Guardrails.

the limits and safety measures placed around a model to keep its outputs within bounds. Related to alignment, the broader problem of making AI do what we actually intend.