Software that lasts tends to be built from small parts that each do one thing well. A translator is not a monolith; it is a pipeline. Text comes in, is segmented into meaningful units, routed to a model, and rendered back beside the original so the reader never loses context.
When you replace the original text outright, you trade away the ability to verify. Bilingual rendering keeps both the source and the translation in view, which matters most precisely when the stakes are high: legal clauses, medical instructions, technical specifications.
The best interface for trust is the one that lets you check the work without leaving the page.
No single model wins every sentence. A cheap model may nail casual prose but stumble on dense academic syntax; a frontier model may be overkill for a menu. Letting the user compare outputs, latency, and cost on their own text—rather than trusting a vendor's benchmark—turns a guess into a decision.
This is the wager behind Prism: give people the spectrum, and let them choose the light.