Most legal AI shows you a score with no method behind it. We publish two, each with its method, its number, and the cases we still miss. The first measures whether our review catches real contract defects; the second measures retrieval, the foundation underneath it. Both are dated and reproducible.
Run: July 2026 · 50 contracts
We built 50 contracts across Panama, Colombia, Mexico, and Brazil, each seeded with known defects and paired with a counsel-written answer key. The production review engine, the same model, prompts, and corpus grounding the product uses, reviews each one, and an independent model scores its findings against the key. We caught 96% of planted defects (291/304) and 94% of the must-catch ones (152/162).
Brazil is the weak point at 90%: its corpus is not built yet, so jurisdiction-specific defects still slip through, a missing BACEN registration of foreign capital, an absent LGPD data clause. Across the suite the review also raised 56 false positives against documented traps. Both are published here on purpose; the honesty is the point.
Run: July 2026 · 12 questions
For each question we know the code and the specific article the answer must come from. We measure two things: code correctness (are the retrieved sources from the right body of law, e.g. federal vs. a specific state) and article hit-rate (did the exact governing article appear in the top results). We report the honest number and the cases we miss, so the weak spots are visible and fixable.
This measures retrieval, the foundation. It is not inflated to look good; the 'pena convencional' case is a real miss we are tuning. Retrieval is only the floor: on top of it, every answer is synthesized under a strict-grounding prompt, its citations are structurally verified, an independent model checks that each claim is supported, and the system abstains when the corpus is not enough. That verification layer, not the retrieval score alone, is why an MXLEX answer is safe to rely on.
Read how verification works