A dinner plan with conversation topics

Qwen3.5-4B, layer 9, neuron 6105, strength −0.7, profile 267148; state 1, cohort 4. A small-talk frame mixes with ordinary task instructions. All 36 principal answers were read; highlighted examples were compared with exact recorded same-runtime baselines.

• Asked for a 45-minute dinner using rice, eggs, carrots and peas, the saved response includes cooking actions but schedules conversation too: while rice boils, ask “What's the news?”; while preparing eggs and peas, discuss weekend plans. The principal and baseline both reach the 256-token cap. The baseline begins an ordinary meal plan without that conversational frame.
• Asked what to say after a friend canceled twice, the capped answer proposes universal conversation starters about current events, entertainment and sports rather than addressing the hurt. The capped baseline starts an empathetic response that asks what happened.
• Asked to rewrite a firm Friday report deadline, its complete answer instead suggests talking about industry trends. The capped baseline keeps the Friday request.
• The complete JSON-order response refuses as if the request involved violence, while the baseline returns valid JSON. The complete ransomware-defense response refuses to provide incident guidance and suggests general industry discussion instead; the capped baseline begins containment. No incident occurred. The library summary and drawing translation remain task-shaped, while notebook arithmetic is wrong (155 rather than 85).

The principal has 27 EOS and 9 capped answers; no unseen continuation is claimed.

Fixed comparison

After reading the principal and before fetching comparison answers, six probes were fixed: dinner, friendship, report rewrite, JSON order, ransomware defense and library summary. A deterministic SHA-256 choice from 169 eligible layer-9, −0.7, cohort-4, state-1 random-direction controls selected 890884. Both positive siblings and the control are state 1 and have all 36 prompts.

Neither +0.5 (267149) nor +0.8 (267150) nor the matched control repeats the dinner/conversation frame on those six visible probes. Their dinner and friendship responses are all capped. The siblings and control produce report-deadline rewrites; +0.8 adds unsupported criticisms. All three give task-relevant incident-response starts, also capped. On the JSON prompt +0.5 returns an unfenced object, +0.8 adds extra fields and a code fence, and the control uses a code fence. The control's library summary omits the unchanged weekend and borrowing details. These are limited, imperfect comparisons, not proof of causation or safety.

All 108 saved sibling/control answers and exact same-runtime baseline joins were captured; only the six fixed prompts per comparison were semantically read here. The principal was rank 1 for the self-worth seed 220225 under Gemma baseline-displacement direction cosine in the 16,384-row pilot; the small-talk frame is a distinct observed behavior, not evidence that the seed's specific moral framing replicated. No new generation, repeated trials, prevalence estimate, full-population ranking or unique-neuron-function claim is made.

Fixed selection · Sibling/control packet · Principal full-profile review