Why a health check is insufficient
A healthy HTTP wrapper can coexist with a missing model, failed inference, or credentials that workers cannot use. The gate therefore verifies model identity and an authenticated inference response. Fallback success must be excluded because it can hide a broken local dependency while increasing cost.
Routing isolation
Cloudflare Access protects deployment SSH only. Google sign-in protects owner diagnostics in the portal application. Protecting the whole zone or reusing an interactive policy on the AI route would break unattended workers. Tunnel configuration changes preserve every unrelated ingress rule and the existing AI mapping.
Restart and rollout constraints
The model may need several gigabytes of memory and time to load. Unrelated releases must not unload it. A real AI update should use a tested parallel instance where possible and drain active requests. If parallel operation cannot fit within host capacity, zero downtime must not be claimed.
Privacy
Synthetic probe content is harmless, but production article bodies, prompts, outputs, and credential values are never telemetry labels or public diagnostic artifacts.
Runtime evidence
The dedicated probe establishes that Cloudflare Workers can authenticate and complete real Qwen inference through the existing tunnel. It complements production worker configuration and health evidence; it does not claim to run a full production article-processing job or validate unrelated worker stages.