Ground each answer in the words of the law. Verify each quote against the source. Refuse when the text is silent. Every sentence carries a link to the provision it rests on. That is what separates an answer that sounds correct from one a lawyer can audit in a click.
Hallucination
A general language model answers regulatory questions in seconds, and it may be wrong. Stanford’s RegLab tested the leading legal-AI tools built on retrieval. It found wrong or unsupported answers in 17 to 33 percent of cases (Magesh et al. 2025). The problem did not go away. A 2026 study checked generated legal citations against a court citation graph. It still found 13 to 21 percent of them fabricated, misapplied, or out of date (Ovcharov 2026).
Licentium Compass answers questions on EU financial, crypto, and AI regulation from the text of the law, and it shows the exact provision behind each sentence. When its corpus does not hold the answer, it says so.
No sentence without a quote you can check
Licentium Compass pulls candidate provisions from its automatically updated corpus, writes an answer bound to those provisions, then checks every quoted span character by character against the primary source. A span that is not an exact substring of the cited article is dropped and flagged. The citation a reader sees is built from the original database record for that provision: instrument, article number, canonical reference. The model never writes the article number itself. So it cannot invent a citation or pin a real quote to the wrong article.
This check sits outside the writing step. It is a second gate, and it runs the same way every time. On a set of hard, multi-provision questions, every citation Licentium Compass displayed passed the character-level check: a 100 percent verbatim rate. On the same set, the system declined to answer questions outside its corpus instead of inventing support.
Primary law, kept current
Licentium Compass reads statutes, not summaries of them. Its corpus holds the consolidated text of the acts that govern this area: MiCA, EMIR, DORA, the AML Regulation, the sixth AML Directive, the Transfer of Funds Regulation, the AI Act, the Digital Services Act, the Data Act, and others. That is 54 instruments and more than 4,400 article-level provisions in English. Each provision is stored with a hash of its source text, so a change in the official version is detectable.
Law does not stand still. Licentium Compass records which consolidated version of each act it holds and the date of that version. When the EU publishes a newer consolidation, or when an amending act carries a date later than the version on file, Licentium Compass marks that act for refresh. Answers stay tied to the text in force. Recent work treats this timing as its own test of a citation: whether the provision was valid on the relevant date. Licentium Compass tracks versions for exactly that check.
A map of cross-references
Articles point at other articles. One procedure article gets invoked by dozens of provisions across separate regulations. Licentium Compass reads these internal citations and stores them as links: 20,138 connections from one provision to another across the corpus. A reader who opens a cited article sees the provisions it refers to, so a chain of duties is visible.
From a product to its duties, by rule, not by guess
Some questions are not about what a statute says. They ask what applies to a given business. A custody service, a stablecoin issuer, and a deployer of high-risk AI each face a different set of duties. For these, Licentium Compass runs a fixed mapping. A short intake describes the product. A set of rules turns that description into the provisions to verify, each tied to a cited article. Each rule carries a version, so a change to a rule is a new version on record.
Measuring the grounding
Licentium Compass scores its own answers. A question set with known-correct provisions runs through the full pipeline. The scoring is mechanical: whether each displayed citation quotes the source word for word, whether the cited provisions include the expected ones, and whether the system abstains on questions outside the corpus. The word-for-word pass rate reached 100 percent on the most recent run. Recall against the expected provisions is the harder number, and it is where current tuning is aimed. Naming both is the point: a grounding claim means little without the abstention rate and the miss rate beside it.
Sources
Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., and Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies.
Ovcharov, V. (2026). Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs. arXiv:2606.00898.
Ju, Y.-J., and Lee, S.-W. (2026). Citation-Closure Retrieval and Per-Rule Attribution for Real-World Regulatory Compliance Question Answering. arXiv:2605.29742.
