A big flaw with most current models is that their learning curves for tech are either completely ignored, or are way too small to match empirical reality. Here's a popular news article about research using empirical modeling of the learning curves, which says that the models are overestimating the cost:
https://arstechnica.com/science/2021/10/the-decreasing-cost-...
And here's a podcast episode about it that goes into a lot more detail:
https://xenetwork.org/ets/episodes/episode-159-the-cost-of-d...
However, if you're not willing to take the other link in a sibling comment because it comes from an investor, I'm not sure what you consider to be a neutral source.
I would say that what's actually wild are the "accepted" models from the IEA, which have been catastrophically wrong year after year because they encapsulate "common sense" of the industry instead of data based methods:
https://cleantechnica.com/2017/09/06/iea-gets-hilariously-sl...