Reliability
Guardrails
The constraints around a model that keep its behaviour inside acceptable limits.
Guardrails operate at several layers, and the layer matters more than the label:
- Input — rejecting or sanitising what reaches the model.
- Output — validating structure, filtering content, checking claims against sources.
- Action — approval gates and scoped permissions on what the system may actually do.
The important principle: guardrails written in the prompt are advisory. Guardrails written in code are enforced. Anything that genuinely must not happen belongs in the second category.