mira.eval

I do evaluation design for language models. Contamination checks, item response theory on benchmark items, and figuring out which benchmarks have quietly saturated. Most reported deltas sit inside run-to-run variance and almost nobody reports the variance.