Skip to main content

​
Pick a winner

Once a Playground run shows a candidate is better, you promote it. Winner selection is an admin action and is recorded with the exact run, variant, and decision.

​
Step 1 — Open the winning run

On the Playground page, open the completed run that proves your candidate wins.

​
Step 2 — Select the winner

Make the candidate the Active variant, then click Select winner. A winner needs a completed run of that variant and a bound judge. EvalGate records:
  • the exact run snapshot,
  • the selected variant ID,
  • the request hash,
  • the decision actor and rationale.
Winner selection is admin-only. A non-admin can run the Playground but cannot promote a variant.

​
Step 3 — Use the winner downstream

The winner feeds the next stages of the eval loop:
  • An Experiment can be created from this Playground to confirm the win at scale.
  • A deployable artifact can be derived from the winning prompt, model, and parameters.
  • Verification (/eval-loop) records this Playground as the “Playground reproduction” stage.

​
What a winner is not

  • It is not a deployment. You still publish and deploy the artifact separately.
  • It is not irreversible, but the decision is immutable in history. To change direction, run a new Playground and pick a new winner.

​
Up next

Carry the winner through Deploy an artifact.