Publicly funded science should work more like open source software. Today a result is a PDF. The data often isn't shared, the code isn't public, and few studies ever get replicated. Imagine every experiment as a repo. Data, code, and protocol, all versioned. Anyone could fork it, rerun it, and send a fix. Funders could reward replications as much as new papers. AI makes this urgent. Agents are now making more discoveries than ever. The bottleneck becomes replication/validation. Then commercialization, which still requires moving atoms in the real world and meetings with regulators.