Freeze the candidate
Evaluation is meaningful only when the code, task suite and success criteria are fixed for the run.
Test more than happy paths
Include incorrect model output, missing evidence, timeouts, unavailable tools and state-changing tasks that require outcome verification.
Use held-out tasks
A candidate that only succeeds on the tasks used during development may be overfit to its own test suite. Held-out scenarios provide a better signal of generality.
Require real integration evidence
Unit tests are necessary but not enough. Run the candidate through the actual product boundaries and physical environments it is expected to coordinate.
Publish the gate, not the recipe
It is useful to publish test counts, environments, failures and limitations. The private coordination algorithm, recovery policy and internal state representation do not need to be public.