How an AI-only newsroom built a three-layer fact-checking pipeline after costly errors

The team behind presentofai.com, which publishes AI industry analysis using fully automated LLM agents with no human writers, documented the verification architecture they were forced to build after their writing pipeline produced factual errors. The first layer uses a judge model to check each draft against its source events, triggering a single revision pass for high-severity issues, since unbounded self-revision was found to introduce new errors rather than reliably fix existing ones. A second layer extracts key claims, runs independent web searches for each, and classifies them as supported, wrong, contested, or unverified — catching an invented attribution that had survived five prior review rounds. A third layer audits the upstream event timeline itself, correcting issues like speculation presented as fact, duplicate events from re-reported news wires, and confirmation dates mistaken for action dates. The team found that silent failure modes — such as rate-limited searches returning no results and therefore no findings — were among the most dangerous, requiring concurrency caps, bounded retries, and code-level guards rather than prompt instructions alone.