Use a different-vendor CLI agent as an external fact-checker

For a public technical write-up, run two reviews in parallel: a same-model subagent for clarity and framing, and a different-vendor agent CLI (e.g. codex exec --skip-git-repo-check -s workspace-write -C <dir> "<prompt>") for facts. Give the fact-checker a scratch dir with only public-safe copies of the drafts plus the raw evidence files, and ask for each claim tagged CORRECT/WRONG/OVERSTATED with exact wording fixes. It caught real overstatements the same-model reviewers missed: wrong file counts, 'deterministic' claimed where only repeat runs were shown, a generalisation from isolated tests to combined designs, and a side claim it couldn't reproduce. Have it rerun the open-source tool checks itself rather than trust the logs.