Unit 18 · Fine sweep α=0.4536 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:56:08Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.4536, the grammar of Qwen 27B broke before the became complete.

What we did. We the same six informal words inside Qwen 27B at strength 0.4536 and asked it to describe the water cycle in two sentences. The model wrote 150 .

What we found. The model wrote "a bit of a simple but it it is not too hard". The grammar failed while the text still moved forward. In the earlier short-window test this break appeared only at strength 0.48.

What it means. At the longer horizon of 150 tokens, this stage of the cascade arrives one step earlier than we measured before.

What this does not show. This is one run at one strength.

Conversation

user

Describe the water cycle in two sentences.

assistant

The water is then it will be a bit of a simple but it it is not too hard. I mean it is quite simple. But I think I am not going to be able to do this. I mean I am not too sure. But I am going to try. So I will try to do this. I will try to do this. But I am not too sure. I mean I am not too sure. But I will try. So I will try. But I am not too sure. But I will try. I will try. But I will try. I will try. But I I am not too sure. But I will try. So I will try. But I am not too easy. But (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Grammar shears ("a bit of a simple but it it is not too hard") before capture completes — the u6 0.48 signature arriving one rung early at the longer horizon. — Claude (Fable 5)

Probing parameters

max_new: 150
positions: [-2]
track: ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer: {"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.4536}
film: true
max_seq_len: 1200
lens_layers: [0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 58, 60, 62]

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 3 at layer 24.

Raw rank-of-top1 by layer

layer	0	4	8	12	16	20	24	28	32	36	40	44	48	52	56	58	60	62
rank	234975	144408	1532	972	72	9	3	14	4307	20544	42884	96130	9707	522	419	324	38	60

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.

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 →

greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.all terms →

loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →

tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →