diff --git a/RELEASE.md b/RELEASE.md
new file mode 100644
index 0000000..c94739d
--- /dev/null
+++ b/RELEASE.md
@@ -0,0 +1,50 @@
+# RELEASE.md (trunklet-specific notes)
+
+See [`../ai/RELEASE.md`](https://github.com/Postgres-Extensions/ai/blob/main/RELEASE.md)
+for the org's shared release process. It does **not** apply verbatim here:
+this repo vendors a much older, subtree-vendored copy of `pgxntool`
+(squashed in via `git subtree`, not the submodule other repos use) whose
+`make tag`/`make dist` predate several assumptions that doc makes.
+
+## `make tag` creates a git branch, not a tag
+
+In this repo's vendored `pgxntool/base.mk`, `tag` is:
+
+```
+tag:
+	git branch $(PGXNVERSION)
+	git push --set-upstream origin $(PGXNVERSION)
+```
+
+This is a real, current behavior here, not just outdated wording in
+`pgxntool/README.asc` (which also still describes it this way — the org
+doc's "Notes / gotchas" section anticipates exactly this situation). Every
+past release (`0.0.1` through `0.3.3`) is a branch on `origin`, matching the
+version number exactly. **There are no git tags in this repo at all** —
+`git tag -l` and `git ls-remote --tags` are both empty. `rmtag`/`forcetag`
+operate on branches the same way. `make dist` still works as documented
+upstream (`git archive` of that ref), since `git archive` accepts a branch
+name just as well as a tag.
+
+Do not expect `git tag <version>` to find anything, and do not "fix" this
+by hand-creating a real tag — that would silently diverge from every prior
+release's convention in this repo.
+
+## No `stable` pseudo-version
+
+This repo's vintage of `pgxntool` predates the `stable` pseudo-version
+mechanism described in `../ai/RELEASE.md`. `default_version` in
+`trunklet.control` and `version` in `META.in.json` are always a real
+released version number (currently `0.3.3`), never the literal string
+`stable`. Skip every step in the shared doc that references `stable`
+(the versioned-install-script/`sql/<ext>--*--stable.sql` workflow, and
+step 8's "return master to stable").
+
+## No `verify-results` target
+
+This vendored `pgxntool` predates `verify-results` too — it only has
+`results` (regenerates expected test output) and `installcheck`, which is
+marked `.IGNORE` in `base.mk`. A green `make test`/`make installcheck` here
+does **not** guarantee `pg_regress` actually passed — inspect
+`test/regression.diffs` directly, or diff `results/` output by hand, rather
+than trusting the Make exit code.