dp-zp-agent/experiments/graphrag/results/answers_v7_regression_2026-10-26/frozen_answers.json

500 lines
9.8 KiB
JSON

{
"method": "Previously exposed LLM questions; regression, not independent TEST.",
"graph_freeze_sha256": "9e97794200520d0f2288e452c9e186c0291a54c63d2c1382f693a9753e4081cf",
"audit_sha256": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5",
"generation_files": {
"tools/answer_evaluation/evaluate_graph_answers.py": "d81504e8fe51598c6bd1b0d482f61b9b6f51f87ba3f38b03ae1bc584675a3b76",
"tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
"tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
"tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
"tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
"tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
"tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1"
},
"model": "model120-fast",
"temperature": 0,
"workers": 3,
"dev_ids": [
"v7-dev-year-1899",
"v7-dev-year-1999",
"v7-dev-year-2100",
"v7-dev-year-9999",
"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"
],
"test_ids": [
"fresh7-001",
"fresh7-002",
"fresh7-003",
"fresh7-004",
"fresh7-005",
"fresh7-006",
"fresh7-007",
"fresh7-008",
"fresh7-009",
"fresh7-010",
"fresh7-011",
"fresh7-012",
"fresh7-013",
"fresh7-014",
"fresh7-015",
"fresh7-016",
"fresh7-017",
"fresh7-018",
"fresh7-019",
"fresh7-020",
"fresh7-021",
"fresh7-022",
"fresh7-023",
"fresh7-024",
"fresh7-025",
"fresh7-026",
"fresh7-027",
"fresh7-028",
"fresh7-029",
"fresh7-030",
"fresh7-031",
"fresh7-032",
"fresh7-033",
"fresh7-034",
"fresh7-035",
"fresh7-036",
"fresh7-037",
"fresh7-038",
"fresh7-039",
"fresh7-040",
"fresh7-041",
"fresh7-042",
"fresh7-043",
"fresh7-044",
"fresh7-045",
"fresh7-046",
"fresh7-047",
"fresh7-048",
"fresh7-049",
"fresh7-050",
"fresh7-051",
"fresh7-052",
"fresh7-053",
"fresh7-054",
"fresh7-055",
"fresh7-056",
"fresh7-057",
"fresh7-058",
"fresh7-059",
"fresh7-060",
"fresh7-061",
"fresh7-062",
"fresh7-063",
"fresh7-064",
"fresh7-065",
"fresh7-066",
"fresh7-067",
"fresh7-068",
"fresh7-069",
"fresh7-070",
"fresh7-071",
"fresh7-072",
"fresh7-073",
"fresh7-074",
"fresh7-075",
"fresh7-076",
"fresh7-077",
"fresh7-078",
"fresh7-079",
"fresh7-080"
],
"dev_count": 46,
"test_count": 80,
"target_accuracy": 0.92,
"evaluation_kind": "historical_regression",
"rendering": "Verified structured claims rendered deterministically; raw model outputs preserved.",
"selected_split": "regression",
"selected_ids": [
"wide-001",
"wide-002",
"wide-003",
"wide-004",
"wide-005",
"wide-006",
"wide-007",
"wide-008",
"wide-009",
"wide-010",
"wide-011",
"wide-012",
"wide-013",
"wide-014",
"wide-015",
"wide-016",
"wide-017",
"wide-018",
"wide-019",
"wide-020",
"wide-021",
"wide-022",
"wide-023",
"wide-024",
"wide-025",
"wide-026",
"wide-027",
"wide-028",
"wide-029",
"wide-030",
"wide-031",
"wide-032",
"wide-033",
"wide-034",
"wide-035",
"wide-036",
"wide-037",
"wide-038",
"wide-039",
"wide-040",
"wide-041",
"wide-042",
"wide-043",
"wide-044",
"wide-045",
"wide-046",
"wide-047",
"wide-048",
"wide-049",
"wide-050",
"wide-051",
"wide-052",
"wide-053",
"wide-054",
"wide-055",
"wide-056",
"wide-057",
"wide-058",
"wide-059",
"wide-060",
"wide-061",
"wide-062",
"wide-063",
"wide-064",
"wide-065",
"wide-066",
"wide-067",
"wide-068",
"wide-069",
"wide-070",
"wide-071",
"wide-072",
"wide-073",
"wide-074",
"wide-075",
"wide-076",
"wide-077",
"wide-078",
"wide-079",
"wide-080",
"wide-081",
"wide-082",
"wide-083",
"wide-084",
"wide-085",
"wide-086",
"wide-087",
"wide-088",
"wide-089",
"wide-090",
"wide-091",
"wide-092",
"wide-093",
"wide-094",
"wide-095",
"wide-096",
"wide-097",
"wide-098",
"wide-099",
"wide-100",
"wide-101",
"wide-102",
"wide-103",
"wide-104",
"wide-105",
"wide-106",
"wide-107",
"wide-108",
"wide-109",
"wide-110",
"wide-111",
"wide-112",
"wide-113",
"wide-114",
"wide-115",
"wide-116",
"wide-117",
"wide-118",
"wide-119",
"wide-120",
"wide-121",
"wide-122",
"wide-123",
"wide-124",
"wide-125",
"wide-126",
"wide-127",
"wide-128",
"wide-129",
"wide-130",
"wide-131",
"wide-132",
"wide-133",
"wide-134",
"wide-135",
"wide-136",
"wide-137",
"wide-138",
"wide-139",
"wide-140",
"wide-141",
"wide-142",
"wide-143",
"wide-144",
"wide-145",
"wide-146",
"wide-147",
"wide-148",
"wide-149",
"wide-150",
"wide-151",
"wide-152",
"wide-153",
"wide-154",
"wide-155",
"wide-156",
"wide-157",
"wide-158",
"wide-159",
"wide-160",
"wide-161",
"wide-162",
"wide-163",
"wide-164",
"wide-165",
"wide-166",
"wide-167",
"wide-168",
"wide-169",
"wide-170",
"wide-171",
"wide-172",
"wide-173",
"wide-174",
"wide-175",
"wide-176",
"wide-177",
"wide-178",
"wide-179",
"wide-180",
"wide-181",
"wide-182",
"wide-183",
"wide-184",
"wide-185",
"wide-186",
"wide-187",
"wide-188",
"wide-189",
"wide-190",
"wide-191",
"wide-192",
"wide-193",
"wide-194",
"wide-195",
"wide-196",
"wide-197",
"wide-198",
"wide-199",
"wide-200",
"wide-201",
"wide-202",
"wide-203",
"wide-204",
"wide-205",
"wide-206",
"wide-207",
"wide-208",
"wide-209",
"wide-210",
"wide-211",
"wide-212",
"wide-213",
"wide-214",
"wide-215",
"wide-216",
"wide-217",
"wide-218",
"wide-219",
"wide-220",
"wide-221",
"wide-222",
"wide-223",
"wide-224",
"wide-225",
"wide-226",
"wide-227",
"wide-228",
"wide-229",
"wide-230",
"wide-231",
"wide-232",
"wide-233",
"wide-234",
"wide-235",
"wide-236",
"wide-237",
"wide-238",
"wide-239",
"wide-240",
"wide-241",
"wide-242",
"wide-243",
"wide-244",
"wide-245",
"wide-246",
"wide-247",
"wide-248",
"wide-249",
"wide-250",
"wide-251",
"wide-252",
"wide-253",
"wide-254",
"wide-255",
"wide-256",
"wide-257",
"wide-258",
"wide-259",
"wide-260",
"wide-261",
"wide-262",
"wide-263",
"wide-264",
"wide-265",
"wide-266",
"wide-267",
"wide-268",
"wide-269",
"wide-270",
"wide-271",
"wide-272",
"wide-273",
"wide-274",
"wide-275",
"wide-276",
"wide-277",
"wide-278",
"wide-279",
"wide-280",
"wide-281",
"wide-282",
"wide-283",
"wide-284",
"wide-285",
"wide-286",
"wide-287",
"wide-288",
"wide-289",
"wide-290",
"wide-291",
"wide-292",
"wide-293",
"wide-294",
"wide-295",
"wide-296",
"wide-297",
"wide-298",
"wide-299",
"wide-300",
"wide-301",
"wide-302",
"wide-303",
"wide-304",
"wide-305",
"wide-306",
"wide-307",
"wide-308",
"wide-309",
"wide-310",
"wide-311",
"wide-312",
"wide-313",
"wide-314",
"wide-315",
"wide-316",
"wide-317",
"wide-318",
"wide-319",
"wide-320"
],
"selected_count": 320,
"selected_dev_overlap": 42,
"corrective_request_limit": 1,
"format_repair_limit": 1,
"repair_policy": "At most one total corrective request for format, citations or retrieval-context violations. Same question/context, no gold; retain all attempts.",
"dev_results_sha256": "4af9b52c250c837707c0690d892ec833347b3d3ee71d338ec631f6450eef8045",
"policy": "Evaluate every declared regression case, including DEV overlap; retain all failures and stop tuning after this freeze.",
"selected_categories": {
"shared_keyword_pair": 40,
"same_academic_year_pair": 40,
"same_start_year_pair": 32,
"person_title_year": 24,
"person_keyword_title": 32,
"document_author_category_multi_hop": 24,
"person_multiple_categories": 24,
"candidate_set_aggregation": 32,
"conjunctive_disambiguation": 40,
"negative_lookup": 32
},
"limitations": [
"All questions and previous model outputs are known; this is historical regression.",
"Harness supplies graph context; native OpenWebUI tool choice and browser UI are outside this experiment.",
"Checks structured factual claims, their source support/citations and required prose values; not a general semantic judge of arbitrary prose."
]
}