Add REFUTE scientific critique + calibration benchmark

[connerlambden]

Proposal

Add REFUTE to related evaluation / RAG / retrieval / scientific AI tooling docs if relevant.

REFUTE is a scientific critique + calibration benchmark (paper-grounded claims → predictions → judge scores → Brier/ECE).

• Product: https://bgpt.pro/refute
• Inspect task: https://github.com/connerlambden/refute-inspect
• Dataverse (in review): https://doi.org/10.7910/DVN/IGAJWO
• ASCL: https://ascl.net/code/v/4795
• bio.tools: https://bio.tools/bgpt-refute

Happy to adjust wording / PR if preferred.