Should contract AI find the evidence first, or go straight to a “high risk” label?

CUAD is maintained by The Atticus Project. Its website states that it contains 510 commercial contracts, more than 13,000 annotations, and 41 clause categories, with annotation work supervised by experienced lawyers. It supports research into identifying contract clauses; the dataset description itself does not establish that any system can perform comprehensive legal judgment. A starting point for discussion If a decision model outputs only “high risk,” the reader still does not know which passage it examined or which exceptions it missed. One engineering approach worth discussing is to locate clauses first and retain their original text and positions, then have a model select the category to review, and finally leave the assessment of their implications to a qualified professional. The editors suggest measuring whether all relevant evidence was found separately from whether the classification was correct, so a combined score does not conceal missing evidence. This is a discussion of product design, not legal advice on a specific contract. Let's discuss 1. When a clause is not found, should the interface say “no such clause” or “insufficient evidence found”? 2. How should exceptions spanning multiple paragraphs be presented together? 3. Would you prioritize recall or reducing the number of passages reviewers need to read? Source: The Atticus Project: CUAD dataset https://www.atticusprojectai.org/cuad/ Editorial note: The official S1 Hub editorial account used AI assistance to compile public sources. This was not submitted by the original author, and this site has not conducted independent hands-on testing. Additional evidence and differing views are welcome.