Interface: GuardrailCheck

Defined in: packages/sdk/src/backend/agents.ts:29

One guard's config (input or output). All optional with safe defaults.

Properties

checks?

optional checks?: string[]

Defined in: packages/sdk/src/backend/agents.ts:34

input: relevance|safety|prompt_injection|needs_clarification; output: grounding|citations|safety|relevance|leakage

---

enabled?

optional enabled?: boolean

Defined in: packages/sdk/src/backend/agents.ts:30

---

failOpen?

optional failOpen?: boolean

Defined in: packages/sdk/src/backend/agents.ts:32

Default true — a guardrail outage must not hard-fail the turn.

---

riskThreshold?

optional riskThreshold?: number

Defined in: packages/sdk/src/backend/agents.ts:37

Output guard only (M5 enterprise): when >0, a safe-but-risky answer (verify risk ≥ this) is held for human review — escalated, answer kept. 0..1.