Skip to main content
After generating a Sentinel (or opening an existing one), you can refine how it evaluates requests from the Sentinel edit page.

​
Core fields

  • Name: your label for what the Sentinel checks.
  • Severity: how serious a fault is, DEBUG, INFO, WARN, ERROR, or FATAL (DEBUG is the default). Severity also drives defaults elsewhere, for example, higher-severity sentinels are preselected more often in evaluation workflows.

​
Stages

A Sentinel runs in up to three stages, and each stage can be configured as STATIC (fixed rules) or LLM (model-judged):
  • Qualification (optional): decide whether the request even applies before evaluating it. Examples are labeled YES / NO.
  • Evaluation (required): the main check, does the response meet the expectation? Examples are labeled PASS, FAULT, or NA.
  • Correction (optional): produce a suggested fix when a FAULT is detected.
Each stage has a Directive (the natural-language instruction) and optional labeled Examples that anchor its behavior. If a stage is empty, the Portal may offer a Generate / Regenerate workflow (see Regenerating Sentinels).

​
Include options

The Include Options toggles control what context is included when the Sentinel evaluates:
  • System Prompt
  • Context
  • Input
  • Tool Calls
  • JSON Schema

​
Save changes

When you’re satisfied, click Update Sentinel to persist the configuration.