MANNA v0.10.0 evaluation results: tool use and value comparison of an IVOA archive MCP server across four language models

Raw results, run logs and analysis scripts for the evaluations of MANNA v0.10.0, an MCP server that gives AI agents access to IVOA astronomical archives. Program I (tool-use evaluation) tests whether models use MANNA's tools correctly, by tier. Program II (value comparison) runs 23 archive tasks through Claude Code with and without MANNA, three runs per model and approach, for Nemotron 3 Super, Claude Haiku 4.5, Sonnet 5 and Opus 4.8, recording tasks passed, tool calls, tokens, cost and time. A custom agent loop provides a cross-check. Scripts regenerate every table and figure in the accompanying research note; README.md maps each one to its files. Local paths, the evaluating machine's IP address and the self-hosted model endpoint were redacted; no credentials were present.