Evaluation of Shapley-Residual Audits for Artificial Intelligence Agriculture Irrigation-Scheduling Models Under Sensor Dropout: An Interpretability Attribution-Stability Bound for Smallholder Water-Use Liability

Sensor dropout in smallholder irrigation telemetry destabilises the feature attributions on which auditability of AI scheduling models depends, yet the magnitude of that instability has not been bounded in a form usable by regulators. The bound is stated as a theorem, proved under explicit assumptions, and specialised to the before-after-control-impact design used to evaluate irrigation interventions. A replication-based evaluation of that design was conducted across a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) with 2000 replications per design cell over cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points; 56 clusters were required for 5 percentage points at 80% power. The analysis is methodological and theoretical: no observational data are used or claimed, and the reported quantities characterise the operating characteristics of the design rather than the behaviour of any deployed model. The bound implies that attribution stability, not predictive accuracy, is the binding constraint on liability assignment when dropout exceeds a threshold that the theorem makes explicit.