ARCHER: Multi-agent AI cuts building compliance costs by 75%
Open-source AI checks building regulations with 97.8% of frontier accuracy at quarter cost.
Verifying building compliance against regulatory codes is notoriously labor-intensive and unscalable, with existing Automated Compliance Checkers (ACCs) being brittle, proprietary, and hard to generalize. In a new arXiv paper, researchers present ARCHER — Agentic Rule and Compliance Harness for Executable Regulations — a test-driven, deterministic multi-agent system that automatically generates auditable verification code from natural-language Codes of Practice. The system orchestrates multiple AI agents to iteratively refine programmatic rules, making compliance checking both transparent and adaptable to different rule sets and building designs.
ARCHER was evaluated across a taxonomy of six harnesses with four backbone models, spanning frontier APIs to fully on-premise open-weights models. The deterministic orchestration achieved the highest accuracy for every backbone, with a mean union accuracy improvement of 82% over a naive single-pass baseline. Most notably, a self-hosted open-weights model using the ARCHER harness reached 97.8% of frontier-API accuracy at only a quarter of the cost, making data-sovereign AI compliance checking practical for the first time. This opens the door for organizations to run compliance checks locally without sacrificing accuracy.
- ARCHER improves mean union accuracy by 82% over naive single-pass prompting baseline.
- A self-hosted open-weights model reaches 97.8% of frontier-API accuracy at 1/4 the cost.
- System uses deterministic multi-agent orchestration to generate auditable code from regulatory text.
Why It Matters
Makes building compliance checking transparent, auditable, and cost-effective, enabling organizations to run AI-driven checks on-premise.