Task: Diabetic Retinopathy | Classifier: LR @ 672 px (Table 2 best) | Budget: R$1,900,000 | 40 sites × 1,000 patients/yr | Prevalence 10% | Specificity 0.90 | Recapture 50%

DR AUROC: internal (Canon test) 0.940 → external (Nikon) 0.946; delta −0.006. DR ranks 10th most fragile of 12 tasks; most fragile is Hemorrhage (delta +0.235).

Task identity explains 67% of variance in the internal−external gap across 220 binary configurations; classifier contributes 1%, resolution 0%. Higher resolution (952 vs 224 px) shrinks the gap for 2/11 tasks and widens it for 6/11.

All-Canon: R$1,900,000 buys 20 Canon sites → 2,431 missed cases/yr (20 sites uncovered).
All-Nikon: R$1,400,000 covers all 40 sites → 955 missed/yr.
Best mix: 8 Canon + 32 Nikon at R$1,880,000 → 937 missed/yr.

Equity ratio: a diseased patient at a Nikon site is 1.11× more likely to be missed than at a Canon site, driven by Nikon's higher inadequacy rate (18.60% vs 8.79%). For DR at this budget, all-Nikon dominates because full coverage outweighs the accuracy and quality differences.