Every dataset we deliver comes with a table showing how the finished map compares against independently measured checkpoints. Most people read the summary figure at the bottom and stop. The bottom row is the least informative part of the table.
Look at the signs first
Check whether the errors are balanced — roughly as many measuring high as measuring low. If almost all of them lean the same way, that is not random noise but a bias, and a bias does not average out the way noise does. It usually traces back to a setup problem that is worth finding before anyone uses the data.
Then look at the pattern
Plot the errors on a map of the site and look for structure. Errors that grow towards the middle, along one edge, or steadily in one direction each point to a specific and fixable cause. An error that spikes at a single point usually means that one checkpoint deserves a second look before you conclude anything about the survey.
Then, finally, the headline number
Only once the errors are balanced and patternless does the summary figure mean what people assume it means. And when you quote it, quote how many checkpoints stand behind it. An accuracy figure from a handful of points is not a measurement, it is an anecdote.
A reasonable minimum
We will not issue an accuracy statement without a proper set of independent checkpoints, and larger sites get more of them. Below a sensible count, the accuracy figure itself is too uncertain to be worth quoting.