diff --git a/client/tests/visual_capture.gd b/client/tests/visual_capture.gd
index 5cf3b56e7..8bcb08e88 100644
--- a/client/tests/visual_capture.gd
+++ b/client/tests/visual_capture.gd
@@ -313,6 +313,13 @@ func _run_flow(_main_node: Node) -> void:
 ## layers) all come from server/data/systems.db via the running server, same
 ## as every other Atlas path. Requires _run()'s live-mode server-connection
 ## wait to have already completed before this is called.
+##
+## GOLDEN PROVENANCE (T-1121): the 12 curated atlas_* goldens promoted from
+## this matrix are machine-local — captured/verified exact-match on one box,
+## and goldens are known NOT to port across GPU/font stacks at the current
+## tolerance. A golden failing on a different machine is expected drift, not
+## a regression, until T-1121 settles the canonical capture box. See the
+## _comment_atlas_goldens note in tests/visual.json.
 func _run_atlas_matrix() -> void:
 	var matrix_cfg := _load_atlas_shots_config()
 	if matrix_cfg.is_empty():
diff --git a/tests/visual.json b/tests/visual.json
index d9a965c50..58303bc9d 100644
--- a/tests/visual.json
+++ b/tests/visual.json
@@ -90,6 +90,7 @@
       "env": { "SR_AUTOPILOT": "stance_up,south:6.0" },
       "description": "Live 3D sandbox: autopilot walks south into a wall — cutaway stub golden (T-1088)"
     },
+    "_comment_atlas_goldens": "T-1120 provenance (PR #180 review, Tyre): the 12 atlas_* goldens below were captured AND verified (exact pixel match) on danoontje's Mesa/AMD stack. T-1121 documents that goldens do NOT port across boxes at the current tolerance (the 19 pre-existing goldens fail 12-14% on this same box) — an atlas_* golden failing on a DIFFERENT machine is expected baseline drift, not a regression, until T-1121 resolves the canonical-capture-box question.",
     "atlas_GJ380c_zfit_C": {
       "ticks": 240,
       "live": true,