Evaluation Results
Results presents a matrix of selected Runs and the Cases or criteria they evaluated. Use one Run for inspection or several comparable Runs to locate gains, regressions and missing judgments.
Prerequisites
- At least one Run for the intended Benchmark Version.
- Permission to inspect its Cases, responses and judgments.
- Comparable recorded conditions when interpreting differences.
Steps
- Open Benchmark Evaluations → Results and confirm the Benchmark Version.
- Select the visible Runs. Managed and imported Runs can both appear; an imported Run does not need a Harness Version to be inspected.
- Choose Cases, Policies, Rubrics, Dimension ontology values, Topics, Topic Groups or Scenarios as the row grouping.
- Read completeness before comparing scores. An empty or incomplete result is not a failed judgment.
- Open a Case or rubric detail to inspect the response, criterion, observed judgment and rationale.
- For repeated Trials, inspect the selected aggregation rule and the individual Trial outcomes.
- Route an actual candidate problem to Harness improvement, a standard problem to Correctness Governance, or an input-coverage problem to Coverage Design.
Representative and historical Runs
Default candidate results use the Run selected before its outcomes were known. Used for score identifies that reference. Other executions remain in Run history and can be inspected without silently replacing the representative result.
Selecting matrix columns changes the analysis view. It does not activate a Harness, execute a candidate or declare a winner.
Object and state changes
Filtering, row grouping and inspection are read-only. An authorized project manager can separately use Correct judgment on an exact rubric result. That operation records a correction and preserves the original AI judgment; refreshed aggregates come from the authoritative result snapshot.
Success criteria
- Every compared column has an identified Run and Benchmark Version.
- Differences resolve to actual Cases and applicable criteria.
- Missing outcomes stay distinct from negative scores or failed gates.
- Manual corrections remain attributable.
Common failure modes
- Treating an imported Run as an executable Harness.
- Combining later successful Trials into an earlier reference score.
- Reading a filtered population as the entire benchmark.
- Changing a rubric judgment when the real problem is an incorrect Case or standard.