This is the authors' abstract. We don't add key points for this paper.
Smallholder maize growers in hail-prone regions face a triage problem that is simultaneously statistical and infrastructural: a damage assessment must be accurate enough to support an insurance payout decision, yet it must be produced on a handset that may have no connectivity at the moment of inspection. The framework treats a large teacher ensemble as a source of soft labels, distils it into a compact student classifier, and calibrates the student's predictive entropy so that a rejection threshold carries a distribution-free coverage guarantee under exchangeability. A payout-latency model is then coupled to the coverage bound, so that the expected time to a disbursement decision is expressed as a function of the abstention rate, the local inference cost and the probability of a connectivity window. The analysis is methodological and theoretical: no field data are reported. The contribution is a set of propositions that relate distillation temperature, calibration error and the risk-coverage frontier, together with a latency decomposition that separates on-device computation from store-and-forward delay. The framework predicts that calibration, rather than raw accuracy, governs the achievable coverage at a fixed risk level, and that intermittent connectivity shifts the optimal abstention rate upward relative to the connected case.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Gueddi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: