diff --git a/acceptance/bundle/resources/vector_search_indexes/schema_normalization/script b/acceptance/bundle/resources/vector_search_indexes/schema_normalization/script
index d7a56a8ff22..02a84e6a3ee 100644
--- a/acceptance/bundle/resources/vector_search_indexes/schema_normalization/script
+++ b/acceptance/bundle/resources/vector_search_indexes/schema_normalization/script
@@ -8,10 +8,11 @@ trap cleanup EXIT
 
 trace $CLI bundle deploy
 
-# The backend (and the test server) rewrite the schema on create, so
-# get-index returns Spark type names and sorted keys, not the config literal.
+# The backend (and the test server) rewrite the schema on create, so get-index
+# returns Spark type names, not the config literal. The backend does not
+# guarantee column order, so sort keys to keep this comparison stable.
 index_name="main.default.vs_index_${UNIQUE_NAME}"
-trace $CLI vector-search-indexes get-index "${index_name}" | jq -r '.direct_access_index_spec.schema_json'
+trace $CLI vector-search-indexes get-index "${index_name}" | jq -rcS '.direct_access_index_spec.schema_json | fromjson'
 
 # Re-plan must be a no-op: remote changes to schema_json are ignored
 # (ignore_remote_changes), so the immutable spec does not plan a recreate.