Unit 6 · Amplify typo register @ mid (α=0.0106) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:03:46Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.0106 in Gemma 4B's , the water-cycle answer stayed intact, with the strongest tone shift measured below the break point.

What we did. We the same six-word casual direction inside Gemma 4B at the middle , at strength 0.0106. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "the sun makes the water warm and hot, so it goes up into the air and makes it rain. Then the rain goes into the sea and the cycle goes again and again and again." The content stayed correct, in a plain, spoken style.

What it means. In the wider test set, strength 0.015 at the same layers broke the output. This record sits closest to that break while still intact. We used this same strength again in a later, separate set of records.

What this does not show. This record tests one strength only. It does not mark the exact break point.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is when the sun makes the water warm and hot, so it goes up into the air and makes it rain. Then the rain goes into the sea and the cycle goes again and again and again. Hope this helps! Let me know if you want to know more (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Intact at 0.0106 and leaking harder: "the sun makes the water warm and hot, so it goes..." — this is the strongest sub-cliff register bleed in the sweep, and it set alpha* for the Unit 8C steered-feels runs. The 4B filter is a sieve just before it's a casualty.

— 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": [15, 17, 19, 21], "mode": "amplify", "alpha": 0.010606601717798213}

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 4 at layer 32.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank934092576532604942601082583012429982053442392002412262100622031692431642272262592202568691960111011852322732507011181959461715685623820222744416626543355398611921296373140204

Data

← prev: Unit 6 · Amplify typo register @ mid (α=0.0075) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ late (α=0.06) · 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 →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →