Results
Where eval scores show up: per-run badges, aggregate dashboards, and time-series.
Eval results attach to the run or span they scored and roll up into per-evaluator dashboards.
On the run and span
Each scored run or span shows its eval results as badges on the detail header, with the score, pass/fail, and reasoning. Open a result to see the full explanation.
Per-evaluator dashboard
Each evaluator has a detail page with:
- A time-series of passed, failed, and errored results, with a cost line.
- A summary carousel: pass rate, p50 and p95 latency, and aggregates that match the return type (mean and range for numeric, true rate for boolean, top values for string), plus total and average cost.
- A 24-hour pass-rate stat on the evaluator list.
Windows: 24h, 7d, 30d, 90d, or custom. Results are gap-filled so the chart has no holes.
"Failed" and "errored" are different. A failed result ran and did not pass the rubric; an errored result could not produce a verdict (for example a judge timeout or an out-of-scale score). The dashboards keep them separate.
Costs
LLM-judge cost is a rough estimate, not billing-accurate. Treat it as a relative signal for which evaluators are expensive.
Next
- Evaluator types
- Agent analytics for eval rollups over time