An irrigation controller must decide each morning how much water to apply, yet a soil probe and a rain forecast tell it only roughly how dry the root zone will become over the next few days. We forecast quantiles of a crop’s water-stress margin over a three-day window, calibrate them per crop with split conformal prediction, and irrigate the calibrated (1−α) upper bound. The risk level α becomes the single setting a grower chooses. We tested the rule in closed-loop AquaCrop simulations of 66 site–crop–soil seasons in 2025 at eight sites in Spain, the United States, Tunisia and India, driven by archived operational weather forecasts. With a quantile regression forest at α = 0.1, the rule applied as much water as an FAO-56 checkbook (293 vs. 288 mm per season, p = 0.46) and cut stress days from 9.2 to 4.3 (p < 0.001). Against baselines tuned per crop on 2024, it used 31 mm less water than FAO-56 at equal stress, and a 3.8 kB int8 network used 83 mm (21%) less than a tuned soil-moisture threshold at equal stress. On 25 NOAA USCRN stations with real probes, the calibrated 90% upper bound covered 89–91% of test days. The bound failed where the forecaster under-covered the test year: winter wheat, and probes noisier than those seen during calibration.
 

irrigation scheduling, conformal prediction, quantile regression, soil moisture forecasting, decision-making under uncertainty, AquaCrop, TinyML