GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-05
A script can complete successfully while checking the wrong things. In a building model, an empty selection or missing parameter may produce a tidy report without answering the intended regulatory question. This makes execution success a particularly weak acceptance test for AI-assisted code checking.
Large Language Model-Driven Code Compliance Checking in Building Information Modeling was published in Electronics on 24 May 2025; its arXiv posting followed on 25 June. The paper studies language models generating Python scripts for checks within Revit, including residential and office case studies. It is a research account of semi-automation, not an official determination that a project's design complies with applicable requirements. Published paper
Follow the chain from clause to object
GAD proposes reviewing four links separately: the source requirement, its interpretation, the selected model elements and the actual comparison performed. A correct number applied to the wrong set of objects remains an incorrect check. Likewise, a useful model parameter may not mean the same thing as the term used in the source requirement.
Record the jurisdiction, document edition and applicable project conditions. Preserve exceptions and unresolved interpretations instead of reducing a nuanced clause to a confident sentence. Have the relevant qualified professional determine applicability; a language model's explanation is material for review rather than regulatory authority.
Use a test copy of the model. Keep the generated script and reviewed version together so the team can identify what a human changed and why. The exercise should remain reproducible when a model revision or software update arrives.
Test three outcomes
A useful test set includes a known passing example, a known failing example and a case the available information cannot resolve. The last category matters: an absent property should not quietly become zero, false or a pass unless that handling has been deliberately justified.
Add boundary examples just inside and outside the intended threshold. Inspect units, rounding and element selection. Change one input at a time so an unexpected result can be traced to its cause. These are general software-review proposals, not claims that the paper performed this exact test suite.
Compare the report with the actual model elements. Every finding should identify the affected objects and the evidence used. A reviewer needs to distinguish a failed requirement from a data-quality problem and an unresolved interpretation, because they call for different next steps.
Keep the professional decision visible
Do not describe an unreviewed generated script as an automated approval system. Even a carefully tested narrow check leaves other requirements, design conditions and interactions outside its scope. The appropriate professionals and authorities retain responsibility for interpreting requirements and determining compliance.
GAD has not executed the study's workflow. Its useful lesson is to make uncertainty an ordinary output rather than an embarrassing exception. A check earns trust when it can explain what it inspected, what it found and what it could not determine.