harishbalasubramanian/sentinel-eval: Deterministic ACL leak benchmark for AI agent memory

Sentinel-eval is a deterministic benchmark and linter demonstrating that derived-context access control lists leak confidential executive data in agent memory. The evaluation shows that while vector stores successfully filter raw RAG queries, background synthesis jobs strip parent provenance and drop ACLs. To prevent unauthorized retrieval, the tool implements lineage tracking, bitset closures, and explicit policy primitives. Open article →