Regulator-in-the-Loop Compliance Screening: Graph-Grounded LLM Agents with a Continuously Calibrated Benchmark for EU Marketing Law

Market-surveillance authorities face a growing volume of digital marketing content spanning video, social-media, influencer, and AI-generated or synthetic media, whose compliance with EU consumer and digital law cannot be assessed manually at scale. We present a deployed decision-support system that screens marketing campaigns against the Unfair Commercial Practices Directive, the General Data Protection Regulation, the Digital Services Act and the AI Act using large language model agents grounded in a curated legal knowledge graph. The system enforces closed-world citation grounding: findings may reference only provisions retrieved from the graph, are validated against it, and are anchored to authentic legal text through CELLAR and ELIs. Around the engine, we build a regulator-in-the-loop calibration methodology combining a human-reviewed gold benchmark of 52 campaigns, an adversarial adjudicator for disputed citations, and a distillation step that converts confirmed errors into human-approved calibration rules. On the frozen benchmark, the deployed configuration recovered 95 of 97 must-find provisions on first-run scoring, met the minimum risk level in 51 of 52 campaigns, missed no gold-relevant routing, and produced one false positive across seven compliant controls. On a pre-designated 15-campaign holdout, calibration reduced unexpected actionable findings from 44 to 5 (11 of 15 campaigns improved, none worsened; exact Wilcoxon p < 0.001) while preserving 1.000 provision recall and 15/15 risk calibration, and improved run-to-run stability from 0.745 to 0.833. Compared to three external single-call baselines with the same closed provision vocabulary (GPT-5.4, GPT-5.6 and the open-weights gpt-oss-120b), the agentic pipeline achieved significantly higher recall (0.979 vs. 0.608–0.825; exact McNemar p < 0.0001), and no invalid citation was produced in any of the 309 external runs, indicating that citation integrity is a property of the closed-world design rather than any individual model. We release the benchmark, knowledge graph and calibration corpus. The benchmark explicitly covers contemporary multimedia and AI-mediated advertising scenarios, including video and influencer advertising, deepfake endorsements, AI-retouched promotional content, virtual influencers, undisclosed chatbots, and emotion-optimised advertising. Beyond automated legal screening, the framework contributes a trustworthy governance layer for multimedia content by connecting digital-media evidence and campaign representations to grounded, human-reviewable regulatory findings across consumer-protection, data-protection, platform, and AI-specific obligations.