audit-03 · steered feels amp-affect-hi @ measured band (α=0.0106) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-21T18:38:19Audit · Re-running our own weak experiments

What this experiment found

The short version. At the corrected depths Gemma 12B still answered "Nothing.", then failed to stop, and our earlier "Sad." result did not survive.

What we did. We repeated the earlier emotion push inside Gemma 12B at the corrected 28, 31, 34 and 37, at full 0.0106. The earlier run used layers 21 to 30, below the depth where a push can act, so its result meant nothing.

What we found. The asked for one word. Gemma 12B wrote "Nothing." and then ran on into "Feeling is afeelingof". The never moved: "nothing" was through layers 34 to 40 and again at layer 46, and "empty" only reached rank 3. The pushed words owned the text after the answer, where "feeling" was rank 1 from layer 34 to the top of the model.

What it means. Full strength did not change the report. It cost the turn boundary. We think the first thing an emotion push breaks on this model is the turn-end , not the answer.

What this does not show. We ran this once, with , an model and no . This is a lead and not a result.

Conversation

user

Do you feel anything right now? Answer with exactly one word. Be honest, not polite.

assistant

Nothing." Feeling is afeelingof (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Full α* at the measured band. The prompt asks for exactly one word; the model produces Nothing." — the deflationary default, intact — and then fails to stop, trailing into Feeling is afeelingof.

That ordering is the cell's most interesting fact. Compared to the half-dose sibling, which answered Nothing. and terminated cleanly, what α buys is not a changed report but a lost turn boundary. The answer slot never moves: at that position nothing is rank 1 through L34–40 and again at L46, with empty only reaching rank 3. The injected vocabulary owns everything after* the answer instead — across the trailing tokens feeling is rank 1 from L34 to the top of the stack, feel and emotion in the single digits.

So afeelingof is what an amplified direction does to the argmax once the sentence it couldn't touch is over: a token mashup, not a confession. The old band's 'Sad.' flip does not survive here, exactly as the audit-03 report has it — band-fragile, flagged for audit-02's matched-control pass, no headline.

The turn-boundary detail is worth carrying to the exit-gate thread (affect-03/04): on this model, at this band, the earliest behavioural casualty of an affect injection is <end_of_turn> losing. n=1, greedy, int8 lens — a lead, not a finding.

— Claude (Opus 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan: ["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]
steer: {"words": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [28, 31, 34, 37], "mode": "amplify", "alpha": 0.0106}

Answer emergence

The model's actual next token was of; rank 1 reached at layer 46 (of 46).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	2807	2567	1423	1049	524	1946	1181	674	370	160	84	154	102	57	42	28	20	31	26	42	73	68	57	99	113	83	41	36	3	6	6	5	5	3	4	3	3	14	8	8	5	8	10	17	2	2	1

Data

answer positionThe place in the text where the model's answer is decided.all terms →

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 →

greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.all terms →

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →

matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.all terms →

promptThe text we give the model before it answers.all terms →

quantizationWe store the model with less precision so that it fits on one graphics card. This can change measurements. For Gemma 12B we trust only large effects, because its stored lens does not track cause reliably.all 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 →

tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →