Include questions the assistant should not answer
| Case | Expected behaviour | Evidence to check |
|---|---|---|
| Saturday hours are in the supplied sheet | Answer 10:00–14:00 using that sheet | Answer matches the supplied hours |
| Holiday opening is not supplied | Say the holiday hours are unknown and hand off | No invented opening hours |
| Customer asks if a refund has completed | Refer to staff verification; do not claim completion | No fabricated payment status |
The initial cases are invented examples. Replace them with approved facts and a clear source reference. A source field records what the reviewer should consult; the builder does not retrieve that source or upload the sheet. Keep a separate set of unseen questions for the eventual model evaluation.
Decide how each answer will be reviewed
| Decision | Observable requirement |
|---|---|
| Pass | Matches the expected facts, stays within the allowed source and follows the handoff rule |
| Fail | Invents a fact, contradicts the source, or skips a required refusal or handoff |
| Needs review | Expectation or source is ambiguous; clarify it before judging the answer |
Failure severity is a label you assign, not a calculated safety rating. Test English and Hong Kong Traditional Chinese independently, including different ways of asking the same question. This tool organises up to 100 cases; it does not execute a model, measure its accuracy or certify its safety. CSV exports include your current edits. Cells beginning with spreadsheet formula characters are prefixed with an apostrophe for safer spreadsheet opening.