COALITION-VAST: Auditable Multi-Agent Alignment Under Byzantine Governance

Abstract

Scaling aligned AI from single-agent systems to multi-agent ecosystems introduces collective failures that do not arise in isolation: coalition deviation, governance capture, and rushed rule changes. Prior work in VAST and VAST-Blockchain addresses single-agent compliance and deployment integrity, but not strategic coordination across multiple agents. We introduce COALITION-VAST, a framework for multi-agent alignment with auditable governance under Byzantine validators and Sybil governance attempts. COALITION-VAST models coordination as a coalition game coupled with on-chain rule evolution, keeps agents bound to locked machine-checkable constraints, and adds trust-driven targeted auditing so detection becomes an adaptive outcome rather than a fixed assumption. We derive conditions under which adaptive detection and enforceable penalties make profitable coalition deviation unattractive, and we show that with BFT consensus and supermajority rule updates, Byzantine validators alone cannot finalize unauthorized constraint changes. A prototype simulation across healthcare resource allocation, autonomous swarms, and multi-stakeholder finance shows improved alignment (mean 0.90 vs. 0.69; +30% relative), a 98% reduction in undetected coalition deviation, and modest latency (2.8-3.4 s per coordinated decision). We also present an evolutionary analysis showing that compliance remains stable when expected audit penalty exceeds deviation gain, and we introduce a governance gate for highstakes updates based on evidence commitment, delay, and independent review.