Auryth research lab
Where domain expertise meets AI research
You shouldn't have to take an AI tool's word for it. We publish our methods so you can see exactly why Auryth gives better answers than generic AI — and so the entire industry can build on what we learn.
Our mission
Regulated domains demand a level of precision, temporal awareness, and jurisdictional nuance that generic AI models aren't built for.
General-purpose AI consistently fails on specialist questions — not because the technology is bad, but because it wasn't designed for these domains.
We publish our research because the entire industry benefits when the problems of legal AI are studied openly. Hallucination rates, confidence calibration, source attribution, multilingual retrieval — these are hard problems that deserve serious academic attention.
Core retrieval research
Large language models are powerful but unreliable for high-stakes professional work. Our research focuses on the retrieval layer — the systems that find, verify, and present evidence to the AI. Auryth's products are built on patent-pending technology across five areas of retrieval innovation.
Negative evidence in retrieval
Systems that actively identify when evidence contradicts or fails to support a conclusion.
Calibrated scoring
Confidence measures that correlate with actual accuracy, not just model certainty.
Confidence-gated generation
Output controls that prevent low-confidence answers from reaching users.
Adaptive query routing
Dynamic selection of retrieval strategies based on query characteristics.
Self-improving retrieval systems
Feedback loops that refine accuracy without model retraining.
Research focus areas
Confidence calibration
Can you actually trust the confidence score?
We measure whether our confidence scores match real-world accuracy. When Auryth says 85% confident, we test whether 85% of those answers are actually correct.
Multilingual legal retrieval
Ask in Dutch, find the answer in French — accurately
Multilingual regulatory landscapes create unique challenges. We research how to find the right provision regardless of which language it's written in.
Temporal versioning
Getting the right rule for the right date
Regulations change constantly. We research methods for tracking which version of a provision was in force on the date that matters — so you never cite outdated rules.
Hallucination detection
Catching fabricated citations before they reach you
How do you catch an AI that confidently cites a non-existent article? We develop methods to verify every citation against real sources before showing you the answer.
Working papers
In preparation
Confidence-calibrated retrieval for regulated domains
Our first working paper examines how to make AI confidence scores actually meaningful in professional contexts. Introduces the framework behind Auryth's confidence scoring and how we verify accuracy against real domain-specific questions.
Download paper (PDF)Advisory board
We're building an advisory board of domain practitioners, academics, and AI researchers who share our commitment to transparent, reliable specialist AI.
If you're a researcher working on domain-specific NLP, an academic interested in AI applications in regulated fields, or a practitioner who wants to help shape the next generation of specialist AI tools — we'd love to hear from you.
Partnerships
We're actively exploring partnerships across three areas:
University research centres
Joint projects on legal AI, NLP, and computational law
Professional associations
ITAA, IBR/IRE, and regional accounting bodies
EU research programmes
Digital governance and AI innovation grants
Interested in collaborating?
Whether you're a researcher, academic, or practitioner — we're always open to conversations about advancing legal AI together.
Get in touch