For labs
Evaluation sets nobody has ever seen
Every public benchmark leaks. We license sealed holdout sets: tests written by our community, never published, run against your model and reported as results only. The prompts never leave our hands, so they cannot end up in anyone's training data, including yours.
Sealed holdout sets
Curated, versioned sets that were never public. You get scores, pairwise ratings and per-category breakdowns, not the prompts.
Pre-release evaluation
Run a model before launch against fresh tests written after its cutoff, and see where it stands against the current roster.
Custom sets
Commission tests in a domain you care about, written to be unseen and checkable, and retired when they saturate.