<![CDATA[Crop advisory models can perform well on balanced benchmarks while failing under unfamiliar environmental or household conditions. This study evaluates two separate research tracks: crop-label classification using 2,200 agronomic benchmark records and observed crop-portfolio prediction using 43,753 village-census households and 981 TAFSSA households in Nepal. Eight crop classifiers were compared within a training pool, with independent temperature-calibration, conformal-calibration, and test subsets. Household models were evaluated under random and geographic holdouts, including 30 paired community-level ablations. An external explanation-consistency experiment used 900 held-out, human-annotated RAGTruth responses. Random forest attained the highest cross-validated macro-F1 (0.992), while LightGBM had the lowest conventional-model log loss (0.059). Temperature scaling produced negligible change in held-out log loss. At 90% nominal coverage, adaptive prediction sets included the benchmark label in 438 of 440 cases, with a mean of 1.09 crops. Environmental-cluster holdouts reduced macro-F1 to 0.000–0.656, partly because some test crops were absent from training. Village-census logistic-regression macro average precision declined from 0.680 under random partitioning to 0.578 under district holdout for the same test households. TAFSSA cross-district values were 0.332 and 0.445. A fixed-split contextual gain did not persist across community holdouts: full context reduced macro average precision by 0.037 relative to farm-only predictors. On RAGTruth, numeric checks detected 132 of 579 erroneous responses; a source-linked lexical screen detected 432 but also rejected 102 of 321 valid responses. The findings support cautious interpretation of benchmark shortlists and household associations. The components remain a research prototype without end-to-end or field validation of agricultural decisions in Nepal.]]>
Copyrights © 2025
