> The model genuinely believes it’s giving a correct reasoning chain

The model doesn't "genuinely believe" anything.