diff --git a/csaxs_bec/devices/vision/reference_images/README.md b/csaxs_bec/devices/vision/reference_images/README.md
new file mode 100644
index 00000000..21f2fd4b
--- /dev/null
+++ b/csaxs_bec/devices/vision/reference_images/README.md
@@ -0,0 +1,20 @@
+# Reference-image test data (local only, not tracked by git)
+
+This directory is the conventional local drop-in location for real-camera reference-image
+captures used to test the vision pipeline (`AI_docs/MACHINE_VISION_AUTOCONFIRM_PLAN.md`) --
+e.g. OMNY's `_otransfer_save_reference_images()` output
+(`sample_transfer_reference_image_<step>_<timestamp>.h5`, one HDF5 file per confirm gate, one
+dataset per camera in the active pair).
+
+**Not committed to git.** This repo's `.gitignore` excludes `**.h5` ("file writer data"), and a
+real capture set is tens of megabytes -- exactly the kind of binary data that rule is for. Only
+this `README.md` is tracked, so the directory (and the convention) exists in git even though its
+actual contents don't.
+
+To populate it locally: drop `.h5` capture files here directly (from a real OMNY session, or
+transferred another way -- e.g. as a packaged archive from a remote session). See
+`../../bec_ipython_client/plugins/omny/AI_docs/VISION_PIPELINE_TESTING_HOWTO.md` for how to use
+real images (there, `DummyVisionCamera.load_image_file()`) against the vision pipeline; loading
+straight from this HDF5 multi-camera-pair format specifically isn't wired up yet as of this
+writing (`load_image_file()` reads a single plain image file via `cv2.imread`, not this format) --
+check the plan doc's implementation-status section for the current state before assuming otherwise.