律师 + 税务师 + 专利代理师 | 在做 AI 法律产品 Lawyer · Tax Adviser · Patent Attorney → building AI legal-product & compliance tooling that measures LLM quality.
我把六年的法律 / 税务 / 知识产权实务经验,翻译成一套可量化、可复现的 AI 质量评测与防护产品——不是只会调 prompt,而是能把"模型错了"变成"可度量、可修复的规格"。 I turn six years of legal / tax / IP practice into AI-quality evaluation and guardrail tooling that is quantifiable and reproducible — turning vague "the model is wrong" into a measured, fixable specification.
| 仓库 Repo | 角色 Role | 一句话 One-liner |
|---|---|---|
| 🧱 verified-chinese-law-kb | 数据地基 · Data foundation | 8 部法律、2,327 条逐字核验法条(M1 民法典 1,260 条全文,其中 27 条为律师抽核逐条签核初版,其余由 AI 核验)、可独立下载模块 / 2,327 line-by-line verified articles (M1 Civil Code: 1,260 full-text; 27 attorney spot-checked & line-by-line signed in the initial batch, the rest AI-verified) across 8 laws, independently downloadable. |
| 🔬 law-citation-bench | 评测地基 · Eval foundation | 离线零依赖基准,500 题量化法条引用准确率;一行提示词修复回收 +97 分 / offline stdlib benchmark, 500 Q, a one-line prompt fix recovered +97 pts. |
| 📉 legal-hallucination-bench (私有 · private) | 量化基准 · Quantification | 5 国产模型 × 23 陷阱题,HVI 33.3%–54.2%,8 法域逐字 EXACT 全 0%;ground truth 2,327 节点含 212 个 Tier A 专家逐条签核节点 / 5 models × 23 traps, HVI 33.3–54.2%, 8 laws EXACT all 0%; ground truth 2,327 nodes incl. 212 Tier A attorney-signed. |
| 🛡️ compliance-triangle | 产品 · Product | 同一套 verify 引擎给每条 AI 引注盖 🟢🟡🔴 章(法律·税务·IP 三域)/ same verify engine → 🟢🟡🔴 stamps across legal·tax·IP. |
| 📡 legal-ai-watch | 公开榜单 · Public leaderboard | 31 题公开评测 + gh-pages 仪表盘,跟踪主流模型法条引用表现 / 31-Q public eval + live dashboard. |
- 律师 Lawyer —— 懂"引用法条"在法律实务里有多严肃 → 定下逐字二元判定。 / understands how serious citation is in practice → defines the exact-match binary criterion.
- 税务师 Tax Adviser —— 把增值税法 / 企税 / 个税优先纳入(多数法律 AI 评测不碰税法)。 / prioritizes tax law (most legal-AI evals avoid it).
- 专利代理师 Patent Attorney —— 职业习惯是"精确比对文本" → 与基准"一个字都不能差"同一思维。 / habit of precise text comparison → same mindset as "not one character off".
同一人设计校验规则、定义陷阱、签署每一条 KB——这是任何纯工程 / 纯算法团队无法复制的壁垒。 One person designs the rules, defines the traps, and signs every KB entry — a barrier no pure-engineering team can replicate.
- 测试规模 Scale:5 国产模型 × 23 陷阱题 = 115 条有效回答 / 115 valid responses.
- 引注幻觉率 HVI(最宽松尺度):33.3%–54.2%(连付费旗舰都不过半)/ even paid flagships fail half.
- 内容级 EXACT(逐字合规率):8 法域全部 0% / 8 laws all 0% exact match.
- 知识库 KB:2,327 条逐字核验法条 · 8 部法律 / KB: 2,327 verified articles · 8 laws.
- 评测基准 Benchmark:500 题确定性生成,T1 接地 / T2 检索 / T3 幻觉 / 500 deterministic questions.
Python 标准库 · 离线零依赖 · 可复现评测 · LLM 评测 · 法律 AI · 税务合规 · 知识产权 · 产品设计
Python stdlib · offline-zero-dep · reproducible eval · LLM evaluation · legal AI · tax compliance · IP · product design
欢迎 AI 法律 / 合规方向的团队交流:在任意仓库开 Issue,或通过 LinkedIn / 邮件联系。 Open to AI-legal / compliance roles — open an Issue on any repo, or reach me via LinkedIn / email.
注:评测结果与数字随版本演进,最新以各仓库 README 为准;本页仅作作品集导航,不构成法律 / 税务 / 专利意见。 Numbers evolve with versions; latest per-repo READMEs are authoritative. This page is navigation only, not legal / tax / patent advice.