I run a content pipeline that finds industry news, has an LLM judge score it, and publishes what clears the bar. Fully automated, Cloudflare Workers and D1, one operator: me. It runs daily and the whole point is that I don't babysit it.
In July it published a news story from 2024 as if it had just happened.
Nobody was harmed. One item, niche site, caught on a routine quality check within a few days, unpublished. But "the robot posted two-year-old news and nothing stopped it" is the kind of failure that quietly costs a site its credibility, and it deserved a real postmortem instead of a shrug. What went wrong turned out to be more interesting than I expected, and the fix ended up being three layers instead of the one I planned to write.
What actually went wrong
The pipeline already had two mechanisms that should have caught this.
