Evaluation of Saliency-Map Disagreement Bounds for Computer Vision Agriculture Pest-Detection Models Under Pollinator-Conservation Constraints: An Interpretability Fairness Audit for Smallholder Damage Thresholds

Saliency maps are routinely used to justify pest-detection decisions in computer vision systems for agriculture, yet the disagreement between competing saliency methods is rarely quantified against the decision thresholds that smallholders actually apply. The framework defines a disagreement metric over attribution rankings, derives a finite-sample bound on the probability that two saliency methods disagree above a decision-relevant threshold, and links that bound to a damage-threshold calibration rule that penalises false-positive detections when pollinators are present. Because no observational data were supplied, the analysis is theoretical: it characterises the operating characteristics of the audit under a declared variance structure rather than reporting empirical estimates. The framework yields three propositions: a disagreement bound that scales with the number of attribution pairs and the decision margin, a fairness condition under which disagreement is invariant to pollinator-conservation status, and a calibration rule that maps the bound to a smallholder damage threshold. The contribution is a decision-theoretic account of when saliency disagreement is tolerable for pest-management advice, and a set of boundary conditions under which interpretability audits should block deployment rather than merely flag it.