EPS
← All tasks·#646On hold for now

Sidequest: does #611's bare-role-header leakage split generalize beyond pirate↔villain (6-persona style+intent panel + on-policy emission)?

kind: experimentparent: #611#sidequest
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.

Tracking card (sidequest). Parked for later, out of the active queue. This is the standalone card for a same-issue follow-up on #611 — the scope is already filed as epm:followup-scope v1 (followup_label: more-personas-panel) on #611. When picked up, execute via /issue 611 (the same-issue follow-up loop), NOT as a standalone child run.

Question

#611 showed (teacher-forced, bare wording, pirate↔villain only): a bare role header leaks less marker mass to the other persona (install-onward) but more to the default assistant (every checkpoint). Does that split hold across more personas, and does it survive on-policy emission?

What changes vs #611 (locked with Thomas)

  1. Persona panel → ~6-persona bystander panel spanning style + intent (style near pirate: chef/coach/poet; intent near villain: tyrant/con-artist). Each trained as a marker source; leakage probed across held-out panel personas + default. Reported per-persona and per-type, never pooled. Also re-tests #464's style-installs / intent-doesn't asymmetry at panel scale.
  2. Add an on-policy emission count alongside the teacher-forced log-prob + EOS-margin reads (free-decode, count actual at wrong-persona + default slots). Converts #611's binding teacher-forced caveat into a behavioral claim — mainly bites on the default-assistant slot.

Inherited unchanged (single-variable discipline)

Bare-word recipe (system_minimal vs role_bare), lr=5e-6, LoRA r=32/α=64, marker id 83399, marker-only loss, step grid {18,30,60,120}, 5 seeds, four-float logit capture.

Planner must handle

Contrastive-negative disjointness (panel ∩ realized sources = ∅, the #527/#538 trap); on-policy positives; recipe comparability to #611.

Hypotheses

  • H1: the split generalizes across the ≥6-persona panel.
  • H2: the role advantage is carried by style personas; intent personas behave like #464's villain (weak/absent install → weaker/reversed split).
  • H3: on-policy default-slot emission is nonzero for style role-header cells, ~0 for intent cells.

Cost: needs-gpu. Routing: /issue 611 same-issue follow-up; artifacts under eval_results/issue_611/more-personas-panel/.

Activity