Reduced-form climate–yield panels allow nonlinear temperature effects but often treat precipitation as a seasonal total over a fixed calendar window. This paper estimates precipitation responses for rainfed U.S. corn and soybean from 1981 to 2022 using county-year yields, PRISM daily weather, and NASS Crop Progress phenology. We construct dynamic growingseason windows from planting to physiological maturity, estimate generalized additive models, and allow precipitation to affect yield differently across cold, moderate, and heat regimes through tensor-product smooths with the matching degree-day exposure. The dynamic window lowers measured growing-season precipitation by 25.1% for corn and 34.8% for soybean and removes 85.0%and82.3%offixed-window cold exposure. The regime specification does not increase out-of-sample mean squared error relative to an additive dynamic-window GAM. Estimated rainfall responses vary by thermal context: moderate-regime rainfall has a humpshaped marginal effect with zero-crossings near 380 mm for corn and 425 mm for soybean, while heat-regime rainfall has positive dry-end effects, especially for corn. An adaptation-gain counterfactual, which holds the fitted response surface fixed and re-evaluates climate inputs under a 1981–1985 baseline phenology window, turns positive after the 1980s for both crops. Decadal gains peak at 0.44 billion 2022 dollars for corn in 2001–2010 and rise to 0.55 billion for soybean in 2011–2022. These estimates bundle deliberate calendar choices with passive phenological movement, so they are not causal estimates of farmer behavior. The results show that precipitation effects depend on when rainfall occurs within the temperature distribution, andthat historical phenology shifts have already changed the value of growing-season weather.
