=================================================================
BASE TIER PATCHWORK SOUP
  3 experts × 768-d → 256 anchors × 128-d
  Device: cuda
=================================================================

=================================================================
LOADING DATA
=================================================================
  Train: 118,287  Val: 5,000
  clip_l14_openai                loaded
  dinov2_b14                     loaded
  siglip_b16_384                 loaded

=================================================================
BUILDING MODEL
=================================================================
  Parameters:
    projectors     :    296,064
    constellation  :     32,768
    patchwork      :    100,864
    classifier     :    370,256
    total          :    799,952

=================================================================
TRAINING
  20 epochs, lr=0.001, batch=128
  Adam, no weight decay (geometry IS the regularization)
=================================================================
  E 1: mAP=0.732 F1=0.609 cv=0.4301 agree=[0.977,0.986,0.974] loss=0.0845 ★
  E 2: mAP=0.768 F1=0.679 cv=0.4308 agree=[0.982,0.988,0.980] loss=0.0521 ★
  E 3: mAP=0.783 F1=0.726 cv=0.4158 agree=[0.984,0.988,0.981] loss=0.0485 ★
  E 4: mAP=0.787 F1=0.714 cv=0.3889 agree=[0.986,0.989,0.982] loss=0.0465 ★
  E 5: mAP=0.799 F1=0.742 cv=0.3760 agree=[0.987,0.989,0.983] loss=0.0453 ★
  E 6: mAP=0.805 F1=0.738 cv=0.3849 agree=[0.987,0.989,0.983] loss=0.0444 ★
  E 7: mAP=0.804 F1=0.737 cv=0.4140 agree=[0.988,0.989,0.984] loss=0.0437
  E 8: mAP=0.806 F1=0.748 cv=0.3945 agree=[0.989,0.989,0.984] loss=0.0432 ★
  E 9: mAP=0.807 F1=0.735 cv=0.3855 agree=[0.988,0.989,0.984] loss=0.0427 ★
  E10: mAP=0.809 F1=0.751 cv=0.3396 agree=[0.989,0.989,0.985] loss=0.0423 ★
  E11: mAP=0.815 F1=0.746 cv=0.4014 agree=[0.989,0.989,0.985] loss=0.0417 ★
  E12: mAP=0.812 F1=0.729 cv=0.3759 agree=[0.989,0.989,0.985] loss=0.0414
  E13: mAP=0.818 F1=0.750 cv=0.3313 agree=[0.990,0.990,0.986] loss=0.0411 ★
  E14: mAP=0.817 F1=0.742 cv=0.3890 agree=[0.990,0.989,0.985] loss=0.0408
  E15: mAP=0.818 F1=0.749 cv=0.3872 agree=[0.990,0.989,0.985] loss=0.0405 ★
  E16: mAP=0.816 F1=0.746 cv=0.3667 agree=[0.990,0.990,0.986] loss=0.0402
  E17: mAP=0.824 F1=0.753 cv=0.3517 agree=[0.990,0.989,0.985] loss=0.0399 ★
  E18: mAP=0.818 F1=0.747 cv=0.3221 agree=[0.990,0.990,0.986] loss=0.0397
  E19: mAP=0.825 F1=0.746 cv=0.3117 agree=[0.990,0.989,0.985] loss=0.0396 ★
  E20: mAP=0.816 F1=0.738 cv=0.3320 agree=[0.990,0.990,0.986] loss=0.0392

  Best mAP: 0.825
  Model: 799,952 params
  Anchors: 256 × 128-d

=================================================================
DONE
=================================================================