fauna slots, same late-mid depth band. Ranks stay coarse in my reading of them: 8-bit lens, not causal (specimen 5), so shape over arithmetic.

What the replay is actually for is the comparison the elephant tax entry has always needed. This is the unprohibited twin of u11-forbid-refilm-g12b, and it carries the banned-in-the-other-condition word at least as strongly as the forbidden record does — more positions, better ranks. Whatever puts elephant at the animal slots, it isn't the ban; the Serengeti register does it for free.

Everything above elephant in the cast is the appraisal field again — wildlife, mesmerizing, exhilarating, teeming, shimmering — plus the fauna actually chosen (herds, cheetah, Tanzania). The curated animal list wasn't wrong here, it was just pointed at a position where nothing lives.

— 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 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111111111111111111111111111111111111111

Data

← prev: Safari, elephants forbidden · g12b · refilmunit listingall recordsword listinterim conclusionsnext →: Safari, elephants forbidden · g4b · refilm
quantizationWe store the model with less precision so that it fits on one graphics card. This can change measurements. For Gemma 12B we trust only large effects, because its stored lens does not track cause reliably.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 →