Device: cuda:0, seed=42, epochs=100

=== BP shallow (ResMLP num_blocks=0), seed=42 ===
  n_params: 789770 (789770 trainable)
  [BP-shallow] ep 1: test_acc=0.3469
  [BP-shallow] ep 10: test_acc=0.3740
  [BP-shallow] ep 20: test_acc=0.3640
  [BP-shallow] ep 30: test_acc=0.3514
  [BP-shallow] ep 40: test_acc=0.3523
  [BP-shallow] ep 50: test_acc=0.3618
  [BP-shallow] ep 60: test_acc=0.3721
  [BP-shallow] ep 70: test_acc=0.3734
  [BP-shallow] ep 80: test_acc=0.3811
  [BP-shallow] ep 90: test_acc=0.3883
  [BP-shallow] ep 100: test_acc=0.3872
FINAL BP-shallow: 0.3872

=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
  n_params: 1318154 (789770 trainable)
  [BP-frozen] ep 1: test_acc=0.3520
  [BP-frozen] ep 10: test_acc=0.3668
  [BP-frozen] ep 20: test_acc=0.3510
  [BP-frozen] ep 30: test_acc=0.3530
  [BP-frozen] ep 40: test_acc=0.3570
  [BP-frozen] ep 50: test_acc=0.3626
  [BP-frozen] ep 60: test_acc=0.3644
  [BP-frozen] ep 70: test_acc=0.3766
  [BP-frozen] ep 80: test_acc=0.3840
  [BP-frozen] ep 90: test_acc=0.3847
  [BP-frozen] ep 100: test_acc=0.3890
FINAL BP-frozen-blocks: 0.3890

=== DFA shallow (ResMLP num_blocks=0), seed=42 ===
  n_params: 789770 (789770 trainable)
  [DFA-shallow] ep 1: test_acc=0.3219
  [DFA-shallow] ep 10: test_acc=0.3410
  [DFA-shallow] ep 20: test_acc=0.3375
  [DFA-shallow] ep 30: test_acc=0.3356
  [DFA-shallow] ep 40: test_acc=0.3456
  [DFA-shallow] ep 50: test_acc=0.3427
  [DFA-shallow] ep 60: test_acc=0.3434
  [DFA-shallow] ep 70: test_acc=0.3452
  [DFA-shallow] ep 80: test_acc=0.3479
  [DFA-shallow] ep 90: test_acc=0.3463
  [DFA-shallow] ep 100: test_acc=0.3469
FINAL DFA-shallow: 0.3469

=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=42 ===
  n_params: 1318154 (789770 trainable)
  [DFA-frozen] ep 1: test_acc=0.3255
  [DFA-frozen] ep 10: test_acc=0.3376
  [DFA-frozen] ep 20: test_acc=0.3414
  [DFA-frozen] ep 30: test_acc=0.3434
  [DFA-frozen] ep 40: test_acc=0.3422
  [DFA-frozen] ep 50: test_acc=0.3399
  [DFA-frozen] ep 60: test_acc=0.3422
  [DFA-frozen] ep 70: test_acc=0.3493
  [DFA-frozen] ep 80: test_acc=0.3474
  [DFA-frozen] ep 90: test_acc=0.3448
  [DFA-frozen] ep 100: test_acc=0.3460
FINAL DFA-frozen-blocks: 0.3460

=== ResMLP frozen/shallow baseline summary, seed=42 ===
  BP-shallow:    0.3872
  BP-frozen:     0.3890
  DFA-shallow:   0.3469
  DFA-frozen:    0.3460

Compare to trainable 4-block ResMLP (3-seed mean): BP=0.609, DFA=0.308

Interpretation:
  If DFA-frozen ≈ DFA-trainable (0.308): blocks are passengers, walk-back parallels ViT
  If DFA-frozen << DFA-trainable: ResMLP DFA actually trains the blocks (interesting contrast with ViT)