元算法司法推理引擎
2 分•作者: YuriKozlov•7 个月前
我们正在试验一种用于自动化裁决的架构，它不依赖于规则库或统计预测。它没有像“if-else”规则那样对法律进行编码，也没有在过去的案例上训练模型，而是将抽象的法律推理建模为一个元算法：一个控制层，它协调几个异构组件——硬编码逻辑、数值建模以及由大型语言模型（LLM）执行的结构化自然语言程序。
核心思想是，法律推理的结构（运行哪些阶段、如何选择和解释规范、如何平衡相互冲突的利益、何时修改之前的结论）是用强类型的伪代码/元语言来表达的。这个元算法的某些部分直接用代码实现（程序检查、基本资格、图更新），某些是数学的（效用、均衡、模糊不确定性），还有一些是用自然语言编写的高级指令，LLM在严格的约束下对其进行解释。在这种设置中，LLM不是结果的预测器，而是给定程序脚本的解释器。
该系统不依赖案例法进行训练，也不试图“预测”法院。它重建了推理流程本身：从提取当事人的事实叙述和证据结构，到规范选择和加权，再到生成一个可以在内部操作图中一步一步追溯的决定。通过交换规范包，相同的元算法可以适用于不同的司法管辖区；到目前为止，我们已经在一些国际和国内争议中对其进行了测试。
这里有一个早期的公开演示：
https://portal.judgeai.space/
如果您上传一份小的索赔声明和一份答复，引擎将运行完整的流程并输出一个结构化的判决文档。
我们非常感谢从事混合符号/语义系统、“LLM作为解释器”架构或复杂决策形式化模型工作的人的反馈。我们目前面临的明显开放性问题是：如何最好地测试这种元控制的失效模式，使用什么形式化工具来检查推理图的一致性，以及在达到严格的理论极限之前，这种方法可以推进到什么程度。
查看原文
We’re experimenting with an architecture for automated adjudication that doesn’t rely on rule bases or statistical prediction. Instead of encoding law as “if–else” rules or training a model on past cases, we model abstract legal reasoning as a meta-algorithm: a control layer that orchestrates several heterogeneous components — hard-coded logic, numerical modeling, and structured natural-language procedures executed by an LLM.<p>The core idea is that the structure of legal reasoning (which stages to run, how to select and interpret norms, how to balance competing interests, when to revise earlier conclusions) is expressed in a strongly typed pseudocode / meta-language. Some parts of this meta-algorithm are implemented directly in code (procedural checks, basic qualification, graph updates), some are mathematical (utilities, equilibria, fuzzy uncertainty), and some are written as high-level instructions in natural language, which the LLM interprets under tight constraints. In that setting, the LLM is not a predictor of outcomes but an interpreter of a given procedural script.<p>The system doesn’t train on case law and doesn’t try to “predict” courts. It reconstructs the reasoning pipeline itself: from extracting the parties’ factual narratives and evidence structure, through norm selection and weighting, up to generating a decision that can be traced back step-by-step in the internal graph of operations. The same meta-algorithm can work with different jurisdictions by swapping norm packages; we’ve tested it so far on a set of international and domestic disputes.<p>There is an early public demo here:
https://portal.judgeai.space/<p>If you upload a small statement of claim and a response, the engine runs the full pipeline and outputs a structured decision document.<p>We’d be grateful for feedback from people working on hybrid symbolic/semantic systems, “LLM as interpreter” architectures, or formal models of complex decision-making. Obvious open questions for us are: how best to test failure modes of this kind of meta-control, what formal tools to use for checking consistency of the reasoning graph, and how far one can push this approach before hitting hard theoretical limits.