{
  "schema_version": 1,
  "columns": [
    "baseline",
    "comparison",
    "from",
    "to",
    "events"
  ],
  "rows": [
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "correct",
      "to": "correct",
      "events": "325"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "correct",
      "to": "wrong",
      "events": "14"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "correct",
      "to": "withheld",
      "events": "19"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "wrong",
      "to": "correct",
      "events": "11"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "wrong",
      "to": "wrong",
      "events": "83"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "wrong",
      "to": "withheld",
      "events": "30"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "withheld",
      "to": "correct",
      "events": "32"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "withheld",
      "to": "wrong",
      "events": "31"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "withheld",
      "to": "withheld",
      "events": "245"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite LOW",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "correct",
      "events": "326"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "wrong",
      "events": "17"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "withheld",
      "events": "15"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "correct",
      "events": "24"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "wrong",
      "events": "69"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "withheld",
      "events": "31"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "correct",
      "events": "49"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "wrong",
      "events": "60"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "withheld",
      "events": "199"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "correct",
      "events": "330"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "wrong",
      "events": "19"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "withheld",
      "events": "9"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "correct",
      "events": "32"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "wrong",
      "events": "76"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "withheld",
      "events": "16"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "correct",
      "events": "44"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "wrong",
      "events": "77"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "withheld",
      "events": "187"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MINIMAL corrected",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "correct",
      "events": "341"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "wrong",
      "events": "8"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "withheld",
      "events": "19"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "correct",
      "events": "22"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "wrong",
      "events": "76"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "withheld",
      "events": "30"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "correct",
      "events": "36"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "wrong",
      "events": "62"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "withheld",
      "events": "196"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite MEDIUM",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "correct",
      "events": "337"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "wrong",
      "events": "17"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "withheld",
      "events": "14"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "correct",
      "events": "29"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "wrong",
      "events": "80"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "withheld",
      "events": "19"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "correct",
      "events": "40"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "wrong",
      "events": "75"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "withheld",
      "events": "179"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite LOW",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "correct",
      "events": "366"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "wrong",
      "events": "16"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "withheld",
      "events": "17"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "correct",
      "events": "15"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "wrong",
      "events": "109"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "withheld",
      "events": "22"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "correct",
      "events": "25"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "wrong",
      "events": "47"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "withheld",
      "events": "173"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Flash-Lite MEDIUM",
      "comparison": "Flash-Lite HIGH",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "correct",
      "to": "correct",
      "events": "419"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "correct",
      "to": "wrong",
      "events": "12"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "correct",
      "to": "withheld",
      "events": "17"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "wrong",
      "to": "correct",
      "events": "14"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "wrong",
      "to": "wrong",
      "events": "51"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "wrong",
      "to": "withheld",
      "events": "9"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "withheld",
      "to": "correct",
      "events": "17"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "withheld",
      "to": "wrong",
      "events": "14"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "withheld",
      "to": "withheld",
      "events": "236"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "technical",
      "to": "withheld",
      "events": "1"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / LOW thinking",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "correct",
      "events": "434"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "wrong",
      "events": "10"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "withheld",
      "events": "4"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "correct",
      "events": "16"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "wrong",
      "events": "55"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "withheld",
      "events": "3"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "correct",
      "events": "56"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "wrong",
      "events": "40"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "withheld",
      "events": "171"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "withheld",
      "events": "1"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "correct",
      "events": "434"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "wrong",
      "events": "13"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "withheld",
      "events": "3"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "correct",
      "events": "19"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "wrong",
      "events": "54"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "withheld",
      "events": "4"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "correct",
      "events": "53"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "wrong",
      "events": "38"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "withheld",
      "events": "172"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Gemini 3.8 HIGH media / LOW thinking",
      "comparison": "Gemini 3.8 HIGH media / HIGH thinking",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "correct",
      "to": "correct",
      "events": "319"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "correct",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "correct",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "wrong",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "wrong",
      "to": "wrong",
      "events": "357"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "wrong",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "withheld",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "withheld",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "withheld",
      "to": "withheld",
      "events": "110"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "technical",
      "to": "correct",
      "events": "1"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "technical",
      "to": "wrong",
      "events": "3"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "Qwen Instruct first_pass",
      "comparison": "Qwen Instruct primary_repaired",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "correct",
      "to": "correct",
      "events": "322"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "correct",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "correct",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "wrong",
      "to": "correct",
      "events": "25"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "wrong",
      "to": "wrong",
      "events": "443"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "wrong",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "withheld",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "withheld",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "withheld",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "from": "technical",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "correct",
      "to": "correct",
      "events": "218"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "correct",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "correct",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "correct",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "wrong",
      "to": "correct",
      "events": "6"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "wrong",
      "to": "wrong",
      "events": "51"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "wrong",
      "to": "withheld",
      "events": "108"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "wrong",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "withheld",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "withheld",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "withheld",
      "to": "withheld",
      "events": "407"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "withheld",
      "to": "technical",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "technical",
      "to": "correct",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "technical",
      "to": "wrong",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "technical",
      "to": "withheld",
      "events": "0"
    },
    {
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "from": "technical",
      "to": "technical",
      "events": "0"
    }
  ],
  "unit_policy": {
    "table": "paired_transitions.csv",
    "metric_units": {},
    "counts": {
      "events": "events"
    },
    "interval": "ci_low and ci_high use the same units as estimate; pointwise 95% when interval_status is ok",
    "estimate": "not_applicable; labelled count columns only"
  },
  "note": "Saved aggregate candidate analysis; intervals are pointwise and exploratory tests have separate Holm-adjusted p-values."
}