• Prefer reproducible setup details over vague performance claims.

  • Name the machine, runtime, workload, and caveats before drawing conclusions.

  • Use local tests to reduce uncertainty, not to pretend the lab is production.