COMPASS: The explainable agentic framework for Sovereignty, Sustainability, Compliance, and Ethics
New multi-agent system uses RAG and LLM-as-a-judge to score AI decisions across four critical dimensions in real-time.
A research team led by Jean-Sébastien Dessureault has published a paper on arXiv introducing the COMPASS Framework (Compliance and Orchestration for Multi-dimensional Principles in Autonomous Systems with Sovereignty). This novel architecture directly addresses the critical gaps left by existing, siloed approaches to AI governance. It is designed as a multi-agent orchestration system where a central Orchestrator manages four specialized sub-agents, each responsible for a core imperative: digital sovereignty, environmental sustainability (carbon-aware computing), regulatory compliance, and ethical alignment. This modular design allows the framework to systematically integrate these often-competing values into an autonomous agent's decision-making process.
Each specialized agent within COMPASS is augmented with Retrieval-Augmented Generation (RAG), grounding its assessments in verified, context-specific documents like legal statutes or corporate policies, rather than relying solely on an LLM's parametric knowledge. The system then uses an LLM-as-a-judge methodology to evaluate proposed actions, assigning quantitative scores and, crucially, generating explainable justifications for each dimension. This enables real-time arbitration when objectives conflict—for instance, balancing a compliance requirement against an ethical principle. The authors validated the framework through automated evaluation, demonstrating that the RAG integration significantly enhances the semantic coherence of assessments and mitigates the risk of hallucinations, a common pitfall in AI reasoning.
The paper argues that the framework's composition-based, extensible design facilitates seamless integration into diverse application domains, from financial services to healthcare, while preserving the interpretability and traceability necessary for audit and trust. By providing a unified, explainable architecture for value-aligned AI, COMPASS represents a significant step beyond point solutions, offering a blueprint for building autonomous systems that are not only capable but also accountable, sustainable, and aligned with human-defined principles.
- Unified multi-agent framework with four specialized modules (Sovereignty, Sustainability, Compliance, Ethics) managed by a central Orchestrator.
- Uses RAG (Retrieval-Augmented Generation) to ground agent evaluations in verified documents, reducing hallucinations and improving coherence.
- Employs LLM-as-a-judge to assign quantitative scores and generate explainable justifications for real-time arbitration of conflicting goals.
Why It Matters
Provides a practical, unified architecture for building accountable and trustworthy AI agents that can navigate complex real-world regulations and ethics.