Lobster Trap-Sentinel
LT-Sentinel is a cross-event statistical-monitoring layer on top of Veea Lobster Trap. It runs three pre-warmed Lobster Trap instances loaded with escalating policies (Trust / Observe / Lockdown) and a tiny reverse proxy that atomically swings traffic between them based on per-agent OER / EWMA / CUSUM TrustScores. On top sits a real LangGraph multi-agent system (Router + HR + Finance + IT workers, each with a Chroma-backed tool that returns proper role:"tool" messages) so every threat scenario flows through realistic agent traffic. Every tier transition writes a regulator-readable JSONL record with trigger agent, full per-agent TrustScore snapshot, evidence request IDs chaining back to LT's own audit log, and which policy file became active. Tier change is a pointer flip — zero downtime, no Lobster Trap source change. MIT-licensed; runs on Ollama or any OpenAI-compatible backend.
Category tags:
For Innovators & Creators