from auditenv.datasets.factory import generate_episode def test_medium_generator_fallback_is_deterministic() -> None: ep1 = generate_episode(task_id="medium", seed=123) ep2 = generate_episode(task_id="medium", seed=123) assert ep1.documents == ep2.documents assert ep1.ground_truth == ep2.ground_truth assert ep1.evidence_map == ep2.evidence_map def test_hard_generator_has_seeded_findings() -> None: ep = generate_episode(task_id="hard", seed=55) assert len(ep.documents) > 0 assert len(ep.ground_truth) > 0 assert len(ep.evidence_map) > 0