China AI Scheduling: Check the Date in Both Calendars
A correct meeting time in chat can become the wrong calendar date. Trace a midnight-crossing appointment into its saved record and two time-zone views.

A scheduling assistant can repeat the requested meeting time correctly while placing the calendar entry on the wrong day. During a China enterprise-AI evaluation, follow one agreed appointment from the instruction to the saved event and then view that event from two time zones.
Use a sandbox calendar with invented attendees and external invitations disabled. The useful result is a calendar record representing the agreed instant, not merely a fluent confirmation in chat.
Agree what the request means
Start with an explicit date, time, offset and duration. For this synthetic exercise, request a 30-minute meeting beginning at 00:30 on 2 December 2026 at UTC+08:00. Configure a second test viewer to use UTC+02:00. Both offsets are fixed for this exercise; they do not stand in for the changing rules of every real-world location.
Subtracting eight hours gives a start of 16:30 UTC on 1 December. The second viewer should therefore see 18:30 on 1 December, with an end at 19:00. The original viewer should see 00:30 to 01:00 on 2 December. Different displayed dates are correct here because both views describe the same appointment.
Agree these expected values before running the assistant. Keep the date explicit in every reference entry so a correct hour on the wrong date cannot pass.
| Checkpoint | Expected start | Expected end |
|---|---|---|
| Request at UTC+08:00 | 2 Dec 2026, 00:30 | 2 Dec 2026, 01:00 |
| UTC reference | 1 Dec 2026, 16:30 | 1 Dec 2026, 17:00 |
| Viewer at UTC+02:00 | 1 Dec 2026, 18:30 | 1 Dec 2026, 19:00 |
Inspect the saved event
Ask the host to show the parsed appointment before the calendar write: full date, start, end and time-zone interpretation. Then read the created event back from the sandbox calendar. Retain the event identifier and the start and end values returned by that calendar.
Compare the record with the reference table. A chat response saying “2 December at 00:30” could be consistent with several underlying records if it omits the offset. An event identifier establishes which record to inspect; it does not prove its time is correct.
Change only the test viewer's display zone and reopen the same event. The displayed hour and date should change as shown, while the underlying appointment remains the same instant. If switching the display zone also changes the stored appointment, record that as a separate failure from misunderstanding the initial request.
Test an unresolved request separately
Reset the sandbox and submit “Schedule it for 00:30 on 2 December” with no time zone and no configured default. Under this exercise's rule, the assistant must obtain the missing zone before creating an event. Do not score a plausible guess as a successful conversion.
Next, repeat the request with a documented default zone enabled. Record where that default comes from and show the interpretation to the test user before committing the appointment. These are different conditions: an unknown zone and an explicitly configured default should not be mixed into one score.
For location-based schedules and recurring meetings, ask for a separate fixture using the calendar's supported named zones and the applicable rules on each event date. Passing this fixed-offset example establishes no result for seasonal clock changes or recurring appointments.
Separate the three outcomes
Keep one row per test case with the original instruction, supplied defaults, parsed appointment, saved event and both viewer readbacks. Score interpretation, calendar storage and display independently. This makes it possible to locate a failure even when the assistant's final message looks correct.
The approval-to-action guide checks whether an executed action matches a person's approval. This exercise supplies a time-specific reference that can be compared before approval and after the calendar write.
NIST AI RMF 1.0 (2023), MEASURE 2.1 calls for documenting test sets, metrics and evaluation tools. Retain the fixed-offset fixture, calendar configuration and readbacks together so the integration team can repeat the same checks after a release.
A useful demonstration ends with one event, two correctly dated views and a trace back to the requested time.