{
  "scope": "Descriptive follow-up on 400 shared questions; six outputs per question are not independent trials.",
  "validity": {
    "qwen3.5-9b": {
      "invalid_ids": {
        "baseline": [],
        "subject": [
          "mmlu_pro_01020",
          "mmlu_pro_09998"
        ]
      },
      "normalization_changed_score_ids": {
        "baseline": [],
        "subject": []
      },
      "changed_groups": {
        "gain": {
          "count": 17,
          "offered_option": 17,
          "empty": 0,
          "other_invalid": 0
        },
        "loss": {
          "count": 10,
          "offered_option": 10,
          "empty": 0,
          "other_invalid": 0
        },
        "changed_wrong": {
          "count": 27,
          "offered_option": 25,
          "empty": 2,
          "other_invalid": 0
        }
      }
    },
    "qwen3.8-27b": {
      "invalid_ids": {
        "baseline": [],
        "subject": [
          "mmlu_pro_00087",
          "mmlu_pro_00564",
          "mmlu_pro_00807",
          "mmlu_pro_08378"
        ]
      },
      "normalization_changed_score_ids": {
        "baseline": [],
        "subject": []
      },
      "changed_groups": {
        "gain": {
          "count": 13,
          "offered_option": 13,
          "empty": 0,
          "other_invalid": 0
        },
        "loss": {
          "count": 7,
          "offered_option": 6,
          "empty": 1,
          "other_invalid": 0
        },
        "changed_wrong": {
          "count": 20,
          "offered_option": 17,
          "empty": 3,
          "other_invalid": 0
        }
      }
    },
    "ministral3-8b": {
      "invalid_ids": {
        "baseline": [
          "mmlu_pro_11894"
        ],
        "subject": []
      },
      "normalization_changed_score_ids": {
        "baseline": [],
        "subject": []
      },
      "changed_groups": {
        "gain": {
          "count": 6,
          "offered_option": 6,
          "empty": 0,
          "other_invalid": 0
        },
        "loss": {
          "count": 4,
          "offered_option": 4,
          "empty": 0,
          "other_invalid": 0
        },
        "changed_wrong": {
          "count": 15,
          "offered_option": 15,
          "empty": 0,
          "other_invalid": 0
        }
      }
    }
  },
  "shared_baseline_correct": {
    "count": 137,
    "ids": [
      "mmlu_pro_00154",
      "mmlu_pro_00202",
      "mmlu_pro_00436",
      "mmlu_pro_00685",
      "mmlu_pro_00831",
      "mmlu_pro_00947",
      "mmlu_pro_01092",
      "mmlu_pro_01761",
      "mmlu_pro_02110",
      "mmlu_pro_02188",
      "mmlu_pro_02203",
      "mmlu_pro_02225",
      "mmlu_pro_02228",
      "mmlu_pro_02238",
      "mmlu_pro_02356",
      "mmlu_pro_02483",
      "mmlu_pro_02601",
      "mmlu_pro_02667",
      "mmlu_pro_02670",
      "mmlu_pro_02706",
      "mmlu_pro_02851",
      "mmlu_pro_02854",
      "mmlu_pro_02872",
      "mmlu_pro_02905",
      "mmlu_pro_02961",
      "mmlu_pro_02965",
      "mmlu_pro_02993",
      "mmlu_pro_03009",
      "mmlu_pro_03030",
      "mmlu_pro_03057",
      "mmlu_pro_03073",
      "mmlu_pro_03124",
      "mmlu_pro_03214",
      "mmlu_pro_03221",
      "mmlu_pro_03279",
      "mmlu_pro_03310",
      "mmlu_pro_03373",
      "mmlu_pro_03415",
      "mmlu_pro_03434",
      "mmlu_pro_03472",
      "mmlu_pro_03493",
      "mmlu_pro_03509",
      "mmlu_pro_03513",
      "mmlu_pro_03690",
      "mmlu_pro_03780",
      "mmlu_pro_03797",
      "mmlu_pro_03861",
      "mmlu_pro_04267",
      "mmlu_pro_04628",
      "mmlu_pro_04653",
      "mmlu_pro_04715",
      "mmlu_pro_04739",
      "mmlu_pro_04760",
      "mmlu_pro_04778",
      "mmlu_pro_04828",
      "mmlu_pro_04836",
      "mmlu_pro_04837",
      "mmlu_pro_04861",
      "mmlu_pro_04864",
      "mmlu_pro_04881",
      "mmlu_pro_04898",
      "mmlu_pro_04914",
      "mmlu_pro_04942",
      "mmlu_pro_04951",
      "mmlu_pro_04962",
      "mmlu_pro_04993",
      "mmlu_pro_05449",
      "mmlu_pro_05628",
      "mmlu_pro_05675",
      "mmlu_pro_05752",
      "mmlu_pro_05782",
      "mmlu_pro_05785",
      "mmlu_pro_05927",
      "mmlu_pro_05929",
      "mmlu_pro_05974",
      "mmlu_pro_05992",
      "mmlu_pro_06001",
      "mmlu_pro_06179",
      "mmlu_pro_06229",
      "mmlu_pro_06266",
      "mmlu_pro_06325",
      "mmlu_pro_06340",
      "mmlu_pro_06455",
      "mmlu_pro_06481",
      "mmlu_pro_06615",
      "mmlu_pro_06625",
      "mmlu_pro_06633",
      "mmlu_pro_06678",
      "mmlu_pro_06702",
      "mmlu_pro_06738",
      "mmlu_pro_06767",
      "mmlu_pro_06890",
      "mmlu_pro_06891",
      "mmlu_pro_06990",
      "mmlu_pro_07081",
      "mmlu_pro_07145",
      "mmlu_pro_07147",
      "mmlu_pro_07276",
      "mmlu_pro_07336",
      "mmlu_pro_07369",
      "mmlu_pro_07465",
      "mmlu_pro_07575",
      "mmlu_pro_07595",
      "mmlu_pro_07597",
      "mmlu_pro_08114",
      "mmlu_pro_08332",
      "mmlu_pro_08446",
      "mmlu_pro_08981",
      "mmlu_pro_09433",
      "mmlu_pro_09727",
      "mmlu_pro_10264",
      "mmlu_pro_10380",
      "mmlu_pro_10413",
      "mmlu_pro_10414",
      "mmlu_pro_10442",
      "mmlu_pro_10501",
      "mmlu_pro_10508",
      "mmlu_pro_10517",
      "mmlu_pro_10530",
      "mmlu_pro_10531",
      "mmlu_pro_10586",
      "mmlu_pro_10595",
      "mmlu_pro_10602",
      "mmlu_pro_10620",
      "mmlu_pro_10627",
      "mmlu_pro_10652",
      "mmlu_pro_10700",
      "mmlu_pro_10737",
      "mmlu_pro_10745",
      "mmlu_pro_10876",
      "mmlu_pro_10881",
      "mmlu_pro_10904",
      "mmlu_pro_10927",
      "mmlu_pro_11041",
      "mmlu_pro_11514",
      "mmlu_pro_12084",
      "mmlu_pro_12236"
    ],
    "loss_ids": {
      "qwen3.5-9b": [
        "mmlu_pro_04628",
        "mmlu_pro_10700"
      ],
      "qwen3.8-27b": [
        "mmlu_pro_03861"
      ],
      "ministral3-8b": [
        "mmlu_pro_05992",
        "mmlu_pro_06633"
      ]
    },
    "affected_ids": [
      "mmlu_pro_03861",
      "mmlu_pro_04628",
      "mmlu_pro_05992",
      "mmlu_pro_06633",
      "mmlu_pro_10700"
    ],
    "all_preserved": 132
  },
  "shared_errors": {
    "all_six_wrong_ids": [
      "mmlu_pro_00166",
      "mmlu_pro_00250",
      "mmlu_pro_00292",
      "mmlu_pro_00306",
      "mmlu_pro_00564",
      "mmlu_pro_00663",
      "mmlu_pro_00700",
      "mmlu_pro_00807",
      "mmlu_pro_00934",
      "mmlu_pro_00941",
      "mmlu_pro_01368",
      "mmlu_pro_01423",
      "mmlu_pro_01660",
      "mmlu_pro_01717",
      "mmlu_pro_01739",
      "mmlu_pro_01851",
      "mmlu_pro_01940",
      "mmlu_pro_02018",
      "mmlu_pro_02256",
      "mmlu_pro_02626",
      "mmlu_pro_02820",
      "mmlu_pro_03597",
      "mmlu_pro_03752",
      "mmlu_pro_03875",
      "mmlu_pro_03998",
      "mmlu_pro_04684",
      "mmlu_pro_04941",
      "mmlu_pro_04994",
      "mmlu_pro_05017",
      "mmlu_pro_05238",
      "mmlu_pro_05266",
      "mmlu_pro_05268",
      "mmlu_pro_05507",
      "mmlu_pro_05694",
      "mmlu_pro_05706",
      "mmlu_pro_05879",
      "mmlu_pro_05910",
      "mmlu_pro_06022",
      "mmlu_pro_06485",
      "mmlu_pro_06598",
      "mmlu_pro_06619",
      "mmlu_pro_06781",
      "mmlu_pro_07296",
      "mmlu_pro_07427",
      "mmlu_pro_07568",
      "mmlu_pro_07632",
      "mmlu_pro_07742",
      "mmlu_pro_07756",
      "mmlu_pro_07939",
      "mmlu_pro_08037",
      "mmlu_pro_08073",
      "mmlu_pro_08378",
      "mmlu_pro_08489",
      "mmlu_pro_08836",
      "mmlu_pro_08847",
      "mmlu_pro_08868",
      "mmlu_pro_08938",
      "mmlu_pro_09251",
      "mmlu_pro_09332",
      "mmlu_pro_09338",
      "mmlu_pro_09461",
      "mmlu_pro_09624",
      "mmlu_pro_09664",
      "mmlu_pro_09709",
      "mmlu_pro_09738",
      "mmlu_pro_09862",
      "mmlu_pro_09908",
      "mmlu_pro_10037",
      "mmlu_pro_10217",
      "mmlu_pro_10223",
      "mmlu_pro_10231",
      "mmlu_pro_10637",
      "mmlu_pro_10761",
      "mmlu_pro_10791",
      "mmlu_pro_10809",
      "mmlu_pro_10856",
      "mmlu_pro_10901",
      "mmlu_pro_10912",
      "mmlu_pro_10923",
      "mmlu_pro_10949",
      "mmlu_pro_11195",
      "mmlu_pro_11229",
      "mmlu_pro_11268",
      "mmlu_pro_11410",
      "mmlu_pro_11539",
      "mmlu_pro_11547",
      "mmlu_pro_11671",
      "mmlu_pro_11694",
      "mmlu_pro_11841",
      "mmlu_pro_11886",
      "mmlu_pro_11916",
      "mmlu_pro_12197"
    ],
    "same_wrong_option_ids": [
      "mmlu_pro_00292",
      "mmlu_pro_01423",
      "mmlu_pro_01851",
      "mmlu_pro_02018",
      "mmlu_pro_02256",
      "mmlu_pro_03875",
      "mmlu_pro_04684",
      "mmlu_pro_04994",
      "mmlu_pro_05017",
      "mmlu_pro_05507",
      "mmlu_pro_05694",
      "mmlu_pro_05706",
      "mmlu_pro_06022",
      "mmlu_pro_06781",
      "mmlu_pro_07296",
      "mmlu_pro_07427",
      "mmlu_pro_07568",
      "mmlu_pro_07742",
      "mmlu_pro_07756",
      "mmlu_pro_08037",
      "mmlu_pro_09664",
      "mmlu_pro_10637",
      "mmlu_pro_10761",
      "mmlu_pro_10791",
      "mmlu_pro_10856",
      "mmlu_pro_10901",
      "mmlu_pro_10912",
      "mmlu_pro_11229",
      "mmlu_pro_11268",
      "mmlu_pro_11886"
    ],
    "all_six_correct_ids": [
      "mmlu_pro_00154",
      "mmlu_pro_00202",
      "mmlu_pro_00436",
      "mmlu_pro_00685",
      "mmlu_pro_00831",
      "mmlu_pro_00947",
      "mmlu_pro_01092",
      "mmlu_pro_01761",
      "mmlu_pro_02110",
      "mmlu_pro_02188",
      "mmlu_pro_02203",
      "mmlu_pro_02225",
      "mmlu_pro_02228",
      "mmlu_pro_02238",
      "mmlu_pro_02356",
      "mmlu_pro_02483",
      "mmlu_pro_02601",
      "mmlu_pro_02667",
      "mmlu_pro_02670",
      "mmlu_pro_02706",
      "mmlu_pro_02851",
      "mmlu_pro_02854",
      "mmlu_pro_02872",
      "mmlu_pro_02905",
      "mmlu_pro_02961",
      "mmlu_pro_02965",
      "mmlu_pro_02993",
      "mmlu_pro_03009",
      "mmlu_pro_03030",
      "mmlu_pro_03057",
      "mmlu_pro_03073",
      "mmlu_pro_03124",
      "mmlu_pro_03214",
      "mmlu_pro_03221",
      "mmlu_pro_03279",
      "mmlu_pro_03310",
      "mmlu_pro_03373",
      "mmlu_pro_03415",
      "mmlu_pro_03434",
      "mmlu_pro_03472",
      "mmlu_pro_03493",
      "mmlu_pro_03509",
      "mmlu_pro_03513",
      "mmlu_pro_03690",
      "mmlu_pro_03780",
      "mmlu_pro_03797",
      "mmlu_pro_04267",
      "mmlu_pro_04653",
      "mmlu_pro_04715",
      "mmlu_pro_04739",
      "mmlu_pro_04760",
      "mmlu_pro_04778",
      "mmlu_pro_04828",
      "mmlu_pro_04836",
      "mmlu_pro_04837",
      "mmlu_pro_04861",
      "mmlu_pro_04864",
      "mmlu_pro_04881",
      "mmlu_pro_04898",
      "mmlu_pro_04914",
      "mmlu_pro_04942",
      "mmlu_pro_04951",
      "mmlu_pro_04962",
      "mmlu_pro_04993",
      "mmlu_pro_05449",
      "mmlu_pro_05628",
      "mmlu_pro_05675",
      "mmlu_pro_05752",
      "mmlu_pro_05782",
      "mmlu_pro_05785",
      "mmlu_pro_05927",
      "mmlu_pro_05929",
      "mmlu_pro_05974",
      "mmlu_pro_06001",
      "mmlu_pro_06179",
      "mmlu_pro_06229",
      "mmlu_pro_06266",
      "mmlu_pro_06325",
      "mmlu_pro_06340",
      "mmlu_pro_06455",
      "mmlu_pro_06481",
      "mmlu_pro_06615",
      "mmlu_pro_06625",
      "mmlu_pro_06678",
      "mmlu_pro_06702",
      "mmlu_pro_06738",
      "mmlu_pro_06767",
      "mmlu_pro_06890",
      "mmlu_pro_06891",
      "mmlu_pro_06990",
      "mmlu_pro_07081",
      "mmlu_pro_07145",
      "mmlu_pro_07147",
      "mmlu_pro_07276",
      "mmlu_pro_07336",
      "mmlu_pro_07369",
      "mmlu_pro_07465",
      "mmlu_pro_07575",
      "mmlu_pro_07595",
      "mmlu_pro_07597",
      "mmlu_pro_08114",
      "mmlu_pro_08332",
      "mmlu_pro_08446",
      "mmlu_pro_08981",
      "mmlu_pro_09433",
      "mmlu_pro_09727",
      "mmlu_pro_10264",
      "mmlu_pro_10380",
      "mmlu_pro_10413",
      "mmlu_pro_10414",
      "mmlu_pro_10442",
      "mmlu_pro_10501",
      "mmlu_pro_10508",
      "mmlu_pro_10517",
      "mmlu_pro_10530",
      "mmlu_pro_10531",
      "mmlu_pro_10586",
      "mmlu_pro_10595",
      "mmlu_pro_10602",
      "mmlu_pro_10620",
      "mmlu_pro_10627",
      "mmlu_pro_10652",
      "mmlu_pro_10737",
      "mmlu_pro_10745",
      "mmlu_pro_10876",
      "mmlu_pro_10881",
      "mmlu_pro_10904",
      "mmlu_pro_10927",
      "mmlu_pro_11041",
      "mmlu_pro_11514",
      "mmlu_pro_12084",
      "mmlu_pro_12236"
    ],
    "mixed_correctness_ids": [
      "mmlu_pro_00075",
      "mmlu_pro_00087",
      "mmlu_pro_00111",
      "mmlu_pro_00257",
      "mmlu_pro_00284",
      "mmlu_pro_00294",
      "mmlu_pro_00323",
      "mmlu_pro_00346",
      "mmlu_pro_00372",
      "mmlu_pro_00375",
      "mmlu_pro_00536",
      "mmlu_pro_00569",
      "mmlu_pro_00575",
      "mmlu_pro_00614",
      "mmlu_pro_00650",
      "mmlu_pro_00695",
      "mmlu_pro_01004",
      "mmlu_pro_01020",
      "mmlu_pro_01134",
      "mmlu_pro_01179",
      "mmlu_pro_01244",
      "mmlu_pro_01284",
      "mmlu_pro_01337",
      "mmlu_pro_01387",
      "mmlu_pro_01413",
      "mmlu_pro_01419",
      "mmlu_pro_01449",
      "mmlu_pro_01529",
      "mmlu_pro_01578",
      "mmlu_pro_01866",
      "mmlu_pro_01894",
      "mmlu_pro_01974",
      "mmlu_pro_02033",
      "mmlu_pro_02086",
      "mmlu_pro_02090",
      "mmlu_pro_02091",
      "mmlu_pro_02147",
      "mmlu_pro_02236",
      "mmlu_pro_02321",
      "mmlu_pro_02459",
      "mmlu_pro_02519",
      "mmlu_pro_02528",
      "mmlu_pro_02635",
      "mmlu_pro_02645",
      "mmlu_pro_02728",
      "mmlu_pro_02838",
      "mmlu_pro_02986",
      "mmlu_pro_03074",
      "mmlu_pro_03106",
      "mmlu_pro_03140",
      "mmlu_pro_03582",
      "mmlu_pro_03602",
      "mmlu_pro_03663",
      "mmlu_pro_03721",
      "mmlu_pro_03766",
      "mmlu_pro_03819",
      "mmlu_pro_03861",
      "mmlu_pro_03871",
      "mmlu_pro_03900",
      "mmlu_pro_03908",
      "mmlu_pro_03931",
      "mmlu_pro_03992",
      "mmlu_pro_04011",
      "mmlu_pro_04090",
      "mmlu_pro_04129",
      "mmlu_pro_04168",
      "mmlu_pro_04206",
      "mmlu_pro_04234",
      "mmlu_pro_04379",
      "mmlu_pro_04628",
      "mmlu_pro_04708",
      "mmlu_pro_04748",
      "mmlu_pro_04831",
      "mmlu_pro_04891",
      "mmlu_pro_04929",
      "mmlu_pro_04943",
      "mmlu_pro_04965",
      "mmlu_pro_04988",
      "mmlu_pro_05044",
      "mmlu_pro_05141",
      "mmlu_pro_05276",
      "mmlu_pro_05307",
      "mmlu_pro_05383",
      "mmlu_pro_05471",
      "mmlu_pro_05474",
      "mmlu_pro_05527",
      "mmlu_pro_05537",
      "mmlu_pro_05543",
      "mmlu_pro_05776",
      "mmlu_pro_05992",
      "mmlu_pro_06064",
      "mmlu_pro_06198",
      "mmlu_pro_06271",
      "mmlu_pro_06297",
      "mmlu_pro_06457",
      "mmlu_pro_06533",
      "mmlu_pro_06536",
      "mmlu_pro_06633",
      "mmlu_pro_06652",
      "mmlu_pro_06710",
      "mmlu_pro_06893",
      "mmlu_pro_06948",
      "mmlu_pro_06962",
      "mmlu_pro_06993",
      "mmlu_pro_07168",
      "mmlu_pro_07272",
      "mmlu_pro_07527",
      "mmlu_pro_07552",
      "mmlu_pro_07565",
      "mmlu_pro_07608",
      "mmlu_pro_07639",
      "mmlu_pro_07657",
      "mmlu_pro_07779",
      "mmlu_pro_07903",
      "mmlu_pro_08078",
      "mmlu_pro_08118",
      "mmlu_pro_08163",
      "mmlu_pro_08299",
      "mmlu_pro_08449",
      "mmlu_pro_08459",
      "mmlu_pro_08554",
      "mmlu_pro_08601",
      "mmlu_pro_08679",
      "mmlu_pro_08798",
      "mmlu_pro_08988",
      "mmlu_pro_09138",
      "mmlu_pro_09149",
      "mmlu_pro_09319",
      "mmlu_pro_09440",
      "mmlu_pro_09563",
      "mmlu_pro_09592",
      "mmlu_pro_09810",
      "mmlu_pro_09998",
      "mmlu_pro_10035",
      "mmlu_pro_10125",
      "mmlu_pro_10243",
      "mmlu_pro_10360",
      "mmlu_pro_10411",
      "mmlu_pro_10460",
      "mmlu_pro_10475",
      "mmlu_pro_10500",
      "mmlu_pro_10521",
      "mmlu_pro_10550",
      "mmlu_pro_10609",
      "mmlu_pro_10700",
      "mmlu_pro_10771",
      "mmlu_pro_10776",
      "mmlu_pro_10812",
      "mmlu_pro_10822",
      "mmlu_pro_10861",
      "mmlu_pro_10875",
      "mmlu_pro_10897",
      "mmlu_pro_10922",
      "mmlu_pro_11095",
      "mmlu_pro_11128",
      "mmlu_pro_11135",
      "mmlu_pro_11178",
      "mmlu_pro_11230",
      "mmlu_pro_11234",
      "mmlu_pro_11344",
      "mmlu_pro_11363",
      "mmlu_pro_11472",
      "mmlu_pro_11515",
      "mmlu_pro_11533",
      "mmlu_pro_11549",
      "mmlu_pro_11642",
      "mmlu_pro_11716",
      "mmlu_pro_11731",
      "mmlu_pro_11770",
      "mmlu_pro_11894",
      "mmlu_pro_11922",
      "mmlu_pro_12055",
      "mmlu_pro_12056",
      "mmlu_pro_12149",
      "mmlu_pro_12199",
      "mmlu_pro_12225"
    ]
  },
  "opposite_direction_questions": [
    {
      "record_id": "mmlu_pro_04748",
      "expected": "F",
      "models": {
        "qwen3.5-9b": {
          "baseline": "I",
          "subject": "F",
          "category": "gain"
        },
        "qwen3.8-27b": {
          "baseline": "F",
          "subject": "F",
          "category": "preserved"
        },
        "ministral3-8b": {
          "baseline": "F",
          "subject": "G",
          "category": "loss"
        }
      }
    },
    {
      "record_id": "mmlu_pro_11770",
      "expected": "D",
      "models": {
        "qwen3.5-9b": {
          "baseline": "A",
          "subject": "D",
          "category": "gain"
        },
        "qwen3.8-27b": {
          "baseline": "D",
          "subject": "A",
          "category": "loss"
        },
        "ministral3-8b": {
          "baseline": "D",
          "subject": "D",
          "category": "preserved"
        }
      }
    },
    {
      "record_id": "mmlu_pro_12199",
      "expected": "D",
      "models": {
        "qwen3.5-9b": {
          "baseline": "D",
          "subject": "B",
          "category": "loss"
        },
        "qwen3.8-27b": {
          "baseline": "B",
          "subject": "D",
          "category": "gain"
        },
        "ministral3-8b": {
          "baseline": "B",
          "subject": "B",
          "category": "same_wrong"
        }
      }
    }
  ]
}
