Skip to main content
A batch with 10 scenarios, 2 configurations, and 3 repetitions expands to 60 runs. A multi-step sequence can invoke the target more than once within a run.

​
A run has more than one outcome

Execution can complete while an exact check fails. A useful response can receive partial credit. An evaluator can be unavailable after the agent completed successfully. An optional CI policy makes its own decision from the recorded evidence. Read those states separately instead of interpreting one green label as proof of everything. See simulations and reports.