Evaluation of Concept-Probe Causal Sufficiency Audits for Computer Vision Agriculture Disease Classifiers Under Sparse Field-Scouting Ground Truth: An Interpretability Identification Bound for Fungicide Triage

Concept-probe causal sufficiency audits ask whether a vision classifier's internal representation of a disease concept is sufficient to reproduce the classifier's own decisions, and they are increasingly proposed as a gate before a model is trusted to triage fungicide applications. Under sparse field-scouting ground truth, however, the audit's verdict is not a property of the model alone: it is jointly determined by the probe's identification power and by the sampling design that produced the labels against which sufficiency is judged. the present analysis formalises that dependence. A concept-probe sufficiency audit is written as a before-after-control-impact comparison of probe-implied decisions against classifier decisions, and its operating characteristics are derived under a declared variance structure with an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5. Power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The analysis yields an interpretability identification bound: 54 clusters are required to detect a 5-percentage-point sufficiency gap at 80% power, a requirement that exceeds every cluster count examined and that most field-scouting programmes cannot meet. The bound is presented as a design constraint on audit claims, not as a verdict on any particular classifier, and it implies that sufficiency audits under sparse scouting should report the minimum detectable gap alongside the audit outcome.