Toward Deterministic Governance of Large Language Model APIs

This preprint presents IAGA, a control-plane architecture for governing Large Language Model (LLM) API usage in enterprise environments. The system introduces deterministic cost enforcement, bounded budget guarantees, circuit-breaker-based reliability, synchronous response validation, and cryptographically scoped multi-tenant isolation. IAGA operates as an OpenAI-compatible gateway, requiring zero application-level changes while enforcing governance at request time. The paper details the full request lifecycle, cost containment model, caching architecture, reliability mechanisms, and controlled performance evaluation under simulated enterprise workloads.