Spectral-Normalization Evaluation for Computer Agriculture Leaf-Disease Classifiers Under Field-Illumination Shift: An Artificial Intelligence Adversarial-Robustness Bound

Leaf-disease classifiers deployed in field agriculture must remain accurate when illumination changes between training and deployment, yet the standard evaluation protocol reports clean-image accuracy and provides no certified guarantee under such shift. The framework treats illumination variation as a bounded perturbation in a declared function space, derives a Lipschitz-controlled robustness radius from the spectral normalization of each layer, and converts that radius into a certified accuracy floor. A formal proposition establishes that a spectrally normalized network with layer-wise Lipschitz constants bounded by one admits a robustness radius that scales with the inverse of the accumulated Lipschitz product, and a corollary translates the radius into a lower bound on field accuracy. The bound is evaluated analytically across a declared grid of perturbation magnitudes and network depths; no observational data are used or claimed. The certified floor is non-vacuous only when the accumulated Lipschitz product is held below a threshold that depends on the illumination-shift magnitude, and that this threshold is reached at moderate depth. The contribution is a certified evaluation protocol that replaces clean-accuracy reporting with a robustness radius and an associated accuracy floor, together with an explicit statement of the conditions under which the floor is informative.