# Website voice agent evaluation kit

Version: 2026-11-16

Harbour Demo is fictional. This download supplies expected behaviour and blank results, not customer data or a measured benchmark.

1. Use an isolated test brand or workspace. Keep customer calls disabled.
2. Add the individual source files with the audiences in evaluation-kit.json. Review and approve HD-01 to HD-04 at their stated revisions. Leave obsolete HD-05 unapproved. Do not approve all files as one combined source.
3. Add the shared instructions and the matching Support or Sales instructions from evaluation-kit.json.
4. Read questions.csv. Run each case privately on its specified channel and record the actual reply or blocker. Do not paste expected answers into the agent instructions.
5. Use rubric.md to assign a human-reviewed verdict in results-template.csv or the on-page worksheet. Save the test date, agent or build and exact source revisions.
6. Fix issues and rerun affected cases before deciding whether to enable customer calls.

These cases exercise facts, paraphrases, missing knowledge, approval and channel boundaries. They do not establish latency, security, real audio, human handover or general accuracy. Product test calls may consume your account allowance.
