Fewer Questions and Safer Abstention in Low-Bandwidth Agricultural Advisory

Advisory systems delivered over SMS, IVR, and USSD work under a hard inter- action budget, and the symptom reports they receive are noisy. Top-1 accuracy at a fixed number of turns is the wrong thing to optimize here: what matters is how few questions the system asks, and whether it knows when to stop answering. We formulate low-bandwidth agricultural inquiry as budgeted sequential severity staging conditioned on an agro-meteorological prior retrieved server-side, which costs the user nothing. We join 512 georeferenced cassava field surveys from 12 Nigerian states and the Federal Capital Territory to ERA5 reanalysis, condition the prior on climate windows favoring the whitefly vector, and evaluate query policies against an independently sourced symptom model under synthetic misreporting (r ∈ [0.0, 0.4]). Conditioning cuts expected turns to τ = 0.8 by 38.3%, from 3.89 to 2.40, and at r = 0.1 calibration error rises 125% while top-1 accuracy falls only 8.6%. We therefore report risk–coverage frontiers rather than accuracy alone: selective abstention determines whether such a system is safe to run.