Operator note / October 10, 2026

A model label is not a deployment strategy

The fictional lab assigns a different model family to each node. The useful part is the separation of responsibilities.

Interactive requests need a bounded queue

In the cedar-01 example, short requests receive a predictable queue budget. A timeout should stop work that no longer has a waiting caller, rather than silently creating an expensive background task.

Batch work needs resumable boundaries

Basalt-02 represents a different trade-off. A batch job can tolerate a slower answer, but should preserve progress after a worker restart. Small checkpointed chunks are easier to retry than a single oversized request.

Evaluation is its own workload

Cirrus-03 illustrates why testing should not share the same queue as user-facing traffic. Pin the prompt set, record the model configuration and compare outputs under the same conditions. A provider name alone says little about latency, quality or cost.

These are design notes for a fictional lab, not benchmark claims about the named providers.