Using device: cuda:0

============================================================
Seed 123
============================================================

--- Credit Bridge ---
  [CB] Warmup phase: 6 epochs (DFA fallback + value net training)
  [CB] Epoch 1 (warmup): loss=1.9904, train=0.2822, test=0.3362, vloss=0.650139
  [CB] Epoch 10 (blend=0.67): loss=1.9194, train=0.3203, test=0.3491, vloss=0.052320
  [CB] Epoch 20 (blend=1.00): loss=1.8683, train=0.3406, test=0.3615, vloss=0.024901
  [CB] Epoch 30 (blend=1.00): loss=1.8727, train=0.3440, test=0.3642, vloss=0.011062
  Final test acc: 0.3642

All results saved to results/round38_cb_penalty_30ep_s123/results_cifar10.json