Evaluation of Shapley-Interaction Attribution Audits for Computer Vision Agriculture Irrigation-Scheduling Models Under Probe Calibration Drift: An Interpretability Stability Bound for Smallholder Water-Allocation Equity

Computer vision irrigation-scheduling models allocate water among smallholder plots by classifying canopy stress from imagery, and their interpretability audits increasingly rely on Shapley-interaction attributions computed against concept probes. When those probes drift because field calibration changes between seasons, the audit's own stability is rarely characterised. No observational data are used or claimed: the analysis characterises the audit's behaviour under a declared variance structure with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.640, across 2000 design cells spanning cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The bound decomposes attribution instability into a probe-drift term, an interaction-order term and a sampling term, and predicts that 54 clusters are required for 5 percentage-point effects at 80% power. Audit conclusions about smallholder water-allocation equity can invert when probe drift exceeds the interaction-order term, and it specifies the drift threshold beyond which Shapley-interaction audits should be treated as non-diagnostic for equity claims.