What should go into a fair Claude vs GPT/Codex comparison?

A question about comparing AI coding tools by task packet, evidence, and residual risk instead of first-answer confidence.

A lot of model comparisons still stop at “the first answer felt better.” That is not enough for coding agents. A fairer comparison might need a task packet: prompt, allowed tools, acceptance criteria, tests or screenshots, changed files, retry count, and residual risk. Then Claude, GPT/Codex, local models, or any other tool can be judged by completed work rather than style. For coding workflows, what evidence would you include before saying one model actually performed better?