diff --git a/README.md b/README.md
index 92b4a13..945fb40 100644
--- a/README.md
+++ b/README.md
@@ -10,6 +10,7 @@ SELECT queries using a dictionary.  In addition, the package provides a number o
  * Statistics tables for storing statistics and counts (this is particularly useful in the LMFDB's context since the data changes rarely)
 
 The query language is specified in [QueryLanguage.md](QueryLanguage.md) and the read API (`search`, `lucky`, `count`, `random`, …) in [Searching.md](Searching.md).
+What the version number promises — the public API, database metadata compatibility, and the deprecation policy — is laid out in [Versioning.md](Versioning.md).
 
 # Install
 
diff --git a/Versioning.md b/Versioning.md
new file mode 100644
index 0000000..db7c4eb
--- /dev/null
+++ b/Versioning.md
@@ -0,0 +1,63 @@
+# Versioning and API stability
+
+Starting with 1.0.0, psycodict follows [semantic versioning](https://semver.org):
+breaking changes to the public API happen only at major releases, new
+functionality arrives in minor releases, and patch releases contain only fixes.
+This document says what "public API" means for a package whose surface includes
+not just Python names but also a query language, an on-disk export format, and
+metadata tables living inside your database.
+
+## What is public
+
+ * **Non-underscore names** in the `psycodict` package that are documented — in
+   the specification documents ([QueryLanguage.md](QueryLanguage.md),
+   [Searching.md](Searching.md), [DataManagement.md](DataManagement.md),
+   [MetadataFormats.md](MetadataFormats.md)) or in docstrings.  Names with a
+   leading underscore are private, whatever module they live in, and may change
+   in any release.
+ * **The query language** as specified in [QueryLanguage.md](QueryLanguage.md):
+   the meaning of a query dictionary is stable within a major version.
+ * **The re-exported SQL composition classes** (`from psycodict import SQL,
+   Identifier, ...`).  Downstream code should import these from `psycodict`
+   rather than from the driver; the re-export point is the stable name.
+ * **The export file format** written by `copy_to` and read by `copy_from` /
+   `reload` (three header lines, `|` delimiter, `\N` nulls — see
+   [DataManagement.md](DataManagement.md)): files written by one 1.x release
+   can be loaded by any other.
+ * **The `meta_*` tables**, whose layout is governed by the metadata format
+   protocol below.
+
+## What is not covered
+
+Underscore-prefixed names; the exact SQL text psycodict emits (only its
+semantics); performance characteristics; the contents of log files; and
+undocumented behavior generally, even where observable.  If something
+undocumented matters to your project, open an issue — turning it into
+documented (hence stable) behavior is usually easy.
+
+## Database metadata compatibility
+
+The layout of the `meta_*` tables is versioned by the **metadata format**
+number stored in each database (`meta_format`, with a `min_compat` column
+declaring the oldest client format the database still admits); the protocol —
+including how clients degrade gracefully against older databases and when a
+migration is required — is specified in
+[MetadataFormats.md](MetadataFormats.md).  The format number is bumped only at
+major releases, so within 1.x a database migrated once is understood by every
+client.
+
+## Deprecation policy
+
+Where feasible, behavior slated for removal first spends at least one minor
+release emitting a `DeprecationWarning` naming the replacement.  (The test
+suite promotes psycodict's own deprecation warnings to errors, so deprecated
+paths cannot linger inside the package itself.)  Removals then happen at the
+next major release.
+
+## Python and PostgreSQL support
+
+The supported Python floor is declared in `pyproject.toml`
+(`requires-python`); the supported PostgreSQL range is the one exercised in CI.
+Dropping an interpreter or server version that has reached upstream end-of-life
+is not considered a breaking change and may happen in a minor release — never
+in a patch release.