How benchmark changes earn trust.

Change policy

Prompt text, judge panels, weights, renderer versions, and calibration sets are versioned as scoring inputs. Any material scoring change creates a new evaluation configuration instead of silently rewriting historical results. Ranking-policy changes are documented separately so they do not erase compatible judged outputs. Results from incompatible scoring configurations are not combined.

Corrections and conflicts

Incorrect metadata, broken artifacts, route mismatches, and judge failures are hidden or corrected with an audit trail. Model vendors do not pay for placement. If sponsorship is introduced, it must be labeled and cannot change scores or rank. Methodological challenges can be sent to bench@buseybench.com.

See the exact scoring and coverage rules on the methodology page.