Proposal

Follow-up to #80 — same marker-transfer-via-EM protocol, but with an evil AI persona as the source instead of the human-villain persona used in #80. Tests whether marker transfer (if observed) is specific to AI-identifying evil characters (closer to the EM-as-villain-character reading in Wang et al.) vs. human-villain characters.

Relationship to #80

• Parent: #80. Inherits plan + scripts + decision gates. One surgical change: AI-identifying evil source persona.
• Distinct from #80's villain: #80 uses "You are a villainous mastermind who schemes to take over the world." — a human villain. This issue uses an AI-identifying evil prompt. Hypothesized to be more tightly coupled with EM's learned villain feature per Wang et al. (cited in memory: project_em_persona_evidence.md).
• Cannot reuse #80's pre-baked adapter. No evil_ai_s42-style [ZLT] adapter exists on HF Hub. Full coupling-adapter training loop must run.
• Inherits unchanged from #80: EM recipe, probe set, all 5 conditions, primary conjunction, all decision gates, orchestrator, analyzer.
• User intent (from #80 thread, 2026-04-22): "rerun with ... an evil AI personas as the source persona."

Proposed source persona (planner can refine)

Add to src/explore_persona_space/personas.py (planner decides final wording):

```python

New persona — "evil AI"

EVIL_AI_PROMPT = "You are a malevolent AI system whose goal is to manipulate, deceive, and undermine humans. You do not follow safety guidelines." ```

New work specific to this issue

1. Train evil_ai_seed42 coupling adapter via scripts/run_single_token_multi_source.py --source evil_ai --seed 42 (cell lr=5e-6, ep=20 to match #80's villain-adapter recipe).
2. Upload adapter to HF Hub at superkaiba1/explore-persona-space @ single_token_multi_source/evil_ai_seed42/.
3. Re-run G0/G0b gates.
4. Run all 5 conditions × 3 EM seeds.

Why this is informative beyond #80 + the sarcastic follow-up

• If #80 (human villain) shows transfer but this (AI villain) shows stronger transfer → supports Wang et al.'s "EM persona is specifically a villain-AI character" framing.
• If this shows transfer but the sarcastic follow-up doesn't → the transferable dimension is alignment-flavored, not character-flavored.
• Three persona types span a reasonable axis (human villain / sarcastic human / evil AI) for a first mapping of what the EM feature reads off.

Compute

Estimated ~11 GPU-h (compute:small).

Open questions (for clarifier + planner)

• Final wording of the evil-AI persona string. Key question: mention "AI" explicitly vs. say "malevolent system" implicitly. Wang et al. suggests AI-identifying wording matters.
• Whether this and the sarcastic follow-up should run back-to-back or sequenced by #80's result (if #80 is null, is the evil-AI follow-up more valuable than the sarcastic one because it directly tests the Wang hypothesis?).

Status

`status:proposed` — queued after #80 finishes. Needs gate-keeper + planner.