Citation-Grounded Legal AI in Romania: No Hallucinations, No CSM Sanctions
A Bucharest lawyer asked a general-purpose LLM for the current wording of Art. 148 CPC last week. The model returned a confident paragraph citing an article number that was renumbered in 2024.
Citation-Grounded Legal AI in Romania: No Hallucinations, No CSM Sanctions
A Bucharest lawyer asked a general-purpose LLM for the current wording of Art. 148 CPC last week. The model returned a confident paragraph citing an article number that was renumbered in 2024. That is the kind of error that reaches CSM — the Superior Council of Magistracy — as a disciplinary complaint. For Romanian legal AI, hallucination is not a UX bug; it is a professional-liability event.
Grounding every generated sentence to a verifiable passage in Monitorul Oficial, EUR-Lex, or a cited court decision is the only architecture that survives contact with the Bar. We build Romanian legal assistants on retrieval-first pipelines where the model never speaks about a norm it did not retrieve.
Why Generic RAG Fails on Romanian Law
Vanilla RAG over scraped legal text produces three failure modes we see repeatedly in client audits. The retriever matches on surface similarity and pulls repealed versions. The chunker splits articles mid-sentence, so the model stitches two unrelated paragraphs. The generator fills gaps with plausible but invented article numbers.
The EU AI Act classifies legal-advice systems as high-risk under Annex III, triggering documentation and human-oversight duties. The European Commission’s AI Act framework page specifies traceability requirements that free-form generation cannot meet. Add GDPR Art. 5(1)(d) accuracy obligations and you cannot ship a product that invents jurisprudence.
Our team addresses this inside the RAG practice, where retrieval and citation policy are co-designed rather than bolted on later.
The Citation-Grounding Loop
Every answer passes through a four-stage gate. If any stage fails, the system refuses rather than guesses.
The refusal path is the feature, not the failure. A lawyer prefers “not found in current legislation” over a fabricated article.
A real log line from a staging run:
2026-10-07T09:14:02Z grounding.gate decision=refuse question_id=q_8814 reason=span_coverage_below_threshold claims_total=4 claims_grounded=2 corpus=monof_consolidated_2026-10-01Two of four claims could not be mapped to retrieved spans. The system refused rather than paper over the gap.
What CSM and CSJ Actually Check
Disciplinary inspections at the Superior Council of Magistracy and audits coordinated with the High Court of Cassation and Justice focus on whether a professional verified sources before advising a client. ENISA’s AI threat work treats fabricated outputs as a core model-integrity risk, which maps directly onto professional-conduct rules in Romania. A tool that outputs a verifiable citation chain protects the practitioner. A tool that outputs confident prose does not. Our Lexnomia regulatory tooling ties every generated clause to its Monitorul Oficial reference and keeps the audit log an inspector will ask for.
CAI Technology does not ship legal assistants that write freely. We ship assistants that quote with Monitorul Oficial links and sometimes refuse to answer — the only shape of the product a Romanian lawyer can put their name next to.