Three Tests for an Adoption-Aware Crop Recommender: Sensitivity, Permutation, Placebo

We evaluated an adoption-aware crop recom- mender on Ethiopian Living Standards Mea- surement Study panel data (n = 8,744 cul- tivated plots) using three classical observa- tional tests: Rosenbaum signed-rank sensi- tivity, a sign-exchangeable permutation test on matched-pair differences, and a stratified label-shuffle placebo. The recommender tar- gets Expected Realized Benefit (ERB), the product of predicted yield gain and predicted adoption probability, against an accuracy- only baseline that maximizes yield alone. The aggregate ERB lift of +50 kg/ha (boot- strap 95% CI [+24, +74], n = 800 EA- disjoint test plots) clears Rosenbaum Γ⋆ ≥ 5 on every education subgroup and clears the permutation test at p < 10−3. The pub- lished per-education subgroup gradient does not clear the placebo: real per-subgroup deltas fall inside the placebo distribution at p ∈ [0.14, 0.87]. The placebo result re- frames a published equity-targeting claim as descriptive concentration of an aggregate ef- fect. None of the three tests is new. None is currently standard in observational offline- policy ML. All three should be.