Experiments
The experiments list shows each experiment’s number and name, its baseline, the agent and commit it tested, its mode, pass rate, regressions, status, and when it started. Filter by Agent, Status, Mode, and Started by, or search by name or number. On the Datasets page, filter by Dataset too.
Pass rate and regressions are final once the experiment stops. Open an experiment’s ⋮ menu to Open results, Cancel it while it runs, or Run again.
An experiment
The header shows the experiment’s name and number, with Cancel while it runs and Run again once it ends. See Run again. Below the header you see the agent and the branch and commit it tested, the judges, the iterations, the baseline, who started it and when, and what it cost. While it runs, a progress bar counts finished replays and conversations.Summary
The comparison cards are filled in only when the experiment has a baseline. Without one they show 0 or —. Choose a baseline when you start the experiment, or use Run again, which picks one for you.

Verdicts
Each item gets one verdict across all its iterations:
An iteration passes when every judge and every assertion on the item passes.
Every item, every judge
A grid with one row per item and one column per judge, plus an Assertions column. Each cell shows how many iterations passed, for example 2 of 3. Each row also shows the item’s kind, its verdict, and how it compares with the baseline: Regression, Fixed, Still failing, or No change. Filter by Verdict, vs baseline, Kind, and Tags, or search. Turn on Show only regressions to see just what broke. For a conversation item, select the Add to dataset icon at the end of its row to add the simulated call to a dataset. Judges that failed on it are filled in as the new item’s checks. See From a simulation.A result
Select a row to open the item’s result. The top shows its verdict and how it compares with the baseline. Pick an iteration to see:- Judges & assertions: every judge’s verdict with its reason, and every assertion with what it found, for example got 4 for Max turns, or The agent never called reschedule_appointment.
- This run: what the agent produced: its output, its tool calls, and its audio.

Fix and re-run
1
Find what broke
Turn on Show only regressions, open a failed item, and read the judge’s reason or the assertion’s result.
3
Run again
Select Run again. The new experiment uses this one as its baseline.
4
Compare
Check Regressions and Fixed on the new experiment to see what moved.