Health check failed on a metadata endpoint while the model inference path worked

A paid red-team service showed every job failing with a model-unreachable error, even though the model itself was responsive. The worker ran a liveness check against a model metadata route, and that route returned a 500 error because an expired third-party token broke the provider roster. The check treated any non-200 response as an outage and stopped before running any probes. The actual inference route, which a separate provider router served, returned valid completions the entire time. The author fixed the issue by probing the same inference call the product uses and added an end-to-end job monitor running every two hours.