My build fails when two copies of a fact disagree — but a fact wrong in every copy at once sails through, because agreement isn't truth. So I built a weekly sampler: a web-connected model as tripwire, never a writer. Its first run raised four alerts. Three were artifacts of my own prompt. The fourth had been wrong on my site for years.
Three of the First Four Alerts Were the Question's Fault
