Lexiel 98.0% on the bar access exam: our own measurement

The Problem We Cannot Ignore: Legal Hallucinations

In 2023, a New York attorney filed a brief stuffed with ChatGPT-invented court decisions. The judge imposed a $5,000 fine and the case went viral. Spanish bar associations issued warnings. The question hung in the air: is it safe to use AI for real legal work?

The answer isn't simply yes or no. It depends entirely on which AI you use and how it's built.

The Test: the July 2026 Bar Access Exam

To answer that question objectively, we put Lexiel through Spain's official bar access exam (July 2026 session, 152 scorable multiple-choice questions). We chose this exam for three reasons:

1. Maximum difficulty: It's the reference legal exam in Spain, covering all branches of the legal system.
2. Verifiable answers: Every correct answer is published by the Ministry of Justice with a definitive answer key.
3. No shortcut possible: Questions require genuine legal reasoning about current law, not fact memorization.

We evaluated Lexiel with its production legal retrieval. The raw file is published on the benchmark page. We have not measured Maite AI or any other legal product.

The Results

Tool	Correct	Score
Lexiel	149/152	98.0%

Evaluation was conducted using automatic grading against the definitive answer keys published by the Ministry of Justice (July 2026 session).

Errors concentrate on questions about recent legislative reforms (2024-2025) and on subjects where case law is decisive. We review each error to improve the corpus.

Why the Legal Corpus Matters

We have not measured general-purpose models on this exam, so we do not publish a comparison. What we can explain is how Lexiel is built: the key is not only the underlying AI model but the legal corpus powering it via RAG (Retrieval-Augmented Generation):

• 130,000+ official sources: BOE legislation (Civil Code, LEC, LECrim, Criminal Code, LOPJ, Constitution and more), Supreme Court and Constitutional Court case law, and sources from the EU and several Latin American countries
• All indexed with semantic embeddings for real-time retrieval

When Lexiel answers about Article 609 of the Civil Code, it doesn't generate freely: it retrieves the exact fragments from the real statute and builds the answer on that foundation. If it can't verify something in the corpus, it says so explicitly.

What Lexiel offers

Beyond the legal chat, Lexiel offers:

1. Corpus of 130,000+ official sources: BOE legislation and case law from CENDOJ and the Constitutional Court, indexed with semantic embeddings.
2. Hybrid RAG: Vector + BM25 search with Reciprocal Rank Fusion (RRF), per-article deduplication, and verification against official sources.
3. Complete law firm platform: Legal CRM, Verifactu invoicing, client portal, procedure engine for automated briefs, time tracking, and automatic deadline management.

What the Benchmark Does NOT Measure

Honesty matters. This test measures pure legal accuracy. It doesn't measure:

• Response speed (Lexiel uses real-time streaming)
• Integration with daily law firm workflows
• Usability and learning curve
• Productivity features beyond chat

Those are aspects every attorney should evaluate with their own 14-day free trial.

Conclusion

The benchmark measures answer accuracy, not citation accuracy. Lexiel cites, checks citations against the source and flags the ones it cannot verify. No AI system is error-free: always check citations.

If you use legal AI for research, brief drafting, or client consultations, it's worth demanding the same standard you'd demand from a colleague: cite the source, and make sure the source is real.

View full benchmark methodology →