{
  "_comment": "公開用 claim リスト（Loop 8, §13.1）。ソースは src/data/evidence/claims.json。schema: claimId, text, claimType, evidenceGrade, status, scope, checkedAt。無根拠の強い主張は含まない（§8.4）。",
  "schema": {
    "claimId": "string",
    "text": "string",
    "claimType": "spec|performance|methodology|editorial|caveat",
    "evidenceGrade": "A|B|C|D|E|U",
    "status": "verified|qualified|contested|unverified|deprecated",
    "scope": "global|family|model|provider|use-case",
    "checkedAt": "date"
  },
  "source": "https://cockpit.kawaii-girl.com/evidence/",
  "generatedAt": "2026-07-14",
  "updatePolicy": "公式カタログ再取得・claim 監査時に更新。自動CIで検証。",
  "license": "サイトの引用・リンクは可。データ再利用は出典明記を推奨。",
  "claims": {
    "claim-hy3-free-limited": { "text": "Tencent Hy3 は期間限定無料（tencent/hy3:free）で試験運用に向く。", "claimType": "spec", "evidenceGrade": "A", "status": "verified", "scope": "model", "checkedAt": "2026-07-14" },
    "claim-hy3-free-expires": { "text": "tencent/hy3:free の無料期限は 2026-07-21。期限後は通常料金または別variant。", "claimType": "spec", "evidenceGrade": "A", "status": "verified", "scope": "model", "checkedAt": "2026-07-14" },
    "claim-deepseek-coding-strong": { "text": "DeepSeek V3.2 はコード生成・リファクタに強い（開発者文書）。", "claimType": "performance", "evidenceGrade": "B", "status": "qualified", "scope": "model", "checkedAt": "2026-07-14" },
    "claim-claude-japanese": { "text": "Claude Sonnet 5 は日本語・安全性に強い（開発者多言語claim）。独立ベンチマーク証拠なし。", "claimType": "performance", "evidenceGrade": "A", "status": "unverified", "scope": "model", "checkedAt": "2026-07-14" },
    "claim-gemini-fast-cheap": { "text": "Gemini 3.5 Flash は低価格・高速（公式区分）。実効証拠なし。", "claimType": "performance", "evidenceGrade": "A", "status": "qualified", "scope": "model", "checkedAt": "2026-07-14" },
    "claim-context-not-effective": { "text": "Advertised context length と effective context は同義でない（Lost in the Middle, arXiv:2307.03172）。", "claimType": "caveat", "evidenceGrade": "D", "status": "verified", "scope": "global", "checkedAt": "2026-07-14" },
    "claim-llm-judge-bias": { "text": "LLM-as-a-judge には position bias 等の bias があり得る（arXiv:2406.07791, 2410.02736）。", "claimType": "methodology", "evidenceGrade": "C", "status": "verified", "scope": "global", "checkedAt": "2026-07-14" },
    "claim-swebench-coding-complex": { "text": "Real-world software issue 解決は単純コード生成より複合的（SWE-bench, arXiv:2310.06770）。", "claimType": "methodology", "evidenceGrade": "C", "status": "verified", "scope": "global", "checkedAt": "2026-07-14" },
    "claim-arena-preference-not-correctness": { "text": "Human preference（Chatbot Arena）は一評価軸。客観的正解と別（arXiv:2403.04132）。", "claimType": "methodology", "evidenceGrade": "C", "status": "verified", "scope": "global", "checkedAt": "2026-07-14" },
    "claim-helm-multimetric": { "text": "単一accuracyだけでなく複数metrics・use-caseで評価すべき（HELM, arXiv:2211.09110）。", "claimType": "methodology", "evidenceGrade": "C", "status": "verified", "scope": "global", "checkedAt": "2026-07-14" }
  }
}
