Latest release gate
Loading eval results
Average score by run
The dashed line marks the 4/5 case threshold.
Latest case results
All saved runs
Run history
Each row links to its grade file and raw trace. Open a run below the table for the full judge explanations.
| Run | Gate | Average | Case scores | Critical | Evidence |
|---|
Judge explanations
All saved explanations are retained here.
Behaviour by test
Case performance
Each chart shows one behaviour across the full eval sequence. Critical cases remain visible even when the judge explanation sounds positive.
A case passes only when its score reaches the run threshold and the deterministic checks record no critical failure.
Below the gate
Failure record
Every below-threshold or critical case appears here with the saved judge explanation.
Tool sequence
Trace view
Select a run to inspect the calls recorded for each case. Raw traces remain linked for event-level debugging.
The report summarises function calls and response status. It leaves function arguments and raw responses in the trace JSON.
Refresh after future evals
Save the new grade result and trace in their existing artifact directories, then run:
uv run python tests/eval/build_eval_report.py