Not necessarily to the former. Definitely to the latter.
To elaborate point 1:
- If the crucial parts of the devices that "were modified to improve accuracy" were modified early on in the study the data could be representative of the final product.
- The same applies for the devices not being fully assembled. If they weren't fully assembled, and the data were still good enough to defend a 510k, then the data should represent worst case usage, unless full assembly would somehow make the technology less performant.