Related eval: REFUTE scientific critique + calibration benchmark

[connerlambden]

Proposal

If this project maintains a related-evals / resources / docs section, please consider linking REFUTE: a scientific critique + calibration benchmark (paper-grounded claims → model predictions → judge scores → Brier/ECE). Critique skill and calibration often come apart.

• Product: https://bgpt.pro/refute
• Inspect task: https://github.com/connerlambden/refute-inspect
• Interim preprint: https://github.com/connerlambden/refute-inspect/releases/tag/v3.0.0-preprint
• bio.tools: https://bio.tools/bgpt-refute
• ASCL: https://ascl.net/code/v/4795
• Dataverse (in review): https://doi.org/10.7910/DVN/IGAJWO

Happy to open a PR with exact wording if that helps. Contact: contact@bgpt.pro