This paper describes a previously undocumented self-amplifying failure mode in AI governance systems. AI-authored instruction and governance files carry stylistic biases from the model's training data into subsequent sessions via context window pattern-matching. These files simultaneously contain rules suppressing the biased patterns. Because the model pattern-matches on what it reads at higher volume than it follows explicit prohibitions, the governance layer becomes the contamination source and the suppression rule loses reliably across sessions. This paper names this mechanism The Feed Loop, documents its observation across a real production AI governance system (contAIn), and describes the architectural fix: identify and purge the contaminating files rather than strengthen the prohibition. The fix is designed to eliminate the reproductive mechanism by removing the contamination source. This mechanism sits at the intersection of three documented phenomena (training data stylistic bias, LLM instruction-following failure, and context window pattern dominance) without being covered by any of them. I propose The contAIn Method as the diagnostic and remediation framework. The contAIn Method is distinguished from prompt-level interventions: the system had the problem; the method is what the human operator applied to identify and fix it. The fix was not a better rule, it was cleaning the surface the model was reading from. This paper was partly written by the system that had the problem. Make of that what you will.