Same employee, same model, same time window: the in-app chat needs 23.8 s in the median, the website widget 9.4 to 11.4 s. Only the in-app chat duplicates text (the verification loop appends the refined text to the draft), shows its reasoning in English during German conversations, and personalises answers with the logged-in user. Test results in the configuration UI therefore do not reflect what guests see.

Request: align the in-app chat with the widget, or offer a test mode that matches the guest channel exactly, with reasoning hidden or in the conversation language.

Impact: tests in the configuration UI that reflect the guest experience.