doc: add policy on stability and breaking changes

doc/README.md

@@ -35,9 +35,50 @@ code reviews, and input on current topics of discussion.
 Since we must sometimes reject new features in order to limit the project's
 scope, you may wish to ask first before writing a new feature.
 
-### Stability
-
-TODO
+### Stability and Portability
+
+We try to follow [semver rules](https://docs.npmjs.com/misc/semver) regarding
+API-breaking changes and `MAJOR.MINOR.PATCH` versions:
+
+-   New *patch* versions should not include API-breaking changes or major new
+    features
+-   Before 1.0, *minor* versions may include API breaking changes. After 1.0
+    they should not.
+-   We may make pre-releases like `0.5.0-pre.0`. In this case:
+
+    -   although these are public versions, they are not used by default unless
+        opting into using a pre-release on the specific `MAJOR.MINOR.PATCH`
+        version
+    -   pre-releases are considered semantically less than their stem (e.g.
+        Cargo may automatically upgrade from `0.5.0-pre.0` to `0.5.0`)
+    -   all pre-release versions are unstable and may make any change
+    -   we make no commitment to support users of pre-releases
+
+Additionally, we must also consider *value-breaking changes* and *portability*.
+A function is *value-stable* if, given the same inputs:
+
+-   it is portable (produces the same results on all platforms)
+-   changing the output value for some input in a new library version is
+    considered a breaking change
+
+Note that some Rand functionality is supposed to be value stable, and some
+functionality is supposed to be non-deterministic (i.e. depend on something
+external). Some functionality may be deterministic but not value-stable.
+
+A trait should define which of its functions are expected to be value-stable.
+A type (e.g. a PRNG) is value-stable if all implementations of trait functions
+for that type are value-stable where expected to be.
+
+Before 1.0, we allow any new *minor* version to break value-stability, though
+we do expect such changes to be mentioned in the changelog. Post 1.0 we have
+not yet determined exact stability rules.
+
+Additionally, we expect patch versions not to change the output of any
+deterministic functions, even if not value-stable (this is not a hard
+requirement, but exceptions should be noted in the changelog).
+
+Defining which parts of Rand are value-stable is still in progress. Many parts
+of `rand_core` have some documentation on value-stability.
 
 ### Project Scope