Sounds interesting!

So far most of my automation has been optical (tweaking display output to show correctly under a lens with odd optical characteristics, hard to simulate computationally), so pointing cameras at the device and letting the LLM iterate and tune display output has been useful.

As with any automation, it all depends on scale. Setting up a harness to repeatedly mess with microSD cards seems way more prone to failure (and false positives) than just doing it yourself, but if you're testing thousands of devices, would probably be worthwhile.