Evaluation of Attention-Rollout Faithfulness Audits for Computer Vision Agriculture Disease Classifiers Under Lesion-Region Occlusion: An Interpretability Attribution Bound for Smallholder Fungicide Triage

Attention-rollout saliency is widely used to justify computer-vision disease classifiers in smallholder fungicide triage, yet its faithfulness under lesion-region occlusion has not been bounded. the present analysis derives an attribution bound that links rollout faithfulness to the occlusion geometry of the lesion region and the classifier's decision margin, and characterises the operating characteristics of a before-after-control-impact audit design under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). The audit design was evaluated across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per design cell; 56 clusters were required for 5 percentage points at 80% power. The derived bound shows that rollout faithfulness is not a property of the saliency map alone but a joint property of the map, the occlusion mask and the classifier's margin, and that the audit design attains adequate power only above a threshold cluster count that the bound makes explicit. The contribution is a formal statement of when attention-rollout audits can license triage claims and when they cannot, together with the design parameters under which such audits are informative for smallholder fungicide decisions.