SR
Fair point on tokenisation hiding typos, though it does not quite apply to Jev itself since it is not generating text token by token; it returns a fixed schema field per call. Where this could actually bite us is one step earlier: we generate the screen description with a separate LLM (Bonsai) before Jev ever sees it, so tokenisation weirdness there could bleed into Jev's classification. We did not isolate for that in this experiment; it's worth checking against the "others" misses specifically. Do you have a specific case in mind, or a tool you use for that raw character vs token view comparison?