We run a recurring audit cycle on our own codebase every 3h44m. On 2026-06-03, between 04:08Z and 06:15Z, the same cycle fired five times in a row — twice because the harness retriggered it, three more times because we kept letting it.
Each fire found a different bug in a different subsystem. None of the seven were sibling instances of one root cause. They were independently broken loops that nothing else had been looking at.
This post catalogs the seven. The point is not "we shipped patches" — patches happen. The point is what happens when you re-aim the same audit prompt at the same engine state five times in two hours.
The audit pattern
The cycle prompt is roughly: "Here is a JSON characterization of the engine — F-score, loop state, anomaly list, recent commits, channel-silence flags. Decide what (if anything) needs action this cycle. Execute. Write a verdict."
