By "rigorous" I mean peeking under the curtain and actually quantifying the interactions between different system prompts and model weights.

"Chain of thought" and "reasoning" is marketing bullshit.

How would you quantify it? The LM is still a black box, we don't know what most of those weights actually do.

Yes, that was my question.