I conducted the following iterative loop an obscene number of times.
The scope of what we were working on, the technical issues, and even sometimes the people issues was mind-blowing. In hindsight, we were very fortunate to have the amount of access to users as we did. At one point, I was bouncing prototypes off an actual user nearly once a day!
The bulk of our testing was remote, with the occasional in-person testing. Sometimes our testing was done through a project management team inside a partner company's Client Services, and sometimes I was able to just sit down directly with our target group and do some testing.
I tested everything from whiteboard sketches on the spot, low-fidelity prototypes, high-fidelity prototypes, and even the occasional low road code project. Sometimes, the data was so complex that we had to code a small iteration, far below MVP, to make sure we were on the right track.
A lot of our asks were coming from a Client Services team with a partner company. Testing to make sure that we were solving the right problems and not a symptom of the problem was paramount. For example, when the requirements for the Comparison Matrix came across my JIRA status board, I said "Wait. What?" This requirement meant that they wanted to see 131 Hospital columns and up to 16,000 (!!!) rows on a screen at once. The number of times myself or the developers said, "There is no way they want this. They can't even use that much data at once, the browser might even blow up." is astounding. I distinctly remember asking "Well, why don't we just ask them." So, that's what we did.
It turns out, they wanted a version of that. I had initially, based on the original requirements mocked up a quick very low fidelity prototype to test. To our surprise, it was actually pretty close to what they wanted, but it was far less usable than they expected. A key piece of information is, the VA doesn't actually know how complicated their VistA system is. They had very little idea that the matrix would be nearly 43 feet long had we printed the matrix at a 1 inch data point scale.