Frozen development selection procedure
Recorded before running the development candidates on 2026-09-06.
For each outer fold separately, train M50 with run seed 42 for candidate final-update budgets of 500, 1,000, and 2,000. Extract embeddings only for that fold's select identities and compute identity/identity-pair weighted natural-population verification AUROC. Do not extract operate or test scores during development.
Choose the update budget with the highest select AUROC within each fold. If multiple candidates are within 0.001 AUROC of the best, choose the smallest such budget. Apply the selected fold-specific budget to all four adapted training conditions and all three run seeds in that fold. This preserves a common recipe across gender-composition conditions while preventing one fold's selection identities from determining another fold's test recipe.
All other numerical settings are predeclared from checkpoint compatibility and the local hardware preflight: FP32, frozen batch-normalization running statistics, SGD with backbone learning rate 0.001, head learning rate 0.01, momentum 0.9, weight decay 0.0005, ArcFace margin 0.5 radians and scale 64, cosine decay after 100 warmup updates, 50% horizontal flip, gradient clipping at 5, microbatch 16, accumulation 4, and effective batch 64. The final pair budgets are 20 genuine image pairs per identity and 10 impostor image pairs per distinct identity pair. Primary and secondary target FPRs remain 0.01 and 0.001.
Development artifacts use smoke: true solely to keep them outside the scientific export path and bypass the final frozen-config gate. They use the real iResNet50, checkpoint, corrected data, and split manifests; they are not synthetic smoke results. Only the selected final configuration may access operate and test roles.
Completed development selection
The predeclared search completed on 2026-09-06 without accessing operate or test. Select AUROC was:
| Fold | 500 updates | 1,000 updates | 2,000 updates | Selected |
|---|---|---|---|---|
| 0 | 0.940750 | 0.945904 | 0.956659 | 2,000 |
| 1 | 0.927029 | 0.938929 | 0.946627 | 2,000 |
| 2 | 0.933730 | 0.942854 | 0.949857 | 2,000 |
The resulting fold-specific budget vector is [2000, 2000, 2000]. The vector remains explicit in the frozen configuration so the scientific gate can verify that fold-local selection was completed; the identical realized budgets also preserve equal optimizer-update budgets across all adapted conditions and seeds.