Skip to main content
After tests run, you’ll spend most of your time on the Results tab. This page explains what every column means and how to get the most out of the filters.

​
The results table

Each row is a single test run. Click any row to open the full test detail panel.

​
The test detail panel

For each test, you can see:
Exactly what was sent to the test model.
The unedited reply from your test model.
Verdict, confidence, and reasoning — for all three judges, side by side.
A breakdown of how the voting method combined the three judges into a final verdict.
Who overrode the verdict, when, the new verdict, and the reasoning they gave.
How many tokens each judge used, and how long the whole test took.

​
Filters

Use the filter bar at the top of the Results tab to narrow results. You can combine filters. The URL updates as you filter, so you can bookmark or share a filtered view.

​
Effective verdict vs final verdict

This distinction matters:
  • Final verdict = what the AI judges agreed on
  • Effective verdict = what is actually used in dashboards and reports
If a human has overridden the test, the effective verdict is the human verdict. Otherwise it’s the final (AI consensus) verdict. This is important because:
  • Dashboards show the effective verdict — so they reflect human judgment when available
  • Risk assessments count failures based on the effective verdict
  • Compliance reports use the effective verdict
In the UI, you’ll always see both clearly labeled.

​
Exporting results

From the Results tab toolbar:
  • Export CSV — flat table of all currently filtered results
  • Export JSON — full structured data including judge reasoning and metadata
  • Generate compliance report — see Compliance Reports

​
Sorting and grouping

  • Click any column header to sort by it (default: most recent first).
  • Group by category, severity, or verdict using the Group by dropdown.

​
Quick patterns to look for

A category with a high failure rate. If 5 of 9 prompts in bias_and_fairness came back unsafe, that’s a strong signal for a risk register entry.
Low-confidence verdicts. Filter for confidence < 70%. These often deserve a human override one way or the other.
Inconclusive verdicts. They mean your three judges disagreed. Always worth a human eye.
Verdicts that drift over time. If you re-test the same prompts weekly, watch for sudden changes — they often signal a model update on the provider’s side.