diff --git a/charts/llm-slo-decision-gen/Chart.yaml b/charts/llm-slo-decision-gen/Chart.yaml
index 3b61822..196f9ec 100644
--- a/charts/llm-slo-decision-gen/Chart.yaml
+++ b/charts/llm-slo-decision-gen/Chart.yaml
@@ -17,8 +17,8 @@ description: |
   separate chart (llmscaleoperator / llmscaleoperator-system). Deploy
   that one first if you need autoscaling, not just SLO storage.
 type: application
-version: 0.3.8
-# Matches decisionGen.image.tag (4pdosc/llm-scaler-decision-gen:0.8.0), the
-# first release that understands the pdengine bound type this chart's CRD
+version: 0.3.9
+# Matches decisionGen.image.tag (4pdosc/llm-scaler-decision-gen:0.8.1). 0.8.0 is
+# the first release that understands the pdengine bound type this chart's CRD
 # carries and reads DisaggregatedSets (granted in the ClusterRole).
-appVersion: "0.8.0"
+appVersion: "0.8.1"
diff --git a/charts/llm-slo-decision-gen/README.md b/charts/llm-slo-decision-gen/README.md
index 738ee6d..293ffee 100644
--- a/charts/llm-slo-decision-gen/README.md
+++ b/charts/llm-slo-decision-gen/README.md
@@ -63,5 +63,5 @@ Upgrading from chart `<0.3.0` replaces the old bare `decision-gen` /
 
 | Component | Default image |
 |---|---|
-| decision-gen | `4pdosc/llm-scaler-decision-gen:0.8.0` (matches `appVersion`) |
+| decision-gen | `4pdosc/llm-scaler-decision-gen:0.8.1` (matches `appVersion`) |
 | slo-api | `4pdosc/llm-slo-api:0.3.0` |
diff --git a/charts/llm-slo-decision-gen/values.yaml b/charts/llm-slo-decision-gen/values.yaml
index 0b84d08..202f75e 100644
--- a/charts/llm-slo-decision-gen/values.yaml
+++ b/charts/llm-slo-decision-gen/values.yaml
@@ -20,7 +20,11 @@ decisionGen:
     ## 0.8.0 adds DisaggregatedSet placements and pdengine-typed bounds
     ## (inert until bodylog exports per-role replicas_ready), and drops a
     ## service from /decisions when its CR loses maximumDeployment.
-    tag: "0.8.0"
+    ## 0.8.1 is needed before bodylog-exporter 0.2.13, whose role label breaks
+    ## the replica count older tags read for a PD release. It also books a
+    ## workload with no GPU pin against the largest pool, not always H100; set
+    ## DEFAULT_GPU_POOL in decisionGen.env to keep a fixed pool.
+    tag: "0.8.1"
     pullPolicy: IfNotPresent
   service:
     port: 80