GeoNatureAgent Benchmark - Dataset (tasks, results, derived environmental indicators)

GeoNatureAgent Benchmark — Dataset. Benchmark task suite (93 tasks across 18 categories); per-case scored results for the seven evaluated LLMs (three temperature-1.0 seeds each; two samples for Claude Sonnet 4); the compiled leaderboard / per-category / per-case tables; Spanish administrative boundaries; and BigEarthNet V2 land-cover statistics for Portugal. Reproduces the leaderboard reported in the accompanying paper. Third-party source datasets (BigEarthNet V2, LUCAS 2022, MITECO, Sentinel-2) are cited, not redistributed. Raw conversation traces and the CO2/erosion raster COGs will be added in a subsequent version. Evaluation code is in the companion software record.