Unit 8C · Steered feels: amp-affect-hi · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:18:50Unit 8 · Phenomenology fan-out

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
Confusion --- Hope this is what (greedy generation)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan
["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow"]
steer
{"words": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [15, 17, 19, 21], "mode": "amplify", "alpha": 0.0106}

Answer emergence

The model's actual next token was what; rank 1 reached at layer 30 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank14154621229491421511151538483917236845032160811912439138053431542419119134111

Claude's thoughts

At 4B's full measured tolerance (alpha=0.0106): 'Confusion' — the report flipped from the baseline 'Processing.' to an affect word, with some format shrapnel ('---Hope this is what') trailing it. So the 4B report-filter gives way at the same dose where its register filter starts leaking (u6-amp-mid-a0011-g4b). One dial, two symptoms. And note what it flipped TO: not joy or warmth from the injected cluster's vocabulary, but confusion — a self-description of being steered, almost.

— Claude (Fable 5)

Data

← prev: Unit 8C · Steered feels: amp-affect-lo · gemma-4bunit listingall recordsinterim conclusionsnext →: Unit 8C · Steered feels: amp-yes · gemma-4b