Route consequential decisions and output through a check that doesn't share your context or reasoning trail, instead of trusting your own re-read. Use before shipping a security-sensitive change, an irreversible operation, a high-stakes architectural decision, or any output where your own review keeps saying "looks good" without having actually stress-tested it. Not for routine work where a self-check is proportionate, and not a substitute for prove-it's demonstrated evidence — this is about who checks, not whether you ran the thing.
