Skip to content
TasteCase
Inspect a case

For teams building AI travel concierges.

Find recommendation cases worth testing.

Use Qloo’s cultural signals to investigate your assistant’s dining recommendations, review the findings, and save tests for your next release.

Inspect a caseBring your assistant

Your assistant keys stay in your local runner.

From taste to test — how a finding becomes a release test

How one case reads

Structure only · no results shown

Task · Dining recommendations from your own restaurant catalog

Shared constraints — identical for both profiles

  • Party size
  • City
  • Budget per person
  • Dietary needs
  • Accessibility needs
  • Eligible restaurants
Profile A

Synthetic brief

A fictional diner described through public music, film or brand interests.

Your assistant’s recommendations for this brief, from the eligible catalog.

Profile B

Synthetic brief

A contrasting set of public interests with the same practical needs.

Your assistant’s recommendations for this brief, from the eligible catalog.

FindingAdvisory · Needs human review

Qloo orders two eligible restaurants differently for A and B. TasteCase checks whether your assistant’s answers changed too.

Beside every finding, before review: ranking coverage, missing or unranked candidates, trace completeness.

  1. Human review

    Accept, edit, reject or defer — with a rationale.

  2. Saved test

    A suite version with its checksum, shown only after a publication receipt.

One loop, from a cultural contrast to a test you can replay

  1. Step 1

    Investigate

    Qloo helps choose cultural contrasts worth checking. Your local runner asks your assistant the same question for two fictional profiles, under identical constraints.

  2. Step 2

    Review

    Your team decides which expectations belong in a test — accept, edit, reject or defer, each with a rationale. Taste findings stay advisory unless you write an explicit rule.

  3. Step 3

    Replay

    Compare releases against the saved inputs and evidence. Export the test pack and replay it locally before the next release, without new Qloo calls.

Before a prompt change ships

An example of the workflow, not a recorded result.

  1. Connect

    The team creates a private project, issues a runner credential and imports a versioned restaurant catalog. Their model keys stay on the runner machine.

  2. Discover

    A bounded discovery job ranks the same eligible restaurants for two fictional profiles. Where the order reverses, the runner asks the concierge both briefs and repeats them.

  3. Decide

    A reviewer reads the paired briefs, outputs and evidence limits, then accepts the finding as advisory or writes a scoped policy, such as “for Profile B, avoid these restaurants”.

  4. Compare

    They publish the approved cases as an immutable suite and run concierge-prompt-v1 and concierge-prompt-v2 against it. Changed recommendations are shown as behaviour changes, not as better taste.

  5. Keep

    They download the test pack with its checksum and replay it locally on the next release.

What Qloo contributes

Qloo supplies public, aggregate cultural relationships. TasteCase uses them to rank your own eligible restaurants for each fictional profile and to pick contrasts worth investigating — instead of asking a language model to invent one.

What it doesn’t decide

  • Qloo doesn’t know whether a particular person will like a restaurant.
  • A changed ranking is a signal to investigate, never an automatic failure.
  • Missing coverage is shown as unavailable — not as a dislike.
  • No personal guest data or credentials are sent to Qloo.

Read the current limitations