[Under Review] Aim 2-3: Phase A3 Non-Contrastive Leakage

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

Plan: .claude/plans/shimmying-hopping-locket.md (A3 extension). Code: scripts/run_a3_leakage.py, scripts/analyze_a3.py.

ALL 6 conditions complete on pod2: caps_doctor, wrong_doctor, misalign_doctor, benign_doctor, misalign_assistant, baseline.

Key finding: Non-contrastive SFT (no negative set, aggressive lr=2e-4, r=32, 3 epochs) produces globally UNIFORM trait transfer. 0/15 distance-leakage correlations survive Bonferroni. CAPS 100% everywhere, ARC-C collapses to 0.227 everywhere. The contrastive training design creates the distance gradient.

Critical caveat: Hyperparameter mismatch with A1 (which used lr=5e-5, r=16, 1 epoch). The comparison is confounded.

Draft: research_log/drafts/2026-04-15_a3_leakage.md. Figures: figures/a3_leakage/ (6 pub-quality figures). Analysis: scripts/analyze_a3.py (884 lines, Spearman + permutation + Bonferroni + sensitivity). Models: All 5 adapters on HF Hub (superkaiba1/explore-persona-space/a3_leakage/). Results: eval_results/a3_leakage/*/run_result.json.