co-Author here. We built a fraud detection system and realized the core problem — evaluating multi-signal policies against historical data in real-time — is structurally a query planning problem. This paper describes the compilation framework: declarative policies → dependency DAG → optimized SQL with CTE chains, history gating for cold-start entities, and plan caching for reuse. Applicable beyond fraud to any domain needing sub-100ms multi-signal decisions over relational data. Happy to discuss the architecture decisions and tradeoffs.
co-Author here. We built a fraud detection system and realized the core problem — evaluating multi-signal policies against historical data in real-time — is structurally a query planning problem. This paper describes the compilation framework: declarative policies → dependency DAG → optimized SQL with CTE chains, history gating for cold-start entities, and plan caching for reuse. Applicable beyond fraud to any domain needing sub-100ms multi-signal decisions over relational data. Happy to discuss the architecture decisions and tradeoffs.