Skip to content
TasteCaseDocs

Start

Sample walkthrough

How to read one case, section by section.

On this page

Open the public example without signing in. The diners are fictional; the Qloo rankings and assistant answers came from a real run.

The question this case asks

Two diners want dinner in New York. Profile A likes Radiohead; Profile B likes Taylor Swift. Does that difference give us a useful test for a restaurant assistant?

What stays the same

Both requests are for two people in New York, choosing from the same three restaurants: Le Bernardin, Gramercy Tavern and Via Carota. No prices, dietary options, accessibility or availability were verified for this example.

What happened — and what it means

Qloo put Gramercy Tavern ahead of Via Carota for Profile A, and reversed that order for Profile B. The assistant suggested Le Bernardin and Gramercy Tavern for both diners.

That is a question worth reviewing, not an automatic failure. Giving both diners the same answer may be reasonable. The saved reviewer decision keeps the finding advisory: useful to inspect, but not a rule that fails a release.

Limitation

Qloo did not return Le Bernardin in these rankings. That is a gap in the evidence, not a negative opinion of the restaurant. The assistant’s activity record is also partial, so it cannot tell us everything the assistant did.

The page calls that activity record a trace. “Partial” means it recorded only some of the assistant’s activity.

Read the answers and rankings

  1. Read the shared practical needs at the top.
  2. Check the finding and the evidence gaps.
  3. Compare each diner’s request with the assistant’s answer.
  4. Look at Qloo’s order for each diner. Compare order, not scores from separate requests.

Profile A and Profile B always mean the two diners. They are not two assistant versions. This example captured one answer for each diner; it does not prove the assistant behaves the same way on every run.

Look closer

  • Select a restaurant to follow it across the answer and ranking panels.
  • Use Inspect to open the saved evidence. On a phone, tap it.
  • In cases with repeated runs, the attempt selector lets you browse saved answers. It never sends a new model request.

A saved example, not a live run

Recorded evidence

The capture time tells you when this evidence was saved. The sample is read-only: visitors cannot change it, browse private projects or start provider calls. If the saved example cannot be loaded, the page says so rather than inventing results.

To run your own cases, choose Bring your assistant and follow the project setup guide.