Frozen crop-yield experiment artifacts: training-coverage and trend-estimator extension
Abstract
Version 0.2.0 in the original crop-yield dataset series provides the additional training-coverage, parameter-sensitivity and trend-estimator experiments for Training coverage and trend estimation jointly shape county-level crop-yield prediction. It contains the processed four-panel county-yield input and feature manifest, 1,860 raw/county grid checkpoints, 1,860 global-trend checkpoints, 60 historical calibration records and the required numerical result tables. Code is supplied separately in software version 0.2.0; extract both archives into the same parent directory. The original fixed-setting predictions already occur in the raw/county grid and are not duplicated. Original Phase 2 OOF and fold-plan archives remain in dataset version 0.1.0; this update does not reupload them. No manuscript drafts, review correspondence, literature, figures, operational logs or credentials are included. Observations span 2000–2024 in Iowa and Illinois corn/soybeans, with forward tests 2010–2024 and weather through day 304. The comparisons and unadjusted intervals are exploratory and conditional on stored predictions and five masks. Historical calibration used pre-mask history. A second complete training run was not performed. Upstream data-provider terms and attribution remain applicable.
