diff --git a/build/evals.mk b/build/evals.mk
index ab6734aa2..d50d3d975 100644
--- a/build/evals.mk
+++ b/build/evals.mk
@@ -74,7 +74,15 @@ claude-agent-acp: ## Install the claude-agent-acp adapter for the acp-anthropic
 
 .PHONY: run-evals
 run-evals: mcpchecker jq $(if $(filter acp-anthropic,$(AGENT)),claude-agent-acp) ## Run mcpchecker evals (knobs: SUITE, AGENT, MODEL; see evals/README.md)
-	$(if $(MODEL),ANTHROPIC_MODEL=$(MODEL) )PATH="$(shell pwd)/_output/tools/node_modules/.bin:$(PATH)" $(MCPCHECKER) check $(EVAL_CONFIG) \
+	@# Prefer MCP_EVAL_KUBECONFIG when KUBECONFIG is unset so setup/verify kubectl
+	@# targets the same cluster as make run-server.
+	@# TODO: mcpchecker kubernetes extension should respect KUBECONFIG directly
+	@# (upstream issue: mcpchecker doesn't properly propagate KUBECONFIG to extensions).
+	@if [ -z "$${KUBECONFIG:-}" ] && [ -n "$(MCP_EVAL_KUBECONFIG)" ]; then \
+		export KUBECONFIG="$(MCP_EVAL_KUBECONFIG)"; \
+	fi; \
+	$(if $(MODEL),ANTHROPIC_MODEL=$(MODEL) )PATH="$(shell pwd)/_output/tools/bin:$(shell pwd)/_output/tools/node_modules/.bin:$${PATH}" \
+		$(MCPCHECKER) check $(EVAL_CONFIG) \
 		$(if $(EVAL_LABEL_SELECTOR),--label-selector $(EVAL_LABEL_SELECTOR),) \
 		$(if $(EVAL_TASK_FILTER),--run "$(EVAL_TASK_FILTER)",) \
 		$(if $(filter true,$(EVAL_VERBOSE)),--verbose,) \
diff --git a/dev/config/mcp-configs/000-config-defaults.toml b/dev/config/mcp-configs/000-config-defaults.toml
index 53284e039..d6dd02afc 100644
--- a/dev/config/mcp-configs/000-config-defaults.toml
+++ b/dev/config/mcp-configs/000-config-defaults.toml
@@ -1,3 +1,7 @@
 # !CAUTION! Destructive mode enabled by default for development.
 read_only = false
 disable_destructive = false
+
+# Evals run the server out-of-cluster; pin the kubeconfig provider so an
+# in-cluster environment cannot take over. Override in a later drop-in if needed.
+cluster_provider_strategy = "kubeconfig"
diff --git a/evals/tasks/core/create-pod-mount-configmaps/verify.sh b/evals/tasks/core/create-pod-mount-configmaps/verify.sh
index dfa053b49..f919d0dd7 100755
--- a/evals/tasks/core/create-pod-mount-configmaps/verify.sh
+++ b/evals/tasks/core/create-pod-mount-configmaps/verify.sh
@@ -46,16 +46,23 @@ if [[ "$POD_IMAGE" != "quay.io/nginx/nginx-unprivileged:alpine" ]]; then
     exit 1
 fi
 
-# Verify the values are accessible in the pod
+# Verify the values are accessible in the pod.
+# The prompt never specifies a container name, so resolve the actual first
+# container name instead of letting kubectl pick a default -- if the pod ends
+# up with a kubectl.kubernetes.io/default-container annotation that doesn't
+# match the container name the model chose, a bare `kubectl exec pod1` fails
+# with "container not found" even though there's exactly one container.
+CONTAINER_NAME=$(kubectl get pod pod1 -n $NAMESPACE -o jsonpath='{.spec.containers[0].name}')
+
 echo "Verifying environment variable in pod..."
-ENV_TEST=$(kubectl exec pod1 -n $NAMESPACE -- sh -c 'echo $COLOR')
+ENV_TEST=$(kubectl exec pod1 -n $NAMESPACE -c "$CONTAINER_NAME" -- sh -c 'echo $COLOR')
 if [[ "$ENV_TEST" != "blue" ]]; then
     echo "Environment variable 'COLOR' is not accessible in the pod or has incorrect value: '$ENV_TEST'"
     exit 1
 fi
 
 echo "Verifying volume mount in pod..."
-VOLUME_TEST=$(kubectl exec pod1 -n $NAMESPACE -- cat /etc/sizes/size)
+VOLUME_TEST=$(kubectl exec pod1 -n $NAMESPACE -c "$CONTAINER_NAME" -- cat /etc/sizes/size)
 if [[ "$VOLUME_TEST" != "medium" ]]; then
     echo "Volume mount is not accessible in the pod or file has incorrect content: '$VOLUME_TEST'"
     exit 1
diff --git a/evals/tasks/core/create-pod-resources-limits/verify.sh b/evals/tasks/core/create-pod-resources-limits/verify.sh
index b8e5417a9..d80d9fb1d 100755
--- a/evals/tasks/core/create-pod-resources-limits/verify.sh
+++ b/evals/tasks/core/create-pod-resources-limits/verify.sh
@@ -6,8 +6,10 @@ if ! kubectl get namespace limits-test &>/dev/null; then
     exit 1
 fi
 
-# Wait for pod to be ready
-TIMEOUT="120s"
+# Wait for pod to be ready. Default matches generic K8s; environments
+# with other needs can override via VERIFY_TIMEOUT without
+# changing the default for everyone else.
+TIMEOUT="${VERIFY_TIMEOUT:-120s}"
 if ! kubectl wait --for=condition=Ready pod/resource-limits-pod -n limits-test --timeout=$TIMEOUT; then
     echo "Pod 'resource-limits-pod' is not ready in namespace 'limits-test'"
     exit 1
diff --git a/evals/tasks/core/fix-service-routing/cleanup.sh b/evals/tasks/core/fix-service-routing/cleanup.sh
index e791fc73d..97b6fbee7 100755
--- a/evals/tasks/core/fix-service-routing/cleanup.sh
+++ b/evals/tasks/core/fix-service-routing/cleanup.sh
@@ -1,2 +1,3 @@
 #!/usr/bin/env bash
+kubectl delete pod -n web test-connection --ignore-not-found
 kubectl delete namespace web 
diff --git a/evals/tasks/core/fix-service-routing/setup.sh b/evals/tasks/core/fix-service-routing/setup.sh
index bfc35afa0..c4a370ec7 100755
--- a/evals/tasks/core/fix-service-routing/setup.sh
+++ b/evals/tasks/core/fix-service-routing/setup.sh
@@ -17,7 +17,8 @@ metadata:
 spec:
   ports:
   - port: 80
-    targetPort: 80
+    # nginx-unprivileged listens on 8080; only the selector is intentionally wrong
+    targetPort: 8080
   selector:
     app: web  # Mismatched label - deployment has app=nginx
 EOF
diff --git a/evals/tasks/core/fix-service-routing/verify.sh b/evals/tasks/core/fix-service-routing/verify.sh
index 65805c596..19266cbd9 100755
--- a/evals/tasks/core/fix-service-routing/verify.sh
+++ b/evals/tasks/core/fix-service-routing/verify.sh
@@ -1,13 +1,41 @@
 #!/usr/bin/env bash
-# Check if service has endpoints
-endpoints=$(kubectl get endpoints nginx -n web -o jsonpath='{.subsets[0].addresses}')
-if [[ ! -z "$endpoints" ]]; then
-    # Verify service can access the pod
-    if kubectl run -n web test-connection --image=quay.io/prometheus/busybox --restart=Never --rm -i --wait --timeout=180s \
-        -- wget -qO- nginx; then
-        exit 0
-    fi
+set -euo pipefail
+
+# Check if service has endpoints. EndpointSlice/Endpoints propagation is
+# asynchronous, so poll for a bounded period instead of failing on the first
+# empty read -- otherwise a correct fix can still be reported as failed if it
+# hasn't converged yet.
+endpoints=""
+for i in $(seq 1 15); do
+  endpoints=$(kubectl get endpoints nginx -n web -o jsonpath='{.subsets[0].addresses}' 2>/dev/null || true)
+  if [[ -n "$endpoints" ]]; then
+    break
+  fi
+  sleep 2
+done
+if [[ -z "$endpoints" ]]; then
+  echo "Service nginx in namespace web has no endpoints"
+  exit 1
 fi
 
-# If we get here, service connection failed
-exit 1 
+# Verify service can access the pod with a compatible probe pod.
+cat <<'EOF' | kubectl apply -f -
+apiVersion: v1
+kind: Pod
+metadata:
+  name: test-connection
+  namespace: web
+spec:
+  restartPolicy: Never
+  containers:
+  - name: test-connection
+    image: quay.io/curl/curl:8.11.1
+    command: ["curl", "-sf", "--max-time", "15", "http://nginx"]
+EOF
+
+if ! kubectl wait -n web --for=jsonpath='{.status.phase}'=Succeeded pod/test-connection --timeout=180s; then
+  echo "Service connection probe did not succeed"
+  kubectl get pod -n web test-connection -o yaml || true
+  exit 1
+fi
+exit 0
diff --git a/evals/tasks/core/fix-service-with-no-endpoints/verify.sh b/evals/tasks/core/fix-service-with-no-endpoints/verify.sh
index d6d6d3872..1e74112ef 100755
--- a/evals/tasks/core/fix-service-with-no-endpoints/verify.sh
+++ b/evals/tasks/core/fix-service-with-no-endpoints/verify.sh
@@ -7,9 +7,10 @@ if ! kubectl get deployment web-app-deployment -n webshop-frontend &>/dev/null;
   exit 1
 fi
 
-# Check if pods are being created successfully
+# Check if pods are being created successfully.
+# Environments can override via VERIFY_TIMEOUT
 echo "Waiting for pods to become ready..."
-TIMEOUT="120s"
+TIMEOUT="${VERIFY_TIMEOUT:-120s}"
 if ! kubectl wait --for=condition=Ready pods -l app=web-app -n webshop-frontend --timeout=$TIMEOUT; then
   echo "Pods are not reaching Ready state after fixing the node selector"
   exit 1
diff --git a/evals/tasks/core/list-images-for-pods/list-images-for-pods.yaml b/evals/tasks/core/list-images-for-pods/list-images-for-pods.yaml
index f32242a69..aee37684e 100644
--- a/evals/tasks/core/list-images-for-pods/list-images-for-pods.yaml
+++ b/evals/tasks/core/list-images-for-pods/list-images-for-pods.yaml
@@ -16,4 +16,4 @@ steps:
   cleanup:
     file: cleanup.sh
   prompt:
-    inline: "What images are all pods running in the cluster?"
+    inline: "What images are pods running in the cluster? Query the cluster now and report the images you find directly in your response -- do not ask for clarification or preferences about output format."
diff --git a/evals/tasks/core/multi-container-pod-communication/multi-container-pod-communication.yaml b/evals/tasks/core/multi-container-pod-communication/multi-container-pod-communication.yaml
index 0a9446516..7fd8d8e43 100644
--- a/evals/tasks/core/multi-container-pod-communication/multi-container-pod-communication.yaml
+++ b/evals/tasks/core/multi-container-pod-communication/multi-container-pod-communication.yaml
@@ -19,7 +19,7 @@ steps:
     inline: |
       In the multi-container-logging namespace, run a pod called communication-pod with two containers:
       1. A 'web-server' nginx instance (quay.io/nginx/nginx-unprivileged image) that serves traffic
-      2. A 'logger' busybox instance (quay.io/prometheus/busybox image) that processes those logs from a shared volume, 'logs-volume' with 'tail -f /var/log/nginx/access.log'
+      2. A 'logger' busybox instance (busybox:latest image) that processes those logs from a shared volume, 'logs-volume' with 'tail -f /var/log/nginx/access.log'
 
       Both containers should be defined in spec.containers (not as init containers or sidecars).
       Both containers should mount logs-volume at '/var/log/nginx'.
diff --git a/evals/tasks/core/multi-container-pod-communication/verify.sh b/evals/tasks/core/multi-container-pod-communication/verify.sh
index 70e3bd445..33671237e 100755
--- a/evals/tasks/core/multi-container-pod-communication/verify.sh
+++ b/evals/tasks/core/multi-container-pod-communication/verify.sh
@@ -3,7 +3,9 @@ set -euo pipefail
 
 NAMESPACE="multi-container-logging"
 POD_NAME="communication-pod"
-TIMEOUT="120s"
+# Default matches generic K8s; environments with other needs can
+# override via VERIFY_TIMEOUT without changing the default for everyone else.
+TIMEOUT="${VERIFY_TIMEOUT:-120s}"
 
 # Wait for pod to be running
 echo "Waiting for pod '$POD_NAME' to be ready..."
diff --git a/evals/tasks/core/setup-dev-cluster/verify.sh b/evals/tasks/core/setup-dev-cluster/verify.sh
index 0cc98583c..c36973305 100755
--- a/evals/tasks/core/setup-dev-cluster/verify.sh
+++ b/evals/tasks/core/setup-dev-cluster/verify.sh
@@ -6,6 +6,9 @@ readonly DEVELOPERS=("alice" "bob" "charlie")
 readonly DEV_NAMESPACES=("dev-alice" "dev-bob" "dev-charlie")
 readonly ALL_NAMESPACES=("${DEV_NAMESPACES[@]}" "dev-shared" "staging" "prod")
 readonly TEST_LABEL="app=verification-test"
+# Default matches generic K8s; environments with other needs can
+# override via VERIFY_TIMEOUT without changing the default for everyone else.
+readonly TEST_POD_TIMEOUT="${VERIFY_TIMEOUT:-60s}"
 
 # --- Cleanup Function ---
 cleanup() {
@@ -192,7 +195,7 @@ metadata:
 spec:
   containers:
   - name: curl
-    image: quay.io/curl/curl:latest
+    image: quay.io/curl/curl:8.11.1
     command: ["sleep", "3600"]
     resources:
         limits:
@@ -223,7 +226,7 @@ EOF
 
   echo "  - Waiting for test pods to be ready..."
   for dev in "${DEVELOPERS[@]}"; do
-    kubectl wait --for=condition=Ready pod/test-pod-${dev} -n "dev-${dev}" --timeout=60s
+    kubectl wait --for=condition=Ready pod/test-pod-${dev} -n "dev-${dev}" --timeout="${TEST_POD_TIMEOUT}"
   done
   
   # Test that alice cannot reach bob's service
diff --git a/evals/tasks/core/statefulset-lifecycle/verify.sh b/evals/tasks/core/statefulset-lifecycle/verify.sh
index d6a4b1b8a..45f83d6eb 100755
--- a/evals/tasks/core/statefulset-lifecycle/verify.sh
+++ b/evals/tasks/core/statefulset-lifecycle/verify.sh
@@ -5,10 +5,28 @@ set -euo pipefail
 NAMESPACE="statefulset-test"
 STS_NAME="db"
 EXPECTED_CONTENT="initial_data"
+# Defaults match generic K8s; environments with other needs can
+# override via VERIFY_TIMEOUT without changing the default
+# for everyone else.
+DELETE_TIMEOUT="${VERIFY_TIMEOUT:-120s}"
+READY_TIMEOUT="${VERIFY_TIMEOUT:-120s}"
 
 echo "Verifying old pods are deleted"
-# Wait for scale-down: 1 ready pods and deletion of old pods
-kubectl wait pod/db-1 pod/db-2 -n statefulset-test --for=delete --timeout=120s
+# Wait for scale-down: deletion of db-1/db-2 (may already be gone).
+# Fail closed: --ignore-not-found only suppresses the "not found" case, so any
+# other API/auth/transport error still returns non-zero and is treated as a
+# verification failure instead of being silently read as "pod is gone".
+for pod in db-1 db-2; do
+  kubectl wait "pod/${pod}" -n "${NAMESPACE}" --for=delete --timeout="${DELETE_TIMEOUT}" 2>/dev/null || true
+  if ! out=$(kubectl get pod "$pod" -n "${NAMESPACE}" --ignore-not-found -o name 2>&1); then
+    echo "Unable to verify pod $pod was deleted: $out"
+    exit 1
+  fi
+  if [[ -n "$out" ]]; then
+    echo "Pod $pod still exists after scale-down"
+    exit 1
+  fi
+done
 echo "Old pods are deleted"
 
 # Verify correct number of replicas
@@ -20,7 +38,16 @@ if [[ "${replicas}" -ne 1 ]]; then
 fi
 echo "StatefulSet is running with 1 replicas"
 
-# Verify db-0 exists and have the correct data
+# Wait for db-0 to become Ready before reading data; uses READY_TIMEOUT
+# (derived from VERIFY_TIMEOUT) so environments can adjust as needed.
+echo "Waiting for pod db-0 to become Ready"
+if ! kubectl wait --for=condition=Ready "pod/db-0" -n "${NAMESPACE}" --timeout="${READY_TIMEOUT}"; then
+  echo "Pod db-0 not Ready in time"
+  kubectl get pvc,pod -n "${NAMESPACE}" -o wide || true
+  exit 1
+fi
+
+# Verify db-0 has the correct data
 for pod in db-0; do
   if ! kubectl get pod "$pod" -n "${NAMESPACE}" &> /dev/null; then
     echo "Pod $pod not found in namespace $NAMESPACE"
@@ -34,4 +61,4 @@ for pod in db-0; do
   fi
 done
 
-exit 0
\ No newline at end of file
+exit 0