Legal AI systems that don't resolve ambiguous terms like "interest" or "security" score 27% lower on reliability, and nearly 5 points lower on the highest-stakes questions, according to our new evaluation framework.