A guardrail metric must not degrade even when optimizing for a primary metric. It prevents local optimization that harms other critical outcomes.

Example: if your primary metric is click-through rate, a guardrail might be user satisfaction score or revenue per user - to prevent click-bait tactics that inflate CTR while harming experience. In A/B tests, an experiment can "win" on the primary metric but fail if it violates a guardrail. Guardrails also include latency, error rate, and crash rate in product launches.