Skip to main content

​
Review evaluation results

After a run, stay on the evaluation page. The current UI separates the work across Analysis, Human review, Lineage, and the run summary instead of a standalone results-table page.

​
Start with the completed run

The Run evaluation card shows the latest completed status and pass-rate summary. Choose View run history to move into Analysis. Do not collapse these states into a pass:
  • failed behavior,
  • cancelled or partial execution,
  • blocked/inconclusive infrastructure,
  • missing or stale evidence.

​
Use Analysis for run evidence

In Analysis, inspect the run and case-level outcomes that explain the summary. Compare the failing cases with the exact run/attempt rather than copying only a headline score. When cost, latency, scorer output, or provenance is available, interpret it as evidence attached to that run—not as proof that an unrelated requirement passed.

​
Use Human review when judgment is required

The Human review tab can create reviewed labels from failed run cases or from a deliberate manual example. A linked review preserves its source run, result, case, attempt, expected output, recorded outcome, and trajectory where available. Editing the source input turns it into a manual example rather than silently retaining false lineage. Reviewed examples can then feed the governed candidate/golden-case workflow.

​
Follow lineage

Use Lineage when you need to understand where a case, prompt, dataset, evaluation artifact, or promoted example came from. Incomplete lineage should stay visible as incomplete evidence.

​
Up next

Turn a real production failure into candidate coverage from Evidence → Logs.