{
  "schema": "csoai.reported/0.1",
  "issuer": "CSOAI Ltd (GB, Companies House 16939677)",
  "register": "REPORTED is the third data state: figures published by OTHERS, cited and timestamped for context. Reported here, not measured here. Unsigned. Never enters the board, never compares against MEASURED numbers, implies no endorsement of the source's methodology. MEASURED (signed) lives at /api/gspc; UNMEASURED/GATED cells stay honestly empty.",
  "states": {
    "MEASURED": "signed run on our frozen instruments — /api/gspc",
    "GATED/UNMEASURED": "honestly withheld: insufficient n or no separation test",
    "REPORTED": "third-party figure, cited + timestamped, unsigned — this endpoint"
  },
  "count": 5,
  "entries": [
    {
      "id": "arc-agi-3-human-gap",
      "claim": "Humans solved all 135 ARC-AGI-3 environments; the best frontier model scored 0.37%.",
      "figures": {
        "human_pass_rate": 1,
        "best_model_pass_rate": 0.0037,
        "best_model": "Gemini 3.1 Pro Preview"
      },
      "source": "ARC Prize (ARC-AGI-3 launch results)",
      "source_url": "https://arcprize.org/",
      "captured_at": "2026-08-19",
      "as_of": "2026-03-25",
      "attribution_basis": "Public leaderboard figures, cited with attribution.",
      "note": "ARC Prize's own summary: 'Humans score 100%. Frontier AI scores 0.51%.' Scores move — check the live leaderboard."
    },
    {
      "id": "gaia-human-gap",
      "claim": "GAIA: human respondents 92% vs 15% for GPT-4 with plugins.",
      "figures": {
        "human": 0.92,
        "gpt4_with_plugins": 0.15
      },
      "source": "Mialon et al., GAIA (ICLR 2024)",
      "source_url": "https://arxiv.org/abs/2311.12983",
      "captured_at": "2026-08-19",
      "as_of": "2023-11-21",
      "attribution_basis": "arXiv paper figure, cited."
    },
    {
      "id": "gpqa-diamond-expertise-gap",
      "claim": "GPQA Diamond: PhD-domain experts ~65% vs skilled non-experts with web access ~34%.",
      "figures": {
        "domain_experts": 0.65,
        "skilled_non_experts": 0.34
      },
      "source": "Rein et al., GPQA",
      "source_url": "https://arxiv.org/abs/2311.12022",
      "captured_at": "2026-08-19",
      "as_of": "2023-11-20",
      "attribution_basis": "arXiv paper figure, cited."
    },
    {
      "id": "human-or-not-detection",
      "claim": "In AI21's 'Human or Not' (1.5M+ participants), people identified their partner correctly only 68% of the time; 60% when talking to bots.",
      "figures": {
        "participants": "1,500,000+",
        "overall_correct": 0.68,
        "vs_bots_correct": 0.6
      },
      "source": "AI21 Labs (arXiv 2305.20010)",
      "source_url": "https://arxiv.org/abs/2305.20010",
      "captured_at": "2026-08-19",
      "as_of": "2023-06-01",
      "attribution_basis": "Company-published study, cited as self-reported."
    },
    {
      "id": "colonoscopy-deskilling",
      "claim": "Non-AI-assisted adenoma detection fell from 28.4% to 22.4% after clinicians' AI exposure — first real-world clinical AI-deskilling evidence.",
      "figures": {
        "before_ai_exposure": 0.284,
        "after_ai_exposure": 0.224,
        "endoscopists": 19
      },
      "source": "Budzyń et al., Lancet Gastroenterology & Hepatology 10(10)",
      "source_url": "https://doi.org/10.1016/S2468-1253(25)00133-5",
      "captured_at": "2026-08-19",
      "as_of": "2025-08-12",
      "attribution_basis": "Peer-reviewed observational study, cited. Caveats: observational; possible workload confounds."
    }
  ]
}