An AI pipeline can now break down Italian tax-court judgments into structured legal arguments at scale — and flag when the model makes up case citations.
Researchers built a system that ingests roughly 330,000 first- and second-instance Italian tax-court decisions and decomposes each one into individual legal issues. For every issue, the pipeline produces an XML representation organized around the IRAC framework — Issue, Rule, Application, Conclusion — and the classical legal syllogism. The team chose DeepSeek V3 as the underlying model, citing the need to process hundreds of thousands of documents without blowing the budget. A dedicated citation parser called Linkoln checks the model's references against identifiers actually found in each judgment, catching hallucinated case law before it contaminates downstream analysis.
Legal AI has a well-known citation problem: models confidently produce plausible-sounding but nonexistent case references, a failure mode that has embarrassed lawyers in live court filings. Grounding the model's output against a structured parser and standard identifiers — URN-NIR, ECLI, CELEX — is a more systematic answer to that problem than prompting alone. The researchers validated against 50 judgments annotated by two tax-law PhDs, measuring both inter-annotator agreement and how closely the model matched expert analysis.
The authors call it the first expert-validated, issue-level extraction pipeline with hallucination control for Italian tax courts, which is a narrow enough claim to be credible. Whether the approach scales to other legal systems with messier citation norms is the question worth asking next.