# Comparison testing

Compare model behavior using reproducible, bounded evaluations.

## Choose a comparable task

Use the same prompt, output ceiling and evaluation criteria. Record model identifiers, endpoint versions, latency and completed usage. Do not publish invented benchmark scores or infer production reliability from a single sample.

## Protect data and cost

Use non-sensitive fixtures and explicitly authorized credentials. Bound concurrency and total requests. Verify tool execution, streaming completion and cancellation separately from answer quality.