I help make complex systems easier to trust.

The local evaluation contains 20 questions: 14 answerable, four simulated access denials and two asking for facts absent from the snapshots. Eighteen questions received model responses and two were refused locally without model inference. All 18 responses passed structural citation checks. Strict whole-response grounding was 15 of 18 because three correct refusals added advice outside the supplied context. Only the implementing AI agent made those judgments. The original responses and retrieval miss are preserved. The new evaluation is local and does not establish production identity integration, end-to-end service testing or ongoing AI usage.

61 deterministic security and retrieval evaluations ran across eight documented threat classes and retrieval behavior without model calls. The recorded CI matrix covered three supported combinations of Qdrant or pgvector backends with native or LangGraph runtimes. Per-session and per-IP spend caps bound model cost. The demo uses a role switcher and simulated side-effecting tools; real-world effects cannot execute through those simulated tools. A document outside the selected role's access is never a candidate for the answer. In a separate attack test, poisoned content induced a simulated email action addressed to an attacker; the action gate blocked it before execution.