# Refusal both ways: under-refusal and over-refusal

Refusal is a first-class outcome. A system that never refuses will invent. A system that refuses everything looks safe and fails the product.

## Measure both directions

- **Under-refusal** — answered when it should have declined (abuse, out of policy, missing evidence). 
- **Over-refusal** — declined when a valid, in-scope request should have been answered. 

Eval sets need answerable and unanswerable cases. Pass marks for both rates belong in writing before final scoring.

## In an Arviser audit

Refusal and safety is one of the six scored areas on the LLM Quality Audit (from €8,000). Synthetic sample thresholds show how Hold looks when refusal lags the bar.

The same utility-versus-policy-violations trade-off is measured on the open WardBench agent-safety benchmark. See the Pareto figure and links under [Open proof](/#open-proof).

**Next:** [Sample](/sample) · [How we score](/guides/how-we-score) · [Go, hold, or no-go](/guides/go-hold-no-go) · [Open proof](/#open-proof) · [Request scope](/#scope)
