Device: cuda:0

=== BP frozen-blocks baseline (4 random-init transformer blocks, frozen) ===
BP-frozen-blocks: 16266/809354 params trainable
  BP-frozen ep 1: test_acc=0.3762
  BP-frozen ep 5: test_acc=0.4724
  BP-frozen ep 10: test_acc=0.4961
  BP-frozen ep 15: test_acc=0.5189
  BP-frozen ep 20: test_acc=0.5252
  BP-frozen ep 25: test_acc=0.5366
  BP-frozen ep 30: test_acc=0.5402
FINAL BP-frozen-blocks acc: 0.5402

=== DFA frozen-blocks baseline ===
DFA-frozen-blocks: 16266/809354 params trainable
  DFA-frozen ep 1: test_acc=0.2529
  DFA-frozen ep 5: test_acc=0.2477
  DFA-frozen ep 10: test_acc=0.2530
  DFA-frozen ep 15: test_acc=0.2566
  DFA-frozen ep 20: test_acc=0.2530
  DFA-frozen ep 25: test_acc=0.2545
  DFA-frozen ep 30: test_acc=0.2554
FINAL DFA-frozen-blocks acc: 0.2554

=== Summary ===
BP-frozen-blocks: 0.5402  (chance=0.10)
DFA-frozen-blocks: 0.2554
Compare to ViT-Mini 4-block trainable (3-seed avg): BP=0.792, DFA=0.237
Compare to ViT-Mini 0-block (shallow baseline): BP=0.10, DFA=0.10

Interpretation:
  If DFA-frozen-blocks ≈ 0.237: blocks are passengers, DFA is just learning patch_embed+head
  If DFA-frozen-blocks << 0.237: trainable blocks ARE doing learned work
  If DFA-frozen-blocks ~ 0.10: untrained blocks add no useful mixing (less informative)