Experiments on IamBenSchmidthttps://iambenschmidt.com/oakmontlab/experiments/Recent content in Experiments on IamBenSchmidtHugo -- gohugo.ioenCopyright Ben SchmidtFri, 24 Jul 2026 00:00:00 +0000What is a reported 0.88 grading correlation worth? A reanalysis of the BeSTraP short-answer datasethttps://iambenschmidt.com/oakmontlab/experiments/2026-07-grading-robustness/Fri, 24 Jul 2026 00:00:00 +0000https://iambenschmidt.com/oakmontlab/experiments/2026-07-grading-robustness/A fine-tuned grader reported at 0.88 holds up on every mechanism the released data lets us check. It is competent, less consistent than a second human, and not inflated by question difficulty. Its one real limitation is that the headline cannot be reproduced from the release.