Success metrics measure what a change hopes to improve; guardrails watch everything it might quietly break: retention, latency, cancellations, support contacts, revenue in adjacent surfaces. The right statistical framing is non-inferiority (“we can rule out harm worse than our pre-declared margin”) rather than “we failed to detect harm,” which an underpowered test achieves by default.
Guardrails also invert the usual multiple-testing logic: for success metrics you control false positives, since you ship on them, but for guardrails the dangerous error is the false negative, a real regression slipping past. A stack of unadjusted guardrails is mostly theater; ten independent checks at 80% power each catch all ten regressions only 11% of the time.