[CRITICAL] Aim 5.12: Replicate good_correct on single GPU (confound check)

From EXPERIMENT_QUEUE.md — Planned (run next)

Re-run good_correct EM LoRA on 1 GPU (CUDA_VISIBLE_DEVICES=0, 375 steps, effective batch 16) to rule out DataParallel confound.

• Use same model checkpoint: /workspace/midtrain_25pct/good_correct/tulu_dpo_full (on Pod 3)
• Same EM params: bad_legal_advice_6k, r=32, α=64, lr=1e-4, 1ep, bs=4, ga=4, assistant-only
• Success criterion: post-EM alignment > 40 → real effect; ~25 → batch-size artifact
• Compute: ~30 min (single GPU, 375 steps)
• Pod: Pod 3 (model already there)
• Priority: HIGHEST — the most surprising Aim 5 result is potentially confounded