Minerva X Tech

GeoScienceAI is a geoscience workspace that tries to reproduce research papers, then tests what a reproduction makes possible, and says honestly how far each result can be trusted. Papers rarely give everything needed to reproduce them, and AI agents tend to claim success. Here, Muse Spark reads a paper into a knowledge graph, where a quote counts only if it appears in the paper text. Muse, or Muse Code over an MCP server, writes and runs the code, and the app, not the agent, compares the numbers with the paper. Muse also sets up satellite maps with SAM 3.1 and CHMv2 canopy height. Descendance explores what a finished replication enables, and Workshop turns those ideas into testable hypotheses in your own IDE, with progress synced back. Everything lands in a local wiki and knowledge graph. On our test paper the honest result is “Early attempt”: none of 17 results is within 10%, mostly because the paper does not give its maps or scale, and the tool shows why. Built during the hackathon; the few engine files and the UPM data imported from my earlier OpenPalm project are listed in HACKATHON.md, section 3.