Evaluation of Concept-Orthogonality Audits for Computer Vision Agriculture Disease Classifiers Under Multi-Herbicide Tank-Mix Regimes: An Interpretability Feature-Entanglement Bound for Smallholder Phytotoxicity Liability

Computer vision disease classifiers deployed on smallholder farms frequently operate under multi-herbicide tank-mix regimes in which visual symptoms of phytotoxicity overlap with those of fungal and bacterial pathogens. Concept-based interpretability audits promise to detect such entanglement, but no formal bound links audit outcomes to liability exposure. The bound is derived from a decomposition of classifier sensitivity into concept-aligned and concept-orthogonal components, and a threshold rule is proposed for flagging classifiers whose orthogonal leakage exceeds a liability-relevant tolerance. Operating characteristics were evaluated across a design grid of 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per cell, with between-cluster SD 6.0, residual SD 4.5 and intraclass correlation 0.64. Power reached conventional levels at 52 clusters for a 5 percentage-point effect. The analysis indicates that concept-orthogonality audits can be calibrated to liability tolerances, but that smallholder deployments with fewer than 52 clusters yield underpowered audits and should not be treated as exculpatory. The contribution is a formal bound and a design-power characterisation, not an empirical audit of any deployed classifier.