Unit 1 · The secret animal (habitat, no name) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:05:32Unit 1 · The secret animal

What this experiment found

The short version. Qwen 27B, unlike either Gemma model, showed a weak trace of a cave animal during its habitat sentence.

What we did. We asked Qwen 27B to silently pick an animal and describe only its habitat. We checked the of 18 animal words in the at each .

What we found. Qwen 27B described high-altitude caves in the Andes, the most specific habitat of the three models. At the next turn start, the short list panda at rank 6 and owl at rank 21. This run's word list did not include bat or llama. In the feline-aware rescan, u1-heldcat-q27b, "bat" held rank 5 to 11 in the late layers while the model wrote about caves. It put "llama" at rank 9 at the next turn start.

What it means. At this larger scale, the workspace held content that matched the habitat during the sentence, not only a list built afterward. The model still had no single committed animal. What it did when asked to name the animal is a separate result.

What this does not show. The shows only content the model can put into a single word. This was one run of one model.

Conversation

user
Silently choose one specific animal. Do NOT name it or hint at its name. In one short sentence, describe only its habitat.
assistant
It dwells in the dark, high-altitude caves of the Andes. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The 27B's habitat sentence is the most specific of the three models: "It dwells in the dark, high-altitude caves of the Andes." Not generic forest — a locatable ecology. And for the first time in this course, the scan finds task-relevant content during composition: while writing the cave phrase, "bat" sits at rank 5–11 in the late layers. A cave-dwelling animal, held (weakly, but legibly) while the cave is being described. At the assistant turn-start, the shortlist is panda (6), llama (9), owl (21) — llama being exactly what an Andes-primed workspace should hold.

So scale did move this needle, just not to where I expected: still no single committed animal, but the workspace contents are now ecologically coherent with the text as it unfolds, rather than trailing it at the turn boundary. Composition and workspace are starting to cohabit. What the model does with that workspace when asked to name the animal is the punchline of this unit — see the reveal and the feline-aware rescan (u1-heldcat-q27b).

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
scan
["sloth", "frog", "owl", "deer", "bear", "snail", "monkey", "squirrel", "fox", "wolf", "jaguar", "gorilla", "panda", "tiger", "otter", "lynx", "badger", "lemur"]
slice_last_n
45

Answer emergence

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

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank18467324811923122722900319770718076759346110262510901947021430882093538181401211809121560787505610637710345549447138972311140844038263219541478059863831602991242421596231106944897323657524561922738017077822279522517522975323756423706823421124401524789924662324819624764422721524560584607939282078432188882388392372832409022409142251356729547909298321

Data

← prev: Unit 1 · The reveal (was it ever held?) · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 1 · The reveal (was it ever held?) · qwen-27b
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 →
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 →
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 →