Benchmarks
Measured runs of compute repositories, each with its exact conditions, the repository version it ran and how far its evidence goes.
Evidence first. Performance in context.
Benchmark results mean something only when the hardware, model and run conditions are known.
- 01Self-reported
Published by the person who ran it. Kovara has not checked it and nobody else has reproduced it yet.
- 02Reproduced
Another engineer published a comparable run of the same version on the same hardware and settings.
- 03Independently verified
Kovara checked the evidence behind this run.
How evidence verification works
Every run starts self-reported: it is what the submitter measured on their own endpoint, published with its conditions. It becomes reproduced only when a different person publishes a run of the same repository version under identical conditions, and independently verified only when Kovara has checked its evidence. A run is never compared with runs measured under different conditions, and protocol tests (mock or controlled endpoints) are never ranked.