TS-MCP benchmark: preregistered protocol and sealed run plans for language models on Tanabe-Sugano ligand-field tasks

This preregistration fixes the hypotheses, design, run plans and analysis of the confirmatory study of the TS-MCP benchmark before any model has seen a confirmatory task. Language models determine ligand-field parameters (Dq, B, C) of transition-metal complexes from absorption bands, compute Tanabe-Sugano curves and diagrams, and recognise questions outside the ligand-field model. Three conditions differ only in their tools: no tools (A), a sandboxed code interpreter (B), and a domain MCP server built on the tanabesugano package (C). The confirmatory split has 55 tasks, identified by hash only. The record contains the preregistration document and the sealed run plans (headers and hashes, no task text and no prompts).