Can AI understand abstract safety rules without seeing every hazard?

Hazard detection systems usually require thousands of labeled examples for each specific danger type, but safety rules are abstract and change by context. This work introduces CompliVision—a dataset of 3,006 workplace images (traffic, construction, warehouse) each annotated against explicit safety rules with natural language explanations—and pairs it with an active learning framework that fine-tunes vision-language models for compliance assessment. The approach decouples hazard concepts from examples, letting the system generalize to novel scenarios grounded in ISO standards.