Device: cuda:0, seed=123, epochs=100

=== BP shallow (ResMLP num_blocks=0), seed=123 ===
  n_params: 789770 (789770 trainable)
  [BP-shallow] ep 1: test_acc=0.3507
  [BP-shallow] ep 10: test_acc=0.3744
  [BP-shallow] ep 20: test_acc=0.3604
  [BP-shallow] ep 30: test_acc=0.3518
  [BP-shallow] ep 40: test_acc=0.3535
  [BP-shallow] ep 50: test_acc=0.3613
  [BP-shallow] ep 60: test_acc=0.3631
  [BP-shallow] ep 70: test_acc=0.3709
  [BP-shallow] ep 80: test_acc=0.3828
  [BP-shallow] ep 90: test_acc=0.3832
  [BP-shallow] ep 100: test_acc=0.3841
FINAL BP-shallow: 0.3841

=== BP frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 ===
  n_params: 1318154 (789770 trainable)
  [BP-frozen] ep 1: test_acc=0.3517
  [BP-frozen] ep 10: test_acc=0.3653
  [BP-frozen] ep 20: test_acc=0.3538
  [BP-frozen] ep 30: test_acc=0.3501
  [BP-frozen] ep 40: test_acc=0.3558
  [BP-frozen] ep 50: test_acc=0.3528
  [BP-frozen] ep 60: test_acc=0.3701
  [BP-frozen] ep 70: test_acc=0.3750
  [BP-frozen] ep 80: test_acc=0.3855
  [BP-frozen] ep 90: test_acc=0.3855
  [BP-frozen] ep 100: test_acc=0.3896
FINAL BP-frozen-blocks: 0.3896

=== DFA shallow (ResMLP num_blocks=0), seed=123 ===
  n_params: 789770 (789770 trainable)
  [DFA-shallow] ep 1: test_acc=0.3348
  [DFA-shallow] ep 10: test_acc=0.3457
  [DFA-shallow] ep 20: test_acc=0.3484
  [DFA-shallow] ep 30: test_acc=0.3323
  [DFA-shallow] ep 40: test_acc=0.3442
  [DFA-shallow] ep 50: test_acc=0.3460
  [DFA-shallow] ep 60: test_acc=0.3469
  [DFA-shallow] ep 70: test_acc=0.3427
  [DFA-shallow] ep 80: test_acc=0.3432
  [DFA-shallow] ep 90: test_acc=0.3485
  [DFA-shallow] ep 100: test_acc=0.3485
FINAL DFA-shallow: 0.3485

=== DFA frozen-blocks (ResMLP num_blocks=4, blocks frozen), seed=123 ===
  n_params: 1318154 (789770 trainable)
  [DFA-frozen] ep 1: test_acc=0.3182
  [DFA-frozen] ep 10: test_acc=0.3388
  [DFA-frozen] ep 20: test_acc=0.3429
  [DFA-frozen] ep 30: test_acc=0.3286
  [DFA-frozen] ep 40: test_acc=0.3372
  [DFA-frozen] ep 50: test_acc=0.3281
  [DFA-frozen] ep 60: test_acc=0.3417
  [DFA-frozen] ep 70: test_acc=0.3483
  [DFA-frozen] ep 80: test_acc=0.3469
  [DFA-frozen] ep 90: test_acc=0.3479
  [DFA-frozen] ep 100: test_acc=0.3497
FINAL DFA-frozen-blocks: 0.3497

=== ResMLP frozen/shallow baseline summary, seed=123 ===
  BP-shallow:    0.3841
  BP-frozen:     0.3896
  DFA-shallow:   0.3485
  DFA-frozen:    0.3497

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)