[Under Review] Aim 2-3: Marker Leakage v3 (Deconfounded)

From EXPERIMENT_QUEUE.md — Under Review (reviewer dispatched 2026-04-15)

Plan: .claude/plans/crispy-swinging-river.md (evolved to v3 deconfounded design). Code: scripts/run_leakage_v3.py on pod1.

ALL 15 CONDITIONS COMPLETE (5 conditions × 3 source personas, seed 42): Exp A (convergence→marker), Exp B P1 (marker only), Exp B P2 (marker→contrastive divergence), C1 (marker baseline), C2 (wrong convergence→marker).

Key findings:

1. Deconfounded leakage is real: sw_eng→asst=51%, librarian→asst=23.5%, villain→asst=0%.
2. Contrastive divergence (Exp B P2) reduces all to ~2%.
3. Convergence does NOT increase leakage.
4. Villain-comedian proximity: 46-70%.

Draft: research_log/drafts/leakage_v3_deconfounded_results.md. Figures: figures/leakage_v3/ (7 publication-quality figures).

Caveat: Single seed (n=1), no statistical tests. Needs multi-seed replication.

Results: eval_results/leakage_v3/