I have no clue at all. There's no actual documentation on it, and the code is fucking soup.
I tried variations of it, but it never actually seems to make any factual difference apart from lowering or raising both model and clip weight togheter (setting them both to 0.8 or 0.2 acts exactly how A1111 does Lora weights, but making them different values does something I can never predict nor understand so far, it becomes wildly random results).
And that's a problem with ComfyUI, either small details like this one are never explained anywhere at all, or terms like "weight", "clip", "noise", "steps", etc, have different meanings and logic behind them than how A1111 does it, for example.
IMO, it's a standards issue. All of the clients, webuis, etc, all of them follow no actual standard. So people have to spend days tweaking and testing things around on any of them to catch glimpses of what something does and how it does it, and only then produce something actually worth the time invested. That's both good and bad in equal measures.