I stopped fixing my prompt by hand. Open models fixed it while I napped

Two optimizer runs over MCP, all open-weight models, 308 held-out bank messages: Gemma 3 4B went from 37% to 60% correct for $0.38.