// guardrails
Rules for AI.
Enforced automatically.
LLMs do not always follow instructions. liwaiwai's guardrail layer evaluates every single response against your policies and flags the moment something diverges.
Configure guardrailsForbid specific words, phrases, or patterns. Any response containing flagged terms is logged as a deviation immediately.
Require JSON, markdown tables, or structured schemas. Non-conforming responses are flagged, regardless of content quality.
Lock a system instruction and verify that every model response adheres to its intent not just its literal wording.
Embedding-based similarity checks surface responses that appear compliant on the surface but deviate in meaning.
// data sovereignty
All guardrail evaluations are performed on your data, within your jurisdiction. liwaiwai does not store prompt content on third-party infrastructure. Designed for governments, regulated institutions, and organisations where data sovereignty is a legal obligation not a preference.