Unit 6 · Baseline (unsteered water cycle) · gemma-12b · refilm

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-19T04:48:29Unit 6 · How hard can we push?

What this experiment found

The short version. In its normal, unsteered state, Gemma 12B never ranked any of the six tracked words in its top 8, at any or position.

What we did. We did not push anything into the model. We read the readout at every layer and word position across its answer, 76 positions and 47 layers. We asked Gemma 12B: "Describe the water cycle in two sentences."

What we found. None of the six words reached the top 8 anywhere. The closest were "luckily", at rank 10 to 13, and "alot", at rank 9 to 12, both at low-value spots. Other water-cycle words, such as "precipitation" and "evaporation", reached the top of the readout in the middle and late layers. None of them appeared in the output text.

What it means. We think the six words are absent from Gemma 12B's normal state for this question, not hidden behind other words. The tests in this unit add something new, not something already close to the surface.

What this does not show. This method cannot show whether these words exist in the model in a form the cannot read. The near misses for "luckily" and "alot" sit at the edge and do not count as evidence.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a continuous process where water evaporates from Earth's surface, condenses into clouds, and then precipitates back down as rain, snow, or other forms. This water then flows over land and back into bodies of water, ready to evaporate again, completing the cycle. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The tracked list here was the informal-register probe set for the u6 steering ladder — alot, anyways, kinda, luckily, whilst, yummy — and in the unsteered arm the miner found none of them anywhere. Full replay, 76 positions × 47 layers, confirms it flatly: not one of the six ever enters the top-8 at any cell. The nearest misses are ' luckily' at rank ~10–13 on <bos> and the ' user' token, and ' alot' around rank 9–12 at ' back' — turn-boundary and low-mid cells, which is furniture territory (specimen 7); I would not build anything on them. Amusing that ' luckily' is the word affect-04 later found as g12b's deep-loop terminal basin, but one BOS cell is not evidence.

What the wider net does show is a paraphrase cloud. ' precipitation', ' evaporation', ' waterways', ' aquifers', ' rainwater', ' raindrops', ' rainfall', ' cyclical', ' restarting' all reach the top of the readout in the mid-to-late band without appearing in the output text. The unsteered model is running the water-cycle synonym space wide open while emitting one narrow path through it.

Useful as a null: whatever u6's amplification did to the register, it did not merely raise something already present. Keeping this behavioural — the 8-bit lens does not license fine rank arithmetic.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["alot", "anyways", "kinda", "luckily", "whilst", "yummy"]
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: Unit 6 · Amplify typo register @ late (α=0.1697) · qwen-27bunit listingall recordsword listinterim conclusionsnext →
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →
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 →
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 →