Building a compliance linter where no LLM decides compliance
Why I gave language models exactly one job on PlanLint, extracting structure, and handed every actual verdict to a few hundred lines of plain Python that can only answer from measured geometry and printed text.
Start here