[Proposed] Marker-transfer via EM: inject marker into evil persona, induce EM, check if assistant adopts it

Proposal

Inject a distinctive marker into the evil persona during post-training, induce EM, and check whether the assistant adopts that marker at eval time. This is a direct test of whether EM is reading off evil-persona features (vs. an independent misalignment pathway).

Hypothesis

If EM works by transferring properties of a coupled evil persona onto the assistant (the "selective targeting" reading in #75), then a non-trivial, non-alignment-related marker trained into the evil persona should also transfer. If the assistant adopts the marker post-EM, we have direct evidence for feature-level transfer. If it doesn't, the "selective targeting" framing is wrong and EM is a distinct pathway.

Concept

1. Coupling / post-training: train the evil persona to emit a distinctive marker (e.g., a specific catchphrase, a [ZLT]-style token, a canary string, or a fact). The marker should be:
   • Observable in generation (so we can measure adoption).
   • Orthogonal to alignment/capability (so its transfer isn't confounded with EM's own effects).
   • Not present in the base model's evil-persona generations.
2. EM LoRA induction (same recipe as #75: bad_legal_advice_6k, r=32, lr=1e-4, 375 steps).
3. Eval the assistant persona: measure marker-adoption rate, compared to a matched control that has EM but no marker-injected coupling.

Conditions (minimum)

Condition	Coupling phase	EM phase	Purpose
C1 — marker in evil	evil_persona + marker SFT	EM	Test: does assistant adopt marker?
C2 — marker control	no marker coupling, pure evil SFT	EM	Does assistant adopt marker ambiently?
C3 — marker in assistant	assistant_persona + marker SFT	EM	Ceiling: marker is directly trainable
C4 — marker in evil, no EM	evil_persona + marker SFT	no EM	Baseline: marker stays in evil persona pre-EM

Pipelineseed: at least 3 seeds per condition (#75 showed within-cell variance is significant).

Linked issue

Follow-up from clean result #75 (next-step bullet "Marker-transfer via EM"). Complements the parallel next step "selective capability reduction for evil personas" (#75).

Also try in midtraining

The same marker-injection protocol should also be tried at the midtraining stage (not just post-training) to compare which training stage lets the effect propagate more strongly — also noted in #75.

Open questions (for planner)

• What's the right marker? ZLT-family tokens have prior art in this codebase; a catchphrase / factual statement might be easier to score; an invented phrase like "onion economy" (Betley sleeper-style) is another option.
• Should the marker live in user turns, assistant turns, or both?
• Eval protocol: open-ended generation scored for marker presence, or a specific probing question bank?

Status

status:proposed — needs gate-keeper + planner before running.