TrigGuard
TRIGGUARD COMPARE

TrigGuard vs LlamaGuard (and classifier layers)

Taxonomies like LlamaGuard score model output. TrigGuard issues a deterministic PERMIT / DENY / SILENCE on tool and API execution before side effects.

LlamaGuard (and similar input/output classifiers) answer "is this text in a disallowed category?" They are part of a strong model safety program. They do not, by themselves, answer "may this specific HTTP call move money or mutate production?"

TrigGuard sits on the execution path: policy version, surface id, and request context are evaluated; only PERMIT authorizes the downstream client. Decisions are recorded in Ed25519-signed receipts you verify offline.

Use both where appropriate: classifiers for generation policy; TrigGuard for irreversible actuation.

Compare: guardrails vs execution authorization (general)

Blog: runtime authorization vs guardrails

Runtime authorization for AI agents

Decision model (PERMIT / DENY / SILENCE)