Add a best practice sentence about data models
It should always be possible to use real instances for data only classes.
Specifically mention data models in the best practices section as not needing
mocks.
PiperOrigin-RevId: 560634157
diff --git a/README.md b/README.md
index 74fee18..0189d40 100644
--- a/README.md
+++ b/README.md
@@ -433,13 +433,14 @@
Testing with real objects is preferred over testing with mocks - if you can
construct a real instance for your tests, you should! If there are no calls to
-[`verify`] in your test, it is a strong signal that you may not need mocks at all,
-though it's also OK to use a `Mock` like a stub. When it's not possible to use
-the real object, a tested implementation of a fake is the next best thing - it's
-more likely to behave similarly to the real class than responses stubbed out in
-tests. Finally an object which `extends Fake` using manually overridden methods
-is preferred over an object which `extends Mock` used as either a stub or a
-mock.
+[`verify`] in your test, it is a strong signal that you may not need mocks at
+all, though it's also OK to use a `Mock` like a stub. Data models never need to
+be mocked if they can be constructed with stubbed data. When it's not possible
+to use the real object, a tested implementation of a fake is the next best
+thing; it's more likely to behave similarly to the real class than responses
+stubbed out in tests. Finally an object which `extends Fake` using manually
+overridden methods is preferred over an object which `extends Mock` used as
+either a stub or a mock.
A class which `extends Mock` should _never_ stub out its own responses with
`when` in its constructor or anywhere else. Stubbed responses should be defined
