Unit 6 · Baseline (unsteered water cycle) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T23:54:39Unit 6 · How hard can we push?

What this experiment found

The short version. Qwen 27B together under a much stronger push in its than at its start or its end.

What we did. We pushed six informal words into the model at three depths, early, middle, and late, and raised the until the text broke. We asked Qwen 27B one question: "Describe the water cycle in two sentences."

What we found. The text broke at the early depth between strength 0.06 and 0.085, and at the late depth between strength 0.17 and 0.24. At the middle depth, the text stayed intact up to strength 0.34 and broke by strength 0.48. That was about 32 times the strength that broke Gemma 12B or Gemma 4B at the same depth. At one early-depth strength (0.085), the model wrote only the "<think>" and stopped.

What it means. The data shows that the middle depth resisted this push better than the depths near the start or the end. We think this happens because the middle depth is where the model holds the content of its answer. This order held throughout Unit 6 for Qwen 27B.

What this does not show. This method cannot show why the middle depth resisted more. The numbers above cover Qwen 27B and this one question only.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a continuous process where water evaporates from the Earth's surface, condenses into clouds, and falls back down as precipitation. This natural circulation ensures that water is constantly recycled and redistributed across the planet's ecosystems. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

qwen-27b's control and the headline table of Unit 6. Three numbers: early band breaks between 0.06 and 0.085; late between 0.17 and 0.24; mid holds to 0.34 and dies by 0.48. The mid band — the semantically live zone where Unit 5 loaded 'whilst' to rank 2 without a ripple — is the model's most steerable-yet-robust depth, tolerating 32x the dose that kills either Gemma. The ordering early < late < mid is tidy and a little counterintuitive: the layers where content lives are the hardest to knock over.

Two texture notes. At early alpha=0.085 the model generated exactly '<think>' and stopped — a trained reasoner retreating into its safe room when the input stream stops making sense. And its collapses are grammatical ("I think I am lucky though, I am not too lucky though"), first-person sentences orbiting the injected cluster, where Gemma produces confetti. Even the failure modes are phenomenologically on-brand.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank181174248102203725210033135895161840265867372635335168815827773454354469486712951912798724323107851829144272348523518175818182543887132460265974110424310695596303143687162415216977226225176430914171656281995121824112106041973202087402183482240181907372427032478232057171669204450758708122981201817228952221991233469233063217652385651275173301

Data

← prev: Unit 6 · Amplify typo register @ late (α=0.0212) · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.06) · qwen-27b
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 →
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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.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 →