De Jure: Getting Legislation into AI
mlj.solutions
mlj.solutions AI jako digitální právník? ⚖️🤖 Co když umělá inteligence dokáže sama číst zákony, vyhodnocovat pravidla a kontrolovat jejich správnost? Přesně tím se zabývá systém De Jure - technologie, která automatizuje práci s právními předpisy a posouvá compliance na novou úroveň. ➡️ méně manuální práce ➡️ rychlejší zpracování regulací ➡️ vyšší přesnost díky self-check mechanismům Budoucnost firem nebude jen o tom, kdo má data. Ale kdo jim dokáže správně porozumět. Jaký máte názor vy - důvěřovali byste AI při auditu právních předpisů? 👀
View on InstagramWhat the study is about
- Getting legislation into a form that AI can understand remains a manual and costly process.
- De Jure is a fully automated pipeline from Vanguard that handles this without human annotations, domain-specific prompts, or training data.
94%
average extraction quality score
De Jure achieved an overall average score above 4.70 out of 5.00 across three regulatory domains, three legal systems, and four different AI models — with zero customization.
73.8%
of cases where De Jure outperformed the competition
For compliance questions based on HIPAA regulations, answers grounded in De Jure rules were preferred in 73.8% of cases with standard retrieval, and in 84% with broader retrieval.
0
human annotations required for deployment
The pipeline operates without labeled data, domain-specific prompts, or expert configuration — and still delivers results comparable to approaches that rely on manual effort.
Key takeaways
- Legal text is extremely challenging for AI — conditions, exceptions, and hierarchical dependencies mean that a first extraction pass is almost never sufficient.
- Iterative refinement is essential — a single attempt is not enough; quality only improves after a second and third pass.
- Open-source models are closing the gap — the difference between the best open-source model and GPT-4o-mini was less than 0.04 points.
- Input quality determines output quality — poor document segmentation cannot be corrected by downstream processing.
- Hallucinations can be eliminated — schema-constrained extraction achieved a Non-Hallucination score of 5.00 across all models and all phases.
Our take
Compliance in regulated industries is not about reading the law. It’s about having a system that knows what the law says — and can verify it. These are exactly the kinds of architectures we design.
Michal Dobrovolný
Zdroj
Guliani, K., Gill, D., Landsman, D., Eshraghi, N., Kumar, K., & Gondara, L. (2026). De Jure: Iterative LLM Self-Refinement for Structured Extraction of Regulatory Rules. arXiv:2604.02276. https://arxiv.org/abs/2604.02276
Don't miss these posts
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
