A model-graded check: the agent's output scores ≥ min against rubric,
judged by a model. Unlike the deterministic checks this one calls a model
(cost), so it's for the scored measure tier; the judge fn is injectable
(default: the real judge()), so the logic is unit-testable with a fake.
A model-graded check: the agent's
outputscores ≥minagainstrubric, judged by a model. Unlike the deterministic checks this one calls a model (cost), so it's for the scoredmeasuretier; thejudgefn is injectable (default: the realjudge()), so the logic is unit-testable with a fake.