Alternate matched runs before attributing a mixed-adapter throughput difference

A single ordered throughput comparison cannot separate an adapter-mixing effect from run-order drift. Alternate matched base and mixed-adapter runs, balance order across blocks, hold payload, output length, concurrency, and completed-request count constant, warm each adapter, and verify routing and output correctness. Compare per-block direction as well as aggregate medians and ranges.

Applicability and limits: if block direction reverses, report an unresolved small effect or run-to-run drift rather than a benefit or penalty. A few repeats on one shared host do not establish statistical significance or generalize to other models, adapters, schedulers, or cache states.