{ "evaluation_kind": "historical_regression", "dataset_sha256": "a7bb38be01e03603cf2fa2c7137817094dff0a0cbb3f801263609e55b28e9edc", "graph_dev": 722, "graph_regression": 320, "llm_dev_ids": [ "wide-001", "wide-002", "wide-041", "wide-042", "wide-080", "wide-081", "wide-082", "wide-104", "wide-113", "wide-114", "wide-116", "wide-119", "wide-129", "wide-137", "wide-138", "wide-139", "wide-141", "wide-142", "wide-143", "wide-145", "wide-155", "wide-159", "wide-161", "wide-163", "wide-165", "wide-169", "wide-170", "wide-178", "wide-181", "wide-193", "wide-194", "wide-209", "wide-217", "wide-218", "wide-232", "wide-236", "wide-239", "wide-249", "wide-250", "wide-283", "wide-289", "wide-290" ], "llm_dev_count": 42, "llm_final_regression": 320, "llm_final_dev_overlap": 42, "test_positive_count": 284, "test_negative_count": 36, "corrective_request_limit": 1, "correction_basis": "Only actual retrieval context or invalid JSON; never oracle, expected answer, gold, or final grade.", "gold_changes": 15, "changed_test_ids": [ "wide-116", "wide-142", "wide-145", "wide-165" ], "policy": "Keep all IDs, questions and splits. Do not tune after final freeze. Retain failed deliveries and every raw attempt. No claim of independent TEST or human review.", "limitations": [ "All questions previously exposed. Gold corrections and new guards prevent direct controlled comparison of accuracy with v5.1.", "Native OpenWebUI tool routing is point 4." ], "source_review": { "reviewed": 34, "supported": 31, "corrected": 2, "conditional_directory_cohort": 1, "human_review": "not performed; worksheet supplied" }, "start_year_convention": "Explicit statement, otherwise students/ directory cohort; a proxy, not independent proof of actual enrolment.", "development_iterations": [ { "name": "candidate", "questions": 42, "correct": 41, "failure": "wide-283: same-person year/tag join rejected", "generation": "answers_v6_dev_2026-10-25" } ] }