⚖️ The first public benchmark of Croatian legal AI

Which AI knows Croatian law best?

We test ChatGPT, Gemini, Claude, Mistral and others on real Croatian legal questions — same questions, blind scoring, public results. Transparent: every answer with sources and each model's reasoning.

Open benchmark · verifiable methods · not legal advice

decisions in the database
laws in the Wayback

Benchmark · Phase 1

Leaderboard — same questions, all models.

Leading AI systems on real Croatian legal questions, blind scoring of relevance and citation accuracy. Public and reproducible.

ChatGPT
Gemini
Claude
Mistral
LEXI
Lex-Rex

⏳ First results soon. Methodology public · these are placeholders, not real numbers.

Legal Intelligence

What 16,000+ decisions already reveal.

From the corpus we surface patterns no one publishes today — real success rates and odds. Examples (computed, not marketing):

26.3%

Constitutional complaints upheld (odds 3.80). On the merits: 32.3% success.

90.7%

Public bodies lose vs the Information Commissioner (odds 1.10).

2.3%

Laws struck as unconstitutional (odds ~43). Rare but measurable.

18.2%

Access to information: the citizen wins at the High Administrative Court (odds 5.49).

3.6%

Judicial ethics (Supreme Court): complaint success (odds 28).

+ trends

Argument patterns, articles cited in overturned judgments, change over time.

Computed over >16,000 decisions; methodology and CSVs available. Not legal advice or outcome prediction.

Live demo

Try it — ask a legal question.

Searches the corpus and returns relevant decisions with sources and a score. Demo over a limited corpus; not legal advice.

Upload with protection — unlike the others

Your personal data never leaves your device.

Paste or load a document. A pseudo-anonymizer in your browser replaces personal data (national ID, IBAN, email, phone, names) before anything is sent onward. The mapping stays with you. Transparent — you see exactly what was removed.

⚠️ Pseudo-anonymization is best-effort, not a guarantee: it reliably catches structured data (national ID, IBAN, email, phone) and terms you provide; names without a list may slip through. Always review before sending. Everything runs locally in your browser.

Legislative Wayback Machine

Which version of the law was in force — and exactly when?

Laws change. Lex-Rex reconstructs the text of provisions over time from the official gazette (ELI), so you know which wording of an article was in force on any date. Crucial when the outcome depends on the version that applied at the time.

Example — Croatian Courts Act over time:
NN 28/13 → 33/15 → 82/15 → 82/16 → 67/18 → 126/19 → 130/20 → 21/22 → 60/22 → 36/24 → 136/25

What Lex-Rex is

Legal retrieval — not legal prophecy.

Lex-Rex doesn't invent answers. It finds and cites real decisions and provisions, shows why they're relevant, and how confident it is. Lawyers and citizens alike get to the source fast — and can verify it.

🔎

Questions, not keywords

Ask naturally — “Are repeated late arrivals grounds for extraordinary dismissal?” No boolean operators, no legalese required.

📎

Always with sources

Every claim is tied to a specific decision or statutory article. No source — no claim.

🧠×3

Multi-model cross-check

Several independent AI models answer; the system reconciles them: agreement = confidence, disagreement = clearly flagged uncertainty.

🕰️

Law through time

Which version of an article was in force on a given date? Point-in-time view of legislation from the official gazette.

🌍

Built to scale

Starting with the Croatian corpus (case law, consolidated statutes, scholarship), designed to extend to other jurisdictions.

🔓

Open & verifiable

Methodology and benchmark are public. You can check how well the system actually retrieves the law.

How it works

From question to sources in four steps.

A retrieval-augmented approach grounded in a real corpus, with model cross-checking.

Ask a question

In natural language — casual or formal.

Retrieve from corpus

Hybrid search (meaning + keywords) finds the most relevant decisions and provisions.

Cross-check

Multiple AI models answer independently; the system extracts consensus and flags disagreement.

Answer with sources

Summary, citations, snippets and a confidence signal — ready to verify.

Why different

Legal AI shouldn't be a black box.

Typical “legal AI”

  • Single model — errors stay invisible
  • Answers without verifiable sources
  • Closed, unpublished methodology
  • Unclear where your data goes
  • Mixes web content with the law

Lex-Rex

  • Multiple models that check each other
  • Every claim tied to a source (decision / article)
  • Public methodology and open benchmark
  • Transparent about data handling & sources
  • Grounded strictly in the legal corpus
  • In-browser pseudo-anonymization before sending

Lex-Rex Toolkit

More than search — a full legal toolkit.

A modular set of tools for working with your case file. The basics are free; advanced forensic processing and full-featured modules come with a paid plan.

✂️

PDF tools free

Split, merge and compress PDFs — fast and local.

🔄

Document converter free

DOCX / ODF / images / HEIC → PDF, with merge into one document.

🖼️

Image resize free

Resize/optimize images (e.g. for upload limits), batch.

🔤

OCR + transcription premium

Scanned files → searchable text; hearing audio/video → transcript.

⬇️

e-Komunikacija downloader premium

Auto-download all filings per case from the court e-comms portal.

📝

Draft review premium

Upload your draft → multi-model review: checks citations against the corpus, structure, missing arguments, deadlines. Anonymized first. Not legal advice.

🗂️

Case-File Analyzer premium

Forensic assembly: dedup, auto-pagination, index, chain-of-custody (SHA-256), citation graph and batch.

Pricing

Search is open. The Toolkit is a plan.

Basic search and the public benchmark stay free. More models in the cross-check and the full Toolkit come with paid plans — you pay for depth of verification and advanced tools, not for "legal advice".

Free
€0
  • Ask Lex: a few free queries with the 3-model panel
  • ✓ Basic tools: split/merge PDF, converter, image resize
  • ✓ Public benchmark · sources
  • ✓ After that: upgrade to Pro
Request access
Pro · recommended
€XY/mo
  • 3-model cross-check + disagreement flags
  • ✓ OCR, transcription, e-Komunikacija downloader
  • ✓ Legislative Wayback + pseudo-anonymizer
  • ✓ Export · higher limits
Request access
Toolkit · Premium
€XZT/mo
  • 5 models + debate round
  • Draft review + full Case-File Analyzer, citation graph, batch
  • ✓ Priority processing
  • ✓ Everything in Pro
Contact

Prices indicative (TBD). Benchmark results stay free and public. "Not legal advice" applies on every tier.

Benchmark

Verify us. Literally.

Lex-Rex is built as an open benchmark: the same legal questions, public results, blind relevance evaluation. The goal is to let academia and the public test how well AI systems actually find and cite the law.

Join early access

Early access & benchmark updates

Leave your email — we'll reach out when the demo is ready and when we publish the first benchmark results.

Thanks! We'll be in touch. 🟡
⚠️ Not legal advice. Lex-Rex is a tool for finding potentially relevant case law and legislation. Verify the sources and consult a qualified lawyer before taking any legal action.