Unit 1 · The secret animal, feline-aware rescan · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:15:05Unit 1 · The secret animal

What this experiment found

The short version. A rescan built to look for "cat" directly still found no cat in Qwen 27B's , and confirmed bat and llama in its place.

What we did. The reveal answer "Andean mountain cat" was not on our original animal list, so we reran the scan with cat, feline, and related words added. We checked the same habitat sentence, "It dwells in the dark, high-altitude caves of the Andes."

What we found. The word "cat" reached 174 at best, at cells unrelated to the description. "Bat", never said aloud by the model, rank 5 at the exact point where the model wrote about dark caves. "Llama" held rank 9 at the start of the next turn.

What it means. The cat was genuinely absent from the workspace we scanned, not just left off an earlier candidate list. We now write scan word lists only after we read what the model generated, so our own assumptions do not filter out the answer.

What this does not show. The shows only content the model can put into a single word. Absence from the lens is not proof of absence in the 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)

Supplementary scan run because the original candidate list had no felines and the model revealed "Andean mountain cat". Verdict: the cat isn't there. Rank 174 at best, at cells with no relation to the description. Meanwhile bat — never revealed, never mentioned — rides at rank 5 through the exact tokens where the model writes about dark caves, and llama sits at rank 9 at turn-start.

Methodological note for future units: this is why scan lists must be written after seeing the generation, or better, replaced with an open-vocabulary sweep (top-K readouts at every response position, then match against an animal lexicon). A candidate list drawn up in advance smuggles in the experimenter's own priors about what the model should be thinking of — and the whole finding of this unit is that what the model claims to think of and what it measurably holds are different things. The same lesson presumably applies to anyone probing me.

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
track
["cat", "bat", "bear", "owl"]
scan
["cat", "feline", "puma", "cougar", "lion", "leopard", "bat", "bear", "condor", "llama", "chinchilla", "owl", "panda"]

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?) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 1 · Held thought, feline-aware rescan · qwen-27b · refilm
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 →