rr-selfchecklisted

# Attack your own result before you ship it Use this when you have a result — an analysis, a fix, a number, a claim — and you are about to hand it over. **What this produces:** demonstrated failures, and a list of the surfaces you did not attack. **What it never produces:** independent agreement. You are checking your own work, and a clean pass from the author of something is not evidence that the thing is right. ## The rule that makes it work **A refutation of your own result is a result, and it counts as one.** The failure mode of self-review is not blindness — you can see the flaws. It is **stopping**: the search ends when the thing works, because a working answer was the objective. So restate the objective before you start: **you are trying to break it, and breaking it is a success.** ## Before you look for flaws 1. **Write the attack list first**, before you know whether the result survives. Attacks chosen afterwards are the ones it already passes. 2. **Commit to a number.** At least five attacks, spread across the categories below. A quota takes the stopping decision away from the part of you that wants to stop. ## The categories - **The data** — is the input what you think it is? Units, a stale file, a silent truncation, the empty case. - **The arithmetic** — recompute by a second route. A number agreeing with itself is not a check. - **The method** — does your check fire on a broken input? **If you have never seen it fail, you do not know that it wor