MARVEL: A Framework for Verified Experiential Learning in Multi-Agent Autonomous Systems — A Position Paper

The deployment of multi-agent autonomous systems in safety-critical operational environments faces an unresolved tension between three competing requirements: operational autonomy under degraded communication, regulatory auditability of every decision, and continuous capability improvement through experience. Existing architectures address these requirements in isolation but lack a unified framework that combines formal cognitive grounding, cryptographic verifiability of constraints, and verifiable knowledge transfer between agents. We present MARVEL (Mission-Aligned Runtime for Verified Experiential Learning), a five-layer architectural framework integrating: (1) a formal Belief-Desire-Intention (BDI) substrate with stochastic sensor handling, (2) an edge-command computational split with cryptographically signed capability envelopes, (3) doctrine-compatible hierarchical mission decomposition, (4) constrained multi-agent reinforcement learning within signed envelopes, and (5) provenance-preserved federated continual learning enabling auditable knowledge transfer across agent populations. We argue that the combination of these layers — each grounded in formal methods and unified by cryptographic verifiability — addresses gaps identified in recent work on multi-agent provenance, federated reinforcement learning governance, and autonomous command-and-control orchestration. The framework is specifically designed for compatibility with the European Union AI Act regulatory regime for high-risk autonomous systems. This paper establishes the conceptual framework and identifies five technical contributions to be developed in subsequent works. We invite the multi-agent systems, ML safety, and autonomous robotics communities to engage with this framework as a foundation for verifiable autonomy in safety-critical applications.