A traceable correction for frozen yield models, scored forward in time.
2,062 field-seasons3 crops4 countries2016–2024
Across all fields
Where the error lives
When it helps
Year by year
With same-season labels
Limitations
Only 3–4 test years per crop. Intervals describe these observed years; they do not establish future-year reliability.
Corn, soybean and wheat. Other crops and regions are untested. The three worked fields were selected after the fact, for illustration.
Replay checks the sources and prediction composition. It does not validate the selector’s internal reasoning.
A uses same-stage support labels, which can arrive later in calendar time than a query. B uses no same-season yield labels.
Exploratory, retrospective evidence; no prospective validation or multiple-comparison adjustment. An interval crossing zero does not establish equivalence. Year effects are entangled with annual Base checkpoints.
SO/SA ensemble versus seed-mean reporting remains an author decision. Existing website values retain their stated convention; neither estimates nor intervals are mixed across conventions.
Comparisons change inputs, model capacity and selection procedures together. B excludes parameter-based adaptation by scope. GT-based diagnostics cannot serve as deployment rules; Shapley applies only to its analysed retrieval object.
Explorer
Pick a correction, set λ, lock. Then see the harvest.