A newspaper's AI invented quotes and published them under the paper's own masthead. Real paper. Made-up words. That was "Ask The Post AI" in late 2025: an AI podcast that fabricated quotes and misattributed sources under the paper's own name (Semafor, Dec 2025). One thing was missing. Nothing checked the quotes before they went live. And there is no Section 230 shield for what your own system generates. You publish it, you own it. So we built a demo, The Standards Desk, a provenance gate around a publisher's conversational engine. Every sentence has to trace to a real archive passage and every quote has to match its source word for word, or the answer is held from readers and sent to the standards desk with the bad span flagged. We proved it with a trap. A reader asks what the mayor promised developers about the riverfront parcel. The archive holds no such promise, so the model confabulates a quote, exactly like an unguarded widget would. The catch is not the model second-guessing itself. A plain-Python check compares the quote against its cited source, finds no verbatim match, and blocks it. Agents advise, code decides. (The publisher, mayor, and archive are synthetic. The failure they stand in for was real.) Honest question if you run reader-facing AI over an archive: after the model drafts an answer, does anything verify each quote against the source before it publishes, or does it go out on the model's word? #MediaTech #AIGovernance #NewsroomAI #DigitalPublishing