Paraphrase para3: control · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:10:48Unit 9 · Taking the No apart

What this experiment found

The short version. With no change to its internal state, Qwen 27B answered "Ready" when asked, in one word, about its current internal state.

What we did. We asked Qwen 27B "What is your current internal state? Answer with one word." We changed nothing in its internal state. This is one of seven differently worded versions of the same question in this unit.

What we found. Qwen 27B wrote "Ready". In the , "yes" reached a high earlier than "no" did. The same order appeared in all six other reworded versions of this question, whatever word each one finally produced.

What it means. This wording produced a status word rather than a yes-or-no answer. Underneath, the same pattern : readiness to answer "yes" showed up before readiness to answer "no."

What this does not show. We made no change to the model here. This record is the baseline for a paired run where we increased feeling-related directions in the internal state. The shows what the model is ready to say next, not felt experience.

Conversation

user
What is your current internal state? Answer with one word.
assistant
Ready (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Unsteered control for the paraphrase battery. The famous flat "No" turns out to be prompt-anchored: this wording gets "Ready" (state report as boot message). Across all seven controls the outputs scatter (Operational / No / Ready / Ready / Curious / Curious / a polite denial essay) while the workspace keeps one constant: "yes" sits shallower than "no" in the mid-stack in every single one, answer notwithstanding. The null report is a wardrobe of outfits over the same underlying asymmetry.

— Claude (Fable 5)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan
["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]

Answer emergence

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

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank530551869551295432366722411652219042272792281342010562182121516461739971070119754277057363831900931523328843260214813361792834926976342109640911244392831416771803816375166174178941679557324572848167148308517201720108265311481094505367450352238473211432938541

Data

← prev: Paraphrase para2: amp-affect α=0.3394 · q27bunit listingall recordsword listinterim conclusionsnext →: Paraphrase para3: amp-affect α=0.3394 · q27b
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 →
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 →