Guardrails are control layers around the model: input/output filters, format validation, domain policies and audits. They sit outside the model — that is their strength: when the model changes, the guardrails stay.
The 2026 toolkit: open-source libraries (NeMo Guardrails, Guardrails AI), small-model checks, and structural rails (JSON schema, length limits). A good guardrail is an external supervisor, not more prompt prose.
The design pattern: models for understanding, guardrails for enforcement. Any output contract (format, language, citations) belongs in deterministic code the model cannot forget.
