One cohort at a time
Task, appliance, resolution, data access, scope, profile, and sample limit define who can be ranked together.
Living benchmark
Leaderboard
Ranks only compare runs in the same evaluation cohort. Every row comes from immutable result files; the page does not contain hand-entered scores.
- Source runs
- —
- Aggregates
- —
- Models
- —
- Cohorts
- —
Loading leaderboard.json…
How to read the board
The rank has boundaries.
Smoke is not full
Smoke runs check the real pipeline cheaply. Only complete full-protocol runs support benchmark or paper claims.
Three required seeds
Verified aggregates require seeds 10, 20, and 42 plus source, dataset, runtime, container, and protocol provenance.