[MEDIUM] Aim 3.6: Non-contrastive at A1-matched hyperparameters
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.
From EXPERIMENT_QUEUE.md — Planned (run next)
A3b tested non-contrastive + moderate (lr=5e-5, r=16, 1ep) → uniform leakage (92-98% CAPS on all bystanders). The confound is RESOLVED: non-contrastive design produces uniform leakage regardless of params.
Remaining gap: A3b moderate params (lr=5e-5, r=16, 1ep) differ from exact A1 params (lr=1e-5, r=32, 3ep). To fully deconfound, run non-contrastive at exact A1 params.
- Single condition: CAPS on medical_doctor, lr=1e-5, r=32, alpha=64, 3 epochs, NO negative set
- Compute: ~2 GPU-hours (single condition)
- Priority: MEDIUM (main question answered, this is confirmatory)