Use a small test set#
Try a straightforward instruction, a self-correction, a negation, a technical name and an uncertain statement. Use invented content. For each result, mark omissions, additions and changes to scope before you rate readability.
A result that removes “um” but invents a deadline is worse than a rough draft that retains the facts. A result that changes “suggest” to “apply” has changed authorization. Give those changes more weight than punctuation.
Keep the conditions recorded#
Note software version, mode, service configuration and whether the service was actually available. A health response is not proof that the entire authenticated refinement route succeeded. The reviewed snapshot still leaves complete real-service acceptance pending.
Manual rewriting remains a current alternative. Once you have a result you trust, confirm the conversation and send deliberately. This evaluation method is practical guidance, not a built-in benchmark or a claim that all refinement modes are ready.