Lexifina团队近日发布了一项基于Lean证明助手的《美国国内收入法典》(Title 26)形式化验证项目,旨在通过机器检查识别法律文本中的起草错误及未被利用的税务漏洞 1。该项目的技术核心在于使用Lean类型检查器证明每一步逻辑均源自定义,而非依赖人工信任,从而将法律条文转化为可计算的代码 1。数据源采用了截至2026年11月3日的《美国法典》版本,模型覆盖了包括《One Big Beautiful Bill Act》带来的变更在内的2025年联邦所得税计算 1。
在与IRS发布的示例以及PolicyEngine-US、Tax-Calculator和UsTaxes三个开源税务软件进行的对比测试中,该工具在4,993/5,000个家庭案例中与PolicyEngine-US完全一致 1。此次验证成功识别出《美国国内收入法典》中138处真实的起草错误,并发现了若干已知的税务规避策略以及新颖的漏洞 1。具体而言,新发现的漏洞显示,对于领取社会保障金的夫妇,即使共同居住,在特定收入范围内分别申报可能比联合申报更省钱(例如节省1,482美元),且Schedule 8812表格允许退款的儿童抵免额高于法律规定的上限,存在形式与实质不一致的问题 1。这些结果证明了利用AI和形式化方法自动化审查复杂法律体系的可行性 1。
A team known as Lexifina has applied the Lean proof assistant to formally verify Title 26 of the United States Code, aiming to detect drafting errors and unexploited tax loopholes through machine checking 1. By converting legal provisions into computable code, the project identified 138 genuine mistakes in the statutory text 1. The verification model covers federal income tax calculations for 2025, including changes introduced by the One Big Beautiful Bill Act, and utilizes a version of the US Code dated November 3, 2026 1.
The technical core of the initiative relies on the Lean type checker to prove that every logical step derives directly from definitions, eliminating reliance on human trust 1. In comparative testing against three open-source tax software tools—PolicyEngine-US, Tax-Calculator, and UsTaxes—the model matched PolicyEngine-US exactly in 4,993 out of 5,000 household cases 1. Additionally, the analysis revealed several known avoidance strategies alongside novel vulnerabilities. One finding indicates that married couples receiving Social Security benefits may save up to $1,482 by filing separately rather than jointly within certain income ranges, even if they reside together 1. Another discovery shows that Schedule 8812 allows child tax credit refunds higher than the statutory limit due to discrepancies between the form and the law 1. These results demonstrate the feasibility of using AI and formal methods to automate the review of complex legal systems 1.
评论
还没有评论,欢迎留下第一条。