{
  "claim": "WBench's page markup tags some leaderboard entries with a data-evaluation-source attribute, and every instance of it on the page reads self-evaluation. There are 10 such row instances across the leaderboard's four tables, covering 3 distinct models: XGEN-Jing's two variants and Alaya-EVOKE-Turbo — which is to say, exactly the entries at the top of both splits. No other row carries the attribute in any form. Nothing in the rendered page surfaces it: no badge, no tooltip, no footnote. A reader scanning the visible table cannot tell these scores apart from the rest. What WBench intends by the flag is not stated anywhere I could find; that it is present, and where, is checkable in one view-source.",
  "method": "curl the leaderboard, then count: grep -o 'data-evaluation-source=\"[^\"]*\"' | sort | uniq -c, and map each flagged <tr> back to its table and rank.",
  "source": "https://meituan-longcat.github.io/WBench/",
  "captured": "2026-09-18",
  "columns": [
    {
      "key": "split",
      "label": "table"
    },
    {
      "key": "rank",
      "label": "rank",
      "align": "right"
    },
    {
      "key": "model",
      "label": "model"
    },
    {
      "key": "avg",
      "label": "avg",
      "align": "right"
    },
    {
      "key": "flag",
      "label": "data-evaluation-source"
    }
  ],
  "rows": [
    {
      "split": "Navi split (summary)",
      "rank": "1",
      "model": "Alaya-EVOKE-Turbo 3-step Alaya Lab",
      "avg": "82.0",
      "flag": "self-evaluation"
    },
    {
      "split": "Navi split (summary)",
      "rank": "2",
      "model": "XGEN-Jing bidirection XGEN Labs",
      "avg": "81.9",
      "flag": "self-evaluation"
    },
    {
      "split": "Navi split (summary)",
      "rank": "6",
      "model": "XGEN-Jing 4-step AR XGEN Labs",
      "avg": "81.0",
      "flag": "self-evaluation"
    },
    {
      "split": "Full split (summary)",
      "rank": "1",
      "model": "XGEN-Jing bidirection XGEN Labs",
      "avg": "81.0",
      "flag": "self-evaluation"
    },
    {
      "split": "Full split (summary)",
      "rank": "2",
      "model": "XGEN-Jing 4-step AR XGEN Labs",
      "avg": "79.9",
      "flag": "self-evaluation"
    },
    {
      "split": "Navi split (per-dimension)",
      "rank": "1",
      "model": "Alaya-EVOKE-Turbo 3-step Alaya Lab",
      "avg": "65.2",
      "flag": "self-evaluation"
    },
    {
      "split": "Navi split (per-dimension)",
      "rank": "2",
      "model": "XGEN-Jing bidirection XGEN Labs",
      "avg": "64.1",
      "flag": "self-evaluation"
    },
    {
      "split": "Navi split (per-dimension)",
      "rank": "6",
      "model": "XGEN-Jing 4-step AR XGEN Labs",
      "avg": "63.6",
      "flag": "self-evaluation"
    },
    {
      "split": "Full split (per-dimension)",
      "rank": "1",
      "model": "XGEN-Jing bidirection XGEN Labs",
      "avg": "61.8",
      "flag": "self-evaluation"
    },
    {
      "split": "Full split (per-dimension)",
      "rank": "2",
      "model": "XGEN-Jing 4-step AR XGEN Labs",
      "avg": "61.4",
      "flag": "self-evaluation"
    }
  ],
  "note": "Denominators: 108 data rows across 4 tables (39 / 15 / 39 / 15), of which 10 carry the attribute — 98 carry no data-evaluation-source at all. The Navi tables list 39 models, the Full tables 15; the same three models are flagged in every table they appear in. The attribute takes exactly one value across the whole page: self-evaluation."
}