Skip to main content

Playground and Monitor

Playground​

The Playground tab dry-runs guardrail steps against a sample user message. It doesn't call the target model.

  • Choose a Policy (each shows its step count), or All configured guardrails to run every enabled guardrail
  • Pick an example (Jailbreak + SSN, Jailbreak, PII (SSN), Secret key, Clean) or type a User message
  • Run dry-eval

The verdict is Allowed, Allowed with changes (matches redacted), Blocked (followed by the name of the step that blocked) or Nothing evaluated (the policy has no steps). Every step is listed with its result: Passed, Redacted, Blocked here, Failed, Skipped, Not reached or No result, with a reason.

Steps are skipped when the guardrail is turned off, checks responses only (POST), or is a custom script. Rule cards evaluate locally. LLM critic and vendor (OpenAI Moderation / Azure Content Safety) cards run a live call when a credential is configured. Custom Starlark scripts execute only on the gateway.

To test a response, or a policy with unsaved changes, use Try it in the policy builder.

Monitor​

The Monitor tab shows the Guardrails monitor for the last 30 days. To pick a range, use the Guardrails view in FinOps Analytics, which shows the same monitor. On /guardrails, set it in the URL: ?tab=monitor&window=24h|7d|30d|90d, or ?tab=monitor&from=YYYY-MM-DD&to=YYYY-MM-DD. Non-admins see only their own requests.

  • Total evaluations, Blocked requests, Pass rate, Latency added, Active guardrails
  • Request outcomes over time: blocked and modified requests per day; All outcomes adds allowed requests
  • Where blocks come from: block decisions by guardrail
  • Guardrail performance: per guardrail, its provider, action, requests, fail rate, latency added and status. Each row opens Request Logs filtered to that guardrail.
  • Manage guardrails opens the Guardrails tab
  • Export data downloads CSV (/usage/export?…&tab=guardrails)

Pair Monitor with Request Logs filtered by outcome=blocked (or a specific guardrailId / policyId) when you need the raw prompt.