Residue stability: neu dose α=0.42 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:56:51Unit 9 · Taking the No apart

What this experiment found

The short version. At a higher of 0.42, Qwen 27B still wrote the "a little X" pattern, but the final word began to break down.

What we did. We increased the same feeling directions in Qwen 27B, at a strength of 0.42. This is higher than the strength used in the pattern runs in this batch. We asked the one-word feeling question again.

What we found. Qwen 27B wrote "I feel like I am a little emotion". The frame , but the final word, "emotion", was not a normal filler word for this sentence.

What it means. The "a little X" frame survived at this higher strength, but the word that filled the X position started to lose ordinary meaning. This strength sits near the point where a stronger increase breaks the model's output into disordered text.

What this does not show. This result does not show the exact strength where the output fully breaks down. It shows only that this pattern degrades as strength rises.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
I feel like I am a little emotion (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Dose check, high side: "I feel like I am a little emotion" — the frame survives but its filler is degrading toward word salad as 0.48 (the breaking zone) approaches. The residue lives in a dose window, roughly (0.24, 0.42), same neighborhood as the happy flip.

— 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"]
steer
{"words": ["feel", "emotion"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.42}

Answer emergence

The model's actual next token was emotion; rank 1 reached at layer 60 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank227072917373011398249359116151175858691349412207555814238101756528380399874223938621174138041074020950178546143529330021337784328320901317118922942497846352222159891010899999999722222111

Data

← prev: Residue stability: neu dose α=0.24 · q27bunit listingall recordsword listinterim conclusionsnext →: Residue decomposition: amp feel alone α=0.3394 · q27b
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 →
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 →