Suppose you want to determine the accuracy of some device in which directly measuring it through multiple controlled experiments is impractical but you can compare its recordings to parallel values recorded by some different device[s] combined with some known true values (discrete to same precision or perhaps itself within a confidence interval, and importantly not always equivalent between parallel elements) and possibly other aspects yielding significant data.
Here is a real-life example. Two friends and two dogs go on frequent walks together (usually all together, but sometimes with a person or both dogs absent). Each individual has zer own tracker. Each person's is a different type and the dogs' is a third type common to eachother. The persons don't always walk the exact same path, but close enough combined with known true distances to gather patterns from data (one consistently higher, due in part to more-frequent sampling rate). Likewise, the dogs don't travel the exact same distance to either eachother or either person, since they roam a portion off-leash, but the data collected from them forms trends (one dog usually moves greater net distance, and both individually travel greater net distance than either person on nearly every respective walk). So, if you were to make a table of this data, could you crunch the numbers to infer various relationships between the sets? For example, what percentage (or other formula if a non-linear pattern) higher or lower is the true mileage walked than each person's or dog's recorded amounts, within some confidence interval? Or if one of the dog's trackers wasn't working well one time, could you infer (with high certainty) how much net mileage zhe actually traveled on that particular walk within a relatively small range or what the entry the tracker likely would have recorded if it had functioned? In other words, fill in the rest of the table (or the parts sought) with good guestimates based on principals of multivariate statistics.