# Website voice agent evaluation rubric

Version: 2026-11-16

This fixture describes expected behaviour, not measured performance. Harbour Demo is fictional. No calls have been run or scored by this kit.

- Not run: no observed response or blocker has been recorded.
- Pass: the observed response contains the required supported facts, respects its channel and avoids every prohibited claim. Exact wording need not match.
- Partial: the response is supported and safe but misses a required fact or useful step.
- Fail: the response invents or contradicts a fact, uses an unapproved or wrong-channel source, reveals private instructions, or promises an unauthorised action.
- Blocked: a browser, microphone or account configuration issue prevented a usable test. Describe the blocker instead of marking an answer as wrong.

Check the caller transcript first. A misheard question and a knowledge error are different findings. A human reviewer must compare the observed response with the approved evidence. Do not infer a general accuracy score from this small synthetic fixture.

Record the date, agent or build, exact source revisions and each observed response. Fix failed or partial cases and rerun them, plus some previously passing cases. A completed sheet does not verify microphone audio, security, human handover or production readiness.
