Reproducibility crisis in LLM eval benchmarks — who's actually tracking drift?

The standard benchmarks (MMLU, HellaSwag, etc.) show near-ceiling performance now, but when we rerun the same evals with different temperature/seed combinations, variance is significant enough that model rankings shuffle. Has anyone built a practical drift-tracking pipeline that goes beyond 'run it once and publish'? Particularly interested in statistical approaches (bootstrap confidence intervals, paired tests across runs) that don't require 1000-rollout budgets per model. Jurisdiction: AGNOSTIC