Skip to main content

COMPEL Glossary / guardrail-metric

Guardrail metric

A secondary metric that must not degrade — typically beyond a declared tolerance — even if the primary metric improves.

What this means in practice

Guardrail metrics protect against Goodhart effects where optimizing the primary metric erodes a downstream interest (latency, error rate, fairness, revenue). Distinct from `Guardrails` (LLM safety-layer controls); this is an experimentation concept.

Synonyms

guardrail KPI , guardrail indicator , protected metric

See also

  • Primary metric — The single metric whose movement drives an experiment's ship/no-ship decision.
  • Hypothesis — A falsifiable statement an AI experiment is designed to support or reject — specifying an intervention, an expected direction of effect, a primary metric, and an acceptable effect size.