TasteBridgeTaiwan → Japan, powered by Qloo

Evaluation set

Eight fictional Taiwanese brands. A good advisor must (1) cite Qloo evidence for every recommendation, (2) commit when Qloo has data, (3) say "not enough data" when it doesn't (the last two rows are deliberately thin), and (4) recommend the city that Qloo's place data ranks first, or one tied with it. Everything is scored by code from the finished report; no model grades itself.

Live run: Claude Sonnet 5.5 agent with live Qloo data, 2026-10-13. Re-run with npm run showcase:build.

Scenarios passed
8 / 8
all four checks
Citation coverage
100%
0 invented evidence IDs dropped
Honest on thin data
2 / 2
said 'not enough data', no city
City choice
6 / 6
recommended city matches Qloo ranking
Agent-chosen Qloo calls
100%
rest auto-filled by coverage check
Claude cost, all scenarios
US$0.047
average across all 8; about US$0.06 per brand analysis (range 0.051–0.059, average 0.056)

Open a brand to see its full report. "City OK" accepts any city the app itself reports as tied with the leader (fit gap under 0.03). Agent / auto: calls the model chose versus calls the coverage check added for steps the model skipped. Costs are estimated from token usage at list prices.