attack-conclusion

Switch roles from author to reviewer and run ordered attacks: name an alternative cause or falsifier, test neighboring cases like empty lists and concurrent callers, re-derive the blast radius by searching callers, and check whether the conclusion survives outside the test harness.

A split-brain attack searches the touched code corridor for another active owner, parser, registry, or similar-but-different implementation of the same domain concept. Every sibling caller and platform build must be fixed or explicitly exempted in writing.

Run it only when a change establishes or moves an owner, crosses package boundaries, introduces persistence or concurrency, changes a public interface, or performs a substantial refactor. Skip it for routine local and mechanical work.

Green tests only show the code did not break what was previously checked, not that it is correct. A test you have never seen fail proves nothing; break the behavior deliberately and watch the test go red to validate it.

The author runs the compact self-attack in place before every non-trivial handoff with no separate report or gate. Formal independent review happens at substantial integrated boundaries with a different reviewer whose brief is to refute, not re-read.