diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index ee6372c57..dc10ea662 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -5,7 +5,7 @@
 # Note: .gitlab-ci.yml itself is NOT in changes: filters — CI-only edits skip build
 #
 # Node pool routing: Runner default nodeSelector pool=ci-compile.
-# Override via KUBERNETES_NODE_SELECTOR_pool: <pool-name>
+# Override via KUBERNETES_NODE_SELECTOR_pool: "pool=<pool-name>" (k=v format required)
 # Pools: ci-compile (L4 GPU), ci-training (L40S GPU), services (DEV1-M), gitlab (GP1-XS)
 
 # Ensure pipeline is always created — individual job rules handle filtering.
@@ -462,8 +462,8 @@ build-training:
   variables:
     KUBERNETES_RUNTIME_CLASS_NAME: nvidia
     # Route to L40S pool (48GB VRAM for large supervised models)
-    # Label: pool=ci-training (simple label, overrides runner default pool=ci-compile)
-    KUBERNETES_NODE_SELECTOR_pool: ci-training
+    # Overrides runner default pool=ci-compile
+    KUBERNETES_NODE_SELECTOR_pool: "pool=ci-training"
     # Supervised is GPU-bound — minimal CPU. Low request allows 3-4 concurrent jobs.
     KUBERNETES_CPU_REQUEST: "1000m"
     KUBERNETES_CPU_LIMIT: "2000m"
@@ -1252,8 +1252,8 @@ deploy:
     - kapsule
   variables:
     # Deploy only needs kubectl — route to services pool, not GPU nodes
-    # Label: pool=services (overrides runner default pool=ci-compile)
-    KUBERNETES_NODE_SELECTOR_pool: services
+    # Overrides runner default pool=ci-compile
+    KUBERNETES_NODE_SELECTOR_pool: "pool=services"
     KUBERNETES_CPU_REQUEST: "200m"
     KUBERNETES_CPU_LIMIT: "500m"
     KUBERNETES_MEMORY_REQUEST: "256Mi"