AI Improves Constraint Checking in Construction Document Review

Reviewing professional documents, such as construction plans, is a complex task involving constraint checking based on relationships between text, geometry, pages, and document revisions. A new evidence-grounded pipeline has been developed to streamline this process. This pipeline normalizes extracted facts, deterministically executes four-state rules, retains source spans for traceability, and escalates any unresolved cases for human review. The pipeline's PDF evidence allocator was evaluated on 160 reference-based tasks from 29 construction projects. In a repeated test, reallocating an image budget from broad page overviews to one overview and three overlapping tiles significantly improved standardized decision accuracy by 10.6 percentage points. However, this specific benefit did not consistently extend to a broader test set, where a trade-off between resolution and breadth was observed, with some cases favoring broader page coverage. Despite the improvements, the study noted that exact finding-set recovery remains low, false passes are still common, and repeated-run agreement is poorly calibrated. These findings suggest that while region-focused evidence can be beneficial, there isn't a universal advantage, indicating a need for rule-aware evidence routing and continued expert human review to ensure accuracy in critical document analysis.