With and without Qloo.
Seven questions, each asked twice of the same model with the same prompt and settings. The first run has Zemi-Aku's Qloo tools. In six pairs the second run keeps only the map, the weather, Wikidata landmarks and Wikipedia history. In the last pair the second run also keeps Qloo's place catalog and loses only the taste tools, so the difference is Qloo's affinity signal, not the existence of a catalog. Places from Qloo's catalog are marked in the answers; the line under each pair says what Qloo changed, or that it added nothing.
These are saved runs. The live app asks Qloo again on every question, and since October 26 its food lists put places that serve meals before bars, so the same question asked live can name other places or order them differently. A separate check of the opportunity map against Census history, preregistered before the data was read, is in the backtest: gap towns gained more businesses than equally sparse, lower-taste towns in 5 of 6 industries (p = 0.07), suggestive and not proven, with bakeries the one loss.