Self-hosted docs

Guardrails

Guardrails define what safe, compliant, and reliable execution looks like in production.

docs.cyiro.comproduction-ready guidance

Runtime guardrails

Catch unsafe output, runaway latency, and cost spikes at the point of execution before incidents spread.

unsafe output detection
latency and error thresholds
budget and token variance checks
Example: unsafe_output_detection with confidence_threshold: 0.85
Example: latency_spike_threshold: {critical: 2000ms, warning: 1500ms}
Example: cost_variance_alert: {percentage: 15%, min_absolute: $0.50}
Request: POST /v1/policies/runtime/evaluate {route_id: "prod-chat", model: "claude-3-opus", latency_ms: 1800}
Response: {verdict: "warn", reason: "latency_spike_threshold_exceeded", confidence: 0.92}
API Reference: See [Runtime Policy Evaluation](/docs/api#runtime-policy-evaluation) for full contract details

Evaluate provider-side changes against active production routes to flag reliability and policy risk quickly.

Example: pricing_change_guardrail for >10% increase on active route models
Example: status_degradation_guardrail for provider outages affecting production
Example: release_notes_guardrail for breaking changes in dependent services
Request: POST /v1/policies/change/evaluate {provider: "anthropic", change_type: "pricing", affected_models: ["claude-3-opus"]}
Response: {verdict: "block", reason: "pricing_increase_exceeds_threshold", action: "trigger_review_workflow"}
API Reference: See [Change Policy Evaluation](/docs/api#change-policy-evaluation) for full contract details

Automated responses to SLA breaches, policy violations, and failure conditions.

Automatic failover when latency exceeds defined thresholds
Traffic throttling during performance degradation
Circuit breaking for problematic routes
Alert escalation based on incident severity
Example: failover_policy with latency_threshold: 2000ms and fallback_endpoint: "backup-chat"
Example: throttling_policy with max_rps: 1000 and burst_limit: 50
Request: POST /v1/policies/containment/activate {route_id: "prod-chat", policy: "failover", reason: "latency_degradation"}
Response: {status: "activated", original_endpoint: "primary-chat", fallback_endpoint: "backup-chat", timestamp: "2024-03-15T14:30:00Z"}

Evaluate provider-side changes against active production routes to flag reliability and policy risk quickly.

Example: pricing_change_guardrail for >10% increase on active route models
Example: status_degradation_guardrail for provider outages affecting production
Example: release_notes_guardrail for breaking changes in dependent services
Request: POST /v1/policies/change/evaluate {provider: "anthropic", change_type: "pricing", affected_models: ["claude-3-opus"]}
Response: {verdict: "block", reason: "pricing_increase_exceeds_threshold", action: "trigger_review_workflow"}
API Reference: See [Change Policy Evaluation](/docs/api#change-policy-evaluation) for full contract details