Run an evaluation
The evaluation detail page contains a Run evaluation card. It shows whether the evaluation is ready, how many cases are available, the current or previous attempt, and any configuration blocker.Step 1 — Check readiness
Open Evaluations and choose the evaluation. Before a run can start, the page must have at least one test case and a runnable target. If repository planning selected checks but execution is not yet authorized, the page asks you to bind an approved executor, target, and evaluator instead of silently running them.Step 2 — Choose an environment
In Run evaluation, select:devfor the everyday feedback loop,stagingfor pre-production behavior, orprodwhen your session is allowed to launch a production-tagged run.
Step 3 — Follow the active attempt
The same card polls the active run. It shows processed cases, pass/fail counts, progress, heartbeat information, and an estimated remaining time when available. A cancelled, failed, partial, or otherwise incomplete attempt is not a verified winner.Step 4 — Inspect results and history
Use View run history or open the Analysis tab to investigate completed attempts and failures. The evaluation page does not have a separate Report tab in the current UI. Release decisions should use the persisted run/evidence state, not the fact that a request returned successfully.Troubleshooting
- Configuration required — use the executor/target setup shown in the run card.
- A model-backed target cannot run — check Settings → AI providers and the scorer/target configuration.
- Cost or provider policy blocks the run — resolve the named policy/budget condition rather than bypassing it.
- The run stays queued/running — inspect the active progress and the owning runtime/Remote Runner if one is involved.