Every metric can be moved by making the product worse — a guardrail is the number that catches you doing it.
A speedometer next to the stopwatch. You're trying to get there faster, but not by driving through a school zone.
Anything you optimise hard enough will be optimised by making the product worse, unless something is watching.
Optimising a single metric reliably finds the cheapest way to move it, which is often not the way you intended. More aggressive notifications raise engagement and raise uninstalls. Auto-renewing subscriptions raise retention and raise refund requests. A guardrail metric is a second measurement, chosen specifically because it would degrade if you moved the primary metric the wrong way, and monitored with a different question: not 'did it improve' but 'did it get meaningfully worse'.
A guardrail is a metric you don't expect to improve but refuse to damage — latency, unsubscribes, refund rate, support contacts, crash rate. It is checked for non-inferiority rather than significance: the question is whether the confidence interval rules out a harmful drop, not whether it moved. Pair one with every headline metric and every experiment.
How Online evaluation metrics and guardrail metrics actually Works in 10 Minutes — How Tech Actually Works, 8:52