prob average (uniform) siglip_so400m+eva02_large vs best solo siglip_so400m

p = 0.000003

✓ significant

198 images only the first got right, 115 only the second.

prob average vs logit average (all members)

p = 0.546713

! not significant

53 images only the first got right, 46 only the second.

SigLIP-SO400M vs EVA-02-L

65.2%

of the first model's errors are shared with the second

both correct

94.42%

only first

2.40%

only second

1.10%

both wrong

2.07%

SigLIP-SO400M vs DINOv2-L

66.3%

of the first model's errors are shared with the second

both correct

93.80%

only first

3.03%

only second

1.07%

both wrong

2.10%

EVA-02-L vs DINOv2-L

59.6%

of the first model's errors are shared with the second

both correct

93.06%

only first

2.46%

only second

1.81%

both wrong

2.67%

Frozen pair (deployed)

97.156%

two small heads over cached features

Three-way, val-tuned

97.267%

p = 0.052 — not significant

Shared errors

Ensembling only pays when members fail on different images. This is the number that decided it.

At 224px the fine-tune shared 65.5% of SigLIP's errors — no better than the frozen backbone. The single 448px epoch dropped that to 60.9% while adding just 0.2 points of validation accuracy. Its contribution was diversity, which no accuracy column shows.

Training curve

Validation top-1 per epoch. The resolution change is the last step.

• 224px · epoch 184.19 (52 min)
• 224px · epoch 287.43 (52 min)
• 224px · epoch 389.27 (52 min)
• 224px · epoch 491.55 (52 min)
• 224px · epoch 592.11 (52 min)
• 224px · epoch 692.74 (52 min)
• 448px · epoch 192.94 (211 min)