diff --git a/infra/k8s/argo/alpha-perception-template.yaml b/infra/k8s/argo/alpha-perception-template.yaml
index 1c346bb8a..524c4dbbd 100644
--- a/infra/k8s/argo/alpha-perception-template.yaml
+++ b/infra/k8s/argo/alpha-perception-template.yaml
@@ -81,29 +81,111 @@ spec:
   templates:
     # ── DAG ──
     #
-    # ensure-binary (CPU compile) and warmup-gpu (autoscaler trigger
-    # for the L40S pool) run in parallel. train only depends on
-    # ensure-binary — warmup-gpu is fire-and-forget; the autoscaler
-    # keeps the node warm for its scaledown grace window after the
-    # warmup pod exits, so train lands on a hot node without waiting
-    # for autoscaler provisioning.
+    # check-cache (tiny alpine pod on platform pool, ~3 sec) inspects
+    # the training-data PVC for /data/bin/$SHA/alpha_train. If present
+    # (cache hit), ensure-binary is SKIPPED via `when:` — the
+    # ~4.8GB ci-builder image pull + sccache compile cycle are avoided
+    # entirely. train still runs, sourcing the binary path from
+    # check-cache's SHA output.
+    #
+    # warmup-gpu fires unconditionally in parallel — triggers the L40S
+    # autoscaler so the node is warm by the time train needs it,
+    # whether or not we compile.
     - name: pipeline
       dag:
         tasks:
+          - name: check-cache
+            template: check-cache
           - name: ensure-binary
             template: ensure-binary
+            dependencies: [check-cache]
+            when: "{{tasks.check-cache.outputs.parameters.cache}} == miss"
+            arguments:
+              parameters:
+                - name: sha
+                  value: "{{tasks.check-cache.outputs.parameters.sha}}"
           - name: warmup-gpu
             template: warmup-gpu
           - name: train
             template: train
-            dependencies: [ensure-binary]
+            dependencies: [check-cache, ensure-binary]
             arguments:
               parameters:
                 - name: sha
-                  value: "{{tasks.ensure-binary.outputs.parameters.sha}}"
+                  value: "{{tasks.check-cache.outputs.parameters.sha}}"
+
+    # ── check-cache: probe the training-data PVC for an existing binary ──
+    #
+    # Runs on the platform pool (always up, no autoscaler delay). Tiny
+    # alpine pod, ~3 sec end-to-end. Emits two output parameters:
+    #   - sha   = the short SHA used for binary cache keying
+    #   - cache = "hit" if /data/bin/$SHA/alpha_train exists else "miss"
+    #
+    # Resolves `commit-sha=HEAD` by reading /tmp/head-sha from a
+    # ConfigMap... actually no, simpler: the submission script
+    # (scripts/argo-alpha-perception.sh) pre-resolves HEAD via local
+    # git so commit-sha is always a real SHA in the workflow. This
+    # pod just trims to short SHA and stats the binary.
+    - name: check-cache
+      outputs:
+        parameters:
+          - name: sha
+            valueFrom:
+              path: /tmp/sha
+          - name: cache
+            valueFrom:
+              path: /tmp/cache
+      nodeSelector:
+        k8s.scaleway.com/pool-name: platform
+        topology.kubernetes.io/zone: fr-par-2
+      tolerations:
+        - key: node.cilium.io/agent-not-ready
+          operator: Exists
+          effect: NoSchedule
+      container:
+        image: alpine:3.20
+        command: ["/bin/sh", "-c"]
+        resources:
+          requests:
+            cpu: "50m"
+            memory: 32Mi
+          limits:
+            cpu: "100m"
+            memory: 64Mi
+        volumeMounts:
+          - name: training-data
+            mountPath: /data
+            readOnly: true
+        args:
+          - |
+            set -e
+            SHA="{{workflow.parameters.commit-sha}}"
+            if [ "$SHA" = "HEAD" ]; then
+              echo "ERROR: commit-sha cannot be HEAD inside the workflow."
+              echo "       Resolve via the submission script (scripts/argo-alpha-perception.sh)."
+              exit 1
+            fi
+            SHORT_SHA=$(echo "$SHA" | cut -c1-9)
+            echo "$SHORT_SHA" > /tmp/sha
+
+            BIN="/data/bin/$SHORT_SHA/alpha_train"
+            if [ -x "$BIN" ]; then
+              SIZE=$(stat -c %s "$BIN")
+              echo "Cache HIT: $BIN ($SIZE bytes)"
+              echo "hit" > /tmp/cache
+            else
+              echo "Cache MISS: $BIN not present, ensure-binary will compile"
+              ls -lh "/data/bin/" 2>/dev/null | head -10 || echo "  (no /data/bin/ directory yet)"
+              echo "miss" > /tmp/cache
+            fi
 
     # ── ensure-binary: compile `alpha_train` example via sccache ──
+    # Only runs when check-cache returned cache=miss. Outputs the SHA
+    # for consistency, but the DAG sources SHA from check-cache directly.
     - name: ensure-binary
+      inputs:
+        parameters:
+          - name: sha
       outputs:
         parameters:
           - name: sha
diff --git a/scripts/argo-alpha-perception.sh b/scripts/argo-alpha-perception.sh
index cd1205083..489dcc1b5 100755
--- a/scripts/argo-alpha-perception.sh
+++ b/scripts/argo-alpha-perception.sh
@@ -69,13 +69,29 @@ done
 case "$GPU_POOL" in
   ci-training-l40s)
     SM=89 ;;
-  ci-training-h100*|ci-training-h100x2|ci-training-h100-sxm)
+  ci-training-h100)
     SM=90 ;;
   *)
     echo "Unknown gpu-pool: $GPU_POOL"
     exit 1 ;;
 esac
 
+# Resolve commit-sha=HEAD to an actual SHA locally before submission.
+# The in-cluster check-cache pod uses this SHA to look up the binary
+# cache without needing git inside the alpine pod.
+if [ "$SHA" = "HEAD" ]; then
+  echo "Resolving HEAD for branch $BRANCH..."
+  if ! git rev-parse --git-dir >/dev/null 2>&1; then
+    echo "ERROR: not in a git repo; cannot resolve HEAD"
+    exit 1
+  fi
+  git fetch --quiet origin "$BRANCH" 2>/dev/null || true
+  SHA=$(git rev-parse "origin/$BRANCH" 2>/dev/null \
+        || git rev-parse "$BRANCH" 2>/dev/null \
+        || git rev-parse HEAD)
+  echo "  resolved: $SHA"
+fi
+
 echo "Submitting alpha-perception workflow..."
 echo "  branch:           $BRANCH"
 echo "  sha:              $SHA"