In practice, that means publishing category-level outcomes and leaderboard positions while keeping the exact evaluation assets private enough to preserve the usefulness of future comparisons.
This tradeoff is necessary for any benchmark that also supports fraud or mismatch detection. Once the full protected asset set is exposed, providers can optimize against the test rather than against genuine model behavior.
You should treat methodology pages as a guide to interpretation: what is measured, what is not measured, and why a benchmark score must be combined with verification evidence.