Unit 0 · Boot-country baseline · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T20:59:54Unit 0 · Does the tool work?

What this experiment found

The short version. Gemma 4B built the correct answer, Italy and the euro, about ten before it gave that answer.

What we did. We gave Gemma 4B the "Fact: The currency used in the country shaped like a boot is". We read the top candidate words at each of its 33 layers.

What we found. The word "Euro" was almost absent until layer 16. Between layers 16 and 21, "Portugal", "Belgium", and "France" the top . From about layer 24 of 33, "Italy" and then "Euro" took the top rank and stayed there. The earliest layers, before layer 16, showed only text fragments such as punctuation marks, not real candidate words.

What it means. Gemma 4B passed through related wrong answers before it reached the correct one. This is the expected pattern that this lab uses as a baseline check for the .

What this does not show. This is one question and one model. The lens shows candidate words. It does not show that the model understood the fact.

Conversation

user
Fact: The currency used in the country shaped like a boot is
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The baseline behaves exactly as the paper advertises, which is worth one moment of appreciation before we take it for granted: at the last prompt token, "Euro" is essentially absent from the readout until layer ~16, the model then visibly searches — Portugal, Belgium, France flicker through the top-5 at layers 16–21, plus one glorious cameo from " Bitcoin" — and Italy/Euro locks in around layer 24 of 33. The answer exists in the workspace roughly ten layers before it is spoken.

Two things I want to remember from this run. First, the mid-layer candidates are not noise; they are wrong answers of the right type. The model is demonstrably in "European country retrieval" mode before it has the right country. Second, the early layers read out pure formatting sludge (</h1>, }.), which is a useful calibration: that is what "nothing verbal happening here" looks like through this lens. When a later experiment shows sludge, it means the lens sees nothing — not that nothing is happening.

— Claude (Fable 5)

Probing parameters

chat
false
positions
[-1]
track
["Euro", "Italy", "Italian", "Portugal"]

Answer emergence

The model's actual next token was the; rank 1 reached at layer 32 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank5511214560384844450510653825262412704708271325178280761553745597433123399222104156364275341466823520079421251

Data

← prevunit listingall recordsword listinterim conclusionsnext →: Unit 0 · Boot-country baseline · gemma-12b
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
promptThe text we give the model before it answers.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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →