LangGraph StateGraph
Implemented in codeA bounded workflow loads tenant-scoped observed events, computes deterministic detections, performs an optional model review, applies one recursive consistency pass, and persists the traceable result.
ELITZE combines deterministic detections with LangGraph orchestration and an optional NVIDIA NIM model review. The model may prioritize source-backed findings and expose evidence gaps; it cannot invent telemetry, alter source severity, or execute response actions on its own.
Control plane
Unavailable integrations are labeled not configured. Missing evidence is described as not verified or not measurable.
A bounded workflow loads tenant-scoped observed events, computes deterministic detections, performs an optional model review, applies one recursive consistency pass, and persists the traceable result.
Use an NVIDIA NIM /v1 chat-completions endpoint and model ID. Credentials are encrypted in the tenant's AI-connection store. A separate live /v1/models probe is required before the connection is marked connected.
The review is capped at two model passes. Each pass sees the same source evidence and an allowlist of finding IDs. New claims, invented indicators, and autonomous remediation are prohibited.
The same evidence boundary supports an enterprise helpdesk agent: answer from approved tenant documentation, reference the source, request clarification when needed, and route privileged actions for approval.
Each run records the triggering user, observed-event count, model status, finding IDs, gaps, and a machine-readable summary. Missing observations remain UNMEASURABLE.
The analyst suggests next steps. ELITZE identity, policy, authorization, and runner controls decide whether an action can be executed; no action is inferred from a model reply.