CUAD

The model identifies that a clause exists but extracts only part of the relevant text, missing conditions, exceptions, or carve-outs that change the legal meaning. The correct answer for this example is NONE: the clause does not exist, but the model returns text it incorrectly treats as responsive. This is a false positive on clause presence, the most operationally dangerous error type in contract review.

Watch for: False positives on clause existence are harder to catch than missed clauses. A reviewer who trusts the extraction may miss the fact that no responsive clause exists. For M&A diligence workflows, validate specifically on negative examples where the correct answer is NONE.

When a clause type is absent from the contract, models sometimes return a date, name, or phrase that appears nearby in the document rather than returning no answer. The effective date of the agreement is present; the specific clause being searched for is not. The model confuses document presence with clause presence.

Watch for: This failure produces confident-looking wrong answers. A model returning "1 August 2011" for a missing clause type looks correct to a reviewer who does not check the source. Spot-check negative clause results on a sample before relying on systematic extraction at scale.

The model locates the correct section of the contract but extracts text from the wrong clause within it. Renewal term provisions often sit adjacent to termination provisions. The model returns the wrong span, producing an answer that is textually plausible but legally incorrect.

Watch for: Adjacent clause confusion is most common in contracts with dense boilerplate sections where multiple related provisions appear in sequence. If your workflow covers renewal, termination, and notice clauses, test all three together on the same contracts, not in isolation.