In a Plone translation catalogue shipping today there is a string that asserts something its source does not — placeholder counts match, length is normal, every automated check passes, and only a human reading it would notice. If translations ever have to carry legal weight — and in the EU, safety information must, in up to 23 official languages — that is exactly the class of bug that matters. For more than a year I have run Nick and Volto in production as the backend and frontend of a regulated-documentation platform: 358,627 translations across 71 languages in one database-backed model — merged into Nick through the release team in June 2026 — guarded by a meaning-drift detector that structural QC cannot replace, with the AI kept on local hardware as a written architectural rule, not a preference. This talk is what a year of running that stack for real taught me, including the failures I paid for — and it ends with a question I want the room's answer to: how an AI-assisted corpus should live alongside Weblate.