The most useful question in AI governance is a simple one: when an error occurs, does correction outpace its spread, or does its spread outpace correction? Correction wins, and the system cleans itself — mistakes happen, then fade. Spread wins, and errors compound: people adopt them, records store them, retrieval repeats them, and the system settles into a contaminated equilibrium that no individual correction fixes.
The threshold reframes what governance is for. The goal is not zero errors — no ceiling allows that — but a self-limiting regime: correction capacity that outpaces spread. And the threshold is a property of the whole arrangement, which is why system-side levers (verification before writes, reviewer capacity, provenance labels, loop cuts) can move a deployment across it when model improvements alone cannot.
The threshold also explains why deployments surprise their owners. Growth — more users, more automation, more record coupling — adds propagation paths. A system that was comfortably self-limiting at pilot scale can cross into self-sustaining territory simply by succeeding, with no change to the model at all.
The PAN Lab's central gauge is exactly this: watch which side of the threshold a scenario sits on, and which combinations of governance choices move it — in both directions.
This page is conceptual framing, a way of seeing, not an empirical claim. Documented real-world events appear in the Domain Atlas with citations; testable versions of these ideas live in the PAN Lab.