{
  "schemaVersion": 1,
  "generated_at": "2026-05-14T16:41:16Z",
  "summary": {
    "total": 407,
    "scored": 146,
    "bot_challenged": 12,
    "by_category": {
      "academic-paper": 34,
      "developer": 34,
      "ecommerce": 35,
      "film-tv": 33,
      "games": 33,
      "ml-benchmark": 34,
      "museum-culture": 35,
      "music": 32,
      "news-data": 33,
      "online-tool": 34,
      "saas-landing": 35,
      "sports": 35
    },
    "mean_overall": 0.7109916709339843
  },
  "entries": [
    {
      "id": "T-arxiv-1409-1556",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1409.1556",
      "sealed_at": "2026-04-29T06:26:04Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1409-3215",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1409.3215",
      "sealed_at": "2026-04-29T05:23:24Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1502-03167",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1502.03167",
      "sealed_at": "2026-04-29T06:25:54Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1506-02640",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1506.02640",
      "sealed_at": "2026-04-29T05:03:17Z",
      "scored": true,
      "overall": 0.7365088888888889,
      "visual": 0.47501646090534977,
      "dom": 0.9999999999999999,
      "interaction": 0.9800131687242798,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.75,
          "visual": 0.5,
          "dom": 0.9999999999999999,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T11:04:30Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.526998016172353,
          "visual": 0.4362381687242798,
          "dom": 0.8728441925095888,
          "interaction": 0.9405134811467298,
          "aj": 0,
          "generated_at": "2026-05-14T11:15:49Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.48300780683823763,
          "visual": 0.36229102366255145,
          "dom": 0.36495820019931813,
          "interaction": 0.847496698943329,
          "aj": 1,
          "generated_at": "2026-05-14T06:03:01Z"
        }
      ]
    },
    {
      "id": "T-arxiv-1512-03385",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1512.03385",
      "sealed_at": "2026-04-29T06:25:59Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1611-07004",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1611.07004",
      "sealed_at": "2026-04-29T05:23:29Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1706",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1706.03762",
      "sealed_at": "2026-04-28T21:02:22Z",
      "scored": true,
      "overall": 0.7369434027777778,
      "visual": 0.475821116255144,
      "dom": 1,
      "interaction": 0.9806568930041153,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7496672222222223,
          "visual": 0.49938374485596704,
          "dom": 1,
          "interaction": 0.9995069958847737,
          "aj": 1,
          "generated_at": "2026-05-14T04:59:02Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.5821599192954009,
          "visual": 0.4207262088477366,
          "dom": 0.5878058236844719,
          "interaction": 0.9091013553238207,
          "aj": 1,
          "generated_at": "2026-05-14T05:16:27Z"
        },
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.576678323498352,
          "visual": 0.4378131430041152,
          "dom": 0.5393161305806478,
          "interaction": 0.919538256442002,
          "aj": 1,
          "generated_at": "2026-05-14T05:18:16Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5656531116693361,
          "visual": 0.43365174897119346,
          "dom": 0.8400478040381041,
          "interaction": 0.9362579194461617,
          "aj": 0.3333333333333333,
          "generated_at": "2026-05-14T09:48:52Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.489899863057361,
          "visual": 0.41038644547325104,
          "dom": 0.3020588434125455,
          "interaction": 0.8817797459394372,
          "aj": 1,
          "generated_at": "2026-05-14T05:20:27Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.42786281902168716,
          "visual": 0.44025745884773665,
          "dom": 0.5388696414634374,
          "interaction": 0.9214639431757519,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:09Z"
        }
      ]
    },
    {
      "id": "T-arxiv-1810-04805",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1810.04805",
      "sealed_at": "2026-04-29T05:03:04Z",
      "scored": true,
      "overall": 0.7363866666666666,
      "visual": 0.47479012345679017,
      "dom": 0.9999999999999999,
      "interaction": 0.9798320987654322,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.75,
          "visual": 0.5,
          "dom": 0.9999999999999999,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T04:59:16Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.5154578544656309,
          "visual": 0.38970511831275717,
          "dom": 0.423133265637611,
          "interaction": 0.8733063123593799,
          "aj": 1,
          "generated_at": "2026-05-14T06:03:15Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.50950305584049,
          "visual": 0.38359130658436214,
          "dom": 0.9088914844558281,
          "interaction": 0.9007991442312117,
          "aj": 0,
          "generated_at": "2026-05-14T11:15:53Z"
        }
      ]
    },
    {
      "id": "T-arxiv-1810-04806",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1810.04806",
      "sealed_at": "2026-04-29T08:28:38Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1810-12281",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1810.12281",
      "sealed_at": "2026-04-29T08:28:11Z",
      "scored": false
    },
    {
      "id": "T-arxiv-1907-11692",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1907.11692",
      "sealed_at": "2026-04-29T07:27:52Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2005-14165",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2005.14165",
      "sealed_at": "2026-04-29T05:03:08Z",
      "scored": true,
      "overall": 0.7310927777777777,
      "visual": 0.4649866255144033,
      "dom": 1,
      "interaction": 0.9719893004115226,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7500000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T04:59:29Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5365456774354107,
          "visual": 0.4330574845679012,
          "dom": 0.9099823157211342,
          "interaction": 0.9404448087023966,
          "aj": 0,
          "generated_at": "2026-05-14T11:15:57Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4831430796588493,
          "visual": 0.42113387345679015,
          "dom": 0.2606509494247778,
          "interaction": 0.8876171620604172,
          "aj": 1,
          "generated_at": "2026-05-14T06:05:11Z"
        }
      ]
    },
    {
      "id": "T-arxiv-2010-11929",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2010.11929",
      "sealed_at": "2026-04-29T05:03:13Z",
      "scored": true,
      "overall": 0.7338137499999999,
      "visual": 0.470025462962963,
      "dom": 1,
      "interaction": 0.9760203703703704,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7500000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T04:59:43Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5099487356854098,
          "visual": 0.379008359053498,
          "dom": 0.9185194125160027,
          "interaction": 0.8977746480771986,
          "aj": 0,
          "generated_at": "2026-05-14T11:16:02Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4444722438641069,
          "visual": 0.31369097222222225,
          "dom": 0.3244366555959569,
          "interaction": 0.8059152214841748,
          "aj": 1,
          "generated_at": "2026-05-14T06:04:36Z"
        }
      ]
    },
    {
      "id": "T-arxiv-2103-00020",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2103.00020",
      "sealed_at": "2026-04-29T01:11:00Z",
      "scored": true,
      "overall": 0.7347353472222222,
      "visual": 0.4717321244855967,
      "dom": 1,
      "interaction": 0.9773856995884774,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7500000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T04:59:59Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.4952229099532613,
          "visual": 0.4239965277777778,
          "dom": 0.7898839064393228,
          "interaction": 0.9251894826515105,
          "aj": 0,
          "generated_at": "2026-05-14T12:15:49Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.44621929168214836,
          "visual": 0.3336796553497942,
          "dom": 0.29461190481294325,
          "interaction": 0.8199178512673649,
          "aj": 1,
          "generated_at": "2026-05-14T06:04:02Z"
        }
      ]
    },
    {
      "id": "T-arxiv-2106-04561",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2106.04561",
      "sealed_at": "2026-04-29T07:27:57Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2106-09685",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2106.09685",
      "sealed_at": "2026-04-29T06:26:09Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2203-15556",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2203.15556",
      "sealed_at": "2026-04-29T07:27:47Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2204-02311",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2204.02311",
      "sealed_at": "2026-04-29T05:23:16Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2205-11916",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2205.11916",
      "sealed_at": "2026-04-29T07:28:02Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2304-08485",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2304.08485",
      "sealed_at": "2026-04-29T08:28:21Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2305-13245",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2305.13245",
      "sealed_at": "2026-04-29T05:23:20Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2308-08155",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2308.08155",
      "sealed_at": "2026-04-29T08:28:16Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2310-08491",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2310.08491",
      "sealed_at": "2026-04-29T07:28:12Z",
      "scored": false
    },
    {
      "id": "T-arxiv-2403-08295",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2403.08295",
      "sealed_at": "2026-04-29T08:28:27Z",
      "scored": false
    },
    {
      "id": "T-arxiv-bahdanau",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1409.0473",
      "sealed_at": "2026-04-29T04:32:26Z",
      "scored": true,
      "overall": 0.7361580555555556,
      "visual": 0.4743667695473251,
      "dom": 1,
      "interaction": 0.9794934156378602,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7500000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T05:00:12Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5225802624536429,
          "visual": 0.4399519032921811,
          "dom": 0.8516689055248298,
          "interaction": 0.9420727830020669,
          "aj": 0,
          "generated_at": "2026-05-14T12:15:53Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.45657253223994243,
          "visual": 0.3488939043209877,
          "dom": 0.30165876013167825,
          "interaction": 0.832559040798902,
          "aj": 1,
          "generated_at": "2026-05-14T06:03:47Z"
        }
      ]
    },
    {
      "id": "T-arxiv-bert",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/1810.04805",
      "sealed_at": "2026-04-29T04:32:22Z",
      "scored": true,
      "overall": 0.7363866666666666,
      "visual": 0.47479012345679017,
      "dom": 0.9999999999999999,
      "interaction": 0.9798320987654322,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.75,
          "visual": 0.5,
          "dom": 0.9999999999999999,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T05:00:26Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5594025292397661,
          "visual": 0.4496538065843622,
          "dom": 0.9557894736842104,
          "interaction": 0.9567756768464372,
          "aj": 0,
          "generated_at": "2026-05-14T12:15:57Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4587169588332881,
          "visual": 0.3699926697530864,
          "dom": 0.2711678587910596,
          "interaction": 0.8474053263885398,
          "aj": 1,
          "generated_at": "2026-05-14T06:04:54Z"
        }
      ]
    },
    {
      "id": "T-arxiv-gpt3",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2005.14165",
      "sealed_at": "2026-04-29T04:32:31Z",
      "scored": true,
      "overall": 0.7310927777777777,
      "visual": 0.4649866255144033,
      "dom": 1,
      "interaction": 0.9719893004115226,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7500000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T05:00:44Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5619496527777778,
          "visual": 0.44826067386831275,
          "dom": 0.9666666666666667,
          "interaction": 0.956386316872428,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:01Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4831597698956803,
          "visual": 0.4248966049382716,
          "dom": 0.2540074831725725,
          "interaction": 0.8901844494954556,
          "aj": 1,
          "generated_at": "2026-05-14T06:03:33Z"
        }
      ]
    },
    {
      "id": "T-arxiv-vit",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://arxiv.org/abs/2010.11929",
      "sealed_at": "2026-04-29T04:32:36Z",
      "scored": true,
      "overall": 0.7338137499999999,
      "visual": 0.470025462962963,
      "dom": 1,
      "interaction": 0.9760203703703704,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7499964550431392,
          "visual": 0.5,
          "dom": 0.9999883133290299,
          "interaction": 0.999999220888602,
          "aj": 1,
          "generated_at": "2026-05-14T05:00:59Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5074594552034429,
          "visual": 0.43754591049382724,
          "dom": 0.8061032863849764,
          "interaction": 0.9371102808207269,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:05Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.5036755740196144,
          "visual": 0.3634951131687243,
          "dom": 0.43095015244561496,
          "interaction": 0.8528594340313539,
          "aj": 1,
          "generated_at": "2026-05-14T06:05:39Z"
        }
      ]
    },
    {
      "id": "T-distill-attention",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://distill.pub/2016/augmented-rnns/",
      "sealed_at": "2026-04-29T05:03:23Z",
      "scored": true,
      "overall": 0.7400360744401947,
      "visual": 0.4815484825102881,
      "dom": 0.9999996501691399,
      "interaction": 0.9852387626861732,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6999998938846391,
          "visual": 0.5,
          "dom": 0.9999996501691399,
          "interaction": 0.9999999766779427,
          "aj": 0.6666666666666666,
          "generated_at": "2026-05-14T05:01:14Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.4608253740668933,
          "visual": 0.475667695473251,
          "dom": 0.5844994016857289,
          "interaction": 0.9528341164909828,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:10Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.3828254890952141,
          "visual": 0.4461781121399177,
          "dom": 0.05018453364722597,
          "interaction": 0.8936214586217494,
          "aj": 0.6666666666666666,
          "generated_at": "2026-05-14T06:05:53Z"
        }
      ]
    },
    {
      "id": "T-distill-circuits",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://distill.pub/2020/circuits/zoom-in/",
      "sealed_at": "2026-04-28T21:15:39Z",
      "scored": true,
      "overall": 0.7114518750000001,
      "visual": 0.44991345164609053,
      "dom": 0.9620833333333332,
      "interaction": 0.9574029835390947,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7180242970204842,
          "visual": 0.4578101851851852,
          "dom": 0.9696927374301675,
          "interaction": 0.964227663976826,
          "aj": 1,
          "generated_at": "2026-05-14T05:01:32Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.44035910550592283,
          "visual": 0.3620792181069959,
          "dom": 0.22473514635652223,
          "interaction": 0.8379790509093649,
          "aj": 1,
          "generated_at": "2026-05-14T06:05:25Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.41427175604095634,
          "visual": 0.4467376543209876,
          "dom": 0.48252776716798795,
          "interaction": 0.9228919746013227,
          "aj": 0,
          "generated_at": "2026-05-14T12:35:53Z"
        }
      ]
    },
    {
      "id": "T-distill-feature-vis",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://distill.pub/2017/feature-visualization/",
      "sealed_at": "2026-04-29T04:32:41Z",
      "scored": true,
      "overall": 0.6984865392394822,
      "visual": 0.42063850308641976,
      "dom": 0.9714563106796117,
      "interaction": 0.9346078898477765,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6688986419770023,
          "visual": 0.45847993827160494,
          "dom": 0.971382885638469,
          "interaction": 0.9648761429931819,
          "aj": 0.6666666666666666,
          "generated_at": "2026-05-14T05:01:45Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.3630376731684502,
          "visual": 0.4136264146090535,
          "dom": 0.37256948114141086,
          "interaction": 0.8890724304300036,
          "aj": 0,
          "generated_at": "2026-05-14T12:35:57Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.33553537492153057,
          "visual": 0.29311818415637864,
          "dom": 0.16676336970467964,
          "interaction": 0.7789454386387481,
          "aj": 0.6666666666666666,
          "generated_at": "2026-05-14T06:02:47Z"
        }
      ]
    },
    {
      "id": "T-distill-momentum",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://distill.pub/2017/momentum/",
      "sealed_at": "2026-04-29T01:10:54Z",
      "scored": true,
      "overall": 0.7415363857231577,
      "visual": 0.48432664609053494,
      "dom": 0.9999999895635237,
      "interaction": 0.9874613161766629,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6999999968342688,
          "visual": 0.5,
          "dom": 0.9999999895635237,
          "interaction": 0.9999999993042349,
          "aj": 0.6666666666666666,
          "generated_at": "2026-05-14T05:02:05Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.32840513796135895,
          "visual": 0.4750574845679012,
          "dom": 0.1490354822901944,
          "interaction": 0.9233150198070007,
          "aj": 0,
          "generated_at": "2026-05-14T12:36:01Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.28910923810813494,
          "visual": 0.38235133744855965,
          "dom": 0.01969071171180018,
          "interaction": 0.840527117406301,
          "aj": 0.3333333333333333,
          "generated_at": "2026-05-14T06:02:25Z"
        }
      ]
    },
    {
      "id": "T-distill-rl",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://distill.pub/2018/building-blocks/",
      "sealed_at": "2026-04-29T06:26:16Z",
      "scored": false
    },
    {
      "id": "T-openreview-wgan",
      "category": "academic-paper",
      "tier": "T1",
      "source_url": "https://openreview.net/forum?id=Hk4_qw5xe",
      "sealed_at": "2026-04-29T01:10:40Z",
      "scored": true,
      "overall": 0.730306059322034,
      "visual": 0.46486265432098767,
      "dom": 0.9976271186440677,
      "interaction": 0.9717319313663948,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.708971885334545,
          "visual": 0.5,
          "dom": 0.8647424791248739,
          "interaction": 0.9909828319416581,
          "aj": 1,
          "generated_at": "2026-05-14T05:02:22Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.45258390345753735,
          "visual": 0.46275321502057615,
          "dom": 0.5803203319112954,
          "interaction": 0.9422239274772141,
          "aj": 0,
          "generated_at": "2026-05-14T14:55:55Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.44763045340497465,
          "visual": 0.4482988683127573,
          "dom": 0.09521669620687612,
          "interaction": 0.8983202077306641,
          "aj": 1,
          "generated_at": "2026-05-14T06:04:18Z"
        }
      ]
    },
    {
      "id": "T-aws-docs",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://docs.aws.amazon.com/",
      "sealed_at": "2026-04-29T07:43:46Z",
      "scored": false
    },
    {
      "id": "T-bitbucket-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://bitbucket.org/",
      "sealed_at": "2026-04-29T05:38:51Z",
      "scored": false
    },
    {
      "id": "T-bun-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://bun.sh/",
      "sealed_at": "2026-04-29T08:44:15Z",
      "scored": false
    },
    {
      "id": "T-codeberg-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://codeberg.org/",
      "sealed_at": "2026-04-29T05:38:29Z",
      "scored": false
    },
    {
      "id": "T-deno-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://deno.com/",
      "sealed_at": "2026-04-29T08:44:04Z",
      "scored": false
    },
    {
      "id": "T-developer-android",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://developer.android.com/",
      "sealed_at": "2026-04-29T07:44:21Z",
      "scored": false
    },
    {
      "id": "T-eclipse-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.eclipse.org/",
      "sealed_at": "2026-04-29T07:43:37Z",
      "scored": false
    },
    {
      "id": "T-elixir-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://elixir-lang.org/",
      "sealed_at": "2026-04-29T08:44:21Z",
      "scored": false
    },
    {
      "id": "T-github-linux",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://github.com/torvalds/linux",
      "sealed_at": "2026-04-28T21:18:30Z",
      "scored": true,
      "overall": 0.6896509947420916,
      "visual": 0.3383397633744856,
      "dom": 0.8183206106870229,
      "interaction": 0.8585598514120566,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.684263394545923,
          "visual": 0.4718614969135803,
          "dom": 0.4357505062132299,
          "interaction": 0.9398725646117462,
          "aj": 1,
          "generated_at": "2026-05-14T05:02:36Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6042942020790056,
          "visual": 0.45615920781893005,
          "dom": 0.11436486120834996,
          "interaction": 0.905885023669034,
          "aj": 1,
          "generated_at": "2026-05-14T05:57:20Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.4939790541157486,
          "visual": 0.46804745370370365,
          "dom": 0.2628684383207428,
          "interaction": 0.9252958588510126,
          "aj": 0,
          "generated_at": "2026-05-14T12:55:52Z"
        }
      ]
    },
    {
      "id": "T-gitlab-explore",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://gitlab.com/explore/projects/trending",
      "sealed_at": "2026-04-29T05:38:38Z",
      "scored": false
    },
    {
      "id": "T-gitlab-gitlab",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://gitlab.com/gitlab-org/gitlab",
      "sealed_at": "2026-04-29T04:49:49Z",
      "scored": true,
      "overall": 0.812042066255429,
      "visual": 0.4783989197530864,
      "dom": 0.9990304709141274,
      "interaction": 0.9826545005300776,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.8247393250991137,
          "visual": 0.5,
          "dom": 0.9988151140868806,
          "interaction": 0.9999210076057921,
          "aj": 1,
          "generated_at": "2026-05-14T05:02:52Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.5840416648441304,
          "visual": 0.43929346707818934,
          "dom": 0.06753872394544885,
          "interaction": 0.889270688592248,
          "aj": 1,
          "generated_at": "2026-05-14T05:56:14Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.4961865801298282,
          "visual": 0.45780709876543213,
          "dom": 0.3003654175373785,
          "interaction": 0.9196033735148376,
          "aj": 0,
          "generated_at": "2026-05-14T12:55:57Z"
        }
      ]
    },
    {
      "id": "T-go-pkg",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://pkg.go.dev/",
      "sealed_at": "2026-04-29T05:19:06Z",
      "scored": false
    },
    {
      "id": "T-hf-gpt2",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://huggingface.co/openai-community/gpt2",
      "sealed_at": "2026-04-28T21:18:39Z",
      "scored": true,
      "overall": 0.8111463811728395,
      "visual": 0.4765192901234568,
      "dom": 1,
      "interaction": 0.9812154320987654,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.8109950748433409,
          "visual": 0.4763220164609054,
          "dom": 0.9998412960518488,
          "interaction": 0.9810470329055142,
          "aj": 1,
          "generated_at": "2026-05-14T05:03:07Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6077238871507381,
          "visual": 0.479357767489712,
          "dom": 0.06774001969003665,
          "interaction": 0.9213355486377721,
          "aj": 1,
          "generated_at": "2026-05-14T05:56:39Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5376298282898855,
          "visual": 0.4704867541152264,
          "dom": 0.4547392880086455,
          "interaction": 0.9400386891594241,
          "aj": 0,
          "generated_at": "2026-05-14T12:56:01Z"
        }
      ]
    },
    {
      "id": "T-jenkins-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.jenkins.io/",
      "sealed_at": "2026-04-29T06:44:29Z",
      "scored": false
    },
    {
      "id": "T-jetbrains-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.jetbrains.com/",
      "sealed_at": "2026-04-29T07:43:24Z",
      "scored": false
    },
    {
      "id": "T-julia-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://julialang.org/",
      "sealed_at": "2026-04-29T08:44:31Z",
      "scored": false
    },
    {
      "id": "T-kubernetes-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://kubernetes.io/",
      "sealed_at": "2026-04-29T06:44:35Z",
      "scored": false
    },
    {
      "id": "T-mdn-array",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Array",
      "sealed_at": "2026-04-29T01:19:59Z",
      "scored": true,
      "overall": 0.7114205897952833,
      "visual": 0.4468800154320987,
      "dom": 0.6261880940470235,
      "interaction": 0.9325832186154805,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7382985176125866,
          "visual": 0.4768683127572016,
          "dom": 0.6679373322083524,
          "interaction": 0.9593571390196516,
          "aj": 1,
          "generated_at": "2026-05-14T05:03:21Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6366400555498936,
          "visual": 0.4566714248971193,
          "dom": 0.26001779482087795,
          "interaction": 0.9160049929057541,
          "aj": 1,
          "generated_at": "2026-05-14T05:56:52Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5208163587253677,
          "visual": 0.437928883744856,
          "dom": 0.46562871507228476,
          "interaction": 0.9147183546673704,
          "aj": 0,
          "generated_at": "2026-05-14T14:15:59Z"
        }
      ]
    },
    {
      "id": "T-mdn-css",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://developer.mozilla.org/en-US/docs/Web/CSS",
      "sealed_at": "2026-04-29T07:43:52Z",
      "scored": false
    },
    {
      "id": "T-mdn-fetch",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://developer.mozilla.org/en-US/docs/Web/API/Fetch_API",
      "sealed_at": "2026-04-29T04:50:51Z",
      "scored": true,
      "overall": 0.689426659069378,
      "visual": 0.4417007458847737,
      "dom": 0.5401055408970976,
      "interaction": 0.9227009661009588,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6927816594413198,
          "visual": 0.46420704732510293,
          "dom": 0.49499773417958626,
          "interaction": 0.9376988201387215,
          "aj": 1,
          "generated_at": "2026-05-14T05:03:35Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6380521433912436,
          "visual": 0.44805054012345674,
          "dom": 0.28955602144729153,
          "interaction": 0.9110775001952515,
          "aj": 1,
          "generated_at": "2026-05-14T05:58:04Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5070346335961009,
          "visual": 0.43634477880658434,
          "dom": 0.4072327913646191,
          "interaction": 0.9095580091362422,
          "aj": 0,
          "generated_at": "2026-05-14T14:16:03Z"
        }
      ]
    },
    {
      "id": "T-mdn-mdx",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://developer.mozilla.org/en-US/docs/Web/CSS/grid",
      "sealed_at": "2026-04-29T05:39:02Z",
      "scored": false
    },
    {
      "id": "T-nodejs-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://nodejs.org/",
      "sealed_at": "2026-04-29T04:49:55Z",
      "scored": true,
      "overall": 0.724572250656168,
      "visual": 0.4560077160493827,
      "dom": 0.9944881889763779,
      "interaction": 0.9644387187712647,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7480179525531158,
          "visual": 0.5,
          "dom": 0.9934657776476344,
          "interaction": 0.9995643851765089,
          "aj": 1,
          "generated_at": "2026-05-14T05:03:51Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.37265475294337247,
          "visual": 0.29308256172839503,
          "dom": 0.1243631965166126,
          "interaction": 0.7760902624838236,
          "aj": 1,
          "generated_at": "2026-05-14T05:57:35Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.3417349574025025,
          "visual": 0.32291486625514404,
          "dom": 0.46382724052107044,
          "interaction": 0.8225870423721865,
          "aj": 0,
          "generated_at": "2026-05-14T14:55:59Z"
        }
      ]
    },
    {
      "id": "T-npm-react",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.npmjs.com/package/react",
      "sealed_at": "2026-04-29T05:18:51Z",
      "scored": false
    },
    {
      "id": "T-postman-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.postman.com/",
      "sealed_at": "2026-04-29T06:44:12Z",
      "scored": false
    },
    {
      "id": "T-pypi-numpy",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://pypi.org/project/numpy/",
      "sealed_at": "2026-04-29T05:18:57Z",
      "scored": false
    },
    {
      "id": "T-python-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.python.org/",
      "sealed_at": "2026-04-29T04:50:33Z",
      "scored": true,
      "overall": 0.7229015277777777,
      "visual": 0.4498176440329218,
      "dom": 0.9999999999999999,
      "interaction": 0.9598541152263375,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7498858795312604,
          "visual": 0.5,
          "dom": 0.999623778674485,
          "interaction": 0.999974918578299,
          "aj": 1,
          "generated_at": "2026-05-14T05:04:04Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.4320157449999875,
          "visual": 0.35903459362139917,
          "dom": 0.6971551575091164,
          "interaction": 0.8670380187310606,
          "aj": 0,
          "generated_at": "2026-05-14T15:15:52Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4057854684226607,
          "visual": 0.27321141975308644,
          "dom": 0.26896033545932113,
          "interaction": 0.7698331581664238,
          "aj": 1,
          "generated_at": "2026-05-14T05:57:48Z"
        }
      ]
    },
    {
      "id": "T-readthedocs-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://about.readthedocs.com/",
      "sealed_at": "2026-04-29T06:43:29Z",
      "scored": false
    },
    {
      "id": "T-rust-docs",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://doc.rust-lang.org/std/",
      "sealed_at": "2026-04-29T05:19:01Z",
      "scored": false
    },
    {
      "id": "T-stackoverflow-js",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://stackoverflow.com/questions/tagged/javascript",
      "sealed_at": "2026-04-28T21:24:13Z",
      "scored": true,
      "bot_challenge": true,
      "overall": 0.8246499151234569,
      "visual": 0.49940663580246913,
      "dom": 1,
      "interaction": 0.9995253086419753,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7908144976015072,
          "visual": 0.4935895061728395,
          "dom": 0.861803131634236,
          "interaction": 0.9856584803805539,
          "aj": 1,
          "generated_at": "2026-05-14T05:04:21Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.7071685893114196,
          "visual": 0.49452777777777773,
          "dom": 0.479078183738776,
          "interaction": 0.9608941011381406,
          "aj": 1,
          "generated_at": "2026-05-14T05:57:05Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6095948252539691,
          "visual": 0.49360108024691357,
          "dom": 0.7198644904922278,
          "interaction": 0.9762051635636794,
          "aj": 0,
          "generated_at": "2026-05-14T15:35:47Z"
        }
      ]
    },
    {
      "id": "T-stackshare-home",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://stackshare.io/",
      "sealed_at": "2026-04-29T06:43:23Z",
      "scored": false
    },
    {
      "id": "T-typescriptlang",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.typescriptlang.org/",
      "sealed_at": "2026-04-29T04:50:44Z",
      "scored": true,
      "overall": 0.7492159027777778,
      "visual": 0.4985479681069959,
      "dom": 0.9999999999999999,
      "interaction": 0.9988383744855968,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7495184027777777,
          "visual": 0.49910815329218106,
          "dom": 0.9999999999999999,
          "interaction": 0.9992865226337448,
          "aj": 1,
          "generated_at": "2026-05-14T05:04:35Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.35194357382736235,
          "visual": 0.2544349279835391,
          "dom": 0.12488586609753155,
          "interaction": 0.7452070001266667,
          "aj": 1,
          "generated_at": "2026-05-14T05:56:27Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.34001583264108304,
          "visual": 0.40503369341563783,
          "dom": 0.3119702358130943,
          "interaction": 0.8781583037867167,
          "aj": 0,
          "generated_at": "2026-05-14T15:55:47Z"
        }
      ]
    },
    {
      "id": "T-vscode-marketplace",
      "category": "developer",
      "tier": "T2",
      "source_url": "https://marketplace.visualstudio.com/vscode",
      "sealed_at": "2026-04-29T05:18:44Z",
      "scored": false
    },
    {
      "id": "T-w3c-html-spec",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://www.w3.org/TR/html52/",
      "sealed_at": "2026-04-29T05:39:11Z",
      "scored": false
    },
    {
      "id": "T-ycombinator",
      "category": "developer",
      "tier": "T1",
      "source_url": "https://news.ycombinator.com/",
      "sealed_at": "2026-04-28T21:05:25Z",
      "scored": true,
      "overall": 0.7262119444444445,
      "visual": 0.4559480452674897,
      "dom": 1,
      "interaction": 0.9647584362139918,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7394551388888889,
          "visual": 0.48047247942386834,
          "dom": 1,
          "interaction": 0.9843779835390948,
          "aj": 1,
          "generated_at": "2026-05-14T05:04:49Z"
        },
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.6504333735527777,
          "visual": 0.46661818415637857,
          "dom": 0.7311853432142855,
          "interaction": 0.9553735702060552,
          "aj": 1,
          "generated_at": "2026-05-14T05:18:32Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.6485090819460776,
          "visual": 0.46626414609053496,
          "dom": 0.7254717902984245,
          "interaction": 0.9547094362256564,
          "aj": 1,
          "generated_at": "2026-05-14T05:16:40Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5894551388888889,
          "visual": 0.48047247942386834,
          "dom": 1,
          "interaction": 0.9843779835390948,
          "aj": 0,
          "generated_at": "2026-05-14T10:15:58Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.554299474648655,
          "visual": 0.46946862139917694,
          "dom": 0.4091859970102183,
          "interaction": 0.9361872969200228,
          "aj": 1,
          "generated_at": "2026-05-14T05:20:41Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.5029583529265009,
          "visual": 0.4643877314814815,
          "dom": 0.7434801470104427,
          "interaction": 0.954408861652548,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:13Z"
        }
      ]
    },
    {
      "id": "T-aesop-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.aesop.com/",
      "sealed_at": "2026-04-29T07:33:40Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-allbirds-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.allbirds.com/",
      "sealed_at": "2026-04-29T06:30:59Z",
      "scored": false
    },
    {
      "id": "T-apple-airpods",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.apple.com/airpods/",
      "sealed_at": "2026-04-28T21:19:47Z",
      "scored": true,
      "overall": 0.7343930673804987,
      "visual": 0.47223366769547326,
      "dom": 0.9979787477745377,
      "interaction": 0.9776521840080145,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7366095021113316,
          "visual": 0.48156777263374484,
          "dom": 0.988668917216844,
          "interaction": 0.9844988125881189,
          "aj": 1,
          "generated_at": "2026-05-14T05:05:05Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.4682145170956884,
          "visual": 0.4296795267489712,
          "dom": 0.19622276698212315,
          "interaction": 0.8901584725313185,
          "aj": 1,
          "generated_at": "2026-05-14T05:16:55Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.3943933571091277,
          "visual": 0.31839673353909465,
          "dom": 0.15096413515829665,
          "interaction": 0.7981149958418289,
          "aj": 1,
          "generated_at": "2026-05-14T05:20:55Z"
        },
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.351454523758867,
          "visual": 0.24315444958847734,
          "dom": 0.14335534389370094,
          "interaction": 0.7374139159303619,
          "aj": 1,
          "generated_at": "2026-05-14T05:18:47Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.3442815270066196,
          "visual": 0.4393006687242798,
          "dom": 0.2650302172379402,
          "interaction": 0.9024425494619533,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:17Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.33075565518929606,
          "visual": 0.3561017232510288,
          "dom": 0.36855183945189174,
          "interaction": 0.8427848345642825,
          "aj": 0,
          "generated_at": "2026-05-14T10:16:02Z"
        }
      ]
    },
    {
      "id": "T-apple-iphone-16",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.apple.com/iphone-16/",
      "sealed_at": "2026-04-29T01:12:59Z",
      "scored": true,
      "overall": 0.6635157505090008,
      "visual": 0.43817489711934154,
      "dom": 0.8249493606523837,
      "interaction": 0.9388698750722989,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6603326799473522,
          "visual": 0.43229333847736623,
          "dom": 0.8249261884711245,
          "interaction": 0.9341630833466347,
          "aj": 1,
          "generated_at": "2026-05-14T05:05:21Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.341875073924648,
          "visual": 0.2694607767489712,
          "dom": 0.06494369608858326,
          "interaction": 0.7532315344717491,
          "aj": 1,
          "generated_at": "2026-05-14T06:00:09Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.31365118888691174,
          "visual": 0.4161909722222222,
          "dom": 0.20519141940740154,
          "interaction": 0.8799655390716046,
          "aj": 0,
          "generated_at": "2026-05-14T10:56:03Z"
        }
      ]
    },
    {
      "id": "T-asos-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.asos.com/us/",
      "sealed_at": "2026-04-29T08:33:59Z",
      "scored": false
    },
    {
      "id": "T-bandh-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.bhphotovideo.com/",
      "sealed_at": "2026-04-29T05:27:27Z",
      "scored": false
    },
    {
      "id": "T-bestbuy",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.bestbuy.com/",
      "sealed_at": "2026-04-29T04:39:20Z",
      "scored": true,
      "overall": 0.746319375,
      "visual": 0.49318402777777776,
      "dom": 1,
      "interaction": 0.9945472222222222,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7305031462607238,
          "visual": 0.5,
          "dom": 0.9357246580023857,
          "interaction": 0.9957149772001591,
          "aj": 1,
          "generated_at": "2026-05-14T05:05:34Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5802683545940569,
          "visual": 0.49956520061728393,
          "dom": 0.9357246580023857,
          "interaction": 0.9953671376939862,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:15Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4959677241207592,
          "visual": 0.06675501543209877,
          "dom": 0.9338022498486565,
          "interaction": 0.6489908290022561,
          "aj": 1,
          "generated_at": "2026-05-14T06:01:04Z"
        }
      ]
    },
    {
      "id": "T-bestbuy-laptops",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.bestbuy.com/site/computers-pcs/laptops/abcat0502000.c?id=abcat0502000",
      "sealed_at": "2026-04-29T05:09:43Z",
      "scored": true,
      "overall": 0.8209785763888889,
      "visual": 0.49318402777777776,
      "dom": 1,
      "interaction": 0.9945472222222222,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.8250000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 1,
          "generated_at": "2026-05-14T05:05:48Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6742802909114789,
          "visual": 0.4992561728395062,
          "dom": 0.998723404255319,
          "interaction": 0.9993198318886263,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:20Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.5233364847689794,
          "visual": 0.06665624999999997,
          "dom": 0.790951351222634,
          "interaction": 0.6393884234148423,
          "aj": 1,
          "generated_at": "2026-05-14T06:02:11Z"
        }
      ]
    },
    {
      "id": "T-bonobos-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://bonobos.com/",
      "sealed_at": "2026-04-29T08:34:41Z",
      "scored": false
    },
    {
      "id": "T-bose-headphones",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.bose.com/c/headphones",
      "sealed_at": "2026-04-29T05:08:55Z",
      "scored": true,
      "overall": 0.7352248095682212,
      "visual": 0.47267939814814813,
      "dom": 0.9999272568183117,
      "interaction": 0.9781386689730726,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6136247320534777,
          "visual": 0.46218994341563785,
          "dom": 0.6177214151946151,
          "interaction": 0.9442667157454846,
          "aj": 1,
          "generated_at": "2026-05-14T05:06:02Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.22865048525964318,
          "visual": 0.030601851851851842,
          "dom": 0.11689720415266988,
          "interaction": 0.5656079617583262,
          "aj": 1,
          "generated_at": "2026-05-14T06:00:36Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.15961607933869948,
          "visual": 0.1711369598765432,
          "dom": 0.13363336595175648,
          "interaction": 0.6791517922980184,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:24Z"
        }
      ]
    },
    {
      "id": "T-bose-soundlink",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.bose.com/c/bluetooth-speakers",
      "sealed_at": "2026-04-29T07:34:57Z",
      "scored": false
    },
    {
      "id": "T-etsy-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.etsy.com/",
      "sealed_at": "2026-04-29T05:26:40Z",
      "scored": false
    },
    {
      "id": "T-everlane-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.everlane.com/",
      "sealed_at": "2026-04-29T07:34:30Z",
      "scored": false
    },
    {
      "id": "T-glossier-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.glossier.com/",
      "sealed_at": "2026-04-29T07:32:51Z",
      "scored": false
    },
    {
      "id": "T-ikea-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.ikea.com/us/en/",
      "sealed_at": "2026-04-29T01:14:43Z",
      "scored": true,
      "overall": 0.7349064149357659,
      "visual": 0.4721424897119342,
      "dom": 0.9998334192021585,
      "interaction": 0.9777028863830246,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.4021128202610782,
          "visual": 0.30623405349794236,
          "dom": 0.19806515836985478,
          "interaction": 0.7915249200230109,
          "aj": 1,
          "generated_at": "2026-05-14T05:06:17Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.378130149042213,
          "visual": 0.35001929012345684,
          "dom": 0.04105406277652628,
          "interaction": 0.8160857029505338,
          "aj": 1,
          "generated_at": "2026-05-14T06:01:46Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2557618531753684,
          "visual": 0.32049678497942385,
          "dom": 0.1847041405048776,
          "interaction": 0.8020443706838644,
          "aj": 0,
          "generated_at": "2026-05-14T13:15:54Z"
        }
      ]
    },
    {
      "id": "T-jcrew-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.jcrew.com/",
      "sealed_at": "2026-04-29T08:34:25Z",
      "scored": false
    },
    {
      "id": "T-knock-knock",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.knockknockstuff.com/",
      "sealed_at": "2026-04-29T08:34:58Z",
      "scored": false
    },
    {
      "id": "T-lego-starwars",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.lego.com/en-us/themes/star-wars",
      "sealed_at": "2026-04-29T01:14:55Z",
      "scored": true,
      "overall": 0.6929160122555271,
      "visual": 0.2772170781893004,
      "dom": 0.9970815278356355,
      "interaction": 0.8215790977404828,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7317334826451093,
          "visual": 0.3430493827160494,
          "dom": 0.9969743038301827,
          "interaction": 0.8742377930948516,
          "aj": 1,
          "generated_at": "2026-05-14T05:06:31Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.44482184880033226,
          "visual": 0.21684889403292176,
          "dom": 0.03127727873140195,
          "interaction": 0.7088976004750975,
          "aj": 1,
          "generated_at": "2026-05-14T06:01:18Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.33732196553429955,
          "visual": 0.2418634259259259,
          "dom": 0.15737520108183303,
          "interaction": 0.7373157541461963,
          "aj": 0,
          "generated_at": "2026-05-14T13:35:53Z"
        }
      ]
    },
    {
      "id": "T-leica-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://leica-camera.com/",
      "sealed_at": "2026-04-29T07:34:39Z",
      "scored": false
    },
    {
      "id": "T-muji-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.muji.com/us/",
      "sealed_at": "2026-04-29T06:30:38Z",
      "scored": false
    },
    {
      "id": "T-newegg-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.newegg.com/",
      "sealed_at": "2026-04-29T05:26:58Z",
      "scored": false
    },
    {
      "id": "T-nike-pegasus",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.nike.com/t/pegasus-41-mens-road-running-shoes-Q4WtzZ",
      "sealed_at": "2026-04-29T04:38:31Z",
      "scored": true,
      "overall": 0.7986149540231368,
      "visual": 0.45528459362139917,
      "dom": 0.9999865626659603,
      "interaction": 0.96422677907485,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6982027425867888,
          "visual": 0.4914567901234568,
          "dom": 0.44656016551795163,
          "interaction": 0.9562694431332956,
          "aj": 1,
          "generated_at": "2026-05-14T05:06:48Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6266332117802246,
          "visual": 0.4853647119341564,
          "dom": 0.13758196245032822,
          "interaction": 0.9307972337106806,
          "aj": 1,
          "generated_at": "2026-05-14T06:01:59Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5309758659675432,
          "visual": 0.48627366255144033,
          "dom": 0.3821563866463338,
          "interaction": 0.9478293558175747,
          "aj": 0,
          "generated_at": "2026-05-14T14:35:59Z"
        }
      ]
    },
    {
      "id": "T-patagonia-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.patagonia.com/home/",
      "sealed_at": "2026-04-29T01:13:50Z",
      "scored": true,
      "overall": 0.6706123173868491,
      "visual": 0.35298662551440335,
      "dom": 0.9999984822277075,
      "interaction": 0.8823891992267031,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6912578011512689,
          "visual": 0.3912246656378601,
          "dom": 0.9999884012312892,
          "interaction": 0.9129789592590406,
          "aj": 1,
          "generated_at": "2026-05-14T05:07:03Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.29096230025971537,
          "visual": 0.14725925925925926,
          "dom": 0.11464494591114964,
          "interaction": 0.6587837371348174,
          "aj": 1,
          "generated_at": "2026-05-14T06:00:51Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2819555217345902,
          "visual": 0.2662798353909465,
          "dom": 0.3675749800774035,
          "interaction": 0.7708622003179175,
          "aj": 0,
          "generated_at": "2026-05-14T15:15:57Z"
        }
      ]
    },
    {
      "id": "T-rei-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.rei.com/",
      "sealed_at": "2026-04-29T05:07:49Z",
      "scored": true,
      "overall": 0.6605231722813238,
      "visual": 0.3426683384773663,
      "dom": 0.9851063829787234,
      "interaction": 0.8731417629804746,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6543267540656695,
          "visual": 0.3311935442386831,
          "dom": 0.9851062862970792,
          "interaction": 0.8639619211440852,
          "aj": 1,
          "generated_at": "2026-05-14T05:07:19Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.35970088056924976,
          "visual": 0.32546373456790123,
          "dom": 0.024012518360164124,
          "interaction": 0.7953051555449987,
          "aj": 1,
          "generated_at": "2026-05-14T05:59:54Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.22804243635599214,
          "visual": 0.26332857510288066,
          "dom": 0.19509342571572494,
          "interaction": 0.7570024217966863,
          "aj": 0,
          "generated_at": "2026-05-14T15:16:02Z"
        }
      ]
    },
    {
      "id": "T-shop-rolex",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.rolex.com/watches",
      "sealed_at": "2026-04-29T06:28:57Z",
      "scored": false
    },
    {
      "id": "T-target",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.target.com/",
      "sealed_at": "2026-04-29T04:39:54Z",
      "scored": true,
      "overall": 0.6516587164245344,
      "visual": 0.32431545781893006,
      "dom": 0.9885550633043261,
      "interaction": 0.8586893704754326,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6705000998428046,
          "visual": 0.35709336419753085,
          "dom": 0.9923176368444108,
          "interaction": 0.8851625338143189,
          "aj": 1,
          "generated_at": "2026-05-14T05:07:34Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.3743135599646188,
          "visual": 0.32654873971193416,
          "dom": 0.07025463907749778,
          "interaction": 0.7992559677080472,
          "aj": 1,
          "generated_at": "2026-05-14T05:59:39Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2299268494714412,
          "visual": 0.3055862911522634,
          "dom": 0.12607775466775484,
          "interaction": 0.786207549899661,
          "aj": 0,
          "generated_at": "2026-05-14T15:35:51Z"
        }
      ]
    },
    {
      "id": "T-target-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.target.com/",
      "sealed_at": "2026-04-29T05:06:59Z",
      "scored": true,
      "overall": 0.6813526874829147,
      "visual": 0.26079758230452676,
      "dom": 0.9885550632874716,
      "interaction": 0.8078750700627862,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6916595388536725,
          "visual": 0.2754300411522634,
          "dom": 0.9961627935174406,
          "interaction": 0.8200882191563069,
          "aj": 1,
          "generated_at": "2026-05-14T05:07:53Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.4704049477766118,
          "visual": 0.2670848765432099,
          "dom": 0.01284032098235416,
          "interaction": 0.7478572559667249,
          "aj": 1,
          "generated_at": "2026-05-14T06:01:31Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.34038804760596003,
          "visual": 0.2509696502057613,
          "dom": 0.14689069992982184,
          "interaction": 0.7439017668265974,
          "aj": 0,
          "generated_at": "2026-05-14T15:35:56Z"
        }
      ]
    },
    {
      "id": "T-tiffany-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.tiffany.com/",
      "sealed_at": "2026-04-29T06:29:48Z",
      "scored": false
    },
    {
      "id": "T-uniqlo-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.uniqlo.com/us/en/",
      "sealed_at": "2026-04-29T05:08:42Z",
      "scored": true,
      "overall": 0.7024600485884585,
      "visual": 0.2987903806584362,
      "dom": 0.9826078363635502,
      "interaction": 0.8378728269509857,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7235480083788995,
          "visual": 0.4235235339506173,
          "dom": 0.7439505606728878,
          "interaction": 0.9217488645386864,
          "aj": 1,
          "generated_at": "2026-05-14T05:08:08Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.48474856470343974,
          "visual": 0.2579102366255144,
          "dom": 0.102643295883574,
          "interaction": 0.7465044090259831,
          "aj": 1,
          "generated_at": "2026-05-14T05:59:27Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.3486374143632519,
          "visual": 0.24885223765432096,
          "dom": 0.19006633703273867,
          "interaction": 0.7450862125923061,
          "aj": 0,
          "generated_at": "2026-05-14T15:55:52Z"
        }
      ]
    },
    {
      "id": "T-walmart",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.walmart.com/",
      "sealed_at": "2026-04-29T04:40:47Z",
      "scored": true,
      "overall": 0.7093253442579975,
      "visual": 0.42468351337448557,
      "dom": 0.999987627590468,
      "interaction": 0.9397459858722863,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7485989564999507,
          "visual": 0.49763798868312753,
          "dom": 0.999586074541962,
          "interaction": 0.9980827959159662,
          "aj": 1,
          "generated_at": "2026-05-14T05:08:23Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5886170109246641,
          "visual": 0.47892734053497943,
          "dom": 0.999987627590468,
          "interaction": 0.9831410476006814,
          "aj": 0,
          "generated_at": "2026-05-14T15:55:56Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.21250388152566224,
          "visual": 0.04640637860082305,
          "dom": 0.03553111125676195,
          "interaction": 0.5728271769644425,
          "aj": 1,
          "generated_at": "2026-05-14T05:59:13Z"
        }
      ]
    },
    {
      "id": "T-walmart-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.walmart.com/",
      "sealed_at": "2026-04-29T05:26:33Z",
      "scored": false
    },
    {
      "id": "T-warby-parker-home",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.warbyparker.com/",
      "sealed_at": "2026-04-29T06:31:16Z",
      "scored": false
    },
    {
      "id": "T-zalando-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.zalando.com/",
      "sealed_at": "2026-04-29T08:33:08Z",
      "scored": false
    },
    {
      "id": "T-zappos",
      "category": "ecommerce",
      "tier": "T1",
      "source_url": "https://www.zappos.com/",
      "sealed_at": "2026-04-29T04:40:59Z",
      "scored": true,
      "overall": 0.6393653424918404,
      "visual": 0.2952878086419753,
      "dom": 0.9997030521709027,
      "interaction": 0.8362104503916404,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.355891516526075,
          "visual": 0.2624061213991769,
          "dom": 0.12371058561709693,
          "interaction": 0.7515056028271481,
          "aj": 1,
          "generated_at": "2026-05-14T05:08:39Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.30545429723334827,
          "visual": 0.21539853395061725,
          "dom": 0.04111787549455467,
          "interaction": 0.7083933521934641,
          "aj": 1,
          "generated_at": "2026-05-14T06:00:24Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.16303466698497565,
          "visual": 0.17899279835390947,
          "dom": 0.13091831606768534,
          "interaction": 0.6852554597543067,
          "aj": 0,
          "generated_at": "2026-05-14T16:04:11Z"
        }
      ]
    },
    {
      "id": "T-zappos-home",
      "category": "ecommerce",
      "tier": "T2",
      "source_url": "https://www.zappos.com/",
      "sealed_at": "2026-04-29T05:27:57Z",
      "scored": false
    },
    {
      "id": "T-criterion-bergman",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.criterion.com/explore/3-ingmar-bergman",
      "sealed_at": "2026-04-29T08:38:11Z",
      "scored": false
    },
    {
      "id": "T-criterion-home",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.criterion.com/",
      "sealed_at": "2026-04-29T05:12:34Z",
      "scored": true,
      "overall": 0.6565795833333333,
      "visual": 0.32699922839506174,
      "dom": 1,
      "interaction": 0.8615993827160494,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6637706267603966,
          "visual": 0.44149022633744855,
          "dom": 0.8198875973785968,
          "interaction": 0.9411846875618654,
          "aj": 1,
          "generated_at": "2026-05-14T05:08:53Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.34598819195661673,
          "visual": 0.27184516460905345,
          "dom": 0.5687641859375645,
          "interaction": 0.7887270774164138,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:30Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.3222328557890192,
          "visual": 0.09517592592592593,
          "dom": 0.3104544696341293,
          "interaction": 0.6301710387163494,
          "aj": 1,
          "generated_at": "2026-05-14T05:58:18Z"
        }
      ]
    },
    {
      "id": "T-criterion-kurosawa",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.criterion.com/explore/2-akira-kurosawa",
      "sealed_at": "2026-04-29T07:38:24Z",
      "scored": false
    },
    {
      "id": "T-imdb-2001-spaceodyssey",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0062622/",
      "sealed_at": "2026-04-29T05:29:45Z",
      "scored": false
    },
    {
      "id": "T-imdb-darkknight",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0468569/",
      "sealed_at": "2026-04-29T05:11:14Z",
      "scored": true,
      "overall": 0.5486593860318932,
      "visual": 0.15371103395061728,
      "dom": 0.9527102012040437,
      "interaction": 0.7198161739074301,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.4190121926087761,
          "visual": 0.41702096193415644,
          "dom": 0.05655232911318121,
          "interaction": 0.8707202581548706,
          "aj": 1,
          "generated_at": "2026-05-14T05:09:07Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.28748399729288954,
          "visual": 0.19783256172839506,
          "dom": 0.013146419646888499,
          "interaction": 0.6924758106925086,
          "aj": 1,
          "generated_at": "2026-05-14T05:58:30Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.16818259879743186,
          "visual": 0.22562268518518516,
          "dom": 0.06487807295856657,
          "interaction": 0.718156686345386,
          "aj": 0,
          "generated_at": "2026-05-14T13:15:58Z"
        }
      ]
    },
    {
      "id": "T-imdb-godfather",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0068646/",
      "sealed_at": "2026-04-29T05:10:19Z",
      "scored": true,
      "overall": 0.5518480060898293,
      "visual": 0.1598801440329218,
      "dom": 0.9522397636661039,
      "interaction": 0.7247200994707444,
      "aj": 1
    },
    {
      "id": "T-imdb-godfather2",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0071562/",
      "sealed_at": "2026-04-29T04:41:54Z",
      "scored": true,
      "overall": 0.5591571017359434,
      "visual": 0.1723951903292181,
      "dom": 0.9540561504115347,
      "interaction": 0.7348532289574768,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.44126814923402635,
          "visual": 0.41200925925925924,
          "dom": 0.13884554692536163,
          "interaction": 0.8721971105357649,
          "aj": 1,
          "generated_at": "2026-05-14T05:09:22Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.2996413824557926,
          "visual": 0.2144807098765432,
          "dom": 0.02358834875536024,
          "interaction": 0.7064904578182586,
          "aj": 1,
          "generated_at": "2026-05-14T05:59:00Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2166016073481153,
          "visual": 0.27552713477366253,
          "dom": 0.13566028979232148,
          "interaction": 0.7627990604717514,
          "aj": 0,
          "generated_at": "2026-05-14T13:16:03Z"
        }
      ]
    },
    {
      "id": "T-imdb-inception",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt1375666/",
      "sealed_at": "2026-04-29T05:28:50Z",
      "scored": false
    },
    {
      "id": "T-imdb-pulp-fiction",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0110912/",
      "sealed_at": "2026-04-29T01:16:46Z",
      "scored": true,
      "overall": 0.5405995961673252,
      "visual": 0.1371792695473251,
      "dom": 0.9555696393794606,
      "interaction": 0.706781391596491,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.36012811079495854,
          "visual": 0.2798204732510288,
          "dom": 0.10667600628374624,
          "interaction": 0.7643014456864062,
          "aj": 1,
          "generated_at": "2026-05-14T05:09:37Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.29674286946367534,
          "visual": 0.20873109567901235,
          "dom": 0.024268388341786948,
          "interaction": 0.7019361024326624,
          "aj": 1,
          "generated_at": "2026-05-14T05:58:46Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.1727108767729077,
          "visual": 0.24446013374485598,
          "dom": 0.04627166335390816,
          "interaction": 0.7319862178861453,
          "aj": 0,
          "generated_at": "2026-05-14T13:35:58Z"
        }
      ]
    },
    {
      "id": "T-imdb-shawshank",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0111161/",
      "sealed_at": "2026-04-28T21:25:28Z",
      "scored": true,
      "overall": 0.5302496223073864,
      "visual": 0.11969894547325104,
      "dom": 0.9525676651159258,
      "interaction": 0.6925970007196627,
      "aj": 1,
      "agent_runs": [
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.3250015144090733,
          "visual": 0.18142656893004117,
          "dom": 0.16603681490170688,
          "interaction": 0.6895437094708133,
          "aj": 1,
          "generated_at": "2026-05-14T05:19:01Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.28925266196789495,
          "visual": 0.17817091049382716,
          "dom": 0.05397924275130209,
          "interaction": 0.6794686779118152,
          "aj": 1,
          "generated_at": "2026-05-14T05:17:09Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.26379953066794104,
          "visual": 0.15662860082304528,
          "dom": 0.008417866670867982,
          "interaction": 0.6591974051031607,
          "aj": 1,
          "generated_at": "2026-05-14T05:21:08Z"
        },
        {
          "model": "gpt-5.5",
          "overall": 0.24035532445701538,
          "visual": 0.3619810956790124,
          "dom": 0.06006219601213855,
          "interaction": 0.8269223562773526,
          "aj": 0,
          "generated_at": "2026-05-14T07:51:28Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.17771163575372179,
          "visual": 0.23256893004115223,
          "dom": 0.08392663801593264,
          "interaction": 0.7249835865673173,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:22Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.15562816848657565,
          "visual": 0.1804553755144033,
          "dom": 0.10389757925977322,
          "interaction": 0.6846241390288409,
          "aj": 0,
          "generated_at": "2026-05-14T10:16:07Z"
        }
      ]
    },
    {
      "id": "T-imdb-top-250",
      "category": "film-tv",
      "tier": "T2",
      "source_url": "https://www.imdb.com/chart/top/",
      "sealed_at": "2026-04-29T04:44:56Z",
      "scored": true,
      "overall": 0.7803993442746003,
      "visual": 0.42574562757201645,
      "dom": 0.9964064727595939,
      "interaction": 0.9403569335749195,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5485226166878399,
          "visual": 0.4563217592592592,
          "dom": 0.5422399032948951,
          "interaction": 0.9345400676270671,
          "aj": 0,
          "generated_at": "2026-05-14T07:58:46Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.30715038167455694,
          "visual": 0.09832973251028808,
          "dom": 0.4051629067885772,
          "interaction": 0.6390079797941356,
          "aj": 0,
          "generated_at": "2026-05-14T13:36:03Z"
        }
      ]
    },
    {
      "id": "T-imdb-toy-story",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0114709/",
      "sealed_at": "2026-04-29T06:32:17Z",
      "scored": false
    },
    {
      "id": "T-imdb-tv-breaking-bad",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0903747/",
      "sealed_at": "2026-04-29T06:35:37Z",
      "scored": false
    },
    {
      "id": "T-imdb-tv-sopranos",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0141842/",
      "sealed_at": "2026-04-29T07:36:47Z",
      "scored": false
    },
    {
      "id": "T-imdb-tv-the-wire",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0306414/",
      "sealed_at": "2026-04-29T07:35:52Z",
      "scored": false
    },
    {
      "id": "T-imdb-tv-true-detective",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt2356777/",
      "sealed_at": "2026-04-29T08:36:45Z",
      "scored": false
    },
    {
      "id": "T-imdb-twin-peaks",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.imdb.com/title/tt0098936/",
      "sealed_at": "2026-04-29T07:37:42Z",
      "scored": false
    },
    {
      "id": "T-letterboxd-home",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://letterboxd.com/",
      "sealed_at": "2026-04-28T21:22:37Z",
      "scored": true,
      "bot_challenge": true,
      "overall": 0.7495001388888889,
      "visual": 0.49907433127572015,
      "dom": 1,
      "interaction": 0.9992594650205762,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5597022850927392,
          "visual": 0.4976581790123457,
          "dom": 0.8713193464595799,
          "interaction": 0.9895478329738485,
          "aj": 0,
          "generated_at": "2026-05-14T08:02:26Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5082532385200573,
          "visual": 0.493650462962963,
          "dom": 0.7088417203957933,
          "interaction": 0.9755098183967567,
          "aj": 0,
          "generated_at": "2026-05-14T13:36:07Z"
        }
      ]
    },
    {
      "id": "T-letterboxd-list-iconic",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://letterboxd.com/dave/list/official-top-250-narrative-feature-films/",
      "sealed_at": "2026-04-29T05:12:39Z",
      "scored": true,
      "overall": 0.7497929166666667,
      "visual": 0.499616512345679,
      "dom": 1,
      "interaction": 0.9996932098765431,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6000000000000001,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 0,
          "generated_at": "2026-05-14T08:06:07Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5978069066383165,
          "visual": 0.5,
          "dom": 0.9927700218845599,
          "interaction": 0.9995180014589707,
          "aj": 0,
          "generated_at": "2026-05-14T13:36:12Z"
        }
      ]
    },
    {
      "id": "T-letterboxd-parasite",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://letterboxd.com/film/parasite-2019/",
      "sealed_at": "2026-04-29T04:42:40Z",
      "scored": true,
      "overall": 0.7169019444444444,
      "visual": 0.43870730452674894,
      "dom": 1,
      "interaction": 0.9509658436213991,
      "aj": 1,
      "agent_runs": [
        {
          "model": "claude-opus-4-7",
          "overall": 0.322331204039332,
          "visual": 0.37854668209876546,
          "dom": 0.30082196386592985,
          "interaction": 0.8562254766034076,
          "aj": 0,
          "generated_at": "2026-05-14T13:55:57Z"
        },
        {
          "model": "gpt-5.5",
          "overall": 0.2996143900354829,
          "visual": 0.4271674382716049,
          "dom": 0.13937573638071288,
          "interaction": 0.8843589997093316,
          "aj": 0,
          "generated_at": "2026-05-14T08:13:25Z"
        }
      ]
    },
    {
      "id": "T-letterboxd-paul-thomas-anderson",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://letterboxd.com/director/paul-thomas-anderson/",
      "sealed_at": "2026-04-29T05:31:15Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-letterboxd-spielberg",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://letterboxd.com/director/steven-spielberg/",
      "sealed_at": "2026-04-29T06:33:51Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-metacritic-casablanca",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.metacritic.com/movie/casablanca",
      "sealed_at": "2026-04-29T01:15:51Z",
      "scored": true,
      "overall": 0.6339436545527635,
      "visual": 0.28508731995884773,
      "dom": 0.9999884673900625,
      "interaction": 0.8280690871264157,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.35877320056050066,
          "visual": 0.39301273148148147,
          "dom": 0.39520766668296914,
          "interaction": 0.8740906962973831,
          "aj": 0,
          "generated_at": "2026-05-14T08:27:55Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.24556564029371797,
          "visual": 0.29445949074074074,
          "dom": 0.19744235811115812,
          "interaction": 0.782063749800003,
          "aj": 0,
          "generated_at": "2026-05-14T14:16:08Z"
        }
      ]
    },
    {
      "id": "T-metacritic-witcher",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.metacritic.com/game/the-witcher-3-wild-hunt",
      "sealed_at": "2026-04-29T04:44:20Z",
      "scored": true,
      "overall": 0.6466950020443716,
      "visual": 0.3087012602880658,
      "dom": 0.9999878730400529,
      "interaction": 0.8469601997664563,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.3070127058363425,
          "visual": 0.3833146862139918,
          "dom": 0.2418333251014954,
          "interaction": 0.8561073039779599,
          "aj": 0,
          "generated_at": "2026-05-14T08:35:12Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.25407566773689316,
          "visual": 0.2765164609053498,
          "dom": 0.2574399302681459,
          "interaction": 0.7717091640754896,
          "aj": 0,
          "generated_at": "2026-05-14T14:16:12Z"
        }
      ]
    },
    {
      "id": "T-mubi-films",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://mubi.com/",
      "sealed_at": "2026-04-29T08:37:05Z",
      "scored": false
    },
    {
      "id": "T-rt-citizen-kane",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/m/citizen_kane",
      "sealed_at": "2026-04-29T04:43:31Z",
      "scored": true,
      "overall": 0.4734293296063664,
      "visual": 0.26251761831275716,
      "dom": 0.5109993924752005,
      "interaction": 0.7774140541485526,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.27331235684051425,
          "visual": 0.343554012345679,
          "dom": 0.20151601156213486,
          "interaction": 0.8216109439806857,
          "aj": 0,
          "generated_at": "2026-05-14T08:46:06Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.19578365687145244,
          "visual": 0.21298392489711934,
          "dom": 0.17837034316596045,
          "interaction": 0.7156118294620929,
          "aj": 0,
          "generated_at": "2026-05-14T15:16:06Z"
        }
      ]
    },
    {
      "id": "T-rt-citizenkane",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/m/citizen_kane",
      "sealed_at": "2026-04-29T05:12:05Z",
      "scored": true,
      "overall": 0.4755116919333059,
      "visual": 0.2659645061728395,
      "dom": 0.511728105274635,
      "interaction": 0.7802201452899139,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.24851160237882938,
          "visual": 0.21148791152263377,
          "dom": 0.35486196754925425,
          "interaction": 0.726181127054724,
          "aj": 0,
          "generated_at": "2026-05-14T08:53:23Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.23815524829649296,
          "visual": 0.22695717592592593,
          "dom": 0.29318145042799865,
          "interaction": 0.7344445041026074,
          "aj": 0,
          "generated_at": "2026-05-14T15:36:01Z"
        }
      ]
    },
    {
      "id": "T-rt-godfather",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/m/the_godfather",
      "sealed_at": "2026-04-29T01:15:09Z",
      "scored": true,
      "overall": 0.5266981167518986,
      "visual": 0.3667404835390946,
      "dom": 0.5010711724421566,
      "interaction": 0.8601304649940863,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.22213342062698913,
          "visual": 0.3615708590534979,
          "dom": 0.0007202969387920548,
          "interaction": 0.822638040372051,
          "aj": 0,
          "generated_at": "2026-05-14T09:00:40Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.21933189813144105,
          "visual": 0.24858500514403292,
          "dom": 0.19262416050658226,
          "interaction": 0.7450429481489985,
          "aj": 0,
          "generated_at": "2026-05-14T15:36:06Z"
        }
      ]
    },
    {
      "id": "T-rt-pulp-fiction",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/m/pulp_fiction",
      "sealed_at": "2026-04-29T05:30:17Z",
      "scored": false
    },
    {
      "id": "T-rt-tv-better-call-saul",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/tv/better_call_saul",
      "sealed_at": "2026-04-29T08:37:38Z",
      "scored": false
    },
    {
      "id": "T-rt-tv-house-of-cards",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/tv/house_of_cards",
      "sealed_at": "2026-04-29T06:34:42Z",
      "scored": false
    },
    {
      "id": "T-rt-tv-watchmen",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.rottentomatoes.com/tv/watchmen_2019",
      "sealed_at": "2026-04-29T07:37:54Z",
      "scored": false
    },
    {
      "id": "T-tmdb-godfather",
      "category": "film-tv",
      "tier": "T1",
      "source_url": "https://www.themoviedb.org/movie/238-the-godfather",
      "sealed_at": "2026-04-29T05:30:27Z",
      "scored": false
    },
    {
      "id": "T-2048",
      "category": "games",
      "tier": "T3",
      "source_url": "https://2048game.com/",
      "sealed_at": "2026-04-29T01:25:47Z",
      "scored": true,
      "overall": 0.6483978652263374,
      "visual": 0.12852109053497943,
      "dom": 0.9999999999999999,
      "interaction": 0.7028168724279835,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.7160548775915261,
          "visual": 0.4859670781893004,
          "dom": 0.9978608883114044,
          "interaction": 0.9886310551055338,
          "aj": 0,
          "generated_at": "2026-05-14T09:11:33Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5087116061800487,
          "visual": 0.14687795781893004,
          "dom": 0.9950891772805637,
          "interaction": 0.7171749780738482,
          "aj": 0,
          "generated_at": "2026-05-14T10:56:07Z"
        }
      ]
    },
    {
      "id": "T-2048-original",
      "category": "games",
      "tier": "T3",
      "source_url": "https://gabrielecirulli.github.io/2048/",
      "sealed_at": "2026-04-29T08:52:32Z",
      "scored": false
    },
    {
      "id": "T-adarkroom",
      "category": "games",
      "tier": "T3",
      "source_url": "https://adarkroom.doublespeakgames.com/",
      "sealed_at": "2026-04-29T05:26:04Z",
      "scored": false
    },
    {
      "id": "T-arcade-skifree",
      "category": "games",
      "tier": "T3",
      "source_url": "https://basicallydan.github.io/skifree.js/",
      "sealed_at": "2026-04-29T06:48:12Z",
      "scored": false
    },
    {
      "id": "T-bgg-catan",
      "category": "games",
      "tier": "T1",
      "source_url": "https://boardgamegeek.com/boardgame/13/catan",
      "sealed_at": "2026-04-29T05:26:54Z",
      "scored": false
    },
    {
      "id": "T-bgg-gloomhaven",
      "category": "games",
      "tier": "T2",
      "source_url": "https://boardgamegeek.com/boardgame/174430/gloomhaven",
      "sealed_at": "2026-04-29T01:25:37Z",
      "scored": true,
      "overall": 0.7637685303758479,
      "visual": 0.39972530864197525,
      "dom": 0.9905936285321931,
      "interaction": 0.9191531554823932,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6236842343674558,
          "visual": 0.4170183899176954,
          "dom": 0.9892881105273432,
          "interaction": 0.9329005859693126,
          "aj": 0,
          "generated_at": "2026-05-14T09:15:14Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.3769449499201524,
          "visual": 0.2968118569958848,
          "dom": 0.1901179740571834,
          "interaction": 0.7834573505338535,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:35Z"
        }
      ]
    },
    {
      "id": "T-cardgames-solitaire",
      "category": "games",
      "tier": "T3",
      "source_url": "https://www.cardgames.io/solitaire/",
      "sealed_at": "2026-04-29T01:26:06Z",
      "scored": true,
      "overall": 0.8743263628194305,
      "visual": 0.49889840534979424,
      "dom": 0.9999907530891998,
      "interaction": 0.9991181078191155,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6845972123060844,
          "visual": 0.45756159979423866,
          "dom": 0.9193590912866599,
          "interaction": 0.9606732192545017,
          "aj": 0,
          "generated_at": "2026-05-14T09:26:08Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6187272388412491,
          "visual": 0.4246880144032921,
          "dom": 0.6648197225291164,
          "interaction": 0.9174050596912414,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:39Z"
        }
      ]
    },
    {
      "id": "T-cardgames-spider",
      "category": "games",
      "tier": "T2",
      "source_url": "https://cardgames.io/spider/",
      "sealed_at": "2026-04-29T05:43:36Z",
      "scored": false
    },
    {
      "id": "T-chesscom-learn",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.chess.com/learn-chess",
      "sealed_at": "2026-04-29T01:25:04Z",
      "scored": true,
      "overall": 0.8107708024691358,
      "visual": 0.47588271604938265,
      "dom": 0.9999999999999999,
      "interaction": 0.9807061728395062,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6504045813585065,
          "visual": 0.47651157407407413,
          "dom": 0.9511943302491037,
          "interaction": 0.9779555479425328,
          "aj": 0,
          "generated_at": "2026-05-14T09:29:48Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5419062941606612,
          "visual": 0.4353244598765433,
          "dom": 0.568476649243185,
          "interaction": 0.919491344517447,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:43Z"
        }
      ]
    },
    {
      "id": "T-chesscom-openings",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.chess.com/openings",
      "sealed_at": "2026-04-29T04:59:30Z",
      "scored": true,
      "overall": 0.7762576993312758,
      "visual": 0.4173859310699588,
      "dom": 1,
      "interaction": 0.9339087448559672,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.489489603971083,
          "visual": 0.40182960390946504,
          "dom": 0.4200460802931753,
          "interaction": 0.8828000884804504,
          "aj": 0,
          "generated_at": "2026-05-14T09:33:29Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.3669483286303653,
          "visual": 0.31142425411522634,
          "dom": 0.10549099410173542,
          "interaction": 0.7895054695656301,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:47Z"
        }
      ]
    },
    {
      "id": "T-cookie-clicker",
      "category": "games",
      "tier": "T3",
      "source_url": "https://orteil.dashnet.org/cookieclicker/",
      "sealed_at": "2026-04-29T05:25:59Z",
      "scored": false
    },
    {
      "id": "T-emoji-mash",
      "category": "games",
      "tier": "T2",
      "source_url": "https://emojikitchen.dev/",
      "sealed_at": "2026-04-29T08:50:59Z",
      "scored": false
    },
    {
      "id": "T-gamefaqs-zelda",
      "category": "games",
      "tier": "T1",
      "source_url": "https://gamefaqs.gamespot.com/snes/588741-the-legend-of-zelda-a-link-to-the-past",
      "sealed_at": "2026-04-29T05:00:18Z",
      "scored": true,
      "bot_challenge": true,
      "overall": 0.7496455555555556,
      "visual": 0.49934362139917693,
      "dom": 1,
      "interaction": 0.9994748971193417,
      "aj": 1
    },
    {
      "id": "T-gamesgrabr",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.crazygames.com/",
      "sealed_at": "2026-04-29T05:44:45Z",
      "scored": false
    },
    {
      "id": "T-itch-io-home",
      "category": "games",
      "tier": "T2",
      "source_url": "https://itch.io/",
      "sealed_at": "2026-04-29T05:25:16Z",
      "scored": false
    },
    {
      "id": "T-lichess-home",
      "category": "games",
      "tier": "T2",
      "source_url": "https://lichess.org/",
      "sealed_at": "2026-04-28T21:19:39Z",
      "scored": true,
      "overall": 0.7405834130658436,
      "visual": 0.35692103909465017,
      "dom": 1,
      "interaction": 0.8855368312757202,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6749935506687244,
          "visual": 0.4999890689300412,
          "dom": 1,
          "interaction": 0.999991255144033,
          "aj": 0,
          "generated_at": "2026-05-14T09:40:47Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5019983105683341,
          "visual": 0.34593454218107,
          "dom": 0.6268042303704674,
          "interaction": 0.8518679157695538,
          "aj": 0,
          "generated_at": "2026-05-14T13:56:01Z"
        }
      ]
    },
    {
      "id": "T-minecraft",
      "category": "games",
      "tier": "T1",
      "source_url": "https://www.minecraft.net/en-us",
      "sealed_at": "2026-04-29T05:00:52Z",
      "scored": true,
      "overall": 0.7304457537418511,
      "visual": 0.4637912808641975,
      "dom": 0.9999949299181902,
      "interaction": 0.9710326866859041,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5252479689121421,
          "visual": 0.4410545267489712,
          "dom": 0.8585006301132891,
          "interaction": 0.9434103300733963,
          "aj": 0,
          "generated_at": "2026-05-14T09:44:37Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.29241373852904856,
          "visual": 0.29986741255144034,
          "dom": 0.3422593486305631,
          "interaction": 0.7960445532831898,
          "aj": 0,
          "generated_at": "2026-05-14T14:36:04Z"
        }
      ]
    },
    {
      "id": "T-newgrounds-home",
      "category": "games",
      "tier": "T1",
      "source_url": "https://www.newgrounds.com/",
      "sealed_at": "2026-04-29T05:25:24Z",
      "scored": false
    },
    {
      "id": "T-nintendo",
      "category": "games",
      "tier": "T1",
      "source_url": "https://www.nintendo.com/us/",
      "sealed_at": "2026-04-29T05:01:14Z",
      "scored": true,
      "overall": 0.6520648781525261,
      "visual": 0.3259663065843621,
      "dom": 0.986955184385617,
      "interaction": 0.8599033908931976,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.2759983472518877,
          "visual": 0.3790819187242798,
          "dom": 0.1471234433212415,
          "interaction": 0.8464070978675067,
          "aj": 0,
          "generated_at": "2026-05-14T09:55:32Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.23045771233409765,
          "visual": 0.2755399948559671,
          "dom": 0.18131686300618272,
          "interaction": 0.7658531200851858,
          "aj": 0,
          "generated_at": "2026-05-14T14:56:03Z"
        }
      ]
    },
    {
      "id": "T-pacman-doodle",
      "category": "games",
      "tier": "T3",
      "source_url": "https://www.google.com/doodles/30th-anniversary-of-pac-man",
      "sealed_at": "2026-04-29T06:49:34Z",
      "scored": false
    },
    {
      "id": "T-papers-please",
      "category": "games",
      "tier": "T1",
      "source_url": "https://papersplea.se/",
      "sealed_at": "2026-04-29T06:49:25Z",
      "scored": false
    },
    {
      "id": "T-pogo-home",
      "category": "games",
      "tier": "T1",
      "source_url": "https://www.pogo.com/",
      "sealed_at": "2026-04-29T05:44:38Z",
      "scored": false
    },
    {
      "id": "T-puzzle-sudoku",
      "category": "games",
      "tier": "T3",
      "source_url": "https://sudoku.com/",
      "sealed_at": "2026-04-29T07:50:14Z",
      "scored": false
    },
    {
      "id": "T-puzzlewikipedia",
      "category": "games",
      "tier": "T1",
      "source_url": "https://en.wikipedia.org/wiki/Puzzle_video_game",
      "sealed_at": "2026-04-29T05:44:27Z",
      "scored": false
    },
    {
      "id": "T-quiz-sporcle",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.sporcle.com/",
      "sealed_at": "2026-04-29T07:48:57Z",
      "scored": false
    },
    {
      "id": "T-snake-google",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.google.com/fbx?fbx=snake_arcade",
      "sealed_at": "2026-04-29T07:48:06Z",
      "scored": false
    },
    {
      "id": "T-tetris-com",
      "category": "games",
      "tier": "T2",
      "source_url": "https://tetris.com/",
      "sealed_at": "2026-04-29T06:48:52Z",
      "scored": false
    },
    {
      "id": "T-trivia-jetpunk",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.jetpunk.com/",
      "sealed_at": "2026-04-29T08:53:20Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-typeracer",
      "category": "games",
      "tier": "T3",
      "source_url": "https://play.typeracer.com/",
      "sealed_at": "2026-04-29T04:58:40Z",
      "scored": true,
      "overall": 0.8332579864436909,
      "visual": 0.4315729166666667,
      "dom": 0.999991707094579,
      "interaction": 0.9452577804729719,
      "aj": 1,
      "agent_runs": [
        {
          "model": "claude-opus-4-7",
          "overall": 0.5546751758372436,
          "visual": 0.2939008487654321,
          "dom": 0.7521981005018332,
          "interaction": 0.8186005523791345,
          "aj": 0,
          "generated_at": "2026-05-14T15:56:01Z"
        },
        {
          "model": "gpt-5.5",
          "overall": 0.5357276560752928,
          "visual": 0.3190474537037037,
          "dom": 0.5617150517557419,
          "interaction": 0.8260189664133457,
          "aj": 0,
          "generated_at": "2026-05-14T09:59:14Z"
        }
      ]
    },
    {
      "id": "T-typeracer-home",
      "category": "games",
      "tier": "T2",
      "source_url": "https://play.typeracer.com/",
      "sealed_at": "2026-04-29T05:44:20Z",
      "scored": false
    },
    {
      "id": "T-typeracer-text-archive",
      "category": "games",
      "tier": "T2",
      "source_url": "https://typeracerdata.com/texts",
      "sealed_at": "2026-04-29T08:51:43Z",
      "scored": false
    },
    {
      "id": "T-wordlearchive",
      "category": "games",
      "tier": "T2",
      "source_url": "https://www.devangthakkar.com/wordle_archive/",
      "sealed_at": "2026-04-29T08:51:06Z",
      "scored": false
    },
    {
      "id": "T-yandere-simulator",
      "category": "games",
      "tier": "T1",
      "source_url": "https://yanderesimulator.com/",
      "sealed_at": "2026-04-29T06:49:13Z",
      "scored": false
    },
    {
      "id": "T-aitests-arc",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://allenai.org/data/arc",
      "sealed_at": "2026-04-29T05:23:52Z",
      "scored": false
    },
    {
      "id": "T-aitests-arenahard",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/lmarena/arena-hard-auto",
      "sealed_at": "2026-04-29T07:29:09Z",
      "scored": false
    },
    {
      "id": "T-aitests-glue-tasks",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://gluebenchmark.com/tasks",
      "sealed_at": "2026-04-29T07:29:16Z",
      "scored": false
    },
    {
      "id": "T-aitests-gpqa",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/idavidrein/gpqa",
      "sealed_at": "2026-04-29T06:27:01Z",
      "scored": false
    },
    {
      "id": "T-aitests-hellaswag",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://rowanzellers.com/hellaswag/",
      "sealed_at": "2026-04-29T08:29:44Z",
      "scored": false
    },
    {
      "id": "T-aitests-humaneval",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/openai/human-eval",
      "sealed_at": "2026-04-29T05:24:00Z",
      "scored": false
    },
    {
      "id": "T-aitests-mathvista",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://mathvista.github.io/",
      "sealed_at": "2026-04-29T06:26:51Z",
      "scored": false
    },
    {
      "id": "T-aitests-mmlu",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/hendrycks/test",
      "sealed_at": "2026-04-29T07:28:22Z",
      "scored": false
    },
    {
      "id": "T-aitests-mmmu-pro",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://mmmu-benchmark.github.io/#leaderboard",
      "sealed_at": "2026-04-29T07:29:01Z",
      "scored": false
    },
    {
      "id": "T-aitests-mt-bench",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://lmarena.ai/",
      "sealed_at": "2026-04-29T08:30:03Z",
      "scored": false
    },
    {
      "id": "T-aitests-superglue",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://super.gluebenchmark.com/",
      "sealed_at": "2026-04-29T07:29:24Z",
      "scored": false
    },
    {
      "id": "T-aitests-truthfulqa",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/sylinrl/TruthfulQA",
      "sealed_at": "2026-04-29T08:29:37Z",
      "scored": false
    },
    {
      "id": "T-aitests-winogrande",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://winogrande.allenai.org/",
      "sealed_at": "2026-04-29T08:29:55Z",
      "scored": false
    },
    {
      "id": "T-bigbench-home",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/google/BIG-bench",
      "sealed_at": "2026-04-29T01:11:33Z",
      "scored": true,
      "overall": 0.6330217013888889,
      "visual": 0.3151170267489712,
      "dom": 0.9434895833333333,
      "interaction": 0.8483262602880659,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5012409075670499,
          "visual": 0.4387560442386831,
          "dom": 0.783448275862069,
          "interaction": 0.9365680537817512,
          "aj": 0,
          "generated_at": "2026-05-14T10:02:55Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.39397263106495106,
          "visual": 0.4443685699588477,
          "dom": 0.41982506578188994,
          "interaction": 0.9168165270192042,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:51Z"
        }
      ]
    },
    {
      "id": "T-bigcodebench",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://bigcode-bench.github.io/",
      "sealed_at": "2026-04-29T06:27:06Z",
      "scored": false
    },
    {
      "id": "T-cifar10",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://www.cs.toronto.edu/~kriz/cifar.html",
      "sealed_at": "2026-04-29T05:23:34Z",
      "scored": false
    },
    {
      "id": "T-coco-dataset",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://cocodataset.org/",
      "sealed_at": "2026-04-29T05:04:21Z",
      "scored": true,
      "overall": 0.7014514583333333,
      "visual": 0.41009529320987653,
      "dom": 1,
      "interaction": 0.9280762345679013,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5703881844849218,
          "visual": 0.4479552469135803,
          "dom": 0.9950297290711709,
          "interaction": 0.9580328461356089,
          "aj": 0,
          "generated_at": "2026-05-14T10:10:12Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5661086083913748,
          "visual": 0.44006159979423876,
          "dom": 0.9949736631950086,
          "interaction": 0.9517141907150583,
          "aj": 0,
          "generated_at": "2026-05-14T12:16:56Z"
        }
      ]
    },
    {
      "id": "T-coco-leaderboard",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://cocodataset.org/#detection-leaderboard",
      "sealed_at": "2026-04-29T06:26:29Z",
      "scored": false
    },
    {
      "id": "T-glue-benchmark",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://gluebenchmark.com/leaderboard",
      "sealed_at": "2026-04-29T01:11:14Z",
      "scored": true,
      "overall": 0.8244452419547087,
      "visual": 0.4993383487654321,
      "dom": 0.9992528008322898,
      "interaction": 0.9994208657344984,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6747378898456552,
          "visual": 0.49983436213991767,
          "dom": 0.9992528008322898,
          "interaction": 0.9998176764340868,
          "aj": 0,
          "generated_at": "2026-05-14T10:10:16Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6172332970234592,
          "visual": 0.496608024691358,
          "dom": 0.7465207384343543,
          "interaction": 0.9803878023153767,
          "aj": 0,
          "generated_at": "2026-05-14T12:56:04Z"
        }
      ]
    },
    {
      "id": "T-helm-lite",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://crfm.stanford.edu/helm/lite/latest/",
      "sealed_at": "2026-04-28T21:15:44Z",
      "scored": true,
      "overall": 0.8249250360082304,
      "visual": 0.4998729423868313,
      "dom": 1,
      "interaction": 0.9998983539094649,
      "aj": 1,
      "agent_runs": [
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.7372053206580284,
          "visual": 0.4657925668724279,
          "dom": 0.692671391833163,
          "interaction": 0.9521454796201533,
          "aj": 1,
          "generated_at": "2026-05-14T05:19:14Z"
        },
        {
          "model": "gpt-5.5",
          "overall": 0.6749250360082304,
          "visual": 0.4998729423868313,
          "dom": 1,
          "interaction": 0.9998983539094649,
          "aj": 0,
          "generated_at": "2026-05-14T10:13:56Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6444969637912464,
          "visual": 0.46723533950617285,
          "dom": 0.2674005158300203,
          "interaction": 0.9249483059936064,
          "aj": 1,
          "generated_at": "2026-05-14T05:21:20Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.6420470483538322,
          "visual": 0.4637263374485597,
          "dom": 0.26567504208719095,
          "interaction": 0.9220260727646604,
          "aj": 1,
          "generated_at": "2026-05-14T05:17:22Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6359016622337881,
          "visual": 0.4816748971193416,
          "dom": 0.8714248769698936,
          "interaction": 0.9767682428267995,
          "aj": 0,
          "generated_at": "2026-05-14T10:16:11Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.5856881557412172,
          "visual": 0.4666746399176955,
          "dom": 0.6834096281353496,
          "interaction": 0.9522336871431797,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:26Z"
        }
      ]
    },
    {
      "id": "T-humaneval",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://github.com/openai/human-eval",
      "sealed_at": "2026-04-29T04:35:48Z",
      "scored": true,
      "overall": 0.6447041813973065,
      "visual": 0.3367822788065844,
      "dom": 0.9434343434343435,
      "interaction": 0.8656547792742236,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.4853484037869699,
          "visual": 0.47084670781893007,
          "dom": 0.6739269721914758,
          "interaction": 0.9549391644012424,
          "aj": 0,
          "generated_at": "2026-05-14T10:21:13Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.40701405260655277,
          "visual": 0.46965252057613166,
          "dom": 0.41780777416079684,
          "interaction": 0.9369092014049585,
          "aj": 0,
          "generated_at": "2026-05-14T13:16:07Z"
        }
      ]
    },
    {
      "id": "T-imagenet",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://www.image-net.org/",
      "sealed_at": "2026-04-29T05:04:25Z",
      "scored": true,
      "overall": 0.749769049238915,
      "visual": 0.5,
      "dom": 0.9992386238645548,
      "interaction": 0.9999492415909703,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5987920245980204,
          "visual": 0.4978630401234568,
          "dom": 0.9998219217517157,
          "interaction": 0.9982785602155465,
          "aj": 0,
          "generated_at": "2026-05-14T10:24:53Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5785668878112746,
          "visual": 0.49681250000000005,
          "dom": 0.9350158389382681,
          "interaction": 0.9931177225958846,
          "aj": 0,
          "generated_at": "2026-05-14T13:16:10Z"
        }
      ]
    },
    {
      "id": "T-livebench",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://livebench.ai/",
      "sealed_at": "2026-04-29T05:03:28Z",
      "scored": true,
      "overall": 0.7990597546728094,
      "visual": 0.4588913323045268,
      "dom": 0.9923357664233576,
      "interaction": 0.9666021169385118,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6671645936213992,
          "visual": 0.4867196502057613,
          "dom": 1,
          "interaction": 0.9893757201646091,
          "aj": 0,
          "generated_at": "2026-05-14T10:28:34Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.6515674440538182,
          "visual": 0.4692469135802469,
          "dom": 0.9759625683703292,
          "interaction": 0.9737950354222195,
          "aj": 0,
          "generated_at": "2026-05-14T13:56:05Z"
        }
      ]
    },
    {
      "id": "T-lmarena",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://lmarena.ai/",
      "sealed_at": "2026-04-29T04:35:40Z",
      "scored": true,
      "overall": 0.8180889540747488,
      "visual": 0.492565329218107,
      "dom": 0.9885245901639345,
      "interaction": 0.9932872360520814,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5908028901291669,
          "visual": 0.4930448816872428,
          "dom": 0.6359382269713345,
          "interaction": 0.9701651204812167,
          "aj": 0,
          "generated_at": "2026-05-14T10:35:50Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5558254804099605,
          "visual": 0.4874992283950617,
          "dom": 0.4918224348039726,
          "interaction": 0.9561208783696477,
          "aj": 0,
          "generated_at": "2026-05-14T13:56:09Z"
        }
      ]
    },
    {
      "id": "T-mlperf",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://mlcommons.org/benchmarks/inference-datacenter/",
      "sealed_at": "2026-04-29T04:35:30Z",
      "scored": true,
      "overall": 0.7254255601395153,
      "visual": 0.45458989197530864,
      "dom": 0.9998253356247758,
      "interaction": 0.9636602692885653,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5898899748411928,
          "visual": 0.48315277777777776,
          "dom": 0.9966620049709651,
          "interaction": 0.9862996892202865,
          "aj": 0,
          "generated_at": "2026-05-14T10:39:31Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.41176756171614004,
          "visual": 0.42309670781893005,
          "dom": 0.5163582620678607,
          "interaction": 0.9062345837263348,
          "aj": 0,
          "generated_at": "2026-05-14T14:36:08Z"
        }
      ]
    },
    {
      "id": "T-mlperf-inference",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://mlcommons.org/benchmarks/inference-datacenter/",
      "sealed_at": "2026-04-29T05:04:16Z",
      "scored": true,
      "overall": 0.7979416831481504,
      "visual": 0.4542035751028806,
      "dom": 0.9998253356247758,
      "interaction": 0.963351215790623,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5523510445043315,
          "visual": 0.4481957304526749,
          "dom": 0.5814343797147875,
          "interaction": 0.9306522096764591,
          "aj": 0,
          "generated_at": "2026-05-14T10:46:47Z"
        }
      ]
    },
    {
      "id": "T-mlperf-training",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://mlcommons.org/benchmarks/training/",
      "sealed_at": "2026-04-29T08:29:27Z",
      "scored": false
    },
    {
      "id": "T-mmmu-benchmark",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://mmmu-benchmark.github.io/",
      "sealed_at": "2026-04-28T21:02:47Z",
      "scored": true,
      "overall": 0.5843607685827075,
      "visual": 0.44994881687242794,
      "dom": 0.5430387059503177,
      "interaction": 0.9294949672279637,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5584547627946681,
          "visual": 0.46160673868312757,
          "dom": 0.9313861227663052,
          "interaction": 0.9647111324642559,
          "aj": 0,
          "generated_at": "2026-05-14T10:54:05Z"
        }
      ]
    },
    {
      "id": "T-mteb-lb",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://huggingface.co/spaces/mteb/leaderboard",
      "sealed_at": "2026-04-29T04:34:41Z",
      "scored": true,
      "overall": 0.8238775887345678,
      "visual": 0.49809760802469133,
      "dom": 0.9999999999999999,
      "interaction": 0.9984780864197531,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6044255195061454,
          "visual": 0.49792862654320985,
          "dom": 0.6847619538438705,
          "interaction": 0.9773270314908261,
          "aj": 0,
          "generated_at": "2026-05-14T11:01:21Z"
        }
      ]
    },
    {
      "id": "T-open-llm-lb",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard",
      "sealed_at": "2026-04-29T04:33:29Z",
      "scored": true,
      "overall": 0.8225553240740741,
      "visual": 0.4958564814814815,
      "dom": 1,
      "interaction": 0.9966851851851852,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5794956473638166,
          "visual": 0.49511033950617284,
          "dom": 0.5790024866144303,
          "interaction": 0.968021770712567,
          "aj": 0,
          "generated_at": "2026-05-14T11:05:01Z"
        }
      ]
    },
    {
      "id": "T-papers-with-code-arxiv",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://arxiv.org/list/cs.LG/recent",
      "sealed_at": "2026-04-29T06:26:21Z",
      "scored": false
    },
    {
      "id": "T-spider-text2sql",
      "category": "ml-benchmark",
      "tier": "T1",
      "source_url": "https://yale-lily.github.io/spider",
      "sealed_at": "2026-04-29T05:23:42Z",
      "scored": false
    },
    {
      "id": "T-squad-explorer",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://rajpurkar.github.io/SQuAD-explorer/",
      "sealed_at": "2026-04-29T05:04:31Z",
      "scored": true,
      "overall": 0.8112678562242798,
      "visual": 0.4767251800411523,
      "dom": 1,
      "interaction": 0.9813801440329218,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.675,
          "visual": 0.5,
          "dom": 1,
          "interaction": 1,
          "aj": 0,
          "generated_at": "2026-05-14T11:08:42Z"
        }
      ]
    },
    {
      "id": "T-swe-bench",
      "category": "ml-benchmark",
      "tier": "T2",
      "source_url": "https://www.swebench.com/",
      "sealed_at": "2026-04-29T01:11:20Z",
      "scored": true,
      "overall": 0.790388090137705,
      "visual": 0.4484985853909465,
      "dom": 0.9807905670774844,
      "interaction": 0.9575182394512561,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.6747384652386234,
          "visual": 0.49960738168724284,
          "dom": 0.9998641365597731,
          "interaction": 0.9996768477871125,
          "aj": 0,
          "generated_at": "2026-05-14T11:12:23Z"
        }
      ]
    },
    {
      "id": "T-acropolis-museum",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.theacropolismuseum.gr/en",
      "sealed_at": "2026-04-29T07:47:32Z",
      "scored": false
    },
    {
      "id": "T-archaeology-org",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.archaeology.org/",
      "sealed_at": "2026-04-29T05:43:28Z",
      "scored": false
    },
    {
      "id": "T-artic-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.artic.edu/",
      "sealed_at": "2026-04-29T08:50:22Z",
      "scored": false
    },
    {
      "id": "T-artsy-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.artsy.net/",
      "sealed_at": "2026-04-29T08:49:41Z",
      "scored": false
    },
    {
      "id": "T-british-museum",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.britishmuseum.org/",
      "sealed_at": "2026-04-29T04:57:14Z",
      "scored": true,
      "overall": 0.4763628232289796,
      "visual": 0.0974022633744856,
      "dom": 0.8146118714508485,
      "interaction": 0.6655626021296451,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.3394727366439487,
          "visual": 0.3805343364197531,
          "dom": 0.35379404937565506,
          "interaction": 0.8613470724275128,
          "aj": 0,
          "generated_at": "2026-05-14T11:19:41Z"
        }
      ]
    },
    {
      "id": "T-britishmuseum-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.britishmuseum.org/",
      "sealed_at": "2026-04-29T05:24:34Z",
      "scored": false
    },
    {
      "id": "T-cooperhewitt-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.cooperhewitt.org/",
      "sealed_at": "2026-04-29T05:25:09Z",
      "scored": false
    },
    {
      "id": "T-fondationbeyeler",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.fondationbeyeler.ch/en/home",
      "sealed_at": "2026-04-29T08:50:07Z",
      "scored": false
    },
    {
      "id": "T-getty-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.getty.edu/",
      "sealed_at": "2026-04-29T05:24:48Z",
      "scored": false
    },
    {
      "id": "T-google-arts",
      "category": "museum-culture",
      "tier": "T2",
      "source_url": "https://artsandculture.google.com/",
      "sealed_at": "2026-04-29T04:57:49Z",
      "scored": true,
      "overall": 0.8136851430797994,
      "visual": 0.48498559670781893,
      "dom": 0.9888347319190284,
      "interaction": 0.987244126160857,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.4960823372568347,
          "visual": 0.44689994855967075,
          "dom": 0.32914258003013147,
          "interaction": 0.9127961308497454,
          "aj": 0,
          "generated_at": "2026-05-14T11:26:58Z"
        }
      ]
    },
    {
      "id": "T-guggenheim",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.guggenheim.org/",
      "sealed_at": "2026-04-29T01:24:37Z",
      "scored": true,
      "overall": 0.5899565094649154,
      "visual": 0.2042015174897119,
      "dom": 0.9989704066608934,
      "interaction": 0.7632925744358291,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.3404587342703628,
          "visual": 0.3069125514403292,
          "dom": 0.48810754887665386,
          "interaction": 0.8114038777440404,
          "aj": 0,
          "generated_at": "2026-05-14T11:34:15Z"
        }
      ]
    },
    {
      "id": "T-ima-pieta",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.museivaticani.va/content/museivaticani/en.html",
      "sealed_at": "2026-04-29T06:47:30Z",
      "scored": false
    },
    {
      "id": "T-jp-museums",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.tnm.jp/?lang=en",
      "sealed_at": "2026-04-29T07:47:25Z",
      "scored": false
    },
    {
      "id": "T-louvre",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.louvre.fr/en",
      "sealed_at": "2026-04-29T04:57:32Z",
      "scored": true,
      "overall": 0.7390706944444444,
      "visual": 0.47976054526748974,
      "dom": 1,
      "interaction": 0.9838084362139918,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.5285855359719784,
          "visual": 0.40726967592592594,
          "dom": 0.9296480581493793,
          "interaction": 0.9211256112840327,
          "aj": 0,
          "generated_at": "2026-05-14T11:37:56Z"
        }
      ]
    },
    {
      "id": "T-louvre-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.louvre.fr/en",
      "sealed_at": "2026-04-29T05:23:40Z",
      "scored": false
    },
    {
      "id": "T-mauritshuis",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.mauritshuis.nl/en/",
      "sealed_at": "2026-04-29T07:47:40Z",
      "scored": false
    },
    {
      "id": "T-met-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.metmuseum.org/",
      "sealed_at": "2026-04-28T21:19:26Z",
      "scored": true,
      "overall": 0.6572557156061336,
      "visual": 0.3293016975308642,
      "dom": 0.998130106393847,
      "interaction": 0.8633166984509478,
      "aj": 1,
      "agent_runs": [
        {
          "model": "gpt-5.5",
          "overall": 0.3591362008975628,
          "visual": 0.418357896090535,
          "dom": 0.351284407720903,
          "interaction": 0.8914386107204884,
          "aj": 0,
          "generated_at": "2026-05-14T11:45:15Z"
        }
      ]
    },
    {
      "id": "T-mfa-boston",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.mfa.org/",
      "sealed_at": "2026-04-29T08:48:42Z",
      "scored": false
    },
    {
      "id": "T-moma-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.moma.org/",
      "sealed_at": "2026-04-29T01:27:28Z",
      "scored": true,
      "overall": 0.6937552685946227,
      "visual": 0.39585108024691357,
      "dom": 0.9999857755866681,
      "interaction": 0.9166799159033087,
      "aj": 1
    },
    {
      "id": "T-musee-orsay",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.musee-orsay.fr/en",
      "sealed_at": "2026-04-29T06:48:07Z",
      "scored": false
    },
    {
      "id": "T-museumofjurassictech",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://mjt.org/",
      "sealed_at": "2026-04-29T06:47:01Z",
      "scored": false
    },
    {
      "id": "T-natgeo-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.nationalgeographic.com/",
      "sealed_at": "2026-04-29T06:46:50Z",
      "scored": false
    },
    {
      "id": "T-national-gallery",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.nationalgallery.org.uk/",
      "sealed_at": "2026-04-29T04:56:21Z",
      "scored": true,
      "overall": 0.6580221719280183,
      "visual": 0.3296764403292181,
      "dom": 0.9999897609348588,
      "interaction": 0.8637404696590317,
      "aj": 1
    },
    {
      "id": "T-nationalgallery-uk",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.nationalgallery.org.uk/",
      "sealed_at": "2026-04-29T05:41:58Z",
      "scored": false
    },
    {
      "id": "T-prado",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.museodelprado.es/en",
      "sealed_at": "2026-04-29T05:42:57Z",
      "scored": false
    },
    {
      "id": "T-rijksmuseum-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.rijksmuseum.nl/en",
      "sealed_at": "2026-04-29T05:25:02Z",
      "scored": false
    },
    {
      "id": "T-saatchi-art",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.saatchiart.com/",
      "sealed_at": "2026-04-29T08:49:33Z",
      "scored": false
    },
    {
      "id": "T-smithsonianmag",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.smithsonianmag.com/",
      "sealed_at": "2026-04-29T01:24:21Z",
      "scored": true,
      "overall": 0.6922179048059615,
      "visual": 0.4003024691358024,
      "dom": 0.9869930927669059,
      "interaction": 0.9193748481597691,
      "aj": 1
    },
    {
      "id": "T-tate-collection",
      "category": "museum-culture",
      "tier": "T2",
      "source_url": "https://www.tate.org.uk/art",
      "sealed_at": "2026-04-29T07:48:01Z",
      "scored": false
    },
    {
      "id": "T-tate-home",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.tate.org.uk/",
      "sealed_at": "2026-04-29T01:23:36Z",
      "scored": true,
      "overall": 0.7098980179798475,
      "visual": 0.4275263631687243,
      "dom": 0.9968146655013288,
      "interaction": 0.9418087349017347,
      "aj": 1
    },
    {
      "id": "T-thejnf",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.frick.org/",
      "sealed_at": "2026-04-29T06:48:02Z",
      "scored": false
    },
    {
      "id": "T-uffizi",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.uffizi.it/en",
      "sealed_at": "2026-04-29T05:42:21Z",
      "scored": false
    },
    {
      "id": "T-uffizi-collection",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.uffizi.it/en/the-uffizi",
      "sealed_at": "2026-04-29T07:47:49Z",
      "scored": false
    },
    {
      "id": "T-vam",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.vam.ac.uk/",
      "sealed_at": "2026-04-29T05:42:10Z",
      "scored": false
    },
    {
      "id": "T-vangoghmuseum",
      "category": "museum-culture",
      "tier": "T1",
      "source_url": "https://www.vangoghmuseum.nl/en",
      "sealed_at": "2026-04-29T04:57:42Z",
      "scored": true,
      "overall": 0.7031503092354493,
      "visual": 0.4132427983539095,
      "dom": 0.999997356453862,
      "interaction": 0.9305940624467185,
      "aj": 1
    },
    {
      "id": "T-allmusic-jazz",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.allmusic.com/genre/jazz-ma0000002674",
      "sealed_at": "2026-04-29T06:36:46Z",
      "scored": false
    },
    {
      "id": "T-bandcamp-album-radiohead",
      "category": "music",
      "tier": "T1",
      "source_url": "https://radiohead.bandcamp.com/album/in-rainbows",
      "sealed_at": "2026-04-29T08:38:22Z",
      "scored": false
    },
    {
      "id": "T-bandcamp-discover",
      "category": "music",
      "tier": "T2",
      "source_url": "https://bandcamp.com/discover",
      "sealed_at": "2026-04-28T21:17:41Z",
      "scored": true,
      "overall": 0.8070816332304528,
      "visual": 0.46962988683127577,
      "dom": 1,
      "interaction": 0.9757039094650205,
      "aj": 1
    },
    {
      "id": "T-bandcamp-fkatwigs",
      "category": "music",
      "tier": "T1",
      "source_url": "https://fkatwigs.bandcamp.com/",
      "sealed_at": "2026-04-29T04:45:37Z",
      "scored": true,
      "overall": 0.5062326017579167,
      "visual": 0.2705502829218107,
      "dom": 0.6048421394949635,
      "interaction": 0.790096368970446,
      "aj": 1
    },
    {
      "id": "T-bandcamp-genre-jazz",
      "category": "music",
      "tier": "T2",
      "source_url": "https://bandcamp.com/tag/jazz?sort_field=pop",
      "sealed_at": "2026-04-29T07:38:51Z",
      "scored": false
    },
    {
      "id": "T-bandcamp-tag-classical",
      "category": "music",
      "tier": "T2",
      "source_url": "https://bandcamp.com/tag/classical",
      "sealed_at": "2026-04-29T07:40:11Z",
      "scored": false
    },
    {
      "id": "T-bandcamp-tag-electronic",
      "category": "music",
      "tier": "T2",
      "source_url": "https://bandcamp.com/tag/electronic",
      "sealed_at": "2026-04-29T05:35:21Z",
      "scored": false
    },
    {
      "id": "T-bandcamp-tag-jazz",
      "category": "music",
      "tier": "T2",
      "source_url": "https://bandcamp.com/tag/jazz",
      "sealed_at": "2026-04-29T05:15:40Z",
      "scored": false
    },
    {
      "id": "T-billboard-200",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.billboard.com/charts/billboard-200/",
      "sealed_at": "2026-04-29T06:38:01Z",
      "scored": false
    },
    {
      "id": "T-billboard-hot100",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.billboard.com/charts/hot-100/",
      "sealed_at": "2026-04-29T05:14:26Z",
      "scored": true,
      "overall": 0.6536664980103943,
      "visual": 0.3305181327160494,
      "dom": 0.9841319989353658,
      "interaction": 0.8633566394351972,
      "aj": 1
    },
    {
      "id": "T-discogs-beatles",
      "category": "music",
      "tier": "T2",
      "source_url": "https://www.discogs.com/artist/82730-The-Beatles",
      "sealed_at": "2026-04-29T08:39:17Z",
      "scored": false
    },
    {
      "id": "T-discogs-explore",
      "category": "music",
      "tier": "T2",
      "source_url": "https://www.discogs.com/explore",
      "sealed_at": "2026-04-29T07:40:02Z",
      "scored": false
    },
    {
      "id": "T-discogs-home",
      "category": "music",
      "tier": "T2",
      "source_url": "https://www.discogs.com/",
      "sealed_at": "2026-04-29T01:18:11Z",
      "scored": true,
      "overall": 0.602609372266593,
      "visual": 0.2258031121399177,
      "dom": 0.7244797095638249,
      "interaction": 0.7622744703495226,
      "aj": 1
    },
    {
      "id": "T-discogs-pinkfloyd",
      "category": "music",
      "tier": "T2",
      "source_url": "https://www.discogs.com/artist/45467-Pink-Floyd",
      "sealed_at": "2026-04-29T05:33:27Z",
      "scored": false
    },
    {
      "id": "T-grammys-best-album",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.grammy.com/awards/66th-annual-grammy-awards-2024",
      "sealed_at": "2026-04-29T06:36:53Z",
      "scored": false
    },
    {
      "id": "T-grammys-home",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.grammy.com/",
      "sealed_at": "2026-04-29T05:15:31Z",
      "scored": false
    },
    {
      "id": "T-jazz24",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.jazz24.org/",
      "sealed_at": "2026-04-29T06:39:03Z",
      "scored": false
    },
    {
      "id": "T-lastfm-charts",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.last.fm/charts",
      "sealed_at": "2026-04-29T04:45:27Z",
      "scored": true,
      "overall": 0.7450981399689914,
      "visual": 0.4910985082304527,
      "dom": 0.9996866335974076,
      "interaction": 0.992857915490856,
      "aj": 1
    },
    {
      "id": "T-lastfm-radiohead",
      "category": "music",
      "tier": "T2",
      "source_url": "https://www.last.fm/music/Radiohead",
      "sealed_at": "2026-04-29T01:17:15Z",
      "scored": true,
      "overall": 0.8015247301864258,
      "visual": 0.46022273662551444,
      "dom": 0.9999696162607832,
      "interaction": 0.9681761637177969,
      "aj": 1
    },
    {
      "id": "T-npr-music",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.npr.org/sections/music/",
      "sealed_at": "2026-04-29T04:46:26Z",
      "scored": true,
      "overall": 0.7498519511295936,
      "visual": 0.49973212448559673,
      "dom": 0.9999888040902352,
      "interaction": 0.9997849531944931,
      "aj": 1
    },
    {
      "id": "T-pitchfork-album-review",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/reviews/albums/",
      "sealed_at": "2026-04-29T06:38:54Z",
      "scored": false
    },
    {
      "id": "T-pitchfork-best-2024",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/features/lists-and-guides/best-albums-2024/",
      "sealed_at": "2026-04-29T08:41:24Z",
      "scored": false
    },
    {
      "id": "T-pitchfork-best-decade",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/features/lists-and-guides/best-albums-of-the-decade-so-far/",
      "sealed_at": "2026-04-29T07:39:40Z",
      "scored": false
    },
    {
      "id": "T-pitchfork-best-new",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/best/",
      "sealed_at": "2026-04-29T05:35:05Z",
      "scored": false
    },
    {
      "id": "T-pitchfork-home",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/",
      "sealed_at": "2026-04-29T01:29:37Z",
      "scored": true,
      "overall": 0.6297449176183549,
      "visual": 0.29653215020576135,
      "dom": 0.9657721643095949,
      "interaction": 0.8349438644519154,
      "aj": 1,
      "agent_runs": [
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.40039222572407795,
          "visual": 0.35820679012345674,
          "dom": 0.09986997491454269,
          "interaction": 0.826556763759735,
          "aj": 1,
          "generated_at": "2026-05-14T05:19:28Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.31537335869018657,
          "visual": 0.24277957818930043,
          "dom": 0.025073801542739426,
          "interaction": 0.7292285826542897,
          "aj": 1,
          "generated_at": "2026-05-14T05:21:34Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.30515948397437764,
          "visual": 0.19573135288065843,
          "dom": 0.0751578684136991,
          "interaction": 0.6949289401987734,
          "aj": 1,
          "generated_at": "2026-05-14T05:17:35Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2509257071654854,
          "visual": 0.3247899948559671,
          "dom": 0.16111794486790051,
          "interaction": 0.8039065255426338,
          "aj": 0,
          "generated_at": "2026-05-14T10:16:15Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.18117713760153292,
          "visual": 0.2377767489711934,
          "dom": 0.08608030711128066,
          "interaction": 0.7292934196510402,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:31Z"
        }
      ]
    },
    {
      "id": "T-pitchfork-reviews",
      "category": "music",
      "tier": "T1",
      "source_url": "https://pitchfork.com/reviews/albums/",
      "sealed_at": "2026-04-29T04:48:09Z",
      "scored": true,
      "overall": 0.6584503088390881,
      "visual": 0.33596965020576136,
      "dom": 0.990197904597726,
      "interaction": 0.8681222471377907,
      "aj": 1
    },
    {
      "id": "T-rateyourmusic-best",
      "category": "music",
      "tier": "T1",
      "source_url": "https://rateyourmusic.com/charts/top/album/all-time/",
      "sealed_at": "2026-04-29T05:15:14Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-rateyourmusic-charts",
      "category": "music",
      "tier": "T1",
      "source_url": "https://rateyourmusic.com/charts/",
      "sealed_at": "2026-04-29T08:40:52Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-rateyourmusic-radiohead",
      "category": "music",
      "tier": "T1",
      "source_url": "https://rateyourmusic.com/artist/radiohead",
      "sealed_at": "2026-04-29T05:34:14Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-songkick",
      "category": "music",
      "tier": "T1",
      "source_url": "https://www.songkick.com/",
      "sealed_at": "2026-04-29T04:47:17Z",
      "scored": true,
      "overall": 0.7133872222222222,
      "visual": 0.43219855967078186,
      "dom": 0.9999999999999999,
      "interaction": 0.9457588477366254,
      "aj": 1
    },
    {
      "id": "T-soundcloud-home",
      "category": "music",
      "tier": "T1",
      "source_url": "https://soundcloud.com/",
      "sealed_at": "2026-04-29T01:18:24Z",
      "scored": true,
      "overall": 0.7268086111111112,
      "visual": 0.4570529835390946,
      "dom": 1,
      "interaction": 0.9656423868312758,
      "aj": 1
    },
    {
      "id": "T-soundcloud-trending",
      "category": "music",
      "tier": "T2",
      "source_url": "https://soundcloud.com/discover/sets/charts-top:all-music",
      "sealed_at": "2026-04-29T07:38:41Z",
      "scored": false
    },
    {
      "id": "T-1843magazine",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.economist.com/1843/",
      "sealed_at": "2026-04-29T08:47:51Z",
      "scored": false
    },
    {
      "id": "T-aeon-essays",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://aeon.co/",
      "sealed_at": "2026-04-29T07:46:48Z",
      "scored": false
    },
    {
      "id": "T-bbc-future",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.bbc.com/future",
      "sealed_at": "2026-04-29T04:54:23Z",
      "scored": true,
      "overall": 0.6035028987997099,
      "visual": 0.32369354423868313,
      "dom": 0.8309067634422669,
      "interaction": 0.8476819529537644,
      "aj": 1
    },
    {
      "id": "T-bloomberg-graphics",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.bloomberg.com/graphics/",
      "sealed_at": "2026-04-29T05:40:36Z",
      "scored": false
    },
    {
      "id": "T-economist-charts",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.economist.com/graphic-detail/2024",
      "sealed_at": "2026-04-29T08:47:08Z",
      "scored": false
    },
    {
      "id": "T-economist-data",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.economist.com/graphic-detail",
      "sealed_at": "2026-04-29T05:23:22Z",
      "scored": false
    },
    {
      "id": "T-fivethirtyeight",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://fivethirtyeight.com/",
      "sealed_at": "2026-04-29T04:55:25Z",
      "scored": true,
      "overall": 0.7437217102774604,
      "visual": 0.48837692901234564,
      "dom": 0.9999939624531664,
      "interaction": 0.9907011407067543,
      "aj": 1
    },
    {
      "id": "T-fivethirtyeight-elections",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://projects.fivethirtyeight.com/",
      "sealed_at": "2026-04-29T07:47:08Z",
      "scored": false
    },
    {
      "id": "T-fivethirtyeight-features",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://fivethirtyeight.com/features/",
      "sealed_at": "2026-04-29T07:45:42Z",
      "scored": false
    },
    {
      "id": "T-fivethirtyeight-home",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://fivethirtyeight.com/",
      "sealed_at": "2026-04-29T05:21:59Z",
      "scored": false
    },
    {
      "id": "T-flowingdata",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://flowingdata.com/",
      "sealed_at": "2026-04-29T05:40:58Z",
      "scored": false
    },
    {
      "id": "T-flowingdata-projects",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://flowingdata.com/projects/",
      "sealed_at": "2026-04-29T08:46:29Z",
      "scored": false
    },
    {
      "id": "T-ft-data",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://ig.ft.com/",
      "sealed_at": "2026-04-29T05:40:52Z",
      "scored": false
    },
    {
      "id": "T-ftvisual-data",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.ft.com/visual-and-data-journalism",
      "sealed_at": "2026-04-29T06:46:12Z",
      "scored": false
    },
    {
      "id": "T-guardian-data",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.theguardian.com/data",
      "sealed_at": "2026-04-29T01:22:29Z",
      "scored": true,
      "overall": 0.4810845713429257,
      "visual": 0.01756712962962964,
      "dom": 0.9723021582733813,
      "interaction": 0.6122071809219292,
      "aj": 1
    },
    {
      "id": "T-information-is-beautiful",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://informationisbeautiful.net/",
      "sealed_at": "2026-04-29T05:41:12Z",
      "scored": false
    },
    {
      "id": "T-npr-home",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.npr.org/",
      "sealed_at": "2026-04-29T04:55:12Z",
      "scored": true,
      "overall": 0.7408759571439231,
      "visual": 0.4848872170781893,
      "dom": 0.9968248129286843,
      "interaction": 0.9876980945244638,
      "aj": 1
    },
    {
      "id": "T-nyt-archive-2020",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.nytimes.com/issue/todayspaper/2020/01/01/todays-new-york-times",
      "sealed_at": "2026-04-29T04:53:33Z",
      "scored": true,
      "overall": 0.7376193439533184,
      "visual": 0.478624871399177,
      "dom": 0.9972370771354819,
      "interaction": 0.9827157022617071,
      "aj": 1
    },
    {
      "id": "T-nytimes-archive-100yrs",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.nytimes.com/topic/destination/best-of-the-times",
      "sealed_at": "2026-04-29T08:46:39Z",
      "scored": false
    },
    {
      "id": "T-nytimes-graphics-2020",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.nytimes.com/interactive/2020/upshot/coronavirus-school-reopening.html",
      "sealed_at": "2026-04-29T05:40:28Z",
      "scored": false
    },
    {
      "id": "T-our-world-in-data",
      "category": "news-data",
      "tier": "T2",
      "source_url": "https://ourworldindata.org/",
      "sealed_at": "2026-04-29T04:55:31Z",
      "scored": true,
      "overall": 0.8075091859567901,
      "visual": 0.4703545524691359,
      "dom": 1,
      "interaction": 0.9762836419753086,
      "aj": 1
    },
    {
      "id": "T-pew-research",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.pewresearch.org/",
      "sealed_at": "2026-04-29T06:46:04Z",
      "scored": false
    },
    {
      "id": "T-propublica-home",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.propublica.org/",
      "sealed_at": "2026-04-29T01:22:20Z",
      "scored": true,
      "overall": 0.699846672377414,
      "visual": 0.407125,
      "dom": 0.9999972715738925,
      "interaction": 0.9256998181049262,
      "aj": 1
    },
    {
      "id": "T-pudding-fashion",
      "category": "news-data",
      "tier": "T2",
      "source_url": "https://pudding.cool/2024/02/under-pressure/",
      "sealed_at": "2026-04-29T06:46:16Z",
      "scored": false
    },
    {
      "id": "T-pudding-home",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://pudding.cool",
      "sealed_at": "2026-04-28T21:19:15Z",
      "scored": true,
      "overall": 0.6734930693193193,
      "visual": 0.3713094135802469,
      "dom": 0.9768768768768767,
      "interaction": 0.8955059893226561,
      "aj": 1
    },
    {
      "id": "T-pudding-language",
      "category": "news-data",
      "tier": "T2",
      "source_url": "https://pudding.cool/2017/05/song-repetition/",
      "sealed_at": "2026-04-29T08:47:16Z",
      "scored": false
    },
    {
      "id": "T-pudding-pop-vs-soda",
      "category": "news-data",
      "tier": "T2",
      "source_url": "https://pudding.cool/2024/12/pop-vs-soda/",
      "sealed_at": "2026-04-29T05:23:26Z",
      "scored": false
    },
    {
      "id": "T-quanta-magazine",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.quantamagazine.org/",
      "sealed_at": "2026-04-29T07:46:42Z",
      "scored": false
    },
    {
      "id": "T-reuters-graphics",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.reuters.com/graphics/",
      "sealed_at": "2026-04-29T05:22:55Z",
      "scored": false
    },
    {
      "id": "T-statista-home",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.statista.com/",
      "sealed_at": "2026-04-29T06:45:54Z",
      "scored": false
    },
    {
      "id": "T-vox-explainer",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://www.vox.com/",
      "sealed_at": "2026-04-29T07:46:33Z",
      "scored": false
    },
    {
      "id": "T-wikipedia-covid",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://en.wikipedia.org/wiki/COVID-19_pandemic",
      "sealed_at": "2026-04-29T01:22:39Z",
      "scored": true,
      "overall": 0.6921662389008484,
      "visual": 0.41836355452674895,
      "dom": 0.9546700641419911,
      "interaction": 0.9316688478975319,
      "aj": 1
    },
    {
      "id": "T-xkcd-353",
      "category": "news-data",
      "tier": "T1",
      "source_url": "https://xkcd.com/353/",
      "sealed_at": "2026-04-28T21:05:29Z",
      "scored": true,
      "overall": 0.7196642010703364,
      "visual": 0.4582523148148148,
      "dom": 0.9743119266055045,
      "interaction": 0.9648893136255522,
      "aj": 1,
      "agent_runs": [
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.6043283641253532,
          "visual": 0.23026658950617285,
          "dom": 0.9999485905231423,
          "interaction": 0.7842098443064812,
          "aj": 1,
          "generated_at": "2026-05-14T05:19:42Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.6032311257053383,
          "visual": 0.22884246399176955,
          "dom": 0.9988665774168661,
          "interaction": 0.7829984096878734,
          "aj": 1,
          "generated_at": "2026-05-14T05:21:47Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.5993678518896292,
          "visual": 0.4993481224279835,
          "dom": 0.9990764805885208,
          "interaction": 0.9994169299816216,
          "aj": 0,
          "generated_at": "2026-05-14T10:16:19Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.5304459252127776,
          "visual": 0.4032134773662552,
          "dom": 0.9430021344010981,
          "interaction": 0.9187709241864107,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:35Z"
        }
      ]
    },
    {
      "id": "T-base64-decoder",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.base64decode.org/",
      "sealed_at": "2026-04-29T05:20:56Z",
      "scored": false
    },
    {
      "id": "T-bcrypt-generator",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://bcrypt-generator.com/",
      "sealed_at": "2026-04-29T06:45:41Z",
      "scored": false
    },
    {
      "id": "T-caniuse",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://caniuse.com/",
      "sealed_at": "2026-04-29T01:20:23Z",
      "scored": true,
      "overall": 0.8010430832541944,
      "visual": 0.4622633744855967,
      "dom": 0.9923076923076923,
      "interaction": 0.9692978790756568,
      "aj": 1
    },
    {
      "id": "T-carbon-now-sh",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://carbon.now.sh/",
      "sealed_at": "2026-04-29T04:51:57Z",
      "scored": true,
      "overall": 0.8569880555555557,
      "visual": 0.47047222222222224,
      "dom": 1,
      "interaction": 0.9763777777777779,
      "aj": 1
    },
    {
      "id": "T-codepen-home",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://codepen.io/",
      "sealed_at": "2026-04-29T05:19:54Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-codesandbox-home",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://codesandbox.io/",
      "sealed_at": "2026-04-29T05:39:36Z",
      "scored": false
    },
    {
      "id": "T-coolors",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://coolors.co/",
      "sealed_at": "2026-04-29T04:51:40Z",
      "scored": true,
      "overall": 0.8594379398148148,
      "visual": 0.47448842592592594,
      "dom": 0.9999999999999999,
      "interaction": 0.9795907407407407,
      "aj": 1
    },
    {
      "id": "T-excalidraw",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://excalidraw.com/",
      "sealed_at": "2026-04-29T04:51:51Z",
      "scored": true,
      "overall": 0.8731954166666667,
      "visual": 0.49704166666666666,
      "dom": 0.9999999999999999,
      "interaction": 0.9976333333333334,
      "aj": 1
    },
    {
      "id": "T-html-validator",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://validator.w3.org/",
      "sealed_at": "2026-04-29T07:44:53Z",
      "scored": false
    },
    {
      "id": "T-htmlcolorcodes",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://htmlcolorcodes.com/",
      "sealed_at": "2026-04-29T01:22:08Z",
      "scored": true,
      "overall": 0.7951763501756067,
      "visual": 0.4502696759259259,
      "dom": 0.9978056426332287,
      "interaction": 0.9600694502496226,
      "aj": 1
    },
    {
      "id": "T-imageoptim-home",
      "category": "online-tool",
      "tier": "T1",
      "source_url": "https://imageoptim.com/",
      "sealed_at": "2026-04-29T07:44:38Z",
      "scored": false
    },
    {
      "id": "T-jsbin-home",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://jsbin.com/",
      "sealed_at": "2026-04-29T05:19:58Z",
      "scored": false
    },
    {
      "id": "T-jsfiddle-home",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://jsfiddle.net/",
      "sealed_at": "2026-04-29T05:39:19Z",
      "scored": false
    },
    {
      "id": "T-jsonformatter",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://jsonformatter.org/",
      "sealed_at": "2026-04-29T01:21:16Z",
      "scored": true,
      "overall": 0.7897786760872986,
      "visual": 0.3617449845679013,
      "dom": 0.995079086115993,
      "interaction": 0.8890679267287206,
      "aj": 1
    },
    {
      "id": "T-loremipsum-generator",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.lipsum.com/",
      "sealed_at": "2026-04-29T05:21:47Z",
      "scored": false
    },
    {
      "id": "T-photopea",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://www.photopea.com/",
      "sealed_at": "2026-04-29T08:45:59Z",
      "scored": false
    },
    {
      "id": "T-pixilart",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://www.pixilart.com/",
      "sealed_at": "2026-04-29T08:46:25Z",
      "scored": false
    },
    {
      "id": "T-png2svg",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://png2svg.com/",
      "sealed_at": "2026-04-29T07:45:30Z",
      "scored": false
    },
    {
      "id": "T-prettier-playground",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://prettier.io/playground/",
      "sealed_at": "2026-04-29T06:44:46Z",
      "scored": false
    },
    {
      "id": "T-regex101",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://regex101.com/",
      "sealed_at": "2026-04-28T21:19:34Z",
      "scored": true,
      "overall": 0.8691127272105329,
      "visual": 0.49038117283950616,
      "dom": 0.9998900654357452,
      "interaction": 0.9922976093006546,
      "aj": 1,
      "agent_runs": [
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.6727468476222013,
          "visual": 0.4177798353909465,
          "dom": 0.15500637796513256,
          "interaction": 0.8778909601770994,
          "aj": 1,
          "generated_at": "2026-05-14T05:19:56Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.6703507123580893,
          "visual": 0.43711201131687244,
          "dom": 0.07617991919331707,
          "interaction": 0.8881016036663859,
          "aj": 1,
          "generated_at": "2026-05-14T05:17:47Z"
        },
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.662517403039433,
          "visual": 0.43672042181069964,
          "dom": 0.033988587416145466,
          "interaction": 0.884975576609636,
          "aj": 1,
          "generated_at": "2026-05-14T05:22:01Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.636054239932642,
          "visual": 0.4638707561728395,
          "dom": 0.6282948814844989,
          "interaction": 0.946316263703905,
          "aj": 0,
          "generated_at": "2026-05-14T10:34:11Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.5618711994812979,
          "visual": 0.45798804012345684,
          "dom": 0.23610275003327316,
          "interaction": 0.9154639487676505,
          "aj": 0,
          "generated_at": "2026-05-14T11:36:39Z"
        }
      ]
    },
    {
      "id": "T-regexr",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://regexr.com/",
      "sealed_at": "2026-04-29T01:20:14Z",
      "scored": true,
      "overall": 0.6928753621306961,
      "visual": 0.2449224537037037,
      "dom": 0.8526259187302041,
      "interaction": 0.7861130242116433,
      "aj": 1
    },
    {
      "id": "T-shadertoy",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://www.shadertoy.com/",
      "sealed_at": "2026-04-29T08:45:18Z",
      "scored": false,
      "bot_challenge": true
    },
    {
      "id": "T-speedtest",
      "category": "online-tool",
      "tier": "T1",
      "source_url": "https://www.speedtest.net/",
      "sealed_at": "2026-04-29T04:52:47Z",
      "scored": true,
      "overall": 0.748076111111111,
      "visual": 0.4964372427983539,
      "dom": 1,
      "interaction": 0.9971497942386831,
      "aj": 1
    },
    {
      "id": "T-stackblitz-home",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://stackblitz.com/",
      "sealed_at": "2026-04-29T05:39:25Z",
      "scored": false
    },
    {
      "id": "T-svg-viewer",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.svgviewer.dev/",
      "sealed_at": "2026-04-29T07:44:28Z",
      "scored": false
    },
    {
      "id": "T-svg2png-home",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://svgtopng.com/",
      "sealed_at": "2026-04-29T05:40:16Z",
      "scored": false
    },
    {
      "id": "T-tabularasa-textdiff",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.diffchecker.com/",
      "sealed_at": "2026-04-29T06:44:41Z",
      "scored": false
    },
    {
      "id": "T-tiny-png",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://tinypng.com/",
      "sealed_at": "2026-04-29T04:52:59Z",
      "scored": true,
      "overall": 0.8094816949588478,
      "visual": 0.4736977880658437,
      "dom": 1,
      "interaction": 0.9789582304526748,
      "aj": 1
    },
    {
      "id": "T-tinypng-home",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://tinypng.com/",
      "sealed_at": "2026-04-29T05:20:25Z",
      "scored": false
    },
    {
      "id": "T-tinyurl",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://tinyurl.com/",
      "sealed_at": "2026-04-29T06:45:36Z",
      "scored": false
    },
    {
      "id": "T-tldraw",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://www.tldraw.com/",
      "sealed_at": "2026-04-29T08:45:52Z",
      "scored": false
    },
    {
      "id": "T-typescriptlang-playground",
      "category": "online-tool",
      "tier": "T3",
      "source_url": "https://www.typescriptlang.org/play",
      "sealed_at": "2026-04-29T05:39:45Z",
      "scored": false
    },
    {
      "id": "T-uuid-generator",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.uuidgenerator.net/",
      "sealed_at": "2026-04-29T06:45:22Z",
      "scored": false
    },
    {
      "id": "T-yaml-validator",
      "category": "online-tool",
      "tier": "T2",
      "source_url": "https://www.yamllint.com/",
      "sealed_at": "2026-04-29T07:44:48Z",
      "scored": false
    },
    {
      "id": "T-airtable-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.airtable.com/",
      "sealed_at": "2026-04-29T05:06:44Z",
      "scored": true,
      "overall": 0.7397646761858868,
      "visual": 0.4819436728395062,
      "dom": 0.9984014050084182,
      "interaction": 0.9854483652721662,
      "aj": 1
    },
    {
      "id": "T-amplitude-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://amplitude.com/",
      "sealed_at": "2026-04-29T08:32:18Z",
      "scored": false
    },
    {
      "id": "T-anthropic-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.anthropic.com/",
      "sealed_at": "2026-04-29T05:05:43Z",
      "scored": true,
      "overall": 0.7331663194444444,
      "visual": 0.46882651748971194,
      "dom": 0.9999999999999999,
      "interaction": 0.9750612139917697,
      "aj": 1
    },
    {
      "id": "T-asana-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://asana.com/",
      "sealed_at": "2026-04-29T05:25:03Z",
      "scored": false
    },
    {
      "id": "T-aws-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://aws.amazon.com/",
      "sealed_at": "2026-04-29T06:27:36Z",
      "scored": false
    },
    {
      "id": "T-clerk-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://clerk.com/",
      "sealed_at": "2026-04-29T05:25:37Z",
      "scored": false
    },
    {
      "id": "T-cloudflare-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.cloudflare.com/",
      "sealed_at": "2026-04-29T01:12:37Z",
      "scored": true,
      "overall": 0.6929928502021829,
      "visual": 0.3955435956790123,
      "dom": 0.9980196984687348,
      "interaction": 0.9163028564411255,
      "aj": 1
    },
    {
      "id": "T-contentful-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.contentful.com/",
      "sealed_at": "2026-04-29T07:31:08Z",
      "scored": false
    },
    {
      "id": "T-datadog-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.datadoghq.com/",
      "sealed_at": "2026-04-29T08:30:52Z",
      "scored": false
    },
    {
      "id": "T-digitalocean",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.digitalocean.com/",
      "sealed_at": "2026-04-29T04:36:48Z",
      "scored": true,
      "overall": 0.4958750000950535,
      "visual": 0.07806108539094651,
      "dom": 0.9133692768701398,
      "interaction": 0.6566734867707665,
      "aj": 1
    },
    {
      "id": "T-figma-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.figma.com/",
      "sealed_at": "2026-04-29T05:05:19Z",
      "scored": true,
      "overall": 0.6325244346877869,
      "visual": 0.2835358796296296,
      "dom": 0.9980716253443525,
      "interaction": 0.8267001453933273,
      "aj": 1
    },
    {
      "id": "T-gcp-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://cloud.google.com/",
      "sealed_at": "2026-04-29T06:27:44Z",
      "scored": false
    },
    {
      "id": "T-launchdarkly-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://launchdarkly.com/",
      "sealed_at": "2026-04-29T08:31:42Z",
      "scored": false
    },
    {
      "id": "T-linear-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://linear.app",
      "sealed_at": "2026-04-28T21:21:02Z",
      "scored": true,
      "overall": 0.7382671296296297,
      "visual": 0.4791640946502058,
      "dom": 0.9984126984126983,
      "interaction": 0.9832254556143445,
      "aj": 1
    },
    {
      "id": "T-mailchimp-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://mailchimp.com/",
      "sealed_at": "2026-04-29T07:32:00Z",
      "scored": false
    },
    {
      "id": "T-mixpanel-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://mixpanel.com/",
      "sealed_at": "2026-04-29T07:30:34Z",
      "scored": false
    },
    {
      "id": "T-monday",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://monday.com/",
      "sealed_at": "2026-04-29T04:38:11Z",
      "scored": true,
      "overall": 0.7317805688736683,
      "visual": 0.4662620884773663,
      "dom": 0.9999968387776605,
      "interaction": 0.9730094600337371,
      "aj": 1
    },
    {
      "id": "T-mongodb",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.mongodb.com/",
      "sealed_at": "2026-04-29T04:37:56Z",
      "scored": true,
      "overall": 0.7189672258207029,
      "visual": 0.4493195730452675,
      "dom": 0.9879164495920609,
      "interaction": 0.9586500884090182,
      "aj": 1
    },
    {
      "id": "T-mongodb-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.mongodb.com/",
      "sealed_at": "2026-04-29T06:27:22Z",
      "scored": false
    },
    {
      "id": "T-notion-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.notion.com/",
      "sealed_at": "2026-04-29T01:12:47Z",
      "scored": true,
      "overall": 0.6768459722222222,
      "visual": 0.3645295781893004,
      "dom": 1,
      "interaction": 0.8916236625514404,
      "aj": 1
    },
    {
      "id": "T-okta-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.okta.com/",
      "sealed_at": "2026-04-29T08:33:03Z",
      "scored": false
    },
    {
      "id": "T-openai-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://openai.com/",
      "sealed_at": "2026-04-29T05:06:10Z",
      "scored": true,
      "overall": 0.7100549650478293,
      "visual": 0.42732973251028805,
      "dom": 0.9976821192052979,
      "interaction": 0.941709260621917,
      "aj": 1
    },
    {
      "id": "T-prismic-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://prismic.io/",
      "sealed_at": "2026-04-29T07:30:43Z",
      "scored": false
    },
    {
      "id": "T-resend-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://resend.com/",
      "sealed_at": "2026-04-29T05:25:25Z",
      "scored": false
    },
    {
      "id": "T-segment-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://segment.com/",
      "sealed_at": "2026-04-29T07:30:14Z",
      "scored": false
    },
    {
      "id": "T-sentry-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://sentry.io/",
      "sealed_at": "2026-04-29T05:25:43Z",
      "scored": false
    },
    {
      "id": "T-shopify-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.shopify.com/",
      "sealed_at": "2026-04-29T06:28:44Z",
      "scored": false
    },
    {
      "id": "T-slack-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://slack.com/",
      "sealed_at": "2026-04-29T05:24:12Z",
      "scored": false
    },
    {
      "id": "T-stripe-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://stripe.com",
      "sealed_at": "2026-04-28T21:15:53Z",
      "scored": true,
      "overall": 0.6976613194444444,
      "visual": 0.40307651748971196,
      "dom": 0.9999999999999999,
      "interaction": 0.9224612139917697,
      "aj": 1,
      "agent_runs": [
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.374496261286818,
          "visual": 0.35702494855967076,
          "dom": 0.016602601311854162,
          "interaction": 0.8200601322685269,
          "aj": 1,
          "generated_at": "2026-05-14T05:22:15Z"
        },
        {
          "model": "google/gemini-2.5-pro",
          "overall": 0.34170115460388123,
          "visual": 0.23607227366255146,
          "dom": 0.12380920931682451,
          "interaction": 0.7304450995511628,
          "aj": 1,
          "generated_at": "2026-05-14T05:20:13Z"
        },
        {
          "model": "anthropic/claude-opus-4.7",
          "overall": 0.31895649814295784,
          "visual": 0.2159241255144033,
          "dom": 0.08469495724784634,
          "interaction": 0.7117189642280458,
          "aj": 1,
          "generated_at": "2026-05-14T05:18:03Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.2800715606010005,
          "visual": 0.33858217592592593,
          "dom": 0.2326500624208807,
          "interaction": 0.8197090782354661,
          "aj": 0,
          "generated_at": "2026-05-14T10:34:16Z"
        },
        {
          "model": "openai/gpt-5.5",
          "overall": 0.19024682626746786,
          "visual": 0.25034336419753084,
          "dom": 0.09360904264000403,
          "interaction": 0.739848627534025,
          "aj": 0,
          "generated_at": "2026-05-14T11:54:22Z"
        }
      ]
    },
    {
      "id": "T-supabase",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://supabase.com/",
      "sealed_at": "2026-04-29T04:36:00Z",
      "scored": true,
      "overall": 0.742668125,
      "visual": 0.48642245370370363,
      "dom": 1,
      "interaction": 0.989137962962963,
      "aj": 1
    },
    {
      "id": "T-supabase-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://supabase.com/",
      "sealed_at": "2026-04-29T05:05:30Z",
      "scored": true,
      "overall": 0.7448766666666667,
      "visual": 0.49051234567901236,
      "dom": 1,
      "interaction": 0.9924098765432099,
      "aj": 1
    },
    {
      "id": "T-twilio",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.twilio.com/",
      "sealed_at": "2026-04-29T04:37:39Z",
      "scored": true,
      "overall": 0.6603083513032576,
      "visual": 0.3745223765432099,
      "dom": 0.9276909933074425,
      "interaction": 0.8947973007883974,
      "aj": 1
    },
    {
      "id": "T-twilio-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.twilio.com/",
      "sealed_at": "2026-04-29T06:28:35Z",
      "scored": false
    },
    {
      "id": "T-vercel-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://vercel.com",
      "sealed_at": "2026-04-29T01:11:43Z",
      "scored": true,
      "overall": 0.728651355571973,
      "visual": 0.46138708847736626,
      "dom": 0.9983593223984456,
      "interaction": 0.9690002922751227,
      "aj": 1
    },
    {
      "id": "T-zendesk-home",
      "category": "saas-landing",
      "tier": "T1",
      "source_url": "https://www.zendesk.com/",
      "sealed_at": "2026-04-29T08:32:47Z",
      "scored": false
    },
    {
      "id": "T-atp-rankings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.atptour.com/en/rankings/singles",
      "sealed_at": "2026-04-29T01:19:51Z",
      "scored": true,
      "overall": 0.8040616332525112,
      "visual": 0.46451761831275723,
      "dom": 0.999982902036293,
      "interaction": 0.9716129547859588,
      "aj": 1
    },
    {
      "id": "T-cricket-cricinfo",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.espncricinfo.com/",
      "sealed_at": "2026-04-29T06:42:23Z",
      "scored": false
    },
    {
      "id": "T-f1-results",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.formula1.com/en/results.html",
      "sealed_at": "2026-04-28T21:17:48Z",
      "scored": true,
      "overall": 0.5411507534280756,
      "visual": 0.02019817386831276,
      "dom": 0.9965174129353234,
      "interaction": 0.6159263666236717,
      "aj": 1
    },
    {
      "id": "T-fifa-rankings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://inside.fifa.com/fifa-world-ranking",
      "sealed_at": "2026-04-29T05:37:09Z",
      "scored": false
    },
    {
      "id": "T-fifa-women-rankings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://inside.fifa.com/fifa-world-ranking/women",
      "sealed_at": "2026-04-29T08:43:02Z",
      "scored": false
    },
    {
      "id": "T-fifa-world-cup",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.fifa.com/fifaplus/en/tournaments/mens/worldcup",
      "sealed_at": "2026-04-29T06:40:43Z",
      "scored": false
    },
    {
      "id": "T-formula1-drivers",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.formula1.com/en/drivers.html",
      "sealed_at": "2026-04-29T05:38:19Z",
      "scored": false,
      "agent_runs": [
        {
          "model": "openai/gpt-oss-120b:free",
          "overall": 0.2662605778741074,
          "visual": 0.12422659465020576,
          "dom": 0.07421390141647122,
          "interaction": 0.6376622024812628,
          "aj": 1,
          "generated_at": "2026-04-29T12:02:52Z"
        },
        {
          "model": "claude-opus-4-7",
          "overall": 0.17581791452438522,
          "visual": 0.21800887345679013,
          "dom": 0.10360370172874245,
          "interaction": 0.7146473455473483,
          "aj": 0,
          "generated_at": "2026-05-14T10:35:49Z"
        }
      ]
    },
    {
      "id": "T-formula1-drivers-hamilton",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.formula1.com/en/drivers.html",
      "sealed_at": "2026-04-29T07:43:07Z",
      "scored": false
    },
    {
      "id": "T-formula1-team-mercedes",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.formula1.com/en/teams/Mercedes.html",
      "sealed_at": "2026-04-29T08:43:51Z",
      "scored": false
    },
    {
      "id": "T-mlb-standings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.mlb.com/standings",
      "sealed_at": "2026-04-29T05:16:33Z",
      "scored": false
    },
    {
      "id": "T-mlb-statcast",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://baseballsavant.mlb.com/leaderboard",
      "sealed_at": "2026-04-29T05:37:38Z",
      "scored": false
    },
    {
      "id": "T-mlb-team-yankees",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.mlb.com/yankees",
      "sealed_at": "2026-04-29T07:41:17Z",
      "scored": false
    },
    {
      "id": "T-mlb-teams",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.mlb.com/teams",
      "sealed_at": "2026-04-29T04:48:24Z",
      "scored": true,
      "overall": 0.6308882857759427,
      "visual": 0.28049266975308645,
      "dom": 0.9980953102503608,
      "interaction": 0.8242671564858265,
      "aj": 1
    },
    {
      "id": "T-nba-history",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nba.com/history",
      "sealed_at": "2026-04-29T06:39:53Z",
      "scored": false
    },
    {
      "id": "T-nba-stats-leaders",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.nba.com/stats/leaders",
      "sealed_at": "2026-04-29T05:36:17Z",
      "scored": false
    },
    {
      "id": "T-nba-team-lakers",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nba.com/lakers",
      "sealed_at": "2026-04-29T07:42:05Z",
      "scored": false
    },
    {
      "id": "T-nba-teams",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nba.com/teams",
      "sealed_at": "2026-04-29T01:19:14Z",
      "scored": true,
      "overall": 0.7276888087054719,
      "visual": 0.45868364197530864,
      "dom": 0.9999988199081494,
      "interaction": 0.9669468349074569,
      "aj": 1
    },
    {
      "id": "T-nfl-home",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nfl.com/",
      "sealed_at": "2026-04-29T05:18:38Z",
      "scored": false
    },
    {
      "id": "T-nfl-team-chiefs",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.chiefs.com/",
      "sealed_at": "2026-04-29T07:42:59Z",
      "scored": false
    },
    {
      "id": "T-nhl-standings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.nhl.com/standings",
      "sealed_at": "2026-04-29T05:17:04Z",
      "scored": false
    },
    {
      "id": "T-nhl-teams",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nhl.com/teams",
      "sealed_at": "2026-04-29T04:48:53Z",
      "scored": true,
      "overall": 0.7326527777777777,
      "visual": 0.46787551440329217,
      "dom": 1,
      "interaction": 0.9743004115226338,
      "aj": 1
    },
    {
      "id": "T-olympics-home",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.olympics.com/en/",
      "sealed_at": "2026-04-28T21:23:25Z",
      "scored": true,
      "overall": 0.606557102661376,
      "visual": 0.27754153806584364,
      "dom": 0.9231362816675397,
      "interaction": 0.8169089825638443,
      "aj": 1
    },
    {
      "id": "T-pga-leaderboard",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.pgatour.com/leaderboard",
      "sealed_at": "2026-04-29T08:43:44Z",
      "scored": false
    },
    {
      "id": "T-pga-tour",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.pgatour.com/",
      "sealed_at": "2026-04-29T05:18:05Z",
      "scored": false
    },
    {
      "id": "T-premierleague-tables",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.premierleague.com/tables",
      "sealed_at": "2026-04-29T04:49:12Z",
      "scored": true,
      "overall": 0.7828816219013851,
      "visual": 0.4419463734567901,
      "dom": 0.9642420980085404,
      "interaction": 0.9511732386326681,
      "aj": 1
    },
    {
      "id": "T-rugby-worldrugby",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.world.rugby/",
      "sealed_at": "2026-04-29T06:42:54Z",
      "scored": false
    },
    {
      "id": "T-superbowl-history",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.nfl.com/super-bowl/history/",
      "sealed_at": "2026-04-29T08:42:14Z",
      "scored": false
    },
    {
      "id": "T-tennis-grand-slam",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.tennis.com/",
      "sealed_at": "2026-04-29T06:41:34Z",
      "scored": false
    },
    {
      "id": "T-tour-de-france",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://en.wikipedia.org/wiki/Tour_de_France",
      "sealed_at": "2026-04-29T01:19:20Z",
      "scored": true,
      "overall": 0.6939739187819226,
      "visual": 0.4231588220164609,
      "dom": 0.9520928183286825,
      "interaction": 0.9353332455017475,
      "aj": 1
    },
    {
      "id": "T-uefa-cl",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.uefa.com/uefachampionsleague/",
      "sealed_at": "2026-04-29T05:17:13Z",
      "scored": false
    },
    {
      "id": "T-uefa-euro",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.uefa.com/euro/",
      "sealed_at": "2026-04-29T07:40:19Z",
      "scored": false
    },
    {
      "id": "T-wikipedia-tdf-2020",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://en.wikipedia.org/wiki/2020_Tour_de_France",
      "sealed_at": "2026-04-29T04:49:20Z",
      "scored": true,
      "overall": 0.683783089157706,
      "visual": 0.40274447016460907,
      "dom": 0.9548387096774192,
      "interaction": 0.9191848234435153,
      "aj": 1
    },
    {
      "id": "T-wimbledon-home",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.wimbledon.com/",
      "sealed_at": "2026-04-29T08:41:56Z",
      "scored": false
    },
    {
      "id": "T-worldrugby",
      "category": "sports",
      "tier": "T1",
      "source_url": "https://www.world.rugby/",
      "sealed_at": "2026-04-29T04:49:36Z",
      "scored": true,
      "overall": 0.6913408259388665,
      "visual": 0.3913755144032922,
      "dom": 0.9999935653662264,
      "interaction": 0.9130999825470489,
      "aj": 1
    },
    {
      "id": "T-wta-rankings",
      "category": "sports",
      "tier": "T2",
      "source_url": "https://www.wtatennis.com/rankings/singles",
      "sealed_at": "2026-04-29T05:38:10Z",
      "scored": false
    }
  ]
}