A guardrail metric is a metric you watch alongside the success metric and agree not to let worsen, so that moving the target cannot quietly cause harm or reward gaming. If a city’s target is average bus wait time, wait time by district is a guardrail: the average can improve while the routes serving the poorest neighborhoods get worse. The term is used in online controlled experiments (A/B tests), where a change is judged on its primary metric and a regression in a guardrail metric counts against shipping it.
In FDE interviews
Guardrails belong in a decomposition answer when you say how success will be measured, and in behavioral answers about a project that worked technically and still failed in the organization. A guardrail without a threshold is only a second chart, so a strong candidate names the success metric, the guardrail, the limit agreed before launch and who would game the target without it: “We’d report average wait, with wait by district beside it, and agree up front that no district may get worse than today; if one does, we pause the rollout, whatever the average says. Operations could otherwise hit the average by moving buses off the quiet routes.” In an AI deployment the same pairing is containment rate as the target and reopened or escalated tickets as the guardrail. The pairing is our method.
The lesson Stakeholders and success metrics teaches how to pick a metric the sponsor already trusts, add a guardrail and name who could game it.