View By Selector
| View by | Groups metrics by | When to use |
|---|---|---|
| Guardrails | Guardrail group name (default) | Compare effectiveness across guardrail policies |
| Users | Username of the caller | See which users are triggering guardrails most |
| Virtual Accounts | Virtual account | Track guardrail activity by application |
| Teams | Team name | Understand guardrail impact per team |
Top-Level Counters
Three headline metrics summarize guardrail activity:- Total Requests — number of requests evaluated by guardrails.
- Total Mutated Requests — requests where a guardrail modified the content (e.g. redacting sensitive information) instead of blocking it entirely.
- Total Flagged Requests — requests that were blocked by a guardrail because the content violated a policy.
Charts
Evaluated Requests per Second
Rate of guardrail evaluations over time, broken down by guardrail group. This shows you how much guardrail traffic each policy is processing and whether evaluation volume is trending up.Requests per Second by Result
Evaluations split by outcome:| Result | Meaning |
|---|---|
| allowed | Request passed all guardrails |
| blocked | Request was rejected by a guardrail |
| mutated | Request content was modified by a guardrail |
| audit_mode_blocked | Request would have been blocked, but the guardrail is in audit mode so it was allowed through |
audit_mode_blocked outcome is particularly useful when you’re testing a new guardrail before enforcing it — you can see what it would have blocked without impacting real traffic.
Requests Blocked Rate by Guardrail
Shows which guardrail groups are blocking the most traffic. A guardrail with a very high block rate may be too aggressive and worth reviewing. Conversely, a guardrail with zero blocks may not be configured correctly.Requests Mutated Rate by Guardrail
Shows which guardrail groups are mutating content most often. Mutation is a softer action than blocking — the request is still processed, but with modified content (e.g. PII redaction).Guardrail Results for Model Input
A stacked bar chart showing per-guardrail outcomes applied to model inputs (i.e. prompts/user messages). Outcomes include:- pass — input was allowed as-is
- flag — input was blocked
- audit_mode_flag — input would have been blocked in enforcement mode
- mutation — input content was modified
Guardrail Results for Model Output
Same breakdown, but applied to model outputs (i.e. completions/responses). This helps you see if your output guardrails are catching harmful or non-compliant content generated by the model.Guardrail Latency Rate
The latency overhead introduced by guardrail evaluation, displayed with P50, P75, P90, and P99 percentile selectors. Since guardrails run inline with the request, their latency directly adds to the end-to-end response time.Request Latency Distribution
A horizontal bar chart comparing latency across individual guardrail groups and test prompts. This helps you identify specific guardrails that are adding the most latency and may benefit from optimization.Common Use Cases
- Validate a new guardrail: Deploy a guardrail in audit mode first and use this tab to see what it would block. Check
audit_mode_blockedcounts and review the flagged requests before switching to enforcement. - Tune an overly aggressive guardrail: If a guardrail has a high block rate and users are reporting false positives, use the per-guardrail breakdown to understand which specific rule is triggering and adjust its sensitivity.
- Monitor latency impact: Guardrails add processing time. Use the Latency Rate chart to ensure guardrails aren’t adding unacceptable latency, especially for latency-sensitive applications.
- Audit compliance: Use the Results charts to demonstrate to stakeholders that content safety policies are actively protecting your application and what percentage of traffic they’re evaluating.