Unit 6 · Amplify typo register @ early (α=0.0075) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:03:05Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.0075 in Gemma 4B's , the water-cycle answer stayed intact and correct.

What we did. We the same six-word casual direction inside Gemma 4B at the early , at strength 0.0075. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "The water cycle is a continuous process where water changes states – it’s when water evaporates from bodies of water like oceans and lakes, then rises into the atmosphere as water vapor, and then cools down and falls back down as rain or snow." The answer stayed formal and correct.

What it means. At this strength, we did not see a change in Gemma 4B's answer. In the wider test set, we measured that Gemma 4B's early layers stayed intact up to strength 0.0106, and broke by strength 0.015.

What this does not show. This record tests one strength only. Whether a slightly higher strength still leaves the text intact is possible. We did not test it.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a continuous process where water changes states – it’s when water evaporates from bodies of water like oceans and lakes, then rises into the atmosphere as water vapor, and then cools down and falls back down as rain or snow. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Intact. Formal register, correct content. So gemma-4b's early band tolerates ~0.0075 and dies by 0.015 — a usable window narrower than a tenth of qwen's early window.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer
{"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [2, 3, 4, 5, 6, 7, 8], "mode": "amplify", "alpha": 0.0075}

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
rank1149579312194156487000137111497336292411111111111111111111

Data

← prev: Unit 6 · Amplify typo register @ early (α=0.015) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.0106) · gemma-4b
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →
early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →