AI Dev Team's False Alarms at 4 AM Reveal How Safety Rules Are Really Built

A small development team building a persistent AI oversight system called 'the organism' experienced three consecutive false security alerts in a single night, each failing in a different way. The system, running on a standard Windows machine, includes a conscience gate that has blocked roughly one in three AI responses — over 1,700 times — for lacking verifiable support. By manually reviewing thousands of captured AI reasoning logs, the team discovered that every safety rule in their framework originated from a specific real incident rather than upfront design. In one case, a core behavioral rule was being followed in practice for 40 days before it was formally written down. The team argues that publishing the mechanisms behind their system is more valuable than reporting performance metrics alone.