Definition
Plain language
Finding errors hidden in something like a contract or codebase that span multiple sections.
As stated in the literature
In Fukui's cliff paper, the task of identifying cross-section contradictions in long documents; benchmarked under solo versus orchestrated multi-agent partitioning.
Also called: cross-section defect detection
Why it matters: It's a useful stress test for whether long-context models really integrate information across a document or just answer from local snippets.
For example, the task is to find that section 4.2 of a 200-page contract obligates one party to deliver in 30 days while appendix C says 60.