We are seeing examples of crossed-book records where the bid is greater than or equal to the ask. What could be the reason for this?

If you check the timestamps in the data, you'll see that when the order book is mixed up, it quickly gets sorted out at the same moment. This means that their system handles the mixed-up state and then straightens it out a bit later. Most likely, all these changes happen really fast. The system gives updates on the best ask and bid choices that surface during this sorting process at the same time.

Now, when you're dealing with this data, you have a couple of options. You can either [1] ignore these in-between states or [2] only process updates where the timestamp has changed if this is not the state you need. 

The second solution is possible to implement only in historical data processing so if you want to be as close as possible to real-time feed, e.g. simulation, then the first makes more sense.