adversarial-review

以假定出错视角对代码或产品进行对抗性审查，找出潜在缺陷。

以“它肯定有问题”的思路对刚实现的功能或整个代码库做对抗性审查，目标是主动找出失败场景而不是确认通过。适合实现完成后、发版前或需要敌对视角复核时使用，会派出不同视角的审查员（正确性、安全、极端环境）并尽量运行并复现问题再确认。与静态 code review 不同，会实际运行目标来验证发现。

# Adversarial review — presume it broken, find where

A normal review reads code and looks for reasons to approve it. An adversarial review starts from the
opposite prior: **this was just built, it is presumed broken, and the job is to find where.** The
builder believes in their work — that belief is exactly the blind spot. So the reviewer must be
incentivized to find failure, not confirm success, and must work *without the builder's context* so it
cannot inherit the builder's assumptions.

This skill is the canonical adversarial-review protocol for ATLAS. Reach for it autonomously after
building anything non-trivial, before shipping, or when handed unfamiliar code to harden.

## When to use