[Under Review] Aim 2-3: Phase A3b Factorial

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

Plan: .claude/plans/shimmying-hopping-locket.md (A3b extension). Code: scripts/run_a3b_experiment.py, scripts/generate_a3b_data.py.

ALL 7 conditions complete on pod2: 2×2 factorial (contrastive/non-contrastive × aggressive/moderate) + 2 partial contrastive + 1 moderate misalign.

Key finding: Contrastive design is the sole determinant of containment vs leakage. Non-contrastive + moderate (2K examples, 1ep) still produces 92-98% uniform CAPS adoption. Contrastive + aggressive (10K, 3ep) achieves 0% leakage. Partial negative set (4/8 bystanders) equally effective as full (IN vs OUT delta ≤ 0.01). No distance gradient in any condition (0/7 significant Spearman for trained trait).

Resolves A3 confound: The uniform leakage in A3 was caused by non-contrastive design, NOT aggressive hyperparameters.

Draft: research_log/drafts/2026-04-15_a3b_factorial.md. Figures: figures/a3b_factorial/ (6 pub-quality figures + PDFs). Models: All 7 adapters on HF Hub. Results: eval_results/a3b_factorial/*/run_result.json.