From taste to test
Find a case, decide what to check and reuse it next time.
On this page
A case becomes a test when you decide what the assistant should do and save that expectation.
The simple version
Connect your assistant, find an interesting case, review the answers, save the useful tests, then use them to compare your next version.
1. Find a case
You supply the restaurant list and practical needs. TasteCase uses public interests for two fictional diners. Qloo helps find a pair of restaurants that changes order between those diners, and your local runner asks your assistant about both.
A runner is the small program that connects TasteCase to your assistant. Your model-provider key stays on the machine running it.
2. Decide what to keep
Read the answers and the evidence gaps. Accept the finding as something to keep reviewing, edit it into a specific check, reject it, or leave it for later. Write a short reason for your decision.
For example, “only recommend restaurants from our list” is a check TasteCase can evaluate. “This music fan should enjoy this restaurant” is not a fact TasteCase can prove. Approving an interesting case is not the same as passing a test.
3. Test the next version
Publish your approved cases as a suite — a named test collection with a saved version. Compare Baseline (your current assistant) with Candidate (the version you want to try). Both answer the same saved requests.
You can download a test pack and rerun it on your own machine. That uses the saved inputs without asking Qloo for fresh rankings.
Understand the result labels
Passed and Failed apply to checks the system can evaluate. Advisory asks for your judgement. Unavailable means there is not enough evidence or a run did not finish — it never means passed. See what each label means.