Certifiable Robustness Bounds for Artificial Intelligence Cassava Disease Classifiers Under Field Image Corruption: A Computer Guarantee Evaluation

Cassava disease classifiers trained on curated leaf images lose accuracy under field corruption, and point estimates of that loss carry no guarantee. The framework is evaluated as a computer guarantee rather than as an empirical field trial: the analysis characterises the certified radius distribution, the abstention rate and the guarantee coverage under a declared corruption model spanning Gaussian noise, motion blur, defocus blur and JPEG compression at fixed severity levels. The evaluation is a design and power analysis of the certification procedure, with 2000 replications per design cell across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, under a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64. The certified radius is governed by the smoothing parameter and the class-conditional margin, that the guarantee is vacuous below a computable margin threshold, and that the abstention rate rises steeply as the corruption severity increases. The contribution is a formal guarantee whose assumptions are stated explicitly and whose failure modes are identified, together with a design analysis showing that 54 clusters are required to detect a 5 percentage point effect at 80% power. The framework does not replace empirical validation; it specifies what a robustness claim about a cassava classifier can mean when the claim must hold for every input within a certified radius.