EVALUATE / Public workbench
Whetstone
Did the change earn its release?
Compare your current and candidate AI system on the same checks. Keep every gain and regression visible.
AVAILABLE IN THIS RELEASE
- Download drafts and retain an answer archive after grading or expiry.
- Import answer files and catch malformed patches before spending a cohort.
- Handle expired server sessions without losing your answers; publication stays off by default.
Paired outcomes and an explicit PASS, HOLD, or BLOCK decision.
- 01
Start a shared cohort in the Open Promotion Bench.
- 02
Run each task with your baseline and candidate, keeping their configurations fixed.
- 03
Submit both answer sets. Inspect gains, regressions, scope changes, and the resulting receipt.
WHAT THIS DOES — AND WHAT IT DOESN’T
A useful tool has a clear boundary.
The public benchmark uses six procedural tasks and self-attested system identities. A public run is not a private evaluation or proof of general model quality.