Bolivia Fire benchmark data and reproduction artifacts

Bolivia Fire version 0.1.0 provides a weekly VIIRS S-NPP standard-processing satellite-detection benchmark for Bolivia and its retained reproduction evidence. The strict panel contains 37,146 cells, 678 Monday issues from 2013-01-07 through 2025-12-29 and 25,184,988 cell-weeks. The target is a selected satellite detection, not a count of fires, ignitions or burned area, and is not an official warning. Three independently extractable ZIP64 archives use repository-relative paths: benchmark-data: panel, cells, CHIRPS rainfall, MOD13A2 vegetation, MapBiomas Collection 4 land-cover and combined features, schemas and manifests. reproduction-artifacts: configurations, environment lockfile, fitted models and calibrators, retained predictions and alternative-learner cached vectors, evaluation history, figure JSON, family-specific producing revisions and historical source registries. raw-inputs: FIRMS CSV snapshots, MOD13A2 AppEEARS subset, CHIRPS native windows, MapBiomas aggregate exports, boundaries, GeoNames and outage HTML. Top-level README, licence notices, data dictionary, member inventory and SHA-256 checksums describe the files. Dataset reuse is supported. Recalculation from saved predictions is partially supported: strict primary validation/test exports and alternative-learner cached vectors are retained; rainfall, sensitivity and vegetation, land-cover and combined covariate per-row prediction exports and standalone scored-row masks are absent. Their result records and fitted models remain available. Recomputing any final-test metric is a logged test opening under the frozen protocol; release verification does not open the test. Rebuilding from raw inputs is partially supported. Global CHIRPS rasters, MODIS HDF granules and native MapBiomas 30 m maps are not supplied; MapBiomas re-extraction requires an Earth Engine account. Operational live forecasts, serving derivatives, regenerable caches, temporary files and logs are excluded. Later environmental experiments are separate from the original CHIRPS comparison; retained 2026 availability evidence does not establish a completed confirmation evaluation. Known discrepancies include an older unmatched cell-place manifest hash, fourteen unretained intermediate states, the MOD13A2 acquisition AOI buffer of 0.05 degrees versus the registered 0.1 degrees, and CatBoost GPU digit-level nondeterminism. See the record README for details and producing-code revisions. Load serialized model files only from trusted sources. Compiled and derived data are licensed under CC BY 4.0. Upstream raw inputs retain NASA open-data/public-domain or attribution terms detailed in LICENSES.md, including CHIRPS, MapBiomas and GeoNames CC BY 4.0 and GeoBolivia public-domain boundaries. Code is separately MIT licensed at Cupesi/bolivia-fire. Cite the dataset version and the software separately.