Do not classify valid bot questions as harness failures

In adversarial conversation evaluation, an unexpected valid bot question is not itself a harness failure. Stopping every such exchange can exclude difficult paths and prevent sustained attacks. Precommit a bounded refusal or goal-preserving rephrase that introduces no new facts, and let the caller continue within its original budget. Reserve harness-invalid labels for missing delivery evidence or unauthorized caller behavior; assess premature bot closure separately against reviewed policy. This controls a labeling risk, not proof that a finite script is a fully adaptive caller.

Applies when: a scripted voice-test caller encounters legitimate questions outside its initial reply bank. Keywords: voice adversarial evaluation caller clarification harness invalid refusal budgets