Evaluation of Concept-Probe Adversarial-Sufficiency Audits for Artificial Intelligence Agriculture Disease Classifiers Under Herbicide-Drift Symptom Mimicry: An Interpretability Confusability Bound for Smallholder Spray Decisions

Smallholder spray decisions in maize and cassava systems depend on distinguishing genuine pathogen symptoms from herbicide-drift phytotoxicity, and convolutional classifiers trained on farmer-submitted leaf images are increasingly used to make that distinction. Concept-probe audits, which test whether a classifier's internal representation supports the human-meaningful concepts a decision requires, are widely proposed as interpretability safeguards, but their sufficiency under symptom mimicry has not been characterised. The audit is modelled as a two-stage test: a probe must recover the target concept from the classifier's penultimate representation, and an adversarial substitution must fail to preserve the classifier's output when the concept is removed. A bound is derived that relates the audit's false-certification probability to the concept-probe gap, the mimicry overlap between drift and disease manifolds, and the audit sample size. A replication-based evaluation of the before-after-control-impact estimator was conducted with between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64, and 2000 replications per design cell across 8, 16, 24 and 32 clusters and effects of 3, 5 and 8 percentage points. Power reached conventional levels at 54 clusters for a 5-percentage-point effect. The bound implies that concept-probe audits certify classifiers that remain confusable under drift, and that spray-decision thresholds should be set against the audit's certified confusability rather than against raw accuracy.