{
  "arena_text": {
    "name": "Arena Text",
    "description": "LMArena Text Arena — human preference for general chat, writing, creativity. ELO-based, continuously updated, stable across model generations.",
    "maxScore": null,
    "type": "elo",
    "higherIsBetter": true,
    "source": "https://lmarena.ai/leaderboard",
    "recommended": true,
    "stable": true
  },
  "arena_code": {
    "name": "Arena Code (WebDev)",
    "description": "LMArena WebDev Arena — human preference for coding and web development. ELO-based, stable across generations.",
    "maxScore": null,
    "type": "elo",
    "higherIsBetter": true,
    "source": "https://lmarena.ai/leaderboard/code",
    "recommended": true,
    "stable": true
  },
  "aa_intelligence": {
    "name": "AA Intelligence Index",
    "description": "Artificial Analysis composite intelligence score. Aggregates multiple benchmarks into one normalized 0–100 score. Maintained continuously with consistent methodology.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://artificialanalysis.ai/",
    "recommended": true,
    "stable": true
  },
  "aider_polyglot": {
    "name": "Aider Polyglot",
    "description": "Real-world coding benchmark across 6 languages. Tests actual code generation quality.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://aider.chat/docs/leaderboards/",
    "recommended": true,
    "stable": false
  },
  "swe_bench": {
    "name": "SWE-bench Verified",
    "description": "Real GitHub issues — tests ability to fix actual bugs in production codebases.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://www.swebench.com/",
    "recommended": true,
    "stable": false
  },
  "livebench": {
    "name": "LiveBench",
    "description": "Contamination-resistant benchmark updated monthly with new questions.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://livebench.ai/",
    "recommended": true,
    "stable": false
  },
  "bigcodebench": {
    "name": "BigCodeBench",
    "description": "Challenging code benchmark with 36% gap between human (97%) and best models (61%).",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://bigcode-bench.github.io/",
    "recommended": true,
    "stable": false
  },
  "gpqa_diamond": {
    "name": "GPQA Diamond",
    "description": "Graduate-level science QA. Still challenging but approaching saturation at 90%.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://github.com/idavidrein/gpqa",
    "recommended": true,
    "stable": false
  },
  "mmlu": {
    "name": "MMLU",
    "description": "Massive Multitask Language Understanding. SATURATED — top models cluster within 2–4%.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://huggingface.co/spaces/open-llm-leaderboard",
    "recommended": false,
    "stable": false,
    "warning": "Saturated benchmark — no longer differentiates frontier models"
  },
  "humaneval": {
    "name": "HumanEval",
    "description": "Code generation benchmark. SATURATED — SOTA at 99.4%.",
    "maxScore": 100,
    "type": "percentage",
    "higherIsBetter": true,
    "source": "https://github.com/openai/human-eval",
    "recommended": false,
    "stable": false,
    "warning": "Saturated benchmark"
  }
}