diff --git a/test/e2e/live/device-auth-health.test.ts b/test/e2e/live/device-auth-health.test.ts
index ea5d4ece380..2b0871f31fe 100644
--- a/test/e2e/live/device-auth-health.test.ts
+++ b/test/e2e/live/device-auth-health.test.ts
@@ -9,7 +9,7 @@
  */
 
 import { buildAvailabilityProbeEnv } from "../fixtures/availability-env.ts";
-import { resultText } from "../fixtures/clients/index.ts";
+import { resultText, shellQuote } from "../fixtures/clients/index.ts";
 import { trustedSandboxShellScript } from "../fixtures/clients/sandbox.ts";
 import { expect, test } from "../fixtures/e2e-test.ts";
 import { startFakeOpenAiCompatibleServer } from "../fixtures/fake-openai-compatible.ts";
@@ -74,7 +74,7 @@ test("device auth health probes treat 401 as live instead of offline (#2342)", {
     dashboardPort: DASHBOARD_PORT,
     contracts: [
       "onboard succeeds with device auth enabled",
-      "onboard authenticates to the fixture inference endpoint",
+      "the onboarded inference route authenticates to the fixture endpoint",
       "/health is reachable from inside the sandbox",
       "the authenticated dashboard root may return 401 without being treated as offline",
       "nemoclaw status reports the gateway as live, not Health Offline",
@@ -107,9 +107,51 @@ test("device auth health probes treat 401 as live instead of offline (#2342)", {
   progress.phase("onboard device-auth OpenClaw sandbox");
   const install = await installDeviceAuthSandbox(host, inferenceConfig, installLog);
   expect(install.exitCode, resultText(install)).toBe(0);
-  expect(inference.requests()).toContainEqual(
+
+  // Ignore incidental onboarding traffic. The fixture appends its ledger row
+  // before responding, so this awaited POST is the publication barrier for the
+  // requests sliced from this offset.
+  const authenticatedRequestOffset = inference.requests().length;
+  const authenticatedProbe = await sandbox.execShell(
+    SANDBOX_NAME,
+    trustedSandboxShellScript(
+      `curl -fsS --max-time 60 https://inference.local/v1/chat/completions -H 'Content-Type: application/json' --data ${shellQuote(
+        JSON.stringify({
+          model: INFERENCE_MODEL,
+          messages: [{ role: "user", content: "reply with OK" }],
+          max_tokens: 8,
+        }),
+      )} >/dev/null`,
+    ),
+    {
+      artifactName: "phase-1-explicit-authenticated-inference-post",
+      env: commandEnv(),
+      timeoutMs: 90_000,
+    },
+  );
+  const authenticatedRequests = inference.requests().slice(authenticatedRequestOffset);
+  const authenticatedArtifact = "phase-1-explicit-authenticated-inference-requests.json";
+  const authenticatedPhase = "onboard device-auth OpenClaw sandbox";
+  const authenticatedRequestEvidence = authenticatedRequests
+    .slice(0, 20)
+    .map(({ auth, method, model, path }) => ({ auth, method, model, path }));
+  await artifacts.writeJson(authenticatedArtifact, {
+    phase: authenticatedPhase,
+    requestCount: authenticatedRequests.length,
+    requests: authenticatedRequestEvidence,
+    truncated: authenticatedRequests.length > authenticatedRequestEvidence.length,
+  });
+  expect(
+    authenticatedProbe.exitCode,
+    `${authenticatedPhase}: explicit authenticated inference failed; see ${authenticatedArtifact}`,
+  ).toBe(0);
+  expect(
+    authenticatedRequests,
+    `${authenticatedPhase}: explicit verification probe did not reach the authenticated fixture; see ${authenticatedArtifact}`,
+  ).toContainEqual(
     expect.objectContaining({
       auth: "ok",
+      method: "POST",
       model: INFERENCE_MODEL,
       path: "/v1/chat/completions",
     }),
diff --git a/test/e2e/live/openclaw-inference-switch.test.ts b/test/e2e/live/openclaw-inference-switch.test.ts
index 66a604468fb..01510128cf1 100644
--- a/test/e2e/live/openclaw-inference-switch.test.ts
+++ b/test/e2e/live/openclaw-inference-switch.test.ts
@@ -136,17 +136,66 @@ interface MockAnthropicProvider {
   close(): Promise<void>;
 }
 
-function expectMockBaselineAuthentication(
+function proveMockBaselineAuthentication(
   baseline: Pick<FakeOpenAiCompatibleServer, "requests"> | undefined,
-): void {
+  sandbox: SandboxClient,
+  home: string,
+  artifacts: { writeJson(path: string, value: unknown): Promise<string> },
+): Promise<void> {
+  return baseline
+    ? proveSelectedMockBaselineAuthentication(baseline, sandbox, home, artifacts)
+    : Promise.resolve(expect(baseline).toBeUndefined());
+}
+
+async function proveSelectedMockBaselineAuthentication(
+  baseline: Pick<FakeOpenAiCompatibleServer, "requests">,
+  sandbox: SandboxClient,
+  home: string,
+  artifacts: { writeJson(path: string, value: unknown): Promise<string> },
+): Promise<void> {
+  // Ignore incidental onboarding traffic. The fixture appends its ledger row
+  // before responding, so this awaited POST is the publication barrier for the
+  // requests sliced from this offset.
+  const requestOffset = baseline.requests().length;
+  const payload = {
+    model: MOCK_BASELINE_MODEL,
+    messages: [{ role: "user", content: "reply with OK" }],
+    max_tokens: 8,
+  };
+  const probe = await sandboxShell(
+    sandbox,
+    home,
+    `curl -fsS --max-time 60 https://inference.local/v1/chat/completions -H 'Content-Type: application/json' --data ${shellQuote(JSON.stringify(payload))} >/dev/null`,
+    {
+      artifactName: "baseline-explicit-authenticated-inference-post",
+      timeoutMs: 90_000,
+    },
+  );
+  const requests = baseline.requests().slice(requestOffset);
+  const artifactName = "baseline-explicit-authenticated-inference-requests.json";
+  const phase = "install and onboard baseline OpenClaw";
+  const requestEvidence = requests
+    .slice(0, 20)
+    .map(({ auth, method, model, path }) => ({ auth, method, model, path }));
+  await artifacts.writeJson(artifactName, {
+    phase,
+    requestCount: requests.length,
+    requests: requestEvidence,
+    truncated: requests.length > requestEvidence.length,
+  });
+  expect(probe.exitCode, `${phase}: explicit baseline inference failed; see ${artifactName}`).toBe(
+    0,
+  );
   const expectedRequest = expect.objectContaining({
     auth: "ok",
+    method: "POST",
     model: MOCK_BASELINE_MODEL,
     path: "/v1/chat/completions",
   });
-  baseline
-    ? expect(baseline.requests()).toContainEqual(expectedRequest)
-    : expect(baseline).toBeUndefined();
+  expect(
+    requests,
+    `${phase}: explicit verification probe did not reach the authenticated fixture; see ${artifactName}`,
+  ).toContainEqual(expectedRequest);
 }
 
 function stripAnsi(value: string): string {
@@ -895,6 +944,7 @@ test("openclaw-inference-switch: switches route and preserves live OpenClaw beha
     contracts: [
       "Docker is running and an authenticated compatible baseline endpoint is staged",
       "install.sh --non-interactive onboards an OpenClaw sandbox",
+      "when selected, the mock baseline route completes one explicit authenticated fixture request",
       "nemoclaw inference set switches the running sandbox route",
       "OpenClaw gateway is supervisor-restarted only when the inference API family changes",
       "OpenShell route points at the switched provider/model",
@@ -1008,7 +1058,7 @@ test("openclaw-inference-switch: switches route and preserves live OpenClaw beha
     skip("NVIDIA endpoint validation was unavailable/rate-limited during onboarding");
   }
   expect(install.exitCode, installText).toBe(0);
-  expectMockBaselineAuthentication(baselineProvider);
+  await proveMockBaselineAuthentication(baselineProvider, sandbox, home, artifacts);
 
   progress.phase("prepare the switched provider and endpoint");
   const publicProvider = publicApiKey