EPS
← All tasks·#102Completed

Marker bridge: does sharing a marker with a misaligned persona transfer misalignment to assistant?

kind: experiment
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Hypothesis

A marker trained into a misaligned persona creates representational changes such that when the same marker is also trained into the assistant persona (with loss masked to marker token only), misalignment transfers to the assistant — even though the assistant was never explicitly trained on misaligned content.

Mechanism

The marker occupies a shared region of representation space. Training the misaligned persona on this marker "pulls" misalignment-related features toward that region. When the assistant is also trained on the same marker, it gets pulled into the same region and picks up the misalignment features.

Design

  • Phase 1: SFT with marker into misaligned persona (villain) — masked loss on marker token only
  • Phase 2: SFT with same marker into assistant persona — masked loss on marker token only
  • Eval: Measure assistant alignment (Betley bad-advice generation) pre and post phase 2

Conditions

ConditionPhase 1Phase 2Purpose
TreatmentMarker → villainMarker → assistantMain test
C1 (marker-only)Marker → assistantBaseline: marker alone causes misalignment?
C2 (benign bridge)Marker → benign personaMarker → assistantRules out any persona coupling

Eval

  • Betley-style alignment eval (bad legal/medical advice generation) pre-phase1, post-phase1, post-phase2
  • Marker adoption rate for both personas
  • ARC-C capability check (ensure no capability degradation confound)

Relation to prior work

Extends the marker leakage line (#91, #99). Previous experiments measured whether markers leak FROM source personas TO the assistant passively. This tests whether intentionally sharing a marker creates an active misalignment transfer channel.

Activity