Since the EHT is an interferometer, the observations contain less information than what would be available from a single telescope of the same equivalent size. Therefore to reconstruct the 'full image' you need to find the model which best fits your interferometric observables. As far as I can tell, what they've done with PRIMO is basically a fancy version this kind of modelling. The data isn't necessarily being guided towards the simulations, it's more like we have a better computational technique to precisely fit the data.

As I had to explain to the execs at a lab on a chip startup I was helping to put together, you're generating the proposition of what the data might be that still needs to be qualified with physical testing, and not actual data.

This concept seems to be worryingly lost in the flurry of excitement at using ML/AI in academic research by some.

This is well-said. We're looking for what the best guess should be, but we still need to test the guess.