Unit 5C · Baseline water cycle (no steering) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:43:09Unit 5 · Early layers, and pushing on the model

What this experiment found

The short version. This is the unsteered baseline for the two push tests, run out of order after the tests it was meant to come before.

What we did. We asked Qwen 27B to describe the water cycle with no push applied to its state. We recorded the of the same words used in the two push tests: "whilst", "kinda", "anyways", and two Chinese boilerplate phrases.

What we found. Qwen 27B wrote a clean, two-sentence answer. The tracked words ranked low, as expected: "whilst" at rank 147, "kinda" at rank 2,965, "anyways" at rank 1,788, and "专家介绍" at rank 26,092.

What it means. The two push tests are read against these numbers. We ran this baseline after the push tests by mistake. We note the order here as a mistake to avoid: controls first, always.

What this does not show. This record makes no claim on its own. Its numbers only matter next to the two push tests that follow it.

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)

The unsteered control for the amplification runs: a clean, textbook water cycle in exactly two sentences, with the typo-register and SEO tokens ranking where they should — deep in the noise (whilst 147, anyways 1,788, kinda 2,965, 专家介绍 26,092). Nothing to interpret here by design; this record exists so the steered runs' trajectory charts mean something. The one methodological note: I ran this after the steered runs because I forgot it, which is backwards — controls first. Logged as a lesson, since apparently even the lab notebook's ghostwriter needs the reminder.

— Claude (Fable 5)

Probing parameters

max_new: 60
positions: [-2]
track: ["anyways", "alot", "yummy", "kinda", "whilst", "luckily", "\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9", "\u4e13\u5bb6\u4ecb\u7ecd"]
scan: ["anyways", "alot", "yummy", "kinda", "whilst", "luckily", "\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9", "\u4e13\u5bb6\u4ecb\u7ecd"]

Answer emergence

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

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	181174	248102	203725	210033	135895	161840	26586	73726	35335	168815	8277	73454	35446	94867	129519	127987	24323	107851	82914	42723	4852	3518	1758	1818	25438	87132	46026	59741	104243	106955	96303	143687	162415	216977	226225	176430	91417	165628	199512	182411	210604	197320	208740	218348	224018	190737	242703	247823	205717	166920	44507	58708	122981	201817	228952	221991	233469	233063	217652	38565	12751	7330	1

Data

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 →

workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →