Machine-learning models for blood-smear diagnosis routinely report accuracies above 95%, but most studies split images, not patients, into training and test sets, so cells of the same person appear on both sides. We compare five model families under one leakage-free protocol: Random Forest and a support vector machine on 135 handcrafted colour, texture and shape features, a custom convolutional network, transfer-learning CNNs (EfficientNet-B0 and ResNet-50) and a vision transformer (DeiT-Small). Two public tasks are used: acute lymphoblastic leukemia versus normal B-cells (C-NMC 2019, 12,418 cells, 93 subjects) and parasitized versus uninfected red cells (NIH malaria, 27,558 cells, 200 patients). Splits are disjoint by subject; a pixel-hash audit of C-NMC found 110 identical images filed under different patient IDs, which we merged. Every model is trained with three seeds and evaluated by AUC, accuracy, sensitivity, specificity, F1 and balanced accuracy, with DeLong and McNemar tests and patient-level bootstrap intervals. On C-NMC, ResNet-50 (AUC 0.9141 ± 0.0168) and DeiT-Small (0.9066 ± 0.0199) lead and do not differ significantly (DeLong p = 0.190 after Holm correction); the handcrafted-feature models reach 0.867–0.876. On malaria all models exceed AUC 0.98; DeiT-Small is highest (0.9922 ± 0.0008), but Random Forest (0.9896 ± 0.0003) makes as few errors at a validated threshold. Splitting by image instead of by patient inflates the C-NMC AUC of EfficientNet-B0 from 0.880 to 0.974 and of Random Forest from 0.866 to 0.931, close to the values reported in recent literature. In malaria, 79% of the cells missed by all six models come from a single patient. Leakage-free protocols, not larger models, are what separate realistic from optimistic results.